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 // i8 and i16 vectors are custom , because the source register and source
842 // source memory operand types are not the same width. f32 vectors are
843 // custom since the immediate controlling the insert encodes additional
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
848 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
851 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
852 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
853 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
855 if (Subtarget->is64Bit()) {
856 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
857 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
861 if (Subtarget->hasSSE42()) {
862 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
865 if (!UseSoftFloat && Subtarget->hasAVX()) {
866 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
867 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
868 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
869 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
871 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
872 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
873 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
874 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
875 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
876 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
877 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
878 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
879 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
880 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
881 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
882 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
883 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
884 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
885 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
887 // Operations to consider commented out -v16i16 v32i8
888 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
889 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
890 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
891 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
892 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
893 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
894 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
895 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
896 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
897 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
898 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
899 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
900 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
901 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
903 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
904 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
905 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
906 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
908 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
909 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
910 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
912 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
914 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
915 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
916 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
917 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
918 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
919 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
922 // Not sure we want to do this since there are no 256-bit integer
925 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
926 // This includes 256-bit vectors
927 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
928 EVT VT = (MVT::SimpleValueType)i;
930 // Do not attempt to custom lower non-power-of-2 vectors
931 if (!isPowerOf2_32(VT.getVectorNumElements()))
934 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
935 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
936 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
939 if (Subtarget->is64Bit()) {
940 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
941 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
946 // Not sure we want to do this since there are no 256-bit integer
949 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
950 // Including 256-bit vectors
951 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
952 EVT VT = (MVT::SimpleValueType)i;
954 if (!VT.is256BitVector()) {
957 setOperationAction(ISD::AND, VT, Promote);
958 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
959 setOperationAction(ISD::OR, VT, Promote);
960 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
961 setOperationAction(ISD::XOR, VT, Promote);
962 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
963 setOperationAction(ISD::LOAD, VT, Promote);
964 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
965 setOperationAction(ISD::SELECT, VT, Promote);
966 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
969 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
973 // We want to custom lower some of our intrinsics.
974 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
976 // Add/Sub/Mul with overflow operations are custom lowered.
977 setOperationAction(ISD::SADDO, MVT::i32, Custom);
978 setOperationAction(ISD::UADDO, MVT::i32, Custom);
979 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
980 setOperationAction(ISD::USUBO, MVT::i32, Custom);
981 setOperationAction(ISD::SMULO, MVT::i32, Custom);
983 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
984 // handle type legalization for these operations here.
986 // FIXME: We really should do custom legalization for addition and
987 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
988 // than generic legalization for 64-bit multiplication-with-overflow, though.
989 if (Subtarget->is64Bit()) {
990 setOperationAction(ISD::SADDO, MVT::i64, Custom);
991 setOperationAction(ISD::UADDO, MVT::i64, Custom);
992 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
993 setOperationAction(ISD::USUBO, MVT::i64, Custom);
994 setOperationAction(ISD::SMULO, MVT::i64, Custom);
997 if (!Subtarget->is64Bit()) {
998 // These libcalls are not available in 32-bit.
999 setLibcallName(RTLIB::SHL_I128, 0);
1000 setLibcallName(RTLIB::SRL_I128, 0);
1001 setLibcallName(RTLIB::SRA_I128, 0);
1004 // We have target-specific dag combine patterns for the following nodes:
1005 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1006 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1007 setTargetDAGCombine(ISD::BUILD_VECTOR);
1008 setTargetDAGCombine(ISD::SELECT);
1009 setTargetDAGCombine(ISD::SHL);
1010 setTargetDAGCombine(ISD::SRA);
1011 setTargetDAGCombine(ISD::SRL);
1012 setTargetDAGCombine(ISD::OR);
1013 setTargetDAGCombine(ISD::STORE);
1014 setTargetDAGCombine(ISD::ZERO_EXTEND);
1015 if (Subtarget->is64Bit())
1016 setTargetDAGCombine(ISD::MUL);
1018 computeRegisterProperties();
1020 // FIXME: These should be based on subtarget info. Plus, the values should
1021 // be smaller when we are in optimizing for size mode.
1022 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1023 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1024 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1025 setPrefLoopAlignment(16);
1026 benefitFromCodePlacementOpt = true;
1030 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1035 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1036 /// the desired ByVal argument alignment.
1037 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1040 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1041 if (VTy->getBitWidth() == 128)
1043 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1044 unsigned EltAlign = 0;
1045 getMaxByValAlign(ATy->getElementType(), EltAlign);
1046 if (EltAlign > MaxAlign)
1047 MaxAlign = EltAlign;
1048 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1049 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1050 unsigned EltAlign = 0;
1051 getMaxByValAlign(STy->getElementType(i), EltAlign);
1052 if (EltAlign > MaxAlign)
1053 MaxAlign = EltAlign;
1061 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1062 /// function arguments in the caller parameter area. For X86, aggregates
1063 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1064 /// are at 4-byte boundaries.
1065 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1066 if (Subtarget->is64Bit()) {
1067 // Max of 8 and alignment of type.
1068 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1075 if (Subtarget->hasSSE1())
1076 getMaxByValAlign(Ty, Align);
1080 /// getOptimalMemOpType - Returns the target specific optimal type for load
1081 /// and store operations as a result of memset, memcpy, and memmove
1082 /// lowering. If DstAlign is zero that means it's safe to destination
1083 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1084 /// means there isn't a need to check it against alignment requirement,
1085 /// probably because the source does not need to be loaded. If
1086 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1087 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1088 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1089 /// constant so it does not need to be loaded.
1090 /// It returns EVT::Other if the type should be determined using generic
1091 /// target-independent logic.
1093 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1094 unsigned DstAlign, unsigned SrcAlign,
1095 bool NonScalarIntSafe,
1097 MachineFunction &MF) const {
1098 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1099 // linux. This is because the stack realignment code can't handle certain
1100 // cases like PR2962. This should be removed when PR2962 is fixed.
1101 const Function *F = MF.getFunction();
1102 if (NonScalarIntSafe &&
1103 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1105 (Subtarget->isUnalignedMemAccessFast() ||
1106 ((DstAlign == 0 || DstAlign >= 16) &&
1107 (SrcAlign == 0 || SrcAlign >= 16))) &&
1108 Subtarget->getStackAlignment() >= 16) {
1109 if (Subtarget->hasSSE2())
1111 if (Subtarget->hasSSE1())
1113 } else if (!MemcpyStrSrc && Size >= 8 &&
1114 !Subtarget->is64Bit() &&
1115 Subtarget->getStackAlignment() >= 8 &&
1116 Subtarget->hasSSE2()) {
1117 // Do not use f64 to lower memcpy if source is string constant. It's
1118 // better to use i32 to avoid the loads.
1122 if (Subtarget->is64Bit() && Size >= 8)
1127 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1128 /// current function. The returned value is a member of the
1129 /// MachineJumpTableInfo::JTEntryKind enum.
1130 unsigned X86TargetLowering::getJumpTableEncoding() const {
1131 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1133 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1134 Subtarget->isPICStyleGOT())
1135 return MachineJumpTableInfo::EK_Custom32;
1137 // Otherwise, use the normal jump table encoding heuristics.
1138 return TargetLowering::getJumpTableEncoding();
1141 /// getPICBaseSymbol - Return the X86-32 PIC base.
1143 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1144 MCContext &Ctx) const {
1145 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1146 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1147 Twine(MF->getFunctionNumber())+"$pb");
1152 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1153 const MachineBasicBlock *MBB,
1154 unsigned uid,MCContext &Ctx) const{
1155 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1156 Subtarget->isPICStyleGOT());
1157 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1159 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1160 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1163 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1165 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1166 SelectionDAG &DAG) const {
1167 if (!Subtarget->is64Bit())
1168 // This doesn't have DebugLoc associated with it, but is not really the
1169 // same as a Register.
1170 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1174 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1175 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1177 const MCExpr *X86TargetLowering::
1178 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1179 MCContext &Ctx) const {
1180 // X86-64 uses RIP relative addressing based on the jump table label.
1181 if (Subtarget->isPICStyleRIPRel())
1182 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1184 // Otherwise, the reference is relative to the PIC base.
1185 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1188 /// getFunctionAlignment - Return the Log2 alignment of this function.
1189 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1190 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1193 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1194 unsigned &Offset) const {
1195 if (!Subtarget->isTargetLinux())
1198 if (Subtarget->is64Bit()) {
1199 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1201 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1214 //===----------------------------------------------------------------------===//
1215 // Return Value Calling Convention Implementation
1216 //===----------------------------------------------------------------------===//
1218 #include "X86GenCallingConv.inc"
1221 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1222 const SmallVectorImpl<ISD::OutputArg> &Outs,
1223 LLVMContext &Context) const {
1224 SmallVector<CCValAssign, 16> RVLocs;
1225 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1227 return CCInfo.CheckReturn(Outs, RetCC_X86);
1231 X86TargetLowering::LowerReturn(SDValue Chain,
1232 CallingConv::ID CallConv, bool isVarArg,
1233 const SmallVectorImpl<ISD::OutputArg> &Outs,
1234 const SmallVectorImpl<SDValue> &OutVals,
1235 DebugLoc dl, SelectionDAG &DAG) const {
1236 MachineFunction &MF = DAG.getMachineFunction();
1237 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1239 SmallVector<CCValAssign, 16> RVLocs;
1240 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1241 RVLocs, *DAG.getContext());
1242 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1244 // Add the regs to the liveout set for the function.
1245 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1246 for (unsigned i = 0; i != RVLocs.size(); ++i)
1247 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1248 MRI.addLiveOut(RVLocs[i].getLocReg());
1252 SmallVector<SDValue, 6> RetOps;
1253 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1254 // Operand #1 = Bytes To Pop
1255 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1258 // Copy the result values into the output registers.
1259 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1260 CCValAssign &VA = RVLocs[i];
1261 assert(VA.isRegLoc() && "Can only return in registers!");
1262 SDValue ValToCopy = OutVals[i];
1263 EVT ValVT = ValToCopy.getValueType();
1265 // If this is x86-64, and we disabled SSE, we can't return FP values
1266 if ((ValVT == MVT::f32 || ValVT == MVT::f64) &&
1267 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1268 report_fatal_error("SSE register return with SSE disabled");
1270 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1271 // llvm-gcc has never done it right and no one has noticed, so this
1272 // should be OK for now.
1273 if (ValVT == MVT::f64 &&
1274 (Subtarget->is64Bit() && !Subtarget->hasSSE2())) {
1275 report_fatal_error("SSE2 register return with SSE2 disabled");
1278 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1279 // the RET instruction and handled by the FP Stackifier.
1280 if (VA.getLocReg() == X86::ST0 ||
1281 VA.getLocReg() == X86::ST1) {
1282 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1283 // change the value to the FP stack register class.
1284 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1285 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1286 RetOps.push_back(ValToCopy);
1287 // Don't emit a copytoreg.
1291 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1292 // which is returned in RAX / RDX.
1293 if (Subtarget->is64Bit()) {
1294 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1295 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1296 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1297 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1302 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1303 Flag = Chain.getValue(1);
1306 // The x86-64 ABI for returning structs by value requires that we copy
1307 // the sret argument into %rax for the return. We saved the argument into
1308 // a virtual register in the entry block, so now we copy the value out
1310 if (Subtarget->is64Bit() &&
1311 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1312 MachineFunction &MF = DAG.getMachineFunction();
1313 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1314 unsigned Reg = FuncInfo->getSRetReturnReg();
1316 "SRetReturnReg should have been set in LowerFormalArguments().");
1317 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1319 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1320 Flag = Chain.getValue(1);
1322 // RAX now acts like a return value.
1323 MRI.addLiveOut(X86::RAX);
1326 RetOps[0] = Chain; // Update chain.
1328 // Add the flag if we have it.
1330 RetOps.push_back(Flag);
1332 return DAG.getNode(X86ISD::RET_FLAG, dl,
1333 MVT::Other, &RetOps[0], RetOps.size());
1336 /// LowerCallResult - Lower the result values of a call into the
1337 /// appropriate copies out of appropriate physical registers.
1340 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1341 CallingConv::ID CallConv, bool isVarArg,
1342 const SmallVectorImpl<ISD::InputArg> &Ins,
1343 DebugLoc dl, SelectionDAG &DAG,
1344 SmallVectorImpl<SDValue> &InVals) const {
1346 // Assign locations to each value returned by this call.
1347 SmallVector<CCValAssign, 16> RVLocs;
1348 bool Is64Bit = Subtarget->is64Bit();
1349 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1350 RVLocs, *DAG.getContext());
1351 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1353 // Copy all of the result registers out of their specified physreg.
1354 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1355 CCValAssign &VA = RVLocs[i];
1356 EVT CopyVT = VA.getValVT();
1358 // If this is x86-64, and we disabled SSE, we can't return FP values
1359 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1360 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1361 report_fatal_error("SSE register return with SSE disabled");
1366 // If this is a call to a function that returns an fp value on the floating
1367 // point stack, we must guarantee the the value is popped from the stack, so
1368 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1369 // if the return value is not used. We use the FpGET_ST0 instructions
1371 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1372 // If we prefer to use the value in xmm registers, copy it out as f80 and
1373 // use a truncate to move it from fp stack reg to xmm reg.
1374 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1375 bool isST0 = VA.getLocReg() == X86::ST0;
1377 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1378 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1379 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1380 SDValue Ops[] = { Chain, InFlag };
1381 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
1383 Val = Chain.getValue(0);
1385 // Round the f80 to the right size, which also moves it to the appropriate
1387 if (CopyVT != VA.getValVT())
1388 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1389 // This truncation won't change the value.
1390 DAG.getIntPtrConstant(1));
1391 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1392 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1393 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1394 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1395 MVT::v2i64, InFlag).getValue(1);
1396 Val = Chain.getValue(0);
1397 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1398 Val, DAG.getConstant(0, MVT::i64));
1400 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1401 MVT::i64, InFlag).getValue(1);
1402 Val = Chain.getValue(0);
1404 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1406 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1407 CopyVT, InFlag).getValue(1);
1408 Val = Chain.getValue(0);
1410 InFlag = Chain.getValue(2);
1411 InVals.push_back(Val);
1418 //===----------------------------------------------------------------------===//
1419 // C & StdCall & Fast Calling Convention implementation
1420 //===----------------------------------------------------------------------===//
1421 // StdCall calling convention seems to be standard for many Windows' API
1422 // routines and around. It differs from C calling convention just a little:
1423 // callee should clean up the stack, not caller. Symbols should be also
1424 // decorated in some fancy way :) It doesn't support any vector arguments.
1425 // For info on fast calling convention see Fast Calling Convention (tail call)
1426 // implementation LowerX86_32FastCCCallTo.
1428 /// CallIsStructReturn - Determines whether a call uses struct return
1430 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1434 return Outs[0].Flags.isSRet();
1437 /// ArgsAreStructReturn - Determines whether a function uses struct
1438 /// return semantics.
1440 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1444 return Ins[0].Flags.isSRet();
1447 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1448 /// given CallingConvention value.
1449 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1450 if (Subtarget->is64Bit()) {
1451 if (CC == CallingConv::GHC)
1452 return CC_X86_64_GHC;
1453 else if (Subtarget->isTargetWin64())
1454 return CC_X86_Win64_C;
1459 if (CC == CallingConv::X86_FastCall)
1460 return CC_X86_32_FastCall;
1461 else if (CC == CallingConv::X86_ThisCall)
1462 return CC_X86_32_ThisCall;
1463 else if (CC == CallingConv::Fast)
1464 return CC_X86_32_FastCC;
1465 else if (CC == CallingConv::GHC)
1466 return CC_X86_32_GHC;
1471 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1472 /// by "Src" to address "Dst" with size and alignment information specified by
1473 /// the specific parameter attribute. The copy will be passed as a byval
1474 /// function parameter.
1476 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1477 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1479 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1480 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1481 /*isVolatile*/false, /*AlwaysInline=*/true,
1485 /// IsTailCallConvention - Return true if the calling convention is one that
1486 /// supports tail call optimization.
1487 static bool IsTailCallConvention(CallingConv::ID CC) {
1488 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1491 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1492 /// a tailcall target by changing its ABI.
1493 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1494 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1498 X86TargetLowering::LowerMemArgument(SDValue Chain,
1499 CallingConv::ID CallConv,
1500 const SmallVectorImpl<ISD::InputArg> &Ins,
1501 DebugLoc dl, SelectionDAG &DAG,
1502 const CCValAssign &VA,
1503 MachineFrameInfo *MFI,
1505 // Create the nodes corresponding to a load from this parameter slot.
1506 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1507 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1508 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1511 // If value is passed by pointer we have address passed instead of the value
1513 if (VA.getLocInfo() == CCValAssign::Indirect)
1514 ValVT = VA.getLocVT();
1516 ValVT = VA.getValVT();
1518 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1519 // changed with more analysis.
1520 // In case of tail call optimization mark all arguments mutable. Since they
1521 // could be overwritten by lowering of arguments in case of a tail call.
1522 if (Flags.isByVal()) {
1523 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1524 VA.getLocMemOffset(), isImmutable);
1525 return DAG.getFrameIndex(FI, getPointerTy());
1527 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1528 VA.getLocMemOffset(), isImmutable);
1529 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1530 return DAG.getLoad(ValVT, dl, Chain, FIN,
1531 PseudoSourceValue::getFixedStack(FI), 0,
1537 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1538 CallingConv::ID CallConv,
1540 const SmallVectorImpl<ISD::InputArg> &Ins,
1543 SmallVectorImpl<SDValue> &InVals)
1545 MachineFunction &MF = DAG.getMachineFunction();
1546 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1548 const Function* Fn = MF.getFunction();
1549 if (Fn->hasExternalLinkage() &&
1550 Subtarget->isTargetCygMing() &&
1551 Fn->getName() == "main")
1552 FuncInfo->setForceFramePointer(true);
1554 MachineFrameInfo *MFI = MF.getFrameInfo();
1555 bool Is64Bit = Subtarget->is64Bit();
1556 bool IsWin64 = Subtarget->isTargetWin64();
1558 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1559 "Var args not supported with calling convention fastcc or ghc");
1561 // Assign locations to all of the incoming arguments.
1562 SmallVector<CCValAssign, 16> ArgLocs;
1563 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1564 ArgLocs, *DAG.getContext());
1565 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1567 unsigned LastVal = ~0U;
1569 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1570 CCValAssign &VA = ArgLocs[i];
1571 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1573 assert(VA.getValNo() != LastVal &&
1574 "Don't support value assigned to multiple locs yet");
1575 LastVal = VA.getValNo();
1577 if (VA.isRegLoc()) {
1578 EVT RegVT = VA.getLocVT();
1579 TargetRegisterClass *RC = NULL;
1580 if (RegVT == MVT::i32)
1581 RC = X86::GR32RegisterClass;
1582 else if (Is64Bit && RegVT == MVT::i64)
1583 RC = X86::GR64RegisterClass;
1584 else if (RegVT == MVT::f32)
1585 RC = X86::FR32RegisterClass;
1586 else if (RegVT == MVT::f64)
1587 RC = X86::FR64RegisterClass;
1588 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1589 RC = X86::VR128RegisterClass;
1590 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1591 RC = X86::VR64RegisterClass;
1593 llvm_unreachable("Unknown argument type!");
1595 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1596 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1598 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1599 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1601 if (VA.getLocInfo() == CCValAssign::SExt)
1602 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1603 DAG.getValueType(VA.getValVT()));
1604 else if (VA.getLocInfo() == CCValAssign::ZExt)
1605 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1606 DAG.getValueType(VA.getValVT()));
1607 else if (VA.getLocInfo() == CCValAssign::BCvt)
1608 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1610 if (VA.isExtInLoc()) {
1611 // Handle MMX values passed in XMM regs.
1612 if (RegVT.isVector()) {
1613 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1614 ArgValue, DAG.getConstant(0, MVT::i64));
1615 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1617 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1620 assert(VA.isMemLoc());
1621 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1624 // If value is passed via pointer - do a load.
1625 if (VA.getLocInfo() == CCValAssign::Indirect)
1626 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1629 InVals.push_back(ArgValue);
1632 // The x86-64 ABI for returning structs by value requires that we copy
1633 // the sret argument into %rax for the return. Save the argument into
1634 // a virtual register so that we can access it from the return points.
1635 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1636 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1637 unsigned Reg = FuncInfo->getSRetReturnReg();
1639 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1640 FuncInfo->setSRetReturnReg(Reg);
1642 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1643 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1646 unsigned StackSize = CCInfo.getNextStackOffset();
1647 // Align stack specially for tail calls.
1648 if (FuncIsMadeTailCallSafe(CallConv))
1649 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1651 // If the function takes variable number of arguments, make a frame index for
1652 // the start of the first vararg value... for expansion of llvm.va_start.
1654 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1655 CallConv != CallingConv::X86_ThisCall)) {
1656 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1659 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1661 // FIXME: We should really autogenerate these arrays
1662 static const unsigned GPR64ArgRegsWin64[] = {
1663 X86::RCX, X86::RDX, X86::R8, X86::R9
1665 static const unsigned XMMArgRegsWin64[] = {
1666 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1668 static const unsigned GPR64ArgRegs64Bit[] = {
1669 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1671 static const unsigned XMMArgRegs64Bit[] = {
1672 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1673 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1675 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1678 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1679 GPR64ArgRegs = GPR64ArgRegsWin64;
1680 XMMArgRegs = XMMArgRegsWin64;
1682 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1683 GPR64ArgRegs = GPR64ArgRegs64Bit;
1684 XMMArgRegs = XMMArgRegs64Bit;
1686 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1688 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1691 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1692 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1693 "SSE register cannot be used when SSE is disabled!");
1694 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1695 "SSE register cannot be used when SSE is disabled!");
1696 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1697 // Kernel mode asks for SSE to be disabled, so don't push them
1699 TotalNumXMMRegs = 0;
1701 // For X86-64, if there are vararg parameters that are passed via
1702 // registers, then we must store them to their spots on the stack so they
1703 // may be loaded by deferencing the result of va_next.
1704 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1705 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1706 FuncInfo->setRegSaveFrameIndex(
1707 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1710 // Store the integer parameter registers.
1711 SmallVector<SDValue, 8> MemOps;
1712 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1714 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1715 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1716 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1717 DAG.getIntPtrConstant(Offset));
1718 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1719 X86::GR64RegisterClass);
1720 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1722 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1723 PseudoSourceValue::getFixedStack(
1724 FuncInfo->getRegSaveFrameIndex()),
1725 Offset, false, false, 0);
1726 MemOps.push_back(Store);
1730 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1731 // Now store the XMM (fp + vector) parameter registers.
1732 SmallVector<SDValue, 11> SaveXMMOps;
1733 SaveXMMOps.push_back(Chain);
1735 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1736 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1737 SaveXMMOps.push_back(ALVal);
1739 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1740 FuncInfo->getRegSaveFrameIndex()));
1741 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1742 FuncInfo->getVarArgsFPOffset()));
1744 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1745 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1746 X86::VR128RegisterClass);
1747 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1748 SaveXMMOps.push_back(Val);
1750 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1752 &SaveXMMOps[0], SaveXMMOps.size()));
1755 if (!MemOps.empty())
1756 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1757 &MemOps[0], MemOps.size());
1761 // Some CCs need callee pop.
1762 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1763 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1765 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1766 // If this is an sret function, the return should pop the hidden pointer.
1767 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1768 FuncInfo->setBytesToPopOnReturn(4);
1772 // RegSaveFrameIndex is X86-64 only.
1773 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1774 if (CallConv == CallingConv::X86_FastCall ||
1775 CallConv == CallingConv::X86_ThisCall)
1776 // fastcc functions can't have varargs.
1777 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1784 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1785 SDValue StackPtr, SDValue Arg,
1786 DebugLoc dl, SelectionDAG &DAG,
1787 const CCValAssign &VA,
1788 ISD::ArgFlagsTy Flags) const {
1789 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1790 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1791 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1792 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1793 if (Flags.isByVal()) {
1794 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1796 return DAG.getStore(Chain, dl, Arg, PtrOff,
1797 PseudoSourceValue::getStack(), LocMemOffset,
1801 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1802 /// optimization is performed and it is required.
1804 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1805 SDValue &OutRetAddr, SDValue Chain,
1806 bool IsTailCall, bool Is64Bit,
1807 int FPDiff, DebugLoc dl) const {
1808 // Adjust the Return address stack slot.
1809 EVT VT = getPointerTy();
1810 OutRetAddr = getReturnAddressFrameIndex(DAG);
1812 // Load the "old" Return address.
1813 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1814 return SDValue(OutRetAddr.getNode(), 1);
1817 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1818 /// optimization is performed and it is required (FPDiff!=0).
1820 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1821 SDValue Chain, SDValue RetAddrFrIdx,
1822 bool Is64Bit, int FPDiff, DebugLoc dl) {
1823 // Store the return address to the appropriate stack slot.
1824 if (!FPDiff) return Chain;
1825 // Calculate the new stack slot for the return address.
1826 int SlotSize = Is64Bit ? 8 : 4;
1827 int NewReturnAddrFI =
1828 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1829 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1830 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1831 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1832 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1838 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1839 CallingConv::ID CallConv, bool isVarArg,
1841 const SmallVectorImpl<ISD::OutputArg> &Outs,
1842 const SmallVectorImpl<SDValue> &OutVals,
1843 const SmallVectorImpl<ISD::InputArg> &Ins,
1844 DebugLoc dl, SelectionDAG &DAG,
1845 SmallVectorImpl<SDValue> &InVals) const {
1846 MachineFunction &MF = DAG.getMachineFunction();
1847 bool Is64Bit = Subtarget->is64Bit();
1848 bool IsStructRet = CallIsStructReturn(Outs);
1849 bool IsSibcall = false;
1852 // Check if it's really possible to do a tail call.
1853 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1854 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1855 Outs, OutVals, Ins, DAG);
1857 // Sibcalls are automatically detected tailcalls which do not require
1859 if (!GuaranteedTailCallOpt && isTailCall)
1866 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1867 "Var args not supported with calling convention fastcc or ghc");
1869 // Analyze operands of the call, assigning locations to each operand.
1870 SmallVector<CCValAssign, 16> ArgLocs;
1871 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1872 ArgLocs, *DAG.getContext());
1873 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1875 // Get a count of how many bytes are to be pushed on the stack.
1876 unsigned NumBytes = CCInfo.getNextStackOffset();
1878 // This is a sibcall. The memory operands are available in caller's
1879 // own caller's stack.
1881 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1882 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1885 if (isTailCall && !IsSibcall) {
1886 // Lower arguments at fp - stackoffset + fpdiff.
1887 unsigned NumBytesCallerPushed =
1888 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1889 FPDiff = NumBytesCallerPushed - NumBytes;
1891 // Set the delta of movement of the returnaddr stackslot.
1892 // But only set if delta is greater than previous delta.
1893 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1894 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1898 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1900 SDValue RetAddrFrIdx;
1901 // Load return adress for tail calls.
1902 if (isTailCall && FPDiff)
1903 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1904 Is64Bit, FPDiff, dl);
1906 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1907 SmallVector<SDValue, 8> MemOpChains;
1910 // Walk the register/memloc assignments, inserting copies/loads. In the case
1911 // of tail call optimization arguments are handle later.
1912 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1913 CCValAssign &VA = ArgLocs[i];
1914 EVT RegVT = VA.getLocVT();
1915 SDValue Arg = OutVals[i];
1916 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1917 bool isByVal = Flags.isByVal();
1919 // Promote the value if needed.
1920 switch (VA.getLocInfo()) {
1921 default: llvm_unreachable("Unknown loc info!");
1922 case CCValAssign::Full: break;
1923 case CCValAssign::SExt:
1924 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1926 case CCValAssign::ZExt:
1927 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1929 case CCValAssign::AExt:
1930 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1931 // Special case: passing MMX values in XMM registers.
1932 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1933 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1934 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1936 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1938 case CCValAssign::BCvt:
1939 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1941 case CCValAssign::Indirect: {
1942 // Store the argument.
1943 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1944 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1945 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1946 PseudoSourceValue::getFixedStack(FI), 0,
1953 if (VA.isRegLoc()) {
1954 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1955 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1956 assert(VA.isMemLoc());
1957 if (StackPtr.getNode() == 0)
1958 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1959 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1960 dl, DAG, VA, Flags));
1964 if (!MemOpChains.empty())
1965 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1966 &MemOpChains[0], MemOpChains.size());
1968 // Build a sequence of copy-to-reg nodes chained together with token chain
1969 // and flag operands which copy the outgoing args into registers.
1971 // Tail call byval lowering might overwrite argument registers so in case of
1972 // tail call optimization the copies to registers are lowered later.
1974 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1975 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1976 RegsToPass[i].second, InFlag);
1977 InFlag = Chain.getValue(1);
1980 if (Subtarget->isPICStyleGOT()) {
1981 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1984 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1985 DAG.getNode(X86ISD::GlobalBaseReg,
1986 DebugLoc(), getPointerTy()),
1988 InFlag = Chain.getValue(1);
1990 // If we are tail calling and generating PIC/GOT style code load the
1991 // address of the callee into ECX. The value in ecx is used as target of
1992 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1993 // for tail calls on PIC/GOT architectures. Normally we would just put the
1994 // address of GOT into ebx and then call target@PLT. But for tail calls
1995 // ebx would be restored (since ebx is callee saved) before jumping to the
1998 // Note: The actual moving to ECX is done further down.
1999 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2000 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2001 !G->getGlobal()->hasProtectedVisibility())
2002 Callee = LowerGlobalAddress(Callee, DAG);
2003 else if (isa<ExternalSymbolSDNode>(Callee))
2004 Callee = LowerExternalSymbol(Callee, DAG);
2008 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
2009 // From AMD64 ABI document:
2010 // For calls that may call functions that use varargs or stdargs
2011 // (prototype-less calls or calls to functions containing ellipsis (...) in
2012 // the declaration) %al is used as hidden argument to specify the number
2013 // of SSE registers used. The contents of %al do not need to match exactly
2014 // the number of registers, but must be an ubound on the number of SSE
2015 // registers used and is in the range 0 - 8 inclusive.
2017 // Count the number of XMM registers allocated.
2018 static const unsigned XMMArgRegs[] = {
2019 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2020 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2022 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2023 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2024 && "SSE registers cannot be used when SSE is disabled");
2026 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2027 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2028 InFlag = Chain.getValue(1);
2032 // For tail calls lower the arguments to the 'real' stack slot.
2034 // Force all the incoming stack arguments to be loaded from the stack
2035 // before any new outgoing arguments are stored to the stack, because the
2036 // outgoing stack slots may alias the incoming argument stack slots, and
2037 // the alias isn't otherwise explicit. This is slightly more conservative
2038 // than necessary, because it means that each store effectively depends
2039 // on every argument instead of just those arguments it would clobber.
2040 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2042 SmallVector<SDValue, 8> MemOpChains2;
2045 // Do not flag preceeding copytoreg stuff together with the following stuff.
2047 if (GuaranteedTailCallOpt) {
2048 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2049 CCValAssign &VA = ArgLocs[i];
2052 assert(VA.isMemLoc());
2053 SDValue Arg = OutVals[i];
2054 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2055 // Create frame index.
2056 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2057 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2058 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2059 FIN = DAG.getFrameIndex(FI, getPointerTy());
2061 if (Flags.isByVal()) {
2062 // Copy relative to framepointer.
2063 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2064 if (StackPtr.getNode() == 0)
2065 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2067 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2069 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2073 // Store relative to framepointer.
2074 MemOpChains2.push_back(
2075 DAG.getStore(ArgChain, dl, Arg, FIN,
2076 PseudoSourceValue::getFixedStack(FI), 0,
2082 if (!MemOpChains2.empty())
2083 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2084 &MemOpChains2[0], MemOpChains2.size());
2086 // Copy arguments to their registers.
2087 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2088 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2089 RegsToPass[i].second, InFlag);
2090 InFlag = Chain.getValue(1);
2094 // Store the return address to the appropriate stack slot.
2095 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2099 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2100 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2101 // In the 64-bit large code model, we have to make all calls
2102 // through a register, since the call instruction's 32-bit
2103 // pc-relative offset may not be large enough to hold the whole
2105 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2106 // If the callee is a GlobalAddress node (quite common, every direct call
2107 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2110 // We should use extra load for direct calls to dllimported functions in
2112 const GlobalValue *GV = G->getGlobal();
2113 if (!GV->hasDLLImportLinkage()) {
2114 unsigned char OpFlags = 0;
2116 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2117 // external symbols most go through the PLT in PIC mode. If the symbol
2118 // has hidden or protected visibility, or if it is static or local, then
2119 // we don't need to use the PLT - we can directly call it.
2120 if (Subtarget->isTargetELF() &&
2121 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2122 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2123 OpFlags = X86II::MO_PLT;
2124 } else if (Subtarget->isPICStyleStubAny() &&
2125 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2126 Subtarget->getDarwinVers() < 9) {
2127 // PC-relative references to external symbols should go through $stub,
2128 // unless we're building with the leopard linker or later, which
2129 // automatically synthesizes these stubs.
2130 OpFlags = X86II::MO_DARWIN_STUB;
2133 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2134 G->getOffset(), OpFlags);
2136 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2137 unsigned char OpFlags = 0;
2139 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2140 // symbols should go through the PLT.
2141 if (Subtarget->isTargetELF() &&
2142 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2143 OpFlags = X86II::MO_PLT;
2144 } else if (Subtarget->isPICStyleStubAny() &&
2145 Subtarget->getDarwinVers() < 9) {
2146 // PC-relative references to external symbols should go through $stub,
2147 // unless we're building with the leopard linker or later, which
2148 // automatically synthesizes these stubs.
2149 OpFlags = X86II::MO_DARWIN_STUB;
2152 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2156 // Returns a chain & a flag for retval copy to use.
2157 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2158 SmallVector<SDValue, 8> Ops;
2160 if (!IsSibcall && isTailCall) {
2161 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2162 DAG.getIntPtrConstant(0, true), InFlag);
2163 InFlag = Chain.getValue(1);
2166 Ops.push_back(Chain);
2167 Ops.push_back(Callee);
2170 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2172 // Add argument registers to the end of the list so that they are known live
2174 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2175 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2176 RegsToPass[i].second.getValueType()));
2178 // Add an implicit use GOT pointer in EBX.
2179 if (!isTailCall && Subtarget->isPICStyleGOT())
2180 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2182 // Add an implicit use of AL for x86 vararg functions.
2183 if (Is64Bit && isVarArg)
2184 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2186 if (InFlag.getNode())
2187 Ops.push_back(InFlag);
2191 //// If this is the first return lowered for this function, add the regs
2192 //// to the liveout set for the function.
2193 // This isn't right, although it's probably harmless on x86; liveouts
2194 // should be computed from returns not tail calls. Consider a void
2195 // function making a tail call to a function returning int.
2196 return DAG.getNode(X86ISD::TC_RETURN, dl,
2197 NodeTys, &Ops[0], Ops.size());
2200 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2201 InFlag = Chain.getValue(1);
2203 // Create the CALLSEQ_END node.
2204 unsigned NumBytesForCalleeToPush;
2205 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2206 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2207 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2208 // If this is a call to a struct-return function, the callee
2209 // pops the hidden struct pointer, so we have to push it back.
2210 // This is common for Darwin/X86, Linux & Mingw32 targets.
2211 NumBytesForCalleeToPush = 4;
2213 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2215 // Returns a flag for retval copy to use.
2217 Chain = DAG.getCALLSEQ_END(Chain,
2218 DAG.getIntPtrConstant(NumBytes, true),
2219 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2222 InFlag = Chain.getValue(1);
2225 // Handle result values, copying them out of physregs into vregs that we
2227 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2228 Ins, dl, DAG, InVals);
2232 //===----------------------------------------------------------------------===//
2233 // Fast Calling Convention (tail call) implementation
2234 //===----------------------------------------------------------------------===//
2236 // Like std call, callee cleans arguments, convention except that ECX is
2237 // reserved for storing the tail called function address. Only 2 registers are
2238 // free for argument passing (inreg). Tail call optimization is performed
2240 // * tailcallopt is enabled
2241 // * caller/callee are fastcc
2242 // On X86_64 architecture with GOT-style position independent code only local
2243 // (within module) calls are supported at the moment.
2244 // To keep the stack aligned according to platform abi the function
2245 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2246 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2247 // If a tail called function callee has more arguments than the caller the
2248 // caller needs to make sure that there is room to move the RETADDR to. This is
2249 // achieved by reserving an area the size of the argument delta right after the
2250 // original REtADDR, but before the saved framepointer or the spilled registers
2251 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2263 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2264 /// for a 16 byte align requirement.
2266 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2267 SelectionDAG& DAG) const {
2268 MachineFunction &MF = DAG.getMachineFunction();
2269 const TargetMachine &TM = MF.getTarget();
2270 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2271 unsigned StackAlignment = TFI.getStackAlignment();
2272 uint64_t AlignMask = StackAlignment - 1;
2273 int64_t Offset = StackSize;
2274 uint64_t SlotSize = TD->getPointerSize();
2275 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2276 // Number smaller than 12 so just add the difference.
2277 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2279 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2280 Offset = ((~AlignMask) & Offset) + StackAlignment +
2281 (StackAlignment-SlotSize);
2286 /// MatchingStackOffset - Return true if the given stack call argument is
2287 /// already available in the same position (relatively) of the caller's
2288 /// incoming argument stack.
2290 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2291 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2292 const X86InstrInfo *TII) {
2293 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2295 if (Arg.getOpcode() == ISD::CopyFromReg) {
2296 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2297 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2299 MachineInstr *Def = MRI->getVRegDef(VR);
2302 if (!Flags.isByVal()) {
2303 if (!TII->isLoadFromStackSlot(Def, FI))
2306 unsigned Opcode = Def->getOpcode();
2307 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2308 Def->getOperand(1).isFI()) {
2309 FI = Def->getOperand(1).getIndex();
2310 Bytes = Flags.getByValSize();
2314 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2315 if (Flags.isByVal())
2316 // ByVal argument is passed in as a pointer but it's now being
2317 // dereferenced. e.g.
2318 // define @foo(%struct.X* %A) {
2319 // tail call @bar(%struct.X* byval %A)
2322 SDValue Ptr = Ld->getBasePtr();
2323 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2326 FI = FINode->getIndex();
2330 assert(FI != INT_MAX);
2331 if (!MFI->isFixedObjectIndex(FI))
2333 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2336 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2337 /// for tail call optimization. Targets which want to do tail call
2338 /// optimization should implement this function.
2340 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2341 CallingConv::ID CalleeCC,
2343 bool isCalleeStructRet,
2344 bool isCallerStructRet,
2345 const SmallVectorImpl<ISD::OutputArg> &Outs,
2346 const SmallVectorImpl<SDValue> &OutVals,
2347 const SmallVectorImpl<ISD::InputArg> &Ins,
2348 SelectionDAG& DAG) const {
2349 if (!IsTailCallConvention(CalleeCC) &&
2350 CalleeCC != CallingConv::C)
2353 // If -tailcallopt is specified, make fastcc functions tail-callable.
2354 const MachineFunction &MF = DAG.getMachineFunction();
2355 const Function *CallerF = DAG.getMachineFunction().getFunction();
2356 CallingConv::ID CallerCC = CallerF->getCallingConv();
2357 bool CCMatch = CallerCC == CalleeCC;
2359 if (GuaranteedTailCallOpt) {
2360 if (IsTailCallConvention(CalleeCC) && CCMatch)
2365 // Look for obvious safe cases to perform tail call optimization that do not
2366 // require ABI changes. This is what gcc calls sibcall.
2368 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2369 // emit a special epilogue.
2370 if (RegInfo->needsStackRealignment(MF))
2373 // Do not sibcall optimize vararg calls unless the call site is not passing
2375 if (isVarArg && !Outs.empty())
2378 // Also avoid sibcall optimization if either caller or callee uses struct
2379 // return semantics.
2380 if (isCalleeStructRet || isCallerStructRet)
2383 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2384 // Therefore if it's not used by the call it is not safe to optimize this into
2386 bool Unused = false;
2387 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2394 SmallVector<CCValAssign, 16> RVLocs;
2395 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2396 RVLocs, *DAG.getContext());
2397 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2398 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2399 CCValAssign &VA = RVLocs[i];
2400 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2405 // If the calling conventions do not match, then we'd better make sure the
2406 // results are returned in the same way as what the caller expects.
2408 SmallVector<CCValAssign, 16> RVLocs1;
2409 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2410 RVLocs1, *DAG.getContext());
2411 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2413 SmallVector<CCValAssign, 16> RVLocs2;
2414 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2415 RVLocs2, *DAG.getContext());
2416 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2418 if (RVLocs1.size() != RVLocs2.size())
2420 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2421 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2423 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2425 if (RVLocs1[i].isRegLoc()) {
2426 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2429 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2435 // If the callee takes no arguments then go on to check the results of the
2437 if (!Outs.empty()) {
2438 // Check if stack adjustment is needed. For now, do not do this if any
2439 // argument is passed on the stack.
2440 SmallVector<CCValAssign, 16> ArgLocs;
2441 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2442 ArgLocs, *DAG.getContext());
2443 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2444 if (CCInfo.getNextStackOffset()) {
2445 MachineFunction &MF = DAG.getMachineFunction();
2446 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2448 if (Subtarget->isTargetWin64())
2449 // Win64 ABI has additional complications.
2452 // Check if the arguments are already laid out in the right way as
2453 // the caller's fixed stack objects.
2454 MachineFrameInfo *MFI = MF.getFrameInfo();
2455 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2456 const X86InstrInfo *TII =
2457 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2458 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2459 CCValAssign &VA = ArgLocs[i];
2460 SDValue Arg = OutVals[i];
2461 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2462 if (VA.getLocInfo() == CCValAssign::Indirect)
2464 if (!VA.isRegLoc()) {
2465 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2472 // If the tailcall address may be in a register, then make sure it's
2473 // possible to register allocate for it. In 32-bit, the call address can
2474 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2475 // callee-saved registers are restored. These happen to be the same
2476 // registers used to pass 'inreg' arguments so watch out for those.
2477 if (!Subtarget->is64Bit() &&
2478 !isa<GlobalAddressSDNode>(Callee) &&
2479 !isa<ExternalSymbolSDNode>(Callee)) {
2480 unsigned NumInRegs = 0;
2481 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2482 CCValAssign &VA = ArgLocs[i];
2485 unsigned Reg = VA.getLocReg();
2488 case X86::EAX: case X86::EDX: case X86::ECX:
2489 if (++NumInRegs == 3)
2501 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2502 return X86::createFastISel(funcInfo);
2506 //===----------------------------------------------------------------------===//
2507 // Other Lowering Hooks
2508 //===----------------------------------------------------------------------===//
2511 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2512 MachineFunction &MF = DAG.getMachineFunction();
2513 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2514 int ReturnAddrIndex = FuncInfo->getRAIndex();
2516 if (ReturnAddrIndex == 0) {
2517 // Set up a frame object for the return address.
2518 uint64_t SlotSize = TD->getPointerSize();
2519 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2521 FuncInfo->setRAIndex(ReturnAddrIndex);
2524 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2528 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2529 bool hasSymbolicDisplacement) {
2530 // Offset should fit into 32 bit immediate field.
2531 if (!isInt<32>(Offset))
2534 // If we don't have a symbolic displacement - we don't have any extra
2536 if (!hasSymbolicDisplacement)
2539 // FIXME: Some tweaks might be needed for medium code model.
2540 if (M != CodeModel::Small && M != CodeModel::Kernel)
2543 // For small code model we assume that latest object is 16MB before end of 31
2544 // bits boundary. We may also accept pretty large negative constants knowing
2545 // that all objects are in the positive half of address space.
2546 if (M == CodeModel::Small && Offset < 16*1024*1024)
2549 // For kernel code model we know that all object resist in the negative half
2550 // of 32bits address space. We may not accept negative offsets, since they may
2551 // be just off and we may accept pretty large positive ones.
2552 if (M == CodeModel::Kernel && Offset > 0)
2558 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2559 /// specific condition code, returning the condition code and the LHS/RHS of the
2560 /// comparison to make.
2561 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2562 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2564 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2565 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2566 // X > -1 -> X == 0, jump !sign.
2567 RHS = DAG.getConstant(0, RHS.getValueType());
2568 return X86::COND_NS;
2569 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2570 // X < 0 -> X == 0, jump on sign.
2572 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2574 RHS = DAG.getConstant(0, RHS.getValueType());
2575 return X86::COND_LE;
2579 switch (SetCCOpcode) {
2580 default: llvm_unreachable("Invalid integer condition!");
2581 case ISD::SETEQ: return X86::COND_E;
2582 case ISD::SETGT: return X86::COND_G;
2583 case ISD::SETGE: return X86::COND_GE;
2584 case ISD::SETLT: return X86::COND_L;
2585 case ISD::SETLE: return X86::COND_LE;
2586 case ISD::SETNE: return X86::COND_NE;
2587 case ISD::SETULT: return X86::COND_B;
2588 case ISD::SETUGT: return X86::COND_A;
2589 case ISD::SETULE: return X86::COND_BE;
2590 case ISD::SETUGE: return X86::COND_AE;
2594 // First determine if it is required or is profitable to flip the operands.
2596 // If LHS is a foldable load, but RHS is not, flip the condition.
2597 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2598 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2599 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2600 std::swap(LHS, RHS);
2603 switch (SetCCOpcode) {
2609 std::swap(LHS, RHS);
2613 // On a floating point condition, the flags are set as follows:
2615 // 0 | 0 | 0 | X > Y
2616 // 0 | 0 | 1 | X < Y
2617 // 1 | 0 | 0 | X == Y
2618 // 1 | 1 | 1 | unordered
2619 switch (SetCCOpcode) {
2620 default: llvm_unreachable("Condcode should be pre-legalized away");
2622 case ISD::SETEQ: return X86::COND_E;
2623 case ISD::SETOLT: // flipped
2625 case ISD::SETGT: return X86::COND_A;
2626 case ISD::SETOLE: // flipped
2628 case ISD::SETGE: return X86::COND_AE;
2629 case ISD::SETUGT: // flipped
2631 case ISD::SETLT: return X86::COND_B;
2632 case ISD::SETUGE: // flipped
2634 case ISD::SETLE: return X86::COND_BE;
2636 case ISD::SETNE: return X86::COND_NE;
2637 case ISD::SETUO: return X86::COND_P;
2638 case ISD::SETO: return X86::COND_NP;
2640 case ISD::SETUNE: return X86::COND_INVALID;
2644 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2645 /// code. Current x86 isa includes the following FP cmov instructions:
2646 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2647 static bool hasFPCMov(unsigned X86CC) {
2663 /// isFPImmLegal - Returns true if the target can instruction select the
2664 /// specified FP immediate natively. If false, the legalizer will
2665 /// materialize the FP immediate as a load from a constant pool.
2666 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2667 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2668 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2674 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2675 /// the specified range (L, H].
2676 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2677 return (Val < 0) || (Val >= Low && Val < Hi);
2680 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2681 /// specified value.
2682 static bool isUndefOrEqual(int Val, int CmpVal) {
2683 if (Val < 0 || Val == CmpVal)
2688 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2689 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2690 /// the second operand.
2691 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2692 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2693 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2694 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2695 return (Mask[0] < 2 && Mask[1] < 2);
2699 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2700 SmallVector<int, 8> M;
2702 return ::isPSHUFDMask(M, N->getValueType(0));
2705 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2706 /// is suitable for input to PSHUFHW.
2707 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2708 if (VT != MVT::v8i16)
2711 // Lower quadword copied in order or undef.
2712 for (int i = 0; i != 4; ++i)
2713 if (Mask[i] >= 0 && Mask[i] != i)
2716 // Upper quadword shuffled.
2717 for (int i = 4; i != 8; ++i)
2718 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2724 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2725 SmallVector<int, 8> M;
2727 return ::isPSHUFHWMask(M, N->getValueType(0));
2730 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2731 /// is suitable for input to PSHUFLW.
2732 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2733 if (VT != MVT::v8i16)
2736 // Upper quadword copied in order.
2737 for (int i = 4; i != 8; ++i)
2738 if (Mask[i] >= 0 && Mask[i] != i)
2741 // Lower quadword shuffled.
2742 for (int i = 0; i != 4; ++i)
2749 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2750 SmallVector<int, 8> M;
2752 return ::isPSHUFLWMask(M, N->getValueType(0));
2755 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2756 /// is suitable for input to PALIGNR.
2757 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2759 int i, e = VT.getVectorNumElements();
2761 // Do not handle v2i64 / v2f64 shuffles with palignr.
2762 if (e < 4 || !hasSSSE3)
2765 for (i = 0; i != e; ++i)
2769 // All undef, not a palignr.
2773 // Determine if it's ok to perform a palignr with only the LHS, since we
2774 // don't have access to the actual shuffle elements to see if RHS is undef.
2775 bool Unary = Mask[i] < (int)e;
2776 bool NeedsUnary = false;
2778 int s = Mask[i] - i;
2780 // Check the rest of the elements to see if they are consecutive.
2781 for (++i; i != e; ++i) {
2786 Unary = Unary && (m < (int)e);
2787 NeedsUnary = NeedsUnary || (m < s);
2789 if (NeedsUnary && !Unary)
2791 if (Unary && m != ((s+i) & (e-1)))
2793 if (!Unary && m != (s+i))
2799 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2800 SmallVector<int, 8> M;
2802 return ::isPALIGNRMask(M, N->getValueType(0), true);
2805 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2806 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2807 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2808 int NumElems = VT.getVectorNumElements();
2809 if (NumElems != 2 && NumElems != 4)
2812 int Half = NumElems / 2;
2813 for (int i = 0; i < Half; ++i)
2814 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2816 for (int i = Half; i < NumElems; ++i)
2817 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2823 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2824 SmallVector<int, 8> M;
2826 return ::isSHUFPMask(M, N->getValueType(0));
2829 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2830 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2831 /// half elements to come from vector 1 (which would equal the dest.) and
2832 /// the upper half to come from vector 2.
2833 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2834 int NumElems = VT.getVectorNumElements();
2836 if (NumElems != 2 && NumElems != 4)
2839 int Half = NumElems / 2;
2840 for (int i = 0; i < Half; ++i)
2841 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2843 for (int i = Half; i < NumElems; ++i)
2844 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2849 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2850 SmallVector<int, 8> M;
2852 return isCommutedSHUFPMask(M, N->getValueType(0));
2855 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2856 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2857 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2858 if (N->getValueType(0).getVectorNumElements() != 4)
2861 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2862 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2863 isUndefOrEqual(N->getMaskElt(1), 7) &&
2864 isUndefOrEqual(N->getMaskElt(2), 2) &&
2865 isUndefOrEqual(N->getMaskElt(3), 3);
2868 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2869 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2871 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2872 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2877 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2878 isUndefOrEqual(N->getMaskElt(1), 3) &&
2879 isUndefOrEqual(N->getMaskElt(2), 2) &&
2880 isUndefOrEqual(N->getMaskElt(3), 3);
2883 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2884 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2885 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2886 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2888 if (NumElems != 2 && NumElems != 4)
2891 for (unsigned i = 0; i < NumElems/2; ++i)
2892 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2895 for (unsigned i = NumElems/2; i < NumElems; ++i)
2896 if (!isUndefOrEqual(N->getMaskElt(i), i))
2902 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2903 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2904 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2905 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2907 if (NumElems != 2 && NumElems != 4)
2910 for (unsigned i = 0; i < NumElems/2; ++i)
2911 if (!isUndefOrEqual(N->getMaskElt(i), i))
2914 for (unsigned i = 0; i < NumElems/2; ++i)
2915 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2921 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2922 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2923 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2924 bool V2IsSplat = false) {
2925 int NumElts = VT.getVectorNumElements();
2926 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2929 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2931 int BitI1 = Mask[i+1];
2932 if (!isUndefOrEqual(BitI, j))
2935 if (!isUndefOrEqual(BitI1, NumElts))
2938 if (!isUndefOrEqual(BitI1, j + NumElts))
2945 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2946 SmallVector<int, 8> M;
2948 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2951 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2952 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2953 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2954 bool V2IsSplat = false) {
2955 int NumElts = VT.getVectorNumElements();
2956 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2959 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2961 int BitI1 = Mask[i+1];
2962 if (!isUndefOrEqual(BitI, j + NumElts/2))
2965 if (isUndefOrEqual(BitI1, NumElts))
2968 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2975 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2976 SmallVector<int, 8> M;
2978 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2981 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2982 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2984 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2985 int NumElems = VT.getVectorNumElements();
2986 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2989 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2991 int BitI1 = Mask[i+1];
2992 if (!isUndefOrEqual(BitI, j))
2994 if (!isUndefOrEqual(BitI1, j))
3000 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3001 SmallVector<int, 8> M;
3003 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3006 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3007 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3009 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3010 int NumElems = VT.getVectorNumElements();
3011 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3014 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3016 int BitI1 = Mask[i+1];
3017 if (!isUndefOrEqual(BitI, j))
3019 if (!isUndefOrEqual(BitI1, j))
3025 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3026 SmallVector<int, 8> M;
3028 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3031 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3032 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3033 /// MOVSD, and MOVD, i.e. setting the lowest element.
3034 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3035 if (VT.getVectorElementType().getSizeInBits() < 32)
3038 int NumElts = VT.getVectorNumElements();
3040 if (!isUndefOrEqual(Mask[0], NumElts))
3043 for (int i = 1; i < NumElts; ++i)
3044 if (!isUndefOrEqual(Mask[i], i))
3050 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3051 SmallVector<int, 8> M;
3053 return ::isMOVLMask(M, N->getValueType(0));
3056 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3057 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3058 /// element of vector 2 and the other elements to come from vector 1 in order.
3059 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3060 bool V2IsSplat = false, bool V2IsUndef = false) {
3061 int NumOps = VT.getVectorNumElements();
3062 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3065 if (!isUndefOrEqual(Mask[0], 0))
3068 for (int i = 1; i < NumOps; ++i)
3069 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3070 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3071 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3077 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3078 bool V2IsUndef = false) {
3079 SmallVector<int, 8> M;
3081 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3084 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3085 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3086 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3087 if (N->getValueType(0).getVectorNumElements() != 4)
3090 // Expect 1, 1, 3, 3
3091 for (unsigned i = 0; i < 2; ++i) {
3092 int Elt = N->getMaskElt(i);
3093 if (Elt >= 0 && Elt != 1)
3098 for (unsigned i = 2; i < 4; ++i) {
3099 int Elt = N->getMaskElt(i);
3100 if (Elt >= 0 && Elt != 3)
3105 // Don't use movshdup if it can be done with a shufps.
3106 // FIXME: verify that matching u, u, 3, 3 is what we want.
3110 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3111 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3112 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3113 if (N->getValueType(0).getVectorNumElements() != 4)
3116 // Expect 0, 0, 2, 2
3117 for (unsigned i = 0; i < 2; ++i)
3118 if (N->getMaskElt(i) > 0)
3122 for (unsigned i = 2; i < 4; ++i) {
3123 int Elt = N->getMaskElt(i);
3124 if (Elt >= 0 && Elt != 2)
3129 // Don't use movsldup if it can be done with a shufps.
3133 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3134 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3135 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3136 int e = N->getValueType(0).getVectorNumElements() / 2;
3138 for (int i = 0; i < e; ++i)
3139 if (!isUndefOrEqual(N->getMaskElt(i), i))
3141 for (int i = 0; i < e; ++i)
3142 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3147 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3148 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3149 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3150 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3151 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3153 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3155 for (int i = 0; i < NumOperands; ++i) {
3156 int Val = SVOp->getMaskElt(NumOperands-i-1);
3157 if (Val < 0) Val = 0;
3158 if (Val >= NumOperands) Val -= NumOperands;
3160 if (i != NumOperands - 1)
3166 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3167 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3168 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3169 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3171 // 8 nodes, but we only care about the last 4.
3172 for (unsigned i = 7; i >= 4; --i) {
3173 int Val = SVOp->getMaskElt(i);
3182 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3183 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3184 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3185 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3187 // 8 nodes, but we only care about the first 4.
3188 for (int i = 3; i >= 0; --i) {
3189 int Val = SVOp->getMaskElt(i);
3198 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3199 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3200 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3201 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3202 EVT VVT = N->getValueType(0);
3203 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3207 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3208 Val = SVOp->getMaskElt(i);
3212 return (Val - i) * EltSize;
3215 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3217 bool X86::isZeroNode(SDValue Elt) {
3218 return ((isa<ConstantSDNode>(Elt) &&
3219 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3220 (isa<ConstantFPSDNode>(Elt) &&
3221 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3224 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3225 /// their permute mask.
3226 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3227 SelectionDAG &DAG) {
3228 EVT VT = SVOp->getValueType(0);
3229 unsigned NumElems = VT.getVectorNumElements();
3230 SmallVector<int, 8> MaskVec;
3232 for (unsigned i = 0; i != NumElems; ++i) {
3233 int idx = SVOp->getMaskElt(i);
3235 MaskVec.push_back(idx);
3236 else if (idx < (int)NumElems)
3237 MaskVec.push_back(idx + NumElems);
3239 MaskVec.push_back(idx - NumElems);
3241 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3242 SVOp->getOperand(0), &MaskVec[0]);
3245 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3246 /// the two vector operands have swapped position.
3247 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3248 unsigned NumElems = VT.getVectorNumElements();
3249 for (unsigned i = 0; i != NumElems; ++i) {
3253 else if (idx < (int)NumElems)
3254 Mask[i] = idx + NumElems;
3256 Mask[i] = idx - NumElems;
3260 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3261 /// match movhlps. The lower half elements should come from upper half of
3262 /// V1 (and in order), and the upper half elements should come from the upper
3263 /// half of V2 (and in order).
3264 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3265 if (Op->getValueType(0).getVectorNumElements() != 4)
3267 for (unsigned i = 0, e = 2; i != e; ++i)
3268 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3270 for (unsigned i = 2; i != 4; ++i)
3271 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3276 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3277 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3279 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3280 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3282 N = N->getOperand(0).getNode();
3283 if (!ISD::isNON_EXTLoad(N))
3286 *LD = cast<LoadSDNode>(N);
3290 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3291 /// match movlp{s|d}. The lower half elements should come from lower half of
3292 /// V1 (and in order), and the upper half elements should come from the upper
3293 /// half of V2 (and in order). And since V1 will become the source of the
3294 /// MOVLP, it must be either a vector load or a scalar load to vector.
3295 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3296 ShuffleVectorSDNode *Op) {
3297 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3299 // Is V2 is a vector load, don't do this transformation. We will try to use
3300 // load folding shufps op.
3301 if (ISD::isNON_EXTLoad(V2))
3304 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3306 if (NumElems != 2 && NumElems != 4)
3308 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3309 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3311 for (unsigned i = NumElems/2; i != NumElems; ++i)
3312 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3317 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3319 static bool isSplatVector(SDNode *N) {
3320 if (N->getOpcode() != ISD::BUILD_VECTOR)
3323 SDValue SplatValue = N->getOperand(0);
3324 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3325 if (N->getOperand(i) != SplatValue)
3330 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3331 /// to an zero vector.
3332 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3333 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3334 SDValue V1 = N->getOperand(0);
3335 SDValue V2 = N->getOperand(1);
3336 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3337 for (unsigned i = 0; i != NumElems; ++i) {
3338 int Idx = N->getMaskElt(i);
3339 if (Idx >= (int)NumElems) {
3340 unsigned Opc = V2.getOpcode();
3341 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3343 if (Opc != ISD::BUILD_VECTOR ||
3344 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3346 } else if (Idx >= 0) {
3347 unsigned Opc = V1.getOpcode();
3348 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3350 if (Opc != ISD::BUILD_VECTOR ||
3351 !X86::isZeroNode(V1.getOperand(Idx)))
3358 /// getZeroVector - Returns a vector of specified type with all zero elements.
3360 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3362 assert(VT.isVector() && "Expected a vector type");
3364 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3365 // type. This ensures they get CSE'd.
3367 if (VT.getSizeInBits() == 64) { // MMX
3368 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3369 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3370 } else if (HasSSE2) { // SSE2
3371 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3372 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3374 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3375 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3377 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3380 /// getOnesVector - Returns a vector of specified type with all bits set.
3382 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3383 assert(VT.isVector() && "Expected a vector type");
3385 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3386 // type. This ensures they get CSE'd.
3387 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3389 if (VT.getSizeInBits() == 64) // MMX
3390 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3392 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3393 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3397 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3398 /// that point to V2 points to its first element.
3399 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3400 EVT VT = SVOp->getValueType(0);
3401 unsigned NumElems = VT.getVectorNumElements();
3403 bool Changed = false;
3404 SmallVector<int, 8> MaskVec;
3405 SVOp->getMask(MaskVec);
3407 for (unsigned i = 0; i != NumElems; ++i) {
3408 if (MaskVec[i] > (int)NumElems) {
3409 MaskVec[i] = NumElems;
3414 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3415 SVOp->getOperand(1), &MaskVec[0]);
3416 return SDValue(SVOp, 0);
3419 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3420 /// operation of specified width.
3421 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3423 unsigned NumElems = VT.getVectorNumElements();
3424 SmallVector<int, 8> Mask;
3425 Mask.push_back(NumElems);
3426 for (unsigned i = 1; i != NumElems; ++i)
3428 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3431 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3432 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3434 unsigned NumElems = VT.getVectorNumElements();
3435 SmallVector<int, 8> Mask;
3436 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3438 Mask.push_back(i + NumElems);
3440 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3443 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3444 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3446 unsigned NumElems = VT.getVectorNumElements();
3447 unsigned Half = NumElems/2;
3448 SmallVector<int, 8> Mask;
3449 for (unsigned i = 0; i != Half; ++i) {
3450 Mask.push_back(i + Half);
3451 Mask.push_back(i + NumElems + Half);
3453 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3456 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3457 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3459 if (SV->getValueType(0).getVectorNumElements() <= 4)
3460 return SDValue(SV, 0);
3462 EVT PVT = MVT::v4f32;
3463 EVT VT = SV->getValueType(0);
3464 DebugLoc dl = SV->getDebugLoc();
3465 SDValue V1 = SV->getOperand(0);
3466 int NumElems = VT.getVectorNumElements();
3467 int EltNo = SV->getSplatIndex();
3469 // unpack elements to the correct location
3470 while (NumElems > 4) {
3471 if (EltNo < NumElems/2) {
3472 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3474 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3475 EltNo -= NumElems/2;
3480 // Perform the splat.
3481 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3482 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3483 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3484 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3487 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3488 /// vector of zero or undef vector. This produces a shuffle where the low
3489 /// element of V2 is swizzled into the zero/undef vector, landing at element
3490 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3491 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3492 bool isZero, bool HasSSE2,
3493 SelectionDAG &DAG) {
3494 EVT VT = V2.getValueType();
3496 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3497 unsigned NumElems = VT.getVectorNumElements();
3498 SmallVector<int, 16> MaskVec;
3499 for (unsigned i = 0; i != NumElems; ++i)
3500 // If this is the insertion idx, put the low elt of V2 here.
3501 MaskVec.push_back(i == Idx ? NumElems : i);
3502 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3505 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3506 /// a shuffle that is zero.
3508 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3509 bool Low, SelectionDAG &DAG) {
3510 unsigned NumZeros = 0;
3511 for (int i = 0; i < NumElems; ++i) {
3512 unsigned Index = Low ? i : NumElems-i-1;
3513 int Idx = SVOp->getMaskElt(Index);
3518 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3519 if (Elt.getNode() && X86::isZeroNode(Elt))
3527 /// isVectorShift - Returns true if the shuffle can be implemented as a
3528 /// logical left or right shift of a vector.
3529 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3530 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3531 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3532 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3535 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3538 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3542 bool SeenV1 = false;
3543 bool SeenV2 = false;
3544 for (unsigned i = NumZeros; i < NumElems; ++i) {
3545 unsigned Val = isLeft ? (i - NumZeros) : i;
3546 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3549 unsigned Idx = (unsigned) Idx_;
3559 if (SeenV1 && SeenV2)
3562 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3568 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3570 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3571 unsigned NumNonZero, unsigned NumZero,
3573 const TargetLowering &TLI) {
3577 DebugLoc dl = Op.getDebugLoc();
3580 for (unsigned i = 0; i < 16; ++i) {
3581 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3582 if (ThisIsNonZero && First) {
3584 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3586 V = DAG.getUNDEF(MVT::v8i16);
3591 SDValue ThisElt(0, 0), LastElt(0, 0);
3592 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3593 if (LastIsNonZero) {
3594 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3595 MVT::i16, Op.getOperand(i-1));
3597 if (ThisIsNonZero) {
3598 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3599 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3600 ThisElt, DAG.getConstant(8, MVT::i8));
3602 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3606 if (ThisElt.getNode())
3607 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3608 DAG.getIntPtrConstant(i/2));
3612 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3615 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3617 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3618 unsigned NumNonZero, unsigned NumZero,
3620 const TargetLowering &TLI) {
3624 DebugLoc dl = Op.getDebugLoc();
3627 for (unsigned i = 0; i < 8; ++i) {
3628 bool isNonZero = (NonZeros & (1 << i)) != 0;
3632 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3634 V = DAG.getUNDEF(MVT::v8i16);
3637 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3638 MVT::v8i16, V, Op.getOperand(i),
3639 DAG.getIntPtrConstant(i));
3646 /// getVShift - Return a vector logical shift node.
3648 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3649 unsigned NumBits, SelectionDAG &DAG,
3650 const TargetLowering &TLI, DebugLoc dl) {
3651 bool isMMX = VT.getSizeInBits() == 64;
3652 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3653 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3654 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3655 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3656 DAG.getNode(Opc, dl, ShVT, SrcOp,
3657 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3661 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3662 SelectionDAG &DAG) const {
3664 // Check if the scalar load can be widened into a vector load. And if
3665 // the address is "base + cst" see if the cst can be "absorbed" into
3666 // the shuffle mask.
3667 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3668 SDValue Ptr = LD->getBasePtr();
3669 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3671 EVT PVT = LD->getValueType(0);
3672 if (PVT != MVT::i32 && PVT != MVT::f32)
3677 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3678 FI = FINode->getIndex();
3680 } else if (Ptr.getOpcode() == ISD::ADD &&
3681 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3682 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3683 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3684 Offset = Ptr.getConstantOperandVal(1);
3685 Ptr = Ptr.getOperand(0);
3690 SDValue Chain = LD->getChain();
3691 // Make sure the stack object alignment is at least 16.
3692 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3693 if (DAG.InferPtrAlignment(Ptr) < 16) {
3694 if (MFI->isFixedObjectIndex(FI)) {
3695 // Can't change the alignment. FIXME: It's possible to compute
3696 // the exact stack offset and reference FI + adjust offset instead.
3697 // If someone *really* cares about this. That's the way to implement it.
3700 MFI->setObjectAlignment(FI, 16);
3704 // (Offset % 16) must be multiple of 4. Then address is then
3705 // Ptr + (Offset & ~15).
3708 if ((Offset % 16) & 3)
3710 int64_t StartOffset = Offset & ~15;
3712 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3713 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3715 int EltNo = (Offset - StartOffset) >> 2;
3716 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3717 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3718 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3720 // Canonicalize it to a v4i32 shuffle.
3721 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3722 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3723 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3724 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3730 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3731 /// vector of type 'VT', see if the elements can be replaced by a single large
3732 /// load which has the same value as a build_vector whose operands are 'elts'.
3734 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3736 /// FIXME: we'd also like to handle the case where the last elements are zero
3737 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3738 /// There's even a handy isZeroNode for that purpose.
3739 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3740 DebugLoc &dl, SelectionDAG &DAG) {
3741 EVT EltVT = VT.getVectorElementType();
3742 unsigned NumElems = Elts.size();
3744 LoadSDNode *LDBase = NULL;
3745 unsigned LastLoadedElt = -1U;
3747 // For each element in the initializer, see if we've found a load or an undef.
3748 // If we don't find an initial load element, or later load elements are
3749 // non-consecutive, bail out.
3750 for (unsigned i = 0; i < NumElems; ++i) {
3751 SDValue Elt = Elts[i];
3753 if (!Elt.getNode() ||
3754 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3757 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3759 LDBase = cast<LoadSDNode>(Elt.getNode());
3763 if (Elt.getOpcode() == ISD::UNDEF)
3766 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3767 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3772 // If we have found an entire vector of loads and undefs, then return a large
3773 // load of the entire vector width starting at the base pointer. If we found
3774 // consecutive loads for the low half, generate a vzext_load node.
3775 if (LastLoadedElt == NumElems - 1) {
3776 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3777 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3778 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3779 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3780 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3781 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3782 LDBase->isVolatile(), LDBase->isNonTemporal(),
3783 LDBase->getAlignment());
3784 } else if (NumElems == 4 && LastLoadedElt == 1) {
3785 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3786 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3787 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3788 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3794 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3795 DebugLoc dl = Op.getDebugLoc();
3796 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3797 if (ISD::isBuildVectorAllZeros(Op.getNode())
3798 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3799 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3800 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3801 // eliminated on x86-32 hosts.
3802 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3805 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3806 return getOnesVector(Op.getValueType(), DAG, dl);
3807 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3810 EVT VT = Op.getValueType();
3811 EVT ExtVT = VT.getVectorElementType();
3812 unsigned EVTBits = ExtVT.getSizeInBits();
3814 unsigned NumElems = Op.getNumOperands();
3815 unsigned NumZero = 0;
3816 unsigned NumNonZero = 0;
3817 unsigned NonZeros = 0;
3818 bool IsAllConstants = true;
3819 SmallSet<SDValue, 8> Values;
3820 for (unsigned i = 0; i < NumElems; ++i) {
3821 SDValue Elt = Op.getOperand(i);
3822 if (Elt.getOpcode() == ISD::UNDEF)
3825 if (Elt.getOpcode() != ISD::Constant &&
3826 Elt.getOpcode() != ISD::ConstantFP)
3827 IsAllConstants = false;
3828 if (X86::isZeroNode(Elt))
3831 NonZeros |= (1 << i);
3836 if (NumNonZero == 0) {
3837 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3838 return DAG.getUNDEF(VT);
3841 // Special case for single non-zero, non-undef, element.
3842 if (NumNonZero == 1) {
3843 unsigned Idx = CountTrailingZeros_32(NonZeros);
3844 SDValue Item = Op.getOperand(Idx);
3846 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3847 // the value are obviously zero, truncate the value to i32 and do the
3848 // insertion that way. Only do this if the value is non-constant or if the
3849 // value is a constant being inserted into element 0. It is cheaper to do
3850 // a constant pool load than it is to do a movd + shuffle.
3851 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3852 (!IsAllConstants || Idx == 0)) {
3853 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3854 // Handle MMX and SSE both.
3855 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3856 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3858 // Truncate the value (which may itself be a constant) to i32, and
3859 // convert it to a vector with movd (S2V+shuffle to zero extend).
3860 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3861 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3862 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3863 Subtarget->hasSSE2(), DAG);
3865 // Now we have our 32-bit value zero extended in the low element of
3866 // a vector. If Idx != 0, swizzle it into place.
3868 SmallVector<int, 4> Mask;
3869 Mask.push_back(Idx);
3870 for (unsigned i = 1; i != VecElts; ++i)
3872 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3873 DAG.getUNDEF(Item.getValueType()),
3876 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3880 // If we have a constant or non-constant insertion into the low element of
3881 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3882 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3883 // depending on what the source datatype is.
3886 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3887 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3888 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3889 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3890 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3891 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3893 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3894 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3895 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3896 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3897 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3898 Subtarget->hasSSE2(), DAG);
3899 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3903 // Is it a vector logical left shift?
3904 if (NumElems == 2 && Idx == 1 &&
3905 X86::isZeroNode(Op.getOperand(0)) &&
3906 !X86::isZeroNode(Op.getOperand(1))) {
3907 unsigned NumBits = VT.getSizeInBits();
3908 return getVShift(true, VT,
3909 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3910 VT, Op.getOperand(1)),
3911 NumBits/2, DAG, *this, dl);
3914 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3917 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3918 // is a non-constant being inserted into an element other than the low one,
3919 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3920 // movd/movss) to move this into the low element, then shuffle it into
3922 if (EVTBits == 32) {
3923 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3925 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3926 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3927 Subtarget->hasSSE2(), DAG);
3928 SmallVector<int, 8> MaskVec;
3929 for (unsigned i = 0; i < NumElems; i++)
3930 MaskVec.push_back(i == Idx ? 0 : 1);
3931 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3935 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3936 if (Values.size() == 1) {
3937 if (EVTBits == 32) {
3938 // Instead of a shuffle like this:
3939 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3940 // Check if it's possible to issue this instead.
3941 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3942 unsigned Idx = CountTrailingZeros_32(NonZeros);
3943 SDValue Item = Op.getOperand(Idx);
3944 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3945 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3950 // A vector full of immediates; various special cases are already
3951 // handled, so this is best done with a single constant-pool load.
3955 // Let legalizer expand 2-wide build_vectors.
3956 if (EVTBits == 64) {
3957 if (NumNonZero == 1) {
3958 // One half is zero or undef.
3959 unsigned Idx = CountTrailingZeros_32(NonZeros);
3960 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3961 Op.getOperand(Idx));
3962 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3963 Subtarget->hasSSE2(), DAG);
3968 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3969 if (EVTBits == 8 && NumElems == 16) {
3970 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3972 if (V.getNode()) return V;
3975 if (EVTBits == 16 && NumElems == 8) {
3976 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3978 if (V.getNode()) return V;
3981 // If element VT is == 32 bits, turn it into a number of shuffles.
3982 SmallVector<SDValue, 8> V;
3984 if (NumElems == 4 && NumZero > 0) {
3985 for (unsigned i = 0; i < 4; ++i) {
3986 bool isZero = !(NonZeros & (1 << i));
3988 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3990 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3993 for (unsigned i = 0; i < 2; ++i) {
3994 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3997 V[i] = V[i*2]; // Must be a zero vector.
4000 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4003 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4006 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4011 SmallVector<int, 8> MaskVec;
4012 bool Reverse = (NonZeros & 0x3) == 2;
4013 for (unsigned i = 0; i < 2; ++i)
4014 MaskVec.push_back(Reverse ? 1-i : i);
4015 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4016 for (unsigned i = 0; i < 2; ++i)
4017 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4018 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4021 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4022 // Check for a build vector of consecutive loads.
4023 for (unsigned i = 0; i < NumElems; ++i)
4024 V[i] = Op.getOperand(i);
4026 // Check for elements which are consecutive loads.
4027 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4031 // For SSE 4.1, use inserts into undef.
4032 if (getSubtarget()->hasSSE41()) {
4033 V[0] = DAG.getUNDEF(VT);
4034 for (unsigned i = 0; i < NumElems; ++i)
4035 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4036 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
4037 Op.getOperand(i), DAG.getIntPtrConstant(i));
4041 // Otherwise, expand into a number of unpckl*
4043 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4044 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4045 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4046 for (unsigned i = 0; i < NumElems; ++i)
4047 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4049 while (NumElems != 0) {
4050 for (unsigned i = 0; i < NumElems; ++i)
4051 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4060 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4061 // We support concatenate two MMX registers and place them in a MMX
4062 // register. This is better than doing a stack convert.
4063 DebugLoc dl = Op.getDebugLoc();
4064 EVT ResVT = Op.getValueType();
4065 assert(Op.getNumOperands() == 2);
4066 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4067 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4069 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4070 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4071 InVec = Op.getOperand(1);
4072 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4073 unsigned NumElts = ResVT.getVectorNumElements();
4074 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4075 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4076 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4078 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4079 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4080 Mask[0] = 0; Mask[1] = 2;
4081 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4083 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4086 // v8i16 shuffles - Prefer shuffles in the following order:
4087 // 1. [all] pshuflw, pshufhw, optional move
4088 // 2. [ssse3] 1 x pshufb
4089 // 3. [ssse3] 2 x pshufb + 1 x por
4090 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4092 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4094 const X86TargetLowering &TLI) {
4095 SDValue V1 = SVOp->getOperand(0);
4096 SDValue V2 = SVOp->getOperand(1);
4097 DebugLoc dl = SVOp->getDebugLoc();
4098 SmallVector<int, 8> MaskVals;
4100 // Determine if more than 1 of the words in each of the low and high quadwords
4101 // of the result come from the same quadword of one of the two inputs. Undef
4102 // mask values count as coming from any quadword, for better codegen.
4103 SmallVector<unsigned, 4> LoQuad(4);
4104 SmallVector<unsigned, 4> HiQuad(4);
4105 BitVector InputQuads(4);
4106 for (unsigned i = 0; i < 8; ++i) {
4107 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4108 int EltIdx = SVOp->getMaskElt(i);
4109 MaskVals.push_back(EltIdx);
4118 InputQuads.set(EltIdx / 4);
4121 int BestLoQuad = -1;
4122 unsigned MaxQuad = 1;
4123 for (unsigned i = 0; i < 4; ++i) {
4124 if (LoQuad[i] > MaxQuad) {
4126 MaxQuad = LoQuad[i];
4130 int BestHiQuad = -1;
4132 for (unsigned i = 0; i < 4; ++i) {
4133 if (HiQuad[i] > MaxQuad) {
4135 MaxQuad = HiQuad[i];
4139 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4140 // of the two input vectors, shuffle them into one input vector so only a
4141 // single pshufb instruction is necessary. If There are more than 2 input
4142 // quads, disable the next transformation since it does not help SSSE3.
4143 bool V1Used = InputQuads[0] || InputQuads[1];
4144 bool V2Used = InputQuads[2] || InputQuads[3];
4145 if (TLI.getSubtarget()->hasSSSE3()) {
4146 if (InputQuads.count() == 2 && V1Used && V2Used) {
4147 BestLoQuad = InputQuads.find_first();
4148 BestHiQuad = InputQuads.find_next(BestLoQuad);
4150 if (InputQuads.count() > 2) {
4156 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4157 // the shuffle mask. If a quad is scored as -1, that means that it contains
4158 // words from all 4 input quadwords.
4160 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4161 SmallVector<int, 8> MaskV;
4162 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4163 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4164 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4165 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4166 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4167 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4169 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4170 // source words for the shuffle, to aid later transformations.
4171 bool AllWordsInNewV = true;
4172 bool InOrder[2] = { true, true };
4173 for (unsigned i = 0; i != 8; ++i) {
4174 int idx = MaskVals[i];
4176 InOrder[i/4] = false;
4177 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4179 AllWordsInNewV = false;
4183 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4184 if (AllWordsInNewV) {
4185 for (int i = 0; i != 8; ++i) {
4186 int idx = MaskVals[i];
4189 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4190 if ((idx != i) && idx < 4)
4192 if ((idx != i) && idx > 3)
4201 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4202 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4203 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4204 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4205 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4209 // If we have SSSE3, and all words of the result are from 1 input vector,
4210 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4211 // is present, fall back to case 4.
4212 if (TLI.getSubtarget()->hasSSSE3()) {
4213 SmallVector<SDValue,16> pshufbMask;
4215 // If we have elements from both input vectors, set the high bit of the
4216 // shuffle mask element to zero out elements that come from V2 in the V1
4217 // mask, and elements that come from V1 in the V2 mask, so that the two
4218 // results can be OR'd together.
4219 bool TwoInputs = V1Used && V2Used;
4220 for (unsigned i = 0; i != 8; ++i) {
4221 int EltIdx = MaskVals[i] * 2;
4222 if (TwoInputs && (EltIdx >= 16)) {
4223 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4224 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4227 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4228 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4230 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4231 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4232 DAG.getNode(ISD::BUILD_VECTOR, dl,
4233 MVT::v16i8, &pshufbMask[0], 16));
4235 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4237 // Calculate the shuffle mask for the second input, shuffle it, and
4238 // OR it with the first shuffled input.
4240 for (unsigned i = 0; i != 8; ++i) {
4241 int EltIdx = MaskVals[i] * 2;
4243 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4244 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4247 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4248 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4250 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4251 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4252 DAG.getNode(ISD::BUILD_VECTOR, dl,
4253 MVT::v16i8, &pshufbMask[0], 16));
4254 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4255 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4258 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4259 // and update MaskVals with new element order.
4260 BitVector InOrder(8);
4261 if (BestLoQuad >= 0) {
4262 SmallVector<int, 8> MaskV;
4263 for (int i = 0; i != 4; ++i) {
4264 int idx = MaskVals[i];
4266 MaskV.push_back(-1);
4268 } else if ((idx / 4) == BestLoQuad) {
4269 MaskV.push_back(idx & 3);
4272 MaskV.push_back(-1);
4275 for (unsigned i = 4; i != 8; ++i)
4277 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4281 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4282 // and update MaskVals with the new element order.
4283 if (BestHiQuad >= 0) {
4284 SmallVector<int, 8> MaskV;
4285 for (unsigned i = 0; i != 4; ++i)
4287 for (unsigned i = 4; i != 8; ++i) {
4288 int idx = MaskVals[i];
4290 MaskV.push_back(-1);
4292 } else if ((idx / 4) == BestHiQuad) {
4293 MaskV.push_back((idx & 3) + 4);
4296 MaskV.push_back(-1);
4299 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4303 // In case BestHi & BestLo were both -1, which means each quadword has a word
4304 // from each of the four input quadwords, calculate the InOrder bitvector now
4305 // before falling through to the insert/extract cleanup.
4306 if (BestLoQuad == -1 && BestHiQuad == -1) {
4308 for (int i = 0; i != 8; ++i)
4309 if (MaskVals[i] < 0 || MaskVals[i] == i)
4313 // The other elements are put in the right place using pextrw and pinsrw.
4314 for (unsigned i = 0; i != 8; ++i) {
4317 int EltIdx = MaskVals[i];
4320 SDValue ExtOp = (EltIdx < 8)
4321 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4322 DAG.getIntPtrConstant(EltIdx))
4323 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4324 DAG.getIntPtrConstant(EltIdx - 8));
4325 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4326 DAG.getIntPtrConstant(i));
4331 // v16i8 shuffles - Prefer shuffles in the following order:
4332 // 1. [ssse3] 1 x pshufb
4333 // 2. [ssse3] 2 x pshufb + 1 x por
4334 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4336 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4338 const X86TargetLowering &TLI) {
4339 SDValue V1 = SVOp->getOperand(0);
4340 SDValue V2 = SVOp->getOperand(1);
4341 DebugLoc dl = SVOp->getDebugLoc();
4342 SmallVector<int, 16> MaskVals;
4343 SVOp->getMask(MaskVals);
4345 // If we have SSSE3, case 1 is generated when all result bytes come from
4346 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4347 // present, fall back to case 3.
4348 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4351 for (unsigned i = 0; i < 16; ++i) {
4352 int EltIdx = MaskVals[i];
4361 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4362 if (TLI.getSubtarget()->hasSSSE3()) {
4363 SmallVector<SDValue,16> pshufbMask;
4365 // If all result elements are from one input vector, then only translate
4366 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4368 // Otherwise, we have elements from both input vectors, and must zero out
4369 // elements that come from V2 in the first mask, and V1 in the second mask
4370 // so that we can OR them together.
4371 bool TwoInputs = !(V1Only || V2Only);
4372 for (unsigned i = 0; i != 16; ++i) {
4373 int EltIdx = MaskVals[i];
4374 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4375 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4378 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4380 // If all the elements are from V2, assign it to V1 and return after
4381 // building the first pshufb.
4384 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4385 DAG.getNode(ISD::BUILD_VECTOR, dl,
4386 MVT::v16i8, &pshufbMask[0], 16));
4390 // Calculate the shuffle mask for the second input, shuffle it, and
4391 // OR it with the first shuffled input.
4393 for (unsigned i = 0; i != 16; ++i) {
4394 int EltIdx = MaskVals[i];
4396 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4399 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4401 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4402 DAG.getNode(ISD::BUILD_VECTOR, dl,
4403 MVT::v16i8, &pshufbMask[0], 16));
4404 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4407 // No SSSE3 - Calculate in place words and then fix all out of place words
4408 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4409 // the 16 different words that comprise the two doublequadword input vectors.
4410 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4411 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4412 SDValue NewV = V2Only ? V2 : V1;
4413 for (int i = 0; i != 8; ++i) {
4414 int Elt0 = MaskVals[i*2];
4415 int Elt1 = MaskVals[i*2+1];
4417 // This word of the result is all undef, skip it.
4418 if (Elt0 < 0 && Elt1 < 0)
4421 // This word of the result is already in the correct place, skip it.
4422 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4424 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4427 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4428 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4431 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4432 // using a single extract together, load it and store it.
4433 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4434 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4435 DAG.getIntPtrConstant(Elt1 / 2));
4436 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4437 DAG.getIntPtrConstant(i));
4441 // If Elt1 is defined, extract it from the appropriate source. If the
4442 // source byte is not also odd, shift the extracted word left 8 bits
4443 // otherwise clear the bottom 8 bits if we need to do an or.
4445 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4446 DAG.getIntPtrConstant(Elt1 / 2));
4447 if ((Elt1 & 1) == 0)
4448 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4449 DAG.getConstant(8, TLI.getShiftAmountTy()));
4451 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4452 DAG.getConstant(0xFF00, MVT::i16));
4454 // If Elt0 is defined, extract it from the appropriate source. If the
4455 // source byte is not also even, shift the extracted word right 8 bits. If
4456 // Elt1 was also defined, OR the extracted values together before
4457 // inserting them in the result.
4459 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4460 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4461 if ((Elt0 & 1) != 0)
4462 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4463 DAG.getConstant(8, TLI.getShiftAmountTy()));
4465 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4466 DAG.getConstant(0x00FF, MVT::i16));
4467 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4470 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4471 DAG.getIntPtrConstant(i));
4473 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4476 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4477 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4478 /// done when every pair / quad of shuffle mask elements point to elements in
4479 /// the right sequence. e.g.
4480 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4482 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4484 const TargetLowering &TLI, DebugLoc dl) {
4485 EVT VT = SVOp->getValueType(0);
4486 SDValue V1 = SVOp->getOperand(0);
4487 SDValue V2 = SVOp->getOperand(1);
4488 unsigned NumElems = VT.getVectorNumElements();
4489 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4490 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4492 switch (VT.getSimpleVT().SimpleTy) {
4493 default: assert(false && "Unexpected!");
4494 case MVT::v4f32: NewVT = MVT::v2f64; break;
4495 case MVT::v4i32: NewVT = MVT::v2i64; break;
4496 case MVT::v8i16: NewVT = MVT::v4i32; break;
4497 case MVT::v16i8: NewVT = MVT::v4i32; break;
4500 if (NewWidth == 2) {
4506 int Scale = NumElems / NewWidth;
4507 SmallVector<int, 8> MaskVec;
4508 for (unsigned i = 0; i < NumElems; i += Scale) {
4510 for (int j = 0; j < Scale; ++j) {
4511 int EltIdx = SVOp->getMaskElt(i+j);
4515 StartIdx = EltIdx - (EltIdx % Scale);
4516 if (EltIdx != StartIdx + j)
4520 MaskVec.push_back(-1);
4522 MaskVec.push_back(StartIdx / Scale);
4525 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4526 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4527 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4530 /// getVZextMovL - Return a zero-extending vector move low node.
4532 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4533 SDValue SrcOp, SelectionDAG &DAG,
4534 const X86Subtarget *Subtarget, DebugLoc dl) {
4535 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4536 LoadSDNode *LD = NULL;
4537 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4538 LD = dyn_cast<LoadSDNode>(SrcOp);
4540 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4542 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4543 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4544 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4545 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4546 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4548 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4549 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4550 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4551 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4559 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4560 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4561 DAG.getNode(ISD::BIT_CONVERT, dl,
4565 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4568 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4569 SDValue V1 = SVOp->getOperand(0);
4570 SDValue V2 = SVOp->getOperand(1);
4571 DebugLoc dl = SVOp->getDebugLoc();
4572 EVT VT = SVOp->getValueType(0);
4574 SmallVector<std::pair<int, int>, 8> Locs;
4576 SmallVector<int, 8> Mask1(4U, -1);
4577 SmallVector<int, 8> PermMask;
4578 SVOp->getMask(PermMask);
4582 for (unsigned i = 0; i != 4; ++i) {
4583 int Idx = PermMask[i];
4585 Locs[i] = std::make_pair(-1, -1);
4587 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4589 Locs[i] = std::make_pair(0, NumLo);
4593 Locs[i] = std::make_pair(1, NumHi);
4595 Mask1[2+NumHi] = Idx;
4601 if (NumLo <= 2 && NumHi <= 2) {
4602 // If no more than two elements come from either vector. This can be
4603 // implemented with two shuffles. First shuffle gather the elements.
4604 // The second shuffle, which takes the first shuffle as both of its
4605 // vector operands, put the elements into the right order.
4606 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4608 SmallVector<int, 8> Mask2(4U, -1);
4610 for (unsigned i = 0; i != 4; ++i) {
4611 if (Locs[i].first == -1)
4614 unsigned Idx = (i < 2) ? 0 : 4;
4615 Idx += Locs[i].first * 2 + Locs[i].second;
4620 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4621 } else if (NumLo == 3 || NumHi == 3) {
4622 // Otherwise, we must have three elements from one vector, call it X, and
4623 // one element from the other, call it Y. First, use a shufps to build an
4624 // intermediate vector with the one element from Y and the element from X
4625 // that will be in the same half in the final destination (the indexes don't
4626 // matter). Then, use a shufps to build the final vector, taking the half
4627 // containing the element from Y from the intermediate, and the other half
4630 // Normalize it so the 3 elements come from V1.
4631 CommuteVectorShuffleMask(PermMask, VT);
4635 // Find the element from V2.
4637 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4638 int Val = PermMask[HiIndex];
4645 Mask1[0] = PermMask[HiIndex];
4647 Mask1[2] = PermMask[HiIndex^1];
4649 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4652 Mask1[0] = PermMask[0];
4653 Mask1[1] = PermMask[1];
4654 Mask1[2] = HiIndex & 1 ? 6 : 4;
4655 Mask1[3] = HiIndex & 1 ? 4 : 6;
4656 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4658 Mask1[0] = HiIndex & 1 ? 2 : 0;
4659 Mask1[1] = HiIndex & 1 ? 0 : 2;
4660 Mask1[2] = PermMask[2];
4661 Mask1[3] = PermMask[3];
4666 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4670 // Break it into (shuffle shuffle_hi, shuffle_lo).
4672 SmallVector<int,8> LoMask(4U, -1);
4673 SmallVector<int,8> HiMask(4U, -1);
4675 SmallVector<int,8> *MaskPtr = &LoMask;
4676 unsigned MaskIdx = 0;
4679 for (unsigned i = 0; i != 4; ++i) {
4686 int Idx = PermMask[i];
4688 Locs[i] = std::make_pair(-1, -1);
4689 } else if (Idx < 4) {
4690 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4691 (*MaskPtr)[LoIdx] = Idx;
4694 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4695 (*MaskPtr)[HiIdx] = Idx;
4700 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4701 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4702 SmallVector<int, 8> MaskOps;
4703 for (unsigned i = 0; i != 4; ++i) {
4704 if (Locs[i].first == -1) {
4705 MaskOps.push_back(-1);
4707 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4708 MaskOps.push_back(Idx);
4711 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4715 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4716 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4717 SDValue V1 = Op.getOperand(0);
4718 SDValue V2 = Op.getOperand(1);
4719 EVT VT = Op.getValueType();
4720 DebugLoc dl = Op.getDebugLoc();
4721 unsigned NumElems = VT.getVectorNumElements();
4722 bool isMMX = VT.getSizeInBits() == 64;
4723 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4724 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4725 bool V1IsSplat = false;
4726 bool V2IsSplat = false;
4728 if (isZeroShuffle(SVOp))
4729 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4731 // Promote splats to v4f32.
4732 if (SVOp->isSplat()) {
4733 if (isMMX || NumElems < 4)
4735 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4738 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4740 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4741 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4742 if (NewOp.getNode())
4743 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4744 LowerVECTOR_SHUFFLE(NewOp, DAG));
4745 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4746 // FIXME: Figure out a cleaner way to do this.
4747 // Try to make use of movq to zero out the top part.
4748 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4749 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4750 if (NewOp.getNode()) {
4751 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4752 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4753 DAG, Subtarget, dl);
4755 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4756 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4757 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4758 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4759 DAG, Subtarget, dl);
4763 if (X86::isPSHUFDMask(SVOp))
4766 // Check if this can be converted into a logical shift.
4767 bool isLeft = false;
4770 bool isShift = getSubtarget()->hasSSE2() &&
4771 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4772 if (isShift && ShVal.hasOneUse()) {
4773 // If the shifted value has multiple uses, it may be cheaper to use
4774 // v_set0 + movlhps or movhlps, etc.
4775 EVT EltVT = VT.getVectorElementType();
4776 ShAmt *= EltVT.getSizeInBits();
4777 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4780 if (X86::isMOVLMask(SVOp)) {
4783 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4784 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4789 // FIXME: fold these into legal mask.
4790 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4791 X86::isMOVSLDUPMask(SVOp) ||
4792 X86::isMOVHLPSMask(SVOp) ||
4793 X86::isMOVLHPSMask(SVOp) ||
4794 X86::isMOVLPMask(SVOp)))
4797 if (ShouldXformToMOVHLPS(SVOp) ||
4798 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4799 return CommuteVectorShuffle(SVOp, DAG);
4802 // No better options. Use a vshl / vsrl.
4803 EVT EltVT = VT.getVectorElementType();
4804 ShAmt *= EltVT.getSizeInBits();
4805 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4808 bool Commuted = false;
4809 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4810 // 1,1,1,1 -> v8i16 though.
4811 V1IsSplat = isSplatVector(V1.getNode());
4812 V2IsSplat = isSplatVector(V2.getNode());
4814 // Canonicalize the splat or undef, if present, to be on the RHS.
4815 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4816 Op = CommuteVectorShuffle(SVOp, DAG);
4817 SVOp = cast<ShuffleVectorSDNode>(Op);
4818 V1 = SVOp->getOperand(0);
4819 V2 = SVOp->getOperand(1);
4820 std::swap(V1IsSplat, V2IsSplat);
4821 std::swap(V1IsUndef, V2IsUndef);
4825 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4826 // Shuffling low element of v1 into undef, just return v1.
4829 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4830 // the instruction selector will not match, so get a canonical MOVL with
4831 // swapped operands to undo the commute.
4832 return getMOVL(DAG, dl, VT, V2, V1);
4835 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4836 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4837 X86::isUNPCKLMask(SVOp) ||
4838 X86::isUNPCKHMask(SVOp))
4842 // Normalize mask so all entries that point to V2 points to its first
4843 // element then try to match unpck{h|l} again. If match, return a
4844 // new vector_shuffle with the corrected mask.
4845 SDValue NewMask = NormalizeMask(SVOp, DAG);
4846 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4847 if (NSVOp != SVOp) {
4848 if (X86::isUNPCKLMask(NSVOp, true)) {
4850 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4857 // Commute is back and try unpck* again.
4858 // FIXME: this seems wrong.
4859 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4860 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4861 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4862 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4863 X86::isUNPCKLMask(NewSVOp) ||
4864 X86::isUNPCKHMask(NewSVOp))
4868 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4870 // Normalize the node to match x86 shuffle ops if needed
4871 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4872 return CommuteVectorShuffle(SVOp, DAG);
4874 // Check for legal shuffle and return?
4875 SmallVector<int, 16> PermMask;
4876 SVOp->getMask(PermMask);
4877 if (isShuffleMaskLegal(PermMask, VT))
4880 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4881 if (VT == MVT::v8i16) {
4882 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4883 if (NewOp.getNode())
4887 if (VT == MVT::v16i8) {
4888 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4889 if (NewOp.getNode())
4893 // Handle all 4 wide cases with a number of shuffles except for MMX.
4894 if (NumElems == 4 && !isMMX)
4895 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4901 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4902 SelectionDAG &DAG) const {
4903 EVT VT = Op.getValueType();
4904 DebugLoc dl = Op.getDebugLoc();
4905 if (VT.getSizeInBits() == 8) {
4906 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4907 Op.getOperand(0), Op.getOperand(1));
4908 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4909 DAG.getValueType(VT));
4910 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4911 } else if (VT.getSizeInBits() == 16) {
4912 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4913 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4915 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4916 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4917 DAG.getNode(ISD::BIT_CONVERT, dl,
4921 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4922 Op.getOperand(0), Op.getOperand(1));
4923 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4924 DAG.getValueType(VT));
4925 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4926 } else if (VT == MVT::f32) {
4927 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4928 // the result back to FR32 register. It's only worth matching if the
4929 // result has a single use which is a store or a bitcast to i32. And in
4930 // the case of a store, it's not worth it if the index is a constant 0,
4931 // because a MOVSSmr can be used instead, which is smaller and faster.
4932 if (!Op.hasOneUse())
4934 SDNode *User = *Op.getNode()->use_begin();
4935 if ((User->getOpcode() != ISD::STORE ||
4936 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4937 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4938 (User->getOpcode() != ISD::BIT_CONVERT ||
4939 User->getValueType(0) != MVT::i32))
4941 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4942 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4945 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4946 } else if (VT == MVT::i32) {
4947 // ExtractPS works with constant index.
4948 if (isa<ConstantSDNode>(Op.getOperand(1)))
4956 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4957 SelectionDAG &DAG) const {
4958 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4961 if (Subtarget->hasSSE41()) {
4962 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4967 EVT VT = Op.getValueType();
4968 DebugLoc dl = Op.getDebugLoc();
4969 // TODO: handle v16i8.
4970 if (VT.getSizeInBits() == 16) {
4971 SDValue Vec = Op.getOperand(0);
4972 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4974 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4975 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4976 DAG.getNode(ISD::BIT_CONVERT, dl,
4979 // Transform it so it match pextrw which produces a 32-bit result.
4980 EVT EltVT = MVT::i32;
4981 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4982 Op.getOperand(0), Op.getOperand(1));
4983 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4984 DAG.getValueType(VT));
4985 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4986 } else if (VT.getSizeInBits() == 32) {
4987 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4991 // SHUFPS the element to the lowest double word, then movss.
4992 int Mask[4] = { Idx, -1, -1, -1 };
4993 EVT VVT = Op.getOperand(0).getValueType();
4994 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4995 DAG.getUNDEF(VVT), Mask);
4996 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4997 DAG.getIntPtrConstant(0));
4998 } else if (VT.getSizeInBits() == 64) {
4999 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5000 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5001 // to match extract_elt for f64.
5002 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5006 // UNPCKHPD the element to the lowest double word, then movsd.
5007 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5008 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5009 int Mask[2] = { 1, -1 };
5010 EVT VVT = Op.getOperand(0).getValueType();
5011 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5012 DAG.getUNDEF(VVT), Mask);
5013 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5014 DAG.getIntPtrConstant(0));
5021 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5022 SelectionDAG &DAG) const {
5023 EVT VT = Op.getValueType();
5024 EVT EltVT = VT.getVectorElementType();
5025 DebugLoc dl = Op.getDebugLoc();
5027 SDValue N0 = Op.getOperand(0);
5028 SDValue N1 = Op.getOperand(1);
5029 SDValue N2 = Op.getOperand(2);
5031 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5032 isa<ConstantSDNode>(N2)) {
5034 if (VT == MVT::v8i16)
5035 Opc = X86ISD::PINSRW;
5036 else if (VT == MVT::v4i16)
5037 Opc = X86ISD::MMX_PINSRW;
5038 else if (VT == MVT::v16i8)
5039 Opc = X86ISD::PINSRB;
5041 Opc = X86ISD::PINSRB;
5043 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5045 if (N1.getValueType() != MVT::i32)
5046 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5047 if (N2.getValueType() != MVT::i32)
5048 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5049 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5050 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5051 // Bits [7:6] of the constant are the source select. This will always be
5052 // zero here. The DAG Combiner may combine an extract_elt index into these
5053 // bits. For example (insert (extract, 3), 2) could be matched by putting
5054 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5055 // Bits [5:4] of the constant are the destination select. This is the
5056 // value of the incoming immediate.
5057 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5058 // combine either bitwise AND or insert of float 0.0 to set these bits.
5059 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5060 // Create this as a scalar to vector..
5061 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5062 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5063 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5064 // PINSR* works with constant index.
5071 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5072 EVT VT = Op.getValueType();
5073 EVT EltVT = VT.getVectorElementType();
5075 if (Subtarget->hasSSE41())
5076 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5078 if (EltVT == MVT::i8)
5081 DebugLoc dl = Op.getDebugLoc();
5082 SDValue N0 = Op.getOperand(0);
5083 SDValue N1 = Op.getOperand(1);
5084 SDValue N2 = Op.getOperand(2);
5086 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5087 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5088 // as its second argument.
5089 if (N1.getValueType() != MVT::i32)
5090 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5091 if (N2.getValueType() != MVT::i32)
5092 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5093 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5094 dl, VT, N0, N1, N2);
5100 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5101 DebugLoc dl = Op.getDebugLoc();
5103 if (Op.getValueType() == MVT::v1i64 &&
5104 Op.getOperand(0).getValueType() == MVT::i64)
5105 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5107 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5108 EVT VT = MVT::v2i32;
5109 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5116 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5117 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5120 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5121 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5122 // one of the above mentioned nodes. It has to be wrapped because otherwise
5123 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5124 // be used to form addressing mode. These wrapped nodes will be selected
5127 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5128 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5130 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5132 unsigned char OpFlag = 0;
5133 unsigned WrapperKind = X86ISD::Wrapper;
5134 CodeModel::Model M = getTargetMachine().getCodeModel();
5136 if (Subtarget->isPICStyleRIPRel() &&
5137 (M == CodeModel::Small || M == CodeModel::Kernel))
5138 WrapperKind = X86ISD::WrapperRIP;
5139 else if (Subtarget->isPICStyleGOT())
5140 OpFlag = X86II::MO_GOTOFF;
5141 else if (Subtarget->isPICStyleStubPIC())
5142 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5144 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5146 CP->getOffset(), OpFlag);
5147 DebugLoc DL = CP->getDebugLoc();
5148 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5149 // With PIC, the address is actually $g + Offset.
5151 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5152 DAG.getNode(X86ISD::GlobalBaseReg,
5153 DebugLoc(), getPointerTy()),
5160 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5161 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5163 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5165 unsigned char OpFlag = 0;
5166 unsigned WrapperKind = X86ISD::Wrapper;
5167 CodeModel::Model M = getTargetMachine().getCodeModel();
5169 if (Subtarget->isPICStyleRIPRel() &&
5170 (M == CodeModel::Small || M == CodeModel::Kernel))
5171 WrapperKind = X86ISD::WrapperRIP;
5172 else if (Subtarget->isPICStyleGOT())
5173 OpFlag = X86II::MO_GOTOFF;
5174 else if (Subtarget->isPICStyleStubPIC())
5175 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5177 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5179 DebugLoc DL = JT->getDebugLoc();
5180 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5182 // With PIC, the address is actually $g + Offset.
5184 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5185 DAG.getNode(X86ISD::GlobalBaseReg,
5186 DebugLoc(), getPointerTy()),
5194 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5195 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5197 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5199 unsigned char OpFlag = 0;
5200 unsigned WrapperKind = X86ISD::Wrapper;
5201 CodeModel::Model M = getTargetMachine().getCodeModel();
5203 if (Subtarget->isPICStyleRIPRel() &&
5204 (M == CodeModel::Small || M == CodeModel::Kernel))
5205 WrapperKind = X86ISD::WrapperRIP;
5206 else if (Subtarget->isPICStyleGOT())
5207 OpFlag = X86II::MO_GOTOFF;
5208 else if (Subtarget->isPICStyleStubPIC())
5209 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5211 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5213 DebugLoc DL = Op.getDebugLoc();
5214 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5217 // With PIC, the address is actually $g + Offset.
5218 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5219 !Subtarget->is64Bit()) {
5220 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5221 DAG.getNode(X86ISD::GlobalBaseReg,
5222 DebugLoc(), getPointerTy()),
5230 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5231 // Create the TargetBlockAddressAddress node.
5232 unsigned char OpFlags =
5233 Subtarget->ClassifyBlockAddressReference();
5234 CodeModel::Model M = getTargetMachine().getCodeModel();
5235 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5236 DebugLoc dl = Op.getDebugLoc();
5237 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5238 /*isTarget=*/true, OpFlags);
5240 if (Subtarget->isPICStyleRIPRel() &&
5241 (M == CodeModel::Small || M == CodeModel::Kernel))
5242 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5244 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5246 // With PIC, the address is actually $g + Offset.
5247 if (isGlobalRelativeToPICBase(OpFlags)) {
5248 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5249 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5257 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5259 SelectionDAG &DAG) const {
5260 // Create the TargetGlobalAddress node, folding in the constant
5261 // offset if it is legal.
5262 unsigned char OpFlags =
5263 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5264 CodeModel::Model M = getTargetMachine().getCodeModel();
5266 if (OpFlags == X86II::MO_NO_FLAG &&
5267 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5268 // A direct static reference to a global.
5269 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5272 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5275 if (Subtarget->isPICStyleRIPRel() &&
5276 (M == CodeModel::Small || M == CodeModel::Kernel))
5277 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5279 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5281 // With PIC, the address is actually $g + Offset.
5282 if (isGlobalRelativeToPICBase(OpFlags)) {
5283 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5284 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5288 // For globals that require a load from a stub to get the address, emit the
5290 if (isGlobalStubReference(OpFlags))
5291 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5292 PseudoSourceValue::getGOT(), 0, false, false, 0);
5294 // If there was a non-zero offset that we didn't fold, create an explicit
5297 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5298 DAG.getConstant(Offset, getPointerTy()));
5304 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5305 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5306 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5307 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5311 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5312 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5313 unsigned char OperandFlags) {
5314 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5315 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5316 DebugLoc dl = GA->getDebugLoc();
5317 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5318 GA->getValueType(0),
5322 SDValue Ops[] = { Chain, TGA, *InFlag };
5323 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5325 SDValue Ops[] = { Chain, TGA };
5326 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5329 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5330 MFI->setAdjustsStack(true);
5332 SDValue Flag = Chain.getValue(1);
5333 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5336 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5338 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5341 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5342 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5343 DAG.getNode(X86ISD::GlobalBaseReg,
5344 DebugLoc(), PtrVT), InFlag);
5345 InFlag = Chain.getValue(1);
5347 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5350 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5352 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5354 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5355 X86::RAX, X86II::MO_TLSGD);
5358 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5359 // "local exec" model.
5360 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5361 const EVT PtrVT, TLSModel::Model model,
5363 DebugLoc dl = GA->getDebugLoc();
5364 // Get the Thread Pointer
5365 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5367 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5370 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5371 NULL, 0, false, false, 0);
5373 unsigned char OperandFlags = 0;
5374 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5376 unsigned WrapperKind = X86ISD::Wrapper;
5377 if (model == TLSModel::LocalExec) {
5378 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5379 } else if (is64Bit) {
5380 assert(model == TLSModel::InitialExec);
5381 OperandFlags = X86II::MO_GOTTPOFF;
5382 WrapperKind = X86ISD::WrapperRIP;
5384 assert(model == TLSModel::InitialExec);
5385 OperandFlags = X86II::MO_INDNTPOFF;
5388 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5390 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5391 GA->getValueType(0),
5392 GA->getOffset(), OperandFlags);
5393 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5395 if (model == TLSModel::InitialExec)
5396 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5397 PseudoSourceValue::getGOT(), 0, false, false, 0);
5399 // The address of the thread local variable is the add of the thread
5400 // pointer with the offset of the variable.
5401 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5405 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5407 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5408 const GlobalValue *GV = GA->getGlobal();
5410 if (Subtarget->isTargetELF()) {
5411 // TODO: implement the "local dynamic" model
5412 // TODO: implement the "initial exec"model for pic executables
5414 // If GV is an alias then use the aliasee for determining
5415 // thread-localness.
5416 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5417 GV = GA->resolveAliasedGlobal(false);
5419 TLSModel::Model model
5420 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5423 case TLSModel::GeneralDynamic:
5424 case TLSModel::LocalDynamic: // not implemented
5425 if (Subtarget->is64Bit())
5426 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5427 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5429 case TLSModel::InitialExec:
5430 case TLSModel::LocalExec:
5431 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5432 Subtarget->is64Bit());
5434 } else if (Subtarget->isTargetDarwin()) {
5435 // Darwin only has one model of TLS. Lower to that.
5436 unsigned char OpFlag = 0;
5437 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5438 X86ISD::WrapperRIP : X86ISD::Wrapper;
5440 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5442 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5443 !Subtarget->is64Bit();
5445 OpFlag = X86II::MO_TLVP_PIC_BASE;
5447 OpFlag = X86II::MO_TLVP;
5448 DebugLoc DL = Op.getDebugLoc();
5449 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
5451 GA->getOffset(), OpFlag);
5452 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5454 // With PIC32, the address is actually $g + Offset.
5456 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5457 DAG.getNode(X86ISD::GlobalBaseReg,
5458 DebugLoc(), getPointerTy()),
5461 // Lowering the machine isd will make sure everything is in the right
5463 SDValue Args[] = { Offset };
5464 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5466 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5467 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5468 MFI->setAdjustsStack(true);
5470 // And our return value (tls address) is in the standard call return value
5472 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5473 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5477 "TLS not implemented for this target.");
5479 llvm_unreachable("Unreachable");
5484 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5485 /// take a 2 x i32 value to shift plus a shift amount.
5486 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5487 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5488 EVT VT = Op.getValueType();
5489 unsigned VTBits = VT.getSizeInBits();
5490 DebugLoc dl = Op.getDebugLoc();
5491 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5492 SDValue ShOpLo = Op.getOperand(0);
5493 SDValue ShOpHi = Op.getOperand(1);
5494 SDValue ShAmt = Op.getOperand(2);
5495 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5496 DAG.getConstant(VTBits - 1, MVT::i8))
5497 : DAG.getConstant(0, VT);
5500 if (Op.getOpcode() == ISD::SHL_PARTS) {
5501 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5502 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5504 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5505 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5508 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5509 DAG.getConstant(VTBits, MVT::i8));
5510 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5511 AndNode, DAG.getConstant(0, MVT::i8));
5514 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5515 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5516 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5518 if (Op.getOpcode() == ISD::SHL_PARTS) {
5519 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5520 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5522 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5523 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5526 SDValue Ops[2] = { Lo, Hi };
5527 return DAG.getMergeValues(Ops, 2, dl);
5530 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5531 SelectionDAG &DAG) const {
5532 EVT SrcVT = Op.getOperand(0).getValueType();
5534 if (SrcVT.isVector()) {
5535 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5541 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5542 "Unknown SINT_TO_FP to lower!");
5544 // These are really Legal; return the operand so the caller accepts it as
5546 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5548 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5549 Subtarget->is64Bit()) {
5553 DebugLoc dl = Op.getDebugLoc();
5554 unsigned Size = SrcVT.getSizeInBits()/8;
5555 MachineFunction &MF = DAG.getMachineFunction();
5556 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5557 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5558 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5560 PseudoSourceValue::getFixedStack(SSFI), 0,
5562 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5565 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5567 SelectionDAG &DAG) const {
5569 DebugLoc dl = Op.getDebugLoc();
5571 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5573 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5575 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5576 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5577 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5578 Tys, Ops, array_lengthof(Ops));
5581 Chain = Result.getValue(1);
5582 SDValue InFlag = Result.getValue(2);
5584 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5585 // shouldn't be necessary except that RFP cannot be live across
5586 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5587 MachineFunction &MF = DAG.getMachineFunction();
5588 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5589 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5590 Tys = DAG.getVTList(MVT::Other);
5592 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5594 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5595 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5596 PseudoSourceValue::getFixedStack(SSFI), 0,
5603 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5604 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5605 SelectionDAG &DAG) const {
5606 // This algorithm is not obvious. Here it is in C code, more or less:
5608 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5609 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5610 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5612 // Copy ints to xmm registers.
5613 __m128i xh = _mm_cvtsi32_si128( hi );
5614 __m128i xl = _mm_cvtsi32_si128( lo );
5616 // Combine into low half of a single xmm register.
5617 __m128i x = _mm_unpacklo_epi32( xh, xl );
5621 // Merge in appropriate exponents to give the integer bits the right
5623 x = _mm_unpacklo_epi32( x, exp );
5625 // Subtract away the biases to deal with the IEEE-754 double precision
5627 d = _mm_sub_pd( (__m128d) x, bias );
5629 // All conversions up to here are exact. The correctly rounded result is
5630 // calculated using the current rounding mode using the following
5632 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5633 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5634 // store doesn't really need to be here (except
5635 // maybe to zero the other double)
5640 DebugLoc dl = Op.getDebugLoc();
5641 LLVMContext *Context = DAG.getContext();
5643 // Build some magic constants.
5644 std::vector<Constant*> CV0;
5645 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5646 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5647 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5648 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5649 Constant *C0 = ConstantVector::get(CV0);
5650 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5652 std::vector<Constant*> CV1;
5654 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5656 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5657 Constant *C1 = ConstantVector::get(CV1);
5658 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5660 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5661 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5663 DAG.getIntPtrConstant(1)));
5664 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5665 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5667 DAG.getIntPtrConstant(0)));
5668 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5669 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5670 PseudoSourceValue::getConstantPool(), 0,
5672 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5673 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5674 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5675 PseudoSourceValue::getConstantPool(), 0,
5677 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5679 // Add the halves; easiest way is to swap them into another reg first.
5680 int ShufMask[2] = { 1, -1 };
5681 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5682 DAG.getUNDEF(MVT::v2f64), ShufMask);
5683 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5684 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5685 DAG.getIntPtrConstant(0));
5688 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5689 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5690 SelectionDAG &DAG) const {
5691 DebugLoc dl = Op.getDebugLoc();
5692 // FP constant to bias correct the final result.
5693 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5696 // Load the 32-bit value into an XMM register.
5697 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5698 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5700 DAG.getIntPtrConstant(0)));
5702 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5703 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5704 DAG.getIntPtrConstant(0));
5706 // Or the load with the bias.
5707 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5708 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5709 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5711 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5712 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5713 MVT::v2f64, Bias)));
5714 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5715 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5716 DAG.getIntPtrConstant(0));
5718 // Subtract the bias.
5719 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5721 // Handle final rounding.
5722 EVT DestVT = Op.getValueType();
5724 if (DestVT.bitsLT(MVT::f64)) {
5725 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5726 DAG.getIntPtrConstant(0));
5727 } else if (DestVT.bitsGT(MVT::f64)) {
5728 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5731 // Handle final rounding.
5735 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5736 SelectionDAG &DAG) const {
5737 SDValue N0 = Op.getOperand(0);
5738 DebugLoc dl = Op.getDebugLoc();
5740 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5741 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5742 // the optimization here.
5743 if (DAG.SignBitIsZero(N0))
5744 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5746 EVT SrcVT = N0.getValueType();
5747 EVT DstVT = Op.getValueType();
5748 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5749 return LowerUINT_TO_FP_i64(Op, DAG);
5750 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5751 return LowerUINT_TO_FP_i32(Op, DAG);
5753 // Make a 64-bit buffer, and use it to build an FILD.
5754 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5755 if (SrcVT == MVT::i32) {
5756 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5757 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5758 getPointerTy(), StackSlot, WordOff);
5759 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5760 StackSlot, NULL, 0, false, false, 0);
5761 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5762 OffsetSlot, NULL, 0, false, false, 0);
5763 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5767 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5768 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5769 StackSlot, NULL, 0, false, false, 0);
5770 // For i64 source, we need to add the appropriate power of 2 if the input
5771 // was negative. This is the same as the optimization in
5772 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5773 // we must be careful to do the computation in x87 extended precision, not
5774 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5775 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5776 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5777 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5779 APInt FF(32, 0x5F800000ULL);
5781 // Check whether the sign bit is set.
5782 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5783 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5786 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5787 SDValue FudgePtr = DAG.getConstantPool(
5788 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5791 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5792 SDValue Zero = DAG.getIntPtrConstant(0);
5793 SDValue Four = DAG.getIntPtrConstant(4);
5794 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5796 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5798 // Load the value out, extending it from f32 to f80.
5799 // FIXME: Avoid the extend by constructing the right constant pool?
5800 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
5801 FudgePtr, PseudoSourceValue::getConstantPool(),
5802 0, MVT::f32, false, false, 4);
5803 // Extend everything to 80 bits to force it to be done on x87.
5804 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5805 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5808 std::pair<SDValue,SDValue> X86TargetLowering::
5809 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5810 DebugLoc dl = Op.getDebugLoc();
5812 EVT DstTy = Op.getValueType();
5815 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5819 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5820 DstTy.getSimpleVT() >= MVT::i16 &&
5821 "Unknown FP_TO_SINT to lower!");
5823 // These are really Legal.
5824 if (DstTy == MVT::i32 &&
5825 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5826 return std::make_pair(SDValue(), SDValue());
5827 if (Subtarget->is64Bit() &&
5828 DstTy == MVT::i64 &&
5829 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5830 return std::make_pair(SDValue(), SDValue());
5832 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5834 MachineFunction &MF = DAG.getMachineFunction();
5835 unsigned MemSize = DstTy.getSizeInBits()/8;
5836 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5837 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5840 switch (DstTy.getSimpleVT().SimpleTy) {
5841 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5842 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5843 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5844 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5847 SDValue Chain = DAG.getEntryNode();
5848 SDValue Value = Op.getOperand(0);
5849 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5850 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5851 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5852 PseudoSourceValue::getFixedStack(SSFI), 0,
5854 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5856 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5858 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5859 Chain = Value.getValue(1);
5860 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5861 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5864 // Build the FP_TO_INT*_IN_MEM
5865 SDValue Ops[] = { Chain, Value, StackSlot };
5866 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5868 return std::make_pair(FIST, StackSlot);
5871 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5872 SelectionDAG &DAG) const {
5873 if (Op.getValueType().isVector()) {
5874 if (Op.getValueType() == MVT::v2i32 &&
5875 Op.getOperand(0).getValueType() == MVT::v2f64) {
5881 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5882 SDValue FIST = Vals.first, StackSlot = Vals.second;
5883 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5884 if (FIST.getNode() == 0) return Op;
5887 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5888 FIST, StackSlot, NULL, 0, false, false, 0);
5891 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5892 SelectionDAG &DAG) const {
5893 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5894 SDValue FIST = Vals.first, StackSlot = Vals.second;
5895 assert(FIST.getNode() && "Unexpected failure");
5898 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5899 FIST, StackSlot, NULL, 0, false, false, 0);
5902 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5903 SelectionDAG &DAG) const {
5904 LLVMContext *Context = DAG.getContext();
5905 DebugLoc dl = Op.getDebugLoc();
5906 EVT VT = Op.getValueType();
5909 EltVT = VT.getVectorElementType();
5910 std::vector<Constant*> CV;
5911 if (EltVT == MVT::f64) {
5912 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5916 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5922 Constant *C = ConstantVector::get(CV);
5923 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5924 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5925 PseudoSourceValue::getConstantPool(), 0,
5927 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5930 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5931 LLVMContext *Context = DAG.getContext();
5932 DebugLoc dl = Op.getDebugLoc();
5933 EVT VT = Op.getValueType();
5936 EltVT = VT.getVectorElementType();
5937 std::vector<Constant*> CV;
5938 if (EltVT == MVT::f64) {
5939 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5943 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5949 Constant *C = ConstantVector::get(CV);
5950 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5951 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5952 PseudoSourceValue::getConstantPool(), 0,
5954 if (VT.isVector()) {
5955 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5956 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5957 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5959 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5961 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5965 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5966 LLVMContext *Context = DAG.getContext();
5967 SDValue Op0 = Op.getOperand(0);
5968 SDValue Op1 = Op.getOperand(1);
5969 DebugLoc dl = Op.getDebugLoc();
5970 EVT VT = Op.getValueType();
5971 EVT SrcVT = Op1.getValueType();
5973 // If second operand is smaller, extend it first.
5974 if (SrcVT.bitsLT(VT)) {
5975 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5978 // And if it is bigger, shrink it first.
5979 if (SrcVT.bitsGT(VT)) {
5980 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5984 // At this point the operands and the result should have the same
5985 // type, and that won't be f80 since that is not custom lowered.
5987 // First get the sign bit of second operand.
5988 std::vector<Constant*> CV;
5989 if (SrcVT == MVT::f64) {
5990 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5991 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5993 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5994 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5995 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5996 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5998 Constant *C = ConstantVector::get(CV);
5999 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6000 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6001 PseudoSourceValue::getConstantPool(), 0,
6003 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6005 // Shift sign bit right or left if the two operands have different types.
6006 if (SrcVT.bitsGT(VT)) {
6007 // Op0 is MVT::f32, Op1 is MVT::f64.
6008 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6009 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6010 DAG.getConstant(32, MVT::i32));
6011 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
6012 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6013 DAG.getIntPtrConstant(0));
6016 // Clear first operand sign bit.
6018 if (VT == MVT::f64) {
6019 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6020 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6022 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6023 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6024 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6025 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6027 C = ConstantVector::get(CV);
6028 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6029 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6030 PseudoSourceValue::getConstantPool(), 0,
6032 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6034 // Or the value with the sign bit.
6035 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6038 /// Emit nodes that will be selected as "test Op0,Op0", or something
6040 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6041 SelectionDAG &DAG) const {
6042 DebugLoc dl = Op.getDebugLoc();
6044 // CF and OF aren't always set the way we want. Determine which
6045 // of these we need.
6046 bool NeedCF = false;
6047 bool NeedOF = false;
6050 case X86::COND_A: case X86::COND_AE:
6051 case X86::COND_B: case X86::COND_BE:
6054 case X86::COND_G: case X86::COND_GE:
6055 case X86::COND_L: case X86::COND_LE:
6056 case X86::COND_O: case X86::COND_NO:
6061 // See if we can use the EFLAGS value from the operand instead of
6062 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6063 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6064 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6065 // Emit a CMP with 0, which is the TEST pattern.
6066 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6067 DAG.getConstant(0, Op.getValueType()));
6069 unsigned Opcode = 0;
6070 unsigned NumOperands = 0;
6071 switch (Op.getNode()->getOpcode()) {
6073 // Due to an isel shortcoming, be conservative if this add is likely to be
6074 // selected as part of a load-modify-store instruction. When the root node
6075 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6076 // uses of other nodes in the match, such as the ADD in this case. This
6077 // leads to the ADD being left around and reselected, with the result being
6078 // two adds in the output. Alas, even if none our users are stores, that
6079 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6080 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6081 // climbing the DAG back to the root, and it doesn't seem to be worth the
6083 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6084 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6085 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6088 if (ConstantSDNode *C =
6089 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6090 // An add of one will be selected as an INC.
6091 if (C->getAPIntValue() == 1) {
6092 Opcode = X86ISD::INC;
6097 // An add of negative one (subtract of one) will be selected as a DEC.
6098 if (C->getAPIntValue().isAllOnesValue()) {
6099 Opcode = X86ISD::DEC;
6105 // Otherwise use a regular EFLAGS-setting add.
6106 Opcode = X86ISD::ADD;
6110 // If the primary and result isn't used, don't bother using X86ISD::AND,
6111 // because a TEST instruction will be better.
6112 bool NonFlagUse = false;
6113 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6114 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6116 unsigned UOpNo = UI.getOperandNo();
6117 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6118 // Look pass truncate.
6119 UOpNo = User->use_begin().getOperandNo();
6120 User = *User->use_begin();
6123 if (User->getOpcode() != ISD::BRCOND &&
6124 User->getOpcode() != ISD::SETCC &&
6125 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6138 // Due to the ISEL shortcoming noted above, be conservative if this op is
6139 // likely to be selected as part of a load-modify-store instruction.
6140 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6141 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6142 if (UI->getOpcode() == ISD::STORE)
6145 // Otherwise use a regular EFLAGS-setting instruction.
6146 switch (Op.getNode()->getOpcode()) {
6147 default: llvm_unreachable("unexpected operator!");
6148 case ISD::SUB: Opcode = X86ISD::SUB; break;
6149 case ISD::OR: Opcode = X86ISD::OR; break;
6150 case ISD::XOR: Opcode = X86ISD::XOR; break;
6151 case ISD::AND: Opcode = X86ISD::AND; break;
6163 return SDValue(Op.getNode(), 1);
6170 // Emit a CMP with 0, which is the TEST pattern.
6171 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6172 DAG.getConstant(0, Op.getValueType()));
6174 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6175 SmallVector<SDValue, 4> Ops;
6176 for (unsigned i = 0; i != NumOperands; ++i)
6177 Ops.push_back(Op.getOperand(i));
6179 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6180 DAG.ReplaceAllUsesWith(Op, New);
6181 return SDValue(New.getNode(), 1);
6184 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6186 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6187 SelectionDAG &DAG) const {
6188 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6189 if (C->getAPIntValue() == 0)
6190 return EmitTest(Op0, X86CC, DAG);
6192 DebugLoc dl = Op0.getDebugLoc();
6193 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6196 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6197 /// if it's possible.
6198 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6199 DebugLoc dl, SelectionDAG &DAG) const {
6200 SDValue Op0 = And.getOperand(0);
6201 SDValue Op1 = And.getOperand(1);
6202 if (Op0.getOpcode() == ISD::TRUNCATE)
6203 Op0 = Op0.getOperand(0);
6204 if (Op1.getOpcode() == ISD::TRUNCATE)
6205 Op1 = Op1.getOperand(0);
6208 if (Op1.getOpcode() == ISD::SHL)
6209 std::swap(Op0, Op1);
6210 if (Op0.getOpcode() == ISD::SHL) {
6211 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6212 if (And00C->getZExtValue() == 1) {
6213 // If we looked past a truncate, check that it's only truncating away
6215 unsigned BitWidth = Op0.getValueSizeInBits();
6216 unsigned AndBitWidth = And.getValueSizeInBits();
6217 if (BitWidth > AndBitWidth) {
6218 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6219 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6220 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6224 RHS = Op0.getOperand(1);
6226 } else if (Op1.getOpcode() == ISD::Constant) {
6227 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6228 SDValue AndLHS = Op0;
6229 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6230 LHS = AndLHS.getOperand(0);
6231 RHS = AndLHS.getOperand(1);
6235 if (LHS.getNode()) {
6236 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6237 // instruction. Since the shift amount is in-range-or-undefined, we know
6238 // that doing a bittest on the i32 value is ok. We extend to i32 because
6239 // the encoding for the i16 version is larger than the i32 version.
6240 // Also promote i16 to i32 for performance / code size reason.
6241 if (LHS.getValueType() == MVT::i8 ||
6242 LHS.getValueType() == MVT::i16)
6243 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6245 // If the operand types disagree, extend the shift amount to match. Since
6246 // BT ignores high bits (like shifts) we can use anyextend.
6247 if (LHS.getValueType() != RHS.getValueType())
6248 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6250 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6251 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6252 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6253 DAG.getConstant(Cond, MVT::i8), BT);
6259 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6260 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6261 SDValue Op0 = Op.getOperand(0);
6262 SDValue Op1 = Op.getOperand(1);
6263 DebugLoc dl = Op.getDebugLoc();
6264 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6266 // Optimize to BT if possible.
6267 // Lower (X & (1 << N)) == 0 to BT(X, N).
6268 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6269 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6270 if (Op0.getOpcode() == ISD::AND &&
6272 Op1.getOpcode() == ISD::Constant &&
6273 cast<ConstantSDNode>(Op1)->isNullValue() &&
6274 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6275 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6276 if (NewSetCC.getNode())
6280 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6281 if (Op0.getOpcode() == X86ISD::SETCC &&
6282 Op1.getOpcode() == ISD::Constant &&
6283 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6284 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6285 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6286 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6287 bool Invert = (CC == ISD::SETNE) ^
6288 cast<ConstantSDNode>(Op1)->isNullValue();
6290 CCode = X86::GetOppositeBranchCondition(CCode);
6291 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6292 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6295 bool isFP = Op1.getValueType().isFloatingPoint();
6296 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6297 if (X86CC == X86::COND_INVALID)
6300 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6302 // Use sbb x, x to materialize carry bit into a GPR.
6303 if (X86CC == X86::COND_B)
6304 return DAG.getNode(ISD::AND, dl, MVT::i8,
6305 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6306 DAG.getConstant(X86CC, MVT::i8), Cond),
6307 DAG.getConstant(1, MVT::i8));
6309 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6310 DAG.getConstant(X86CC, MVT::i8), Cond);
6313 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6315 SDValue Op0 = Op.getOperand(0);
6316 SDValue Op1 = Op.getOperand(1);
6317 SDValue CC = Op.getOperand(2);
6318 EVT VT = Op.getValueType();
6319 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6320 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6321 DebugLoc dl = Op.getDebugLoc();
6325 EVT VT0 = Op0.getValueType();
6326 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6327 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6330 switch (SetCCOpcode) {
6333 case ISD::SETEQ: SSECC = 0; break;
6335 case ISD::SETGT: Swap = true; // Fallthrough
6337 case ISD::SETOLT: SSECC = 1; break;
6339 case ISD::SETGE: Swap = true; // Fallthrough
6341 case ISD::SETOLE: SSECC = 2; break;
6342 case ISD::SETUO: SSECC = 3; break;
6344 case ISD::SETNE: SSECC = 4; break;
6345 case ISD::SETULE: Swap = true;
6346 case ISD::SETUGE: SSECC = 5; break;
6347 case ISD::SETULT: Swap = true;
6348 case ISD::SETUGT: SSECC = 6; break;
6349 case ISD::SETO: SSECC = 7; break;
6352 std::swap(Op0, Op1);
6354 // In the two special cases we can't handle, emit two comparisons.
6356 if (SetCCOpcode == ISD::SETUEQ) {
6358 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6359 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6360 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6362 else if (SetCCOpcode == ISD::SETONE) {
6364 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6365 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6366 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6368 llvm_unreachable("Illegal FP comparison");
6370 // Handle all other FP comparisons here.
6371 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6374 // We are handling one of the integer comparisons here. Since SSE only has
6375 // GT and EQ comparisons for integer, swapping operands and multiple
6376 // operations may be required for some comparisons.
6377 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6378 bool Swap = false, Invert = false, FlipSigns = false;
6380 switch (VT.getSimpleVT().SimpleTy) {
6383 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6385 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6387 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6388 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6391 switch (SetCCOpcode) {
6393 case ISD::SETNE: Invert = true;
6394 case ISD::SETEQ: Opc = EQOpc; break;
6395 case ISD::SETLT: Swap = true;
6396 case ISD::SETGT: Opc = GTOpc; break;
6397 case ISD::SETGE: Swap = true;
6398 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6399 case ISD::SETULT: Swap = true;
6400 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6401 case ISD::SETUGE: Swap = true;
6402 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6405 std::swap(Op0, Op1);
6407 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6408 // bits of the inputs before performing those operations.
6410 EVT EltVT = VT.getVectorElementType();
6411 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6413 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6414 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6416 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6417 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6420 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6422 // If the logical-not of the result is required, perform that now.
6424 Result = DAG.getNOT(dl, Result, VT);
6429 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6430 static bool isX86LogicalCmp(SDValue Op) {
6431 unsigned Opc = Op.getNode()->getOpcode();
6432 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6434 if (Op.getResNo() == 1 &&
6435 (Opc == X86ISD::ADD ||
6436 Opc == X86ISD::SUB ||
6437 Opc == X86ISD::SMUL ||
6438 Opc == X86ISD::UMUL ||
6439 Opc == X86ISD::INC ||
6440 Opc == X86ISD::DEC ||
6441 Opc == X86ISD::OR ||
6442 Opc == X86ISD::XOR ||
6443 Opc == X86ISD::AND))
6449 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6450 bool addTest = true;
6451 SDValue Cond = Op.getOperand(0);
6452 DebugLoc dl = Op.getDebugLoc();
6455 if (Cond.getOpcode() == ISD::SETCC) {
6456 SDValue NewCond = LowerSETCC(Cond, DAG);
6457 if (NewCond.getNode())
6461 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6462 SDValue Op1 = Op.getOperand(1);
6463 SDValue Op2 = Op.getOperand(2);
6464 if (Cond.getOpcode() == X86ISD::SETCC &&
6465 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6466 SDValue Cmp = Cond.getOperand(1);
6467 if (Cmp.getOpcode() == X86ISD::CMP) {
6468 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6469 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6470 ConstantSDNode *RHSC =
6471 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6472 if (N1C && N1C->isAllOnesValue() &&
6473 N2C && N2C->isNullValue() &&
6474 RHSC && RHSC->isNullValue()) {
6475 SDValue CmpOp0 = Cmp.getOperand(0);
6476 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6477 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6478 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6479 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6484 // Look pass (and (setcc_carry (cmp ...)), 1).
6485 if (Cond.getOpcode() == ISD::AND &&
6486 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6487 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6488 if (C && C->getAPIntValue() == 1)
6489 Cond = Cond.getOperand(0);
6492 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6493 // setting operand in place of the X86ISD::SETCC.
6494 if (Cond.getOpcode() == X86ISD::SETCC ||
6495 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6496 CC = Cond.getOperand(0);
6498 SDValue Cmp = Cond.getOperand(1);
6499 unsigned Opc = Cmp.getOpcode();
6500 EVT VT = Op.getValueType();
6502 bool IllegalFPCMov = false;
6503 if (VT.isFloatingPoint() && !VT.isVector() &&
6504 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6505 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6507 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6508 Opc == X86ISD::BT) { // FIXME
6515 // Look pass the truncate.
6516 if (Cond.getOpcode() == ISD::TRUNCATE)
6517 Cond = Cond.getOperand(0);
6519 // We know the result of AND is compared against zero. Try to match
6521 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6522 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6523 if (NewSetCC.getNode()) {
6524 CC = NewSetCC.getOperand(0);
6525 Cond = NewSetCC.getOperand(1);
6532 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6533 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6536 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6537 // condition is true.
6538 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6539 SDValue Ops[] = { Op2, Op1, CC, Cond };
6540 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6543 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6544 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6545 // from the AND / OR.
6546 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6547 Opc = Op.getOpcode();
6548 if (Opc != ISD::OR && Opc != ISD::AND)
6550 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6551 Op.getOperand(0).hasOneUse() &&
6552 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6553 Op.getOperand(1).hasOneUse());
6556 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6557 // 1 and that the SETCC node has a single use.
6558 static bool isXor1OfSetCC(SDValue Op) {
6559 if (Op.getOpcode() != ISD::XOR)
6561 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6562 if (N1C && N1C->getAPIntValue() == 1) {
6563 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6564 Op.getOperand(0).hasOneUse();
6569 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6570 bool addTest = true;
6571 SDValue Chain = Op.getOperand(0);
6572 SDValue Cond = Op.getOperand(1);
6573 SDValue Dest = Op.getOperand(2);
6574 DebugLoc dl = Op.getDebugLoc();
6577 if (Cond.getOpcode() == ISD::SETCC) {
6578 SDValue NewCond = LowerSETCC(Cond, DAG);
6579 if (NewCond.getNode())
6583 // FIXME: LowerXALUO doesn't handle these!!
6584 else if (Cond.getOpcode() == X86ISD::ADD ||
6585 Cond.getOpcode() == X86ISD::SUB ||
6586 Cond.getOpcode() == X86ISD::SMUL ||
6587 Cond.getOpcode() == X86ISD::UMUL)
6588 Cond = LowerXALUO(Cond, DAG);
6591 // Look pass (and (setcc_carry (cmp ...)), 1).
6592 if (Cond.getOpcode() == ISD::AND &&
6593 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6594 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6595 if (C && C->getAPIntValue() == 1)
6596 Cond = Cond.getOperand(0);
6599 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6600 // setting operand in place of the X86ISD::SETCC.
6601 if (Cond.getOpcode() == X86ISD::SETCC ||
6602 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6603 CC = Cond.getOperand(0);
6605 SDValue Cmp = Cond.getOperand(1);
6606 unsigned Opc = Cmp.getOpcode();
6607 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6608 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6612 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6616 // These can only come from an arithmetic instruction with overflow,
6617 // e.g. SADDO, UADDO.
6618 Cond = Cond.getNode()->getOperand(1);
6625 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6626 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6627 if (CondOpc == ISD::OR) {
6628 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6629 // two branches instead of an explicit OR instruction with a
6631 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6632 isX86LogicalCmp(Cmp)) {
6633 CC = Cond.getOperand(0).getOperand(0);
6634 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6635 Chain, Dest, CC, Cmp);
6636 CC = Cond.getOperand(1).getOperand(0);
6640 } else { // ISD::AND
6641 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6642 // two branches instead of an explicit AND instruction with a
6643 // separate test. However, we only do this if this block doesn't
6644 // have a fall-through edge, because this requires an explicit
6645 // jmp when the condition is false.
6646 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6647 isX86LogicalCmp(Cmp) &&
6648 Op.getNode()->hasOneUse()) {
6649 X86::CondCode CCode =
6650 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6651 CCode = X86::GetOppositeBranchCondition(CCode);
6652 CC = DAG.getConstant(CCode, MVT::i8);
6653 SDNode *User = *Op.getNode()->use_begin();
6654 // Look for an unconditional branch following this conditional branch.
6655 // We need this because we need to reverse the successors in order
6656 // to implement FCMP_OEQ.
6657 if (User->getOpcode() == ISD::BR) {
6658 SDValue FalseBB = User->getOperand(1);
6660 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6661 assert(NewBR == User);
6665 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6666 Chain, Dest, CC, Cmp);
6667 X86::CondCode CCode =
6668 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6669 CCode = X86::GetOppositeBranchCondition(CCode);
6670 CC = DAG.getConstant(CCode, MVT::i8);
6676 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6677 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6678 // It should be transformed during dag combiner except when the condition
6679 // is set by a arithmetics with overflow node.
6680 X86::CondCode CCode =
6681 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6682 CCode = X86::GetOppositeBranchCondition(CCode);
6683 CC = DAG.getConstant(CCode, MVT::i8);
6684 Cond = Cond.getOperand(0).getOperand(1);
6690 // Look pass the truncate.
6691 if (Cond.getOpcode() == ISD::TRUNCATE)
6692 Cond = Cond.getOperand(0);
6694 // We know the result of AND is compared against zero. Try to match
6696 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6697 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6698 if (NewSetCC.getNode()) {
6699 CC = NewSetCC.getOperand(0);
6700 Cond = NewSetCC.getOperand(1);
6707 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6708 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6710 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6711 Chain, Dest, CC, Cond);
6715 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6716 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6717 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6718 // that the guard pages used by the OS virtual memory manager are allocated in
6719 // correct sequence.
6721 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6722 SelectionDAG &DAG) const {
6723 assert(Subtarget->isTargetCygMing() &&
6724 "This should be used only on Cygwin/Mingw targets");
6725 DebugLoc dl = Op.getDebugLoc();
6728 SDValue Chain = Op.getOperand(0);
6729 SDValue Size = Op.getOperand(1);
6730 // FIXME: Ensure alignment here
6734 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6736 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6737 Flag = Chain.getValue(1);
6739 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6741 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6742 Flag = Chain.getValue(1);
6744 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6746 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6747 return DAG.getMergeValues(Ops1, 2, dl);
6750 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6751 MachineFunction &MF = DAG.getMachineFunction();
6752 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6754 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6755 DebugLoc dl = Op.getDebugLoc();
6757 if (!Subtarget->is64Bit()) {
6758 // vastart just stores the address of the VarArgsFrameIndex slot into the
6759 // memory location argument.
6760 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6762 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6767 // gp_offset (0 - 6 * 8)
6768 // fp_offset (48 - 48 + 8 * 16)
6769 // overflow_arg_area (point to parameters coming in memory).
6771 SmallVector<SDValue, 8> MemOps;
6772 SDValue FIN = Op.getOperand(1);
6774 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6775 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6777 FIN, SV, 0, false, false, 0);
6778 MemOps.push_back(Store);
6781 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6782 FIN, DAG.getIntPtrConstant(4));
6783 Store = DAG.getStore(Op.getOperand(0), dl,
6784 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6786 FIN, SV, 4, false, false, 0);
6787 MemOps.push_back(Store);
6789 // Store ptr to overflow_arg_area
6790 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6791 FIN, DAG.getIntPtrConstant(4));
6792 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6794 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 8,
6796 MemOps.push_back(Store);
6798 // Store ptr to reg_save_area.
6799 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6800 FIN, DAG.getIntPtrConstant(8));
6801 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6803 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 16,
6805 MemOps.push_back(Store);
6806 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6807 &MemOps[0], MemOps.size());
6810 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6811 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6812 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6814 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6818 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6819 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6820 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6821 SDValue Chain = Op.getOperand(0);
6822 SDValue DstPtr = Op.getOperand(1);
6823 SDValue SrcPtr = Op.getOperand(2);
6824 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6825 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6826 DebugLoc dl = Op.getDebugLoc();
6828 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6829 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6830 false, DstSV, 0, SrcSV, 0);
6834 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6835 DebugLoc dl = Op.getDebugLoc();
6836 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6838 default: return SDValue(); // Don't custom lower most intrinsics.
6839 // Comparison intrinsics.
6840 case Intrinsic::x86_sse_comieq_ss:
6841 case Intrinsic::x86_sse_comilt_ss:
6842 case Intrinsic::x86_sse_comile_ss:
6843 case Intrinsic::x86_sse_comigt_ss:
6844 case Intrinsic::x86_sse_comige_ss:
6845 case Intrinsic::x86_sse_comineq_ss:
6846 case Intrinsic::x86_sse_ucomieq_ss:
6847 case Intrinsic::x86_sse_ucomilt_ss:
6848 case Intrinsic::x86_sse_ucomile_ss:
6849 case Intrinsic::x86_sse_ucomigt_ss:
6850 case Intrinsic::x86_sse_ucomige_ss:
6851 case Intrinsic::x86_sse_ucomineq_ss:
6852 case Intrinsic::x86_sse2_comieq_sd:
6853 case Intrinsic::x86_sse2_comilt_sd:
6854 case Intrinsic::x86_sse2_comile_sd:
6855 case Intrinsic::x86_sse2_comigt_sd:
6856 case Intrinsic::x86_sse2_comige_sd:
6857 case Intrinsic::x86_sse2_comineq_sd:
6858 case Intrinsic::x86_sse2_ucomieq_sd:
6859 case Intrinsic::x86_sse2_ucomilt_sd:
6860 case Intrinsic::x86_sse2_ucomile_sd:
6861 case Intrinsic::x86_sse2_ucomigt_sd:
6862 case Intrinsic::x86_sse2_ucomige_sd:
6863 case Intrinsic::x86_sse2_ucomineq_sd: {
6865 ISD::CondCode CC = ISD::SETCC_INVALID;
6868 case Intrinsic::x86_sse_comieq_ss:
6869 case Intrinsic::x86_sse2_comieq_sd:
6873 case Intrinsic::x86_sse_comilt_ss:
6874 case Intrinsic::x86_sse2_comilt_sd:
6878 case Intrinsic::x86_sse_comile_ss:
6879 case Intrinsic::x86_sse2_comile_sd:
6883 case Intrinsic::x86_sse_comigt_ss:
6884 case Intrinsic::x86_sse2_comigt_sd:
6888 case Intrinsic::x86_sse_comige_ss:
6889 case Intrinsic::x86_sse2_comige_sd:
6893 case Intrinsic::x86_sse_comineq_ss:
6894 case Intrinsic::x86_sse2_comineq_sd:
6898 case Intrinsic::x86_sse_ucomieq_ss:
6899 case Intrinsic::x86_sse2_ucomieq_sd:
6900 Opc = X86ISD::UCOMI;
6903 case Intrinsic::x86_sse_ucomilt_ss:
6904 case Intrinsic::x86_sse2_ucomilt_sd:
6905 Opc = X86ISD::UCOMI;
6908 case Intrinsic::x86_sse_ucomile_ss:
6909 case Intrinsic::x86_sse2_ucomile_sd:
6910 Opc = X86ISD::UCOMI;
6913 case Intrinsic::x86_sse_ucomigt_ss:
6914 case Intrinsic::x86_sse2_ucomigt_sd:
6915 Opc = X86ISD::UCOMI;
6918 case Intrinsic::x86_sse_ucomige_ss:
6919 case Intrinsic::x86_sse2_ucomige_sd:
6920 Opc = X86ISD::UCOMI;
6923 case Intrinsic::x86_sse_ucomineq_ss:
6924 case Intrinsic::x86_sse2_ucomineq_sd:
6925 Opc = X86ISD::UCOMI;
6930 SDValue LHS = Op.getOperand(1);
6931 SDValue RHS = Op.getOperand(2);
6932 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6933 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6934 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6935 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6936 DAG.getConstant(X86CC, MVT::i8), Cond);
6937 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6939 // ptest intrinsics. The intrinsic these come from are designed to return
6940 // an integer value, not just an instruction so lower it to the ptest
6941 // pattern and a setcc for the result.
6942 case Intrinsic::x86_sse41_ptestz:
6943 case Intrinsic::x86_sse41_ptestc:
6944 case Intrinsic::x86_sse41_ptestnzc:{
6947 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6948 case Intrinsic::x86_sse41_ptestz:
6950 X86CC = X86::COND_E;
6952 case Intrinsic::x86_sse41_ptestc:
6954 X86CC = X86::COND_B;
6956 case Intrinsic::x86_sse41_ptestnzc:
6958 X86CC = X86::COND_A;
6962 SDValue LHS = Op.getOperand(1);
6963 SDValue RHS = Op.getOperand(2);
6964 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6965 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6966 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6967 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6970 // Fix vector shift instructions where the last operand is a non-immediate
6972 case Intrinsic::x86_sse2_pslli_w:
6973 case Intrinsic::x86_sse2_pslli_d:
6974 case Intrinsic::x86_sse2_pslli_q:
6975 case Intrinsic::x86_sse2_psrli_w:
6976 case Intrinsic::x86_sse2_psrli_d:
6977 case Intrinsic::x86_sse2_psrli_q:
6978 case Intrinsic::x86_sse2_psrai_w:
6979 case Intrinsic::x86_sse2_psrai_d:
6980 case Intrinsic::x86_mmx_pslli_w:
6981 case Intrinsic::x86_mmx_pslli_d:
6982 case Intrinsic::x86_mmx_pslli_q:
6983 case Intrinsic::x86_mmx_psrli_w:
6984 case Intrinsic::x86_mmx_psrli_d:
6985 case Intrinsic::x86_mmx_psrli_q:
6986 case Intrinsic::x86_mmx_psrai_w:
6987 case Intrinsic::x86_mmx_psrai_d: {
6988 SDValue ShAmt = Op.getOperand(2);
6989 if (isa<ConstantSDNode>(ShAmt))
6992 unsigned NewIntNo = 0;
6993 EVT ShAmtVT = MVT::v4i32;
6995 case Intrinsic::x86_sse2_pslli_w:
6996 NewIntNo = Intrinsic::x86_sse2_psll_w;
6998 case Intrinsic::x86_sse2_pslli_d:
6999 NewIntNo = Intrinsic::x86_sse2_psll_d;
7001 case Intrinsic::x86_sse2_pslli_q:
7002 NewIntNo = Intrinsic::x86_sse2_psll_q;
7004 case Intrinsic::x86_sse2_psrli_w:
7005 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7007 case Intrinsic::x86_sse2_psrli_d:
7008 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7010 case Intrinsic::x86_sse2_psrli_q:
7011 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7013 case Intrinsic::x86_sse2_psrai_w:
7014 NewIntNo = Intrinsic::x86_sse2_psra_w;
7016 case Intrinsic::x86_sse2_psrai_d:
7017 NewIntNo = Intrinsic::x86_sse2_psra_d;
7020 ShAmtVT = MVT::v2i32;
7022 case Intrinsic::x86_mmx_pslli_w:
7023 NewIntNo = Intrinsic::x86_mmx_psll_w;
7025 case Intrinsic::x86_mmx_pslli_d:
7026 NewIntNo = Intrinsic::x86_mmx_psll_d;
7028 case Intrinsic::x86_mmx_pslli_q:
7029 NewIntNo = Intrinsic::x86_mmx_psll_q;
7031 case Intrinsic::x86_mmx_psrli_w:
7032 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7034 case Intrinsic::x86_mmx_psrli_d:
7035 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7037 case Intrinsic::x86_mmx_psrli_q:
7038 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7040 case Intrinsic::x86_mmx_psrai_w:
7041 NewIntNo = Intrinsic::x86_mmx_psra_w;
7043 case Intrinsic::x86_mmx_psrai_d:
7044 NewIntNo = Intrinsic::x86_mmx_psra_d;
7046 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7052 // The vector shift intrinsics with scalars uses 32b shift amounts but
7053 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7057 ShOps[1] = DAG.getConstant(0, MVT::i32);
7058 if (ShAmtVT == MVT::v4i32) {
7059 ShOps[2] = DAG.getUNDEF(MVT::i32);
7060 ShOps[3] = DAG.getUNDEF(MVT::i32);
7061 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7063 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7066 EVT VT = Op.getValueType();
7067 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7068 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7069 DAG.getConstant(NewIntNo, MVT::i32),
7070 Op.getOperand(1), ShAmt);
7075 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7076 SelectionDAG &DAG) const {
7077 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7078 MFI->setReturnAddressIsTaken(true);
7080 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7081 DebugLoc dl = Op.getDebugLoc();
7084 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7086 DAG.getConstant(TD->getPointerSize(),
7087 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7088 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7089 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7091 NULL, 0, false, false, 0);
7094 // Just load the return address.
7095 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7096 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7097 RetAddrFI, NULL, 0, false, false, 0);
7100 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7101 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7102 MFI->setFrameAddressIsTaken(true);
7104 EVT VT = Op.getValueType();
7105 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7106 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7107 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7108 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7110 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7115 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7116 SelectionDAG &DAG) const {
7117 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7120 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7121 MachineFunction &MF = DAG.getMachineFunction();
7122 SDValue Chain = Op.getOperand(0);
7123 SDValue Offset = Op.getOperand(1);
7124 SDValue Handler = Op.getOperand(2);
7125 DebugLoc dl = Op.getDebugLoc();
7127 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7129 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7131 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7132 DAG.getIntPtrConstant(-TD->getPointerSize()));
7133 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7134 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7135 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7136 MF.getRegInfo().addLiveOut(StoreAddrReg);
7138 return DAG.getNode(X86ISD::EH_RETURN, dl,
7140 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7143 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7144 SelectionDAG &DAG) const {
7145 SDValue Root = Op.getOperand(0);
7146 SDValue Trmp = Op.getOperand(1); // trampoline
7147 SDValue FPtr = Op.getOperand(2); // nested function
7148 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7149 DebugLoc dl = Op.getDebugLoc();
7151 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7153 if (Subtarget->is64Bit()) {
7154 SDValue OutChains[6];
7156 // Large code-model.
7157 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7158 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7160 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7161 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7163 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7165 // Load the pointer to the nested function into R11.
7166 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7167 SDValue Addr = Trmp;
7168 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7169 Addr, TrmpAddr, 0, false, false, 0);
7171 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7172 DAG.getConstant(2, MVT::i64));
7173 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7176 // Load the 'nest' parameter value into R10.
7177 // R10 is specified in X86CallingConv.td
7178 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7179 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7180 DAG.getConstant(10, MVT::i64));
7181 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7182 Addr, TrmpAddr, 10, false, false, 0);
7184 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7185 DAG.getConstant(12, MVT::i64));
7186 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7189 // Jump to the nested function.
7190 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7191 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7192 DAG.getConstant(20, MVT::i64));
7193 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7194 Addr, TrmpAddr, 20, false, false, 0);
7196 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7197 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7198 DAG.getConstant(22, MVT::i64));
7199 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7200 TrmpAddr, 22, false, false, 0);
7203 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7204 return DAG.getMergeValues(Ops, 2, dl);
7206 const Function *Func =
7207 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7208 CallingConv::ID CC = Func->getCallingConv();
7213 llvm_unreachable("Unsupported calling convention");
7214 case CallingConv::C:
7215 case CallingConv::X86_StdCall: {
7216 // Pass 'nest' parameter in ECX.
7217 // Must be kept in sync with X86CallingConv.td
7220 // Check that ECX wasn't needed by an 'inreg' parameter.
7221 const FunctionType *FTy = Func->getFunctionType();
7222 const AttrListPtr &Attrs = Func->getAttributes();
7224 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7225 unsigned InRegCount = 0;
7228 for (FunctionType::param_iterator I = FTy->param_begin(),
7229 E = FTy->param_end(); I != E; ++I, ++Idx)
7230 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7231 // FIXME: should only count parameters that are lowered to integers.
7232 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7234 if (InRegCount > 2) {
7235 report_fatal_error("Nest register in use - reduce number of inreg"
7241 case CallingConv::X86_FastCall:
7242 case CallingConv::X86_ThisCall:
7243 case CallingConv::Fast:
7244 // Pass 'nest' parameter in EAX.
7245 // Must be kept in sync with X86CallingConv.td
7250 SDValue OutChains[4];
7253 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7254 DAG.getConstant(10, MVT::i32));
7255 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7257 // This is storing the opcode for MOV32ri.
7258 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7259 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7260 OutChains[0] = DAG.getStore(Root, dl,
7261 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7262 Trmp, TrmpAddr, 0, false, false, 0);
7264 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7265 DAG.getConstant(1, MVT::i32));
7266 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7269 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7270 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7271 DAG.getConstant(5, MVT::i32));
7272 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7273 TrmpAddr, 5, false, false, 1);
7275 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7276 DAG.getConstant(6, MVT::i32));
7277 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7281 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7282 return DAG.getMergeValues(Ops, 2, dl);
7286 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7287 SelectionDAG &DAG) const {
7289 The rounding mode is in bits 11:10 of FPSR, and has the following
7296 FLT_ROUNDS, on the other hand, expects the following:
7303 To perform the conversion, we do:
7304 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7307 MachineFunction &MF = DAG.getMachineFunction();
7308 const TargetMachine &TM = MF.getTarget();
7309 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7310 unsigned StackAlignment = TFI.getStackAlignment();
7311 EVT VT = Op.getValueType();
7312 DebugLoc dl = Op.getDebugLoc();
7314 // Save FP Control Word to stack slot
7315 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7316 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7318 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7319 DAG.getEntryNode(), StackSlot);
7321 // Load FP Control Word from stack slot
7322 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7325 // Transform as necessary
7327 DAG.getNode(ISD::SRL, dl, MVT::i16,
7328 DAG.getNode(ISD::AND, dl, MVT::i16,
7329 CWD, DAG.getConstant(0x800, MVT::i16)),
7330 DAG.getConstant(11, MVT::i8));
7332 DAG.getNode(ISD::SRL, dl, MVT::i16,
7333 DAG.getNode(ISD::AND, dl, MVT::i16,
7334 CWD, DAG.getConstant(0x400, MVT::i16)),
7335 DAG.getConstant(9, MVT::i8));
7338 DAG.getNode(ISD::AND, dl, MVT::i16,
7339 DAG.getNode(ISD::ADD, dl, MVT::i16,
7340 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7341 DAG.getConstant(1, MVT::i16)),
7342 DAG.getConstant(3, MVT::i16));
7345 return DAG.getNode((VT.getSizeInBits() < 16 ?
7346 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7349 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7350 EVT VT = Op.getValueType();
7352 unsigned NumBits = VT.getSizeInBits();
7353 DebugLoc dl = Op.getDebugLoc();
7355 Op = Op.getOperand(0);
7356 if (VT == MVT::i8) {
7357 // Zero extend to i32 since there is not an i8 bsr.
7359 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7362 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7363 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7364 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7366 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7369 DAG.getConstant(NumBits+NumBits-1, OpVT),
7370 DAG.getConstant(X86::COND_E, MVT::i8),
7373 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7375 // Finally xor with NumBits-1.
7376 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7379 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7383 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7384 EVT VT = Op.getValueType();
7386 unsigned NumBits = VT.getSizeInBits();
7387 DebugLoc dl = Op.getDebugLoc();
7389 Op = Op.getOperand(0);
7390 if (VT == MVT::i8) {
7392 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7395 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7396 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7397 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7399 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7402 DAG.getConstant(NumBits, OpVT),
7403 DAG.getConstant(X86::COND_E, MVT::i8),
7406 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7409 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7413 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7414 EVT VT = Op.getValueType();
7415 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7416 DebugLoc dl = Op.getDebugLoc();
7418 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7419 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7420 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7421 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7422 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7424 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7425 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7426 // return AloBlo + AloBhi + AhiBlo;
7428 SDValue A = Op.getOperand(0);
7429 SDValue B = Op.getOperand(1);
7431 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7432 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7433 A, DAG.getConstant(32, MVT::i32));
7434 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7435 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7436 B, DAG.getConstant(32, MVT::i32));
7437 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7438 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7440 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7441 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7443 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7444 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7446 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7447 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7448 AloBhi, DAG.getConstant(32, MVT::i32));
7449 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7450 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7451 AhiBlo, DAG.getConstant(32, MVT::i32));
7452 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7453 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7458 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7459 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7460 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7461 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7462 // has only one use.
7463 SDNode *N = Op.getNode();
7464 SDValue LHS = N->getOperand(0);
7465 SDValue RHS = N->getOperand(1);
7466 unsigned BaseOp = 0;
7468 DebugLoc dl = Op.getDebugLoc();
7470 switch (Op.getOpcode()) {
7471 default: llvm_unreachable("Unknown ovf instruction!");
7473 // A subtract of one will be selected as a INC. Note that INC doesn't
7474 // set CF, so we can't do this for UADDO.
7475 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7476 if (C->getAPIntValue() == 1) {
7477 BaseOp = X86ISD::INC;
7481 BaseOp = X86ISD::ADD;
7485 BaseOp = X86ISD::ADD;
7489 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7490 // set CF, so we can't do this for USUBO.
7491 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7492 if (C->getAPIntValue() == 1) {
7493 BaseOp = X86ISD::DEC;
7497 BaseOp = X86ISD::SUB;
7501 BaseOp = X86ISD::SUB;
7505 BaseOp = X86ISD::SMUL;
7509 BaseOp = X86ISD::UMUL;
7514 // Also sets EFLAGS.
7515 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7516 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7519 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7520 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7522 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7526 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
7527 DebugLoc dl = Op.getDebugLoc();
7529 if (!Subtarget->hasSSE2())
7530 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
7531 DAG.getConstant(0, MVT::i32));
7533 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
7535 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
7537 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7538 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
7539 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
7540 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
7542 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
7543 if (!Op1 && !Op2 && !Op3 && Op4)
7544 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
7546 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
7547 if (Op1 && !Op2 && !Op3 && !Op4)
7548 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
7550 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
7552 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
7556 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7557 EVT T = Op.getValueType();
7558 DebugLoc dl = Op.getDebugLoc();
7561 switch(T.getSimpleVT().SimpleTy) {
7563 assert(false && "Invalid value type!");
7564 case MVT::i8: Reg = X86::AL; size = 1; break;
7565 case MVT::i16: Reg = X86::AX; size = 2; break;
7566 case MVT::i32: Reg = X86::EAX; size = 4; break;
7568 assert(Subtarget->is64Bit() && "Node not type legal!");
7569 Reg = X86::RAX; size = 8;
7572 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7573 Op.getOperand(2), SDValue());
7574 SDValue Ops[] = { cpIn.getValue(0),
7577 DAG.getTargetConstant(size, MVT::i8),
7579 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7580 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7582 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7586 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7587 SelectionDAG &DAG) const {
7588 assert(Subtarget->is64Bit() && "Result not type legalized?");
7589 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7590 SDValue TheChain = Op.getOperand(0);
7591 DebugLoc dl = Op.getDebugLoc();
7592 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7593 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7594 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7596 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7597 DAG.getConstant(32, MVT::i8));
7599 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7602 return DAG.getMergeValues(Ops, 2, dl);
7605 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7606 SelectionDAG &DAG) const {
7607 EVT SrcVT = Op.getOperand(0).getValueType();
7608 EVT DstVT = Op.getValueType();
7609 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7610 Subtarget->hasMMX() && !DisableMMX) &&
7611 "Unexpected custom BIT_CONVERT");
7612 assert((DstVT == MVT::i64 ||
7613 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7614 "Unexpected custom BIT_CONVERT");
7615 // i64 <=> MMX conversions are Legal.
7616 if (SrcVT==MVT::i64 && DstVT.isVector())
7618 if (DstVT==MVT::i64 && SrcVT.isVector())
7620 // MMX <=> MMX conversions are Legal.
7621 if (SrcVT.isVector() && DstVT.isVector())
7623 // All other conversions need to be expanded.
7626 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7627 SDNode *Node = Op.getNode();
7628 DebugLoc dl = Node->getDebugLoc();
7629 EVT T = Node->getValueType(0);
7630 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7631 DAG.getConstant(0, T), Node->getOperand(2));
7632 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7633 cast<AtomicSDNode>(Node)->getMemoryVT(),
7634 Node->getOperand(0),
7635 Node->getOperand(1), negOp,
7636 cast<AtomicSDNode>(Node)->getSrcValue(),
7637 cast<AtomicSDNode>(Node)->getAlignment());
7640 /// LowerOperation - Provide custom lowering hooks for some operations.
7642 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7643 switch (Op.getOpcode()) {
7644 default: llvm_unreachable("Should not custom lower this!");
7645 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
7646 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7647 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7648 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7649 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7650 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7651 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7652 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7653 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7654 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7655 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7656 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7657 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7658 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7659 case ISD::SHL_PARTS:
7660 case ISD::SRA_PARTS:
7661 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7662 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7663 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7664 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7665 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7666 case ISD::FABS: return LowerFABS(Op, DAG);
7667 case ISD::FNEG: return LowerFNEG(Op, DAG);
7668 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7669 case ISD::SETCC: return LowerSETCC(Op, DAG);
7670 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7671 case ISD::SELECT: return LowerSELECT(Op, DAG);
7672 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7673 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7674 case ISD::VASTART: return LowerVASTART(Op, DAG);
7675 case ISD::VAARG: return LowerVAARG(Op, DAG);
7676 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7677 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7678 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7679 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7680 case ISD::FRAME_TO_ARGS_OFFSET:
7681 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7682 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7683 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7684 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7685 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7686 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7687 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7688 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7694 case ISD::UMULO: return LowerXALUO(Op, DAG);
7695 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7696 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7700 void X86TargetLowering::
7701 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7702 SelectionDAG &DAG, unsigned NewOp) const {
7703 EVT T = Node->getValueType(0);
7704 DebugLoc dl = Node->getDebugLoc();
7705 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7707 SDValue Chain = Node->getOperand(0);
7708 SDValue In1 = Node->getOperand(1);
7709 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7710 Node->getOperand(2), DAG.getIntPtrConstant(0));
7711 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7712 Node->getOperand(2), DAG.getIntPtrConstant(1));
7713 SDValue Ops[] = { Chain, In1, In2L, In2H };
7714 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7716 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7717 cast<MemSDNode>(Node)->getMemOperand());
7718 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7719 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7720 Results.push_back(Result.getValue(2));
7723 /// ReplaceNodeResults - Replace a node with an illegal result type
7724 /// with a new node built out of custom code.
7725 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7726 SmallVectorImpl<SDValue>&Results,
7727 SelectionDAG &DAG) const {
7728 DebugLoc dl = N->getDebugLoc();
7729 switch (N->getOpcode()) {
7731 assert(false && "Do not know how to custom type legalize this operation!");
7733 case ISD::FP_TO_SINT: {
7734 std::pair<SDValue,SDValue> Vals =
7735 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7736 SDValue FIST = Vals.first, StackSlot = Vals.second;
7737 if (FIST.getNode() != 0) {
7738 EVT VT = N->getValueType(0);
7739 // Return a load from the stack slot.
7740 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7745 case ISD::READCYCLECOUNTER: {
7746 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7747 SDValue TheChain = N->getOperand(0);
7748 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7749 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7751 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7753 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7754 SDValue Ops[] = { eax, edx };
7755 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7756 Results.push_back(edx.getValue(1));
7759 case ISD::ATOMIC_CMP_SWAP: {
7760 EVT T = N->getValueType(0);
7761 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7762 SDValue cpInL, cpInH;
7763 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7764 DAG.getConstant(0, MVT::i32));
7765 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7766 DAG.getConstant(1, MVT::i32));
7767 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7768 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7770 SDValue swapInL, swapInH;
7771 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7772 DAG.getConstant(0, MVT::i32));
7773 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7774 DAG.getConstant(1, MVT::i32));
7775 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7777 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7778 swapInL.getValue(1));
7779 SDValue Ops[] = { swapInH.getValue(0),
7781 swapInH.getValue(1) };
7782 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7783 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7784 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7785 MVT::i32, Result.getValue(1));
7786 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7787 MVT::i32, cpOutL.getValue(2));
7788 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7789 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7790 Results.push_back(cpOutH.getValue(1));
7793 case ISD::ATOMIC_LOAD_ADD:
7794 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7796 case ISD::ATOMIC_LOAD_AND:
7797 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7799 case ISD::ATOMIC_LOAD_NAND:
7800 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7802 case ISD::ATOMIC_LOAD_OR:
7803 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7805 case ISD::ATOMIC_LOAD_SUB:
7806 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7808 case ISD::ATOMIC_LOAD_XOR:
7809 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7811 case ISD::ATOMIC_SWAP:
7812 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7817 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7819 default: return NULL;
7820 case X86ISD::BSF: return "X86ISD::BSF";
7821 case X86ISD::BSR: return "X86ISD::BSR";
7822 case X86ISD::SHLD: return "X86ISD::SHLD";
7823 case X86ISD::SHRD: return "X86ISD::SHRD";
7824 case X86ISD::FAND: return "X86ISD::FAND";
7825 case X86ISD::FOR: return "X86ISD::FOR";
7826 case X86ISD::FXOR: return "X86ISD::FXOR";
7827 case X86ISD::FSRL: return "X86ISD::FSRL";
7828 case X86ISD::FILD: return "X86ISD::FILD";
7829 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7830 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7831 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7832 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7833 case X86ISD::FLD: return "X86ISD::FLD";
7834 case X86ISD::FST: return "X86ISD::FST";
7835 case X86ISD::CALL: return "X86ISD::CALL";
7836 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7837 case X86ISD::BT: return "X86ISD::BT";
7838 case X86ISD::CMP: return "X86ISD::CMP";
7839 case X86ISD::COMI: return "X86ISD::COMI";
7840 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7841 case X86ISD::SETCC: return "X86ISD::SETCC";
7842 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7843 case X86ISD::CMOV: return "X86ISD::CMOV";
7844 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7845 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7846 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7847 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7848 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7849 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7850 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7851 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7852 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7853 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7854 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7855 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7856 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7857 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7858 case X86ISD::FMAX: return "X86ISD::FMAX";
7859 case X86ISD::FMIN: return "X86ISD::FMIN";
7860 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7861 case X86ISD::FRCP: return "X86ISD::FRCP";
7862 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7863 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
7864 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7865 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7866 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7867 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7868 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7869 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7870 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7871 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7872 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7873 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7874 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7875 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7876 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7877 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7878 case X86ISD::VSHL: return "X86ISD::VSHL";
7879 case X86ISD::VSRL: return "X86ISD::VSRL";
7880 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7881 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7882 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7883 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7884 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7885 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7886 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7887 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7888 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7889 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7890 case X86ISD::ADD: return "X86ISD::ADD";
7891 case X86ISD::SUB: return "X86ISD::SUB";
7892 case X86ISD::SMUL: return "X86ISD::SMUL";
7893 case X86ISD::UMUL: return "X86ISD::UMUL";
7894 case X86ISD::INC: return "X86ISD::INC";
7895 case X86ISD::DEC: return "X86ISD::DEC";
7896 case X86ISD::OR: return "X86ISD::OR";
7897 case X86ISD::XOR: return "X86ISD::XOR";
7898 case X86ISD::AND: return "X86ISD::AND";
7899 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7900 case X86ISD::PTEST: return "X86ISD::PTEST";
7901 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7902 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7906 // isLegalAddressingMode - Return true if the addressing mode represented
7907 // by AM is legal for this target, for a load/store of the specified type.
7908 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7909 const Type *Ty) const {
7910 // X86 supports extremely general addressing modes.
7911 CodeModel::Model M = getTargetMachine().getCodeModel();
7913 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7914 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7919 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7921 // If a reference to this global requires an extra load, we can't fold it.
7922 if (isGlobalStubReference(GVFlags))
7925 // If BaseGV requires a register for the PIC base, we cannot also have a
7926 // BaseReg specified.
7927 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7930 // If lower 4G is not available, then we must use rip-relative addressing.
7931 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7941 // These scales always work.
7946 // These scales are formed with basereg+scalereg. Only accept if there is
7951 default: // Other stuff never works.
7959 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7960 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7962 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7963 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7964 if (NumBits1 <= NumBits2)
7969 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7970 if (!VT1.isInteger() || !VT2.isInteger())
7972 unsigned NumBits1 = VT1.getSizeInBits();
7973 unsigned NumBits2 = VT2.getSizeInBits();
7974 if (NumBits1 <= NumBits2)
7979 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7980 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7981 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7984 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7985 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7986 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7989 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7990 // i16 instructions are longer (0x66 prefix) and potentially slower.
7991 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7994 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7995 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7996 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7997 /// are assumed to be legal.
7999 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
8001 // Very little shuffling can be done for 64-bit vectors right now.
8002 if (VT.getSizeInBits() == 64)
8003 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8005 // FIXME: pshufb, blends, shifts.
8006 return (VT.getVectorNumElements() == 2 ||
8007 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8008 isMOVLMask(M, VT) ||
8009 isSHUFPMask(M, VT) ||
8010 isPSHUFDMask(M, VT) ||
8011 isPSHUFHWMask(M, VT) ||
8012 isPSHUFLWMask(M, VT) ||
8013 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8014 isUNPCKLMask(M, VT) ||
8015 isUNPCKHMask(M, VT) ||
8016 isUNPCKL_v_undef_Mask(M, VT) ||
8017 isUNPCKH_v_undef_Mask(M, VT));
8021 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8023 unsigned NumElts = VT.getVectorNumElements();
8024 // FIXME: This collection of masks seems suspect.
8027 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8028 return (isMOVLMask(Mask, VT) ||
8029 isCommutedMOVLMask(Mask, VT, true) ||
8030 isSHUFPMask(Mask, VT) ||
8031 isCommutedSHUFPMask(Mask, VT));
8036 //===----------------------------------------------------------------------===//
8037 // X86 Scheduler Hooks
8038 //===----------------------------------------------------------------------===//
8040 // private utility function
8042 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8043 MachineBasicBlock *MBB,
8050 TargetRegisterClass *RC,
8051 bool invSrc) const {
8052 // For the atomic bitwise operator, we generate
8055 // ld t1 = [bitinstr.addr]
8056 // op t2 = t1, [bitinstr.val]
8058 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8060 // fallthrough -->nextMBB
8061 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8062 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8063 MachineFunction::iterator MBBIter = MBB;
8066 /// First build the CFG
8067 MachineFunction *F = MBB->getParent();
8068 MachineBasicBlock *thisMBB = MBB;
8069 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8070 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8071 F->insert(MBBIter, newMBB);
8072 F->insert(MBBIter, nextMBB);
8074 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8075 nextMBB->splice(nextMBB->begin(), thisMBB,
8076 llvm::next(MachineBasicBlock::iterator(bInstr)),
8078 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8080 // Update thisMBB to fall through to newMBB
8081 thisMBB->addSuccessor(newMBB);
8083 // newMBB jumps to itself and fall through to nextMBB
8084 newMBB->addSuccessor(nextMBB);
8085 newMBB->addSuccessor(newMBB);
8087 // Insert instructions into newMBB based on incoming instruction
8088 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8089 "unexpected number of operands");
8090 DebugLoc dl = bInstr->getDebugLoc();
8091 MachineOperand& destOper = bInstr->getOperand(0);
8092 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8093 int numArgs = bInstr->getNumOperands() - 1;
8094 for (int i=0; i < numArgs; ++i)
8095 argOpers[i] = &bInstr->getOperand(i+1);
8097 // x86 address has 4 operands: base, index, scale, and displacement
8098 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8099 int valArgIndx = lastAddrIndx + 1;
8101 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8102 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8103 for (int i=0; i <= lastAddrIndx; ++i)
8104 (*MIB).addOperand(*argOpers[i]);
8106 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8108 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8113 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8114 assert((argOpers[valArgIndx]->isReg() ||
8115 argOpers[valArgIndx]->isImm()) &&
8117 if (argOpers[valArgIndx]->isReg())
8118 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8120 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8122 (*MIB).addOperand(*argOpers[valArgIndx]);
8124 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
8127 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8128 for (int i=0; i <= lastAddrIndx; ++i)
8129 (*MIB).addOperand(*argOpers[i]);
8131 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8132 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8133 bInstr->memoperands_end());
8135 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8139 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8141 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8145 // private utility function: 64 bit atomics on 32 bit host.
8147 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8148 MachineBasicBlock *MBB,
8153 bool invSrc) const {
8154 // For the atomic bitwise operator, we generate
8155 // thisMBB (instructions are in pairs, except cmpxchg8b)
8156 // ld t1,t2 = [bitinstr.addr]
8158 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8159 // op t5, t6 <- out1, out2, [bitinstr.val]
8160 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8161 // mov ECX, EBX <- t5, t6
8162 // mov EAX, EDX <- t1, t2
8163 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8164 // mov t3, t4 <- EAX, EDX
8166 // result in out1, out2
8167 // fallthrough -->nextMBB
8169 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8170 const unsigned LoadOpc = X86::MOV32rm;
8171 const unsigned NotOpc = X86::NOT32r;
8172 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8173 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8174 MachineFunction::iterator MBBIter = MBB;
8177 /// First build the CFG
8178 MachineFunction *F = MBB->getParent();
8179 MachineBasicBlock *thisMBB = MBB;
8180 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8181 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8182 F->insert(MBBIter, newMBB);
8183 F->insert(MBBIter, nextMBB);
8185 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8186 nextMBB->splice(nextMBB->begin(), thisMBB,
8187 llvm::next(MachineBasicBlock::iterator(bInstr)),
8189 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8191 // Update thisMBB to fall through to newMBB
8192 thisMBB->addSuccessor(newMBB);
8194 // newMBB jumps to itself and fall through to nextMBB
8195 newMBB->addSuccessor(nextMBB);
8196 newMBB->addSuccessor(newMBB);
8198 DebugLoc dl = bInstr->getDebugLoc();
8199 // Insert instructions into newMBB based on incoming instruction
8200 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8201 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
8202 "unexpected number of operands");
8203 MachineOperand& dest1Oper = bInstr->getOperand(0);
8204 MachineOperand& dest2Oper = bInstr->getOperand(1);
8205 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8206 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
8207 argOpers[i] = &bInstr->getOperand(i+2);
8209 // We use some of the operands multiple times, so conservatively just
8210 // clear any kill flags that might be present.
8211 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8212 argOpers[i]->setIsKill(false);
8215 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8216 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8218 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8219 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8220 for (int i=0; i <= lastAddrIndx; ++i)
8221 (*MIB).addOperand(*argOpers[i]);
8222 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8223 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8224 // add 4 to displacement.
8225 for (int i=0; i <= lastAddrIndx-2; ++i)
8226 (*MIB).addOperand(*argOpers[i]);
8227 MachineOperand newOp3 = *(argOpers[3]);
8229 newOp3.setImm(newOp3.getImm()+4);
8231 newOp3.setOffset(newOp3.getOffset()+4);
8232 (*MIB).addOperand(newOp3);
8233 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8235 // t3/4 are defined later, at the bottom of the loop
8236 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8237 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8238 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8239 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8240 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8241 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8243 // The subsequent operations should be using the destination registers of
8244 //the PHI instructions.
8246 t1 = F->getRegInfo().createVirtualRegister(RC);
8247 t2 = F->getRegInfo().createVirtualRegister(RC);
8248 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8249 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8251 t1 = dest1Oper.getReg();
8252 t2 = dest2Oper.getReg();
8255 int valArgIndx = lastAddrIndx + 1;
8256 assert((argOpers[valArgIndx]->isReg() ||
8257 argOpers[valArgIndx]->isImm()) &&
8259 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8260 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8261 if (argOpers[valArgIndx]->isReg())
8262 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8264 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8265 if (regOpcL != X86::MOV32rr)
8267 (*MIB).addOperand(*argOpers[valArgIndx]);
8268 assert(argOpers[valArgIndx + 1]->isReg() ==
8269 argOpers[valArgIndx]->isReg());
8270 assert(argOpers[valArgIndx + 1]->isImm() ==
8271 argOpers[valArgIndx]->isImm());
8272 if (argOpers[valArgIndx + 1]->isReg())
8273 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8275 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8276 if (regOpcH != X86::MOV32rr)
8278 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8280 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8282 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
8285 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
8287 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
8290 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8291 for (int i=0; i <= lastAddrIndx; ++i)
8292 (*MIB).addOperand(*argOpers[i]);
8294 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8295 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8296 bInstr->memoperands_end());
8298 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
8299 MIB.addReg(X86::EAX);
8300 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
8301 MIB.addReg(X86::EDX);
8304 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8306 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8310 // private utility function
8312 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8313 MachineBasicBlock *MBB,
8314 unsigned cmovOpc) const {
8315 // For the atomic min/max operator, we generate
8318 // ld t1 = [min/max.addr]
8319 // mov t2 = [min/max.val]
8321 // cmov[cond] t2 = t1
8323 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8325 // fallthrough -->nextMBB
8327 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8328 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8329 MachineFunction::iterator MBBIter = MBB;
8332 /// First build the CFG
8333 MachineFunction *F = MBB->getParent();
8334 MachineBasicBlock *thisMBB = MBB;
8335 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8336 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8337 F->insert(MBBIter, newMBB);
8338 F->insert(MBBIter, nextMBB);
8340 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8341 nextMBB->splice(nextMBB->begin(), thisMBB,
8342 llvm::next(MachineBasicBlock::iterator(mInstr)),
8344 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8346 // Update thisMBB to fall through to newMBB
8347 thisMBB->addSuccessor(newMBB);
8349 // newMBB jumps to newMBB and fall through to nextMBB
8350 newMBB->addSuccessor(nextMBB);
8351 newMBB->addSuccessor(newMBB);
8353 DebugLoc dl = mInstr->getDebugLoc();
8354 // Insert instructions into newMBB based on incoming instruction
8355 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8356 "unexpected number of operands");
8357 MachineOperand& destOper = mInstr->getOperand(0);
8358 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8359 int numArgs = mInstr->getNumOperands() - 1;
8360 for (int i=0; i < numArgs; ++i)
8361 argOpers[i] = &mInstr->getOperand(i+1);
8363 // x86 address has 4 operands: base, index, scale, and displacement
8364 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8365 int valArgIndx = lastAddrIndx + 1;
8367 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8368 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8369 for (int i=0; i <= lastAddrIndx; ++i)
8370 (*MIB).addOperand(*argOpers[i]);
8372 // We only support register and immediate values
8373 assert((argOpers[valArgIndx]->isReg() ||
8374 argOpers[valArgIndx]->isImm()) &&
8377 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8378 if (argOpers[valArgIndx]->isReg())
8379 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
8381 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8382 (*MIB).addOperand(*argOpers[valArgIndx]);
8384 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8387 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8392 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8393 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8397 // Cmp and exchange if none has modified the memory location
8398 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8399 for (int i=0; i <= lastAddrIndx; ++i)
8400 (*MIB).addOperand(*argOpers[i]);
8402 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8403 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8404 mInstr->memoperands_end());
8406 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8407 MIB.addReg(X86::EAX);
8410 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8412 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
8416 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8417 // all of this code can be replaced with that in the .td file.
8419 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8420 unsigned numArgs, bool memArg) const {
8422 DebugLoc dl = MI->getDebugLoc();
8423 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8427 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8429 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8431 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8433 for (unsigned i = 0; i < numArgs; ++i) {
8434 MachineOperand &Op = MI->getOperand(i+1);
8436 if (!(Op.isReg() && Op.isImplicit()))
8440 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8443 MI->eraseFromParent();
8449 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8451 MachineBasicBlock *MBB) const {
8452 // Emit code to save XMM registers to the stack. The ABI says that the
8453 // number of registers to save is given in %al, so it's theoretically
8454 // possible to do an indirect jump trick to avoid saving all of them,
8455 // however this code takes a simpler approach and just executes all
8456 // of the stores if %al is non-zero. It's less code, and it's probably
8457 // easier on the hardware branch predictor, and stores aren't all that
8458 // expensive anyway.
8460 // Create the new basic blocks. One block contains all the XMM stores,
8461 // and one block is the final destination regardless of whether any
8462 // stores were performed.
8463 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8464 MachineFunction *F = MBB->getParent();
8465 MachineFunction::iterator MBBIter = MBB;
8467 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8468 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8469 F->insert(MBBIter, XMMSaveMBB);
8470 F->insert(MBBIter, EndMBB);
8472 // Transfer the remainder of MBB and its successor edges to EndMBB.
8473 EndMBB->splice(EndMBB->begin(), MBB,
8474 llvm::next(MachineBasicBlock::iterator(MI)),
8476 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
8478 // The original block will now fall through to the XMM save block.
8479 MBB->addSuccessor(XMMSaveMBB);
8480 // The XMMSaveMBB will fall through to the end block.
8481 XMMSaveMBB->addSuccessor(EndMBB);
8483 // Now add the instructions.
8484 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8485 DebugLoc DL = MI->getDebugLoc();
8487 unsigned CountReg = MI->getOperand(0).getReg();
8488 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8489 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8491 if (!Subtarget->isTargetWin64()) {
8492 // If %al is 0, branch around the XMM save block.
8493 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8494 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8495 MBB->addSuccessor(EndMBB);
8498 // In the XMM save block, save all the XMM argument registers.
8499 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8500 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8501 MachineMemOperand *MMO =
8502 F->getMachineMemOperand(
8503 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8504 MachineMemOperand::MOStore, Offset,
8505 /*Size=*/16, /*Align=*/16);
8506 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8507 .addFrameIndex(RegSaveFrameIndex)
8508 .addImm(/*Scale=*/1)
8509 .addReg(/*IndexReg=*/0)
8510 .addImm(/*Disp=*/Offset)
8511 .addReg(/*Segment=*/0)
8512 .addReg(MI->getOperand(i).getReg())
8513 .addMemOperand(MMO);
8516 MI->eraseFromParent(); // The pseudo instruction is gone now.
8522 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8523 MachineBasicBlock *BB) const {
8524 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8525 DebugLoc DL = MI->getDebugLoc();
8527 // To "insert" a SELECT_CC instruction, we actually have to insert the
8528 // diamond control-flow pattern. The incoming instruction knows the
8529 // destination vreg to set, the condition code register to branch on, the
8530 // true/false values to select between, and a branch opcode to use.
8531 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8532 MachineFunction::iterator It = BB;
8538 // cmpTY ccX, r1, r2
8540 // fallthrough --> copy0MBB
8541 MachineBasicBlock *thisMBB = BB;
8542 MachineFunction *F = BB->getParent();
8543 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8544 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8545 F->insert(It, copy0MBB);
8546 F->insert(It, sinkMBB);
8548 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8549 // live into the sink and copy blocks.
8550 const MachineFunction *MF = BB->getParent();
8551 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8552 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8554 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
8555 const MachineOperand &MO = MI->getOperand(I);
8556 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
8557 unsigned Reg = MO.getReg();
8558 if (Reg != X86::EFLAGS) continue;
8559 copy0MBB->addLiveIn(Reg);
8560 sinkMBB->addLiveIn(Reg);
8563 // Transfer the remainder of BB and its successor edges to sinkMBB.
8564 sinkMBB->splice(sinkMBB->begin(), BB,
8565 llvm::next(MachineBasicBlock::iterator(MI)),
8567 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
8569 // Add the true and fallthrough blocks as its successors.
8570 BB->addSuccessor(copy0MBB);
8571 BB->addSuccessor(sinkMBB);
8573 // Create the conditional branch instruction.
8575 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8576 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8579 // %FalseValue = ...
8580 // # fallthrough to sinkMBB
8581 copy0MBB->addSuccessor(sinkMBB);
8584 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8586 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
8587 TII->get(X86::PHI), MI->getOperand(0).getReg())
8588 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8589 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8591 MI->eraseFromParent(); // The pseudo instruction is gone now.
8596 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8597 MachineBasicBlock *BB) const {
8598 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8599 DebugLoc DL = MI->getDebugLoc();
8601 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8602 // non-trivial part is impdef of ESP.
8603 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8606 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
8607 .addExternalSymbol("_alloca")
8608 .addReg(X86::EAX, RegState::Implicit)
8609 .addReg(X86::ESP, RegState::Implicit)
8610 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8611 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8613 MI->eraseFromParent(); // The pseudo instruction is gone now.
8618 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8619 MachineBasicBlock *BB) const {
8620 // This is pretty easy. We're taking the value that we received from
8621 // our load from the relocation, sticking it in either RDI (x86-64)
8622 // or EAX and doing an indirect call. The return value will then
8623 // be in the normal return register.
8624 const X86InstrInfo *TII
8625 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8626 DebugLoc DL = MI->getDebugLoc();
8627 MachineFunction *F = BB->getParent();
8629 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8631 if (Subtarget->is64Bit()) {
8632 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8633 TII->get(X86::MOV64rm), X86::RDI)
8635 .addImm(0).addReg(0)
8636 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8637 MI->getOperand(3).getTargetFlags())
8639 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
8640 addDirectMem(MIB, X86::RDI);
8641 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8642 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8643 TII->get(X86::MOV32rm), X86::EAX)
8645 .addImm(0).addReg(0)
8646 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8647 MI->getOperand(3).getTargetFlags())
8649 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8650 addDirectMem(MIB, X86::EAX);
8652 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8653 TII->get(X86::MOV32rm), X86::EAX)
8654 .addReg(TII->getGlobalBaseReg(F))
8655 .addImm(0).addReg(0)
8656 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8657 MI->getOperand(3).getTargetFlags())
8659 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8660 addDirectMem(MIB, X86::EAX);
8663 MI->eraseFromParent(); // The pseudo instruction is gone now.
8668 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8669 MachineBasicBlock *BB) const {
8670 switch (MI->getOpcode()) {
8671 default: assert(false && "Unexpected instr type to insert");
8672 case X86::MINGW_ALLOCA:
8673 return EmitLoweredMingwAlloca(MI, BB);
8674 case X86::TLSCall_32:
8675 case X86::TLSCall_64:
8676 return EmitLoweredTLSCall(MI, BB);
8678 case X86::CMOV_V1I64:
8679 case X86::CMOV_FR32:
8680 case X86::CMOV_FR64:
8681 case X86::CMOV_V4F32:
8682 case X86::CMOV_V2F64:
8683 case X86::CMOV_V2I64:
8684 case X86::CMOV_GR16:
8685 case X86::CMOV_GR32:
8686 case X86::CMOV_RFP32:
8687 case X86::CMOV_RFP64:
8688 case X86::CMOV_RFP80:
8689 return EmitLoweredSelect(MI, BB);
8691 case X86::FP32_TO_INT16_IN_MEM:
8692 case X86::FP32_TO_INT32_IN_MEM:
8693 case X86::FP32_TO_INT64_IN_MEM:
8694 case X86::FP64_TO_INT16_IN_MEM:
8695 case X86::FP64_TO_INT32_IN_MEM:
8696 case X86::FP64_TO_INT64_IN_MEM:
8697 case X86::FP80_TO_INT16_IN_MEM:
8698 case X86::FP80_TO_INT32_IN_MEM:
8699 case X86::FP80_TO_INT64_IN_MEM: {
8700 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8701 DebugLoc DL = MI->getDebugLoc();
8703 // Change the floating point control register to use "round towards zero"
8704 // mode when truncating to an integer value.
8705 MachineFunction *F = BB->getParent();
8706 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8707 addFrameReference(BuildMI(*BB, MI, DL,
8708 TII->get(X86::FNSTCW16m)), CWFrameIdx);
8710 // Load the old value of the high byte of the control word...
8712 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8713 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
8716 // Set the high part to be round to zero...
8717 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8720 // Reload the modified control word now...
8721 addFrameReference(BuildMI(*BB, MI, DL,
8722 TII->get(X86::FLDCW16m)), CWFrameIdx);
8724 // Restore the memory image of control word to original value
8725 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8728 // Get the X86 opcode to use.
8730 switch (MI->getOpcode()) {
8731 default: llvm_unreachable("illegal opcode!");
8732 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8733 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8734 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8735 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8736 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8737 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8738 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8739 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8740 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8744 MachineOperand &Op = MI->getOperand(0);
8746 AM.BaseType = X86AddressMode::RegBase;
8747 AM.Base.Reg = Op.getReg();
8749 AM.BaseType = X86AddressMode::FrameIndexBase;
8750 AM.Base.FrameIndex = Op.getIndex();
8752 Op = MI->getOperand(1);
8754 AM.Scale = Op.getImm();
8755 Op = MI->getOperand(2);
8757 AM.IndexReg = Op.getImm();
8758 Op = MI->getOperand(3);
8759 if (Op.isGlobal()) {
8760 AM.GV = Op.getGlobal();
8762 AM.Disp = Op.getImm();
8764 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
8765 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
8767 // Reload the original control word now.
8768 addFrameReference(BuildMI(*BB, MI, DL,
8769 TII->get(X86::FLDCW16m)), CWFrameIdx);
8771 MI->eraseFromParent(); // The pseudo instruction is gone now.
8774 // String/text processing lowering.
8775 case X86::PCMPISTRM128REG:
8776 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8777 case X86::PCMPISTRM128MEM:
8778 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8779 case X86::PCMPESTRM128REG:
8780 return EmitPCMP(MI, BB, 5, false /* in mem */);
8781 case X86::PCMPESTRM128MEM:
8782 return EmitPCMP(MI, BB, 5, true /* in mem */);
8785 case X86::ATOMAND32:
8786 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8787 X86::AND32ri, X86::MOV32rm,
8789 X86::NOT32r, X86::EAX,
8790 X86::GR32RegisterClass);
8792 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8793 X86::OR32ri, X86::MOV32rm,
8795 X86::NOT32r, X86::EAX,
8796 X86::GR32RegisterClass);
8797 case X86::ATOMXOR32:
8798 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8799 X86::XOR32ri, X86::MOV32rm,
8801 X86::NOT32r, X86::EAX,
8802 X86::GR32RegisterClass);
8803 case X86::ATOMNAND32:
8804 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8805 X86::AND32ri, X86::MOV32rm,
8807 X86::NOT32r, X86::EAX,
8808 X86::GR32RegisterClass, true);
8809 case X86::ATOMMIN32:
8810 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8811 case X86::ATOMMAX32:
8812 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8813 case X86::ATOMUMIN32:
8814 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8815 case X86::ATOMUMAX32:
8816 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8818 case X86::ATOMAND16:
8819 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8820 X86::AND16ri, X86::MOV16rm,
8822 X86::NOT16r, X86::AX,
8823 X86::GR16RegisterClass);
8825 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8826 X86::OR16ri, X86::MOV16rm,
8828 X86::NOT16r, X86::AX,
8829 X86::GR16RegisterClass);
8830 case X86::ATOMXOR16:
8831 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8832 X86::XOR16ri, X86::MOV16rm,
8834 X86::NOT16r, X86::AX,
8835 X86::GR16RegisterClass);
8836 case X86::ATOMNAND16:
8837 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8838 X86::AND16ri, X86::MOV16rm,
8840 X86::NOT16r, X86::AX,
8841 X86::GR16RegisterClass, true);
8842 case X86::ATOMMIN16:
8843 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8844 case X86::ATOMMAX16:
8845 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8846 case X86::ATOMUMIN16:
8847 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8848 case X86::ATOMUMAX16:
8849 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8852 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8853 X86::AND8ri, X86::MOV8rm,
8855 X86::NOT8r, X86::AL,
8856 X86::GR8RegisterClass);
8858 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8859 X86::OR8ri, X86::MOV8rm,
8861 X86::NOT8r, X86::AL,
8862 X86::GR8RegisterClass);
8864 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8865 X86::XOR8ri, X86::MOV8rm,
8867 X86::NOT8r, X86::AL,
8868 X86::GR8RegisterClass);
8869 case X86::ATOMNAND8:
8870 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8871 X86::AND8ri, X86::MOV8rm,
8873 X86::NOT8r, X86::AL,
8874 X86::GR8RegisterClass, true);
8875 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8876 // This group is for 64-bit host.
8877 case X86::ATOMAND64:
8878 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8879 X86::AND64ri32, X86::MOV64rm,
8881 X86::NOT64r, X86::RAX,
8882 X86::GR64RegisterClass);
8884 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8885 X86::OR64ri32, X86::MOV64rm,
8887 X86::NOT64r, X86::RAX,
8888 X86::GR64RegisterClass);
8889 case X86::ATOMXOR64:
8890 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8891 X86::XOR64ri32, X86::MOV64rm,
8893 X86::NOT64r, X86::RAX,
8894 X86::GR64RegisterClass);
8895 case X86::ATOMNAND64:
8896 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8897 X86::AND64ri32, X86::MOV64rm,
8899 X86::NOT64r, X86::RAX,
8900 X86::GR64RegisterClass, true);
8901 case X86::ATOMMIN64:
8902 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8903 case X86::ATOMMAX64:
8904 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8905 case X86::ATOMUMIN64:
8906 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8907 case X86::ATOMUMAX64:
8908 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8910 // This group does 64-bit operations on a 32-bit host.
8911 case X86::ATOMAND6432:
8912 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8913 X86::AND32rr, X86::AND32rr,
8914 X86::AND32ri, X86::AND32ri,
8916 case X86::ATOMOR6432:
8917 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8918 X86::OR32rr, X86::OR32rr,
8919 X86::OR32ri, X86::OR32ri,
8921 case X86::ATOMXOR6432:
8922 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8923 X86::XOR32rr, X86::XOR32rr,
8924 X86::XOR32ri, X86::XOR32ri,
8926 case X86::ATOMNAND6432:
8927 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8928 X86::AND32rr, X86::AND32rr,
8929 X86::AND32ri, X86::AND32ri,
8931 case X86::ATOMADD6432:
8932 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8933 X86::ADD32rr, X86::ADC32rr,
8934 X86::ADD32ri, X86::ADC32ri,
8936 case X86::ATOMSUB6432:
8937 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8938 X86::SUB32rr, X86::SBB32rr,
8939 X86::SUB32ri, X86::SBB32ri,
8941 case X86::ATOMSWAP6432:
8942 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8943 X86::MOV32rr, X86::MOV32rr,
8944 X86::MOV32ri, X86::MOV32ri,
8946 case X86::VASTART_SAVE_XMM_REGS:
8947 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8951 //===----------------------------------------------------------------------===//
8952 // X86 Optimization Hooks
8953 //===----------------------------------------------------------------------===//
8955 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8959 const SelectionDAG &DAG,
8960 unsigned Depth) const {
8961 unsigned Opc = Op.getOpcode();
8962 assert((Opc >= ISD::BUILTIN_OP_END ||
8963 Opc == ISD::INTRINSIC_WO_CHAIN ||
8964 Opc == ISD::INTRINSIC_W_CHAIN ||
8965 Opc == ISD::INTRINSIC_VOID) &&
8966 "Should use MaskedValueIsZero if you don't know whether Op"
8967 " is a target node!");
8969 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8981 // These nodes' second result is a boolean.
8982 if (Op.getResNo() == 0)
8986 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8987 Mask.getBitWidth() - 1);
8992 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8993 /// node is a GlobalAddress + offset.
8994 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8995 const GlobalValue* &GA,
8996 int64_t &Offset) const {
8997 if (N->getOpcode() == X86ISD::Wrapper) {
8998 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8999 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
9000 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
9004 return TargetLowering::isGAPlusOffset(N, GA, Offset);
9007 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
9008 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
9009 /// if the load addresses are consecutive, non-overlapping, and in the right
9011 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
9012 const TargetLowering &TLI) {
9013 DebugLoc dl = N->getDebugLoc();
9014 EVT VT = N->getValueType(0);
9015 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9017 if (VT.getSizeInBits() != 128)
9020 SmallVector<SDValue, 16> Elts;
9021 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
9022 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
9024 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
9027 /// PerformShuffleCombine - Detect vector gather/scatter index generation
9028 /// and convert it from being a bunch of shuffles and extracts to a simple
9029 /// store and scalar loads to extract the elements.
9030 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
9031 const TargetLowering &TLI) {
9032 SDValue InputVector = N->getOperand(0);
9034 // Only operate on vectors of 4 elements, where the alternative shuffling
9035 // gets to be more expensive.
9036 if (InputVector.getValueType() != MVT::v4i32)
9039 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
9040 // single use which is a sign-extend or zero-extend, and all elements are
9042 SmallVector<SDNode *, 4> Uses;
9043 unsigned ExtractedElements = 0;
9044 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
9045 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
9046 if (UI.getUse().getResNo() != InputVector.getResNo())
9049 SDNode *Extract = *UI;
9050 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9053 if (Extract->getValueType(0) != MVT::i32)
9055 if (!Extract->hasOneUse())
9057 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9058 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9060 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9063 // Record which element was extracted.
9064 ExtractedElements |=
9065 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9067 Uses.push_back(Extract);
9070 // If not all the elements were used, this may not be worthwhile.
9071 if (ExtractedElements != 15)
9074 // Ok, we've now decided to do the transformation.
9075 DebugLoc dl = InputVector.getDebugLoc();
9077 // Store the value to a temporary stack slot.
9078 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9079 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL,
9080 0, false, false, 0);
9082 // Replace each use (extract) with a load of the appropriate element.
9083 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9084 UE = Uses.end(); UI != UE; ++UI) {
9085 SDNode *Extract = *UI;
9087 // Compute the element's address.
9088 SDValue Idx = Extract->getOperand(1);
9090 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9091 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9092 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9094 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
9095 OffsetVal, StackPtr);
9098 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
9099 ScalarAddr, NULL, 0, false, false, 0);
9101 // Replace the exact with the load.
9102 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9105 // The replacement was made in place; don't return anything.
9109 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9110 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9111 const X86Subtarget *Subtarget) {
9112 DebugLoc DL = N->getDebugLoc();
9113 SDValue Cond = N->getOperand(0);
9114 // Get the LHS/RHS of the select.
9115 SDValue LHS = N->getOperand(1);
9116 SDValue RHS = N->getOperand(2);
9118 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9119 // instructions match the semantics of the common C idiom x<y?x:y but not
9120 // x<=y?x:y, because of how they handle negative zero (which can be
9121 // ignored in unsafe-math mode).
9122 if (Subtarget->hasSSE2() &&
9123 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9124 Cond.getOpcode() == ISD::SETCC) {
9125 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9127 unsigned Opcode = 0;
9128 // Check for x CC y ? x : y.
9129 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9130 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9134 // Converting this to a min would handle NaNs incorrectly, and swapping
9135 // the operands would cause it to handle comparisons between positive
9136 // and negative zero incorrectly.
9137 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9138 if (!UnsafeFPMath &&
9139 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9141 std::swap(LHS, RHS);
9143 Opcode = X86ISD::FMIN;
9146 // Converting this to a min would handle comparisons between positive
9147 // and negative zero incorrectly.
9148 if (!UnsafeFPMath &&
9149 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9151 Opcode = X86ISD::FMIN;
9154 // Converting this to a min would handle both negative zeros and NaNs
9155 // incorrectly, but we can swap the operands to fix both.
9156 std::swap(LHS, RHS);
9160 Opcode = X86ISD::FMIN;
9164 // Converting this to a max would handle comparisons between positive
9165 // and negative zero incorrectly.
9166 if (!UnsafeFPMath &&
9167 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9169 Opcode = X86ISD::FMAX;
9172 // Converting this to a max would handle NaNs incorrectly, and swapping
9173 // the operands would cause it to handle comparisons between positive
9174 // and negative zero incorrectly.
9175 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9176 if (!UnsafeFPMath &&
9177 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9179 std::swap(LHS, RHS);
9181 Opcode = X86ISD::FMAX;
9184 // Converting this to a max would handle both negative zeros and NaNs
9185 // incorrectly, but we can swap the operands to fix both.
9186 std::swap(LHS, RHS);
9190 Opcode = X86ISD::FMAX;
9193 // Check for x CC y ? y : x -- a min/max with reversed arms.
9194 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9195 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9199 // Converting this to a min would handle comparisons between positive
9200 // and negative zero incorrectly, and swapping the operands would
9201 // cause it to handle NaNs incorrectly.
9202 if (!UnsafeFPMath &&
9203 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9204 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9206 std::swap(LHS, RHS);
9208 Opcode = X86ISD::FMIN;
9211 // Converting this to a min would handle NaNs incorrectly.
9212 if (!UnsafeFPMath &&
9213 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9215 Opcode = X86ISD::FMIN;
9218 // Converting this to a min would handle both negative zeros and NaNs
9219 // incorrectly, but we can swap the operands to fix both.
9220 std::swap(LHS, RHS);
9224 Opcode = X86ISD::FMIN;
9228 // Converting this to a max would handle NaNs incorrectly.
9229 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9231 Opcode = X86ISD::FMAX;
9234 // Converting this to a max would handle comparisons between positive
9235 // and negative zero incorrectly, and swapping the operands would
9236 // cause it to handle NaNs incorrectly.
9237 if (!UnsafeFPMath &&
9238 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9239 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9241 std::swap(LHS, RHS);
9243 Opcode = X86ISD::FMAX;
9246 // Converting this to a max would handle both negative zeros and NaNs
9247 // incorrectly, but we can swap the operands to fix both.
9248 std::swap(LHS, RHS);
9252 Opcode = X86ISD::FMAX;
9258 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9261 // If this is a select between two integer constants, try to do some
9263 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9264 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9265 // Don't do this for crazy integer types.
9266 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9267 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9268 // so that TrueC (the true value) is larger than FalseC.
9269 bool NeedsCondInvert = false;
9271 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9272 // Efficiently invertible.
9273 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9274 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9275 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9276 NeedsCondInvert = true;
9277 std::swap(TrueC, FalseC);
9280 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9281 if (FalseC->getAPIntValue() == 0 &&
9282 TrueC->getAPIntValue().isPowerOf2()) {
9283 if (NeedsCondInvert) // Invert the condition if needed.
9284 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9285 DAG.getConstant(1, Cond.getValueType()));
9287 // Zero extend the condition if needed.
9288 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9290 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9291 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9292 DAG.getConstant(ShAmt, MVT::i8));
9295 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9296 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9297 if (NeedsCondInvert) // Invert the condition if needed.
9298 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9299 DAG.getConstant(1, Cond.getValueType()));
9301 // Zero extend the condition if needed.
9302 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9303 FalseC->getValueType(0), Cond);
9304 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9305 SDValue(FalseC, 0));
9308 // Optimize cases that will turn into an LEA instruction. This requires
9309 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9310 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9311 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9312 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9314 bool isFastMultiplier = false;
9316 switch ((unsigned char)Diff) {
9318 case 1: // result = add base, cond
9319 case 2: // result = lea base( , cond*2)
9320 case 3: // result = lea base(cond, cond*2)
9321 case 4: // result = lea base( , cond*4)
9322 case 5: // result = lea base(cond, cond*4)
9323 case 8: // result = lea base( , cond*8)
9324 case 9: // result = lea base(cond, cond*8)
9325 isFastMultiplier = true;
9330 if (isFastMultiplier) {
9331 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9332 if (NeedsCondInvert) // Invert the condition if needed.
9333 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9334 DAG.getConstant(1, Cond.getValueType()));
9336 // Zero extend the condition if needed.
9337 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9339 // Scale the condition by the difference.
9341 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9342 DAG.getConstant(Diff, Cond.getValueType()));
9344 // Add the base if non-zero.
9345 if (FalseC->getAPIntValue() != 0)
9346 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9347 SDValue(FalseC, 0));
9357 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9358 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9359 TargetLowering::DAGCombinerInfo &DCI) {
9360 DebugLoc DL = N->getDebugLoc();
9362 // If the flag operand isn't dead, don't touch this CMOV.
9363 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9366 // If this is a select between two integer constants, try to do some
9367 // optimizations. Note that the operands are ordered the opposite of SELECT
9369 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9370 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9371 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9372 // larger than FalseC (the false value).
9373 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9375 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9376 CC = X86::GetOppositeBranchCondition(CC);
9377 std::swap(TrueC, FalseC);
9380 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9381 // This is efficient for any integer data type (including i8/i16) and
9383 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9384 SDValue Cond = N->getOperand(3);
9385 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9386 DAG.getConstant(CC, MVT::i8), Cond);
9388 // Zero extend the condition if needed.
9389 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9391 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9392 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9393 DAG.getConstant(ShAmt, MVT::i8));
9394 if (N->getNumValues() == 2) // Dead flag value?
9395 return DCI.CombineTo(N, Cond, SDValue());
9399 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9400 // for any integer data type, including i8/i16.
9401 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9402 SDValue Cond = N->getOperand(3);
9403 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9404 DAG.getConstant(CC, MVT::i8), Cond);
9406 // Zero extend the condition if needed.
9407 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9408 FalseC->getValueType(0), Cond);
9409 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9410 SDValue(FalseC, 0));
9412 if (N->getNumValues() == 2) // Dead flag value?
9413 return DCI.CombineTo(N, Cond, SDValue());
9417 // Optimize cases that will turn into an LEA instruction. This requires
9418 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9419 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9420 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9421 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9423 bool isFastMultiplier = false;
9425 switch ((unsigned char)Diff) {
9427 case 1: // result = add base, cond
9428 case 2: // result = lea base( , cond*2)
9429 case 3: // result = lea base(cond, cond*2)
9430 case 4: // result = lea base( , cond*4)
9431 case 5: // result = lea base(cond, cond*4)
9432 case 8: // result = lea base( , cond*8)
9433 case 9: // result = lea base(cond, cond*8)
9434 isFastMultiplier = true;
9439 if (isFastMultiplier) {
9440 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9441 SDValue Cond = N->getOperand(3);
9442 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9443 DAG.getConstant(CC, MVT::i8), Cond);
9444 // Zero extend the condition if needed.
9445 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9447 // Scale the condition by the difference.
9449 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9450 DAG.getConstant(Diff, Cond.getValueType()));
9452 // Add the base if non-zero.
9453 if (FalseC->getAPIntValue() != 0)
9454 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9455 SDValue(FalseC, 0));
9456 if (N->getNumValues() == 2) // Dead flag value?
9457 return DCI.CombineTo(N, Cond, SDValue());
9467 /// PerformMulCombine - Optimize a single multiply with constant into two
9468 /// in order to implement it with two cheaper instructions, e.g.
9469 /// LEA + SHL, LEA + LEA.
9470 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9471 TargetLowering::DAGCombinerInfo &DCI) {
9472 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9475 EVT VT = N->getValueType(0);
9479 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9482 uint64_t MulAmt = C->getZExtValue();
9483 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9486 uint64_t MulAmt1 = 0;
9487 uint64_t MulAmt2 = 0;
9488 if ((MulAmt % 9) == 0) {
9490 MulAmt2 = MulAmt / 9;
9491 } else if ((MulAmt % 5) == 0) {
9493 MulAmt2 = MulAmt / 5;
9494 } else if ((MulAmt % 3) == 0) {
9496 MulAmt2 = MulAmt / 3;
9499 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9500 DebugLoc DL = N->getDebugLoc();
9502 if (isPowerOf2_64(MulAmt2) &&
9503 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9504 // If second multiplifer is pow2, issue it first. We want the multiply by
9505 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9507 std::swap(MulAmt1, MulAmt2);
9510 if (isPowerOf2_64(MulAmt1))
9511 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9512 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9514 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9515 DAG.getConstant(MulAmt1, VT));
9517 if (isPowerOf2_64(MulAmt2))
9518 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9519 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9521 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9522 DAG.getConstant(MulAmt2, VT));
9524 // Do not add new nodes to DAG combiner worklist.
9525 DCI.CombineTo(N, NewMul, false);
9530 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9531 SDValue N0 = N->getOperand(0);
9532 SDValue N1 = N->getOperand(1);
9533 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9534 EVT VT = N0.getValueType();
9536 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9537 // since the result of setcc_c is all zero's or all ones.
9538 if (N1C && N0.getOpcode() == ISD::AND &&
9539 N0.getOperand(1).getOpcode() == ISD::Constant) {
9540 SDValue N00 = N0.getOperand(0);
9541 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9542 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9543 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9544 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9545 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9546 APInt ShAmt = N1C->getAPIntValue();
9547 Mask = Mask.shl(ShAmt);
9549 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9550 N00, DAG.getConstant(Mask, VT));
9557 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9559 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9560 const X86Subtarget *Subtarget) {
9561 EVT VT = N->getValueType(0);
9562 if (!VT.isVector() && VT.isInteger() &&
9563 N->getOpcode() == ISD::SHL)
9564 return PerformSHLCombine(N, DAG);
9566 // On X86 with SSE2 support, we can transform this to a vector shift if
9567 // all elements are shifted by the same amount. We can't do this in legalize
9568 // because the a constant vector is typically transformed to a constant pool
9569 // so we have no knowledge of the shift amount.
9570 if (!Subtarget->hasSSE2())
9573 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9576 SDValue ShAmtOp = N->getOperand(1);
9577 EVT EltVT = VT.getVectorElementType();
9578 DebugLoc DL = N->getDebugLoc();
9579 SDValue BaseShAmt = SDValue();
9580 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9581 unsigned NumElts = VT.getVectorNumElements();
9583 for (; i != NumElts; ++i) {
9584 SDValue Arg = ShAmtOp.getOperand(i);
9585 if (Arg.getOpcode() == ISD::UNDEF) continue;
9589 for (; i != NumElts; ++i) {
9590 SDValue Arg = ShAmtOp.getOperand(i);
9591 if (Arg.getOpcode() == ISD::UNDEF) continue;
9592 if (Arg != BaseShAmt) {
9596 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9597 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9598 SDValue InVec = ShAmtOp.getOperand(0);
9599 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9600 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9602 for (; i != NumElts; ++i) {
9603 SDValue Arg = InVec.getOperand(i);
9604 if (Arg.getOpcode() == ISD::UNDEF) continue;
9608 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9609 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9610 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9611 if (C->getZExtValue() == SplatIdx)
9612 BaseShAmt = InVec.getOperand(1);
9615 if (BaseShAmt.getNode() == 0)
9616 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9617 DAG.getIntPtrConstant(0));
9621 // The shift amount is an i32.
9622 if (EltVT.bitsGT(MVT::i32))
9623 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9624 else if (EltVT.bitsLT(MVT::i32))
9625 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9627 // The shift amount is identical so we can do a vector shift.
9628 SDValue ValOp = N->getOperand(0);
9629 switch (N->getOpcode()) {
9631 llvm_unreachable("Unknown shift opcode!");
9634 if (VT == MVT::v2i64)
9635 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9636 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9638 if (VT == MVT::v4i32)
9639 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9640 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9642 if (VT == MVT::v8i16)
9643 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9644 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9648 if (VT == MVT::v4i32)
9649 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9650 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9652 if (VT == MVT::v8i16)
9653 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9654 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9658 if (VT == MVT::v2i64)
9659 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9660 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9662 if (VT == MVT::v4i32)
9663 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9664 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9666 if (VT == MVT::v8i16)
9667 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9668 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9675 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9676 TargetLowering::DAGCombinerInfo &DCI,
9677 const X86Subtarget *Subtarget) {
9678 if (DCI.isBeforeLegalizeOps())
9681 EVT VT = N->getValueType(0);
9682 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9685 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9686 SDValue N0 = N->getOperand(0);
9687 SDValue N1 = N->getOperand(1);
9688 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9690 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9692 if (!N0.hasOneUse() || !N1.hasOneUse())
9695 SDValue ShAmt0 = N0.getOperand(1);
9696 if (ShAmt0.getValueType() != MVT::i8)
9698 SDValue ShAmt1 = N1.getOperand(1);
9699 if (ShAmt1.getValueType() != MVT::i8)
9701 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9702 ShAmt0 = ShAmt0.getOperand(0);
9703 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9704 ShAmt1 = ShAmt1.getOperand(0);
9706 DebugLoc DL = N->getDebugLoc();
9707 unsigned Opc = X86ISD::SHLD;
9708 SDValue Op0 = N0.getOperand(0);
9709 SDValue Op1 = N1.getOperand(0);
9710 if (ShAmt0.getOpcode() == ISD::SUB) {
9712 std::swap(Op0, Op1);
9713 std::swap(ShAmt0, ShAmt1);
9716 unsigned Bits = VT.getSizeInBits();
9717 if (ShAmt1.getOpcode() == ISD::SUB) {
9718 SDValue Sum = ShAmt1.getOperand(0);
9719 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9720 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
9721 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
9722 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
9723 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
9724 return DAG.getNode(Opc, DL, VT,
9726 DAG.getNode(ISD::TRUNCATE, DL,
9729 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9730 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9732 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9733 return DAG.getNode(Opc, DL, VT,
9734 N0.getOperand(0), N1.getOperand(0),
9735 DAG.getNode(ISD::TRUNCATE, DL,
9742 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9743 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9744 const X86Subtarget *Subtarget) {
9745 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9746 // the FP state in cases where an emms may be missing.
9747 // A preferable solution to the general problem is to figure out the right
9748 // places to insert EMMS. This qualifies as a quick hack.
9750 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9751 StoreSDNode *St = cast<StoreSDNode>(N);
9752 EVT VT = St->getValue().getValueType();
9753 if (VT.getSizeInBits() != 64)
9756 const Function *F = DAG.getMachineFunction().getFunction();
9757 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9758 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9759 && Subtarget->hasSSE2();
9760 if ((VT.isVector() ||
9761 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9762 isa<LoadSDNode>(St->getValue()) &&
9763 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9764 St->getChain().hasOneUse() && !St->isVolatile()) {
9765 SDNode* LdVal = St->getValue().getNode();
9767 int TokenFactorIndex = -1;
9768 SmallVector<SDValue, 8> Ops;
9769 SDNode* ChainVal = St->getChain().getNode();
9770 // Must be a store of a load. We currently handle two cases: the load
9771 // is a direct child, and it's under an intervening TokenFactor. It is
9772 // possible to dig deeper under nested TokenFactors.
9773 if (ChainVal == LdVal)
9774 Ld = cast<LoadSDNode>(St->getChain());
9775 else if (St->getValue().hasOneUse() &&
9776 ChainVal->getOpcode() == ISD::TokenFactor) {
9777 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9778 if (ChainVal->getOperand(i).getNode() == LdVal) {
9779 TokenFactorIndex = i;
9780 Ld = cast<LoadSDNode>(St->getValue());
9782 Ops.push_back(ChainVal->getOperand(i));
9786 if (!Ld || !ISD::isNormalLoad(Ld))
9789 // If this is not the MMX case, i.e. we are just turning i64 load/store
9790 // into f64 load/store, avoid the transformation if there are multiple
9791 // uses of the loaded value.
9792 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9795 DebugLoc LdDL = Ld->getDebugLoc();
9796 DebugLoc StDL = N->getDebugLoc();
9797 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9798 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9800 if (Subtarget->is64Bit() || F64IsLegal) {
9801 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9802 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9803 Ld->getBasePtr(), Ld->getSrcValue(),
9804 Ld->getSrcValueOffset(), Ld->isVolatile(),
9805 Ld->isNonTemporal(), Ld->getAlignment());
9806 SDValue NewChain = NewLd.getValue(1);
9807 if (TokenFactorIndex != -1) {
9808 Ops.push_back(NewChain);
9809 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9812 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9813 St->getSrcValue(), St->getSrcValueOffset(),
9814 St->isVolatile(), St->isNonTemporal(),
9815 St->getAlignment());
9818 // Otherwise, lower to two pairs of 32-bit loads / stores.
9819 SDValue LoAddr = Ld->getBasePtr();
9820 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9821 DAG.getConstant(4, MVT::i32));
9823 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9824 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9825 Ld->isVolatile(), Ld->isNonTemporal(),
9826 Ld->getAlignment());
9827 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9828 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9829 Ld->isVolatile(), Ld->isNonTemporal(),
9830 MinAlign(Ld->getAlignment(), 4));
9832 SDValue NewChain = LoLd.getValue(1);
9833 if (TokenFactorIndex != -1) {
9834 Ops.push_back(LoLd);
9835 Ops.push_back(HiLd);
9836 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9840 LoAddr = St->getBasePtr();
9841 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9842 DAG.getConstant(4, MVT::i32));
9844 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9845 St->getSrcValue(), St->getSrcValueOffset(),
9846 St->isVolatile(), St->isNonTemporal(),
9847 St->getAlignment());
9848 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9850 St->getSrcValueOffset() + 4,
9852 St->isNonTemporal(),
9853 MinAlign(St->getAlignment(), 4));
9854 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9859 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9860 /// X86ISD::FXOR nodes.
9861 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9862 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9863 // F[X]OR(0.0, x) -> x
9864 // F[X]OR(x, 0.0) -> x
9865 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9866 if (C->getValueAPF().isPosZero())
9867 return N->getOperand(1);
9868 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9869 if (C->getValueAPF().isPosZero())
9870 return N->getOperand(0);
9874 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9875 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9876 // FAND(0.0, x) -> 0.0
9877 // FAND(x, 0.0) -> 0.0
9878 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9879 if (C->getValueAPF().isPosZero())
9880 return N->getOperand(0);
9881 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9882 if (C->getValueAPF().isPosZero())
9883 return N->getOperand(1);
9887 static SDValue PerformBTCombine(SDNode *N,
9889 TargetLowering::DAGCombinerInfo &DCI) {
9890 // BT ignores high bits in the bit index operand.
9891 SDValue Op1 = N->getOperand(1);
9892 if (Op1.hasOneUse()) {
9893 unsigned BitWidth = Op1.getValueSizeInBits();
9894 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9895 APInt KnownZero, KnownOne;
9896 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9897 !DCI.isBeforeLegalizeOps());
9898 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9899 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9900 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9901 DCI.CommitTargetLoweringOpt(TLO);
9906 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9907 SDValue Op = N->getOperand(0);
9908 if (Op.getOpcode() == ISD::BIT_CONVERT)
9909 Op = Op.getOperand(0);
9910 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9911 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9912 VT.getVectorElementType().getSizeInBits() ==
9913 OpVT.getVectorElementType().getSizeInBits()) {
9914 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9919 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9920 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9921 // (and (i32 x86isd::setcc_carry), 1)
9922 // This eliminates the zext. This transformation is necessary because
9923 // ISD::SETCC is always legalized to i8.
9924 DebugLoc dl = N->getDebugLoc();
9925 SDValue N0 = N->getOperand(0);
9926 EVT VT = N->getValueType(0);
9927 if (N0.getOpcode() == ISD::AND &&
9929 N0.getOperand(0).hasOneUse()) {
9930 SDValue N00 = N0.getOperand(0);
9931 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9933 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9934 if (!C || C->getZExtValue() != 1)
9936 return DAG.getNode(ISD::AND, dl, VT,
9937 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9938 N00.getOperand(0), N00.getOperand(1)),
9939 DAG.getConstant(1, VT));
9945 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9946 DAGCombinerInfo &DCI) const {
9947 SelectionDAG &DAG = DCI.DAG;
9948 switch (N->getOpcode()) {
9950 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9951 case ISD::EXTRACT_VECTOR_ELT:
9952 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9953 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9954 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9955 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9958 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9959 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
9960 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9962 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9963 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9964 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9965 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9966 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9972 /// isTypeDesirableForOp - Return true if the target has native support for
9973 /// the specified value type and it is 'desirable' to use the type for the
9974 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9975 /// instruction encodings are longer and some i16 instructions are slow.
9976 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9977 if (!isTypeLegal(VT))
9986 case ISD::SIGN_EXTEND:
9987 case ISD::ZERO_EXTEND:
9988 case ISD::ANY_EXTEND:
10001 static bool MayFoldLoad(SDValue Op) {
10002 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
10005 static bool MayFoldIntoStore(SDValue Op) {
10006 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
10009 /// IsDesirableToPromoteOp - This method query the target whether it is
10010 /// beneficial for dag combiner to promote the specified node. If true, it
10011 /// should return the desired promotion type by reference.
10012 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
10013 EVT VT = Op.getValueType();
10014 if (VT != MVT::i16)
10017 bool Promote = false;
10018 bool Commute = false;
10019 switch (Op.getOpcode()) {
10022 LoadSDNode *LD = cast<LoadSDNode>(Op);
10023 // If the non-extending load has a single use and it's not live out, then it
10024 // might be folded.
10025 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
10026 Op.hasOneUse()*/) {
10027 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
10028 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
10029 // The only case where we'd want to promote LOAD (rather then it being
10030 // promoted as an operand is when it's only use is liveout.
10031 if (UI->getOpcode() != ISD::CopyToReg)
10038 case ISD::SIGN_EXTEND:
10039 case ISD::ZERO_EXTEND:
10040 case ISD::ANY_EXTEND:
10045 SDValue N0 = Op.getOperand(0);
10046 // Look out for (store (shl (load), x)).
10047 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10060 SDValue N0 = Op.getOperand(0);
10061 SDValue N1 = Op.getOperand(1);
10062 if (!Commute && MayFoldLoad(N1))
10064 // Avoid disabling potential load folding opportunities.
10065 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10067 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10077 //===----------------------------------------------------------------------===//
10078 // X86 Inline Assembly Support
10079 //===----------------------------------------------------------------------===//
10081 static bool LowerToBSwap(CallInst *CI) {
10082 // FIXME: this should verify that we are targetting a 486 or better. If not,
10083 // we will turn this bswap into something that will be lowered to logical ops
10084 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10085 // so don't worry about this.
10087 // Verify this is a simple bswap.
10088 if (CI->getNumArgOperands() != 1 ||
10089 CI->getType() != CI->getArgOperand(0)->getType() ||
10090 !CI->getType()->isIntegerTy())
10093 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10094 if (!Ty || Ty->getBitWidth() % 16 != 0)
10097 // Okay, we can do this xform, do so now.
10098 const Type *Tys[] = { Ty };
10099 Module *M = CI->getParent()->getParent()->getParent();
10100 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10102 Value *Op = CI->getArgOperand(0);
10103 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10105 CI->replaceAllUsesWith(Op);
10106 CI->eraseFromParent();
10110 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10111 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10112 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10114 std::string AsmStr = IA->getAsmString();
10116 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10117 SmallVector<StringRef, 4> AsmPieces;
10118 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10120 switch (AsmPieces.size()) {
10121 default: return false;
10123 AsmStr = AsmPieces[0];
10125 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10128 if (AsmPieces.size() == 2 &&
10129 (AsmPieces[0] == "bswap" ||
10130 AsmPieces[0] == "bswapq" ||
10131 AsmPieces[0] == "bswapl") &&
10132 (AsmPieces[1] == "$0" ||
10133 AsmPieces[1] == "${0:q}")) {
10134 // No need to check constraints, nothing other than the equivalent of
10135 // "=r,0" would be valid here.
10136 return LowerToBSwap(CI);
10138 // rorw $$8, ${0:w} --> llvm.bswap.i16
10139 if (CI->getType()->isIntegerTy(16) &&
10140 AsmPieces.size() == 3 &&
10141 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10142 AsmPieces[1] == "$$8," &&
10143 AsmPieces[2] == "${0:w}" &&
10144 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10146 const std::string &Constraints = IA->getConstraintString();
10147 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10148 std::sort(AsmPieces.begin(), AsmPieces.end());
10149 if (AsmPieces.size() == 4 &&
10150 AsmPieces[0] == "~{cc}" &&
10151 AsmPieces[1] == "~{dirflag}" &&
10152 AsmPieces[2] == "~{flags}" &&
10153 AsmPieces[3] == "~{fpsr}") {
10154 return LowerToBSwap(CI);
10159 if (CI->getType()->isIntegerTy(64) &&
10160 Constraints.size() >= 2 &&
10161 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10162 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10163 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10164 SmallVector<StringRef, 4> Words;
10165 SplitString(AsmPieces[0], Words, " \t");
10166 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10168 SplitString(AsmPieces[1], Words, " \t");
10169 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10171 SplitString(AsmPieces[2], Words, " \t,");
10172 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10173 Words[2] == "%edx") {
10174 return LowerToBSwap(CI);
10186 /// getConstraintType - Given a constraint letter, return the type of
10187 /// constraint it is for this target.
10188 X86TargetLowering::ConstraintType
10189 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10190 if (Constraint.size() == 1) {
10191 switch (Constraint[0]) {
10203 return C_RegisterClass;
10211 return TargetLowering::getConstraintType(Constraint);
10214 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10215 /// with another that has more specific requirements based on the type of the
10216 /// corresponding operand.
10217 const char *X86TargetLowering::
10218 LowerXConstraint(EVT ConstraintVT) const {
10219 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10220 // 'f' like normal targets.
10221 if (ConstraintVT.isFloatingPoint()) {
10222 if (Subtarget->hasSSE2())
10224 if (Subtarget->hasSSE1())
10228 return TargetLowering::LowerXConstraint(ConstraintVT);
10231 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10232 /// vector. If it is invalid, don't add anything to Ops.
10233 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10235 std::vector<SDValue>&Ops,
10236 SelectionDAG &DAG) const {
10237 SDValue Result(0, 0);
10239 switch (Constraint) {
10242 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10243 if (C->getZExtValue() <= 31) {
10244 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10250 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10251 if (C->getZExtValue() <= 63) {
10252 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10258 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10259 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10260 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10266 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10267 if (C->getZExtValue() <= 255) {
10268 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10274 // 32-bit signed value
10275 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10276 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10277 C->getSExtValue())) {
10278 // Widen to 64 bits here to get it sign extended.
10279 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10282 // FIXME gcc accepts some relocatable values here too, but only in certain
10283 // memory models; it's complicated.
10288 // 32-bit unsigned value
10289 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10290 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10291 C->getZExtValue())) {
10292 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10296 // FIXME gcc accepts some relocatable values here too, but only in certain
10297 // memory models; it's complicated.
10301 // Literal immediates are always ok.
10302 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10303 // Widen to 64 bits here to get it sign extended.
10304 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10308 // In any sort of PIC mode addresses need to be computed at runtime by
10309 // adding in a register or some sort of table lookup. These can't
10310 // be used as immediates.
10311 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
10314 // If we are in non-pic codegen mode, we allow the address of a global (with
10315 // an optional displacement) to be used with 'i'.
10316 GlobalAddressSDNode *GA = 0;
10317 int64_t Offset = 0;
10319 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10321 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10322 Offset += GA->getOffset();
10324 } else if (Op.getOpcode() == ISD::ADD) {
10325 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10326 Offset += C->getZExtValue();
10327 Op = Op.getOperand(0);
10330 } else if (Op.getOpcode() == ISD::SUB) {
10331 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10332 Offset += -C->getZExtValue();
10333 Op = Op.getOperand(0);
10338 // Otherwise, this isn't something we can handle, reject it.
10342 const GlobalValue *GV = GA->getGlobal();
10343 // If we require an extra load to get this address, as in PIC mode, we
10344 // can't accept it.
10345 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10346 getTargetMachine())))
10349 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
10350 GA->getValueType(0), Offset);
10355 if (Result.getNode()) {
10356 Ops.push_back(Result);
10359 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10362 std::vector<unsigned> X86TargetLowering::
10363 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10365 if (Constraint.size() == 1) {
10366 // FIXME: not handling fp-stack yet!
10367 switch (Constraint[0]) { // GCC X86 Constraint Letters
10368 default: break; // Unknown constraint letter
10369 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10370 if (Subtarget->is64Bit()) {
10371 if (VT == MVT::i32)
10372 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10373 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10374 X86::R10D,X86::R11D,X86::R12D,
10375 X86::R13D,X86::R14D,X86::R15D,
10376 X86::EBP, X86::ESP, 0);
10377 else if (VT == MVT::i16)
10378 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10379 X86::SI, X86::DI, X86::R8W,X86::R9W,
10380 X86::R10W,X86::R11W,X86::R12W,
10381 X86::R13W,X86::R14W,X86::R15W,
10382 X86::BP, X86::SP, 0);
10383 else if (VT == MVT::i8)
10384 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10385 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10386 X86::R10B,X86::R11B,X86::R12B,
10387 X86::R13B,X86::R14B,X86::R15B,
10388 X86::BPL, X86::SPL, 0);
10390 else if (VT == MVT::i64)
10391 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10392 X86::RSI, X86::RDI, X86::R8, X86::R9,
10393 X86::R10, X86::R11, X86::R12,
10394 X86::R13, X86::R14, X86::R15,
10395 X86::RBP, X86::RSP, 0);
10399 // 32-bit fallthrough
10400 case 'Q': // Q_REGS
10401 if (VT == MVT::i32)
10402 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10403 else if (VT == MVT::i16)
10404 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10405 else if (VT == MVT::i8)
10406 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10407 else if (VT == MVT::i64)
10408 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10413 return std::vector<unsigned>();
10416 std::pair<unsigned, const TargetRegisterClass*>
10417 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10419 // First, see if this is a constraint that directly corresponds to an LLVM
10421 if (Constraint.size() == 1) {
10422 // GCC Constraint Letters
10423 switch (Constraint[0]) {
10425 case 'r': // GENERAL_REGS
10426 case 'l': // INDEX_REGS
10428 return std::make_pair(0U, X86::GR8RegisterClass);
10429 if (VT == MVT::i16)
10430 return std::make_pair(0U, X86::GR16RegisterClass);
10431 if (VT == MVT::i32 || !Subtarget->is64Bit())
10432 return std::make_pair(0U, X86::GR32RegisterClass);
10433 return std::make_pair(0U, X86::GR64RegisterClass);
10434 case 'R': // LEGACY_REGS
10436 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10437 if (VT == MVT::i16)
10438 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10439 if (VT == MVT::i32 || !Subtarget->is64Bit())
10440 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10441 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10442 case 'f': // FP Stack registers.
10443 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10444 // value to the correct fpstack register class.
10445 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10446 return std::make_pair(0U, X86::RFP32RegisterClass);
10447 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10448 return std::make_pair(0U, X86::RFP64RegisterClass);
10449 return std::make_pair(0U, X86::RFP80RegisterClass);
10450 case 'y': // MMX_REGS if MMX allowed.
10451 if (!Subtarget->hasMMX()) break;
10452 return std::make_pair(0U, X86::VR64RegisterClass);
10453 case 'Y': // SSE_REGS if SSE2 allowed
10454 if (!Subtarget->hasSSE2()) break;
10456 case 'x': // SSE_REGS if SSE1 allowed
10457 if (!Subtarget->hasSSE1()) break;
10459 switch (VT.getSimpleVT().SimpleTy) {
10461 // Scalar SSE types.
10464 return std::make_pair(0U, X86::FR32RegisterClass);
10467 return std::make_pair(0U, X86::FR64RegisterClass);
10475 return std::make_pair(0U, X86::VR128RegisterClass);
10481 // Use the default implementation in TargetLowering to convert the register
10482 // constraint into a member of a register class.
10483 std::pair<unsigned, const TargetRegisterClass*> Res;
10484 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10486 // Not found as a standard register?
10487 if (Res.second == 0) {
10488 // Map st(0) -> st(7) -> ST0
10489 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10490 tolower(Constraint[1]) == 's' &&
10491 tolower(Constraint[2]) == 't' &&
10492 Constraint[3] == '(' &&
10493 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10494 Constraint[5] == ')' &&
10495 Constraint[6] == '}') {
10497 Res.first = X86::ST0+Constraint[4]-'0';
10498 Res.second = X86::RFP80RegisterClass;
10502 // GCC allows "st(0)" to be called just plain "st".
10503 if (StringRef("{st}").equals_lower(Constraint)) {
10504 Res.first = X86::ST0;
10505 Res.second = X86::RFP80RegisterClass;
10510 if (StringRef("{flags}").equals_lower(Constraint)) {
10511 Res.first = X86::EFLAGS;
10512 Res.second = X86::CCRRegisterClass;
10516 // 'A' means EAX + EDX.
10517 if (Constraint == "A") {
10518 Res.first = X86::EAX;
10519 Res.second = X86::GR32_ADRegisterClass;
10525 // Otherwise, check to see if this is a register class of the wrong value
10526 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10527 // turn into {ax},{dx}.
10528 if (Res.second->hasType(VT))
10529 return Res; // Correct type already, nothing to do.
10531 // All of the single-register GCC register classes map their values onto
10532 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10533 // really want an 8-bit or 32-bit register, map to the appropriate register
10534 // class and return the appropriate register.
10535 if (Res.second == X86::GR16RegisterClass) {
10536 if (VT == MVT::i8) {
10537 unsigned DestReg = 0;
10538 switch (Res.first) {
10540 case X86::AX: DestReg = X86::AL; break;
10541 case X86::DX: DestReg = X86::DL; break;
10542 case X86::CX: DestReg = X86::CL; break;
10543 case X86::BX: DestReg = X86::BL; break;
10546 Res.first = DestReg;
10547 Res.second = X86::GR8RegisterClass;
10549 } else if (VT == MVT::i32) {
10550 unsigned DestReg = 0;
10551 switch (Res.first) {
10553 case X86::AX: DestReg = X86::EAX; break;
10554 case X86::DX: DestReg = X86::EDX; break;
10555 case X86::CX: DestReg = X86::ECX; break;
10556 case X86::BX: DestReg = X86::EBX; break;
10557 case X86::SI: DestReg = X86::ESI; break;
10558 case X86::DI: DestReg = X86::EDI; break;
10559 case X86::BP: DestReg = X86::EBP; break;
10560 case X86::SP: DestReg = X86::ESP; break;
10563 Res.first = DestReg;
10564 Res.second = X86::GR32RegisterClass;
10566 } else if (VT == MVT::i64) {
10567 unsigned DestReg = 0;
10568 switch (Res.first) {
10570 case X86::AX: DestReg = X86::RAX; break;
10571 case X86::DX: DestReg = X86::RDX; break;
10572 case X86::CX: DestReg = X86::RCX; break;
10573 case X86::BX: DestReg = X86::RBX; break;
10574 case X86::SI: DestReg = X86::RSI; break;
10575 case X86::DI: DestReg = X86::RDI; break;
10576 case X86::BP: DestReg = X86::RBP; break;
10577 case X86::SP: DestReg = X86::RSP; break;
10580 Res.first = DestReg;
10581 Res.second = X86::GR64RegisterClass;
10584 } else if (Res.second == X86::FR32RegisterClass ||
10585 Res.second == X86::FR64RegisterClass ||
10586 Res.second == X86::VR128RegisterClass) {
10587 // Handle references to XMM physical registers that got mapped into the
10588 // wrong class. This can happen with constraints like {xmm0} where the
10589 // target independent register mapper will just pick the first match it can
10590 // find, ignoring the required type.
10591 if (VT == MVT::f32)
10592 Res.second = X86::FR32RegisterClass;
10593 else if (VT == MVT::f64)
10594 Res.second = X86::FR64RegisterClass;
10595 else if (X86::VR128RegisterClass->hasType(VT))
10596 Res.second = X86::VR128RegisterClass;