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 non-Windows x86 64-bit vararg functions.
2234 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64())
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 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4156 SelectionDAG &DAG) const {
4157 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
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 (Subtarget->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 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE)
4231 NewV = LowerVECTOR_SHUFFLE(NewV, DAG);
4232 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4234 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4235 // source words for the shuffle, to aid later transformations.
4236 bool AllWordsInNewV = true;
4237 bool InOrder[2] = { true, true };
4238 for (unsigned i = 0; i != 8; ++i) {
4239 int idx = MaskVals[i];
4241 InOrder[i/4] = false;
4242 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4244 AllWordsInNewV = false;
4248 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4249 if (AllWordsInNewV) {
4250 for (int i = 0; i != 8; ++i) {
4251 int idx = MaskVals[i];
4254 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4255 if ((idx != i) && idx < 4)
4257 if ((idx != i) && idx > 3)
4266 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4267 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4268 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4269 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4270 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4274 // If we have SSSE3, and all words of the result are from 1 input vector,
4275 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4276 // is present, fall back to case 4.
4277 if (Subtarget->hasSSSE3()) {
4278 SmallVector<SDValue,16> pshufbMask;
4280 // If we have elements from both input vectors, set the high bit of the
4281 // shuffle mask element to zero out elements that come from V2 in the V1
4282 // mask, and elements that come from V1 in the V2 mask, so that the two
4283 // results can be OR'd together.
4284 bool TwoInputs = V1Used && V2Used;
4285 for (unsigned i = 0; i != 8; ++i) {
4286 int EltIdx = MaskVals[i] * 2;
4287 if (TwoInputs && (EltIdx >= 16)) {
4288 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4289 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4292 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4293 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4295 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4296 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4297 DAG.getNode(ISD::BUILD_VECTOR, dl,
4298 MVT::v16i8, &pshufbMask[0], 16));
4300 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4302 // Calculate the shuffle mask for the second input, shuffle it, and
4303 // OR it with the first shuffled input.
4305 for (unsigned i = 0; i != 8; ++i) {
4306 int EltIdx = MaskVals[i] * 2;
4308 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4309 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4312 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4313 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4315 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4316 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4317 DAG.getNode(ISD::BUILD_VECTOR, dl,
4318 MVT::v16i8, &pshufbMask[0], 16));
4319 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4320 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4323 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4324 // and update MaskVals with new element order.
4325 BitVector InOrder(8);
4326 if (BestLoQuad >= 0) {
4327 SmallVector<int, 8> MaskV;
4328 for (int i = 0; i != 4; ++i) {
4329 int idx = MaskVals[i];
4331 MaskV.push_back(-1);
4333 } else if ((idx / 4) == BestLoQuad) {
4334 MaskV.push_back(idx & 3);
4337 MaskV.push_back(-1);
4340 for (unsigned i = 4; i != 8; ++i)
4342 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4346 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4347 // and update MaskVals with the new element order.
4348 if (BestHiQuad >= 0) {
4349 SmallVector<int, 8> MaskV;
4350 for (unsigned i = 0; i != 4; ++i)
4352 for (unsigned i = 4; i != 8; ++i) {
4353 int idx = MaskVals[i];
4355 MaskV.push_back(-1);
4357 } else if ((idx / 4) == BestHiQuad) {
4358 MaskV.push_back((idx & 3) + 4);
4361 MaskV.push_back(-1);
4364 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4368 // In case BestHi & BestLo were both -1, which means each quadword has a word
4369 // from each of the four input quadwords, calculate the InOrder bitvector now
4370 // before falling through to the insert/extract cleanup.
4371 if (BestLoQuad == -1 && BestHiQuad == -1) {
4373 for (int i = 0; i != 8; ++i)
4374 if (MaskVals[i] < 0 || MaskVals[i] == i)
4378 // The other elements are put in the right place using pextrw and pinsrw.
4379 for (unsigned i = 0; i != 8; ++i) {
4382 int EltIdx = MaskVals[i];
4385 SDValue ExtOp = (EltIdx < 8)
4386 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4387 DAG.getIntPtrConstant(EltIdx))
4388 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4389 DAG.getIntPtrConstant(EltIdx - 8));
4390 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4391 DAG.getIntPtrConstant(i));
4396 // v16i8 shuffles - Prefer shuffles in the following order:
4397 // 1. [ssse3] 1 x pshufb
4398 // 2. [ssse3] 2 x pshufb + 1 x por
4399 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4401 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4403 const X86TargetLowering &TLI) {
4404 SDValue V1 = SVOp->getOperand(0);
4405 SDValue V2 = SVOp->getOperand(1);
4406 DebugLoc dl = SVOp->getDebugLoc();
4407 SmallVector<int, 16> MaskVals;
4408 SVOp->getMask(MaskVals);
4410 // If we have SSSE3, case 1 is generated when all result bytes come from
4411 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4412 // present, fall back to case 3.
4413 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4416 for (unsigned i = 0; i < 16; ++i) {
4417 int EltIdx = MaskVals[i];
4426 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4427 if (TLI.getSubtarget()->hasSSSE3()) {
4428 SmallVector<SDValue,16> pshufbMask;
4430 // If all result elements are from one input vector, then only translate
4431 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4433 // Otherwise, we have elements from both input vectors, and must zero out
4434 // elements that come from V2 in the first mask, and V1 in the second mask
4435 // so that we can OR them together.
4436 bool TwoInputs = !(V1Only || V2Only);
4437 for (unsigned i = 0; i != 16; ++i) {
4438 int EltIdx = MaskVals[i];
4439 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4440 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4443 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4445 // If all the elements are from V2, assign it to V1 and return after
4446 // building the first pshufb.
4449 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4450 DAG.getNode(ISD::BUILD_VECTOR, dl,
4451 MVT::v16i8, &pshufbMask[0], 16));
4455 // Calculate the shuffle mask for the second input, shuffle it, and
4456 // OR it with the first shuffled input.
4458 for (unsigned i = 0; i != 16; ++i) {
4459 int EltIdx = MaskVals[i];
4461 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4464 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4466 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4467 DAG.getNode(ISD::BUILD_VECTOR, dl,
4468 MVT::v16i8, &pshufbMask[0], 16));
4469 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4472 // No SSSE3 - Calculate in place words and then fix all out of place words
4473 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4474 // the 16 different words that comprise the two doublequadword input vectors.
4475 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4476 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4477 SDValue NewV = V2Only ? V2 : V1;
4478 for (int i = 0; i != 8; ++i) {
4479 int Elt0 = MaskVals[i*2];
4480 int Elt1 = MaskVals[i*2+1];
4482 // This word of the result is all undef, skip it.
4483 if (Elt0 < 0 && Elt1 < 0)
4486 // This word of the result is already in the correct place, skip it.
4487 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4489 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4492 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4493 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4496 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4497 // using a single extract together, load it and store it.
4498 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4499 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4500 DAG.getIntPtrConstant(Elt1 / 2));
4501 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4502 DAG.getIntPtrConstant(i));
4506 // If Elt1 is defined, extract it from the appropriate source. If the
4507 // source byte is not also odd, shift the extracted word left 8 bits
4508 // otherwise clear the bottom 8 bits if we need to do an or.
4510 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4511 DAG.getIntPtrConstant(Elt1 / 2));
4512 if ((Elt1 & 1) == 0)
4513 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4514 DAG.getConstant(8, TLI.getShiftAmountTy()));
4516 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4517 DAG.getConstant(0xFF00, MVT::i16));
4519 // If Elt0 is defined, extract it from the appropriate source. If the
4520 // source byte is not also even, shift the extracted word right 8 bits. If
4521 // Elt1 was also defined, OR the extracted values together before
4522 // inserting them in the result.
4524 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4525 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4526 if ((Elt0 & 1) != 0)
4527 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4528 DAG.getConstant(8, TLI.getShiftAmountTy()));
4530 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4531 DAG.getConstant(0x00FF, MVT::i16));
4532 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4535 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4536 DAG.getIntPtrConstant(i));
4538 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4541 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4542 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4543 /// done when every pair / quad of shuffle mask elements point to elements in
4544 /// the right sequence. e.g.
4545 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4547 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4549 const TargetLowering &TLI, DebugLoc dl) {
4550 EVT VT = SVOp->getValueType(0);
4551 SDValue V1 = SVOp->getOperand(0);
4552 SDValue V2 = SVOp->getOperand(1);
4553 unsigned NumElems = VT.getVectorNumElements();
4554 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4555 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4557 switch (VT.getSimpleVT().SimpleTy) {
4558 default: assert(false && "Unexpected!");
4559 case MVT::v4f32: NewVT = MVT::v2f64; break;
4560 case MVT::v4i32: NewVT = MVT::v2i64; break;
4561 case MVT::v8i16: NewVT = MVT::v4i32; break;
4562 case MVT::v16i8: NewVT = MVT::v4i32; break;
4565 if (NewWidth == 2) {
4571 int Scale = NumElems / NewWidth;
4572 SmallVector<int, 8> MaskVec;
4573 for (unsigned i = 0; i < NumElems; i += Scale) {
4575 for (int j = 0; j < Scale; ++j) {
4576 int EltIdx = SVOp->getMaskElt(i+j);
4580 StartIdx = EltIdx - (EltIdx % Scale);
4581 if (EltIdx != StartIdx + j)
4585 MaskVec.push_back(-1);
4587 MaskVec.push_back(StartIdx / Scale);
4590 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4591 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4592 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4595 /// getVZextMovL - Return a zero-extending vector move low node.
4597 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4598 SDValue SrcOp, SelectionDAG &DAG,
4599 const X86Subtarget *Subtarget, DebugLoc dl) {
4600 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4601 LoadSDNode *LD = NULL;
4602 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4603 LD = dyn_cast<LoadSDNode>(SrcOp);
4605 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4607 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4608 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4609 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4610 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4611 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4613 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4614 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4615 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4616 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4624 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4625 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4626 DAG.getNode(ISD::BIT_CONVERT, dl,
4630 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4633 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4634 SDValue V1 = SVOp->getOperand(0);
4635 SDValue V2 = SVOp->getOperand(1);
4636 DebugLoc dl = SVOp->getDebugLoc();
4637 EVT VT = SVOp->getValueType(0);
4639 SmallVector<std::pair<int, int>, 8> Locs;
4641 SmallVector<int, 8> Mask1(4U, -1);
4642 SmallVector<int, 8> PermMask;
4643 SVOp->getMask(PermMask);
4647 for (unsigned i = 0; i != 4; ++i) {
4648 int Idx = PermMask[i];
4650 Locs[i] = std::make_pair(-1, -1);
4652 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4654 Locs[i] = std::make_pair(0, NumLo);
4658 Locs[i] = std::make_pair(1, NumHi);
4660 Mask1[2+NumHi] = Idx;
4666 if (NumLo <= 2 && NumHi <= 2) {
4667 // If no more than two elements come from either vector. This can be
4668 // implemented with two shuffles. First shuffle gather the elements.
4669 // The second shuffle, which takes the first shuffle as both of its
4670 // vector operands, put the elements into the right order.
4671 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4673 SmallVector<int, 8> Mask2(4U, -1);
4675 for (unsigned i = 0; i != 4; ++i) {
4676 if (Locs[i].first == -1)
4679 unsigned Idx = (i < 2) ? 0 : 4;
4680 Idx += Locs[i].first * 2 + Locs[i].second;
4685 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4686 } else if (NumLo == 3 || NumHi == 3) {
4687 // Otherwise, we must have three elements from one vector, call it X, and
4688 // one element from the other, call it Y. First, use a shufps to build an
4689 // intermediate vector with the one element from Y and the element from X
4690 // that will be in the same half in the final destination (the indexes don't
4691 // matter). Then, use a shufps to build the final vector, taking the half
4692 // containing the element from Y from the intermediate, and the other half
4695 // Normalize it so the 3 elements come from V1.
4696 CommuteVectorShuffleMask(PermMask, VT);
4700 // Find the element from V2.
4702 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4703 int Val = PermMask[HiIndex];
4710 Mask1[0] = PermMask[HiIndex];
4712 Mask1[2] = PermMask[HiIndex^1];
4714 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4717 Mask1[0] = PermMask[0];
4718 Mask1[1] = PermMask[1];
4719 Mask1[2] = HiIndex & 1 ? 6 : 4;
4720 Mask1[3] = HiIndex & 1 ? 4 : 6;
4721 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4723 Mask1[0] = HiIndex & 1 ? 2 : 0;
4724 Mask1[1] = HiIndex & 1 ? 0 : 2;
4725 Mask1[2] = PermMask[2];
4726 Mask1[3] = PermMask[3];
4731 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4735 // Break it into (shuffle shuffle_hi, shuffle_lo).
4737 SmallVector<int,8> LoMask(4U, -1);
4738 SmallVector<int,8> HiMask(4U, -1);
4740 SmallVector<int,8> *MaskPtr = &LoMask;
4741 unsigned MaskIdx = 0;
4744 for (unsigned i = 0; i != 4; ++i) {
4751 int Idx = PermMask[i];
4753 Locs[i] = std::make_pair(-1, -1);
4754 } else if (Idx < 4) {
4755 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4756 (*MaskPtr)[LoIdx] = Idx;
4759 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4760 (*MaskPtr)[HiIdx] = Idx;
4765 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4766 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4767 SmallVector<int, 8> MaskOps;
4768 for (unsigned i = 0; i != 4; ++i) {
4769 if (Locs[i].first == -1) {
4770 MaskOps.push_back(-1);
4772 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4773 MaskOps.push_back(Idx);
4776 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4780 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4781 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4782 SDValue V1 = Op.getOperand(0);
4783 SDValue V2 = Op.getOperand(1);
4784 EVT VT = Op.getValueType();
4785 DebugLoc dl = Op.getDebugLoc();
4786 unsigned NumElems = VT.getVectorNumElements();
4787 bool isMMX = VT.getSizeInBits() == 64;
4788 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4789 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4790 bool V1IsSplat = false;
4791 bool V2IsSplat = false;
4793 if (isZeroShuffle(SVOp))
4794 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4796 // Promote splats to v4f32.
4797 if (SVOp->isSplat()) {
4798 if (isMMX || NumElems < 4)
4800 return PromoteSplat(SVOp, DAG);
4803 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4805 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4806 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4807 if (NewOp.getNode())
4808 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4809 LowerVECTOR_SHUFFLE(NewOp, DAG));
4810 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4811 // FIXME: Figure out a cleaner way to do this.
4812 // Try to make use of movq to zero out the top part.
4813 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4814 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4815 if (NewOp.getNode()) {
4816 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4817 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4818 DAG, Subtarget, dl);
4820 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4821 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4822 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4823 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4824 DAG, Subtarget, dl);
4828 if (X86::isPSHUFDMask(SVOp))
4831 // Check if this can be converted into a logical shift.
4832 bool isLeft = false;
4835 bool isShift = getSubtarget()->hasSSE2() &&
4836 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4837 if (isShift && ShVal.hasOneUse()) {
4838 // If the shifted value has multiple uses, it may be cheaper to use
4839 // v_set0 + movlhps or movhlps, etc.
4840 EVT EltVT = VT.getVectorElementType();
4841 ShAmt *= EltVT.getSizeInBits();
4842 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4845 if (X86::isMOVLMask(SVOp)) {
4848 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4849 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4854 // FIXME: fold these into legal mask.
4855 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4856 X86::isMOVSLDUPMask(SVOp) ||
4857 X86::isMOVHLPSMask(SVOp) ||
4858 X86::isMOVLHPSMask(SVOp) ||
4859 X86::isMOVLPMask(SVOp)))
4862 if (ShouldXformToMOVHLPS(SVOp) ||
4863 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4864 return CommuteVectorShuffle(SVOp, DAG);
4867 // No better options. Use a vshl / vsrl.
4868 EVT EltVT = VT.getVectorElementType();
4869 ShAmt *= EltVT.getSizeInBits();
4870 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4873 bool Commuted = false;
4874 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4875 // 1,1,1,1 -> v8i16 though.
4876 V1IsSplat = isSplatVector(V1.getNode());
4877 V2IsSplat = isSplatVector(V2.getNode());
4879 // Canonicalize the splat or undef, if present, to be on the RHS.
4880 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4881 Op = CommuteVectorShuffle(SVOp, DAG);
4882 SVOp = cast<ShuffleVectorSDNode>(Op);
4883 V1 = SVOp->getOperand(0);
4884 V2 = SVOp->getOperand(1);
4885 std::swap(V1IsSplat, V2IsSplat);
4886 std::swap(V1IsUndef, V2IsUndef);
4890 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4891 // Shuffling low element of v1 into undef, just return v1.
4894 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4895 // the instruction selector will not match, so get a canonical MOVL with
4896 // swapped operands to undo the commute.
4897 return getMOVL(DAG, dl, VT, V2, V1);
4900 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4901 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4902 X86::isUNPCKLMask(SVOp) ||
4903 X86::isUNPCKHMask(SVOp))
4907 // Normalize mask so all entries that point to V2 points to its first
4908 // element then try to match unpck{h|l} again. If match, return a
4909 // new vector_shuffle with the corrected mask.
4910 SDValue NewMask = NormalizeMask(SVOp, DAG);
4911 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4912 if (NSVOp != SVOp) {
4913 if (X86::isUNPCKLMask(NSVOp, true)) {
4915 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4922 // Commute is back and try unpck* again.
4923 // FIXME: this seems wrong.
4924 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4925 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4926 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4927 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4928 X86::isUNPCKLMask(NewSVOp) ||
4929 X86::isUNPCKHMask(NewSVOp))
4933 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4935 // Normalize the node to match x86 shuffle ops if needed
4936 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4937 return CommuteVectorShuffle(SVOp, DAG);
4939 // Check for legal shuffle and return?
4940 SmallVector<int, 16> PermMask;
4941 SVOp->getMask(PermMask);
4942 if (isShuffleMaskLegal(PermMask, VT))
4945 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4946 if (VT == MVT::v8i16) {
4947 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
4948 if (NewOp.getNode())
4952 if (VT == MVT::v16i8) {
4953 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4954 if (NewOp.getNode())
4958 // Handle all 4 wide cases with a number of shuffles except for MMX.
4959 if (NumElems == 4 && !isMMX)
4960 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4966 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4967 SelectionDAG &DAG) const {
4968 EVT VT = Op.getValueType();
4969 DebugLoc dl = Op.getDebugLoc();
4970 if (VT.getSizeInBits() == 8) {
4971 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4972 Op.getOperand(0), Op.getOperand(1));
4973 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4974 DAG.getValueType(VT));
4975 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4976 } else if (VT.getSizeInBits() == 16) {
4977 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4978 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4980 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4981 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4982 DAG.getNode(ISD::BIT_CONVERT, dl,
4986 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4987 Op.getOperand(0), Op.getOperand(1));
4988 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4989 DAG.getValueType(VT));
4990 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4991 } else if (VT == MVT::f32) {
4992 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4993 // the result back to FR32 register. It's only worth matching if the
4994 // result has a single use which is a store or a bitcast to i32. And in
4995 // the case of a store, it's not worth it if the index is a constant 0,
4996 // because a MOVSSmr can be used instead, which is smaller and faster.
4997 if (!Op.hasOneUse())
4999 SDNode *User = *Op.getNode()->use_begin();
5000 if ((User->getOpcode() != ISD::STORE ||
5001 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5002 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5003 (User->getOpcode() != ISD::BIT_CONVERT ||
5004 User->getValueType(0) != MVT::i32))
5006 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5007 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
5010 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
5011 } else if (VT == MVT::i32) {
5012 // ExtractPS works with constant index.
5013 if (isa<ConstantSDNode>(Op.getOperand(1)))
5021 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5022 SelectionDAG &DAG) const {
5023 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5026 if (Subtarget->hasSSE41()) {
5027 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5032 EVT VT = Op.getValueType();
5033 DebugLoc dl = Op.getDebugLoc();
5034 // TODO: handle v16i8.
5035 if (VT.getSizeInBits() == 16) {
5036 SDValue Vec = Op.getOperand(0);
5037 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5039 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5040 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5041 DAG.getNode(ISD::BIT_CONVERT, dl,
5044 // Transform it so it match pextrw which produces a 32-bit result.
5045 EVT EltVT = MVT::i32;
5046 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5047 Op.getOperand(0), Op.getOperand(1));
5048 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5049 DAG.getValueType(VT));
5050 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5051 } else if (VT.getSizeInBits() == 32) {
5052 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5056 // SHUFPS the element to the lowest double word, then movss.
5057 int Mask[4] = { Idx, -1, -1, -1 };
5058 EVT VVT = Op.getOperand(0).getValueType();
5059 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5060 DAG.getUNDEF(VVT), Mask);
5061 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5062 DAG.getIntPtrConstant(0));
5063 } else if (VT.getSizeInBits() == 64) {
5064 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5065 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5066 // to match extract_elt for f64.
5067 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5071 // UNPCKHPD the element to the lowest double word, then movsd.
5072 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5073 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5074 int Mask[2] = { 1, -1 };
5075 EVT VVT = Op.getOperand(0).getValueType();
5076 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5077 DAG.getUNDEF(VVT), Mask);
5078 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5079 DAG.getIntPtrConstant(0));
5086 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5087 SelectionDAG &DAG) const {
5088 EVT VT = Op.getValueType();
5089 EVT EltVT = VT.getVectorElementType();
5090 DebugLoc dl = Op.getDebugLoc();
5092 SDValue N0 = Op.getOperand(0);
5093 SDValue N1 = Op.getOperand(1);
5094 SDValue N2 = Op.getOperand(2);
5096 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5097 isa<ConstantSDNode>(N2)) {
5099 if (VT == MVT::v8i16)
5100 Opc = X86ISD::PINSRW;
5101 else if (VT == MVT::v4i16)
5102 Opc = X86ISD::MMX_PINSRW;
5103 else if (VT == MVT::v16i8)
5104 Opc = X86ISD::PINSRB;
5106 Opc = X86ISD::PINSRB;
5108 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5110 if (N1.getValueType() != MVT::i32)
5111 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5112 if (N2.getValueType() != MVT::i32)
5113 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5114 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5115 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5116 // Bits [7:6] of the constant are the source select. This will always be
5117 // zero here. The DAG Combiner may combine an extract_elt index into these
5118 // bits. For example (insert (extract, 3), 2) could be matched by putting
5119 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5120 // Bits [5:4] of the constant are the destination select. This is the
5121 // value of the incoming immediate.
5122 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5123 // combine either bitwise AND or insert of float 0.0 to set these bits.
5124 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5125 // Create this as a scalar to vector..
5126 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5127 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5128 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5129 // PINSR* works with constant index.
5136 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5137 EVT VT = Op.getValueType();
5138 EVT EltVT = VT.getVectorElementType();
5140 if (Subtarget->hasSSE41())
5141 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5143 if (EltVT == MVT::i8)
5146 DebugLoc dl = Op.getDebugLoc();
5147 SDValue N0 = Op.getOperand(0);
5148 SDValue N1 = Op.getOperand(1);
5149 SDValue N2 = Op.getOperand(2);
5151 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5152 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5153 // as its second argument.
5154 if (N1.getValueType() != MVT::i32)
5155 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5156 if (N2.getValueType() != MVT::i32)
5157 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5158 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5159 dl, VT, N0, N1, N2);
5165 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5166 DebugLoc dl = Op.getDebugLoc();
5168 if (Op.getValueType() == MVT::v1i64 &&
5169 Op.getOperand(0).getValueType() == MVT::i64)
5170 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5172 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5173 EVT VT = MVT::v2i32;
5174 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5181 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5182 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5185 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5186 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5187 // one of the above mentioned nodes. It has to be wrapped because otherwise
5188 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5189 // be used to form addressing mode. These wrapped nodes will be selected
5192 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5193 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5195 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5197 unsigned char OpFlag = 0;
5198 unsigned WrapperKind = X86ISD::Wrapper;
5199 CodeModel::Model M = getTargetMachine().getCodeModel();
5201 if (Subtarget->isPICStyleRIPRel() &&
5202 (M == CodeModel::Small || M == CodeModel::Kernel))
5203 WrapperKind = X86ISD::WrapperRIP;
5204 else if (Subtarget->isPICStyleGOT())
5205 OpFlag = X86II::MO_GOTOFF;
5206 else if (Subtarget->isPICStyleStubPIC())
5207 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5209 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5211 CP->getOffset(), OpFlag);
5212 DebugLoc DL = CP->getDebugLoc();
5213 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5214 // With PIC, the address is actually $g + Offset.
5216 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5217 DAG.getNode(X86ISD::GlobalBaseReg,
5218 DebugLoc(), getPointerTy()),
5225 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5226 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5228 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5230 unsigned char OpFlag = 0;
5231 unsigned WrapperKind = X86ISD::Wrapper;
5232 CodeModel::Model M = getTargetMachine().getCodeModel();
5234 if (Subtarget->isPICStyleRIPRel() &&
5235 (M == CodeModel::Small || M == CodeModel::Kernel))
5236 WrapperKind = X86ISD::WrapperRIP;
5237 else if (Subtarget->isPICStyleGOT())
5238 OpFlag = X86II::MO_GOTOFF;
5239 else if (Subtarget->isPICStyleStubPIC())
5240 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5242 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5244 DebugLoc DL = JT->getDebugLoc();
5245 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5247 // With PIC, the address is actually $g + Offset.
5249 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5250 DAG.getNode(X86ISD::GlobalBaseReg,
5251 DebugLoc(), getPointerTy()),
5259 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5260 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5262 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5264 unsigned char OpFlag = 0;
5265 unsigned WrapperKind = X86ISD::Wrapper;
5266 CodeModel::Model M = getTargetMachine().getCodeModel();
5268 if (Subtarget->isPICStyleRIPRel() &&
5269 (M == CodeModel::Small || M == CodeModel::Kernel))
5270 WrapperKind = X86ISD::WrapperRIP;
5271 else if (Subtarget->isPICStyleGOT())
5272 OpFlag = X86II::MO_GOTOFF;
5273 else if (Subtarget->isPICStyleStubPIC())
5274 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5276 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5278 DebugLoc DL = Op.getDebugLoc();
5279 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5282 // With PIC, the address is actually $g + Offset.
5283 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5284 !Subtarget->is64Bit()) {
5285 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5286 DAG.getNode(X86ISD::GlobalBaseReg,
5287 DebugLoc(), getPointerTy()),
5295 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5296 // Create the TargetBlockAddressAddress node.
5297 unsigned char OpFlags =
5298 Subtarget->ClassifyBlockAddressReference();
5299 CodeModel::Model M = getTargetMachine().getCodeModel();
5300 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5301 DebugLoc dl = Op.getDebugLoc();
5302 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5303 /*isTarget=*/true, OpFlags);
5305 if (Subtarget->isPICStyleRIPRel() &&
5306 (M == CodeModel::Small || M == CodeModel::Kernel))
5307 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5309 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5311 // With PIC, the address is actually $g + Offset.
5312 if (isGlobalRelativeToPICBase(OpFlags)) {
5313 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5314 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5322 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5324 SelectionDAG &DAG) const {
5325 // Create the TargetGlobalAddress node, folding in the constant
5326 // offset if it is legal.
5327 unsigned char OpFlags =
5328 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5329 CodeModel::Model M = getTargetMachine().getCodeModel();
5331 if (OpFlags == X86II::MO_NO_FLAG &&
5332 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5333 // A direct static reference to a global.
5334 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5337 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5340 if (Subtarget->isPICStyleRIPRel() &&
5341 (M == CodeModel::Small || M == CodeModel::Kernel))
5342 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5344 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5346 // With PIC, the address is actually $g + Offset.
5347 if (isGlobalRelativeToPICBase(OpFlags)) {
5348 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5349 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5353 // For globals that require a load from a stub to get the address, emit the
5355 if (isGlobalStubReference(OpFlags))
5356 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5357 PseudoSourceValue::getGOT(), 0, false, false, 0);
5359 // If there was a non-zero offset that we didn't fold, create an explicit
5362 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5363 DAG.getConstant(Offset, getPointerTy()));
5369 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5370 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5371 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5372 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5376 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5377 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5378 unsigned char OperandFlags) {
5379 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5380 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5381 DebugLoc dl = GA->getDebugLoc();
5382 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5383 GA->getValueType(0),
5387 SDValue Ops[] = { Chain, TGA, *InFlag };
5388 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5390 SDValue Ops[] = { Chain, TGA };
5391 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5394 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5395 MFI->setAdjustsStack(true);
5397 SDValue Flag = Chain.getValue(1);
5398 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5401 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5403 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5406 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5407 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5408 DAG.getNode(X86ISD::GlobalBaseReg,
5409 DebugLoc(), PtrVT), InFlag);
5410 InFlag = Chain.getValue(1);
5412 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5415 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5417 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5419 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5420 X86::RAX, X86II::MO_TLSGD);
5423 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5424 // "local exec" model.
5425 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5426 const EVT PtrVT, TLSModel::Model model,
5428 DebugLoc dl = GA->getDebugLoc();
5429 // Get the Thread Pointer
5430 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5432 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5435 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5436 NULL, 0, false, false, 0);
5438 unsigned char OperandFlags = 0;
5439 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5441 unsigned WrapperKind = X86ISD::Wrapper;
5442 if (model == TLSModel::LocalExec) {
5443 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5444 } else if (is64Bit) {
5445 assert(model == TLSModel::InitialExec);
5446 OperandFlags = X86II::MO_GOTTPOFF;
5447 WrapperKind = X86ISD::WrapperRIP;
5449 assert(model == TLSModel::InitialExec);
5450 OperandFlags = X86II::MO_INDNTPOFF;
5453 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5455 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5456 GA->getValueType(0),
5457 GA->getOffset(), OperandFlags);
5458 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5460 if (model == TLSModel::InitialExec)
5461 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5462 PseudoSourceValue::getGOT(), 0, false, false, 0);
5464 // The address of the thread local variable is the add of the thread
5465 // pointer with the offset of the variable.
5466 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5470 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5472 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5473 const GlobalValue *GV = GA->getGlobal();
5475 if (Subtarget->isTargetELF()) {
5476 // TODO: implement the "local dynamic" model
5477 // TODO: implement the "initial exec"model for pic executables
5479 // If GV is an alias then use the aliasee for determining
5480 // thread-localness.
5481 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5482 GV = GA->resolveAliasedGlobal(false);
5484 TLSModel::Model model
5485 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5488 case TLSModel::GeneralDynamic:
5489 case TLSModel::LocalDynamic: // not implemented
5490 if (Subtarget->is64Bit())
5491 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5492 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5494 case TLSModel::InitialExec:
5495 case TLSModel::LocalExec:
5496 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5497 Subtarget->is64Bit());
5499 } else if (Subtarget->isTargetDarwin()) {
5500 // Darwin only has one model of TLS. Lower to that.
5501 unsigned char OpFlag = 0;
5502 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5503 X86ISD::WrapperRIP : X86ISD::Wrapper;
5505 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5507 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5508 !Subtarget->is64Bit();
5510 OpFlag = X86II::MO_TLVP_PIC_BASE;
5512 OpFlag = X86II::MO_TLVP;
5513 DebugLoc DL = Op.getDebugLoc();
5514 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
5516 GA->getOffset(), OpFlag);
5517 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5519 // With PIC32, the address is actually $g + Offset.
5521 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5522 DAG.getNode(X86ISD::GlobalBaseReg,
5523 DebugLoc(), getPointerTy()),
5526 // Lowering the machine isd will make sure everything is in the right
5528 SDValue Args[] = { Offset };
5529 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5531 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5532 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5533 MFI->setAdjustsStack(true);
5535 // And our return value (tls address) is in the standard call return value
5537 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5538 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5542 "TLS not implemented for this target.");
5544 llvm_unreachable("Unreachable");
5549 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5550 /// take a 2 x i32 value to shift plus a shift amount.
5551 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5552 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5553 EVT VT = Op.getValueType();
5554 unsigned VTBits = VT.getSizeInBits();
5555 DebugLoc dl = Op.getDebugLoc();
5556 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5557 SDValue ShOpLo = Op.getOperand(0);
5558 SDValue ShOpHi = Op.getOperand(1);
5559 SDValue ShAmt = Op.getOperand(2);
5560 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5561 DAG.getConstant(VTBits - 1, MVT::i8))
5562 : DAG.getConstant(0, VT);
5565 if (Op.getOpcode() == ISD::SHL_PARTS) {
5566 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5567 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5569 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5570 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5573 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5574 DAG.getConstant(VTBits, MVT::i8));
5575 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5576 AndNode, DAG.getConstant(0, MVT::i8));
5579 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5580 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5581 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5583 if (Op.getOpcode() == ISD::SHL_PARTS) {
5584 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5585 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5587 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5588 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5591 SDValue Ops[2] = { Lo, Hi };
5592 return DAG.getMergeValues(Ops, 2, dl);
5595 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5596 SelectionDAG &DAG) const {
5597 EVT SrcVT = Op.getOperand(0).getValueType();
5599 if (SrcVT.isVector()) {
5600 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5606 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5607 "Unknown SINT_TO_FP to lower!");
5609 // These are really Legal; return the operand so the caller accepts it as
5611 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5613 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5614 Subtarget->is64Bit()) {
5618 DebugLoc dl = Op.getDebugLoc();
5619 unsigned Size = SrcVT.getSizeInBits()/8;
5620 MachineFunction &MF = DAG.getMachineFunction();
5621 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5622 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5623 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5625 PseudoSourceValue::getFixedStack(SSFI), 0,
5627 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5630 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5632 SelectionDAG &DAG) const {
5634 DebugLoc dl = Op.getDebugLoc();
5636 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5638 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5640 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5641 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5642 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5643 Tys, Ops, array_lengthof(Ops));
5646 Chain = Result.getValue(1);
5647 SDValue InFlag = Result.getValue(2);
5649 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5650 // shouldn't be necessary except that RFP cannot be live across
5651 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5652 MachineFunction &MF = DAG.getMachineFunction();
5653 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5654 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5655 Tys = DAG.getVTList(MVT::Other);
5657 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5659 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5660 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5661 PseudoSourceValue::getFixedStack(SSFI), 0,
5668 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5669 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5670 SelectionDAG &DAG) const {
5671 // This algorithm is not obvious. Here it is in C code, more or less:
5673 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5674 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5675 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5677 // Copy ints to xmm registers.
5678 __m128i xh = _mm_cvtsi32_si128( hi );
5679 __m128i xl = _mm_cvtsi32_si128( lo );
5681 // Combine into low half of a single xmm register.
5682 __m128i x = _mm_unpacklo_epi32( xh, xl );
5686 // Merge in appropriate exponents to give the integer bits the right
5688 x = _mm_unpacklo_epi32( x, exp );
5690 // Subtract away the biases to deal with the IEEE-754 double precision
5692 d = _mm_sub_pd( (__m128d) x, bias );
5694 // All conversions up to here are exact. The correctly rounded result is
5695 // calculated using the current rounding mode using the following
5697 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5698 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5699 // store doesn't really need to be here (except
5700 // maybe to zero the other double)
5705 DebugLoc dl = Op.getDebugLoc();
5706 LLVMContext *Context = DAG.getContext();
5708 // Build some magic constants.
5709 std::vector<Constant*> CV0;
5710 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5711 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5712 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5713 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5714 Constant *C0 = ConstantVector::get(CV0);
5715 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5717 std::vector<Constant*> CV1;
5719 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5721 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5722 Constant *C1 = ConstantVector::get(CV1);
5723 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5725 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5726 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5728 DAG.getIntPtrConstant(1)));
5729 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5730 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5732 DAG.getIntPtrConstant(0)));
5733 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5734 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5735 PseudoSourceValue::getConstantPool(), 0,
5737 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5738 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5739 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5740 PseudoSourceValue::getConstantPool(), 0,
5742 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5744 // Add the halves; easiest way is to swap them into another reg first.
5745 int ShufMask[2] = { 1, -1 };
5746 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5747 DAG.getUNDEF(MVT::v2f64), ShufMask);
5748 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5749 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5750 DAG.getIntPtrConstant(0));
5753 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5754 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5755 SelectionDAG &DAG) const {
5756 DebugLoc dl = Op.getDebugLoc();
5757 // FP constant to bias correct the final result.
5758 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5761 // Load the 32-bit value into an XMM register.
5762 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5763 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5765 DAG.getIntPtrConstant(0)));
5767 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5768 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5769 DAG.getIntPtrConstant(0));
5771 // Or the load with the bias.
5772 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5773 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5774 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5776 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5777 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5778 MVT::v2f64, Bias)));
5779 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5780 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5781 DAG.getIntPtrConstant(0));
5783 // Subtract the bias.
5784 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5786 // Handle final rounding.
5787 EVT DestVT = Op.getValueType();
5789 if (DestVT.bitsLT(MVT::f64)) {
5790 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5791 DAG.getIntPtrConstant(0));
5792 } else if (DestVT.bitsGT(MVT::f64)) {
5793 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5796 // Handle final rounding.
5800 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5801 SelectionDAG &DAG) const {
5802 SDValue N0 = Op.getOperand(0);
5803 DebugLoc dl = Op.getDebugLoc();
5805 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5806 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5807 // the optimization here.
5808 if (DAG.SignBitIsZero(N0))
5809 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5811 EVT SrcVT = N0.getValueType();
5812 EVT DstVT = Op.getValueType();
5813 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5814 return LowerUINT_TO_FP_i64(Op, DAG);
5815 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5816 return LowerUINT_TO_FP_i32(Op, DAG);
5818 // Make a 64-bit buffer, and use it to build an FILD.
5819 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5820 if (SrcVT == MVT::i32) {
5821 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5822 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5823 getPointerTy(), StackSlot, WordOff);
5824 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5825 StackSlot, NULL, 0, false, false, 0);
5826 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5827 OffsetSlot, NULL, 0, false, false, 0);
5828 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5832 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5833 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5834 StackSlot, NULL, 0, false, false, 0);
5835 // For i64 source, we need to add the appropriate power of 2 if the input
5836 // was negative. This is the same as the optimization in
5837 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5838 // we must be careful to do the computation in x87 extended precision, not
5839 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5840 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5841 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5842 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5844 APInt FF(32, 0x5F800000ULL);
5846 // Check whether the sign bit is set.
5847 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5848 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5851 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5852 SDValue FudgePtr = DAG.getConstantPool(
5853 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5856 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5857 SDValue Zero = DAG.getIntPtrConstant(0);
5858 SDValue Four = DAG.getIntPtrConstant(4);
5859 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5861 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5863 // Load the value out, extending it from f32 to f80.
5864 // FIXME: Avoid the extend by constructing the right constant pool?
5865 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
5866 FudgePtr, PseudoSourceValue::getConstantPool(),
5867 0, MVT::f32, false, false, 4);
5868 // Extend everything to 80 bits to force it to be done on x87.
5869 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5870 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5873 std::pair<SDValue,SDValue> X86TargetLowering::
5874 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5875 DebugLoc dl = Op.getDebugLoc();
5877 EVT DstTy = Op.getValueType();
5880 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5884 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5885 DstTy.getSimpleVT() >= MVT::i16 &&
5886 "Unknown FP_TO_SINT to lower!");
5888 // These are really Legal.
5889 if (DstTy == MVT::i32 &&
5890 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5891 return std::make_pair(SDValue(), SDValue());
5892 if (Subtarget->is64Bit() &&
5893 DstTy == MVT::i64 &&
5894 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5895 return std::make_pair(SDValue(), SDValue());
5897 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5899 MachineFunction &MF = DAG.getMachineFunction();
5900 unsigned MemSize = DstTy.getSizeInBits()/8;
5901 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5902 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5905 switch (DstTy.getSimpleVT().SimpleTy) {
5906 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5907 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5908 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5909 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5912 SDValue Chain = DAG.getEntryNode();
5913 SDValue Value = Op.getOperand(0);
5914 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5915 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5916 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5917 PseudoSourceValue::getFixedStack(SSFI), 0,
5919 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5921 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5923 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5924 Chain = Value.getValue(1);
5925 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5926 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5929 // Build the FP_TO_INT*_IN_MEM
5930 SDValue Ops[] = { Chain, Value, StackSlot };
5931 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5933 return std::make_pair(FIST, StackSlot);
5936 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5937 SelectionDAG &DAG) const {
5938 if (Op.getValueType().isVector()) {
5939 if (Op.getValueType() == MVT::v2i32 &&
5940 Op.getOperand(0).getValueType() == MVT::v2f64) {
5946 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5947 SDValue FIST = Vals.first, StackSlot = Vals.second;
5948 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5949 if (FIST.getNode() == 0) return Op;
5952 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5953 FIST, StackSlot, NULL, 0, false, false, 0);
5956 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5957 SelectionDAG &DAG) const {
5958 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5959 SDValue FIST = Vals.first, StackSlot = Vals.second;
5960 assert(FIST.getNode() && "Unexpected failure");
5963 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5964 FIST, StackSlot, NULL, 0, false, false, 0);
5967 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5968 SelectionDAG &DAG) const {
5969 LLVMContext *Context = DAG.getContext();
5970 DebugLoc dl = Op.getDebugLoc();
5971 EVT VT = Op.getValueType();
5974 EltVT = VT.getVectorElementType();
5975 std::vector<Constant*> CV;
5976 if (EltVT == MVT::f64) {
5977 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5981 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5987 Constant *C = ConstantVector::get(CV);
5988 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5989 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5990 PseudoSourceValue::getConstantPool(), 0,
5992 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5995 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5996 LLVMContext *Context = DAG.getContext();
5997 DebugLoc dl = Op.getDebugLoc();
5998 EVT VT = Op.getValueType();
6001 EltVT = VT.getVectorElementType();
6002 std::vector<Constant*> CV;
6003 if (EltVT == MVT::f64) {
6004 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
6008 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
6014 Constant *C = ConstantVector::get(CV);
6015 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6016 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6017 PseudoSourceValue::getConstantPool(), 0,
6019 if (VT.isVector()) {
6020 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6021 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6022 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6024 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
6026 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6030 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6031 LLVMContext *Context = DAG.getContext();
6032 SDValue Op0 = Op.getOperand(0);
6033 SDValue Op1 = Op.getOperand(1);
6034 DebugLoc dl = Op.getDebugLoc();
6035 EVT VT = Op.getValueType();
6036 EVT SrcVT = Op1.getValueType();
6038 // If second operand is smaller, extend it first.
6039 if (SrcVT.bitsLT(VT)) {
6040 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6043 // And if it is bigger, shrink it first.
6044 if (SrcVT.bitsGT(VT)) {
6045 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6049 // At this point the operands and the result should have the same
6050 // type, and that won't be f80 since that is not custom lowered.
6052 // First get the sign bit of second operand.
6053 std::vector<Constant*> CV;
6054 if (SrcVT == MVT::f64) {
6055 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6056 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6058 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6059 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6060 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6061 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6063 Constant *C = ConstantVector::get(CV);
6064 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6065 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6066 PseudoSourceValue::getConstantPool(), 0,
6068 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6070 // Shift sign bit right or left if the two operands have different types.
6071 if (SrcVT.bitsGT(VT)) {
6072 // Op0 is MVT::f32, Op1 is MVT::f64.
6073 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6074 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6075 DAG.getConstant(32, MVT::i32));
6076 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
6077 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6078 DAG.getIntPtrConstant(0));
6081 // Clear first operand sign bit.
6083 if (VT == MVT::f64) {
6084 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6085 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6087 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6088 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6089 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6090 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6092 C = ConstantVector::get(CV);
6093 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6094 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6095 PseudoSourceValue::getConstantPool(), 0,
6097 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6099 // Or the value with the sign bit.
6100 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6103 /// Emit nodes that will be selected as "test Op0,Op0", or something
6105 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6106 SelectionDAG &DAG) const {
6107 DebugLoc dl = Op.getDebugLoc();
6109 // CF and OF aren't always set the way we want. Determine which
6110 // of these we need.
6111 bool NeedCF = false;
6112 bool NeedOF = false;
6115 case X86::COND_A: case X86::COND_AE:
6116 case X86::COND_B: case X86::COND_BE:
6119 case X86::COND_G: case X86::COND_GE:
6120 case X86::COND_L: case X86::COND_LE:
6121 case X86::COND_O: case X86::COND_NO:
6126 // See if we can use the EFLAGS value from the operand instead of
6127 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6128 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6129 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6130 // Emit a CMP with 0, which is the TEST pattern.
6131 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6132 DAG.getConstant(0, Op.getValueType()));
6134 unsigned Opcode = 0;
6135 unsigned NumOperands = 0;
6136 switch (Op.getNode()->getOpcode()) {
6138 // Due to an isel shortcoming, be conservative if this add is likely to be
6139 // selected as part of a load-modify-store instruction. When the root node
6140 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6141 // uses of other nodes in the match, such as the ADD in this case. This
6142 // leads to the ADD being left around and reselected, with the result being
6143 // two adds in the output. Alas, even if none our users are stores, that
6144 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6145 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6146 // climbing the DAG back to the root, and it doesn't seem to be worth the
6148 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6149 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6150 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6153 if (ConstantSDNode *C =
6154 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6155 // An add of one will be selected as an INC.
6156 if (C->getAPIntValue() == 1) {
6157 Opcode = X86ISD::INC;
6162 // An add of negative one (subtract of one) will be selected as a DEC.
6163 if (C->getAPIntValue().isAllOnesValue()) {
6164 Opcode = X86ISD::DEC;
6170 // Otherwise use a regular EFLAGS-setting add.
6171 Opcode = X86ISD::ADD;
6175 // If the primary and result isn't used, don't bother using X86ISD::AND,
6176 // because a TEST instruction will be better.
6177 bool NonFlagUse = false;
6178 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6179 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6181 unsigned UOpNo = UI.getOperandNo();
6182 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6183 // Look pass truncate.
6184 UOpNo = User->use_begin().getOperandNo();
6185 User = *User->use_begin();
6188 if (User->getOpcode() != ISD::BRCOND &&
6189 User->getOpcode() != ISD::SETCC &&
6190 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6203 // Due to the ISEL shortcoming noted above, be conservative if this op is
6204 // likely to be selected as part of a load-modify-store instruction.
6205 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6206 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6207 if (UI->getOpcode() == ISD::STORE)
6210 // Otherwise use a regular EFLAGS-setting instruction.
6211 switch (Op.getNode()->getOpcode()) {
6212 default: llvm_unreachable("unexpected operator!");
6213 case ISD::SUB: Opcode = X86ISD::SUB; break;
6214 case ISD::OR: Opcode = X86ISD::OR; break;
6215 case ISD::XOR: Opcode = X86ISD::XOR; break;
6216 case ISD::AND: Opcode = X86ISD::AND; break;
6228 return SDValue(Op.getNode(), 1);
6235 // Emit a CMP with 0, which is the TEST pattern.
6236 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6237 DAG.getConstant(0, Op.getValueType()));
6239 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6240 SmallVector<SDValue, 4> Ops;
6241 for (unsigned i = 0; i != NumOperands; ++i)
6242 Ops.push_back(Op.getOperand(i));
6244 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6245 DAG.ReplaceAllUsesWith(Op, New);
6246 return SDValue(New.getNode(), 1);
6249 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6251 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6252 SelectionDAG &DAG) const {
6253 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6254 if (C->getAPIntValue() == 0)
6255 return EmitTest(Op0, X86CC, DAG);
6257 DebugLoc dl = Op0.getDebugLoc();
6258 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6261 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6262 /// if it's possible.
6263 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6264 DebugLoc dl, SelectionDAG &DAG) const {
6265 SDValue Op0 = And.getOperand(0);
6266 SDValue Op1 = And.getOperand(1);
6267 if (Op0.getOpcode() == ISD::TRUNCATE)
6268 Op0 = Op0.getOperand(0);
6269 if (Op1.getOpcode() == ISD::TRUNCATE)
6270 Op1 = Op1.getOperand(0);
6273 if (Op1.getOpcode() == ISD::SHL)
6274 std::swap(Op0, Op1);
6275 if (Op0.getOpcode() == ISD::SHL) {
6276 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6277 if (And00C->getZExtValue() == 1) {
6278 // If we looked past a truncate, check that it's only truncating away
6280 unsigned BitWidth = Op0.getValueSizeInBits();
6281 unsigned AndBitWidth = And.getValueSizeInBits();
6282 if (BitWidth > AndBitWidth) {
6283 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6284 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6285 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6289 RHS = Op0.getOperand(1);
6291 } else if (Op1.getOpcode() == ISD::Constant) {
6292 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6293 SDValue AndLHS = Op0;
6294 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6295 LHS = AndLHS.getOperand(0);
6296 RHS = AndLHS.getOperand(1);
6300 if (LHS.getNode()) {
6301 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6302 // instruction. Since the shift amount is in-range-or-undefined, we know
6303 // that doing a bittest on the i32 value is ok. We extend to i32 because
6304 // the encoding for the i16 version is larger than the i32 version.
6305 // Also promote i16 to i32 for performance / code size reason.
6306 if (LHS.getValueType() == MVT::i8 ||
6307 LHS.getValueType() == MVT::i16)
6308 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6310 // If the operand types disagree, extend the shift amount to match. Since
6311 // BT ignores high bits (like shifts) we can use anyextend.
6312 if (LHS.getValueType() != RHS.getValueType())
6313 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6315 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6316 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6317 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6318 DAG.getConstant(Cond, MVT::i8), BT);
6324 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6325 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6326 SDValue Op0 = Op.getOperand(0);
6327 SDValue Op1 = Op.getOperand(1);
6328 DebugLoc dl = Op.getDebugLoc();
6329 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6331 // Optimize to BT if possible.
6332 // Lower (X & (1 << N)) == 0 to BT(X, N).
6333 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6334 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6335 if (Op0.getOpcode() == ISD::AND &&
6337 Op1.getOpcode() == ISD::Constant &&
6338 cast<ConstantSDNode>(Op1)->isNullValue() &&
6339 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6340 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6341 if (NewSetCC.getNode())
6345 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6346 if (Op0.getOpcode() == X86ISD::SETCC &&
6347 Op1.getOpcode() == ISD::Constant &&
6348 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6349 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6350 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6351 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6352 bool Invert = (CC == ISD::SETNE) ^
6353 cast<ConstantSDNode>(Op1)->isNullValue();
6355 CCode = X86::GetOppositeBranchCondition(CCode);
6356 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6357 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6360 bool isFP = Op1.getValueType().isFloatingPoint();
6361 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6362 if (X86CC == X86::COND_INVALID)
6365 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6367 // Use sbb x, x to materialize carry bit into a GPR.
6368 if (X86CC == X86::COND_B)
6369 return DAG.getNode(ISD::AND, dl, MVT::i8,
6370 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6371 DAG.getConstant(X86CC, MVT::i8), Cond),
6372 DAG.getConstant(1, MVT::i8));
6374 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6375 DAG.getConstant(X86CC, MVT::i8), Cond);
6378 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6380 SDValue Op0 = Op.getOperand(0);
6381 SDValue Op1 = Op.getOperand(1);
6382 SDValue CC = Op.getOperand(2);
6383 EVT VT = Op.getValueType();
6384 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6385 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6386 DebugLoc dl = Op.getDebugLoc();
6390 EVT VT0 = Op0.getValueType();
6391 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6392 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6395 switch (SetCCOpcode) {
6398 case ISD::SETEQ: SSECC = 0; break;
6400 case ISD::SETGT: Swap = true; // Fallthrough
6402 case ISD::SETOLT: SSECC = 1; break;
6404 case ISD::SETGE: Swap = true; // Fallthrough
6406 case ISD::SETOLE: SSECC = 2; break;
6407 case ISD::SETUO: SSECC = 3; break;
6409 case ISD::SETNE: SSECC = 4; break;
6410 case ISD::SETULE: Swap = true;
6411 case ISD::SETUGE: SSECC = 5; break;
6412 case ISD::SETULT: Swap = true;
6413 case ISD::SETUGT: SSECC = 6; break;
6414 case ISD::SETO: SSECC = 7; break;
6417 std::swap(Op0, Op1);
6419 // In the two special cases we can't handle, emit two comparisons.
6421 if (SetCCOpcode == ISD::SETUEQ) {
6423 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6424 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6425 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6427 else if (SetCCOpcode == ISD::SETONE) {
6429 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6430 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6431 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6433 llvm_unreachable("Illegal FP comparison");
6435 // Handle all other FP comparisons here.
6436 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6439 // We are handling one of the integer comparisons here. Since SSE only has
6440 // GT and EQ comparisons for integer, swapping operands and multiple
6441 // operations may be required for some comparisons.
6442 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6443 bool Swap = false, Invert = false, FlipSigns = false;
6445 switch (VT.getSimpleVT().SimpleTy) {
6448 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6450 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6452 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6453 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6456 switch (SetCCOpcode) {
6458 case ISD::SETNE: Invert = true;
6459 case ISD::SETEQ: Opc = EQOpc; break;
6460 case ISD::SETLT: Swap = true;
6461 case ISD::SETGT: Opc = GTOpc; break;
6462 case ISD::SETGE: Swap = true;
6463 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6464 case ISD::SETULT: Swap = true;
6465 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6466 case ISD::SETUGE: Swap = true;
6467 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6470 std::swap(Op0, Op1);
6472 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6473 // bits of the inputs before performing those operations.
6475 EVT EltVT = VT.getVectorElementType();
6476 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6478 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6479 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6481 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6482 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6485 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6487 // If the logical-not of the result is required, perform that now.
6489 Result = DAG.getNOT(dl, Result, VT);
6494 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6495 static bool isX86LogicalCmp(SDValue Op) {
6496 unsigned Opc = Op.getNode()->getOpcode();
6497 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6499 if (Op.getResNo() == 1 &&
6500 (Opc == X86ISD::ADD ||
6501 Opc == X86ISD::SUB ||
6502 Opc == X86ISD::SMUL ||
6503 Opc == X86ISD::UMUL ||
6504 Opc == X86ISD::INC ||
6505 Opc == X86ISD::DEC ||
6506 Opc == X86ISD::OR ||
6507 Opc == X86ISD::XOR ||
6508 Opc == X86ISD::AND))
6514 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6515 bool addTest = true;
6516 SDValue Cond = Op.getOperand(0);
6517 DebugLoc dl = Op.getDebugLoc();
6520 if (Cond.getOpcode() == ISD::SETCC) {
6521 SDValue NewCond = LowerSETCC(Cond, DAG);
6522 if (NewCond.getNode())
6526 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6527 SDValue Op1 = Op.getOperand(1);
6528 SDValue Op2 = Op.getOperand(2);
6529 if (Cond.getOpcode() == X86ISD::SETCC &&
6530 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6531 SDValue Cmp = Cond.getOperand(1);
6532 if (Cmp.getOpcode() == X86ISD::CMP) {
6533 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6534 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6535 ConstantSDNode *RHSC =
6536 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6537 if (N1C && N1C->isAllOnesValue() &&
6538 N2C && N2C->isNullValue() &&
6539 RHSC && RHSC->isNullValue()) {
6540 SDValue CmpOp0 = Cmp.getOperand(0);
6541 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6542 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6543 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6544 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6549 // Look pass (and (setcc_carry (cmp ...)), 1).
6550 if (Cond.getOpcode() == ISD::AND &&
6551 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6552 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6553 if (C && C->getAPIntValue() == 1)
6554 Cond = Cond.getOperand(0);
6557 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6558 // setting operand in place of the X86ISD::SETCC.
6559 if (Cond.getOpcode() == X86ISD::SETCC ||
6560 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6561 CC = Cond.getOperand(0);
6563 SDValue Cmp = Cond.getOperand(1);
6564 unsigned Opc = Cmp.getOpcode();
6565 EVT VT = Op.getValueType();
6567 bool IllegalFPCMov = false;
6568 if (VT.isFloatingPoint() && !VT.isVector() &&
6569 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6570 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6572 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6573 Opc == X86ISD::BT) { // FIXME
6580 // Look pass the truncate.
6581 if (Cond.getOpcode() == ISD::TRUNCATE)
6582 Cond = Cond.getOperand(0);
6584 // We know the result of AND is compared against zero. Try to match
6586 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6587 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6588 if (NewSetCC.getNode()) {
6589 CC = NewSetCC.getOperand(0);
6590 Cond = NewSetCC.getOperand(1);
6597 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6598 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6601 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6602 // condition is true.
6603 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6604 SDValue Ops[] = { Op2, Op1, CC, Cond };
6605 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6608 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6609 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6610 // from the AND / OR.
6611 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6612 Opc = Op.getOpcode();
6613 if (Opc != ISD::OR && Opc != ISD::AND)
6615 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6616 Op.getOperand(0).hasOneUse() &&
6617 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6618 Op.getOperand(1).hasOneUse());
6621 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6622 // 1 and that the SETCC node has a single use.
6623 static bool isXor1OfSetCC(SDValue Op) {
6624 if (Op.getOpcode() != ISD::XOR)
6626 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6627 if (N1C && N1C->getAPIntValue() == 1) {
6628 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6629 Op.getOperand(0).hasOneUse();
6634 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6635 bool addTest = true;
6636 SDValue Chain = Op.getOperand(0);
6637 SDValue Cond = Op.getOperand(1);
6638 SDValue Dest = Op.getOperand(2);
6639 DebugLoc dl = Op.getDebugLoc();
6642 if (Cond.getOpcode() == ISD::SETCC) {
6643 SDValue NewCond = LowerSETCC(Cond, DAG);
6644 if (NewCond.getNode())
6648 // FIXME: LowerXALUO doesn't handle these!!
6649 else if (Cond.getOpcode() == X86ISD::ADD ||
6650 Cond.getOpcode() == X86ISD::SUB ||
6651 Cond.getOpcode() == X86ISD::SMUL ||
6652 Cond.getOpcode() == X86ISD::UMUL)
6653 Cond = LowerXALUO(Cond, DAG);
6656 // Look pass (and (setcc_carry (cmp ...)), 1).
6657 if (Cond.getOpcode() == ISD::AND &&
6658 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6659 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6660 if (C && C->getAPIntValue() == 1)
6661 Cond = Cond.getOperand(0);
6664 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6665 // setting operand in place of the X86ISD::SETCC.
6666 if (Cond.getOpcode() == X86ISD::SETCC ||
6667 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6668 CC = Cond.getOperand(0);
6670 SDValue Cmp = Cond.getOperand(1);
6671 unsigned Opc = Cmp.getOpcode();
6672 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6673 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6677 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6681 // These can only come from an arithmetic instruction with overflow,
6682 // e.g. SADDO, UADDO.
6683 Cond = Cond.getNode()->getOperand(1);
6690 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6691 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6692 if (CondOpc == ISD::OR) {
6693 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6694 // two branches instead of an explicit OR instruction with a
6696 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6697 isX86LogicalCmp(Cmp)) {
6698 CC = Cond.getOperand(0).getOperand(0);
6699 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6700 Chain, Dest, CC, Cmp);
6701 CC = Cond.getOperand(1).getOperand(0);
6705 } else { // ISD::AND
6706 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6707 // two branches instead of an explicit AND instruction with a
6708 // separate test. However, we only do this if this block doesn't
6709 // have a fall-through edge, because this requires an explicit
6710 // jmp when the condition is false.
6711 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6712 isX86LogicalCmp(Cmp) &&
6713 Op.getNode()->hasOneUse()) {
6714 X86::CondCode CCode =
6715 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6716 CCode = X86::GetOppositeBranchCondition(CCode);
6717 CC = DAG.getConstant(CCode, MVT::i8);
6718 SDNode *User = *Op.getNode()->use_begin();
6719 // Look for an unconditional branch following this conditional branch.
6720 // We need this because we need to reverse the successors in order
6721 // to implement FCMP_OEQ.
6722 if (User->getOpcode() == ISD::BR) {
6723 SDValue FalseBB = User->getOperand(1);
6725 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6726 assert(NewBR == User);
6730 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6731 Chain, Dest, CC, Cmp);
6732 X86::CondCode CCode =
6733 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6734 CCode = X86::GetOppositeBranchCondition(CCode);
6735 CC = DAG.getConstant(CCode, MVT::i8);
6741 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6742 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6743 // It should be transformed during dag combiner except when the condition
6744 // is set by a arithmetics with overflow node.
6745 X86::CondCode CCode =
6746 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6747 CCode = X86::GetOppositeBranchCondition(CCode);
6748 CC = DAG.getConstant(CCode, MVT::i8);
6749 Cond = Cond.getOperand(0).getOperand(1);
6755 // Look pass the truncate.
6756 if (Cond.getOpcode() == ISD::TRUNCATE)
6757 Cond = Cond.getOperand(0);
6759 // We know the result of AND is compared against zero. Try to match
6761 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6762 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6763 if (NewSetCC.getNode()) {
6764 CC = NewSetCC.getOperand(0);
6765 Cond = NewSetCC.getOperand(1);
6772 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6773 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6775 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6776 Chain, Dest, CC, Cond);
6780 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6781 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6782 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6783 // that the guard pages used by the OS virtual memory manager are allocated in
6784 // correct sequence.
6786 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6787 SelectionDAG &DAG) const {
6788 assert(Subtarget->isTargetCygMing() &&
6789 "This should be used only on Cygwin/Mingw targets");
6790 DebugLoc dl = Op.getDebugLoc();
6793 SDValue Chain = Op.getOperand(0);
6794 SDValue Size = Op.getOperand(1);
6795 // FIXME: Ensure alignment here
6799 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6801 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6802 Flag = Chain.getValue(1);
6804 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6806 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6807 Flag = Chain.getValue(1);
6809 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6811 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6812 return DAG.getMergeValues(Ops1, 2, dl);
6815 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6816 MachineFunction &MF = DAG.getMachineFunction();
6817 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6819 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6820 DebugLoc dl = Op.getDebugLoc();
6822 if (!Subtarget->is64Bit()) {
6823 // vastart just stores the address of the VarArgsFrameIndex slot into the
6824 // memory location argument.
6825 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6827 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6832 // gp_offset (0 - 6 * 8)
6833 // fp_offset (48 - 48 + 8 * 16)
6834 // overflow_arg_area (point to parameters coming in memory).
6836 SmallVector<SDValue, 8> MemOps;
6837 SDValue FIN = Op.getOperand(1);
6839 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6840 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6842 FIN, SV, 0, false, false, 0);
6843 MemOps.push_back(Store);
6846 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6847 FIN, DAG.getIntPtrConstant(4));
6848 Store = DAG.getStore(Op.getOperand(0), dl,
6849 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6851 FIN, SV, 4, false, false, 0);
6852 MemOps.push_back(Store);
6854 // Store ptr to overflow_arg_area
6855 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6856 FIN, DAG.getIntPtrConstant(4));
6857 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6859 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 8,
6861 MemOps.push_back(Store);
6863 // Store ptr to reg_save_area.
6864 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6865 FIN, DAG.getIntPtrConstant(8));
6866 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6868 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 16,
6870 MemOps.push_back(Store);
6871 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6872 &MemOps[0], MemOps.size());
6875 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6876 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6877 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6879 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6883 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6884 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6885 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6886 SDValue Chain = Op.getOperand(0);
6887 SDValue DstPtr = Op.getOperand(1);
6888 SDValue SrcPtr = Op.getOperand(2);
6889 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6890 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6891 DebugLoc dl = Op.getDebugLoc();
6893 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6894 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6895 false, DstSV, 0, SrcSV, 0);
6899 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6900 DebugLoc dl = Op.getDebugLoc();
6901 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6903 default: return SDValue(); // Don't custom lower most intrinsics.
6904 // Comparison intrinsics.
6905 case Intrinsic::x86_sse_comieq_ss:
6906 case Intrinsic::x86_sse_comilt_ss:
6907 case Intrinsic::x86_sse_comile_ss:
6908 case Intrinsic::x86_sse_comigt_ss:
6909 case Intrinsic::x86_sse_comige_ss:
6910 case Intrinsic::x86_sse_comineq_ss:
6911 case Intrinsic::x86_sse_ucomieq_ss:
6912 case Intrinsic::x86_sse_ucomilt_ss:
6913 case Intrinsic::x86_sse_ucomile_ss:
6914 case Intrinsic::x86_sse_ucomigt_ss:
6915 case Intrinsic::x86_sse_ucomige_ss:
6916 case Intrinsic::x86_sse_ucomineq_ss:
6917 case Intrinsic::x86_sse2_comieq_sd:
6918 case Intrinsic::x86_sse2_comilt_sd:
6919 case Intrinsic::x86_sse2_comile_sd:
6920 case Intrinsic::x86_sse2_comigt_sd:
6921 case Intrinsic::x86_sse2_comige_sd:
6922 case Intrinsic::x86_sse2_comineq_sd:
6923 case Intrinsic::x86_sse2_ucomieq_sd:
6924 case Intrinsic::x86_sse2_ucomilt_sd:
6925 case Intrinsic::x86_sse2_ucomile_sd:
6926 case Intrinsic::x86_sse2_ucomigt_sd:
6927 case Intrinsic::x86_sse2_ucomige_sd:
6928 case Intrinsic::x86_sse2_ucomineq_sd: {
6930 ISD::CondCode CC = ISD::SETCC_INVALID;
6933 case Intrinsic::x86_sse_comieq_ss:
6934 case Intrinsic::x86_sse2_comieq_sd:
6938 case Intrinsic::x86_sse_comilt_ss:
6939 case Intrinsic::x86_sse2_comilt_sd:
6943 case Intrinsic::x86_sse_comile_ss:
6944 case Intrinsic::x86_sse2_comile_sd:
6948 case Intrinsic::x86_sse_comigt_ss:
6949 case Intrinsic::x86_sse2_comigt_sd:
6953 case Intrinsic::x86_sse_comige_ss:
6954 case Intrinsic::x86_sse2_comige_sd:
6958 case Intrinsic::x86_sse_comineq_ss:
6959 case Intrinsic::x86_sse2_comineq_sd:
6963 case Intrinsic::x86_sse_ucomieq_ss:
6964 case Intrinsic::x86_sse2_ucomieq_sd:
6965 Opc = X86ISD::UCOMI;
6968 case Intrinsic::x86_sse_ucomilt_ss:
6969 case Intrinsic::x86_sse2_ucomilt_sd:
6970 Opc = X86ISD::UCOMI;
6973 case Intrinsic::x86_sse_ucomile_ss:
6974 case Intrinsic::x86_sse2_ucomile_sd:
6975 Opc = X86ISD::UCOMI;
6978 case Intrinsic::x86_sse_ucomigt_ss:
6979 case Intrinsic::x86_sse2_ucomigt_sd:
6980 Opc = X86ISD::UCOMI;
6983 case Intrinsic::x86_sse_ucomige_ss:
6984 case Intrinsic::x86_sse2_ucomige_sd:
6985 Opc = X86ISD::UCOMI;
6988 case Intrinsic::x86_sse_ucomineq_ss:
6989 case Intrinsic::x86_sse2_ucomineq_sd:
6990 Opc = X86ISD::UCOMI;
6995 SDValue LHS = Op.getOperand(1);
6996 SDValue RHS = Op.getOperand(2);
6997 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6998 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6999 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
7000 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7001 DAG.getConstant(X86CC, MVT::i8), Cond);
7002 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7004 // ptest and testp intrinsics. The intrinsic these come from are designed to
7005 // return an integer value, not just an instruction so lower it to the ptest
7006 // or testp pattern and a setcc for the result.
7007 case Intrinsic::x86_sse41_ptestz:
7008 case Intrinsic::x86_sse41_ptestc:
7009 case Intrinsic::x86_sse41_ptestnzc:
7010 case Intrinsic::x86_avx_ptestz_256:
7011 case Intrinsic::x86_avx_ptestc_256:
7012 case Intrinsic::x86_avx_ptestnzc_256:
7013 case Intrinsic::x86_avx_vtestz_ps:
7014 case Intrinsic::x86_avx_vtestc_ps:
7015 case Intrinsic::x86_avx_vtestnzc_ps:
7016 case Intrinsic::x86_avx_vtestz_pd:
7017 case Intrinsic::x86_avx_vtestc_pd:
7018 case Intrinsic::x86_avx_vtestnzc_pd:
7019 case Intrinsic::x86_avx_vtestz_ps_256:
7020 case Intrinsic::x86_avx_vtestc_ps_256:
7021 case Intrinsic::x86_avx_vtestnzc_ps_256:
7022 case Intrinsic::x86_avx_vtestz_pd_256:
7023 case Intrinsic::x86_avx_vtestc_pd_256:
7024 case Intrinsic::x86_avx_vtestnzc_pd_256: {
7025 bool IsTestPacked = false;
7028 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
7029 case Intrinsic::x86_avx_vtestz_ps:
7030 case Intrinsic::x86_avx_vtestz_pd:
7031 case Intrinsic::x86_avx_vtestz_ps_256:
7032 case Intrinsic::x86_avx_vtestz_pd_256:
7033 IsTestPacked = true; // Fallthrough
7034 case Intrinsic::x86_sse41_ptestz:
7035 case Intrinsic::x86_avx_ptestz_256:
7037 X86CC = X86::COND_E;
7039 case Intrinsic::x86_avx_vtestc_ps:
7040 case Intrinsic::x86_avx_vtestc_pd:
7041 case Intrinsic::x86_avx_vtestc_ps_256:
7042 case Intrinsic::x86_avx_vtestc_pd_256:
7043 IsTestPacked = true; // Fallthrough
7044 case Intrinsic::x86_sse41_ptestc:
7045 case Intrinsic::x86_avx_ptestc_256:
7047 X86CC = X86::COND_B;
7049 case Intrinsic::x86_avx_vtestnzc_ps:
7050 case Intrinsic::x86_avx_vtestnzc_pd:
7051 case Intrinsic::x86_avx_vtestnzc_ps_256:
7052 case Intrinsic::x86_avx_vtestnzc_pd_256:
7053 IsTestPacked = true; // Fallthrough
7054 case Intrinsic::x86_sse41_ptestnzc:
7055 case Intrinsic::x86_avx_ptestnzc_256:
7057 X86CC = X86::COND_A;
7061 SDValue LHS = Op.getOperand(1);
7062 SDValue RHS = Op.getOperand(2);
7063 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
7064 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
7065 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7066 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7067 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7070 // Fix vector shift instructions where the last operand is a non-immediate
7072 case Intrinsic::x86_sse2_pslli_w:
7073 case Intrinsic::x86_sse2_pslli_d:
7074 case Intrinsic::x86_sse2_pslli_q:
7075 case Intrinsic::x86_sse2_psrli_w:
7076 case Intrinsic::x86_sse2_psrli_d:
7077 case Intrinsic::x86_sse2_psrli_q:
7078 case Intrinsic::x86_sse2_psrai_w:
7079 case Intrinsic::x86_sse2_psrai_d:
7080 case Intrinsic::x86_mmx_pslli_w:
7081 case Intrinsic::x86_mmx_pslli_d:
7082 case Intrinsic::x86_mmx_pslli_q:
7083 case Intrinsic::x86_mmx_psrli_w:
7084 case Intrinsic::x86_mmx_psrli_d:
7085 case Intrinsic::x86_mmx_psrli_q:
7086 case Intrinsic::x86_mmx_psrai_w:
7087 case Intrinsic::x86_mmx_psrai_d: {
7088 SDValue ShAmt = Op.getOperand(2);
7089 if (isa<ConstantSDNode>(ShAmt))
7092 unsigned NewIntNo = 0;
7093 EVT ShAmtVT = MVT::v4i32;
7095 case Intrinsic::x86_sse2_pslli_w:
7096 NewIntNo = Intrinsic::x86_sse2_psll_w;
7098 case Intrinsic::x86_sse2_pslli_d:
7099 NewIntNo = Intrinsic::x86_sse2_psll_d;
7101 case Intrinsic::x86_sse2_pslli_q:
7102 NewIntNo = Intrinsic::x86_sse2_psll_q;
7104 case Intrinsic::x86_sse2_psrli_w:
7105 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7107 case Intrinsic::x86_sse2_psrli_d:
7108 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7110 case Intrinsic::x86_sse2_psrli_q:
7111 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7113 case Intrinsic::x86_sse2_psrai_w:
7114 NewIntNo = Intrinsic::x86_sse2_psra_w;
7116 case Intrinsic::x86_sse2_psrai_d:
7117 NewIntNo = Intrinsic::x86_sse2_psra_d;
7120 ShAmtVT = MVT::v2i32;
7122 case Intrinsic::x86_mmx_pslli_w:
7123 NewIntNo = Intrinsic::x86_mmx_psll_w;
7125 case Intrinsic::x86_mmx_pslli_d:
7126 NewIntNo = Intrinsic::x86_mmx_psll_d;
7128 case Intrinsic::x86_mmx_pslli_q:
7129 NewIntNo = Intrinsic::x86_mmx_psll_q;
7131 case Intrinsic::x86_mmx_psrli_w:
7132 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7134 case Intrinsic::x86_mmx_psrli_d:
7135 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7137 case Intrinsic::x86_mmx_psrli_q:
7138 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7140 case Intrinsic::x86_mmx_psrai_w:
7141 NewIntNo = Intrinsic::x86_mmx_psra_w;
7143 case Intrinsic::x86_mmx_psrai_d:
7144 NewIntNo = Intrinsic::x86_mmx_psra_d;
7146 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7152 // The vector shift intrinsics with scalars uses 32b shift amounts but
7153 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7157 ShOps[1] = DAG.getConstant(0, MVT::i32);
7158 if (ShAmtVT == MVT::v4i32) {
7159 ShOps[2] = DAG.getUNDEF(MVT::i32);
7160 ShOps[3] = DAG.getUNDEF(MVT::i32);
7161 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7163 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7166 EVT VT = Op.getValueType();
7167 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7168 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7169 DAG.getConstant(NewIntNo, MVT::i32),
7170 Op.getOperand(1), ShAmt);
7175 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7176 SelectionDAG &DAG) const {
7177 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7178 MFI->setReturnAddressIsTaken(true);
7180 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7181 DebugLoc dl = Op.getDebugLoc();
7184 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7186 DAG.getConstant(TD->getPointerSize(),
7187 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7188 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7189 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7191 NULL, 0, false, false, 0);
7194 // Just load the return address.
7195 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7196 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7197 RetAddrFI, NULL, 0, false, false, 0);
7200 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7201 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7202 MFI->setFrameAddressIsTaken(true);
7204 EVT VT = Op.getValueType();
7205 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7206 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7207 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7208 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7210 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7215 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7216 SelectionDAG &DAG) const {
7217 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7220 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7221 MachineFunction &MF = DAG.getMachineFunction();
7222 SDValue Chain = Op.getOperand(0);
7223 SDValue Offset = Op.getOperand(1);
7224 SDValue Handler = Op.getOperand(2);
7225 DebugLoc dl = Op.getDebugLoc();
7227 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
7228 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7230 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7232 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
7233 DAG.getIntPtrConstant(TD->getPointerSize()));
7234 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7235 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7236 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7237 MF.getRegInfo().addLiveOut(StoreAddrReg);
7239 return DAG.getNode(X86ISD::EH_RETURN, dl,
7241 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7244 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7245 SelectionDAG &DAG) const {
7246 SDValue Root = Op.getOperand(0);
7247 SDValue Trmp = Op.getOperand(1); // trampoline
7248 SDValue FPtr = Op.getOperand(2); // nested function
7249 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7250 DebugLoc dl = Op.getDebugLoc();
7252 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7254 if (Subtarget->is64Bit()) {
7255 SDValue OutChains[6];
7257 // Large code-model.
7258 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7259 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7261 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7262 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7264 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7266 // Load the pointer to the nested function into R11.
7267 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7268 SDValue Addr = Trmp;
7269 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7270 Addr, TrmpAddr, 0, false, false, 0);
7272 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7273 DAG.getConstant(2, MVT::i64));
7274 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7277 // Load the 'nest' parameter value into R10.
7278 // R10 is specified in X86CallingConv.td
7279 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7280 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7281 DAG.getConstant(10, MVT::i64));
7282 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7283 Addr, TrmpAddr, 10, false, false, 0);
7285 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7286 DAG.getConstant(12, MVT::i64));
7287 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7290 // Jump to the nested function.
7291 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7292 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7293 DAG.getConstant(20, MVT::i64));
7294 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7295 Addr, TrmpAddr, 20, false, false, 0);
7297 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7298 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7299 DAG.getConstant(22, MVT::i64));
7300 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7301 TrmpAddr, 22, false, false, 0);
7304 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7305 return DAG.getMergeValues(Ops, 2, dl);
7307 const Function *Func =
7308 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7309 CallingConv::ID CC = Func->getCallingConv();
7314 llvm_unreachable("Unsupported calling convention");
7315 case CallingConv::C:
7316 case CallingConv::X86_StdCall: {
7317 // Pass 'nest' parameter in ECX.
7318 // Must be kept in sync with X86CallingConv.td
7321 // Check that ECX wasn't needed by an 'inreg' parameter.
7322 const FunctionType *FTy = Func->getFunctionType();
7323 const AttrListPtr &Attrs = Func->getAttributes();
7325 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7326 unsigned InRegCount = 0;
7329 for (FunctionType::param_iterator I = FTy->param_begin(),
7330 E = FTy->param_end(); I != E; ++I, ++Idx)
7331 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7332 // FIXME: should only count parameters that are lowered to integers.
7333 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7335 if (InRegCount > 2) {
7336 report_fatal_error("Nest register in use - reduce number of inreg"
7342 case CallingConv::X86_FastCall:
7343 case CallingConv::X86_ThisCall:
7344 case CallingConv::Fast:
7345 // Pass 'nest' parameter in EAX.
7346 // Must be kept in sync with X86CallingConv.td
7351 SDValue OutChains[4];
7354 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7355 DAG.getConstant(10, MVT::i32));
7356 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7358 // This is storing the opcode for MOV32ri.
7359 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7360 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7361 OutChains[0] = DAG.getStore(Root, dl,
7362 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7363 Trmp, TrmpAddr, 0, false, false, 0);
7365 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7366 DAG.getConstant(1, MVT::i32));
7367 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7370 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7371 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7372 DAG.getConstant(5, MVT::i32));
7373 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7374 TrmpAddr, 5, false, false, 1);
7376 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7377 DAG.getConstant(6, MVT::i32));
7378 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7382 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7383 return DAG.getMergeValues(Ops, 2, dl);
7387 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7388 SelectionDAG &DAG) const {
7390 The rounding mode is in bits 11:10 of FPSR, and has the following
7397 FLT_ROUNDS, on the other hand, expects the following:
7404 To perform the conversion, we do:
7405 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7408 MachineFunction &MF = DAG.getMachineFunction();
7409 const TargetMachine &TM = MF.getTarget();
7410 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7411 unsigned StackAlignment = TFI.getStackAlignment();
7412 EVT VT = Op.getValueType();
7413 DebugLoc dl = Op.getDebugLoc();
7415 // Save FP Control Word to stack slot
7416 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7417 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7419 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7420 DAG.getEntryNode(), StackSlot);
7422 // Load FP Control Word from stack slot
7423 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7426 // Transform as necessary
7428 DAG.getNode(ISD::SRL, dl, MVT::i16,
7429 DAG.getNode(ISD::AND, dl, MVT::i16,
7430 CWD, DAG.getConstant(0x800, MVT::i16)),
7431 DAG.getConstant(11, MVT::i8));
7433 DAG.getNode(ISD::SRL, dl, MVT::i16,
7434 DAG.getNode(ISD::AND, dl, MVT::i16,
7435 CWD, DAG.getConstant(0x400, MVT::i16)),
7436 DAG.getConstant(9, MVT::i8));
7439 DAG.getNode(ISD::AND, dl, MVT::i16,
7440 DAG.getNode(ISD::ADD, dl, MVT::i16,
7441 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7442 DAG.getConstant(1, MVT::i16)),
7443 DAG.getConstant(3, MVT::i16));
7446 return DAG.getNode((VT.getSizeInBits() < 16 ?
7447 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7450 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7451 EVT VT = Op.getValueType();
7453 unsigned NumBits = VT.getSizeInBits();
7454 DebugLoc dl = Op.getDebugLoc();
7456 Op = Op.getOperand(0);
7457 if (VT == MVT::i8) {
7458 // Zero extend to i32 since there is not an i8 bsr.
7460 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7463 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7464 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7465 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7467 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7470 DAG.getConstant(NumBits+NumBits-1, OpVT),
7471 DAG.getConstant(X86::COND_E, MVT::i8),
7474 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7476 // Finally xor with NumBits-1.
7477 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7480 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7484 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7485 EVT VT = Op.getValueType();
7487 unsigned NumBits = VT.getSizeInBits();
7488 DebugLoc dl = Op.getDebugLoc();
7490 Op = Op.getOperand(0);
7491 if (VT == MVT::i8) {
7493 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7496 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7497 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7498 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7500 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7503 DAG.getConstant(NumBits, OpVT),
7504 DAG.getConstant(X86::COND_E, MVT::i8),
7507 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7510 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7514 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7515 EVT VT = Op.getValueType();
7516 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7517 DebugLoc dl = Op.getDebugLoc();
7519 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7520 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7521 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7522 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7523 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7525 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7526 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7527 // return AloBlo + AloBhi + AhiBlo;
7529 SDValue A = Op.getOperand(0);
7530 SDValue B = Op.getOperand(1);
7532 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7533 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7534 A, DAG.getConstant(32, MVT::i32));
7535 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7536 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7537 B, DAG.getConstant(32, MVT::i32));
7538 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7539 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7541 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7542 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7544 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7545 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7547 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7548 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7549 AloBhi, DAG.getConstant(32, MVT::i32));
7550 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7551 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7552 AhiBlo, DAG.getConstant(32, MVT::i32));
7553 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7554 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7558 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
7559 EVT VT = Op.getValueType();
7560 DebugLoc dl = Op.getDebugLoc();
7561 SDValue R = Op.getOperand(0);
7563 LLVMContext *Context = DAG.getContext();
7565 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
7567 if (VT == MVT::v4i32) {
7568 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7569 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
7570 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
7572 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
7574 std::vector<Constant*> CV(4, CI);
7575 Constant *C = ConstantVector::get(CV);
7576 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7577 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7578 PseudoSourceValue::getConstantPool(), 0,
7581 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
7582 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, Op);
7583 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
7584 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
7586 if (VT == MVT::v16i8) {
7588 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7589 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
7590 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
7592 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
7593 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
7595 std::vector<Constant*> CVM1(16, CM1);
7596 std::vector<Constant*> CVM2(16, CM2);
7597 Constant *C = ConstantVector::get(CVM1);
7598 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7599 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7600 PseudoSourceValue::getConstantPool(), 0,
7603 // r = pblendv(r, psllw(r & (char16)15, 4), a);
7604 M = DAG.getNode(ISD::AND, dl, VT, R, M);
7605 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7606 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
7607 DAG.getConstant(4, MVT::i32));
7608 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7609 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
7612 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
7614 C = ConstantVector::get(CVM2);
7615 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7616 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7617 PseudoSourceValue::getConstantPool(), 0, false, false, 16);
7619 // r = pblendv(r, psllw(r & (char16)63, 2), a);
7620 M = DAG.getNode(ISD::AND, dl, VT, R, M);
7621 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7622 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
7623 DAG.getConstant(2, MVT::i32));
7624 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7625 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
7628 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
7630 // return pblendv(r, r+r, a);
7631 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7632 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
7633 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
7639 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7640 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7641 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7642 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7643 // has only one use.
7644 SDNode *N = Op.getNode();
7645 SDValue LHS = N->getOperand(0);
7646 SDValue RHS = N->getOperand(1);
7647 unsigned BaseOp = 0;
7649 DebugLoc dl = Op.getDebugLoc();
7651 switch (Op.getOpcode()) {
7652 default: llvm_unreachable("Unknown ovf instruction!");
7654 // A subtract of one will be selected as a INC. Note that INC doesn't
7655 // set CF, so we can't do this for UADDO.
7656 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7657 if (C->getAPIntValue() == 1) {
7658 BaseOp = X86ISD::INC;
7662 BaseOp = X86ISD::ADD;
7666 BaseOp = X86ISD::ADD;
7670 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7671 // set CF, so we can't do this for USUBO.
7672 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7673 if (C->getAPIntValue() == 1) {
7674 BaseOp = X86ISD::DEC;
7678 BaseOp = X86ISD::SUB;
7682 BaseOp = X86ISD::SUB;
7686 BaseOp = X86ISD::SMUL;
7690 BaseOp = X86ISD::UMUL;
7695 // Also sets EFLAGS.
7696 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7697 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7700 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7701 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7703 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7707 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
7708 DebugLoc dl = Op.getDebugLoc();
7710 if (!Subtarget->hasSSE2()) {
7711 SDValue Chain = Op.getOperand(0);
7712 SDValue Zero = DAG.getConstant(0,
7713 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7715 DAG.getRegister(X86::ESP, MVT::i32), // Base
7716 DAG.getTargetConstant(1, MVT::i8), // Scale
7717 DAG.getRegister(0, MVT::i32), // Index
7718 DAG.getTargetConstant(0, MVT::i32), // Disp
7719 DAG.getRegister(0, MVT::i32), // Segment.
7724 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
7725 array_lengthof(Ops));
7726 return SDValue(Res, 0);
7729 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
7731 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
7733 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7734 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
7735 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
7736 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
7738 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
7739 if (!Op1 && !Op2 && !Op3 && Op4)
7740 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
7742 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
7743 if (Op1 && !Op2 && !Op3 && !Op4)
7744 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
7746 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
7748 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
7751 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7752 EVT T = Op.getValueType();
7753 DebugLoc dl = Op.getDebugLoc();
7756 switch(T.getSimpleVT().SimpleTy) {
7758 assert(false && "Invalid value type!");
7759 case MVT::i8: Reg = X86::AL; size = 1; break;
7760 case MVT::i16: Reg = X86::AX; size = 2; break;
7761 case MVT::i32: Reg = X86::EAX; size = 4; break;
7763 assert(Subtarget->is64Bit() && "Node not type legal!");
7764 Reg = X86::RAX; size = 8;
7767 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7768 Op.getOperand(2), SDValue());
7769 SDValue Ops[] = { cpIn.getValue(0),
7772 DAG.getTargetConstant(size, MVT::i8),
7774 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7775 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7777 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7781 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7782 SelectionDAG &DAG) const {
7783 assert(Subtarget->is64Bit() && "Result not type legalized?");
7784 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7785 SDValue TheChain = Op.getOperand(0);
7786 DebugLoc dl = Op.getDebugLoc();
7787 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7788 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7789 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7791 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7792 DAG.getConstant(32, MVT::i8));
7794 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7797 return DAG.getMergeValues(Ops, 2, dl);
7800 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7801 SelectionDAG &DAG) const {
7802 EVT SrcVT = Op.getOperand(0).getValueType();
7803 EVT DstVT = Op.getValueType();
7804 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7805 Subtarget->hasMMX() && !DisableMMX) &&
7806 "Unexpected custom BIT_CONVERT");
7807 assert((DstVT == MVT::i64 ||
7808 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7809 "Unexpected custom BIT_CONVERT");
7810 // i64 <=> MMX conversions are Legal.
7811 if (SrcVT==MVT::i64 && DstVT.isVector())
7813 if (DstVT==MVT::i64 && SrcVT.isVector())
7815 // MMX <=> MMX conversions are Legal.
7816 if (SrcVT.isVector() && DstVT.isVector())
7818 // All other conversions need to be expanded.
7821 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7822 SDNode *Node = Op.getNode();
7823 DebugLoc dl = Node->getDebugLoc();
7824 EVT T = Node->getValueType(0);
7825 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7826 DAG.getConstant(0, T), Node->getOperand(2));
7827 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7828 cast<AtomicSDNode>(Node)->getMemoryVT(),
7829 Node->getOperand(0),
7830 Node->getOperand(1), negOp,
7831 cast<AtomicSDNode>(Node)->getSrcValue(),
7832 cast<AtomicSDNode>(Node)->getAlignment());
7835 /// LowerOperation - Provide custom lowering hooks for some operations.
7837 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7838 switch (Op.getOpcode()) {
7839 default: llvm_unreachable("Should not custom lower this!");
7840 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
7841 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7842 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7843 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7844 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7845 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7846 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7847 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7848 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7849 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7850 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7851 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7852 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7853 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7854 case ISD::SHL_PARTS:
7855 case ISD::SRA_PARTS:
7856 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7857 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7858 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7859 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7860 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7861 case ISD::FABS: return LowerFABS(Op, DAG);
7862 case ISD::FNEG: return LowerFNEG(Op, DAG);
7863 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7864 case ISD::SETCC: return LowerSETCC(Op, DAG);
7865 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7866 case ISD::SELECT: return LowerSELECT(Op, DAG);
7867 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7868 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7869 case ISD::VASTART: return LowerVASTART(Op, DAG);
7870 case ISD::VAARG: return LowerVAARG(Op, DAG);
7871 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7872 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7873 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7874 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7875 case ISD::FRAME_TO_ARGS_OFFSET:
7876 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7877 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7878 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7879 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7880 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7881 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7882 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7883 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7884 case ISD::SHL: return LowerSHL(Op, DAG);
7890 case ISD::UMULO: return LowerXALUO(Op, DAG);
7891 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7892 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7896 void X86TargetLowering::
7897 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7898 SelectionDAG &DAG, unsigned NewOp) const {
7899 EVT T = Node->getValueType(0);
7900 DebugLoc dl = Node->getDebugLoc();
7901 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7903 SDValue Chain = Node->getOperand(0);
7904 SDValue In1 = Node->getOperand(1);
7905 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7906 Node->getOperand(2), DAG.getIntPtrConstant(0));
7907 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7908 Node->getOperand(2), DAG.getIntPtrConstant(1));
7909 SDValue Ops[] = { Chain, In1, In2L, In2H };
7910 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7912 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7913 cast<MemSDNode>(Node)->getMemOperand());
7914 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7915 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7916 Results.push_back(Result.getValue(2));
7919 /// ReplaceNodeResults - Replace a node with an illegal result type
7920 /// with a new node built out of custom code.
7921 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7922 SmallVectorImpl<SDValue>&Results,
7923 SelectionDAG &DAG) const {
7924 DebugLoc dl = N->getDebugLoc();
7925 switch (N->getOpcode()) {
7927 assert(false && "Do not know how to custom type legalize this operation!");
7929 case ISD::FP_TO_SINT: {
7930 std::pair<SDValue,SDValue> Vals =
7931 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7932 SDValue FIST = Vals.first, StackSlot = Vals.second;
7933 if (FIST.getNode() != 0) {
7934 EVT VT = N->getValueType(0);
7935 // Return a load from the stack slot.
7936 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7941 case ISD::READCYCLECOUNTER: {
7942 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7943 SDValue TheChain = N->getOperand(0);
7944 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7945 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7947 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7949 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7950 SDValue Ops[] = { eax, edx };
7951 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7952 Results.push_back(edx.getValue(1));
7955 case ISD::ATOMIC_CMP_SWAP: {
7956 EVT T = N->getValueType(0);
7957 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7958 SDValue cpInL, cpInH;
7959 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7960 DAG.getConstant(0, MVT::i32));
7961 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7962 DAG.getConstant(1, MVT::i32));
7963 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7964 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7966 SDValue swapInL, swapInH;
7967 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7968 DAG.getConstant(0, MVT::i32));
7969 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7970 DAG.getConstant(1, MVT::i32));
7971 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7973 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7974 swapInL.getValue(1));
7975 SDValue Ops[] = { swapInH.getValue(0),
7977 swapInH.getValue(1) };
7978 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7979 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7980 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7981 MVT::i32, Result.getValue(1));
7982 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7983 MVT::i32, cpOutL.getValue(2));
7984 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7985 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7986 Results.push_back(cpOutH.getValue(1));
7989 case ISD::ATOMIC_LOAD_ADD:
7990 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7992 case ISD::ATOMIC_LOAD_AND:
7993 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7995 case ISD::ATOMIC_LOAD_NAND:
7996 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7998 case ISD::ATOMIC_LOAD_OR:
7999 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
8001 case ISD::ATOMIC_LOAD_SUB:
8002 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
8004 case ISD::ATOMIC_LOAD_XOR:
8005 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
8007 case ISD::ATOMIC_SWAP:
8008 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
8013 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
8015 default: return NULL;
8016 case X86ISD::BSF: return "X86ISD::BSF";
8017 case X86ISD::BSR: return "X86ISD::BSR";
8018 case X86ISD::SHLD: return "X86ISD::SHLD";
8019 case X86ISD::SHRD: return "X86ISD::SHRD";
8020 case X86ISD::FAND: return "X86ISD::FAND";
8021 case X86ISD::FOR: return "X86ISD::FOR";
8022 case X86ISD::FXOR: return "X86ISD::FXOR";
8023 case X86ISD::FSRL: return "X86ISD::FSRL";
8024 case X86ISD::FILD: return "X86ISD::FILD";
8025 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
8026 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
8027 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
8028 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
8029 case X86ISD::FLD: return "X86ISD::FLD";
8030 case X86ISD::FST: return "X86ISD::FST";
8031 case X86ISD::CALL: return "X86ISD::CALL";
8032 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
8033 case X86ISD::BT: return "X86ISD::BT";
8034 case X86ISD::CMP: return "X86ISD::CMP";
8035 case X86ISD::COMI: return "X86ISD::COMI";
8036 case X86ISD::UCOMI: return "X86ISD::UCOMI";
8037 case X86ISD::SETCC: return "X86ISD::SETCC";
8038 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
8039 case X86ISD::CMOV: return "X86ISD::CMOV";
8040 case X86ISD::BRCOND: return "X86ISD::BRCOND";
8041 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
8042 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
8043 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
8044 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
8045 case X86ISD::Wrapper: return "X86ISD::Wrapper";
8046 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
8047 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
8048 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
8049 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
8050 case X86ISD::PINSRB: return "X86ISD::PINSRB";
8051 case X86ISD::PINSRW: return "X86ISD::PINSRW";
8052 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
8053 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
8054 case X86ISD::FMAX: return "X86ISD::FMAX";
8055 case X86ISD::FMIN: return "X86ISD::FMIN";
8056 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
8057 case X86ISD::FRCP: return "X86ISD::FRCP";
8058 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
8059 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
8060 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
8061 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
8062 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
8063 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
8064 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
8065 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
8066 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
8067 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
8068 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
8069 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
8070 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
8071 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
8072 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
8073 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
8074 case X86ISD::VSHL: return "X86ISD::VSHL";
8075 case X86ISD::VSRL: return "X86ISD::VSRL";
8076 case X86ISD::CMPPD: return "X86ISD::CMPPD";
8077 case X86ISD::CMPPS: return "X86ISD::CMPPS";
8078 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
8079 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
8080 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
8081 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
8082 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
8083 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
8084 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
8085 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
8086 case X86ISD::ADD: return "X86ISD::ADD";
8087 case X86ISD::SUB: return "X86ISD::SUB";
8088 case X86ISD::SMUL: return "X86ISD::SMUL";
8089 case X86ISD::UMUL: return "X86ISD::UMUL";
8090 case X86ISD::INC: return "X86ISD::INC";
8091 case X86ISD::DEC: return "X86ISD::DEC";
8092 case X86ISD::OR: return "X86ISD::OR";
8093 case X86ISD::XOR: return "X86ISD::XOR";
8094 case X86ISD::AND: return "X86ISD::AND";
8095 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
8096 case X86ISD::PTEST: return "X86ISD::PTEST";
8097 case X86ISD::TESTP: return "X86ISD::TESTP";
8098 case X86ISD::PALIGN: return "X86ISD::PALIGN";
8099 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
8100 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
8101 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
8102 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
8103 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
8104 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
8105 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
8106 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
8107 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
8108 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
8109 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
8110 case X86ISD::MOVHPS: return "X86ISD::MOVHPS";
8111 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
8112 case X86ISD::MOVHPD: return "X86ISD::MOVHPD";
8113 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
8114 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
8115 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
8116 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
8117 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
8118 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
8119 case X86ISD::MOVSD: return "X86ISD::MOVSD";
8120 case X86ISD::MOVSS: return "X86ISD::MOVSS";
8121 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
8122 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
8123 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
8124 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
8125 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
8126 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
8127 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
8128 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
8129 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
8130 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
8131 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
8132 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
8133 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
8134 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
8138 // isLegalAddressingMode - Return true if the addressing mode represented
8139 // by AM is legal for this target, for a load/store of the specified type.
8140 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
8141 const Type *Ty) const {
8142 // X86 supports extremely general addressing modes.
8143 CodeModel::Model M = getTargetMachine().getCodeModel();
8145 // X86 allows a sign-extended 32-bit immediate field as a displacement.
8146 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
8151 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
8153 // If a reference to this global requires an extra load, we can't fold it.
8154 if (isGlobalStubReference(GVFlags))
8157 // If BaseGV requires a register for the PIC base, we cannot also have a
8158 // BaseReg specified.
8159 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
8162 // If lower 4G is not available, then we must use rip-relative addressing.
8163 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
8173 // These scales always work.
8178 // These scales are formed with basereg+scalereg. Only accept if there is
8183 default: // Other stuff never works.
8191 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
8192 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
8194 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
8195 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
8196 if (NumBits1 <= NumBits2)
8201 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
8202 if (!VT1.isInteger() || !VT2.isInteger())
8204 unsigned NumBits1 = VT1.getSizeInBits();
8205 unsigned NumBits2 = VT2.getSizeInBits();
8206 if (NumBits1 <= NumBits2)
8211 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
8212 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8213 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
8216 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
8217 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8218 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
8221 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
8222 // i16 instructions are longer (0x66 prefix) and potentially slower.
8223 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
8226 /// isShuffleMaskLegal - Targets can use this to indicate that they only
8227 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
8228 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
8229 /// are assumed to be legal.
8231 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
8233 // Very little shuffling can be done for 64-bit vectors right now.
8234 if (VT.getSizeInBits() == 64)
8235 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8237 // FIXME: pshufb, blends, shifts.
8238 return (VT.getVectorNumElements() == 2 ||
8239 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8240 isMOVLMask(M, VT) ||
8241 isSHUFPMask(M, VT) ||
8242 isPSHUFDMask(M, VT) ||
8243 isPSHUFHWMask(M, VT) ||
8244 isPSHUFLWMask(M, VT) ||
8245 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8246 isUNPCKLMask(M, VT) ||
8247 isUNPCKHMask(M, VT) ||
8248 isUNPCKL_v_undef_Mask(M, VT) ||
8249 isUNPCKH_v_undef_Mask(M, VT));
8253 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8255 unsigned NumElts = VT.getVectorNumElements();
8256 // FIXME: This collection of masks seems suspect.
8259 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8260 return (isMOVLMask(Mask, VT) ||
8261 isCommutedMOVLMask(Mask, VT, true) ||
8262 isSHUFPMask(Mask, VT) ||
8263 isCommutedSHUFPMask(Mask, VT));
8268 //===----------------------------------------------------------------------===//
8269 // X86 Scheduler Hooks
8270 //===----------------------------------------------------------------------===//
8272 // private utility function
8274 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8275 MachineBasicBlock *MBB,
8282 TargetRegisterClass *RC,
8283 bool invSrc) const {
8284 // For the atomic bitwise operator, we generate
8287 // ld t1 = [bitinstr.addr]
8288 // op t2 = t1, [bitinstr.val]
8290 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8292 // fallthrough -->nextMBB
8293 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8294 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8295 MachineFunction::iterator MBBIter = MBB;
8298 /// First build the CFG
8299 MachineFunction *F = MBB->getParent();
8300 MachineBasicBlock *thisMBB = MBB;
8301 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8302 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8303 F->insert(MBBIter, newMBB);
8304 F->insert(MBBIter, nextMBB);
8306 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8307 nextMBB->splice(nextMBB->begin(), thisMBB,
8308 llvm::next(MachineBasicBlock::iterator(bInstr)),
8310 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8312 // Update thisMBB to fall through to newMBB
8313 thisMBB->addSuccessor(newMBB);
8315 // newMBB jumps to itself and fall through to nextMBB
8316 newMBB->addSuccessor(nextMBB);
8317 newMBB->addSuccessor(newMBB);
8319 // Insert instructions into newMBB based on incoming instruction
8320 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8321 "unexpected number of operands");
8322 DebugLoc dl = bInstr->getDebugLoc();
8323 MachineOperand& destOper = bInstr->getOperand(0);
8324 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8325 int numArgs = bInstr->getNumOperands() - 1;
8326 for (int i=0; i < numArgs; ++i)
8327 argOpers[i] = &bInstr->getOperand(i+1);
8329 // x86 address has 4 operands: base, index, scale, and displacement
8330 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8331 int valArgIndx = lastAddrIndx + 1;
8333 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8334 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8335 for (int i=0; i <= lastAddrIndx; ++i)
8336 (*MIB).addOperand(*argOpers[i]);
8338 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8340 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8345 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8346 assert((argOpers[valArgIndx]->isReg() ||
8347 argOpers[valArgIndx]->isImm()) &&
8349 if (argOpers[valArgIndx]->isReg())
8350 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8352 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8354 (*MIB).addOperand(*argOpers[valArgIndx]);
8356 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
8359 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8360 for (int i=0; i <= lastAddrIndx; ++i)
8361 (*MIB).addOperand(*argOpers[i]);
8363 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8364 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8365 bInstr->memoperands_end());
8367 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8371 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8373 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8377 // private utility function: 64 bit atomics on 32 bit host.
8379 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8380 MachineBasicBlock *MBB,
8385 bool invSrc) const {
8386 // For the atomic bitwise operator, we generate
8387 // thisMBB (instructions are in pairs, except cmpxchg8b)
8388 // ld t1,t2 = [bitinstr.addr]
8390 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8391 // op t5, t6 <- out1, out2, [bitinstr.val]
8392 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8393 // mov ECX, EBX <- t5, t6
8394 // mov EAX, EDX <- t1, t2
8395 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8396 // mov t3, t4 <- EAX, EDX
8398 // result in out1, out2
8399 // fallthrough -->nextMBB
8401 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8402 const unsigned LoadOpc = X86::MOV32rm;
8403 const unsigned NotOpc = X86::NOT32r;
8404 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8405 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8406 MachineFunction::iterator MBBIter = MBB;
8409 /// First build the CFG
8410 MachineFunction *F = MBB->getParent();
8411 MachineBasicBlock *thisMBB = MBB;
8412 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8413 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8414 F->insert(MBBIter, newMBB);
8415 F->insert(MBBIter, nextMBB);
8417 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8418 nextMBB->splice(nextMBB->begin(), thisMBB,
8419 llvm::next(MachineBasicBlock::iterator(bInstr)),
8421 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8423 // Update thisMBB to fall through to newMBB
8424 thisMBB->addSuccessor(newMBB);
8426 // newMBB jumps to itself and fall through to nextMBB
8427 newMBB->addSuccessor(nextMBB);
8428 newMBB->addSuccessor(newMBB);
8430 DebugLoc dl = bInstr->getDebugLoc();
8431 // Insert instructions into newMBB based on incoming instruction
8432 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8433 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
8434 "unexpected number of operands");
8435 MachineOperand& dest1Oper = bInstr->getOperand(0);
8436 MachineOperand& dest2Oper = bInstr->getOperand(1);
8437 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8438 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
8439 argOpers[i] = &bInstr->getOperand(i+2);
8441 // We use some of the operands multiple times, so conservatively just
8442 // clear any kill flags that might be present.
8443 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8444 argOpers[i]->setIsKill(false);
8447 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8448 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8450 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8451 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8452 for (int i=0; i <= lastAddrIndx; ++i)
8453 (*MIB).addOperand(*argOpers[i]);
8454 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8455 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8456 // add 4 to displacement.
8457 for (int i=0; i <= lastAddrIndx-2; ++i)
8458 (*MIB).addOperand(*argOpers[i]);
8459 MachineOperand newOp3 = *(argOpers[3]);
8461 newOp3.setImm(newOp3.getImm()+4);
8463 newOp3.setOffset(newOp3.getOffset()+4);
8464 (*MIB).addOperand(newOp3);
8465 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8467 // t3/4 are defined later, at the bottom of the loop
8468 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8469 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8470 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8471 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8472 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8473 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8475 // The subsequent operations should be using the destination registers of
8476 //the PHI instructions.
8478 t1 = F->getRegInfo().createVirtualRegister(RC);
8479 t2 = F->getRegInfo().createVirtualRegister(RC);
8480 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8481 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8483 t1 = dest1Oper.getReg();
8484 t2 = dest2Oper.getReg();
8487 int valArgIndx = lastAddrIndx + 1;
8488 assert((argOpers[valArgIndx]->isReg() ||
8489 argOpers[valArgIndx]->isImm()) &&
8491 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8492 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8493 if (argOpers[valArgIndx]->isReg())
8494 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8496 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8497 if (regOpcL != X86::MOV32rr)
8499 (*MIB).addOperand(*argOpers[valArgIndx]);
8500 assert(argOpers[valArgIndx + 1]->isReg() ==
8501 argOpers[valArgIndx]->isReg());
8502 assert(argOpers[valArgIndx + 1]->isImm() ==
8503 argOpers[valArgIndx]->isImm());
8504 if (argOpers[valArgIndx + 1]->isReg())
8505 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8507 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8508 if (regOpcH != X86::MOV32rr)
8510 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8512 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8514 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
8517 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
8519 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
8522 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8523 for (int i=0; i <= lastAddrIndx; ++i)
8524 (*MIB).addOperand(*argOpers[i]);
8526 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8527 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8528 bInstr->memoperands_end());
8530 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
8531 MIB.addReg(X86::EAX);
8532 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
8533 MIB.addReg(X86::EDX);
8536 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8538 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8542 // private utility function
8544 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8545 MachineBasicBlock *MBB,
8546 unsigned cmovOpc) const {
8547 // For the atomic min/max operator, we generate
8550 // ld t1 = [min/max.addr]
8551 // mov t2 = [min/max.val]
8553 // cmov[cond] t2 = t1
8555 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8557 // fallthrough -->nextMBB
8559 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8560 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8561 MachineFunction::iterator MBBIter = MBB;
8564 /// First build the CFG
8565 MachineFunction *F = MBB->getParent();
8566 MachineBasicBlock *thisMBB = MBB;
8567 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8568 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8569 F->insert(MBBIter, newMBB);
8570 F->insert(MBBIter, nextMBB);
8572 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8573 nextMBB->splice(nextMBB->begin(), thisMBB,
8574 llvm::next(MachineBasicBlock::iterator(mInstr)),
8576 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8578 // Update thisMBB to fall through to newMBB
8579 thisMBB->addSuccessor(newMBB);
8581 // newMBB jumps to newMBB and fall through to nextMBB
8582 newMBB->addSuccessor(nextMBB);
8583 newMBB->addSuccessor(newMBB);
8585 DebugLoc dl = mInstr->getDebugLoc();
8586 // Insert instructions into newMBB based on incoming instruction
8587 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8588 "unexpected number of operands");
8589 MachineOperand& destOper = mInstr->getOperand(0);
8590 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8591 int numArgs = mInstr->getNumOperands() - 1;
8592 for (int i=0; i < numArgs; ++i)
8593 argOpers[i] = &mInstr->getOperand(i+1);
8595 // x86 address has 4 operands: base, index, scale, and displacement
8596 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8597 int valArgIndx = lastAddrIndx + 1;
8599 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8600 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8601 for (int i=0; i <= lastAddrIndx; ++i)
8602 (*MIB).addOperand(*argOpers[i]);
8604 // We only support register and immediate values
8605 assert((argOpers[valArgIndx]->isReg() ||
8606 argOpers[valArgIndx]->isImm()) &&
8609 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8610 if (argOpers[valArgIndx]->isReg())
8611 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
8613 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8614 (*MIB).addOperand(*argOpers[valArgIndx]);
8616 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8619 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8624 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8625 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8629 // Cmp and exchange if none has modified the memory location
8630 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8631 for (int i=0; i <= lastAddrIndx; ++i)
8632 (*MIB).addOperand(*argOpers[i]);
8634 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8635 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8636 mInstr->memoperands_end());
8638 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8639 MIB.addReg(X86::EAX);
8642 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8644 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
8648 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8649 // or XMM0_V32I8 in AVX all of this code can be replaced with that
8652 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8653 unsigned numArgs, bool memArg) const {
8655 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
8656 "Target must have SSE4.2 or AVX features enabled");
8658 DebugLoc dl = MI->getDebugLoc();
8659 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8663 if (!Subtarget->hasAVX()) {
8665 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8667 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8670 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
8672 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
8675 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8677 for (unsigned i = 0; i < numArgs; ++i) {
8678 MachineOperand &Op = MI->getOperand(i+1);
8680 if (!(Op.isReg() && Op.isImplicit()))
8684 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8687 MI->eraseFromParent();
8693 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8695 MachineBasicBlock *MBB) const {
8696 // Emit code to save XMM registers to the stack. The ABI says that the
8697 // number of registers to save is given in %al, so it's theoretically
8698 // possible to do an indirect jump trick to avoid saving all of them,
8699 // however this code takes a simpler approach and just executes all
8700 // of the stores if %al is non-zero. It's less code, and it's probably
8701 // easier on the hardware branch predictor, and stores aren't all that
8702 // expensive anyway.
8704 // Create the new basic blocks. One block contains all the XMM stores,
8705 // and one block is the final destination regardless of whether any
8706 // stores were performed.
8707 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8708 MachineFunction *F = MBB->getParent();
8709 MachineFunction::iterator MBBIter = MBB;
8711 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8712 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8713 F->insert(MBBIter, XMMSaveMBB);
8714 F->insert(MBBIter, EndMBB);
8716 // Transfer the remainder of MBB and its successor edges to EndMBB.
8717 EndMBB->splice(EndMBB->begin(), MBB,
8718 llvm::next(MachineBasicBlock::iterator(MI)),
8720 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
8722 // The original block will now fall through to the XMM save block.
8723 MBB->addSuccessor(XMMSaveMBB);
8724 // The XMMSaveMBB will fall through to the end block.
8725 XMMSaveMBB->addSuccessor(EndMBB);
8727 // Now add the instructions.
8728 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8729 DebugLoc DL = MI->getDebugLoc();
8731 unsigned CountReg = MI->getOperand(0).getReg();
8732 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8733 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8735 if (!Subtarget->isTargetWin64()) {
8736 // If %al is 0, branch around the XMM save block.
8737 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8738 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8739 MBB->addSuccessor(EndMBB);
8742 // In the XMM save block, save all the XMM argument registers.
8743 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8744 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8745 MachineMemOperand *MMO =
8746 F->getMachineMemOperand(
8747 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8748 MachineMemOperand::MOStore, Offset,
8749 /*Size=*/16, /*Align=*/16);
8750 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8751 .addFrameIndex(RegSaveFrameIndex)
8752 .addImm(/*Scale=*/1)
8753 .addReg(/*IndexReg=*/0)
8754 .addImm(/*Disp=*/Offset)
8755 .addReg(/*Segment=*/0)
8756 .addReg(MI->getOperand(i).getReg())
8757 .addMemOperand(MMO);
8760 MI->eraseFromParent(); // The pseudo instruction is gone now.
8766 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8767 MachineBasicBlock *BB) const {
8768 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8769 DebugLoc DL = MI->getDebugLoc();
8771 // To "insert" a SELECT_CC instruction, we actually have to insert the
8772 // diamond control-flow pattern. The incoming instruction knows the
8773 // destination vreg to set, the condition code register to branch on, the
8774 // true/false values to select between, and a branch opcode to use.
8775 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8776 MachineFunction::iterator It = BB;
8782 // cmpTY ccX, r1, r2
8784 // fallthrough --> copy0MBB
8785 MachineBasicBlock *thisMBB = BB;
8786 MachineFunction *F = BB->getParent();
8787 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8788 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8789 F->insert(It, copy0MBB);
8790 F->insert(It, sinkMBB);
8792 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8793 // live into the sink and copy blocks.
8794 const MachineFunction *MF = BB->getParent();
8795 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8796 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8798 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
8799 const MachineOperand &MO = MI->getOperand(I);
8800 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
8801 unsigned Reg = MO.getReg();
8802 if (Reg != X86::EFLAGS) continue;
8803 copy0MBB->addLiveIn(Reg);
8804 sinkMBB->addLiveIn(Reg);
8807 // Transfer the remainder of BB and its successor edges to sinkMBB.
8808 sinkMBB->splice(sinkMBB->begin(), BB,
8809 llvm::next(MachineBasicBlock::iterator(MI)),
8811 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
8813 // Add the true and fallthrough blocks as its successors.
8814 BB->addSuccessor(copy0MBB);
8815 BB->addSuccessor(sinkMBB);
8817 // Create the conditional branch instruction.
8819 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8820 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8823 // %FalseValue = ...
8824 // # fallthrough to sinkMBB
8825 copy0MBB->addSuccessor(sinkMBB);
8828 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8830 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
8831 TII->get(X86::PHI), MI->getOperand(0).getReg())
8832 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8833 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8835 MI->eraseFromParent(); // The pseudo instruction is gone now.
8840 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8841 MachineBasicBlock *BB) const {
8842 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8843 DebugLoc DL = MI->getDebugLoc();
8845 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8846 // non-trivial part is impdef of ESP.
8847 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8850 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
8851 .addExternalSymbol("_alloca")
8852 .addReg(X86::EAX, RegState::Implicit)
8853 .addReg(X86::ESP, RegState::Implicit)
8854 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8855 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8857 MI->eraseFromParent(); // The pseudo instruction is gone now.
8862 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8863 MachineBasicBlock *BB) const {
8864 // This is pretty easy. We're taking the value that we received from
8865 // our load from the relocation, sticking it in either RDI (x86-64)
8866 // or EAX and doing an indirect call. The return value will then
8867 // be in the normal return register.
8868 const X86InstrInfo *TII
8869 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8870 DebugLoc DL = MI->getDebugLoc();
8871 MachineFunction *F = BB->getParent();
8872 bool IsWin64 = Subtarget->isTargetWin64();
8874 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8876 if (Subtarget->is64Bit()) {
8877 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8878 TII->get(X86::MOV64rm), X86::RDI)
8880 .addImm(0).addReg(0)
8881 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8882 MI->getOperand(3).getTargetFlags())
8884 MIB = BuildMI(*BB, MI, DL, TII->get(IsWin64 ? X86::WINCALL64m : X86::CALL64m));
8885 addDirectMem(MIB, X86::RDI);
8886 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8887 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8888 TII->get(X86::MOV32rm), X86::EAX)
8890 .addImm(0).addReg(0)
8891 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8892 MI->getOperand(3).getTargetFlags())
8894 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8895 addDirectMem(MIB, X86::EAX);
8897 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8898 TII->get(X86::MOV32rm), X86::EAX)
8899 .addReg(TII->getGlobalBaseReg(F))
8900 .addImm(0).addReg(0)
8901 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8902 MI->getOperand(3).getTargetFlags())
8904 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8905 addDirectMem(MIB, X86::EAX);
8908 MI->eraseFromParent(); // The pseudo instruction is gone now.
8913 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8914 MachineBasicBlock *BB) const {
8915 switch (MI->getOpcode()) {
8916 default: assert(false && "Unexpected instr type to insert");
8917 case X86::MINGW_ALLOCA:
8918 return EmitLoweredMingwAlloca(MI, BB);
8919 case X86::TLSCall_32:
8920 case X86::TLSCall_64:
8921 return EmitLoweredTLSCall(MI, BB);
8923 case X86::CMOV_V1I64:
8924 case X86::CMOV_FR32:
8925 case X86::CMOV_FR64:
8926 case X86::CMOV_V4F32:
8927 case X86::CMOV_V2F64:
8928 case X86::CMOV_V2I64:
8929 case X86::CMOV_GR16:
8930 case X86::CMOV_GR32:
8931 case X86::CMOV_RFP32:
8932 case X86::CMOV_RFP64:
8933 case X86::CMOV_RFP80:
8934 return EmitLoweredSelect(MI, BB);
8936 case X86::FP32_TO_INT16_IN_MEM:
8937 case X86::FP32_TO_INT32_IN_MEM:
8938 case X86::FP32_TO_INT64_IN_MEM:
8939 case X86::FP64_TO_INT16_IN_MEM:
8940 case X86::FP64_TO_INT32_IN_MEM:
8941 case X86::FP64_TO_INT64_IN_MEM:
8942 case X86::FP80_TO_INT16_IN_MEM:
8943 case X86::FP80_TO_INT32_IN_MEM:
8944 case X86::FP80_TO_INT64_IN_MEM: {
8945 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8946 DebugLoc DL = MI->getDebugLoc();
8948 // Change the floating point control register to use "round towards zero"
8949 // mode when truncating to an integer value.
8950 MachineFunction *F = BB->getParent();
8951 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8952 addFrameReference(BuildMI(*BB, MI, DL,
8953 TII->get(X86::FNSTCW16m)), CWFrameIdx);
8955 // Load the old value of the high byte of the control word...
8957 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8958 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
8961 // Set the high part to be round to zero...
8962 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8965 // Reload the modified control word now...
8966 addFrameReference(BuildMI(*BB, MI, DL,
8967 TII->get(X86::FLDCW16m)), CWFrameIdx);
8969 // Restore the memory image of control word to original value
8970 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8973 // Get the X86 opcode to use.
8975 switch (MI->getOpcode()) {
8976 default: llvm_unreachable("illegal opcode!");
8977 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8978 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8979 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8980 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8981 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8982 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8983 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8984 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8985 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8989 MachineOperand &Op = MI->getOperand(0);
8991 AM.BaseType = X86AddressMode::RegBase;
8992 AM.Base.Reg = Op.getReg();
8994 AM.BaseType = X86AddressMode::FrameIndexBase;
8995 AM.Base.FrameIndex = Op.getIndex();
8997 Op = MI->getOperand(1);
8999 AM.Scale = Op.getImm();
9000 Op = MI->getOperand(2);
9002 AM.IndexReg = Op.getImm();
9003 Op = MI->getOperand(3);
9004 if (Op.isGlobal()) {
9005 AM.GV = Op.getGlobal();
9007 AM.Disp = Op.getImm();
9009 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
9010 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
9012 // Reload the original control word now.
9013 addFrameReference(BuildMI(*BB, MI, DL,
9014 TII->get(X86::FLDCW16m)), CWFrameIdx);
9016 MI->eraseFromParent(); // The pseudo instruction is gone now.
9019 // String/text processing lowering.
9020 case X86::PCMPISTRM128REG:
9021 case X86::VPCMPISTRM128REG:
9022 return EmitPCMP(MI, BB, 3, false /* in-mem */);
9023 case X86::PCMPISTRM128MEM:
9024 case X86::VPCMPISTRM128MEM:
9025 return EmitPCMP(MI, BB, 3, true /* in-mem */);
9026 case X86::PCMPESTRM128REG:
9027 case X86::VPCMPESTRM128REG:
9028 return EmitPCMP(MI, BB, 5, false /* in mem */);
9029 case X86::PCMPESTRM128MEM:
9030 case X86::VPCMPESTRM128MEM:
9031 return EmitPCMP(MI, BB, 5, true /* in mem */);
9034 case X86::ATOMAND32:
9035 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9036 X86::AND32ri, X86::MOV32rm,
9038 X86::NOT32r, X86::EAX,
9039 X86::GR32RegisterClass);
9041 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
9042 X86::OR32ri, X86::MOV32rm,
9044 X86::NOT32r, X86::EAX,
9045 X86::GR32RegisterClass);
9046 case X86::ATOMXOR32:
9047 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
9048 X86::XOR32ri, X86::MOV32rm,
9050 X86::NOT32r, X86::EAX,
9051 X86::GR32RegisterClass);
9052 case X86::ATOMNAND32:
9053 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9054 X86::AND32ri, X86::MOV32rm,
9056 X86::NOT32r, X86::EAX,
9057 X86::GR32RegisterClass, true);
9058 case X86::ATOMMIN32:
9059 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
9060 case X86::ATOMMAX32:
9061 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
9062 case X86::ATOMUMIN32:
9063 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
9064 case X86::ATOMUMAX32:
9065 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
9067 case X86::ATOMAND16:
9068 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9069 X86::AND16ri, X86::MOV16rm,
9071 X86::NOT16r, X86::AX,
9072 X86::GR16RegisterClass);
9074 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
9075 X86::OR16ri, X86::MOV16rm,
9077 X86::NOT16r, X86::AX,
9078 X86::GR16RegisterClass);
9079 case X86::ATOMXOR16:
9080 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
9081 X86::XOR16ri, X86::MOV16rm,
9083 X86::NOT16r, X86::AX,
9084 X86::GR16RegisterClass);
9085 case X86::ATOMNAND16:
9086 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9087 X86::AND16ri, X86::MOV16rm,
9089 X86::NOT16r, X86::AX,
9090 X86::GR16RegisterClass, true);
9091 case X86::ATOMMIN16:
9092 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
9093 case X86::ATOMMAX16:
9094 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
9095 case X86::ATOMUMIN16:
9096 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
9097 case X86::ATOMUMAX16:
9098 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
9101 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9102 X86::AND8ri, X86::MOV8rm,
9104 X86::NOT8r, X86::AL,
9105 X86::GR8RegisterClass);
9107 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
9108 X86::OR8ri, X86::MOV8rm,
9110 X86::NOT8r, X86::AL,
9111 X86::GR8RegisterClass);
9113 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
9114 X86::XOR8ri, X86::MOV8rm,
9116 X86::NOT8r, X86::AL,
9117 X86::GR8RegisterClass);
9118 case X86::ATOMNAND8:
9119 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9120 X86::AND8ri, X86::MOV8rm,
9122 X86::NOT8r, X86::AL,
9123 X86::GR8RegisterClass, true);
9124 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
9125 // This group is for 64-bit host.
9126 case X86::ATOMAND64:
9127 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9128 X86::AND64ri32, X86::MOV64rm,
9130 X86::NOT64r, X86::RAX,
9131 X86::GR64RegisterClass);
9133 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
9134 X86::OR64ri32, X86::MOV64rm,
9136 X86::NOT64r, X86::RAX,
9137 X86::GR64RegisterClass);
9138 case X86::ATOMXOR64:
9139 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
9140 X86::XOR64ri32, X86::MOV64rm,
9142 X86::NOT64r, X86::RAX,
9143 X86::GR64RegisterClass);
9144 case X86::ATOMNAND64:
9145 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9146 X86::AND64ri32, X86::MOV64rm,
9148 X86::NOT64r, X86::RAX,
9149 X86::GR64RegisterClass, true);
9150 case X86::ATOMMIN64:
9151 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
9152 case X86::ATOMMAX64:
9153 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
9154 case X86::ATOMUMIN64:
9155 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
9156 case X86::ATOMUMAX64:
9157 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
9159 // This group does 64-bit operations on a 32-bit host.
9160 case X86::ATOMAND6432:
9161 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9162 X86::AND32rr, X86::AND32rr,
9163 X86::AND32ri, X86::AND32ri,
9165 case X86::ATOMOR6432:
9166 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9167 X86::OR32rr, X86::OR32rr,
9168 X86::OR32ri, X86::OR32ri,
9170 case X86::ATOMXOR6432:
9171 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9172 X86::XOR32rr, X86::XOR32rr,
9173 X86::XOR32ri, X86::XOR32ri,
9175 case X86::ATOMNAND6432:
9176 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9177 X86::AND32rr, X86::AND32rr,
9178 X86::AND32ri, X86::AND32ri,
9180 case X86::ATOMADD6432:
9181 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9182 X86::ADD32rr, X86::ADC32rr,
9183 X86::ADD32ri, X86::ADC32ri,
9185 case X86::ATOMSUB6432:
9186 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9187 X86::SUB32rr, X86::SBB32rr,
9188 X86::SUB32ri, X86::SBB32ri,
9190 case X86::ATOMSWAP6432:
9191 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9192 X86::MOV32rr, X86::MOV32rr,
9193 X86::MOV32ri, X86::MOV32ri,
9195 case X86::VASTART_SAVE_XMM_REGS:
9196 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
9200 //===----------------------------------------------------------------------===//
9201 // X86 Optimization Hooks
9202 //===----------------------------------------------------------------------===//
9204 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9208 const SelectionDAG &DAG,
9209 unsigned Depth) const {
9210 unsigned Opc = Op.getOpcode();
9211 assert((Opc >= ISD::BUILTIN_OP_END ||
9212 Opc == ISD::INTRINSIC_WO_CHAIN ||
9213 Opc == ISD::INTRINSIC_W_CHAIN ||
9214 Opc == ISD::INTRINSIC_VOID) &&
9215 "Should use MaskedValueIsZero if you don't know whether Op"
9216 " is a target node!");
9218 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
9230 // These nodes' second result is a boolean.
9231 if (Op.getResNo() == 0)
9235 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
9236 Mask.getBitWidth() - 1);
9241 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
9242 /// node is a GlobalAddress + offset.
9243 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
9244 const GlobalValue* &GA,
9245 int64_t &Offset) const {
9246 if (N->getOpcode() == X86ISD::Wrapper) {
9247 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
9248 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
9249 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
9253 return TargetLowering::isGAPlusOffset(N, GA, Offset);
9256 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
9257 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
9258 /// if the load addresses are consecutive, non-overlapping, and in the right
9260 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
9261 const TargetLowering &TLI) {
9262 DebugLoc dl = N->getDebugLoc();
9263 EVT VT = N->getValueType(0);
9264 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9266 if (VT.getSizeInBits() != 128)
9269 SmallVector<SDValue, 16> Elts;
9270 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
9271 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
9273 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
9276 /// PerformShuffleCombine - Detect vector gather/scatter index generation
9277 /// and convert it from being a bunch of shuffles and extracts to a simple
9278 /// store and scalar loads to extract the elements.
9279 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
9280 const TargetLowering &TLI) {
9281 SDValue InputVector = N->getOperand(0);
9283 // Only operate on vectors of 4 elements, where the alternative shuffling
9284 // gets to be more expensive.
9285 if (InputVector.getValueType() != MVT::v4i32)
9288 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
9289 // single use which is a sign-extend or zero-extend, and all elements are
9291 SmallVector<SDNode *, 4> Uses;
9292 unsigned ExtractedElements = 0;
9293 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
9294 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
9295 if (UI.getUse().getResNo() != InputVector.getResNo())
9298 SDNode *Extract = *UI;
9299 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9302 if (Extract->getValueType(0) != MVT::i32)
9304 if (!Extract->hasOneUse())
9306 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9307 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9309 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9312 // Record which element was extracted.
9313 ExtractedElements |=
9314 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9316 Uses.push_back(Extract);
9319 // If not all the elements were used, this may not be worthwhile.
9320 if (ExtractedElements != 15)
9323 // Ok, we've now decided to do the transformation.
9324 DebugLoc dl = InputVector.getDebugLoc();
9326 // Store the value to a temporary stack slot.
9327 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9328 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL,
9329 0, false, false, 0);
9331 // Replace each use (extract) with a load of the appropriate element.
9332 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9333 UE = Uses.end(); UI != UE; ++UI) {
9334 SDNode *Extract = *UI;
9336 // Compute the element's address.
9337 SDValue Idx = Extract->getOperand(1);
9339 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9340 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9341 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9343 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
9344 OffsetVal, StackPtr);
9347 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
9348 ScalarAddr, NULL, 0, false, false, 0);
9350 // Replace the exact with the load.
9351 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9354 // The replacement was made in place; don't return anything.
9358 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9359 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9360 const X86Subtarget *Subtarget) {
9361 DebugLoc DL = N->getDebugLoc();
9362 SDValue Cond = N->getOperand(0);
9363 // Get the LHS/RHS of the select.
9364 SDValue LHS = N->getOperand(1);
9365 SDValue RHS = N->getOperand(2);
9367 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9368 // instructions match the semantics of the common C idiom x<y?x:y but not
9369 // x<=y?x:y, because of how they handle negative zero (which can be
9370 // ignored in unsafe-math mode).
9371 if (Subtarget->hasSSE2() &&
9372 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9373 Cond.getOpcode() == ISD::SETCC) {
9374 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9376 unsigned Opcode = 0;
9377 // Check for x CC y ? x : y.
9378 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9379 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9383 // Converting this to a min would handle NaNs incorrectly, and swapping
9384 // the operands would cause it to handle comparisons between positive
9385 // and negative zero incorrectly.
9386 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9387 if (!UnsafeFPMath &&
9388 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9390 std::swap(LHS, RHS);
9392 Opcode = X86ISD::FMIN;
9395 // Converting this to a min would handle comparisons between positive
9396 // and negative zero incorrectly.
9397 if (!UnsafeFPMath &&
9398 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9400 Opcode = X86ISD::FMIN;
9403 // Converting this to a min would handle both negative zeros and NaNs
9404 // incorrectly, but we can swap the operands to fix both.
9405 std::swap(LHS, RHS);
9409 Opcode = X86ISD::FMIN;
9413 // Converting this to a max would handle comparisons between positive
9414 // and negative zero incorrectly.
9415 if (!UnsafeFPMath &&
9416 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9418 Opcode = X86ISD::FMAX;
9421 // Converting this to a max would handle NaNs incorrectly, and swapping
9422 // the operands would cause it to handle comparisons between positive
9423 // and negative zero incorrectly.
9424 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9425 if (!UnsafeFPMath &&
9426 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9428 std::swap(LHS, RHS);
9430 Opcode = X86ISD::FMAX;
9433 // Converting this to a max would handle both negative zeros and NaNs
9434 // incorrectly, but we can swap the operands to fix both.
9435 std::swap(LHS, RHS);
9439 Opcode = X86ISD::FMAX;
9442 // Check for x CC y ? y : x -- a min/max with reversed arms.
9443 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9444 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9448 // Converting this to a min would handle comparisons between positive
9449 // and negative zero incorrectly, and swapping the operands would
9450 // cause it to handle NaNs incorrectly.
9451 if (!UnsafeFPMath &&
9452 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9453 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9455 std::swap(LHS, RHS);
9457 Opcode = X86ISD::FMIN;
9460 // Converting this to a min would handle NaNs incorrectly.
9461 if (!UnsafeFPMath &&
9462 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9464 Opcode = X86ISD::FMIN;
9467 // Converting this to a min would handle both negative zeros and NaNs
9468 // incorrectly, but we can swap the operands to fix both.
9469 std::swap(LHS, RHS);
9473 Opcode = X86ISD::FMIN;
9477 // Converting this to a max would handle NaNs incorrectly.
9478 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9480 Opcode = X86ISD::FMAX;
9483 // Converting this to a max would handle comparisons between positive
9484 // and negative zero incorrectly, and swapping the operands would
9485 // cause it to handle NaNs incorrectly.
9486 if (!UnsafeFPMath &&
9487 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9488 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9490 std::swap(LHS, RHS);
9492 Opcode = X86ISD::FMAX;
9495 // Converting this to a max would handle both negative zeros and NaNs
9496 // incorrectly, but we can swap the operands to fix both.
9497 std::swap(LHS, RHS);
9501 Opcode = X86ISD::FMAX;
9507 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9510 // If this is a select between two integer constants, try to do some
9512 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9513 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9514 // Don't do this for crazy integer types.
9515 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9516 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9517 // so that TrueC (the true value) is larger than FalseC.
9518 bool NeedsCondInvert = false;
9520 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9521 // Efficiently invertible.
9522 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9523 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9524 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9525 NeedsCondInvert = true;
9526 std::swap(TrueC, FalseC);
9529 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9530 if (FalseC->getAPIntValue() == 0 &&
9531 TrueC->getAPIntValue().isPowerOf2()) {
9532 if (NeedsCondInvert) // Invert the condition if needed.
9533 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9534 DAG.getConstant(1, Cond.getValueType()));
9536 // Zero extend the condition if needed.
9537 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9539 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9540 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9541 DAG.getConstant(ShAmt, MVT::i8));
9544 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9545 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9546 if (NeedsCondInvert) // Invert the condition if needed.
9547 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9548 DAG.getConstant(1, Cond.getValueType()));
9550 // Zero extend the condition if needed.
9551 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9552 FalseC->getValueType(0), Cond);
9553 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9554 SDValue(FalseC, 0));
9557 // Optimize cases that will turn into an LEA instruction. This requires
9558 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9559 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9560 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9561 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9563 bool isFastMultiplier = false;
9565 switch ((unsigned char)Diff) {
9567 case 1: // result = add base, cond
9568 case 2: // result = lea base( , cond*2)
9569 case 3: // result = lea base(cond, cond*2)
9570 case 4: // result = lea base( , cond*4)
9571 case 5: // result = lea base(cond, cond*4)
9572 case 8: // result = lea base( , cond*8)
9573 case 9: // result = lea base(cond, cond*8)
9574 isFastMultiplier = true;
9579 if (isFastMultiplier) {
9580 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9581 if (NeedsCondInvert) // Invert the condition if needed.
9582 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9583 DAG.getConstant(1, Cond.getValueType()));
9585 // Zero extend the condition if needed.
9586 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9588 // Scale the condition by the difference.
9590 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9591 DAG.getConstant(Diff, Cond.getValueType()));
9593 // Add the base if non-zero.
9594 if (FalseC->getAPIntValue() != 0)
9595 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9596 SDValue(FalseC, 0));
9606 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9607 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9608 TargetLowering::DAGCombinerInfo &DCI) {
9609 DebugLoc DL = N->getDebugLoc();
9611 // If the flag operand isn't dead, don't touch this CMOV.
9612 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9615 // If this is a select between two integer constants, try to do some
9616 // optimizations. Note that the operands are ordered the opposite of SELECT
9618 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9619 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9620 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9621 // larger than FalseC (the false value).
9622 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9624 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9625 CC = X86::GetOppositeBranchCondition(CC);
9626 std::swap(TrueC, FalseC);
9629 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9630 // This is efficient for any integer data type (including i8/i16) and
9632 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9633 SDValue Cond = N->getOperand(3);
9634 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9635 DAG.getConstant(CC, MVT::i8), Cond);
9637 // Zero extend the condition if needed.
9638 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9640 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9641 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9642 DAG.getConstant(ShAmt, MVT::i8));
9643 if (N->getNumValues() == 2) // Dead flag value?
9644 return DCI.CombineTo(N, Cond, SDValue());
9648 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9649 // for any integer data type, including i8/i16.
9650 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9651 SDValue Cond = N->getOperand(3);
9652 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9653 DAG.getConstant(CC, MVT::i8), Cond);
9655 // Zero extend the condition if needed.
9656 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9657 FalseC->getValueType(0), Cond);
9658 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9659 SDValue(FalseC, 0));
9661 if (N->getNumValues() == 2) // Dead flag value?
9662 return DCI.CombineTo(N, Cond, SDValue());
9666 // Optimize cases that will turn into an LEA instruction. This requires
9667 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9668 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9669 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9670 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9672 bool isFastMultiplier = false;
9674 switch ((unsigned char)Diff) {
9676 case 1: // result = add base, cond
9677 case 2: // result = lea base( , cond*2)
9678 case 3: // result = lea base(cond, cond*2)
9679 case 4: // result = lea base( , cond*4)
9680 case 5: // result = lea base(cond, cond*4)
9681 case 8: // result = lea base( , cond*8)
9682 case 9: // result = lea base(cond, cond*8)
9683 isFastMultiplier = true;
9688 if (isFastMultiplier) {
9689 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9690 SDValue Cond = N->getOperand(3);
9691 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9692 DAG.getConstant(CC, MVT::i8), Cond);
9693 // Zero extend the condition if needed.
9694 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9696 // Scale the condition by the difference.
9698 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9699 DAG.getConstant(Diff, Cond.getValueType()));
9701 // Add the base if non-zero.
9702 if (FalseC->getAPIntValue() != 0)
9703 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9704 SDValue(FalseC, 0));
9705 if (N->getNumValues() == 2) // Dead flag value?
9706 return DCI.CombineTo(N, Cond, SDValue());
9716 /// PerformMulCombine - Optimize a single multiply with constant into two
9717 /// in order to implement it with two cheaper instructions, e.g.
9718 /// LEA + SHL, LEA + LEA.
9719 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9720 TargetLowering::DAGCombinerInfo &DCI) {
9721 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9724 EVT VT = N->getValueType(0);
9728 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9731 uint64_t MulAmt = C->getZExtValue();
9732 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9735 uint64_t MulAmt1 = 0;
9736 uint64_t MulAmt2 = 0;
9737 if ((MulAmt % 9) == 0) {
9739 MulAmt2 = MulAmt / 9;
9740 } else if ((MulAmt % 5) == 0) {
9742 MulAmt2 = MulAmt / 5;
9743 } else if ((MulAmt % 3) == 0) {
9745 MulAmt2 = MulAmt / 3;
9748 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9749 DebugLoc DL = N->getDebugLoc();
9751 if (isPowerOf2_64(MulAmt2) &&
9752 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9753 // If second multiplifer is pow2, issue it first. We want the multiply by
9754 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9756 std::swap(MulAmt1, MulAmt2);
9759 if (isPowerOf2_64(MulAmt1))
9760 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9761 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9763 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9764 DAG.getConstant(MulAmt1, VT));
9766 if (isPowerOf2_64(MulAmt2))
9767 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9768 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9770 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9771 DAG.getConstant(MulAmt2, VT));
9773 // Do not add new nodes to DAG combiner worklist.
9774 DCI.CombineTo(N, NewMul, false);
9779 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9780 SDValue N0 = N->getOperand(0);
9781 SDValue N1 = N->getOperand(1);
9782 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9783 EVT VT = N0.getValueType();
9785 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9786 // since the result of setcc_c is all zero's or all ones.
9787 if (N1C && N0.getOpcode() == ISD::AND &&
9788 N0.getOperand(1).getOpcode() == ISD::Constant) {
9789 SDValue N00 = N0.getOperand(0);
9790 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9791 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9792 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9793 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9794 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9795 APInt ShAmt = N1C->getAPIntValue();
9796 Mask = Mask.shl(ShAmt);
9798 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9799 N00, DAG.getConstant(Mask, VT));
9806 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9808 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9809 const X86Subtarget *Subtarget) {
9810 EVT VT = N->getValueType(0);
9811 if (!VT.isVector() && VT.isInteger() &&
9812 N->getOpcode() == ISD::SHL)
9813 return PerformSHLCombine(N, DAG);
9815 // On X86 with SSE2 support, we can transform this to a vector shift if
9816 // all elements are shifted by the same amount. We can't do this in legalize
9817 // because the a constant vector is typically transformed to a constant pool
9818 // so we have no knowledge of the shift amount.
9819 if (!Subtarget->hasSSE2())
9822 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9825 SDValue ShAmtOp = N->getOperand(1);
9826 EVT EltVT = VT.getVectorElementType();
9827 DebugLoc DL = N->getDebugLoc();
9828 SDValue BaseShAmt = SDValue();
9829 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9830 unsigned NumElts = VT.getVectorNumElements();
9832 for (; i != NumElts; ++i) {
9833 SDValue Arg = ShAmtOp.getOperand(i);
9834 if (Arg.getOpcode() == ISD::UNDEF) continue;
9838 for (; i != NumElts; ++i) {
9839 SDValue Arg = ShAmtOp.getOperand(i);
9840 if (Arg.getOpcode() == ISD::UNDEF) continue;
9841 if (Arg != BaseShAmt) {
9845 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9846 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9847 SDValue InVec = ShAmtOp.getOperand(0);
9848 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9849 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9851 for (; i != NumElts; ++i) {
9852 SDValue Arg = InVec.getOperand(i);
9853 if (Arg.getOpcode() == ISD::UNDEF) continue;
9857 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9858 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9859 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9860 if (C->getZExtValue() == SplatIdx)
9861 BaseShAmt = InVec.getOperand(1);
9864 if (BaseShAmt.getNode() == 0)
9865 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9866 DAG.getIntPtrConstant(0));
9870 // The shift amount is an i32.
9871 if (EltVT.bitsGT(MVT::i32))
9872 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9873 else if (EltVT.bitsLT(MVT::i32))
9874 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9876 // The shift amount is identical so we can do a vector shift.
9877 SDValue ValOp = N->getOperand(0);
9878 switch (N->getOpcode()) {
9880 llvm_unreachable("Unknown shift opcode!");
9883 if (VT == MVT::v2i64)
9884 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9885 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9887 if (VT == MVT::v4i32)
9888 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9889 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9891 if (VT == MVT::v8i16)
9892 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9893 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9897 if (VT == MVT::v4i32)
9898 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9899 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9901 if (VT == MVT::v8i16)
9902 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9903 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9907 if (VT == MVT::v2i64)
9908 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9909 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9911 if (VT == MVT::v4i32)
9912 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9913 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9915 if (VT == MVT::v8i16)
9916 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9917 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9924 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9925 TargetLowering::DAGCombinerInfo &DCI,
9926 const X86Subtarget *Subtarget) {
9927 if (DCI.isBeforeLegalizeOps())
9930 EVT VT = N->getValueType(0);
9931 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9934 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9935 SDValue N0 = N->getOperand(0);
9936 SDValue N1 = N->getOperand(1);
9937 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9939 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9941 if (!N0.hasOneUse() || !N1.hasOneUse())
9944 SDValue ShAmt0 = N0.getOperand(1);
9945 if (ShAmt0.getValueType() != MVT::i8)
9947 SDValue ShAmt1 = N1.getOperand(1);
9948 if (ShAmt1.getValueType() != MVT::i8)
9950 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9951 ShAmt0 = ShAmt0.getOperand(0);
9952 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9953 ShAmt1 = ShAmt1.getOperand(0);
9955 DebugLoc DL = N->getDebugLoc();
9956 unsigned Opc = X86ISD::SHLD;
9957 SDValue Op0 = N0.getOperand(0);
9958 SDValue Op1 = N1.getOperand(0);
9959 if (ShAmt0.getOpcode() == ISD::SUB) {
9961 std::swap(Op0, Op1);
9962 std::swap(ShAmt0, ShAmt1);
9965 unsigned Bits = VT.getSizeInBits();
9966 if (ShAmt1.getOpcode() == ISD::SUB) {
9967 SDValue Sum = ShAmt1.getOperand(0);
9968 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9969 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
9970 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
9971 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
9972 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
9973 return DAG.getNode(Opc, DL, VT,
9975 DAG.getNode(ISD::TRUNCATE, DL,
9978 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9979 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9981 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9982 return DAG.getNode(Opc, DL, VT,
9983 N0.getOperand(0), N1.getOperand(0),
9984 DAG.getNode(ISD::TRUNCATE, DL,
9991 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9992 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9993 const X86Subtarget *Subtarget) {
9994 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9995 // the FP state in cases where an emms may be missing.
9996 // A preferable solution to the general problem is to figure out the right
9997 // places to insert EMMS. This qualifies as a quick hack.
9999 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
10000 StoreSDNode *St = cast<StoreSDNode>(N);
10001 EVT VT = St->getValue().getValueType();
10002 if (VT.getSizeInBits() != 64)
10005 const Function *F = DAG.getMachineFunction().getFunction();
10006 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
10007 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
10008 && Subtarget->hasSSE2();
10009 if ((VT.isVector() ||
10010 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
10011 isa<LoadSDNode>(St->getValue()) &&
10012 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
10013 St->getChain().hasOneUse() && !St->isVolatile()) {
10014 SDNode* LdVal = St->getValue().getNode();
10015 LoadSDNode *Ld = 0;
10016 int TokenFactorIndex = -1;
10017 SmallVector<SDValue, 8> Ops;
10018 SDNode* ChainVal = St->getChain().getNode();
10019 // Must be a store of a load. We currently handle two cases: the load
10020 // is a direct child, and it's under an intervening TokenFactor. It is
10021 // possible to dig deeper under nested TokenFactors.
10022 if (ChainVal == LdVal)
10023 Ld = cast<LoadSDNode>(St->getChain());
10024 else if (St->getValue().hasOneUse() &&
10025 ChainVal->getOpcode() == ISD::TokenFactor) {
10026 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
10027 if (ChainVal->getOperand(i).getNode() == LdVal) {
10028 TokenFactorIndex = i;
10029 Ld = cast<LoadSDNode>(St->getValue());
10031 Ops.push_back(ChainVal->getOperand(i));
10035 if (!Ld || !ISD::isNormalLoad(Ld))
10038 // If this is not the MMX case, i.e. we are just turning i64 load/store
10039 // into f64 load/store, avoid the transformation if there are multiple
10040 // uses of the loaded value.
10041 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
10044 DebugLoc LdDL = Ld->getDebugLoc();
10045 DebugLoc StDL = N->getDebugLoc();
10046 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
10047 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
10049 if (Subtarget->is64Bit() || F64IsLegal) {
10050 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
10051 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
10052 Ld->getBasePtr(), Ld->getSrcValue(),
10053 Ld->getSrcValueOffset(), Ld->isVolatile(),
10054 Ld->isNonTemporal(), Ld->getAlignment());
10055 SDValue NewChain = NewLd.getValue(1);
10056 if (TokenFactorIndex != -1) {
10057 Ops.push_back(NewChain);
10058 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10061 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
10062 St->getSrcValue(), St->getSrcValueOffset(),
10063 St->isVolatile(), St->isNonTemporal(),
10064 St->getAlignment());
10067 // Otherwise, lower to two pairs of 32-bit loads / stores.
10068 SDValue LoAddr = Ld->getBasePtr();
10069 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
10070 DAG.getConstant(4, MVT::i32));
10072 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
10073 Ld->getSrcValue(), Ld->getSrcValueOffset(),
10074 Ld->isVolatile(), Ld->isNonTemporal(),
10075 Ld->getAlignment());
10076 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
10077 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
10078 Ld->isVolatile(), Ld->isNonTemporal(),
10079 MinAlign(Ld->getAlignment(), 4));
10081 SDValue NewChain = LoLd.getValue(1);
10082 if (TokenFactorIndex != -1) {
10083 Ops.push_back(LoLd);
10084 Ops.push_back(HiLd);
10085 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10089 LoAddr = St->getBasePtr();
10090 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
10091 DAG.getConstant(4, MVT::i32));
10093 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
10094 St->getSrcValue(), St->getSrcValueOffset(),
10095 St->isVolatile(), St->isNonTemporal(),
10096 St->getAlignment());
10097 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
10099 St->getSrcValueOffset() + 4,
10101 St->isNonTemporal(),
10102 MinAlign(St->getAlignment(), 4));
10103 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
10108 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
10109 /// X86ISD::FXOR nodes.
10110 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
10111 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
10112 // F[X]OR(0.0, x) -> x
10113 // F[X]OR(x, 0.0) -> x
10114 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10115 if (C->getValueAPF().isPosZero())
10116 return N->getOperand(1);
10117 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10118 if (C->getValueAPF().isPosZero())
10119 return N->getOperand(0);
10123 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
10124 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
10125 // FAND(0.0, x) -> 0.0
10126 // FAND(x, 0.0) -> 0.0
10127 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10128 if (C->getValueAPF().isPosZero())
10129 return N->getOperand(0);
10130 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10131 if (C->getValueAPF().isPosZero())
10132 return N->getOperand(1);
10136 static SDValue PerformBTCombine(SDNode *N,
10138 TargetLowering::DAGCombinerInfo &DCI) {
10139 // BT ignores high bits in the bit index operand.
10140 SDValue Op1 = N->getOperand(1);
10141 if (Op1.hasOneUse()) {
10142 unsigned BitWidth = Op1.getValueSizeInBits();
10143 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
10144 APInt KnownZero, KnownOne;
10145 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
10146 !DCI.isBeforeLegalizeOps());
10147 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10148 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
10149 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
10150 DCI.CommitTargetLoweringOpt(TLO);
10155 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
10156 SDValue Op = N->getOperand(0);
10157 if (Op.getOpcode() == ISD::BIT_CONVERT)
10158 Op = Op.getOperand(0);
10159 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
10160 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
10161 VT.getVectorElementType().getSizeInBits() ==
10162 OpVT.getVectorElementType().getSizeInBits()) {
10163 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
10168 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
10169 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
10170 // (and (i32 x86isd::setcc_carry), 1)
10171 // This eliminates the zext. This transformation is necessary because
10172 // ISD::SETCC is always legalized to i8.
10173 DebugLoc dl = N->getDebugLoc();
10174 SDValue N0 = N->getOperand(0);
10175 EVT VT = N->getValueType(0);
10176 if (N0.getOpcode() == ISD::AND &&
10178 N0.getOperand(0).hasOneUse()) {
10179 SDValue N00 = N0.getOperand(0);
10180 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
10182 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
10183 if (!C || C->getZExtValue() != 1)
10185 return DAG.getNode(ISD::AND, dl, VT,
10186 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
10187 N00.getOperand(0), N00.getOperand(1)),
10188 DAG.getConstant(1, VT));
10194 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
10195 DAGCombinerInfo &DCI) const {
10196 SelectionDAG &DAG = DCI.DAG;
10197 switch (N->getOpcode()) {
10199 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
10200 case ISD::EXTRACT_VECTOR_ELT:
10201 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
10202 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
10203 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
10204 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
10207 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
10208 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
10209 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
10211 case X86ISD::FOR: return PerformFORCombine(N, DAG);
10212 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
10213 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
10214 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
10215 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
10221 /// isTypeDesirableForOp - Return true if the target has native support for
10222 /// the specified value type and it is 'desirable' to use the type for the
10223 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
10224 /// instruction encodings are longer and some i16 instructions are slow.
10225 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
10226 if (!isTypeLegal(VT))
10228 if (VT != MVT::i16)
10235 case ISD::SIGN_EXTEND:
10236 case ISD::ZERO_EXTEND:
10237 case ISD::ANY_EXTEND:
10250 static bool MayFoldLoad(SDValue Op) {
10251 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
10254 static bool MayFoldIntoStore(SDValue Op) {
10255 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
10258 /// IsDesirableToPromoteOp - This method query the target whether it is
10259 /// beneficial for dag combiner to promote the specified node. If true, it
10260 /// should return the desired promotion type by reference.
10261 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
10262 EVT VT = Op.getValueType();
10263 if (VT != MVT::i16)
10266 bool Promote = false;
10267 bool Commute = false;
10268 switch (Op.getOpcode()) {
10271 LoadSDNode *LD = cast<LoadSDNode>(Op);
10272 // If the non-extending load has a single use and it's not live out, then it
10273 // might be folded.
10274 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
10275 Op.hasOneUse()*/) {
10276 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
10277 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
10278 // The only case where we'd want to promote LOAD (rather then it being
10279 // promoted as an operand is when it's only use is liveout.
10280 if (UI->getOpcode() != ISD::CopyToReg)
10287 case ISD::SIGN_EXTEND:
10288 case ISD::ZERO_EXTEND:
10289 case ISD::ANY_EXTEND:
10294 SDValue N0 = Op.getOperand(0);
10295 // Look out for (store (shl (load), x)).
10296 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10309 SDValue N0 = Op.getOperand(0);
10310 SDValue N1 = Op.getOperand(1);
10311 if (!Commute && MayFoldLoad(N1))
10313 // Avoid disabling potential load folding opportunities.
10314 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10316 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10326 //===----------------------------------------------------------------------===//
10327 // X86 Inline Assembly Support
10328 //===----------------------------------------------------------------------===//
10330 static bool LowerToBSwap(CallInst *CI) {
10331 // FIXME: this should verify that we are targetting a 486 or better. If not,
10332 // we will turn this bswap into something that will be lowered to logical ops
10333 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10334 // so don't worry about this.
10336 // Verify this is a simple bswap.
10337 if (CI->getNumArgOperands() != 1 ||
10338 CI->getType() != CI->getArgOperand(0)->getType() ||
10339 !CI->getType()->isIntegerTy())
10342 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10343 if (!Ty || Ty->getBitWidth() % 16 != 0)
10346 // Okay, we can do this xform, do so now.
10347 const Type *Tys[] = { Ty };
10348 Module *M = CI->getParent()->getParent()->getParent();
10349 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10351 Value *Op = CI->getArgOperand(0);
10352 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10354 CI->replaceAllUsesWith(Op);
10355 CI->eraseFromParent();
10359 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10360 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10361 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10363 std::string AsmStr = IA->getAsmString();
10365 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10366 SmallVector<StringRef, 4> AsmPieces;
10367 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10369 switch (AsmPieces.size()) {
10370 default: return false;
10372 AsmStr = AsmPieces[0];
10374 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10377 if (AsmPieces.size() == 2 &&
10378 (AsmPieces[0] == "bswap" ||
10379 AsmPieces[0] == "bswapq" ||
10380 AsmPieces[0] == "bswapl") &&
10381 (AsmPieces[1] == "$0" ||
10382 AsmPieces[1] == "${0:q}")) {
10383 // No need to check constraints, nothing other than the equivalent of
10384 // "=r,0" would be valid here.
10385 return LowerToBSwap(CI);
10387 // rorw $$8, ${0:w} --> llvm.bswap.i16
10388 if (CI->getType()->isIntegerTy(16) &&
10389 AsmPieces.size() == 3 &&
10390 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10391 AsmPieces[1] == "$$8," &&
10392 AsmPieces[2] == "${0:w}" &&
10393 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10395 const std::string &Constraints = IA->getConstraintString();
10396 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10397 std::sort(AsmPieces.begin(), AsmPieces.end());
10398 if (AsmPieces.size() == 4 &&
10399 AsmPieces[0] == "~{cc}" &&
10400 AsmPieces[1] == "~{dirflag}" &&
10401 AsmPieces[2] == "~{flags}" &&
10402 AsmPieces[3] == "~{fpsr}") {
10403 return LowerToBSwap(CI);
10408 if (CI->getType()->isIntegerTy(64) &&
10409 Constraints.size() >= 2 &&
10410 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10411 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10412 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10413 SmallVector<StringRef, 4> Words;
10414 SplitString(AsmPieces[0], Words, " \t");
10415 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10417 SplitString(AsmPieces[1], Words, " \t");
10418 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10420 SplitString(AsmPieces[2], Words, " \t,");
10421 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10422 Words[2] == "%edx") {
10423 return LowerToBSwap(CI);
10435 /// getConstraintType - Given a constraint letter, return the type of
10436 /// constraint it is for this target.
10437 X86TargetLowering::ConstraintType
10438 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10439 if (Constraint.size() == 1) {
10440 switch (Constraint[0]) {
10452 return C_RegisterClass;
10460 return TargetLowering::getConstraintType(Constraint);
10463 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10464 /// with another that has more specific requirements based on the type of the
10465 /// corresponding operand.
10466 const char *X86TargetLowering::
10467 LowerXConstraint(EVT ConstraintVT) const {
10468 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10469 // 'f' like normal targets.
10470 if (ConstraintVT.isFloatingPoint()) {
10471 if (Subtarget->hasSSE2())
10473 if (Subtarget->hasSSE1())
10477 return TargetLowering::LowerXConstraint(ConstraintVT);
10480 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10481 /// vector. If it is invalid, don't add anything to Ops.
10482 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10484 std::vector<SDValue>&Ops,
10485 SelectionDAG &DAG) const {
10486 SDValue Result(0, 0);
10488 switch (Constraint) {
10491 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10492 if (C->getZExtValue() <= 31) {
10493 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10499 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10500 if (C->getZExtValue() <= 63) {
10501 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10507 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10508 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10509 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10515 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10516 if (C->getZExtValue() <= 255) {
10517 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10523 // 32-bit signed value
10524 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10525 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10526 C->getSExtValue())) {
10527 // Widen to 64 bits here to get it sign extended.
10528 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10531 // FIXME gcc accepts some relocatable values here too, but only in certain
10532 // memory models; it's complicated.
10537 // 32-bit unsigned value
10538 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10539 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10540 C->getZExtValue())) {
10541 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10545 // FIXME gcc accepts some relocatable values here too, but only in certain
10546 // memory models; it's complicated.
10550 // Literal immediates are always ok.
10551 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10552 // Widen to 64 bits here to get it sign extended.
10553 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10557 // In any sort of PIC mode addresses need to be computed at runtime by
10558 // adding in a register or some sort of table lookup. These can't
10559 // be used as immediates.
10560 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
10563 // If we are in non-pic codegen mode, we allow the address of a global (with
10564 // an optional displacement) to be used with 'i'.
10565 GlobalAddressSDNode *GA = 0;
10566 int64_t Offset = 0;
10568 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10570 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10571 Offset += GA->getOffset();
10573 } else if (Op.getOpcode() == ISD::ADD) {
10574 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10575 Offset += C->getZExtValue();
10576 Op = Op.getOperand(0);
10579 } else if (Op.getOpcode() == ISD::SUB) {
10580 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10581 Offset += -C->getZExtValue();
10582 Op = Op.getOperand(0);
10587 // Otherwise, this isn't something we can handle, reject it.
10591 const GlobalValue *GV = GA->getGlobal();
10592 // If we require an extra load to get this address, as in PIC mode, we
10593 // can't accept it.
10594 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10595 getTargetMachine())))
10598 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
10599 GA->getValueType(0), Offset);
10604 if (Result.getNode()) {
10605 Ops.push_back(Result);
10608 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10611 std::vector<unsigned> X86TargetLowering::
10612 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10614 if (Constraint.size() == 1) {
10615 // FIXME: not handling fp-stack yet!
10616 switch (Constraint[0]) { // GCC X86 Constraint Letters
10617 default: break; // Unknown constraint letter
10618 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10619 if (Subtarget->is64Bit()) {
10620 if (VT == MVT::i32)
10621 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10622 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10623 X86::R10D,X86::R11D,X86::R12D,
10624 X86::R13D,X86::R14D,X86::R15D,
10625 X86::EBP, X86::ESP, 0);
10626 else if (VT == MVT::i16)
10627 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10628 X86::SI, X86::DI, X86::R8W,X86::R9W,
10629 X86::R10W,X86::R11W,X86::R12W,
10630 X86::R13W,X86::R14W,X86::R15W,
10631 X86::BP, X86::SP, 0);
10632 else if (VT == MVT::i8)
10633 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10634 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10635 X86::R10B,X86::R11B,X86::R12B,
10636 X86::R13B,X86::R14B,X86::R15B,
10637 X86::BPL, X86::SPL, 0);
10639 else if (VT == MVT::i64)
10640 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10641 X86::RSI, X86::RDI, X86::R8, X86::R9,
10642 X86::R10, X86::R11, X86::R12,
10643 X86::R13, X86::R14, X86::R15,
10644 X86::RBP, X86::RSP, 0);
10648 // 32-bit fallthrough
10649 case 'Q': // Q_REGS
10650 if (VT == MVT::i32)
10651 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10652 else if (VT == MVT::i16)
10653 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10654 else if (VT == MVT::i8)
10655 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10656 else if (VT == MVT::i64)
10657 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10662 return std::vector<unsigned>();
10665 std::pair<unsigned, const TargetRegisterClass*>
10666 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10668 // First, see if this is a constraint that directly corresponds to an LLVM
10670 if (Constraint.size() == 1) {
10671 // GCC Constraint Letters
10672 switch (Constraint[0]) {
10674 case 'r': // GENERAL_REGS
10675 case 'l': // INDEX_REGS
10677 return std::make_pair(0U, X86::GR8RegisterClass);
10678 if (VT == MVT::i16)
10679 return std::make_pair(0U, X86::GR16RegisterClass);
10680 if (VT == MVT::i32 || !Subtarget->is64Bit())
10681 return std::make_pair(0U, X86::GR32RegisterClass);
10682 return std::make_pair(0U, X86::GR64RegisterClass);
10683 case 'R': // LEGACY_REGS
10685 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10686 if (VT == MVT::i16)
10687 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10688 if (VT == MVT::i32 || !Subtarget->is64Bit())
10689 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10690 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10691 case 'f': // FP Stack registers.
10692 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10693 // value to the correct fpstack register class.
10694 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10695 return std::make_pair(0U, X86::RFP32RegisterClass);
10696 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10697 return std::make_pair(0U, X86::RFP64RegisterClass);
10698 return std::make_pair(0U, X86::RFP80RegisterClass);
10699 case 'y': // MMX_REGS if MMX allowed.
10700 if (!Subtarget->hasMMX()) break;
10701 return std::make_pair(0U, X86::VR64RegisterClass);
10702 case 'Y': // SSE_REGS if SSE2 allowed
10703 if (!Subtarget->hasSSE2()) break;
10705 case 'x': // SSE_REGS if SSE1 allowed
10706 if (!Subtarget->hasSSE1()) break;
10708 switch (VT.getSimpleVT().SimpleTy) {
10710 // Scalar SSE types.
10713 return std::make_pair(0U, X86::FR32RegisterClass);
10716 return std::make_pair(0U, X86::FR64RegisterClass);
10724 return std::make_pair(0U, X86::VR128RegisterClass);
10730 // Use the default implementation in TargetLowering to convert the register
10731 // constraint into a member of a register class.
10732 std::pair<unsigned, const TargetRegisterClass*> Res;
10733 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10735 // Not found as a standard register?
10736 if (Res.second == 0) {
10737 // Map st(0) -> st(7) -> ST0
10738 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10739 tolower(Constraint[1]) == 's' &&
10740 tolower(Constraint[2]) == 't' &&
10741 Constraint[3] == '(' &&
10742 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10743 Constraint[5] == ')' &&
10744 Constraint[6] == '}') {
10746 Res.first = X86::ST0+Constraint[4]-'0';
10747 Res.second = X86::RFP80RegisterClass;
10751 // GCC allows "st(0)" to be called just plain "st".
10752 if (StringRef("{st}").equals_lower(Constraint)) {
10753 Res.first = X86::ST0;
10754 Res.second = X86::RFP80RegisterClass;
10759 if (StringRef("{flags}").equals_lower(Constraint)) {
10760 Res.first = X86::EFLAGS;
10761 Res.second = X86::CCRRegisterClass;
10765 // 'A' means EAX + EDX.
10766 if (Constraint == "A") {
10767 Res.first = X86::EAX;
10768 Res.second = X86::GR32_ADRegisterClass;
10774 // Otherwise, check to see if this is a register class of the wrong value
10775 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10776 // turn into {ax},{dx}.
10777 if (Res.second->hasType(VT))
10778 return Res; // Correct type already, nothing to do.
10780 // All of the single-register GCC register classes map their values onto
10781 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10782 // really want an 8-bit or 32-bit register, map to the appropriate register
10783 // class and return the appropriate register.
10784 if (Res.second == X86::GR16RegisterClass) {
10785 if (VT == MVT::i8) {
10786 unsigned DestReg = 0;
10787 switch (Res.first) {
10789 case X86::AX: DestReg = X86::AL; break;
10790 case X86::DX: DestReg = X86::DL; break;
10791 case X86::CX: DestReg = X86::CL; break;
10792 case X86::BX: DestReg = X86::BL; break;
10795 Res.first = DestReg;
10796 Res.second = X86::GR8RegisterClass;
10798 } else if (VT == MVT::i32) {
10799 unsigned DestReg = 0;
10800 switch (Res.first) {
10802 case X86::AX: DestReg = X86::EAX; break;
10803 case X86::DX: DestReg = X86::EDX; break;
10804 case X86::CX: DestReg = X86::ECX; break;
10805 case X86::BX: DestReg = X86::EBX; break;
10806 case X86::SI: DestReg = X86::ESI; break;
10807 case X86::DI: DestReg = X86::EDI; break;
10808 case X86::BP: DestReg = X86::EBP; break;
10809 case X86::SP: DestReg = X86::ESP; break;
10812 Res.first = DestReg;
10813 Res.second = X86::GR32RegisterClass;
10815 } else if (VT == MVT::i64) {
10816 unsigned DestReg = 0;
10817 switch (Res.first) {
10819 case X86::AX: DestReg = X86::RAX; break;
10820 case X86::DX: DestReg = X86::RDX; break;
10821 case X86::CX: DestReg = X86::RCX; break;
10822 case X86::BX: DestReg = X86::RBX; break;
10823 case X86::SI: DestReg = X86::RSI; break;
10824 case X86::DI: DestReg = X86::RDI; break;
10825 case X86::BP: DestReg = X86::RBP; break;
10826 case X86::SP: DestReg = X86::RSP; break;
10829 Res.first = DestReg;
10830 Res.second = X86::GR64RegisterClass;
10833 } else if (Res.second == X86::FR32RegisterClass ||
10834 Res.second == X86::FR64RegisterClass ||
10835 Res.second == X86::VR128RegisterClass) {
10836 // Handle references to XMM physical registers that got mapped into the
10837 // wrong class. This can happen with constraints like {xmm0} where the
10838 // target independent register mapper will just pick the first match it can
10839 // find, ignoring the required type.
10840 if (VT == MVT::f32)
10841 Res.second = X86::FR32RegisterClass;
10842 else if (VT == MVT::f64)
10843 Res.second = X86::FR64RegisterClass;
10844 else if (X86::VR128RegisterClass->hasType(VT))
10845 Res.second = X86::VR128RegisterClass;