1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalAlias.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/LLVMContext.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/Dwarf.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/raw_ostream.h"
53 using namespace dwarf;
55 STATISTIC(NumTailCalls, "Number of tail calls");
58 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
60 // Forward declarations.
61 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
64 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
66 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
68 if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
69 if (is64Bit) return new X8664_MachoTargetObjectFile();
70 return new TargetLoweringObjectFileMachO();
71 } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
72 if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
73 return new X8632_ELFTargetObjectFile(TM);
74 } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
75 return new TargetLoweringObjectFileCOFF();
77 llvm_unreachable("unknown subtarget type");
80 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
81 : TargetLowering(TM, createTLOF(TM)) {
82 Subtarget = &TM.getSubtarget<X86Subtarget>();
83 X86ScalarSSEf64 = Subtarget->hasSSE2();
84 X86ScalarSSEf32 = Subtarget->hasSSE1();
85 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
87 RegInfo = TM.getRegisterInfo();
90 // Set up the TargetLowering object.
92 // X86 is weird, it always uses i8 for shift amounts and setcc results.
93 setShiftAmountType(MVT::i8);
94 setBooleanContents(ZeroOrOneBooleanContent);
95 setSchedulingPreference(Sched::RegPressure);
96 setStackPointerRegisterToSaveRestore(X86StackPtr);
98 if (Subtarget->isTargetDarwin()) {
99 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
100 setUseUnderscoreSetJmp(false);
101 setUseUnderscoreLongJmp(false);
102 } else if (Subtarget->isTargetMingw()) {
103 // MS runtime is weird: it exports _setjmp, but longjmp!
104 setUseUnderscoreSetJmp(true);
105 setUseUnderscoreLongJmp(false);
107 setUseUnderscoreSetJmp(true);
108 setUseUnderscoreLongJmp(true);
111 // Set up the register classes.
112 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
113 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
114 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
115 if (Subtarget->is64Bit())
116 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
118 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
120 // We don't accept any truncstore of integer registers.
121 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
122 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
123 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
124 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
125 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
128 // SETOEQ and SETUNE require checking two conditions.
129 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
130 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
131 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
132 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
133 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
136 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
138 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
139 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
140 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
142 if (Subtarget->is64Bit()) {
143 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
144 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
145 } else if (!UseSoftFloat) {
146 // We have an algorithm for SSE2->double, and we turn this into a
147 // 64-bit FILD followed by conditional FADD for other targets.
148 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
149 // We have an algorithm for SSE2, and we turn this into a 64-bit
150 // FILD for other targets.
151 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
154 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
156 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
157 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
160 // SSE has no i16 to fp conversion, only i32
161 if (X86ScalarSSEf32) {
162 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
163 // f32 and f64 cases are Legal, f80 case is not
164 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
166 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
167 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
170 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
171 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
174 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
175 // are Legal, f80 is custom lowered.
176 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
179 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
181 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
182 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
184 if (X86ScalarSSEf32) {
185 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
186 // f32 and f64 cases are Legal, f80 case is not
187 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
189 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
190 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
193 // Handle FP_TO_UINT by promoting the destination to a larger signed
195 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
196 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
197 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
199 if (Subtarget->is64Bit()) {
200 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
201 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
202 } else if (!UseSoftFloat) {
203 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
204 // Expand FP_TO_UINT into a select.
205 // FIXME: We would like to use a Custom expander here eventually to do
206 // the optimal thing for SSE vs. the default expansion in the legalizer.
207 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
209 // With SSE3 we can use fisttpll to convert to a signed i64; without
210 // SSE, we're stuck with a fistpll.
211 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
214 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
215 if (!X86ScalarSSEf64) {
216 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
217 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
218 if (Subtarget->is64Bit()) {
219 setOperationAction(ISD::BIT_CONVERT , MVT::f64 , Expand);
220 // Without SSE, i64->f64 goes through memory; i64->MMX is Legal.
221 if (Subtarget->hasMMX() && !DisableMMX)
222 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Custom);
224 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Expand);
228 // Scalar integer divide and remainder are lowered to use operations that
229 // produce two results, to match the available instructions. This exposes
230 // the two-result form to trivial CSE, which is able to combine x/y and x%y
231 // into a single instruction.
233 // Scalar integer multiply-high is also lowered to use two-result
234 // operations, to match the available instructions. However, plain multiply
235 // (low) operations are left as Legal, as there are single-result
236 // instructions for this in x86. Using the two-result multiply instructions
237 // when both high and low results are needed must be arranged by dagcombine.
238 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
239 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
240 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
241 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
242 setOperationAction(ISD::SREM , MVT::i8 , Expand);
243 setOperationAction(ISD::UREM , MVT::i8 , Expand);
244 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
245 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
246 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
247 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
248 setOperationAction(ISD::SREM , MVT::i16 , Expand);
249 setOperationAction(ISD::UREM , MVT::i16 , Expand);
250 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
251 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
252 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
253 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
254 setOperationAction(ISD::SREM , MVT::i32 , Expand);
255 setOperationAction(ISD::UREM , MVT::i32 , Expand);
256 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
257 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
258 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
259 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
260 setOperationAction(ISD::SREM , MVT::i64 , Expand);
261 setOperationAction(ISD::UREM , MVT::i64 , Expand);
263 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
264 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
265 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
266 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
267 if (Subtarget->is64Bit())
268 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
269 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
270 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
272 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
273 setOperationAction(ISD::FREM , MVT::f32 , Expand);
274 setOperationAction(ISD::FREM , MVT::f64 , Expand);
275 setOperationAction(ISD::FREM , MVT::f80 , Expand);
276 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
278 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
279 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
280 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
281 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
282 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
283 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
287 if (Subtarget->is64Bit()) {
288 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
293 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
294 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
296 // These should be promoted to a larger select which is supported.
297 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
298 // X86 wants to expand cmov itself.
299 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
300 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
301 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
302 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
303 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
304 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
305 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
306 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
307 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
308 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
309 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
310 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
311 if (Subtarget->is64Bit()) {
312 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
313 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
315 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
318 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
319 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
320 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
321 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
322 if (Subtarget->is64Bit())
323 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
324 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
325 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
326 if (Subtarget->is64Bit()) {
327 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
328 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
329 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
330 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
331 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
333 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
334 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
335 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
336 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
337 if (Subtarget->is64Bit()) {
338 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
339 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
340 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
343 if (Subtarget->hasSSE1())
344 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
346 // We may not have a libcall for MEMBARRIER so we should lower this.
347 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
349 // On X86 and X86-64, atomic operations are lowered to locked instructions.
350 // Locked instructions, in turn, have implicit fence semantics (all memory
351 // operations are flushed before issuing the locked instruction, and they
352 // are not buffered), so we can fold away the common pattern of
353 // fence-atomic-fence.
354 setShouldFoldAtomicFences(true);
356 // Expand certain atomics
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
360 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
365 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
367 if (!Subtarget->is64Bit()) {
368 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
374 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
377 // FIXME - use subtarget debug flags
378 if (!Subtarget->isTargetDarwin() &&
379 !Subtarget->isTargetELF() &&
380 !Subtarget->isTargetCygMing()) {
381 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
384 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
385 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
386 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
387 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
388 if (Subtarget->is64Bit()) {
389 setExceptionPointerRegister(X86::RAX);
390 setExceptionSelectorRegister(X86::RDX);
392 setExceptionPointerRegister(X86::EAX);
393 setExceptionSelectorRegister(X86::EDX);
395 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
396 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
398 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
400 setOperationAction(ISD::TRAP, MVT::Other, Legal);
402 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
403 setOperationAction(ISD::VASTART , MVT::Other, Custom);
404 setOperationAction(ISD::VAEND , MVT::Other, Expand);
405 if (Subtarget->is64Bit()) {
406 setOperationAction(ISD::VAARG , MVT::Other, Custom);
407 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
409 setOperationAction(ISD::VAARG , MVT::Other, Expand);
410 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
413 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
414 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
415 if (Subtarget->is64Bit())
416 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
417 if (Subtarget->isTargetCygMing())
418 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
420 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
422 if (!UseSoftFloat && X86ScalarSSEf64) {
423 // f32 and f64 use SSE.
424 // Set up the FP register classes.
425 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
426 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
428 // Use ANDPD to simulate FABS.
429 setOperationAction(ISD::FABS , MVT::f64, Custom);
430 setOperationAction(ISD::FABS , MVT::f32, Custom);
432 // Use XORP to simulate FNEG.
433 setOperationAction(ISD::FNEG , MVT::f64, Custom);
434 setOperationAction(ISD::FNEG , MVT::f32, Custom);
436 // Use ANDPD and ORPD to simulate FCOPYSIGN.
437 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
438 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
440 // We don't support sin/cos/fmod
441 setOperationAction(ISD::FSIN , MVT::f64, Expand);
442 setOperationAction(ISD::FCOS , MVT::f64, Expand);
443 setOperationAction(ISD::FSIN , MVT::f32, Expand);
444 setOperationAction(ISD::FCOS , MVT::f32, Expand);
446 // Expand FP immediates into loads from the stack, except for the special
448 addLegalFPImmediate(APFloat(+0.0)); // xorpd
449 addLegalFPImmediate(APFloat(+0.0f)); // xorps
450 } else if (!UseSoftFloat && X86ScalarSSEf32) {
451 // Use SSE for f32, x87 for f64.
452 // Set up the FP register classes.
453 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
454 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
456 // Use ANDPS to simulate FABS.
457 setOperationAction(ISD::FABS , MVT::f32, Custom);
459 // Use XORP to simulate FNEG.
460 setOperationAction(ISD::FNEG , MVT::f32, Custom);
462 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
464 // Use ANDPS and ORPS to simulate FCOPYSIGN.
465 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
466 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
468 // We don't support sin/cos/fmod
469 setOperationAction(ISD::FSIN , MVT::f32, Expand);
470 setOperationAction(ISD::FCOS , MVT::f32, Expand);
472 // Special cases we handle for FP constants.
473 addLegalFPImmediate(APFloat(+0.0f)); // xorps
474 addLegalFPImmediate(APFloat(+0.0)); // FLD0
475 addLegalFPImmediate(APFloat(+1.0)); // FLD1
476 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
477 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
480 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
481 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
483 } else if (!UseSoftFloat) {
484 // f32 and f64 in x87.
485 // Set up the FP register classes.
486 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
487 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
489 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
490 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
491 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
492 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
495 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
496 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
498 addLegalFPImmediate(APFloat(+0.0)); // FLD0
499 addLegalFPImmediate(APFloat(+1.0)); // FLD1
500 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
501 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
502 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
503 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
504 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
505 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
508 // Long double always uses X87.
510 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
511 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
512 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
515 APFloat TmpFlt(+0.0);
516 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
518 addLegalFPImmediate(TmpFlt); // FLD0
520 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
521 APFloat TmpFlt2(+1.0);
522 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
524 addLegalFPImmediate(TmpFlt2); // FLD1
525 TmpFlt2.changeSign();
526 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
530 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
531 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
535 // Always use a library call for pow.
536 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
537 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
538 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
540 setOperationAction(ISD::FLOG, MVT::f80, Expand);
541 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
542 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
543 setOperationAction(ISD::FEXP, MVT::f80, Expand);
544 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
546 // First set operation action for all vector types to either promote
547 // (for widening) or expand (for scalarization). Then we will selectively
548 // turn on ones that can be effectively codegen'd.
549 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
550 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
551 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
566 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
567 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
600 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
604 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
605 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
606 setTruncStoreAction((MVT::SimpleValueType)VT,
607 (MVT::SimpleValueType)InnerVT, Expand);
608 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
609 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
610 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
613 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
614 // with -msoft-float, disable use of MMX as well.
615 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
616 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
617 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
618 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
620 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
622 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
623 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
624 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
625 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
627 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
628 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
629 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
630 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
632 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
633 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
635 setOperationAction(ISD::AND, MVT::v8i8, Promote);
636 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
637 setOperationAction(ISD::AND, MVT::v4i16, Promote);
638 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
639 setOperationAction(ISD::AND, MVT::v2i32, Promote);
640 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
641 setOperationAction(ISD::AND, MVT::v1i64, Legal);
643 setOperationAction(ISD::OR, MVT::v8i8, Promote);
644 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
645 setOperationAction(ISD::OR, MVT::v4i16, Promote);
646 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
647 setOperationAction(ISD::OR, MVT::v2i32, Promote);
648 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
649 setOperationAction(ISD::OR, MVT::v1i64, Legal);
651 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
652 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
653 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
654 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
655 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
656 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
657 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
659 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
660 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
661 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
662 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
663 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
664 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
665 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
667 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
669 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
670 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
672 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
674 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
675 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
677 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
678 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
679 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
681 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
683 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
684 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
685 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
686 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
689 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
691 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
692 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
693 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
694 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
695 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
699 if (!UseSoftFloat && Subtarget->hasSSE1()) {
700 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
702 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
703 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
704 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
705 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
706 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
707 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
708 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
709 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
710 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
711 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
712 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
713 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
716 if (!UseSoftFloat && Subtarget->hasSSE2()) {
717 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
719 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
720 // registers cannot be used even for integer operations.
721 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
722 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
723 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
724 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
726 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
727 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
728 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
729 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
730 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
731 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
732 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
733 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
734 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
735 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
736 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
737 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
738 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
739 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
740 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
741 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
743 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
744 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
745 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
746 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
748 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
749 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
750 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
751 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
752 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
754 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
755 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
756 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
757 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
760 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
761 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
762 EVT VT = (MVT::SimpleValueType)i;
763 // Do not attempt to custom lower non-power-of-2 vectors
764 if (!isPowerOf2_32(VT.getVectorNumElements()))
766 // Do not attempt to custom lower non-128-bit vectors
767 if (!VT.is128BitVector())
769 setOperationAction(ISD::BUILD_VECTOR,
770 VT.getSimpleVT().SimpleTy, Custom);
771 setOperationAction(ISD::VECTOR_SHUFFLE,
772 VT.getSimpleVT().SimpleTy, Custom);
773 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
774 VT.getSimpleVT().SimpleTy, Custom);
777 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
778 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
779 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
780 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
781 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
782 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
784 if (Subtarget->is64Bit()) {
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
789 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
790 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
791 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
794 // Do not attempt to promote non-128-bit vectors
795 if (!VT.is128BitVector())
798 setOperationAction(ISD::AND, SVT, Promote);
799 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
800 setOperationAction(ISD::OR, SVT, Promote);
801 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
802 setOperationAction(ISD::XOR, SVT, Promote);
803 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
804 setOperationAction(ISD::LOAD, SVT, Promote);
805 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
806 setOperationAction(ISD::SELECT, SVT, Promote);
807 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
810 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
812 // Custom lower v2i64 and v2f64 selects.
813 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
814 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
815 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
816 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
818 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
819 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
820 if (!DisableMMX && Subtarget->hasMMX()) {
821 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
822 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
826 if (Subtarget->hasSSE41()) {
827 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
828 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
829 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
830 setOperationAction(ISD::FRINT, MVT::f32, Legal);
831 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
832 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
833 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
834 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
835 setOperationAction(ISD::FRINT, MVT::f64, Legal);
836 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
838 // FIXME: Do we need to handle scalar-to-vector here?
839 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
841 // Can turn SHL into an integer multiply.
842 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
843 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
845 // i8 and i16 vectors are custom , because the source register and source
846 // source memory operand types are not the same width. f32 vectors are
847 // custom since the immediate controlling the insert encodes additional
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
850 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
851 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
852 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
854 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
855 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
856 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
857 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
859 if (Subtarget->is64Bit()) {
860 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
861 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
865 if (Subtarget->hasSSE42()) {
866 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
869 if (!UseSoftFloat && Subtarget->hasAVX()) {
870 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
871 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
872 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
873 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
874 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
876 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
877 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
878 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
879 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
880 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
881 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
882 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
883 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
884 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
885 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
886 setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
887 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
888 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
889 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
890 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
892 // Operations to consider commented out -v16i16 v32i8
893 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
894 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
895 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
896 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
897 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
898 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
899 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
900 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
901 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
902 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
903 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
904 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
905 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
906 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
908 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
909 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
910 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
911 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
913 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
914 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
915 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
916 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
917 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
919 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
920 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
921 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
922 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
923 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
924 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
927 // Not sure we want to do this since there are no 256-bit integer
930 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
931 // This includes 256-bit vectors
932 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
933 EVT VT = (MVT::SimpleValueType)i;
935 // Do not attempt to custom lower non-power-of-2 vectors
936 if (!isPowerOf2_32(VT.getVectorNumElements()))
939 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
940 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
941 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
944 if (Subtarget->is64Bit()) {
945 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
946 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
951 // Not sure we want to do this since there are no 256-bit integer
954 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
955 // Including 256-bit vectors
956 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
957 EVT VT = (MVT::SimpleValueType)i;
959 if (!VT.is256BitVector()) {
962 setOperationAction(ISD::AND, VT, Promote);
963 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
964 setOperationAction(ISD::OR, VT, Promote);
965 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
966 setOperationAction(ISD::XOR, VT, Promote);
967 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
968 setOperationAction(ISD::LOAD, VT, Promote);
969 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
970 setOperationAction(ISD::SELECT, VT, Promote);
971 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
974 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
978 // We want to custom lower some of our intrinsics.
979 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
981 // Add/Sub/Mul with overflow operations are custom lowered.
982 setOperationAction(ISD::SADDO, MVT::i32, Custom);
983 setOperationAction(ISD::UADDO, MVT::i32, Custom);
984 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
985 setOperationAction(ISD::USUBO, MVT::i32, Custom);
986 setOperationAction(ISD::SMULO, MVT::i32, Custom);
988 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
989 // handle type legalization for these operations here.
991 // FIXME: We really should do custom legalization for addition and
992 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
993 // than generic legalization for 64-bit multiplication-with-overflow, though.
994 if (Subtarget->is64Bit()) {
995 setOperationAction(ISD::SADDO, MVT::i64, Custom);
996 setOperationAction(ISD::UADDO, MVT::i64, Custom);
997 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
998 setOperationAction(ISD::USUBO, MVT::i64, Custom);
999 setOperationAction(ISD::SMULO, MVT::i64, Custom);
1002 if (!Subtarget->is64Bit()) {
1003 // These libcalls are not available in 32-bit.
1004 setLibcallName(RTLIB::SHL_I128, 0);
1005 setLibcallName(RTLIB::SRL_I128, 0);
1006 setLibcallName(RTLIB::SRA_I128, 0);
1009 // We have target-specific dag combine patterns for the following nodes:
1010 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1011 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1012 setTargetDAGCombine(ISD::BUILD_VECTOR);
1013 setTargetDAGCombine(ISD::SELECT);
1014 setTargetDAGCombine(ISD::SHL);
1015 setTargetDAGCombine(ISD::SRA);
1016 setTargetDAGCombine(ISD::SRL);
1017 setTargetDAGCombine(ISD::OR);
1018 setTargetDAGCombine(ISD::STORE);
1019 setTargetDAGCombine(ISD::ZERO_EXTEND);
1020 if (Subtarget->is64Bit())
1021 setTargetDAGCombine(ISD::MUL);
1023 computeRegisterProperties();
1025 // FIXME: These should be based on subtarget info. Plus, the values should
1026 // be smaller when we are in optimizing for size mode.
1027 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1028 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1029 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1030 setPrefLoopAlignment(16);
1031 benefitFromCodePlacementOpt = true;
1035 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1040 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1041 /// the desired ByVal argument alignment.
1042 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1045 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1046 if (VTy->getBitWidth() == 128)
1048 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1049 unsigned EltAlign = 0;
1050 getMaxByValAlign(ATy->getElementType(), EltAlign);
1051 if (EltAlign > MaxAlign)
1052 MaxAlign = EltAlign;
1053 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1054 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1055 unsigned EltAlign = 0;
1056 getMaxByValAlign(STy->getElementType(i), EltAlign);
1057 if (EltAlign > MaxAlign)
1058 MaxAlign = EltAlign;
1066 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1067 /// function arguments in the caller parameter area. For X86, aggregates
1068 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1069 /// are at 4-byte boundaries.
1070 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1071 if (Subtarget->is64Bit()) {
1072 // Max of 8 and alignment of type.
1073 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1080 if (Subtarget->hasSSE1())
1081 getMaxByValAlign(Ty, Align);
1085 /// getOptimalMemOpType - Returns the target specific optimal type for load
1086 /// and store operations as a result of memset, memcpy, and memmove
1087 /// lowering. If DstAlign is zero that means it's safe to destination
1088 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1089 /// means there isn't a need to check it against alignment requirement,
1090 /// probably because the source does not need to be loaded. If
1091 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1092 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1093 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1094 /// constant so it does not need to be loaded.
1095 /// It returns EVT::Other if the type should be determined using generic
1096 /// target-independent logic.
1098 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1099 unsigned DstAlign, unsigned SrcAlign,
1100 bool NonScalarIntSafe,
1102 MachineFunction &MF) const {
1103 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1104 // linux. This is because the stack realignment code can't handle certain
1105 // cases like PR2962. This should be removed when PR2962 is fixed.
1106 const Function *F = MF.getFunction();
1107 if (NonScalarIntSafe &&
1108 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1110 (Subtarget->isUnalignedMemAccessFast() ||
1111 ((DstAlign == 0 || DstAlign >= 16) &&
1112 (SrcAlign == 0 || SrcAlign >= 16))) &&
1113 Subtarget->getStackAlignment() >= 16) {
1114 if (Subtarget->hasSSE2())
1116 if (Subtarget->hasSSE1())
1118 } else if (!MemcpyStrSrc && Size >= 8 &&
1119 !Subtarget->is64Bit() &&
1120 Subtarget->getStackAlignment() >= 8 &&
1121 Subtarget->hasSSE2()) {
1122 // Do not use f64 to lower memcpy if source is string constant. It's
1123 // better to use i32 to avoid the loads.
1127 if (Subtarget->is64Bit() && Size >= 8)
1132 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1133 /// current function. The returned value is a member of the
1134 /// MachineJumpTableInfo::JTEntryKind enum.
1135 unsigned X86TargetLowering::getJumpTableEncoding() const {
1136 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1138 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1139 Subtarget->isPICStyleGOT())
1140 return MachineJumpTableInfo::EK_Custom32;
1142 // Otherwise, use the normal jump table encoding heuristics.
1143 return TargetLowering::getJumpTableEncoding();
1146 /// getPICBaseSymbol - Return the X86-32 PIC base.
1148 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1149 MCContext &Ctx) const {
1150 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1151 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1152 Twine(MF->getFunctionNumber())+"$pb");
1157 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1158 const MachineBasicBlock *MBB,
1159 unsigned uid,MCContext &Ctx) const{
1160 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1161 Subtarget->isPICStyleGOT());
1162 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1164 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1165 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1168 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1170 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1171 SelectionDAG &DAG) const {
1172 if (!Subtarget->is64Bit())
1173 // This doesn't have DebugLoc associated with it, but is not really the
1174 // same as a Register.
1175 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1179 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1180 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1182 const MCExpr *X86TargetLowering::
1183 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1184 MCContext &Ctx) const {
1185 // X86-64 uses RIP relative addressing based on the jump table label.
1186 if (Subtarget->isPICStyleRIPRel())
1187 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1189 // Otherwise, the reference is relative to the PIC base.
1190 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1193 /// getFunctionAlignment - Return the Log2 alignment of this function.
1194 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1195 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1198 std::pair<const TargetRegisterClass*, uint8_t>
1199 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1200 const TargetRegisterClass *RRC = 0;
1202 switch (VT.getSimpleVT().SimpleTy) {
1204 return TargetLowering::findRepresentativeClass(VT);
1205 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1206 RRC = (Subtarget->is64Bit()
1207 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1209 case MVT::v8i8: case MVT::v4i16:
1210 case MVT::v2i32: case MVT::v1i64:
1211 RRC = X86::VR64RegisterClass;
1213 case MVT::f32: case MVT::f64:
1214 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1215 case MVT::v4f32: case MVT::v2f64:
1216 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1218 RRC = X86::VR128RegisterClass;
1221 return std::make_pair(RRC, Cost);
1225 X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
1226 MachineFunction &MF) const {
1227 unsigned FPDiff = RegInfo->hasFP(MF) ? 1 : 0;
1228 switch (RC->getID()) {
1231 case X86::GR32RegClassID:
1233 case X86::GR64RegClassID:
1235 case X86::VR128RegClassID:
1236 return Subtarget->is64Bit() ? 10 : 4;
1237 case X86::VR64RegClassID:
1242 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1243 unsigned &Offset) const {
1244 if (!Subtarget->isTargetLinux())
1247 if (Subtarget->is64Bit()) {
1248 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1250 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1263 //===----------------------------------------------------------------------===//
1264 // Return Value Calling Convention Implementation
1265 //===----------------------------------------------------------------------===//
1267 #include "X86GenCallingConv.inc"
1270 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1271 const SmallVectorImpl<ISD::OutputArg> &Outs,
1272 LLVMContext &Context) const {
1273 SmallVector<CCValAssign, 16> RVLocs;
1274 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1276 return CCInfo.CheckReturn(Outs, RetCC_X86);
1280 X86TargetLowering::LowerReturn(SDValue Chain,
1281 CallingConv::ID CallConv, bool isVarArg,
1282 const SmallVectorImpl<ISD::OutputArg> &Outs,
1283 const SmallVectorImpl<SDValue> &OutVals,
1284 DebugLoc dl, SelectionDAG &DAG) const {
1285 MachineFunction &MF = DAG.getMachineFunction();
1286 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1288 SmallVector<CCValAssign, 16> RVLocs;
1289 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1290 RVLocs, *DAG.getContext());
1291 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1293 // Add the regs to the liveout set for the function.
1294 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1295 for (unsigned i = 0; i != RVLocs.size(); ++i)
1296 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1297 MRI.addLiveOut(RVLocs[i].getLocReg());
1301 SmallVector<SDValue, 6> RetOps;
1302 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1303 // Operand #1 = Bytes To Pop
1304 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1307 // Copy the result values into the output registers.
1308 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1309 CCValAssign &VA = RVLocs[i];
1310 assert(VA.isRegLoc() && "Can only return in registers!");
1311 SDValue ValToCopy = OutVals[i];
1312 EVT ValVT = ValToCopy.getValueType();
1314 // If this is x86-64, and we disabled SSE, we can't return FP values
1315 if ((ValVT == MVT::f32 || ValVT == MVT::f64) &&
1316 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1317 report_fatal_error("SSE register return with SSE disabled");
1319 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1320 // llvm-gcc has never done it right and no one has noticed, so this
1321 // should be OK for now.
1322 if (ValVT == MVT::f64 &&
1323 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1324 report_fatal_error("SSE2 register return with SSE2 disabled");
1326 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1327 // the RET instruction and handled by the FP Stackifier.
1328 if (VA.getLocReg() == X86::ST0 ||
1329 VA.getLocReg() == X86::ST1) {
1330 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1331 // change the value to the FP stack register class.
1332 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1333 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1334 RetOps.push_back(ValToCopy);
1335 // Don't emit a copytoreg.
1339 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1340 // which is returned in RAX / RDX.
1341 if (Subtarget->is64Bit()) {
1342 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1343 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1344 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1345 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1348 // If we don't have SSE2 available, convert to v4f32 so the generated
1349 // register is legal.
1350 if (!Subtarget->hasSSE2())
1351 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32,ValToCopy);
1356 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1357 Flag = Chain.getValue(1);
1360 // The x86-64 ABI for returning structs by value requires that we copy
1361 // the sret argument into %rax for the return. We saved the argument into
1362 // a virtual register in the entry block, so now we copy the value out
1364 if (Subtarget->is64Bit() &&
1365 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1366 MachineFunction &MF = DAG.getMachineFunction();
1367 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1368 unsigned Reg = FuncInfo->getSRetReturnReg();
1370 "SRetReturnReg should have been set in LowerFormalArguments().");
1371 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1373 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1374 Flag = Chain.getValue(1);
1376 // RAX now acts like a return value.
1377 MRI.addLiveOut(X86::RAX);
1380 RetOps[0] = Chain; // Update chain.
1382 // Add the flag if we have it.
1384 RetOps.push_back(Flag);
1386 return DAG.getNode(X86ISD::RET_FLAG, dl,
1387 MVT::Other, &RetOps[0], RetOps.size());
1390 /// LowerCallResult - Lower the result values of a call into the
1391 /// appropriate copies out of appropriate physical registers.
1394 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1395 CallingConv::ID CallConv, bool isVarArg,
1396 const SmallVectorImpl<ISD::InputArg> &Ins,
1397 DebugLoc dl, SelectionDAG &DAG,
1398 SmallVectorImpl<SDValue> &InVals) const {
1400 // Assign locations to each value returned by this call.
1401 SmallVector<CCValAssign, 16> RVLocs;
1402 bool Is64Bit = Subtarget->is64Bit();
1403 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1404 RVLocs, *DAG.getContext());
1405 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1407 // Copy all of the result registers out of their specified physreg.
1408 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1409 CCValAssign &VA = RVLocs[i];
1410 EVT CopyVT = VA.getValVT();
1412 // If this is x86-64, and we disabled SSE, we can't return FP values
1413 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1414 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1415 report_fatal_error("SSE register return with SSE disabled");
1420 // If this is a call to a function that returns an fp value on the floating
1421 // point stack, we must guarantee the the value is popped from the stack, so
1422 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1423 // if the return value is not used. We use the FpGET_ST0 instructions
1425 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1426 // If we prefer to use the value in xmm registers, copy it out as f80 and
1427 // use a truncate to move it from fp stack reg to xmm reg.
1428 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1429 bool isST0 = VA.getLocReg() == X86::ST0;
1431 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1432 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1433 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1434 SDValue Ops[] = { Chain, InFlag };
1435 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
1437 Val = Chain.getValue(0);
1439 // Round the f80 to the right size, which also moves it to the appropriate
1441 if (CopyVT != VA.getValVT())
1442 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1443 // This truncation won't change the value.
1444 DAG.getIntPtrConstant(1));
1445 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1446 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1447 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1448 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1449 MVT::v2i64, InFlag).getValue(1);
1450 Val = Chain.getValue(0);
1451 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1452 Val, DAG.getConstant(0, MVT::i64));
1454 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1455 MVT::i64, InFlag).getValue(1);
1456 Val = Chain.getValue(0);
1458 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1460 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1461 CopyVT, InFlag).getValue(1);
1462 Val = Chain.getValue(0);
1464 InFlag = Chain.getValue(2);
1465 InVals.push_back(Val);
1472 //===----------------------------------------------------------------------===//
1473 // C & StdCall & Fast Calling Convention implementation
1474 //===----------------------------------------------------------------------===//
1475 // StdCall calling convention seems to be standard for many Windows' API
1476 // routines and around. It differs from C calling convention just a little:
1477 // callee should clean up the stack, not caller. Symbols should be also
1478 // decorated in some fancy way :) It doesn't support any vector arguments.
1479 // For info on fast calling convention see Fast Calling Convention (tail call)
1480 // implementation LowerX86_32FastCCCallTo.
1482 /// CallIsStructReturn - Determines whether a call uses struct return
1484 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1488 return Outs[0].Flags.isSRet();
1491 /// ArgsAreStructReturn - Determines whether a function uses struct
1492 /// return semantics.
1494 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1498 return Ins[0].Flags.isSRet();
1501 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1502 /// given CallingConvention value.
1503 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1504 if (Subtarget->is64Bit()) {
1505 if (CC == CallingConv::GHC)
1506 return CC_X86_64_GHC;
1507 else if (Subtarget->isTargetWin64())
1508 return CC_X86_Win64_C;
1513 if (CC == CallingConv::X86_FastCall)
1514 return CC_X86_32_FastCall;
1515 else if (CC == CallingConv::X86_ThisCall)
1516 return CC_X86_32_ThisCall;
1517 else if (CC == CallingConv::Fast)
1518 return CC_X86_32_FastCC;
1519 else if (CC == CallingConv::GHC)
1520 return CC_X86_32_GHC;
1525 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1526 /// by "Src" to address "Dst" with size and alignment information specified by
1527 /// the specific parameter attribute. The copy will be passed as a byval
1528 /// function parameter.
1530 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1531 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1533 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1534 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1535 /*isVolatile*/false, /*AlwaysInline=*/true,
1539 /// IsTailCallConvention - Return true if the calling convention is one that
1540 /// supports tail call optimization.
1541 static bool IsTailCallConvention(CallingConv::ID CC) {
1542 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1545 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1546 /// a tailcall target by changing its ABI.
1547 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1548 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1552 X86TargetLowering::LowerMemArgument(SDValue Chain,
1553 CallingConv::ID CallConv,
1554 const SmallVectorImpl<ISD::InputArg> &Ins,
1555 DebugLoc dl, SelectionDAG &DAG,
1556 const CCValAssign &VA,
1557 MachineFrameInfo *MFI,
1559 // Create the nodes corresponding to a load from this parameter slot.
1560 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1561 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1562 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1565 // If value is passed by pointer we have address passed instead of the value
1567 if (VA.getLocInfo() == CCValAssign::Indirect)
1568 ValVT = VA.getLocVT();
1570 ValVT = VA.getValVT();
1572 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1573 // changed with more analysis.
1574 // In case of tail call optimization mark all arguments mutable. Since they
1575 // could be overwritten by lowering of arguments in case of a tail call.
1576 if (Flags.isByVal()) {
1577 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1578 VA.getLocMemOffset(), isImmutable);
1579 return DAG.getFrameIndex(FI, getPointerTy());
1581 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1582 VA.getLocMemOffset(), isImmutable);
1583 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1584 return DAG.getLoad(ValVT, dl, Chain, FIN,
1585 PseudoSourceValue::getFixedStack(FI), 0,
1591 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1592 CallingConv::ID CallConv,
1594 const SmallVectorImpl<ISD::InputArg> &Ins,
1597 SmallVectorImpl<SDValue> &InVals)
1599 MachineFunction &MF = DAG.getMachineFunction();
1600 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1602 const Function* Fn = MF.getFunction();
1603 if (Fn->hasExternalLinkage() &&
1604 Subtarget->isTargetCygMing() &&
1605 Fn->getName() == "main")
1606 FuncInfo->setForceFramePointer(true);
1608 MachineFrameInfo *MFI = MF.getFrameInfo();
1609 bool Is64Bit = Subtarget->is64Bit();
1610 bool IsWin64 = Subtarget->isTargetWin64();
1612 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1613 "Var args not supported with calling convention fastcc or ghc");
1615 // Assign locations to all of the incoming arguments.
1616 SmallVector<CCValAssign, 16> ArgLocs;
1617 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1618 ArgLocs, *DAG.getContext());
1619 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1621 unsigned LastVal = ~0U;
1623 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1624 CCValAssign &VA = ArgLocs[i];
1625 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1627 assert(VA.getValNo() != LastVal &&
1628 "Don't support value assigned to multiple locs yet");
1629 LastVal = VA.getValNo();
1631 if (VA.isRegLoc()) {
1632 EVT RegVT = VA.getLocVT();
1633 TargetRegisterClass *RC = NULL;
1634 if (RegVT == MVT::i32)
1635 RC = X86::GR32RegisterClass;
1636 else if (Is64Bit && RegVT == MVT::i64)
1637 RC = X86::GR64RegisterClass;
1638 else if (RegVT == MVT::f32)
1639 RC = X86::FR32RegisterClass;
1640 else if (RegVT == MVT::f64)
1641 RC = X86::FR64RegisterClass;
1642 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1643 RC = X86::VR256RegisterClass;
1644 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1645 RC = X86::VR128RegisterClass;
1646 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1647 RC = X86::VR64RegisterClass;
1649 llvm_unreachable("Unknown argument type!");
1651 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1652 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1654 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1655 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1657 if (VA.getLocInfo() == CCValAssign::SExt)
1658 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1659 DAG.getValueType(VA.getValVT()));
1660 else if (VA.getLocInfo() == CCValAssign::ZExt)
1661 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1662 DAG.getValueType(VA.getValVT()));
1663 else if (VA.getLocInfo() == CCValAssign::BCvt)
1664 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1666 if (VA.isExtInLoc()) {
1667 // Handle MMX values passed in XMM regs.
1668 if (RegVT.isVector()) {
1669 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1670 ArgValue, DAG.getConstant(0, MVT::i64));
1671 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1673 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1676 assert(VA.isMemLoc());
1677 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1680 // If value is passed via pointer - do a load.
1681 if (VA.getLocInfo() == CCValAssign::Indirect)
1682 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1685 InVals.push_back(ArgValue);
1688 // The x86-64 ABI for returning structs by value requires that we copy
1689 // the sret argument into %rax for the return. Save the argument into
1690 // a virtual register so that we can access it from the return points.
1691 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1692 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1693 unsigned Reg = FuncInfo->getSRetReturnReg();
1695 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1696 FuncInfo->setSRetReturnReg(Reg);
1698 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1699 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1702 unsigned StackSize = CCInfo.getNextStackOffset();
1703 // Align stack specially for tail calls.
1704 if (FuncIsMadeTailCallSafe(CallConv))
1705 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1707 // If the function takes variable number of arguments, make a frame index for
1708 // the start of the first vararg value... for expansion of llvm.va_start.
1710 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1711 CallConv != CallingConv::X86_ThisCall)) {
1712 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1715 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1717 // FIXME: We should really autogenerate these arrays
1718 static const unsigned GPR64ArgRegsWin64[] = {
1719 X86::RCX, X86::RDX, X86::R8, X86::R9
1721 static const unsigned XMMArgRegsWin64[] = {
1722 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1724 static const unsigned GPR64ArgRegs64Bit[] = {
1725 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1727 static const unsigned XMMArgRegs64Bit[] = {
1728 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1729 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1731 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1734 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1735 GPR64ArgRegs = GPR64ArgRegsWin64;
1736 XMMArgRegs = XMMArgRegsWin64;
1738 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1739 GPR64ArgRegs = GPR64ArgRegs64Bit;
1740 XMMArgRegs = XMMArgRegs64Bit;
1742 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1744 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1747 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1748 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1749 "SSE register cannot be used when SSE is disabled!");
1750 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1751 "SSE register cannot be used when SSE is disabled!");
1752 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1753 // Kernel mode asks for SSE to be disabled, so don't push them
1755 TotalNumXMMRegs = 0;
1757 // For X86-64, if there are vararg parameters that are passed via
1758 // registers, then we must store them to their spots on the stack so they
1759 // may be loaded by deferencing the result of va_next.
1760 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1761 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1762 FuncInfo->setRegSaveFrameIndex(
1763 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1766 // Store the integer parameter registers.
1767 SmallVector<SDValue, 8> MemOps;
1768 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1770 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1771 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1772 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1773 DAG.getIntPtrConstant(Offset));
1774 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1775 X86::GR64RegisterClass);
1776 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1778 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1779 PseudoSourceValue::getFixedStack(
1780 FuncInfo->getRegSaveFrameIndex()),
1781 Offset, false, false, 0);
1782 MemOps.push_back(Store);
1786 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1787 // Now store the XMM (fp + vector) parameter registers.
1788 SmallVector<SDValue, 11> SaveXMMOps;
1789 SaveXMMOps.push_back(Chain);
1791 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1792 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1793 SaveXMMOps.push_back(ALVal);
1795 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1796 FuncInfo->getRegSaveFrameIndex()));
1797 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1798 FuncInfo->getVarArgsFPOffset()));
1800 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1801 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1802 X86::VR128RegisterClass);
1803 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1804 SaveXMMOps.push_back(Val);
1806 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1808 &SaveXMMOps[0], SaveXMMOps.size()));
1811 if (!MemOps.empty())
1812 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1813 &MemOps[0], MemOps.size());
1817 // Some CCs need callee pop.
1818 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1819 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1821 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1822 // If this is an sret function, the return should pop the hidden pointer.
1823 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1824 FuncInfo->setBytesToPopOnReturn(4);
1828 // RegSaveFrameIndex is X86-64 only.
1829 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1830 if (CallConv == CallingConv::X86_FastCall ||
1831 CallConv == CallingConv::X86_ThisCall)
1832 // fastcc functions can't have varargs.
1833 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1840 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1841 SDValue StackPtr, SDValue Arg,
1842 DebugLoc dl, SelectionDAG &DAG,
1843 const CCValAssign &VA,
1844 ISD::ArgFlagsTy Flags) const {
1845 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1846 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1847 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1848 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1849 if (Flags.isByVal()) {
1850 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1852 return DAG.getStore(Chain, dl, Arg, PtrOff,
1853 PseudoSourceValue::getStack(), LocMemOffset,
1857 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1858 /// optimization is performed and it is required.
1860 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1861 SDValue &OutRetAddr, SDValue Chain,
1862 bool IsTailCall, bool Is64Bit,
1863 int FPDiff, DebugLoc dl) const {
1864 // Adjust the Return address stack slot.
1865 EVT VT = getPointerTy();
1866 OutRetAddr = getReturnAddressFrameIndex(DAG);
1868 // Load the "old" Return address.
1869 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1870 return SDValue(OutRetAddr.getNode(), 1);
1873 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1874 /// optimization is performed and it is required (FPDiff!=0).
1876 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1877 SDValue Chain, SDValue RetAddrFrIdx,
1878 bool Is64Bit, int FPDiff, DebugLoc dl) {
1879 // Store the return address to the appropriate stack slot.
1880 if (!FPDiff) return Chain;
1881 // Calculate the new stack slot for the return address.
1882 int SlotSize = Is64Bit ? 8 : 4;
1883 int NewReturnAddrFI =
1884 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1885 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1886 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1887 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1888 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1894 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1895 CallingConv::ID CallConv, bool isVarArg,
1897 const SmallVectorImpl<ISD::OutputArg> &Outs,
1898 const SmallVectorImpl<SDValue> &OutVals,
1899 const SmallVectorImpl<ISD::InputArg> &Ins,
1900 DebugLoc dl, SelectionDAG &DAG,
1901 SmallVectorImpl<SDValue> &InVals) const {
1902 MachineFunction &MF = DAG.getMachineFunction();
1903 bool Is64Bit = Subtarget->is64Bit();
1904 bool IsStructRet = CallIsStructReturn(Outs);
1905 bool IsSibcall = false;
1908 // Check if it's really possible to do a tail call.
1909 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1910 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1911 Outs, OutVals, Ins, DAG);
1913 // Sibcalls are automatically detected tailcalls which do not require
1915 if (!GuaranteedTailCallOpt && isTailCall)
1922 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1923 "Var args not supported with calling convention fastcc or ghc");
1925 // Analyze operands of the call, assigning locations to each operand.
1926 SmallVector<CCValAssign, 16> ArgLocs;
1927 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1928 ArgLocs, *DAG.getContext());
1929 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1931 // Get a count of how many bytes are to be pushed on the stack.
1932 unsigned NumBytes = CCInfo.getNextStackOffset();
1934 // This is a sibcall. The memory operands are available in caller's
1935 // own caller's stack.
1937 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1938 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1941 if (isTailCall && !IsSibcall) {
1942 // Lower arguments at fp - stackoffset + fpdiff.
1943 unsigned NumBytesCallerPushed =
1944 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1945 FPDiff = NumBytesCallerPushed - NumBytes;
1947 // Set the delta of movement of the returnaddr stackslot.
1948 // But only set if delta is greater than previous delta.
1949 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1950 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1954 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1956 SDValue RetAddrFrIdx;
1957 // Load return adress for tail calls.
1958 if (isTailCall && FPDiff)
1959 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1960 Is64Bit, FPDiff, dl);
1962 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1963 SmallVector<SDValue, 8> MemOpChains;
1966 // Walk the register/memloc assignments, inserting copies/loads. In the case
1967 // of tail call optimization arguments are handle later.
1968 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1969 CCValAssign &VA = ArgLocs[i];
1970 EVT RegVT = VA.getLocVT();
1971 SDValue Arg = OutVals[i];
1972 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1973 bool isByVal = Flags.isByVal();
1975 // Promote the value if needed.
1976 switch (VA.getLocInfo()) {
1977 default: llvm_unreachable("Unknown loc info!");
1978 case CCValAssign::Full: break;
1979 case CCValAssign::SExt:
1980 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1982 case CCValAssign::ZExt:
1983 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1985 case CCValAssign::AExt:
1986 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1987 // Special case: passing MMX values in XMM registers.
1988 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1989 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1990 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1992 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1994 case CCValAssign::BCvt:
1995 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1997 case CCValAssign::Indirect: {
1998 // Store the argument.
1999 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2000 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2001 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2002 PseudoSourceValue::getFixedStack(FI), 0,
2009 if (VA.isRegLoc()) {
2010 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2011 if (isVarArg && Subtarget->isTargetWin64()) {
2012 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2013 // shadow reg if callee is a varargs function.
2014 unsigned ShadowReg = 0;
2015 switch (VA.getLocReg()) {
2016 case X86::XMM0: ShadowReg = X86::RCX; break;
2017 case X86::XMM1: ShadowReg = X86::RDX; break;
2018 case X86::XMM2: ShadowReg = X86::R8; break;
2019 case X86::XMM3: ShadowReg = X86::R9; break;
2022 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2024 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2025 assert(VA.isMemLoc());
2026 if (StackPtr.getNode() == 0)
2027 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2028 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2029 dl, DAG, VA, Flags));
2033 if (!MemOpChains.empty())
2034 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2035 &MemOpChains[0], MemOpChains.size());
2037 // Build a sequence of copy-to-reg nodes chained together with token chain
2038 // and flag operands which copy the outgoing args into registers.
2040 // Tail call byval lowering might overwrite argument registers so in case of
2041 // tail call optimization the copies to registers are lowered later.
2043 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2044 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2045 RegsToPass[i].second, InFlag);
2046 InFlag = Chain.getValue(1);
2049 if (Subtarget->isPICStyleGOT()) {
2050 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2053 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2054 DAG.getNode(X86ISD::GlobalBaseReg,
2055 DebugLoc(), getPointerTy()),
2057 InFlag = Chain.getValue(1);
2059 // If we are tail calling and generating PIC/GOT style code load the
2060 // address of the callee into ECX. The value in ecx is used as target of
2061 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2062 // for tail calls on PIC/GOT architectures. Normally we would just put the
2063 // address of GOT into ebx and then call target@PLT. But for tail calls
2064 // ebx would be restored (since ebx is callee saved) before jumping to the
2067 // Note: The actual moving to ECX is done further down.
2068 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2069 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2070 !G->getGlobal()->hasProtectedVisibility())
2071 Callee = LowerGlobalAddress(Callee, DAG);
2072 else if (isa<ExternalSymbolSDNode>(Callee))
2073 Callee = LowerExternalSymbol(Callee, DAG);
2077 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
2078 // From AMD64 ABI document:
2079 // For calls that may call functions that use varargs or stdargs
2080 // (prototype-less calls or calls to functions containing ellipsis (...) in
2081 // the declaration) %al is used as hidden argument to specify the number
2082 // of SSE registers used. The contents of %al do not need to match exactly
2083 // the number of registers, but must be an ubound on the number of SSE
2084 // registers used and is in the range 0 - 8 inclusive.
2086 // Count the number of XMM registers allocated.
2087 static const unsigned XMMArgRegs[] = {
2088 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2089 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2091 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2092 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2093 && "SSE registers cannot be used when SSE is disabled");
2095 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2096 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2097 InFlag = Chain.getValue(1);
2101 // For tail calls lower the arguments to the 'real' stack slot.
2103 // Force all the incoming stack arguments to be loaded from the stack
2104 // before any new outgoing arguments are stored to the stack, because the
2105 // outgoing stack slots may alias the incoming argument stack slots, and
2106 // the alias isn't otherwise explicit. This is slightly more conservative
2107 // than necessary, because it means that each store effectively depends
2108 // on every argument instead of just those arguments it would clobber.
2109 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2111 SmallVector<SDValue, 8> MemOpChains2;
2114 // Do not flag preceeding copytoreg stuff together with the following stuff.
2116 if (GuaranteedTailCallOpt) {
2117 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2118 CCValAssign &VA = ArgLocs[i];
2121 assert(VA.isMemLoc());
2122 SDValue Arg = OutVals[i];
2123 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2124 // Create frame index.
2125 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2126 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2127 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2128 FIN = DAG.getFrameIndex(FI, getPointerTy());
2130 if (Flags.isByVal()) {
2131 // Copy relative to framepointer.
2132 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2133 if (StackPtr.getNode() == 0)
2134 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2136 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2138 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2142 // Store relative to framepointer.
2143 MemOpChains2.push_back(
2144 DAG.getStore(ArgChain, dl, Arg, FIN,
2145 PseudoSourceValue::getFixedStack(FI), 0,
2151 if (!MemOpChains2.empty())
2152 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2153 &MemOpChains2[0], MemOpChains2.size());
2155 // Copy arguments to their registers.
2156 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2157 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2158 RegsToPass[i].second, InFlag);
2159 InFlag = Chain.getValue(1);
2163 // Store the return address to the appropriate stack slot.
2164 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2168 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2169 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2170 // In the 64-bit large code model, we have to make all calls
2171 // through a register, since the call instruction's 32-bit
2172 // pc-relative offset may not be large enough to hold the whole
2174 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2175 // If the callee is a GlobalAddress node (quite common, every direct call
2176 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2179 // We should use extra load for direct calls to dllimported functions in
2181 const GlobalValue *GV = G->getGlobal();
2182 if (!GV->hasDLLImportLinkage()) {
2183 unsigned char OpFlags = 0;
2185 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2186 // external symbols most go through the PLT in PIC mode. If the symbol
2187 // has hidden or protected visibility, or if it is static or local, then
2188 // we don't need to use the PLT - we can directly call it.
2189 if (Subtarget->isTargetELF() &&
2190 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2191 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2192 OpFlags = X86II::MO_PLT;
2193 } else if (Subtarget->isPICStyleStubAny() &&
2194 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2195 Subtarget->getDarwinVers() < 9) {
2196 // PC-relative references to external symbols should go through $stub,
2197 // unless we're building with the leopard linker or later, which
2198 // automatically synthesizes these stubs.
2199 OpFlags = X86II::MO_DARWIN_STUB;
2202 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2203 G->getOffset(), OpFlags);
2205 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2206 unsigned char OpFlags = 0;
2208 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2209 // symbols should go through the PLT.
2210 if (Subtarget->isTargetELF() &&
2211 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2212 OpFlags = X86II::MO_PLT;
2213 } else if (Subtarget->isPICStyleStubAny() &&
2214 Subtarget->getDarwinVers() < 9) {
2215 // PC-relative references to external symbols should go through $stub,
2216 // unless we're building with the leopard linker or later, which
2217 // automatically synthesizes these stubs.
2218 OpFlags = X86II::MO_DARWIN_STUB;
2221 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2225 // Returns a chain & a flag for retval copy to use.
2226 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2227 SmallVector<SDValue, 8> Ops;
2229 if (!IsSibcall && isTailCall) {
2230 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2231 DAG.getIntPtrConstant(0, true), InFlag);
2232 InFlag = Chain.getValue(1);
2235 Ops.push_back(Chain);
2236 Ops.push_back(Callee);
2239 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2241 // Add argument registers to the end of the list so that they are known live
2243 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2244 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2245 RegsToPass[i].second.getValueType()));
2247 // Add an implicit use GOT pointer in EBX.
2248 if (!isTailCall && Subtarget->isPICStyleGOT())
2249 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2251 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2252 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64())
2253 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2255 if (InFlag.getNode())
2256 Ops.push_back(InFlag);
2260 //// If this is the first return lowered for this function, add the regs
2261 //// to the liveout set for the function.
2262 // This isn't right, although it's probably harmless on x86; liveouts
2263 // should be computed from returns not tail calls. Consider a void
2264 // function making a tail call to a function returning int.
2265 return DAG.getNode(X86ISD::TC_RETURN, dl,
2266 NodeTys, &Ops[0], Ops.size());
2269 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2270 InFlag = Chain.getValue(1);
2272 // Create the CALLSEQ_END node.
2273 unsigned NumBytesForCalleeToPush;
2274 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2275 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2276 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2277 // If this is a call to a struct-return function, the callee
2278 // pops the hidden struct pointer, so we have to push it back.
2279 // This is common for Darwin/X86, Linux & Mingw32 targets.
2280 NumBytesForCalleeToPush = 4;
2282 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2284 // Returns a flag for retval copy to use.
2286 Chain = DAG.getCALLSEQ_END(Chain,
2287 DAG.getIntPtrConstant(NumBytes, true),
2288 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2291 InFlag = Chain.getValue(1);
2294 // Handle result values, copying them out of physregs into vregs that we
2296 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2297 Ins, dl, DAG, InVals);
2301 //===----------------------------------------------------------------------===//
2302 // Fast Calling Convention (tail call) implementation
2303 //===----------------------------------------------------------------------===//
2305 // Like std call, callee cleans arguments, convention except that ECX is
2306 // reserved for storing the tail called function address. Only 2 registers are
2307 // free for argument passing (inreg). Tail call optimization is performed
2309 // * tailcallopt is enabled
2310 // * caller/callee are fastcc
2311 // On X86_64 architecture with GOT-style position independent code only local
2312 // (within module) calls are supported at the moment.
2313 // To keep the stack aligned according to platform abi the function
2314 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2315 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2316 // If a tail called function callee has more arguments than the caller the
2317 // caller needs to make sure that there is room to move the RETADDR to. This is
2318 // achieved by reserving an area the size of the argument delta right after the
2319 // original REtADDR, but before the saved framepointer or the spilled registers
2320 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2332 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2333 /// for a 16 byte align requirement.
2335 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2336 SelectionDAG& DAG) const {
2337 MachineFunction &MF = DAG.getMachineFunction();
2338 const TargetMachine &TM = MF.getTarget();
2339 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2340 unsigned StackAlignment = TFI.getStackAlignment();
2341 uint64_t AlignMask = StackAlignment - 1;
2342 int64_t Offset = StackSize;
2343 uint64_t SlotSize = TD->getPointerSize();
2344 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2345 // Number smaller than 12 so just add the difference.
2346 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2348 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2349 Offset = ((~AlignMask) & Offset) + StackAlignment +
2350 (StackAlignment-SlotSize);
2355 /// MatchingStackOffset - Return true if the given stack call argument is
2356 /// already available in the same position (relatively) of the caller's
2357 /// incoming argument stack.
2359 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2360 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2361 const X86InstrInfo *TII) {
2362 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2364 if (Arg.getOpcode() == ISD::CopyFromReg) {
2365 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2366 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2368 MachineInstr *Def = MRI->getVRegDef(VR);
2371 if (!Flags.isByVal()) {
2372 if (!TII->isLoadFromStackSlot(Def, FI))
2375 unsigned Opcode = Def->getOpcode();
2376 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2377 Def->getOperand(1).isFI()) {
2378 FI = Def->getOperand(1).getIndex();
2379 Bytes = Flags.getByValSize();
2383 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2384 if (Flags.isByVal())
2385 // ByVal argument is passed in as a pointer but it's now being
2386 // dereferenced. e.g.
2387 // define @foo(%struct.X* %A) {
2388 // tail call @bar(%struct.X* byval %A)
2391 SDValue Ptr = Ld->getBasePtr();
2392 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2395 FI = FINode->getIndex();
2399 assert(FI != INT_MAX);
2400 if (!MFI->isFixedObjectIndex(FI))
2402 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2405 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2406 /// for tail call optimization. Targets which want to do tail call
2407 /// optimization should implement this function.
2409 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2410 CallingConv::ID CalleeCC,
2412 bool isCalleeStructRet,
2413 bool isCallerStructRet,
2414 const SmallVectorImpl<ISD::OutputArg> &Outs,
2415 const SmallVectorImpl<SDValue> &OutVals,
2416 const SmallVectorImpl<ISD::InputArg> &Ins,
2417 SelectionDAG& DAG) const {
2418 if (!IsTailCallConvention(CalleeCC) &&
2419 CalleeCC != CallingConv::C)
2422 // If -tailcallopt is specified, make fastcc functions tail-callable.
2423 const MachineFunction &MF = DAG.getMachineFunction();
2424 const Function *CallerF = DAG.getMachineFunction().getFunction();
2425 CallingConv::ID CallerCC = CallerF->getCallingConv();
2426 bool CCMatch = CallerCC == CalleeCC;
2428 if (GuaranteedTailCallOpt) {
2429 if (IsTailCallConvention(CalleeCC) && CCMatch)
2434 // Look for obvious safe cases to perform tail call optimization that do not
2435 // require ABI changes. This is what gcc calls sibcall.
2437 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2438 // emit a special epilogue.
2439 if (RegInfo->needsStackRealignment(MF))
2442 // Do not sibcall optimize vararg calls unless the call site is not passing
2444 if (isVarArg && !Outs.empty())
2447 // Also avoid sibcall optimization if either caller or callee uses struct
2448 // return semantics.
2449 if (isCalleeStructRet || isCallerStructRet)
2452 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2453 // Therefore if it's not used by the call it is not safe to optimize this into
2455 bool Unused = false;
2456 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2463 SmallVector<CCValAssign, 16> RVLocs;
2464 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2465 RVLocs, *DAG.getContext());
2466 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2467 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2468 CCValAssign &VA = RVLocs[i];
2469 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2474 // If the calling conventions do not match, then we'd better make sure the
2475 // results are returned in the same way as what the caller expects.
2477 SmallVector<CCValAssign, 16> RVLocs1;
2478 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2479 RVLocs1, *DAG.getContext());
2480 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2482 SmallVector<CCValAssign, 16> RVLocs2;
2483 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2484 RVLocs2, *DAG.getContext());
2485 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2487 if (RVLocs1.size() != RVLocs2.size())
2489 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2490 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2492 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2494 if (RVLocs1[i].isRegLoc()) {
2495 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2498 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2504 // If the callee takes no arguments then go on to check the results of the
2506 if (!Outs.empty()) {
2507 // Check if stack adjustment is needed. For now, do not do this if any
2508 // argument is passed on the stack.
2509 SmallVector<CCValAssign, 16> ArgLocs;
2510 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2511 ArgLocs, *DAG.getContext());
2512 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2513 if (CCInfo.getNextStackOffset()) {
2514 MachineFunction &MF = DAG.getMachineFunction();
2515 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2517 if (Subtarget->isTargetWin64())
2518 // Win64 ABI has additional complications.
2521 // Check if the arguments are already laid out in the right way as
2522 // the caller's fixed stack objects.
2523 MachineFrameInfo *MFI = MF.getFrameInfo();
2524 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2525 const X86InstrInfo *TII =
2526 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2527 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2528 CCValAssign &VA = ArgLocs[i];
2529 SDValue Arg = OutVals[i];
2530 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2531 if (VA.getLocInfo() == CCValAssign::Indirect)
2533 if (!VA.isRegLoc()) {
2534 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2541 // If the tailcall address may be in a register, then make sure it's
2542 // possible to register allocate for it. In 32-bit, the call address can
2543 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2544 // callee-saved registers are restored. These happen to be the same
2545 // registers used to pass 'inreg' arguments so watch out for those.
2546 if (!Subtarget->is64Bit() &&
2547 !isa<GlobalAddressSDNode>(Callee) &&
2548 !isa<ExternalSymbolSDNode>(Callee)) {
2549 unsigned NumInRegs = 0;
2550 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2551 CCValAssign &VA = ArgLocs[i];
2554 unsigned Reg = VA.getLocReg();
2557 case X86::EAX: case X86::EDX: case X86::ECX:
2558 if (++NumInRegs == 3)
2570 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2571 return X86::createFastISel(funcInfo);
2575 //===----------------------------------------------------------------------===//
2576 // Other Lowering Hooks
2577 //===----------------------------------------------------------------------===//
2579 static bool MayFoldLoad(SDValue Op) {
2580 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2583 static bool MayFoldIntoStore(SDValue Op) {
2584 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2587 static bool isTargetShuffle(unsigned Opcode) {
2589 default: return false;
2590 case X86ISD::PSHUFD:
2591 case X86ISD::PSHUFHW:
2592 case X86ISD::PSHUFLW:
2593 case X86ISD::SHUFPD:
2594 case X86ISD::SHUFPS:
2595 case X86ISD::MOVLHPS:
2596 case X86ISD::MOVLHPD:
2597 case X86ISD::MOVHLPS:
2598 case X86ISD::MOVLPS:
2599 case X86ISD::MOVLPD:
2600 case X86ISD::MOVSHDUP:
2601 case X86ISD::MOVSLDUP:
2604 case X86ISD::PUNPCKLDQ:
2610 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2611 SDValue V1, SelectionDAG &DAG) {
2613 default: llvm_unreachable("Unknown x86 shuffle node");
2614 case X86ISD::MOVSHDUP:
2615 case X86ISD::MOVSLDUP:
2616 return DAG.getNode(Opc, dl, VT, V1);
2622 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2623 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2625 default: llvm_unreachable("Unknown x86 shuffle node");
2626 case X86ISD::PSHUFD:
2627 case X86ISD::PSHUFHW:
2628 case X86ISD::PSHUFLW:
2629 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2635 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2636 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2638 default: llvm_unreachable("Unknown x86 shuffle node");
2639 case X86ISD::SHUFPD:
2640 case X86ISD::SHUFPS:
2641 return DAG.getNode(Opc, dl, VT, V1, V2,
2642 DAG.getConstant(TargetMask, MVT::i8));
2647 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2648 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2650 default: llvm_unreachable("Unknown x86 shuffle node");
2651 case X86ISD::MOVLHPS:
2652 case X86ISD::MOVLHPD:
2653 case X86ISD::MOVHLPS:
2654 case X86ISD::MOVLPS:
2655 case X86ISD::MOVLPD:
2658 case X86ISD::PUNPCKLDQ:
2659 return DAG.getNode(Opc, dl, VT, V1, V2);
2664 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2665 MachineFunction &MF = DAG.getMachineFunction();
2666 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2667 int ReturnAddrIndex = FuncInfo->getRAIndex();
2669 if (ReturnAddrIndex == 0) {
2670 // Set up a frame object for the return address.
2671 uint64_t SlotSize = TD->getPointerSize();
2672 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2674 FuncInfo->setRAIndex(ReturnAddrIndex);
2677 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2681 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2682 bool hasSymbolicDisplacement) {
2683 // Offset should fit into 32 bit immediate field.
2684 if (!isInt<32>(Offset))
2687 // If we don't have a symbolic displacement - we don't have any extra
2689 if (!hasSymbolicDisplacement)
2692 // FIXME: Some tweaks might be needed for medium code model.
2693 if (M != CodeModel::Small && M != CodeModel::Kernel)
2696 // For small code model we assume that latest object is 16MB before end of 31
2697 // bits boundary. We may also accept pretty large negative constants knowing
2698 // that all objects are in the positive half of address space.
2699 if (M == CodeModel::Small && Offset < 16*1024*1024)
2702 // For kernel code model we know that all object resist in the negative half
2703 // of 32bits address space. We may not accept negative offsets, since they may
2704 // be just off and we may accept pretty large positive ones.
2705 if (M == CodeModel::Kernel && Offset > 0)
2711 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2712 /// specific condition code, returning the condition code and the LHS/RHS of the
2713 /// comparison to make.
2714 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2715 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2717 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2718 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2719 // X > -1 -> X == 0, jump !sign.
2720 RHS = DAG.getConstant(0, RHS.getValueType());
2721 return X86::COND_NS;
2722 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2723 // X < 0 -> X == 0, jump on sign.
2725 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2727 RHS = DAG.getConstant(0, RHS.getValueType());
2728 return X86::COND_LE;
2732 switch (SetCCOpcode) {
2733 default: llvm_unreachable("Invalid integer condition!");
2734 case ISD::SETEQ: return X86::COND_E;
2735 case ISD::SETGT: return X86::COND_G;
2736 case ISD::SETGE: return X86::COND_GE;
2737 case ISD::SETLT: return X86::COND_L;
2738 case ISD::SETLE: return X86::COND_LE;
2739 case ISD::SETNE: return X86::COND_NE;
2740 case ISD::SETULT: return X86::COND_B;
2741 case ISD::SETUGT: return X86::COND_A;
2742 case ISD::SETULE: return X86::COND_BE;
2743 case ISD::SETUGE: return X86::COND_AE;
2747 // First determine if it is required or is profitable to flip the operands.
2749 // If LHS is a foldable load, but RHS is not, flip the condition.
2750 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2751 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2752 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2753 std::swap(LHS, RHS);
2756 switch (SetCCOpcode) {
2762 std::swap(LHS, RHS);
2766 // On a floating point condition, the flags are set as follows:
2768 // 0 | 0 | 0 | X > Y
2769 // 0 | 0 | 1 | X < Y
2770 // 1 | 0 | 0 | X == Y
2771 // 1 | 1 | 1 | unordered
2772 switch (SetCCOpcode) {
2773 default: llvm_unreachable("Condcode should be pre-legalized away");
2775 case ISD::SETEQ: return X86::COND_E;
2776 case ISD::SETOLT: // flipped
2778 case ISD::SETGT: return X86::COND_A;
2779 case ISD::SETOLE: // flipped
2781 case ISD::SETGE: return X86::COND_AE;
2782 case ISD::SETUGT: // flipped
2784 case ISD::SETLT: return X86::COND_B;
2785 case ISD::SETUGE: // flipped
2787 case ISD::SETLE: return X86::COND_BE;
2789 case ISD::SETNE: return X86::COND_NE;
2790 case ISD::SETUO: return X86::COND_P;
2791 case ISD::SETO: return X86::COND_NP;
2793 case ISD::SETUNE: return X86::COND_INVALID;
2797 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2798 /// code. Current x86 isa includes the following FP cmov instructions:
2799 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2800 static bool hasFPCMov(unsigned X86CC) {
2816 /// isFPImmLegal - Returns true if the target can instruction select the
2817 /// specified FP immediate natively. If false, the legalizer will
2818 /// materialize the FP immediate as a load from a constant pool.
2819 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2820 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2821 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2827 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2828 /// the specified range (L, H].
2829 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2830 return (Val < 0) || (Val >= Low && Val < Hi);
2833 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2834 /// specified value.
2835 static bool isUndefOrEqual(int Val, int CmpVal) {
2836 if (Val < 0 || Val == CmpVal)
2841 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2842 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2843 /// the second operand.
2844 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2845 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2846 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2847 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2848 return (Mask[0] < 2 && Mask[1] < 2);
2852 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2853 SmallVector<int, 8> M;
2855 return ::isPSHUFDMask(M, N->getValueType(0));
2858 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2859 /// is suitable for input to PSHUFHW.
2860 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2861 if (VT != MVT::v8i16)
2864 // Lower quadword copied in order or undef.
2865 for (int i = 0; i != 4; ++i)
2866 if (Mask[i] >= 0 && Mask[i] != i)
2869 // Upper quadword shuffled.
2870 for (int i = 4; i != 8; ++i)
2871 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2877 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2878 SmallVector<int, 8> M;
2880 return ::isPSHUFHWMask(M, N->getValueType(0));
2883 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2884 /// is suitable for input to PSHUFLW.
2885 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2886 if (VT != MVT::v8i16)
2889 // Upper quadword copied in order.
2890 for (int i = 4; i != 8; ++i)
2891 if (Mask[i] >= 0 && Mask[i] != i)
2894 // Lower quadword shuffled.
2895 for (int i = 0; i != 4; ++i)
2902 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2903 SmallVector<int, 8> M;
2905 return ::isPSHUFLWMask(M, N->getValueType(0));
2908 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2909 /// is suitable for input to PALIGNR.
2910 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2912 int i, e = VT.getVectorNumElements();
2914 // Do not handle v2i64 / v2f64 shuffles with palignr.
2915 if (e < 4 || !hasSSSE3)
2918 for (i = 0; i != e; ++i)
2922 // All undef, not a palignr.
2926 // Determine if it's ok to perform a palignr with only the LHS, since we
2927 // don't have access to the actual shuffle elements to see if RHS is undef.
2928 bool Unary = Mask[i] < (int)e;
2929 bool NeedsUnary = false;
2931 int s = Mask[i] - i;
2933 // Check the rest of the elements to see if they are consecutive.
2934 for (++i; i != e; ++i) {
2939 Unary = Unary && (m < (int)e);
2940 NeedsUnary = NeedsUnary || (m < s);
2942 if (NeedsUnary && !Unary)
2944 if (Unary && m != ((s+i) & (e-1)))
2946 if (!Unary && m != (s+i))
2952 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2953 SmallVector<int, 8> M;
2955 return ::isPALIGNRMask(M, N->getValueType(0), true);
2958 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2959 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2960 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2961 int NumElems = VT.getVectorNumElements();
2962 if (NumElems != 2 && NumElems != 4)
2965 int Half = NumElems / 2;
2966 for (int i = 0; i < Half; ++i)
2967 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2969 for (int i = Half; i < NumElems; ++i)
2970 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2976 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2977 SmallVector<int, 8> M;
2979 return ::isSHUFPMask(M, N->getValueType(0));
2982 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2983 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2984 /// half elements to come from vector 1 (which would equal the dest.) and
2985 /// the upper half to come from vector 2.
2986 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2987 int NumElems = VT.getVectorNumElements();
2989 if (NumElems != 2 && NumElems != 4)
2992 int Half = NumElems / 2;
2993 for (int i = 0; i < Half; ++i)
2994 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2996 for (int i = Half; i < NumElems; ++i)
2997 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3002 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3003 SmallVector<int, 8> M;
3005 return isCommutedSHUFPMask(M, N->getValueType(0));
3008 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3009 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3010 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3011 if (N->getValueType(0).getVectorNumElements() != 4)
3014 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3015 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3016 isUndefOrEqual(N->getMaskElt(1), 7) &&
3017 isUndefOrEqual(N->getMaskElt(2), 2) &&
3018 isUndefOrEqual(N->getMaskElt(3), 3);
3021 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3022 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3024 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3025 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3030 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3031 isUndefOrEqual(N->getMaskElt(1), 3) &&
3032 isUndefOrEqual(N->getMaskElt(2), 2) &&
3033 isUndefOrEqual(N->getMaskElt(3), 3);
3036 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3037 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3038 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3039 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3041 if (NumElems != 2 && NumElems != 4)
3044 for (unsigned i = 0; i < NumElems/2; ++i)
3045 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3048 for (unsigned i = NumElems/2; i < NumElems; ++i)
3049 if (!isUndefOrEqual(N->getMaskElt(i), i))
3055 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3056 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3057 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3058 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3060 if (NumElems != 2 && NumElems != 4)
3063 for (unsigned i = 0; i < NumElems/2; ++i)
3064 if (!isUndefOrEqual(N->getMaskElt(i), i))
3067 for (unsigned i = 0; i < NumElems/2; ++i)
3068 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3074 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3075 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3076 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3077 bool V2IsSplat = false) {
3078 int NumElts = VT.getVectorNumElements();
3079 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3082 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3084 int BitI1 = Mask[i+1];
3085 if (!isUndefOrEqual(BitI, j))
3088 if (!isUndefOrEqual(BitI1, NumElts))
3091 if (!isUndefOrEqual(BitI1, j + NumElts))
3098 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3099 SmallVector<int, 8> M;
3101 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3104 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3105 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3106 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3107 bool V2IsSplat = false) {
3108 int NumElts = VT.getVectorNumElements();
3109 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3112 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3114 int BitI1 = Mask[i+1];
3115 if (!isUndefOrEqual(BitI, j + NumElts/2))
3118 if (isUndefOrEqual(BitI1, NumElts))
3121 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3128 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3129 SmallVector<int, 8> M;
3131 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3134 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3135 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3137 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3138 int NumElems = VT.getVectorNumElements();
3139 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3142 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3144 int BitI1 = Mask[i+1];
3145 if (!isUndefOrEqual(BitI, j))
3147 if (!isUndefOrEqual(BitI1, j))
3153 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3154 SmallVector<int, 8> M;
3156 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3159 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3160 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3162 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3163 int NumElems = VT.getVectorNumElements();
3164 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3167 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3169 int BitI1 = Mask[i+1];
3170 if (!isUndefOrEqual(BitI, j))
3172 if (!isUndefOrEqual(BitI1, j))
3178 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3179 SmallVector<int, 8> M;
3181 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3184 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3185 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3186 /// MOVSD, and MOVD, i.e. setting the lowest element.
3187 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3188 if (VT.getVectorElementType().getSizeInBits() < 32)
3191 int NumElts = VT.getVectorNumElements();
3193 if (!isUndefOrEqual(Mask[0], NumElts))
3196 for (int i = 1; i < NumElts; ++i)
3197 if (!isUndefOrEqual(Mask[i], i))
3203 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3204 SmallVector<int, 8> M;
3206 return ::isMOVLMask(M, N->getValueType(0));
3209 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3210 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3211 /// element of vector 2 and the other elements to come from vector 1 in order.
3212 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3213 bool V2IsSplat = false, bool V2IsUndef = false) {
3214 int NumOps = VT.getVectorNumElements();
3215 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3218 if (!isUndefOrEqual(Mask[0], 0))
3221 for (int i = 1; i < NumOps; ++i)
3222 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3223 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3224 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3230 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3231 bool V2IsUndef = false) {
3232 SmallVector<int, 8> M;
3234 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3237 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3238 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3239 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3240 if (N->getValueType(0).getVectorNumElements() != 4)
3243 // Expect 1, 1, 3, 3
3244 for (unsigned i = 0; i < 2; ++i) {
3245 int Elt = N->getMaskElt(i);
3246 if (Elt >= 0 && Elt != 1)
3251 for (unsigned i = 2; i < 4; ++i) {
3252 int Elt = N->getMaskElt(i);
3253 if (Elt >= 0 && Elt != 3)
3258 // Don't use movshdup if it can be done with a shufps.
3259 // FIXME: verify that matching u, u, 3, 3 is what we want.
3263 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3264 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3265 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3266 if (N->getValueType(0).getVectorNumElements() != 4)
3269 // Expect 0, 0, 2, 2
3270 for (unsigned i = 0; i < 2; ++i)
3271 if (N->getMaskElt(i) > 0)
3275 for (unsigned i = 2; i < 4; ++i) {
3276 int Elt = N->getMaskElt(i);
3277 if (Elt >= 0 && Elt != 2)
3282 // Don't use movsldup if it can be done with a shufps.
3286 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3287 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3288 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3289 int e = N->getValueType(0).getVectorNumElements() / 2;
3291 for (int i = 0; i < e; ++i)
3292 if (!isUndefOrEqual(N->getMaskElt(i), i))
3294 for (int i = 0; i < e; ++i)
3295 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3300 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3301 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3302 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3303 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3304 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3306 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3308 for (int i = 0; i < NumOperands; ++i) {
3309 int Val = SVOp->getMaskElt(NumOperands-i-1);
3310 if (Val < 0) Val = 0;
3311 if (Val >= NumOperands) Val -= NumOperands;
3313 if (i != NumOperands - 1)
3319 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3320 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3321 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3322 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3324 // 8 nodes, but we only care about the last 4.
3325 for (unsigned i = 7; i >= 4; --i) {
3326 int Val = SVOp->getMaskElt(i);
3335 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3336 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3337 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3338 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3340 // 8 nodes, but we only care about the first 4.
3341 for (int i = 3; i >= 0; --i) {
3342 int Val = SVOp->getMaskElt(i);
3351 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3352 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3353 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3354 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3355 EVT VVT = N->getValueType(0);
3356 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3360 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3361 Val = SVOp->getMaskElt(i);
3365 return (Val - i) * EltSize;
3368 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3370 bool X86::isZeroNode(SDValue Elt) {
3371 return ((isa<ConstantSDNode>(Elt) &&
3372 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3373 (isa<ConstantFPSDNode>(Elt) &&
3374 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3377 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3378 /// their permute mask.
3379 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3380 SelectionDAG &DAG) {
3381 EVT VT = SVOp->getValueType(0);
3382 unsigned NumElems = VT.getVectorNumElements();
3383 SmallVector<int, 8> MaskVec;
3385 for (unsigned i = 0; i != NumElems; ++i) {
3386 int idx = SVOp->getMaskElt(i);
3388 MaskVec.push_back(idx);
3389 else if (idx < (int)NumElems)
3390 MaskVec.push_back(idx + NumElems);
3392 MaskVec.push_back(idx - NumElems);
3394 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3395 SVOp->getOperand(0), &MaskVec[0]);
3398 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3399 /// the two vector operands have swapped position.
3400 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3401 unsigned NumElems = VT.getVectorNumElements();
3402 for (unsigned i = 0; i != NumElems; ++i) {
3406 else if (idx < (int)NumElems)
3407 Mask[i] = idx + NumElems;
3409 Mask[i] = idx - NumElems;
3413 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3414 /// match movhlps. The lower half elements should come from upper half of
3415 /// V1 (and in order), and the upper half elements should come from the upper
3416 /// half of V2 (and in order).
3417 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3418 if (Op->getValueType(0).getVectorNumElements() != 4)
3420 for (unsigned i = 0, e = 2; i != e; ++i)
3421 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3423 for (unsigned i = 2; i != 4; ++i)
3424 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3429 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3430 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3432 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3433 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3435 N = N->getOperand(0).getNode();
3436 if (!ISD::isNON_EXTLoad(N))
3439 *LD = cast<LoadSDNode>(N);
3443 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3444 /// match movlp{s|d}. The lower half elements should come from lower half of
3445 /// V1 (and in order), and the upper half elements should come from the upper
3446 /// half of V2 (and in order). And since V1 will become the source of the
3447 /// MOVLP, it must be either a vector load or a scalar load to vector.
3448 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3449 ShuffleVectorSDNode *Op) {
3450 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3452 // Is V2 is a vector load, don't do this transformation. We will try to use
3453 // load folding shufps op.
3454 if (ISD::isNON_EXTLoad(V2))
3457 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3459 if (NumElems != 2 && NumElems != 4)
3461 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3462 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3464 for (unsigned i = NumElems/2; i != NumElems; ++i)
3465 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3470 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3472 static bool isSplatVector(SDNode *N) {
3473 if (N->getOpcode() != ISD::BUILD_VECTOR)
3476 SDValue SplatValue = N->getOperand(0);
3477 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3478 if (N->getOperand(i) != SplatValue)
3483 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3484 /// to an zero vector.
3485 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3486 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3487 SDValue V1 = N->getOperand(0);
3488 SDValue V2 = N->getOperand(1);
3489 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3490 for (unsigned i = 0; i != NumElems; ++i) {
3491 int Idx = N->getMaskElt(i);
3492 if (Idx >= (int)NumElems) {
3493 unsigned Opc = V2.getOpcode();
3494 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3496 if (Opc != ISD::BUILD_VECTOR ||
3497 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3499 } else if (Idx >= 0) {
3500 unsigned Opc = V1.getOpcode();
3501 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3503 if (Opc != ISD::BUILD_VECTOR ||
3504 !X86::isZeroNode(V1.getOperand(Idx)))
3511 /// getZeroVector - Returns a vector of specified type with all zero elements.
3513 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3515 assert(VT.isVector() && "Expected a vector type");
3517 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted
3518 // to their dest type. This ensures they get CSE'd.
3520 if (VT.getSizeInBits() == 64) { // MMX
3521 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3522 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3523 } else if (VT.getSizeInBits() == 128) {
3524 if (HasSSE2) { // SSE2
3525 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3526 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3528 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3529 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3531 } else if (VT.getSizeInBits() == 256) { // AVX
3532 // 256-bit logic and arithmetic instructions in AVX are
3533 // all floating-point, no support for integer ops. Default
3534 // to emitting fp zeroed vectors then.
3535 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3536 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3537 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3539 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3542 /// getOnesVector - Returns a vector of specified type with all bits set.
3544 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3545 assert(VT.isVector() && "Expected a vector type");
3547 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3548 // type. This ensures they get CSE'd.
3549 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3551 if (VT.getSizeInBits() == 64) // MMX
3552 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3554 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3555 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3559 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3560 /// that point to V2 points to its first element.
3561 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3562 EVT VT = SVOp->getValueType(0);
3563 unsigned NumElems = VT.getVectorNumElements();
3565 bool Changed = false;
3566 SmallVector<int, 8> MaskVec;
3567 SVOp->getMask(MaskVec);
3569 for (unsigned i = 0; i != NumElems; ++i) {
3570 if (MaskVec[i] > (int)NumElems) {
3571 MaskVec[i] = NumElems;
3576 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3577 SVOp->getOperand(1), &MaskVec[0]);
3578 return SDValue(SVOp, 0);
3581 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3582 /// operation of specified width.
3583 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3585 unsigned NumElems = VT.getVectorNumElements();
3586 SmallVector<int, 8> Mask;
3587 Mask.push_back(NumElems);
3588 for (unsigned i = 1; i != NumElems; ++i)
3590 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3593 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3594 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3596 unsigned NumElems = VT.getVectorNumElements();
3597 SmallVector<int, 8> Mask;
3598 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3600 Mask.push_back(i + NumElems);
3602 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3605 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3606 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3608 unsigned NumElems = VT.getVectorNumElements();
3609 unsigned Half = NumElems/2;
3610 SmallVector<int, 8> Mask;
3611 for (unsigned i = 0; i != Half; ++i) {
3612 Mask.push_back(i + Half);
3613 Mask.push_back(i + NumElems + Half);
3615 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3618 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3619 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3620 if (SV->getValueType(0).getVectorNumElements() <= 4)
3621 return SDValue(SV, 0);
3623 EVT PVT = MVT::v4f32;
3624 EVT VT = SV->getValueType(0);
3625 DebugLoc dl = SV->getDebugLoc();
3626 SDValue V1 = SV->getOperand(0);
3627 int NumElems = VT.getVectorNumElements();
3628 int EltNo = SV->getSplatIndex();
3630 // unpack elements to the correct location
3631 while (NumElems > 4) {
3632 if (EltNo < NumElems/2) {
3633 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3635 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3636 EltNo -= NumElems/2;
3641 // Perform the splat.
3642 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3643 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3644 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3645 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3648 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3649 /// vector of zero or undef vector. This produces a shuffle where the low
3650 /// element of V2 is swizzled into the zero/undef vector, landing at element
3651 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3652 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3653 bool isZero, bool HasSSE2,
3654 SelectionDAG &DAG) {
3655 EVT VT = V2.getValueType();
3657 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3658 unsigned NumElems = VT.getVectorNumElements();
3659 SmallVector<int, 16> MaskVec;
3660 for (unsigned i = 0; i != NumElems; ++i)
3661 // If this is the insertion idx, put the low elt of V2 here.
3662 MaskVec.push_back(i == Idx ? NumElems : i);
3663 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3666 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
3667 /// element of the result of the vector shuffle.
3668 SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG) {
3669 SDValue V = SDValue(N, 0);
3670 EVT VT = V.getValueType();
3671 unsigned Opcode = V.getOpcode();
3673 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
3674 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
3675 Index = SV->getMaskElt(Index);
3678 return DAG.getUNDEF(VT.getVectorElementType());
3680 int NumElems = VT.getVectorNumElements();
3681 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
3682 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG);
3685 // Recurse into target specific vector shuffles to find scalars.
3686 if (isTargetShuffle(Opcode)) {
3689 case X86ISD::MOVSD: {
3690 // The index 0 always comes from the first element of the second source,
3691 // this is why MOVSS and MOVSD are used in the first place. The other
3692 // elements come from the other positions of the first source vector.
3693 unsigned OpNum = (Index == 0) ? 1 : 0;
3694 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG);
3697 assert("not implemented for target shuffle node");
3702 // Actual nodes that may contain scalar elements
3703 if (Opcode == ISD::BIT_CONVERT) {
3704 V = V.getOperand(0);
3705 EVT SrcVT = V.getValueType();
3706 unsigned NumElems = VT.getVectorNumElements();
3708 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
3712 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
3713 return (Index == 0) ? V.getOperand(0)
3714 : DAG.getUNDEF(VT.getVectorElementType());
3716 if (V.getOpcode() == ISD::BUILD_VECTOR)
3717 return V.getOperand(Index);
3722 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
3723 /// shuffle operation which come from a consecutively from a zero. The
3724 /// search can start in two diferent directions, from left or right.
3726 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
3727 bool ZerosFromLeft, SelectionDAG &DAG) {
3730 while (i < NumElems) {
3731 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
3732 SDValue Elt = getShuffleScalarElt(N, Index, DAG);
3733 if (!(Elt.getNode() &&
3734 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
3742 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
3743 /// MaskE correspond consecutively to elements from one of the vector operands,
3744 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
3746 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
3747 int OpIdx, int NumElems, unsigned &OpNum) {
3748 bool SeenV1 = false;
3749 bool SeenV2 = false;
3751 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
3752 int Idx = SVOp->getMaskElt(i);
3753 // Ignore undef indicies
3762 // Only accept consecutive elements from the same vector
3763 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
3767 OpNum = SeenV1 ? 0 : 1;
3771 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
3772 /// logical left shift of a vector.
3773 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3774 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3775 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3776 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3777 false /* check zeros from right */, DAG);
3783 // Considering the elements in the mask that are not consecutive zeros,
3784 // check if they consecutively come from only one of the source vectors.
3786 // V1 = {X, A, B, C} 0
3788 // vector_shuffle V1, V2 <1, 2, 3, X>
3790 if (!isShuffleMaskConsecutive(SVOp,
3791 0, // Mask Start Index
3792 NumElems-NumZeros-1, // Mask End Index
3793 NumZeros, // Where to start looking in the src vector
3794 NumElems, // Number of elements in vector
3795 OpSrc)) // Which source operand ?
3800 ShVal = SVOp->getOperand(OpSrc);
3804 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
3805 /// logical left shift of a vector.
3806 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3807 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3808 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3809 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3810 true /* check zeros from left */, DAG);
3816 // Considering the elements in the mask that are not consecutive zeros,
3817 // check if they consecutively come from only one of the source vectors.
3819 // 0 { A, B, X, X } = V2
3821 // vector_shuffle V1, V2 <X, X, 4, 5>
3823 if (!isShuffleMaskConsecutive(SVOp,
3824 NumZeros, // Mask Start Index
3825 NumElems-1, // Mask End Index
3826 0, // Where to start looking in the src vector
3827 NumElems, // Number of elements in vector
3828 OpSrc)) // Which source operand ?
3833 ShVal = SVOp->getOperand(OpSrc);
3837 /// isVectorShift - Returns true if the shuffle can be implemented as a
3838 /// logical left or right shift of a vector.
3839 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3840 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3841 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
3842 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
3848 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3850 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3851 unsigned NumNonZero, unsigned NumZero,
3853 const TargetLowering &TLI) {
3857 DebugLoc dl = Op.getDebugLoc();
3860 for (unsigned i = 0; i < 16; ++i) {
3861 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3862 if (ThisIsNonZero && First) {
3864 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3866 V = DAG.getUNDEF(MVT::v8i16);
3871 SDValue ThisElt(0, 0), LastElt(0, 0);
3872 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3873 if (LastIsNonZero) {
3874 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3875 MVT::i16, Op.getOperand(i-1));
3877 if (ThisIsNonZero) {
3878 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3879 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3880 ThisElt, DAG.getConstant(8, MVT::i8));
3882 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3886 if (ThisElt.getNode())
3887 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3888 DAG.getIntPtrConstant(i/2));
3892 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3895 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3897 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3898 unsigned NumNonZero, unsigned NumZero,
3900 const TargetLowering &TLI) {
3904 DebugLoc dl = Op.getDebugLoc();
3907 for (unsigned i = 0; i < 8; ++i) {
3908 bool isNonZero = (NonZeros & (1 << i)) != 0;
3912 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3914 V = DAG.getUNDEF(MVT::v8i16);
3917 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3918 MVT::v8i16, V, Op.getOperand(i),
3919 DAG.getIntPtrConstant(i));
3926 /// getVShift - Return a vector logical shift node.
3928 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3929 unsigned NumBits, SelectionDAG &DAG,
3930 const TargetLowering &TLI, DebugLoc dl) {
3931 bool isMMX = VT.getSizeInBits() == 64;
3932 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3933 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3934 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3935 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3936 DAG.getNode(Opc, dl, ShVT, SrcOp,
3937 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3941 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3942 SelectionDAG &DAG) const {
3944 // Check if the scalar load can be widened into a vector load. And if
3945 // the address is "base + cst" see if the cst can be "absorbed" into
3946 // the shuffle mask.
3947 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3948 SDValue Ptr = LD->getBasePtr();
3949 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3951 EVT PVT = LD->getValueType(0);
3952 if (PVT != MVT::i32 && PVT != MVT::f32)
3957 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3958 FI = FINode->getIndex();
3960 } else if (Ptr.getOpcode() == ISD::ADD &&
3961 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3962 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3963 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3964 Offset = Ptr.getConstantOperandVal(1);
3965 Ptr = Ptr.getOperand(0);
3970 SDValue Chain = LD->getChain();
3971 // Make sure the stack object alignment is at least 16.
3972 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3973 if (DAG.InferPtrAlignment(Ptr) < 16) {
3974 if (MFI->isFixedObjectIndex(FI)) {
3975 // Can't change the alignment. FIXME: It's possible to compute
3976 // the exact stack offset and reference FI + adjust offset instead.
3977 // If someone *really* cares about this. That's the way to implement it.
3980 MFI->setObjectAlignment(FI, 16);
3984 // (Offset % 16) must be multiple of 4. Then address is then
3985 // Ptr + (Offset & ~15).
3988 if ((Offset % 16) & 3)
3990 int64_t StartOffset = Offset & ~15;
3992 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3993 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3995 int EltNo = (Offset - StartOffset) >> 2;
3996 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3997 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3998 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
4000 // Canonicalize it to a v4i32 shuffle.
4001 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
4002 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4003 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4004 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
4010 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4011 /// vector of type 'VT', see if the elements can be replaced by a single large
4012 /// load which has the same value as a build_vector whose operands are 'elts'.
4014 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4016 /// FIXME: we'd also like to handle the case where the last elements are zero
4017 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4018 /// There's even a handy isZeroNode for that purpose.
4019 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4020 DebugLoc &dl, SelectionDAG &DAG) {
4021 EVT EltVT = VT.getVectorElementType();
4022 unsigned NumElems = Elts.size();
4024 LoadSDNode *LDBase = NULL;
4025 unsigned LastLoadedElt = -1U;
4027 // For each element in the initializer, see if we've found a load or an undef.
4028 // If we don't find an initial load element, or later load elements are
4029 // non-consecutive, bail out.
4030 for (unsigned i = 0; i < NumElems; ++i) {
4031 SDValue Elt = Elts[i];
4033 if (!Elt.getNode() ||
4034 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4037 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4039 LDBase = cast<LoadSDNode>(Elt.getNode());
4043 if (Elt.getOpcode() == ISD::UNDEF)
4046 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4047 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4052 // If we have found an entire vector of loads and undefs, then return a large
4053 // load of the entire vector width starting at the base pointer. If we found
4054 // consecutive loads for the low half, generate a vzext_load node.
4055 if (LastLoadedElt == NumElems - 1) {
4056 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4057 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
4058 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
4059 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4060 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
4061 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
4062 LDBase->isVolatile(), LDBase->isNonTemporal(),
4063 LDBase->getAlignment());
4064 } else if (NumElems == 4 && LastLoadedElt == 1) {
4065 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4066 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4067 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
4068 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
4074 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4075 DebugLoc dl = Op.getDebugLoc();
4076 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
4077 // All one's are handled with pcmpeqd. In AVX, zero's are handled with
4078 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
4079 // is present, so AllOnes is ignored.
4080 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4081 (Op.getValueType().getSizeInBits() != 256 &&
4082 ISD::isBuildVectorAllOnes(Op.getNode()))) {
4083 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
4084 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4085 // eliminated on x86-32 hosts.
4086 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
4089 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4090 return getOnesVector(Op.getValueType(), DAG, dl);
4091 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4094 EVT VT = Op.getValueType();
4095 EVT ExtVT = VT.getVectorElementType();
4096 unsigned EVTBits = ExtVT.getSizeInBits();
4098 unsigned NumElems = Op.getNumOperands();
4099 unsigned NumZero = 0;
4100 unsigned NumNonZero = 0;
4101 unsigned NonZeros = 0;
4102 bool IsAllConstants = true;
4103 SmallSet<SDValue, 8> Values;
4104 for (unsigned i = 0; i < NumElems; ++i) {
4105 SDValue Elt = Op.getOperand(i);
4106 if (Elt.getOpcode() == ISD::UNDEF)
4109 if (Elt.getOpcode() != ISD::Constant &&
4110 Elt.getOpcode() != ISD::ConstantFP)
4111 IsAllConstants = false;
4112 if (X86::isZeroNode(Elt))
4115 NonZeros |= (1 << i);
4120 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4121 if (NumNonZero == 0)
4122 return DAG.getUNDEF(VT);
4124 // Special case for single non-zero, non-undef, element.
4125 if (NumNonZero == 1) {
4126 unsigned Idx = CountTrailingZeros_32(NonZeros);
4127 SDValue Item = Op.getOperand(Idx);
4129 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4130 // the value are obviously zero, truncate the value to i32 and do the
4131 // insertion that way. Only do this if the value is non-constant or if the
4132 // value is a constant being inserted into element 0. It is cheaper to do
4133 // a constant pool load than it is to do a movd + shuffle.
4134 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4135 (!IsAllConstants || Idx == 0)) {
4136 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4137 // Handle MMX and SSE both.
4138 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
4139 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
4141 // Truncate the value (which may itself be a constant) to i32, and
4142 // convert it to a vector with movd (S2V+shuffle to zero extend).
4143 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4144 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4145 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4146 Subtarget->hasSSE2(), DAG);
4148 // Now we have our 32-bit value zero extended in the low element of
4149 // a vector. If Idx != 0, swizzle it into place.
4151 SmallVector<int, 4> Mask;
4152 Mask.push_back(Idx);
4153 for (unsigned i = 1; i != VecElts; ++i)
4155 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4156 DAG.getUNDEF(Item.getValueType()),
4159 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
4163 // If we have a constant or non-constant insertion into the low element of
4164 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4165 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4166 // depending on what the source datatype is.
4169 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4170 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4171 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4172 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4173 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4174 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4176 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4177 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4178 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
4179 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4180 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4181 Subtarget->hasSSE2(), DAG);
4182 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
4186 // Is it a vector logical left shift?
4187 if (NumElems == 2 && Idx == 1 &&
4188 X86::isZeroNode(Op.getOperand(0)) &&
4189 !X86::isZeroNode(Op.getOperand(1))) {
4190 unsigned NumBits = VT.getSizeInBits();
4191 return getVShift(true, VT,
4192 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4193 VT, Op.getOperand(1)),
4194 NumBits/2, DAG, *this, dl);
4197 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4200 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4201 // is a non-constant being inserted into an element other than the low one,
4202 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4203 // movd/movss) to move this into the low element, then shuffle it into
4205 if (EVTBits == 32) {
4206 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4208 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4209 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4210 Subtarget->hasSSE2(), DAG);
4211 SmallVector<int, 8> MaskVec;
4212 for (unsigned i = 0; i < NumElems; i++)
4213 MaskVec.push_back(i == Idx ? 0 : 1);
4214 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4218 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4219 if (Values.size() == 1) {
4220 if (EVTBits == 32) {
4221 // Instead of a shuffle like this:
4222 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4223 // Check if it's possible to issue this instead.
4224 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4225 unsigned Idx = CountTrailingZeros_32(NonZeros);
4226 SDValue Item = Op.getOperand(Idx);
4227 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4228 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4233 // A vector full of immediates; various special cases are already
4234 // handled, so this is best done with a single constant-pool load.
4238 // Let legalizer expand 2-wide build_vectors.
4239 if (EVTBits == 64) {
4240 if (NumNonZero == 1) {
4241 // One half is zero or undef.
4242 unsigned Idx = CountTrailingZeros_32(NonZeros);
4243 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4244 Op.getOperand(Idx));
4245 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4246 Subtarget->hasSSE2(), DAG);
4251 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4252 if (EVTBits == 8 && NumElems == 16) {
4253 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4255 if (V.getNode()) return V;
4258 if (EVTBits == 16 && NumElems == 8) {
4259 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4261 if (V.getNode()) return V;
4264 // If element VT is == 32 bits, turn it into a number of shuffles.
4265 SmallVector<SDValue, 8> V;
4267 if (NumElems == 4 && NumZero > 0) {
4268 for (unsigned i = 0; i < 4; ++i) {
4269 bool isZero = !(NonZeros & (1 << i));
4271 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4273 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4276 for (unsigned i = 0; i < 2; ++i) {
4277 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4280 V[i] = V[i*2]; // Must be a zero vector.
4283 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4286 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4289 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4294 SmallVector<int, 8> MaskVec;
4295 bool Reverse = (NonZeros & 0x3) == 2;
4296 for (unsigned i = 0; i < 2; ++i)
4297 MaskVec.push_back(Reverse ? 1-i : i);
4298 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4299 for (unsigned i = 0; i < 2; ++i)
4300 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4301 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4304 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4305 // Check for a build vector of consecutive loads.
4306 for (unsigned i = 0; i < NumElems; ++i)
4307 V[i] = Op.getOperand(i);
4309 // Check for elements which are consecutive loads.
4310 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4314 // For SSE 4.1, use insertps to put the high elements into the low element.
4315 if (getSubtarget()->hasSSE41()) {
4317 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4318 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4320 Result = DAG.getUNDEF(VT);
4322 for (unsigned i = 1; i < NumElems; ++i) {
4323 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4324 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4325 Op.getOperand(i), DAG.getIntPtrConstant(i));
4330 // Otherwise, expand into a number of unpckl*, start by extending each of
4331 // our (non-undef) elements to the full vector width with the element in the
4332 // bottom slot of the vector (which generates no code for SSE).
4333 for (unsigned i = 0; i < NumElems; ++i) {
4334 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4335 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4337 V[i] = DAG.getUNDEF(VT);
4340 // Next, we iteratively mix elements, e.g. for v4f32:
4341 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4342 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4343 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4344 unsigned EltStride = NumElems >> 1;
4345 while (EltStride != 0) {
4346 for (unsigned i = 0; i < EltStride; ++i) {
4347 // If V[i+EltStride] is undef and this is the first round of mixing,
4348 // then it is safe to just drop this shuffle: V[i] is already in the
4349 // right place, the one element (since it's the first round) being
4350 // inserted as undef can be dropped. This isn't safe for successive
4351 // rounds because they will permute elements within both vectors.
4352 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4353 EltStride == NumElems/2)
4356 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4366 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4367 // We support concatenate two MMX registers and place them in a MMX
4368 // register. This is better than doing a stack convert.
4369 DebugLoc dl = Op.getDebugLoc();
4370 EVT ResVT = Op.getValueType();
4371 assert(Op.getNumOperands() == 2);
4372 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4373 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4375 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4376 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4377 InVec = Op.getOperand(1);
4378 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4379 unsigned NumElts = ResVT.getVectorNumElements();
4380 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4381 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4382 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4384 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4385 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4386 Mask[0] = 0; Mask[1] = 2;
4387 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4389 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4392 // v8i16 shuffles - Prefer shuffles in the following order:
4393 // 1. [all] pshuflw, pshufhw, optional move
4394 // 2. [ssse3] 1 x pshufb
4395 // 3. [ssse3] 2 x pshufb + 1 x por
4396 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4398 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4399 SelectionDAG &DAG) const {
4400 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4401 SDValue V1 = SVOp->getOperand(0);
4402 SDValue V2 = SVOp->getOperand(1);
4403 DebugLoc dl = SVOp->getDebugLoc();
4404 SmallVector<int, 8> MaskVals;
4406 // Determine if more than 1 of the words in each of the low and high quadwords
4407 // of the result come from the same quadword of one of the two inputs. Undef
4408 // mask values count as coming from any quadword, for better codegen.
4409 SmallVector<unsigned, 4> LoQuad(4);
4410 SmallVector<unsigned, 4> HiQuad(4);
4411 BitVector InputQuads(4);
4412 for (unsigned i = 0; i < 8; ++i) {
4413 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4414 int EltIdx = SVOp->getMaskElt(i);
4415 MaskVals.push_back(EltIdx);
4424 InputQuads.set(EltIdx / 4);
4427 int BestLoQuad = -1;
4428 unsigned MaxQuad = 1;
4429 for (unsigned i = 0; i < 4; ++i) {
4430 if (LoQuad[i] > MaxQuad) {
4432 MaxQuad = LoQuad[i];
4436 int BestHiQuad = -1;
4438 for (unsigned i = 0; i < 4; ++i) {
4439 if (HiQuad[i] > MaxQuad) {
4441 MaxQuad = HiQuad[i];
4445 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4446 // of the two input vectors, shuffle them into one input vector so only a
4447 // single pshufb instruction is necessary. If There are more than 2 input
4448 // quads, disable the next transformation since it does not help SSSE3.
4449 bool V1Used = InputQuads[0] || InputQuads[1];
4450 bool V2Used = InputQuads[2] || InputQuads[3];
4451 if (Subtarget->hasSSSE3()) {
4452 if (InputQuads.count() == 2 && V1Used && V2Used) {
4453 BestLoQuad = InputQuads.find_first();
4454 BestHiQuad = InputQuads.find_next(BestLoQuad);
4456 if (InputQuads.count() > 2) {
4462 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4463 // the shuffle mask. If a quad is scored as -1, that means that it contains
4464 // words from all 4 input quadwords.
4466 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4467 SmallVector<int, 8> MaskV;
4468 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4469 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4470 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4471 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4472 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4473 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4475 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4476 // source words for the shuffle, to aid later transformations.
4477 bool AllWordsInNewV = true;
4478 bool InOrder[2] = { true, true };
4479 for (unsigned i = 0; i != 8; ++i) {
4480 int idx = MaskVals[i];
4482 InOrder[i/4] = false;
4483 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4485 AllWordsInNewV = false;
4489 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4490 if (AllWordsInNewV) {
4491 for (int i = 0; i != 8; ++i) {
4492 int idx = MaskVals[i];
4495 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4496 if ((idx != i) && idx < 4)
4498 if ((idx != i) && idx > 3)
4507 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4508 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4509 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4510 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4511 unsigned TargetMask = 0;
4512 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4513 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4514 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4515 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4516 V1 = NewV.getOperand(0);
4517 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4521 // If we have SSSE3, and all words of the result are from 1 input vector,
4522 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4523 // is present, fall back to case 4.
4524 if (Subtarget->hasSSSE3()) {
4525 SmallVector<SDValue,16> pshufbMask;
4527 // If we have elements from both input vectors, set the high bit of the
4528 // shuffle mask element to zero out elements that come from V2 in the V1
4529 // mask, and elements that come from V1 in the V2 mask, so that the two
4530 // results can be OR'd together.
4531 bool TwoInputs = V1Used && V2Used;
4532 for (unsigned i = 0; i != 8; ++i) {
4533 int EltIdx = MaskVals[i] * 2;
4534 if (TwoInputs && (EltIdx >= 16)) {
4535 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4536 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4539 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4540 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4542 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4543 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4544 DAG.getNode(ISD::BUILD_VECTOR, dl,
4545 MVT::v16i8, &pshufbMask[0], 16));
4547 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4549 // Calculate the shuffle mask for the second input, shuffle it, and
4550 // OR it with the first shuffled input.
4552 for (unsigned i = 0; i != 8; ++i) {
4553 int EltIdx = MaskVals[i] * 2;
4555 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4556 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4559 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4560 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4562 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4563 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4564 DAG.getNode(ISD::BUILD_VECTOR, dl,
4565 MVT::v16i8, &pshufbMask[0], 16));
4566 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4567 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4570 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4571 // and update MaskVals with new element order.
4572 BitVector InOrder(8);
4573 if (BestLoQuad >= 0) {
4574 SmallVector<int, 8> MaskV;
4575 for (int i = 0; i != 4; ++i) {
4576 int idx = MaskVals[i];
4578 MaskV.push_back(-1);
4580 } else if ((idx / 4) == BestLoQuad) {
4581 MaskV.push_back(idx & 3);
4584 MaskV.push_back(-1);
4587 for (unsigned i = 4; i != 8; ++i)
4589 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4592 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4593 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4595 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4599 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4600 // and update MaskVals with the new element order.
4601 if (BestHiQuad >= 0) {
4602 SmallVector<int, 8> MaskV;
4603 for (unsigned i = 0; i != 4; ++i)
4605 for (unsigned i = 4; i != 8; ++i) {
4606 int idx = MaskVals[i];
4608 MaskV.push_back(-1);
4610 } else if ((idx / 4) == BestHiQuad) {
4611 MaskV.push_back((idx & 3) + 4);
4614 MaskV.push_back(-1);
4617 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4620 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4621 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
4623 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
4627 // In case BestHi & BestLo were both -1, which means each quadword has a word
4628 // from each of the four input quadwords, calculate the InOrder bitvector now
4629 // before falling through to the insert/extract cleanup.
4630 if (BestLoQuad == -1 && BestHiQuad == -1) {
4632 for (int i = 0; i != 8; ++i)
4633 if (MaskVals[i] < 0 || MaskVals[i] == i)
4637 // The other elements are put in the right place using pextrw and pinsrw.
4638 for (unsigned i = 0; i != 8; ++i) {
4641 int EltIdx = MaskVals[i];
4644 SDValue ExtOp = (EltIdx < 8)
4645 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4646 DAG.getIntPtrConstant(EltIdx))
4647 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4648 DAG.getIntPtrConstant(EltIdx - 8));
4649 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4650 DAG.getIntPtrConstant(i));
4655 // v16i8 shuffles - Prefer shuffles in the following order:
4656 // 1. [ssse3] 1 x pshufb
4657 // 2. [ssse3] 2 x pshufb + 1 x por
4658 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4660 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4662 const X86TargetLowering &TLI) {
4663 SDValue V1 = SVOp->getOperand(0);
4664 SDValue V2 = SVOp->getOperand(1);
4665 DebugLoc dl = SVOp->getDebugLoc();
4666 SmallVector<int, 16> MaskVals;
4667 SVOp->getMask(MaskVals);
4669 // If we have SSSE3, case 1 is generated when all result bytes come from
4670 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4671 // present, fall back to case 3.
4672 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4675 for (unsigned i = 0; i < 16; ++i) {
4676 int EltIdx = MaskVals[i];
4685 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4686 if (TLI.getSubtarget()->hasSSSE3()) {
4687 SmallVector<SDValue,16> pshufbMask;
4689 // If all result elements are from one input vector, then only translate
4690 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4692 // Otherwise, we have elements from both input vectors, and must zero out
4693 // elements that come from V2 in the first mask, and V1 in the second mask
4694 // so that we can OR them together.
4695 bool TwoInputs = !(V1Only || V2Only);
4696 for (unsigned i = 0; i != 16; ++i) {
4697 int EltIdx = MaskVals[i];
4698 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4699 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4702 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4704 // If all the elements are from V2, assign it to V1 and return after
4705 // building the first pshufb.
4708 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4709 DAG.getNode(ISD::BUILD_VECTOR, dl,
4710 MVT::v16i8, &pshufbMask[0], 16));
4714 // Calculate the shuffle mask for the second input, shuffle it, and
4715 // OR it with the first shuffled input.
4717 for (unsigned i = 0; i != 16; ++i) {
4718 int EltIdx = MaskVals[i];
4720 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4723 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4725 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4726 DAG.getNode(ISD::BUILD_VECTOR, dl,
4727 MVT::v16i8, &pshufbMask[0], 16));
4728 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4731 // No SSSE3 - Calculate in place words and then fix all out of place words
4732 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4733 // the 16 different words that comprise the two doublequadword input vectors.
4734 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4735 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4736 SDValue NewV = V2Only ? V2 : V1;
4737 for (int i = 0; i != 8; ++i) {
4738 int Elt0 = MaskVals[i*2];
4739 int Elt1 = MaskVals[i*2+1];
4741 // This word of the result is all undef, skip it.
4742 if (Elt0 < 0 && Elt1 < 0)
4745 // This word of the result is already in the correct place, skip it.
4746 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4748 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4751 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4752 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4755 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4756 // using a single extract together, load it and store it.
4757 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4758 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4759 DAG.getIntPtrConstant(Elt1 / 2));
4760 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4761 DAG.getIntPtrConstant(i));
4765 // If Elt1 is defined, extract it from the appropriate source. If the
4766 // source byte is not also odd, shift the extracted word left 8 bits
4767 // otherwise clear the bottom 8 bits if we need to do an or.
4769 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4770 DAG.getIntPtrConstant(Elt1 / 2));
4771 if ((Elt1 & 1) == 0)
4772 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4773 DAG.getConstant(8, TLI.getShiftAmountTy()));
4775 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4776 DAG.getConstant(0xFF00, MVT::i16));
4778 // If Elt0 is defined, extract it from the appropriate source. If the
4779 // source byte is not also even, shift the extracted word right 8 bits. If
4780 // Elt1 was also defined, OR the extracted values together before
4781 // inserting them in the result.
4783 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4784 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4785 if ((Elt0 & 1) != 0)
4786 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4787 DAG.getConstant(8, TLI.getShiftAmountTy()));
4789 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4790 DAG.getConstant(0x00FF, MVT::i16));
4791 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4794 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4795 DAG.getIntPtrConstant(i));
4797 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4800 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4801 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4802 /// done when every pair / quad of shuffle mask elements point to elements in
4803 /// the right sequence. e.g.
4804 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4806 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4808 const TargetLowering &TLI, DebugLoc dl) {
4809 EVT VT = SVOp->getValueType(0);
4810 SDValue V1 = SVOp->getOperand(0);
4811 SDValue V2 = SVOp->getOperand(1);
4812 unsigned NumElems = VT.getVectorNumElements();
4813 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4814 EVT MaskVT = (NewWidth == 4) ? MVT::v4i16 : MVT::v2i32;
4816 switch (VT.getSimpleVT().SimpleTy) {
4817 default: assert(false && "Unexpected!");
4818 case MVT::v4f32: NewVT = MVT::v2f64; break;
4819 case MVT::v4i32: NewVT = MVT::v2i64; break;
4820 case MVT::v8i16: NewVT = MVT::v4i32; break;
4821 case MVT::v16i8: NewVT = MVT::v4i32; break;
4824 if (NewWidth == 2) {
4830 int Scale = NumElems / NewWidth;
4831 SmallVector<int, 8> MaskVec;
4832 for (unsigned i = 0; i < NumElems; i += Scale) {
4834 for (int j = 0; j < Scale; ++j) {
4835 int EltIdx = SVOp->getMaskElt(i+j);
4839 StartIdx = EltIdx - (EltIdx % Scale);
4840 if (EltIdx != StartIdx + j)
4844 MaskVec.push_back(-1);
4846 MaskVec.push_back(StartIdx / Scale);
4849 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4850 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4851 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4854 /// getVZextMovL - Return a zero-extending vector move low node.
4856 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4857 SDValue SrcOp, SelectionDAG &DAG,
4858 const X86Subtarget *Subtarget, DebugLoc dl) {
4859 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4860 LoadSDNode *LD = NULL;
4861 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4862 LD = dyn_cast<LoadSDNode>(SrcOp);
4864 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4866 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4867 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4868 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4869 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4870 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4872 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4873 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4874 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4875 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4883 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4884 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4885 DAG.getNode(ISD::BIT_CONVERT, dl,
4889 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4892 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4893 SDValue V1 = SVOp->getOperand(0);
4894 SDValue V2 = SVOp->getOperand(1);
4895 DebugLoc dl = SVOp->getDebugLoc();
4896 EVT VT = SVOp->getValueType(0);
4898 SmallVector<std::pair<int, int>, 8> Locs;
4900 SmallVector<int, 8> Mask1(4U, -1);
4901 SmallVector<int, 8> PermMask;
4902 SVOp->getMask(PermMask);
4906 for (unsigned i = 0; i != 4; ++i) {
4907 int Idx = PermMask[i];
4909 Locs[i] = std::make_pair(-1, -1);
4911 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4913 Locs[i] = std::make_pair(0, NumLo);
4917 Locs[i] = std::make_pair(1, NumHi);
4919 Mask1[2+NumHi] = Idx;
4925 if (NumLo <= 2 && NumHi <= 2) {
4926 // If no more than two elements come from either vector. This can be
4927 // implemented with two shuffles. First shuffle gather the elements.
4928 // The second shuffle, which takes the first shuffle as both of its
4929 // vector operands, put the elements into the right order.
4930 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4932 SmallVector<int, 8> Mask2(4U, -1);
4934 for (unsigned i = 0; i != 4; ++i) {
4935 if (Locs[i].first == -1)
4938 unsigned Idx = (i < 2) ? 0 : 4;
4939 Idx += Locs[i].first * 2 + Locs[i].second;
4944 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4945 } else if (NumLo == 3 || NumHi == 3) {
4946 // Otherwise, we must have three elements from one vector, call it X, and
4947 // one element from the other, call it Y. First, use a shufps to build an
4948 // intermediate vector with the one element from Y and the element from X
4949 // that will be in the same half in the final destination (the indexes don't
4950 // matter). Then, use a shufps to build the final vector, taking the half
4951 // containing the element from Y from the intermediate, and the other half
4954 // Normalize it so the 3 elements come from V1.
4955 CommuteVectorShuffleMask(PermMask, VT);
4959 // Find the element from V2.
4961 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4962 int Val = PermMask[HiIndex];
4969 Mask1[0] = PermMask[HiIndex];
4971 Mask1[2] = PermMask[HiIndex^1];
4973 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4976 Mask1[0] = PermMask[0];
4977 Mask1[1] = PermMask[1];
4978 Mask1[2] = HiIndex & 1 ? 6 : 4;
4979 Mask1[3] = HiIndex & 1 ? 4 : 6;
4980 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4982 Mask1[0] = HiIndex & 1 ? 2 : 0;
4983 Mask1[1] = HiIndex & 1 ? 0 : 2;
4984 Mask1[2] = PermMask[2];
4985 Mask1[3] = PermMask[3];
4990 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4994 // Break it into (shuffle shuffle_hi, shuffle_lo).
4996 SmallVector<int,8> LoMask(4U, -1);
4997 SmallVector<int,8> HiMask(4U, -1);
4999 SmallVector<int,8> *MaskPtr = &LoMask;
5000 unsigned MaskIdx = 0;
5003 for (unsigned i = 0; i != 4; ++i) {
5010 int Idx = PermMask[i];
5012 Locs[i] = std::make_pair(-1, -1);
5013 } else if (Idx < 4) {
5014 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5015 (*MaskPtr)[LoIdx] = Idx;
5018 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5019 (*MaskPtr)[HiIdx] = Idx;
5024 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5025 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5026 SmallVector<int, 8> MaskOps;
5027 for (unsigned i = 0; i != 4; ++i) {
5028 if (Locs[i].first == -1) {
5029 MaskOps.push_back(-1);
5031 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5032 MaskOps.push_back(Idx);
5035 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5039 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5041 SDValue V1 = Op.getOperand(0);
5042 SDValue V2 = Op.getOperand(1);
5043 EVT VT = Op.getValueType();
5045 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5047 if (HasSSE2 && VT == MVT::v2f64)
5048 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5051 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5055 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5056 SDValue V1 = Op.getOperand(0);
5057 SDValue V2 = Op.getOperand(1);
5058 EVT VT = Op.getValueType();
5060 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5061 "unsupported shuffle type");
5063 if (V2.getOpcode() == ISD::UNDEF)
5067 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5071 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5072 SDValue V1 = Op.getOperand(0);
5073 SDValue V2 = Op.getOperand(1);
5074 EVT VT = Op.getValueType();
5075 unsigned NumElems = VT.getVectorNumElements();
5077 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5078 // operand of these instructions is only memory, so check if there's a
5079 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5081 bool CanFoldLoad = false;
5085 // Trivial case, when V2 comes from a load.
5086 if (TmpV2.hasOneUse() && TmpV2.getOpcode() == ISD::BIT_CONVERT)
5087 TmpV2 = TmpV2.getOperand(0);
5088 if (TmpV2.hasOneUse() && TmpV2.getOpcode() == ISD::SCALAR_TO_VECTOR)
5089 TmpV2 = TmpV2.getOperand(0);
5090 if (MayFoldLoad(TmpV2))
5093 // When V1 is a load, it can be folded later into a store in isel, example:
5094 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5096 // (MOVLPSmr addr:$src1, VR128:$src2)
5097 // So, recognize this potential and also use MOVLPS or MOVLPD
5098 if (TmpV1.hasOneUse() && TmpV1.getOpcode() == ISD::BIT_CONVERT)
5099 TmpV1 = TmpV1.getOperand(0);
5100 if (MayFoldLoad(TmpV1) && MayFoldIntoStore(Op))
5104 if (HasSSE2 && NumElems == 2)
5105 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5108 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5111 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5112 // movl and movlp will both match v2i64, but v2i64 is never matched by
5113 // movl earlier because we make it strict to avoid messing with the movlp load
5114 // folding logic (see the code above getMOVLP call). Match it here then,
5115 // this is horrible, but will stay like this until we move all shuffle
5116 // matching to x86 specific nodes. Note that for the 1st condition all
5117 // types are matched with movsd.
5118 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5119 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5121 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5124 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5126 // Invert the operand order and use SHUFPS to match it.
5127 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5128 X86::getShuffleSHUFImmediate(SVOp), DAG);
5132 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5133 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5134 SDValue V1 = Op.getOperand(0);
5135 SDValue V2 = Op.getOperand(1);
5136 EVT VT = Op.getValueType();
5137 DebugLoc dl = Op.getDebugLoc();
5138 unsigned NumElems = VT.getVectorNumElements();
5139 bool isMMX = VT.getSizeInBits() == 64;
5140 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5141 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5142 bool V1IsSplat = false;
5143 bool V2IsSplat = false;
5144 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5145 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5146 MachineFunction &MF = DAG.getMachineFunction();
5147 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5149 if (isZeroShuffle(SVOp))
5150 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5152 // Promote splats to v4f32.
5153 if (SVOp->isSplat()) {
5154 if (isMMX || NumElems < 4)
5156 return PromoteSplat(SVOp, DAG);
5159 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5161 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5162 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5163 if (NewOp.getNode())
5164 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5165 LowerVECTOR_SHUFFLE(NewOp, DAG));
5166 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5167 // FIXME: Figure out a cleaner way to do this.
5168 // Try to make use of movq to zero out the top part.
5169 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5170 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5171 if (NewOp.getNode()) {
5172 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5173 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5174 DAG, Subtarget, dl);
5176 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5177 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5178 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5179 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5180 DAG, Subtarget, dl);
5184 if (X86::isPSHUFDMask(SVOp)) {
5185 // The actual implementation will match the mask in the if above and then
5186 // during isel it can match several different instructions, not only pshufd
5187 // as its name says, sad but true, emulate the behavior for now...
5188 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5189 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5191 if (OptForSize && HasSSE2 && X86::isUNPCKL_v_undef_Mask(SVOp) &&
5193 return getTargetShuffleNode(X86ISD::PUNPCKLDQ, dl, VT, V1, V1, DAG);
5195 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5197 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5198 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5200 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5201 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5204 if (VT == MVT::v4f32)
5205 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5209 // Check if this can be converted into a logical shift.
5210 bool isLeft = false;
5213 bool isShift = getSubtarget()->hasSSE2() &&
5214 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5215 if (isShift && ShVal.hasOneUse()) {
5216 // If the shifted value has multiple uses, it may be cheaper to use
5217 // v_set0 + movlhps or movhlps, etc.
5218 EVT EltVT = VT.getVectorElementType();
5219 ShAmt *= EltVT.getSizeInBits();
5220 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5223 if (X86::isMOVLMask(SVOp)) {
5226 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5227 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5228 if (!isMMX && !X86::isMOVLPMask(SVOp)) {
5229 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5230 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5232 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5233 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5237 // FIXME: fold these into legal mask.
5239 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5240 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5242 if (X86::isMOVHLPSMask(SVOp))
5243 return getMOVHighToLow(Op, dl, DAG);
5245 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5246 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5248 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5249 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5251 if (X86::isMOVLPMask(SVOp))
5252 return getMOVLP(Op, dl, DAG, HasSSE2);
5255 if (ShouldXformToMOVHLPS(SVOp) ||
5256 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5257 return CommuteVectorShuffle(SVOp, DAG);
5260 // No better options. Use a vshl / vsrl.
5261 EVT EltVT = VT.getVectorElementType();
5262 ShAmt *= EltVT.getSizeInBits();
5263 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5266 bool Commuted = false;
5267 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5268 // 1,1,1,1 -> v8i16 though.
5269 V1IsSplat = isSplatVector(V1.getNode());
5270 V2IsSplat = isSplatVector(V2.getNode());
5272 // Canonicalize the splat or undef, if present, to be on the RHS.
5273 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5274 Op = CommuteVectorShuffle(SVOp, DAG);
5275 SVOp = cast<ShuffleVectorSDNode>(Op);
5276 V1 = SVOp->getOperand(0);
5277 V2 = SVOp->getOperand(1);
5278 std::swap(V1IsSplat, V2IsSplat);
5279 std::swap(V1IsUndef, V2IsUndef);
5283 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5284 // Shuffling low element of v1 into undef, just return v1.
5287 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5288 // the instruction selector will not match, so get a canonical MOVL with
5289 // swapped operands to undo the commute.
5290 return getMOVL(DAG, dl, VT, V2, V1);
5293 if (X86::isUNPCKH_v_undef_Mask(SVOp) ||
5294 X86::isUNPCKLMask(SVOp) ||
5295 X86::isUNPCKHMask(SVOp))
5299 // Normalize mask so all entries that point to V2 points to its first
5300 // element then try to match unpck{h|l} again. If match, return a
5301 // new vector_shuffle with the corrected mask.
5302 SDValue NewMask = NormalizeMask(SVOp, DAG);
5303 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5304 if (NSVOp != SVOp) {
5305 if (X86::isUNPCKLMask(NSVOp, true)) {
5307 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5314 // Commute is back and try unpck* again.
5315 // FIXME: this seems wrong.
5316 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5317 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5318 if (X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
5319 X86::isUNPCKLMask(NewSVOp) ||
5320 X86::isUNPCKHMask(NewSVOp))
5324 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
5326 // Normalize the node to match x86 shuffle ops if needed
5327 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5328 return CommuteVectorShuffle(SVOp, DAG);
5330 // Check for legal shuffle and return?
5331 SmallVector<int, 16> PermMask;
5332 SVOp->getMask(PermMask);
5333 if (isShuffleMaskLegal(PermMask, VT))
5336 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5337 if (VT == MVT::v8i16) {
5338 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5339 if (NewOp.getNode())
5343 if (VT == MVT::v16i8) {
5344 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5345 if (NewOp.getNode())
5349 // Handle all 4 wide cases with a number of shuffles except for MMX.
5350 if (NumElems == 4 && !isMMX)
5351 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5357 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5358 SelectionDAG &DAG) const {
5359 EVT VT = Op.getValueType();
5360 DebugLoc dl = Op.getDebugLoc();
5361 if (VT.getSizeInBits() == 8) {
5362 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
5363 Op.getOperand(0), Op.getOperand(1));
5364 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5365 DAG.getValueType(VT));
5366 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5367 } else if (VT.getSizeInBits() == 16) {
5368 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5369 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
5371 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5372 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5373 DAG.getNode(ISD::BIT_CONVERT, dl,
5377 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
5378 Op.getOperand(0), Op.getOperand(1));
5379 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5380 DAG.getValueType(VT));
5381 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5382 } else if (VT == MVT::f32) {
5383 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
5384 // the result back to FR32 register. It's only worth matching if the
5385 // result has a single use which is a store or a bitcast to i32. And in
5386 // the case of a store, it's not worth it if the index is a constant 0,
5387 // because a MOVSSmr can be used instead, which is smaller and faster.
5388 if (!Op.hasOneUse())
5390 SDNode *User = *Op.getNode()->use_begin();
5391 if ((User->getOpcode() != ISD::STORE ||
5392 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5393 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5394 (User->getOpcode() != ISD::BIT_CONVERT ||
5395 User->getValueType(0) != MVT::i32))
5397 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5398 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
5401 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
5402 } else if (VT == MVT::i32) {
5403 // ExtractPS works with constant index.
5404 if (isa<ConstantSDNode>(Op.getOperand(1)))
5412 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5413 SelectionDAG &DAG) const {
5414 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5417 if (Subtarget->hasSSE41()) {
5418 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5423 EVT VT = Op.getValueType();
5424 DebugLoc dl = Op.getDebugLoc();
5425 // TODO: handle v16i8.
5426 if (VT.getSizeInBits() == 16) {
5427 SDValue Vec = Op.getOperand(0);
5428 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5430 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5431 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5432 DAG.getNode(ISD::BIT_CONVERT, dl,
5435 // Transform it so it match pextrw which produces a 32-bit result.
5436 EVT EltVT = MVT::i32;
5437 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5438 Op.getOperand(0), Op.getOperand(1));
5439 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5440 DAG.getValueType(VT));
5441 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5442 } else if (VT.getSizeInBits() == 32) {
5443 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5447 // SHUFPS the element to the lowest double word, then movss.
5448 int Mask[4] = { Idx, -1, -1, -1 };
5449 EVT VVT = Op.getOperand(0).getValueType();
5450 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5451 DAG.getUNDEF(VVT), Mask);
5452 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5453 DAG.getIntPtrConstant(0));
5454 } else if (VT.getSizeInBits() == 64) {
5455 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5456 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5457 // to match extract_elt for f64.
5458 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5462 // UNPCKHPD the element to the lowest double word, then movsd.
5463 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5464 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5465 int Mask[2] = { 1, -1 };
5466 EVT VVT = Op.getOperand(0).getValueType();
5467 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5468 DAG.getUNDEF(VVT), Mask);
5469 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5470 DAG.getIntPtrConstant(0));
5477 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5478 SelectionDAG &DAG) const {
5479 EVT VT = Op.getValueType();
5480 EVT EltVT = VT.getVectorElementType();
5481 DebugLoc dl = Op.getDebugLoc();
5483 SDValue N0 = Op.getOperand(0);
5484 SDValue N1 = Op.getOperand(1);
5485 SDValue N2 = Op.getOperand(2);
5487 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5488 isa<ConstantSDNode>(N2)) {
5490 if (VT == MVT::v8i16)
5491 Opc = X86ISD::PINSRW;
5492 else if (VT == MVT::v4i16)
5493 Opc = X86ISD::MMX_PINSRW;
5494 else if (VT == MVT::v16i8)
5495 Opc = X86ISD::PINSRB;
5497 Opc = X86ISD::PINSRB;
5499 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5501 if (N1.getValueType() != MVT::i32)
5502 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5503 if (N2.getValueType() != MVT::i32)
5504 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5505 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5506 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5507 // Bits [7:6] of the constant are the source select. This will always be
5508 // zero here. The DAG Combiner may combine an extract_elt index into these
5509 // bits. For example (insert (extract, 3), 2) could be matched by putting
5510 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5511 // Bits [5:4] of the constant are the destination select. This is the
5512 // value of the incoming immediate.
5513 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5514 // combine either bitwise AND or insert of float 0.0 to set these bits.
5515 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5516 // Create this as a scalar to vector..
5517 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5518 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5519 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5520 // PINSR* works with constant index.
5527 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5528 EVT VT = Op.getValueType();
5529 EVT EltVT = VT.getVectorElementType();
5531 if (Subtarget->hasSSE41())
5532 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5534 if (EltVT == MVT::i8)
5537 DebugLoc dl = Op.getDebugLoc();
5538 SDValue N0 = Op.getOperand(0);
5539 SDValue N1 = Op.getOperand(1);
5540 SDValue N2 = Op.getOperand(2);
5542 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5543 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5544 // as its second argument.
5545 if (N1.getValueType() != MVT::i32)
5546 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5547 if (N2.getValueType() != MVT::i32)
5548 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5549 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5550 dl, VT, N0, N1, N2);
5556 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5557 DebugLoc dl = Op.getDebugLoc();
5559 if (Op.getValueType() == MVT::v1i64 &&
5560 Op.getOperand(0).getValueType() == MVT::i64)
5561 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5563 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5564 EVT VT = MVT::v2i32;
5565 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5572 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5573 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5576 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5577 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5578 // one of the above mentioned nodes. It has to be wrapped because otherwise
5579 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5580 // be used to form addressing mode. These wrapped nodes will be selected
5583 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5584 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5586 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5588 unsigned char OpFlag = 0;
5589 unsigned WrapperKind = X86ISD::Wrapper;
5590 CodeModel::Model M = getTargetMachine().getCodeModel();
5592 if (Subtarget->isPICStyleRIPRel() &&
5593 (M == CodeModel::Small || M == CodeModel::Kernel))
5594 WrapperKind = X86ISD::WrapperRIP;
5595 else if (Subtarget->isPICStyleGOT())
5596 OpFlag = X86II::MO_GOTOFF;
5597 else if (Subtarget->isPICStyleStubPIC())
5598 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5600 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5602 CP->getOffset(), OpFlag);
5603 DebugLoc DL = CP->getDebugLoc();
5604 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5605 // With PIC, the address is actually $g + Offset.
5607 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5608 DAG.getNode(X86ISD::GlobalBaseReg,
5609 DebugLoc(), getPointerTy()),
5616 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5617 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5619 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5621 unsigned char OpFlag = 0;
5622 unsigned WrapperKind = X86ISD::Wrapper;
5623 CodeModel::Model M = getTargetMachine().getCodeModel();
5625 if (Subtarget->isPICStyleRIPRel() &&
5626 (M == CodeModel::Small || M == CodeModel::Kernel))
5627 WrapperKind = X86ISD::WrapperRIP;
5628 else if (Subtarget->isPICStyleGOT())
5629 OpFlag = X86II::MO_GOTOFF;
5630 else if (Subtarget->isPICStyleStubPIC())
5631 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5633 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5635 DebugLoc DL = JT->getDebugLoc();
5636 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5638 // With PIC, the address is actually $g + Offset.
5640 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5641 DAG.getNode(X86ISD::GlobalBaseReg,
5642 DebugLoc(), getPointerTy()),
5650 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5651 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5653 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5655 unsigned char OpFlag = 0;
5656 unsigned WrapperKind = X86ISD::Wrapper;
5657 CodeModel::Model M = getTargetMachine().getCodeModel();
5659 if (Subtarget->isPICStyleRIPRel() &&
5660 (M == CodeModel::Small || M == CodeModel::Kernel))
5661 WrapperKind = X86ISD::WrapperRIP;
5662 else if (Subtarget->isPICStyleGOT())
5663 OpFlag = X86II::MO_GOTOFF;
5664 else if (Subtarget->isPICStyleStubPIC())
5665 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5667 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5669 DebugLoc DL = Op.getDebugLoc();
5670 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5673 // With PIC, the address is actually $g + Offset.
5674 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5675 !Subtarget->is64Bit()) {
5676 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5677 DAG.getNode(X86ISD::GlobalBaseReg,
5678 DebugLoc(), getPointerTy()),
5686 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5687 // Create the TargetBlockAddressAddress node.
5688 unsigned char OpFlags =
5689 Subtarget->ClassifyBlockAddressReference();
5690 CodeModel::Model M = getTargetMachine().getCodeModel();
5691 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5692 DebugLoc dl = Op.getDebugLoc();
5693 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5694 /*isTarget=*/true, OpFlags);
5696 if (Subtarget->isPICStyleRIPRel() &&
5697 (M == CodeModel::Small || M == CodeModel::Kernel))
5698 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5700 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5702 // With PIC, the address is actually $g + Offset.
5703 if (isGlobalRelativeToPICBase(OpFlags)) {
5704 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5705 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5713 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5715 SelectionDAG &DAG) const {
5716 // Create the TargetGlobalAddress node, folding in the constant
5717 // offset if it is legal.
5718 unsigned char OpFlags =
5719 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5720 CodeModel::Model M = getTargetMachine().getCodeModel();
5722 if (OpFlags == X86II::MO_NO_FLAG &&
5723 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5724 // A direct static reference to a global.
5725 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5728 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5731 if (Subtarget->isPICStyleRIPRel() &&
5732 (M == CodeModel::Small || M == CodeModel::Kernel))
5733 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5735 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5737 // With PIC, the address is actually $g + Offset.
5738 if (isGlobalRelativeToPICBase(OpFlags)) {
5739 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5740 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5744 // For globals that require a load from a stub to get the address, emit the
5746 if (isGlobalStubReference(OpFlags))
5747 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5748 PseudoSourceValue::getGOT(), 0, false, false, 0);
5750 // If there was a non-zero offset that we didn't fold, create an explicit
5753 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5754 DAG.getConstant(Offset, getPointerTy()));
5760 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5761 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5762 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5763 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5767 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5768 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5769 unsigned char OperandFlags) {
5770 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5771 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5772 DebugLoc dl = GA->getDebugLoc();
5773 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5774 GA->getValueType(0),
5778 SDValue Ops[] = { Chain, TGA, *InFlag };
5779 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5781 SDValue Ops[] = { Chain, TGA };
5782 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5785 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5786 MFI->setAdjustsStack(true);
5788 SDValue Flag = Chain.getValue(1);
5789 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5792 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5794 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5797 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5798 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5799 DAG.getNode(X86ISD::GlobalBaseReg,
5800 DebugLoc(), PtrVT), InFlag);
5801 InFlag = Chain.getValue(1);
5803 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5806 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5808 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5810 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5811 X86::RAX, X86II::MO_TLSGD);
5814 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5815 // "local exec" model.
5816 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5817 const EVT PtrVT, TLSModel::Model model,
5819 DebugLoc dl = GA->getDebugLoc();
5820 // Get the Thread Pointer
5821 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5823 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5826 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5827 NULL, 0, false, false, 0);
5829 unsigned char OperandFlags = 0;
5830 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5832 unsigned WrapperKind = X86ISD::Wrapper;
5833 if (model == TLSModel::LocalExec) {
5834 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5835 } else if (is64Bit) {
5836 assert(model == TLSModel::InitialExec);
5837 OperandFlags = X86II::MO_GOTTPOFF;
5838 WrapperKind = X86ISD::WrapperRIP;
5840 assert(model == TLSModel::InitialExec);
5841 OperandFlags = X86II::MO_INDNTPOFF;
5844 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5846 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5847 GA->getValueType(0),
5848 GA->getOffset(), OperandFlags);
5849 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5851 if (model == TLSModel::InitialExec)
5852 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5853 PseudoSourceValue::getGOT(), 0, false, false, 0);
5855 // The address of the thread local variable is the add of the thread
5856 // pointer with the offset of the variable.
5857 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5861 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5863 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5864 const GlobalValue *GV = GA->getGlobal();
5866 if (Subtarget->isTargetELF()) {
5867 // TODO: implement the "local dynamic" model
5868 // TODO: implement the "initial exec"model for pic executables
5870 // If GV is an alias then use the aliasee for determining
5871 // thread-localness.
5872 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5873 GV = GA->resolveAliasedGlobal(false);
5875 TLSModel::Model model
5876 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5879 case TLSModel::GeneralDynamic:
5880 case TLSModel::LocalDynamic: // not implemented
5881 if (Subtarget->is64Bit())
5882 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5883 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5885 case TLSModel::InitialExec:
5886 case TLSModel::LocalExec:
5887 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5888 Subtarget->is64Bit());
5890 } else if (Subtarget->isTargetDarwin()) {
5891 // Darwin only has one model of TLS. Lower to that.
5892 unsigned char OpFlag = 0;
5893 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5894 X86ISD::WrapperRIP : X86ISD::Wrapper;
5896 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5898 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5899 !Subtarget->is64Bit();
5901 OpFlag = X86II::MO_TLVP_PIC_BASE;
5903 OpFlag = X86II::MO_TLVP;
5904 DebugLoc DL = Op.getDebugLoc();
5905 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
5907 GA->getOffset(), OpFlag);
5908 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5910 // With PIC32, the address is actually $g + Offset.
5912 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5913 DAG.getNode(X86ISD::GlobalBaseReg,
5914 DebugLoc(), getPointerTy()),
5917 // Lowering the machine isd will make sure everything is in the right
5919 SDValue Args[] = { Offset };
5920 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5922 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5923 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5924 MFI->setAdjustsStack(true);
5926 // And our return value (tls address) is in the standard call return value
5928 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5929 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5933 "TLS not implemented for this target.");
5935 llvm_unreachable("Unreachable");
5940 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5941 /// take a 2 x i32 value to shift plus a shift amount.
5942 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5943 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5944 EVT VT = Op.getValueType();
5945 unsigned VTBits = VT.getSizeInBits();
5946 DebugLoc dl = Op.getDebugLoc();
5947 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5948 SDValue ShOpLo = Op.getOperand(0);
5949 SDValue ShOpHi = Op.getOperand(1);
5950 SDValue ShAmt = Op.getOperand(2);
5951 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5952 DAG.getConstant(VTBits - 1, MVT::i8))
5953 : DAG.getConstant(0, VT);
5956 if (Op.getOpcode() == ISD::SHL_PARTS) {
5957 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5958 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5960 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5961 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5964 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5965 DAG.getConstant(VTBits, MVT::i8));
5966 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5967 AndNode, DAG.getConstant(0, MVT::i8));
5970 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5971 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5972 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5974 if (Op.getOpcode() == ISD::SHL_PARTS) {
5975 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5976 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5978 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5979 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5982 SDValue Ops[2] = { Lo, Hi };
5983 return DAG.getMergeValues(Ops, 2, dl);
5986 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5987 SelectionDAG &DAG) const {
5988 EVT SrcVT = Op.getOperand(0).getValueType();
5990 if (SrcVT.isVector()) {
5991 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5997 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5998 "Unknown SINT_TO_FP to lower!");
6000 // These are really Legal; return the operand so the caller accepts it as
6002 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6004 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6005 Subtarget->is64Bit()) {
6009 DebugLoc dl = Op.getDebugLoc();
6010 unsigned Size = SrcVT.getSizeInBits()/8;
6011 MachineFunction &MF = DAG.getMachineFunction();
6012 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6013 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6014 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6016 PseudoSourceValue::getFixedStack(SSFI), 0,
6018 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6021 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6023 SelectionDAG &DAG) const {
6025 DebugLoc dl = Op.getDebugLoc();
6027 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6029 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
6031 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6032 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6033 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
6034 Tys, Ops, array_lengthof(Ops));
6037 Chain = Result.getValue(1);
6038 SDValue InFlag = Result.getValue(2);
6040 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6041 // shouldn't be necessary except that RFP cannot be live across
6042 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6043 MachineFunction &MF = DAG.getMachineFunction();
6044 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
6045 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6046 Tys = DAG.getVTList(MVT::Other);
6048 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6050 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
6051 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
6052 PseudoSourceValue::getFixedStack(SSFI), 0,
6059 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6060 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6061 SelectionDAG &DAG) const {
6062 // This algorithm is not obvious. Here it is in C code, more or less:
6064 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6065 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6066 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6068 // Copy ints to xmm registers.
6069 __m128i xh = _mm_cvtsi32_si128( hi );
6070 __m128i xl = _mm_cvtsi32_si128( lo );
6072 // Combine into low half of a single xmm register.
6073 __m128i x = _mm_unpacklo_epi32( xh, xl );
6077 // Merge in appropriate exponents to give the integer bits the right
6079 x = _mm_unpacklo_epi32( x, exp );
6081 // Subtract away the biases to deal with the IEEE-754 double precision
6083 d = _mm_sub_pd( (__m128d) x, bias );
6085 // All conversions up to here are exact. The correctly rounded result is
6086 // calculated using the current rounding mode using the following
6088 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6089 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6090 // store doesn't really need to be here (except
6091 // maybe to zero the other double)
6096 DebugLoc dl = Op.getDebugLoc();
6097 LLVMContext *Context = DAG.getContext();
6099 // Build some magic constants.
6100 std::vector<Constant*> CV0;
6101 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6102 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6103 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6104 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6105 Constant *C0 = ConstantVector::get(CV0);
6106 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6108 std::vector<Constant*> CV1;
6110 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6112 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6113 Constant *C1 = ConstantVector::get(CV1);
6114 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6116 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6117 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6119 DAG.getIntPtrConstant(1)));
6120 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6121 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6123 DAG.getIntPtrConstant(0)));
6124 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6125 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6126 PseudoSourceValue::getConstantPool(), 0,
6128 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6129 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
6130 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6131 PseudoSourceValue::getConstantPool(), 0,
6133 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6135 // Add the halves; easiest way is to swap them into another reg first.
6136 int ShufMask[2] = { 1, -1 };
6137 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6138 DAG.getUNDEF(MVT::v2f64), ShufMask);
6139 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6140 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6141 DAG.getIntPtrConstant(0));
6144 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6145 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6146 SelectionDAG &DAG) const {
6147 DebugLoc dl = Op.getDebugLoc();
6148 // FP constant to bias correct the final result.
6149 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6152 // Load the 32-bit value into an XMM register.
6153 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6154 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6156 DAG.getIntPtrConstant(0)));
6158 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6159 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
6160 DAG.getIntPtrConstant(0));
6162 // Or the load with the bias.
6163 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6164 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6165 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6167 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6168 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6169 MVT::v2f64, Bias)));
6170 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6171 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
6172 DAG.getIntPtrConstant(0));
6174 // Subtract the bias.
6175 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
6177 // Handle final rounding.
6178 EVT DestVT = Op.getValueType();
6180 if (DestVT.bitsLT(MVT::f64)) {
6181 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
6182 DAG.getIntPtrConstant(0));
6183 } else if (DestVT.bitsGT(MVT::f64)) {
6184 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
6187 // Handle final rounding.
6191 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
6192 SelectionDAG &DAG) const {
6193 SDValue N0 = Op.getOperand(0);
6194 DebugLoc dl = Op.getDebugLoc();
6196 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
6197 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
6198 // the optimization here.
6199 if (DAG.SignBitIsZero(N0))
6200 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
6202 EVT SrcVT = N0.getValueType();
6203 EVT DstVT = Op.getValueType();
6204 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
6205 return LowerUINT_TO_FP_i64(Op, DAG);
6206 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
6207 return LowerUINT_TO_FP_i32(Op, DAG);
6209 // Make a 64-bit buffer, and use it to build an FILD.
6210 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
6211 if (SrcVT == MVT::i32) {
6212 SDValue WordOff = DAG.getConstant(4, getPointerTy());
6213 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
6214 getPointerTy(), StackSlot, WordOff);
6215 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6216 StackSlot, NULL, 0, false, false, 0);
6217 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
6218 OffsetSlot, NULL, 0, false, false, 0);
6219 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
6223 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
6224 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6225 StackSlot, NULL, 0, false, false, 0);
6226 // For i64 source, we need to add the appropriate power of 2 if the input
6227 // was negative. This is the same as the optimization in
6228 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
6229 // we must be careful to do the computation in x87 extended precision, not
6230 // in SSE. (The generic code can't know it's OK to do this, or how to.)
6231 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
6232 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
6233 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
6235 APInt FF(32, 0x5F800000ULL);
6237 // Check whether the sign bit is set.
6238 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
6239 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
6242 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
6243 SDValue FudgePtr = DAG.getConstantPool(
6244 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
6247 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
6248 SDValue Zero = DAG.getIntPtrConstant(0);
6249 SDValue Four = DAG.getIntPtrConstant(4);
6250 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
6252 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
6254 // Load the value out, extending it from f32 to f80.
6255 // FIXME: Avoid the extend by constructing the right constant pool?
6256 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
6257 FudgePtr, PseudoSourceValue::getConstantPool(),
6258 0, MVT::f32, false, false, 4);
6259 // Extend everything to 80 bits to force it to be done on x87.
6260 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
6261 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
6264 std::pair<SDValue,SDValue> X86TargetLowering::
6265 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
6266 DebugLoc dl = Op.getDebugLoc();
6268 EVT DstTy = Op.getValueType();
6271 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
6275 assert(DstTy.getSimpleVT() <= MVT::i64 &&
6276 DstTy.getSimpleVT() >= MVT::i16 &&
6277 "Unknown FP_TO_SINT to lower!");
6279 // These are really Legal.
6280 if (DstTy == MVT::i32 &&
6281 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6282 return std::make_pair(SDValue(), SDValue());
6283 if (Subtarget->is64Bit() &&
6284 DstTy == MVT::i64 &&
6285 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6286 return std::make_pair(SDValue(), SDValue());
6288 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
6290 MachineFunction &MF = DAG.getMachineFunction();
6291 unsigned MemSize = DstTy.getSizeInBits()/8;
6292 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6293 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6296 switch (DstTy.getSimpleVT().SimpleTy) {
6297 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
6298 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
6299 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
6300 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
6303 SDValue Chain = DAG.getEntryNode();
6304 SDValue Value = Op.getOperand(0);
6305 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
6306 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
6307 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
6308 PseudoSourceValue::getFixedStack(SSFI), 0,
6310 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
6312 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
6314 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
6315 Chain = Value.getValue(1);
6316 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6317 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6320 // Build the FP_TO_INT*_IN_MEM
6321 SDValue Ops[] = { Chain, Value, StackSlot };
6322 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
6324 return std::make_pair(FIST, StackSlot);
6327 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
6328 SelectionDAG &DAG) const {
6329 if (Op.getValueType().isVector()) {
6330 if (Op.getValueType() == MVT::v2i32 &&
6331 Op.getOperand(0).getValueType() == MVT::v2f64) {
6337 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
6338 SDValue FIST = Vals.first, StackSlot = Vals.second;
6339 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
6340 if (FIST.getNode() == 0) return Op;
6343 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6344 FIST, StackSlot, NULL, 0, false, false, 0);
6347 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
6348 SelectionDAG &DAG) const {
6349 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
6350 SDValue FIST = Vals.first, StackSlot = Vals.second;
6351 assert(FIST.getNode() && "Unexpected failure");
6354 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6355 FIST, StackSlot, NULL, 0, false, false, 0);
6358 SDValue X86TargetLowering::LowerFABS(SDValue Op,
6359 SelectionDAG &DAG) const {
6360 LLVMContext *Context = DAG.getContext();
6361 DebugLoc dl = Op.getDebugLoc();
6362 EVT VT = Op.getValueType();
6365 EltVT = VT.getVectorElementType();
6366 std::vector<Constant*> CV;
6367 if (EltVT == MVT::f64) {
6368 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
6372 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
6378 Constant *C = ConstantVector::get(CV);
6379 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6380 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6381 PseudoSourceValue::getConstantPool(), 0,
6383 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
6386 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
6387 LLVMContext *Context = DAG.getContext();
6388 DebugLoc dl = Op.getDebugLoc();
6389 EVT VT = Op.getValueType();
6392 EltVT = VT.getVectorElementType();
6393 std::vector<Constant*> CV;
6394 if (EltVT == MVT::f64) {
6395 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
6399 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
6405 Constant *C = ConstantVector::get(CV);
6406 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6407 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6408 PseudoSourceValue::getConstantPool(), 0,
6410 if (VT.isVector()) {
6411 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6412 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6413 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6415 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
6417 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6421 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6422 LLVMContext *Context = DAG.getContext();
6423 SDValue Op0 = Op.getOperand(0);
6424 SDValue Op1 = Op.getOperand(1);
6425 DebugLoc dl = Op.getDebugLoc();
6426 EVT VT = Op.getValueType();
6427 EVT SrcVT = Op1.getValueType();
6429 // If second operand is smaller, extend it first.
6430 if (SrcVT.bitsLT(VT)) {
6431 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6434 // And if it is bigger, shrink it first.
6435 if (SrcVT.bitsGT(VT)) {
6436 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6440 // At this point the operands and the result should have the same
6441 // type, and that won't be f80 since that is not custom lowered.
6443 // First get the sign bit of second operand.
6444 std::vector<Constant*> CV;
6445 if (SrcVT == MVT::f64) {
6446 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6447 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6449 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6450 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6451 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6452 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6454 Constant *C = ConstantVector::get(CV);
6455 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6456 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6457 PseudoSourceValue::getConstantPool(), 0,
6459 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6461 // Shift sign bit right or left if the two operands have different types.
6462 if (SrcVT.bitsGT(VT)) {
6463 // Op0 is MVT::f32, Op1 is MVT::f64.
6464 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6465 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6466 DAG.getConstant(32, MVT::i32));
6467 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
6468 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6469 DAG.getIntPtrConstant(0));
6472 // Clear first operand sign bit.
6474 if (VT == MVT::f64) {
6475 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6476 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6478 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6479 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6480 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6481 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6483 C = ConstantVector::get(CV);
6484 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6485 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6486 PseudoSourceValue::getConstantPool(), 0,
6488 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6490 // Or the value with the sign bit.
6491 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6494 /// Emit nodes that will be selected as "test Op0,Op0", or something
6496 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6497 SelectionDAG &DAG) const {
6498 DebugLoc dl = Op.getDebugLoc();
6500 // CF and OF aren't always set the way we want. Determine which
6501 // of these we need.
6502 bool NeedCF = false;
6503 bool NeedOF = false;
6506 case X86::COND_A: case X86::COND_AE:
6507 case X86::COND_B: case X86::COND_BE:
6510 case X86::COND_G: case X86::COND_GE:
6511 case X86::COND_L: case X86::COND_LE:
6512 case X86::COND_O: case X86::COND_NO:
6517 // See if we can use the EFLAGS value from the operand instead of
6518 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6519 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6520 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6521 // Emit a CMP with 0, which is the TEST pattern.
6522 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6523 DAG.getConstant(0, Op.getValueType()));
6525 unsigned Opcode = 0;
6526 unsigned NumOperands = 0;
6527 switch (Op.getNode()->getOpcode()) {
6529 // Due to an isel shortcoming, be conservative if this add is likely to be
6530 // selected as part of a load-modify-store instruction. When the root node
6531 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6532 // uses of other nodes in the match, such as the ADD in this case. This
6533 // leads to the ADD being left around and reselected, with the result being
6534 // two adds in the output. Alas, even if none our users are stores, that
6535 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6536 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6537 // climbing the DAG back to the root, and it doesn't seem to be worth the
6539 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6540 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6541 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6544 if (ConstantSDNode *C =
6545 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6546 // An add of one will be selected as an INC.
6547 if (C->getAPIntValue() == 1) {
6548 Opcode = X86ISD::INC;
6553 // An add of negative one (subtract of one) will be selected as a DEC.
6554 if (C->getAPIntValue().isAllOnesValue()) {
6555 Opcode = X86ISD::DEC;
6561 // Otherwise use a regular EFLAGS-setting add.
6562 Opcode = X86ISD::ADD;
6566 // If the primary and result isn't used, don't bother using X86ISD::AND,
6567 // because a TEST instruction will be better.
6568 bool NonFlagUse = false;
6569 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6570 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6572 unsigned UOpNo = UI.getOperandNo();
6573 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6574 // Look pass truncate.
6575 UOpNo = User->use_begin().getOperandNo();
6576 User = *User->use_begin();
6579 if (User->getOpcode() != ISD::BRCOND &&
6580 User->getOpcode() != ISD::SETCC &&
6581 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6594 // Due to the ISEL shortcoming noted above, be conservative if this op is
6595 // likely to be selected as part of a load-modify-store instruction.
6596 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6597 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6598 if (UI->getOpcode() == ISD::STORE)
6601 // Otherwise use a regular EFLAGS-setting instruction.
6602 switch (Op.getNode()->getOpcode()) {
6603 default: llvm_unreachable("unexpected operator!");
6604 case ISD::SUB: Opcode = X86ISD::SUB; break;
6605 case ISD::OR: Opcode = X86ISD::OR; break;
6606 case ISD::XOR: Opcode = X86ISD::XOR; break;
6607 case ISD::AND: Opcode = X86ISD::AND; break;
6619 return SDValue(Op.getNode(), 1);
6626 // Emit a CMP with 0, which is the TEST pattern.
6627 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6628 DAG.getConstant(0, Op.getValueType()));
6630 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6631 SmallVector<SDValue, 4> Ops;
6632 for (unsigned i = 0; i != NumOperands; ++i)
6633 Ops.push_back(Op.getOperand(i));
6635 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6636 DAG.ReplaceAllUsesWith(Op, New);
6637 return SDValue(New.getNode(), 1);
6640 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6642 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6643 SelectionDAG &DAG) const {
6644 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6645 if (C->getAPIntValue() == 0)
6646 return EmitTest(Op0, X86CC, DAG);
6648 DebugLoc dl = Op0.getDebugLoc();
6649 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6652 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6653 /// if it's possible.
6654 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6655 DebugLoc dl, SelectionDAG &DAG) const {
6656 SDValue Op0 = And.getOperand(0);
6657 SDValue Op1 = And.getOperand(1);
6658 if (Op0.getOpcode() == ISD::TRUNCATE)
6659 Op0 = Op0.getOperand(0);
6660 if (Op1.getOpcode() == ISD::TRUNCATE)
6661 Op1 = Op1.getOperand(0);
6664 if (Op1.getOpcode() == ISD::SHL)
6665 std::swap(Op0, Op1);
6666 if (Op0.getOpcode() == ISD::SHL) {
6667 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6668 if (And00C->getZExtValue() == 1) {
6669 // If we looked past a truncate, check that it's only truncating away
6671 unsigned BitWidth = Op0.getValueSizeInBits();
6672 unsigned AndBitWidth = And.getValueSizeInBits();
6673 if (BitWidth > AndBitWidth) {
6674 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6675 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6676 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6680 RHS = Op0.getOperand(1);
6682 } else if (Op1.getOpcode() == ISD::Constant) {
6683 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6684 SDValue AndLHS = Op0;
6685 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6686 LHS = AndLHS.getOperand(0);
6687 RHS = AndLHS.getOperand(1);
6691 if (LHS.getNode()) {
6692 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6693 // instruction. Since the shift amount is in-range-or-undefined, we know
6694 // that doing a bittest on the i32 value is ok. We extend to i32 because
6695 // the encoding for the i16 version is larger than the i32 version.
6696 // Also promote i16 to i32 for performance / code size reason.
6697 if (LHS.getValueType() == MVT::i8 ||
6698 LHS.getValueType() == MVT::i16)
6699 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6701 // If the operand types disagree, extend the shift amount to match. Since
6702 // BT ignores high bits (like shifts) we can use anyextend.
6703 if (LHS.getValueType() != RHS.getValueType())
6704 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6706 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6707 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6708 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6709 DAG.getConstant(Cond, MVT::i8), BT);
6715 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6716 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6717 SDValue Op0 = Op.getOperand(0);
6718 SDValue Op1 = Op.getOperand(1);
6719 DebugLoc dl = Op.getDebugLoc();
6720 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6722 // Optimize to BT if possible.
6723 // Lower (X & (1 << N)) == 0 to BT(X, N).
6724 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6725 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6726 if (Op0.getOpcode() == ISD::AND &&
6728 Op1.getOpcode() == ISD::Constant &&
6729 cast<ConstantSDNode>(Op1)->isNullValue() &&
6730 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6731 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6732 if (NewSetCC.getNode())
6736 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6737 if (Op0.getOpcode() == X86ISD::SETCC &&
6738 Op1.getOpcode() == ISD::Constant &&
6739 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6740 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6741 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6742 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6743 bool Invert = (CC == ISD::SETNE) ^
6744 cast<ConstantSDNode>(Op1)->isNullValue();
6746 CCode = X86::GetOppositeBranchCondition(CCode);
6747 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6748 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6751 bool isFP = Op1.getValueType().isFloatingPoint();
6752 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6753 if (X86CC == X86::COND_INVALID)
6756 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6758 // Use sbb x, x to materialize carry bit into a GPR.
6759 if (X86CC == X86::COND_B)
6760 return DAG.getNode(ISD::AND, dl, MVT::i8,
6761 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6762 DAG.getConstant(X86CC, MVT::i8), Cond),
6763 DAG.getConstant(1, MVT::i8));
6765 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6766 DAG.getConstant(X86CC, MVT::i8), Cond);
6769 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6771 SDValue Op0 = Op.getOperand(0);
6772 SDValue Op1 = Op.getOperand(1);
6773 SDValue CC = Op.getOperand(2);
6774 EVT VT = Op.getValueType();
6775 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6776 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6777 DebugLoc dl = Op.getDebugLoc();
6781 EVT VT0 = Op0.getValueType();
6782 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6783 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6786 switch (SetCCOpcode) {
6789 case ISD::SETEQ: SSECC = 0; break;
6791 case ISD::SETGT: Swap = true; // Fallthrough
6793 case ISD::SETOLT: SSECC = 1; break;
6795 case ISD::SETGE: Swap = true; // Fallthrough
6797 case ISD::SETOLE: SSECC = 2; break;
6798 case ISD::SETUO: SSECC = 3; break;
6800 case ISD::SETNE: SSECC = 4; break;
6801 case ISD::SETULE: Swap = true;
6802 case ISD::SETUGE: SSECC = 5; break;
6803 case ISD::SETULT: Swap = true;
6804 case ISD::SETUGT: SSECC = 6; break;
6805 case ISD::SETO: SSECC = 7; break;
6808 std::swap(Op0, Op1);
6810 // In the two special cases we can't handle, emit two comparisons.
6812 if (SetCCOpcode == ISD::SETUEQ) {
6814 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6815 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6816 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6818 else if (SetCCOpcode == ISD::SETONE) {
6820 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6821 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6822 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6824 llvm_unreachable("Illegal FP comparison");
6826 // Handle all other FP comparisons here.
6827 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6830 // We are handling one of the integer comparisons here. Since SSE only has
6831 // GT and EQ comparisons for integer, swapping operands and multiple
6832 // operations may be required for some comparisons.
6833 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6834 bool Swap = false, Invert = false, FlipSigns = false;
6836 switch (VT.getSimpleVT().SimpleTy) {
6839 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6841 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6843 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6844 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6847 switch (SetCCOpcode) {
6849 case ISD::SETNE: Invert = true;
6850 case ISD::SETEQ: Opc = EQOpc; break;
6851 case ISD::SETLT: Swap = true;
6852 case ISD::SETGT: Opc = GTOpc; break;
6853 case ISD::SETGE: Swap = true;
6854 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6855 case ISD::SETULT: Swap = true;
6856 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6857 case ISD::SETUGE: Swap = true;
6858 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6861 std::swap(Op0, Op1);
6863 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6864 // bits of the inputs before performing those operations.
6866 EVT EltVT = VT.getVectorElementType();
6867 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6869 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6870 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6872 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6873 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6876 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6878 // If the logical-not of the result is required, perform that now.
6880 Result = DAG.getNOT(dl, Result, VT);
6885 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6886 static bool isX86LogicalCmp(SDValue Op) {
6887 unsigned Opc = Op.getNode()->getOpcode();
6888 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6890 if (Op.getResNo() == 1 &&
6891 (Opc == X86ISD::ADD ||
6892 Opc == X86ISD::SUB ||
6893 Opc == X86ISD::SMUL ||
6894 Opc == X86ISD::UMUL ||
6895 Opc == X86ISD::INC ||
6896 Opc == X86ISD::DEC ||
6897 Opc == X86ISD::OR ||
6898 Opc == X86ISD::XOR ||
6899 Opc == X86ISD::AND))
6905 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6906 bool addTest = true;
6907 SDValue Cond = Op.getOperand(0);
6908 DebugLoc dl = Op.getDebugLoc();
6911 if (Cond.getOpcode() == ISD::SETCC) {
6912 SDValue NewCond = LowerSETCC(Cond, DAG);
6913 if (NewCond.getNode())
6917 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6918 SDValue Op1 = Op.getOperand(1);
6919 SDValue Op2 = Op.getOperand(2);
6920 if (Cond.getOpcode() == X86ISD::SETCC &&
6921 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6922 SDValue Cmp = Cond.getOperand(1);
6923 if (Cmp.getOpcode() == X86ISD::CMP) {
6924 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6925 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6926 ConstantSDNode *RHSC =
6927 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6928 if (N1C && N1C->isAllOnesValue() &&
6929 N2C && N2C->isNullValue() &&
6930 RHSC && RHSC->isNullValue()) {
6931 SDValue CmpOp0 = Cmp.getOperand(0);
6932 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6933 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6934 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6935 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6940 // Look pass (and (setcc_carry (cmp ...)), 1).
6941 if (Cond.getOpcode() == ISD::AND &&
6942 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6943 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6944 if (C && C->getAPIntValue() == 1)
6945 Cond = Cond.getOperand(0);
6948 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6949 // setting operand in place of the X86ISD::SETCC.
6950 if (Cond.getOpcode() == X86ISD::SETCC ||
6951 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6952 CC = Cond.getOperand(0);
6954 SDValue Cmp = Cond.getOperand(1);
6955 unsigned Opc = Cmp.getOpcode();
6956 EVT VT = Op.getValueType();
6958 bool IllegalFPCMov = false;
6959 if (VT.isFloatingPoint() && !VT.isVector() &&
6960 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6961 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6963 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6964 Opc == X86ISD::BT) { // FIXME
6971 // Look pass the truncate.
6972 if (Cond.getOpcode() == ISD::TRUNCATE)
6973 Cond = Cond.getOperand(0);
6975 // We know the result of AND is compared against zero. Try to match
6977 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6978 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6979 if (NewSetCC.getNode()) {
6980 CC = NewSetCC.getOperand(0);
6981 Cond = NewSetCC.getOperand(1);
6988 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6989 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6992 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6993 // condition is true.
6994 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6995 SDValue Ops[] = { Op2, Op1, CC, Cond };
6996 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6999 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7000 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7001 // from the AND / OR.
7002 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7003 Opc = Op.getOpcode();
7004 if (Opc != ISD::OR && Opc != ISD::AND)
7006 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7007 Op.getOperand(0).hasOneUse() &&
7008 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7009 Op.getOperand(1).hasOneUse());
7012 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7013 // 1 and that the SETCC node has a single use.
7014 static bool isXor1OfSetCC(SDValue Op) {
7015 if (Op.getOpcode() != ISD::XOR)
7017 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7018 if (N1C && N1C->getAPIntValue() == 1) {
7019 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7020 Op.getOperand(0).hasOneUse();
7025 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7026 bool addTest = true;
7027 SDValue Chain = Op.getOperand(0);
7028 SDValue Cond = Op.getOperand(1);
7029 SDValue Dest = Op.getOperand(2);
7030 DebugLoc dl = Op.getDebugLoc();
7033 if (Cond.getOpcode() == ISD::SETCC) {
7034 SDValue NewCond = LowerSETCC(Cond, DAG);
7035 if (NewCond.getNode())
7039 // FIXME: LowerXALUO doesn't handle these!!
7040 else if (Cond.getOpcode() == X86ISD::ADD ||
7041 Cond.getOpcode() == X86ISD::SUB ||
7042 Cond.getOpcode() == X86ISD::SMUL ||
7043 Cond.getOpcode() == X86ISD::UMUL)
7044 Cond = LowerXALUO(Cond, DAG);
7047 // Look pass (and (setcc_carry (cmp ...)), 1).
7048 if (Cond.getOpcode() == ISD::AND &&
7049 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7050 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7051 if (C && C->getAPIntValue() == 1)
7052 Cond = Cond.getOperand(0);
7055 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7056 // setting operand in place of the X86ISD::SETCC.
7057 if (Cond.getOpcode() == X86ISD::SETCC ||
7058 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7059 CC = Cond.getOperand(0);
7061 SDValue Cmp = Cond.getOperand(1);
7062 unsigned Opc = Cmp.getOpcode();
7063 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7064 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7068 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7072 // These can only come from an arithmetic instruction with overflow,
7073 // e.g. SADDO, UADDO.
7074 Cond = Cond.getNode()->getOperand(1);
7081 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7082 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7083 if (CondOpc == ISD::OR) {
7084 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7085 // two branches instead of an explicit OR instruction with a
7087 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7088 isX86LogicalCmp(Cmp)) {
7089 CC = Cond.getOperand(0).getOperand(0);
7090 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7091 Chain, Dest, CC, Cmp);
7092 CC = Cond.getOperand(1).getOperand(0);
7096 } else { // ISD::AND
7097 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7098 // two branches instead of an explicit AND instruction with a
7099 // separate test. However, we only do this if this block doesn't
7100 // have a fall-through edge, because this requires an explicit
7101 // jmp when the condition is false.
7102 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7103 isX86LogicalCmp(Cmp) &&
7104 Op.getNode()->hasOneUse()) {
7105 X86::CondCode CCode =
7106 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7107 CCode = X86::GetOppositeBranchCondition(CCode);
7108 CC = DAG.getConstant(CCode, MVT::i8);
7109 SDNode *User = *Op.getNode()->use_begin();
7110 // Look for an unconditional branch following this conditional branch.
7111 // We need this because we need to reverse the successors in order
7112 // to implement FCMP_OEQ.
7113 if (User->getOpcode() == ISD::BR) {
7114 SDValue FalseBB = User->getOperand(1);
7116 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
7117 assert(NewBR == User);
7121 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7122 Chain, Dest, CC, Cmp);
7123 X86::CondCode CCode =
7124 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
7125 CCode = X86::GetOppositeBranchCondition(CCode);
7126 CC = DAG.getConstant(CCode, MVT::i8);
7132 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
7133 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
7134 // It should be transformed during dag combiner except when the condition
7135 // is set by a arithmetics with overflow node.
7136 X86::CondCode CCode =
7137 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7138 CCode = X86::GetOppositeBranchCondition(CCode);
7139 CC = DAG.getConstant(CCode, MVT::i8);
7140 Cond = Cond.getOperand(0).getOperand(1);
7146 // Look pass the truncate.
7147 if (Cond.getOpcode() == ISD::TRUNCATE)
7148 Cond = Cond.getOperand(0);
7150 // We know the result of AND is compared against zero. Try to match
7152 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7153 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7154 if (NewSetCC.getNode()) {
7155 CC = NewSetCC.getOperand(0);
7156 Cond = NewSetCC.getOperand(1);
7163 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7164 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7166 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7167 Chain, Dest, CC, Cond);
7171 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
7172 // Calls to _alloca is needed to probe the stack when allocating more than 4k
7173 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
7174 // that the guard pages used by the OS virtual memory manager are allocated in
7175 // correct sequence.
7177 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
7178 SelectionDAG &DAG) const {
7179 assert(Subtarget->isTargetCygMing() &&
7180 "This should be used only on Cygwin/Mingw targets");
7181 DebugLoc dl = Op.getDebugLoc();
7184 SDValue Chain = Op.getOperand(0);
7185 SDValue Size = Op.getOperand(1);
7186 // FIXME: Ensure alignment here
7190 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
7192 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
7193 Flag = Chain.getValue(1);
7195 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
7197 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
7198 Flag = Chain.getValue(1);
7200 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
7202 SDValue Ops1[2] = { Chain.getValue(0), Chain };
7203 return DAG.getMergeValues(Ops1, 2, dl);
7206 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
7207 MachineFunction &MF = DAG.getMachineFunction();
7208 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
7210 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7211 DebugLoc dl = Op.getDebugLoc();
7213 if (!Subtarget->is64Bit()) {
7214 // vastart just stores the address of the VarArgsFrameIndex slot into the
7215 // memory location argument.
7216 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7218 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
7223 // gp_offset (0 - 6 * 8)
7224 // fp_offset (48 - 48 + 8 * 16)
7225 // overflow_arg_area (point to parameters coming in memory).
7227 SmallVector<SDValue, 8> MemOps;
7228 SDValue FIN = Op.getOperand(1);
7230 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
7231 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
7233 FIN, SV, 0, false, false, 0);
7234 MemOps.push_back(Store);
7237 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7238 FIN, DAG.getIntPtrConstant(4));
7239 Store = DAG.getStore(Op.getOperand(0), dl,
7240 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
7242 FIN, SV, 4, false, false, 0);
7243 MemOps.push_back(Store);
7245 // Store ptr to overflow_arg_area
7246 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7247 FIN, DAG.getIntPtrConstant(4));
7248 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7250 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 8,
7252 MemOps.push_back(Store);
7254 // Store ptr to reg_save_area.
7255 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7256 FIN, DAG.getIntPtrConstant(8));
7257 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
7259 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 16,
7261 MemOps.push_back(Store);
7262 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
7263 &MemOps[0], MemOps.size());
7266 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
7267 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7268 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
7270 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
7274 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
7275 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7276 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
7277 SDValue Chain = Op.getOperand(0);
7278 SDValue DstPtr = Op.getOperand(1);
7279 SDValue SrcPtr = Op.getOperand(2);
7280 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
7281 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7282 DebugLoc dl = Op.getDebugLoc();
7284 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
7285 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
7286 false, DstSV, 0, SrcSV, 0);
7290 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
7291 DebugLoc dl = Op.getDebugLoc();
7292 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7294 default: return SDValue(); // Don't custom lower most intrinsics.
7295 // Comparison intrinsics.
7296 case Intrinsic::x86_sse_comieq_ss:
7297 case Intrinsic::x86_sse_comilt_ss:
7298 case Intrinsic::x86_sse_comile_ss:
7299 case Intrinsic::x86_sse_comigt_ss:
7300 case Intrinsic::x86_sse_comige_ss:
7301 case Intrinsic::x86_sse_comineq_ss:
7302 case Intrinsic::x86_sse_ucomieq_ss:
7303 case Intrinsic::x86_sse_ucomilt_ss:
7304 case Intrinsic::x86_sse_ucomile_ss:
7305 case Intrinsic::x86_sse_ucomigt_ss:
7306 case Intrinsic::x86_sse_ucomige_ss:
7307 case Intrinsic::x86_sse_ucomineq_ss:
7308 case Intrinsic::x86_sse2_comieq_sd:
7309 case Intrinsic::x86_sse2_comilt_sd:
7310 case Intrinsic::x86_sse2_comile_sd:
7311 case Intrinsic::x86_sse2_comigt_sd:
7312 case Intrinsic::x86_sse2_comige_sd:
7313 case Intrinsic::x86_sse2_comineq_sd:
7314 case Intrinsic::x86_sse2_ucomieq_sd:
7315 case Intrinsic::x86_sse2_ucomilt_sd:
7316 case Intrinsic::x86_sse2_ucomile_sd:
7317 case Intrinsic::x86_sse2_ucomigt_sd:
7318 case Intrinsic::x86_sse2_ucomige_sd:
7319 case Intrinsic::x86_sse2_ucomineq_sd: {
7321 ISD::CondCode CC = ISD::SETCC_INVALID;
7324 case Intrinsic::x86_sse_comieq_ss:
7325 case Intrinsic::x86_sse2_comieq_sd:
7329 case Intrinsic::x86_sse_comilt_ss:
7330 case Intrinsic::x86_sse2_comilt_sd:
7334 case Intrinsic::x86_sse_comile_ss:
7335 case Intrinsic::x86_sse2_comile_sd:
7339 case Intrinsic::x86_sse_comigt_ss:
7340 case Intrinsic::x86_sse2_comigt_sd:
7344 case Intrinsic::x86_sse_comige_ss:
7345 case Intrinsic::x86_sse2_comige_sd:
7349 case Intrinsic::x86_sse_comineq_ss:
7350 case Intrinsic::x86_sse2_comineq_sd:
7354 case Intrinsic::x86_sse_ucomieq_ss:
7355 case Intrinsic::x86_sse2_ucomieq_sd:
7356 Opc = X86ISD::UCOMI;
7359 case Intrinsic::x86_sse_ucomilt_ss:
7360 case Intrinsic::x86_sse2_ucomilt_sd:
7361 Opc = X86ISD::UCOMI;
7364 case Intrinsic::x86_sse_ucomile_ss:
7365 case Intrinsic::x86_sse2_ucomile_sd:
7366 Opc = X86ISD::UCOMI;
7369 case Intrinsic::x86_sse_ucomigt_ss:
7370 case Intrinsic::x86_sse2_ucomigt_sd:
7371 Opc = X86ISD::UCOMI;
7374 case Intrinsic::x86_sse_ucomige_ss:
7375 case Intrinsic::x86_sse2_ucomige_sd:
7376 Opc = X86ISD::UCOMI;
7379 case Intrinsic::x86_sse_ucomineq_ss:
7380 case Intrinsic::x86_sse2_ucomineq_sd:
7381 Opc = X86ISD::UCOMI;
7386 SDValue LHS = Op.getOperand(1);
7387 SDValue RHS = Op.getOperand(2);
7388 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
7389 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
7390 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
7391 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7392 DAG.getConstant(X86CC, MVT::i8), Cond);
7393 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7395 // ptest and testp intrinsics. The intrinsic these come from are designed to
7396 // return an integer value, not just an instruction so lower it to the ptest
7397 // or testp pattern and a setcc for the result.
7398 case Intrinsic::x86_sse41_ptestz:
7399 case Intrinsic::x86_sse41_ptestc:
7400 case Intrinsic::x86_sse41_ptestnzc:
7401 case Intrinsic::x86_avx_ptestz_256:
7402 case Intrinsic::x86_avx_ptestc_256:
7403 case Intrinsic::x86_avx_ptestnzc_256:
7404 case Intrinsic::x86_avx_vtestz_ps:
7405 case Intrinsic::x86_avx_vtestc_ps:
7406 case Intrinsic::x86_avx_vtestnzc_ps:
7407 case Intrinsic::x86_avx_vtestz_pd:
7408 case Intrinsic::x86_avx_vtestc_pd:
7409 case Intrinsic::x86_avx_vtestnzc_pd:
7410 case Intrinsic::x86_avx_vtestz_ps_256:
7411 case Intrinsic::x86_avx_vtestc_ps_256:
7412 case Intrinsic::x86_avx_vtestnzc_ps_256:
7413 case Intrinsic::x86_avx_vtestz_pd_256:
7414 case Intrinsic::x86_avx_vtestc_pd_256:
7415 case Intrinsic::x86_avx_vtestnzc_pd_256: {
7416 bool IsTestPacked = false;
7419 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
7420 case Intrinsic::x86_avx_vtestz_ps:
7421 case Intrinsic::x86_avx_vtestz_pd:
7422 case Intrinsic::x86_avx_vtestz_ps_256:
7423 case Intrinsic::x86_avx_vtestz_pd_256:
7424 IsTestPacked = true; // Fallthrough
7425 case Intrinsic::x86_sse41_ptestz:
7426 case Intrinsic::x86_avx_ptestz_256:
7428 X86CC = X86::COND_E;
7430 case Intrinsic::x86_avx_vtestc_ps:
7431 case Intrinsic::x86_avx_vtestc_pd:
7432 case Intrinsic::x86_avx_vtestc_ps_256:
7433 case Intrinsic::x86_avx_vtestc_pd_256:
7434 IsTestPacked = true; // Fallthrough
7435 case Intrinsic::x86_sse41_ptestc:
7436 case Intrinsic::x86_avx_ptestc_256:
7438 X86CC = X86::COND_B;
7440 case Intrinsic::x86_avx_vtestnzc_ps:
7441 case Intrinsic::x86_avx_vtestnzc_pd:
7442 case Intrinsic::x86_avx_vtestnzc_ps_256:
7443 case Intrinsic::x86_avx_vtestnzc_pd_256:
7444 IsTestPacked = true; // Fallthrough
7445 case Intrinsic::x86_sse41_ptestnzc:
7446 case Intrinsic::x86_avx_ptestnzc_256:
7448 X86CC = X86::COND_A;
7452 SDValue LHS = Op.getOperand(1);
7453 SDValue RHS = Op.getOperand(2);
7454 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
7455 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
7456 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7457 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7458 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7461 // Fix vector shift instructions where the last operand is a non-immediate
7463 case Intrinsic::x86_sse2_pslli_w:
7464 case Intrinsic::x86_sse2_pslli_d:
7465 case Intrinsic::x86_sse2_pslli_q:
7466 case Intrinsic::x86_sse2_psrli_w:
7467 case Intrinsic::x86_sse2_psrli_d:
7468 case Intrinsic::x86_sse2_psrli_q:
7469 case Intrinsic::x86_sse2_psrai_w:
7470 case Intrinsic::x86_sse2_psrai_d:
7471 case Intrinsic::x86_mmx_pslli_w:
7472 case Intrinsic::x86_mmx_pslli_d:
7473 case Intrinsic::x86_mmx_pslli_q:
7474 case Intrinsic::x86_mmx_psrli_w:
7475 case Intrinsic::x86_mmx_psrli_d:
7476 case Intrinsic::x86_mmx_psrli_q:
7477 case Intrinsic::x86_mmx_psrai_w:
7478 case Intrinsic::x86_mmx_psrai_d: {
7479 SDValue ShAmt = Op.getOperand(2);
7480 if (isa<ConstantSDNode>(ShAmt))
7483 unsigned NewIntNo = 0;
7484 EVT ShAmtVT = MVT::v4i32;
7486 case Intrinsic::x86_sse2_pslli_w:
7487 NewIntNo = Intrinsic::x86_sse2_psll_w;
7489 case Intrinsic::x86_sse2_pslli_d:
7490 NewIntNo = Intrinsic::x86_sse2_psll_d;
7492 case Intrinsic::x86_sse2_pslli_q:
7493 NewIntNo = Intrinsic::x86_sse2_psll_q;
7495 case Intrinsic::x86_sse2_psrli_w:
7496 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7498 case Intrinsic::x86_sse2_psrli_d:
7499 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7501 case Intrinsic::x86_sse2_psrli_q:
7502 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7504 case Intrinsic::x86_sse2_psrai_w:
7505 NewIntNo = Intrinsic::x86_sse2_psra_w;
7507 case Intrinsic::x86_sse2_psrai_d:
7508 NewIntNo = Intrinsic::x86_sse2_psra_d;
7511 ShAmtVT = MVT::v2i32;
7513 case Intrinsic::x86_mmx_pslli_w:
7514 NewIntNo = Intrinsic::x86_mmx_psll_w;
7516 case Intrinsic::x86_mmx_pslli_d:
7517 NewIntNo = Intrinsic::x86_mmx_psll_d;
7519 case Intrinsic::x86_mmx_pslli_q:
7520 NewIntNo = Intrinsic::x86_mmx_psll_q;
7522 case Intrinsic::x86_mmx_psrli_w:
7523 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7525 case Intrinsic::x86_mmx_psrli_d:
7526 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7528 case Intrinsic::x86_mmx_psrli_q:
7529 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7531 case Intrinsic::x86_mmx_psrai_w:
7532 NewIntNo = Intrinsic::x86_mmx_psra_w;
7534 case Intrinsic::x86_mmx_psrai_d:
7535 NewIntNo = Intrinsic::x86_mmx_psra_d;
7537 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7543 // The vector shift intrinsics with scalars uses 32b shift amounts but
7544 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7548 ShOps[1] = DAG.getConstant(0, MVT::i32);
7549 if (ShAmtVT == MVT::v4i32) {
7550 ShOps[2] = DAG.getUNDEF(MVT::i32);
7551 ShOps[3] = DAG.getUNDEF(MVT::i32);
7552 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7554 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7557 EVT VT = Op.getValueType();
7558 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7559 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7560 DAG.getConstant(NewIntNo, MVT::i32),
7561 Op.getOperand(1), ShAmt);
7566 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7567 SelectionDAG &DAG) const {
7568 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7569 MFI->setReturnAddressIsTaken(true);
7571 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7572 DebugLoc dl = Op.getDebugLoc();
7575 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7577 DAG.getConstant(TD->getPointerSize(),
7578 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7579 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7580 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7582 NULL, 0, false, false, 0);
7585 // Just load the return address.
7586 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7587 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7588 RetAddrFI, NULL, 0, false, false, 0);
7591 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7592 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7593 MFI->setFrameAddressIsTaken(true);
7595 EVT VT = Op.getValueType();
7596 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7597 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7598 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7599 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7601 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7606 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7607 SelectionDAG &DAG) const {
7608 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7611 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7612 MachineFunction &MF = DAG.getMachineFunction();
7613 SDValue Chain = Op.getOperand(0);
7614 SDValue Offset = Op.getOperand(1);
7615 SDValue Handler = Op.getOperand(2);
7616 DebugLoc dl = Op.getDebugLoc();
7618 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
7619 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7621 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7623 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
7624 DAG.getIntPtrConstant(TD->getPointerSize()));
7625 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7626 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7627 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7628 MF.getRegInfo().addLiveOut(StoreAddrReg);
7630 return DAG.getNode(X86ISD::EH_RETURN, dl,
7632 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7635 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7636 SelectionDAG &DAG) const {
7637 SDValue Root = Op.getOperand(0);
7638 SDValue Trmp = Op.getOperand(1); // trampoline
7639 SDValue FPtr = Op.getOperand(2); // nested function
7640 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7641 DebugLoc dl = Op.getDebugLoc();
7643 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7645 if (Subtarget->is64Bit()) {
7646 SDValue OutChains[6];
7648 // Large code-model.
7649 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7650 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7652 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7653 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7655 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7657 // Load the pointer to the nested function into R11.
7658 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7659 SDValue Addr = Trmp;
7660 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7661 Addr, TrmpAddr, 0, false, false, 0);
7663 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7664 DAG.getConstant(2, MVT::i64));
7665 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7668 // Load the 'nest' parameter value into R10.
7669 // R10 is specified in X86CallingConv.td
7670 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7671 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7672 DAG.getConstant(10, MVT::i64));
7673 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7674 Addr, TrmpAddr, 10, false, false, 0);
7676 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7677 DAG.getConstant(12, MVT::i64));
7678 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7681 // Jump to the nested function.
7682 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7683 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7684 DAG.getConstant(20, MVT::i64));
7685 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7686 Addr, TrmpAddr, 20, false, false, 0);
7688 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7689 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7690 DAG.getConstant(22, MVT::i64));
7691 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7692 TrmpAddr, 22, false, false, 0);
7695 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7696 return DAG.getMergeValues(Ops, 2, dl);
7698 const Function *Func =
7699 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7700 CallingConv::ID CC = Func->getCallingConv();
7705 llvm_unreachable("Unsupported calling convention");
7706 case CallingConv::C:
7707 case CallingConv::X86_StdCall: {
7708 // Pass 'nest' parameter in ECX.
7709 // Must be kept in sync with X86CallingConv.td
7712 // Check that ECX wasn't needed by an 'inreg' parameter.
7713 const FunctionType *FTy = Func->getFunctionType();
7714 const AttrListPtr &Attrs = Func->getAttributes();
7716 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7717 unsigned InRegCount = 0;
7720 for (FunctionType::param_iterator I = FTy->param_begin(),
7721 E = FTy->param_end(); I != E; ++I, ++Idx)
7722 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7723 // FIXME: should only count parameters that are lowered to integers.
7724 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7726 if (InRegCount > 2) {
7727 report_fatal_error("Nest register in use - reduce number of inreg"
7733 case CallingConv::X86_FastCall:
7734 case CallingConv::X86_ThisCall:
7735 case CallingConv::Fast:
7736 // Pass 'nest' parameter in EAX.
7737 // Must be kept in sync with X86CallingConv.td
7742 SDValue OutChains[4];
7745 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7746 DAG.getConstant(10, MVT::i32));
7747 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7749 // This is storing the opcode for MOV32ri.
7750 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7751 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7752 OutChains[0] = DAG.getStore(Root, dl,
7753 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7754 Trmp, TrmpAddr, 0, false, false, 0);
7756 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7757 DAG.getConstant(1, MVT::i32));
7758 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7761 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7762 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7763 DAG.getConstant(5, MVT::i32));
7764 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7765 TrmpAddr, 5, false, false, 1);
7767 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7768 DAG.getConstant(6, MVT::i32));
7769 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7773 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7774 return DAG.getMergeValues(Ops, 2, dl);
7778 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7779 SelectionDAG &DAG) const {
7781 The rounding mode is in bits 11:10 of FPSR, and has the following
7788 FLT_ROUNDS, on the other hand, expects the following:
7795 To perform the conversion, we do:
7796 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7799 MachineFunction &MF = DAG.getMachineFunction();
7800 const TargetMachine &TM = MF.getTarget();
7801 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7802 unsigned StackAlignment = TFI.getStackAlignment();
7803 EVT VT = Op.getValueType();
7804 DebugLoc dl = Op.getDebugLoc();
7806 // Save FP Control Word to stack slot
7807 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7808 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7810 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7811 DAG.getEntryNode(), StackSlot);
7813 // Load FP Control Word from stack slot
7814 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7817 // Transform as necessary
7819 DAG.getNode(ISD::SRL, dl, MVT::i16,
7820 DAG.getNode(ISD::AND, dl, MVT::i16,
7821 CWD, DAG.getConstant(0x800, MVT::i16)),
7822 DAG.getConstant(11, MVT::i8));
7824 DAG.getNode(ISD::SRL, dl, MVT::i16,
7825 DAG.getNode(ISD::AND, dl, MVT::i16,
7826 CWD, DAG.getConstant(0x400, MVT::i16)),
7827 DAG.getConstant(9, MVT::i8));
7830 DAG.getNode(ISD::AND, dl, MVT::i16,
7831 DAG.getNode(ISD::ADD, dl, MVT::i16,
7832 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7833 DAG.getConstant(1, MVT::i16)),
7834 DAG.getConstant(3, MVT::i16));
7837 return DAG.getNode((VT.getSizeInBits() < 16 ?
7838 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7841 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7842 EVT VT = Op.getValueType();
7844 unsigned NumBits = VT.getSizeInBits();
7845 DebugLoc dl = Op.getDebugLoc();
7847 Op = Op.getOperand(0);
7848 if (VT == MVT::i8) {
7849 // Zero extend to i32 since there is not an i8 bsr.
7851 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7854 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7855 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7856 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7858 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7861 DAG.getConstant(NumBits+NumBits-1, OpVT),
7862 DAG.getConstant(X86::COND_E, MVT::i8),
7865 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7867 // Finally xor with NumBits-1.
7868 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7871 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7875 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7876 EVT VT = Op.getValueType();
7878 unsigned NumBits = VT.getSizeInBits();
7879 DebugLoc dl = Op.getDebugLoc();
7881 Op = Op.getOperand(0);
7882 if (VT == MVT::i8) {
7884 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7887 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7888 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7889 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7891 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7894 DAG.getConstant(NumBits, OpVT),
7895 DAG.getConstant(X86::COND_E, MVT::i8),
7898 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7901 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7905 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7906 EVT VT = Op.getValueType();
7907 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7908 DebugLoc dl = Op.getDebugLoc();
7910 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7911 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7912 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7913 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7914 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7916 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7917 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7918 // return AloBlo + AloBhi + AhiBlo;
7920 SDValue A = Op.getOperand(0);
7921 SDValue B = Op.getOperand(1);
7923 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7924 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7925 A, DAG.getConstant(32, MVT::i32));
7926 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7927 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7928 B, DAG.getConstant(32, MVT::i32));
7929 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7930 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7932 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7933 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7935 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7936 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7938 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7939 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7940 AloBhi, DAG.getConstant(32, MVT::i32));
7941 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7942 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7943 AhiBlo, DAG.getConstant(32, MVT::i32));
7944 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7945 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7949 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
7950 EVT VT = Op.getValueType();
7951 DebugLoc dl = Op.getDebugLoc();
7952 SDValue R = Op.getOperand(0);
7954 LLVMContext *Context = DAG.getContext();
7956 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
7958 if (VT == MVT::v4i32) {
7959 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7960 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
7961 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
7963 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
7965 std::vector<Constant*> CV(4, CI);
7966 Constant *C = ConstantVector::get(CV);
7967 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7968 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7969 PseudoSourceValue::getConstantPool(), 0,
7972 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
7973 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, Op);
7974 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
7975 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
7977 if (VT == MVT::v16i8) {
7979 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7980 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
7981 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
7983 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
7984 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
7986 std::vector<Constant*> CVM1(16, CM1);
7987 std::vector<Constant*> CVM2(16, CM2);
7988 Constant *C = ConstantVector::get(CVM1);
7989 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7990 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7991 PseudoSourceValue::getConstantPool(), 0,
7994 // r = pblendv(r, psllw(r & (char16)15, 4), a);
7995 M = DAG.getNode(ISD::AND, dl, VT, R, M);
7996 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7997 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
7998 DAG.getConstant(4, MVT::i32));
7999 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8000 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8003 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8005 C = ConstantVector::get(CVM2);
8006 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8007 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8008 PseudoSourceValue::getConstantPool(), 0, false, false, 16);
8010 // r = pblendv(r, psllw(r & (char16)63, 2), a);
8011 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8012 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8013 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8014 DAG.getConstant(2, MVT::i32));
8015 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8016 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8019 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8021 // return pblendv(r, r+r, a);
8022 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8023 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8024 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
8030 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
8031 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
8032 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
8033 // looks for this combo and may remove the "setcc" instruction if the "setcc"
8034 // has only one use.
8035 SDNode *N = Op.getNode();
8036 SDValue LHS = N->getOperand(0);
8037 SDValue RHS = N->getOperand(1);
8038 unsigned BaseOp = 0;
8040 DebugLoc dl = Op.getDebugLoc();
8042 switch (Op.getOpcode()) {
8043 default: llvm_unreachable("Unknown ovf instruction!");
8045 // A subtract of one will be selected as a INC. Note that INC doesn't
8046 // set CF, so we can't do this for UADDO.
8047 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8048 if (C->getAPIntValue() == 1) {
8049 BaseOp = X86ISD::INC;
8053 BaseOp = X86ISD::ADD;
8057 BaseOp = X86ISD::ADD;
8061 // A subtract of one will be selected as a DEC. Note that DEC doesn't
8062 // set CF, so we can't do this for USUBO.
8063 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8064 if (C->getAPIntValue() == 1) {
8065 BaseOp = X86ISD::DEC;
8069 BaseOp = X86ISD::SUB;
8073 BaseOp = X86ISD::SUB;
8077 BaseOp = X86ISD::SMUL;
8081 BaseOp = X86ISD::UMUL;
8086 // Also sets EFLAGS.
8087 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
8088 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
8091 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
8092 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
8094 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8098 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
8099 DebugLoc dl = Op.getDebugLoc();
8101 if (!Subtarget->hasSSE2()) {
8102 SDValue Chain = Op.getOperand(0);
8103 SDValue Zero = DAG.getConstant(0,
8104 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8106 DAG.getRegister(X86::ESP, MVT::i32), // Base
8107 DAG.getTargetConstant(1, MVT::i8), // Scale
8108 DAG.getRegister(0, MVT::i32), // Index
8109 DAG.getTargetConstant(0, MVT::i32), // Disp
8110 DAG.getRegister(0, MVT::i32), // Segment.
8115 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
8116 array_lengthof(Ops));
8117 return SDValue(Res, 0);
8120 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
8122 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
8124 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
8125 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
8126 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
8127 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
8129 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
8130 if (!Op1 && !Op2 && !Op3 && Op4)
8131 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
8133 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
8134 if (Op1 && !Op2 && !Op3 && !Op4)
8135 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
8137 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
8139 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
8142 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
8143 EVT T = Op.getValueType();
8144 DebugLoc dl = Op.getDebugLoc();
8147 switch(T.getSimpleVT().SimpleTy) {
8149 assert(false && "Invalid value type!");
8150 case MVT::i8: Reg = X86::AL; size = 1; break;
8151 case MVT::i16: Reg = X86::AX; size = 2; break;
8152 case MVT::i32: Reg = X86::EAX; size = 4; break;
8154 assert(Subtarget->is64Bit() && "Node not type legal!");
8155 Reg = X86::RAX; size = 8;
8158 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
8159 Op.getOperand(2), SDValue());
8160 SDValue Ops[] = { cpIn.getValue(0),
8163 DAG.getTargetConstant(size, MVT::i8),
8165 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8166 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
8168 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
8172 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
8173 SelectionDAG &DAG) const {
8174 assert(Subtarget->is64Bit() && "Result not type legalized?");
8175 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8176 SDValue TheChain = Op.getOperand(0);
8177 DebugLoc dl = Op.getDebugLoc();
8178 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8179 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
8180 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
8182 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
8183 DAG.getConstant(32, MVT::i8));
8185 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
8188 return DAG.getMergeValues(Ops, 2, dl);
8191 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
8192 SelectionDAG &DAG) const {
8193 EVT SrcVT = Op.getOperand(0).getValueType();
8194 EVT DstVT = Op.getValueType();
8195 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
8196 Subtarget->hasMMX() && !DisableMMX) &&
8197 "Unexpected custom BIT_CONVERT");
8198 assert((DstVT == MVT::i64 ||
8199 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
8200 "Unexpected custom BIT_CONVERT");
8201 // i64 <=> MMX conversions are Legal.
8202 if (SrcVT==MVT::i64 && DstVT.isVector())
8204 if (DstVT==MVT::i64 && SrcVT.isVector())
8206 // MMX <=> MMX conversions are Legal.
8207 if (SrcVT.isVector() && DstVT.isVector())
8209 // All other conversions need to be expanded.
8212 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
8213 SDNode *Node = Op.getNode();
8214 DebugLoc dl = Node->getDebugLoc();
8215 EVT T = Node->getValueType(0);
8216 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
8217 DAG.getConstant(0, T), Node->getOperand(2));
8218 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
8219 cast<AtomicSDNode>(Node)->getMemoryVT(),
8220 Node->getOperand(0),
8221 Node->getOperand(1), negOp,
8222 cast<AtomicSDNode>(Node)->getSrcValue(),
8223 cast<AtomicSDNode>(Node)->getAlignment());
8226 /// LowerOperation - Provide custom lowering hooks for some operations.
8228 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
8229 switch (Op.getOpcode()) {
8230 default: llvm_unreachable("Should not custom lower this!");
8231 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
8232 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
8233 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
8234 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
8235 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
8236 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
8237 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
8238 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
8239 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
8240 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
8241 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
8242 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
8243 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
8244 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
8245 case ISD::SHL_PARTS:
8246 case ISD::SRA_PARTS:
8247 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
8248 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
8249 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
8250 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
8251 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
8252 case ISD::FABS: return LowerFABS(Op, DAG);
8253 case ISD::FNEG: return LowerFNEG(Op, DAG);
8254 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
8255 case ISD::SETCC: return LowerSETCC(Op, DAG);
8256 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
8257 case ISD::SELECT: return LowerSELECT(Op, DAG);
8258 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
8259 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
8260 case ISD::VASTART: return LowerVASTART(Op, DAG);
8261 case ISD::VAARG: return LowerVAARG(Op, DAG);
8262 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
8263 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
8264 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
8265 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
8266 case ISD::FRAME_TO_ARGS_OFFSET:
8267 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
8268 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
8269 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
8270 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
8271 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
8272 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
8273 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
8274 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
8275 case ISD::SHL: return LowerSHL(Op, DAG);
8281 case ISD::UMULO: return LowerXALUO(Op, DAG);
8282 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
8283 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
8287 void X86TargetLowering::
8288 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
8289 SelectionDAG &DAG, unsigned NewOp) const {
8290 EVT T = Node->getValueType(0);
8291 DebugLoc dl = Node->getDebugLoc();
8292 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
8294 SDValue Chain = Node->getOperand(0);
8295 SDValue In1 = Node->getOperand(1);
8296 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8297 Node->getOperand(2), DAG.getIntPtrConstant(0));
8298 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8299 Node->getOperand(2), DAG.getIntPtrConstant(1));
8300 SDValue Ops[] = { Chain, In1, In2L, In2H };
8301 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
8303 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
8304 cast<MemSDNode>(Node)->getMemOperand());
8305 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
8306 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8307 Results.push_back(Result.getValue(2));
8310 /// ReplaceNodeResults - Replace a node with an illegal result type
8311 /// with a new node built out of custom code.
8312 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
8313 SmallVectorImpl<SDValue>&Results,
8314 SelectionDAG &DAG) const {
8315 DebugLoc dl = N->getDebugLoc();
8316 switch (N->getOpcode()) {
8318 assert(false && "Do not know how to custom type legalize this operation!");
8320 case ISD::FP_TO_SINT: {
8321 std::pair<SDValue,SDValue> Vals =
8322 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
8323 SDValue FIST = Vals.first, StackSlot = Vals.second;
8324 if (FIST.getNode() != 0) {
8325 EVT VT = N->getValueType(0);
8326 // Return a load from the stack slot.
8327 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
8332 case ISD::READCYCLECOUNTER: {
8333 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8334 SDValue TheChain = N->getOperand(0);
8335 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8336 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
8338 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
8340 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
8341 SDValue Ops[] = { eax, edx };
8342 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
8343 Results.push_back(edx.getValue(1));
8346 case ISD::ATOMIC_CMP_SWAP: {
8347 EVT T = N->getValueType(0);
8348 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
8349 SDValue cpInL, cpInH;
8350 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8351 DAG.getConstant(0, MVT::i32));
8352 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8353 DAG.getConstant(1, MVT::i32));
8354 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
8355 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
8357 SDValue swapInL, swapInH;
8358 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8359 DAG.getConstant(0, MVT::i32));
8360 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8361 DAG.getConstant(1, MVT::i32));
8362 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
8364 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
8365 swapInL.getValue(1));
8366 SDValue Ops[] = { swapInH.getValue(0),
8368 swapInH.getValue(1) };
8369 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8370 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
8371 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
8372 MVT::i32, Result.getValue(1));
8373 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
8374 MVT::i32, cpOutL.getValue(2));
8375 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
8376 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8377 Results.push_back(cpOutH.getValue(1));
8380 case ISD::ATOMIC_LOAD_ADD:
8381 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
8383 case ISD::ATOMIC_LOAD_AND:
8384 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
8386 case ISD::ATOMIC_LOAD_NAND:
8387 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
8389 case ISD::ATOMIC_LOAD_OR:
8390 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
8392 case ISD::ATOMIC_LOAD_SUB:
8393 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
8395 case ISD::ATOMIC_LOAD_XOR:
8396 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
8398 case ISD::ATOMIC_SWAP:
8399 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
8404 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
8406 default: return NULL;
8407 case X86ISD::BSF: return "X86ISD::BSF";
8408 case X86ISD::BSR: return "X86ISD::BSR";
8409 case X86ISD::SHLD: return "X86ISD::SHLD";
8410 case X86ISD::SHRD: return "X86ISD::SHRD";
8411 case X86ISD::FAND: return "X86ISD::FAND";
8412 case X86ISD::FOR: return "X86ISD::FOR";
8413 case X86ISD::FXOR: return "X86ISD::FXOR";
8414 case X86ISD::FSRL: return "X86ISD::FSRL";
8415 case X86ISD::FILD: return "X86ISD::FILD";
8416 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
8417 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
8418 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
8419 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
8420 case X86ISD::FLD: return "X86ISD::FLD";
8421 case X86ISD::FST: return "X86ISD::FST";
8422 case X86ISD::CALL: return "X86ISD::CALL";
8423 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
8424 case X86ISD::BT: return "X86ISD::BT";
8425 case X86ISD::CMP: return "X86ISD::CMP";
8426 case X86ISD::COMI: return "X86ISD::COMI";
8427 case X86ISD::UCOMI: return "X86ISD::UCOMI";
8428 case X86ISD::SETCC: return "X86ISD::SETCC";
8429 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
8430 case X86ISD::CMOV: return "X86ISD::CMOV";
8431 case X86ISD::BRCOND: return "X86ISD::BRCOND";
8432 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
8433 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
8434 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
8435 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
8436 case X86ISD::Wrapper: return "X86ISD::Wrapper";
8437 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
8438 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
8439 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
8440 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
8441 case X86ISD::PINSRB: return "X86ISD::PINSRB";
8442 case X86ISD::PINSRW: return "X86ISD::PINSRW";
8443 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
8444 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
8445 case X86ISD::FMAX: return "X86ISD::FMAX";
8446 case X86ISD::FMIN: return "X86ISD::FMIN";
8447 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
8448 case X86ISD::FRCP: return "X86ISD::FRCP";
8449 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
8450 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
8451 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
8452 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
8453 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
8454 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
8455 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
8456 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
8457 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
8458 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
8459 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
8460 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
8461 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
8462 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
8463 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
8464 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
8465 case X86ISD::VSHL: return "X86ISD::VSHL";
8466 case X86ISD::VSRL: return "X86ISD::VSRL";
8467 case X86ISD::CMPPD: return "X86ISD::CMPPD";
8468 case X86ISD::CMPPS: return "X86ISD::CMPPS";
8469 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
8470 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
8471 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
8472 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
8473 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
8474 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
8475 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
8476 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
8477 case X86ISD::ADD: return "X86ISD::ADD";
8478 case X86ISD::SUB: return "X86ISD::SUB";
8479 case X86ISD::SMUL: return "X86ISD::SMUL";
8480 case X86ISD::UMUL: return "X86ISD::UMUL";
8481 case X86ISD::INC: return "X86ISD::INC";
8482 case X86ISD::DEC: return "X86ISD::DEC";
8483 case X86ISD::OR: return "X86ISD::OR";
8484 case X86ISD::XOR: return "X86ISD::XOR";
8485 case X86ISD::AND: return "X86ISD::AND";
8486 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
8487 case X86ISD::PTEST: return "X86ISD::PTEST";
8488 case X86ISD::TESTP: return "X86ISD::TESTP";
8489 case X86ISD::PALIGN: return "X86ISD::PALIGN";
8490 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
8491 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
8492 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
8493 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
8494 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
8495 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
8496 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
8497 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
8498 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
8499 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
8500 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
8501 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
8502 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
8503 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
8504 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
8505 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
8506 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
8507 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
8508 case X86ISD::MOVSD: return "X86ISD::MOVSD";
8509 case X86ISD::MOVSS: return "X86ISD::MOVSS";
8510 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
8511 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
8512 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
8513 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
8514 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
8515 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
8516 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
8517 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
8518 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
8519 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
8520 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
8521 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
8522 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
8523 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
8527 // isLegalAddressingMode - Return true if the addressing mode represented
8528 // by AM is legal for this target, for a load/store of the specified type.
8529 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
8530 const Type *Ty) const {
8531 // X86 supports extremely general addressing modes.
8532 CodeModel::Model M = getTargetMachine().getCodeModel();
8533 Reloc::Model R = getTargetMachine().getRelocationModel();
8535 // X86 allows a sign-extended 32-bit immediate field as a displacement.
8536 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
8541 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
8543 // If a reference to this global requires an extra load, we can't fold it.
8544 if (isGlobalStubReference(GVFlags))
8547 // If BaseGV requires a register for the PIC base, we cannot also have a
8548 // BaseReg specified.
8549 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
8552 // If lower 4G is not available, then we must use rip-relative addressing.
8553 if ((M != CodeModel::Small || R != Reloc::Static) &&
8554 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
8564 // These scales always work.
8569 // These scales are formed with basereg+scalereg. Only accept if there is
8574 default: // Other stuff never works.
8582 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
8583 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
8585 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
8586 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
8587 if (NumBits1 <= NumBits2)
8592 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
8593 if (!VT1.isInteger() || !VT2.isInteger())
8595 unsigned NumBits1 = VT1.getSizeInBits();
8596 unsigned NumBits2 = VT2.getSizeInBits();
8597 if (NumBits1 <= NumBits2)
8602 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
8603 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8604 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
8607 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
8608 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8609 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
8612 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
8613 // i16 instructions are longer (0x66 prefix) and potentially slower.
8614 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
8617 /// isShuffleMaskLegal - Targets can use this to indicate that they only
8618 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
8619 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
8620 /// are assumed to be legal.
8622 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
8624 // Very little shuffling can be done for 64-bit vectors right now.
8625 if (VT.getSizeInBits() == 64)
8626 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8628 // FIXME: pshufb, blends, shifts.
8629 return (VT.getVectorNumElements() == 2 ||
8630 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8631 isMOVLMask(M, VT) ||
8632 isSHUFPMask(M, VT) ||
8633 isPSHUFDMask(M, VT) ||
8634 isPSHUFHWMask(M, VT) ||
8635 isPSHUFLWMask(M, VT) ||
8636 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8637 isUNPCKLMask(M, VT) ||
8638 isUNPCKHMask(M, VT) ||
8639 isUNPCKL_v_undef_Mask(M, VT) ||
8640 isUNPCKH_v_undef_Mask(M, VT));
8644 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8646 unsigned NumElts = VT.getVectorNumElements();
8647 // FIXME: This collection of masks seems suspect.
8650 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8651 return (isMOVLMask(Mask, VT) ||
8652 isCommutedMOVLMask(Mask, VT, true) ||
8653 isSHUFPMask(Mask, VT) ||
8654 isCommutedSHUFPMask(Mask, VT));
8659 //===----------------------------------------------------------------------===//
8660 // X86 Scheduler Hooks
8661 //===----------------------------------------------------------------------===//
8663 // private utility function
8665 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8666 MachineBasicBlock *MBB,
8673 TargetRegisterClass *RC,
8674 bool invSrc) const {
8675 // For the atomic bitwise operator, we generate
8678 // ld t1 = [bitinstr.addr]
8679 // op t2 = t1, [bitinstr.val]
8681 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8683 // fallthrough -->nextMBB
8684 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8685 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8686 MachineFunction::iterator MBBIter = MBB;
8689 /// First build the CFG
8690 MachineFunction *F = MBB->getParent();
8691 MachineBasicBlock *thisMBB = MBB;
8692 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8693 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8694 F->insert(MBBIter, newMBB);
8695 F->insert(MBBIter, nextMBB);
8697 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8698 nextMBB->splice(nextMBB->begin(), thisMBB,
8699 llvm::next(MachineBasicBlock::iterator(bInstr)),
8701 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8703 // Update thisMBB to fall through to newMBB
8704 thisMBB->addSuccessor(newMBB);
8706 // newMBB jumps to itself and fall through to nextMBB
8707 newMBB->addSuccessor(nextMBB);
8708 newMBB->addSuccessor(newMBB);
8710 // Insert instructions into newMBB based on incoming instruction
8711 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8712 "unexpected number of operands");
8713 DebugLoc dl = bInstr->getDebugLoc();
8714 MachineOperand& destOper = bInstr->getOperand(0);
8715 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8716 int numArgs = bInstr->getNumOperands() - 1;
8717 for (int i=0; i < numArgs; ++i)
8718 argOpers[i] = &bInstr->getOperand(i+1);
8720 // x86 address has 4 operands: base, index, scale, and displacement
8721 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8722 int valArgIndx = lastAddrIndx + 1;
8724 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8725 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8726 for (int i=0; i <= lastAddrIndx; ++i)
8727 (*MIB).addOperand(*argOpers[i]);
8729 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8731 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8736 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8737 assert((argOpers[valArgIndx]->isReg() ||
8738 argOpers[valArgIndx]->isImm()) &&
8740 if (argOpers[valArgIndx]->isReg())
8741 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8743 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8745 (*MIB).addOperand(*argOpers[valArgIndx]);
8747 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
8750 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8751 for (int i=0; i <= lastAddrIndx; ++i)
8752 (*MIB).addOperand(*argOpers[i]);
8754 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8755 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8756 bInstr->memoperands_end());
8758 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8762 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8764 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8768 // private utility function: 64 bit atomics on 32 bit host.
8770 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8771 MachineBasicBlock *MBB,
8776 bool invSrc) const {
8777 // For the atomic bitwise operator, we generate
8778 // thisMBB (instructions are in pairs, except cmpxchg8b)
8779 // ld t1,t2 = [bitinstr.addr]
8781 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8782 // op t5, t6 <- out1, out2, [bitinstr.val]
8783 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8784 // mov ECX, EBX <- t5, t6
8785 // mov EAX, EDX <- t1, t2
8786 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8787 // mov t3, t4 <- EAX, EDX
8789 // result in out1, out2
8790 // fallthrough -->nextMBB
8792 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8793 const unsigned LoadOpc = X86::MOV32rm;
8794 const unsigned NotOpc = X86::NOT32r;
8795 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8796 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8797 MachineFunction::iterator MBBIter = MBB;
8800 /// First build the CFG
8801 MachineFunction *F = MBB->getParent();
8802 MachineBasicBlock *thisMBB = MBB;
8803 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8804 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8805 F->insert(MBBIter, newMBB);
8806 F->insert(MBBIter, nextMBB);
8808 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8809 nextMBB->splice(nextMBB->begin(), thisMBB,
8810 llvm::next(MachineBasicBlock::iterator(bInstr)),
8812 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8814 // Update thisMBB to fall through to newMBB
8815 thisMBB->addSuccessor(newMBB);
8817 // newMBB jumps to itself and fall through to nextMBB
8818 newMBB->addSuccessor(nextMBB);
8819 newMBB->addSuccessor(newMBB);
8821 DebugLoc dl = bInstr->getDebugLoc();
8822 // Insert instructions into newMBB based on incoming instruction
8823 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8824 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
8825 "unexpected number of operands");
8826 MachineOperand& dest1Oper = bInstr->getOperand(0);
8827 MachineOperand& dest2Oper = bInstr->getOperand(1);
8828 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8829 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
8830 argOpers[i] = &bInstr->getOperand(i+2);
8832 // We use some of the operands multiple times, so conservatively just
8833 // clear any kill flags that might be present.
8834 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8835 argOpers[i]->setIsKill(false);
8838 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8839 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8841 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8842 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8843 for (int i=0; i <= lastAddrIndx; ++i)
8844 (*MIB).addOperand(*argOpers[i]);
8845 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8846 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8847 // add 4 to displacement.
8848 for (int i=0; i <= lastAddrIndx-2; ++i)
8849 (*MIB).addOperand(*argOpers[i]);
8850 MachineOperand newOp3 = *(argOpers[3]);
8852 newOp3.setImm(newOp3.getImm()+4);
8854 newOp3.setOffset(newOp3.getOffset()+4);
8855 (*MIB).addOperand(newOp3);
8856 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8858 // t3/4 are defined later, at the bottom of the loop
8859 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8860 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8861 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8862 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8863 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8864 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8866 // The subsequent operations should be using the destination registers of
8867 //the PHI instructions.
8869 t1 = F->getRegInfo().createVirtualRegister(RC);
8870 t2 = F->getRegInfo().createVirtualRegister(RC);
8871 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8872 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8874 t1 = dest1Oper.getReg();
8875 t2 = dest2Oper.getReg();
8878 int valArgIndx = lastAddrIndx + 1;
8879 assert((argOpers[valArgIndx]->isReg() ||
8880 argOpers[valArgIndx]->isImm()) &&
8882 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8883 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8884 if (argOpers[valArgIndx]->isReg())
8885 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8887 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8888 if (regOpcL != X86::MOV32rr)
8890 (*MIB).addOperand(*argOpers[valArgIndx]);
8891 assert(argOpers[valArgIndx + 1]->isReg() ==
8892 argOpers[valArgIndx]->isReg());
8893 assert(argOpers[valArgIndx + 1]->isImm() ==
8894 argOpers[valArgIndx]->isImm());
8895 if (argOpers[valArgIndx + 1]->isReg())
8896 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8898 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8899 if (regOpcH != X86::MOV32rr)
8901 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8903 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8905 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
8908 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
8910 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
8913 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8914 for (int i=0; i <= lastAddrIndx; ++i)
8915 (*MIB).addOperand(*argOpers[i]);
8917 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8918 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8919 bInstr->memoperands_end());
8921 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
8922 MIB.addReg(X86::EAX);
8923 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
8924 MIB.addReg(X86::EDX);
8927 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8929 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8933 // private utility function
8935 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8936 MachineBasicBlock *MBB,
8937 unsigned cmovOpc) const {
8938 // For the atomic min/max operator, we generate
8941 // ld t1 = [min/max.addr]
8942 // mov t2 = [min/max.val]
8944 // cmov[cond] t2 = t1
8946 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8948 // fallthrough -->nextMBB
8950 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8951 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8952 MachineFunction::iterator MBBIter = MBB;
8955 /// First build the CFG
8956 MachineFunction *F = MBB->getParent();
8957 MachineBasicBlock *thisMBB = MBB;
8958 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8959 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8960 F->insert(MBBIter, newMBB);
8961 F->insert(MBBIter, nextMBB);
8963 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8964 nextMBB->splice(nextMBB->begin(), thisMBB,
8965 llvm::next(MachineBasicBlock::iterator(mInstr)),
8967 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8969 // Update thisMBB to fall through to newMBB
8970 thisMBB->addSuccessor(newMBB);
8972 // newMBB jumps to newMBB and fall through to nextMBB
8973 newMBB->addSuccessor(nextMBB);
8974 newMBB->addSuccessor(newMBB);
8976 DebugLoc dl = mInstr->getDebugLoc();
8977 // Insert instructions into newMBB based on incoming instruction
8978 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8979 "unexpected number of operands");
8980 MachineOperand& destOper = mInstr->getOperand(0);
8981 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8982 int numArgs = mInstr->getNumOperands() - 1;
8983 for (int i=0; i < numArgs; ++i)
8984 argOpers[i] = &mInstr->getOperand(i+1);
8986 // x86 address has 4 operands: base, index, scale, and displacement
8987 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8988 int valArgIndx = lastAddrIndx + 1;
8990 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8991 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8992 for (int i=0; i <= lastAddrIndx; ++i)
8993 (*MIB).addOperand(*argOpers[i]);
8995 // We only support register and immediate values
8996 assert((argOpers[valArgIndx]->isReg() ||
8997 argOpers[valArgIndx]->isImm()) &&
9000 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9001 if (argOpers[valArgIndx]->isReg())
9002 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
9004 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
9005 (*MIB).addOperand(*argOpers[valArgIndx]);
9007 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9010 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
9015 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9016 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
9020 // Cmp and exchange if none has modified the memory location
9021 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
9022 for (int i=0; i <= lastAddrIndx; ++i)
9023 (*MIB).addOperand(*argOpers[i]);
9025 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9026 (*MIB).setMemRefs(mInstr->memoperands_begin(),
9027 mInstr->memoperands_end());
9029 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9030 MIB.addReg(X86::EAX);
9033 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9035 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
9039 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
9040 // or XMM0_V32I8 in AVX all of this code can be replaced with that
9043 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
9044 unsigned numArgs, bool memArg) const {
9046 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
9047 "Target must have SSE4.2 or AVX features enabled");
9049 DebugLoc dl = MI->getDebugLoc();
9050 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9054 if (!Subtarget->hasAVX()) {
9056 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
9058 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
9061 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
9063 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
9066 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
9068 for (unsigned i = 0; i < numArgs; ++i) {
9069 MachineOperand &Op = MI->getOperand(i+1);
9071 if (!(Op.isReg() && Op.isImplicit()))
9075 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
9078 MI->eraseFromParent();
9084 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
9086 MachineBasicBlock *MBB) const {
9087 // Emit code to save XMM registers to the stack. The ABI says that the
9088 // number of registers to save is given in %al, so it's theoretically
9089 // possible to do an indirect jump trick to avoid saving all of them,
9090 // however this code takes a simpler approach and just executes all
9091 // of the stores if %al is non-zero. It's less code, and it's probably
9092 // easier on the hardware branch predictor, and stores aren't all that
9093 // expensive anyway.
9095 // Create the new basic blocks. One block contains all the XMM stores,
9096 // and one block is the final destination regardless of whether any
9097 // stores were performed.
9098 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9099 MachineFunction *F = MBB->getParent();
9100 MachineFunction::iterator MBBIter = MBB;
9102 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
9103 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
9104 F->insert(MBBIter, XMMSaveMBB);
9105 F->insert(MBBIter, EndMBB);
9107 // Transfer the remainder of MBB and its successor edges to EndMBB.
9108 EndMBB->splice(EndMBB->begin(), MBB,
9109 llvm::next(MachineBasicBlock::iterator(MI)),
9111 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
9113 // The original block will now fall through to the XMM save block.
9114 MBB->addSuccessor(XMMSaveMBB);
9115 // The XMMSaveMBB will fall through to the end block.
9116 XMMSaveMBB->addSuccessor(EndMBB);
9118 // Now add the instructions.
9119 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9120 DebugLoc DL = MI->getDebugLoc();
9122 unsigned CountReg = MI->getOperand(0).getReg();
9123 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
9124 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
9126 if (!Subtarget->isTargetWin64()) {
9127 // If %al is 0, branch around the XMM save block.
9128 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
9129 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
9130 MBB->addSuccessor(EndMBB);
9133 // In the XMM save block, save all the XMM argument registers.
9134 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
9135 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
9136 MachineMemOperand *MMO =
9137 F->getMachineMemOperand(
9138 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
9139 MachineMemOperand::MOStore, Offset,
9140 /*Size=*/16, /*Align=*/16);
9141 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
9142 .addFrameIndex(RegSaveFrameIndex)
9143 .addImm(/*Scale=*/1)
9144 .addReg(/*IndexReg=*/0)
9145 .addImm(/*Disp=*/Offset)
9146 .addReg(/*Segment=*/0)
9147 .addReg(MI->getOperand(i).getReg())
9148 .addMemOperand(MMO);
9151 MI->eraseFromParent(); // The pseudo instruction is gone now.
9157 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
9158 MachineBasicBlock *BB) const {
9159 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9160 DebugLoc DL = MI->getDebugLoc();
9162 // To "insert" a SELECT_CC instruction, we actually have to insert the
9163 // diamond control-flow pattern. The incoming instruction knows the
9164 // destination vreg to set, the condition code register to branch on, the
9165 // true/false values to select between, and a branch opcode to use.
9166 const BasicBlock *LLVM_BB = BB->getBasicBlock();
9167 MachineFunction::iterator It = BB;
9173 // cmpTY ccX, r1, r2
9175 // fallthrough --> copy0MBB
9176 MachineBasicBlock *thisMBB = BB;
9177 MachineFunction *F = BB->getParent();
9178 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
9179 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
9180 F->insert(It, copy0MBB);
9181 F->insert(It, sinkMBB);
9183 // If the EFLAGS register isn't dead in the terminator, then claim that it's
9184 // live into the sink and copy blocks.
9185 const MachineFunction *MF = BB->getParent();
9186 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
9187 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
9189 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
9190 const MachineOperand &MO = MI->getOperand(I);
9191 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
9192 unsigned Reg = MO.getReg();
9193 if (Reg != X86::EFLAGS) continue;
9194 copy0MBB->addLiveIn(Reg);
9195 sinkMBB->addLiveIn(Reg);
9198 // Transfer the remainder of BB and its successor edges to sinkMBB.
9199 sinkMBB->splice(sinkMBB->begin(), BB,
9200 llvm::next(MachineBasicBlock::iterator(MI)),
9202 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
9204 // Add the true and fallthrough blocks as its successors.
9205 BB->addSuccessor(copy0MBB);
9206 BB->addSuccessor(sinkMBB);
9208 // Create the conditional branch instruction.
9210 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
9211 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
9214 // %FalseValue = ...
9215 // # fallthrough to sinkMBB
9216 copy0MBB->addSuccessor(sinkMBB);
9219 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
9221 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
9222 TII->get(X86::PHI), MI->getOperand(0).getReg())
9223 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
9224 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
9226 MI->eraseFromParent(); // The pseudo instruction is gone now.
9231 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
9232 MachineBasicBlock *BB) const {
9233 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9234 DebugLoc DL = MI->getDebugLoc();
9236 // The lowering is pretty easy: we're just emitting the call to _alloca. The
9237 // non-trivial part is impdef of ESP.
9238 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
9241 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
9242 .addExternalSymbol("_alloca")
9243 .addReg(X86::EAX, RegState::Implicit)
9244 .addReg(X86::ESP, RegState::Implicit)
9245 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
9246 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
9247 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
9249 MI->eraseFromParent(); // The pseudo instruction is gone now.
9254 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
9255 MachineBasicBlock *BB) const {
9256 // This is pretty easy. We're taking the value that we received from
9257 // our load from the relocation, sticking it in either RDI (x86-64)
9258 // or EAX and doing an indirect call. The return value will then
9259 // be in the normal return register.
9260 const X86InstrInfo *TII
9261 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
9262 DebugLoc DL = MI->getDebugLoc();
9263 MachineFunction *F = BB->getParent();
9264 bool IsWin64 = Subtarget->isTargetWin64();
9266 assert(MI->getOperand(3).isGlobal() && "This should be a global");
9268 if (Subtarget->is64Bit()) {
9269 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9270 TII->get(X86::MOV64rm), X86::RDI)
9272 .addImm(0).addReg(0)
9273 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9274 MI->getOperand(3).getTargetFlags())
9276 MIB = BuildMI(*BB, MI, DL, TII->get(IsWin64 ? X86::WINCALL64m : X86::CALL64m));
9277 addDirectMem(MIB, X86::RDI);
9278 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
9279 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9280 TII->get(X86::MOV32rm), X86::EAX)
9282 .addImm(0).addReg(0)
9283 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9284 MI->getOperand(3).getTargetFlags())
9286 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9287 addDirectMem(MIB, X86::EAX);
9289 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9290 TII->get(X86::MOV32rm), X86::EAX)
9291 .addReg(TII->getGlobalBaseReg(F))
9292 .addImm(0).addReg(0)
9293 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9294 MI->getOperand(3).getTargetFlags())
9296 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9297 addDirectMem(MIB, X86::EAX);
9300 MI->eraseFromParent(); // The pseudo instruction is gone now.
9305 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
9306 MachineBasicBlock *BB) const {
9307 switch (MI->getOpcode()) {
9308 default: assert(false && "Unexpected instr type to insert");
9309 case X86::MINGW_ALLOCA:
9310 return EmitLoweredMingwAlloca(MI, BB);
9311 case X86::TLSCall_32:
9312 case X86::TLSCall_64:
9313 return EmitLoweredTLSCall(MI, BB);
9315 case X86::CMOV_V1I64:
9316 case X86::CMOV_FR32:
9317 case X86::CMOV_FR64:
9318 case X86::CMOV_V4F32:
9319 case X86::CMOV_V2F64:
9320 case X86::CMOV_V2I64:
9321 case X86::CMOV_GR16:
9322 case X86::CMOV_GR32:
9323 case X86::CMOV_RFP32:
9324 case X86::CMOV_RFP64:
9325 case X86::CMOV_RFP80:
9326 return EmitLoweredSelect(MI, BB);
9328 case X86::FP32_TO_INT16_IN_MEM:
9329 case X86::FP32_TO_INT32_IN_MEM:
9330 case X86::FP32_TO_INT64_IN_MEM:
9331 case X86::FP64_TO_INT16_IN_MEM:
9332 case X86::FP64_TO_INT32_IN_MEM:
9333 case X86::FP64_TO_INT64_IN_MEM:
9334 case X86::FP80_TO_INT16_IN_MEM:
9335 case X86::FP80_TO_INT32_IN_MEM:
9336 case X86::FP80_TO_INT64_IN_MEM: {
9337 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9338 DebugLoc DL = MI->getDebugLoc();
9340 // Change the floating point control register to use "round towards zero"
9341 // mode when truncating to an integer value.
9342 MachineFunction *F = BB->getParent();
9343 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
9344 addFrameReference(BuildMI(*BB, MI, DL,
9345 TII->get(X86::FNSTCW16m)), CWFrameIdx);
9347 // Load the old value of the high byte of the control word...
9349 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
9350 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
9353 // Set the high part to be round to zero...
9354 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
9357 // Reload the modified control word now...
9358 addFrameReference(BuildMI(*BB, MI, DL,
9359 TII->get(X86::FLDCW16m)), CWFrameIdx);
9361 // Restore the memory image of control word to original value
9362 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
9365 // Get the X86 opcode to use.
9367 switch (MI->getOpcode()) {
9368 default: llvm_unreachable("illegal opcode!");
9369 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
9370 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
9371 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
9372 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
9373 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
9374 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
9375 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
9376 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
9377 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
9381 MachineOperand &Op = MI->getOperand(0);
9383 AM.BaseType = X86AddressMode::RegBase;
9384 AM.Base.Reg = Op.getReg();
9386 AM.BaseType = X86AddressMode::FrameIndexBase;
9387 AM.Base.FrameIndex = Op.getIndex();
9389 Op = MI->getOperand(1);
9391 AM.Scale = Op.getImm();
9392 Op = MI->getOperand(2);
9394 AM.IndexReg = Op.getImm();
9395 Op = MI->getOperand(3);
9396 if (Op.isGlobal()) {
9397 AM.GV = Op.getGlobal();
9399 AM.Disp = Op.getImm();
9401 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
9402 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
9404 // Reload the original control word now.
9405 addFrameReference(BuildMI(*BB, MI, DL,
9406 TII->get(X86::FLDCW16m)), CWFrameIdx);
9408 MI->eraseFromParent(); // The pseudo instruction is gone now.
9411 // String/text processing lowering.
9412 case X86::PCMPISTRM128REG:
9413 case X86::VPCMPISTRM128REG:
9414 return EmitPCMP(MI, BB, 3, false /* in-mem */);
9415 case X86::PCMPISTRM128MEM:
9416 case X86::VPCMPISTRM128MEM:
9417 return EmitPCMP(MI, BB, 3, true /* in-mem */);
9418 case X86::PCMPESTRM128REG:
9419 case X86::VPCMPESTRM128REG:
9420 return EmitPCMP(MI, BB, 5, false /* in mem */);
9421 case X86::PCMPESTRM128MEM:
9422 case X86::VPCMPESTRM128MEM:
9423 return EmitPCMP(MI, BB, 5, true /* in mem */);
9426 case X86::ATOMAND32:
9427 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9428 X86::AND32ri, X86::MOV32rm,
9430 X86::NOT32r, X86::EAX,
9431 X86::GR32RegisterClass);
9433 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
9434 X86::OR32ri, X86::MOV32rm,
9436 X86::NOT32r, X86::EAX,
9437 X86::GR32RegisterClass);
9438 case X86::ATOMXOR32:
9439 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
9440 X86::XOR32ri, X86::MOV32rm,
9442 X86::NOT32r, X86::EAX,
9443 X86::GR32RegisterClass);
9444 case X86::ATOMNAND32:
9445 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9446 X86::AND32ri, X86::MOV32rm,
9448 X86::NOT32r, X86::EAX,
9449 X86::GR32RegisterClass, true);
9450 case X86::ATOMMIN32:
9451 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
9452 case X86::ATOMMAX32:
9453 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
9454 case X86::ATOMUMIN32:
9455 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
9456 case X86::ATOMUMAX32:
9457 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
9459 case X86::ATOMAND16:
9460 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9461 X86::AND16ri, X86::MOV16rm,
9463 X86::NOT16r, X86::AX,
9464 X86::GR16RegisterClass);
9466 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
9467 X86::OR16ri, X86::MOV16rm,
9469 X86::NOT16r, X86::AX,
9470 X86::GR16RegisterClass);
9471 case X86::ATOMXOR16:
9472 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
9473 X86::XOR16ri, X86::MOV16rm,
9475 X86::NOT16r, X86::AX,
9476 X86::GR16RegisterClass);
9477 case X86::ATOMNAND16:
9478 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9479 X86::AND16ri, X86::MOV16rm,
9481 X86::NOT16r, X86::AX,
9482 X86::GR16RegisterClass, true);
9483 case X86::ATOMMIN16:
9484 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
9485 case X86::ATOMMAX16:
9486 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
9487 case X86::ATOMUMIN16:
9488 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
9489 case X86::ATOMUMAX16:
9490 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
9493 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9494 X86::AND8ri, X86::MOV8rm,
9496 X86::NOT8r, X86::AL,
9497 X86::GR8RegisterClass);
9499 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
9500 X86::OR8ri, X86::MOV8rm,
9502 X86::NOT8r, X86::AL,
9503 X86::GR8RegisterClass);
9505 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
9506 X86::XOR8ri, X86::MOV8rm,
9508 X86::NOT8r, X86::AL,
9509 X86::GR8RegisterClass);
9510 case X86::ATOMNAND8:
9511 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9512 X86::AND8ri, X86::MOV8rm,
9514 X86::NOT8r, X86::AL,
9515 X86::GR8RegisterClass, true);
9516 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
9517 // This group is for 64-bit host.
9518 case X86::ATOMAND64:
9519 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9520 X86::AND64ri32, X86::MOV64rm,
9522 X86::NOT64r, X86::RAX,
9523 X86::GR64RegisterClass);
9525 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
9526 X86::OR64ri32, X86::MOV64rm,
9528 X86::NOT64r, X86::RAX,
9529 X86::GR64RegisterClass);
9530 case X86::ATOMXOR64:
9531 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
9532 X86::XOR64ri32, X86::MOV64rm,
9534 X86::NOT64r, X86::RAX,
9535 X86::GR64RegisterClass);
9536 case X86::ATOMNAND64:
9537 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9538 X86::AND64ri32, X86::MOV64rm,
9540 X86::NOT64r, X86::RAX,
9541 X86::GR64RegisterClass, true);
9542 case X86::ATOMMIN64:
9543 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
9544 case X86::ATOMMAX64:
9545 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
9546 case X86::ATOMUMIN64:
9547 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
9548 case X86::ATOMUMAX64:
9549 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
9551 // This group does 64-bit operations on a 32-bit host.
9552 case X86::ATOMAND6432:
9553 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9554 X86::AND32rr, X86::AND32rr,
9555 X86::AND32ri, X86::AND32ri,
9557 case X86::ATOMOR6432:
9558 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9559 X86::OR32rr, X86::OR32rr,
9560 X86::OR32ri, X86::OR32ri,
9562 case X86::ATOMXOR6432:
9563 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9564 X86::XOR32rr, X86::XOR32rr,
9565 X86::XOR32ri, X86::XOR32ri,
9567 case X86::ATOMNAND6432:
9568 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9569 X86::AND32rr, X86::AND32rr,
9570 X86::AND32ri, X86::AND32ri,
9572 case X86::ATOMADD6432:
9573 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9574 X86::ADD32rr, X86::ADC32rr,
9575 X86::ADD32ri, X86::ADC32ri,
9577 case X86::ATOMSUB6432:
9578 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9579 X86::SUB32rr, X86::SBB32rr,
9580 X86::SUB32ri, X86::SBB32ri,
9582 case X86::ATOMSWAP6432:
9583 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9584 X86::MOV32rr, X86::MOV32rr,
9585 X86::MOV32ri, X86::MOV32ri,
9587 case X86::VASTART_SAVE_XMM_REGS:
9588 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
9592 //===----------------------------------------------------------------------===//
9593 // X86 Optimization Hooks
9594 //===----------------------------------------------------------------------===//
9596 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9600 const SelectionDAG &DAG,
9601 unsigned Depth) const {
9602 unsigned Opc = Op.getOpcode();
9603 assert((Opc >= ISD::BUILTIN_OP_END ||
9604 Opc == ISD::INTRINSIC_WO_CHAIN ||
9605 Opc == ISD::INTRINSIC_W_CHAIN ||
9606 Opc == ISD::INTRINSIC_VOID) &&
9607 "Should use MaskedValueIsZero if you don't know whether Op"
9608 " is a target node!");
9610 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
9622 // These nodes' second result is a boolean.
9623 if (Op.getResNo() == 0)
9627 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
9628 Mask.getBitWidth() - 1);
9633 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
9634 /// node is a GlobalAddress + offset.
9635 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
9636 const GlobalValue* &GA,
9637 int64_t &Offset) const {
9638 if (N->getOpcode() == X86ISD::Wrapper) {
9639 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
9640 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
9641 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
9645 return TargetLowering::isGAPlusOffset(N, GA, Offset);
9648 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
9649 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
9650 /// if the load addresses are consecutive, non-overlapping, and in the right
9652 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
9653 const TargetLowering &TLI) {
9654 DebugLoc dl = N->getDebugLoc();
9655 EVT VT = N->getValueType(0);
9657 if (VT.getSizeInBits() != 128)
9660 SmallVector<SDValue, 16> Elts;
9661 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
9662 Elts.push_back(getShuffleScalarElt(N, i, DAG));
9664 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
9667 /// PerformShuffleCombine - Detect vector gather/scatter index generation
9668 /// and convert it from being a bunch of shuffles and extracts to a simple
9669 /// store and scalar loads to extract the elements.
9670 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
9671 const TargetLowering &TLI) {
9672 SDValue InputVector = N->getOperand(0);
9674 // Only operate on vectors of 4 elements, where the alternative shuffling
9675 // gets to be more expensive.
9676 if (InputVector.getValueType() != MVT::v4i32)
9679 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
9680 // single use which is a sign-extend or zero-extend, and all elements are
9682 SmallVector<SDNode *, 4> Uses;
9683 unsigned ExtractedElements = 0;
9684 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
9685 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
9686 if (UI.getUse().getResNo() != InputVector.getResNo())
9689 SDNode *Extract = *UI;
9690 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9693 if (Extract->getValueType(0) != MVT::i32)
9695 if (!Extract->hasOneUse())
9697 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9698 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9700 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9703 // Record which element was extracted.
9704 ExtractedElements |=
9705 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9707 Uses.push_back(Extract);
9710 // If not all the elements were used, this may not be worthwhile.
9711 if (ExtractedElements != 15)
9714 // Ok, we've now decided to do the transformation.
9715 DebugLoc dl = InputVector.getDebugLoc();
9717 // Store the value to a temporary stack slot.
9718 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9719 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL,
9720 0, false, false, 0);
9722 // Replace each use (extract) with a load of the appropriate element.
9723 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9724 UE = Uses.end(); UI != UE; ++UI) {
9725 SDNode *Extract = *UI;
9727 // Compute the element's address.
9728 SDValue Idx = Extract->getOperand(1);
9730 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9731 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9732 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9734 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
9735 OffsetVal, StackPtr);
9738 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
9739 ScalarAddr, NULL, 0, false, false, 0);
9741 // Replace the exact with the load.
9742 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9745 // The replacement was made in place; don't return anything.
9749 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9750 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9751 const X86Subtarget *Subtarget) {
9752 DebugLoc DL = N->getDebugLoc();
9753 SDValue Cond = N->getOperand(0);
9754 // Get the LHS/RHS of the select.
9755 SDValue LHS = N->getOperand(1);
9756 SDValue RHS = N->getOperand(2);
9758 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9759 // instructions match the semantics of the common C idiom x<y?x:y but not
9760 // x<=y?x:y, because of how they handle negative zero (which can be
9761 // ignored in unsafe-math mode).
9762 if (Subtarget->hasSSE2() &&
9763 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9764 Cond.getOpcode() == ISD::SETCC) {
9765 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9767 unsigned Opcode = 0;
9768 // Check for x CC y ? x : y.
9769 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9770 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9774 // Converting this to a min would handle NaNs incorrectly, and swapping
9775 // the operands would cause it to handle comparisons between positive
9776 // and negative zero incorrectly.
9777 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9778 if (!UnsafeFPMath &&
9779 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9781 std::swap(LHS, RHS);
9783 Opcode = X86ISD::FMIN;
9786 // Converting this to a min would handle comparisons between positive
9787 // and negative zero incorrectly.
9788 if (!UnsafeFPMath &&
9789 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9791 Opcode = X86ISD::FMIN;
9794 // Converting this to a min would handle both negative zeros and NaNs
9795 // incorrectly, but we can swap the operands to fix both.
9796 std::swap(LHS, RHS);
9800 Opcode = X86ISD::FMIN;
9804 // Converting this to a max would handle comparisons between positive
9805 // and negative zero incorrectly.
9806 if (!UnsafeFPMath &&
9807 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9809 Opcode = X86ISD::FMAX;
9812 // Converting this to a max would handle NaNs incorrectly, and swapping
9813 // the operands would cause it to handle comparisons between positive
9814 // and negative zero incorrectly.
9815 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9816 if (!UnsafeFPMath &&
9817 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9819 std::swap(LHS, RHS);
9821 Opcode = X86ISD::FMAX;
9824 // Converting this to a max would handle both negative zeros and NaNs
9825 // incorrectly, but we can swap the operands to fix both.
9826 std::swap(LHS, RHS);
9830 Opcode = X86ISD::FMAX;
9833 // Check for x CC y ? y : x -- a min/max with reversed arms.
9834 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9835 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9839 // Converting this to a min would handle comparisons between positive
9840 // and negative zero incorrectly, and swapping the operands would
9841 // cause it to handle NaNs incorrectly.
9842 if (!UnsafeFPMath &&
9843 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9844 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9846 std::swap(LHS, RHS);
9848 Opcode = X86ISD::FMIN;
9851 // Converting this to a min would handle NaNs incorrectly.
9852 if (!UnsafeFPMath &&
9853 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9855 Opcode = X86ISD::FMIN;
9858 // Converting this to a min would handle both negative zeros and NaNs
9859 // incorrectly, but we can swap the operands to fix both.
9860 std::swap(LHS, RHS);
9864 Opcode = X86ISD::FMIN;
9868 // Converting this to a max would handle NaNs incorrectly.
9869 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9871 Opcode = X86ISD::FMAX;
9874 // Converting this to a max would handle comparisons between positive
9875 // and negative zero incorrectly, and swapping the operands would
9876 // cause it to handle NaNs incorrectly.
9877 if (!UnsafeFPMath &&
9878 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9879 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9881 std::swap(LHS, RHS);
9883 Opcode = X86ISD::FMAX;
9886 // Converting this to a max would handle both negative zeros and NaNs
9887 // incorrectly, but we can swap the operands to fix both.
9888 std::swap(LHS, RHS);
9892 Opcode = X86ISD::FMAX;
9898 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9901 // If this is a select between two integer constants, try to do some
9903 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9904 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9905 // Don't do this for crazy integer types.
9906 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9907 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9908 // so that TrueC (the true value) is larger than FalseC.
9909 bool NeedsCondInvert = false;
9911 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9912 // Efficiently invertible.
9913 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9914 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9915 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9916 NeedsCondInvert = true;
9917 std::swap(TrueC, FalseC);
9920 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9921 if (FalseC->getAPIntValue() == 0 &&
9922 TrueC->getAPIntValue().isPowerOf2()) {
9923 if (NeedsCondInvert) // Invert the condition if needed.
9924 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9925 DAG.getConstant(1, Cond.getValueType()));
9927 // Zero extend the condition if needed.
9928 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9930 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9931 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9932 DAG.getConstant(ShAmt, MVT::i8));
9935 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9936 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9937 if (NeedsCondInvert) // Invert the condition if needed.
9938 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9939 DAG.getConstant(1, Cond.getValueType()));
9941 // Zero extend the condition if needed.
9942 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9943 FalseC->getValueType(0), Cond);
9944 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9945 SDValue(FalseC, 0));
9948 // Optimize cases that will turn into an LEA instruction. This requires
9949 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9950 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9951 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9952 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9954 bool isFastMultiplier = false;
9956 switch ((unsigned char)Diff) {
9958 case 1: // result = add base, cond
9959 case 2: // result = lea base( , cond*2)
9960 case 3: // result = lea base(cond, cond*2)
9961 case 4: // result = lea base( , cond*4)
9962 case 5: // result = lea base(cond, cond*4)
9963 case 8: // result = lea base( , cond*8)
9964 case 9: // result = lea base(cond, cond*8)
9965 isFastMultiplier = true;
9970 if (isFastMultiplier) {
9971 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9972 if (NeedsCondInvert) // Invert the condition if needed.
9973 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9974 DAG.getConstant(1, Cond.getValueType()));
9976 // Zero extend the condition if needed.
9977 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9979 // Scale the condition by the difference.
9981 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9982 DAG.getConstant(Diff, Cond.getValueType()));
9984 // Add the base if non-zero.
9985 if (FalseC->getAPIntValue() != 0)
9986 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9987 SDValue(FalseC, 0));
9997 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9998 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9999 TargetLowering::DAGCombinerInfo &DCI) {
10000 DebugLoc DL = N->getDebugLoc();
10002 // If the flag operand isn't dead, don't touch this CMOV.
10003 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
10006 // If this is a select between two integer constants, try to do some
10007 // optimizations. Note that the operands are ordered the opposite of SELECT
10009 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
10010 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
10011 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
10012 // larger than FalseC (the false value).
10013 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
10015 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
10016 CC = X86::GetOppositeBranchCondition(CC);
10017 std::swap(TrueC, FalseC);
10020 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
10021 // This is efficient for any integer data type (including i8/i16) and
10023 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
10024 SDValue Cond = N->getOperand(3);
10025 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10026 DAG.getConstant(CC, MVT::i8), Cond);
10028 // Zero extend the condition if needed.
10029 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
10031 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10032 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
10033 DAG.getConstant(ShAmt, MVT::i8));
10034 if (N->getNumValues() == 2) // Dead flag value?
10035 return DCI.CombineTo(N, Cond, SDValue());
10039 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
10040 // for any integer data type, including i8/i16.
10041 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10042 SDValue Cond = N->getOperand(3);
10043 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10044 DAG.getConstant(CC, MVT::i8), Cond);
10046 // Zero extend the condition if needed.
10047 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10048 FalseC->getValueType(0), Cond);
10049 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10050 SDValue(FalseC, 0));
10052 if (N->getNumValues() == 2) // Dead flag value?
10053 return DCI.CombineTo(N, Cond, SDValue());
10057 // Optimize cases that will turn into an LEA instruction. This requires
10058 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10059 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10060 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10061 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10063 bool isFastMultiplier = false;
10065 switch ((unsigned char)Diff) {
10067 case 1: // result = add base, cond
10068 case 2: // result = lea base( , cond*2)
10069 case 3: // result = lea base(cond, cond*2)
10070 case 4: // result = lea base( , cond*4)
10071 case 5: // result = lea base(cond, cond*4)
10072 case 8: // result = lea base( , cond*8)
10073 case 9: // result = lea base(cond, cond*8)
10074 isFastMultiplier = true;
10079 if (isFastMultiplier) {
10080 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10081 SDValue Cond = N->getOperand(3);
10082 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10083 DAG.getConstant(CC, MVT::i8), Cond);
10084 // Zero extend the condition if needed.
10085 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10087 // Scale the condition by the difference.
10089 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10090 DAG.getConstant(Diff, Cond.getValueType()));
10092 // Add the base if non-zero.
10093 if (FalseC->getAPIntValue() != 0)
10094 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10095 SDValue(FalseC, 0));
10096 if (N->getNumValues() == 2) // Dead flag value?
10097 return DCI.CombineTo(N, Cond, SDValue());
10107 /// PerformMulCombine - Optimize a single multiply with constant into two
10108 /// in order to implement it with two cheaper instructions, e.g.
10109 /// LEA + SHL, LEA + LEA.
10110 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
10111 TargetLowering::DAGCombinerInfo &DCI) {
10112 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
10115 EVT VT = N->getValueType(0);
10116 if (VT != MVT::i64)
10119 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
10122 uint64_t MulAmt = C->getZExtValue();
10123 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
10126 uint64_t MulAmt1 = 0;
10127 uint64_t MulAmt2 = 0;
10128 if ((MulAmt % 9) == 0) {
10130 MulAmt2 = MulAmt / 9;
10131 } else if ((MulAmt % 5) == 0) {
10133 MulAmt2 = MulAmt / 5;
10134 } else if ((MulAmt % 3) == 0) {
10136 MulAmt2 = MulAmt / 3;
10139 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
10140 DebugLoc DL = N->getDebugLoc();
10142 if (isPowerOf2_64(MulAmt2) &&
10143 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
10144 // If second multiplifer is pow2, issue it first. We want the multiply by
10145 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
10147 std::swap(MulAmt1, MulAmt2);
10150 if (isPowerOf2_64(MulAmt1))
10151 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
10152 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
10154 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
10155 DAG.getConstant(MulAmt1, VT));
10157 if (isPowerOf2_64(MulAmt2))
10158 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
10159 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
10161 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
10162 DAG.getConstant(MulAmt2, VT));
10164 // Do not add new nodes to DAG combiner worklist.
10165 DCI.CombineTo(N, NewMul, false);
10170 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
10171 SDValue N0 = N->getOperand(0);
10172 SDValue N1 = N->getOperand(1);
10173 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
10174 EVT VT = N0.getValueType();
10176 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
10177 // since the result of setcc_c is all zero's or all ones.
10178 if (N1C && N0.getOpcode() == ISD::AND &&
10179 N0.getOperand(1).getOpcode() == ISD::Constant) {
10180 SDValue N00 = N0.getOperand(0);
10181 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
10182 ((N00.getOpcode() == ISD::ANY_EXTEND ||
10183 N00.getOpcode() == ISD::ZERO_EXTEND) &&
10184 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
10185 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
10186 APInt ShAmt = N1C->getAPIntValue();
10187 Mask = Mask.shl(ShAmt);
10189 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
10190 N00, DAG.getConstant(Mask, VT));
10197 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
10199 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
10200 const X86Subtarget *Subtarget) {
10201 EVT VT = N->getValueType(0);
10202 if (!VT.isVector() && VT.isInteger() &&
10203 N->getOpcode() == ISD::SHL)
10204 return PerformSHLCombine(N, DAG);
10206 // On X86 with SSE2 support, we can transform this to a vector shift if
10207 // all elements are shifted by the same amount. We can't do this in legalize
10208 // because the a constant vector is typically transformed to a constant pool
10209 // so we have no knowledge of the shift amount.
10210 if (!Subtarget->hasSSE2())
10213 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
10216 SDValue ShAmtOp = N->getOperand(1);
10217 EVT EltVT = VT.getVectorElementType();
10218 DebugLoc DL = N->getDebugLoc();
10219 SDValue BaseShAmt = SDValue();
10220 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
10221 unsigned NumElts = VT.getVectorNumElements();
10223 for (; i != NumElts; ++i) {
10224 SDValue Arg = ShAmtOp.getOperand(i);
10225 if (Arg.getOpcode() == ISD::UNDEF) continue;
10229 for (; i != NumElts; ++i) {
10230 SDValue Arg = ShAmtOp.getOperand(i);
10231 if (Arg.getOpcode() == ISD::UNDEF) continue;
10232 if (Arg != BaseShAmt) {
10236 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
10237 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
10238 SDValue InVec = ShAmtOp.getOperand(0);
10239 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
10240 unsigned NumElts = InVec.getValueType().getVectorNumElements();
10242 for (; i != NumElts; ++i) {
10243 SDValue Arg = InVec.getOperand(i);
10244 if (Arg.getOpcode() == ISD::UNDEF) continue;
10248 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
10249 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
10250 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
10251 if (C->getZExtValue() == SplatIdx)
10252 BaseShAmt = InVec.getOperand(1);
10255 if (BaseShAmt.getNode() == 0)
10256 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
10257 DAG.getIntPtrConstant(0));
10261 // The shift amount is an i32.
10262 if (EltVT.bitsGT(MVT::i32))
10263 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
10264 else if (EltVT.bitsLT(MVT::i32))
10265 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
10267 // The shift amount is identical so we can do a vector shift.
10268 SDValue ValOp = N->getOperand(0);
10269 switch (N->getOpcode()) {
10271 llvm_unreachable("Unknown shift opcode!");
10274 if (VT == MVT::v2i64)
10275 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10276 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
10278 if (VT == MVT::v4i32)
10279 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10280 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
10282 if (VT == MVT::v8i16)
10283 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10284 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
10288 if (VT == MVT::v4i32)
10289 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10290 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
10292 if (VT == MVT::v8i16)
10293 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10294 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
10298 if (VT == MVT::v2i64)
10299 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10300 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
10302 if (VT == MVT::v4i32)
10303 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10304 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
10306 if (VT == MVT::v8i16)
10307 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10308 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
10315 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
10316 TargetLowering::DAGCombinerInfo &DCI,
10317 const X86Subtarget *Subtarget) {
10318 if (DCI.isBeforeLegalizeOps())
10321 EVT VT = N->getValueType(0);
10322 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
10325 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
10326 SDValue N0 = N->getOperand(0);
10327 SDValue N1 = N->getOperand(1);
10328 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
10330 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
10332 if (!N0.hasOneUse() || !N1.hasOneUse())
10335 SDValue ShAmt0 = N0.getOperand(1);
10336 if (ShAmt0.getValueType() != MVT::i8)
10338 SDValue ShAmt1 = N1.getOperand(1);
10339 if (ShAmt1.getValueType() != MVT::i8)
10341 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
10342 ShAmt0 = ShAmt0.getOperand(0);
10343 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
10344 ShAmt1 = ShAmt1.getOperand(0);
10346 DebugLoc DL = N->getDebugLoc();
10347 unsigned Opc = X86ISD::SHLD;
10348 SDValue Op0 = N0.getOperand(0);
10349 SDValue Op1 = N1.getOperand(0);
10350 if (ShAmt0.getOpcode() == ISD::SUB) {
10351 Opc = X86ISD::SHRD;
10352 std::swap(Op0, Op1);
10353 std::swap(ShAmt0, ShAmt1);
10356 unsigned Bits = VT.getSizeInBits();
10357 if (ShAmt1.getOpcode() == ISD::SUB) {
10358 SDValue Sum = ShAmt1.getOperand(0);
10359 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
10360 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
10361 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
10362 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
10363 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
10364 return DAG.getNode(Opc, DL, VT,
10366 DAG.getNode(ISD::TRUNCATE, DL,
10369 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
10370 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
10372 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
10373 return DAG.getNode(Opc, DL, VT,
10374 N0.getOperand(0), N1.getOperand(0),
10375 DAG.getNode(ISD::TRUNCATE, DL,
10382 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
10383 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
10384 const X86Subtarget *Subtarget) {
10385 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
10386 // the FP state in cases where an emms may be missing.
10387 // A preferable solution to the general problem is to figure out the right
10388 // places to insert EMMS. This qualifies as a quick hack.
10390 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
10391 StoreSDNode *St = cast<StoreSDNode>(N);
10392 EVT VT = St->getValue().getValueType();
10393 if (VT.getSizeInBits() != 64)
10396 const Function *F = DAG.getMachineFunction().getFunction();
10397 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
10398 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
10399 && Subtarget->hasSSE2();
10400 if ((VT.isVector() ||
10401 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
10402 isa<LoadSDNode>(St->getValue()) &&
10403 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
10404 St->getChain().hasOneUse() && !St->isVolatile()) {
10405 SDNode* LdVal = St->getValue().getNode();
10406 LoadSDNode *Ld = 0;
10407 int TokenFactorIndex = -1;
10408 SmallVector<SDValue, 8> Ops;
10409 SDNode* ChainVal = St->getChain().getNode();
10410 // Must be a store of a load. We currently handle two cases: the load
10411 // is a direct child, and it's under an intervening TokenFactor. It is
10412 // possible to dig deeper under nested TokenFactors.
10413 if (ChainVal == LdVal)
10414 Ld = cast<LoadSDNode>(St->getChain());
10415 else if (St->getValue().hasOneUse() &&
10416 ChainVal->getOpcode() == ISD::TokenFactor) {
10417 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
10418 if (ChainVal->getOperand(i).getNode() == LdVal) {
10419 TokenFactorIndex = i;
10420 Ld = cast<LoadSDNode>(St->getValue());
10422 Ops.push_back(ChainVal->getOperand(i));
10426 if (!Ld || !ISD::isNormalLoad(Ld))
10429 // If this is not the MMX case, i.e. we are just turning i64 load/store
10430 // into f64 load/store, avoid the transformation if there are multiple
10431 // uses of the loaded value.
10432 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
10435 DebugLoc LdDL = Ld->getDebugLoc();
10436 DebugLoc StDL = N->getDebugLoc();
10437 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
10438 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
10440 if (Subtarget->is64Bit() || F64IsLegal) {
10441 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
10442 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
10443 Ld->getBasePtr(), Ld->getSrcValue(),
10444 Ld->getSrcValueOffset(), Ld->isVolatile(),
10445 Ld->isNonTemporal(), Ld->getAlignment());
10446 SDValue NewChain = NewLd.getValue(1);
10447 if (TokenFactorIndex != -1) {
10448 Ops.push_back(NewChain);
10449 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10452 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
10453 St->getSrcValue(), St->getSrcValueOffset(),
10454 St->isVolatile(), St->isNonTemporal(),
10455 St->getAlignment());
10458 // Otherwise, lower to two pairs of 32-bit loads / stores.
10459 SDValue LoAddr = Ld->getBasePtr();
10460 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
10461 DAG.getConstant(4, MVT::i32));
10463 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
10464 Ld->getSrcValue(), Ld->getSrcValueOffset(),
10465 Ld->isVolatile(), Ld->isNonTemporal(),
10466 Ld->getAlignment());
10467 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
10468 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
10469 Ld->isVolatile(), Ld->isNonTemporal(),
10470 MinAlign(Ld->getAlignment(), 4));
10472 SDValue NewChain = LoLd.getValue(1);
10473 if (TokenFactorIndex != -1) {
10474 Ops.push_back(LoLd);
10475 Ops.push_back(HiLd);
10476 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10480 LoAddr = St->getBasePtr();
10481 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
10482 DAG.getConstant(4, MVT::i32));
10484 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
10485 St->getSrcValue(), St->getSrcValueOffset(),
10486 St->isVolatile(), St->isNonTemporal(),
10487 St->getAlignment());
10488 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
10490 St->getSrcValueOffset() + 4,
10492 St->isNonTemporal(),
10493 MinAlign(St->getAlignment(), 4));
10494 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
10499 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
10500 /// X86ISD::FXOR nodes.
10501 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
10502 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
10503 // F[X]OR(0.0, x) -> x
10504 // F[X]OR(x, 0.0) -> x
10505 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10506 if (C->getValueAPF().isPosZero())
10507 return N->getOperand(1);
10508 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10509 if (C->getValueAPF().isPosZero())
10510 return N->getOperand(0);
10514 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
10515 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
10516 // FAND(0.0, x) -> 0.0
10517 // FAND(x, 0.0) -> 0.0
10518 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10519 if (C->getValueAPF().isPosZero())
10520 return N->getOperand(0);
10521 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10522 if (C->getValueAPF().isPosZero())
10523 return N->getOperand(1);
10527 static SDValue PerformBTCombine(SDNode *N,
10529 TargetLowering::DAGCombinerInfo &DCI) {
10530 // BT ignores high bits in the bit index operand.
10531 SDValue Op1 = N->getOperand(1);
10532 if (Op1.hasOneUse()) {
10533 unsigned BitWidth = Op1.getValueSizeInBits();
10534 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
10535 APInt KnownZero, KnownOne;
10536 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
10537 !DCI.isBeforeLegalizeOps());
10538 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10539 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
10540 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
10541 DCI.CommitTargetLoweringOpt(TLO);
10546 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
10547 SDValue Op = N->getOperand(0);
10548 if (Op.getOpcode() == ISD::BIT_CONVERT)
10549 Op = Op.getOperand(0);
10550 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
10551 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
10552 VT.getVectorElementType().getSizeInBits() ==
10553 OpVT.getVectorElementType().getSizeInBits()) {
10554 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
10559 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
10560 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
10561 // (and (i32 x86isd::setcc_carry), 1)
10562 // This eliminates the zext. This transformation is necessary because
10563 // ISD::SETCC is always legalized to i8.
10564 DebugLoc dl = N->getDebugLoc();
10565 SDValue N0 = N->getOperand(0);
10566 EVT VT = N->getValueType(0);
10567 if (N0.getOpcode() == ISD::AND &&
10569 N0.getOperand(0).hasOneUse()) {
10570 SDValue N00 = N0.getOperand(0);
10571 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
10573 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
10574 if (!C || C->getZExtValue() != 1)
10576 return DAG.getNode(ISD::AND, dl, VT,
10577 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
10578 N00.getOperand(0), N00.getOperand(1)),
10579 DAG.getConstant(1, VT));
10585 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
10586 DAGCombinerInfo &DCI) const {
10587 SelectionDAG &DAG = DCI.DAG;
10588 switch (N->getOpcode()) {
10590 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
10591 case ISD::EXTRACT_VECTOR_ELT:
10592 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
10593 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
10594 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
10595 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
10598 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
10599 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
10600 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
10602 case X86ISD::FOR: return PerformFORCombine(N, DAG);
10603 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
10604 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
10605 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
10606 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
10612 /// isTypeDesirableForOp - Return true if the target has native support for
10613 /// the specified value type and it is 'desirable' to use the type for the
10614 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
10615 /// instruction encodings are longer and some i16 instructions are slow.
10616 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
10617 if (!isTypeLegal(VT))
10619 if (VT != MVT::i16)
10626 case ISD::SIGN_EXTEND:
10627 case ISD::ZERO_EXTEND:
10628 case ISD::ANY_EXTEND:
10641 /// IsDesirableToPromoteOp - This method query the target whether it is
10642 /// beneficial for dag combiner to promote the specified node. If true, it
10643 /// should return the desired promotion type by reference.
10644 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
10645 EVT VT = Op.getValueType();
10646 if (VT != MVT::i16)
10649 bool Promote = false;
10650 bool Commute = false;
10651 switch (Op.getOpcode()) {
10654 LoadSDNode *LD = cast<LoadSDNode>(Op);
10655 // If the non-extending load has a single use and it's not live out, then it
10656 // might be folded.
10657 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
10658 Op.hasOneUse()*/) {
10659 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
10660 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
10661 // The only case where we'd want to promote LOAD (rather then it being
10662 // promoted as an operand is when it's only use is liveout.
10663 if (UI->getOpcode() != ISD::CopyToReg)
10670 case ISD::SIGN_EXTEND:
10671 case ISD::ZERO_EXTEND:
10672 case ISD::ANY_EXTEND:
10677 SDValue N0 = Op.getOperand(0);
10678 // Look out for (store (shl (load), x)).
10679 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10692 SDValue N0 = Op.getOperand(0);
10693 SDValue N1 = Op.getOperand(1);
10694 if (!Commute && MayFoldLoad(N1))
10696 // Avoid disabling potential load folding opportunities.
10697 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10699 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10709 //===----------------------------------------------------------------------===//
10710 // X86 Inline Assembly Support
10711 //===----------------------------------------------------------------------===//
10713 static bool LowerToBSwap(CallInst *CI) {
10714 // FIXME: this should verify that we are targetting a 486 or better. If not,
10715 // we will turn this bswap into something that will be lowered to logical ops
10716 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10717 // so don't worry about this.
10719 // Verify this is a simple bswap.
10720 if (CI->getNumArgOperands() != 1 ||
10721 CI->getType() != CI->getArgOperand(0)->getType() ||
10722 !CI->getType()->isIntegerTy())
10725 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10726 if (!Ty || Ty->getBitWidth() % 16 != 0)
10729 // Okay, we can do this xform, do so now.
10730 const Type *Tys[] = { Ty };
10731 Module *M = CI->getParent()->getParent()->getParent();
10732 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10734 Value *Op = CI->getArgOperand(0);
10735 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10737 CI->replaceAllUsesWith(Op);
10738 CI->eraseFromParent();
10742 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10743 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10744 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10746 std::string AsmStr = IA->getAsmString();
10748 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10749 SmallVector<StringRef, 4> AsmPieces;
10750 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10752 switch (AsmPieces.size()) {
10753 default: return false;
10755 AsmStr = AsmPieces[0];
10757 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10760 if (AsmPieces.size() == 2 &&
10761 (AsmPieces[0] == "bswap" ||
10762 AsmPieces[0] == "bswapq" ||
10763 AsmPieces[0] == "bswapl") &&
10764 (AsmPieces[1] == "$0" ||
10765 AsmPieces[1] == "${0:q}")) {
10766 // No need to check constraints, nothing other than the equivalent of
10767 // "=r,0" would be valid here.
10768 return LowerToBSwap(CI);
10770 // rorw $$8, ${0:w} --> llvm.bswap.i16
10771 if (CI->getType()->isIntegerTy(16) &&
10772 AsmPieces.size() == 3 &&
10773 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10774 AsmPieces[1] == "$$8," &&
10775 AsmPieces[2] == "${0:w}" &&
10776 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10778 const std::string &Constraints = IA->getConstraintString();
10779 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10780 std::sort(AsmPieces.begin(), AsmPieces.end());
10781 if (AsmPieces.size() == 4 &&
10782 AsmPieces[0] == "~{cc}" &&
10783 AsmPieces[1] == "~{dirflag}" &&
10784 AsmPieces[2] == "~{flags}" &&
10785 AsmPieces[3] == "~{fpsr}") {
10786 return LowerToBSwap(CI);
10791 if (CI->getType()->isIntegerTy(64) &&
10792 Constraints.size() >= 2 &&
10793 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10794 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10795 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10796 SmallVector<StringRef, 4> Words;
10797 SplitString(AsmPieces[0], Words, " \t");
10798 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10800 SplitString(AsmPieces[1], Words, " \t");
10801 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10803 SplitString(AsmPieces[2], Words, " \t,");
10804 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10805 Words[2] == "%edx") {
10806 return LowerToBSwap(CI);
10818 /// getConstraintType - Given a constraint letter, return the type of
10819 /// constraint it is for this target.
10820 X86TargetLowering::ConstraintType
10821 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10822 if (Constraint.size() == 1) {
10823 switch (Constraint[0]) {
10835 return C_RegisterClass;
10843 return TargetLowering::getConstraintType(Constraint);
10846 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10847 /// with another that has more specific requirements based on the type of the
10848 /// corresponding operand.
10849 const char *X86TargetLowering::
10850 LowerXConstraint(EVT ConstraintVT) const {
10851 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10852 // 'f' like normal targets.
10853 if (ConstraintVT.isFloatingPoint()) {
10854 if (Subtarget->hasSSE2())
10856 if (Subtarget->hasSSE1())
10860 return TargetLowering::LowerXConstraint(ConstraintVT);
10863 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10864 /// vector. If it is invalid, don't add anything to Ops.
10865 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10867 std::vector<SDValue>&Ops,
10868 SelectionDAG &DAG) const {
10869 SDValue Result(0, 0);
10871 switch (Constraint) {
10874 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10875 if (C->getZExtValue() <= 31) {
10876 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10882 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10883 if (C->getZExtValue() <= 63) {
10884 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10890 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10891 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10892 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10898 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10899 if (C->getZExtValue() <= 255) {
10900 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10906 // 32-bit signed value
10907 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10908 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10909 C->getSExtValue())) {
10910 // Widen to 64 bits here to get it sign extended.
10911 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10914 // FIXME gcc accepts some relocatable values here too, but only in certain
10915 // memory models; it's complicated.
10920 // 32-bit unsigned value
10921 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10922 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10923 C->getZExtValue())) {
10924 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10928 // FIXME gcc accepts some relocatable values here too, but only in certain
10929 // memory models; it's complicated.
10933 // Literal immediates are always ok.
10934 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10935 // Widen to 64 bits here to get it sign extended.
10936 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10940 // In any sort of PIC mode addresses need to be computed at runtime by
10941 // adding in a register or some sort of table lookup. These can't
10942 // be used as immediates.
10943 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
10946 // If we are in non-pic codegen mode, we allow the address of a global (with
10947 // an optional displacement) to be used with 'i'.
10948 GlobalAddressSDNode *GA = 0;
10949 int64_t Offset = 0;
10951 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10953 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10954 Offset += GA->getOffset();
10956 } else if (Op.getOpcode() == ISD::ADD) {
10957 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10958 Offset += C->getZExtValue();
10959 Op = Op.getOperand(0);
10962 } else if (Op.getOpcode() == ISD::SUB) {
10963 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10964 Offset += -C->getZExtValue();
10965 Op = Op.getOperand(0);
10970 // Otherwise, this isn't something we can handle, reject it.
10974 const GlobalValue *GV = GA->getGlobal();
10975 // If we require an extra load to get this address, as in PIC mode, we
10976 // can't accept it.
10977 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10978 getTargetMachine())))
10981 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
10982 GA->getValueType(0), Offset);
10987 if (Result.getNode()) {
10988 Ops.push_back(Result);
10991 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10994 std::vector<unsigned> X86TargetLowering::
10995 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10997 if (Constraint.size() == 1) {
10998 // FIXME: not handling fp-stack yet!
10999 switch (Constraint[0]) { // GCC X86 Constraint Letters
11000 default: break; // Unknown constraint letter
11001 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
11002 if (Subtarget->is64Bit()) {
11003 if (VT == MVT::i32)
11004 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
11005 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
11006 X86::R10D,X86::R11D,X86::R12D,
11007 X86::R13D,X86::R14D,X86::R15D,
11008 X86::EBP, X86::ESP, 0);
11009 else if (VT == MVT::i16)
11010 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
11011 X86::SI, X86::DI, X86::R8W,X86::R9W,
11012 X86::R10W,X86::R11W,X86::R12W,
11013 X86::R13W,X86::R14W,X86::R15W,
11014 X86::BP, X86::SP, 0);
11015 else if (VT == MVT::i8)
11016 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
11017 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
11018 X86::R10B,X86::R11B,X86::R12B,
11019 X86::R13B,X86::R14B,X86::R15B,
11020 X86::BPL, X86::SPL, 0);
11022 else if (VT == MVT::i64)
11023 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
11024 X86::RSI, X86::RDI, X86::R8, X86::R9,
11025 X86::R10, X86::R11, X86::R12,
11026 X86::R13, X86::R14, X86::R15,
11027 X86::RBP, X86::RSP, 0);
11031 // 32-bit fallthrough
11032 case 'Q': // Q_REGS
11033 if (VT == MVT::i32)
11034 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
11035 else if (VT == MVT::i16)
11036 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
11037 else if (VT == MVT::i8)
11038 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
11039 else if (VT == MVT::i64)
11040 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
11045 return std::vector<unsigned>();
11048 std::pair<unsigned, const TargetRegisterClass*>
11049 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
11051 // First, see if this is a constraint that directly corresponds to an LLVM
11053 if (Constraint.size() == 1) {
11054 // GCC Constraint Letters
11055 switch (Constraint[0]) {
11057 case 'r': // GENERAL_REGS
11058 case 'l': // INDEX_REGS
11060 return std::make_pair(0U, X86::GR8RegisterClass);
11061 if (VT == MVT::i16)
11062 return std::make_pair(0U, X86::GR16RegisterClass);
11063 if (VT == MVT::i32 || !Subtarget->is64Bit())
11064 return std::make_pair(0U, X86::GR32RegisterClass);
11065 return std::make_pair(0U, X86::GR64RegisterClass);
11066 case 'R': // LEGACY_REGS
11068 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
11069 if (VT == MVT::i16)
11070 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
11071 if (VT == MVT::i32 || !Subtarget->is64Bit())
11072 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
11073 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
11074 case 'f': // FP Stack registers.
11075 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
11076 // value to the correct fpstack register class.
11077 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
11078 return std::make_pair(0U, X86::RFP32RegisterClass);
11079 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
11080 return std::make_pair(0U, X86::RFP64RegisterClass);
11081 return std::make_pair(0U, X86::RFP80RegisterClass);
11082 case 'y': // MMX_REGS if MMX allowed.
11083 if (!Subtarget->hasMMX()) break;
11084 return std::make_pair(0U, X86::VR64RegisterClass);
11085 case 'Y': // SSE_REGS if SSE2 allowed
11086 if (!Subtarget->hasSSE2()) break;
11088 case 'x': // SSE_REGS if SSE1 allowed
11089 if (!Subtarget->hasSSE1()) break;
11091 switch (VT.getSimpleVT().SimpleTy) {
11093 // Scalar SSE types.
11096 return std::make_pair(0U, X86::FR32RegisterClass);
11099 return std::make_pair(0U, X86::FR64RegisterClass);
11107 return std::make_pair(0U, X86::VR128RegisterClass);
11113 // Use the default implementation in TargetLowering to convert the register
11114 // constraint into a member of a register class.
11115 std::pair<unsigned, const TargetRegisterClass*> Res;
11116 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
11118 // Not found as a standard register?
11119 if (Res.second == 0) {
11120 // Map st(0) -> st(7) -> ST0
11121 if (Constraint.size() == 7 && Constraint[0] == '{' &&
11122 tolower(Constraint[1]) == 's' &&
11123 tolower(Constraint[2]) == 't' &&
11124 Constraint[3] == '(' &&
11125 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
11126 Constraint[5] == ')' &&
11127 Constraint[6] == '}') {
11129 Res.first = X86::ST0+Constraint[4]-'0';
11130 Res.second = X86::RFP80RegisterClass;
11134 // GCC allows "st(0)" to be called just plain "st".
11135 if (StringRef("{st}").equals_lower(Constraint)) {
11136 Res.first = X86::ST0;
11137 Res.second = X86::RFP80RegisterClass;
11142 if (StringRef("{flags}").equals_lower(Constraint)) {
11143 Res.first = X86::EFLAGS;
11144 Res.second = X86::CCRRegisterClass;
11148 // 'A' means EAX + EDX.
11149 if (Constraint == "A") {
11150 Res.first = X86::EAX;
11151 Res.second = X86::GR32_ADRegisterClass;
11157 // Otherwise, check to see if this is a register class of the wrong value
11158 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
11159 // turn into {ax},{dx}.
11160 if (Res.second->hasType(VT))
11161 return Res; // Correct type already, nothing to do.
11163 // All of the single-register GCC register classes map their values onto
11164 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
11165 // really want an 8-bit or 32-bit register, map to the appropriate register
11166 // class and return the appropriate register.
11167 if (Res.second == X86::GR16RegisterClass) {
11168 if (VT == MVT::i8) {
11169 unsigned DestReg = 0;
11170 switch (Res.first) {
11172 case X86::AX: DestReg = X86::AL; break;
11173 case X86::DX: DestReg = X86::DL; break;
11174 case X86::CX: DestReg = X86::CL; break;
11175 case X86::BX: DestReg = X86::BL; break;
11178 Res.first = DestReg;
11179 Res.second = X86::GR8RegisterClass;
11181 } else if (VT == MVT::i32) {
11182 unsigned DestReg = 0;
11183 switch (Res.first) {
11185 case X86::AX: DestReg = X86::EAX; break;
11186 case X86::DX: DestReg = X86::EDX; break;
11187 case X86::CX: DestReg = X86::ECX; break;
11188 case X86::BX: DestReg = X86::EBX; break;
11189 case X86::SI: DestReg = X86::ESI; break;
11190 case X86::DI: DestReg = X86::EDI; break;
11191 case X86::BP: DestReg = X86::EBP; break;
11192 case X86::SP: DestReg = X86::ESP; break;
11195 Res.first = DestReg;
11196 Res.second = X86::GR32RegisterClass;
11198 } else if (VT == MVT::i64) {
11199 unsigned DestReg = 0;
11200 switch (Res.first) {
11202 case X86::AX: DestReg = X86::RAX; break;
11203 case X86::DX: DestReg = X86::RDX; break;
11204 case X86::CX: DestReg = X86::RCX; break;
11205 case X86::BX: DestReg = X86::RBX; break;
11206 case X86::SI: DestReg = X86::RSI; break;
11207 case X86::DI: DestReg = X86::RDI; break;
11208 case X86::BP: DestReg = X86::RBP; break;
11209 case X86::SP: DestReg = X86::RSP; break;
11212 Res.first = DestReg;
11213 Res.second = X86::GR64RegisterClass;
11216 } else if (Res.second == X86::FR32RegisterClass ||
11217 Res.second == X86::FR64RegisterClass ||
11218 Res.second == X86::VR128RegisterClass) {
11219 // Handle references to XMM physical registers that got mapped into the
11220 // wrong class. This can happen with constraints like {xmm0} where the
11221 // target independent register mapper will just pick the first match it can
11222 // find, ignoring the required type.
11223 if (VT == MVT::f32)
11224 Res.second = X86::FR32RegisterClass;
11225 else if (VT == MVT::f64)
11226 Res.second = X86::FR64RegisterClass;
11227 else if (X86::VR128RegisterClass->hasType(VT))
11228 Res.second = X86::VR128RegisterClass;