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);
844 // i8 and i16 vectors are custom , because the source register and source
845 // source memory operand types are not the same width. f32 vectors are
846 // custom since the immediate controlling the insert encodes additional
848 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
850 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
851 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
853 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
854 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
855 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
856 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
858 if (Subtarget->is64Bit()) {
859 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
860 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
864 if (Subtarget->hasSSE42()) {
865 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
868 if (!UseSoftFloat && Subtarget->hasAVX()) {
869 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
870 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
871 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
872 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
874 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
875 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
876 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
877 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
878 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
879 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
880 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
881 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
882 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
883 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
884 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
885 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
886 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
887 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
888 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
890 // Operations to consider commented out -v16i16 v32i8
891 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
892 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
893 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
894 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
895 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
896 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
897 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
898 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
899 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
900 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
901 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
902 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
903 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
904 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
906 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
907 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
908 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
909 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
911 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
912 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
913 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
914 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
915 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
917 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
918 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
919 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
920 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
921 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
922 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
925 // Not sure we want to do this since there are no 256-bit integer
928 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
929 // This includes 256-bit vectors
930 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
931 EVT VT = (MVT::SimpleValueType)i;
933 // Do not attempt to custom lower non-power-of-2 vectors
934 if (!isPowerOf2_32(VT.getVectorNumElements()))
937 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
938 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
939 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
942 if (Subtarget->is64Bit()) {
943 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
944 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
949 // Not sure we want to do this since there are no 256-bit integer
952 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
953 // Including 256-bit vectors
954 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
955 EVT VT = (MVT::SimpleValueType)i;
957 if (!VT.is256BitVector()) {
960 setOperationAction(ISD::AND, VT, Promote);
961 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
962 setOperationAction(ISD::OR, VT, Promote);
963 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
964 setOperationAction(ISD::XOR, VT, Promote);
965 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
966 setOperationAction(ISD::LOAD, VT, Promote);
967 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
968 setOperationAction(ISD::SELECT, VT, Promote);
969 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
972 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
976 // We want to custom lower some of our intrinsics.
977 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
979 // Add/Sub/Mul with overflow operations are custom lowered.
980 setOperationAction(ISD::SADDO, MVT::i32, Custom);
981 setOperationAction(ISD::UADDO, MVT::i32, Custom);
982 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
983 setOperationAction(ISD::USUBO, MVT::i32, Custom);
984 setOperationAction(ISD::SMULO, MVT::i32, Custom);
986 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
987 // handle type legalization for these operations here.
989 // FIXME: We really should do custom legalization for addition and
990 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
991 // than generic legalization for 64-bit multiplication-with-overflow, though.
992 if (Subtarget->is64Bit()) {
993 setOperationAction(ISD::SADDO, MVT::i64, Custom);
994 setOperationAction(ISD::UADDO, MVT::i64, Custom);
995 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
996 setOperationAction(ISD::USUBO, MVT::i64, Custom);
997 setOperationAction(ISD::SMULO, MVT::i64, Custom);
1000 if (!Subtarget->is64Bit()) {
1001 // These libcalls are not available in 32-bit.
1002 setLibcallName(RTLIB::SHL_I128, 0);
1003 setLibcallName(RTLIB::SRL_I128, 0);
1004 setLibcallName(RTLIB::SRA_I128, 0);
1007 // We have target-specific dag combine patterns for the following nodes:
1008 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1009 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1010 setTargetDAGCombine(ISD::BUILD_VECTOR);
1011 setTargetDAGCombine(ISD::SELECT);
1012 setTargetDAGCombine(ISD::SHL);
1013 setTargetDAGCombine(ISD::SRA);
1014 setTargetDAGCombine(ISD::SRL);
1015 setTargetDAGCombine(ISD::OR);
1016 setTargetDAGCombine(ISD::STORE);
1017 setTargetDAGCombine(ISD::ZERO_EXTEND);
1018 if (Subtarget->is64Bit())
1019 setTargetDAGCombine(ISD::MUL);
1021 computeRegisterProperties();
1023 // FIXME: These should be based on subtarget info. Plus, the values should
1024 // be smaller when we are in optimizing for size mode.
1025 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1026 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1027 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1028 setPrefLoopAlignment(16);
1029 benefitFromCodePlacementOpt = true;
1033 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1038 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1039 /// the desired ByVal argument alignment.
1040 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1043 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1044 if (VTy->getBitWidth() == 128)
1046 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1047 unsigned EltAlign = 0;
1048 getMaxByValAlign(ATy->getElementType(), EltAlign);
1049 if (EltAlign > MaxAlign)
1050 MaxAlign = EltAlign;
1051 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1052 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1053 unsigned EltAlign = 0;
1054 getMaxByValAlign(STy->getElementType(i), EltAlign);
1055 if (EltAlign > MaxAlign)
1056 MaxAlign = EltAlign;
1064 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1065 /// function arguments in the caller parameter area. For X86, aggregates
1066 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1067 /// are at 4-byte boundaries.
1068 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1069 if (Subtarget->is64Bit()) {
1070 // Max of 8 and alignment of type.
1071 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1078 if (Subtarget->hasSSE1())
1079 getMaxByValAlign(Ty, Align);
1083 /// getOptimalMemOpType - Returns the target specific optimal type for load
1084 /// and store operations as a result of memset, memcpy, and memmove
1085 /// lowering. If DstAlign is zero that means it's safe to destination
1086 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1087 /// means there isn't a need to check it against alignment requirement,
1088 /// probably because the source does not need to be loaded. If
1089 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1090 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1091 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1092 /// constant so it does not need to be loaded.
1093 /// It returns EVT::Other if the type should be determined using generic
1094 /// target-independent logic.
1096 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1097 unsigned DstAlign, unsigned SrcAlign,
1098 bool NonScalarIntSafe,
1100 MachineFunction &MF) const {
1101 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1102 // linux. This is because the stack realignment code can't handle certain
1103 // cases like PR2962. This should be removed when PR2962 is fixed.
1104 const Function *F = MF.getFunction();
1105 if (NonScalarIntSafe &&
1106 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1108 (Subtarget->isUnalignedMemAccessFast() ||
1109 ((DstAlign == 0 || DstAlign >= 16) &&
1110 (SrcAlign == 0 || SrcAlign >= 16))) &&
1111 Subtarget->getStackAlignment() >= 16) {
1112 if (Subtarget->hasSSE2())
1114 if (Subtarget->hasSSE1())
1116 } else if (!MemcpyStrSrc && Size >= 8 &&
1117 !Subtarget->is64Bit() &&
1118 Subtarget->getStackAlignment() >= 8 &&
1119 Subtarget->hasSSE2()) {
1120 // Do not use f64 to lower memcpy if source is string constant. It's
1121 // better to use i32 to avoid the loads.
1125 if (Subtarget->is64Bit() && Size >= 8)
1130 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1131 /// current function. The returned value is a member of the
1132 /// MachineJumpTableInfo::JTEntryKind enum.
1133 unsigned X86TargetLowering::getJumpTableEncoding() const {
1134 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1136 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1137 Subtarget->isPICStyleGOT())
1138 return MachineJumpTableInfo::EK_Custom32;
1140 // Otherwise, use the normal jump table encoding heuristics.
1141 return TargetLowering::getJumpTableEncoding();
1144 /// getPICBaseSymbol - Return the X86-32 PIC base.
1146 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1147 MCContext &Ctx) const {
1148 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1149 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1150 Twine(MF->getFunctionNumber())+"$pb");
1155 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1156 const MachineBasicBlock *MBB,
1157 unsigned uid,MCContext &Ctx) const{
1158 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1159 Subtarget->isPICStyleGOT());
1160 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1162 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1163 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1166 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1168 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1169 SelectionDAG &DAG) const {
1170 if (!Subtarget->is64Bit())
1171 // This doesn't have DebugLoc associated with it, but is not really the
1172 // same as a Register.
1173 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1177 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1178 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1180 const MCExpr *X86TargetLowering::
1181 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1182 MCContext &Ctx) const {
1183 // X86-64 uses RIP relative addressing based on the jump table label.
1184 if (Subtarget->isPICStyleRIPRel())
1185 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1187 // Otherwise, the reference is relative to the PIC base.
1188 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1191 /// getFunctionAlignment - Return the Log2 alignment of this function.
1192 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1193 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1196 std::pair<const TargetRegisterClass*, uint8_t>
1197 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1198 const TargetRegisterClass *RRC = 0;
1200 switch (VT.getSimpleVT().SimpleTy) {
1202 return TargetLowering::findRepresentativeClass(VT);
1203 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1204 RRC = (Subtarget->is64Bit()
1205 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1207 case MVT::v8i8: case MVT::v4i16:
1208 case MVT::v2i32: case MVT::v1i64:
1209 RRC = X86::VR64RegisterClass;
1211 case MVT::f32: case MVT::f64:
1212 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1213 case MVT::v4f32: case MVT::v2f64:
1214 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1216 RRC = X86::VR128RegisterClass;
1219 return std::make_pair(RRC, Cost);
1223 X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
1224 MachineFunction &MF) const {
1225 unsigned FPDiff = RegInfo->hasFP(MF) ? 1 : 0;
1226 switch (RC->getID()) {
1229 case X86::GR32RegClassID:
1231 case X86::GR64RegClassID:
1233 case X86::VR128RegClassID:
1234 return Subtarget->is64Bit() ? 10 : 4;
1235 case X86::VR64RegClassID:
1240 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1241 unsigned &Offset) const {
1242 if (!Subtarget->isTargetLinux())
1245 if (Subtarget->is64Bit()) {
1246 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1248 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1261 //===----------------------------------------------------------------------===//
1262 // Return Value Calling Convention Implementation
1263 //===----------------------------------------------------------------------===//
1265 #include "X86GenCallingConv.inc"
1268 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1269 const SmallVectorImpl<ISD::OutputArg> &Outs,
1270 LLVMContext &Context) const {
1271 SmallVector<CCValAssign, 16> RVLocs;
1272 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1274 return CCInfo.CheckReturn(Outs, RetCC_X86);
1278 X86TargetLowering::LowerReturn(SDValue Chain,
1279 CallingConv::ID CallConv, bool isVarArg,
1280 const SmallVectorImpl<ISD::OutputArg> &Outs,
1281 const SmallVectorImpl<SDValue> &OutVals,
1282 DebugLoc dl, SelectionDAG &DAG) const {
1283 MachineFunction &MF = DAG.getMachineFunction();
1284 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1286 SmallVector<CCValAssign, 16> RVLocs;
1287 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1288 RVLocs, *DAG.getContext());
1289 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1291 // Add the regs to the liveout set for the function.
1292 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1293 for (unsigned i = 0; i != RVLocs.size(); ++i)
1294 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1295 MRI.addLiveOut(RVLocs[i].getLocReg());
1299 SmallVector<SDValue, 6> RetOps;
1300 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1301 // Operand #1 = Bytes To Pop
1302 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1305 // Copy the result values into the output registers.
1306 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1307 CCValAssign &VA = RVLocs[i];
1308 assert(VA.isRegLoc() && "Can only return in registers!");
1309 SDValue ValToCopy = OutVals[i];
1310 EVT ValVT = ValToCopy.getValueType();
1312 // If this is x86-64, and we disabled SSE, we can't return FP values
1313 if ((ValVT == MVT::f32 || ValVT == MVT::f64) &&
1314 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1315 report_fatal_error("SSE register return with SSE disabled");
1317 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1318 // llvm-gcc has never done it right and no one has noticed, so this
1319 // should be OK for now.
1320 if (ValVT == MVT::f64 &&
1321 (Subtarget->is64Bit() && !Subtarget->hasSSE2())) {
1322 report_fatal_error("SSE2 register return with SSE2 disabled");
1325 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1326 // the RET instruction and handled by the FP Stackifier.
1327 if (VA.getLocReg() == X86::ST0 ||
1328 VA.getLocReg() == X86::ST1) {
1329 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1330 // change the value to the FP stack register class.
1331 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1332 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1333 RetOps.push_back(ValToCopy);
1334 // Don't emit a copytoreg.
1338 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1339 // which is returned in RAX / RDX.
1340 if (Subtarget->is64Bit()) {
1341 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1342 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1343 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1344 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1349 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1350 Flag = Chain.getValue(1);
1353 // The x86-64 ABI for returning structs by value requires that we copy
1354 // the sret argument into %rax for the return. We saved the argument into
1355 // a virtual register in the entry block, so now we copy the value out
1357 if (Subtarget->is64Bit() &&
1358 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1359 MachineFunction &MF = DAG.getMachineFunction();
1360 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1361 unsigned Reg = FuncInfo->getSRetReturnReg();
1363 "SRetReturnReg should have been set in LowerFormalArguments().");
1364 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1366 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1367 Flag = Chain.getValue(1);
1369 // RAX now acts like a return value.
1370 MRI.addLiveOut(X86::RAX);
1373 RetOps[0] = Chain; // Update chain.
1375 // Add the flag if we have it.
1377 RetOps.push_back(Flag);
1379 return DAG.getNode(X86ISD::RET_FLAG, dl,
1380 MVT::Other, &RetOps[0], RetOps.size());
1383 /// LowerCallResult - Lower the result values of a call into the
1384 /// appropriate copies out of appropriate physical registers.
1387 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1388 CallingConv::ID CallConv, bool isVarArg,
1389 const SmallVectorImpl<ISD::InputArg> &Ins,
1390 DebugLoc dl, SelectionDAG &DAG,
1391 SmallVectorImpl<SDValue> &InVals) const {
1393 // Assign locations to each value returned by this call.
1394 SmallVector<CCValAssign, 16> RVLocs;
1395 bool Is64Bit = Subtarget->is64Bit();
1396 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1397 RVLocs, *DAG.getContext());
1398 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1400 // Copy all of the result registers out of their specified physreg.
1401 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1402 CCValAssign &VA = RVLocs[i];
1403 EVT CopyVT = VA.getValVT();
1405 // If this is x86-64, and we disabled SSE, we can't return FP values
1406 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1407 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1408 report_fatal_error("SSE register return with SSE disabled");
1413 // If this is a call to a function that returns an fp value on the floating
1414 // point stack, we must guarantee the the value is popped from the stack, so
1415 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1416 // if the return value is not used. We use the FpGET_ST0 instructions
1418 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1419 // If we prefer to use the value in xmm registers, copy it out as f80 and
1420 // use a truncate to move it from fp stack reg to xmm reg.
1421 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1422 bool isST0 = VA.getLocReg() == X86::ST0;
1424 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1425 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1426 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1427 SDValue Ops[] = { Chain, InFlag };
1428 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
1430 Val = Chain.getValue(0);
1432 // Round the f80 to the right size, which also moves it to the appropriate
1434 if (CopyVT != VA.getValVT())
1435 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1436 // This truncation won't change the value.
1437 DAG.getIntPtrConstant(1));
1438 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1439 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1440 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1441 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1442 MVT::v2i64, InFlag).getValue(1);
1443 Val = Chain.getValue(0);
1444 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1445 Val, DAG.getConstant(0, MVT::i64));
1447 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1448 MVT::i64, InFlag).getValue(1);
1449 Val = Chain.getValue(0);
1451 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1453 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1454 CopyVT, InFlag).getValue(1);
1455 Val = Chain.getValue(0);
1457 InFlag = Chain.getValue(2);
1458 InVals.push_back(Val);
1465 //===----------------------------------------------------------------------===//
1466 // C & StdCall & Fast Calling Convention implementation
1467 //===----------------------------------------------------------------------===//
1468 // StdCall calling convention seems to be standard for many Windows' API
1469 // routines and around. It differs from C calling convention just a little:
1470 // callee should clean up the stack, not caller. Symbols should be also
1471 // decorated in some fancy way :) It doesn't support any vector arguments.
1472 // For info on fast calling convention see Fast Calling Convention (tail call)
1473 // implementation LowerX86_32FastCCCallTo.
1475 /// CallIsStructReturn - Determines whether a call uses struct return
1477 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1481 return Outs[0].Flags.isSRet();
1484 /// ArgsAreStructReturn - Determines whether a function uses struct
1485 /// return semantics.
1487 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1491 return Ins[0].Flags.isSRet();
1494 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1495 /// given CallingConvention value.
1496 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1497 if (Subtarget->is64Bit()) {
1498 if (CC == CallingConv::GHC)
1499 return CC_X86_64_GHC;
1500 else if (Subtarget->isTargetWin64())
1501 return CC_X86_Win64_C;
1506 if (CC == CallingConv::X86_FastCall)
1507 return CC_X86_32_FastCall;
1508 else if (CC == CallingConv::X86_ThisCall)
1509 return CC_X86_32_ThisCall;
1510 else if (CC == CallingConv::Fast)
1511 return CC_X86_32_FastCC;
1512 else if (CC == CallingConv::GHC)
1513 return CC_X86_32_GHC;
1518 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1519 /// by "Src" to address "Dst" with size and alignment information specified by
1520 /// the specific parameter attribute. The copy will be passed as a byval
1521 /// function parameter.
1523 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1524 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1526 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1527 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1528 /*isVolatile*/false, /*AlwaysInline=*/true,
1532 /// IsTailCallConvention - Return true if the calling convention is one that
1533 /// supports tail call optimization.
1534 static bool IsTailCallConvention(CallingConv::ID CC) {
1535 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1538 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1539 /// a tailcall target by changing its ABI.
1540 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1541 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1545 X86TargetLowering::LowerMemArgument(SDValue Chain,
1546 CallingConv::ID CallConv,
1547 const SmallVectorImpl<ISD::InputArg> &Ins,
1548 DebugLoc dl, SelectionDAG &DAG,
1549 const CCValAssign &VA,
1550 MachineFrameInfo *MFI,
1552 // Create the nodes corresponding to a load from this parameter slot.
1553 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1554 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1555 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1558 // If value is passed by pointer we have address passed instead of the value
1560 if (VA.getLocInfo() == CCValAssign::Indirect)
1561 ValVT = VA.getLocVT();
1563 ValVT = VA.getValVT();
1565 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1566 // changed with more analysis.
1567 // In case of tail call optimization mark all arguments mutable. Since they
1568 // could be overwritten by lowering of arguments in case of a tail call.
1569 if (Flags.isByVal()) {
1570 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1571 VA.getLocMemOffset(), isImmutable);
1572 return DAG.getFrameIndex(FI, getPointerTy());
1574 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1575 VA.getLocMemOffset(), isImmutable);
1576 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1577 return DAG.getLoad(ValVT, dl, Chain, FIN,
1578 PseudoSourceValue::getFixedStack(FI), 0,
1584 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1585 CallingConv::ID CallConv,
1587 const SmallVectorImpl<ISD::InputArg> &Ins,
1590 SmallVectorImpl<SDValue> &InVals)
1592 MachineFunction &MF = DAG.getMachineFunction();
1593 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1595 const Function* Fn = MF.getFunction();
1596 if (Fn->hasExternalLinkage() &&
1597 Subtarget->isTargetCygMing() &&
1598 Fn->getName() == "main")
1599 FuncInfo->setForceFramePointer(true);
1601 MachineFrameInfo *MFI = MF.getFrameInfo();
1602 bool Is64Bit = Subtarget->is64Bit();
1603 bool IsWin64 = Subtarget->isTargetWin64();
1605 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1606 "Var args not supported with calling convention fastcc or ghc");
1608 // Assign locations to all of the incoming arguments.
1609 SmallVector<CCValAssign, 16> ArgLocs;
1610 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1611 ArgLocs, *DAG.getContext());
1612 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1614 unsigned LastVal = ~0U;
1616 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1617 CCValAssign &VA = ArgLocs[i];
1618 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1620 assert(VA.getValNo() != LastVal &&
1621 "Don't support value assigned to multiple locs yet");
1622 LastVal = VA.getValNo();
1624 if (VA.isRegLoc()) {
1625 EVT RegVT = VA.getLocVT();
1626 TargetRegisterClass *RC = NULL;
1627 if (RegVT == MVT::i32)
1628 RC = X86::GR32RegisterClass;
1629 else if (Is64Bit && RegVT == MVT::i64)
1630 RC = X86::GR64RegisterClass;
1631 else if (RegVT == MVT::f32)
1632 RC = X86::FR32RegisterClass;
1633 else if (RegVT == MVT::f64)
1634 RC = X86::FR64RegisterClass;
1635 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1636 RC = X86::VR128RegisterClass;
1637 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1638 RC = X86::VR64RegisterClass;
1640 llvm_unreachable("Unknown argument type!");
1642 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1643 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1645 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1646 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1648 if (VA.getLocInfo() == CCValAssign::SExt)
1649 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1650 DAG.getValueType(VA.getValVT()));
1651 else if (VA.getLocInfo() == CCValAssign::ZExt)
1652 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1653 DAG.getValueType(VA.getValVT()));
1654 else if (VA.getLocInfo() == CCValAssign::BCvt)
1655 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1657 if (VA.isExtInLoc()) {
1658 // Handle MMX values passed in XMM regs.
1659 if (RegVT.isVector()) {
1660 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1661 ArgValue, DAG.getConstant(0, MVT::i64));
1662 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1664 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1667 assert(VA.isMemLoc());
1668 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1671 // If value is passed via pointer - do a load.
1672 if (VA.getLocInfo() == CCValAssign::Indirect)
1673 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1676 InVals.push_back(ArgValue);
1679 // The x86-64 ABI for returning structs by value requires that we copy
1680 // the sret argument into %rax for the return. Save the argument into
1681 // a virtual register so that we can access it from the return points.
1682 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1683 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1684 unsigned Reg = FuncInfo->getSRetReturnReg();
1686 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1687 FuncInfo->setSRetReturnReg(Reg);
1689 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1690 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1693 unsigned StackSize = CCInfo.getNextStackOffset();
1694 // Align stack specially for tail calls.
1695 if (FuncIsMadeTailCallSafe(CallConv))
1696 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1698 // If the function takes variable number of arguments, make a frame index for
1699 // the start of the first vararg value... for expansion of llvm.va_start.
1701 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1702 CallConv != CallingConv::X86_ThisCall)) {
1703 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1706 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1708 // FIXME: We should really autogenerate these arrays
1709 static const unsigned GPR64ArgRegsWin64[] = {
1710 X86::RCX, X86::RDX, X86::R8, X86::R9
1712 static const unsigned XMMArgRegsWin64[] = {
1713 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1715 static const unsigned GPR64ArgRegs64Bit[] = {
1716 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1718 static const unsigned XMMArgRegs64Bit[] = {
1719 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1720 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1722 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1725 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1726 GPR64ArgRegs = GPR64ArgRegsWin64;
1727 XMMArgRegs = XMMArgRegsWin64;
1729 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1730 GPR64ArgRegs = GPR64ArgRegs64Bit;
1731 XMMArgRegs = XMMArgRegs64Bit;
1733 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1735 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1738 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1739 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1740 "SSE register cannot be used when SSE is disabled!");
1741 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1742 "SSE register cannot be used when SSE is disabled!");
1743 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1744 // Kernel mode asks for SSE to be disabled, so don't push them
1746 TotalNumXMMRegs = 0;
1748 // For X86-64, if there are vararg parameters that are passed via
1749 // registers, then we must store them to their spots on the stack so they
1750 // may be loaded by deferencing the result of va_next.
1751 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1752 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1753 FuncInfo->setRegSaveFrameIndex(
1754 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1757 // Store the integer parameter registers.
1758 SmallVector<SDValue, 8> MemOps;
1759 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1761 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1762 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1763 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1764 DAG.getIntPtrConstant(Offset));
1765 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1766 X86::GR64RegisterClass);
1767 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1769 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1770 PseudoSourceValue::getFixedStack(
1771 FuncInfo->getRegSaveFrameIndex()),
1772 Offset, false, false, 0);
1773 MemOps.push_back(Store);
1777 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1778 // Now store the XMM (fp + vector) parameter registers.
1779 SmallVector<SDValue, 11> SaveXMMOps;
1780 SaveXMMOps.push_back(Chain);
1782 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1783 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1784 SaveXMMOps.push_back(ALVal);
1786 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1787 FuncInfo->getRegSaveFrameIndex()));
1788 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1789 FuncInfo->getVarArgsFPOffset()));
1791 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1792 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1793 X86::VR128RegisterClass);
1794 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1795 SaveXMMOps.push_back(Val);
1797 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1799 &SaveXMMOps[0], SaveXMMOps.size()));
1802 if (!MemOps.empty())
1803 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1804 &MemOps[0], MemOps.size());
1808 // Some CCs need callee pop.
1809 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1810 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1812 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1813 // If this is an sret function, the return should pop the hidden pointer.
1814 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1815 FuncInfo->setBytesToPopOnReturn(4);
1819 // RegSaveFrameIndex is X86-64 only.
1820 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1821 if (CallConv == CallingConv::X86_FastCall ||
1822 CallConv == CallingConv::X86_ThisCall)
1823 // fastcc functions can't have varargs.
1824 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1831 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1832 SDValue StackPtr, SDValue Arg,
1833 DebugLoc dl, SelectionDAG &DAG,
1834 const CCValAssign &VA,
1835 ISD::ArgFlagsTy Flags) const {
1836 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1837 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1838 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1839 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1840 if (Flags.isByVal()) {
1841 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1843 return DAG.getStore(Chain, dl, Arg, PtrOff,
1844 PseudoSourceValue::getStack(), LocMemOffset,
1848 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1849 /// optimization is performed and it is required.
1851 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1852 SDValue &OutRetAddr, SDValue Chain,
1853 bool IsTailCall, bool Is64Bit,
1854 int FPDiff, DebugLoc dl) const {
1855 // Adjust the Return address stack slot.
1856 EVT VT = getPointerTy();
1857 OutRetAddr = getReturnAddressFrameIndex(DAG);
1859 // Load the "old" Return address.
1860 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1861 return SDValue(OutRetAddr.getNode(), 1);
1864 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1865 /// optimization is performed and it is required (FPDiff!=0).
1867 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1868 SDValue Chain, SDValue RetAddrFrIdx,
1869 bool Is64Bit, int FPDiff, DebugLoc dl) {
1870 // Store the return address to the appropriate stack slot.
1871 if (!FPDiff) return Chain;
1872 // Calculate the new stack slot for the return address.
1873 int SlotSize = Is64Bit ? 8 : 4;
1874 int NewReturnAddrFI =
1875 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1876 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1877 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1878 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1879 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1885 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1886 CallingConv::ID CallConv, bool isVarArg,
1888 const SmallVectorImpl<ISD::OutputArg> &Outs,
1889 const SmallVectorImpl<SDValue> &OutVals,
1890 const SmallVectorImpl<ISD::InputArg> &Ins,
1891 DebugLoc dl, SelectionDAG &DAG,
1892 SmallVectorImpl<SDValue> &InVals) const {
1893 MachineFunction &MF = DAG.getMachineFunction();
1894 bool Is64Bit = Subtarget->is64Bit();
1895 bool IsStructRet = CallIsStructReturn(Outs);
1896 bool IsSibcall = false;
1899 // Check if it's really possible to do a tail call.
1900 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1901 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1902 Outs, OutVals, Ins, DAG);
1904 // Sibcalls are automatically detected tailcalls which do not require
1906 if (!GuaranteedTailCallOpt && isTailCall)
1913 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1914 "Var args not supported with calling convention fastcc or ghc");
1916 // Analyze operands of the call, assigning locations to each operand.
1917 SmallVector<CCValAssign, 16> ArgLocs;
1918 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1919 ArgLocs, *DAG.getContext());
1920 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1922 // Get a count of how many bytes are to be pushed on the stack.
1923 unsigned NumBytes = CCInfo.getNextStackOffset();
1925 // This is a sibcall. The memory operands are available in caller's
1926 // own caller's stack.
1928 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1929 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1932 if (isTailCall && !IsSibcall) {
1933 // Lower arguments at fp - stackoffset + fpdiff.
1934 unsigned NumBytesCallerPushed =
1935 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1936 FPDiff = NumBytesCallerPushed - NumBytes;
1938 // Set the delta of movement of the returnaddr stackslot.
1939 // But only set if delta is greater than previous delta.
1940 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1941 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1945 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1947 SDValue RetAddrFrIdx;
1948 // Load return adress for tail calls.
1949 if (isTailCall && FPDiff)
1950 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1951 Is64Bit, FPDiff, dl);
1953 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1954 SmallVector<SDValue, 8> MemOpChains;
1957 // Walk the register/memloc assignments, inserting copies/loads. In the case
1958 // of tail call optimization arguments are handle later.
1959 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1960 CCValAssign &VA = ArgLocs[i];
1961 EVT RegVT = VA.getLocVT();
1962 SDValue Arg = OutVals[i];
1963 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1964 bool isByVal = Flags.isByVal();
1966 // Promote the value if needed.
1967 switch (VA.getLocInfo()) {
1968 default: llvm_unreachable("Unknown loc info!");
1969 case CCValAssign::Full: break;
1970 case CCValAssign::SExt:
1971 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1973 case CCValAssign::ZExt:
1974 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1976 case CCValAssign::AExt:
1977 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1978 // Special case: passing MMX values in XMM registers.
1979 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1980 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1981 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1983 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1985 case CCValAssign::BCvt:
1986 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1988 case CCValAssign::Indirect: {
1989 // Store the argument.
1990 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1991 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1992 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1993 PseudoSourceValue::getFixedStack(FI), 0,
2000 if (VA.isRegLoc()) {
2001 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2002 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2003 assert(VA.isMemLoc());
2004 if (StackPtr.getNode() == 0)
2005 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2006 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2007 dl, DAG, VA, Flags));
2011 if (!MemOpChains.empty())
2012 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2013 &MemOpChains[0], MemOpChains.size());
2015 // Build a sequence of copy-to-reg nodes chained together with token chain
2016 // and flag operands which copy the outgoing args into registers.
2018 // Tail call byval lowering might overwrite argument registers so in case of
2019 // tail call optimization the copies to registers are lowered later.
2021 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2022 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2023 RegsToPass[i].second, InFlag);
2024 InFlag = Chain.getValue(1);
2027 if (Subtarget->isPICStyleGOT()) {
2028 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2031 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2032 DAG.getNode(X86ISD::GlobalBaseReg,
2033 DebugLoc(), getPointerTy()),
2035 InFlag = Chain.getValue(1);
2037 // If we are tail calling and generating PIC/GOT style code load the
2038 // address of the callee into ECX. The value in ecx is used as target of
2039 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2040 // for tail calls on PIC/GOT architectures. Normally we would just put the
2041 // address of GOT into ebx and then call target@PLT. But for tail calls
2042 // ebx would be restored (since ebx is callee saved) before jumping to the
2045 // Note: The actual moving to ECX is done further down.
2046 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2047 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2048 !G->getGlobal()->hasProtectedVisibility())
2049 Callee = LowerGlobalAddress(Callee, DAG);
2050 else if (isa<ExternalSymbolSDNode>(Callee))
2051 Callee = LowerExternalSymbol(Callee, DAG);
2055 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
2056 // From AMD64 ABI document:
2057 // For calls that may call functions that use varargs or stdargs
2058 // (prototype-less calls or calls to functions containing ellipsis (...) in
2059 // the declaration) %al is used as hidden argument to specify the number
2060 // of SSE registers used. The contents of %al do not need to match exactly
2061 // the number of registers, but must be an ubound on the number of SSE
2062 // registers used and is in the range 0 - 8 inclusive.
2064 // Count the number of XMM registers allocated.
2065 static const unsigned XMMArgRegs[] = {
2066 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2067 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2069 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2070 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2071 && "SSE registers cannot be used when SSE is disabled");
2073 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2074 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2075 InFlag = Chain.getValue(1);
2079 // For tail calls lower the arguments to the 'real' stack slot.
2081 // Force all the incoming stack arguments to be loaded from the stack
2082 // before any new outgoing arguments are stored to the stack, because the
2083 // outgoing stack slots may alias the incoming argument stack slots, and
2084 // the alias isn't otherwise explicit. This is slightly more conservative
2085 // than necessary, because it means that each store effectively depends
2086 // on every argument instead of just those arguments it would clobber.
2087 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2089 SmallVector<SDValue, 8> MemOpChains2;
2092 // Do not flag preceeding copytoreg stuff together with the following stuff.
2094 if (GuaranteedTailCallOpt) {
2095 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2096 CCValAssign &VA = ArgLocs[i];
2099 assert(VA.isMemLoc());
2100 SDValue Arg = OutVals[i];
2101 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2102 // Create frame index.
2103 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2104 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2105 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2106 FIN = DAG.getFrameIndex(FI, getPointerTy());
2108 if (Flags.isByVal()) {
2109 // Copy relative to framepointer.
2110 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2111 if (StackPtr.getNode() == 0)
2112 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2114 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2116 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2120 // Store relative to framepointer.
2121 MemOpChains2.push_back(
2122 DAG.getStore(ArgChain, dl, Arg, FIN,
2123 PseudoSourceValue::getFixedStack(FI), 0,
2129 if (!MemOpChains2.empty())
2130 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2131 &MemOpChains2[0], MemOpChains2.size());
2133 // Copy arguments to their registers.
2134 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2135 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2136 RegsToPass[i].second, InFlag);
2137 InFlag = Chain.getValue(1);
2141 // Store the return address to the appropriate stack slot.
2142 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2146 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2147 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2148 // In the 64-bit large code model, we have to make all calls
2149 // through a register, since the call instruction's 32-bit
2150 // pc-relative offset may not be large enough to hold the whole
2152 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2153 // If the callee is a GlobalAddress node (quite common, every direct call
2154 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2157 // We should use extra load for direct calls to dllimported functions in
2159 const GlobalValue *GV = G->getGlobal();
2160 if (!GV->hasDLLImportLinkage()) {
2161 unsigned char OpFlags = 0;
2163 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2164 // external symbols most go through the PLT in PIC mode. If the symbol
2165 // has hidden or protected visibility, or if it is static or local, then
2166 // we don't need to use the PLT - we can directly call it.
2167 if (Subtarget->isTargetELF() &&
2168 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2169 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2170 OpFlags = X86II::MO_PLT;
2171 } else if (Subtarget->isPICStyleStubAny() &&
2172 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2173 Subtarget->getDarwinVers() < 9) {
2174 // PC-relative references to external symbols should go through $stub,
2175 // unless we're building with the leopard linker or later, which
2176 // automatically synthesizes these stubs.
2177 OpFlags = X86II::MO_DARWIN_STUB;
2180 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2181 G->getOffset(), OpFlags);
2183 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2184 unsigned char OpFlags = 0;
2186 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2187 // symbols should go through the PLT.
2188 if (Subtarget->isTargetELF() &&
2189 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2190 OpFlags = X86II::MO_PLT;
2191 } else if (Subtarget->isPICStyleStubAny() &&
2192 Subtarget->getDarwinVers() < 9) {
2193 // PC-relative references to external symbols should go through $stub,
2194 // unless we're building with the leopard linker or later, which
2195 // automatically synthesizes these stubs.
2196 OpFlags = X86II::MO_DARWIN_STUB;
2199 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2203 // Returns a chain & a flag for retval copy to use.
2204 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2205 SmallVector<SDValue, 8> Ops;
2207 if (!IsSibcall && isTailCall) {
2208 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2209 DAG.getIntPtrConstant(0, true), InFlag);
2210 InFlag = Chain.getValue(1);
2213 Ops.push_back(Chain);
2214 Ops.push_back(Callee);
2217 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2219 // Add argument registers to the end of the list so that they are known live
2221 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2222 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2223 RegsToPass[i].second.getValueType()));
2225 // Add an implicit use GOT pointer in EBX.
2226 if (!isTailCall && Subtarget->isPICStyleGOT())
2227 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2229 // Add an implicit use of AL for x86 vararg functions.
2230 if (Is64Bit && isVarArg)
2231 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2233 if (InFlag.getNode())
2234 Ops.push_back(InFlag);
2238 //// If this is the first return lowered for this function, add the regs
2239 //// to the liveout set for the function.
2240 // This isn't right, although it's probably harmless on x86; liveouts
2241 // should be computed from returns not tail calls. Consider a void
2242 // function making a tail call to a function returning int.
2243 return DAG.getNode(X86ISD::TC_RETURN, dl,
2244 NodeTys, &Ops[0], Ops.size());
2247 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2248 InFlag = Chain.getValue(1);
2250 // Create the CALLSEQ_END node.
2251 unsigned NumBytesForCalleeToPush;
2252 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2253 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2254 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2255 // If this is a call to a struct-return function, the callee
2256 // pops the hidden struct pointer, so we have to push it back.
2257 // This is common for Darwin/X86, Linux & Mingw32 targets.
2258 NumBytesForCalleeToPush = 4;
2260 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2262 // Returns a flag for retval copy to use.
2264 Chain = DAG.getCALLSEQ_END(Chain,
2265 DAG.getIntPtrConstant(NumBytes, true),
2266 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2269 InFlag = Chain.getValue(1);
2272 // Handle result values, copying them out of physregs into vregs that we
2274 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2275 Ins, dl, DAG, InVals);
2279 //===----------------------------------------------------------------------===//
2280 // Fast Calling Convention (tail call) implementation
2281 //===----------------------------------------------------------------------===//
2283 // Like std call, callee cleans arguments, convention except that ECX is
2284 // reserved for storing the tail called function address. Only 2 registers are
2285 // free for argument passing (inreg). Tail call optimization is performed
2287 // * tailcallopt is enabled
2288 // * caller/callee are fastcc
2289 // On X86_64 architecture with GOT-style position independent code only local
2290 // (within module) calls are supported at the moment.
2291 // To keep the stack aligned according to platform abi the function
2292 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2293 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2294 // If a tail called function callee has more arguments than the caller the
2295 // caller needs to make sure that there is room to move the RETADDR to. This is
2296 // achieved by reserving an area the size of the argument delta right after the
2297 // original REtADDR, but before the saved framepointer or the spilled registers
2298 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2310 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2311 /// for a 16 byte align requirement.
2313 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2314 SelectionDAG& DAG) const {
2315 MachineFunction &MF = DAG.getMachineFunction();
2316 const TargetMachine &TM = MF.getTarget();
2317 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2318 unsigned StackAlignment = TFI.getStackAlignment();
2319 uint64_t AlignMask = StackAlignment - 1;
2320 int64_t Offset = StackSize;
2321 uint64_t SlotSize = TD->getPointerSize();
2322 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2323 // Number smaller than 12 so just add the difference.
2324 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2326 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2327 Offset = ((~AlignMask) & Offset) + StackAlignment +
2328 (StackAlignment-SlotSize);
2333 /// MatchingStackOffset - Return true if the given stack call argument is
2334 /// already available in the same position (relatively) of the caller's
2335 /// incoming argument stack.
2337 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2338 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2339 const X86InstrInfo *TII) {
2340 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2342 if (Arg.getOpcode() == ISD::CopyFromReg) {
2343 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2344 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2346 MachineInstr *Def = MRI->getVRegDef(VR);
2349 if (!Flags.isByVal()) {
2350 if (!TII->isLoadFromStackSlot(Def, FI))
2353 unsigned Opcode = Def->getOpcode();
2354 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2355 Def->getOperand(1).isFI()) {
2356 FI = Def->getOperand(1).getIndex();
2357 Bytes = Flags.getByValSize();
2361 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2362 if (Flags.isByVal())
2363 // ByVal argument is passed in as a pointer but it's now being
2364 // dereferenced. e.g.
2365 // define @foo(%struct.X* %A) {
2366 // tail call @bar(%struct.X* byval %A)
2369 SDValue Ptr = Ld->getBasePtr();
2370 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2373 FI = FINode->getIndex();
2377 assert(FI != INT_MAX);
2378 if (!MFI->isFixedObjectIndex(FI))
2380 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2383 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2384 /// for tail call optimization. Targets which want to do tail call
2385 /// optimization should implement this function.
2387 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2388 CallingConv::ID CalleeCC,
2390 bool isCalleeStructRet,
2391 bool isCallerStructRet,
2392 const SmallVectorImpl<ISD::OutputArg> &Outs,
2393 const SmallVectorImpl<SDValue> &OutVals,
2394 const SmallVectorImpl<ISD::InputArg> &Ins,
2395 SelectionDAG& DAG) const {
2396 if (!IsTailCallConvention(CalleeCC) &&
2397 CalleeCC != CallingConv::C)
2400 // If -tailcallopt is specified, make fastcc functions tail-callable.
2401 const MachineFunction &MF = DAG.getMachineFunction();
2402 const Function *CallerF = DAG.getMachineFunction().getFunction();
2403 CallingConv::ID CallerCC = CallerF->getCallingConv();
2404 bool CCMatch = CallerCC == CalleeCC;
2406 if (GuaranteedTailCallOpt) {
2407 if (IsTailCallConvention(CalleeCC) && CCMatch)
2412 // Look for obvious safe cases to perform tail call optimization that do not
2413 // require ABI changes. This is what gcc calls sibcall.
2415 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2416 // emit a special epilogue.
2417 if (RegInfo->needsStackRealignment(MF))
2420 // Do not sibcall optimize vararg calls unless the call site is not passing
2422 if (isVarArg && !Outs.empty())
2425 // Also avoid sibcall optimization if either caller or callee uses struct
2426 // return semantics.
2427 if (isCalleeStructRet || isCallerStructRet)
2430 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2431 // Therefore if it's not used by the call it is not safe to optimize this into
2433 bool Unused = false;
2434 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2441 SmallVector<CCValAssign, 16> RVLocs;
2442 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2443 RVLocs, *DAG.getContext());
2444 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2445 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2446 CCValAssign &VA = RVLocs[i];
2447 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2452 // If the calling conventions do not match, then we'd better make sure the
2453 // results are returned in the same way as what the caller expects.
2455 SmallVector<CCValAssign, 16> RVLocs1;
2456 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2457 RVLocs1, *DAG.getContext());
2458 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2460 SmallVector<CCValAssign, 16> RVLocs2;
2461 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2462 RVLocs2, *DAG.getContext());
2463 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2465 if (RVLocs1.size() != RVLocs2.size())
2467 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2468 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2470 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2472 if (RVLocs1[i].isRegLoc()) {
2473 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2476 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2482 // If the callee takes no arguments then go on to check the results of the
2484 if (!Outs.empty()) {
2485 // Check if stack adjustment is needed. For now, do not do this if any
2486 // argument is passed on the stack.
2487 SmallVector<CCValAssign, 16> ArgLocs;
2488 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2489 ArgLocs, *DAG.getContext());
2490 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2491 if (CCInfo.getNextStackOffset()) {
2492 MachineFunction &MF = DAG.getMachineFunction();
2493 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2495 if (Subtarget->isTargetWin64())
2496 // Win64 ABI has additional complications.
2499 // Check if the arguments are already laid out in the right way as
2500 // the caller's fixed stack objects.
2501 MachineFrameInfo *MFI = MF.getFrameInfo();
2502 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2503 const X86InstrInfo *TII =
2504 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2505 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2506 CCValAssign &VA = ArgLocs[i];
2507 SDValue Arg = OutVals[i];
2508 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2509 if (VA.getLocInfo() == CCValAssign::Indirect)
2511 if (!VA.isRegLoc()) {
2512 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2519 // If the tailcall address may be in a register, then make sure it's
2520 // possible to register allocate for it. In 32-bit, the call address can
2521 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2522 // callee-saved registers are restored. These happen to be the same
2523 // registers used to pass 'inreg' arguments so watch out for those.
2524 if (!Subtarget->is64Bit() &&
2525 !isa<GlobalAddressSDNode>(Callee) &&
2526 !isa<ExternalSymbolSDNode>(Callee)) {
2527 unsigned NumInRegs = 0;
2528 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2529 CCValAssign &VA = ArgLocs[i];
2532 unsigned Reg = VA.getLocReg();
2535 case X86::EAX: case X86::EDX: case X86::ECX:
2536 if (++NumInRegs == 3)
2548 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2549 return X86::createFastISel(funcInfo);
2553 //===----------------------------------------------------------------------===//
2554 // Other Lowering Hooks
2555 //===----------------------------------------------------------------------===//
2558 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2559 MachineFunction &MF = DAG.getMachineFunction();
2560 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2561 int ReturnAddrIndex = FuncInfo->getRAIndex();
2563 if (ReturnAddrIndex == 0) {
2564 // Set up a frame object for the return address.
2565 uint64_t SlotSize = TD->getPointerSize();
2566 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2568 FuncInfo->setRAIndex(ReturnAddrIndex);
2571 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2575 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2576 bool hasSymbolicDisplacement) {
2577 // Offset should fit into 32 bit immediate field.
2578 if (!isInt<32>(Offset))
2581 // If we don't have a symbolic displacement - we don't have any extra
2583 if (!hasSymbolicDisplacement)
2586 // FIXME: Some tweaks might be needed for medium code model.
2587 if (M != CodeModel::Small && M != CodeModel::Kernel)
2590 // For small code model we assume that latest object is 16MB before end of 31
2591 // bits boundary. We may also accept pretty large negative constants knowing
2592 // that all objects are in the positive half of address space.
2593 if (M == CodeModel::Small && Offset < 16*1024*1024)
2596 // For kernel code model we know that all object resist in the negative half
2597 // of 32bits address space. We may not accept negative offsets, since they may
2598 // be just off and we may accept pretty large positive ones.
2599 if (M == CodeModel::Kernel && Offset > 0)
2605 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2606 /// specific condition code, returning the condition code and the LHS/RHS of the
2607 /// comparison to make.
2608 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2609 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2611 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2612 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2613 // X > -1 -> X == 0, jump !sign.
2614 RHS = DAG.getConstant(0, RHS.getValueType());
2615 return X86::COND_NS;
2616 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2617 // X < 0 -> X == 0, jump on sign.
2619 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2621 RHS = DAG.getConstant(0, RHS.getValueType());
2622 return X86::COND_LE;
2626 switch (SetCCOpcode) {
2627 default: llvm_unreachable("Invalid integer condition!");
2628 case ISD::SETEQ: return X86::COND_E;
2629 case ISD::SETGT: return X86::COND_G;
2630 case ISD::SETGE: return X86::COND_GE;
2631 case ISD::SETLT: return X86::COND_L;
2632 case ISD::SETLE: return X86::COND_LE;
2633 case ISD::SETNE: return X86::COND_NE;
2634 case ISD::SETULT: return X86::COND_B;
2635 case ISD::SETUGT: return X86::COND_A;
2636 case ISD::SETULE: return X86::COND_BE;
2637 case ISD::SETUGE: return X86::COND_AE;
2641 // First determine if it is required or is profitable to flip the operands.
2643 // If LHS is a foldable load, but RHS is not, flip the condition.
2644 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2645 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2646 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2647 std::swap(LHS, RHS);
2650 switch (SetCCOpcode) {
2656 std::swap(LHS, RHS);
2660 // On a floating point condition, the flags are set as follows:
2662 // 0 | 0 | 0 | X > Y
2663 // 0 | 0 | 1 | X < Y
2664 // 1 | 0 | 0 | X == Y
2665 // 1 | 1 | 1 | unordered
2666 switch (SetCCOpcode) {
2667 default: llvm_unreachable("Condcode should be pre-legalized away");
2669 case ISD::SETEQ: return X86::COND_E;
2670 case ISD::SETOLT: // flipped
2672 case ISD::SETGT: return X86::COND_A;
2673 case ISD::SETOLE: // flipped
2675 case ISD::SETGE: return X86::COND_AE;
2676 case ISD::SETUGT: // flipped
2678 case ISD::SETLT: return X86::COND_B;
2679 case ISD::SETUGE: // flipped
2681 case ISD::SETLE: return X86::COND_BE;
2683 case ISD::SETNE: return X86::COND_NE;
2684 case ISD::SETUO: return X86::COND_P;
2685 case ISD::SETO: return X86::COND_NP;
2687 case ISD::SETUNE: return X86::COND_INVALID;
2691 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2692 /// code. Current x86 isa includes the following FP cmov instructions:
2693 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2694 static bool hasFPCMov(unsigned X86CC) {
2710 /// isFPImmLegal - Returns true if the target can instruction select the
2711 /// specified FP immediate natively. If false, the legalizer will
2712 /// materialize the FP immediate as a load from a constant pool.
2713 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2714 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2715 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2721 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2722 /// the specified range (L, H].
2723 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2724 return (Val < 0) || (Val >= Low && Val < Hi);
2727 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2728 /// specified value.
2729 static bool isUndefOrEqual(int Val, int CmpVal) {
2730 if (Val < 0 || Val == CmpVal)
2735 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2736 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2737 /// the second operand.
2738 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2739 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2740 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2741 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2742 return (Mask[0] < 2 && Mask[1] < 2);
2746 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2747 SmallVector<int, 8> M;
2749 return ::isPSHUFDMask(M, N->getValueType(0));
2752 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2753 /// is suitable for input to PSHUFHW.
2754 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2755 if (VT != MVT::v8i16)
2758 // Lower quadword copied in order or undef.
2759 for (int i = 0; i != 4; ++i)
2760 if (Mask[i] >= 0 && Mask[i] != i)
2763 // Upper quadword shuffled.
2764 for (int i = 4; i != 8; ++i)
2765 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2771 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2772 SmallVector<int, 8> M;
2774 return ::isPSHUFHWMask(M, N->getValueType(0));
2777 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2778 /// is suitable for input to PSHUFLW.
2779 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2780 if (VT != MVT::v8i16)
2783 // Upper quadword copied in order.
2784 for (int i = 4; i != 8; ++i)
2785 if (Mask[i] >= 0 && Mask[i] != i)
2788 // Lower quadword shuffled.
2789 for (int i = 0; i != 4; ++i)
2796 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2797 SmallVector<int, 8> M;
2799 return ::isPSHUFLWMask(M, N->getValueType(0));
2802 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2803 /// is suitable for input to PALIGNR.
2804 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2806 int i, e = VT.getVectorNumElements();
2808 // Do not handle v2i64 / v2f64 shuffles with palignr.
2809 if (e < 4 || !hasSSSE3)
2812 for (i = 0; i != e; ++i)
2816 // All undef, not a palignr.
2820 // Determine if it's ok to perform a palignr with only the LHS, since we
2821 // don't have access to the actual shuffle elements to see if RHS is undef.
2822 bool Unary = Mask[i] < (int)e;
2823 bool NeedsUnary = false;
2825 int s = Mask[i] - i;
2827 // Check the rest of the elements to see if they are consecutive.
2828 for (++i; i != e; ++i) {
2833 Unary = Unary && (m < (int)e);
2834 NeedsUnary = NeedsUnary || (m < s);
2836 if (NeedsUnary && !Unary)
2838 if (Unary && m != ((s+i) & (e-1)))
2840 if (!Unary && m != (s+i))
2846 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2847 SmallVector<int, 8> M;
2849 return ::isPALIGNRMask(M, N->getValueType(0), true);
2852 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2853 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2854 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2855 int NumElems = VT.getVectorNumElements();
2856 if (NumElems != 2 && NumElems != 4)
2859 int Half = NumElems / 2;
2860 for (int i = 0; i < Half; ++i)
2861 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2863 for (int i = Half; i < NumElems; ++i)
2864 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2870 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2871 SmallVector<int, 8> M;
2873 return ::isSHUFPMask(M, N->getValueType(0));
2876 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2877 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2878 /// half elements to come from vector 1 (which would equal the dest.) and
2879 /// the upper half to come from vector 2.
2880 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2881 int NumElems = VT.getVectorNumElements();
2883 if (NumElems != 2 && NumElems != 4)
2886 int Half = NumElems / 2;
2887 for (int i = 0; i < Half; ++i)
2888 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2890 for (int i = Half; i < NumElems; ++i)
2891 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2896 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2897 SmallVector<int, 8> M;
2899 return isCommutedSHUFPMask(M, N->getValueType(0));
2902 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2903 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2904 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2905 if (N->getValueType(0).getVectorNumElements() != 4)
2908 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2909 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2910 isUndefOrEqual(N->getMaskElt(1), 7) &&
2911 isUndefOrEqual(N->getMaskElt(2), 2) &&
2912 isUndefOrEqual(N->getMaskElt(3), 3);
2915 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2916 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2918 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2919 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2924 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2925 isUndefOrEqual(N->getMaskElt(1), 3) &&
2926 isUndefOrEqual(N->getMaskElt(2), 2) &&
2927 isUndefOrEqual(N->getMaskElt(3), 3);
2930 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2931 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2932 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2933 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2935 if (NumElems != 2 && NumElems != 4)
2938 for (unsigned i = 0; i < NumElems/2; ++i)
2939 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2942 for (unsigned i = NumElems/2; i < NumElems; ++i)
2943 if (!isUndefOrEqual(N->getMaskElt(i), i))
2949 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2950 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2951 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2952 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2954 if (NumElems != 2 && NumElems != 4)
2957 for (unsigned i = 0; i < NumElems/2; ++i)
2958 if (!isUndefOrEqual(N->getMaskElt(i), i))
2961 for (unsigned i = 0; i < NumElems/2; ++i)
2962 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2968 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2969 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2970 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2971 bool V2IsSplat = false) {
2972 int NumElts = VT.getVectorNumElements();
2973 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2976 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2978 int BitI1 = Mask[i+1];
2979 if (!isUndefOrEqual(BitI, j))
2982 if (!isUndefOrEqual(BitI1, NumElts))
2985 if (!isUndefOrEqual(BitI1, j + NumElts))
2992 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2993 SmallVector<int, 8> M;
2995 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2998 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2999 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3000 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3001 bool V2IsSplat = false) {
3002 int NumElts = VT.getVectorNumElements();
3003 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3006 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3008 int BitI1 = Mask[i+1];
3009 if (!isUndefOrEqual(BitI, j + NumElts/2))
3012 if (isUndefOrEqual(BitI1, NumElts))
3015 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3022 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3023 SmallVector<int, 8> M;
3025 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3028 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3029 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3031 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3032 int NumElems = VT.getVectorNumElements();
3033 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3036 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3038 int BitI1 = Mask[i+1];
3039 if (!isUndefOrEqual(BitI, j))
3041 if (!isUndefOrEqual(BitI1, j))
3047 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3048 SmallVector<int, 8> M;
3050 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3053 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3054 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3056 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3057 int NumElems = VT.getVectorNumElements();
3058 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3061 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3063 int BitI1 = Mask[i+1];
3064 if (!isUndefOrEqual(BitI, j))
3066 if (!isUndefOrEqual(BitI1, j))
3072 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3073 SmallVector<int, 8> M;
3075 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3078 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3079 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3080 /// MOVSD, and MOVD, i.e. setting the lowest element.
3081 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3082 if (VT.getVectorElementType().getSizeInBits() < 32)
3085 int NumElts = VT.getVectorNumElements();
3087 if (!isUndefOrEqual(Mask[0], NumElts))
3090 for (int i = 1; i < NumElts; ++i)
3091 if (!isUndefOrEqual(Mask[i], i))
3097 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3098 SmallVector<int, 8> M;
3100 return ::isMOVLMask(M, N->getValueType(0));
3103 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3104 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3105 /// element of vector 2 and the other elements to come from vector 1 in order.
3106 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3107 bool V2IsSplat = false, bool V2IsUndef = false) {
3108 int NumOps = VT.getVectorNumElements();
3109 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3112 if (!isUndefOrEqual(Mask[0], 0))
3115 for (int i = 1; i < NumOps; ++i)
3116 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3117 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3118 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3124 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3125 bool V2IsUndef = false) {
3126 SmallVector<int, 8> M;
3128 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3131 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3132 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3133 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3134 if (N->getValueType(0).getVectorNumElements() != 4)
3137 // Expect 1, 1, 3, 3
3138 for (unsigned i = 0; i < 2; ++i) {
3139 int Elt = N->getMaskElt(i);
3140 if (Elt >= 0 && Elt != 1)
3145 for (unsigned i = 2; i < 4; ++i) {
3146 int Elt = N->getMaskElt(i);
3147 if (Elt >= 0 && Elt != 3)
3152 // Don't use movshdup if it can be done with a shufps.
3153 // FIXME: verify that matching u, u, 3, 3 is what we want.
3157 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3158 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3159 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3160 if (N->getValueType(0).getVectorNumElements() != 4)
3163 // Expect 0, 0, 2, 2
3164 for (unsigned i = 0; i < 2; ++i)
3165 if (N->getMaskElt(i) > 0)
3169 for (unsigned i = 2; i < 4; ++i) {
3170 int Elt = N->getMaskElt(i);
3171 if (Elt >= 0 && Elt != 2)
3176 // Don't use movsldup if it can be done with a shufps.
3180 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3181 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3182 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3183 int e = N->getValueType(0).getVectorNumElements() / 2;
3185 for (int i = 0; i < e; ++i)
3186 if (!isUndefOrEqual(N->getMaskElt(i), i))
3188 for (int i = 0; i < e; ++i)
3189 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3194 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3195 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3196 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3197 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3198 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3200 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3202 for (int i = 0; i < NumOperands; ++i) {
3203 int Val = SVOp->getMaskElt(NumOperands-i-1);
3204 if (Val < 0) Val = 0;
3205 if (Val >= NumOperands) Val -= NumOperands;
3207 if (i != NumOperands - 1)
3213 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3214 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3215 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3216 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3218 // 8 nodes, but we only care about the last 4.
3219 for (unsigned i = 7; i >= 4; --i) {
3220 int Val = SVOp->getMaskElt(i);
3229 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3230 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3231 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3232 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3234 // 8 nodes, but we only care about the first 4.
3235 for (int i = 3; i >= 0; --i) {
3236 int Val = SVOp->getMaskElt(i);
3245 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3246 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3247 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3248 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3249 EVT VVT = N->getValueType(0);
3250 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3254 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3255 Val = SVOp->getMaskElt(i);
3259 return (Val - i) * EltSize;
3262 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3264 bool X86::isZeroNode(SDValue Elt) {
3265 return ((isa<ConstantSDNode>(Elt) &&
3266 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3267 (isa<ConstantFPSDNode>(Elt) &&
3268 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3271 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3272 /// their permute mask.
3273 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3274 SelectionDAG &DAG) {
3275 EVT VT = SVOp->getValueType(0);
3276 unsigned NumElems = VT.getVectorNumElements();
3277 SmallVector<int, 8> MaskVec;
3279 for (unsigned i = 0; i != NumElems; ++i) {
3280 int idx = SVOp->getMaskElt(i);
3282 MaskVec.push_back(idx);
3283 else if (idx < (int)NumElems)
3284 MaskVec.push_back(idx + NumElems);
3286 MaskVec.push_back(idx - NumElems);
3288 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3289 SVOp->getOperand(0), &MaskVec[0]);
3292 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3293 /// the two vector operands have swapped position.
3294 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3295 unsigned NumElems = VT.getVectorNumElements();
3296 for (unsigned i = 0; i != NumElems; ++i) {
3300 else if (idx < (int)NumElems)
3301 Mask[i] = idx + NumElems;
3303 Mask[i] = idx - NumElems;
3307 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3308 /// match movhlps. The lower half elements should come from upper half of
3309 /// V1 (and in order), and the upper half elements should come from the upper
3310 /// half of V2 (and in order).
3311 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3312 if (Op->getValueType(0).getVectorNumElements() != 4)
3314 for (unsigned i = 0, e = 2; i != e; ++i)
3315 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3317 for (unsigned i = 2; i != 4; ++i)
3318 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3323 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3324 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3326 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3327 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3329 N = N->getOperand(0).getNode();
3330 if (!ISD::isNON_EXTLoad(N))
3333 *LD = cast<LoadSDNode>(N);
3337 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3338 /// match movlp{s|d}. The lower half elements should come from lower half of
3339 /// V1 (and in order), and the upper half elements should come from the upper
3340 /// half of V2 (and in order). And since V1 will become the source of the
3341 /// MOVLP, it must be either a vector load or a scalar load to vector.
3342 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3343 ShuffleVectorSDNode *Op) {
3344 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3346 // Is V2 is a vector load, don't do this transformation. We will try to use
3347 // load folding shufps op.
3348 if (ISD::isNON_EXTLoad(V2))
3351 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3353 if (NumElems != 2 && NumElems != 4)
3355 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3356 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3358 for (unsigned i = NumElems/2; i != NumElems; ++i)
3359 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3364 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3366 static bool isSplatVector(SDNode *N) {
3367 if (N->getOpcode() != ISD::BUILD_VECTOR)
3370 SDValue SplatValue = N->getOperand(0);
3371 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3372 if (N->getOperand(i) != SplatValue)
3377 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3378 /// to an zero vector.
3379 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3380 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3381 SDValue V1 = N->getOperand(0);
3382 SDValue V2 = N->getOperand(1);
3383 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3384 for (unsigned i = 0; i != NumElems; ++i) {
3385 int Idx = N->getMaskElt(i);
3386 if (Idx >= (int)NumElems) {
3387 unsigned Opc = V2.getOpcode();
3388 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3390 if (Opc != ISD::BUILD_VECTOR ||
3391 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3393 } else if (Idx >= 0) {
3394 unsigned Opc = V1.getOpcode();
3395 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3397 if (Opc != ISD::BUILD_VECTOR ||
3398 !X86::isZeroNode(V1.getOperand(Idx)))
3405 /// getZeroVector - Returns a vector of specified type with all zero elements.
3407 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3409 assert(VT.isVector() && "Expected a vector type");
3411 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3412 // type. This ensures they get CSE'd.
3414 if (VT.getSizeInBits() == 64) { // MMX
3415 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3416 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3417 } else if (HasSSE2) { // SSE2
3418 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3419 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3421 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3422 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3424 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3427 /// getOnesVector - Returns a vector of specified type with all bits set.
3429 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3430 assert(VT.isVector() && "Expected a vector type");
3432 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3433 // type. This ensures they get CSE'd.
3434 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3436 if (VT.getSizeInBits() == 64) // MMX
3437 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3439 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3440 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3444 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3445 /// that point to V2 points to its first element.
3446 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3447 EVT VT = SVOp->getValueType(0);
3448 unsigned NumElems = VT.getVectorNumElements();
3450 bool Changed = false;
3451 SmallVector<int, 8> MaskVec;
3452 SVOp->getMask(MaskVec);
3454 for (unsigned i = 0; i != NumElems; ++i) {
3455 if (MaskVec[i] > (int)NumElems) {
3456 MaskVec[i] = NumElems;
3461 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3462 SVOp->getOperand(1), &MaskVec[0]);
3463 return SDValue(SVOp, 0);
3466 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3467 /// operation of specified width.
3468 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3470 unsigned NumElems = VT.getVectorNumElements();
3471 SmallVector<int, 8> Mask;
3472 Mask.push_back(NumElems);
3473 for (unsigned i = 1; i != NumElems; ++i)
3475 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3478 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3479 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3481 unsigned NumElems = VT.getVectorNumElements();
3482 SmallVector<int, 8> Mask;
3483 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3485 Mask.push_back(i + NumElems);
3487 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3490 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3491 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3493 unsigned NumElems = VT.getVectorNumElements();
3494 unsigned Half = NumElems/2;
3495 SmallVector<int, 8> Mask;
3496 for (unsigned i = 0; i != Half; ++i) {
3497 Mask.push_back(i + Half);
3498 Mask.push_back(i + NumElems + Half);
3500 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3503 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3504 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3506 if (SV->getValueType(0).getVectorNumElements() <= 4)
3507 return SDValue(SV, 0);
3509 EVT PVT = MVT::v4f32;
3510 EVT VT = SV->getValueType(0);
3511 DebugLoc dl = SV->getDebugLoc();
3512 SDValue V1 = SV->getOperand(0);
3513 int NumElems = VT.getVectorNumElements();
3514 int EltNo = SV->getSplatIndex();
3516 // unpack elements to the correct location
3517 while (NumElems > 4) {
3518 if (EltNo < NumElems/2) {
3519 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3521 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3522 EltNo -= NumElems/2;
3527 // Perform the splat.
3528 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3529 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3530 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3531 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3534 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3535 /// vector of zero or undef vector. This produces a shuffle where the low
3536 /// element of V2 is swizzled into the zero/undef vector, landing at element
3537 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3538 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3539 bool isZero, bool HasSSE2,
3540 SelectionDAG &DAG) {
3541 EVT VT = V2.getValueType();
3543 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3544 unsigned NumElems = VT.getVectorNumElements();
3545 SmallVector<int, 16> MaskVec;
3546 for (unsigned i = 0; i != NumElems; ++i)
3547 // If this is the insertion idx, put the low elt of V2 here.
3548 MaskVec.push_back(i == Idx ? NumElems : i);
3549 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3552 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3553 /// a shuffle that is zero.
3555 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3556 bool Low, SelectionDAG &DAG) {
3557 unsigned NumZeros = 0;
3558 for (int i = 0; i < NumElems; ++i) {
3559 unsigned Index = Low ? i : NumElems-i-1;
3560 int Idx = SVOp->getMaskElt(Index);
3565 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3566 if (Elt.getNode() && X86::isZeroNode(Elt))
3574 /// isVectorShift - Returns true if the shuffle can be implemented as a
3575 /// logical left or right shift of a vector.
3576 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3577 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3578 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3579 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3582 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3585 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3589 bool SeenV1 = false;
3590 bool SeenV2 = false;
3591 for (unsigned i = NumZeros; i < NumElems; ++i) {
3592 unsigned Val = isLeft ? (i - NumZeros) : i;
3593 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3596 unsigned Idx = (unsigned) Idx_;
3606 if (SeenV1 && SeenV2)
3609 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3615 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3617 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3618 unsigned NumNonZero, unsigned NumZero,
3620 const TargetLowering &TLI) {
3624 DebugLoc dl = Op.getDebugLoc();
3627 for (unsigned i = 0; i < 16; ++i) {
3628 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3629 if (ThisIsNonZero && First) {
3631 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3633 V = DAG.getUNDEF(MVT::v8i16);
3638 SDValue ThisElt(0, 0), LastElt(0, 0);
3639 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3640 if (LastIsNonZero) {
3641 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3642 MVT::i16, Op.getOperand(i-1));
3644 if (ThisIsNonZero) {
3645 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3646 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3647 ThisElt, DAG.getConstant(8, MVT::i8));
3649 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3653 if (ThisElt.getNode())
3654 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3655 DAG.getIntPtrConstant(i/2));
3659 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3662 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3664 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3665 unsigned NumNonZero, unsigned NumZero,
3667 const TargetLowering &TLI) {
3671 DebugLoc dl = Op.getDebugLoc();
3674 for (unsigned i = 0; i < 8; ++i) {
3675 bool isNonZero = (NonZeros & (1 << i)) != 0;
3679 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3681 V = DAG.getUNDEF(MVT::v8i16);
3684 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3685 MVT::v8i16, V, Op.getOperand(i),
3686 DAG.getIntPtrConstant(i));
3693 /// getVShift - Return a vector logical shift node.
3695 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3696 unsigned NumBits, SelectionDAG &DAG,
3697 const TargetLowering &TLI, DebugLoc dl) {
3698 bool isMMX = VT.getSizeInBits() == 64;
3699 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3700 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3701 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3702 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3703 DAG.getNode(Opc, dl, ShVT, SrcOp,
3704 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3708 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3709 SelectionDAG &DAG) const {
3711 // Check if the scalar load can be widened into a vector load. And if
3712 // the address is "base + cst" see if the cst can be "absorbed" into
3713 // the shuffle mask.
3714 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3715 SDValue Ptr = LD->getBasePtr();
3716 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3718 EVT PVT = LD->getValueType(0);
3719 if (PVT != MVT::i32 && PVT != MVT::f32)
3724 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3725 FI = FINode->getIndex();
3727 } else if (Ptr.getOpcode() == ISD::ADD &&
3728 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3729 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3730 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3731 Offset = Ptr.getConstantOperandVal(1);
3732 Ptr = Ptr.getOperand(0);
3737 SDValue Chain = LD->getChain();
3738 // Make sure the stack object alignment is at least 16.
3739 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3740 if (DAG.InferPtrAlignment(Ptr) < 16) {
3741 if (MFI->isFixedObjectIndex(FI)) {
3742 // Can't change the alignment. FIXME: It's possible to compute
3743 // the exact stack offset and reference FI + adjust offset instead.
3744 // If someone *really* cares about this. That's the way to implement it.
3747 MFI->setObjectAlignment(FI, 16);
3751 // (Offset % 16) must be multiple of 4. Then address is then
3752 // Ptr + (Offset & ~15).
3755 if ((Offset % 16) & 3)
3757 int64_t StartOffset = Offset & ~15;
3759 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3760 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3762 int EltNo = (Offset - StartOffset) >> 2;
3763 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3764 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3765 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3767 // Canonicalize it to a v4i32 shuffle.
3768 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3769 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3770 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3771 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3777 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3778 /// vector of type 'VT', see if the elements can be replaced by a single large
3779 /// load which has the same value as a build_vector whose operands are 'elts'.
3781 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3783 /// FIXME: we'd also like to handle the case where the last elements are zero
3784 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3785 /// There's even a handy isZeroNode for that purpose.
3786 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3787 DebugLoc &dl, SelectionDAG &DAG) {
3788 EVT EltVT = VT.getVectorElementType();
3789 unsigned NumElems = Elts.size();
3791 LoadSDNode *LDBase = NULL;
3792 unsigned LastLoadedElt = -1U;
3794 // For each element in the initializer, see if we've found a load or an undef.
3795 // If we don't find an initial load element, or later load elements are
3796 // non-consecutive, bail out.
3797 for (unsigned i = 0; i < NumElems; ++i) {
3798 SDValue Elt = Elts[i];
3800 if (!Elt.getNode() ||
3801 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3804 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3806 LDBase = cast<LoadSDNode>(Elt.getNode());
3810 if (Elt.getOpcode() == ISD::UNDEF)
3813 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3814 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3819 // If we have found an entire vector of loads and undefs, then return a large
3820 // load of the entire vector width starting at the base pointer. If we found
3821 // consecutive loads for the low half, generate a vzext_load node.
3822 if (LastLoadedElt == NumElems - 1) {
3823 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3824 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3825 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3826 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3827 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3828 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3829 LDBase->isVolatile(), LDBase->isNonTemporal(),
3830 LDBase->getAlignment());
3831 } else if (NumElems == 4 && LastLoadedElt == 1) {
3832 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3833 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3834 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3835 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3841 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3842 DebugLoc dl = Op.getDebugLoc();
3843 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3844 if (ISD::isBuildVectorAllZeros(Op.getNode())
3845 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3846 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3847 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3848 // eliminated on x86-32 hosts.
3849 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3852 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3853 return getOnesVector(Op.getValueType(), DAG, dl);
3854 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3857 EVT VT = Op.getValueType();
3858 EVT ExtVT = VT.getVectorElementType();
3859 unsigned EVTBits = ExtVT.getSizeInBits();
3861 unsigned NumElems = Op.getNumOperands();
3862 unsigned NumZero = 0;
3863 unsigned NumNonZero = 0;
3864 unsigned NonZeros = 0;
3865 bool IsAllConstants = true;
3866 SmallSet<SDValue, 8> Values;
3867 for (unsigned i = 0; i < NumElems; ++i) {
3868 SDValue Elt = Op.getOperand(i);
3869 if (Elt.getOpcode() == ISD::UNDEF)
3872 if (Elt.getOpcode() != ISD::Constant &&
3873 Elt.getOpcode() != ISD::ConstantFP)
3874 IsAllConstants = false;
3875 if (X86::isZeroNode(Elt))
3878 NonZeros |= (1 << i);
3883 if (NumNonZero == 0) {
3884 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3885 return DAG.getUNDEF(VT);
3888 // Special case for single non-zero, non-undef, element.
3889 if (NumNonZero == 1) {
3890 unsigned Idx = CountTrailingZeros_32(NonZeros);
3891 SDValue Item = Op.getOperand(Idx);
3893 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3894 // the value are obviously zero, truncate the value to i32 and do the
3895 // insertion that way. Only do this if the value is non-constant or if the
3896 // value is a constant being inserted into element 0. It is cheaper to do
3897 // a constant pool load than it is to do a movd + shuffle.
3898 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3899 (!IsAllConstants || Idx == 0)) {
3900 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3901 // Handle MMX and SSE both.
3902 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3903 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3905 // Truncate the value (which may itself be a constant) to i32, and
3906 // convert it to a vector with movd (S2V+shuffle to zero extend).
3907 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3908 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3909 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3910 Subtarget->hasSSE2(), DAG);
3912 // Now we have our 32-bit value zero extended in the low element of
3913 // a vector. If Idx != 0, swizzle it into place.
3915 SmallVector<int, 4> Mask;
3916 Mask.push_back(Idx);
3917 for (unsigned i = 1; i != VecElts; ++i)
3919 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3920 DAG.getUNDEF(Item.getValueType()),
3923 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3927 // If we have a constant or non-constant insertion into the low element of
3928 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3929 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3930 // depending on what the source datatype is.
3933 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3934 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3935 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3936 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3937 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3938 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3940 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3941 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3942 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3943 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3944 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3945 Subtarget->hasSSE2(), DAG);
3946 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3950 // Is it a vector logical left shift?
3951 if (NumElems == 2 && Idx == 1 &&
3952 X86::isZeroNode(Op.getOperand(0)) &&
3953 !X86::isZeroNode(Op.getOperand(1))) {
3954 unsigned NumBits = VT.getSizeInBits();
3955 return getVShift(true, VT,
3956 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3957 VT, Op.getOperand(1)),
3958 NumBits/2, DAG, *this, dl);
3961 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3964 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3965 // is a non-constant being inserted into an element other than the low one,
3966 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3967 // movd/movss) to move this into the low element, then shuffle it into
3969 if (EVTBits == 32) {
3970 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3972 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3973 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3974 Subtarget->hasSSE2(), DAG);
3975 SmallVector<int, 8> MaskVec;
3976 for (unsigned i = 0; i < NumElems; i++)
3977 MaskVec.push_back(i == Idx ? 0 : 1);
3978 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3982 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3983 if (Values.size() == 1) {
3984 if (EVTBits == 32) {
3985 // Instead of a shuffle like this:
3986 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3987 // Check if it's possible to issue this instead.
3988 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3989 unsigned Idx = CountTrailingZeros_32(NonZeros);
3990 SDValue Item = Op.getOperand(Idx);
3991 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3992 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3997 // A vector full of immediates; various special cases are already
3998 // handled, so this is best done with a single constant-pool load.
4002 // Let legalizer expand 2-wide build_vectors.
4003 if (EVTBits == 64) {
4004 if (NumNonZero == 1) {
4005 // One half is zero or undef.
4006 unsigned Idx = CountTrailingZeros_32(NonZeros);
4007 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4008 Op.getOperand(Idx));
4009 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4010 Subtarget->hasSSE2(), DAG);
4015 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4016 if (EVTBits == 8 && NumElems == 16) {
4017 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4019 if (V.getNode()) return V;
4022 if (EVTBits == 16 && NumElems == 8) {
4023 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4025 if (V.getNode()) return V;
4028 // If element VT is == 32 bits, turn it into a number of shuffles.
4029 SmallVector<SDValue, 8> V;
4031 if (NumElems == 4 && NumZero > 0) {
4032 for (unsigned i = 0; i < 4; ++i) {
4033 bool isZero = !(NonZeros & (1 << i));
4035 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4037 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4040 for (unsigned i = 0; i < 2; ++i) {
4041 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4044 V[i] = V[i*2]; // Must be a zero vector.
4047 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4050 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4053 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4058 SmallVector<int, 8> MaskVec;
4059 bool Reverse = (NonZeros & 0x3) == 2;
4060 for (unsigned i = 0; i < 2; ++i)
4061 MaskVec.push_back(Reverse ? 1-i : i);
4062 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4063 for (unsigned i = 0; i < 2; ++i)
4064 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4065 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4068 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4069 // Check for a build vector of consecutive loads.
4070 for (unsigned i = 0; i < NumElems; ++i)
4071 V[i] = Op.getOperand(i);
4073 // Check for elements which are consecutive loads.
4074 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4078 // For SSE 4.1, use inserts into undef.
4079 if (getSubtarget()->hasSSE41()) {
4080 V[0] = DAG.getUNDEF(VT);
4081 for (unsigned i = 0; i < NumElems; ++i)
4082 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4083 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
4084 Op.getOperand(i), DAG.getIntPtrConstant(i));
4088 // Otherwise, expand into a number of unpckl*
4090 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4091 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4092 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4093 for (unsigned i = 0; i < NumElems; ++i)
4094 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4096 while (NumElems != 0) {
4097 for (unsigned i = 0; i < NumElems; ++i)
4098 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4107 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4108 // We support concatenate two MMX registers and place them in a MMX
4109 // register. This is better than doing a stack convert.
4110 DebugLoc dl = Op.getDebugLoc();
4111 EVT ResVT = Op.getValueType();
4112 assert(Op.getNumOperands() == 2);
4113 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4114 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4116 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4117 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4118 InVec = Op.getOperand(1);
4119 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4120 unsigned NumElts = ResVT.getVectorNumElements();
4121 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4122 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4123 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4125 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4126 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4127 Mask[0] = 0; Mask[1] = 2;
4128 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4130 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4133 // v8i16 shuffles - Prefer shuffles in the following order:
4134 // 1. [all] pshuflw, pshufhw, optional move
4135 // 2. [ssse3] 1 x pshufb
4136 // 3. [ssse3] 2 x pshufb + 1 x por
4137 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4139 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4141 const X86TargetLowering &TLI) {
4142 SDValue V1 = SVOp->getOperand(0);
4143 SDValue V2 = SVOp->getOperand(1);
4144 DebugLoc dl = SVOp->getDebugLoc();
4145 SmallVector<int, 8> MaskVals;
4147 // Determine if more than 1 of the words in each of the low and high quadwords
4148 // of the result come from the same quadword of one of the two inputs. Undef
4149 // mask values count as coming from any quadword, for better codegen.
4150 SmallVector<unsigned, 4> LoQuad(4);
4151 SmallVector<unsigned, 4> HiQuad(4);
4152 BitVector InputQuads(4);
4153 for (unsigned i = 0; i < 8; ++i) {
4154 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4155 int EltIdx = SVOp->getMaskElt(i);
4156 MaskVals.push_back(EltIdx);
4165 InputQuads.set(EltIdx / 4);
4168 int BestLoQuad = -1;
4169 unsigned MaxQuad = 1;
4170 for (unsigned i = 0; i < 4; ++i) {
4171 if (LoQuad[i] > MaxQuad) {
4173 MaxQuad = LoQuad[i];
4177 int BestHiQuad = -1;
4179 for (unsigned i = 0; i < 4; ++i) {
4180 if (HiQuad[i] > MaxQuad) {
4182 MaxQuad = HiQuad[i];
4186 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4187 // of the two input vectors, shuffle them into one input vector so only a
4188 // single pshufb instruction is necessary. If There are more than 2 input
4189 // quads, disable the next transformation since it does not help SSSE3.
4190 bool V1Used = InputQuads[0] || InputQuads[1];
4191 bool V2Used = InputQuads[2] || InputQuads[3];
4192 if (TLI.getSubtarget()->hasSSSE3()) {
4193 if (InputQuads.count() == 2 && V1Used && V2Used) {
4194 BestLoQuad = InputQuads.find_first();
4195 BestHiQuad = InputQuads.find_next(BestLoQuad);
4197 if (InputQuads.count() > 2) {
4203 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4204 // the shuffle mask. If a quad is scored as -1, that means that it contains
4205 // words from all 4 input quadwords.
4207 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4208 SmallVector<int, 8> MaskV;
4209 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4210 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4211 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4212 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4213 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4214 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4216 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4217 // source words for the shuffle, to aid later transformations.
4218 bool AllWordsInNewV = true;
4219 bool InOrder[2] = { true, true };
4220 for (unsigned i = 0; i != 8; ++i) {
4221 int idx = MaskVals[i];
4223 InOrder[i/4] = false;
4224 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4226 AllWordsInNewV = false;
4230 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4231 if (AllWordsInNewV) {
4232 for (int i = 0; i != 8; ++i) {
4233 int idx = MaskVals[i];
4236 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4237 if ((idx != i) && idx < 4)
4239 if ((idx != i) && idx > 3)
4248 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4249 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4250 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4251 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4252 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4256 // If we have SSSE3, and all words of the result are from 1 input vector,
4257 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4258 // is present, fall back to case 4.
4259 if (TLI.getSubtarget()->hasSSSE3()) {
4260 SmallVector<SDValue,16> pshufbMask;
4262 // If we have elements from both input vectors, set the high bit of the
4263 // shuffle mask element to zero out elements that come from V2 in the V1
4264 // mask, and elements that come from V1 in the V2 mask, so that the two
4265 // results can be OR'd together.
4266 bool TwoInputs = V1Used && V2Used;
4267 for (unsigned i = 0; i != 8; ++i) {
4268 int EltIdx = MaskVals[i] * 2;
4269 if (TwoInputs && (EltIdx >= 16)) {
4270 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4271 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4274 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4275 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4277 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4278 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4279 DAG.getNode(ISD::BUILD_VECTOR, dl,
4280 MVT::v16i8, &pshufbMask[0], 16));
4282 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4284 // Calculate the shuffle mask for the second input, shuffle it, and
4285 // OR it with the first shuffled input.
4287 for (unsigned i = 0; i != 8; ++i) {
4288 int EltIdx = MaskVals[i] * 2;
4290 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4291 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4294 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4295 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4297 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4298 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4299 DAG.getNode(ISD::BUILD_VECTOR, dl,
4300 MVT::v16i8, &pshufbMask[0], 16));
4301 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4302 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4305 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4306 // and update MaskVals with new element order.
4307 BitVector InOrder(8);
4308 if (BestLoQuad >= 0) {
4309 SmallVector<int, 8> MaskV;
4310 for (int i = 0; i != 4; ++i) {
4311 int idx = MaskVals[i];
4313 MaskV.push_back(-1);
4315 } else if ((idx / 4) == BestLoQuad) {
4316 MaskV.push_back(idx & 3);
4319 MaskV.push_back(-1);
4322 for (unsigned i = 4; i != 8; ++i)
4324 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4328 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4329 // and update MaskVals with the new element order.
4330 if (BestHiQuad >= 0) {
4331 SmallVector<int, 8> MaskV;
4332 for (unsigned i = 0; i != 4; ++i)
4334 for (unsigned i = 4; i != 8; ++i) {
4335 int idx = MaskVals[i];
4337 MaskV.push_back(-1);
4339 } else if ((idx / 4) == BestHiQuad) {
4340 MaskV.push_back((idx & 3) + 4);
4343 MaskV.push_back(-1);
4346 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4350 // In case BestHi & BestLo were both -1, which means each quadword has a word
4351 // from each of the four input quadwords, calculate the InOrder bitvector now
4352 // before falling through to the insert/extract cleanup.
4353 if (BestLoQuad == -1 && BestHiQuad == -1) {
4355 for (int i = 0; i != 8; ++i)
4356 if (MaskVals[i] < 0 || MaskVals[i] == i)
4360 // The other elements are put in the right place using pextrw and pinsrw.
4361 for (unsigned i = 0; i != 8; ++i) {
4364 int EltIdx = MaskVals[i];
4367 SDValue ExtOp = (EltIdx < 8)
4368 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4369 DAG.getIntPtrConstant(EltIdx))
4370 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4371 DAG.getIntPtrConstant(EltIdx - 8));
4372 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4373 DAG.getIntPtrConstant(i));
4378 // v16i8 shuffles - Prefer shuffles in the following order:
4379 // 1. [ssse3] 1 x pshufb
4380 // 2. [ssse3] 2 x pshufb + 1 x por
4381 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4383 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4385 const X86TargetLowering &TLI) {
4386 SDValue V1 = SVOp->getOperand(0);
4387 SDValue V2 = SVOp->getOperand(1);
4388 DebugLoc dl = SVOp->getDebugLoc();
4389 SmallVector<int, 16> MaskVals;
4390 SVOp->getMask(MaskVals);
4392 // If we have SSSE3, case 1 is generated when all result bytes come from
4393 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4394 // present, fall back to case 3.
4395 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4398 for (unsigned i = 0; i < 16; ++i) {
4399 int EltIdx = MaskVals[i];
4408 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4409 if (TLI.getSubtarget()->hasSSSE3()) {
4410 SmallVector<SDValue,16> pshufbMask;
4412 // If all result elements are from one input vector, then only translate
4413 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4415 // Otherwise, we have elements from both input vectors, and must zero out
4416 // elements that come from V2 in the first mask, and V1 in the second mask
4417 // so that we can OR them together.
4418 bool TwoInputs = !(V1Only || V2Only);
4419 for (unsigned i = 0; i != 16; ++i) {
4420 int EltIdx = MaskVals[i];
4421 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4422 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4425 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4427 // If all the elements are from V2, assign it to V1 and return after
4428 // building the first pshufb.
4431 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4432 DAG.getNode(ISD::BUILD_VECTOR, dl,
4433 MVT::v16i8, &pshufbMask[0], 16));
4437 // Calculate the shuffle mask for the second input, shuffle it, and
4438 // OR it with the first shuffled input.
4440 for (unsigned i = 0; i != 16; ++i) {
4441 int EltIdx = MaskVals[i];
4443 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4446 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4448 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4449 DAG.getNode(ISD::BUILD_VECTOR, dl,
4450 MVT::v16i8, &pshufbMask[0], 16));
4451 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4454 // No SSSE3 - Calculate in place words and then fix all out of place words
4455 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4456 // the 16 different words that comprise the two doublequadword input vectors.
4457 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4458 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4459 SDValue NewV = V2Only ? V2 : V1;
4460 for (int i = 0; i != 8; ++i) {
4461 int Elt0 = MaskVals[i*2];
4462 int Elt1 = MaskVals[i*2+1];
4464 // This word of the result is all undef, skip it.
4465 if (Elt0 < 0 && Elt1 < 0)
4468 // This word of the result is already in the correct place, skip it.
4469 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4471 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4474 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4475 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4478 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4479 // using a single extract together, load it and store it.
4480 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4481 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4482 DAG.getIntPtrConstant(Elt1 / 2));
4483 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4484 DAG.getIntPtrConstant(i));
4488 // If Elt1 is defined, extract it from the appropriate source. If the
4489 // source byte is not also odd, shift the extracted word left 8 bits
4490 // otherwise clear the bottom 8 bits if we need to do an or.
4492 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4493 DAG.getIntPtrConstant(Elt1 / 2));
4494 if ((Elt1 & 1) == 0)
4495 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4496 DAG.getConstant(8, TLI.getShiftAmountTy()));
4498 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4499 DAG.getConstant(0xFF00, MVT::i16));
4501 // If Elt0 is defined, extract it from the appropriate source. If the
4502 // source byte is not also even, shift the extracted word right 8 bits. If
4503 // Elt1 was also defined, OR the extracted values together before
4504 // inserting them in the result.
4506 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4507 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4508 if ((Elt0 & 1) != 0)
4509 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4510 DAG.getConstant(8, TLI.getShiftAmountTy()));
4512 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4513 DAG.getConstant(0x00FF, MVT::i16));
4514 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4517 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4518 DAG.getIntPtrConstant(i));
4520 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4523 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4524 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4525 /// done when every pair / quad of shuffle mask elements point to elements in
4526 /// the right sequence. e.g.
4527 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4529 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4531 const TargetLowering &TLI, DebugLoc dl) {
4532 EVT VT = SVOp->getValueType(0);
4533 SDValue V1 = SVOp->getOperand(0);
4534 SDValue V2 = SVOp->getOperand(1);
4535 unsigned NumElems = VT.getVectorNumElements();
4536 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4537 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4539 switch (VT.getSimpleVT().SimpleTy) {
4540 default: assert(false && "Unexpected!");
4541 case MVT::v4f32: NewVT = MVT::v2f64; break;
4542 case MVT::v4i32: NewVT = MVT::v2i64; break;
4543 case MVT::v8i16: NewVT = MVT::v4i32; break;
4544 case MVT::v16i8: NewVT = MVT::v4i32; break;
4547 if (NewWidth == 2) {
4553 int Scale = NumElems / NewWidth;
4554 SmallVector<int, 8> MaskVec;
4555 for (unsigned i = 0; i < NumElems; i += Scale) {
4557 for (int j = 0; j < Scale; ++j) {
4558 int EltIdx = SVOp->getMaskElt(i+j);
4562 StartIdx = EltIdx - (EltIdx % Scale);
4563 if (EltIdx != StartIdx + j)
4567 MaskVec.push_back(-1);
4569 MaskVec.push_back(StartIdx / Scale);
4572 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4573 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4574 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4577 /// getVZextMovL - Return a zero-extending vector move low node.
4579 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4580 SDValue SrcOp, SelectionDAG &DAG,
4581 const X86Subtarget *Subtarget, DebugLoc dl) {
4582 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4583 LoadSDNode *LD = NULL;
4584 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4585 LD = dyn_cast<LoadSDNode>(SrcOp);
4587 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4589 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4590 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4591 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4592 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4593 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4595 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4596 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4597 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4598 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4606 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4607 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4608 DAG.getNode(ISD::BIT_CONVERT, dl,
4612 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4615 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4616 SDValue V1 = SVOp->getOperand(0);
4617 SDValue V2 = SVOp->getOperand(1);
4618 DebugLoc dl = SVOp->getDebugLoc();
4619 EVT VT = SVOp->getValueType(0);
4621 SmallVector<std::pair<int, int>, 8> Locs;
4623 SmallVector<int, 8> Mask1(4U, -1);
4624 SmallVector<int, 8> PermMask;
4625 SVOp->getMask(PermMask);
4629 for (unsigned i = 0; i != 4; ++i) {
4630 int Idx = PermMask[i];
4632 Locs[i] = std::make_pair(-1, -1);
4634 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4636 Locs[i] = std::make_pair(0, NumLo);
4640 Locs[i] = std::make_pair(1, NumHi);
4642 Mask1[2+NumHi] = Idx;
4648 if (NumLo <= 2 && NumHi <= 2) {
4649 // If no more than two elements come from either vector. This can be
4650 // implemented with two shuffles. First shuffle gather the elements.
4651 // The second shuffle, which takes the first shuffle as both of its
4652 // vector operands, put the elements into the right order.
4653 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4655 SmallVector<int, 8> Mask2(4U, -1);
4657 for (unsigned i = 0; i != 4; ++i) {
4658 if (Locs[i].first == -1)
4661 unsigned Idx = (i < 2) ? 0 : 4;
4662 Idx += Locs[i].first * 2 + Locs[i].second;
4667 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4668 } else if (NumLo == 3 || NumHi == 3) {
4669 // Otherwise, we must have three elements from one vector, call it X, and
4670 // one element from the other, call it Y. First, use a shufps to build an
4671 // intermediate vector with the one element from Y and the element from X
4672 // that will be in the same half in the final destination (the indexes don't
4673 // matter). Then, use a shufps to build the final vector, taking the half
4674 // containing the element from Y from the intermediate, and the other half
4677 // Normalize it so the 3 elements come from V1.
4678 CommuteVectorShuffleMask(PermMask, VT);
4682 // Find the element from V2.
4684 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4685 int Val = PermMask[HiIndex];
4692 Mask1[0] = PermMask[HiIndex];
4694 Mask1[2] = PermMask[HiIndex^1];
4696 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4699 Mask1[0] = PermMask[0];
4700 Mask1[1] = PermMask[1];
4701 Mask1[2] = HiIndex & 1 ? 6 : 4;
4702 Mask1[3] = HiIndex & 1 ? 4 : 6;
4703 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4705 Mask1[0] = HiIndex & 1 ? 2 : 0;
4706 Mask1[1] = HiIndex & 1 ? 0 : 2;
4707 Mask1[2] = PermMask[2];
4708 Mask1[3] = PermMask[3];
4713 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4717 // Break it into (shuffle shuffle_hi, shuffle_lo).
4719 SmallVector<int,8> LoMask(4U, -1);
4720 SmallVector<int,8> HiMask(4U, -1);
4722 SmallVector<int,8> *MaskPtr = &LoMask;
4723 unsigned MaskIdx = 0;
4726 for (unsigned i = 0; i != 4; ++i) {
4733 int Idx = PermMask[i];
4735 Locs[i] = std::make_pair(-1, -1);
4736 } else if (Idx < 4) {
4737 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4738 (*MaskPtr)[LoIdx] = Idx;
4741 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4742 (*MaskPtr)[HiIdx] = Idx;
4747 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4748 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4749 SmallVector<int, 8> MaskOps;
4750 for (unsigned i = 0; i != 4; ++i) {
4751 if (Locs[i].first == -1) {
4752 MaskOps.push_back(-1);
4754 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4755 MaskOps.push_back(Idx);
4758 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4762 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4763 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4764 SDValue V1 = Op.getOperand(0);
4765 SDValue V2 = Op.getOperand(1);
4766 EVT VT = Op.getValueType();
4767 DebugLoc dl = Op.getDebugLoc();
4768 unsigned NumElems = VT.getVectorNumElements();
4769 bool isMMX = VT.getSizeInBits() == 64;
4770 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4771 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4772 bool V1IsSplat = false;
4773 bool V2IsSplat = false;
4775 if (isZeroShuffle(SVOp))
4776 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4778 // Promote splats to v4f32.
4779 if (SVOp->isSplat()) {
4780 if (isMMX || NumElems < 4)
4782 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4785 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4787 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4788 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4789 if (NewOp.getNode())
4790 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4791 LowerVECTOR_SHUFFLE(NewOp, DAG));
4792 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4793 // FIXME: Figure out a cleaner way to do this.
4794 // Try to make use of movq to zero out the top part.
4795 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4796 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4797 if (NewOp.getNode()) {
4798 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4799 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4800 DAG, Subtarget, dl);
4802 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4803 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4804 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4805 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4806 DAG, Subtarget, dl);
4810 if (X86::isPSHUFDMask(SVOp))
4813 // Check if this can be converted into a logical shift.
4814 bool isLeft = false;
4817 bool isShift = getSubtarget()->hasSSE2() &&
4818 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4819 if (isShift && ShVal.hasOneUse()) {
4820 // If the shifted value has multiple uses, it may be cheaper to use
4821 // v_set0 + movlhps or movhlps, etc.
4822 EVT EltVT = VT.getVectorElementType();
4823 ShAmt *= EltVT.getSizeInBits();
4824 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4827 if (X86::isMOVLMask(SVOp)) {
4830 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4831 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4836 // FIXME: fold these into legal mask.
4837 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4838 X86::isMOVSLDUPMask(SVOp) ||
4839 X86::isMOVHLPSMask(SVOp) ||
4840 X86::isMOVLHPSMask(SVOp) ||
4841 X86::isMOVLPMask(SVOp)))
4844 if (ShouldXformToMOVHLPS(SVOp) ||
4845 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4846 return CommuteVectorShuffle(SVOp, DAG);
4849 // No better options. Use a vshl / vsrl.
4850 EVT EltVT = VT.getVectorElementType();
4851 ShAmt *= EltVT.getSizeInBits();
4852 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4855 bool Commuted = false;
4856 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4857 // 1,1,1,1 -> v8i16 though.
4858 V1IsSplat = isSplatVector(V1.getNode());
4859 V2IsSplat = isSplatVector(V2.getNode());
4861 // Canonicalize the splat or undef, if present, to be on the RHS.
4862 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4863 Op = CommuteVectorShuffle(SVOp, DAG);
4864 SVOp = cast<ShuffleVectorSDNode>(Op);
4865 V1 = SVOp->getOperand(0);
4866 V2 = SVOp->getOperand(1);
4867 std::swap(V1IsSplat, V2IsSplat);
4868 std::swap(V1IsUndef, V2IsUndef);
4872 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4873 // Shuffling low element of v1 into undef, just return v1.
4876 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4877 // the instruction selector will not match, so get a canonical MOVL with
4878 // swapped operands to undo the commute.
4879 return getMOVL(DAG, dl, VT, V2, V1);
4882 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4883 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4884 X86::isUNPCKLMask(SVOp) ||
4885 X86::isUNPCKHMask(SVOp))
4889 // Normalize mask so all entries that point to V2 points to its first
4890 // element then try to match unpck{h|l} again. If match, return a
4891 // new vector_shuffle with the corrected mask.
4892 SDValue NewMask = NormalizeMask(SVOp, DAG);
4893 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4894 if (NSVOp != SVOp) {
4895 if (X86::isUNPCKLMask(NSVOp, true)) {
4897 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4904 // Commute is back and try unpck* again.
4905 // FIXME: this seems wrong.
4906 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4907 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4908 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4909 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4910 X86::isUNPCKLMask(NewSVOp) ||
4911 X86::isUNPCKHMask(NewSVOp))
4915 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4917 // Normalize the node to match x86 shuffle ops if needed
4918 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4919 return CommuteVectorShuffle(SVOp, DAG);
4921 // Check for legal shuffle and return?
4922 SmallVector<int, 16> PermMask;
4923 SVOp->getMask(PermMask);
4924 if (isShuffleMaskLegal(PermMask, VT))
4927 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4928 if (VT == MVT::v8i16) {
4929 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4930 if (NewOp.getNode())
4934 if (VT == MVT::v16i8) {
4935 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4936 if (NewOp.getNode())
4940 // Handle all 4 wide cases with a number of shuffles except for MMX.
4941 if (NumElems == 4 && !isMMX)
4942 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4948 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4949 SelectionDAG &DAG) const {
4950 EVT VT = Op.getValueType();
4951 DebugLoc dl = Op.getDebugLoc();
4952 if (VT.getSizeInBits() == 8) {
4953 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4954 Op.getOperand(0), Op.getOperand(1));
4955 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4956 DAG.getValueType(VT));
4957 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4958 } else if (VT.getSizeInBits() == 16) {
4959 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4960 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4962 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4963 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4964 DAG.getNode(ISD::BIT_CONVERT, dl,
4968 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4969 Op.getOperand(0), Op.getOperand(1));
4970 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4971 DAG.getValueType(VT));
4972 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4973 } else if (VT == MVT::f32) {
4974 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4975 // the result back to FR32 register. It's only worth matching if the
4976 // result has a single use which is a store or a bitcast to i32. And in
4977 // the case of a store, it's not worth it if the index is a constant 0,
4978 // because a MOVSSmr can be used instead, which is smaller and faster.
4979 if (!Op.hasOneUse())
4981 SDNode *User = *Op.getNode()->use_begin();
4982 if ((User->getOpcode() != ISD::STORE ||
4983 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4984 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4985 (User->getOpcode() != ISD::BIT_CONVERT ||
4986 User->getValueType(0) != MVT::i32))
4988 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4989 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4992 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4993 } else if (VT == MVT::i32) {
4994 // ExtractPS works with constant index.
4995 if (isa<ConstantSDNode>(Op.getOperand(1)))
5003 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5004 SelectionDAG &DAG) const {
5005 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5008 if (Subtarget->hasSSE41()) {
5009 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5014 EVT VT = Op.getValueType();
5015 DebugLoc dl = Op.getDebugLoc();
5016 // TODO: handle v16i8.
5017 if (VT.getSizeInBits() == 16) {
5018 SDValue Vec = Op.getOperand(0);
5019 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5021 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5022 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5023 DAG.getNode(ISD::BIT_CONVERT, dl,
5026 // Transform it so it match pextrw which produces a 32-bit result.
5027 EVT EltVT = MVT::i32;
5028 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5029 Op.getOperand(0), Op.getOperand(1));
5030 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5031 DAG.getValueType(VT));
5032 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5033 } else if (VT.getSizeInBits() == 32) {
5034 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5038 // SHUFPS the element to the lowest double word, then movss.
5039 int Mask[4] = { Idx, -1, -1, -1 };
5040 EVT VVT = Op.getOperand(0).getValueType();
5041 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5042 DAG.getUNDEF(VVT), Mask);
5043 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5044 DAG.getIntPtrConstant(0));
5045 } else if (VT.getSizeInBits() == 64) {
5046 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5047 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5048 // to match extract_elt for f64.
5049 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5053 // UNPCKHPD the element to the lowest double word, then movsd.
5054 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5055 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5056 int Mask[2] = { 1, -1 };
5057 EVT VVT = Op.getOperand(0).getValueType();
5058 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5059 DAG.getUNDEF(VVT), Mask);
5060 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5061 DAG.getIntPtrConstant(0));
5068 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5069 SelectionDAG &DAG) const {
5070 EVT VT = Op.getValueType();
5071 EVT EltVT = VT.getVectorElementType();
5072 DebugLoc dl = Op.getDebugLoc();
5074 SDValue N0 = Op.getOperand(0);
5075 SDValue N1 = Op.getOperand(1);
5076 SDValue N2 = Op.getOperand(2);
5078 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5079 isa<ConstantSDNode>(N2)) {
5081 if (VT == MVT::v8i16)
5082 Opc = X86ISD::PINSRW;
5083 else if (VT == MVT::v4i16)
5084 Opc = X86ISD::MMX_PINSRW;
5085 else if (VT == MVT::v16i8)
5086 Opc = X86ISD::PINSRB;
5088 Opc = X86ISD::PINSRB;
5090 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5092 if (N1.getValueType() != MVT::i32)
5093 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5094 if (N2.getValueType() != MVT::i32)
5095 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5096 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5097 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5098 // Bits [7:6] of the constant are the source select. This will always be
5099 // zero here. The DAG Combiner may combine an extract_elt index into these
5100 // bits. For example (insert (extract, 3), 2) could be matched by putting
5101 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5102 // Bits [5:4] of the constant are the destination select. This is the
5103 // value of the incoming immediate.
5104 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5105 // combine either bitwise AND or insert of float 0.0 to set these bits.
5106 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5107 // Create this as a scalar to vector..
5108 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5109 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5110 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5111 // PINSR* works with constant index.
5118 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5119 EVT VT = Op.getValueType();
5120 EVT EltVT = VT.getVectorElementType();
5122 if (Subtarget->hasSSE41())
5123 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5125 if (EltVT == MVT::i8)
5128 DebugLoc dl = Op.getDebugLoc();
5129 SDValue N0 = Op.getOperand(0);
5130 SDValue N1 = Op.getOperand(1);
5131 SDValue N2 = Op.getOperand(2);
5133 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5134 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5135 // as its second argument.
5136 if (N1.getValueType() != MVT::i32)
5137 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5138 if (N2.getValueType() != MVT::i32)
5139 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5140 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5141 dl, VT, N0, N1, N2);
5147 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5148 DebugLoc dl = Op.getDebugLoc();
5150 if (Op.getValueType() == MVT::v1i64 &&
5151 Op.getOperand(0).getValueType() == MVT::i64)
5152 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5154 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5155 EVT VT = MVT::v2i32;
5156 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5163 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5164 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5167 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5168 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5169 // one of the above mentioned nodes. It has to be wrapped because otherwise
5170 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5171 // be used to form addressing mode. These wrapped nodes will be selected
5174 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5175 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5177 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5179 unsigned char OpFlag = 0;
5180 unsigned WrapperKind = X86ISD::Wrapper;
5181 CodeModel::Model M = getTargetMachine().getCodeModel();
5183 if (Subtarget->isPICStyleRIPRel() &&
5184 (M == CodeModel::Small || M == CodeModel::Kernel))
5185 WrapperKind = X86ISD::WrapperRIP;
5186 else if (Subtarget->isPICStyleGOT())
5187 OpFlag = X86II::MO_GOTOFF;
5188 else if (Subtarget->isPICStyleStubPIC())
5189 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5191 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5193 CP->getOffset(), OpFlag);
5194 DebugLoc DL = CP->getDebugLoc();
5195 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5196 // With PIC, the address is actually $g + Offset.
5198 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5199 DAG.getNode(X86ISD::GlobalBaseReg,
5200 DebugLoc(), getPointerTy()),
5207 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5208 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5210 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5212 unsigned char OpFlag = 0;
5213 unsigned WrapperKind = X86ISD::Wrapper;
5214 CodeModel::Model M = getTargetMachine().getCodeModel();
5216 if (Subtarget->isPICStyleRIPRel() &&
5217 (M == CodeModel::Small || M == CodeModel::Kernel))
5218 WrapperKind = X86ISD::WrapperRIP;
5219 else if (Subtarget->isPICStyleGOT())
5220 OpFlag = X86II::MO_GOTOFF;
5221 else if (Subtarget->isPICStyleStubPIC())
5222 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5224 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5226 DebugLoc DL = JT->getDebugLoc();
5227 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5229 // With PIC, the address is actually $g + Offset.
5231 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5232 DAG.getNode(X86ISD::GlobalBaseReg,
5233 DebugLoc(), getPointerTy()),
5241 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5242 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5244 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5246 unsigned char OpFlag = 0;
5247 unsigned WrapperKind = X86ISD::Wrapper;
5248 CodeModel::Model M = getTargetMachine().getCodeModel();
5250 if (Subtarget->isPICStyleRIPRel() &&
5251 (M == CodeModel::Small || M == CodeModel::Kernel))
5252 WrapperKind = X86ISD::WrapperRIP;
5253 else if (Subtarget->isPICStyleGOT())
5254 OpFlag = X86II::MO_GOTOFF;
5255 else if (Subtarget->isPICStyleStubPIC())
5256 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5258 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5260 DebugLoc DL = Op.getDebugLoc();
5261 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5264 // With PIC, the address is actually $g + Offset.
5265 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5266 !Subtarget->is64Bit()) {
5267 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5268 DAG.getNode(X86ISD::GlobalBaseReg,
5269 DebugLoc(), getPointerTy()),
5277 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5278 // Create the TargetBlockAddressAddress node.
5279 unsigned char OpFlags =
5280 Subtarget->ClassifyBlockAddressReference();
5281 CodeModel::Model M = getTargetMachine().getCodeModel();
5282 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5283 DebugLoc dl = Op.getDebugLoc();
5284 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5285 /*isTarget=*/true, OpFlags);
5287 if (Subtarget->isPICStyleRIPRel() &&
5288 (M == CodeModel::Small || M == CodeModel::Kernel))
5289 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5291 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5293 // With PIC, the address is actually $g + Offset.
5294 if (isGlobalRelativeToPICBase(OpFlags)) {
5295 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5296 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5304 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5306 SelectionDAG &DAG) const {
5307 // Create the TargetGlobalAddress node, folding in the constant
5308 // offset if it is legal.
5309 unsigned char OpFlags =
5310 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5311 CodeModel::Model M = getTargetMachine().getCodeModel();
5313 if (OpFlags == X86II::MO_NO_FLAG &&
5314 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5315 // A direct static reference to a global.
5316 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5319 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5322 if (Subtarget->isPICStyleRIPRel() &&
5323 (M == CodeModel::Small || M == CodeModel::Kernel))
5324 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5326 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5328 // With PIC, the address is actually $g + Offset.
5329 if (isGlobalRelativeToPICBase(OpFlags)) {
5330 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5331 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5335 // For globals that require a load from a stub to get the address, emit the
5337 if (isGlobalStubReference(OpFlags))
5338 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5339 PseudoSourceValue::getGOT(), 0, false, false, 0);
5341 // If there was a non-zero offset that we didn't fold, create an explicit
5344 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5345 DAG.getConstant(Offset, getPointerTy()));
5351 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5352 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5353 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5354 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5358 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5359 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5360 unsigned char OperandFlags) {
5361 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5362 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5363 DebugLoc dl = GA->getDebugLoc();
5364 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5365 GA->getValueType(0),
5369 SDValue Ops[] = { Chain, TGA, *InFlag };
5370 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5372 SDValue Ops[] = { Chain, TGA };
5373 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5376 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5377 MFI->setAdjustsStack(true);
5379 SDValue Flag = Chain.getValue(1);
5380 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5383 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5385 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5388 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5389 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5390 DAG.getNode(X86ISD::GlobalBaseReg,
5391 DebugLoc(), PtrVT), InFlag);
5392 InFlag = Chain.getValue(1);
5394 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5397 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5399 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5401 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5402 X86::RAX, X86II::MO_TLSGD);
5405 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5406 // "local exec" model.
5407 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5408 const EVT PtrVT, TLSModel::Model model,
5410 DebugLoc dl = GA->getDebugLoc();
5411 // Get the Thread Pointer
5412 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5414 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5417 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5418 NULL, 0, false, false, 0);
5420 unsigned char OperandFlags = 0;
5421 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5423 unsigned WrapperKind = X86ISD::Wrapper;
5424 if (model == TLSModel::LocalExec) {
5425 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5426 } else if (is64Bit) {
5427 assert(model == TLSModel::InitialExec);
5428 OperandFlags = X86II::MO_GOTTPOFF;
5429 WrapperKind = X86ISD::WrapperRIP;
5431 assert(model == TLSModel::InitialExec);
5432 OperandFlags = X86II::MO_INDNTPOFF;
5435 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5437 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5438 GA->getValueType(0),
5439 GA->getOffset(), OperandFlags);
5440 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5442 if (model == TLSModel::InitialExec)
5443 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5444 PseudoSourceValue::getGOT(), 0, false, false, 0);
5446 // The address of the thread local variable is the add of the thread
5447 // pointer with the offset of the variable.
5448 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5452 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5454 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5455 const GlobalValue *GV = GA->getGlobal();
5457 if (Subtarget->isTargetELF()) {
5458 // TODO: implement the "local dynamic" model
5459 // TODO: implement the "initial exec"model for pic executables
5461 // If GV is an alias then use the aliasee for determining
5462 // thread-localness.
5463 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5464 GV = GA->resolveAliasedGlobal(false);
5466 TLSModel::Model model
5467 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5470 case TLSModel::GeneralDynamic:
5471 case TLSModel::LocalDynamic: // not implemented
5472 if (Subtarget->is64Bit())
5473 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5474 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5476 case TLSModel::InitialExec:
5477 case TLSModel::LocalExec:
5478 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5479 Subtarget->is64Bit());
5481 } else if (Subtarget->isTargetDarwin()) {
5482 // Darwin only has one model of TLS. Lower to that.
5483 unsigned char OpFlag = 0;
5484 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5485 X86ISD::WrapperRIP : X86ISD::Wrapper;
5487 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5489 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5490 !Subtarget->is64Bit();
5492 OpFlag = X86II::MO_TLVP_PIC_BASE;
5494 OpFlag = X86II::MO_TLVP;
5495 DebugLoc DL = Op.getDebugLoc();
5496 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
5498 GA->getOffset(), OpFlag);
5499 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5501 // With PIC32, the address is actually $g + Offset.
5503 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5504 DAG.getNode(X86ISD::GlobalBaseReg,
5505 DebugLoc(), getPointerTy()),
5508 // Lowering the machine isd will make sure everything is in the right
5510 SDValue Args[] = { Offset };
5511 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5513 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5514 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5515 MFI->setAdjustsStack(true);
5517 // And our return value (tls address) is in the standard call return value
5519 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5520 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5524 "TLS not implemented for this target.");
5526 llvm_unreachable("Unreachable");
5531 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5532 /// take a 2 x i32 value to shift plus a shift amount.
5533 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5534 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5535 EVT VT = Op.getValueType();
5536 unsigned VTBits = VT.getSizeInBits();
5537 DebugLoc dl = Op.getDebugLoc();
5538 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5539 SDValue ShOpLo = Op.getOperand(0);
5540 SDValue ShOpHi = Op.getOperand(1);
5541 SDValue ShAmt = Op.getOperand(2);
5542 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5543 DAG.getConstant(VTBits - 1, MVT::i8))
5544 : DAG.getConstant(0, VT);
5547 if (Op.getOpcode() == ISD::SHL_PARTS) {
5548 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5549 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5551 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5552 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5555 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5556 DAG.getConstant(VTBits, MVT::i8));
5557 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5558 AndNode, DAG.getConstant(0, MVT::i8));
5561 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5562 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5563 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5565 if (Op.getOpcode() == ISD::SHL_PARTS) {
5566 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5567 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5569 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5570 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5573 SDValue Ops[2] = { Lo, Hi };
5574 return DAG.getMergeValues(Ops, 2, dl);
5577 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5578 SelectionDAG &DAG) const {
5579 EVT SrcVT = Op.getOperand(0).getValueType();
5581 if (SrcVT.isVector()) {
5582 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5588 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5589 "Unknown SINT_TO_FP to lower!");
5591 // These are really Legal; return the operand so the caller accepts it as
5593 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5595 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5596 Subtarget->is64Bit()) {
5600 DebugLoc dl = Op.getDebugLoc();
5601 unsigned Size = SrcVT.getSizeInBits()/8;
5602 MachineFunction &MF = DAG.getMachineFunction();
5603 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5604 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5605 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5607 PseudoSourceValue::getFixedStack(SSFI), 0,
5609 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5612 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5614 SelectionDAG &DAG) const {
5616 DebugLoc dl = Op.getDebugLoc();
5618 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5620 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5622 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5623 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5624 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5625 Tys, Ops, array_lengthof(Ops));
5628 Chain = Result.getValue(1);
5629 SDValue InFlag = Result.getValue(2);
5631 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5632 // shouldn't be necessary except that RFP cannot be live across
5633 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5634 MachineFunction &MF = DAG.getMachineFunction();
5635 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5636 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5637 Tys = DAG.getVTList(MVT::Other);
5639 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5641 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5642 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5643 PseudoSourceValue::getFixedStack(SSFI), 0,
5650 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5651 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5652 SelectionDAG &DAG) const {
5653 // This algorithm is not obvious. Here it is in C code, more or less:
5655 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5656 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5657 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5659 // Copy ints to xmm registers.
5660 __m128i xh = _mm_cvtsi32_si128( hi );
5661 __m128i xl = _mm_cvtsi32_si128( lo );
5663 // Combine into low half of a single xmm register.
5664 __m128i x = _mm_unpacklo_epi32( xh, xl );
5668 // Merge in appropriate exponents to give the integer bits the right
5670 x = _mm_unpacklo_epi32( x, exp );
5672 // Subtract away the biases to deal with the IEEE-754 double precision
5674 d = _mm_sub_pd( (__m128d) x, bias );
5676 // All conversions up to here are exact. The correctly rounded result is
5677 // calculated using the current rounding mode using the following
5679 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5680 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5681 // store doesn't really need to be here (except
5682 // maybe to zero the other double)
5687 DebugLoc dl = Op.getDebugLoc();
5688 LLVMContext *Context = DAG.getContext();
5690 // Build some magic constants.
5691 std::vector<Constant*> CV0;
5692 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5693 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5694 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5695 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5696 Constant *C0 = ConstantVector::get(CV0);
5697 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5699 std::vector<Constant*> CV1;
5701 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5703 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5704 Constant *C1 = ConstantVector::get(CV1);
5705 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5707 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5708 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5710 DAG.getIntPtrConstant(1)));
5711 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5712 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5714 DAG.getIntPtrConstant(0)));
5715 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5716 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5717 PseudoSourceValue::getConstantPool(), 0,
5719 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5720 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5721 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5722 PseudoSourceValue::getConstantPool(), 0,
5724 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5726 // Add the halves; easiest way is to swap them into another reg first.
5727 int ShufMask[2] = { 1, -1 };
5728 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5729 DAG.getUNDEF(MVT::v2f64), ShufMask);
5730 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5731 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5732 DAG.getIntPtrConstant(0));
5735 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5736 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5737 SelectionDAG &DAG) const {
5738 DebugLoc dl = Op.getDebugLoc();
5739 // FP constant to bias correct the final result.
5740 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5743 // Load the 32-bit value into an XMM register.
5744 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5745 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5747 DAG.getIntPtrConstant(0)));
5749 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5750 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5751 DAG.getIntPtrConstant(0));
5753 // Or the load with the bias.
5754 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5755 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5756 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5758 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5759 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5760 MVT::v2f64, Bias)));
5761 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5762 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5763 DAG.getIntPtrConstant(0));
5765 // Subtract the bias.
5766 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5768 // Handle final rounding.
5769 EVT DestVT = Op.getValueType();
5771 if (DestVT.bitsLT(MVT::f64)) {
5772 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5773 DAG.getIntPtrConstant(0));
5774 } else if (DestVT.bitsGT(MVT::f64)) {
5775 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5778 // Handle final rounding.
5782 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5783 SelectionDAG &DAG) const {
5784 SDValue N0 = Op.getOperand(0);
5785 DebugLoc dl = Op.getDebugLoc();
5787 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5788 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5789 // the optimization here.
5790 if (DAG.SignBitIsZero(N0))
5791 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5793 EVT SrcVT = N0.getValueType();
5794 EVT DstVT = Op.getValueType();
5795 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5796 return LowerUINT_TO_FP_i64(Op, DAG);
5797 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5798 return LowerUINT_TO_FP_i32(Op, DAG);
5800 // Make a 64-bit buffer, and use it to build an FILD.
5801 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5802 if (SrcVT == MVT::i32) {
5803 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5804 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5805 getPointerTy(), StackSlot, WordOff);
5806 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5807 StackSlot, NULL, 0, false, false, 0);
5808 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5809 OffsetSlot, NULL, 0, false, false, 0);
5810 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5814 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5815 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5816 StackSlot, NULL, 0, false, false, 0);
5817 // For i64 source, we need to add the appropriate power of 2 if the input
5818 // was negative. This is the same as the optimization in
5819 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5820 // we must be careful to do the computation in x87 extended precision, not
5821 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5822 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5823 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5824 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5826 APInt FF(32, 0x5F800000ULL);
5828 // Check whether the sign bit is set.
5829 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5830 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5833 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5834 SDValue FudgePtr = DAG.getConstantPool(
5835 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5838 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5839 SDValue Zero = DAG.getIntPtrConstant(0);
5840 SDValue Four = DAG.getIntPtrConstant(4);
5841 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5843 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5845 // Load the value out, extending it from f32 to f80.
5846 // FIXME: Avoid the extend by constructing the right constant pool?
5847 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
5848 FudgePtr, PseudoSourceValue::getConstantPool(),
5849 0, MVT::f32, false, false, 4);
5850 // Extend everything to 80 bits to force it to be done on x87.
5851 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5852 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5855 std::pair<SDValue,SDValue> X86TargetLowering::
5856 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5857 DebugLoc dl = Op.getDebugLoc();
5859 EVT DstTy = Op.getValueType();
5862 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5866 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5867 DstTy.getSimpleVT() >= MVT::i16 &&
5868 "Unknown FP_TO_SINT to lower!");
5870 // These are really Legal.
5871 if (DstTy == MVT::i32 &&
5872 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5873 return std::make_pair(SDValue(), SDValue());
5874 if (Subtarget->is64Bit() &&
5875 DstTy == MVT::i64 &&
5876 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5877 return std::make_pair(SDValue(), SDValue());
5879 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5881 MachineFunction &MF = DAG.getMachineFunction();
5882 unsigned MemSize = DstTy.getSizeInBits()/8;
5883 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5884 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5887 switch (DstTy.getSimpleVT().SimpleTy) {
5888 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5889 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5890 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5891 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5894 SDValue Chain = DAG.getEntryNode();
5895 SDValue Value = Op.getOperand(0);
5896 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5897 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5898 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5899 PseudoSourceValue::getFixedStack(SSFI), 0,
5901 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5903 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5905 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5906 Chain = Value.getValue(1);
5907 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5908 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5911 // Build the FP_TO_INT*_IN_MEM
5912 SDValue Ops[] = { Chain, Value, StackSlot };
5913 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5915 return std::make_pair(FIST, StackSlot);
5918 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5919 SelectionDAG &DAG) const {
5920 if (Op.getValueType().isVector()) {
5921 if (Op.getValueType() == MVT::v2i32 &&
5922 Op.getOperand(0).getValueType() == MVT::v2f64) {
5928 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5929 SDValue FIST = Vals.first, StackSlot = Vals.second;
5930 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5931 if (FIST.getNode() == 0) return Op;
5934 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5935 FIST, StackSlot, NULL, 0, false, false, 0);
5938 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5939 SelectionDAG &DAG) const {
5940 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5941 SDValue FIST = Vals.first, StackSlot = Vals.second;
5942 assert(FIST.getNode() && "Unexpected failure");
5945 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5946 FIST, StackSlot, NULL, 0, false, false, 0);
5949 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5950 SelectionDAG &DAG) const {
5951 LLVMContext *Context = DAG.getContext();
5952 DebugLoc dl = Op.getDebugLoc();
5953 EVT VT = Op.getValueType();
5956 EltVT = VT.getVectorElementType();
5957 std::vector<Constant*> CV;
5958 if (EltVT == MVT::f64) {
5959 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5963 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5969 Constant *C = ConstantVector::get(CV);
5970 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5971 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5972 PseudoSourceValue::getConstantPool(), 0,
5974 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5977 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5978 LLVMContext *Context = DAG.getContext();
5979 DebugLoc dl = Op.getDebugLoc();
5980 EVT VT = Op.getValueType();
5983 EltVT = VT.getVectorElementType();
5984 std::vector<Constant*> CV;
5985 if (EltVT == MVT::f64) {
5986 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5990 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5996 Constant *C = ConstantVector::get(CV);
5997 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5998 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5999 PseudoSourceValue::getConstantPool(), 0,
6001 if (VT.isVector()) {
6002 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6003 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6004 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6006 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
6008 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6012 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6013 LLVMContext *Context = DAG.getContext();
6014 SDValue Op0 = Op.getOperand(0);
6015 SDValue Op1 = Op.getOperand(1);
6016 DebugLoc dl = Op.getDebugLoc();
6017 EVT VT = Op.getValueType();
6018 EVT SrcVT = Op1.getValueType();
6020 // If second operand is smaller, extend it first.
6021 if (SrcVT.bitsLT(VT)) {
6022 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6025 // And if it is bigger, shrink it first.
6026 if (SrcVT.bitsGT(VT)) {
6027 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6031 // At this point the operands and the result should have the same
6032 // type, and that won't be f80 since that is not custom lowered.
6034 // First get the sign bit of second operand.
6035 std::vector<Constant*> CV;
6036 if (SrcVT == MVT::f64) {
6037 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6038 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6040 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6041 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6042 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6043 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6045 Constant *C = ConstantVector::get(CV);
6046 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6047 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6048 PseudoSourceValue::getConstantPool(), 0,
6050 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6052 // Shift sign bit right or left if the two operands have different types.
6053 if (SrcVT.bitsGT(VT)) {
6054 // Op0 is MVT::f32, Op1 is MVT::f64.
6055 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6056 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6057 DAG.getConstant(32, MVT::i32));
6058 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
6059 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6060 DAG.getIntPtrConstant(0));
6063 // Clear first operand sign bit.
6065 if (VT == MVT::f64) {
6066 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6067 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6069 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6070 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6071 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6072 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6074 C = ConstantVector::get(CV);
6075 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6076 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6077 PseudoSourceValue::getConstantPool(), 0,
6079 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6081 // Or the value with the sign bit.
6082 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6085 /// Emit nodes that will be selected as "test Op0,Op0", or something
6087 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6088 SelectionDAG &DAG) const {
6089 DebugLoc dl = Op.getDebugLoc();
6091 // CF and OF aren't always set the way we want. Determine which
6092 // of these we need.
6093 bool NeedCF = false;
6094 bool NeedOF = false;
6097 case X86::COND_A: case X86::COND_AE:
6098 case X86::COND_B: case X86::COND_BE:
6101 case X86::COND_G: case X86::COND_GE:
6102 case X86::COND_L: case X86::COND_LE:
6103 case X86::COND_O: case X86::COND_NO:
6108 // See if we can use the EFLAGS value from the operand instead of
6109 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6110 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6111 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6112 // Emit a CMP with 0, which is the TEST pattern.
6113 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6114 DAG.getConstant(0, Op.getValueType()));
6116 unsigned Opcode = 0;
6117 unsigned NumOperands = 0;
6118 switch (Op.getNode()->getOpcode()) {
6120 // Due to an isel shortcoming, be conservative if this add is likely to be
6121 // selected as part of a load-modify-store instruction. When the root node
6122 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6123 // uses of other nodes in the match, such as the ADD in this case. This
6124 // leads to the ADD being left around and reselected, with the result being
6125 // two adds in the output. Alas, even if none our users are stores, that
6126 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6127 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6128 // climbing the DAG back to the root, and it doesn't seem to be worth the
6130 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6131 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6132 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6135 if (ConstantSDNode *C =
6136 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6137 // An add of one will be selected as an INC.
6138 if (C->getAPIntValue() == 1) {
6139 Opcode = X86ISD::INC;
6144 // An add of negative one (subtract of one) will be selected as a DEC.
6145 if (C->getAPIntValue().isAllOnesValue()) {
6146 Opcode = X86ISD::DEC;
6152 // Otherwise use a regular EFLAGS-setting add.
6153 Opcode = X86ISD::ADD;
6157 // If the primary and result isn't used, don't bother using X86ISD::AND,
6158 // because a TEST instruction will be better.
6159 bool NonFlagUse = false;
6160 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6161 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6163 unsigned UOpNo = UI.getOperandNo();
6164 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6165 // Look pass truncate.
6166 UOpNo = User->use_begin().getOperandNo();
6167 User = *User->use_begin();
6170 if (User->getOpcode() != ISD::BRCOND &&
6171 User->getOpcode() != ISD::SETCC &&
6172 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6185 // Due to the ISEL shortcoming noted above, be conservative if this op is
6186 // likely to be selected as part of a load-modify-store instruction.
6187 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6188 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6189 if (UI->getOpcode() == ISD::STORE)
6192 // Otherwise use a regular EFLAGS-setting instruction.
6193 switch (Op.getNode()->getOpcode()) {
6194 default: llvm_unreachable("unexpected operator!");
6195 case ISD::SUB: Opcode = X86ISD::SUB; break;
6196 case ISD::OR: Opcode = X86ISD::OR; break;
6197 case ISD::XOR: Opcode = X86ISD::XOR; break;
6198 case ISD::AND: Opcode = X86ISD::AND; break;
6210 return SDValue(Op.getNode(), 1);
6217 // Emit a CMP with 0, which is the TEST pattern.
6218 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6219 DAG.getConstant(0, Op.getValueType()));
6221 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6222 SmallVector<SDValue, 4> Ops;
6223 for (unsigned i = 0; i != NumOperands; ++i)
6224 Ops.push_back(Op.getOperand(i));
6226 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6227 DAG.ReplaceAllUsesWith(Op, New);
6228 return SDValue(New.getNode(), 1);
6231 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6233 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6234 SelectionDAG &DAG) const {
6235 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6236 if (C->getAPIntValue() == 0)
6237 return EmitTest(Op0, X86CC, DAG);
6239 DebugLoc dl = Op0.getDebugLoc();
6240 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6243 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6244 /// if it's possible.
6245 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6246 DebugLoc dl, SelectionDAG &DAG) const {
6247 SDValue Op0 = And.getOperand(0);
6248 SDValue Op1 = And.getOperand(1);
6249 if (Op0.getOpcode() == ISD::TRUNCATE)
6250 Op0 = Op0.getOperand(0);
6251 if (Op1.getOpcode() == ISD::TRUNCATE)
6252 Op1 = Op1.getOperand(0);
6255 if (Op1.getOpcode() == ISD::SHL)
6256 std::swap(Op0, Op1);
6257 if (Op0.getOpcode() == ISD::SHL) {
6258 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6259 if (And00C->getZExtValue() == 1) {
6260 // If we looked past a truncate, check that it's only truncating away
6262 unsigned BitWidth = Op0.getValueSizeInBits();
6263 unsigned AndBitWidth = And.getValueSizeInBits();
6264 if (BitWidth > AndBitWidth) {
6265 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6266 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6267 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6271 RHS = Op0.getOperand(1);
6273 } else if (Op1.getOpcode() == ISD::Constant) {
6274 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6275 SDValue AndLHS = Op0;
6276 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6277 LHS = AndLHS.getOperand(0);
6278 RHS = AndLHS.getOperand(1);
6282 if (LHS.getNode()) {
6283 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6284 // instruction. Since the shift amount is in-range-or-undefined, we know
6285 // that doing a bittest on the i32 value is ok. We extend to i32 because
6286 // the encoding for the i16 version is larger than the i32 version.
6287 // Also promote i16 to i32 for performance / code size reason.
6288 if (LHS.getValueType() == MVT::i8 ||
6289 LHS.getValueType() == MVT::i16)
6290 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6292 // If the operand types disagree, extend the shift amount to match. Since
6293 // BT ignores high bits (like shifts) we can use anyextend.
6294 if (LHS.getValueType() != RHS.getValueType())
6295 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6297 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6298 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6299 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6300 DAG.getConstant(Cond, MVT::i8), BT);
6306 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6307 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6308 SDValue Op0 = Op.getOperand(0);
6309 SDValue Op1 = Op.getOperand(1);
6310 DebugLoc dl = Op.getDebugLoc();
6311 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6313 // Optimize to BT if possible.
6314 // Lower (X & (1 << N)) == 0 to BT(X, N).
6315 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6316 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6317 if (Op0.getOpcode() == ISD::AND &&
6319 Op1.getOpcode() == ISD::Constant &&
6320 cast<ConstantSDNode>(Op1)->isNullValue() &&
6321 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6322 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6323 if (NewSetCC.getNode())
6327 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6328 if (Op0.getOpcode() == X86ISD::SETCC &&
6329 Op1.getOpcode() == ISD::Constant &&
6330 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6331 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6332 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6333 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6334 bool Invert = (CC == ISD::SETNE) ^
6335 cast<ConstantSDNode>(Op1)->isNullValue();
6337 CCode = X86::GetOppositeBranchCondition(CCode);
6338 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6339 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6342 bool isFP = Op1.getValueType().isFloatingPoint();
6343 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6344 if (X86CC == X86::COND_INVALID)
6347 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6349 // Use sbb x, x to materialize carry bit into a GPR.
6350 if (X86CC == X86::COND_B)
6351 return DAG.getNode(ISD::AND, dl, MVT::i8,
6352 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6353 DAG.getConstant(X86CC, MVT::i8), Cond),
6354 DAG.getConstant(1, MVT::i8));
6356 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6357 DAG.getConstant(X86CC, MVT::i8), Cond);
6360 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6362 SDValue Op0 = Op.getOperand(0);
6363 SDValue Op1 = Op.getOperand(1);
6364 SDValue CC = Op.getOperand(2);
6365 EVT VT = Op.getValueType();
6366 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6367 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6368 DebugLoc dl = Op.getDebugLoc();
6372 EVT VT0 = Op0.getValueType();
6373 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6374 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6377 switch (SetCCOpcode) {
6380 case ISD::SETEQ: SSECC = 0; break;
6382 case ISD::SETGT: Swap = true; // Fallthrough
6384 case ISD::SETOLT: SSECC = 1; break;
6386 case ISD::SETGE: Swap = true; // Fallthrough
6388 case ISD::SETOLE: SSECC = 2; break;
6389 case ISD::SETUO: SSECC = 3; break;
6391 case ISD::SETNE: SSECC = 4; break;
6392 case ISD::SETULE: Swap = true;
6393 case ISD::SETUGE: SSECC = 5; break;
6394 case ISD::SETULT: Swap = true;
6395 case ISD::SETUGT: SSECC = 6; break;
6396 case ISD::SETO: SSECC = 7; break;
6399 std::swap(Op0, Op1);
6401 // In the two special cases we can't handle, emit two comparisons.
6403 if (SetCCOpcode == ISD::SETUEQ) {
6405 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6406 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6407 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6409 else if (SetCCOpcode == ISD::SETONE) {
6411 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6412 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6413 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6415 llvm_unreachable("Illegal FP comparison");
6417 // Handle all other FP comparisons here.
6418 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6421 // We are handling one of the integer comparisons here. Since SSE only has
6422 // GT and EQ comparisons for integer, swapping operands and multiple
6423 // operations may be required for some comparisons.
6424 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6425 bool Swap = false, Invert = false, FlipSigns = false;
6427 switch (VT.getSimpleVT().SimpleTy) {
6430 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6432 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6434 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6435 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6438 switch (SetCCOpcode) {
6440 case ISD::SETNE: Invert = true;
6441 case ISD::SETEQ: Opc = EQOpc; break;
6442 case ISD::SETLT: Swap = true;
6443 case ISD::SETGT: Opc = GTOpc; break;
6444 case ISD::SETGE: Swap = true;
6445 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6446 case ISD::SETULT: Swap = true;
6447 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6448 case ISD::SETUGE: Swap = true;
6449 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6452 std::swap(Op0, Op1);
6454 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6455 // bits of the inputs before performing those operations.
6457 EVT EltVT = VT.getVectorElementType();
6458 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6460 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6461 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6463 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6464 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6467 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6469 // If the logical-not of the result is required, perform that now.
6471 Result = DAG.getNOT(dl, Result, VT);
6476 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6477 static bool isX86LogicalCmp(SDValue Op) {
6478 unsigned Opc = Op.getNode()->getOpcode();
6479 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6481 if (Op.getResNo() == 1 &&
6482 (Opc == X86ISD::ADD ||
6483 Opc == X86ISD::SUB ||
6484 Opc == X86ISD::SMUL ||
6485 Opc == X86ISD::UMUL ||
6486 Opc == X86ISD::INC ||
6487 Opc == X86ISD::DEC ||
6488 Opc == X86ISD::OR ||
6489 Opc == X86ISD::XOR ||
6490 Opc == X86ISD::AND))
6496 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6497 bool addTest = true;
6498 SDValue Cond = Op.getOperand(0);
6499 DebugLoc dl = Op.getDebugLoc();
6502 if (Cond.getOpcode() == ISD::SETCC) {
6503 SDValue NewCond = LowerSETCC(Cond, DAG);
6504 if (NewCond.getNode())
6508 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6509 SDValue Op1 = Op.getOperand(1);
6510 SDValue Op2 = Op.getOperand(2);
6511 if (Cond.getOpcode() == X86ISD::SETCC &&
6512 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6513 SDValue Cmp = Cond.getOperand(1);
6514 if (Cmp.getOpcode() == X86ISD::CMP) {
6515 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6516 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6517 ConstantSDNode *RHSC =
6518 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6519 if (N1C && N1C->isAllOnesValue() &&
6520 N2C && N2C->isNullValue() &&
6521 RHSC && RHSC->isNullValue()) {
6522 SDValue CmpOp0 = Cmp.getOperand(0);
6523 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6524 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6525 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6526 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6531 // Look pass (and (setcc_carry (cmp ...)), 1).
6532 if (Cond.getOpcode() == ISD::AND &&
6533 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6534 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6535 if (C && C->getAPIntValue() == 1)
6536 Cond = Cond.getOperand(0);
6539 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6540 // setting operand in place of the X86ISD::SETCC.
6541 if (Cond.getOpcode() == X86ISD::SETCC ||
6542 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6543 CC = Cond.getOperand(0);
6545 SDValue Cmp = Cond.getOperand(1);
6546 unsigned Opc = Cmp.getOpcode();
6547 EVT VT = Op.getValueType();
6549 bool IllegalFPCMov = false;
6550 if (VT.isFloatingPoint() && !VT.isVector() &&
6551 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6552 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6554 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6555 Opc == X86ISD::BT) { // FIXME
6562 // Look pass the truncate.
6563 if (Cond.getOpcode() == ISD::TRUNCATE)
6564 Cond = Cond.getOperand(0);
6566 // We know the result of AND is compared against zero. Try to match
6568 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6569 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6570 if (NewSetCC.getNode()) {
6571 CC = NewSetCC.getOperand(0);
6572 Cond = NewSetCC.getOperand(1);
6579 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6580 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6583 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6584 // condition is true.
6585 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6586 SDValue Ops[] = { Op2, Op1, CC, Cond };
6587 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6590 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6591 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6592 // from the AND / OR.
6593 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6594 Opc = Op.getOpcode();
6595 if (Opc != ISD::OR && Opc != ISD::AND)
6597 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6598 Op.getOperand(0).hasOneUse() &&
6599 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6600 Op.getOperand(1).hasOneUse());
6603 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6604 // 1 and that the SETCC node has a single use.
6605 static bool isXor1OfSetCC(SDValue Op) {
6606 if (Op.getOpcode() != ISD::XOR)
6608 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6609 if (N1C && N1C->getAPIntValue() == 1) {
6610 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6611 Op.getOperand(0).hasOneUse();
6616 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6617 bool addTest = true;
6618 SDValue Chain = Op.getOperand(0);
6619 SDValue Cond = Op.getOperand(1);
6620 SDValue Dest = Op.getOperand(2);
6621 DebugLoc dl = Op.getDebugLoc();
6624 if (Cond.getOpcode() == ISD::SETCC) {
6625 SDValue NewCond = LowerSETCC(Cond, DAG);
6626 if (NewCond.getNode())
6630 // FIXME: LowerXALUO doesn't handle these!!
6631 else if (Cond.getOpcode() == X86ISD::ADD ||
6632 Cond.getOpcode() == X86ISD::SUB ||
6633 Cond.getOpcode() == X86ISD::SMUL ||
6634 Cond.getOpcode() == X86ISD::UMUL)
6635 Cond = LowerXALUO(Cond, DAG);
6638 // Look pass (and (setcc_carry (cmp ...)), 1).
6639 if (Cond.getOpcode() == ISD::AND &&
6640 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6641 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6642 if (C && C->getAPIntValue() == 1)
6643 Cond = Cond.getOperand(0);
6646 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6647 // setting operand in place of the X86ISD::SETCC.
6648 if (Cond.getOpcode() == X86ISD::SETCC ||
6649 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6650 CC = Cond.getOperand(0);
6652 SDValue Cmp = Cond.getOperand(1);
6653 unsigned Opc = Cmp.getOpcode();
6654 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6655 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6659 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6663 // These can only come from an arithmetic instruction with overflow,
6664 // e.g. SADDO, UADDO.
6665 Cond = Cond.getNode()->getOperand(1);
6672 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6673 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6674 if (CondOpc == ISD::OR) {
6675 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6676 // two branches instead of an explicit OR instruction with a
6678 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6679 isX86LogicalCmp(Cmp)) {
6680 CC = Cond.getOperand(0).getOperand(0);
6681 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6682 Chain, Dest, CC, Cmp);
6683 CC = Cond.getOperand(1).getOperand(0);
6687 } else { // ISD::AND
6688 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6689 // two branches instead of an explicit AND instruction with a
6690 // separate test. However, we only do this if this block doesn't
6691 // have a fall-through edge, because this requires an explicit
6692 // jmp when the condition is false.
6693 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6694 isX86LogicalCmp(Cmp) &&
6695 Op.getNode()->hasOneUse()) {
6696 X86::CondCode CCode =
6697 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6698 CCode = X86::GetOppositeBranchCondition(CCode);
6699 CC = DAG.getConstant(CCode, MVT::i8);
6700 SDNode *User = *Op.getNode()->use_begin();
6701 // Look for an unconditional branch following this conditional branch.
6702 // We need this because we need to reverse the successors in order
6703 // to implement FCMP_OEQ.
6704 if (User->getOpcode() == ISD::BR) {
6705 SDValue FalseBB = User->getOperand(1);
6707 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6708 assert(NewBR == User);
6712 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6713 Chain, Dest, CC, Cmp);
6714 X86::CondCode CCode =
6715 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6716 CCode = X86::GetOppositeBranchCondition(CCode);
6717 CC = DAG.getConstant(CCode, MVT::i8);
6723 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6724 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6725 // It should be transformed during dag combiner except when the condition
6726 // is set by a arithmetics with overflow node.
6727 X86::CondCode CCode =
6728 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6729 CCode = X86::GetOppositeBranchCondition(CCode);
6730 CC = DAG.getConstant(CCode, MVT::i8);
6731 Cond = Cond.getOperand(0).getOperand(1);
6737 // Look pass the truncate.
6738 if (Cond.getOpcode() == ISD::TRUNCATE)
6739 Cond = Cond.getOperand(0);
6741 // We know the result of AND is compared against zero. Try to match
6743 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6744 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6745 if (NewSetCC.getNode()) {
6746 CC = NewSetCC.getOperand(0);
6747 Cond = NewSetCC.getOperand(1);
6754 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6755 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6757 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6758 Chain, Dest, CC, Cond);
6762 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6763 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6764 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6765 // that the guard pages used by the OS virtual memory manager are allocated in
6766 // correct sequence.
6768 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6769 SelectionDAG &DAG) const {
6770 assert(Subtarget->isTargetCygMing() &&
6771 "This should be used only on Cygwin/Mingw targets");
6772 DebugLoc dl = Op.getDebugLoc();
6775 SDValue Chain = Op.getOperand(0);
6776 SDValue Size = Op.getOperand(1);
6777 // FIXME: Ensure alignment here
6781 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6783 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6784 Flag = Chain.getValue(1);
6786 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6788 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6789 Flag = Chain.getValue(1);
6791 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6793 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6794 return DAG.getMergeValues(Ops1, 2, dl);
6797 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6798 MachineFunction &MF = DAG.getMachineFunction();
6799 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6801 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6802 DebugLoc dl = Op.getDebugLoc();
6804 if (!Subtarget->is64Bit()) {
6805 // vastart just stores the address of the VarArgsFrameIndex slot into the
6806 // memory location argument.
6807 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6809 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6814 // gp_offset (0 - 6 * 8)
6815 // fp_offset (48 - 48 + 8 * 16)
6816 // overflow_arg_area (point to parameters coming in memory).
6818 SmallVector<SDValue, 8> MemOps;
6819 SDValue FIN = Op.getOperand(1);
6821 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6822 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6824 FIN, SV, 0, false, false, 0);
6825 MemOps.push_back(Store);
6828 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6829 FIN, DAG.getIntPtrConstant(4));
6830 Store = DAG.getStore(Op.getOperand(0), dl,
6831 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6833 FIN, SV, 4, false, false, 0);
6834 MemOps.push_back(Store);
6836 // Store ptr to overflow_arg_area
6837 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6838 FIN, DAG.getIntPtrConstant(4));
6839 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6841 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 8,
6843 MemOps.push_back(Store);
6845 // Store ptr to reg_save_area.
6846 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6847 FIN, DAG.getIntPtrConstant(8));
6848 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6850 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 16,
6852 MemOps.push_back(Store);
6853 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6854 &MemOps[0], MemOps.size());
6857 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6858 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6859 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6861 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6865 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6866 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6867 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6868 SDValue Chain = Op.getOperand(0);
6869 SDValue DstPtr = Op.getOperand(1);
6870 SDValue SrcPtr = Op.getOperand(2);
6871 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6872 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6873 DebugLoc dl = Op.getDebugLoc();
6875 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6876 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6877 false, DstSV, 0, SrcSV, 0);
6881 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6882 DebugLoc dl = Op.getDebugLoc();
6883 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6885 default: return SDValue(); // Don't custom lower most intrinsics.
6886 // Comparison intrinsics.
6887 case Intrinsic::x86_sse_comieq_ss:
6888 case Intrinsic::x86_sse_comilt_ss:
6889 case Intrinsic::x86_sse_comile_ss:
6890 case Intrinsic::x86_sse_comigt_ss:
6891 case Intrinsic::x86_sse_comige_ss:
6892 case Intrinsic::x86_sse_comineq_ss:
6893 case Intrinsic::x86_sse_ucomieq_ss:
6894 case Intrinsic::x86_sse_ucomilt_ss:
6895 case Intrinsic::x86_sse_ucomile_ss:
6896 case Intrinsic::x86_sse_ucomigt_ss:
6897 case Intrinsic::x86_sse_ucomige_ss:
6898 case Intrinsic::x86_sse_ucomineq_ss:
6899 case Intrinsic::x86_sse2_comieq_sd:
6900 case Intrinsic::x86_sse2_comilt_sd:
6901 case Intrinsic::x86_sse2_comile_sd:
6902 case Intrinsic::x86_sse2_comigt_sd:
6903 case Intrinsic::x86_sse2_comige_sd:
6904 case Intrinsic::x86_sse2_comineq_sd:
6905 case Intrinsic::x86_sse2_ucomieq_sd:
6906 case Intrinsic::x86_sse2_ucomilt_sd:
6907 case Intrinsic::x86_sse2_ucomile_sd:
6908 case Intrinsic::x86_sse2_ucomigt_sd:
6909 case Intrinsic::x86_sse2_ucomige_sd:
6910 case Intrinsic::x86_sse2_ucomineq_sd: {
6912 ISD::CondCode CC = ISD::SETCC_INVALID;
6915 case Intrinsic::x86_sse_comieq_ss:
6916 case Intrinsic::x86_sse2_comieq_sd:
6920 case Intrinsic::x86_sse_comilt_ss:
6921 case Intrinsic::x86_sse2_comilt_sd:
6925 case Intrinsic::x86_sse_comile_ss:
6926 case Intrinsic::x86_sse2_comile_sd:
6930 case Intrinsic::x86_sse_comigt_ss:
6931 case Intrinsic::x86_sse2_comigt_sd:
6935 case Intrinsic::x86_sse_comige_ss:
6936 case Intrinsic::x86_sse2_comige_sd:
6940 case Intrinsic::x86_sse_comineq_ss:
6941 case Intrinsic::x86_sse2_comineq_sd:
6945 case Intrinsic::x86_sse_ucomieq_ss:
6946 case Intrinsic::x86_sse2_ucomieq_sd:
6947 Opc = X86ISD::UCOMI;
6950 case Intrinsic::x86_sse_ucomilt_ss:
6951 case Intrinsic::x86_sse2_ucomilt_sd:
6952 Opc = X86ISD::UCOMI;
6955 case Intrinsic::x86_sse_ucomile_ss:
6956 case Intrinsic::x86_sse2_ucomile_sd:
6957 Opc = X86ISD::UCOMI;
6960 case Intrinsic::x86_sse_ucomigt_ss:
6961 case Intrinsic::x86_sse2_ucomigt_sd:
6962 Opc = X86ISD::UCOMI;
6965 case Intrinsic::x86_sse_ucomige_ss:
6966 case Intrinsic::x86_sse2_ucomige_sd:
6967 Opc = X86ISD::UCOMI;
6970 case Intrinsic::x86_sse_ucomineq_ss:
6971 case Intrinsic::x86_sse2_ucomineq_sd:
6972 Opc = X86ISD::UCOMI;
6977 SDValue LHS = Op.getOperand(1);
6978 SDValue RHS = Op.getOperand(2);
6979 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6980 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6981 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6982 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6983 DAG.getConstant(X86CC, MVT::i8), Cond);
6984 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6986 // ptest intrinsics. The intrinsic these come from are designed to return
6987 // an integer value, not just an instruction so lower it to the ptest
6988 // pattern and a setcc for the result.
6989 case Intrinsic::x86_sse41_ptestz:
6990 case Intrinsic::x86_sse41_ptestc:
6991 case Intrinsic::x86_sse41_ptestnzc:{
6994 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6995 case Intrinsic::x86_sse41_ptestz:
6997 X86CC = X86::COND_E;
6999 case Intrinsic::x86_sse41_ptestc:
7001 X86CC = X86::COND_B;
7003 case Intrinsic::x86_sse41_ptestnzc:
7005 X86CC = X86::COND_A;
7009 SDValue LHS = Op.getOperand(1);
7010 SDValue RHS = Op.getOperand(2);
7011 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
7012 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7013 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7014 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7017 // Fix vector shift instructions where the last operand is a non-immediate
7019 case Intrinsic::x86_sse2_pslli_w:
7020 case Intrinsic::x86_sse2_pslli_d:
7021 case Intrinsic::x86_sse2_pslli_q:
7022 case Intrinsic::x86_sse2_psrli_w:
7023 case Intrinsic::x86_sse2_psrli_d:
7024 case Intrinsic::x86_sse2_psrli_q:
7025 case Intrinsic::x86_sse2_psrai_w:
7026 case Intrinsic::x86_sse2_psrai_d:
7027 case Intrinsic::x86_mmx_pslli_w:
7028 case Intrinsic::x86_mmx_pslli_d:
7029 case Intrinsic::x86_mmx_pslli_q:
7030 case Intrinsic::x86_mmx_psrli_w:
7031 case Intrinsic::x86_mmx_psrli_d:
7032 case Intrinsic::x86_mmx_psrli_q:
7033 case Intrinsic::x86_mmx_psrai_w:
7034 case Intrinsic::x86_mmx_psrai_d: {
7035 SDValue ShAmt = Op.getOperand(2);
7036 if (isa<ConstantSDNode>(ShAmt))
7039 unsigned NewIntNo = 0;
7040 EVT ShAmtVT = MVT::v4i32;
7042 case Intrinsic::x86_sse2_pslli_w:
7043 NewIntNo = Intrinsic::x86_sse2_psll_w;
7045 case Intrinsic::x86_sse2_pslli_d:
7046 NewIntNo = Intrinsic::x86_sse2_psll_d;
7048 case Intrinsic::x86_sse2_pslli_q:
7049 NewIntNo = Intrinsic::x86_sse2_psll_q;
7051 case Intrinsic::x86_sse2_psrli_w:
7052 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7054 case Intrinsic::x86_sse2_psrli_d:
7055 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7057 case Intrinsic::x86_sse2_psrli_q:
7058 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7060 case Intrinsic::x86_sse2_psrai_w:
7061 NewIntNo = Intrinsic::x86_sse2_psra_w;
7063 case Intrinsic::x86_sse2_psrai_d:
7064 NewIntNo = Intrinsic::x86_sse2_psra_d;
7067 ShAmtVT = MVT::v2i32;
7069 case Intrinsic::x86_mmx_pslli_w:
7070 NewIntNo = Intrinsic::x86_mmx_psll_w;
7072 case Intrinsic::x86_mmx_pslli_d:
7073 NewIntNo = Intrinsic::x86_mmx_psll_d;
7075 case Intrinsic::x86_mmx_pslli_q:
7076 NewIntNo = Intrinsic::x86_mmx_psll_q;
7078 case Intrinsic::x86_mmx_psrli_w:
7079 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7081 case Intrinsic::x86_mmx_psrli_d:
7082 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7084 case Intrinsic::x86_mmx_psrli_q:
7085 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7087 case Intrinsic::x86_mmx_psrai_w:
7088 NewIntNo = Intrinsic::x86_mmx_psra_w;
7090 case Intrinsic::x86_mmx_psrai_d:
7091 NewIntNo = Intrinsic::x86_mmx_psra_d;
7093 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7099 // The vector shift intrinsics with scalars uses 32b shift amounts but
7100 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7104 ShOps[1] = DAG.getConstant(0, MVT::i32);
7105 if (ShAmtVT == MVT::v4i32) {
7106 ShOps[2] = DAG.getUNDEF(MVT::i32);
7107 ShOps[3] = DAG.getUNDEF(MVT::i32);
7108 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7110 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7113 EVT VT = Op.getValueType();
7114 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7115 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7116 DAG.getConstant(NewIntNo, MVT::i32),
7117 Op.getOperand(1), ShAmt);
7122 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7123 SelectionDAG &DAG) const {
7124 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7125 MFI->setReturnAddressIsTaken(true);
7127 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7128 DebugLoc dl = Op.getDebugLoc();
7131 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7133 DAG.getConstant(TD->getPointerSize(),
7134 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7135 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7136 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7138 NULL, 0, false, false, 0);
7141 // Just load the return address.
7142 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7143 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7144 RetAddrFI, NULL, 0, false, false, 0);
7147 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7148 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7149 MFI->setFrameAddressIsTaken(true);
7151 EVT VT = Op.getValueType();
7152 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7153 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7154 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7155 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7157 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7162 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7163 SelectionDAG &DAG) const {
7164 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7167 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7168 MachineFunction &MF = DAG.getMachineFunction();
7169 SDValue Chain = Op.getOperand(0);
7170 SDValue Offset = Op.getOperand(1);
7171 SDValue Handler = Op.getOperand(2);
7172 DebugLoc dl = Op.getDebugLoc();
7174 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7176 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7178 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7179 DAG.getIntPtrConstant(-TD->getPointerSize()));
7180 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7181 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7182 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7183 MF.getRegInfo().addLiveOut(StoreAddrReg);
7185 return DAG.getNode(X86ISD::EH_RETURN, dl,
7187 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7190 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7191 SelectionDAG &DAG) const {
7192 SDValue Root = Op.getOperand(0);
7193 SDValue Trmp = Op.getOperand(1); // trampoline
7194 SDValue FPtr = Op.getOperand(2); // nested function
7195 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7196 DebugLoc dl = Op.getDebugLoc();
7198 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7200 if (Subtarget->is64Bit()) {
7201 SDValue OutChains[6];
7203 // Large code-model.
7204 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7205 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7207 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7208 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7210 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7212 // Load the pointer to the nested function into R11.
7213 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7214 SDValue Addr = Trmp;
7215 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7216 Addr, TrmpAddr, 0, false, false, 0);
7218 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7219 DAG.getConstant(2, MVT::i64));
7220 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7223 // Load the 'nest' parameter value into R10.
7224 // R10 is specified in X86CallingConv.td
7225 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7226 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7227 DAG.getConstant(10, MVT::i64));
7228 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7229 Addr, TrmpAddr, 10, false, false, 0);
7231 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7232 DAG.getConstant(12, MVT::i64));
7233 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7236 // Jump to the nested function.
7237 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7238 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7239 DAG.getConstant(20, MVT::i64));
7240 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7241 Addr, TrmpAddr, 20, false, false, 0);
7243 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7244 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7245 DAG.getConstant(22, MVT::i64));
7246 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7247 TrmpAddr, 22, false, false, 0);
7250 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7251 return DAG.getMergeValues(Ops, 2, dl);
7253 const Function *Func =
7254 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7255 CallingConv::ID CC = Func->getCallingConv();
7260 llvm_unreachable("Unsupported calling convention");
7261 case CallingConv::C:
7262 case CallingConv::X86_StdCall: {
7263 // Pass 'nest' parameter in ECX.
7264 // Must be kept in sync with X86CallingConv.td
7267 // Check that ECX wasn't needed by an 'inreg' parameter.
7268 const FunctionType *FTy = Func->getFunctionType();
7269 const AttrListPtr &Attrs = Func->getAttributes();
7271 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7272 unsigned InRegCount = 0;
7275 for (FunctionType::param_iterator I = FTy->param_begin(),
7276 E = FTy->param_end(); I != E; ++I, ++Idx)
7277 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7278 // FIXME: should only count parameters that are lowered to integers.
7279 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7281 if (InRegCount > 2) {
7282 report_fatal_error("Nest register in use - reduce number of inreg"
7288 case CallingConv::X86_FastCall:
7289 case CallingConv::X86_ThisCall:
7290 case CallingConv::Fast:
7291 // Pass 'nest' parameter in EAX.
7292 // Must be kept in sync with X86CallingConv.td
7297 SDValue OutChains[4];
7300 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7301 DAG.getConstant(10, MVT::i32));
7302 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7304 // This is storing the opcode for MOV32ri.
7305 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7306 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7307 OutChains[0] = DAG.getStore(Root, dl,
7308 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7309 Trmp, TrmpAddr, 0, false, false, 0);
7311 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7312 DAG.getConstant(1, MVT::i32));
7313 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7316 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7317 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7318 DAG.getConstant(5, MVT::i32));
7319 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7320 TrmpAddr, 5, false, false, 1);
7322 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7323 DAG.getConstant(6, MVT::i32));
7324 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7328 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7329 return DAG.getMergeValues(Ops, 2, dl);
7333 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7334 SelectionDAG &DAG) const {
7336 The rounding mode is in bits 11:10 of FPSR, and has the following
7343 FLT_ROUNDS, on the other hand, expects the following:
7350 To perform the conversion, we do:
7351 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7354 MachineFunction &MF = DAG.getMachineFunction();
7355 const TargetMachine &TM = MF.getTarget();
7356 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7357 unsigned StackAlignment = TFI.getStackAlignment();
7358 EVT VT = Op.getValueType();
7359 DebugLoc dl = Op.getDebugLoc();
7361 // Save FP Control Word to stack slot
7362 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7363 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7365 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7366 DAG.getEntryNode(), StackSlot);
7368 // Load FP Control Word from stack slot
7369 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7372 // Transform as necessary
7374 DAG.getNode(ISD::SRL, dl, MVT::i16,
7375 DAG.getNode(ISD::AND, dl, MVT::i16,
7376 CWD, DAG.getConstant(0x800, MVT::i16)),
7377 DAG.getConstant(11, MVT::i8));
7379 DAG.getNode(ISD::SRL, dl, MVT::i16,
7380 DAG.getNode(ISD::AND, dl, MVT::i16,
7381 CWD, DAG.getConstant(0x400, MVT::i16)),
7382 DAG.getConstant(9, MVT::i8));
7385 DAG.getNode(ISD::AND, dl, MVT::i16,
7386 DAG.getNode(ISD::ADD, dl, MVT::i16,
7387 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7388 DAG.getConstant(1, MVT::i16)),
7389 DAG.getConstant(3, MVT::i16));
7392 return DAG.getNode((VT.getSizeInBits() < 16 ?
7393 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7396 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7397 EVT VT = Op.getValueType();
7399 unsigned NumBits = VT.getSizeInBits();
7400 DebugLoc dl = Op.getDebugLoc();
7402 Op = Op.getOperand(0);
7403 if (VT == MVT::i8) {
7404 // Zero extend to i32 since there is not an i8 bsr.
7406 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7409 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7410 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7411 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7413 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7416 DAG.getConstant(NumBits+NumBits-1, OpVT),
7417 DAG.getConstant(X86::COND_E, MVT::i8),
7420 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7422 // Finally xor with NumBits-1.
7423 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7426 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7430 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7431 EVT VT = Op.getValueType();
7433 unsigned NumBits = VT.getSizeInBits();
7434 DebugLoc dl = Op.getDebugLoc();
7436 Op = Op.getOperand(0);
7437 if (VT == MVT::i8) {
7439 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7442 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7443 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7444 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7446 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7449 DAG.getConstant(NumBits, OpVT),
7450 DAG.getConstant(X86::COND_E, MVT::i8),
7453 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7456 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7460 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7461 EVT VT = Op.getValueType();
7462 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7463 DebugLoc dl = Op.getDebugLoc();
7465 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7466 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7467 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7468 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7469 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7471 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7472 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7473 // return AloBlo + AloBhi + AhiBlo;
7475 SDValue A = Op.getOperand(0);
7476 SDValue B = Op.getOperand(1);
7478 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7479 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7480 A, DAG.getConstant(32, MVT::i32));
7481 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7482 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7483 B, DAG.getConstant(32, MVT::i32));
7484 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7485 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7487 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7488 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7490 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7491 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7493 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7494 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7495 AloBhi, DAG.getConstant(32, MVT::i32));
7496 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7497 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7498 AhiBlo, DAG.getConstant(32, MVT::i32));
7499 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7500 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7504 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
7505 EVT VT = Op.getValueType();
7506 DebugLoc dl = Op.getDebugLoc();
7507 SDValue R = Op.getOperand(0);
7509 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
7510 assert(VT == MVT::v4i32 && "Only know how to lower v4i32");
7512 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7513 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
7514 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
7516 std::vector<Constant*> CV;
7517 LLVMContext *Context = DAG.getContext();
7518 CV.push_back(ConstantInt::get(*Context, APInt(32, 0x3f800000U)));
7519 CV.push_back(ConstantInt::get(*Context, APInt(32, 0x3f800000U)));
7520 CV.push_back(ConstantInt::get(*Context, APInt(32, 0x3f800000U)));
7521 CV.push_back(ConstantInt::get(*Context, APInt(32, 0x3f800000U)));
7522 Constant *C = ConstantVector::get(CV);
7523 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7524 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7525 PseudoSourceValue::getConstantPool(), 0,
7528 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
7529 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, Op);
7530 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
7531 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
7534 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7535 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7536 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7537 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7538 // has only one use.
7539 SDNode *N = Op.getNode();
7540 SDValue LHS = N->getOperand(0);
7541 SDValue RHS = N->getOperand(1);
7542 unsigned BaseOp = 0;
7544 DebugLoc dl = Op.getDebugLoc();
7546 switch (Op.getOpcode()) {
7547 default: llvm_unreachable("Unknown ovf instruction!");
7549 // A subtract of one will be selected as a INC. Note that INC doesn't
7550 // set CF, so we can't do this for UADDO.
7551 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7552 if (C->getAPIntValue() == 1) {
7553 BaseOp = X86ISD::INC;
7557 BaseOp = X86ISD::ADD;
7561 BaseOp = X86ISD::ADD;
7565 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7566 // set CF, so we can't do this for USUBO.
7567 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7568 if (C->getAPIntValue() == 1) {
7569 BaseOp = X86ISD::DEC;
7573 BaseOp = X86ISD::SUB;
7577 BaseOp = X86ISD::SUB;
7581 BaseOp = X86ISD::SMUL;
7585 BaseOp = X86ISD::UMUL;
7590 // Also sets EFLAGS.
7591 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7592 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7595 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7596 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7598 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7602 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
7603 DebugLoc dl = Op.getDebugLoc();
7605 if (!Subtarget->hasSSE2())
7606 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
7607 DAG.getConstant(0, MVT::i32));
7609 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
7611 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
7613 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7614 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
7615 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
7616 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
7618 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
7619 if (!Op1 && !Op2 && !Op3 && Op4)
7620 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
7622 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
7623 if (Op1 && !Op2 && !Op3 && !Op4)
7624 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
7626 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
7628 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
7632 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7633 EVT T = Op.getValueType();
7634 DebugLoc dl = Op.getDebugLoc();
7637 switch(T.getSimpleVT().SimpleTy) {
7639 assert(false && "Invalid value type!");
7640 case MVT::i8: Reg = X86::AL; size = 1; break;
7641 case MVT::i16: Reg = X86::AX; size = 2; break;
7642 case MVT::i32: Reg = X86::EAX; size = 4; break;
7644 assert(Subtarget->is64Bit() && "Node not type legal!");
7645 Reg = X86::RAX; size = 8;
7648 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7649 Op.getOperand(2), SDValue());
7650 SDValue Ops[] = { cpIn.getValue(0),
7653 DAG.getTargetConstant(size, MVT::i8),
7655 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7656 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7658 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7662 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7663 SelectionDAG &DAG) const {
7664 assert(Subtarget->is64Bit() && "Result not type legalized?");
7665 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7666 SDValue TheChain = Op.getOperand(0);
7667 DebugLoc dl = Op.getDebugLoc();
7668 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7669 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7670 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7672 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7673 DAG.getConstant(32, MVT::i8));
7675 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7678 return DAG.getMergeValues(Ops, 2, dl);
7681 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7682 SelectionDAG &DAG) const {
7683 EVT SrcVT = Op.getOperand(0).getValueType();
7684 EVT DstVT = Op.getValueType();
7685 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7686 Subtarget->hasMMX() && !DisableMMX) &&
7687 "Unexpected custom BIT_CONVERT");
7688 assert((DstVT == MVT::i64 ||
7689 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7690 "Unexpected custom BIT_CONVERT");
7691 // i64 <=> MMX conversions are Legal.
7692 if (SrcVT==MVT::i64 && DstVT.isVector())
7694 if (DstVT==MVT::i64 && SrcVT.isVector())
7696 // MMX <=> MMX conversions are Legal.
7697 if (SrcVT.isVector() && DstVT.isVector())
7699 // All other conversions need to be expanded.
7702 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7703 SDNode *Node = Op.getNode();
7704 DebugLoc dl = Node->getDebugLoc();
7705 EVT T = Node->getValueType(0);
7706 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7707 DAG.getConstant(0, T), Node->getOperand(2));
7708 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7709 cast<AtomicSDNode>(Node)->getMemoryVT(),
7710 Node->getOperand(0),
7711 Node->getOperand(1), negOp,
7712 cast<AtomicSDNode>(Node)->getSrcValue(),
7713 cast<AtomicSDNode>(Node)->getAlignment());
7716 /// LowerOperation - Provide custom lowering hooks for some operations.
7718 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7719 switch (Op.getOpcode()) {
7720 default: llvm_unreachable("Should not custom lower this!");
7721 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
7722 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7723 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7724 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7725 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7726 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7727 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7728 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7729 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7730 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7731 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7732 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7733 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7734 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7735 case ISD::SHL_PARTS:
7736 case ISD::SRA_PARTS:
7737 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7738 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7739 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7740 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7741 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7742 case ISD::FABS: return LowerFABS(Op, DAG);
7743 case ISD::FNEG: return LowerFNEG(Op, DAG);
7744 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7745 case ISD::SETCC: return LowerSETCC(Op, DAG);
7746 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7747 case ISD::SELECT: return LowerSELECT(Op, DAG);
7748 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7749 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7750 case ISD::VASTART: return LowerVASTART(Op, DAG);
7751 case ISD::VAARG: return LowerVAARG(Op, DAG);
7752 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7753 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7754 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7755 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7756 case ISD::FRAME_TO_ARGS_OFFSET:
7757 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7758 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7759 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7760 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7761 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7762 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7763 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7764 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7765 case ISD::SHL: return LowerSHL(Op, DAG);
7771 case ISD::UMULO: return LowerXALUO(Op, DAG);
7772 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7773 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7777 void X86TargetLowering::
7778 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7779 SelectionDAG &DAG, unsigned NewOp) const {
7780 EVT T = Node->getValueType(0);
7781 DebugLoc dl = Node->getDebugLoc();
7782 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7784 SDValue Chain = Node->getOperand(0);
7785 SDValue In1 = Node->getOperand(1);
7786 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7787 Node->getOperand(2), DAG.getIntPtrConstant(0));
7788 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7789 Node->getOperand(2), DAG.getIntPtrConstant(1));
7790 SDValue Ops[] = { Chain, In1, In2L, In2H };
7791 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7793 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7794 cast<MemSDNode>(Node)->getMemOperand());
7795 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7796 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7797 Results.push_back(Result.getValue(2));
7800 /// ReplaceNodeResults - Replace a node with an illegal result type
7801 /// with a new node built out of custom code.
7802 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7803 SmallVectorImpl<SDValue>&Results,
7804 SelectionDAG &DAG) const {
7805 DebugLoc dl = N->getDebugLoc();
7806 switch (N->getOpcode()) {
7808 assert(false && "Do not know how to custom type legalize this operation!");
7810 case ISD::FP_TO_SINT: {
7811 std::pair<SDValue,SDValue> Vals =
7812 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7813 SDValue FIST = Vals.first, StackSlot = Vals.second;
7814 if (FIST.getNode() != 0) {
7815 EVT VT = N->getValueType(0);
7816 // Return a load from the stack slot.
7817 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7822 case ISD::READCYCLECOUNTER: {
7823 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7824 SDValue TheChain = N->getOperand(0);
7825 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7826 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7828 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7830 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7831 SDValue Ops[] = { eax, edx };
7832 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7833 Results.push_back(edx.getValue(1));
7836 case ISD::ATOMIC_CMP_SWAP: {
7837 EVT T = N->getValueType(0);
7838 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7839 SDValue cpInL, cpInH;
7840 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7841 DAG.getConstant(0, MVT::i32));
7842 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7843 DAG.getConstant(1, MVT::i32));
7844 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7845 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7847 SDValue swapInL, swapInH;
7848 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7849 DAG.getConstant(0, MVT::i32));
7850 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7851 DAG.getConstant(1, MVT::i32));
7852 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7854 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7855 swapInL.getValue(1));
7856 SDValue Ops[] = { swapInH.getValue(0),
7858 swapInH.getValue(1) };
7859 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7860 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7861 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7862 MVT::i32, Result.getValue(1));
7863 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7864 MVT::i32, cpOutL.getValue(2));
7865 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7866 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7867 Results.push_back(cpOutH.getValue(1));
7870 case ISD::ATOMIC_LOAD_ADD:
7871 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7873 case ISD::ATOMIC_LOAD_AND:
7874 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7876 case ISD::ATOMIC_LOAD_NAND:
7877 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7879 case ISD::ATOMIC_LOAD_OR:
7880 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7882 case ISD::ATOMIC_LOAD_SUB:
7883 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7885 case ISD::ATOMIC_LOAD_XOR:
7886 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7888 case ISD::ATOMIC_SWAP:
7889 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7894 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7896 default: return NULL;
7897 case X86ISD::BSF: return "X86ISD::BSF";
7898 case X86ISD::BSR: return "X86ISD::BSR";
7899 case X86ISD::SHLD: return "X86ISD::SHLD";
7900 case X86ISD::SHRD: return "X86ISD::SHRD";
7901 case X86ISD::FAND: return "X86ISD::FAND";
7902 case X86ISD::FOR: return "X86ISD::FOR";
7903 case X86ISD::FXOR: return "X86ISD::FXOR";
7904 case X86ISD::FSRL: return "X86ISD::FSRL";
7905 case X86ISD::FILD: return "X86ISD::FILD";
7906 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7907 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7908 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7909 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7910 case X86ISD::FLD: return "X86ISD::FLD";
7911 case X86ISD::FST: return "X86ISD::FST";
7912 case X86ISD::CALL: return "X86ISD::CALL";
7913 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7914 case X86ISD::BT: return "X86ISD::BT";
7915 case X86ISD::CMP: return "X86ISD::CMP";
7916 case X86ISD::COMI: return "X86ISD::COMI";
7917 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7918 case X86ISD::SETCC: return "X86ISD::SETCC";
7919 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7920 case X86ISD::CMOV: return "X86ISD::CMOV";
7921 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7922 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7923 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7924 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7925 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7926 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7927 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7928 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7929 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7930 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7931 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7932 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7933 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7934 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7935 case X86ISD::FMAX: return "X86ISD::FMAX";
7936 case X86ISD::FMIN: return "X86ISD::FMIN";
7937 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7938 case X86ISD::FRCP: return "X86ISD::FRCP";
7939 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7940 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
7941 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7942 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7943 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7944 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7945 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7946 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7947 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7948 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7949 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7950 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7951 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7952 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7953 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7954 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7955 case X86ISD::VSHL: return "X86ISD::VSHL";
7956 case X86ISD::VSRL: return "X86ISD::VSRL";
7957 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7958 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7959 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7960 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7961 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7962 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7963 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7964 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7965 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7966 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7967 case X86ISD::ADD: return "X86ISD::ADD";
7968 case X86ISD::SUB: return "X86ISD::SUB";
7969 case X86ISD::SMUL: return "X86ISD::SMUL";
7970 case X86ISD::UMUL: return "X86ISD::UMUL";
7971 case X86ISD::INC: return "X86ISD::INC";
7972 case X86ISD::DEC: return "X86ISD::DEC";
7973 case X86ISD::OR: return "X86ISD::OR";
7974 case X86ISD::XOR: return "X86ISD::XOR";
7975 case X86ISD::AND: return "X86ISD::AND";
7976 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7977 case X86ISD::PTEST: return "X86ISD::PTEST";
7978 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7979 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7983 // isLegalAddressingMode - Return true if the addressing mode represented
7984 // by AM is legal for this target, for a load/store of the specified type.
7985 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7986 const Type *Ty) const {
7987 // X86 supports extremely general addressing modes.
7988 CodeModel::Model M = getTargetMachine().getCodeModel();
7990 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7991 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7996 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7998 // If a reference to this global requires an extra load, we can't fold it.
7999 if (isGlobalStubReference(GVFlags))
8002 // If BaseGV requires a register for the PIC base, we cannot also have a
8003 // BaseReg specified.
8004 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
8007 // If lower 4G is not available, then we must use rip-relative addressing.
8008 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
8018 // These scales always work.
8023 // These scales are formed with basereg+scalereg. Only accept if there is
8028 default: // Other stuff never works.
8036 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
8037 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
8039 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
8040 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
8041 if (NumBits1 <= NumBits2)
8046 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
8047 if (!VT1.isInteger() || !VT2.isInteger())
8049 unsigned NumBits1 = VT1.getSizeInBits();
8050 unsigned NumBits2 = VT2.getSizeInBits();
8051 if (NumBits1 <= NumBits2)
8056 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
8057 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8058 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
8061 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
8062 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8063 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
8066 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
8067 // i16 instructions are longer (0x66 prefix) and potentially slower.
8068 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
8071 /// isShuffleMaskLegal - Targets can use this to indicate that they only
8072 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
8073 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
8074 /// are assumed to be legal.
8076 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
8078 // Very little shuffling can be done for 64-bit vectors right now.
8079 if (VT.getSizeInBits() == 64)
8080 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8082 // FIXME: pshufb, blends, shifts.
8083 return (VT.getVectorNumElements() == 2 ||
8084 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8085 isMOVLMask(M, VT) ||
8086 isSHUFPMask(M, VT) ||
8087 isPSHUFDMask(M, VT) ||
8088 isPSHUFHWMask(M, VT) ||
8089 isPSHUFLWMask(M, VT) ||
8090 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8091 isUNPCKLMask(M, VT) ||
8092 isUNPCKHMask(M, VT) ||
8093 isUNPCKL_v_undef_Mask(M, VT) ||
8094 isUNPCKH_v_undef_Mask(M, VT));
8098 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8100 unsigned NumElts = VT.getVectorNumElements();
8101 // FIXME: This collection of masks seems suspect.
8104 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8105 return (isMOVLMask(Mask, VT) ||
8106 isCommutedMOVLMask(Mask, VT, true) ||
8107 isSHUFPMask(Mask, VT) ||
8108 isCommutedSHUFPMask(Mask, VT));
8113 //===----------------------------------------------------------------------===//
8114 // X86 Scheduler Hooks
8115 //===----------------------------------------------------------------------===//
8117 // private utility function
8119 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8120 MachineBasicBlock *MBB,
8127 TargetRegisterClass *RC,
8128 bool invSrc) const {
8129 // For the atomic bitwise operator, we generate
8132 // ld t1 = [bitinstr.addr]
8133 // op t2 = t1, [bitinstr.val]
8135 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8137 // fallthrough -->nextMBB
8138 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8139 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8140 MachineFunction::iterator MBBIter = MBB;
8143 /// First build the CFG
8144 MachineFunction *F = MBB->getParent();
8145 MachineBasicBlock *thisMBB = MBB;
8146 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8147 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8148 F->insert(MBBIter, newMBB);
8149 F->insert(MBBIter, nextMBB);
8151 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8152 nextMBB->splice(nextMBB->begin(), thisMBB,
8153 llvm::next(MachineBasicBlock::iterator(bInstr)),
8155 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8157 // Update thisMBB to fall through to newMBB
8158 thisMBB->addSuccessor(newMBB);
8160 // newMBB jumps to itself and fall through to nextMBB
8161 newMBB->addSuccessor(nextMBB);
8162 newMBB->addSuccessor(newMBB);
8164 // Insert instructions into newMBB based on incoming instruction
8165 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8166 "unexpected number of operands");
8167 DebugLoc dl = bInstr->getDebugLoc();
8168 MachineOperand& destOper = bInstr->getOperand(0);
8169 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8170 int numArgs = bInstr->getNumOperands() - 1;
8171 for (int i=0; i < numArgs; ++i)
8172 argOpers[i] = &bInstr->getOperand(i+1);
8174 // x86 address has 4 operands: base, index, scale, and displacement
8175 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8176 int valArgIndx = lastAddrIndx + 1;
8178 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8179 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8180 for (int i=0; i <= lastAddrIndx; ++i)
8181 (*MIB).addOperand(*argOpers[i]);
8183 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8185 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8190 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8191 assert((argOpers[valArgIndx]->isReg() ||
8192 argOpers[valArgIndx]->isImm()) &&
8194 if (argOpers[valArgIndx]->isReg())
8195 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8197 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8199 (*MIB).addOperand(*argOpers[valArgIndx]);
8201 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
8204 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8205 for (int i=0; i <= lastAddrIndx; ++i)
8206 (*MIB).addOperand(*argOpers[i]);
8208 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8209 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8210 bInstr->memoperands_end());
8212 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8216 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8218 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8222 // private utility function: 64 bit atomics on 32 bit host.
8224 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8225 MachineBasicBlock *MBB,
8230 bool invSrc) const {
8231 // For the atomic bitwise operator, we generate
8232 // thisMBB (instructions are in pairs, except cmpxchg8b)
8233 // ld t1,t2 = [bitinstr.addr]
8235 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8236 // op t5, t6 <- out1, out2, [bitinstr.val]
8237 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8238 // mov ECX, EBX <- t5, t6
8239 // mov EAX, EDX <- t1, t2
8240 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8241 // mov t3, t4 <- EAX, EDX
8243 // result in out1, out2
8244 // fallthrough -->nextMBB
8246 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8247 const unsigned LoadOpc = X86::MOV32rm;
8248 const unsigned NotOpc = X86::NOT32r;
8249 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8250 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8251 MachineFunction::iterator MBBIter = MBB;
8254 /// First build the CFG
8255 MachineFunction *F = MBB->getParent();
8256 MachineBasicBlock *thisMBB = MBB;
8257 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8258 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8259 F->insert(MBBIter, newMBB);
8260 F->insert(MBBIter, nextMBB);
8262 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8263 nextMBB->splice(nextMBB->begin(), thisMBB,
8264 llvm::next(MachineBasicBlock::iterator(bInstr)),
8266 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8268 // Update thisMBB to fall through to newMBB
8269 thisMBB->addSuccessor(newMBB);
8271 // newMBB jumps to itself and fall through to nextMBB
8272 newMBB->addSuccessor(nextMBB);
8273 newMBB->addSuccessor(newMBB);
8275 DebugLoc dl = bInstr->getDebugLoc();
8276 // Insert instructions into newMBB based on incoming instruction
8277 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8278 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
8279 "unexpected number of operands");
8280 MachineOperand& dest1Oper = bInstr->getOperand(0);
8281 MachineOperand& dest2Oper = bInstr->getOperand(1);
8282 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8283 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
8284 argOpers[i] = &bInstr->getOperand(i+2);
8286 // We use some of the operands multiple times, so conservatively just
8287 // clear any kill flags that might be present.
8288 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8289 argOpers[i]->setIsKill(false);
8292 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8293 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8295 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8296 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8297 for (int i=0; i <= lastAddrIndx; ++i)
8298 (*MIB).addOperand(*argOpers[i]);
8299 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8300 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8301 // add 4 to displacement.
8302 for (int i=0; i <= lastAddrIndx-2; ++i)
8303 (*MIB).addOperand(*argOpers[i]);
8304 MachineOperand newOp3 = *(argOpers[3]);
8306 newOp3.setImm(newOp3.getImm()+4);
8308 newOp3.setOffset(newOp3.getOffset()+4);
8309 (*MIB).addOperand(newOp3);
8310 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8312 // t3/4 are defined later, at the bottom of the loop
8313 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8314 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8315 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8316 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8317 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8318 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8320 // The subsequent operations should be using the destination registers of
8321 //the PHI instructions.
8323 t1 = F->getRegInfo().createVirtualRegister(RC);
8324 t2 = F->getRegInfo().createVirtualRegister(RC);
8325 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8326 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8328 t1 = dest1Oper.getReg();
8329 t2 = dest2Oper.getReg();
8332 int valArgIndx = lastAddrIndx + 1;
8333 assert((argOpers[valArgIndx]->isReg() ||
8334 argOpers[valArgIndx]->isImm()) &&
8336 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8337 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8338 if (argOpers[valArgIndx]->isReg())
8339 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8341 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8342 if (regOpcL != X86::MOV32rr)
8344 (*MIB).addOperand(*argOpers[valArgIndx]);
8345 assert(argOpers[valArgIndx + 1]->isReg() ==
8346 argOpers[valArgIndx]->isReg());
8347 assert(argOpers[valArgIndx + 1]->isImm() ==
8348 argOpers[valArgIndx]->isImm());
8349 if (argOpers[valArgIndx + 1]->isReg())
8350 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8352 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8353 if (regOpcH != X86::MOV32rr)
8355 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8357 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8359 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
8362 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
8364 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
8367 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8368 for (int i=0; i <= lastAddrIndx; ++i)
8369 (*MIB).addOperand(*argOpers[i]);
8371 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8372 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8373 bInstr->memoperands_end());
8375 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
8376 MIB.addReg(X86::EAX);
8377 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
8378 MIB.addReg(X86::EDX);
8381 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8383 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8387 // private utility function
8389 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8390 MachineBasicBlock *MBB,
8391 unsigned cmovOpc) const {
8392 // For the atomic min/max operator, we generate
8395 // ld t1 = [min/max.addr]
8396 // mov t2 = [min/max.val]
8398 // cmov[cond] t2 = t1
8400 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8402 // fallthrough -->nextMBB
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(mInstr)),
8421 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8423 // Update thisMBB to fall through to newMBB
8424 thisMBB->addSuccessor(newMBB);
8426 // newMBB jumps to newMBB and fall through to nextMBB
8427 newMBB->addSuccessor(nextMBB);
8428 newMBB->addSuccessor(newMBB);
8430 DebugLoc dl = mInstr->getDebugLoc();
8431 // Insert instructions into newMBB based on incoming instruction
8432 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8433 "unexpected number of operands");
8434 MachineOperand& destOper = mInstr->getOperand(0);
8435 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8436 int numArgs = mInstr->getNumOperands() - 1;
8437 for (int i=0; i < numArgs; ++i)
8438 argOpers[i] = &mInstr->getOperand(i+1);
8440 // x86 address has 4 operands: base, index, scale, and displacement
8441 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8442 int valArgIndx = lastAddrIndx + 1;
8444 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8445 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8446 for (int i=0; i <= lastAddrIndx; ++i)
8447 (*MIB).addOperand(*argOpers[i]);
8449 // We only support register and immediate values
8450 assert((argOpers[valArgIndx]->isReg() ||
8451 argOpers[valArgIndx]->isImm()) &&
8454 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8455 if (argOpers[valArgIndx]->isReg())
8456 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
8458 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8459 (*MIB).addOperand(*argOpers[valArgIndx]);
8461 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8464 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8469 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8470 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8474 // Cmp and exchange if none has modified the memory location
8475 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8476 for (int i=0; i <= lastAddrIndx; ++i)
8477 (*MIB).addOperand(*argOpers[i]);
8479 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8480 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8481 mInstr->memoperands_end());
8483 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8484 MIB.addReg(X86::EAX);
8487 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8489 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
8493 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8494 // all of this code can be replaced with that in the .td file.
8496 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8497 unsigned numArgs, bool memArg) const {
8499 DebugLoc dl = MI->getDebugLoc();
8500 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8504 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8506 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8508 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8510 for (unsigned i = 0; i < numArgs; ++i) {
8511 MachineOperand &Op = MI->getOperand(i+1);
8513 if (!(Op.isReg() && Op.isImplicit()))
8517 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8520 MI->eraseFromParent();
8526 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8528 MachineBasicBlock *MBB) const {
8529 // Emit code to save XMM registers to the stack. The ABI says that the
8530 // number of registers to save is given in %al, so it's theoretically
8531 // possible to do an indirect jump trick to avoid saving all of them,
8532 // however this code takes a simpler approach and just executes all
8533 // of the stores if %al is non-zero. It's less code, and it's probably
8534 // easier on the hardware branch predictor, and stores aren't all that
8535 // expensive anyway.
8537 // Create the new basic blocks. One block contains all the XMM stores,
8538 // and one block is the final destination regardless of whether any
8539 // stores were performed.
8540 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8541 MachineFunction *F = MBB->getParent();
8542 MachineFunction::iterator MBBIter = MBB;
8544 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8545 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8546 F->insert(MBBIter, XMMSaveMBB);
8547 F->insert(MBBIter, EndMBB);
8549 // Transfer the remainder of MBB and its successor edges to EndMBB.
8550 EndMBB->splice(EndMBB->begin(), MBB,
8551 llvm::next(MachineBasicBlock::iterator(MI)),
8553 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
8555 // The original block will now fall through to the XMM save block.
8556 MBB->addSuccessor(XMMSaveMBB);
8557 // The XMMSaveMBB will fall through to the end block.
8558 XMMSaveMBB->addSuccessor(EndMBB);
8560 // Now add the instructions.
8561 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8562 DebugLoc DL = MI->getDebugLoc();
8564 unsigned CountReg = MI->getOperand(0).getReg();
8565 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8566 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8568 if (!Subtarget->isTargetWin64()) {
8569 // If %al is 0, branch around the XMM save block.
8570 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8571 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8572 MBB->addSuccessor(EndMBB);
8575 // In the XMM save block, save all the XMM argument registers.
8576 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8577 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8578 MachineMemOperand *MMO =
8579 F->getMachineMemOperand(
8580 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8581 MachineMemOperand::MOStore, Offset,
8582 /*Size=*/16, /*Align=*/16);
8583 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8584 .addFrameIndex(RegSaveFrameIndex)
8585 .addImm(/*Scale=*/1)
8586 .addReg(/*IndexReg=*/0)
8587 .addImm(/*Disp=*/Offset)
8588 .addReg(/*Segment=*/0)
8589 .addReg(MI->getOperand(i).getReg())
8590 .addMemOperand(MMO);
8593 MI->eraseFromParent(); // The pseudo instruction is gone now.
8599 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8600 MachineBasicBlock *BB) const {
8601 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8602 DebugLoc DL = MI->getDebugLoc();
8604 // To "insert" a SELECT_CC instruction, we actually have to insert the
8605 // diamond control-flow pattern. The incoming instruction knows the
8606 // destination vreg to set, the condition code register to branch on, the
8607 // true/false values to select between, and a branch opcode to use.
8608 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8609 MachineFunction::iterator It = BB;
8615 // cmpTY ccX, r1, r2
8617 // fallthrough --> copy0MBB
8618 MachineBasicBlock *thisMBB = BB;
8619 MachineFunction *F = BB->getParent();
8620 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8621 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8622 F->insert(It, copy0MBB);
8623 F->insert(It, sinkMBB);
8625 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8626 // live into the sink and copy blocks.
8627 const MachineFunction *MF = BB->getParent();
8628 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8629 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8631 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
8632 const MachineOperand &MO = MI->getOperand(I);
8633 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
8634 unsigned Reg = MO.getReg();
8635 if (Reg != X86::EFLAGS) continue;
8636 copy0MBB->addLiveIn(Reg);
8637 sinkMBB->addLiveIn(Reg);
8640 // Transfer the remainder of BB and its successor edges to sinkMBB.
8641 sinkMBB->splice(sinkMBB->begin(), BB,
8642 llvm::next(MachineBasicBlock::iterator(MI)),
8644 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
8646 // Add the true and fallthrough blocks as its successors.
8647 BB->addSuccessor(copy0MBB);
8648 BB->addSuccessor(sinkMBB);
8650 // Create the conditional branch instruction.
8652 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8653 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8656 // %FalseValue = ...
8657 // # fallthrough to sinkMBB
8658 copy0MBB->addSuccessor(sinkMBB);
8661 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8663 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
8664 TII->get(X86::PHI), MI->getOperand(0).getReg())
8665 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8666 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8668 MI->eraseFromParent(); // The pseudo instruction is gone now.
8673 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8674 MachineBasicBlock *BB) const {
8675 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8676 DebugLoc DL = MI->getDebugLoc();
8678 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8679 // non-trivial part is impdef of ESP.
8680 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8683 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
8684 .addExternalSymbol("_alloca")
8685 .addReg(X86::EAX, RegState::Implicit)
8686 .addReg(X86::ESP, RegState::Implicit)
8687 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8688 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8690 MI->eraseFromParent(); // The pseudo instruction is gone now.
8695 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8696 MachineBasicBlock *BB) const {
8697 // This is pretty easy. We're taking the value that we received from
8698 // our load from the relocation, sticking it in either RDI (x86-64)
8699 // or EAX and doing an indirect call. The return value will then
8700 // be in the normal return register.
8701 const X86InstrInfo *TII
8702 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8703 DebugLoc DL = MI->getDebugLoc();
8704 MachineFunction *F = BB->getParent();
8706 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8708 if (Subtarget->is64Bit()) {
8709 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8710 TII->get(X86::MOV64rm), X86::RDI)
8712 .addImm(0).addReg(0)
8713 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8714 MI->getOperand(3).getTargetFlags())
8716 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
8717 addDirectMem(MIB, X86::RDI);
8718 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8719 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8720 TII->get(X86::MOV32rm), X86::EAX)
8722 .addImm(0).addReg(0)
8723 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8724 MI->getOperand(3).getTargetFlags())
8726 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8727 addDirectMem(MIB, X86::EAX);
8729 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8730 TII->get(X86::MOV32rm), X86::EAX)
8731 .addReg(TII->getGlobalBaseReg(F))
8732 .addImm(0).addReg(0)
8733 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8734 MI->getOperand(3).getTargetFlags())
8736 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8737 addDirectMem(MIB, X86::EAX);
8740 MI->eraseFromParent(); // The pseudo instruction is gone now.
8745 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8746 MachineBasicBlock *BB) const {
8747 switch (MI->getOpcode()) {
8748 default: assert(false && "Unexpected instr type to insert");
8749 case X86::MINGW_ALLOCA:
8750 return EmitLoweredMingwAlloca(MI, BB);
8751 case X86::TLSCall_32:
8752 case X86::TLSCall_64:
8753 return EmitLoweredTLSCall(MI, BB);
8755 case X86::CMOV_V1I64:
8756 case X86::CMOV_FR32:
8757 case X86::CMOV_FR64:
8758 case X86::CMOV_V4F32:
8759 case X86::CMOV_V2F64:
8760 case X86::CMOV_V2I64:
8761 case X86::CMOV_GR16:
8762 case X86::CMOV_GR32:
8763 case X86::CMOV_RFP32:
8764 case X86::CMOV_RFP64:
8765 case X86::CMOV_RFP80:
8766 return EmitLoweredSelect(MI, BB);
8768 case X86::FP32_TO_INT16_IN_MEM:
8769 case X86::FP32_TO_INT32_IN_MEM:
8770 case X86::FP32_TO_INT64_IN_MEM:
8771 case X86::FP64_TO_INT16_IN_MEM:
8772 case X86::FP64_TO_INT32_IN_MEM:
8773 case X86::FP64_TO_INT64_IN_MEM:
8774 case X86::FP80_TO_INT16_IN_MEM:
8775 case X86::FP80_TO_INT32_IN_MEM:
8776 case X86::FP80_TO_INT64_IN_MEM: {
8777 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8778 DebugLoc DL = MI->getDebugLoc();
8780 // Change the floating point control register to use "round towards zero"
8781 // mode when truncating to an integer value.
8782 MachineFunction *F = BB->getParent();
8783 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8784 addFrameReference(BuildMI(*BB, MI, DL,
8785 TII->get(X86::FNSTCW16m)), CWFrameIdx);
8787 // Load the old value of the high byte of the control word...
8789 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8790 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
8793 // Set the high part to be round to zero...
8794 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8797 // Reload the modified control word now...
8798 addFrameReference(BuildMI(*BB, MI, DL,
8799 TII->get(X86::FLDCW16m)), CWFrameIdx);
8801 // Restore the memory image of control word to original value
8802 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8805 // Get the X86 opcode to use.
8807 switch (MI->getOpcode()) {
8808 default: llvm_unreachable("illegal opcode!");
8809 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8810 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8811 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8812 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8813 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8814 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8815 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8816 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8817 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8821 MachineOperand &Op = MI->getOperand(0);
8823 AM.BaseType = X86AddressMode::RegBase;
8824 AM.Base.Reg = Op.getReg();
8826 AM.BaseType = X86AddressMode::FrameIndexBase;
8827 AM.Base.FrameIndex = Op.getIndex();
8829 Op = MI->getOperand(1);
8831 AM.Scale = Op.getImm();
8832 Op = MI->getOperand(2);
8834 AM.IndexReg = Op.getImm();
8835 Op = MI->getOperand(3);
8836 if (Op.isGlobal()) {
8837 AM.GV = Op.getGlobal();
8839 AM.Disp = Op.getImm();
8841 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
8842 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
8844 // Reload the original control word now.
8845 addFrameReference(BuildMI(*BB, MI, DL,
8846 TII->get(X86::FLDCW16m)), CWFrameIdx);
8848 MI->eraseFromParent(); // The pseudo instruction is gone now.
8851 // String/text processing lowering.
8852 case X86::PCMPISTRM128REG:
8853 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8854 case X86::PCMPISTRM128MEM:
8855 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8856 case X86::PCMPESTRM128REG:
8857 return EmitPCMP(MI, BB, 5, false /* in mem */);
8858 case X86::PCMPESTRM128MEM:
8859 return EmitPCMP(MI, BB, 5, true /* in mem */);
8862 case X86::ATOMAND32:
8863 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8864 X86::AND32ri, X86::MOV32rm,
8866 X86::NOT32r, X86::EAX,
8867 X86::GR32RegisterClass);
8869 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8870 X86::OR32ri, X86::MOV32rm,
8872 X86::NOT32r, X86::EAX,
8873 X86::GR32RegisterClass);
8874 case X86::ATOMXOR32:
8875 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8876 X86::XOR32ri, X86::MOV32rm,
8878 X86::NOT32r, X86::EAX,
8879 X86::GR32RegisterClass);
8880 case X86::ATOMNAND32:
8881 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8882 X86::AND32ri, X86::MOV32rm,
8884 X86::NOT32r, X86::EAX,
8885 X86::GR32RegisterClass, true);
8886 case X86::ATOMMIN32:
8887 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8888 case X86::ATOMMAX32:
8889 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8890 case X86::ATOMUMIN32:
8891 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8892 case X86::ATOMUMAX32:
8893 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8895 case X86::ATOMAND16:
8896 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8897 X86::AND16ri, X86::MOV16rm,
8899 X86::NOT16r, X86::AX,
8900 X86::GR16RegisterClass);
8902 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8903 X86::OR16ri, X86::MOV16rm,
8905 X86::NOT16r, X86::AX,
8906 X86::GR16RegisterClass);
8907 case X86::ATOMXOR16:
8908 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8909 X86::XOR16ri, X86::MOV16rm,
8911 X86::NOT16r, X86::AX,
8912 X86::GR16RegisterClass);
8913 case X86::ATOMNAND16:
8914 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8915 X86::AND16ri, X86::MOV16rm,
8917 X86::NOT16r, X86::AX,
8918 X86::GR16RegisterClass, true);
8919 case X86::ATOMMIN16:
8920 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8921 case X86::ATOMMAX16:
8922 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8923 case X86::ATOMUMIN16:
8924 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8925 case X86::ATOMUMAX16:
8926 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8929 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8930 X86::AND8ri, X86::MOV8rm,
8932 X86::NOT8r, X86::AL,
8933 X86::GR8RegisterClass);
8935 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8936 X86::OR8ri, X86::MOV8rm,
8938 X86::NOT8r, X86::AL,
8939 X86::GR8RegisterClass);
8941 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8942 X86::XOR8ri, X86::MOV8rm,
8944 X86::NOT8r, X86::AL,
8945 X86::GR8RegisterClass);
8946 case X86::ATOMNAND8:
8947 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8948 X86::AND8ri, X86::MOV8rm,
8950 X86::NOT8r, X86::AL,
8951 X86::GR8RegisterClass, true);
8952 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8953 // This group is for 64-bit host.
8954 case X86::ATOMAND64:
8955 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8956 X86::AND64ri32, X86::MOV64rm,
8958 X86::NOT64r, X86::RAX,
8959 X86::GR64RegisterClass);
8961 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8962 X86::OR64ri32, X86::MOV64rm,
8964 X86::NOT64r, X86::RAX,
8965 X86::GR64RegisterClass);
8966 case X86::ATOMXOR64:
8967 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8968 X86::XOR64ri32, X86::MOV64rm,
8970 X86::NOT64r, X86::RAX,
8971 X86::GR64RegisterClass);
8972 case X86::ATOMNAND64:
8973 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8974 X86::AND64ri32, X86::MOV64rm,
8976 X86::NOT64r, X86::RAX,
8977 X86::GR64RegisterClass, true);
8978 case X86::ATOMMIN64:
8979 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8980 case X86::ATOMMAX64:
8981 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8982 case X86::ATOMUMIN64:
8983 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8984 case X86::ATOMUMAX64:
8985 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8987 // This group does 64-bit operations on a 32-bit host.
8988 case X86::ATOMAND6432:
8989 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8990 X86::AND32rr, X86::AND32rr,
8991 X86::AND32ri, X86::AND32ri,
8993 case X86::ATOMOR6432:
8994 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8995 X86::OR32rr, X86::OR32rr,
8996 X86::OR32ri, X86::OR32ri,
8998 case X86::ATOMXOR6432:
8999 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9000 X86::XOR32rr, X86::XOR32rr,
9001 X86::XOR32ri, X86::XOR32ri,
9003 case X86::ATOMNAND6432:
9004 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9005 X86::AND32rr, X86::AND32rr,
9006 X86::AND32ri, X86::AND32ri,
9008 case X86::ATOMADD6432:
9009 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9010 X86::ADD32rr, X86::ADC32rr,
9011 X86::ADD32ri, X86::ADC32ri,
9013 case X86::ATOMSUB6432:
9014 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9015 X86::SUB32rr, X86::SBB32rr,
9016 X86::SUB32ri, X86::SBB32ri,
9018 case X86::ATOMSWAP6432:
9019 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9020 X86::MOV32rr, X86::MOV32rr,
9021 X86::MOV32ri, X86::MOV32ri,
9023 case X86::VASTART_SAVE_XMM_REGS:
9024 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
9028 //===----------------------------------------------------------------------===//
9029 // X86 Optimization Hooks
9030 //===----------------------------------------------------------------------===//
9032 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9036 const SelectionDAG &DAG,
9037 unsigned Depth) const {
9038 unsigned Opc = Op.getOpcode();
9039 assert((Opc >= ISD::BUILTIN_OP_END ||
9040 Opc == ISD::INTRINSIC_WO_CHAIN ||
9041 Opc == ISD::INTRINSIC_W_CHAIN ||
9042 Opc == ISD::INTRINSIC_VOID) &&
9043 "Should use MaskedValueIsZero if you don't know whether Op"
9044 " is a target node!");
9046 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
9058 // These nodes' second result is a boolean.
9059 if (Op.getResNo() == 0)
9063 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
9064 Mask.getBitWidth() - 1);
9069 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
9070 /// node is a GlobalAddress + offset.
9071 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
9072 const GlobalValue* &GA,
9073 int64_t &Offset) const {
9074 if (N->getOpcode() == X86ISD::Wrapper) {
9075 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
9076 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
9077 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
9081 return TargetLowering::isGAPlusOffset(N, GA, Offset);
9084 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
9085 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
9086 /// if the load addresses are consecutive, non-overlapping, and in the right
9088 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
9089 const TargetLowering &TLI) {
9090 DebugLoc dl = N->getDebugLoc();
9091 EVT VT = N->getValueType(0);
9092 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9094 if (VT.getSizeInBits() != 128)
9097 SmallVector<SDValue, 16> Elts;
9098 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
9099 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
9101 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
9104 /// PerformShuffleCombine - Detect vector gather/scatter index generation
9105 /// and convert it from being a bunch of shuffles and extracts to a simple
9106 /// store and scalar loads to extract the elements.
9107 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
9108 const TargetLowering &TLI) {
9109 SDValue InputVector = N->getOperand(0);
9111 // Only operate on vectors of 4 elements, where the alternative shuffling
9112 // gets to be more expensive.
9113 if (InputVector.getValueType() != MVT::v4i32)
9116 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
9117 // single use which is a sign-extend or zero-extend, and all elements are
9119 SmallVector<SDNode *, 4> Uses;
9120 unsigned ExtractedElements = 0;
9121 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
9122 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
9123 if (UI.getUse().getResNo() != InputVector.getResNo())
9126 SDNode *Extract = *UI;
9127 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9130 if (Extract->getValueType(0) != MVT::i32)
9132 if (!Extract->hasOneUse())
9134 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9135 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9137 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9140 // Record which element was extracted.
9141 ExtractedElements |=
9142 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9144 Uses.push_back(Extract);
9147 // If not all the elements were used, this may not be worthwhile.
9148 if (ExtractedElements != 15)
9151 // Ok, we've now decided to do the transformation.
9152 DebugLoc dl = InputVector.getDebugLoc();
9154 // Store the value to a temporary stack slot.
9155 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9156 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL,
9157 0, false, false, 0);
9159 // Replace each use (extract) with a load of the appropriate element.
9160 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9161 UE = Uses.end(); UI != UE; ++UI) {
9162 SDNode *Extract = *UI;
9164 // Compute the element's address.
9165 SDValue Idx = Extract->getOperand(1);
9167 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9168 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9169 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9171 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
9172 OffsetVal, StackPtr);
9175 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
9176 ScalarAddr, NULL, 0, false, false, 0);
9178 // Replace the exact with the load.
9179 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9182 // The replacement was made in place; don't return anything.
9186 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9187 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9188 const X86Subtarget *Subtarget) {
9189 DebugLoc DL = N->getDebugLoc();
9190 SDValue Cond = N->getOperand(0);
9191 // Get the LHS/RHS of the select.
9192 SDValue LHS = N->getOperand(1);
9193 SDValue RHS = N->getOperand(2);
9195 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9196 // instructions match the semantics of the common C idiom x<y?x:y but not
9197 // x<=y?x:y, because of how they handle negative zero (which can be
9198 // ignored in unsafe-math mode).
9199 if (Subtarget->hasSSE2() &&
9200 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9201 Cond.getOpcode() == ISD::SETCC) {
9202 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9204 unsigned Opcode = 0;
9205 // Check for x CC y ? x : y.
9206 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9207 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9211 // Converting this to a min would handle NaNs incorrectly, and swapping
9212 // the operands would cause it to handle comparisons between positive
9213 // and negative zero incorrectly.
9214 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9215 if (!UnsafeFPMath &&
9216 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9218 std::swap(LHS, RHS);
9220 Opcode = X86ISD::FMIN;
9223 // Converting this to a min would handle comparisons between positive
9224 // and negative zero incorrectly.
9225 if (!UnsafeFPMath &&
9226 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9228 Opcode = X86ISD::FMIN;
9231 // Converting this to a min would handle both negative zeros and NaNs
9232 // incorrectly, but we can swap the operands to fix both.
9233 std::swap(LHS, RHS);
9237 Opcode = X86ISD::FMIN;
9241 // Converting this to a max would handle comparisons between positive
9242 // and negative zero incorrectly.
9243 if (!UnsafeFPMath &&
9244 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9246 Opcode = X86ISD::FMAX;
9249 // Converting this to a max would handle NaNs incorrectly, and swapping
9250 // the operands would cause it to handle comparisons between positive
9251 // and negative zero incorrectly.
9252 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9253 if (!UnsafeFPMath &&
9254 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9256 std::swap(LHS, RHS);
9258 Opcode = X86ISD::FMAX;
9261 // Converting this to a max would handle both negative zeros and NaNs
9262 // incorrectly, but we can swap the operands to fix both.
9263 std::swap(LHS, RHS);
9267 Opcode = X86ISD::FMAX;
9270 // Check for x CC y ? y : x -- a min/max with reversed arms.
9271 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9272 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9276 // Converting this to a min would handle comparisons between positive
9277 // and negative zero incorrectly, and swapping the operands would
9278 // cause it to handle NaNs incorrectly.
9279 if (!UnsafeFPMath &&
9280 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9281 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9283 std::swap(LHS, RHS);
9285 Opcode = X86ISD::FMIN;
9288 // Converting this to a min would handle NaNs incorrectly.
9289 if (!UnsafeFPMath &&
9290 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9292 Opcode = X86ISD::FMIN;
9295 // Converting this to a min would handle both negative zeros and NaNs
9296 // incorrectly, but we can swap the operands to fix both.
9297 std::swap(LHS, RHS);
9301 Opcode = X86ISD::FMIN;
9305 // Converting this to a max would handle NaNs incorrectly.
9306 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9308 Opcode = X86ISD::FMAX;
9311 // Converting this to a max would handle comparisons between positive
9312 // and negative zero incorrectly, and swapping the operands would
9313 // cause it to handle NaNs incorrectly.
9314 if (!UnsafeFPMath &&
9315 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9316 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9318 std::swap(LHS, RHS);
9320 Opcode = X86ISD::FMAX;
9323 // Converting this to a max would handle both negative zeros and NaNs
9324 // incorrectly, but we can swap the operands to fix both.
9325 std::swap(LHS, RHS);
9329 Opcode = X86ISD::FMAX;
9335 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9338 // If this is a select between two integer constants, try to do some
9340 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9341 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9342 // Don't do this for crazy integer types.
9343 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9344 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9345 // so that TrueC (the true value) is larger than FalseC.
9346 bool NeedsCondInvert = false;
9348 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9349 // Efficiently invertible.
9350 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9351 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9352 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9353 NeedsCondInvert = true;
9354 std::swap(TrueC, FalseC);
9357 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9358 if (FalseC->getAPIntValue() == 0 &&
9359 TrueC->getAPIntValue().isPowerOf2()) {
9360 if (NeedsCondInvert) // Invert the condition if needed.
9361 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9362 DAG.getConstant(1, Cond.getValueType()));
9364 // Zero extend the condition if needed.
9365 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9367 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9368 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9369 DAG.getConstant(ShAmt, MVT::i8));
9372 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9373 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9374 if (NeedsCondInvert) // Invert the condition if needed.
9375 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9376 DAG.getConstant(1, Cond.getValueType()));
9378 // Zero extend the condition if needed.
9379 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9380 FalseC->getValueType(0), Cond);
9381 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9382 SDValue(FalseC, 0));
9385 // Optimize cases that will turn into an LEA instruction. This requires
9386 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9387 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9388 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9389 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9391 bool isFastMultiplier = false;
9393 switch ((unsigned char)Diff) {
9395 case 1: // result = add base, cond
9396 case 2: // result = lea base( , cond*2)
9397 case 3: // result = lea base(cond, cond*2)
9398 case 4: // result = lea base( , cond*4)
9399 case 5: // result = lea base(cond, cond*4)
9400 case 8: // result = lea base( , cond*8)
9401 case 9: // result = lea base(cond, cond*8)
9402 isFastMultiplier = true;
9407 if (isFastMultiplier) {
9408 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9409 if (NeedsCondInvert) // Invert the condition if needed.
9410 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9411 DAG.getConstant(1, Cond.getValueType()));
9413 // Zero extend the condition if needed.
9414 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9416 // Scale the condition by the difference.
9418 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9419 DAG.getConstant(Diff, Cond.getValueType()));
9421 // Add the base if non-zero.
9422 if (FalseC->getAPIntValue() != 0)
9423 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9424 SDValue(FalseC, 0));
9434 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9435 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9436 TargetLowering::DAGCombinerInfo &DCI) {
9437 DebugLoc DL = N->getDebugLoc();
9439 // If the flag operand isn't dead, don't touch this CMOV.
9440 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9443 // If this is a select between two integer constants, try to do some
9444 // optimizations. Note that the operands are ordered the opposite of SELECT
9446 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9447 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9448 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9449 // larger than FalseC (the false value).
9450 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9452 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9453 CC = X86::GetOppositeBranchCondition(CC);
9454 std::swap(TrueC, FalseC);
9457 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9458 // This is efficient for any integer data type (including i8/i16) and
9460 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9461 SDValue Cond = N->getOperand(3);
9462 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9463 DAG.getConstant(CC, MVT::i8), Cond);
9465 // Zero extend the condition if needed.
9466 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9468 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9469 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9470 DAG.getConstant(ShAmt, MVT::i8));
9471 if (N->getNumValues() == 2) // Dead flag value?
9472 return DCI.CombineTo(N, Cond, SDValue());
9476 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9477 // for any integer data type, including i8/i16.
9478 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9479 SDValue Cond = N->getOperand(3);
9480 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9481 DAG.getConstant(CC, MVT::i8), Cond);
9483 // Zero extend the condition if needed.
9484 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9485 FalseC->getValueType(0), Cond);
9486 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9487 SDValue(FalseC, 0));
9489 if (N->getNumValues() == 2) // Dead flag value?
9490 return DCI.CombineTo(N, Cond, SDValue());
9494 // Optimize cases that will turn into an LEA instruction. This requires
9495 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9496 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9497 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9498 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9500 bool isFastMultiplier = false;
9502 switch ((unsigned char)Diff) {
9504 case 1: // result = add base, cond
9505 case 2: // result = lea base( , cond*2)
9506 case 3: // result = lea base(cond, cond*2)
9507 case 4: // result = lea base( , cond*4)
9508 case 5: // result = lea base(cond, cond*4)
9509 case 8: // result = lea base( , cond*8)
9510 case 9: // result = lea base(cond, cond*8)
9511 isFastMultiplier = true;
9516 if (isFastMultiplier) {
9517 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9518 SDValue Cond = N->getOperand(3);
9519 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9520 DAG.getConstant(CC, MVT::i8), Cond);
9521 // Zero extend the condition if needed.
9522 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9524 // Scale the condition by the difference.
9526 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9527 DAG.getConstant(Diff, Cond.getValueType()));
9529 // Add the base if non-zero.
9530 if (FalseC->getAPIntValue() != 0)
9531 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9532 SDValue(FalseC, 0));
9533 if (N->getNumValues() == 2) // Dead flag value?
9534 return DCI.CombineTo(N, Cond, SDValue());
9544 /// PerformMulCombine - Optimize a single multiply with constant into two
9545 /// in order to implement it with two cheaper instructions, e.g.
9546 /// LEA + SHL, LEA + LEA.
9547 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9548 TargetLowering::DAGCombinerInfo &DCI) {
9549 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9552 EVT VT = N->getValueType(0);
9556 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9559 uint64_t MulAmt = C->getZExtValue();
9560 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9563 uint64_t MulAmt1 = 0;
9564 uint64_t MulAmt2 = 0;
9565 if ((MulAmt % 9) == 0) {
9567 MulAmt2 = MulAmt / 9;
9568 } else if ((MulAmt % 5) == 0) {
9570 MulAmt2 = MulAmt / 5;
9571 } else if ((MulAmt % 3) == 0) {
9573 MulAmt2 = MulAmt / 3;
9576 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9577 DebugLoc DL = N->getDebugLoc();
9579 if (isPowerOf2_64(MulAmt2) &&
9580 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9581 // If second multiplifer is pow2, issue it first. We want the multiply by
9582 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9584 std::swap(MulAmt1, MulAmt2);
9587 if (isPowerOf2_64(MulAmt1))
9588 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9589 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9591 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9592 DAG.getConstant(MulAmt1, VT));
9594 if (isPowerOf2_64(MulAmt2))
9595 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9596 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9598 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9599 DAG.getConstant(MulAmt2, VT));
9601 // Do not add new nodes to DAG combiner worklist.
9602 DCI.CombineTo(N, NewMul, false);
9607 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9608 SDValue N0 = N->getOperand(0);
9609 SDValue N1 = N->getOperand(1);
9610 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9611 EVT VT = N0.getValueType();
9613 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9614 // since the result of setcc_c is all zero's or all ones.
9615 if (N1C && N0.getOpcode() == ISD::AND &&
9616 N0.getOperand(1).getOpcode() == ISD::Constant) {
9617 SDValue N00 = N0.getOperand(0);
9618 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9619 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9620 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9621 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9622 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9623 APInt ShAmt = N1C->getAPIntValue();
9624 Mask = Mask.shl(ShAmt);
9626 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9627 N00, DAG.getConstant(Mask, VT));
9634 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9636 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9637 const X86Subtarget *Subtarget) {
9638 EVT VT = N->getValueType(0);
9639 if (!VT.isVector() && VT.isInteger() &&
9640 N->getOpcode() == ISD::SHL)
9641 return PerformSHLCombine(N, DAG);
9643 // On X86 with SSE2 support, we can transform this to a vector shift if
9644 // all elements are shifted by the same amount. We can't do this in legalize
9645 // because the a constant vector is typically transformed to a constant pool
9646 // so we have no knowledge of the shift amount.
9647 if (!Subtarget->hasSSE2())
9650 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9653 SDValue ShAmtOp = N->getOperand(1);
9654 EVT EltVT = VT.getVectorElementType();
9655 DebugLoc DL = N->getDebugLoc();
9656 SDValue BaseShAmt = SDValue();
9657 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9658 unsigned NumElts = VT.getVectorNumElements();
9660 for (; i != NumElts; ++i) {
9661 SDValue Arg = ShAmtOp.getOperand(i);
9662 if (Arg.getOpcode() == ISD::UNDEF) continue;
9666 for (; i != NumElts; ++i) {
9667 SDValue Arg = ShAmtOp.getOperand(i);
9668 if (Arg.getOpcode() == ISD::UNDEF) continue;
9669 if (Arg != BaseShAmt) {
9673 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9674 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9675 SDValue InVec = ShAmtOp.getOperand(0);
9676 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9677 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9679 for (; i != NumElts; ++i) {
9680 SDValue Arg = InVec.getOperand(i);
9681 if (Arg.getOpcode() == ISD::UNDEF) continue;
9685 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9686 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9687 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9688 if (C->getZExtValue() == SplatIdx)
9689 BaseShAmt = InVec.getOperand(1);
9692 if (BaseShAmt.getNode() == 0)
9693 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9694 DAG.getIntPtrConstant(0));
9698 // The shift amount is an i32.
9699 if (EltVT.bitsGT(MVT::i32))
9700 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9701 else if (EltVT.bitsLT(MVT::i32))
9702 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9704 // The shift amount is identical so we can do a vector shift.
9705 SDValue ValOp = N->getOperand(0);
9706 switch (N->getOpcode()) {
9708 llvm_unreachable("Unknown shift opcode!");
9711 if (VT == MVT::v2i64)
9712 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9713 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9715 if (VT == MVT::v4i32)
9716 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9717 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9719 if (VT == MVT::v8i16)
9720 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9721 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9725 if (VT == MVT::v4i32)
9726 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9727 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9729 if (VT == MVT::v8i16)
9730 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9731 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9735 if (VT == MVT::v2i64)
9736 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9737 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9739 if (VT == MVT::v4i32)
9740 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9741 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9743 if (VT == MVT::v8i16)
9744 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9745 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9752 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9753 TargetLowering::DAGCombinerInfo &DCI,
9754 const X86Subtarget *Subtarget) {
9755 if (DCI.isBeforeLegalizeOps())
9758 EVT VT = N->getValueType(0);
9759 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9762 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9763 SDValue N0 = N->getOperand(0);
9764 SDValue N1 = N->getOperand(1);
9765 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9767 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9769 if (!N0.hasOneUse() || !N1.hasOneUse())
9772 SDValue ShAmt0 = N0.getOperand(1);
9773 if (ShAmt0.getValueType() != MVT::i8)
9775 SDValue ShAmt1 = N1.getOperand(1);
9776 if (ShAmt1.getValueType() != MVT::i8)
9778 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9779 ShAmt0 = ShAmt0.getOperand(0);
9780 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9781 ShAmt1 = ShAmt1.getOperand(0);
9783 DebugLoc DL = N->getDebugLoc();
9784 unsigned Opc = X86ISD::SHLD;
9785 SDValue Op0 = N0.getOperand(0);
9786 SDValue Op1 = N1.getOperand(0);
9787 if (ShAmt0.getOpcode() == ISD::SUB) {
9789 std::swap(Op0, Op1);
9790 std::swap(ShAmt0, ShAmt1);
9793 unsigned Bits = VT.getSizeInBits();
9794 if (ShAmt1.getOpcode() == ISD::SUB) {
9795 SDValue Sum = ShAmt1.getOperand(0);
9796 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9797 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
9798 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
9799 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
9800 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
9801 return DAG.getNode(Opc, DL, VT,
9803 DAG.getNode(ISD::TRUNCATE, DL,
9806 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9807 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9809 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9810 return DAG.getNode(Opc, DL, VT,
9811 N0.getOperand(0), N1.getOperand(0),
9812 DAG.getNode(ISD::TRUNCATE, DL,
9819 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9820 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9821 const X86Subtarget *Subtarget) {
9822 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9823 // the FP state in cases where an emms may be missing.
9824 // A preferable solution to the general problem is to figure out the right
9825 // places to insert EMMS. This qualifies as a quick hack.
9827 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9828 StoreSDNode *St = cast<StoreSDNode>(N);
9829 EVT VT = St->getValue().getValueType();
9830 if (VT.getSizeInBits() != 64)
9833 const Function *F = DAG.getMachineFunction().getFunction();
9834 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9835 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9836 && Subtarget->hasSSE2();
9837 if ((VT.isVector() ||
9838 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9839 isa<LoadSDNode>(St->getValue()) &&
9840 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9841 St->getChain().hasOneUse() && !St->isVolatile()) {
9842 SDNode* LdVal = St->getValue().getNode();
9844 int TokenFactorIndex = -1;
9845 SmallVector<SDValue, 8> Ops;
9846 SDNode* ChainVal = St->getChain().getNode();
9847 // Must be a store of a load. We currently handle two cases: the load
9848 // is a direct child, and it's under an intervening TokenFactor. It is
9849 // possible to dig deeper under nested TokenFactors.
9850 if (ChainVal == LdVal)
9851 Ld = cast<LoadSDNode>(St->getChain());
9852 else if (St->getValue().hasOneUse() &&
9853 ChainVal->getOpcode() == ISD::TokenFactor) {
9854 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9855 if (ChainVal->getOperand(i).getNode() == LdVal) {
9856 TokenFactorIndex = i;
9857 Ld = cast<LoadSDNode>(St->getValue());
9859 Ops.push_back(ChainVal->getOperand(i));
9863 if (!Ld || !ISD::isNormalLoad(Ld))
9866 // If this is not the MMX case, i.e. we are just turning i64 load/store
9867 // into f64 load/store, avoid the transformation if there are multiple
9868 // uses of the loaded value.
9869 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9872 DebugLoc LdDL = Ld->getDebugLoc();
9873 DebugLoc StDL = N->getDebugLoc();
9874 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9875 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9877 if (Subtarget->is64Bit() || F64IsLegal) {
9878 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9879 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9880 Ld->getBasePtr(), Ld->getSrcValue(),
9881 Ld->getSrcValueOffset(), Ld->isVolatile(),
9882 Ld->isNonTemporal(), Ld->getAlignment());
9883 SDValue NewChain = NewLd.getValue(1);
9884 if (TokenFactorIndex != -1) {
9885 Ops.push_back(NewChain);
9886 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9889 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9890 St->getSrcValue(), St->getSrcValueOffset(),
9891 St->isVolatile(), St->isNonTemporal(),
9892 St->getAlignment());
9895 // Otherwise, lower to two pairs of 32-bit loads / stores.
9896 SDValue LoAddr = Ld->getBasePtr();
9897 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9898 DAG.getConstant(4, MVT::i32));
9900 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9901 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9902 Ld->isVolatile(), Ld->isNonTemporal(),
9903 Ld->getAlignment());
9904 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9905 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9906 Ld->isVolatile(), Ld->isNonTemporal(),
9907 MinAlign(Ld->getAlignment(), 4));
9909 SDValue NewChain = LoLd.getValue(1);
9910 if (TokenFactorIndex != -1) {
9911 Ops.push_back(LoLd);
9912 Ops.push_back(HiLd);
9913 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9917 LoAddr = St->getBasePtr();
9918 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9919 DAG.getConstant(4, MVT::i32));
9921 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9922 St->getSrcValue(), St->getSrcValueOffset(),
9923 St->isVolatile(), St->isNonTemporal(),
9924 St->getAlignment());
9925 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9927 St->getSrcValueOffset() + 4,
9929 St->isNonTemporal(),
9930 MinAlign(St->getAlignment(), 4));
9931 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9936 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9937 /// X86ISD::FXOR nodes.
9938 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9939 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9940 // F[X]OR(0.0, x) -> x
9941 // F[X]OR(x, 0.0) -> x
9942 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9943 if (C->getValueAPF().isPosZero())
9944 return N->getOperand(1);
9945 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9946 if (C->getValueAPF().isPosZero())
9947 return N->getOperand(0);
9951 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9952 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9953 // FAND(0.0, x) -> 0.0
9954 // FAND(x, 0.0) -> 0.0
9955 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9956 if (C->getValueAPF().isPosZero())
9957 return N->getOperand(0);
9958 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9959 if (C->getValueAPF().isPosZero())
9960 return N->getOperand(1);
9964 static SDValue PerformBTCombine(SDNode *N,
9966 TargetLowering::DAGCombinerInfo &DCI) {
9967 // BT ignores high bits in the bit index operand.
9968 SDValue Op1 = N->getOperand(1);
9969 if (Op1.hasOneUse()) {
9970 unsigned BitWidth = Op1.getValueSizeInBits();
9971 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9972 APInt KnownZero, KnownOne;
9973 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9974 !DCI.isBeforeLegalizeOps());
9975 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9976 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9977 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9978 DCI.CommitTargetLoweringOpt(TLO);
9983 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9984 SDValue Op = N->getOperand(0);
9985 if (Op.getOpcode() == ISD::BIT_CONVERT)
9986 Op = Op.getOperand(0);
9987 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9988 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9989 VT.getVectorElementType().getSizeInBits() ==
9990 OpVT.getVectorElementType().getSizeInBits()) {
9991 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9996 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9997 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9998 // (and (i32 x86isd::setcc_carry), 1)
9999 // This eliminates the zext. This transformation is necessary because
10000 // ISD::SETCC is always legalized to i8.
10001 DebugLoc dl = N->getDebugLoc();
10002 SDValue N0 = N->getOperand(0);
10003 EVT VT = N->getValueType(0);
10004 if (N0.getOpcode() == ISD::AND &&
10006 N0.getOperand(0).hasOneUse()) {
10007 SDValue N00 = N0.getOperand(0);
10008 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
10010 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
10011 if (!C || C->getZExtValue() != 1)
10013 return DAG.getNode(ISD::AND, dl, VT,
10014 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
10015 N00.getOperand(0), N00.getOperand(1)),
10016 DAG.getConstant(1, VT));
10022 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
10023 DAGCombinerInfo &DCI) const {
10024 SelectionDAG &DAG = DCI.DAG;
10025 switch (N->getOpcode()) {
10027 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
10028 case ISD::EXTRACT_VECTOR_ELT:
10029 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
10030 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
10031 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
10032 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
10035 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
10036 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
10037 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
10039 case X86ISD::FOR: return PerformFORCombine(N, DAG);
10040 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
10041 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
10042 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
10043 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
10049 /// isTypeDesirableForOp - Return true if the target has native support for
10050 /// the specified value type and it is 'desirable' to use the type for the
10051 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
10052 /// instruction encodings are longer and some i16 instructions are slow.
10053 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
10054 if (!isTypeLegal(VT))
10056 if (VT != MVT::i16)
10063 case ISD::SIGN_EXTEND:
10064 case ISD::ZERO_EXTEND:
10065 case ISD::ANY_EXTEND:
10078 static bool MayFoldLoad(SDValue Op) {
10079 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
10082 static bool MayFoldIntoStore(SDValue Op) {
10083 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
10086 /// IsDesirableToPromoteOp - This method query the target whether it is
10087 /// beneficial for dag combiner to promote the specified node. If true, it
10088 /// should return the desired promotion type by reference.
10089 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
10090 EVT VT = Op.getValueType();
10091 if (VT != MVT::i16)
10094 bool Promote = false;
10095 bool Commute = false;
10096 switch (Op.getOpcode()) {
10099 LoadSDNode *LD = cast<LoadSDNode>(Op);
10100 // If the non-extending load has a single use and it's not live out, then it
10101 // might be folded.
10102 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
10103 Op.hasOneUse()*/) {
10104 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
10105 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
10106 // The only case where we'd want to promote LOAD (rather then it being
10107 // promoted as an operand is when it's only use is liveout.
10108 if (UI->getOpcode() != ISD::CopyToReg)
10115 case ISD::SIGN_EXTEND:
10116 case ISD::ZERO_EXTEND:
10117 case ISD::ANY_EXTEND:
10122 SDValue N0 = Op.getOperand(0);
10123 // Look out for (store (shl (load), x)).
10124 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10137 SDValue N0 = Op.getOperand(0);
10138 SDValue N1 = Op.getOperand(1);
10139 if (!Commute && MayFoldLoad(N1))
10141 // Avoid disabling potential load folding opportunities.
10142 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10144 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10154 //===----------------------------------------------------------------------===//
10155 // X86 Inline Assembly Support
10156 //===----------------------------------------------------------------------===//
10158 static bool LowerToBSwap(CallInst *CI) {
10159 // FIXME: this should verify that we are targetting a 486 or better. If not,
10160 // we will turn this bswap into something that will be lowered to logical ops
10161 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10162 // so don't worry about this.
10164 // Verify this is a simple bswap.
10165 if (CI->getNumArgOperands() != 1 ||
10166 CI->getType() != CI->getArgOperand(0)->getType() ||
10167 !CI->getType()->isIntegerTy())
10170 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10171 if (!Ty || Ty->getBitWidth() % 16 != 0)
10174 // Okay, we can do this xform, do so now.
10175 const Type *Tys[] = { Ty };
10176 Module *M = CI->getParent()->getParent()->getParent();
10177 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10179 Value *Op = CI->getArgOperand(0);
10180 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10182 CI->replaceAllUsesWith(Op);
10183 CI->eraseFromParent();
10187 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10188 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10189 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10191 std::string AsmStr = IA->getAsmString();
10193 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10194 SmallVector<StringRef, 4> AsmPieces;
10195 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10197 switch (AsmPieces.size()) {
10198 default: return false;
10200 AsmStr = AsmPieces[0];
10202 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10205 if (AsmPieces.size() == 2 &&
10206 (AsmPieces[0] == "bswap" ||
10207 AsmPieces[0] == "bswapq" ||
10208 AsmPieces[0] == "bswapl") &&
10209 (AsmPieces[1] == "$0" ||
10210 AsmPieces[1] == "${0:q}")) {
10211 // No need to check constraints, nothing other than the equivalent of
10212 // "=r,0" would be valid here.
10213 return LowerToBSwap(CI);
10215 // rorw $$8, ${0:w} --> llvm.bswap.i16
10216 if (CI->getType()->isIntegerTy(16) &&
10217 AsmPieces.size() == 3 &&
10218 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10219 AsmPieces[1] == "$$8," &&
10220 AsmPieces[2] == "${0:w}" &&
10221 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10223 const std::string &Constraints = IA->getConstraintString();
10224 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10225 std::sort(AsmPieces.begin(), AsmPieces.end());
10226 if (AsmPieces.size() == 4 &&
10227 AsmPieces[0] == "~{cc}" &&
10228 AsmPieces[1] == "~{dirflag}" &&
10229 AsmPieces[2] == "~{flags}" &&
10230 AsmPieces[3] == "~{fpsr}") {
10231 return LowerToBSwap(CI);
10236 if (CI->getType()->isIntegerTy(64) &&
10237 Constraints.size() >= 2 &&
10238 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10239 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10240 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10241 SmallVector<StringRef, 4> Words;
10242 SplitString(AsmPieces[0], Words, " \t");
10243 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10245 SplitString(AsmPieces[1], Words, " \t");
10246 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10248 SplitString(AsmPieces[2], Words, " \t,");
10249 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10250 Words[2] == "%edx") {
10251 return LowerToBSwap(CI);
10263 /// getConstraintType - Given a constraint letter, return the type of
10264 /// constraint it is for this target.
10265 X86TargetLowering::ConstraintType
10266 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10267 if (Constraint.size() == 1) {
10268 switch (Constraint[0]) {
10280 return C_RegisterClass;
10288 return TargetLowering::getConstraintType(Constraint);
10291 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10292 /// with another that has more specific requirements based on the type of the
10293 /// corresponding operand.
10294 const char *X86TargetLowering::
10295 LowerXConstraint(EVT ConstraintVT) const {
10296 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10297 // 'f' like normal targets.
10298 if (ConstraintVT.isFloatingPoint()) {
10299 if (Subtarget->hasSSE2())
10301 if (Subtarget->hasSSE1())
10305 return TargetLowering::LowerXConstraint(ConstraintVT);
10308 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10309 /// vector. If it is invalid, don't add anything to Ops.
10310 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10312 std::vector<SDValue>&Ops,
10313 SelectionDAG &DAG) const {
10314 SDValue Result(0, 0);
10316 switch (Constraint) {
10319 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10320 if (C->getZExtValue() <= 31) {
10321 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10327 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10328 if (C->getZExtValue() <= 63) {
10329 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10335 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10336 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10337 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10343 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10344 if (C->getZExtValue() <= 255) {
10345 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10351 // 32-bit signed value
10352 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10353 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10354 C->getSExtValue())) {
10355 // Widen to 64 bits here to get it sign extended.
10356 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10359 // FIXME gcc accepts some relocatable values here too, but only in certain
10360 // memory models; it's complicated.
10365 // 32-bit unsigned value
10366 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10367 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10368 C->getZExtValue())) {
10369 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10373 // FIXME gcc accepts some relocatable values here too, but only in certain
10374 // memory models; it's complicated.
10378 // Literal immediates are always ok.
10379 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10380 // Widen to 64 bits here to get it sign extended.
10381 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10385 // In any sort of PIC mode addresses need to be computed at runtime by
10386 // adding in a register or some sort of table lookup. These can't
10387 // be used as immediates.
10388 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
10391 // If we are in non-pic codegen mode, we allow the address of a global (with
10392 // an optional displacement) to be used with 'i'.
10393 GlobalAddressSDNode *GA = 0;
10394 int64_t Offset = 0;
10396 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10398 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10399 Offset += GA->getOffset();
10401 } else if (Op.getOpcode() == ISD::ADD) {
10402 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10403 Offset += C->getZExtValue();
10404 Op = Op.getOperand(0);
10407 } else if (Op.getOpcode() == ISD::SUB) {
10408 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10409 Offset += -C->getZExtValue();
10410 Op = Op.getOperand(0);
10415 // Otherwise, this isn't something we can handle, reject it.
10419 const GlobalValue *GV = GA->getGlobal();
10420 // If we require an extra load to get this address, as in PIC mode, we
10421 // can't accept it.
10422 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10423 getTargetMachine())))
10426 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
10427 GA->getValueType(0), Offset);
10432 if (Result.getNode()) {
10433 Ops.push_back(Result);
10436 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10439 std::vector<unsigned> X86TargetLowering::
10440 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10442 if (Constraint.size() == 1) {
10443 // FIXME: not handling fp-stack yet!
10444 switch (Constraint[0]) { // GCC X86 Constraint Letters
10445 default: break; // Unknown constraint letter
10446 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10447 if (Subtarget->is64Bit()) {
10448 if (VT == MVT::i32)
10449 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10450 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10451 X86::R10D,X86::R11D,X86::R12D,
10452 X86::R13D,X86::R14D,X86::R15D,
10453 X86::EBP, X86::ESP, 0);
10454 else if (VT == MVT::i16)
10455 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10456 X86::SI, X86::DI, X86::R8W,X86::R9W,
10457 X86::R10W,X86::R11W,X86::R12W,
10458 X86::R13W,X86::R14W,X86::R15W,
10459 X86::BP, X86::SP, 0);
10460 else if (VT == MVT::i8)
10461 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10462 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10463 X86::R10B,X86::R11B,X86::R12B,
10464 X86::R13B,X86::R14B,X86::R15B,
10465 X86::BPL, X86::SPL, 0);
10467 else if (VT == MVT::i64)
10468 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10469 X86::RSI, X86::RDI, X86::R8, X86::R9,
10470 X86::R10, X86::R11, X86::R12,
10471 X86::R13, X86::R14, X86::R15,
10472 X86::RBP, X86::RSP, 0);
10476 // 32-bit fallthrough
10477 case 'Q': // Q_REGS
10478 if (VT == MVT::i32)
10479 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10480 else if (VT == MVT::i16)
10481 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10482 else if (VT == MVT::i8)
10483 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10484 else if (VT == MVT::i64)
10485 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10490 return std::vector<unsigned>();
10493 std::pair<unsigned, const TargetRegisterClass*>
10494 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10496 // First, see if this is a constraint that directly corresponds to an LLVM
10498 if (Constraint.size() == 1) {
10499 // GCC Constraint Letters
10500 switch (Constraint[0]) {
10502 case 'r': // GENERAL_REGS
10503 case 'l': // INDEX_REGS
10505 return std::make_pair(0U, X86::GR8RegisterClass);
10506 if (VT == MVT::i16)
10507 return std::make_pair(0U, X86::GR16RegisterClass);
10508 if (VT == MVT::i32 || !Subtarget->is64Bit())
10509 return std::make_pair(0U, X86::GR32RegisterClass);
10510 return std::make_pair(0U, X86::GR64RegisterClass);
10511 case 'R': // LEGACY_REGS
10513 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10514 if (VT == MVT::i16)
10515 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10516 if (VT == MVT::i32 || !Subtarget->is64Bit())
10517 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10518 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10519 case 'f': // FP Stack registers.
10520 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10521 // value to the correct fpstack register class.
10522 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10523 return std::make_pair(0U, X86::RFP32RegisterClass);
10524 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10525 return std::make_pair(0U, X86::RFP64RegisterClass);
10526 return std::make_pair(0U, X86::RFP80RegisterClass);
10527 case 'y': // MMX_REGS if MMX allowed.
10528 if (!Subtarget->hasMMX()) break;
10529 return std::make_pair(0U, X86::VR64RegisterClass);
10530 case 'Y': // SSE_REGS if SSE2 allowed
10531 if (!Subtarget->hasSSE2()) break;
10533 case 'x': // SSE_REGS if SSE1 allowed
10534 if (!Subtarget->hasSSE1()) break;
10536 switch (VT.getSimpleVT().SimpleTy) {
10538 // Scalar SSE types.
10541 return std::make_pair(0U, X86::FR32RegisterClass);
10544 return std::make_pair(0U, X86::FR64RegisterClass);
10552 return std::make_pair(0U, X86::VR128RegisterClass);
10558 // Use the default implementation in TargetLowering to convert the register
10559 // constraint into a member of a register class.
10560 std::pair<unsigned, const TargetRegisterClass*> Res;
10561 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10563 // Not found as a standard register?
10564 if (Res.second == 0) {
10565 // Map st(0) -> st(7) -> ST0
10566 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10567 tolower(Constraint[1]) == 's' &&
10568 tolower(Constraint[2]) == 't' &&
10569 Constraint[3] == '(' &&
10570 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10571 Constraint[5] == ')' &&
10572 Constraint[6] == '}') {
10574 Res.first = X86::ST0+Constraint[4]-'0';
10575 Res.second = X86::RFP80RegisterClass;
10579 // GCC allows "st(0)" to be called just plain "st".
10580 if (StringRef("{st}").equals_lower(Constraint)) {
10581 Res.first = X86::ST0;
10582 Res.second = X86::RFP80RegisterClass;
10587 if (StringRef("{flags}").equals_lower(Constraint)) {
10588 Res.first = X86::EFLAGS;
10589 Res.second = X86::CCRRegisterClass;
10593 // 'A' means EAX + EDX.
10594 if (Constraint == "A") {
10595 Res.first = X86::EAX;
10596 Res.second = X86::GR32_ADRegisterClass;
10602 // Otherwise, check to see if this is a register class of the wrong value
10603 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10604 // turn into {ax},{dx}.
10605 if (Res.second->hasType(VT))
10606 return Res; // Correct type already, nothing to do.
10608 // All of the single-register GCC register classes map their values onto
10609 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10610 // really want an 8-bit or 32-bit register, map to the appropriate register
10611 // class and return the appropriate register.
10612 if (Res.second == X86::GR16RegisterClass) {
10613 if (VT == MVT::i8) {
10614 unsigned DestReg = 0;
10615 switch (Res.first) {
10617 case X86::AX: DestReg = X86::AL; break;
10618 case X86::DX: DestReg = X86::DL; break;
10619 case X86::CX: DestReg = X86::CL; break;
10620 case X86::BX: DestReg = X86::BL; break;
10623 Res.first = DestReg;
10624 Res.second = X86::GR8RegisterClass;
10626 } else if (VT == MVT::i32) {
10627 unsigned DestReg = 0;
10628 switch (Res.first) {
10630 case X86::AX: DestReg = X86::EAX; break;
10631 case X86::DX: DestReg = X86::EDX; break;
10632 case X86::CX: DestReg = X86::ECX; break;
10633 case X86::BX: DestReg = X86::EBX; break;
10634 case X86::SI: DestReg = X86::ESI; break;
10635 case X86::DI: DestReg = X86::EDI; break;
10636 case X86::BP: DestReg = X86::EBP; break;
10637 case X86::SP: DestReg = X86::ESP; break;
10640 Res.first = DestReg;
10641 Res.second = X86::GR32RegisterClass;
10643 } else if (VT == MVT::i64) {
10644 unsigned DestReg = 0;
10645 switch (Res.first) {
10647 case X86::AX: DestReg = X86::RAX; break;
10648 case X86::DX: DestReg = X86::RDX; break;
10649 case X86::CX: DestReg = X86::RCX; break;
10650 case X86::BX: DestReg = X86::RBX; break;
10651 case X86::SI: DestReg = X86::RSI; break;
10652 case X86::DI: DestReg = X86::RDI; break;
10653 case X86::BP: DestReg = X86::RBP; break;
10654 case X86::SP: DestReg = X86::RSP; break;
10657 Res.first = DestReg;
10658 Res.second = X86::GR64RegisterClass;
10661 } else if (Res.second == X86::FR32RegisterClass ||
10662 Res.second == X86::FR64RegisterClass ||
10663 Res.second == X86::VR128RegisterClass) {
10664 // Handle references to XMM physical registers that got mapped into the
10665 // wrong class. This can happen with constraints like {xmm0} where the
10666 // target independent register mapper will just pick the first match it can
10667 // find, ignoring the required type.
10668 if (VT == MVT::f32)
10669 Res.second = X86::FR32RegisterClass;
10670 else if (VT == MVT::f64)
10671 Res.second = X86::FR64RegisterClass;
10672 else if (X86::VR128RegisterClass->hasType(VT))
10673 Res.second = X86::VR128RegisterClass;