1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalAlias.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/LLVMContext.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/Dwarf.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/raw_ostream.h"
53 using namespace dwarf;
55 STATISTIC(NumTailCalls, "Number of tail calls");
58 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
60 // Forward declarations.
61 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
64 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
66 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
68 if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
69 if (is64Bit) return new X8664_MachoTargetObjectFile();
70 return new TargetLoweringObjectFileMachO();
71 } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
72 if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
73 return new X8632_ELFTargetObjectFile(TM);
74 } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
75 return new TargetLoweringObjectFileCOFF();
77 llvm_unreachable("unknown subtarget type");
80 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
81 : TargetLowering(TM, createTLOF(TM)) {
82 Subtarget = &TM.getSubtarget<X86Subtarget>();
83 X86ScalarSSEf64 = Subtarget->hasSSE2();
84 X86ScalarSSEf32 = Subtarget->hasSSE1();
85 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
87 RegInfo = TM.getRegisterInfo();
90 // Set up the TargetLowering object.
92 // X86 is weird, it always uses i8 for shift amounts and setcc results.
93 setShiftAmountType(MVT::i8);
94 setBooleanContents(ZeroOrOneBooleanContent);
95 setSchedulingPreference(Sched::RegPressure);
96 setStackPointerRegisterToSaveRestore(X86StackPtr);
98 if (Subtarget->isTargetDarwin()) {
99 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
100 setUseUnderscoreSetJmp(false);
101 setUseUnderscoreLongJmp(false);
102 } else if (Subtarget->isTargetMingw()) {
103 // MS runtime is weird: it exports _setjmp, but longjmp!
104 setUseUnderscoreSetJmp(true);
105 setUseUnderscoreLongJmp(false);
107 setUseUnderscoreSetJmp(true);
108 setUseUnderscoreLongJmp(true);
111 // Set up the register classes.
112 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
113 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
114 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
115 if (Subtarget->is64Bit())
116 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
118 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
120 // We don't accept any truncstore of integer registers.
121 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
122 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
123 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
124 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
125 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
128 // SETOEQ and SETUNE require checking two conditions.
129 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
130 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
131 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
132 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
133 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
136 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
138 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
139 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
140 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
142 if (Subtarget->is64Bit()) {
143 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
144 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
145 } else if (!UseSoftFloat) {
146 // We have an algorithm for SSE2->double, and we turn this into a
147 // 64-bit FILD followed by conditional FADD for other targets.
148 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
149 // We have an algorithm for SSE2, and we turn this into a 64-bit
150 // FILD for other targets.
151 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
154 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
156 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
157 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
160 // SSE has no i16 to fp conversion, only i32
161 if (X86ScalarSSEf32) {
162 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
163 // f32 and f64 cases are Legal, f80 case is not
164 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
166 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
167 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
170 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
171 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
174 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
175 // are Legal, f80 is custom lowered.
176 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
179 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
181 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
182 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
184 if (X86ScalarSSEf32) {
185 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
186 // f32 and f64 cases are Legal, f80 case is not
187 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
189 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
190 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
193 // Handle FP_TO_UINT by promoting the destination to a larger signed
195 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
196 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
197 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
199 if (Subtarget->is64Bit()) {
200 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
201 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
202 } else if (!UseSoftFloat) {
203 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
204 // Expand FP_TO_UINT into a select.
205 // FIXME: We would like to use a Custom expander here eventually to do
206 // the optimal thing for SSE vs. the default expansion in the legalizer.
207 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
209 // With SSE3 we can use fisttpll to convert to a signed i64; without
210 // SSE, we're stuck with a fistpll.
211 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
214 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
215 if (!X86ScalarSSEf64) {
216 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
217 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
218 if (Subtarget->is64Bit()) {
219 setOperationAction(ISD::BIT_CONVERT , MVT::f64 , Expand);
220 // Without SSE, i64->f64 goes through memory; i64->MMX is Legal.
221 if (Subtarget->hasMMX() && !DisableMMX)
222 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Custom);
224 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Expand);
228 // Scalar integer divide and remainder are lowered to use operations that
229 // produce two results, to match the available instructions. This exposes
230 // the two-result form to trivial CSE, which is able to combine x/y and x%y
231 // into a single instruction.
233 // Scalar integer multiply-high is also lowered to use two-result
234 // operations, to match the available instructions. However, plain multiply
235 // (low) operations are left as Legal, as there are single-result
236 // instructions for this in x86. Using the two-result multiply instructions
237 // when both high and low results are needed must be arranged by dagcombine.
238 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
239 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
240 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
241 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
242 setOperationAction(ISD::SREM , MVT::i8 , Expand);
243 setOperationAction(ISD::UREM , MVT::i8 , Expand);
244 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
245 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
246 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
247 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
248 setOperationAction(ISD::SREM , MVT::i16 , Expand);
249 setOperationAction(ISD::UREM , MVT::i16 , Expand);
250 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
251 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
252 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
253 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
254 setOperationAction(ISD::SREM , MVT::i32 , Expand);
255 setOperationAction(ISD::UREM , MVT::i32 , Expand);
256 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
257 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
258 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
259 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
260 setOperationAction(ISD::SREM , MVT::i64 , Expand);
261 setOperationAction(ISD::UREM , MVT::i64 , Expand);
263 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
264 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
265 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
266 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
267 if (Subtarget->is64Bit())
268 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
269 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
270 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
272 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
273 setOperationAction(ISD::FREM , MVT::f32 , Expand);
274 setOperationAction(ISD::FREM , MVT::f64 , Expand);
275 setOperationAction(ISD::FREM , MVT::f80 , Expand);
276 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
278 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
279 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
280 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
281 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
282 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
283 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
287 if (Subtarget->is64Bit()) {
288 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
293 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
294 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
296 // These should be promoted to a larger select which is supported.
297 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
298 // X86 wants to expand cmov itself.
299 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
300 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
301 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
302 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
303 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
304 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
305 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
306 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
307 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
308 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
309 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
310 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
311 if (Subtarget->is64Bit()) {
312 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
313 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
315 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
318 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
319 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
320 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
321 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
322 if (Subtarget->is64Bit())
323 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
324 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
325 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
326 if (Subtarget->is64Bit()) {
327 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
328 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
329 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
330 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
331 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
333 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
334 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
335 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
336 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
337 if (Subtarget->is64Bit()) {
338 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
339 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
340 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
343 if (Subtarget->hasSSE1())
344 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
346 // We may not have a libcall for MEMBARRIER so we should lower this.
347 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
349 // On X86 and X86-64, atomic operations are lowered to locked instructions.
350 // Locked instructions, in turn, have implicit fence semantics (all memory
351 // operations are flushed before issuing the locked instruction, and they
352 // are not buffered), so we can fold away the common pattern of
353 // fence-atomic-fence.
354 setShouldFoldAtomicFences(true);
356 // Expand certain atomics
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
360 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
365 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
367 if (!Subtarget->is64Bit()) {
368 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
374 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
377 // FIXME - use subtarget debug flags
378 if (!Subtarget->isTargetDarwin() &&
379 !Subtarget->isTargetELF() &&
380 !Subtarget->isTargetCygMing()) {
381 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
384 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
385 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
386 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
387 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
388 if (Subtarget->is64Bit()) {
389 setExceptionPointerRegister(X86::RAX);
390 setExceptionSelectorRegister(X86::RDX);
392 setExceptionPointerRegister(X86::EAX);
393 setExceptionSelectorRegister(X86::EDX);
395 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
396 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
398 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
400 setOperationAction(ISD::TRAP, MVT::Other, Legal);
402 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
403 setOperationAction(ISD::VASTART , MVT::Other, Custom);
404 setOperationAction(ISD::VAEND , MVT::Other, Expand);
405 if (Subtarget->is64Bit()) {
406 setOperationAction(ISD::VAARG , MVT::Other, Custom);
407 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
409 setOperationAction(ISD::VAARG , MVT::Other, Expand);
410 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
413 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
414 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
415 if (Subtarget->is64Bit())
416 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
417 if (Subtarget->isTargetCygMing())
418 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
420 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
422 if (!UseSoftFloat && X86ScalarSSEf64) {
423 // f32 and f64 use SSE.
424 // Set up the FP register classes.
425 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
426 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
428 // Use ANDPD to simulate FABS.
429 setOperationAction(ISD::FABS , MVT::f64, Custom);
430 setOperationAction(ISD::FABS , MVT::f32, Custom);
432 // Use XORP to simulate FNEG.
433 setOperationAction(ISD::FNEG , MVT::f64, Custom);
434 setOperationAction(ISD::FNEG , MVT::f32, Custom);
436 // Use ANDPD and ORPD to simulate FCOPYSIGN.
437 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
438 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
440 // We don't support sin/cos/fmod
441 setOperationAction(ISD::FSIN , MVT::f64, Expand);
442 setOperationAction(ISD::FCOS , MVT::f64, Expand);
443 setOperationAction(ISD::FSIN , MVT::f32, Expand);
444 setOperationAction(ISD::FCOS , MVT::f32, Expand);
446 // Expand FP immediates into loads from the stack, except for the special
448 addLegalFPImmediate(APFloat(+0.0)); // xorpd
449 addLegalFPImmediate(APFloat(+0.0f)); // xorps
450 } else if (!UseSoftFloat && X86ScalarSSEf32) {
451 // Use SSE for f32, x87 for f64.
452 // Set up the FP register classes.
453 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
454 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
456 // Use ANDPS to simulate FABS.
457 setOperationAction(ISD::FABS , MVT::f32, Custom);
459 // Use XORP to simulate FNEG.
460 setOperationAction(ISD::FNEG , MVT::f32, Custom);
462 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
464 // Use ANDPS and ORPS to simulate FCOPYSIGN.
465 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
466 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
468 // We don't support sin/cos/fmod
469 setOperationAction(ISD::FSIN , MVT::f32, Expand);
470 setOperationAction(ISD::FCOS , MVT::f32, Expand);
472 // Special cases we handle for FP constants.
473 addLegalFPImmediate(APFloat(+0.0f)); // xorps
474 addLegalFPImmediate(APFloat(+0.0)); // FLD0
475 addLegalFPImmediate(APFloat(+1.0)); // FLD1
476 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
477 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
480 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
481 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
483 } else if (!UseSoftFloat) {
484 // f32 and f64 in x87.
485 // Set up the FP register classes.
486 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
487 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
489 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
490 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
491 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
492 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
495 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
496 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
498 addLegalFPImmediate(APFloat(+0.0)); // FLD0
499 addLegalFPImmediate(APFloat(+1.0)); // FLD1
500 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
501 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
502 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
503 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
504 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
505 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
508 // Long double always uses X87.
510 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
511 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
512 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
515 APFloat TmpFlt(+0.0);
516 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
518 addLegalFPImmediate(TmpFlt); // FLD0
520 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
521 APFloat TmpFlt2(+1.0);
522 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
524 addLegalFPImmediate(TmpFlt2); // FLD1
525 TmpFlt2.changeSign();
526 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
530 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
531 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
535 // Always use a library call for pow.
536 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
537 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
538 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
540 setOperationAction(ISD::FLOG, MVT::f80, Expand);
541 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
542 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
543 setOperationAction(ISD::FEXP, MVT::f80, Expand);
544 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
546 // First set operation action for all vector types to either promote
547 // (for widening) or expand (for scalarization). Then we will selectively
548 // turn on ones that can be effectively codegen'd.
549 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
550 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
551 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
566 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
567 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
600 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
604 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
605 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
606 setTruncStoreAction((MVT::SimpleValueType)VT,
607 (MVT::SimpleValueType)InnerVT, Expand);
608 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
609 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
610 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
613 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
614 // with -msoft-float, disable use of MMX as well.
615 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
616 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
617 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
618 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
620 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
622 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
623 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
624 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
625 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
627 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
628 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
629 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
630 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
632 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
633 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
635 setOperationAction(ISD::AND, MVT::v8i8, Promote);
636 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
637 setOperationAction(ISD::AND, MVT::v4i16, Promote);
638 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
639 setOperationAction(ISD::AND, MVT::v2i32, Promote);
640 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
641 setOperationAction(ISD::AND, MVT::v1i64, Legal);
643 setOperationAction(ISD::OR, MVT::v8i8, Promote);
644 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
645 setOperationAction(ISD::OR, MVT::v4i16, Promote);
646 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
647 setOperationAction(ISD::OR, MVT::v2i32, Promote);
648 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
649 setOperationAction(ISD::OR, MVT::v1i64, Legal);
651 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
652 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
653 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
654 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
655 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
656 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
657 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
659 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
660 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
661 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
662 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
663 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
664 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
665 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
667 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
669 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
670 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
672 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
674 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
675 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
677 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
678 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
679 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
681 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
683 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
684 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
685 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
686 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
689 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
691 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
692 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
693 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
694 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
695 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
699 if (!UseSoftFloat && Subtarget->hasSSE1()) {
700 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
702 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
703 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
704 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
705 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
706 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
707 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
708 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
709 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
710 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
711 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
712 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
713 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
716 if (!UseSoftFloat && Subtarget->hasSSE2()) {
717 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
719 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
720 // registers cannot be used even for integer operations.
721 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
722 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
723 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
724 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
726 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
727 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
728 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
729 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
730 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
731 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
732 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
733 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
734 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
735 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
736 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
737 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
738 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
739 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
740 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
741 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
743 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
744 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
745 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
746 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
748 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
749 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
750 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
751 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
752 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
754 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
755 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
756 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
757 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
760 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
761 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
762 EVT VT = (MVT::SimpleValueType)i;
763 // Do not attempt to custom lower non-power-of-2 vectors
764 if (!isPowerOf2_32(VT.getVectorNumElements()))
766 // Do not attempt to custom lower non-128-bit vectors
767 if (!VT.is128BitVector())
769 setOperationAction(ISD::BUILD_VECTOR,
770 VT.getSimpleVT().SimpleTy, Custom);
771 setOperationAction(ISD::VECTOR_SHUFFLE,
772 VT.getSimpleVT().SimpleTy, Custom);
773 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
774 VT.getSimpleVT().SimpleTy, Custom);
777 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
778 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
779 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
780 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
781 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
782 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
784 if (Subtarget->is64Bit()) {
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
789 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
790 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
791 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
794 // Do not attempt to promote non-128-bit vectors
795 if (!VT.is128BitVector())
798 setOperationAction(ISD::AND, SVT, Promote);
799 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
800 setOperationAction(ISD::OR, SVT, Promote);
801 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
802 setOperationAction(ISD::XOR, SVT, Promote);
803 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
804 setOperationAction(ISD::LOAD, SVT, Promote);
805 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
806 setOperationAction(ISD::SELECT, SVT, Promote);
807 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
810 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
812 // Custom lower v2i64 and v2f64 selects.
813 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
814 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
815 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
816 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
818 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
819 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
820 if (!DisableMMX && Subtarget->hasMMX()) {
821 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
822 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
826 if (Subtarget->hasSSE41()) {
827 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
828 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
829 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
830 setOperationAction(ISD::FRINT, MVT::f32, Legal);
831 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
832 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
833 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
834 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
835 setOperationAction(ISD::FRINT, MVT::f64, Legal);
836 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
838 // FIXME: Do we need to handle scalar-to-vector here?
839 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
841 // Can turn SHL into an integer multiply.
842 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
843 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
845 // i8 and i16 vectors are custom , because the source register and source
846 // source memory operand types are not the same width. f32 vectors are
847 // custom since the immediate controlling the insert encodes additional
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
850 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
851 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
852 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
854 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
855 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
856 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
857 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
859 if (Subtarget->is64Bit()) {
860 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
861 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
865 if (Subtarget->hasSSE42()) {
866 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
869 if (!UseSoftFloat && Subtarget->hasAVX()) {
870 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
871 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
872 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
873 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
874 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
876 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
877 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
878 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
879 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
880 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
881 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
882 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
883 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
884 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
885 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
886 setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
887 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
888 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
889 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
890 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
892 // Operations to consider commented out -v16i16 v32i8
893 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
894 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
895 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
896 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
897 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
898 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
899 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
900 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
901 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
902 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
903 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
904 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
905 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
906 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
908 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
909 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
910 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
911 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
913 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
914 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
915 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
916 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
917 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
919 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
920 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
921 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
922 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
923 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
924 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
927 // Not sure we want to do this since there are no 256-bit integer
930 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
931 // This includes 256-bit vectors
932 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
933 EVT VT = (MVT::SimpleValueType)i;
935 // Do not attempt to custom lower non-power-of-2 vectors
936 if (!isPowerOf2_32(VT.getVectorNumElements()))
939 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
940 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
941 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
944 if (Subtarget->is64Bit()) {
945 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
946 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
951 // Not sure we want to do this since there are no 256-bit integer
954 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
955 // Including 256-bit vectors
956 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
957 EVT VT = (MVT::SimpleValueType)i;
959 if (!VT.is256BitVector()) {
962 setOperationAction(ISD::AND, VT, Promote);
963 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
964 setOperationAction(ISD::OR, VT, Promote);
965 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
966 setOperationAction(ISD::XOR, VT, Promote);
967 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
968 setOperationAction(ISD::LOAD, VT, Promote);
969 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
970 setOperationAction(ISD::SELECT, VT, Promote);
971 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
974 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
978 // We want to custom lower some of our intrinsics.
979 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
981 // Add/Sub/Mul with overflow operations are custom lowered.
982 setOperationAction(ISD::SADDO, MVT::i32, Custom);
983 setOperationAction(ISD::UADDO, MVT::i32, Custom);
984 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
985 setOperationAction(ISD::USUBO, MVT::i32, Custom);
986 setOperationAction(ISD::SMULO, MVT::i32, Custom);
988 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
989 // handle type legalization for these operations here.
991 // FIXME: We really should do custom legalization for addition and
992 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
993 // than generic legalization for 64-bit multiplication-with-overflow, though.
994 if (Subtarget->is64Bit()) {
995 setOperationAction(ISD::SADDO, MVT::i64, Custom);
996 setOperationAction(ISD::UADDO, MVT::i64, Custom);
997 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
998 setOperationAction(ISD::USUBO, MVT::i64, Custom);
999 setOperationAction(ISD::SMULO, MVT::i64, Custom);
1002 if (!Subtarget->is64Bit()) {
1003 // These libcalls are not available in 32-bit.
1004 setLibcallName(RTLIB::SHL_I128, 0);
1005 setLibcallName(RTLIB::SRL_I128, 0);
1006 setLibcallName(RTLIB::SRA_I128, 0);
1009 // We have target-specific dag combine patterns for the following nodes:
1010 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1011 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1012 setTargetDAGCombine(ISD::BUILD_VECTOR);
1013 setTargetDAGCombine(ISD::SELECT);
1014 setTargetDAGCombine(ISD::SHL);
1015 setTargetDAGCombine(ISD::SRA);
1016 setTargetDAGCombine(ISD::SRL);
1017 setTargetDAGCombine(ISD::OR);
1018 setTargetDAGCombine(ISD::STORE);
1019 setTargetDAGCombine(ISD::ZERO_EXTEND);
1020 if (Subtarget->is64Bit())
1021 setTargetDAGCombine(ISD::MUL);
1023 computeRegisterProperties();
1025 // FIXME: These should be based on subtarget info. Plus, the values should
1026 // be smaller when we are in optimizing for size mode.
1027 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1028 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1029 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1030 setPrefLoopAlignment(16);
1031 benefitFromCodePlacementOpt = true;
1035 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1040 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1041 /// the desired ByVal argument alignment.
1042 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1045 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1046 if (VTy->getBitWidth() == 128)
1048 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1049 unsigned EltAlign = 0;
1050 getMaxByValAlign(ATy->getElementType(), EltAlign);
1051 if (EltAlign > MaxAlign)
1052 MaxAlign = EltAlign;
1053 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1054 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1055 unsigned EltAlign = 0;
1056 getMaxByValAlign(STy->getElementType(i), EltAlign);
1057 if (EltAlign > MaxAlign)
1058 MaxAlign = EltAlign;
1066 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1067 /// function arguments in the caller parameter area. For X86, aggregates
1068 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1069 /// are at 4-byte boundaries.
1070 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1071 if (Subtarget->is64Bit()) {
1072 // Max of 8 and alignment of type.
1073 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1080 if (Subtarget->hasSSE1())
1081 getMaxByValAlign(Ty, Align);
1085 /// getOptimalMemOpType - Returns the target specific optimal type for load
1086 /// and store operations as a result of memset, memcpy, and memmove
1087 /// lowering. If DstAlign is zero that means it's safe to destination
1088 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1089 /// means there isn't a need to check it against alignment requirement,
1090 /// probably because the source does not need to be loaded. If
1091 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1092 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1093 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1094 /// constant so it does not need to be loaded.
1095 /// It returns EVT::Other if the type should be determined using generic
1096 /// target-independent logic.
1098 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1099 unsigned DstAlign, unsigned SrcAlign,
1100 bool NonScalarIntSafe,
1102 MachineFunction &MF) const {
1103 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1104 // linux. This is because the stack realignment code can't handle certain
1105 // cases like PR2962. This should be removed when PR2962 is fixed.
1106 const Function *F = MF.getFunction();
1107 if (NonScalarIntSafe &&
1108 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1110 (Subtarget->isUnalignedMemAccessFast() ||
1111 ((DstAlign == 0 || DstAlign >= 16) &&
1112 (SrcAlign == 0 || SrcAlign >= 16))) &&
1113 Subtarget->getStackAlignment() >= 16) {
1114 if (Subtarget->hasSSE2())
1116 if (Subtarget->hasSSE1())
1118 } else if (!MemcpyStrSrc && Size >= 8 &&
1119 !Subtarget->is64Bit() &&
1120 Subtarget->getStackAlignment() >= 8 &&
1121 Subtarget->hasSSE2()) {
1122 // Do not use f64 to lower memcpy if source is string constant. It's
1123 // better to use i32 to avoid the loads.
1127 if (Subtarget->is64Bit() && Size >= 8)
1132 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1133 /// current function. The returned value is a member of the
1134 /// MachineJumpTableInfo::JTEntryKind enum.
1135 unsigned X86TargetLowering::getJumpTableEncoding() const {
1136 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1138 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1139 Subtarget->isPICStyleGOT())
1140 return MachineJumpTableInfo::EK_Custom32;
1142 // Otherwise, use the normal jump table encoding heuristics.
1143 return TargetLowering::getJumpTableEncoding();
1146 /// getPICBaseSymbol - Return the X86-32 PIC base.
1148 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1149 MCContext &Ctx) const {
1150 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1151 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1152 Twine(MF->getFunctionNumber())+"$pb");
1157 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1158 const MachineBasicBlock *MBB,
1159 unsigned uid,MCContext &Ctx) const{
1160 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1161 Subtarget->isPICStyleGOT());
1162 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1164 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1165 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1168 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1170 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1171 SelectionDAG &DAG) const {
1172 if (!Subtarget->is64Bit())
1173 // This doesn't have DebugLoc associated with it, but is not really the
1174 // same as a Register.
1175 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1179 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1180 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1182 const MCExpr *X86TargetLowering::
1183 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1184 MCContext &Ctx) const {
1185 // X86-64 uses RIP relative addressing based on the jump table label.
1186 if (Subtarget->isPICStyleRIPRel())
1187 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1189 // Otherwise, the reference is relative to the PIC base.
1190 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1193 /// getFunctionAlignment - Return the Log2 alignment of this function.
1194 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1195 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1198 std::pair<const TargetRegisterClass*, uint8_t>
1199 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1200 const TargetRegisterClass *RRC = 0;
1202 switch (VT.getSimpleVT().SimpleTy) {
1204 return TargetLowering::findRepresentativeClass(VT);
1205 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1206 RRC = (Subtarget->is64Bit()
1207 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1209 case MVT::v8i8: case MVT::v4i16:
1210 case MVT::v2i32: case MVT::v1i64:
1211 RRC = X86::VR64RegisterClass;
1213 case MVT::f32: case MVT::f64:
1214 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1215 case MVT::v4f32: case MVT::v2f64:
1216 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1218 RRC = X86::VR128RegisterClass;
1221 return std::make_pair(RRC, Cost);
1225 X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
1226 MachineFunction &MF) const {
1227 unsigned FPDiff = RegInfo->hasFP(MF) ? 1 : 0;
1228 switch (RC->getID()) {
1231 case X86::GR32RegClassID:
1233 case X86::GR64RegClassID:
1235 case X86::VR128RegClassID:
1236 return Subtarget->is64Bit() ? 10 : 4;
1237 case X86::VR64RegClassID:
1242 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1243 unsigned &Offset) const {
1244 if (!Subtarget->isTargetLinux())
1247 if (Subtarget->is64Bit()) {
1248 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1250 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1263 //===----------------------------------------------------------------------===//
1264 // Return Value Calling Convention Implementation
1265 //===----------------------------------------------------------------------===//
1267 #include "X86GenCallingConv.inc"
1270 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1271 const SmallVectorImpl<ISD::OutputArg> &Outs,
1272 LLVMContext &Context) const {
1273 SmallVector<CCValAssign, 16> RVLocs;
1274 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1276 return CCInfo.CheckReturn(Outs, RetCC_X86);
1280 X86TargetLowering::LowerReturn(SDValue Chain,
1281 CallingConv::ID CallConv, bool isVarArg,
1282 const SmallVectorImpl<ISD::OutputArg> &Outs,
1283 const SmallVectorImpl<SDValue> &OutVals,
1284 DebugLoc dl, SelectionDAG &DAG) const {
1285 MachineFunction &MF = DAG.getMachineFunction();
1286 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1288 SmallVector<CCValAssign, 16> RVLocs;
1289 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1290 RVLocs, *DAG.getContext());
1291 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1293 // Add the regs to the liveout set for the function.
1294 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1295 for (unsigned i = 0; i != RVLocs.size(); ++i)
1296 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1297 MRI.addLiveOut(RVLocs[i].getLocReg());
1301 SmallVector<SDValue, 6> RetOps;
1302 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1303 // Operand #1 = Bytes To Pop
1304 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1307 // Copy the result values into the output registers.
1308 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1309 CCValAssign &VA = RVLocs[i];
1310 assert(VA.isRegLoc() && "Can only return in registers!");
1311 SDValue ValToCopy = OutVals[i];
1312 EVT ValVT = ValToCopy.getValueType();
1314 // If this is x86-64, and we disabled SSE, we can't return FP values
1315 if ((ValVT == MVT::f32 || ValVT == MVT::f64) &&
1316 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1317 report_fatal_error("SSE register return with SSE disabled");
1319 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1320 // llvm-gcc has never done it right and no one has noticed, so this
1321 // should be OK for now.
1322 if (ValVT == MVT::f64 &&
1323 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1324 report_fatal_error("SSE2 register return with SSE2 disabled");
1326 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1327 // the RET instruction and handled by the FP Stackifier.
1328 if (VA.getLocReg() == X86::ST0 ||
1329 VA.getLocReg() == X86::ST1) {
1330 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1331 // change the value to the FP stack register class.
1332 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1333 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1334 RetOps.push_back(ValToCopy);
1335 // Don't emit a copytoreg.
1339 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1340 // which is returned in RAX / RDX.
1341 if (Subtarget->is64Bit()) {
1342 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1343 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1344 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1345 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1348 // If we don't have SSE2 available, convert to v4f32 so the generated
1349 // register is legal.
1350 if (!Subtarget->hasSSE2())
1351 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32,ValToCopy);
1356 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1357 Flag = Chain.getValue(1);
1360 // The x86-64 ABI for returning structs by value requires that we copy
1361 // the sret argument into %rax for the return. We saved the argument into
1362 // a virtual register in the entry block, so now we copy the value out
1364 if (Subtarget->is64Bit() &&
1365 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1366 MachineFunction &MF = DAG.getMachineFunction();
1367 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1368 unsigned Reg = FuncInfo->getSRetReturnReg();
1370 "SRetReturnReg should have been set in LowerFormalArguments().");
1371 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1373 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1374 Flag = Chain.getValue(1);
1376 // RAX now acts like a return value.
1377 MRI.addLiveOut(X86::RAX);
1380 RetOps[0] = Chain; // Update chain.
1382 // Add the flag if we have it.
1384 RetOps.push_back(Flag);
1386 return DAG.getNode(X86ISD::RET_FLAG, dl,
1387 MVT::Other, &RetOps[0], RetOps.size());
1390 /// LowerCallResult - Lower the result values of a call into the
1391 /// appropriate copies out of appropriate physical registers.
1394 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1395 CallingConv::ID CallConv, bool isVarArg,
1396 const SmallVectorImpl<ISD::InputArg> &Ins,
1397 DebugLoc dl, SelectionDAG &DAG,
1398 SmallVectorImpl<SDValue> &InVals) const {
1400 // Assign locations to each value returned by this call.
1401 SmallVector<CCValAssign, 16> RVLocs;
1402 bool Is64Bit = Subtarget->is64Bit();
1403 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1404 RVLocs, *DAG.getContext());
1405 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1407 // Copy all of the result registers out of their specified physreg.
1408 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1409 CCValAssign &VA = RVLocs[i];
1410 EVT CopyVT = VA.getValVT();
1412 // If this is x86-64, and we disabled SSE, we can't return FP values
1413 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1414 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1415 report_fatal_error("SSE register return with SSE disabled");
1420 // If this is a call to a function that returns an fp value on the floating
1421 // point stack, we must guarantee the the value is popped from the stack, so
1422 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1423 // if the return value is not used. We use the FpGET_ST0 instructions
1425 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1426 // If we prefer to use the value in xmm registers, copy it out as f80 and
1427 // use a truncate to move it from fp stack reg to xmm reg.
1428 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1429 bool isST0 = VA.getLocReg() == X86::ST0;
1431 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1432 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1433 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1434 SDValue Ops[] = { Chain, InFlag };
1435 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
1437 Val = Chain.getValue(0);
1439 // Round the f80 to the right size, which also moves it to the appropriate
1441 if (CopyVT != VA.getValVT())
1442 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1443 // This truncation won't change the value.
1444 DAG.getIntPtrConstant(1));
1445 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1446 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1447 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1448 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1449 MVT::v2i64, InFlag).getValue(1);
1450 Val = Chain.getValue(0);
1451 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1452 Val, DAG.getConstant(0, MVT::i64));
1454 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1455 MVT::i64, InFlag).getValue(1);
1456 Val = Chain.getValue(0);
1458 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1460 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1461 CopyVT, InFlag).getValue(1);
1462 Val = Chain.getValue(0);
1464 InFlag = Chain.getValue(2);
1465 InVals.push_back(Val);
1472 //===----------------------------------------------------------------------===//
1473 // C & StdCall & Fast Calling Convention implementation
1474 //===----------------------------------------------------------------------===//
1475 // StdCall calling convention seems to be standard for many Windows' API
1476 // routines and around. It differs from C calling convention just a little:
1477 // callee should clean up the stack, not caller. Symbols should be also
1478 // decorated in some fancy way :) It doesn't support any vector arguments.
1479 // For info on fast calling convention see Fast Calling Convention (tail call)
1480 // implementation LowerX86_32FastCCCallTo.
1482 /// CallIsStructReturn - Determines whether a call uses struct return
1484 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1488 return Outs[0].Flags.isSRet();
1491 /// ArgsAreStructReturn - Determines whether a function uses struct
1492 /// return semantics.
1494 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1498 return Ins[0].Flags.isSRet();
1501 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1502 /// given CallingConvention value.
1503 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1504 if (Subtarget->is64Bit()) {
1505 if (CC == CallingConv::GHC)
1506 return CC_X86_64_GHC;
1507 else if (Subtarget->isTargetWin64())
1508 return CC_X86_Win64_C;
1513 if (CC == CallingConv::X86_FastCall)
1514 return CC_X86_32_FastCall;
1515 else if (CC == CallingConv::X86_ThisCall)
1516 return CC_X86_32_ThisCall;
1517 else if (CC == CallingConv::Fast)
1518 return CC_X86_32_FastCC;
1519 else if (CC == CallingConv::GHC)
1520 return CC_X86_32_GHC;
1525 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1526 /// by "Src" to address "Dst" with size and alignment information specified by
1527 /// the specific parameter attribute. The copy will be passed as a byval
1528 /// function parameter.
1530 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1531 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1533 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1534 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1535 /*isVolatile*/false, /*AlwaysInline=*/true,
1539 /// IsTailCallConvention - Return true if the calling convention is one that
1540 /// supports tail call optimization.
1541 static bool IsTailCallConvention(CallingConv::ID CC) {
1542 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1545 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1546 /// a tailcall target by changing its ABI.
1547 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1548 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1552 X86TargetLowering::LowerMemArgument(SDValue Chain,
1553 CallingConv::ID CallConv,
1554 const SmallVectorImpl<ISD::InputArg> &Ins,
1555 DebugLoc dl, SelectionDAG &DAG,
1556 const CCValAssign &VA,
1557 MachineFrameInfo *MFI,
1559 // Create the nodes corresponding to a load from this parameter slot.
1560 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1561 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1562 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1565 // If value is passed by pointer we have address passed instead of the value
1567 if (VA.getLocInfo() == CCValAssign::Indirect)
1568 ValVT = VA.getLocVT();
1570 ValVT = VA.getValVT();
1572 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1573 // changed with more analysis.
1574 // In case of tail call optimization mark all arguments mutable. Since they
1575 // could be overwritten by lowering of arguments in case of a tail call.
1576 if (Flags.isByVal()) {
1577 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1578 VA.getLocMemOffset(), isImmutable);
1579 return DAG.getFrameIndex(FI, getPointerTy());
1581 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1582 VA.getLocMemOffset(), isImmutable);
1583 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1584 return DAG.getLoad(ValVT, dl, Chain, FIN,
1585 PseudoSourceValue::getFixedStack(FI), 0,
1591 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1592 CallingConv::ID CallConv,
1594 const SmallVectorImpl<ISD::InputArg> &Ins,
1597 SmallVectorImpl<SDValue> &InVals)
1599 MachineFunction &MF = DAG.getMachineFunction();
1600 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1602 const Function* Fn = MF.getFunction();
1603 if (Fn->hasExternalLinkage() &&
1604 Subtarget->isTargetCygMing() &&
1605 Fn->getName() == "main")
1606 FuncInfo->setForceFramePointer(true);
1608 MachineFrameInfo *MFI = MF.getFrameInfo();
1609 bool Is64Bit = Subtarget->is64Bit();
1610 bool IsWin64 = Subtarget->isTargetWin64();
1612 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1613 "Var args not supported with calling convention fastcc or ghc");
1615 // Assign locations to all of the incoming arguments.
1616 SmallVector<CCValAssign, 16> ArgLocs;
1617 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1618 ArgLocs, *DAG.getContext());
1619 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1621 unsigned LastVal = ~0U;
1623 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1624 CCValAssign &VA = ArgLocs[i];
1625 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1627 assert(VA.getValNo() != LastVal &&
1628 "Don't support value assigned to multiple locs yet");
1629 LastVal = VA.getValNo();
1631 if (VA.isRegLoc()) {
1632 EVT RegVT = VA.getLocVT();
1633 TargetRegisterClass *RC = NULL;
1634 if (RegVT == MVT::i32)
1635 RC = X86::GR32RegisterClass;
1636 else if (Is64Bit && RegVT == MVT::i64)
1637 RC = X86::GR64RegisterClass;
1638 else if (RegVT == MVT::f32)
1639 RC = X86::FR32RegisterClass;
1640 else if (RegVT == MVT::f64)
1641 RC = X86::FR64RegisterClass;
1642 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1643 RC = X86::VR256RegisterClass;
1644 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1645 RC = X86::VR128RegisterClass;
1646 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1647 RC = X86::VR64RegisterClass;
1649 llvm_unreachable("Unknown argument type!");
1651 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1652 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1654 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1655 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1657 if (VA.getLocInfo() == CCValAssign::SExt)
1658 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1659 DAG.getValueType(VA.getValVT()));
1660 else if (VA.getLocInfo() == CCValAssign::ZExt)
1661 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1662 DAG.getValueType(VA.getValVT()));
1663 else if (VA.getLocInfo() == CCValAssign::BCvt)
1664 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1666 if (VA.isExtInLoc()) {
1667 // Handle MMX values passed in XMM regs.
1668 if (RegVT.isVector()) {
1669 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1670 ArgValue, DAG.getConstant(0, MVT::i64));
1671 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1673 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1676 assert(VA.isMemLoc());
1677 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1680 // If value is passed via pointer - do a load.
1681 if (VA.getLocInfo() == CCValAssign::Indirect)
1682 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1685 InVals.push_back(ArgValue);
1688 // The x86-64 ABI for returning structs by value requires that we copy
1689 // the sret argument into %rax for the return. Save the argument into
1690 // a virtual register so that we can access it from the return points.
1691 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1692 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1693 unsigned Reg = FuncInfo->getSRetReturnReg();
1695 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1696 FuncInfo->setSRetReturnReg(Reg);
1698 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1699 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1702 unsigned StackSize = CCInfo.getNextStackOffset();
1703 // Align stack specially for tail calls.
1704 if (FuncIsMadeTailCallSafe(CallConv))
1705 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1707 // If the function takes variable number of arguments, make a frame index for
1708 // the start of the first vararg value... for expansion of llvm.va_start.
1710 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1711 CallConv != CallingConv::X86_ThisCall)) {
1712 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1715 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1717 // FIXME: We should really autogenerate these arrays
1718 static const unsigned GPR64ArgRegsWin64[] = {
1719 X86::RCX, X86::RDX, X86::R8, X86::R9
1721 static const unsigned XMMArgRegsWin64[] = {
1722 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1724 static const unsigned GPR64ArgRegs64Bit[] = {
1725 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1727 static const unsigned XMMArgRegs64Bit[] = {
1728 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1729 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1731 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1734 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1735 GPR64ArgRegs = GPR64ArgRegsWin64;
1736 XMMArgRegs = XMMArgRegsWin64;
1738 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1739 GPR64ArgRegs = GPR64ArgRegs64Bit;
1740 XMMArgRegs = XMMArgRegs64Bit;
1742 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1744 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1747 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1748 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1749 "SSE register cannot be used when SSE is disabled!");
1750 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1751 "SSE register cannot be used when SSE is disabled!");
1752 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1753 // Kernel mode asks for SSE to be disabled, so don't push them
1755 TotalNumXMMRegs = 0;
1757 // For X86-64, if there are vararg parameters that are passed via
1758 // registers, then we must store them to their spots on the stack so they
1759 // may be loaded by deferencing the result of va_next.
1760 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1761 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1762 FuncInfo->setRegSaveFrameIndex(
1763 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1766 // Store the integer parameter registers.
1767 SmallVector<SDValue, 8> MemOps;
1768 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1770 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1771 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1772 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1773 DAG.getIntPtrConstant(Offset));
1774 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1775 X86::GR64RegisterClass);
1776 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1778 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1779 PseudoSourceValue::getFixedStack(
1780 FuncInfo->getRegSaveFrameIndex()),
1781 Offset, false, false, 0);
1782 MemOps.push_back(Store);
1786 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1787 // Now store the XMM (fp + vector) parameter registers.
1788 SmallVector<SDValue, 11> SaveXMMOps;
1789 SaveXMMOps.push_back(Chain);
1791 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1792 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1793 SaveXMMOps.push_back(ALVal);
1795 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1796 FuncInfo->getRegSaveFrameIndex()));
1797 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1798 FuncInfo->getVarArgsFPOffset()));
1800 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1801 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1802 X86::VR128RegisterClass);
1803 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1804 SaveXMMOps.push_back(Val);
1806 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1808 &SaveXMMOps[0], SaveXMMOps.size()));
1811 if (!MemOps.empty())
1812 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1813 &MemOps[0], MemOps.size());
1817 // Some CCs need callee pop.
1818 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1819 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1821 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1822 // If this is an sret function, the return should pop the hidden pointer.
1823 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1824 FuncInfo->setBytesToPopOnReturn(4);
1828 // RegSaveFrameIndex is X86-64 only.
1829 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1830 if (CallConv == CallingConv::X86_FastCall ||
1831 CallConv == CallingConv::X86_ThisCall)
1832 // fastcc functions can't have varargs.
1833 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1840 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1841 SDValue StackPtr, SDValue Arg,
1842 DebugLoc dl, SelectionDAG &DAG,
1843 const CCValAssign &VA,
1844 ISD::ArgFlagsTy Flags) const {
1845 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1846 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1847 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1848 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1849 if (Flags.isByVal()) {
1850 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1852 return DAG.getStore(Chain, dl, Arg, PtrOff,
1853 PseudoSourceValue::getStack(), LocMemOffset,
1857 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1858 /// optimization is performed and it is required.
1860 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1861 SDValue &OutRetAddr, SDValue Chain,
1862 bool IsTailCall, bool Is64Bit,
1863 int FPDiff, DebugLoc dl) const {
1864 // Adjust the Return address stack slot.
1865 EVT VT = getPointerTy();
1866 OutRetAddr = getReturnAddressFrameIndex(DAG);
1868 // Load the "old" Return address.
1869 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1870 return SDValue(OutRetAddr.getNode(), 1);
1873 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1874 /// optimization is performed and it is required (FPDiff!=0).
1876 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1877 SDValue Chain, SDValue RetAddrFrIdx,
1878 bool Is64Bit, int FPDiff, DebugLoc dl) {
1879 // Store the return address to the appropriate stack slot.
1880 if (!FPDiff) return Chain;
1881 // Calculate the new stack slot for the return address.
1882 int SlotSize = Is64Bit ? 8 : 4;
1883 int NewReturnAddrFI =
1884 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1885 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1886 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1887 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1888 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1894 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1895 CallingConv::ID CallConv, bool isVarArg,
1897 const SmallVectorImpl<ISD::OutputArg> &Outs,
1898 const SmallVectorImpl<SDValue> &OutVals,
1899 const SmallVectorImpl<ISD::InputArg> &Ins,
1900 DebugLoc dl, SelectionDAG &DAG,
1901 SmallVectorImpl<SDValue> &InVals) const {
1902 MachineFunction &MF = DAG.getMachineFunction();
1903 bool Is64Bit = Subtarget->is64Bit();
1904 bool IsStructRet = CallIsStructReturn(Outs);
1905 bool IsSibcall = false;
1908 // Check if it's really possible to do a tail call.
1909 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1910 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1911 Outs, OutVals, Ins, DAG);
1913 // Sibcalls are automatically detected tailcalls which do not require
1915 if (!GuaranteedTailCallOpt && isTailCall)
1922 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1923 "Var args not supported with calling convention fastcc or ghc");
1925 // Analyze operands of the call, assigning locations to each operand.
1926 SmallVector<CCValAssign, 16> ArgLocs;
1927 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1928 ArgLocs, *DAG.getContext());
1929 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1931 // Get a count of how many bytes are to be pushed on the stack.
1932 unsigned NumBytes = CCInfo.getNextStackOffset();
1934 // This is a sibcall. The memory operands are available in caller's
1935 // own caller's stack.
1937 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1938 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1941 if (isTailCall && !IsSibcall) {
1942 // Lower arguments at fp - stackoffset + fpdiff.
1943 unsigned NumBytesCallerPushed =
1944 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1945 FPDiff = NumBytesCallerPushed - NumBytes;
1947 // Set the delta of movement of the returnaddr stackslot.
1948 // But only set if delta is greater than previous delta.
1949 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1950 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1954 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1956 SDValue RetAddrFrIdx;
1957 // Load return adress for tail calls.
1958 if (isTailCall && FPDiff)
1959 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1960 Is64Bit, FPDiff, dl);
1962 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1963 SmallVector<SDValue, 8> MemOpChains;
1966 // Walk the register/memloc assignments, inserting copies/loads. In the case
1967 // of tail call optimization arguments are handle later.
1968 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1969 CCValAssign &VA = ArgLocs[i];
1970 EVT RegVT = VA.getLocVT();
1971 SDValue Arg = OutVals[i];
1972 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1973 bool isByVal = Flags.isByVal();
1975 // Promote the value if needed.
1976 switch (VA.getLocInfo()) {
1977 default: llvm_unreachable("Unknown loc info!");
1978 case CCValAssign::Full: break;
1979 case CCValAssign::SExt:
1980 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1982 case CCValAssign::ZExt:
1983 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1985 case CCValAssign::AExt:
1986 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1987 // Special case: passing MMX values in XMM registers.
1988 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1989 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1990 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1992 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1994 case CCValAssign::BCvt:
1995 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1997 case CCValAssign::Indirect: {
1998 // Store the argument.
1999 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2000 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2001 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2002 PseudoSourceValue::getFixedStack(FI), 0,
2009 if (VA.isRegLoc()) {
2010 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2011 if (isVarArg && Subtarget->isTargetWin64()) {
2012 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2013 // shadow reg if callee is a varargs function.
2014 unsigned ShadowReg = 0;
2015 switch (VA.getLocReg()) {
2016 case X86::XMM0: ShadowReg = X86::RCX; break;
2017 case X86::XMM1: ShadowReg = X86::RDX; break;
2018 case X86::XMM2: ShadowReg = X86::R8; break;
2019 case X86::XMM3: ShadowReg = X86::R9; break;
2022 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2024 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2025 assert(VA.isMemLoc());
2026 if (StackPtr.getNode() == 0)
2027 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2028 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2029 dl, DAG, VA, Flags));
2033 if (!MemOpChains.empty())
2034 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2035 &MemOpChains[0], MemOpChains.size());
2037 // Build a sequence of copy-to-reg nodes chained together with token chain
2038 // and flag operands which copy the outgoing args into registers.
2040 // Tail call byval lowering might overwrite argument registers so in case of
2041 // tail call optimization the copies to registers are lowered later.
2043 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2044 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2045 RegsToPass[i].second, InFlag);
2046 InFlag = Chain.getValue(1);
2049 if (Subtarget->isPICStyleGOT()) {
2050 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2053 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2054 DAG.getNode(X86ISD::GlobalBaseReg,
2055 DebugLoc(), getPointerTy()),
2057 InFlag = Chain.getValue(1);
2059 // If we are tail calling and generating PIC/GOT style code load the
2060 // address of the callee into ECX. The value in ecx is used as target of
2061 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2062 // for tail calls on PIC/GOT architectures. Normally we would just put the
2063 // address of GOT into ebx and then call target@PLT. But for tail calls
2064 // ebx would be restored (since ebx is callee saved) before jumping to the
2067 // Note: The actual moving to ECX is done further down.
2068 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2069 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2070 !G->getGlobal()->hasProtectedVisibility())
2071 Callee = LowerGlobalAddress(Callee, DAG);
2072 else if (isa<ExternalSymbolSDNode>(Callee))
2073 Callee = LowerExternalSymbol(Callee, DAG);
2077 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
2078 // From AMD64 ABI document:
2079 // For calls that may call functions that use varargs or stdargs
2080 // (prototype-less calls or calls to functions containing ellipsis (...) in
2081 // the declaration) %al is used as hidden argument to specify the number
2082 // of SSE registers used. The contents of %al do not need to match exactly
2083 // the number of registers, but must be an ubound on the number of SSE
2084 // registers used and is in the range 0 - 8 inclusive.
2086 // Count the number of XMM registers allocated.
2087 static const unsigned XMMArgRegs[] = {
2088 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2089 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2091 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2092 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2093 && "SSE registers cannot be used when SSE is disabled");
2095 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2096 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2097 InFlag = Chain.getValue(1);
2101 // For tail calls lower the arguments to the 'real' stack slot.
2103 // Force all the incoming stack arguments to be loaded from the stack
2104 // before any new outgoing arguments are stored to the stack, because the
2105 // outgoing stack slots may alias the incoming argument stack slots, and
2106 // the alias isn't otherwise explicit. This is slightly more conservative
2107 // than necessary, because it means that each store effectively depends
2108 // on every argument instead of just those arguments it would clobber.
2109 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2111 SmallVector<SDValue, 8> MemOpChains2;
2114 // Do not flag preceeding copytoreg stuff together with the following stuff.
2116 if (GuaranteedTailCallOpt) {
2117 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2118 CCValAssign &VA = ArgLocs[i];
2121 assert(VA.isMemLoc());
2122 SDValue Arg = OutVals[i];
2123 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2124 // Create frame index.
2125 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2126 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2127 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2128 FIN = DAG.getFrameIndex(FI, getPointerTy());
2130 if (Flags.isByVal()) {
2131 // Copy relative to framepointer.
2132 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2133 if (StackPtr.getNode() == 0)
2134 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2136 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2138 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2142 // Store relative to framepointer.
2143 MemOpChains2.push_back(
2144 DAG.getStore(ArgChain, dl, Arg, FIN,
2145 PseudoSourceValue::getFixedStack(FI), 0,
2151 if (!MemOpChains2.empty())
2152 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2153 &MemOpChains2[0], MemOpChains2.size());
2155 // Copy arguments to their registers.
2156 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2157 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2158 RegsToPass[i].second, InFlag);
2159 InFlag = Chain.getValue(1);
2163 // Store the return address to the appropriate stack slot.
2164 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2168 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2169 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2170 // In the 64-bit large code model, we have to make all calls
2171 // through a register, since the call instruction's 32-bit
2172 // pc-relative offset may not be large enough to hold the whole
2174 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2175 // If the callee is a GlobalAddress node (quite common, every direct call
2176 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2179 // We should use extra load for direct calls to dllimported functions in
2181 const GlobalValue *GV = G->getGlobal();
2182 if (!GV->hasDLLImportLinkage()) {
2183 unsigned char OpFlags = 0;
2185 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2186 // external symbols most go through the PLT in PIC mode. If the symbol
2187 // has hidden or protected visibility, or if it is static or local, then
2188 // we don't need to use the PLT - we can directly call it.
2189 if (Subtarget->isTargetELF() &&
2190 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2191 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2192 OpFlags = X86II::MO_PLT;
2193 } else if (Subtarget->isPICStyleStubAny() &&
2194 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2195 Subtarget->getDarwinVers() < 9) {
2196 // PC-relative references to external symbols should go through $stub,
2197 // unless we're building with the leopard linker or later, which
2198 // automatically synthesizes these stubs.
2199 OpFlags = X86II::MO_DARWIN_STUB;
2202 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2203 G->getOffset(), OpFlags);
2205 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2206 unsigned char OpFlags = 0;
2208 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2209 // symbols should go through the PLT.
2210 if (Subtarget->isTargetELF() &&
2211 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2212 OpFlags = X86II::MO_PLT;
2213 } else if (Subtarget->isPICStyleStubAny() &&
2214 Subtarget->getDarwinVers() < 9) {
2215 // PC-relative references to external symbols should go through $stub,
2216 // unless we're building with the leopard linker or later, which
2217 // automatically synthesizes these stubs.
2218 OpFlags = X86II::MO_DARWIN_STUB;
2221 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2225 // Returns a chain & a flag for retval copy to use.
2226 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2227 SmallVector<SDValue, 8> Ops;
2229 if (!IsSibcall && isTailCall) {
2230 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2231 DAG.getIntPtrConstant(0, true), InFlag);
2232 InFlag = Chain.getValue(1);
2235 Ops.push_back(Chain);
2236 Ops.push_back(Callee);
2239 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2241 // Add argument registers to the end of the list so that they are known live
2243 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2244 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2245 RegsToPass[i].second.getValueType()));
2247 // Add an implicit use GOT pointer in EBX.
2248 if (!isTailCall && Subtarget->isPICStyleGOT())
2249 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2251 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2252 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64())
2253 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2255 if (InFlag.getNode())
2256 Ops.push_back(InFlag);
2260 //// If this is the first return lowered for this function, add the regs
2261 //// to the liveout set for the function.
2262 // This isn't right, although it's probably harmless on x86; liveouts
2263 // should be computed from returns not tail calls. Consider a void
2264 // function making a tail call to a function returning int.
2265 return DAG.getNode(X86ISD::TC_RETURN, dl,
2266 NodeTys, &Ops[0], Ops.size());
2269 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2270 InFlag = Chain.getValue(1);
2272 // Create the CALLSEQ_END node.
2273 unsigned NumBytesForCalleeToPush;
2274 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2275 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2276 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2277 // If this is a call to a struct-return function, the callee
2278 // pops the hidden struct pointer, so we have to push it back.
2279 // This is common for Darwin/X86, Linux & Mingw32 targets.
2280 NumBytesForCalleeToPush = 4;
2282 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2284 // Returns a flag for retval copy to use.
2286 Chain = DAG.getCALLSEQ_END(Chain,
2287 DAG.getIntPtrConstant(NumBytes, true),
2288 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2291 InFlag = Chain.getValue(1);
2294 // Handle result values, copying them out of physregs into vregs that we
2296 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2297 Ins, dl, DAG, InVals);
2301 //===----------------------------------------------------------------------===//
2302 // Fast Calling Convention (tail call) implementation
2303 //===----------------------------------------------------------------------===//
2305 // Like std call, callee cleans arguments, convention except that ECX is
2306 // reserved for storing the tail called function address. Only 2 registers are
2307 // free for argument passing (inreg). Tail call optimization is performed
2309 // * tailcallopt is enabled
2310 // * caller/callee are fastcc
2311 // On X86_64 architecture with GOT-style position independent code only local
2312 // (within module) calls are supported at the moment.
2313 // To keep the stack aligned according to platform abi the function
2314 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2315 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2316 // If a tail called function callee has more arguments than the caller the
2317 // caller needs to make sure that there is room to move the RETADDR to. This is
2318 // achieved by reserving an area the size of the argument delta right after the
2319 // original REtADDR, but before the saved framepointer or the spilled registers
2320 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2332 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2333 /// for a 16 byte align requirement.
2335 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2336 SelectionDAG& DAG) const {
2337 MachineFunction &MF = DAG.getMachineFunction();
2338 const TargetMachine &TM = MF.getTarget();
2339 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2340 unsigned StackAlignment = TFI.getStackAlignment();
2341 uint64_t AlignMask = StackAlignment - 1;
2342 int64_t Offset = StackSize;
2343 uint64_t SlotSize = TD->getPointerSize();
2344 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2345 // Number smaller than 12 so just add the difference.
2346 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2348 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2349 Offset = ((~AlignMask) & Offset) + StackAlignment +
2350 (StackAlignment-SlotSize);
2355 /// MatchingStackOffset - Return true if the given stack call argument is
2356 /// already available in the same position (relatively) of the caller's
2357 /// incoming argument stack.
2359 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2360 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2361 const X86InstrInfo *TII) {
2362 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2364 if (Arg.getOpcode() == ISD::CopyFromReg) {
2365 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2366 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2368 MachineInstr *Def = MRI->getVRegDef(VR);
2371 if (!Flags.isByVal()) {
2372 if (!TII->isLoadFromStackSlot(Def, FI))
2375 unsigned Opcode = Def->getOpcode();
2376 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2377 Def->getOperand(1).isFI()) {
2378 FI = Def->getOperand(1).getIndex();
2379 Bytes = Flags.getByValSize();
2383 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2384 if (Flags.isByVal())
2385 // ByVal argument is passed in as a pointer but it's now being
2386 // dereferenced. e.g.
2387 // define @foo(%struct.X* %A) {
2388 // tail call @bar(%struct.X* byval %A)
2391 SDValue Ptr = Ld->getBasePtr();
2392 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2395 FI = FINode->getIndex();
2399 assert(FI != INT_MAX);
2400 if (!MFI->isFixedObjectIndex(FI))
2402 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2405 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2406 /// for tail call optimization. Targets which want to do tail call
2407 /// optimization should implement this function.
2409 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2410 CallingConv::ID CalleeCC,
2412 bool isCalleeStructRet,
2413 bool isCallerStructRet,
2414 const SmallVectorImpl<ISD::OutputArg> &Outs,
2415 const SmallVectorImpl<SDValue> &OutVals,
2416 const SmallVectorImpl<ISD::InputArg> &Ins,
2417 SelectionDAG& DAG) const {
2418 if (!IsTailCallConvention(CalleeCC) &&
2419 CalleeCC != CallingConv::C)
2422 // If -tailcallopt is specified, make fastcc functions tail-callable.
2423 const MachineFunction &MF = DAG.getMachineFunction();
2424 const Function *CallerF = DAG.getMachineFunction().getFunction();
2425 CallingConv::ID CallerCC = CallerF->getCallingConv();
2426 bool CCMatch = CallerCC == CalleeCC;
2428 if (GuaranteedTailCallOpt) {
2429 if (IsTailCallConvention(CalleeCC) && CCMatch)
2434 // Look for obvious safe cases to perform tail call optimization that do not
2435 // require ABI changes. This is what gcc calls sibcall.
2437 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2438 // emit a special epilogue.
2439 if (RegInfo->needsStackRealignment(MF))
2442 // Do not sibcall optimize vararg calls unless the call site is not passing
2444 if (isVarArg && !Outs.empty())
2447 // Also avoid sibcall optimization if either caller or callee uses struct
2448 // return semantics.
2449 if (isCalleeStructRet || isCallerStructRet)
2452 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2453 // Therefore if it's not used by the call it is not safe to optimize this into
2455 bool Unused = false;
2456 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2463 SmallVector<CCValAssign, 16> RVLocs;
2464 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2465 RVLocs, *DAG.getContext());
2466 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2467 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2468 CCValAssign &VA = RVLocs[i];
2469 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2474 // If the calling conventions do not match, then we'd better make sure the
2475 // results are returned in the same way as what the caller expects.
2477 SmallVector<CCValAssign, 16> RVLocs1;
2478 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2479 RVLocs1, *DAG.getContext());
2480 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2482 SmallVector<CCValAssign, 16> RVLocs2;
2483 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2484 RVLocs2, *DAG.getContext());
2485 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2487 if (RVLocs1.size() != RVLocs2.size())
2489 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2490 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2492 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2494 if (RVLocs1[i].isRegLoc()) {
2495 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2498 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2504 // If the callee takes no arguments then go on to check the results of the
2506 if (!Outs.empty()) {
2507 // Check if stack adjustment is needed. For now, do not do this if any
2508 // argument is passed on the stack.
2509 SmallVector<CCValAssign, 16> ArgLocs;
2510 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2511 ArgLocs, *DAG.getContext());
2512 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2513 if (CCInfo.getNextStackOffset()) {
2514 MachineFunction &MF = DAG.getMachineFunction();
2515 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2517 if (Subtarget->isTargetWin64())
2518 // Win64 ABI has additional complications.
2521 // Check if the arguments are already laid out in the right way as
2522 // the caller's fixed stack objects.
2523 MachineFrameInfo *MFI = MF.getFrameInfo();
2524 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2525 const X86InstrInfo *TII =
2526 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2527 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2528 CCValAssign &VA = ArgLocs[i];
2529 SDValue Arg = OutVals[i];
2530 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2531 if (VA.getLocInfo() == CCValAssign::Indirect)
2533 if (!VA.isRegLoc()) {
2534 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2541 // If the tailcall address may be in a register, then make sure it's
2542 // possible to register allocate for it. In 32-bit, the call address can
2543 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2544 // callee-saved registers are restored. These happen to be the same
2545 // registers used to pass 'inreg' arguments so watch out for those.
2546 if (!Subtarget->is64Bit() &&
2547 !isa<GlobalAddressSDNode>(Callee) &&
2548 !isa<ExternalSymbolSDNode>(Callee)) {
2549 unsigned NumInRegs = 0;
2550 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2551 CCValAssign &VA = ArgLocs[i];
2554 unsigned Reg = VA.getLocReg();
2557 case X86::EAX: case X86::EDX: case X86::ECX:
2558 if (++NumInRegs == 3)
2570 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2571 return X86::createFastISel(funcInfo);
2575 //===----------------------------------------------------------------------===//
2576 // Other Lowering Hooks
2577 //===----------------------------------------------------------------------===//
2579 static bool MayFoldLoad(SDValue Op) {
2580 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2583 static bool MayFoldIntoStore(SDValue Op) {
2584 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2587 static bool isTargetShuffle(unsigned Opcode) {
2589 default: return false;
2590 case X86ISD::PSHUFD:
2591 case X86ISD::PSHUFHW:
2592 case X86ISD::PSHUFLW:
2593 case X86ISD::SHUFPD:
2594 case X86ISD::SHUFPS:
2595 case X86ISD::MOVLHPS:
2596 case X86ISD::MOVLHPD:
2597 case X86ISD::MOVHLPS:
2598 case X86ISD::MOVLPS:
2599 case X86ISD::MOVLPD:
2600 case X86ISD::MOVSHDUP:
2601 case X86ISD::MOVSLDUP:
2604 case X86ISD::UNPCKLPS:
2605 case X86ISD::PUNPCKLWD:
2606 case X86ISD::PUNPCKLBW:
2607 case X86ISD::PUNPCKLDQ:
2608 case X86ISD::UNPCKHPS:
2609 case X86ISD::PUNPCKHWD:
2610 case X86ISD::PUNPCKHBW:
2611 case X86ISD::PUNPCKHDQ:
2617 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2618 SDValue V1, SelectionDAG &DAG) {
2620 default: llvm_unreachable("Unknown x86 shuffle node");
2621 case X86ISD::MOVSHDUP:
2622 case X86ISD::MOVSLDUP:
2623 return DAG.getNode(Opc, dl, VT, V1);
2629 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2630 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2632 default: llvm_unreachable("Unknown x86 shuffle node");
2633 case X86ISD::PSHUFD:
2634 case X86ISD::PSHUFHW:
2635 case X86ISD::PSHUFLW:
2636 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2642 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2643 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2645 default: llvm_unreachable("Unknown x86 shuffle node");
2646 case X86ISD::SHUFPD:
2647 case X86ISD::SHUFPS:
2648 return DAG.getNode(Opc, dl, VT, V1, V2,
2649 DAG.getConstant(TargetMask, MVT::i8));
2654 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2655 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2657 default: llvm_unreachable("Unknown x86 shuffle node");
2658 case X86ISD::MOVLHPS:
2659 case X86ISD::MOVLHPD:
2660 case X86ISD::MOVHLPS:
2661 case X86ISD::MOVLPS:
2662 case X86ISD::MOVLPD:
2665 case X86ISD::UNPCKLPS:
2666 case X86ISD::PUNPCKLWD:
2667 case X86ISD::PUNPCKLBW:
2668 case X86ISD::PUNPCKLDQ:
2669 case X86ISD::UNPCKHPS:
2670 case X86ISD::PUNPCKHWD:
2671 case X86ISD::PUNPCKHBW:
2672 case X86ISD::PUNPCKHDQ:
2673 return DAG.getNode(Opc, dl, VT, V1, V2);
2678 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2679 MachineFunction &MF = DAG.getMachineFunction();
2680 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2681 int ReturnAddrIndex = FuncInfo->getRAIndex();
2683 if (ReturnAddrIndex == 0) {
2684 // Set up a frame object for the return address.
2685 uint64_t SlotSize = TD->getPointerSize();
2686 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2688 FuncInfo->setRAIndex(ReturnAddrIndex);
2691 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2695 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2696 bool hasSymbolicDisplacement) {
2697 // Offset should fit into 32 bit immediate field.
2698 if (!isInt<32>(Offset))
2701 // If we don't have a symbolic displacement - we don't have any extra
2703 if (!hasSymbolicDisplacement)
2706 // FIXME: Some tweaks might be needed for medium code model.
2707 if (M != CodeModel::Small && M != CodeModel::Kernel)
2710 // For small code model we assume that latest object is 16MB before end of 31
2711 // bits boundary. We may also accept pretty large negative constants knowing
2712 // that all objects are in the positive half of address space.
2713 if (M == CodeModel::Small && Offset < 16*1024*1024)
2716 // For kernel code model we know that all object resist in the negative half
2717 // of 32bits address space. We may not accept negative offsets, since they may
2718 // be just off and we may accept pretty large positive ones.
2719 if (M == CodeModel::Kernel && Offset > 0)
2725 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2726 /// specific condition code, returning the condition code and the LHS/RHS of the
2727 /// comparison to make.
2728 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2729 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2731 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2732 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2733 // X > -1 -> X == 0, jump !sign.
2734 RHS = DAG.getConstant(0, RHS.getValueType());
2735 return X86::COND_NS;
2736 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2737 // X < 0 -> X == 0, jump on sign.
2739 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2741 RHS = DAG.getConstant(0, RHS.getValueType());
2742 return X86::COND_LE;
2746 switch (SetCCOpcode) {
2747 default: llvm_unreachable("Invalid integer condition!");
2748 case ISD::SETEQ: return X86::COND_E;
2749 case ISD::SETGT: return X86::COND_G;
2750 case ISD::SETGE: return X86::COND_GE;
2751 case ISD::SETLT: return X86::COND_L;
2752 case ISD::SETLE: return X86::COND_LE;
2753 case ISD::SETNE: return X86::COND_NE;
2754 case ISD::SETULT: return X86::COND_B;
2755 case ISD::SETUGT: return X86::COND_A;
2756 case ISD::SETULE: return X86::COND_BE;
2757 case ISD::SETUGE: return X86::COND_AE;
2761 // First determine if it is required or is profitable to flip the operands.
2763 // If LHS is a foldable load, but RHS is not, flip the condition.
2764 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2765 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2766 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2767 std::swap(LHS, RHS);
2770 switch (SetCCOpcode) {
2776 std::swap(LHS, RHS);
2780 // On a floating point condition, the flags are set as follows:
2782 // 0 | 0 | 0 | X > Y
2783 // 0 | 0 | 1 | X < Y
2784 // 1 | 0 | 0 | X == Y
2785 // 1 | 1 | 1 | unordered
2786 switch (SetCCOpcode) {
2787 default: llvm_unreachable("Condcode should be pre-legalized away");
2789 case ISD::SETEQ: return X86::COND_E;
2790 case ISD::SETOLT: // flipped
2792 case ISD::SETGT: return X86::COND_A;
2793 case ISD::SETOLE: // flipped
2795 case ISD::SETGE: return X86::COND_AE;
2796 case ISD::SETUGT: // flipped
2798 case ISD::SETLT: return X86::COND_B;
2799 case ISD::SETUGE: // flipped
2801 case ISD::SETLE: return X86::COND_BE;
2803 case ISD::SETNE: return X86::COND_NE;
2804 case ISD::SETUO: return X86::COND_P;
2805 case ISD::SETO: return X86::COND_NP;
2807 case ISD::SETUNE: return X86::COND_INVALID;
2811 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2812 /// code. Current x86 isa includes the following FP cmov instructions:
2813 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2814 static bool hasFPCMov(unsigned X86CC) {
2830 /// isFPImmLegal - Returns true if the target can instruction select the
2831 /// specified FP immediate natively. If false, the legalizer will
2832 /// materialize the FP immediate as a load from a constant pool.
2833 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2834 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2835 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2841 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2842 /// the specified range (L, H].
2843 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2844 return (Val < 0) || (Val >= Low && Val < Hi);
2847 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2848 /// specified value.
2849 static bool isUndefOrEqual(int Val, int CmpVal) {
2850 if (Val < 0 || Val == CmpVal)
2855 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2856 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2857 /// the second operand.
2858 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2859 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2860 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2861 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2862 return (Mask[0] < 2 && Mask[1] < 2);
2866 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2867 SmallVector<int, 8> M;
2869 return ::isPSHUFDMask(M, N->getValueType(0));
2872 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2873 /// is suitable for input to PSHUFHW.
2874 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2875 if (VT != MVT::v8i16)
2878 // Lower quadword copied in order or undef.
2879 for (int i = 0; i != 4; ++i)
2880 if (Mask[i] >= 0 && Mask[i] != i)
2883 // Upper quadword shuffled.
2884 for (int i = 4; i != 8; ++i)
2885 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2891 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2892 SmallVector<int, 8> M;
2894 return ::isPSHUFHWMask(M, N->getValueType(0));
2897 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2898 /// is suitable for input to PSHUFLW.
2899 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2900 if (VT != MVT::v8i16)
2903 // Upper quadword copied in order.
2904 for (int i = 4; i != 8; ++i)
2905 if (Mask[i] >= 0 && Mask[i] != i)
2908 // Lower quadword shuffled.
2909 for (int i = 0; i != 4; ++i)
2916 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2917 SmallVector<int, 8> M;
2919 return ::isPSHUFLWMask(M, N->getValueType(0));
2922 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2923 /// is suitable for input to PALIGNR.
2924 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2926 int i, e = VT.getVectorNumElements();
2928 // Do not handle v2i64 / v2f64 shuffles with palignr.
2929 if (e < 4 || !hasSSSE3)
2932 for (i = 0; i != e; ++i)
2936 // All undef, not a palignr.
2940 // Determine if it's ok to perform a palignr with only the LHS, since we
2941 // don't have access to the actual shuffle elements to see if RHS is undef.
2942 bool Unary = Mask[i] < (int)e;
2943 bool NeedsUnary = false;
2945 int s = Mask[i] - i;
2947 // Check the rest of the elements to see if they are consecutive.
2948 for (++i; i != e; ++i) {
2953 Unary = Unary && (m < (int)e);
2954 NeedsUnary = NeedsUnary || (m < s);
2956 if (NeedsUnary && !Unary)
2958 if (Unary && m != ((s+i) & (e-1)))
2960 if (!Unary && m != (s+i))
2966 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2967 SmallVector<int, 8> M;
2969 return ::isPALIGNRMask(M, N->getValueType(0), true);
2972 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2973 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2974 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2975 int NumElems = VT.getVectorNumElements();
2976 if (NumElems != 2 && NumElems != 4)
2979 int Half = NumElems / 2;
2980 for (int i = 0; i < Half; ++i)
2981 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2983 for (int i = Half; i < NumElems; ++i)
2984 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2990 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2991 SmallVector<int, 8> M;
2993 return ::isSHUFPMask(M, N->getValueType(0));
2996 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2997 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2998 /// half elements to come from vector 1 (which would equal the dest.) and
2999 /// the upper half to come from vector 2.
3000 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3001 int NumElems = VT.getVectorNumElements();
3003 if (NumElems != 2 && NumElems != 4)
3006 int Half = NumElems / 2;
3007 for (int i = 0; i < Half; ++i)
3008 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3010 for (int i = Half; i < NumElems; ++i)
3011 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3016 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3017 SmallVector<int, 8> M;
3019 return isCommutedSHUFPMask(M, N->getValueType(0));
3022 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3023 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3024 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3025 if (N->getValueType(0).getVectorNumElements() != 4)
3028 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3029 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3030 isUndefOrEqual(N->getMaskElt(1), 7) &&
3031 isUndefOrEqual(N->getMaskElt(2), 2) &&
3032 isUndefOrEqual(N->getMaskElt(3), 3);
3035 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3036 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3038 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3039 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3044 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3045 isUndefOrEqual(N->getMaskElt(1), 3) &&
3046 isUndefOrEqual(N->getMaskElt(2), 2) &&
3047 isUndefOrEqual(N->getMaskElt(3), 3);
3050 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3051 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3052 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3053 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3055 if (NumElems != 2 && NumElems != 4)
3058 for (unsigned i = 0; i < NumElems/2; ++i)
3059 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3062 for (unsigned i = NumElems/2; i < NumElems; ++i)
3063 if (!isUndefOrEqual(N->getMaskElt(i), i))
3069 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3070 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3071 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3072 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3074 if (NumElems != 2 && NumElems != 4)
3077 for (unsigned i = 0; i < NumElems/2; ++i)
3078 if (!isUndefOrEqual(N->getMaskElt(i), i))
3081 for (unsigned i = 0; i < NumElems/2; ++i)
3082 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3088 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3089 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3090 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3091 bool V2IsSplat = false) {
3092 int NumElts = VT.getVectorNumElements();
3093 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3096 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3098 int BitI1 = Mask[i+1];
3099 if (!isUndefOrEqual(BitI, j))
3102 if (!isUndefOrEqual(BitI1, NumElts))
3105 if (!isUndefOrEqual(BitI1, j + NumElts))
3112 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3113 SmallVector<int, 8> M;
3115 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3118 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3119 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3120 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3121 bool V2IsSplat = false) {
3122 int NumElts = VT.getVectorNumElements();
3123 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3126 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3128 int BitI1 = Mask[i+1];
3129 if (!isUndefOrEqual(BitI, j + NumElts/2))
3132 if (isUndefOrEqual(BitI1, NumElts))
3135 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3142 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3143 SmallVector<int, 8> M;
3145 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3148 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3149 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3151 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3152 int NumElems = VT.getVectorNumElements();
3153 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3156 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3158 int BitI1 = Mask[i+1];
3159 if (!isUndefOrEqual(BitI, j))
3161 if (!isUndefOrEqual(BitI1, j))
3167 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3168 SmallVector<int, 8> M;
3170 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3173 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3174 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3176 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3177 int NumElems = VT.getVectorNumElements();
3178 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3181 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3183 int BitI1 = Mask[i+1];
3184 if (!isUndefOrEqual(BitI, j))
3186 if (!isUndefOrEqual(BitI1, j))
3192 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3193 SmallVector<int, 8> M;
3195 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3198 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3199 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3200 /// MOVSD, and MOVD, i.e. setting the lowest element.
3201 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3202 if (VT.getVectorElementType().getSizeInBits() < 32)
3205 int NumElts = VT.getVectorNumElements();
3207 if (!isUndefOrEqual(Mask[0], NumElts))
3210 for (int i = 1; i < NumElts; ++i)
3211 if (!isUndefOrEqual(Mask[i], i))
3217 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3218 SmallVector<int, 8> M;
3220 return ::isMOVLMask(M, N->getValueType(0));
3223 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3224 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3225 /// element of vector 2 and the other elements to come from vector 1 in order.
3226 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3227 bool V2IsSplat = false, bool V2IsUndef = false) {
3228 int NumOps = VT.getVectorNumElements();
3229 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3232 if (!isUndefOrEqual(Mask[0], 0))
3235 for (int i = 1; i < NumOps; ++i)
3236 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3237 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3238 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3244 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3245 bool V2IsUndef = false) {
3246 SmallVector<int, 8> M;
3248 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3251 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3252 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3253 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3254 if (N->getValueType(0).getVectorNumElements() != 4)
3257 // Expect 1, 1, 3, 3
3258 for (unsigned i = 0; i < 2; ++i) {
3259 int Elt = N->getMaskElt(i);
3260 if (Elt >= 0 && Elt != 1)
3265 for (unsigned i = 2; i < 4; ++i) {
3266 int Elt = N->getMaskElt(i);
3267 if (Elt >= 0 && Elt != 3)
3272 // Don't use movshdup if it can be done with a shufps.
3273 // FIXME: verify that matching u, u, 3, 3 is what we want.
3277 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3278 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3279 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3280 if (N->getValueType(0).getVectorNumElements() != 4)
3283 // Expect 0, 0, 2, 2
3284 for (unsigned i = 0; i < 2; ++i)
3285 if (N->getMaskElt(i) > 0)
3289 for (unsigned i = 2; i < 4; ++i) {
3290 int Elt = N->getMaskElt(i);
3291 if (Elt >= 0 && Elt != 2)
3296 // Don't use movsldup if it can be done with a shufps.
3300 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3301 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3302 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3303 int e = N->getValueType(0).getVectorNumElements() / 2;
3305 for (int i = 0; i < e; ++i)
3306 if (!isUndefOrEqual(N->getMaskElt(i), i))
3308 for (int i = 0; i < e; ++i)
3309 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3314 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3315 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3316 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3317 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3318 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3320 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3322 for (int i = 0; i < NumOperands; ++i) {
3323 int Val = SVOp->getMaskElt(NumOperands-i-1);
3324 if (Val < 0) Val = 0;
3325 if (Val >= NumOperands) Val -= NumOperands;
3327 if (i != NumOperands - 1)
3333 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3334 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3335 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3336 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3338 // 8 nodes, but we only care about the last 4.
3339 for (unsigned i = 7; i >= 4; --i) {
3340 int Val = SVOp->getMaskElt(i);
3349 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3350 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3351 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3352 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3354 // 8 nodes, but we only care about the first 4.
3355 for (int i = 3; i >= 0; --i) {
3356 int Val = SVOp->getMaskElt(i);
3365 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3366 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3367 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3368 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3369 EVT VVT = N->getValueType(0);
3370 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3374 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3375 Val = SVOp->getMaskElt(i);
3379 return (Val - i) * EltSize;
3382 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3384 bool X86::isZeroNode(SDValue Elt) {
3385 return ((isa<ConstantSDNode>(Elt) &&
3386 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3387 (isa<ConstantFPSDNode>(Elt) &&
3388 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3391 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3392 /// their permute mask.
3393 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3394 SelectionDAG &DAG) {
3395 EVT VT = SVOp->getValueType(0);
3396 unsigned NumElems = VT.getVectorNumElements();
3397 SmallVector<int, 8> MaskVec;
3399 for (unsigned i = 0; i != NumElems; ++i) {
3400 int idx = SVOp->getMaskElt(i);
3402 MaskVec.push_back(idx);
3403 else if (idx < (int)NumElems)
3404 MaskVec.push_back(idx + NumElems);
3406 MaskVec.push_back(idx - NumElems);
3408 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3409 SVOp->getOperand(0), &MaskVec[0]);
3412 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3413 /// the two vector operands have swapped position.
3414 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3415 unsigned NumElems = VT.getVectorNumElements();
3416 for (unsigned i = 0; i != NumElems; ++i) {
3420 else if (idx < (int)NumElems)
3421 Mask[i] = idx + NumElems;
3423 Mask[i] = idx - NumElems;
3427 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3428 /// match movhlps. The lower half elements should come from upper half of
3429 /// V1 (and in order), and the upper half elements should come from the upper
3430 /// half of V2 (and in order).
3431 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3432 if (Op->getValueType(0).getVectorNumElements() != 4)
3434 for (unsigned i = 0, e = 2; i != e; ++i)
3435 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3437 for (unsigned i = 2; i != 4; ++i)
3438 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3443 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3444 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3446 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3447 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3449 N = N->getOperand(0).getNode();
3450 if (!ISD::isNON_EXTLoad(N))
3453 *LD = cast<LoadSDNode>(N);
3457 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3458 /// match movlp{s|d}. The lower half elements should come from lower half of
3459 /// V1 (and in order), and the upper half elements should come from the upper
3460 /// half of V2 (and in order). And since V1 will become the source of the
3461 /// MOVLP, it must be either a vector load or a scalar load to vector.
3462 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3463 ShuffleVectorSDNode *Op) {
3464 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3466 // Is V2 is a vector load, don't do this transformation. We will try to use
3467 // load folding shufps op.
3468 if (ISD::isNON_EXTLoad(V2))
3471 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3473 if (NumElems != 2 && NumElems != 4)
3475 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3476 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3478 for (unsigned i = NumElems/2; i != NumElems; ++i)
3479 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3484 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3486 static bool isSplatVector(SDNode *N) {
3487 if (N->getOpcode() != ISD::BUILD_VECTOR)
3490 SDValue SplatValue = N->getOperand(0);
3491 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3492 if (N->getOperand(i) != SplatValue)
3497 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3498 /// to an zero vector.
3499 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3500 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3501 SDValue V1 = N->getOperand(0);
3502 SDValue V2 = N->getOperand(1);
3503 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3504 for (unsigned i = 0; i != NumElems; ++i) {
3505 int Idx = N->getMaskElt(i);
3506 if (Idx >= (int)NumElems) {
3507 unsigned Opc = V2.getOpcode();
3508 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3510 if (Opc != ISD::BUILD_VECTOR ||
3511 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3513 } else if (Idx >= 0) {
3514 unsigned Opc = V1.getOpcode();
3515 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3517 if (Opc != ISD::BUILD_VECTOR ||
3518 !X86::isZeroNode(V1.getOperand(Idx)))
3525 /// getZeroVector - Returns a vector of specified type with all zero elements.
3527 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3529 assert(VT.isVector() && "Expected a vector type");
3531 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted
3532 // to their dest type. This ensures they get CSE'd.
3534 if (VT.getSizeInBits() == 64) { // MMX
3535 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3536 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3537 } else if (VT.getSizeInBits() == 128) {
3538 if (HasSSE2) { // SSE2
3539 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3540 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3542 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3543 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3545 } else if (VT.getSizeInBits() == 256) { // AVX
3546 // 256-bit logic and arithmetic instructions in AVX are
3547 // all floating-point, no support for integer ops. Default
3548 // to emitting fp zeroed vectors then.
3549 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3550 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3551 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3553 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3556 /// getOnesVector - Returns a vector of specified type with all bits set.
3558 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3559 assert(VT.isVector() && "Expected a vector type");
3561 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3562 // type. This ensures they get CSE'd.
3563 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3565 if (VT.getSizeInBits() == 64) // MMX
3566 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3568 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3569 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3573 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3574 /// that point to V2 points to its first element.
3575 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3576 EVT VT = SVOp->getValueType(0);
3577 unsigned NumElems = VT.getVectorNumElements();
3579 bool Changed = false;
3580 SmallVector<int, 8> MaskVec;
3581 SVOp->getMask(MaskVec);
3583 for (unsigned i = 0; i != NumElems; ++i) {
3584 if (MaskVec[i] > (int)NumElems) {
3585 MaskVec[i] = NumElems;
3590 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3591 SVOp->getOperand(1), &MaskVec[0]);
3592 return SDValue(SVOp, 0);
3595 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3596 /// operation of specified width.
3597 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3599 unsigned NumElems = VT.getVectorNumElements();
3600 SmallVector<int, 8> Mask;
3601 Mask.push_back(NumElems);
3602 for (unsigned i = 1; i != NumElems; ++i)
3604 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3607 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3608 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3610 unsigned NumElems = VT.getVectorNumElements();
3611 SmallVector<int, 8> Mask;
3612 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3614 Mask.push_back(i + NumElems);
3616 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3619 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3620 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3622 unsigned NumElems = VT.getVectorNumElements();
3623 unsigned Half = NumElems/2;
3624 SmallVector<int, 8> Mask;
3625 for (unsigned i = 0; i != Half; ++i) {
3626 Mask.push_back(i + Half);
3627 Mask.push_back(i + NumElems + Half);
3629 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3632 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3633 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3634 if (SV->getValueType(0).getVectorNumElements() <= 4)
3635 return SDValue(SV, 0);
3637 EVT PVT = MVT::v4f32;
3638 EVT VT = SV->getValueType(0);
3639 DebugLoc dl = SV->getDebugLoc();
3640 SDValue V1 = SV->getOperand(0);
3641 int NumElems = VT.getVectorNumElements();
3642 int EltNo = SV->getSplatIndex();
3644 // unpack elements to the correct location
3645 while (NumElems > 4) {
3646 if (EltNo < NumElems/2) {
3647 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3649 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3650 EltNo -= NumElems/2;
3655 // Perform the splat.
3656 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3657 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3658 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3659 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3662 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3663 /// vector of zero or undef vector. This produces a shuffle where the low
3664 /// element of V2 is swizzled into the zero/undef vector, landing at element
3665 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3666 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3667 bool isZero, bool HasSSE2,
3668 SelectionDAG &DAG) {
3669 EVT VT = V2.getValueType();
3671 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3672 unsigned NumElems = VT.getVectorNumElements();
3673 SmallVector<int, 16> MaskVec;
3674 for (unsigned i = 0; i != NumElems; ++i)
3675 // If this is the insertion idx, put the low elt of V2 here.
3676 MaskVec.push_back(i == Idx ? NumElems : i);
3677 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3680 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
3681 /// element of the result of the vector shuffle.
3682 SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG) {
3683 SDValue V = SDValue(N, 0);
3684 EVT VT = V.getValueType();
3685 unsigned Opcode = V.getOpcode();
3687 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
3688 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
3689 Index = SV->getMaskElt(Index);
3692 return DAG.getUNDEF(VT.getVectorElementType());
3694 int NumElems = VT.getVectorNumElements();
3695 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
3696 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG);
3699 // Recurse into target specific vector shuffles to find scalars.
3700 if (isTargetShuffle(Opcode)) {
3703 case X86ISD::MOVSD: {
3704 // The index 0 always comes from the first element of the second source,
3705 // this is why MOVSS and MOVSD are used in the first place. The other
3706 // elements come from the other positions of the first source vector.
3707 unsigned OpNum = (Index == 0) ? 1 : 0;
3708 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG);
3711 assert("not implemented for target shuffle node");
3716 // Actual nodes that may contain scalar elements
3717 if (Opcode == ISD::BIT_CONVERT) {
3718 V = V.getOperand(0);
3719 EVT SrcVT = V.getValueType();
3720 unsigned NumElems = VT.getVectorNumElements();
3722 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
3726 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
3727 return (Index == 0) ? V.getOperand(0)
3728 : DAG.getUNDEF(VT.getVectorElementType());
3730 if (V.getOpcode() == ISD::BUILD_VECTOR)
3731 return V.getOperand(Index);
3736 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
3737 /// shuffle operation which come from a consecutively from a zero. The
3738 /// search can start in two diferent directions, from left or right.
3740 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
3741 bool ZerosFromLeft, SelectionDAG &DAG) {
3744 while (i < NumElems) {
3745 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
3746 SDValue Elt = getShuffleScalarElt(N, Index, DAG);
3747 if (!(Elt.getNode() &&
3748 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
3756 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
3757 /// MaskE correspond consecutively to elements from one of the vector operands,
3758 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
3760 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
3761 int OpIdx, int NumElems, unsigned &OpNum) {
3762 bool SeenV1 = false;
3763 bool SeenV2 = false;
3765 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
3766 int Idx = SVOp->getMaskElt(i);
3767 // Ignore undef indicies
3776 // Only accept consecutive elements from the same vector
3777 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
3781 OpNum = SeenV1 ? 0 : 1;
3785 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
3786 /// logical left shift of a vector.
3787 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3788 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3789 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3790 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3791 false /* check zeros from right */, DAG);
3797 // Considering the elements in the mask that are not consecutive zeros,
3798 // check if they consecutively come from only one of the source vectors.
3800 // V1 = {X, A, B, C} 0
3802 // vector_shuffle V1, V2 <1, 2, 3, X>
3804 if (!isShuffleMaskConsecutive(SVOp,
3805 0, // Mask Start Index
3806 NumElems-NumZeros-1, // Mask End Index
3807 NumZeros, // Where to start looking in the src vector
3808 NumElems, // Number of elements in vector
3809 OpSrc)) // Which source operand ?
3814 ShVal = SVOp->getOperand(OpSrc);
3818 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
3819 /// logical left shift of a vector.
3820 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3821 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3822 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3823 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3824 true /* check zeros from left */, DAG);
3830 // Considering the elements in the mask that are not consecutive zeros,
3831 // check if they consecutively come from only one of the source vectors.
3833 // 0 { A, B, X, X } = V2
3835 // vector_shuffle V1, V2 <X, X, 4, 5>
3837 if (!isShuffleMaskConsecutive(SVOp,
3838 NumZeros, // Mask Start Index
3839 NumElems-1, // Mask End Index
3840 0, // Where to start looking in the src vector
3841 NumElems, // Number of elements in vector
3842 OpSrc)) // Which source operand ?
3847 ShVal = SVOp->getOperand(OpSrc);
3851 /// isVectorShift - Returns true if the shuffle can be implemented as a
3852 /// logical left or right shift of a vector.
3853 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3854 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3855 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
3856 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
3862 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3864 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3865 unsigned NumNonZero, unsigned NumZero,
3867 const TargetLowering &TLI) {
3871 DebugLoc dl = Op.getDebugLoc();
3874 for (unsigned i = 0; i < 16; ++i) {
3875 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3876 if (ThisIsNonZero && First) {
3878 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3880 V = DAG.getUNDEF(MVT::v8i16);
3885 SDValue ThisElt(0, 0), LastElt(0, 0);
3886 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3887 if (LastIsNonZero) {
3888 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3889 MVT::i16, Op.getOperand(i-1));
3891 if (ThisIsNonZero) {
3892 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3893 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3894 ThisElt, DAG.getConstant(8, MVT::i8));
3896 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3900 if (ThisElt.getNode())
3901 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3902 DAG.getIntPtrConstant(i/2));
3906 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3909 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3911 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3912 unsigned NumNonZero, unsigned NumZero,
3914 const TargetLowering &TLI) {
3918 DebugLoc dl = Op.getDebugLoc();
3921 for (unsigned i = 0; i < 8; ++i) {
3922 bool isNonZero = (NonZeros & (1 << i)) != 0;
3926 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3928 V = DAG.getUNDEF(MVT::v8i16);
3931 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3932 MVT::v8i16, V, Op.getOperand(i),
3933 DAG.getIntPtrConstant(i));
3940 /// getVShift - Return a vector logical shift node.
3942 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3943 unsigned NumBits, SelectionDAG &DAG,
3944 const TargetLowering &TLI, DebugLoc dl) {
3945 bool isMMX = VT.getSizeInBits() == 64;
3946 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3947 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3948 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3949 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3950 DAG.getNode(Opc, dl, ShVT, SrcOp,
3951 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3955 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3956 SelectionDAG &DAG) const {
3958 // Check if the scalar load can be widened into a vector load. And if
3959 // the address is "base + cst" see if the cst can be "absorbed" into
3960 // the shuffle mask.
3961 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3962 SDValue Ptr = LD->getBasePtr();
3963 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3965 EVT PVT = LD->getValueType(0);
3966 if (PVT != MVT::i32 && PVT != MVT::f32)
3971 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3972 FI = FINode->getIndex();
3974 } else if (Ptr.getOpcode() == ISD::ADD &&
3975 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3976 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3977 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3978 Offset = Ptr.getConstantOperandVal(1);
3979 Ptr = Ptr.getOperand(0);
3984 SDValue Chain = LD->getChain();
3985 // Make sure the stack object alignment is at least 16.
3986 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3987 if (DAG.InferPtrAlignment(Ptr) < 16) {
3988 if (MFI->isFixedObjectIndex(FI)) {
3989 // Can't change the alignment. FIXME: It's possible to compute
3990 // the exact stack offset and reference FI + adjust offset instead.
3991 // If someone *really* cares about this. That's the way to implement it.
3994 MFI->setObjectAlignment(FI, 16);
3998 // (Offset % 16) must be multiple of 4. Then address is then
3999 // Ptr + (Offset & ~15).
4002 if ((Offset % 16) & 3)
4004 int64_t StartOffset = Offset & ~15;
4006 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4007 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4009 int EltNo = (Offset - StartOffset) >> 2;
4010 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4011 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4012 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
4014 // Canonicalize it to a v4i32 shuffle.
4015 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
4016 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4017 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4018 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
4024 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4025 /// vector of type 'VT', see if the elements can be replaced by a single large
4026 /// load which has the same value as a build_vector whose operands are 'elts'.
4028 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4030 /// FIXME: we'd also like to handle the case where the last elements are zero
4031 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4032 /// There's even a handy isZeroNode for that purpose.
4033 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4034 DebugLoc &dl, SelectionDAG &DAG) {
4035 EVT EltVT = VT.getVectorElementType();
4036 unsigned NumElems = Elts.size();
4038 LoadSDNode *LDBase = NULL;
4039 unsigned LastLoadedElt = -1U;
4041 // For each element in the initializer, see if we've found a load or an undef.
4042 // If we don't find an initial load element, or later load elements are
4043 // non-consecutive, bail out.
4044 for (unsigned i = 0; i < NumElems; ++i) {
4045 SDValue Elt = Elts[i];
4047 if (!Elt.getNode() ||
4048 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4051 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4053 LDBase = cast<LoadSDNode>(Elt.getNode());
4057 if (Elt.getOpcode() == ISD::UNDEF)
4060 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4061 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4066 // If we have found an entire vector of loads and undefs, then return a large
4067 // load of the entire vector width starting at the base pointer. If we found
4068 // consecutive loads for the low half, generate a vzext_load node.
4069 if (LastLoadedElt == NumElems - 1) {
4070 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4071 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
4072 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
4073 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4074 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
4075 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
4076 LDBase->isVolatile(), LDBase->isNonTemporal(),
4077 LDBase->getAlignment());
4078 } else if (NumElems == 4 && LastLoadedElt == 1) {
4079 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4080 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4081 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
4082 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
4088 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4089 DebugLoc dl = Op.getDebugLoc();
4090 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
4091 // All one's are handled with pcmpeqd. In AVX, zero's are handled with
4092 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
4093 // is present, so AllOnes is ignored.
4094 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4095 (Op.getValueType().getSizeInBits() != 256 &&
4096 ISD::isBuildVectorAllOnes(Op.getNode()))) {
4097 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
4098 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4099 // eliminated on x86-32 hosts.
4100 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
4103 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4104 return getOnesVector(Op.getValueType(), DAG, dl);
4105 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4108 EVT VT = Op.getValueType();
4109 EVT ExtVT = VT.getVectorElementType();
4110 unsigned EVTBits = ExtVT.getSizeInBits();
4112 unsigned NumElems = Op.getNumOperands();
4113 unsigned NumZero = 0;
4114 unsigned NumNonZero = 0;
4115 unsigned NonZeros = 0;
4116 bool IsAllConstants = true;
4117 SmallSet<SDValue, 8> Values;
4118 for (unsigned i = 0; i < NumElems; ++i) {
4119 SDValue Elt = Op.getOperand(i);
4120 if (Elt.getOpcode() == ISD::UNDEF)
4123 if (Elt.getOpcode() != ISD::Constant &&
4124 Elt.getOpcode() != ISD::ConstantFP)
4125 IsAllConstants = false;
4126 if (X86::isZeroNode(Elt))
4129 NonZeros |= (1 << i);
4134 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4135 if (NumNonZero == 0)
4136 return DAG.getUNDEF(VT);
4138 // Special case for single non-zero, non-undef, element.
4139 if (NumNonZero == 1) {
4140 unsigned Idx = CountTrailingZeros_32(NonZeros);
4141 SDValue Item = Op.getOperand(Idx);
4143 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4144 // the value are obviously zero, truncate the value to i32 and do the
4145 // insertion that way. Only do this if the value is non-constant or if the
4146 // value is a constant being inserted into element 0. It is cheaper to do
4147 // a constant pool load than it is to do a movd + shuffle.
4148 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4149 (!IsAllConstants || Idx == 0)) {
4150 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4151 // Handle MMX and SSE both.
4152 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
4153 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
4155 // Truncate the value (which may itself be a constant) to i32, and
4156 // convert it to a vector with movd (S2V+shuffle to zero extend).
4157 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4158 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4159 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4160 Subtarget->hasSSE2(), DAG);
4162 // Now we have our 32-bit value zero extended in the low element of
4163 // a vector. If Idx != 0, swizzle it into place.
4165 SmallVector<int, 4> Mask;
4166 Mask.push_back(Idx);
4167 for (unsigned i = 1; i != VecElts; ++i)
4169 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4170 DAG.getUNDEF(Item.getValueType()),
4173 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
4177 // If we have a constant or non-constant insertion into the low element of
4178 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4179 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4180 // depending on what the source datatype is.
4183 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4184 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4185 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4186 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4187 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4188 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4190 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4191 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4192 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
4193 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4194 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4195 Subtarget->hasSSE2(), DAG);
4196 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
4200 // Is it a vector logical left shift?
4201 if (NumElems == 2 && Idx == 1 &&
4202 X86::isZeroNode(Op.getOperand(0)) &&
4203 !X86::isZeroNode(Op.getOperand(1))) {
4204 unsigned NumBits = VT.getSizeInBits();
4205 return getVShift(true, VT,
4206 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4207 VT, Op.getOperand(1)),
4208 NumBits/2, DAG, *this, dl);
4211 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4214 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4215 // is a non-constant being inserted into an element other than the low one,
4216 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4217 // movd/movss) to move this into the low element, then shuffle it into
4219 if (EVTBits == 32) {
4220 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4222 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4223 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4224 Subtarget->hasSSE2(), DAG);
4225 SmallVector<int, 8> MaskVec;
4226 for (unsigned i = 0; i < NumElems; i++)
4227 MaskVec.push_back(i == Idx ? 0 : 1);
4228 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4232 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4233 if (Values.size() == 1) {
4234 if (EVTBits == 32) {
4235 // Instead of a shuffle like this:
4236 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4237 // Check if it's possible to issue this instead.
4238 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4239 unsigned Idx = CountTrailingZeros_32(NonZeros);
4240 SDValue Item = Op.getOperand(Idx);
4241 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4242 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4247 // A vector full of immediates; various special cases are already
4248 // handled, so this is best done with a single constant-pool load.
4252 // Let legalizer expand 2-wide build_vectors.
4253 if (EVTBits == 64) {
4254 if (NumNonZero == 1) {
4255 // One half is zero or undef.
4256 unsigned Idx = CountTrailingZeros_32(NonZeros);
4257 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4258 Op.getOperand(Idx));
4259 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4260 Subtarget->hasSSE2(), DAG);
4265 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4266 if (EVTBits == 8 && NumElems == 16) {
4267 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4269 if (V.getNode()) return V;
4272 if (EVTBits == 16 && NumElems == 8) {
4273 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4275 if (V.getNode()) return V;
4278 // If element VT is == 32 bits, turn it into a number of shuffles.
4279 SmallVector<SDValue, 8> V;
4281 if (NumElems == 4 && NumZero > 0) {
4282 for (unsigned i = 0; i < 4; ++i) {
4283 bool isZero = !(NonZeros & (1 << i));
4285 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4287 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4290 for (unsigned i = 0; i < 2; ++i) {
4291 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4294 V[i] = V[i*2]; // Must be a zero vector.
4297 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4300 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4303 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4308 SmallVector<int, 8> MaskVec;
4309 bool Reverse = (NonZeros & 0x3) == 2;
4310 for (unsigned i = 0; i < 2; ++i)
4311 MaskVec.push_back(Reverse ? 1-i : i);
4312 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4313 for (unsigned i = 0; i < 2; ++i)
4314 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4315 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4318 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4319 // Check for a build vector of consecutive loads.
4320 for (unsigned i = 0; i < NumElems; ++i)
4321 V[i] = Op.getOperand(i);
4323 // Check for elements which are consecutive loads.
4324 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4328 // For SSE 4.1, use insertps to put the high elements into the low element.
4329 if (getSubtarget()->hasSSE41()) {
4331 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4332 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4334 Result = DAG.getUNDEF(VT);
4336 for (unsigned i = 1; i < NumElems; ++i) {
4337 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4338 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4339 Op.getOperand(i), DAG.getIntPtrConstant(i));
4344 // Otherwise, expand into a number of unpckl*, start by extending each of
4345 // our (non-undef) elements to the full vector width with the element in the
4346 // bottom slot of the vector (which generates no code for SSE).
4347 for (unsigned i = 0; i < NumElems; ++i) {
4348 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4349 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4351 V[i] = DAG.getUNDEF(VT);
4354 // Next, we iteratively mix elements, e.g. for v4f32:
4355 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4356 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4357 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4358 unsigned EltStride = NumElems >> 1;
4359 while (EltStride != 0) {
4360 for (unsigned i = 0; i < EltStride; ++i) {
4361 // If V[i+EltStride] is undef and this is the first round of mixing,
4362 // then it is safe to just drop this shuffle: V[i] is already in the
4363 // right place, the one element (since it's the first round) being
4364 // inserted as undef can be dropped. This isn't safe for successive
4365 // rounds because they will permute elements within both vectors.
4366 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4367 EltStride == NumElems/2)
4370 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4380 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4381 // We support concatenate two MMX registers and place them in a MMX
4382 // register. This is better than doing a stack convert.
4383 DebugLoc dl = Op.getDebugLoc();
4384 EVT ResVT = Op.getValueType();
4385 assert(Op.getNumOperands() == 2);
4386 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4387 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4389 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4390 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4391 InVec = Op.getOperand(1);
4392 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4393 unsigned NumElts = ResVT.getVectorNumElements();
4394 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4395 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4396 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4398 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4399 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4400 Mask[0] = 0; Mask[1] = 2;
4401 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4403 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4406 // v8i16 shuffles - Prefer shuffles in the following order:
4407 // 1. [all] pshuflw, pshufhw, optional move
4408 // 2. [ssse3] 1 x pshufb
4409 // 3. [ssse3] 2 x pshufb + 1 x por
4410 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4412 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4413 SelectionDAG &DAG) const {
4414 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4415 SDValue V1 = SVOp->getOperand(0);
4416 SDValue V2 = SVOp->getOperand(1);
4417 DebugLoc dl = SVOp->getDebugLoc();
4418 SmallVector<int, 8> MaskVals;
4420 // Determine if more than 1 of the words in each of the low and high quadwords
4421 // of the result come from the same quadword of one of the two inputs. Undef
4422 // mask values count as coming from any quadword, for better codegen.
4423 SmallVector<unsigned, 4> LoQuad(4);
4424 SmallVector<unsigned, 4> HiQuad(4);
4425 BitVector InputQuads(4);
4426 for (unsigned i = 0; i < 8; ++i) {
4427 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4428 int EltIdx = SVOp->getMaskElt(i);
4429 MaskVals.push_back(EltIdx);
4438 InputQuads.set(EltIdx / 4);
4441 int BestLoQuad = -1;
4442 unsigned MaxQuad = 1;
4443 for (unsigned i = 0; i < 4; ++i) {
4444 if (LoQuad[i] > MaxQuad) {
4446 MaxQuad = LoQuad[i];
4450 int BestHiQuad = -1;
4452 for (unsigned i = 0; i < 4; ++i) {
4453 if (HiQuad[i] > MaxQuad) {
4455 MaxQuad = HiQuad[i];
4459 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4460 // of the two input vectors, shuffle them into one input vector so only a
4461 // single pshufb instruction is necessary. If There are more than 2 input
4462 // quads, disable the next transformation since it does not help SSSE3.
4463 bool V1Used = InputQuads[0] || InputQuads[1];
4464 bool V2Used = InputQuads[2] || InputQuads[3];
4465 if (Subtarget->hasSSSE3()) {
4466 if (InputQuads.count() == 2 && V1Used && V2Used) {
4467 BestLoQuad = InputQuads.find_first();
4468 BestHiQuad = InputQuads.find_next(BestLoQuad);
4470 if (InputQuads.count() > 2) {
4476 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4477 // the shuffle mask. If a quad is scored as -1, that means that it contains
4478 // words from all 4 input quadwords.
4480 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4481 SmallVector<int, 8> MaskV;
4482 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4483 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4484 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4485 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4486 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4487 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4489 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4490 // source words for the shuffle, to aid later transformations.
4491 bool AllWordsInNewV = true;
4492 bool InOrder[2] = { true, true };
4493 for (unsigned i = 0; i != 8; ++i) {
4494 int idx = MaskVals[i];
4496 InOrder[i/4] = false;
4497 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4499 AllWordsInNewV = false;
4503 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4504 if (AllWordsInNewV) {
4505 for (int i = 0; i != 8; ++i) {
4506 int idx = MaskVals[i];
4509 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4510 if ((idx != i) && idx < 4)
4512 if ((idx != i) && idx > 3)
4521 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4522 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4523 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4524 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4525 unsigned TargetMask = 0;
4526 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4527 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4528 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4529 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4530 V1 = NewV.getOperand(0);
4531 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4535 // If we have SSSE3, and all words of the result are from 1 input vector,
4536 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4537 // is present, fall back to case 4.
4538 if (Subtarget->hasSSSE3()) {
4539 SmallVector<SDValue,16> pshufbMask;
4541 // If we have elements from both input vectors, set the high bit of the
4542 // shuffle mask element to zero out elements that come from V2 in the V1
4543 // mask, and elements that come from V1 in the V2 mask, so that the two
4544 // results can be OR'd together.
4545 bool TwoInputs = V1Used && V2Used;
4546 for (unsigned i = 0; i != 8; ++i) {
4547 int EltIdx = MaskVals[i] * 2;
4548 if (TwoInputs && (EltIdx >= 16)) {
4549 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4550 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4553 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4554 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4556 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4557 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4558 DAG.getNode(ISD::BUILD_VECTOR, dl,
4559 MVT::v16i8, &pshufbMask[0], 16));
4561 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4563 // Calculate the shuffle mask for the second input, shuffle it, and
4564 // OR it with the first shuffled input.
4566 for (unsigned i = 0; i != 8; ++i) {
4567 int EltIdx = MaskVals[i] * 2;
4569 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4570 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4573 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4574 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4576 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4577 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4578 DAG.getNode(ISD::BUILD_VECTOR, dl,
4579 MVT::v16i8, &pshufbMask[0], 16));
4580 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4581 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4584 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4585 // and update MaskVals with new element order.
4586 BitVector InOrder(8);
4587 if (BestLoQuad >= 0) {
4588 SmallVector<int, 8> MaskV;
4589 for (int i = 0; i != 4; ++i) {
4590 int idx = MaskVals[i];
4592 MaskV.push_back(-1);
4594 } else if ((idx / 4) == BestLoQuad) {
4595 MaskV.push_back(idx & 3);
4598 MaskV.push_back(-1);
4601 for (unsigned i = 4; i != 8; ++i)
4603 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4606 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4607 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4609 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4613 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4614 // and update MaskVals with the new element order.
4615 if (BestHiQuad >= 0) {
4616 SmallVector<int, 8> MaskV;
4617 for (unsigned i = 0; i != 4; ++i)
4619 for (unsigned i = 4; i != 8; ++i) {
4620 int idx = MaskVals[i];
4622 MaskV.push_back(-1);
4624 } else if ((idx / 4) == BestHiQuad) {
4625 MaskV.push_back((idx & 3) + 4);
4628 MaskV.push_back(-1);
4631 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4634 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4635 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
4637 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
4641 // In case BestHi & BestLo were both -1, which means each quadword has a word
4642 // from each of the four input quadwords, calculate the InOrder bitvector now
4643 // before falling through to the insert/extract cleanup.
4644 if (BestLoQuad == -1 && BestHiQuad == -1) {
4646 for (int i = 0; i != 8; ++i)
4647 if (MaskVals[i] < 0 || MaskVals[i] == i)
4651 // The other elements are put in the right place using pextrw and pinsrw.
4652 for (unsigned i = 0; i != 8; ++i) {
4655 int EltIdx = MaskVals[i];
4658 SDValue ExtOp = (EltIdx < 8)
4659 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4660 DAG.getIntPtrConstant(EltIdx))
4661 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4662 DAG.getIntPtrConstant(EltIdx - 8));
4663 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4664 DAG.getIntPtrConstant(i));
4669 // v16i8 shuffles - Prefer shuffles in the following order:
4670 // 1. [ssse3] 1 x pshufb
4671 // 2. [ssse3] 2 x pshufb + 1 x por
4672 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4674 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4676 const X86TargetLowering &TLI) {
4677 SDValue V1 = SVOp->getOperand(0);
4678 SDValue V2 = SVOp->getOperand(1);
4679 DebugLoc dl = SVOp->getDebugLoc();
4680 SmallVector<int, 16> MaskVals;
4681 SVOp->getMask(MaskVals);
4683 // If we have SSSE3, case 1 is generated when all result bytes come from
4684 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4685 // present, fall back to case 3.
4686 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4689 for (unsigned i = 0; i < 16; ++i) {
4690 int EltIdx = MaskVals[i];
4699 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4700 if (TLI.getSubtarget()->hasSSSE3()) {
4701 SmallVector<SDValue,16> pshufbMask;
4703 // If all result elements are from one input vector, then only translate
4704 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4706 // Otherwise, we have elements from both input vectors, and must zero out
4707 // elements that come from V2 in the first mask, and V1 in the second mask
4708 // so that we can OR them together.
4709 bool TwoInputs = !(V1Only || V2Only);
4710 for (unsigned i = 0; i != 16; ++i) {
4711 int EltIdx = MaskVals[i];
4712 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4713 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4716 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4718 // If all the elements are from V2, assign it to V1 and return after
4719 // building the first pshufb.
4722 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4723 DAG.getNode(ISD::BUILD_VECTOR, dl,
4724 MVT::v16i8, &pshufbMask[0], 16));
4728 // Calculate the shuffle mask for the second input, shuffle it, and
4729 // OR it with the first shuffled input.
4731 for (unsigned i = 0; i != 16; ++i) {
4732 int EltIdx = MaskVals[i];
4734 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4737 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4739 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4740 DAG.getNode(ISD::BUILD_VECTOR, dl,
4741 MVT::v16i8, &pshufbMask[0], 16));
4742 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4745 // No SSSE3 - Calculate in place words and then fix all out of place words
4746 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4747 // the 16 different words that comprise the two doublequadword input vectors.
4748 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4749 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4750 SDValue NewV = V2Only ? V2 : V1;
4751 for (int i = 0; i != 8; ++i) {
4752 int Elt0 = MaskVals[i*2];
4753 int Elt1 = MaskVals[i*2+1];
4755 // This word of the result is all undef, skip it.
4756 if (Elt0 < 0 && Elt1 < 0)
4759 // This word of the result is already in the correct place, skip it.
4760 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4762 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4765 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4766 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4769 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4770 // using a single extract together, load it and store it.
4771 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4772 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4773 DAG.getIntPtrConstant(Elt1 / 2));
4774 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4775 DAG.getIntPtrConstant(i));
4779 // If Elt1 is defined, extract it from the appropriate source. If the
4780 // source byte is not also odd, shift the extracted word left 8 bits
4781 // otherwise clear the bottom 8 bits if we need to do an or.
4783 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4784 DAG.getIntPtrConstant(Elt1 / 2));
4785 if ((Elt1 & 1) == 0)
4786 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4787 DAG.getConstant(8, TLI.getShiftAmountTy()));
4789 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4790 DAG.getConstant(0xFF00, MVT::i16));
4792 // If Elt0 is defined, extract it from the appropriate source. If the
4793 // source byte is not also even, shift the extracted word right 8 bits. If
4794 // Elt1 was also defined, OR the extracted values together before
4795 // inserting them in the result.
4797 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4798 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4799 if ((Elt0 & 1) != 0)
4800 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4801 DAG.getConstant(8, TLI.getShiftAmountTy()));
4803 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4804 DAG.getConstant(0x00FF, MVT::i16));
4805 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4808 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4809 DAG.getIntPtrConstant(i));
4811 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4814 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4815 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4816 /// done when every pair / quad of shuffle mask elements point to elements in
4817 /// the right sequence. e.g.
4818 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4820 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4822 const TargetLowering &TLI, DebugLoc dl) {
4823 EVT VT = SVOp->getValueType(0);
4824 SDValue V1 = SVOp->getOperand(0);
4825 SDValue V2 = SVOp->getOperand(1);
4826 unsigned NumElems = VT.getVectorNumElements();
4827 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4828 EVT MaskVT = (NewWidth == 4) ? MVT::v4i16 : MVT::v2i32;
4830 switch (VT.getSimpleVT().SimpleTy) {
4831 default: assert(false && "Unexpected!");
4832 case MVT::v4f32: NewVT = MVT::v2f64; break;
4833 case MVT::v4i32: NewVT = MVT::v2i64; break;
4834 case MVT::v8i16: NewVT = MVT::v4i32; break;
4835 case MVT::v16i8: NewVT = MVT::v4i32; break;
4838 if (NewWidth == 2) {
4844 int Scale = NumElems / NewWidth;
4845 SmallVector<int, 8> MaskVec;
4846 for (unsigned i = 0; i < NumElems; i += Scale) {
4848 for (int j = 0; j < Scale; ++j) {
4849 int EltIdx = SVOp->getMaskElt(i+j);
4853 StartIdx = EltIdx - (EltIdx % Scale);
4854 if (EltIdx != StartIdx + j)
4858 MaskVec.push_back(-1);
4860 MaskVec.push_back(StartIdx / Scale);
4863 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4864 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4865 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4868 /// getVZextMovL - Return a zero-extending vector move low node.
4870 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4871 SDValue SrcOp, SelectionDAG &DAG,
4872 const X86Subtarget *Subtarget, DebugLoc dl) {
4873 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4874 LoadSDNode *LD = NULL;
4875 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4876 LD = dyn_cast<LoadSDNode>(SrcOp);
4878 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4880 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4881 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4882 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4883 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4884 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4886 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4887 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4888 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4889 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4897 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4898 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4899 DAG.getNode(ISD::BIT_CONVERT, dl,
4903 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4906 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4907 SDValue V1 = SVOp->getOperand(0);
4908 SDValue V2 = SVOp->getOperand(1);
4909 DebugLoc dl = SVOp->getDebugLoc();
4910 EVT VT = SVOp->getValueType(0);
4912 SmallVector<std::pair<int, int>, 8> Locs;
4914 SmallVector<int, 8> Mask1(4U, -1);
4915 SmallVector<int, 8> PermMask;
4916 SVOp->getMask(PermMask);
4920 for (unsigned i = 0; i != 4; ++i) {
4921 int Idx = PermMask[i];
4923 Locs[i] = std::make_pair(-1, -1);
4925 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4927 Locs[i] = std::make_pair(0, NumLo);
4931 Locs[i] = std::make_pair(1, NumHi);
4933 Mask1[2+NumHi] = Idx;
4939 if (NumLo <= 2 && NumHi <= 2) {
4940 // If no more than two elements come from either vector. This can be
4941 // implemented with two shuffles. First shuffle gather the elements.
4942 // The second shuffle, which takes the first shuffle as both of its
4943 // vector operands, put the elements into the right order.
4944 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4946 SmallVector<int, 8> Mask2(4U, -1);
4948 for (unsigned i = 0; i != 4; ++i) {
4949 if (Locs[i].first == -1)
4952 unsigned Idx = (i < 2) ? 0 : 4;
4953 Idx += Locs[i].first * 2 + Locs[i].second;
4958 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4959 } else if (NumLo == 3 || NumHi == 3) {
4960 // Otherwise, we must have three elements from one vector, call it X, and
4961 // one element from the other, call it Y. First, use a shufps to build an
4962 // intermediate vector with the one element from Y and the element from X
4963 // that will be in the same half in the final destination (the indexes don't
4964 // matter). Then, use a shufps to build the final vector, taking the half
4965 // containing the element from Y from the intermediate, and the other half
4968 // Normalize it so the 3 elements come from V1.
4969 CommuteVectorShuffleMask(PermMask, VT);
4973 // Find the element from V2.
4975 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4976 int Val = PermMask[HiIndex];
4983 Mask1[0] = PermMask[HiIndex];
4985 Mask1[2] = PermMask[HiIndex^1];
4987 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4990 Mask1[0] = PermMask[0];
4991 Mask1[1] = PermMask[1];
4992 Mask1[2] = HiIndex & 1 ? 6 : 4;
4993 Mask1[3] = HiIndex & 1 ? 4 : 6;
4994 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4996 Mask1[0] = HiIndex & 1 ? 2 : 0;
4997 Mask1[1] = HiIndex & 1 ? 0 : 2;
4998 Mask1[2] = PermMask[2];
4999 Mask1[3] = PermMask[3];
5004 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5008 // Break it into (shuffle shuffle_hi, shuffle_lo).
5010 SmallVector<int,8> LoMask(4U, -1);
5011 SmallVector<int,8> HiMask(4U, -1);
5013 SmallVector<int,8> *MaskPtr = &LoMask;
5014 unsigned MaskIdx = 0;
5017 for (unsigned i = 0; i != 4; ++i) {
5024 int Idx = PermMask[i];
5026 Locs[i] = std::make_pair(-1, -1);
5027 } else if (Idx < 4) {
5028 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5029 (*MaskPtr)[LoIdx] = Idx;
5032 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5033 (*MaskPtr)[HiIdx] = Idx;
5038 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5039 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5040 SmallVector<int, 8> MaskOps;
5041 for (unsigned i = 0; i != 4; ++i) {
5042 if (Locs[i].first == -1) {
5043 MaskOps.push_back(-1);
5045 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5046 MaskOps.push_back(Idx);
5049 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5053 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5055 SDValue V1 = Op.getOperand(0);
5056 SDValue V2 = Op.getOperand(1);
5057 EVT VT = Op.getValueType();
5059 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5061 if (HasSSE2 && VT == MVT::v2f64)
5062 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5065 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5069 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5070 SDValue V1 = Op.getOperand(0);
5071 SDValue V2 = Op.getOperand(1);
5072 EVT VT = Op.getValueType();
5074 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5075 "unsupported shuffle type");
5077 if (V2.getOpcode() == ISD::UNDEF)
5081 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5085 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5086 SDValue V1 = Op.getOperand(0);
5087 SDValue V2 = Op.getOperand(1);
5088 EVT VT = Op.getValueType();
5089 unsigned NumElems = VT.getVectorNumElements();
5091 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5092 // operand of these instructions is only memory, so check if there's a
5093 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5095 bool CanFoldLoad = false;
5099 // Trivial case, when V2 comes from a load.
5100 if (TmpV2.hasOneUse() && TmpV2.getOpcode() == ISD::BIT_CONVERT)
5101 TmpV2 = TmpV2.getOperand(0);
5102 if (TmpV2.hasOneUse() && TmpV2.getOpcode() == ISD::SCALAR_TO_VECTOR)
5103 TmpV2 = TmpV2.getOperand(0);
5104 if (MayFoldLoad(TmpV2))
5107 // When V1 is a load, it can be folded later into a store in isel, example:
5108 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5110 // (MOVLPSmr addr:$src1, VR128:$src2)
5111 // So, recognize this potential and also use MOVLPS or MOVLPD
5112 if (TmpV1.hasOneUse() && TmpV1.getOpcode() == ISD::BIT_CONVERT)
5113 TmpV1 = TmpV1.getOperand(0);
5114 if (MayFoldLoad(TmpV1) && MayFoldIntoStore(Op))
5118 if (HasSSE2 && NumElems == 2)
5119 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5122 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5125 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5126 // movl and movlp will both match v2i64, but v2i64 is never matched by
5127 // movl earlier because we make it strict to avoid messing with the movlp load
5128 // folding logic (see the code above getMOVLP call). Match it here then,
5129 // this is horrible, but will stay like this until we move all shuffle
5130 // matching to x86 specific nodes. Note that for the 1st condition all
5131 // types are matched with movsd.
5132 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5133 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5135 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5138 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5140 // Invert the operand order and use SHUFPS to match it.
5141 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5142 X86::getShuffleSHUFImmediate(SVOp), DAG);
5146 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5147 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5148 SDValue V1 = Op.getOperand(0);
5149 SDValue V2 = Op.getOperand(1);
5150 EVT VT = Op.getValueType();
5151 DebugLoc dl = Op.getDebugLoc();
5152 unsigned NumElems = VT.getVectorNumElements();
5153 bool isMMX = VT.getSizeInBits() == 64;
5154 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5155 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5156 bool V1IsSplat = false;
5157 bool V2IsSplat = false;
5158 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5159 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5160 MachineFunction &MF = DAG.getMachineFunction();
5161 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5163 if (isZeroShuffle(SVOp))
5164 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5166 // Promote splats to v4f32.
5167 if (SVOp->isSplat()) {
5168 if (isMMX || NumElems < 4)
5170 return PromoteSplat(SVOp, DAG);
5173 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5175 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5176 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5177 if (NewOp.getNode())
5178 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5179 LowerVECTOR_SHUFFLE(NewOp, DAG));
5180 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5181 // FIXME: Figure out a cleaner way to do this.
5182 // Try to make use of movq to zero out the top part.
5183 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5184 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5185 if (NewOp.getNode()) {
5186 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5187 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5188 DAG, Subtarget, dl);
5190 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5191 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5192 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5193 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5194 DAG, Subtarget, dl);
5198 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp)) {
5199 // NOTE: isPSHUFDMask can also match this mask, if speed is more
5200 // important than size here, this will be matched by pshufd
5201 if (VT == MVT::v4f32)
5202 return getTargetShuffleNode(X86ISD::UNPCKLPS, dl, VT, V1, V1, DAG);
5203 if (HasSSE2 && VT == MVT::v16i8)
5204 return getTargetShuffleNode(X86ISD::PUNPCKLBW, dl, VT, V1, V1, DAG);
5205 if (HasSSE2 && VT == MVT::v8i16)
5206 return getTargetShuffleNode(X86ISD::PUNPCKLWD, dl, VT, V1, V1, DAG);
5207 if (HasSSE2 && VT == MVT::v4i32)
5208 return getTargetShuffleNode(X86ISD::PUNPCKLDQ, dl, VT, V1, V1, DAG);
5211 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp)) {
5212 // NOTE: isPSHUFDMask can also match this mask, if speed is more
5213 // important than size here, this will be matched by pshufd
5214 if (VT == MVT::v4f32)
5215 return getTargetShuffleNode(X86ISD::UNPCKHPS, dl, VT, V1, V1, DAG);
5216 if (HasSSE2 && VT == MVT::v16i8)
5217 return getTargetShuffleNode(X86ISD::PUNPCKHBW, dl, VT, V1, V1, DAG);
5218 if (HasSSE2 && VT == MVT::v8i16)
5219 return getTargetShuffleNode(X86ISD::PUNPCKHWD, dl, VT, V1, V1, DAG);
5220 if (HasSSE2 && VT == MVT::v4i32)
5221 return getTargetShuffleNode(X86ISD::PUNPCKHDQ, dl, VT, V1, V1, DAG);
5224 if (X86::isPSHUFDMask(SVOp)) {
5225 // The actual implementation will match the mask in the if above and then
5226 // during isel it can match several different instructions, not only pshufd
5227 // as its name says, sad but true, emulate the behavior for now...
5228 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5229 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5231 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5233 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5234 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5236 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5237 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5240 if (VT == MVT::v4f32)
5241 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5245 // Check if this can be converted into a logical shift.
5246 bool isLeft = false;
5249 bool isShift = getSubtarget()->hasSSE2() &&
5250 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5251 if (isShift && ShVal.hasOneUse()) {
5252 // If the shifted value has multiple uses, it may be cheaper to use
5253 // v_set0 + movlhps or movhlps, etc.
5254 EVT EltVT = VT.getVectorElementType();
5255 ShAmt *= EltVT.getSizeInBits();
5256 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5259 if (X86::isMOVLMask(SVOp)) {
5262 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5263 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5264 if (!isMMX && !X86::isMOVLPMask(SVOp)) {
5265 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5266 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5268 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5269 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5273 // FIXME: fold these into legal mask.
5275 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5276 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5278 if (X86::isMOVHLPSMask(SVOp))
5279 return getMOVHighToLow(Op, dl, DAG);
5281 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5282 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5284 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5285 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5287 if (X86::isMOVLPMask(SVOp))
5288 return getMOVLP(Op, dl, DAG, HasSSE2);
5291 if (ShouldXformToMOVHLPS(SVOp) ||
5292 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5293 return CommuteVectorShuffle(SVOp, DAG);
5296 // No better options. Use a vshl / vsrl.
5297 EVT EltVT = VT.getVectorElementType();
5298 ShAmt *= EltVT.getSizeInBits();
5299 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5302 bool Commuted = false;
5303 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5304 // 1,1,1,1 -> v8i16 though.
5305 V1IsSplat = isSplatVector(V1.getNode());
5306 V2IsSplat = isSplatVector(V2.getNode());
5308 // Canonicalize the splat or undef, if present, to be on the RHS.
5309 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5310 Op = CommuteVectorShuffle(SVOp, DAG);
5311 SVOp = cast<ShuffleVectorSDNode>(Op);
5312 V1 = SVOp->getOperand(0);
5313 V2 = SVOp->getOperand(1);
5314 std::swap(V1IsSplat, V2IsSplat);
5315 std::swap(V1IsUndef, V2IsUndef);
5319 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5320 // Shuffling low element of v1 into undef, just return v1.
5323 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5324 // the instruction selector will not match, so get a canonical MOVL with
5325 // swapped operands to undo the commute.
5326 return getMOVL(DAG, dl, VT, V2, V1);
5329 if (X86::isUNPCKLMask(SVOp) ||
5330 X86::isUNPCKHMask(SVOp))
5334 // Normalize mask so all entries that point to V2 points to its first
5335 // element then try to match unpck{h|l} again. If match, return a
5336 // new vector_shuffle with the corrected mask.
5337 SDValue NewMask = NormalizeMask(SVOp, DAG);
5338 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5339 if (NSVOp != SVOp) {
5340 if (X86::isUNPCKLMask(NSVOp, true)) {
5342 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5349 // Commute is back and try unpck* again.
5350 // FIXME: this seems wrong.
5351 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5352 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5353 if (X86::isUNPCKLMask(NewSVOp) ||
5354 X86::isUNPCKHMask(NewSVOp))
5358 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
5360 // Normalize the node to match x86 shuffle ops if needed
5361 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5362 return CommuteVectorShuffle(SVOp, DAG);
5364 // Check for legal shuffle and return?
5365 SmallVector<int, 16> PermMask;
5366 SVOp->getMask(PermMask);
5367 if (isShuffleMaskLegal(PermMask, VT))
5370 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5371 if (VT == MVT::v8i16) {
5372 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5373 if (NewOp.getNode())
5377 if (VT == MVT::v16i8) {
5378 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5379 if (NewOp.getNode())
5383 // Handle all 4 wide cases with a number of shuffles except for MMX.
5384 if (NumElems == 4 && !isMMX)
5385 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5391 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5392 SelectionDAG &DAG) const {
5393 EVT VT = Op.getValueType();
5394 DebugLoc dl = Op.getDebugLoc();
5395 if (VT.getSizeInBits() == 8) {
5396 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
5397 Op.getOperand(0), Op.getOperand(1));
5398 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5399 DAG.getValueType(VT));
5400 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5401 } else if (VT.getSizeInBits() == 16) {
5402 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5403 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
5405 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5406 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5407 DAG.getNode(ISD::BIT_CONVERT, dl,
5411 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
5412 Op.getOperand(0), Op.getOperand(1));
5413 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5414 DAG.getValueType(VT));
5415 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5416 } else if (VT == MVT::f32) {
5417 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
5418 // the result back to FR32 register. It's only worth matching if the
5419 // result has a single use which is a store or a bitcast to i32. And in
5420 // the case of a store, it's not worth it if the index is a constant 0,
5421 // because a MOVSSmr can be used instead, which is smaller and faster.
5422 if (!Op.hasOneUse())
5424 SDNode *User = *Op.getNode()->use_begin();
5425 if ((User->getOpcode() != ISD::STORE ||
5426 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5427 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5428 (User->getOpcode() != ISD::BIT_CONVERT ||
5429 User->getValueType(0) != MVT::i32))
5431 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5432 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
5435 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
5436 } else if (VT == MVT::i32) {
5437 // ExtractPS works with constant index.
5438 if (isa<ConstantSDNode>(Op.getOperand(1)))
5446 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5447 SelectionDAG &DAG) const {
5448 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5451 if (Subtarget->hasSSE41()) {
5452 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5457 EVT VT = Op.getValueType();
5458 DebugLoc dl = Op.getDebugLoc();
5459 // TODO: handle v16i8.
5460 if (VT.getSizeInBits() == 16) {
5461 SDValue Vec = Op.getOperand(0);
5462 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5464 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5465 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5466 DAG.getNode(ISD::BIT_CONVERT, dl,
5469 // Transform it so it match pextrw which produces a 32-bit result.
5470 EVT EltVT = MVT::i32;
5471 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5472 Op.getOperand(0), Op.getOperand(1));
5473 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5474 DAG.getValueType(VT));
5475 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5476 } else if (VT.getSizeInBits() == 32) {
5477 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5481 // SHUFPS the element to the lowest double word, then movss.
5482 int Mask[4] = { Idx, -1, -1, -1 };
5483 EVT VVT = Op.getOperand(0).getValueType();
5484 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5485 DAG.getUNDEF(VVT), Mask);
5486 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5487 DAG.getIntPtrConstant(0));
5488 } else if (VT.getSizeInBits() == 64) {
5489 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5490 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5491 // to match extract_elt for f64.
5492 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5496 // UNPCKHPD the element to the lowest double word, then movsd.
5497 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5498 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5499 int Mask[2] = { 1, -1 };
5500 EVT VVT = Op.getOperand(0).getValueType();
5501 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5502 DAG.getUNDEF(VVT), Mask);
5503 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5504 DAG.getIntPtrConstant(0));
5511 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5512 SelectionDAG &DAG) const {
5513 EVT VT = Op.getValueType();
5514 EVT EltVT = VT.getVectorElementType();
5515 DebugLoc dl = Op.getDebugLoc();
5517 SDValue N0 = Op.getOperand(0);
5518 SDValue N1 = Op.getOperand(1);
5519 SDValue N2 = Op.getOperand(2);
5521 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5522 isa<ConstantSDNode>(N2)) {
5524 if (VT == MVT::v8i16)
5525 Opc = X86ISD::PINSRW;
5526 else if (VT == MVT::v4i16)
5527 Opc = X86ISD::MMX_PINSRW;
5528 else if (VT == MVT::v16i8)
5529 Opc = X86ISD::PINSRB;
5531 Opc = X86ISD::PINSRB;
5533 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5535 if (N1.getValueType() != MVT::i32)
5536 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5537 if (N2.getValueType() != MVT::i32)
5538 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5539 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5540 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5541 // Bits [7:6] of the constant are the source select. This will always be
5542 // zero here. The DAG Combiner may combine an extract_elt index into these
5543 // bits. For example (insert (extract, 3), 2) could be matched by putting
5544 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5545 // Bits [5:4] of the constant are the destination select. This is the
5546 // value of the incoming immediate.
5547 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5548 // combine either bitwise AND or insert of float 0.0 to set these bits.
5549 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5550 // Create this as a scalar to vector..
5551 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5552 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5553 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5554 // PINSR* works with constant index.
5561 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5562 EVT VT = Op.getValueType();
5563 EVT EltVT = VT.getVectorElementType();
5565 if (Subtarget->hasSSE41())
5566 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5568 if (EltVT == MVT::i8)
5571 DebugLoc dl = Op.getDebugLoc();
5572 SDValue N0 = Op.getOperand(0);
5573 SDValue N1 = Op.getOperand(1);
5574 SDValue N2 = Op.getOperand(2);
5576 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5577 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5578 // as its second argument.
5579 if (N1.getValueType() != MVT::i32)
5580 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5581 if (N2.getValueType() != MVT::i32)
5582 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5583 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5584 dl, VT, N0, N1, N2);
5590 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5591 DebugLoc dl = Op.getDebugLoc();
5593 if (Op.getValueType() == MVT::v1i64 &&
5594 Op.getOperand(0).getValueType() == MVT::i64)
5595 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5597 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5598 EVT VT = MVT::v2i32;
5599 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5606 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5607 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5610 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5611 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5612 // one of the above mentioned nodes. It has to be wrapped because otherwise
5613 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5614 // be used to form addressing mode. These wrapped nodes will be selected
5617 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5618 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5620 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5622 unsigned char OpFlag = 0;
5623 unsigned WrapperKind = X86ISD::Wrapper;
5624 CodeModel::Model M = getTargetMachine().getCodeModel();
5626 if (Subtarget->isPICStyleRIPRel() &&
5627 (M == CodeModel::Small || M == CodeModel::Kernel))
5628 WrapperKind = X86ISD::WrapperRIP;
5629 else if (Subtarget->isPICStyleGOT())
5630 OpFlag = X86II::MO_GOTOFF;
5631 else if (Subtarget->isPICStyleStubPIC())
5632 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5634 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5636 CP->getOffset(), OpFlag);
5637 DebugLoc DL = CP->getDebugLoc();
5638 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5639 // With PIC, the address is actually $g + Offset.
5641 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5642 DAG.getNode(X86ISD::GlobalBaseReg,
5643 DebugLoc(), getPointerTy()),
5650 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5651 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
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.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5669 DebugLoc DL = JT->getDebugLoc();
5670 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5672 // With PIC, the address is actually $g + Offset.
5674 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5675 DAG.getNode(X86ISD::GlobalBaseReg,
5676 DebugLoc(), getPointerTy()),
5684 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5685 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5687 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5689 unsigned char OpFlag = 0;
5690 unsigned WrapperKind = X86ISD::Wrapper;
5691 CodeModel::Model M = getTargetMachine().getCodeModel();
5693 if (Subtarget->isPICStyleRIPRel() &&
5694 (M == CodeModel::Small || M == CodeModel::Kernel))
5695 WrapperKind = X86ISD::WrapperRIP;
5696 else if (Subtarget->isPICStyleGOT())
5697 OpFlag = X86II::MO_GOTOFF;
5698 else if (Subtarget->isPICStyleStubPIC())
5699 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5701 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5703 DebugLoc DL = Op.getDebugLoc();
5704 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5707 // With PIC, the address is actually $g + Offset.
5708 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5709 !Subtarget->is64Bit()) {
5710 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5711 DAG.getNode(X86ISD::GlobalBaseReg,
5712 DebugLoc(), getPointerTy()),
5720 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5721 // Create the TargetBlockAddressAddress node.
5722 unsigned char OpFlags =
5723 Subtarget->ClassifyBlockAddressReference();
5724 CodeModel::Model M = getTargetMachine().getCodeModel();
5725 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5726 DebugLoc dl = Op.getDebugLoc();
5727 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5728 /*isTarget=*/true, OpFlags);
5730 if (Subtarget->isPICStyleRIPRel() &&
5731 (M == CodeModel::Small || M == CodeModel::Kernel))
5732 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5734 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5736 // With PIC, the address is actually $g + Offset.
5737 if (isGlobalRelativeToPICBase(OpFlags)) {
5738 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5739 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5747 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5749 SelectionDAG &DAG) const {
5750 // Create the TargetGlobalAddress node, folding in the constant
5751 // offset if it is legal.
5752 unsigned char OpFlags =
5753 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5754 CodeModel::Model M = getTargetMachine().getCodeModel();
5756 if (OpFlags == X86II::MO_NO_FLAG &&
5757 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5758 // A direct static reference to a global.
5759 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5762 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5765 if (Subtarget->isPICStyleRIPRel() &&
5766 (M == CodeModel::Small || M == CodeModel::Kernel))
5767 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5769 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5771 // With PIC, the address is actually $g + Offset.
5772 if (isGlobalRelativeToPICBase(OpFlags)) {
5773 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5774 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5778 // For globals that require a load from a stub to get the address, emit the
5780 if (isGlobalStubReference(OpFlags))
5781 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5782 PseudoSourceValue::getGOT(), 0, false, false, 0);
5784 // If there was a non-zero offset that we didn't fold, create an explicit
5787 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5788 DAG.getConstant(Offset, getPointerTy()));
5794 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5795 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5796 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5797 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5801 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5802 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5803 unsigned char OperandFlags) {
5804 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5805 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5806 DebugLoc dl = GA->getDebugLoc();
5807 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5808 GA->getValueType(0),
5812 SDValue Ops[] = { Chain, TGA, *InFlag };
5813 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5815 SDValue Ops[] = { Chain, TGA };
5816 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5819 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5820 MFI->setAdjustsStack(true);
5822 SDValue Flag = Chain.getValue(1);
5823 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5826 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5828 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5831 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5832 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5833 DAG.getNode(X86ISD::GlobalBaseReg,
5834 DebugLoc(), PtrVT), InFlag);
5835 InFlag = Chain.getValue(1);
5837 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5840 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5842 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5844 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5845 X86::RAX, X86II::MO_TLSGD);
5848 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5849 // "local exec" model.
5850 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5851 const EVT PtrVT, TLSModel::Model model,
5853 DebugLoc dl = GA->getDebugLoc();
5854 // Get the Thread Pointer
5855 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5857 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5860 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5861 NULL, 0, false, false, 0);
5863 unsigned char OperandFlags = 0;
5864 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5866 unsigned WrapperKind = X86ISD::Wrapper;
5867 if (model == TLSModel::LocalExec) {
5868 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5869 } else if (is64Bit) {
5870 assert(model == TLSModel::InitialExec);
5871 OperandFlags = X86II::MO_GOTTPOFF;
5872 WrapperKind = X86ISD::WrapperRIP;
5874 assert(model == TLSModel::InitialExec);
5875 OperandFlags = X86II::MO_INDNTPOFF;
5878 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5880 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5881 GA->getValueType(0),
5882 GA->getOffset(), OperandFlags);
5883 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5885 if (model == TLSModel::InitialExec)
5886 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5887 PseudoSourceValue::getGOT(), 0, false, false, 0);
5889 // The address of the thread local variable is the add of the thread
5890 // pointer with the offset of the variable.
5891 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5895 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5897 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5898 const GlobalValue *GV = GA->getGlobal();
5900 if (Subtarget->isTargetELF()) {
5901 // TODO: implement the "local dynamic" model
5902 // TODO: implement the "initial exec"model for pic executables
5904 // If GV is an alias then use the aliasee for determining
5905 // thread-localness.
5906 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5907 GV = GA->resolveAliasedGlobal(false);
5909 TLSModel::Model model
5910 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5913 case TLSModel::GeneralDynamic:
5914 case TLSModel::LocalDynamic: // not implemented
5915 if (Subtarget->is64Bit())
5916 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5917 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5919 case TLSModel::InitialExec:
5920 case TLSModel::LocalExec:
5921 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5922 Subtarget->is64Bit());
5924 } else if (Subtarget->isTargetDarwin()) {
5925 // Darwin only has one model of TLS. Lower to that.
5926 unsigned char OpFlag = 0;
5927 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5928 X86ISD::WrapperRIP : X86ISD::Wrapper;
5930 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5932 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5933 !Subtarget->is64Bit();
5935 OpFlag = X86II::MO_TLVP_PIC_BASE;
5937 OpFlag = X86II::MO_TLVP;
5938 DebugLoc DL = Op.getDebugLoc();
5939 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
5941 GA->getOffset(), OpFlag);
5942 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5944 // With PIC32, the address is actually $g + Offset.
5946 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5947 DAG.getNode(X86ISD::GlobalBaseReg,
5948 DebugLoc(), getPointerTy()),
5951 // Lowering the machine isd will make sure everything is in the right
5953 SDValue Args[] = { Offset };
5954 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5956 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5957 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5958 MFI->setAdjustsStack(true);
5960 // And our return value (tls address) is in the standard call return value
5962 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5963 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5967 "TLS not implemented for this target.");
5969 llvm_unreachable("Unreachable");
5974 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5975 /// take a 2 x i32 value to shift plus a shift amount.
5976 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5977 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5978 EVT VT = Op.getValueType();
5979 unsigned VTBits = VT.getSizeInBits();
5980 DebugLoc dl = Op.getDebugLoc();
5981 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5982 SDValue ShOpLo = Op.getOperand(0);
5983 SDValue ShOpHi = Op.getOperand(1);
5984 SDValue ShAmt = Op.getOperand(2);
5985 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5986 DAG.getConstant(VTBits - 1, MVT::i8))
5987 : DAG.getConstant(0, VT);
5990 if (Op.getOpcode() == ISD::SHL_PARTS) {
5991 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5992 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5994 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5995 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5998 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5999 DAG.getConstant(VTBits, MVT::i8));
6000 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6001 AndNode, DAG.getConstant(0, MVT::i8));
6004 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6005 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6006 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6008 if (Op.getOpcode() == ISD::SHL_PARTS) {
6009 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6010 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6012 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6013 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6016 SDValue Ops[2] = { Lo, Hi };
6017 return DAG.getMergeValues(Ops, 2, dl);
6020 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6021 SelectionDAG &DAG) const {
6022 EVT SrcVT = Op.getOperand(0).getValueType();
6024 if (SrcVT.isVector()) {
6025 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
6031 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6032 "Unknown SINT_TO_FP to lower!");
6034 // These are really Legal; return the operand so the caller accepts it as
6036 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6038 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6039 Subtarget->is64Bit()) {
6043 DebugLoc dl = Op.getDebugLoc();
6044 unsigned Size = SrcVT.getSizeInBits()/8;
6045 MachineFunction &MF = DAG.getMachineFunction();
6046 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6047 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6048 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6050 PseudoSourceValue::getFixedStack(SSFI), 0,
6052 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6055 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6057 SelectionDAG &DAG) const {
6059 DebugLoc dl = Op.getDebugLoc();
6061 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6063 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
6065 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6066 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6067 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
6068 Tys, Ops, array_lengthof(Ops));
6071 Chain = Result.getValue(1);
6072 SDValue InFlag = Result.getValue(2);
6074 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6075 // shouldn't be necessary except that RFP cannot be live across
6076 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6077 MachineFunction &MF = DAG.getMachineFunction();
6078 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
6079 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6080 Tys = DAG.getVTList(MVT::Other);
6082 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6084 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
6085 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
6086 PseudoSourceValue::getFixedStack(SSFI), 0,
6093 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6094 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6095 SelectionDAG &DAG) const {
6096 // This algorithm is not obvious. Here it is in C code, more or less:
6098 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6099 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6100 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6102 // Copy ints to xmm registers.
6103 __m128i xh = _mm_cvtsi32_si128( hi );
6104 __m128i xl = _mm_cvtsi32_si128( lo );
6106 // Combine into low half of a single xmm register.
6107 __m128i x = _mm_unpacklo_epi32( xh, xl );
6111 // Merge in appropriate exponents to give the integer bits the right
6113 x = _mm_unpacklo_epi32( x, exp );
6115 // Subtract away the biases to deal with the IEEE-754 double precision
6117 d = _mm_sub_pd( (__m128d) x, bias );
6119 // All conversions up to here are exact. The correctly rounded result is
6120 // calculated using the current rounding mode using the following
6122 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6123 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6124 // store doesn't really need to be here (except
6125 // maybe to zero the other double)
6130 DebugLoc dl = Op.getDebugLoc();
6131 LLVMContext *Context = DAG.getContext();
6133 // Build some magic constants.
6134 std::vector<Constant*> CV0;
6135 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6136 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6137 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6138 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6139 Constant *C0 = ConstantVector::get(CV0);
6140 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6142 std::vector<Constant*> CV1;
6144 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6146 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6147 Constant *C1 = ConstantVector::get(CV1);
6148 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6150 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6151 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6153 DAG.getIntPtrConstant(1)));
6154 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6155 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6157 DAG.getIntPtrConstant(0)));
6158 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6159 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6160 PseudoSourceValue::getConstantPool(), 0,
6162 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6163 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
6164 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6165 PseudoSourceValue::getConstantPool(), 0,
6167 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6169 // Add the halves; easiest way is to swap them into another reg first.
6170 int ShufMask[2] = { 1, -1 };
6171 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6172 DAG.getUNDEF(MVT::v2f64), ShufMask);
6173 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6174 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6175 DAG.getIntPtrConstant(0));
6178 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6179 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6180 SelectionDAG &DAG) const {
6181 DebugLoc dl = Op.getDebugLoc();
6182 // FP constant to bias correct the final result.
6183 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6186 // Load the 32-bit value into an XMM register.
6187 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6188 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6190 DAG.getIntPtrConstant(0)));
6192 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6193 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
6194 DAG.getIntPtrConstant(0));
6196 // Or the load with the bias.
6197 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6198 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6199 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6201 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6202 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6203 MVT::v2f64, Bias)));
6204 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6205 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
6206 DAG.getIntPtrConstant(0));
6208 // Subtract the bias.
6209 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
6211 // Handle final rounding.
6212 EVT DestVT = Op.getValueType();
6214 if (DestVT.bitsLT(MVT::f64)) {
6215 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
6216 DAG.getIntPtrConstant(0));
6217 } else if (DestVT.bitsGT(MVT::f64)) {
6218 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
6221 // Handle final rounding.
6225 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
6226 SelectionDAG &DAG) const {
6227 SDValue N0 = Op.getOperand(0);
6228 DebugLoc dl = Op.getDebugLoc();
6230 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
6231 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
6232 // the optimization here.
6233 if (DAG.SignBitIsZero(N0))
6234 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
6236 EVT SrcVT = N0.getValueType();
6237 EVT DstVT = Op.getValueType();
6238 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
6239 return LowerUINT_TO_FP_i64(Op, DAG);
6240 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
6241 return LowerUINT_TO_FP_i32(Op, DAG);
6243 // Make a 64-bit buffer, and use it to build an FILD.
6244 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
6245 if (SrcVT == MVT::i32) {
6246 SDValue WordOff = DAG.getConstant(4, getPointerTy());
6247 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
6248 getPointerTy(), StackSlot, WordOff);
6249 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6250 StackSlot, NULL, 0, false, false, 0);
6251 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
6252 OffsetSlot, NULL, 0, false, false, 0);
6253 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
6257 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
6258 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6259 StackSlot, NULL, 0, false, false, 0);
6260 // For i64 source, we need to add the appropriate power of 2 if the input
6261 // was negative. This is the same as the optimization in
6262 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
6263 // we must be careful to do the computation in x87 extended precision, not
6264 // in SSE. (The generic code can't know it's OK to do this, or how to.)
6265 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
6266 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
6267 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
6269 APInt FF(32, 0x5F800000ULL);
6271 // Check whether the sign bit is set.
6272 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
6273 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
6276 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
6277 SDValue FudgePtr = DAG.getConstantPool(
6278 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
6281 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
6282 SDValue Zero = DAG.getIntPtrConstant(0);
6283 SDValue Four = DAG.getIntPtrConstant(4);
6284 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
6286 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
6288 // Load the value out, extending it from f32 to f80.
6289 // FIXME: Avoid the extend by constructing the right constant pool?
6290 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
6291 FudgePtr, PseudoSourceValue::getConstantPool(),
6292 0, MVT::f32, false, false, 4);
6293 // Extend everything to 80 bits to force it to be done on x87.
6294 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
6295 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
6298 std::pair<SDValue,SDValue> X86TargetLowering::
6299 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
6300 DebugLoc dl = Op.getDebugLoc();
6302 EVT DstTy = Op.getValueType();
6305 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
6309 assert(DstTy.getSimpleVT() <= MVT::i64 &&
6310 DstTy.getSimpleVT() >= MVT::i16 &&
6311 "Unknown FP_TO_SINT to lower!");
6313 // These are really Legal.
6314 if (DstTy == MVT::i32 &&
6315 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6316 return std::make_pair(SDValue(), SDValue());
6317 if (Subtarget->is64Bit() &&
6318 DstTy == MVT::i64 &&
6319 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6320 return std::make_pair(SDValue(), SDValue());
6322 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
6324 MachineFunction &MF = DAG.getMachineFunction();
6325 unsigned MemSize = DstTy.getSizeInBits()/8;
6326 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6327 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6330 switch (DstTy.getSimpleVT().SimpleTy) {
6331 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
6332 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
6333 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
6334 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
6337 SDValue Chain = DAG.getEntryNode();
6338 SDValue Value = Op.getOperand(0);
6339 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
6340 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
6341 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
6342 PseudoSourceValue::getFixedStack(SSFI), 0,
6344 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
6346 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
6348 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
6349 Chain = Value.getValue(1);
6350 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6351 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6354 // Build the FP_TO_INT*_IN_MEM
6355 SDValue Ops[] = { Chain, Value, StackSlot };
6356 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
6358 return std::make_pair(FIST, StackSlot);
6361 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
6362 SelectionDAG &DAG) const {
6363 if (Op.getValueType().isVector()) {
6364 if (Op.getValueType() == MVT::v2i32 &&
6365 Op.getOperand(0).getValueType() == MVT::v2f64) {
6371 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
6372 SDValue FIST = Vals.first, StackSlot = Vals.second;
6373 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
6374 if (FIST.getNode() == 0) return Op;
6377 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6378 FIST, StackSlot, NULL, 0, false, false, 0);
6381 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
6382 SelectionDAG &DAG) const {
6383 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
6384 SDValue FIST = Vals.first, StackSlot = Vals.second;
6385 assert(FIST.getNode() && "Unexpected failure");
6388 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6389 FIST, StackSlot, NULL, 0, false, false, 0);
6392 SDValue X86TargetLowering::LowerFABS(SDValue Op,
6393 SelectionDAG &DAG) const {
6394 LLVMContext *Context = DAG.getContext();
6395 DebugLoc dl = Op.getDebugLoc();
6396 EVT VT = Op.getValueType();
6399 EltVT = VT.getVectorElementType();
6400 std::vector<Constant*> CV;
6401 if (EltVT == MVT::f64) {
6402 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
6406 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
6412 Constant *C = ConstantVector::get(CV);
6413 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6414 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6415 PseudoSourceValue::getConstantPool(), 0,
6417 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
6420 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
6421 LLVMContext *Context = DAG.getContext();
6422 DebugLoc dl = Op.getDebugLoc();
6423 EVT VT = Op.getValueType();
6426 EltVT = VT.getVectorElementType();
6427 std::vector<Constant*> CV;
6428 if (EltVT == MVT::f64) {
6429 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
6433 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
6439 Constant *C = ConstantVector::get(CV);
6440 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6441 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6442 PseudoSourceValue::getConstantPool(), 0,
6444 if (VT.isVector()) {
6445 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6446 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6447 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6449 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
6451 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6455 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6456 LLVMContext *Context = DAG.getContext();
6457 SDValue Op0 = Op.getOperand(0);
6458 SDValue Op1 = Op.getOperand(1);
6459 DebugLoc dl = Op.getDebugLoc();
6460 EVT VT = Op.getValueType();
6461 EVT SrcVT = Op1.getValueType();
6463 // If second operand is smaller, extend it first.
6464 if (SrcVT.bitsLT(VT)) {
6465 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6468 // And if it is bigger, shrink it first.
6469 if (SrcVT.bitsGT(VT)) {
6470 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6474 // At this point the operands and the result should have the same
6475 // type, and that won't be f80 since that is not custom lowered.
6477 // First get the sign bit of second operand.
6478 std::vector<Constant*> CV;
6479 if (SrcVT == MVT::f64) {
6480 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6481 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6483 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6484 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6485 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6486 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6488 Constant *C = ConstantVector::get(CV);
6489 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6490 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6491 PseudoSourceValue::getConstantPool(), 0,
6493 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6495 // Shift sign bit right or left if the two operands have different types.
6496 if (SrcVT.bitsGT(VT)) {
6497 // Op0 is MVT::f32, Op1 is MVT::f64.
6498 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6499 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6500 DAG.getConstant(32, MVT::i32));
6501 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
6502 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6503 DAG.getIntPtrConstant(0));
6506 // Clear first operand sign bit.
6508 if (VT == MVT::f64) {
6509 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6510 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6512 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6513 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6514 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6515 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6517 C = ConstantVector::get(CV);
6518 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6519 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6520 PseudoSourceValue::getConstantPool(), 0,
6522 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6524 // Or the value with the sign bit.
6525 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6528 /// Emit nodes that will be selected as "test Op0,Op0", or something
6530 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6531 SelectionDAG &DAG) const {
6532 DebugLoc dl = Op.getDebugLoc();
6534 // CF and OF aren't always set the way we want. Determine which
6535 // of these we need.
6536 bool NeedCF = false;
6537 bool NeedOF = false;
6540 case X86::COND_A: case X86::COND_AE:
6541 case X86::COND_B: case X86::COND_BE:
6544 case X86::COND_G: case X86::COND_GE:
6545 case X86::COND_L: case X86::COND_LE:
6546 case X86::COND_O: case X86::COND_NO:
6551 // See if we can use the EFLAGS value from the operand instead of
6552 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6553 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6554 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6555 // Emit a CMP with 0, which is the TEST pattern.
6556 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6557 DAG.getConstant(0, Op.getValueType()));
6559 unsigned Opcode = 0;
6560 unsigned NumOperands = 0;
6561 switch (Op.getNode()->getOpcode()) {
6563 // Due to an isel shortcoming, be conservative if this add is likely to be
6564 // selected as part of a load-modify-store instruction. When the root node
6565 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6566 // uses of other nodes in the match, such as the ADD in this case. This
6567 // leads to the ADD being left around and reselected, with the result being
6568 // two adds in the output. Alas, even if none our users are stores, that
6569 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6570 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6571 // climbing the DAG back to the root, and it doesn't seem to be worth the
6573 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6574 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6575 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6578 if (ConstantSDNode *C =
6579 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6580 // An add of one will be selected as an INC.
6581 if (C->getAPIntValue() == 1) {
6582 Opcode = X86ISD::INC;
6587 // An add of negative one (subtract of one) will be selected as a DEC.
6588 if (C->getAPIntValue().isAllOnesValue()) {
6589 Opcode = X86ISD::DEC;
6595 // Otherwise use a regular EFLAGS-setting add.
6596 Opcode = X86ISD::ADD;
6600 // If the primary and result isn't used, don't bother using X86ISD::AND,
6601 // because a TEST instruction will be better.
6602 bool NonFlagUse = false;
6603 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6604 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6606 unsigned UOpNo = UI.getOperandNo();
6607 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6608 // Look pass truncate.
6609 UOpNo = User->use_begin().getOperandNo();
6610 User = *User->use_begin();
6613 if (User->getOpcode() != ISD::BRCOND &&
6614 User->getOpcode() != ISD::SETCC &&
6615 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6628 // Due to the ISEL shortcoming noted above, be conservative if this op is
6629 // likely to be selected as part of a load-modify-store instruction.
6630 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6631 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6632 if (UI->getOpcode() == ISD::STORE)
6635 // Otherwise use a regular EFLAGS-setting instruction.
6636 switch (Op.getNode()->getOpcode()) {
6637 default: llvm_unreachable("unexpected operator!");
6638 case ISD::SUB: Opcode = X86ISD::SUB; break;
6639 case ISD::OR: Opcode = X86ISD::OR; break;
6640 case ISD::XOR: Opcode = X86ISD::XOR; break;
6641 case ISD::AND: Opcode = X86ISD::AND; break;
6653 return SDValue(Op.getNode(), 1);
6660 // Emit a CMP with 0, which is the TEST pattern.
6661 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6662 DAG.getConstant(0, Op.getValueType()));
6664 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6665 SmallVector<SDValue, 4> Ops;
6666 for (unsigned i = 0; i != NumOperands; ++i)
6667 Ops.push_back(Op.getOperand(i));
6669 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6670 DAG.ReplaceAllUsesWith(Op, New);
6671 return SDValue(New.getNode(), 1);
6674 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6676 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6677 SelectionDAG &DAG) const {
6678 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6679 if (C->getAPIntValue() == 0)
6680 return EmitTest(Op0, X86CC, DAG);
6682 DebugLoc dl = Op0.getDebugLoc();
6683 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6686 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6687 /// if it's possible.
6688 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6689 DebugLoc dl, SelectionDAG &DAG) const {
6690 SDValue Op0 = And.getOperand(0);
6691 SDValue Op1 = And.getOperand(1);
6692 if (Op0.getOpcode() == ISD::TRUNCATE)
6693 Op0 = Op0.getOperand(0);
6694 if (Op1.getOpcode() == ISD::TRUNCATE)
6695 Op1 = Op1.getOperand(0);
6698 if (Op1.getOpcode() == ISD::SHL)
6699 std::swap(Op0, Op1);
6700 if (Op0.getOpcode() == ISD::SHL) {
6701 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6702 if (And00C->getZExtValue() == 1) {
6703 // If we looked past a truncate, check that it's only truncating away
6705 unsigned BitWidth = Op0.getValueSizeInBits();
6706 unsigned AndBitWidth = And.getValueSizeInBits();
6707 if (BitWidth > AndBitWidth) {
6708 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6709 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6710 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6714 RHS = Op0.getOperand(1);
6716 } else if (Op1.getOpcode() == ISD::Constant) {
6717 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6718 SDValue AndLHS = Op0;
6719 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6720 LHS = AndLHS.getOperand(0);
6721 RHS = AndLHS.getOperand(1);
6725 if (LHS.getNode()) {
6726 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6727 // instruction. Since the shift amount is in-range-or-undefined, we know
6728 // that doing a bittest on the i32 value is ok. We extend to i32 because
6729 // the encoding for the i16 version is larger than the i32 version.
6730 // Also promote i16 to i32 for performance / code size reason.
6731 if (LHS.getValueType() == MVT::i8 ||
6732 LHS.getValueType() == MVT::i16)
6733 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6735 // If the operand types disagree, extend the shift amount to match. Since
6736 // BT ignores high bits (like shifts) we can use anyextend.
6737 if (LHS.getValueType() != RHS.getValueType())
6738 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6740 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6741 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6742 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6743 DAG.getConstant(Cond, MVT::i8), BT);
6749 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6750 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6751 SDValue Op0 = Op.getOperand(0);
6752 SDValue Op1 = Op.getOperand(1);
6753 DebugLoc dl = Op.getDebugLoc();
6754 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6756 // Optimize to BT if possible.
6757 // Lower (X & (1 << N)) == 0 to BT(X, N).
6758 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6759 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6760 if (Op0.getOpcode() == ISD::AND &&
6762 Op1.getOpcode() == ISD::Constant &&
6763 cast<ConstantSDNode>(Op1)->isNullValue() &&
6764 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6765 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6766 if (NewSetCC.getNode())
6770 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6771 if (Op0.getOpcode() == X86ISD::SETCC &&
6772 Op1.getOpcode() == ISD::Constant &&
6773 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6774 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6775 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6776 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6777 bool Invert = (CC == ISD::SETNE) ^
6778 cast<ConstantSDNode>(Op1)->isNullValue();
6780 CCode = X86::GetOppositeBranchCondition(CCode);
6781 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6782 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6785 bool isFP = Op1.getValueType().isFloatingPoint();
6786 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6787 if (X86CC == X86::COND_INVALID)
6790 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6792 // Use sbb x, x to materialize carry bit into a GPR.
6793 if (X86CC == X86::COND_B)
6794 return DAG.getNode(ISD::AND, dl, MVT::i8,
6795 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6796 DAG.getConstant(X86CC, MVT::i8), Cond),
6797 DAG.getConstant(1, MVT::i8));
6799 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6800 DAG.getConstant(X86CC, MVT::i8), Cond);
6803 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6805 SDValue Op0 = Op.getOperand(0);
6806 SDValue Op1 = Op.getOperand(1);
6807 SDValue CC = Op.getOperand(2);
6808 EVT VT = Op.getValueType();
6809 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6810 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6811 DebugLoc dl = Op.getDebugLoc();
6815 EVT VT0 = Op0.getValueType();
6816 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6817 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6820 switch (SetCCOpcode) {
6823 case ISD::SETEQ: SSECC = 0; break;
6825 case ISD::SETGT: Swap = true; // Fallthrough
6827 case ISD::SETOLT: SSECC = 1; break;
6829 case ISD::SETGE: Swap = true; // Fallthrough
6831 case ISD::SETOLE: SSECC = 2; break;
6832 case ISD::SETUO: SSECC = 3; break;
6834 case ISD::SETNE: SSECC = 4; break;
6835 case ISD::SETULE: Swap = true;
6836 case ISD::SETUGE: SSECC = 5; break;
6837 case ISD::SETULT: Swap = true;
6838 case ISD::SETUGT: SSECC = 6; break;
6839 case ISD::SETO: SSECC = 7; break;
6842 std::swap(Op0, Op1);
6844 // In the two special cases we can't handle, emit two comparisons.
6846 if (SetCCOpcode == ISD::SETUEQ) {
6848 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6849 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6850 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6852 else if (SetCCOpcode == ISD::SETONE) {
6854 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6855 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6856 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6858 llvm_unreachable("Illegal FP comparison");
6860 // Handle all other FP comparisons here.
6861 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6864 // We are handling one of the integer comparisons here. Since SSE only has
6865 // GT and EQ comparisons for integer, swapping operands and multiple
6866 // operations may be required for some comparisons.
6867 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6868 bool Swap = false, Invert = false, FlipSigns = false;
6870 switch (VT.getSimpleVT().SimpleTy) {
6873 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6875 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6877 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6878 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6881 switch (SetCCOpcode) {
6883 case ISD::SETNE: Invert = true;
6884 case ISD::SETEQ: Opc = EQOpc; break;
6885 case ISD::SETLT: Swap = true;
6886 case ISD::SETGT: Opc = GTOpc; break;
6887 case ISD::SETGE: Swap = true;
6888 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6889 case ISD::SETULT: Swap = true;
6890 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6891 case ISD::SETUGE: Swap = true;
6892 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6895 std::swap(Op0, Op1);
6897 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6898 // bits of the inputs before performing those operations.
6900 EVT EltVT = VT.getVectorElementType();
6901 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6903 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6904 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6906 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6907 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6910 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6912 // If the logical-not of the result is required, perform that now.
6914 Result = DAG.getNOT(dl, Result, VT);
6919 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6920 static bool isX86LogicalCmp(SDValue Op) {
6921 unsigned Opc = Op.getNode()->getOpcode();
6922 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6924 if (Op.getResNo() == 1 &&
6925 (Opc == X86ISD::ADD ||
6926 Opc == X86ISD::SUB ||
6927 Opc == X86ISD::SMUL ||
6928 Opc == X86ISD::UMUL ||
6929 Opc == X86ISD::INC ||
6930 Opc == X86ISD::DEC ||
6931 Opc == X86ISD::OR ||
6932 Opc == X86ISD::XOR ||
6933 Opc == X86ISD::AND))
6939 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6940 bool addTest = true;
6941 SDValue Cond = Op.getOperand(0);
6942 DebugLoc dl = Op.getDebugLoc();
6945 if (Cond.getOpcode() == ISD::SETCC) {
6946 SDValue NewCond = LowerSETCC(Cond, DAG);
6947 if (NewCond.getNode())
6951 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6952 SDValue Op1 = Op.getOperand(1);
6953 SDValue Op2 = Op.getOperand(2);
6954 if (Cond.getOpcode() == X86ISD::SETCC &&
6955 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6956 SDValue Cmp = Cond.getOperand(1);
6957 if (Cmp.getOpcode() == X86ISD::CMP) {
6958 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6959 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6960 ConstantSDNode *RHSC =
6961 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6962 if (N1C && N1C->isAllOnesValue() &&
6963 N2C && N2C->isNullValue() &&
6964 RHSC && RHSC->isNullValue()) {
6965 SDValue CmpOp0 = Cmp.getOperand(0);
6966 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6967 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6968 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6969 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6974 // Look pass (and (setcc_carry (cmp ...)), 1).
6975 if (Cond.getOpcode() == ISD::AND &&
6976 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6977 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6978 if (C && C->getAPIntValue() == 1)
6979 Cond = Cond.getOperand(0);
6982 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6983 // setting operand in place of the X86ISD::SETCC.
6984 if (Cond.getOpcode() == X86ISD::SETCC ||
6985 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6986 CC = Cond.getOperand(0);
6988 SDValue Cmp = Cond.getOperand(1);
6989 unsigned Opc = Cmp.getOpcode();
6990 EVT VT = Op.getValueType();
6992 bool IllegalFPCMov = false;
6993 if (VT.isFloatingPoint() && !VT.isVector() &&
6994 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6995 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6997 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6998 Opc == X86ISD::BT) { // FIXME
7005 // Look pass the truncate.
7006 if (Cond.getOpcode() == ISD::TRUNCATE)
7007 Cond = Cond.getOperand(0);
7009 // We know the result of AND is compared against zero. Try to match
7011 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7012 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7013 if (NewSetCC.getNode()) {
7014 CC = NewSetCC.getOperand(0);
7015 Cond = NewSetCC.getOperand(1);
7022 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7023 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7026 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7027 // condition is true.
7028 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
7029 SDValue Ops[] = { Op2, Op1, CC, Cond };
7030 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
7033 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7034 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7035 // from the AND / OR.
7036 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7037 Opc = Op.getOpcode();
7038 if (Opc != ISD::OR && Opc != ISD::AND)
7040 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7041 Op.getOperand(0).hasOneUse() &&
7042 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7043 Op.getOperand(1).hasOneUse());
7046 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7047 // 1 and that the SETCC node has a single use.
7048 static bool isXor1OfSetCC(SDValue Op) {
7049 if (Op.getOpcode() != ISD::XOR)
7051 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7052 if (N1C && N1C->getAPIntValue() == 1) {
7053 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7054 Op.getOperand(0).hasOneUse();
7059 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7060 bool addTest = true;
7061 SDValue Chain = Op.getOperand(0);
7062 SDValue Cond = Op.getOperand(1);
7063 SDValue Dest = Op.getOperand(2);
7064 DebugLoc dl = Op.getDebugLoc();
7067 if (Cond.getOpcode() == ISD::SETCC) {
7068 SDValue NewCond = LowerSETCC(Cond, DAG);
7069 if (NewCond.getNode())
7073 // FIXME: LowerXALUO doesn't handle these!!
7074 else if (Cond.getOpcode() == X86ISD::ADD ||
7075 Cond.getOpcode() == X86ISD::SUB ||
7076 Cond.getOpcode() == X86ISD::SMUL ||
7077 Cond.getOpcode() == X86ISD::UMUL)
7078 Cond = LowerXALUO(Cond, DAG);
7081 // Look pass (and (setcc_carry (cmp ...)), 1).
7082 if (Cond.getOpcode() == ISD::AND &&
7083 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7084 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7085 if (C && C->getAPIntValue() == 1)
7086 Cond = Cond.getOperand(0);
7089 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7090 // setting operand in place of the X86ISD::SETCC.
7091 if (Cond.getOpcode() == X86ISD::SETCC ||
7092 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7093 CC = Cond.getOperand(0);
7095 SDValue Cmp = Cond.getOperand(1);
7096 unsigned Opc = Cmp.getOpcode();
7097 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7098 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7102 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7106 // These can only come from an arithmetic instruction with overflow,
7107 // e.g. SADDO, UADDO.
7108 Cond = Cond.getNode()->getOperand(1);
7115 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7116 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7117 if (CondOpc == ISD::OR) {
7118 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7119 // two branches instead of an explicit OR instruction with a
7121 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7122 isX86LogicalCmp(Cmp)) {
7123 CC = Cond.getOperand(0).getOperand(0);
7124 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7125 Chain, Dest, CC, Cmp);
7126 CC = Cond.getOperand(1).getOperand(0);
7130 } else { // ISD::AND
7131 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7132 // two branches instead of an explicit AND instruction with a
7133 // separate test. However, we only do this if this block doesn't
7134 // have a fall-through edge, because this requires an explicit
7135 // jmp when the condition is false.
7136 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7137 isX86LogicalCmp(Cmp) &&
7138 Op.getNode()->hasOneUse()) {
7139 X86::CondCode CCode =
7140 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7141 CCode = X86::GetOppositeBranchCondition(CCode);
7142 CC = DAG.getConstant(CCode, MVT::i8);
7143 SDNode *User = *Op.getNode()->use_begin();
7144 // Look for an unconditional branch following this conditional branch.
7145 // We need this because we need to reverse the successors in order
7146 // to implement FCMP_OEQ.
7147 if (User->getOpcode() == ISD::BR) {
7148 SDValue FalseBB = User->getOperand(1);
7150 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
7151 assert(NewBR == User);
7155 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7156 Chain, Dest, CC, Cmp);
7157 X86::CondCode CCode =
7158 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
7159 CCode = X86::GetOppositeBranchCondition(CCode);
7160 CC = DAG.getConstant(CCode, MVT::i8);
7166 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
7167 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
7168 // It should be transformed during dag combiner except when the condition
7169 // is set by a arithmetics with overflow node.
7170 X86::CondCode CCode =
7171 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7172 CCode = X86::GetOppositeBranchCondition(CCode);
7173 CC = DAG.getConstant(CCode, MVT::i8);
7174 Cond = Cond.getOperand(0).getOperand(1);
7180 // Look pass the truncate.
7181 if (Cond.getOpcode() == ISD::TRUNCATE)
7182 Cond = Cond.getOperand(0);
7184 // We know the result of AND is compared against zero. Try to match
7186 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7187 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7188 if (NewSetCC.getNode()) {
7189 CC = NewSetCC.getOperand(0);
7190 Cond = NewSetCC.getOperand(1);
7197 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7198 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7200 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7201 Chain, Dest, CC, Cond);
7205 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
7206 // Calls to _alloca is needed to probe the stack when allocating more than 4k
7207 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
7208 // that the guard pages used by the OS virtual memory manager are allocated in
7209 // correct sequence.
7211 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
7212 SelectionDAG &DAG) const {
7213 assert(Subtarget->isTargetCygMing() &&
7214 "This should be used only on Cygwin/Mingw targets");
7215 DebugLoc dl = Op.getDebugLoc();
7218 SDValue Chain = Op.getOperand(0);
7219 SDValue Size = Op.getOperand(1);
7220 // FIXME: Ensure alignment here
7224 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
7226 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
7227 Flag = Chain.getValue(1);
7229 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
7231 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
7232 Flag = Chain.getValue(1);
7234 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
7236 SDValue Ops1[2] = { Chain.getValue(0), Chain };
7237 return DAG.getMergeValues(Ops1, 2, dl);
7240 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
7241 MachineFunction &MF = DAG.getMachineFunction();
7242 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
7244 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7245 DebugLoc dl = Op.getDebugLoc();
7247 if (!Subtarget->is64Bit()) {
7248 // vastart just stores the address of the VarArgsFrameIndex slot into the
7249 // memory location argument.
7250 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7252 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
7257 // gp_offset (0 - 6 * 8)
7258 // fp_offset (48 - 48 + 8 * 16)
7259 // overflow_arg_area (point to parameters coming in memory).
7261 SmallVector<SDValue, 8> MemOps;
7262 SDValue FIN = Op.getOperand(1);
7264 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
7265 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
7267 FIN, SV, 0, false, false, 0);
7268 MemOps.push_back(Store);
7271 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7272 FIN, DAG.getIntPtrConstant(4));
7273 Store = DAG.getStore(Op.getOperand(0), dl,
7274 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
7276 FIN, SV, 4, false, false, 0);
7277 MemOps.push_back(Store);
7279 // Store ptr to overflow_arg_area
7280 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7281 FIN, DAG.getIntPtrConstant(4));
7282 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7284 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 8,
7286 MemOps.push_back(Store);
7288 // Store ptr to reg_save_area.
7289 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7290 FIN, DAG.getIntPtrConstant(8));
7291 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
7293 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 16,
7295 MemOps.push_back(Store);
7296 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
7297 &MemOps[0], MemOps.size());
7300 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
7301 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7302 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
7304 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
7308 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
7309 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7310 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
7311 SDValue Chain = Op.getOperand(0);
7312 SDValue DstPtr = Op.getOperand(1);
7313 SDValue SrcPtr = Op.getOperand(2);
7314 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
7315 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7316 DebugLoc dl = Op.getDebugLoc();
7318 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
7319 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
7320 false, DstSV, 0, SrcSV, 0);
7324 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
7325 DebugLoc dl = Op.getDebugLoc();
7326 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7328 default: return SDValue(); // Don't custom lower most intrinsics.
7329 // Comparison intrinsics.
7330 case Intrinsic::x86_sse_comieq_ss:
7331 case Intrinsic::x86_sse_comilt_ss:
7332 case Intrinsic::x86_sse_comile_ss:
7333 case Intrinsic::x86_sse_comigt_ss:
7334 case Intrinsic::x86_sse_comige_ss:
7335 case Intrinsic::x86_sse_comineq_ss:
7336 case Intrinsic::x86_sse_ucomieq_ss:
7337 case Intrinsic::x86_sse_ucomilt_ss:
7338 case Intrinsic::x86_sse_ucomile_ss:
7339 case Intrinsic::x86_sse_ucomigt_ss:
7340 case Intrinsic::x86_sse_ucomige_ss:
7341 case Intrinsic::x86_sse_ucomineq_ss:
7342 case Intrinsic::x86_sse2_comieq_sd:
7343 case Intrinsic::x86_sse2_comilt_sd:
7344 case Intrinsic::x86_sse2_comile_sd:
7345 case Intrinsic::x86_sse2_comigt_sd:
7346 case Intrinsic::x86_sse2_comige_sd:
7347 case Intrinsic::x86_sse2_comineq_sd:
7348 case Intrinsic::x86_sse2_ucomieq_sd:
7349 case Intrinsic::x86_sse2_ucomilt_sd:
7350 case Intrinsic::x86_sse2_ucomile_sd:
7351 case Intrinsic::x86_sse2_ucomigt_sd:
7352 case Intrinsic::x86_sse2_ucomige_sd:
7353 case Intrinsic::x86_sse2_ucomineq_sd: {
7355 ISD::CondCode CC = ISD::SETCC_INVALID;
7358 case Intrinsic::x86_sse_comieq_ss:
7359 case Intrinsic::x86_sse2_comieq_sd:
7363 case Intrinsic::x86_sse_comilt_ss:
7364 case Intrinsic::x86_sse2_comilt_sd:
7368 case Intrinsic::x86_sse_comile_ss:
7369 case Intrinsic::x86_sse2_comile_sd:
7373 case Intrinsic::x86_sse_comigt_ss:
7374 case Intrinsic::x86_sse2_comigt_sd:
7378 case Intrinsic::x86_sse_comige_ss:
7379 case Intrinsic::x86_sse2_comige_sd:
7383 case Intrinsic::x86_sse_comineq_ss:
7384 case Intrinsic::x86_sse2_comineq_sd:
7388 case Intrinsic::x86_sse_ucomieq_ss:
7389 case Intrinsic::x86_sse2_ucomieq_sd:
7390 Opc = X86ISD::UCOMI;
7393 case Intrinsic::x86_sse_ucomilt_ss:
7394 case Intrinsic::x86_sse2_ucomilt_sd:
7395 Opc = X86ISD::UCOMI;
7398 case Intrinsic::x86_sse_ucomile_ss:
7399 case Intrinsic::x86_sse2_ucomile_sd:
7400 Opc = X86ISD::UCOMI;
7403 case Intrinsic::x86_sse_ucomigt_ss:
7404 case Intrinsic::x86_sse2_ucomigt_sd:
7405 Opc = X86ISD::UCOMI;
7408 case Intrinsic::x86_sse_ucomige_ss:
7409 case Intrinsic::x86_sse2_ucomige_sd:
7410 Opc = X86ISD::UCOMI;
7413 case Intrinsic::x86_sse_ucomineq_ss:
7414 case Intrinsic::x86_sse2_ucomineq_sd:
7415 Opc = X86ISD::UCOMI;
7420 SDValue LHS = Op.getOperand(1);
7421 SDValue RHS = Op.getOperand(2);
7422 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
7423 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
7424 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
7425 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7426 DAG.getConstant(X86CC, MVT::i8), Cond);
7427 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7429 // ptest and testp intrinsics. The intrinsic these come from are designed to
7430 // return an integer value, not just an instruction so lower it to the ptest
7431 // or testp pattern and a setcc for the result.
7432 case Intrinsic::x86_sse41_ptestz:
7433 case Intrinsic::x86_sse41_ptestc:
7434 case Intrinsic::x86_sse41_ptestnzc:
7435 case Intrinsic::x86_avx_ptestz_256:
7436 case Intrinsic::x86_avx_ptestc_256:
7437 case Intrinsic::x86_avx_ptestnzc_256:
7438 case Intrinsic::x86_avx_vtestz_ps:
7439 case Intrinsic::x86_avx_vtestc_ps:
7440 case Intrinsic::x86_avx_vtestnzc_ps:
7441 case Intrinsic::x86_avx_vtestz_pd:
7442 case Intrinsic::x86_avx_vtestc_pd:
7443 case Intrinsic::x86_avx_vtestnzc_pd:
7444 case Intrinsic::x86_avx_vtestz_ps_256:
7445 case Intrinsic::x86_avx_vtestc_ps_256:
7446 case Intrinsic::x86_avx_vtestnzc_ps_256:
7447 case Intrinsic::x86_avx_vtestz_pd_256:
7448 case Intrinsic::x86_avx_vtestc_pd_256:
7449 case Intrinsic::x86_avx_vtestnzc_pd_256: {
7450 bool IsTestPacked = false;
7453 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
7454 case Intrinsic::x86_avx_vtestz_ps:
7455 case Intrinsic::x86_avx_vtestz_pd:
7456 case Intrinsic::x86_avx_vtestz_ps_256:
7457 case Intrinsic::x86_avx_vtestz_pd_256:
7458 IsTestPacked = true; // Fallthrough
7459 case Intrinsic::x86_sse41_ptestz:
7460 case Intrinsic::x86_avx_ptestz_256:
7462 X86CC = X86::COND_E;
7464 case Intrinsic::x86_avx_vtestc_ps:
7465 case Intrinsic::x86_avx_vtestc_pd:
7466 case Intrinsic::x86_avx_vtestc_ps_256:
7467 case Intrinsic::x86_avx_vtestc_pd_256:
7468 IsTestPacked = true; // Fallthrough
7469 case Intrinsic::x86_sse41_ptestc:
7470 case Intrinsic::x86_avx_ptestc_256:
7472 X86CC = X86::COND_B;
7474 case Intrinsic::x86_avx_vtestnzc_ps:
7475 case Intrinsic::x86_avx_vtestnzc_pd:
7476 case Intrinsic::x86_avx_vtestnzc_ps_256:
7477 case Intrinsic::x86_avx_vtestnzc_pd_256:
7478 IsTestPacked = true; // Fallthrough
7479 case Intrinsic::x86_sse41_ptestnzc:
7480 case Intrinsic::x86_avx_ptestnzc_256:
7482 X86CC = X86::COND_A;
7486 SDValue LHS = Op.getOperand(1);
7487 SDValue RHS = Op.getOperand(2);
7488 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
7489 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
7490 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7491 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7492 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7495 // Fix vector shift instructions where the last operand is a non-immediate
7497 case Intrinsic::x86_sse2_pslli_w:
7498 case Intrinsic::x86_sse2_pslli_d:
7499 case Intrinsic::x86_sse2_pslli_q:
7500 case Intrinsic::x86_sse2_psrli_w:
7501 case Intrinsic::x86_sse2_psrli_d:
7502 case Intrinsic::x86_sse2_psrli_q:
7503 case Intrinsic::x86_sse2_psrai_w:
7504 case Intrinsic::x86_sse2_psrai_d:
7505 case Intrinsic::x86_mmx_pslli_w:
7506 case Intrinsic::x86_mmx_pslli_d:
7507 case Intrinsic::x86_mmx_pslli_q:
7508 case Intrinsic::x86_mmx_psrli_w:
7509 case Intrinsic::x86_mmx_psrli_d:
7510 case Intrinsic::x86_mmx_psrli_q:
7511 case Intrinsic::x86_mmx_psrai_w:
7512 case Intrinsic::x86_mmx_psrai_d: {
7513 SDValue ShAmt = Op.getOperand(2);
7514 if (isa<ConstantSDNode>(ShAmt))
7517 unsigned NewIntNo = 0;
7518 EVT ShAmtVT = MVT::v4i32;
7520 case Intrinsic::x86_sse2_pslli_w:
7521 NewIntNo = Intrinsic::x86_sse2_psll_w;
7523 case Intrinsic::x86_sse2_pslli_d:
7524 NewIntNo = Intrinsic::x86_sse2_psll_d;
7526 case Intrinsic::x86_sse2_pslli_q:
7527 NewIntNo = Intrinsic::x86_sse2_psll_q;
7529 case Intrinsic::x86_sse2_psrli_w:
7530 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7532 case Intrinsic::x86_sse2_psrli_d:
7533 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7535 case Intrinsic::x86_sse2_psrli_q:
7536 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7538 case Intrinsic::x86_sse2_psrai_w:
7539 NewIntNo = Intrinsic::x86_sse2_psra_w;
7541 case Intrinsic::x86_sse2_psrai_d:
7542 NewIntNo = Intrinsic::x86_sse2_psra_d;
7545 ShAmtVT = MVT::v2i32;
7547 case Intrinsic::x86_mmx_pslli_w:
7548 NewIntNo = Intrinsic::x86_mmx_psll_w;
7550 case Intrinsic::x86_mmx_pslli_d:
7551 NewIntNo = Intrinsic::x86_mmx_psll_d;
7553 case Intrinsic::x86_mmx_pslli_q:
7554 NewIntNo = Intrinsic::x86_mmx_psll_q;
7556 case Intrinsic::x86_mmx_psrli_w:
7557 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7559 case Intrinsic::x86_mmx_psrli_d:
7560 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7562 case Intrinsic::x86_mmx_psrli_q:
7563 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7565 case Intrinsic::x86_mmx_psrai_w:
7566 NewIntNo = Intrinsic::x86_mmx_psra_w;
7568 case Intrinsic::x86_mmx_psrai_d:
7569 NewIntNo = Intrinsic::x86_mmx_psra_d;
7571 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7577 // The vector shift intrinsics with scalars uses 32b shift amounts but
7578 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7582 ShOps[1] = DAG.getConstant(0, MVT::i32);
7583 if (ShAmtVT == MVT::v4i32) {
7584 ShOps[2] = DAG.getUNDEF(MVT::i32);
7585 ShOps[3] = DAG.getUNDEF(MVT::i32);
7586 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7588 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7591 EVT VT = Op.getValueType();
7592 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7593 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7594 DAG.getConstant(NewIntNo, MVT::i32),
7595 Op.getOperand(1), ShAmt);
7600 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7601 SelectionDAG &DAG) const {
7602 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7603 MFI->setReturnAddressIsTaken(true);
7605 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7606 DebugLoc dl = Op.getDebugLoc();
7609 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7611 DAG.getConstant(TD->getPointerSize(),
7612 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7613 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7614 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7616 NULL, 0, false, false, 0);
7619 // Just load the return address.
7620 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7621 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7622 RetAddrFI, NULL, 0, false, false, 0);
7625 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7626 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7627 MFI->setFrameAddressIsTaken(true);
7629 EVT VT = Op.getValueType();
7630 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7631 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7632 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7633 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7635 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7640 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7641 SelectionDAG &DAG) const {
7642 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7645 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7646 MachineFunction &MF = DAG.getMachineFunction();
7647 SDValue Chain = Op.getOperand(0);
7648 SDValue Offset = Op.getOperand(1);
7649 SDValue Handler = Op.getOperand(2);
7650 DebugLoc dl = Op.getDebugLoc();
7652 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
7653 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7655 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7657 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
7658 DAG.getIntPtrConstant(TD->getPointerSize()));
7659 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7660 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7661 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7662 MF.getRegInfo().addLiveOut(StoreAddrReg);
7664 return DAG.getNode(X86ISD::EH_RETURN, dl,
7666 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7669 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7670 SelectionDAG &DAG) const {
7671 SDValue Root = Op.getOperand(0);
7672 SDValue Trmp = Op.getOperand(1); // trampoline
7673 SDValue FPtr = Op.getOperand(2); // nested function
7674 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7675 DebugLoc dl = Op.getDebugLoc();
7677 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7679 if (Subtarget->is64Bit()) {
7680 SDValue OutChains[6];
7682 // Large code-model.
7683 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7684 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7686 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7687 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7689 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7691 // Load the pointer to the nested function into R11.
7692 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7693 SDValue Addr = Trmp;
7694 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7695 Addr, TrmpAddr, 0, false, false, 0);
7697 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7698 DAG.getConstant(2, MVT::i64));
7699 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7702 // Load the 'nest' parameter value into R10.
7703 // R10 is specified in X86CallingConv.td
7704 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7705 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7706 DAG.getConstant(10, MVT::i64));
7707 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7708 Addr, TrmpAddr, 10, false, false, 0);
7710 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7711 DAG.getConstant(12, MVT::i64));
7712 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7715 // Jump to the nested function.
7716 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7717 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7718 DAG.getConstant(20, MVT::i64));
7719 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7720 Addr, TrmpAddr, 20, false, false, 0);
7722 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7723 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7724 DAG.getConstant(22, MVT::i64));
7725 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7726 TrmpAddr, 22, false, false, 0);
7729 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7730 return DAG.getMergeValues(Ops, 2, dl);
7732 const Function *Func =
7733 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7734 CallingConv::ID CC = Func->getCallingConv();
7739 llvm_unreachable("Unsupported calling convention");
7740 case CallingConv::C:
7741 case CallingConv::X86_StdCall: {
7742 // Pass 'nest' parameter in ECX.
7743 // Must be kept in sync with X86CallingConv.td
7746 // Check that ECX wasn't needed by an 'inreg' parameter.
7747 const FunctionType *FTy = Func->getFunctionType();
7748 const AttrListPtr &Attrs = Func->getAttributes();
7750 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7751 unsigned InRegCount = 0;
7754 for (FunctionType::param_iterator I = FTy->param_begin(),
7755 E = FTy->param_end(); I != E; ++I, ++Idx)
7756 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7757 // FIXME: should only count parameters that are lowered to integers.
7758 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7760 if (InRegCount > 2) {
7761 report_fatal_error("Nest register in use - reduce number of inreg"
7767 case CallingConv::X86_FastCall:
7768 case CallingConv::X86_ThisCall:
7769 case CallingConv::Fast:
7770 // Pass 'nest' parameter in EAX.
7771 // Must be kept in sync with X86CallingConv.td
7776 SDValue OutChains[4];
7779 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7780 DAG.getConstant(10, MVT::i32));
7781 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7783 // This is storing the opcode for MOV32ri.
7784 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7785 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7786 OutChains[0] = DAG.getStore(Root, dl,
7787 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7788 Trmp, TrmpAddr, 0, false, false, 0);
7790 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7791 DAG.getConstant(1, MVT::i32));
7792 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7795 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7796 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7797 DAG.getConstant(5, MVT::i32));
7798 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7799 TrmpAddr, 5, false, false, 1);
7801 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7802 DAG.getConstant(6, MVT::i32));
7803 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7807 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7808 return DAG.getMergeValues(Ops, 2, dl);
7812 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7813 SelectionDAG &DAG) const {
7815 The rounding mode is in bits 11:10 of FPSR, and has the following
7822 FLT_ROUNDS, on the other hand, expects the following:
7829 To perform the conversion, we do:
7830 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7833 MachineFunction &MF = DAG.getMachineFunction();
7834 const TargetMachine &TM = MF.getTarget();
7835 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7836 unsigned StackAlignment = TFI.getStackAlignment();
7837 EVT VT = Op.getValueType();
7838 DebugLoc dl = Op.getDebugLoc();
7840 // Save FP Control Word to stack slot
7841 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7842 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7844 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7845 DAG.getEntryNode(), StackSlot);
7847 // Load FP Control Word from stack slot
7848 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7851 // Transform as necessary
7853 DAG.getNode(ISD::SRL, dl, MVT::i16,
7854 DAG.getNode(ISD::AND, dl, MVT::i16,
7855 CWD, DAG.getConstant(0x800, MVT::i16)),
7856 DAG.getConstant(11, MVT::i8));
7858 DAG.getNode(ISD::SRL, dl, MVT::i16,
7859 DAG.getNode(ISD::AND, dl, MVT::i16,
7860 CWD, DAG.getConstant(0x400, MVT::i16)),
7861 DAG.getConstant(9, MVT::i8));
7864 DAG.getNode(ISD::AND, dl, MVT::i16,
7865 DAG.getNode(ISD::ADD, dl, MVT::i16,
7866 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7867 DAG.getConstant(1, MVT::i16)),
7868 DAG.getConstant(3, MVT::i16));
7871 return DAG.getNode((VT.getSizeInBits() < 16 ?
7872 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7875 SDValue X86TargetLowering::LowerCTLZ(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) {
7883 // Zero extend to i32 since there is not an i8 bsr.
7885 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7888 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7889 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7890 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7892 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7895 DAG.getConstant(NumBits+NumBits-1, OpVT),
7896 DAG.getConstant(X86::COND_E, MVT::i8),
7899 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7901 // Finally xor with NumBits-1.
7902 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7905 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7909 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7910 EVT VT = Op.getValueType();
7912 unsigned NumBits = VT.getSizeInBits();
7913 DebugLoc dl = Op.getDebugLoc();
7915 Op = Op.getOperand(0);
7916 if (VT == MVT::i8) {
7918 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7921 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7922 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7923 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7925 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7928 DAG.getConstant(NumBits, OpVT),
7929 DAG.getConstant(X86::COND_E, MVT::i8),
7932 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7935 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7939 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7940 EVT VT = Op.getValueType();
7941 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7942 DebugLoc dl = Op.getDebugLoc();
7944 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7945 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7946 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7947 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7948 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7950 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7951 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7952 // return AloBlo + AloBhi + AhiBlo;
7954 SDValue A = Op.getOperand(0);
7955 SDValue B = Op.getOperand(1);
7957 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7958 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7959 A, DAG.getConstant(32, MVT::i32));
7960 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7961 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7962 B, DAG.getConstant(32, MVT::i32));
7963 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7964 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7966 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7967 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7969 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7970 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7972 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7973 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7974 AloBhi, DAG.getConstant(32, MVT::i32));
7975 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7976 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7977 AhiBlo, DAG.getConstant(32, MVT::i32));
7978 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7979 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7983 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
7984 EVT VT = Op.getValueType();
7985 DebugLoc dl = Op.getDebugLoc();
7986 SDValue R = Op.getOperand(0);
7988 LLVMContext *Context = DAG.getContext();
7990 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
7992 if (VT == MVT::v4i32) {
7993 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7994 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
7995 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
7997 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
7999 std::vector<Constant*> CV(4, CI);
8000 Constant *C = ConstantVector::get(CV);
8001 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8002 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8003 PseudoSourceValue::getConstantPool(), 0,
8006 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
8007 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, Op);
8008 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
8009 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
8011 if (VT == MVT::v16i8) {
8013 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8014 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8015 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
8017 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
8018 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
8020 std::vector<Constant*> CVM1(16, CM1);
8021 std::vector<Constant*> CVM2(16, CM2);
8022 Constant *C = ConstantVector::get(CVM1);
8023 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8024 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8025 PseudoSourceValue::getConstantPool(), 0,
8028 // r = pblendv(r, psllw(r & (char16)15, 4), a);
8029 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8030 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8031 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8032 DAG.getConstant(4, MVT::i32));
8033 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8034 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8037 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8039 C = ConstantVector::get(CVM2);
8040 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8041 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8042 PseudoSourceValue::getConstantPool(), 0, false, false, 16);
8044 // r = pblendv(r, psllw(r & (char16)63, 2), a);
8045 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8046 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8047 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8048 DAG.getConstant(2, MVT::i32));
8049 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8050 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8053 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8055 // return pblendv(r, r+r, a);
8056 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8057 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8058 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
8064 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
8065 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
8066 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
8067 // looks for this combo and may remove the "setcc" instruction if the "setcc"
8068 // has only one use.
8069 SDNode *N = Op.getNode();
8070 SDValue LHS = N->getOperand(0);
8071 SDValue RHS = N->getOperand(1);
8072 unsigned BaseOp = 0;
8074 DebugLoc dl = Op.getDebugLoc();
8076 switch (Op.getOpcode()) {
8077 default: llvm_unreachable("Unknown ovf instruction!");
8079 // A subtract of one will be selected as a INC. Note that INC doesn't
8080 // set CF, so we can't do this for UADDO.
8081 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8082 if (C->getAPIntValue() == 1) {
8083 BaseOp = X86ISD::INC;
8087 BaseOp = X86ISD::ADD;
8091 BaseOp = X86ISD::ADD;
8095 // A subtract of one will be selected as a DEC. Note that DEC doesn't
8096 // set CF, so we can't do this for USUBO.
8097 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8098 if (C->getAPIntValue() == 1) {
8099 BaseOp = X86ISD::DEC;
8103 BaseOp = X86ISD::SUB;
8107 BaseOp = X86ISD::SUB;
8111 BaseOp = X86ISD::SMUL;
8115 BaseOp = X86ISD::UMUL;
8120 // Also sets EFLAGS.
8121 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
8122 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
8125 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
8126 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
8128 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8132 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
8133 DebugLoc dl = Op.getDebugLoc();
8135 if (!Subtarget->hasSSE2()) {
8136 SDValue Chain = Op.getOperand(0);
8137 SDValue Zero = DAG.getConstant(0,
8138 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8140 DAG.getRegister(X86::ESP, MVT::i32), // Base
8141 DAG.getTargetConstant(1, MVT::i8), // Scale
8142 DAG.getRegister(0, MVT::i32), // Index
8143 DAG.getTargetConstant(0, MVT::i32), // Disp
8144 DAG.getRegister(0, MVT::i32), // Segment.
8149 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
8150 array_lengthof(Ops));
8151 return SDValue(Res, 0);
8154 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
8156 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
8158 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
8159 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
8160 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
8161 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
8163 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
8164 if (!Op1 && !Op2 && !Op3 && Op4)
8165 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
8167 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
8168 if (Op1 && !Op2 && !Op3 && !Op4)
8169 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
8171 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
8173 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
8176 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
8177 EVT T = Op.getValueType();
8178 DebugLoc dl = Op.getDebugLoc();
8181 switch(T.getSimpleVT().SimpleTy) {
8183 assert(false && "Invalid value type!");
8184 case MVT::i8: Reg = X86::AL; size = 1; break;
8185 case MVT::i16: Reg = X86::AX; size = 2; break;
8186 case MVT::i32: Reg = X86::EAX; size = 4; break;
8188 assert(Subtarget->is64Bit() && "Node not type legal!");
8189 Reg = X86::RAX; size = 8;
8192 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
8193 Op.getOperand(2), SDValue());
8194 SDValue Ops[] = { cpIn.getValue(0),
8197 DAG.getTargetConstant(size, MVT::i8),
8199 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8200 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
8202 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
8206 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
8207 SelectionDAG &DAG) const {
8208 assert(Subtarget->is64Bit() && "Result not type legalized?");
8209 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8210 SDValue TheChain = Op.getOperand(0);
8211 DebugLoc dl = Op.getDebugLoc();
8212 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8213 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
8214 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
8216 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
8217 DAG.getConstant(32, MVT::i8));
8219 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
8222 return DAG.getMergeValues(Ops, 2, dl);
8225 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
8226 SelectionDAG &DAG) const {
8227 EVT SrcVT = Op.getOperand(0).getValueType();
8228 EVT DstVT = Op.getValueType();
8229 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
8230 Subtarget->hasMMX() && !DisableMMX) &&
8231 "Unexpected custom BIT_CONVERT");
8232 assert((DstVT == MVT::i64 ||
8233 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
8234 "Unexpected custom BIT_CONVERT");
8235 // i64 <=> MMX conversions are Legal.
8236 if (SrcVT==MVT::i64 && DstVT.isVector())
8238 if (DstVT==MVT::i64 && SrcVT.isVector())
8240 // MMX <=> MMX conversions are Legal.
8241 if (SrcVT.isVector() && DstVT.isVector())
8243 // All other conversions need to be expanded.
8246 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
8247 SDNode *Node = Op.getNode();
8248 DebugLoc dl = Node->getDebugLoc();
8249 EVT T = Node->getValueType(0);
8250 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
8251 DAG.getConstant(0, T), Node->getOperand(2));
8252 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
8253 cast<AtomicSDNode>(Node)->getMemoryVT(),
8254 Node->getOperand(0),
8255 Node->getOperand(1), negOp,
8256 cast<AtomicSDNode>(Node)->getSrcValue(),
8257 cast<AtomicSDNode>(Node)->getAlignment());
8260 /// LowerOperation - Provide custom lowering hooks for some operations.
8262 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
8263 switch (Op.getOpcode()) {
8264 default: llvm_unreachable("Should not custom lower this!");
8265 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
8266 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
8267 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
8268 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
8269 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
8270 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
8271 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
8272 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
8273 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
8274 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
8275 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
8276 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
8277 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
8278 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
8279 case ISD::SHL_PARTS:
8280 case ISD::SRA_PARTS:
8281 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
8282 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
8283 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
8284 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
8285 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
8286 case ISD::FABS: return LowerFABS(Op, DAG);
8287 case ISD::FNEG: return LowerFNEG(Op, DAG);
8288 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
8289 case ISD::SETCC: return LowerSETCC(Op, DAG);
8290 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
8291 case ISD::SELECT: return LowerSELECT(Op, DAG);
8292 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
8293 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
8294 case ISD::VASTART: return LowerVASTART(Op, DAG);
8295 case ISD::VAARG: return LowerVAARG(Op, DAG);
8296 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
8297 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
8298 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
8299 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
8300 case ISD::FRAME_TO_ARGS_OFFSET:
8301 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
8302 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
8303 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
8304 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
8305 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
8306 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
8307 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
8308 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
8309 case ISD::SHL: return LowerSHL(Op, DAG);
8315 case ISD::UMULO: return LowerXALUO(Op, DAG);
8316 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
8317 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
8321 void X86TargetLowering::
8322 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
8323 SelectionDAG &DAG, unsigned NewOp) const {
8324 EVT T = Node->getValueType(0);
8325 DebugLoc dl = Node->getDebugLoc();
8326 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
8328 SDValue Chain = Node->getOperand(0);
8329 SDValue In1 = Node->getOperand(1);
8330 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8331 Node->getOperand(2), DAG.getIntPtrConstant(0));
8332 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8333 Node->getOperand(2), DAG.getIntPtrConstant(1));
8334 SDValue Ops[] = { Chain, In1, In2L, In2H };
8335 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
8337 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
8338 cast<MemSDNode>(Node)->getMemOperand());
8339 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
8340 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8341 Results.push_back(Result.getValue(2));
8344 /// ReplaceNodeResults - Replace a node with an illegal result type
8345 /// with a new node built out of custom code.
8346 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
8347 SmallVectorImpl<SDValue>&Results,
8348 SelectionDAG &DAG) const {
8349 DebugLoc dl = N->getDebugLoc();
8350 switch (N->getOpcode()) {
8352 assert(false && "Do not know how to custom type legalize this operation!");
8354 case ISD::FP_TO_SINT: {
8355 std::pair<SDValue,SDValue> Vals =
8356 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
8357 SDValue FIST = Vals.first, StackSlot = Vals.second;
8358 if (FIST.getNode() != 0) {
8359 EVT VT = N->getValueType(0);
8360 // Return a load from the stack slot.
8361 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
8366 case ISD::READCYCLECOUNTER: {
8367 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8368 SDValue TheChain = N->getOperand(0);
8369 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8370 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
8372 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
8374 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
8375 SDValue Ops[] = { eax, edx };
8376 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
8377 Results.push_back(edx.getValue(1));
8380 case ISD::ATOMIC_CMP_SWAP: {
8381 EVT T = N->getValueType(0);
8382 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
8383 SDValue cpInL, cpInH;
8384 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8385 DAG.getConstant(0, MVT::i32));
8386 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8387 DAG.getConstant(1, MVT::i32));
8388 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
8389 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
8391 SDValue swapInL, swapInH;
8392 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8393 DAG.getConstant(0, MVT::i32));
8394 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8395 DAG.getConstant(1, MVT::i32));
8396 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
8398 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
8399 swapInL.getValue(1));
8400 SDValue Ops[] = { swapInH.getValue(0),
8402 swapInH.getValue(1) };
8403 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8404 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
8405 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
8406 MVT::i32, Result.getValue(1));
8407 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
8408 MVT::i32, cpOutL.getValue(2));
8409 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
8410 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8411 Results.push_back(cpOutH.getValue(1));
8414 case ISD::ATOMIC_LOAD_ADD:
8415 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
8417 case ISD::ATOMIC_LOAD_AND:
8418 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
8420 case ISD::ATOMIC_LOAD_NAND:
8421 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
8423 case ISD::ATOMIC_LOAD_OR:
8424 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
8426 case ISD::ATOMIC_LOAD_SUB:
8427 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
8429 case ISD::ATOMIC_LOAD_XOR:
8430 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
8432 case ISD::ATOMIC_SWAP:
8433 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
8438 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
8440 default: return NULL;
8441 case X86ISD::BSF: return "X86ISD::BSF";
8442 case X86ISD::BSR: return "X86ISD::BSR";
8443 case X86ISD::SHLD: return "X86ISD::SHLD";
8444 case X86ISD::SHRD: return "X86ISD::SHRD";
8445 case X86ISD::FAND: return "X86ISD::FAND";
8446 case X86ISD::FOR: return "X86ISD::FOR";
8447 case X86ISD::FXOR: return "X86ISD::FXOR";
8448 case X86ISD::FSRL: return "X86ISD::FSRL";
8449 case X86ISD::FILD: return "X86ISD::FILD";
8450 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
8451 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
8452 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
8453 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
8454 case X86ISD::FLD: return "X86ISD::FLD";
8455 case X86ISD::FST: return "X86ISD::FST";
8456 case X86ISD::CALL: return "X86ISD::CALL";
8457 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
8458 case X86ISD::BT: return "X86ISD::BT";
8459 case X86ISD::CMP: return "X86ISD::CMP";
8460 case X86ISD::COMI: return "X86ISD::COMI";
8461 case X86ISD::UCOMI: return "X86ISD::UCOMI";
8462 case X86ISD::SETCC: return "X86ISD::SETCC";
8463 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
8464 case X86ISD::CMOV: return "X86ISD::CMOV";
8465 case X86ISD::BRCOND: return "X86ISD::BRCOND";
8466 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
8467 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
8468 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
8469 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
8470 case X86ISD::Wrapper: return "X86ISD::Wrapper";
8471 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
8472 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
8473 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
8474 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
8475 case X86ISD::PINSRB: return "X86ISD::PINSRB";
8476 case X86ISD::PINSRW: return "X86ISD::PINSRW";
8477 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
8478 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
8479 case X86ISD::FMAX: return "X86ISD::FMAX";
8480 case X86ISD::FMIN: return "X86ISD::FMIN";
8481 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
8482 case X86ISD::FRCP: return "X86ISD::FRCP";
8483 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
8484 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
8485 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
8486 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
8487 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
8488 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
8489 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
8490 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
8491 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
8492 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
8493 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
8494 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
8495 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
8496 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
8497 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
8498 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
8499 case X86ISD::VSHL: return "X86ISD::VSHL";
8500 case X86ISD::VSRL: return "X86ISD::VSRL";
8501 case X86ISD::CMPPD: return "X86ISD::CMPPD";
8502 case X86ISD::CMPPS: return "X86ISD::CMPPS";
8503 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
8504 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
8505 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
8506 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
8507 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
8508 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
8509 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
8510 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
8511 case X86ISD::ADD: return "X86ISD::ADD";
8512 case X86ISD::SUB: return "X86ISD::SUB";
8513 case X86ISD::SMUL: return "X86ISD::SMUL";
8514 case X86ISD::UMUL: return "X86ISD::UMUL";
8515 case X86ISD::INC: return "X86ISD::INC";
8516 case X86ISD::DEC: return "X86ISD::DEC";
8517 case X86ISD::OR: return "X86ISD::OR";
8518 case X86ISD::XOR: return "X86ISD::XOR";
8519 case X86ISD::AND: return "X86ISD::AND";
8520 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
8521 case X86ISD::PTEST: return "X86ISD::PTEST";
8522 case X86ISD::TESTP: return "X86ISD::TESTP";
8523 case X86ISD::PALIGN: return "X86ISD::PALIGN";
8524 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
8525 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
8526 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
8527 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
8528 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
8529 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
8530 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
8531 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
8532 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
8533 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
8534 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
8535 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
8536 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
8537 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
8538 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
8539 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
8540 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
8541 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
8542 case X86ISD::MOVSD: return "X86ISD::MOVSD";
8543 case X86ISD::MOVSS: return "X86ISD::MOVSS";
8544 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
8545 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
8546 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
8547 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
8548 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
8549 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
8550 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
8551 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
8552 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
8553 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
8554 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
8555 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
8556 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
8557 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
8561 // isLegalAddressingMode - Return true if the addressing mode represented
8562 // by AM is legal for this target, for a load/store of the specified type.
8563 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
8564 const Type *Ty) const {
8565 // X86 supports extremely general addressing modes.
8566 CodeModel::Model M = getTargetMachine().getCodeModel();
8567 Reloc::Model R = getTargetMachine().getRelocationModel();
8569 // X86 allows a sign-extended 32-bit immediate field as a displacement.
8570 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
8575 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
8577 // If a reference to this global requires an extra load, we can't fold it.
8578 if (isGlobalStubReference(GVFlags))
8581 // If BaseGV requires a register for the PIC base, we cannot also have a
8582 // BaseReg specified.
8583 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
8586 // If lower 4G is not available, then we must use rip-relative addressing.
8587 if ((M != CodeModel::Small || R != Reloc::Static) &&
8588 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
8598 // These scales always work.
8603 // These scales are formed with basereg+scalereg. Only accept if there is
8608 default: // Other stuff never works.
8616 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
8617 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
8619 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
8620 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
8621 if (NumBits1 <= NumBits2)
8626 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
8627 if (!VT1.isInteger() || !VT2.isInteger())
8629 unsigned NumBits1 = VT1.getSizeInBits();
8630 unsigned NumBits2 = VT2.getSizeInBits();
8631 if (NumBits1 <= NumBits2)
8636 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
8637 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8638 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
8641 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
8642 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8643 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
8646 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
8647 // i16 instructions are longer (0x66 prefix) and potentially slower.
8648 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
8651 /// isShuffleMaskLegal - Targets can use this to indicate that they only
8652 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
8653 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
8654 /// are assumed to be legal.
8656 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
8658 // Very little shuffling can be done for 64-bit vectors right now.
8659 if (VT.getSizeInBits() == 64)
8660 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8662 // FIXME: pshufb, blends, shifts.
8663 return (VT.getVectorNumElements() == 2 ||
8664 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8665 isMOVLMask(M, VT) ||
8666 isSHUFPMask(M, VT) ||
8667 isPSHUFDMask(M, VT) ||
8668 isPSHUFHWMask(M, VT) ||
8669 isPSHUFLWMask(M, VT) ||
8670 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8671 isUNPCKLMask(M, VT) ||
8672 isUNPCKHMask(M, VT) ||
8673 isUNPCKL_v_undef_Mask(M, VT) ||
8674 isUNPCKH_v_undef_Mask(M, VT));
8678 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8680 unsigned NumElts = VT.getVectorNumElements();
8681 // FIXME: This collection of masks seems suspect.
8684 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8685 return (isMOVLMask(Mask, VT) ||
8686 isCommutedMOVLMask(Mask, VT, true) ||
8687 isSHUFPMask(Mask, VT) ||
8688 isCommutedSHUFPMask(Mask, VT));
8693 //===----------------------------------------------------------------------===//
8694 // X86 Scheduler Hooks
8695 //===----------------------------------------------------------------------===//
8697 // private utility function
8699 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8700 MachineBasicBlock *MBB,
8707 TargetRegisterClass *RC,
8708 bool invSrc) const {
8709 // For the atomic bitwise operator, we generate
8712 // ld t1 = [bitinstr.addr]
8713 // op t2 = t1, [bitinstr.val]
8715 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8717 // fallthrough -->nextMBB
8718 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8719 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8720 MachineFunction::iterator MBBIter = MBB;
8723 /// First build the CFG
8724 MachineFunction *F = MBB->getParent();
8725 MachineBasicBlock *thisMBB = MBB;
8726 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8727 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8728 F->insert(MBBIter, newMBB);
8729 F->insert(MBBIter, nextMBB);
8731 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8732 nextMBB->splice(nextMBB->begin(), thisMBB,
8733 llvm::next(MachineBasicBlock::iterator(bInstr)),
8735 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8737 // Update thisMBB to fall through to newMBB
8738 thisMBB->addSuccessor(newMBB);
8740 // newMBB jumps to itself and fall through to nextMBB
8741 newMBB->addSuccessor(nextMBB);
8742 newMBB->addSuccessor(newMBB);
8744 // Insert instructions into newMBB based on incoming instruction
8745 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8746 "unexpected number of operands");
8747 DebugLoc dl = bInstr->getDebugLoc();
8748 MachineOperand& destOper = bInstr->getOperand(0);
8749 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8750 int numArgs = bInstr->getNumOperands() - 1;
8751 for (int i=0; i < numArgs; ++i)
8752 argOpers[i] = &bInstr->getOperand(i+1);
8754 // x86 address has 4 operands: base, index, scale, and displacement
8755 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8756 int valArgIndx = lastAddrIndx + 1;
8758 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8759 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8760 for (int i=0; i <= lastAddrIndx; ++i)
8761 (*MIB).addOperand(*argOpers[i]);
8763 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8765 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8770 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8771 assert((argOpers[valArgIndx]->isReg() ||
8772 argOpers[valArgIndx]->isImm()) &&
8774 if (argOpers[valArgIndx]->isReg())
8775 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8777 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8779 (*MIB).addOperand(*argOpers[valArgIndx]);
8781 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
8784 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8785 for (int i=0; i <= lastAddrIndx; ++i)
8786 (*MIB).addOperand(*argOpers[i]);
8788 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8789 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8790 bInstr->memoperands_end());
8792 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8796 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8798 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8802 // private utility function: 64 bit atomics on 32 bit host.
8804 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8805 MachineBasicBlock *MBB,
8810 bool invSrc) const {
8811 // For the atomic bitwise operator, we generate
8812 // thisMBB (instructions are in pairs, except cmpxchg8b)
8813 // ld t1,t2 = [bitinstr.addr]
8815 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8816 // op t5, t6 <- out1, out2, [bitinstr.val]
8817 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8818 // mov ECX, EBX <- t5, t6
8819 // mov EAX, EDX <- t1, t2
8820 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8821 // mov t3, t4 <- EAX, EDX
8823 // result in out1, out2
8824 // fallthrough -->nextMBB
8826 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8827 const unsigned LoadOpc = X86::MOV32rm;
8828 const unsigned NotOpc = X86::NOT32r;
8829 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8830 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8831 MachineFunction::iterator MBBIter = MBB;
8834 /// First build the CFG
8835 MachineFunction *F = MBB->getParent();
8836 MachineBasicBlock *thisMBB = MBB;
8837 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8838 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8839 F->insert(MBBIter, newMBB);
8840 F->insert(MBBIter, nextMBB);
8842 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8843 nextMBB->splice(nextMBB->begin(), thisMBB,
8844 llvm::next(MachineBasicBlock::iterator(bInstr)),
8846 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8848 // Update thisMBB to fall through to newMBB
8849 thisMBB->addSuccessor(newMBB);
8851 // newMBB jumps to itself and fall through to nextMBB
8852 newMBB->addSuccessor(nextMBB);
8853 newMBB->addSuccessor(newMBB);
8855 DebugLoc dl = bInstr->getDebugLoc();
8856 // Insert instructions into newMBB based on incoming instruction
8857 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8858 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
8859 "unexpected number of operands");
8860 MachineOperand& dest1Oper = bInstr->getOperand(0);
8861 MachineOperand& dest2Oper = bInstr->getOperand(1);
8862 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8863 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
8864 argOpers[i] = &bInstr->getOperand(i+2);
8866 // We use some of the operands multiple times, so conservatively just
8867 // clear any kill flags that might be present.
8868 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8869 argOpers[i]->setIsKill(false);
8872 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8873 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8875 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8876 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8877 for (int i=0; i <= lastAddrIndx; ++i)
8878 (*MIB).addOperand(*argOpers[i]);
8879 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8880 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8881 // add 4 to displacement.
8882 for (int i=0; i <= lastAddrIndx-2; ++i)
8883 (*MIB).addOperand(*argOpers[i]);
8884 MachineOperand newOp3 = *(argOpers[3]);
8886 newOp3.setImm(newOp3.getImm()+4);
8888 newOp3.setOffset(newOp3.getOffset()+4);
8889 (*MIB).addOperand(newOp3);
8890 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8892 // t3/4 are defined later, at the bottom of the loop
8893 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8894 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8895 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8896 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8897 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8898 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8900 // The subsequent operations should be using the destination registers of
8901 //the PHI instructions.
8903 t1 = F->getRegInfo().createVirtualRegister(RC);
8904 t2 = F->getRegInfo().createVirtualRegister(RC);
8905 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8906 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8908 t1 = dest1Oper.getReg();
8909 t2 = dest2Oper.getReg();
8912 int valArgIndx = lastAddrIndx + 1;
8913 assert((argOpers[valArgIndx]->isReg() ||
8914 argOpers[valArgIndx]->isImm()) &&
8916 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8917 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8918 if (argOpers[valArgIndx]->isReg())
8919 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8921 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8922 if (regOpcL != X86::MOV32rr)
8924 (*MIB).addOperand(*argOpers[valArgIndx]);
8925 assert(argOpers[valArgIndx + 1]->isReg() ==
8926 argOpers[valArgIndx]->isReg());
8927 assert(argOpers[valArgIndx + 1]->isImm() ==
8928 argOpers[valArgIndx]->isImm());
8929 if (argOpers[valArgIndx + 1]->isReg())
8930 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8932 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8933 if (regOpcH != X86::MOV32rr)
8935 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8937 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8939 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
8942 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
8944 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
8947 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8948 for (int i=0; i <= lastAddrIndx; ++i)
8949 (*MIB).addOperand(*argOpers[i]);
8951 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8952 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8953 bInstr->memoperands_end());
8955 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
8956 MIB.addReg(X86::EAX);
8957 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
8958 MIB.addReg(X86::EDX);
8961 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8963 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8967 // private utility function
8969 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8970 MachineBasicBlock *MBB,
8971 unsigned cmovOpc) const {
8972 // For the atomic min/max operator, we generate
8975 // ld t1 = [min/max.addr]
8976 // mov t2 = [min/max.val]
8978 // cmov[cond] t2 = t1
8980 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8982 // fallthrough -->nextMBB
8984 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8985 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8986 MachineFunction::iterator MBBIter = MBB;
8989 /// First build the CFG
8990 MachineFunction *F = MBB->getParent();
8991 MachineBasicBlock *thisMBB = MBB;
8992 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8993 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8994 F->insert(MBBIter, newMBB);
8995 F->insert(MBBIter, nextMBB);
8997 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8998 nextMBB->splice(nextMBB->begin(), thisMBB,
8999 llvm::next(MachineBasicBlock::iterator(mInstr)),
9001 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9003 // Update thisMBB to fall through to newMBB
9004 thisMBB->addSuccessor(newMBB);
9006 // newMBB jumps to newMBB and fall through to nextMBB
9007 newMBB->addSuccessor(nextMBB);
9008 newMBB->addSuccessor(newMBB);
9010 DebugLoc dl = mInstr->getDebugLoc();
9011 // Insert instructions into newMBB based on incoming instruction
9012 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9013 "unexpected number of operands");
9014 MachineOperand& destOper = mInstr->getOperand(0);
9015 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9016 int numArgs = mInstr->getNumOperands() - 1;
9017 for (int i=0; i < numArgs; ++i)
9018 argOpers[i] = &mInstr->getOperand(i+1);
9020 // x86 address has 4 operands: base, index, scale, and displacement
9021 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9022 int valArgIndx = lastAddrIndx + 1;
9024 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9025 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
9026 for (int i=0; i <= lastAddrIndx; ++i)
9027 (*MIB).addOperand(*argOpers[i]);
9029 // We only support register and immediate values
9030 assert((argOpers[valArgIndx]->isReg() ||
9031 argOpers[valArgIndx]->isImm()) &&
9034 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9035 if (argOpers[valArgIndx]->isReg())
9036 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
9038 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
9039 (*MIB).addOperand(*argOpers[valArgIndx]);
9041 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9044 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
9049 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9050 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
9054 // Cmp and exchange if none has modified the memory location
9055 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
9056 for (int i=0; i <= lastAddrIndx; ++i)
9057 (*MIB).addOperand(*argOpers[i]);
9059 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9060 (*MIB).setMemRefs(mInstr->memoperands_begin(),
9061 mInstr->memoperands_end());
9063 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9064 MIB.addReg(X86::EAX);
9067 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9069 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
9073 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
9074 // or XMM0_V32I8 in AVX all of this code can be replaced with that
9077 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
9078 unsigned numArgs, bool memArg) const {
9080 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
9081 "Target must have SSE4.2 or AVX features enabled");
9083 DebugLoc dl = MI->getDebugLoc();
9084 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9088 if (!Subtarget->hasAVX()) {
9090 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
9092 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
9095 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
9097 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
9100 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
9102 for (unsigned i = 0; i < numArgs; ++i) {
9103 MachineOperand &Op = MI->getOperand(i+1);
9105 if (!(Op.isReg() && Op.isImplicit()))
9109 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
9112 MI->eraseFromParent();
9118 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
9120 MachineBasicBlock *MBB) const {
9121 // Emit code to save XMM registers to the stack. The ABI says that the
9122 // number of registers to save is given in %al, so it's theoretically
9123 // possible to do an indirect jump trick to avoid saving all of them,
9124 // however this code takes a simpler approach and just executes all
9125 // of the stores if %al is non-zero. It's less code, and it's probably
9126 // easier on the hardware branch predictor, and stores aren't all that
9127 // expensive anyway.
9129 // Create the new basic blocks. One block contains all the XMM stores,
9130 // and one block is the final destination regardless of whether any
9131 // stores were performed.
9132 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9133 MachineFunction *F = MBB->getParent();
9134 MachineFunction::iterator MBBIter = MBB;
9136 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
9137 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
9138 F->insert(MBBIter, XMMSaveMBB);
9139 F->insert(MBBIter, EndMBB);
9141 // Transfer the remainder of MBB and its successor edges to EndMBB.
9142 EndMBB->splice(EndMBB->begin(), MBB,
9143 llvm::next(MachineBasicBlock::iterator(MI)),
9145 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
9147 // The original block will now fall through to the XMM save block.
9148 MBB->addSuccessor(XMMSaveMBB);
9149 // The XMMSaveMBB will fall through to the end block.
9150 XMMSaveMBB->addSuccessor(EndMBB);
9152 // Now add the instructions.
9153 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9154 DebugLoc DL = MI->getDebugLoc();
9156 unsigned CountReg = MI->getOperand(0).getReg();
9157 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
9158 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
9160 if (!Subtarget->isTargetWin64()) {
9161 // If %al is 0, branch around the XMM save block.
9162 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
9163 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
9164 MBB->addSuccessor(EndMBB);
9167 // In the XMM save block, save all the XMM argument registers.
9168 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
9169 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
9170 MachineMemOperand *MMO =
9171 F->getMachineMemOperand(
9172 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
9173 MachineMemOperand::MOStore, Offset,
9174 /*Size=*/16, /*Align=*/16);
9175 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
9176 .addFrameIndex(RegSaveFrameIndex)
9177 .addImm(/*Scale=*/1)
9178 .addReg(/*IndexReg=*/0)
9179 .addImm(/*Disp=*/Offset)
9180 .addReg(/*Segment=*/0)
9181 .addReg(MI->getOperand(i).getReg())
9182 .addMemOperand(MMO);
9185 MI->eraseFromParent(); // The pseudo instruction is gone now.
9191 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
9192 MachineBasicBlock *BB) const {
9193 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9194 DebugLoc DL = MI->getDebugLoc();
9196 // To "insert" a SELECT_CC instruction, we actually have to insert the
9197 // diamond control-flow pattern. The incoming instruction knows the
9198 // destination vreg to set, the condition code register to branch on, the
9199 // true/false values to select between, and a branch opcode to use.
9200 const BasicBlock *LLVM_BB = BB->getBasicBlock();
9201 MachineFunction::iterator It = BB;
9207 // cmpTY ccX, r1, r2
9209 // fallthrough --> copy0MBB
9210 MachineBasicBlock *thisMBB = BB;
9211 MachineFunction *F = BB->getParent();
9212 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
9213 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
9214 F->insert(It, copy0MBB);
9215 F->insert(It, sinkMBB);
9217 // If the EFLAGS register isn't dead in the terminator, then claim that it's
9218 // live into the sink and copy blocks.
9219 const MachineFunction *MF = BB->getParent();
9220 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
9221 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
9223 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
9224 const MachineOperand &MO = MI->getOperand(I);
9225 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
9226 unsigned Reg = MO.getReg();
9227 if (Reg != X86::EFLAGS) continue;
9228 copy0MBB->addLiveIn(Reg);
9229 sinkMBB->addLiveIn(Reg);
9232 // Transfer the remainder of BB and its successor edges to sinkMBB.
9233 sinkMBB->splice(sinkMBB->begin(), BB,
9234 llvm::next(MachineBasicBlock::iterator(MI)),
9236 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
9238 // Add the true and fallthrough blocks as its successors.
9239 BB->addSuccessor(copy0MBB);
9240 BB->addSuccessor(sinkMBB);
9242 // Create the conditional branch instruction.
9244 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
9245 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
9248 // %FalseValue = ...
9249 // # fallthrough to sinkMBB
9250 copy0MBB->addSuccessor(sinkMBB);
9253 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
9255 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
9256 TII->get(X86::PHI), MI->getOperand(0).getReg())
9257 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
9258 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
9260 MI->eraseFromParent(); // The pseudo instruction is gone now.
9265 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
9266 MachineBasicBlock *BB) const {
9267 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9268 DebugLoc DL = MI->getDebugLoc();
9270 // The lowering is pretty easy: we're just emitting the call to _alloca. The
9271 // non-trivial part is impdef of ESP.
9272 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
9275 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
9276 .addExternalSymbol("_alloca")
9277 .addReg(X86::EAX, RegState::Implicit)
9278 .addReg(X86::ESP, RegState::Implicit)
9279 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
9280 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
9281 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
9283 MI->eraseFromParent(); // The pseudo instruction is gone now.
9288 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
9289 MachineBasicBlock *BB) const {
9290 // This is pretty easy. We're taking the value that we received from
9291 // our load from the relocation, sticking it in either RDI (x86-64)
9292 // or EAX and doing an indirect call. The return value will then
9293 // be in the normal return register.
9294 const X86InstrInfo *TII
9295 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
9296 DebugLoc DL = MI->getDebugLoc();
9297 MachineFunction *F = BB->getParent();
9298 bool IsWin64 = Subtarget->isTargetWin64();
9300 assert(MI->getOperand(3).isGlobal() && "This should be a global");
9302 if (Subtarget->is64Bit()) {
9303 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9304 TII->get(X86::MOV64rm), X86::RDI)
9306 .addImm(0).addReg(0)
9307 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9308 MI->getOperand(3).getTargetFlags())
9310 MIB = BuildMI(*BB, MI, DL, TII->get(IsWin64 ? X86::WINCALL64m : X86::CALL64m));
9311 addDirectMem(MIB, X86::RDI);
9312 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
9313 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9314 TII->get(X86::MOV32rm), X86::EAX)
9316 .addImm(0).addReg(0)
9317 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9318 MI->getOperand(3).getTargetFlags())
9320 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9321 addDirectMem(MIB, X86::EAX);
9323 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9324 TII->get(X86::MOV32rm), X86::EAX)
9325 .addReg(TII->getGlobalBaseReg(F))
9326 .addImm(0).addReg(0)
9327 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9328 MI->getOperand(3).getTargetFlags())
9330 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9331 addDirectMem(MIB, X86::EAX);
9334 MI->eraseFromParent(); // The pseudo instruction is gone now.
9339 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
9340 MachineBasicBlock *BB) const {
9341 switch (MI->getOpcode()) {
9342 default: assert(false && "Unexpected instr type to insert");
9343 case X86::MINGW_ALLOCA:
9344 return EmitLoweredMingwAlloca(MI, BB);
9345 case X86::TLSCall_32:
9346 case X86::TLSCall_64:
9347 return EmitLoweredTLSCall(MI, BB);
9349 case X86::CMOV_V1I64:
9350 case X86::CMOV_FR32:
9351 case X86::CMOV_FR64:
9352 case X86::CMOV_V4F32:
9353 case X86::CMOV_V2F64:
9354 case X86::CMOV_V2I64:
9355 case X86::CMOV_GR16:
9356 case X86::CMOV_GR32:
9357 case X86::CMOV_RFP32:
9358 case X86::CMOV_RFP64:
9359 case X86::CMOV_RFP80:
9360 return EmitLoweredSelect(MI, BB);
9362 case X86::FP32_TO_INT16_IN_MEM:
9363 case X86::FP32_TO_INT32_IN_MEM:
9364 case X86::FP32_TO_INT64_IN_MEM:
9365 case X86::FP64_TO_INT16_IN_MEM:
9366 case X86::FP64_TO_INT32_IN_MEM:
9367 case X86::FP64_TO_INT64_IN_MEM:
9368 case X86::FP80_TO_INT16_IN_MEM:
9369 case X86::FP80_TO_INT32_IN_MEM:
9370 case X86::FP80_TO_INT64_IN_MEM: {
9371 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9372 DebugLoc DL = MI->getDebugLoc();
9374 // Change the floating point control register to use "round towards zero"
9375 // mode when truncating to an integer value.
9376 MachineFunction *F = BB->getParent();
9377 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
9378 addFrameReference(BuildMI(*BB, MI, DL,
9379 TII->get(X86::FNSTCW16m)), CWFrameIdx);
9381 // Load the old value of the high byte of the control word...
9383 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
9384 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
9387 // Set the high part to be round to zero...
9388 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
9391 // Reload the modified control word now...
9392 addFrameReference(BuildMI(*BB, MI, DL,
9393 TII->get(X86::FLDCW16m)), CWFrameIdx);
9395 // Restore the memory image of control word to original value
9396 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
9399 // Get the X86 opcode to use.
9401 switch (MI->getOpcode()) {
9402 default: llvm_unreachable("illegal opcode!");
9403 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
9404 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
9405 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
9406 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
9407 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
9408 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
9409 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
9410 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
9411 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
9415 MachineOperand &Op = MI->getOperand(0);
9417 AM.BaseType = X86AddressMode::RegBase;
9418 AM.Base.Reg = Op.getReg();
9420 AM.BaseType = X86AddressMode::FrameIndexBase;
9421 AM.Base.FrameIndex = Op.getIndex();
9423 Op = MI->getOperand(1);
9425 AM.Scale = Op.getImm();
9426 Op = MI->getOperand(2);
9428 AM.IndexReg = Op.getImm();
9429 Op = MI->getOperand(3);
9430 if (Op.isGlobal()) {
9431 AM.GV = Op.getGlobal();
9433 AM.Disp = Op.getImm();
9435 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
9436 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
9438 // Reload the original control word now.
9439 addFrameReference(BuildMI(*BB, MI, DL,
9440 TII->get(X86::FLDCW16m)), CWFrameIdx);
9442 MI->eraseFromParent(); // The pseudo instruction is gone now.
9445 // String/text processing lowering.
9446 case X86::PCMPISTRM128REG:
9447 case X86::VPCMPISTRM128REG:
9448 return EmitPCMP(MI, BB, 3, false /* in-mem */);
9449 case X86::PCMPISTRM128MEM:
9450 case X86::VPCMPISTRM128MEM:
9451 return EmitPCMP(MI, BB, 3, true /* in-mem */);
9452 case X86::PCMPESTRM128REG:
9453 case X86::VPCMPESTRM128REG:
9454 return EmitPCMP(MI, BB, 5, false /* in mem */);
9455 case X86::PCMPESTRM128MEM:
9456 case X86::VPCMPESTRM128MEM:
9457 return EmitPCMP(MI, BB, 5, true /* in mem */);
9460 case X86::ATOMAND32:
9461 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9462 X86::AND32ri, X86::MOV32rm,
9464 X86::NOT32r, X86::EAX,
9465 X86::GR32RegisterClass);
9467 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
9468 X86::OR32ri, X86::MOV32rm,
9470 X86::NOT32r, X86::EAX,
9471 X86::GR32RegisterClass);
9472 case X86::ATOMXOR32:
9473 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
9474 X86::XOR32ri, X86::MOV32rm,
9476 X86::NOT32r, X86::EAX,
9477 X86::GR32RegisterClass);
9478 case X86::ATOMNAND32:
9479 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9480 X86::AND32ri, X86::MOV32rm,
9482 X86::NOT32r, X86::EAX,
9483 X86::GR32RegisterClass, true);
9484 case X86::ATOMMIN32:
9485 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
9486 case X86::ATOMMAX32:
9487 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
9488 case X86::ATOMUMIN32:
9489 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
9490 case X86::ATOMUMAX32:
9491 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
9493 case X86::ATOMAND16:
9494 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9495 X86::AND16ri, X86::MOV16rm,
9497 X86::NOT16r, X86::AX,
9498 X86::GR16RegisterClass);
9500 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
9501 X86::OR16ri, X86::MOV16rm,
9503 X86::NOT16r, X86::AX,
9504 X86::GR16RegisterClass);
9505 case X86::ATOMXOR16:
9506 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
9507 X86::XOR16ri, X86::MOV16rm,
9509 X86::NOT16r, X86::AX,
9510 X86::GR16RegisterClass);
9511 case X86::ATOMNAND16:
9512 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9513 X86::AND16ri, X86::MOV16rm,
9515 X86::NOT16r, X86::AX,
9516 X86::GR16RegisterClass, true);
9517 case X86::ATOMMIN16:
9518 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
9519 case X86::ATOMMAX16:
9520 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
9521 case X86::ATOMUMIN16:
9522 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
9523 case X86::ATOMUMAX16:
9524 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
9527 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9528 X86::AND8ri, X86::MOV8rm,
9530 X86::NOT8r, X86::AL,
9531 X86::GR8RegisterClass);
9533 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
9534 X86::OR8ri, X86::MOV8rm,
9536 X86::NOT8r, X86::AL,
9537 X86::GR8RegisterClass);
9539 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
9540 X86::XOR8ri, X86::MOV8rm,
9542 X86::NOT8r, X86::AL,
9543 X86::GR8RegisterClass);
9544 case X86::ATOMNAND8:
9545 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9546 X86::AND8ri, X86::MOV8rm,
9548 X86::NOT8r, X86::AL,
9549 X86::GR8RegisterClass, true);
9550 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
9551 // This group is for 64-bit host.
9552 case X86::ATOMAND64:
9553 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9554 X86::AND64ri32, X86::MOV64rm,
9556 X86::NOT64r, X86::RAX,
9557 X86::GR64RegisterClass);
9559 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
9560 X86::OR64ri32, X86::MOV64rm,
9562 X86::NOT64r, X86::RAX,
9563 X86::GR64RegisterClass);
9564 case X86::ATOMXOR64:
9565 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
9566 X86::XOR64ri32, X86::MOV64rm,
9568 X86::NOT64r, X86::RAX,
9569 X86::GR64RegisterClass);
9570 case X86::ATOMNAND64:
9571 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9572 X86::AND64ri32, X86::MOV64rm,
9574 X86::NOT64r, X86::RAX,
9575 X86::GR64RegisterClass, true);
9576 case X86::ATOMMIN64:
9577 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
9578 case X86::ATOMMAX64:
9579 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
9580 case X86::ATOMUMIN64:
9581 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
9582 case X86::ATOMUMAX64:
9583 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
9585 // This group does 64-bit operations on a 32-bit host.
9586 case X86::ATOMAND6432:
9587 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9588 X86::AND32rr, X86::AND32rr,
9589 X86::AND32ri, X86::AND32ri,
9591 case X86::ATOMOR6432:
9592 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9593 X86::OR32rr, X86::OR32rr,
9594 X86::OR32ri, X86::OR32ri,
9596 case X86::ATOMXOR6432:
9597 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9598 X86::XOR32rr, X86::XOR32rr,
9599 X86::XOR32ri, X86::XOR32ri,
9601 case X86::ATOMNAND6432:
9602 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9603 X86::AND32rr, X86::AND32rr,
9604 X86::AND32ri, X86::AND32ri,
9606 case X86::ATOMADD6432:
9607 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9608 X86::ADD32rr, X86::ADC32rr,
9609 X86::ADD32ri, X86::ADC32ri,
9611 case X86::ATOMSUB6432:
9612 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9613 X86::SUB32rr, X86::SBB32rr,
9614 X86::SUB32ri, X86::SBB32ri,
9616 case X86::ATOMSWAP6432:
9617 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9618 X86::MOV32rr, X86::MOV32rr,
9619 X86::MOV32ri, X86::MOV32ri,
9621 case X86::VASTART_SAVE_XMM_REGS:
9622 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
9626 //===----------------------------------------------------------------------===//
9627 // X86 Optimization Hooks
9628 //===----------------------------------------------------------------------===//
9630 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9634 const SelectionDAG &DAG,
9635 unsigned Depth) const {
9636 unsigned Opc = Op.getOpcode();
9637 assert((Opc >= ISD::BUILTIN_OP_END ||
9638 Opc == ISD::INTRINSIC_WO_CHAIN ||
9639 Opc == ISD::INTRINSIC_W_CHAIN ||
9640 Opc == ISD::INTRINSIC_VOID) &&
9641 "Should use MaskedValueIsZero if you don't know whether Op"
9642 " is a target node!");
9644 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
9656 // These nodes' second result is a boolean.
9657 if (Op.getResNo() == 0)
9661 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
9662 Mask.getBitWidth() - 1);
9667 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
9668 /// node is a GlobalAddress + offset.
9669 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
9670 const GlobalValue* &GA,
9671 int64_t &Offset) const {
9672 if (N->getOpcode() == X86ISD::Wrapper) {
9673 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
9674 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
9675 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
9679 return TargetLowering::isGAPlusOffset(N, GA, Offset);
9682 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
9683 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
9684 /// if the load addresses are consecutive, non-overlapping, and in the right
9686 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
9687 const TargetLowering &TLI) {
9688 DebugLoc dl = N->getDebugLoc();
9689 EVT VT = N->getValueType(0);
9691 if (VT.getSizeInBits() != 128)
9694 SmallVector<SDValue, 16> Elts;
9695 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
9696 Elts.push_back(getShuffleScalarElt(N, i, DAG));
9698 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
9701 /// PerformShuffleCombine - Detect vector gather/scatter index generation
9702 /// and convert it from being a bunch of shuffles and extracts to a simple
9703 /// store and scalar loads to extract the elements.
9704 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
9705 const TargetLowering &TLI) {
9706 SDValue InputVector = N->getOperand(0);
9708 // Only operate on vectors of 4 elements, where the alternative shuffling
9709 // gets to be more expensive.
9710 if (InputVector.getValueType() != MVT::v4i32)
9713 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
9714 // single use which is a sign-extend or zero-extend, and all elements are
9716 SmallVector<SDNode *, 4> Uses;
9717 unsigned ExtractedElements = 0;
9718 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
9719 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
9720 if (UI.getUse().getResNo() != InputVector.getResNo())
9723 SDNode *Extract = *UI;
9724 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9727 if (Extract->getValueType(0) != MVT::i32)
9729 if (!Extract->hasOneUse())
9731 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9732 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9734 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9737 // Record which element was extracted.
9738 ExtractedElements |=
9739 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9741 Uses.push_back(Extract);
9744 // If not all the elements were used, this may not be worthwhile.
9745 if (ExtractedElements != 15)
9748 // Ok, we've now decided to do the transformation.
9749 DebugLoc dl = InputVector.getDebugLoc();
9751 // Store the value to a temporary stack slot.
9752 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9753 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL,
9754 0, false, false, 0);
9756 // Replace each use (extract) with a load of the appropriate element.
9757 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9758 UE = Uses.end(); UI != UE; ++UI) {
9759 SDNode *Extract = *UI;
9761 // Compute the element's address.
9762 SDValue Idx = Extract->getOperand(1);
9764 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9765 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9766 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9768 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
9769 OffsetVal, StackPtr);
9772 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
9773 ScalarAddr, NULL, 0, false, false, 0);
9775 // Replace the exact with the load.
9776 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9779 // The replacement was made in place; don't return anything.
9783 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9784 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9785 const X86Subtarget *Subtarget) {
9786 DebugLoc DL = N->getDebugLoc();
9787 SDValue Cond = N->getOperand(0);
9788 // Get the LHS/RHS of the select.
9789 SDValue LHS = N->getOperand(1);
9790 SDValue RHS = N->getOperand(2);
9792 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9793 // instructions match the semantics of the common C idiom x<y?x:y but not
9794 // x<=y?x:y, because of how they handle negative zero (which can be
9795 // ignored in unsafe-math mode).
9796 if (Subtarget->hasSSE2() &&
9797 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9798 Cond.getOpcode() == ISD::SETCC) {
9799 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9801 unsigned Opcode = 0;
9802 // Check for x CC y ? x : y.
9803 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9804 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9808 // Converting this to a min would handle NaNs incorrectly, and swapping
9809 // the operands would cause it to handle comparisons between positive
9810 // and negative zero incorrectly.
9811 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9812 if (!UnsafeFPMath &&
9813 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9815 std::swap(LHS, RHS);
9817 Opcode = X86ISD::FMIN;
9820 // Converting this to a min would handle comparisons between positive
9821 // and negative zero incorrectly.
9822 if (!UnsafeFPMath &&
9823 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9825 Opcode = X86ISD::FMIN;
9828 // Converting this to a min would handle both negative zeros and NaNs
9829 // incorrectly, but we can swap the operands to fix both.
9830 std::swap(LHS, RHS);
9834 Opcode = X86ISD::FMIN;
9838 // Converting this to a max would handle comparisons between positive
9839 // and negative zero incorrectly.
9840 if (!UnsafeFPMath &&
9841 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9843 Opcode = X86ISD::FMAX;
9846 // Converting this to a max would handle NaNs incorrectly, and swapping
9847 // the operands would cause it to handle comparisons between positive
9848 // and negative zero incorrectly.
9849 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9850 if (!UnsafeFPMath &&
9851 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9853 std::swap(LHS, RHS);
9855 Opcode = X86ISD::FMAX;
9858 // Converting this to a max 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::FMAX;
9867 // Check for x CC y ? y : x -- a min/max with reversed arms.
9868 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9869 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9873 // Converting this to a min would handle comparisons between positive
9874 // and negative zero incorrectly, and swapping the operands would
9875 // cause it to handle NaNs incorrectly.
9876 if (!UnsafeFPMath &&
9877 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9878 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9880 std::swap(LHS, RHS);
9882 Opcode = X86ISD::FMIN;
9885 // Converting this to a min would handle NaNs incorrectly.
9886 if (!UnsafeFPMath &&
9887 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9889 Opcode = X86ISD::FMIN;
9892 // Converting this to a min would handle both negative zeros and NaNs
9893 // incorrectly, but we can swap the operands to fix both.
9894 std::swap(LHS, RHS);
9898 Opcode = X86ISD::FMIN;
9902 // Converting this to a max would handle NaNs incorrectly.
9903 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9905 Opcode = X86ISD::FMAX;
9908 // Converting this to a max would handle comparisons between positive
9909 // and negative zero incorrectly, and swapping the operands would
9910 // cause it to handle NaNs incorrectly.
9911 if (!UnsafeFPMath &&
9912 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9913 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9915 std::swap(LHS, RHS);
9917 Opcode = X86ISD::FMAX;
9920 // Converting this to a max would handle both negative zeros and NaNs
9921 // incorrectly, but we can swap the operands to fix both.
9922 std::swap(LHS, RHS);
9926 Opcode = X86ISD::FMAX;
9932 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9935 // If this is a select between two integer constants, try to do some
9937 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9938 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9939 // Don't do this for crazy integer types.
9940 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9941 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9942 // so that TrueC (the true value) is larger than FalseC.
9943 bool NeedsCondInvert = false;
9945 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9946 // Efficiently invertible.
9947 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9948 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9949 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9950 NeedsCondInvert = true;
9951 std::swap(TrueC, FalseC);
9954 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9955 if (FalseC->getAPIntValue() == 0 &&
9956 TrueC->getAPIntValue().isPowerOf2()) {
9957 if (NeedsCondInvert) // Invert the condition if needed.
9958 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9959 DAG.getConstant(1, Cond.getValueType()));
9961 // Zero extend the condition if needed.
9962 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9964 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9965 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9966 DAG.getConstant(ShAmt, MVT::i8));
9969 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9970 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9971 if (NeedsCondInvert) // Invert the condition if needed.
9972 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9973 DAG.getConstant(1, Cond.getValueType()));
9975 // Zero extend the condition if needed.
9976 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9977 FalseC->getValueType(0), Cond);
9978 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9979 SDValue(FalseC, 0));
9982 // Optimize cases that will turn into an LEA instruction. This requires
9983 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9984 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9985 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9986 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9988 bool isFastMultiplier = false;
9990 switch ((unsigned char)Diff) {
9992 case 1: // result = add base, cond
9993 case 2: // result = lea base( , cond*2)
9994 case 3: // result = lea base(cond, cond*2)
9995 case 4: // result = lea base( , cond*4)
9996 case 5: // result = lea base(cond, cond*4)
9997 case 8: // result = lea base( , cond*8)
9998 case 9: // result = lea base(cond, cond*8)
9999 isFastMultiplier = true;
10004 if (isFastMultiplier) {
10005 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10006 if (NeedsCondInvert) // Invert the condition if needed.
10007 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10008 DAG.getConstant(1, Cond.getValueType()));
10010 // Zero extend the condition if needed.
10011 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10013 // Scale the condition by the difference.
10015 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10016 DAG.getConstant(Diff, Cond.getValueType()));
10018 // Add the base if non-zero.
10019 if (FalseC->getAPIntValue() != 0)
10020 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10021 SDValue(FalseC, 0));
10031 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
10032 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
10033 TargetLowering::DAGCombinerInfo &DCI) {
10034 DebugLoc DL = N->getDebugLoc();
10036 // If the flag operand isn't dead, don't touch this CMOV.
10037 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
10040 // If this is a select between two integer constants, try to do some
10041 // optimizations. Note that the operands are ordered the opposite of SELECT
10043 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
10044 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
10045 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
10046 // larger than FalseC (the false value).
10047 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
10049 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
10050 CC = X86::GetOppositeBranchCondition(CC);
10051 std::swap(TrueC, FalseC);
10054 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
10055 // This is efficient for any integer data type (including i8/i16) and
10057 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
10058 SDValue Cond = N->getOperand(3);
10059 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10060 DAG.getConstant(CC, MVT::i8), Cond);
10062 // Zero extend the condition if needed.
10063 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
10065 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10066 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
10067 DAG.getConstant(ShAmt, MVT::i8));
10068 if (N->getNumValues() == 2) // Dead flag value?
10069 return DCI.CombineTo(N, Cond, SDValue());
10073 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
10074 // for any integer data type, including i8/i16.
10075 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10076 SDValue Cond = N->getOperand(3);
10077 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10078 DAG.getConstant(CC, MVT::i8), Cond);
10080 // Zero extend the condition if needed.
10081 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10082 FalseC->getValueType(0), Cond);
10083 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10084 SDValue(FalseC, 0));
10086 if (N->getNumValues() == 2) // Dead flag value?
10087 return DCI.CombineTo(N, Cond, SDValue());
10091 // Optimize cases that will turn into an LEA instruction. This requires
10092 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10093 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10094 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10095 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10097 bool isFastMultiplier = false;
10099 switch ((unsigned char)Diff) {
10101 case 1: // result = add base, cond
10102 case 2: // result = lea base( , cond*2)
10103 case 3: // result = lea base(cond, cond*2)
10104 case 4: // result = lea base( , cond*4)
10105 case 5: // result = lea base(cond, cond*4)
10106 case 8: // result = lea base( , cond*8)
10107 case 9: // result = lea base(cond, cond*8)
10108 isFastMultiplier = true;
10113 if (isFastMultiplier) {
10114 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10115 SDValue Cond = N->getOperand(3);
10116 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10117 DAG.getConstant(CC, MVT::i8), Cond);
10118 // Zero extend the condition if needed.
10119 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10121 // Scale the condition by the difference.
10123 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10124 DAG.getConstant(Diff, Cond.getValueType()));
10126 // Add the base if non-zero.
10127 if (FalseC->getAPIntValue() != 0)
10128 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10129 SDValue(FalseC, 0));
10130 if (N->getNumValues() == 2) // Dead flag value?
10131 return DCI.CombineTo(N, Cond, SDValue());
10141 /// PerformMulCombine - Optimize a single multiply with constant into two
10142 /// in order to implement it with two cheaper instructions, e.g.
10143 /// LEA + SHL, LEA + LEA.
10144 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
10145 TargetLowering::DAGCombinerInfo &DCI) {
10146 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
10149 EVT VT = N->getValueType(0);
10150 if (VT != MVT::i64)
10153 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
10156 uint64_t MulAmt = C->getZExtValue();
10157 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
10160 uint64_t MulAmt1 = 0;
10161 uint64_t MulAmt2 = 0;
10162 if ((MulAmt % 9) == 0) {
10164 MulAmt2 = MulAmt / 9;
10165 } else if ((MulAmt % 5) == 0) {
10167 MulAmt2 = MulAmt / 5;
10168 } else if ((MulAmt % 3) == 0) {
10170 MulAmt2 = MulAmt / 3;
10173 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
10174 DebugLoc DL = N->getDebugLoc();
10176 if (isPowerOf2_64(MulAmt2) &&
10177 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
10178 // If second multiplifer is pow2, issue it first. We want the multiply by
10179 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
10181 std::swap(MulAmt1, MulAmt2);
10184 if (isPowerOf2_64(MulAmt1))
10185 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
10186 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
10188 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
10189 DAG.getConstant(MulAmt1, VT));
10191 if (isPowerOf2_64(MulAmt2))
10192 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
10193 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
10195 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
10196 DAG.getConstant(MulAmt2, VT));
10198 // Do not add new nodes to DAG combiner worklist.
10199 DCI.CombineTo(N, NewMul, false);
10204 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
10205 SDValue N0 = N->getOperand(0);
10206 SDValue N1 = N->getOperand(1);
10207 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
10208 EVT VT = N0.getValueType();
10210 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
10211 // since the result of setcc_c is all zero's or all ones.
10212 if (N1C && N0.getOpcode() == ISD::AND &&
10213 N0.getOperand(1).getOpcode() == ISD::Constant) {
10214 SDValue N00 = N0.getOperand(0);
10215 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
10216 ((N00.getOpcode() == ISD::ANY_EXTEND ||
10217 N00.getOpcode() == ISD::ZERO_EXTEND) &&
10218 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
10219 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
10220 APInt ShAmt = N1C->getAPIntValue();
10221 Mask = Mask.shl(ShAmt);
10223 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
10224 N00, DAG.getConstant(Mask, VT));
10231 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
10233 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
10234 const X86Subtarget *Subtarget) {
10235 EVT VT = N->getValueType(0);
10236 if (!VT.isVector() && VT.isInteger() &&
10237 N->getOpcode() == ISD::SHL)
10238 return PerformSHLCombine(N, DAG);
10240 // On X86 with SSE2 support, we can transform this to a vector shift if
10241 // all elements are shifted by the same amount. We can't do this in legalize
10242 // because the a constant vector is typically transformed to a constant pool
10243 // so we have no knowledge of the shift amount.
10244 if (!Subtarget->hasSSE2())
10247 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
10250 SDValue ShAmtOp = N->getOperand(1);
10251 EVT EltVT = VT.getVectorElementType();
10252 DebugLoc DL = N->getDebugLoc();
10253 SDValue BaseShAmt = SDValue();
10254 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
10255 unsigned NumElts = VT.getVectorNumElements();
10257 for (; i != NumElts; ++i) {
10258 SDValue Arg = ShAmtOp.getOperand(i);
10259 if (Arg.getOpcode() == ISD::UNDEF) continue;
10263 for (; i != NumElts; ++i) {
10264 SDValue Arg = ShAmtOp.getOperand(i);
10265 if (Arg.getOpcode() == ISD::UNDEF) continue;
10266 if (Arg != BaseShAmt) {
10270 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
10271 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
10272 SDValue InVec = ShAmtOp.getOperand(0);
10273 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
10274 unsigned NumElts = InVec.getValueType().getVectorNumElements();
10276 for (; i != NumElts; ++i) {
10277 SDValue Arg = InVec.getOperand(i);
10278 if (Arg.getOpcode() == ISD::UNDEF) continue;
10282 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
10283 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
10284 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
10285 if (C->getZExtValue() == SplatIdx)
10286 BaseShAmt = InVec.getOperand(1);
10289 if (BaseShAmt.getNode() == 0)
10290 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
10291 DAG.getIntPtrConstant(0));
10295 // The shift amount is an i32.
10296 if (EltVT.bitsGT(MVT::i32))
10297 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
10298 else if (EltVT.bitsLT(MVT::i32))
10299 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
10301 // The shift amount is identical so we can do a vector shift.
10302 SDValue ValOp = N->getOperand(0);
10303 switch (N->getOpcode()) {
10305 llvm_unreachable("Unknown shift opcode!");
10308 if (VT == MVT::v2i64)
10309 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10310 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
10312 if (VT == MVT::v4i32)
10313 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10314 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
10316 if (VT == MVT::v8i16)
10317 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10318 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
10322 if (VT == MVT::v4i32)
10323 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10324 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
10326 if (VT == MVT::v8i16)
10327 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10328 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
10332 if (VT == MVT::v2i64)
10333 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10334 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
10336 if (VT == MVT::v4i32)
10337 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10338 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
10340 if (VT == MVT::v8i16)
10341 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10342 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
10349 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
10350 TargetLowering::DAGCombinerInfo &DCI,
10351 const X86Subtarget *Subtarget) {
10352 if (DCI.isBeforeLegalizeOps())
10355 EVT VT = N->getValueType(0);
10356 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
10359 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
10360 SDValue N0 = N->getOperand(0);
10361 SDValue N1 = N->getOperand(1);
10362 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
10364 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
10366 if (!N0.hasOneUse() || !N1.hasOneUse())
10369 SDValue ShAmt0 = N0.getOperand(1);
10370 if (ShAmt0.getValueType() != MVT::i8)
10372 SDValue ShAmt1 = N1.getOperand(1);
10373 if (ShAmt1.getValueType() != MVT::i8)
10375 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
10376 ShAmt0 = ShAmt0.getOperand(0);
10377 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
10378 ShAmt1 = ShAmt1.getOperand(0);
10380 DebugLoc DL = N->getDebugLoc();
10381 unsigned Opc = X86ISD::SHLD;
10382 SDValue Op0 = N0.getOperand(0);
10383 SDValue Op1 = N1.getOperand(0);
10384 if (ShAmt0.getOpcode() == ISD::SUB) {
10385 Opc = X86ISD::SHRD;
10386 std::swap(Op0, Op1);
10387 std::swap(ShAmt0, ShAmt1);
10390 unsigned Bits = VT.getSizeInBits();
10391 if (ShAmt1.getOpcode() == ISD::SUB) {
10392 SDValue Sum = ShAmt1.getOperand(0);
10393 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
10394 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
10395 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
10396 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
10397 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
10398 return DAG.getNode(Opc, DL, VT,
10400 DAG.getNode(ISD::TRUNCATE, DL,
10403 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
10404 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
10406 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
10407 return DAG.getNode(Opc, DL, VT,
10408 N0.getOperand(0), N1.getOperand(0),
10409 DAG.getNode(ISD::TRUNCATE, DL,
10416 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
10417 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
10418 const X86Subtarget *Subtarget) {
10419 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
10420 // the FP state in cases where an emms may be missing.
10421 // A preferable solution to the general problem is to figure out the right
10422 // places to insert EMMS. This qualifies as a quick hack.
10424 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
10425 StoreSDNode *St = cast<StoreSDNode>(N);
10426 EVT VT = St->getValue().getValueType();
10427 if (VT.getSizeInBits() != 64)
10430 const Function *F = DAG.getMachineFunction().getFunction();
10431 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
10432 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
10433 && Subtarget->hasSSE2();
10434 if ((VT.isVector() ||
10435 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
10436 isa<LoadSDNode>(St->getValue()) &&
10437 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
10438 St->getChain().hasOneUse() && !St->isVolatile()) {
10439 SDNode* LdVal = St->getValue().getNode();
10440 LoadSDNode *Ld = 0;
10441 int TokenFactorIndex = -1;
10442 SmallVector<SDValue, 8> Ops;
10443 SDNode* ChainVal = St->getChain().getNode();
10444 // Must be a store of a load. We currently handle two cases: the load
10445 // is a direct child, and it's under an intervening TokenFactor. It is
10446 // possible to dig deeper under nested TokenFactors.
10447 if (ChainVal == LdVal)
10448 Ld = cast<LoadSDNode>(St->getChain());
10449 else if (St->getValue().hasOneUse() &&
10450 ChainVal->getOpcode() == ISD::TokenFactor) {
10451 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
10452 if (ChainVal->getOperand(i).getNode() == LdVal) {
10453 TokenFactorIndex = i;
10454 Ld = cast<LoadSDNode>(St->getValue());
10456 Ops.push_back(ChainVal->getOperand(i));
10460 if (!Ld || !ISD::isNormalLoad(Ld))
10463 // If this is not the MMX case, i.e. we are just turning i64 load/store
10464 // into f64 load/store, avoid the transformation if there are multiple
10465 // uses of the loaded value.
10466 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
10469 DebugLoc LdDL = Ld->getDebugLoc();
10470 DebugLoc StDL = N->getDebugLoc();
10471 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
10472 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
10474 if (Subtarget->is64Bit() || F64IsLegal) {
10475 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
10476 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
10477 Ld->getBasePtr(), Ld->getSrcValue(),
10478 Ld->getSrcValueOffset(), Ld->isVolatile(),
10479 Ld->isNonTemporal(), Ld->getAlignment());
10480 SDValue NewChain = NewLd.getValue(1);
10481 if (TokenFactorIndex != -1) {
10482 Ops.push_back(NewChain);
10483 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10486 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
10487 St->getSrcValue(), St->getSrcValueOffset(),
10488 St->isVolatile(), St->isNonTemporal(),
10489 St->getAlignment());
10492 // Otherwise, lower to two pairs of 32-bit loads / stores.
10493 SDValue LoAddr = Ld->getBasePtr();
10494 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
10495 DAG.getConstant(4, MVT::i32));
10497 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
10498 Ld->getSrcValue(), Ld->getSrcValueOffset(),
10499 Ld->isVolatile(), Ld->isNonTemporal(),
10500 Ld->getAlignment());
10501 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
10502 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
10503 Ld->isVolatile(), Ld->isNonTemporal(),
10504 MinAlign(Ld->getAlignment(), 4));
10506 SDValue NewChain = LoLd.getValue(1);
10507 if (TokenFactorIndex != -1) {
10508 Ops.push_back(LoLd);
10509 Ops.push_back(HiLd);
10510 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10514 LoAddr = St->getBasePtr();
10515 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
10516 DAG.getConstant(4, MVT::i32));
10518 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
10519 St->getSrcValue(), St->getSrcValueOffset(),
10520 St->isVolatile(), St->isNonTemporal(),
10521 St->getAlignment());
10522 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
10524 St->getSrcValueOffset() + 4,
10526 St->isNonTemporal(),
10527 MinAlign(St->getAlignment(), 4));
10528 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
10533 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
10534 /// X86ISD::FXOR nodes.
10535 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
10536 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
10537 // F[X]OR(0.0, x) -> x
10538 // F[X]OR(x, 0.0) -> x
10539 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10540 if (C->getValueAPF().isPosZero())
10541 return N->getOperand(1);
10542 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10543 if (C->getValueAPF().isPosZero())
10544 return N->getOperand(0);
10548 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
10549 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
10550 // FAND(0.0, x) -> 0.0
10551 // FAND(x, 0.0) -> 0.0
10552 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10553 if (C->getValueAPF().isPosZero())
10554 return N->getOperand(0);
10555 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10556 if (C->getValueAPF().isPosZero())
10557 return N->getOperand(1);
10561 static SDValue PerformBTCombine(SDNode *N,
10563 TargetLowering::DAGCombinerInfo &DCI) {
10564 // BT ignores high bits in the bit index operand.
10565 SDValue Op1 = N->getOperand(1);
10566 if (Op1.hasOneUse()) {
10567 unsigned BitWidth = Op1.getValueSizeInBits();
10568 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
10569 APInt KnownZero, KnownOne;
10570 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
10571 !DCI.isBeforeLegalizeOps());
10572 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10573 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
10574 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
10575 DCI.CommitTargetLoweringOpt(TLO);
10580 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
10581 SDValue Op = N->getOperand(0);
10582 if (Op.getOpcode() == ISD::BIT_CONVERT)
10583 Op = Op.getOperand(0);
10584 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
10585 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
10586 VT.getVectorElementType().getSizeInBits() ==
10587 OpVT.getVectorElementType().getSizeInBits()) {
10588 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
10593 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
10594 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
10595 // (and (i32 x86isd::setcc_carry), 1)
10596 // This eliminates the zext. This transformation is necessary because
10597 // ISD::SETCC is always legalized to i8.
10598 DebugLoc dl = N->getDebugLoc();
10599 SDValue N0 = N->getOperand(0);
10600 EVT VT = N->getValueType(0);
10601 if (N0.getOpcode() == ISD::AND &&
10603 N0.getOperand(0).hasOneUse()) {
10604 SDValue N00 = N0.getOperand(0);
10605 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
10607 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
10608 if (!C || C->getZExtValue() != 1)
10610 return DAG.getNode(ISD::AND, dl, VT,
10611 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
10612 N00.getOperand(0), N00.getOperand(1)),
10613 DAG.getConstant(1, VT));
10619 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
10620 DAGCombinerInfo &DCI) const {
10621 SelectionDAG &DAG = DCI.DAG;
10622 switch (N->getOpcode()) {
10624 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
10625 case ISD::EXTRACT_VECTOR_ELT:
10626 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
10627 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
10628 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
10629 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
10632 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
10633 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
10634 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
10636 case X86ISD::FOR: return PerformFORCombine(N, DAG);
10637 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
10638 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
10639 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
10640 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
10646 /// isTypeDesirableForOp - Return true if the target has native support for
10647 /// the specified value type and it is 'desirable' to use the type for the
10648 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
10649 /// instruction encodings are longer and some i16 instructions are slow.
10650 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
10651 if (!isTypeLegal(VT))
10653 if (VT != MVT::i16)
10660 case ISD::SIGN_EXTEND:
10661 case ISD::ZERO_EXTEND:
10662 case ISD::ANY_EXTEND:
10675 /// IsDesirableToPromoteOp - This method query the target whether it is
10676 /// beneficial for dag combiner to promote the specified node. If true, it
10677 /// should return the desired promotion type by reference.
10678 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
10679 EVT VT = Op.getValueType();
10680 if (VT != MVT::i16)
10683 bool Promote = false;
10684 bool Commute = false;
10685 switch (Op.getOpcode()) {
10688 LoadSDNode *LD = cast<LoadSDNode>(Op);
10689 // If the non-extending load has a single use and it's not live out, then it
10690 // might be folded.
10691 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
10692 Op.hasOneUse()*/) {
10693 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
10694 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
10695 // The only case where we'd want to promote LOAD (rather then it being
10696 // promoted as an operand is when it's only use is liveout.
10697 if (UI->getOpcode() != ISD::CopyToReg)
10704 case ISD::SIGN_EXTEND:
10705 case ISD::ZERO_EXTEND:
10706 case ISD::ANY_EXTEND:
10711 SDValue N0 = Op.getOperand(0);
10712 // Look out for (store (shl (load), x)).
10713 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10726 SDValue N0 = Op.getOperand(0);
10727 SDValue N1 = Op.getOperand(1);
10728 if (!Commute && MayFoldLoad(N1))
10730 // Avoid disabling potential load folding opportunities.
10731 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10733 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10743 //===----------------------------------------------------------------------===//
10744 // X86 Inline Assembly Support
10745 //===----------------------------------------------------------------------===//
10747 static bool LowerToBSwap(CallInst *CI) {
10748 // FIXME: this should verify that we are targetting a 486 or better. If not,
10749 // we will turn this bswap into something that will be lowered to logical ops
10750 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10751 // so don't worry about this.
10753 // Verify this is a simple bswap.
10754 if (CI->getNumArgOperands() != 1 ||
10755 CI->getType() != CI->getArgOperand(0)->getType() ||
10756 !CI->getType()->isIntegerTy())
10759 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10760 if (!Ty || Ty->getBitWidth() % 16 != 0)
10763 // Okay, we can do this xform, do so now.
10764 const Type *Tys[] = { Ty };
10765 Module *M = CI->getParent()->getParent()->getParent();
10766 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10768 Value *Op = CI->getArgOperand(0);
10769 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10771 CI->replaceAllUsesWith(Op);
10772 CI->eraseFromParent();
10776 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10777 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10778 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10780 std::string AsmStr = IA->getAsmString();
10782 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10783 SmallVector<StringRef, 4> AsmPieces;
10784 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10786 switch (AsmPieces.size()) {
10787 default: return false;
10789 AsmStr = AsmPieces[0];
10791 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10794 if (AsmPieces.size() == 2 &&
10795 (AsmPieces[0] == "bswap" ||
10796 AsmPieces[0] == "bswapq" ||
10797 AsmPieces[0] == "bswapl") &&
10798 (AsmPieces[1] == "$0" ||
10799 AsmPieces[1] == "${0:q}")) {
10800 // No need to check constraints, nothing other than the equivalent of
10801 // "=r,0" would be valid here.
10802 return LowerToBSwap(CI);
10804 // rorw $$8, ${0:w} --> llvm.bswap.i16
10805 if (CI->getType()->isIntegerTy(16) &&
10806 AsmPieces.size() == 3 &&
10807 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10808 AsmPieces[1] == "$$8," &&
10809 AsmPieces[2] == "${0:w}" &&
10810 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10812 const std::string &Constraints = IA->getConstraintString();
10813 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10814 std::sort(AsmPieces.begin(), AsmPieces.end());
10815 if (AsmPieces.size() == 4 &&
10816 AsmPieces[0] == "~{cc}" &&
10817 AsmPieces[1] == "~{dirflag}" &&
10818 AsmPieces[2] == "~{flags}" &&
10819 AsmPieces[3] == "~{fpsr}") {
10820 return LowerToBSwap(CI);
10825 if (CI->getType()->isIntegerTy(64) &&
10826 Constraints.size() >= 2 &&
10827 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10828 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10829 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10830 SmallVector<StringRef, 4> Words;
10831 SplitString(AsmPieces[0], Words, " \t");
10832 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10834 SplitString(AsmPieces[1], Words, " \t");
10835 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10837 SplitString(AsmPieces[2], Words, " \t,");
10838 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10839 Words[2] == "%edx") {
10840 return LowerToBSwap(CI);
10852 /// getConstraintType - Given a constraint letter, return the type of
10853 /// constraint it is for this target.
10854 X86TargetLowering::ConstraintType
10855 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10856 if (Constraint.size() == 1) {
10857 switch (Constraint[0]) {
10869 return C_RegisterClass;
10877 return TargetLowering::getConstraintType(Constraint);
10880 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10881 /// with another that has more specific requirements based on the type of the
10882 /// corresponding operand.
10883 const char *X86TargetLowering::
10884 LowerXConstraint(EVT ConstraintVT) const {
10885 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10886 // 'f' like normal targets.
10887 if (ConstraintVT.isFloatingPoint()) {
10888 if (Subtarget->hasSSE2())
10890 if (Subtarget->hasSSE1())
10894 return TargetLowering::LowerXConstraint(ConstraintVT);
10897 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10898 /// vector. If it is invalid, don't add anything to Ops.
10899 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10901 std::vector<SDValue>&Ops,
10902 SelectionDAG &DAG) const {
10903 SDValue Result(0, 0);
10905 switch (Constraint) {
10908 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10909 if (C->getZExtValue() <= 31) {
10910 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10916 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10917 if (C->getZExtValue() <= 63) {
10918 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10924 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10925 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10926 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10932 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10933 if (C->getZExtValue() <= 255) {
10934 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10940 // 32-bit signed value
10941 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10942 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10943 C->getSExtValue())) {
10944 // Widen to 64 bits here to get it sign extended.
10945 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10948 // FIXME gcc accepts some relocatable values here too, but only in certain
10949 // memory models; it's complicated.
10954 // 32-bit unsigned value
10955 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10956 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10957 C->getZExtValue())) {
10958 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10962 // FIXME gcc accepts some relocatable values here too, but only in certain
10963 // memory models; it's complicated.
10967 // Literal immediates are always ok.
10968 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10969 // Widen to 64 bits here to get it sign extended.
10970 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10974 // In any sort of PIC mode addresses need to be computed at runtime by
10975 // adding in a register or some sort of table lookup. These can't
10976 // be used as immediates.
10977 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
10980 // If we are in non-pic codegen mode, we allow the address of a global (with
10981 // an optional displacement) to be used with 'i'.
10982 GlobalAddressSDNode *GA = 0;
10983 int64_t Offset = 0;
10985 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10987 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10988 Offset += GA->getOffset();
10990 } else if (Op.getOpcode() == ISD::ADD) {
10991 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10992 Offset += C->getZExtValue();
10993 Op = Op.getOperand(0);
10996 } else if (Op.getOpcode() == ISD::SUB) {
10997 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10998 Offset += -C->getZExtValue();
10999 Op = Op.getOperand(0);
11004 // Otherwise, this isn't something we can handle, reject it.
11008 const GlobalValue *GV = GA->getGlobal();
11009 // If we require an extra load to get this address, as in PIC mode, we
11010 // can't accept it.
11011 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
11012 getTargetMachine())))
11015 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
11016 GA->getValueType(0), Offset);
11021 if (Result.getNode()) {
11022 Ops.push_back(Result);
11025 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11028 std::vector<unsigned> X86TargetLowering::
11029 getRegClassForInlineAsmConstraint(const std::string &Constraint,
11031 if (Constraint.size() == 1) {
11032 // FIXME: not handling fp-stack yet!
11033 switch (Constraint[0]) { // GCC X86 Constraint Letters
11034 default: break; // Unknown constraint letter
11035 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
11036 if (Subtarget->is64Bit()) {
11037 if (VT == MVT::i32)
11038 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
11039 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
11040 X86::R10D,X86::R11D,X86::R12D,
11041 X86::R13D,X86::R14D,X86::R15D,
11042 X86::EBP, X86::ESP, 0);
11043 else if (VT == MVT::i16)
11044 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
11045 X86::SI, X86::DI, X86::R8W,X86::R9W,
11046 X86::R10W,X86::R11W,X86::R12W,
11047 X86::R13W,X86::R14W,X86::R15W,
11048 X86::BP, X86::SP, 0);
11049 else if (VT == MVT::i8)
11050 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
11051 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
11052 X86::R10B,X86::R11B,X86::R12B,
11053 X86::R13B,X86::R14B,X86::R15B,
11054 X86::BPL, X86::SPL, 0);
11056 else if (VT == MVT::i64)
11057 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
11058 X86::RSI, X86::RDI, X86::R8, X86::R9,
11059 X86::R10, X86::R11, X86::R12,
11060 X86::R13, X86::R14, X86::R15,
11061 X86::RBP, X86::RSP, 0);
11065 // 32-bit fallthrough
11066 case 'Q': // Q_REGS
11067 if (VT == MVT::i32)
11068 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
11069 else if (VT == MVT::i16)
11070 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
11071 else if (VT == MVT::i8)
11072 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
11073 else if (VT == MVT::i64)
11074 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
11079 return std::vector<unsigned>();
11082 std::pair<unsigned, const TargetRegisterClass*>
11083 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
11085 // First, see if this is a constraint that directly corresponds to an LLVM
11087 if (Constraint.size() == 1) {
11088 // GCC Constraint Letters
11089 switch (Constraint[0]) {
11091 case 'r': // GENERAL_REGS
11092 case 'l': // INDEX_REGS
11094 return std::make_pair(0U, X86::GR8RegisterClass);
11095 if (VT == MVT::i16)
11096 return std::make_pair(0U, X86::GR16RegisterClass);
11097 if (VT == MVT::i32 || !Subtarget->is64Bit())
11098 return std::make_pair(0U, X86::GR32RegisterClass);
11099 return std::make_pair(0U, X86::GR64RegisterClass);
11100 case 'R': // LEGACY_REGS
11102 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
11103 if (VT == MVT::i16)
11104 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
11105 if (VT == MVT::i32 || !Subtarget->is64Bit())
11106 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
11107 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
11108 case 'f': // FP Stack registers.
11109 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
11110 // value to the correct fpstack register class.
11111 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
11112 return std::make_pair(0U, X86::RFP32RegisterClass);
11113 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
11114 return std::make_pair(0U, X86::RFP64RegisterClass);
11115 return std::make_pair(0U, X86::RFP80RegisterClass);
11116 case 'y': // MMX_REGS if MMX allowed.
11117 if (!Subtarget->hasMMX()) break;
11118 return std::make_pair(0U, X86::VR64RegisterClass);
11119 case 'Y': // SSE_REGS if SSE2 allowed
11120 if (!Subtarget->hasSSE2()) break;
11122 case 'x': // SSE_REGS if SSE1 allowed
11123 if (!Subtarget->hasSSE1()) break;
11125 switch (VT.getSimpleVT().SimpleTy) {
11127 // Scalar SSE types.
11130 return std::make_pair(0U, X86::FR32RegisterClass);
11133 return std::make_pair(0U, X86::FR64RegisterClass);
11141 return std::make_pair(0U, X86::VR128RegisterClass);
11147 // Use the default implementation in TargetLowering to convert the register
11148 // constraint into a member of a register class.
11149 std::pair<unsigned, const TargetRegisterClass*> Res;
11150 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
11152 // Not found as a standard register?
11153 if (Res.second == 0) {
11154 // Map st(0) -> st(7) -> ST0
11155 if (Constraint.size() == 7 && Constraint[0] == '{' &&
11156 tolower(Constraint[1]) == 's' &&
11157 tolower(Constraint[2]) == 't' &&
11158 Constraint[3] == '(' &&
11159 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
11160 Constraint[5] == ')' &&
11161 Constraint[6] == '}') {
11163 Res.first = X86::ST0+Constraint[4]-'0';
11164 Res.second = X86::RFP80RegisterClass;
11168 // GCC allows "st(0)" to be called just plain "st".
11169 if (StringRef("{st}").equals_lower(Constraint)) {
11170 Res.first = X86::ST0;
11171 Res.second = X86::RFP80RegisterClass;
11176 if (StringRef("{flags}").equals_lower(Constraint)) {
11177 Res.first = X86::EFLAGS;
11178 Res.second = X86::CCRRegisterClass;
11182 // 'A' means EAX + EDX.
11183 if (Constraint == "A") {
11184 Res.first = X86::EAX;
11185 Res.second = X86::GR32_ADRegisterClass;
11191 // Otherwise, check to see if this is a register class of the wrong value
11192 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
11193 // turn into {ax},{dx}.
11194 if (Res.second->hasType(VT))
11195 return Res; // Correct type already, nothing to do.
11197 // All of the single-register GCC register classes map their values onto
11198 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
11199 // really want an 8-bit or 32-bit register, map to the appropriate register
11200 // class and return the appropriate register.
11201 if (Res.second == X86::GR16RegisterClass) {
11202 if (VT == MVT::i8) {
11203 unsigned DestReg = 0;
11204 switch (Res.first) {
11206 case X86::AX: DestReg = X86::AL; break;
11207 case X86::DX: DestReg = X86::DL; break;
11208 case X86::CX: DestReg = X86::CL; break;
11209 case X86::BX: DestReg = X86::BL; break;
11212 Res.first = DestReg;
11213 Res.second = X86::GR8RegisterClass;
11215 } else if (VT == MVT::i32) {
11216 unsigned DestReg = 0;
11217 switch (Res.first) {
11219 case X86::AX: DestReg = X86::EAX; break;
11220 case X86::DX: DestReg = X86::EDX; break;
11221 case X86::CX: DestReg = X86::ECX; break;
11222 case X86::BX: DestReg = X86::EBX; break;
11223 case X86::SI: DestReg = X86::ESI; break;
11224 case X86::DI: DestReg = X86::EDI; break;
11225 case X86::BP: DestReg = X86::EBP; break;
11226 case X86::SP: DestReg = X86::ESP; break;
11229 Res.first = DestReg;
11230 Res.second = X86::GR32RegisterClass;
11232 } else if (VT == MVT::i64) {
11233 unsigned DestReg = 0;
11234 switch (Res.first) {
11236 case X86::AX: DestReg = X86::RAX; break;
11237 case X86::DX: DestReg = X86::RDX; break;
11238 case X86::CX: DestReg = X86::RCX; break;
11239 case X86::BX: DestReg = X86::RBX; break;
11240 case X86::SI: DestReg = X86::RSI; break;
11241 case X86::DI: DestReg = X86::RDI; break;
11242 case X86::BP: DestReg = X86::RBP; break;
11243 case X86::SP: DestReg = X86::RSP; break;
11246 Res.first = DestReg;
11247 Res.second = X86::GR64RegisterClass;
11250 } else if (Res.second == X86::FR32RegisterClass ||
11251 Res.second == X86::FR64RegisterClass ||
11252 Res.second == X86::VR128RegisterClass) {
11253 // Handle references to XMM physical registers that got mapped into the
11254 // wrong class. This can happen with constraints like {xmm0} where the
11255 // target independent register mapper will just pick the first match it can
11256 // find, ignoring the required type.
11257 if (VT == MVT::f32)
11258 Res.second = X86::FR32RegisterClass;
11259 else if (VT == MVT::f64)
11260 Res.second = X86::FR64RegisterClass;
11261 else if (X86::VR128RegisterClass->hasType(VT))
11262 Res.second = X86::VR128RegisterClass;