1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalAlias.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/LLVMContext.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/Dwarf.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/raw_ostream.h"
53 using namespace dwarf;
55 STATISTIC(NumTailCalls, "Number of tail calls");
58 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
60 // Forward declarations.
61 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
64 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
66 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
68 if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
69 if (is64Bit) return new X8664_MachoTargetObjectFile();
70 return new TargetLoweringObjectFileMachO();
71 } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
72 if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
73 return new X8632_ELFTargetObjectFile(TM);
74 } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
75 return new TargetLoweringObjectFileCOFF();
77 llvm_unreachable("unknown subtarget type");
80 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
81 : TargetLowering(TM, createTLOF(TM)) {
82 Subtarget = &TM.getSubtarget<X86Subtarget>();
83 X86ScalarSSEf64 = Subtarget->hasSSE2();
84 X86ScalarSSEf32 = Subtarget->hasSSE1();
85 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
87 RegInfo = TM.getRegisterInfo();
90 // Set up the TargetLowering object.
92 // X86 is weird, it always uses i8 for shift amounts and setcc results.
93 setShiftAmountType(MVT::i8);
94 setBooleanContents(ZeroOrOneBooleanContent);
95 setSchedulingPreference(Sched::RegPressure);
96 setStackPointerRegisterToSaveRestore(X86StackPtr);
98 if (Subtarget->isTargetDarwin()) {
99 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
100 setUseUnderscoreSetJmp(false);
101 setUseUnderscoreLongJmp(false);
102 } else if (Subtarget->isTargetMingw()) {
103 // MS runtime is weird: it exports _setjmp, but longjmp!
104 setUseUnderscoreSetJmp(true);
105 setUseUnderscoreLongJmp(false);
107 setUseUnderscoreSetJmp(true);
108 setUseUnderscoreLongJmp(true);
111 // Set up the register classes.
112 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
113 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
114 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
115 if (Subtarget->is64Bit())
116 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
118 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
120 // We don't accept any truncstore of integer registers.
121 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
122 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
123 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
124 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
125 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
128 // SETOEQ and SETUNE require checking two conditions.
129 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
130 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
131 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
132 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
133 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
136 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
138 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
139 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
140 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
142 if (Subtarget->is64Bit()) {
143 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
144 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
145 } else if (!UseSoftFloat) {
146 // We have an algorithm for SSE2->double, and we turn this into a
147 // 64-bit FILD followed by conditional FADD for other targets.
148 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
149 // We have an algorithm for SSE2, and we turn this into a 64-bit
150 // FILD for other targets.
151 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
154 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
156 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
157 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
160 // SSE has no i16 to fp conversion, only i32
161 if (X86ScalarSSEf32) {
162 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
163 // f32 and f64 cases are Legal, f80 case is not
164 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
166 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
167 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
170 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
171 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
174 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
175 // are Legal, f80 is custom lowered.
176 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
179 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
181 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
182 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
184 if (X86ScalarSSEf32) {
185 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
186 // f32 and f64 cases are Legal, f80 case is not
187 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
189 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
190 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
193 // Handle FP_TO_UINT by promoting the destination to a larger signed
195 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
196 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
197 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
199 if (Subtarget->is64Bit()) {
200 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
201 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
202 } else if (!UseSoftFloat) {
203 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
204 // Expand FP_TO_UINT into a select.
205 // FIXME: We would like to use a Custom expander here eventually to do
206 // the optimal thing for SSE vs. the default expansion in the legalizer.
207 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
209 // With SSE3 we can use fisttpll to convert to a signed i64; without
210 // SSE, we're stuck with a fistpll.
211 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
214 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
215 if (!X86ScalarSSEf64) {
216 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
217 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
218 if (Subtarget->is64Bit()) {
219 setOperationAction(ISD::BIT_CONVERT , MVT::f64 , Expand);
220 // Without SSE, i64->f64 goes through memory; i64->MMX is Legal.
221 if (Subtarget->hasMMX() && !DisableMMX)
222 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Custom);
224 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Expand);
228 // Scalar integer divide and remainder are lowered to use operations that
229 // produce two results, to match the available instructions. This exposes
230 // the two-result form to trivial CSE, which is able to combine x/y and x%y
231 // into a single instruction.
233 // Scalar integer multiply-high is also lowered to use two-result
234 // operations, to match the available instructions. However, plain multiply
235 // (low) operations are left as Legal, as there are single-result
236 // instructions for this in x86. Using the two-result multiply instructions
237 // when both high and low results are needed must be arranged by dagcombine.
238 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
239 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
240 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
241 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
242 setOperationAction(ISD::SREM , MVT::i8 , Expand);
243 setOperationAction(ISD::UREM , MVT::i8 , Expand);
244 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
245 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
246 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
247 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
248 setOperationAction(ISD::SREM , MVT::i16 , Expand);
249 setOperationAction(ISD::UREM , MVT::i16 , Expand);
250 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
251 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
252 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
253 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
254 setOperationAction(ISD::SREM , MVT::i32 , Expand);
255 setOperationAction(ISD::UREM , MVT::i32 , Expand);
256 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
257 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
258 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
259 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
260 setOperationAction(ISD::SREM , MVT::i64 , Expand);
261 setOperationAction(ISD::UREM , MVT::i64 , Expand);
263 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
264 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
265 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
266 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
267 if (Subtarget->is64Bit())
268 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
269 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
270 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
272 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
273 setOperationAction(ISD::FREM , MVT::f32 , Expand);
274 setOperationAction(ISD::FREM , MVT::f64 , Expand);
275 setOperationAction(ISD::FREM , MVT::f80 , Expand);
276 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
278 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
279 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
280 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
281 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
282 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
283 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
287 if (Subtarget->is64Bit()) {
288 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
293 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
294 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
296 // These should be promoted to a larger select which is supported.
297 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
298 // X86 wants to expand cmov itself.
299 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
300 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
301 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
302 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
303 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
304 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
305 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
306 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
307 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
308 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
309 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
310 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
311 if (Subtarget->is64Bit()) {
312 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
313 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
315 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
318 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
319 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
320 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
321 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
322 if (Subtarget->is64Bit())
323 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
324 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
325 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
326 if (Subtarget->is64Bit()) {
327 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
328 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
329 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
330 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
331 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
333 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
334 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
335 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
336 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
337 if (Subtarget->is64Bit()) {
338 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
339 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
340 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
343 if (Subtarget->hasSSE1())
344 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
346 // We may not have a libcall for MEMBARRIER so we should lower this.
347 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
349 // On X86 and X86-64, atomic operations are lowered to locked instructions.
350 // Locked instructions, in turn, have implicit fence semantics (all memory
351 // operations are flushed before issuing the locked instruction, and they
352 // are not buffered), so we can fold away the common pattern of
353 // fence-atomic-fence.
354 setShouldFoldAtomicFences(true);
356 // Expand certain atomics
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
360 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
365 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
367 if (!Subtarget->is64Bit()) {
368 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
374 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
377 // FIXME - use subtarget debug flags
378 if (!Subtarget->isTargetDarwin() &&
379 !Subtarget->isTargetELF() &&
380 !Subtarget->isTargetCygMing()) {
381 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
384 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
385 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
386 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
387 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
388 if (Subtarget->is64Bit()) {
389 setExceptionPointerRegister(X86::RAX);
390 setExceptionSelectorRegister(X86::RDX);
392 setExceptionPointerRegister(X86::EAX);
393 setExceptionSelectorRegister(X86::EDX);
395 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
396 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
398 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
400 setOperationAction(ISD::TRAP, MVT::Other, Legal);
402 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
403 setOperationAction(ISD::VASTART , MVT::Other, Custom);
404 setOperationAction(ISD::VAEND , MVT::Other, Expand);
405 if (Subtarget->is64Bit()) {
406 setOperationAction(ISD::VAARG , MVT::Other, Custom);
407 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
409 setOperationAction(ISD::VAARG , MVT::Other, Expand);
410 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
413 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
414 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
415 if (Subtarget->is64Bit())
416 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
417 if (Subtarget->isTargetCygMing())
418 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
420 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
422 if (!UseSoftFloat && X86ScalarSSEf64) {
423 // f32 and f64 use SSE.
424 // Set up the FP register classes.
425 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
426 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
428 // Use ANDPD to simulate FABS.
429 setOperationAction(ISD::FABS , MVT::f64, Custom);
430 setOperationAction(ISD::FABS , MVT::f32, Custom);
432 // Use XORP to simulate FNEG.
433 setOperationAction(ISD::FNEG , MVT::f64, Custom);
434 setOperationAction(ISD::FNEG , MVT::f32, Custom);
436 // Use ANDPD and ORPD to simulate FCOPYSIGN.
437 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
438 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
440 // We don't support sin/cos/fmod
441 setOperationAction(ISD::FSIN , MVT::f64, Expand);
442 setOperationAction(ISD::FCOS , MVT::f64, Expand);
443 setOperationAction(ISD::FSIN , MVT::f32, Expand);
444 setOperationAction(ISD::FCOS , MVT::f32, Expand);
446 // Expand FP immediates into loads from the stack, except for the special
448 addLegalFPImmediate(APFloat(+0.0)); // xorpd
449 addLegalFPImmediate(APFloat(+0.0f)); // xorps
450 } else if (!UseSoftFloat && X86ScalarSSEf32) {
451 // Use SSE for f32, x87 for f64.
452 // Set up the FP register classes.
453 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
454 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
456 // Use ANDPS to simulate FABS.
457 setOperationAction(ISD::FABS , MVT::f32, Custom);
459 // Use XORP to simulate FNEG.
460 setOperationAction(ISD::FNEG , MVT::f32, Custom);
462 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
464 // Use ANDPS and ORPS to simulate FCOPYSIGN.
465 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
466 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
468 // We don't support sin/cos/fmod
469 setOperationAction(ISD::FSIN , MVT::f32, Expand);
470 setOperationAction(ISD::FCOS , MVT::f32, Expand);
472 // Special cases we handle for FP constants.
473 addLegalFPImmediate(APFloat(+0.0f)); // xorps
474 addLegalFPImmediate(APFloat(+0.0)); // FLD0
475 addLegalFPImmediate(APFloat(+1.0)); // FLD1
476 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
477 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
480 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
481 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
483 } else if (!UseSoftFloat) {
484 // f32 and f64 in x87.
485 // Set up the FP register classes.
486 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
487 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
489 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
490 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
491 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
492 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
495 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
496 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
498 addLegalFPImmediate(APFloat(+0.0)); // FLD0
499 addLegalFPImmediate(APFloat(+1.0)); // FLD1
500 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
501 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
502 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
503 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
504 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
505 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
508 // Long double always uses X87.
510 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
511 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
512 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
515 APFloat TmpFlt(+0.0);
516 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
518 addLegalFPImmediate(TmpFlt); // FLD0
520 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
521 APFloat TmpFlt2(+1.0);
522 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
524 addLegalFPImmediate(TmpFlt2); // FLD1
525 TmpFlt2.changeSign();
526 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
530 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
531 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
535 // Always use a library call for pow.
536 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
537 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
538 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
540 setOperationAction(ISD::FLOG, MVT::f80, Expand);
541 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
542 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
543 setOperationAction(ISD::FEXP, MVT::f80, Expand);
544 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
546 // First set operation action for all vector types to either promote
547 // (for widening) or expand (for scalarization). Then we will selectively
548 // turn on ones that can be effectively codegen'd.
549 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
550 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
551 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
566 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
567 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
600 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
604 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
605 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
606 setTruncStoreAction((MVT::SimpleValueType)VT,
607 (MVT::SimpleValueType)InnerVT, Expand);
608 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
609 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
610 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
613 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
614 // with -msoft-float, disable use of MMX as well.
615 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
616 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
617 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
618 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
620 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
622 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
623 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
624 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
625 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
627 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
628 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
629 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
630 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
632 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
633 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
635 setOperationAction(ISD::AND, MVT::v8i8, Promote);
636 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
637 setOperationAction(ISD::AND, MVT::v4i16, Promote);
638 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
639 setOperationAction(ISD::AND, MVT::v2i32, Promote);
640 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
641 setOperationAction(ISD::AND, MVT::v1i64, Legal);
643 setOperationAction(ISD::OR, MVT::v8i8, Promote);
644 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
645 setOperationAction(ISD::OR, MVT::v4i16, Promote);
646 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
647 setOperationAction(ISD::OR, MVT::v2i32, Promote);
648 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
649 setOperationAction(ISD::OR, MVT::v1i64, Legal);
651 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
652 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
653 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
654 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
655 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
656 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
657 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
659 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
660 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
661 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
662 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
663 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
664 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
665 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
667 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
669 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
670 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
672 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
674 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
675 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
677 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
678 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
679 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
681 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
683 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
684 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
685 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
686 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
689 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
691 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
692 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
693 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
694 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
695 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
699 if (!UseSoftFloat && Subtarget->hasSSE1()) {
700 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
702 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
703 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
704 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
705 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
706 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
707 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
708 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
709 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
710 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
711 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
712 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
713 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
716 if (!UseSoftFloat && Subtarget->hasSSE2()) {
717 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
719 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
720 // registers cannot be used even for integer operations.
721 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
722 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
723 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
724 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
726 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
727 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
728 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
729 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
730 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
731 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
732 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
733 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
734 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
735 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
736 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
737 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
738 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
739 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
740 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
741 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
743 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
744 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
745 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
746 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
748 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
749 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
750 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
751 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
752 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
754 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
755 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
756 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
757 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
760 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
761 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
762 EVT VT = (MVT::SimpleValueType)i;
763 // Do not attempt to custom lower non-power-of-2 vectors
764 if (!isPowerOf2_32(VT.getVectorNumElements()))
766 // Do not attempt to custom lower non-128-bit vectors
767 if (!VT.is128BitVector())
769 setOperationAction(ISD::BUILD_VECTOR,
770 VT.getSimpleVT().SimpleTy, Custom);
771 setOperationAction(ISD::VECTOR_SHUFFLE,
772 VT.getSimpleVT().SimpleTy, Custom);
773 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
774 VT.getSimpleVT().SimpleTy, Custom);
777 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
778 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
779 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
780 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
781 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
782 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
784 if (Subtarget->is64Bit()) {
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
789 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
790 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
791 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
794 // Do not attempt to promote non-128-bit vectors
795 if (!VT.is128BitVector())
798 setOperationAction(ISD::AND, SVT, Promote);
799 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
800 setOperationAction(ISD::OR, SVT, Promote);
801 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
802 setOperationAction(ISD::XOR, SVT, Promote);
803 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
804 setOperationAction(ISD::LOAD, SVT, Promote);
805 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
806 setOperationAction(ISD::SELECT, SVT, Promote);
807 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
810 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
812 // Custom lower v2i64 and v2f64 selects.
813 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
814 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
815 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
816 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
818 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
819 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
820 if (!DisableMMX && Subtarget->hasMMX()) {
821 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
822 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
826 if (Subtarget->hasSSE41()) {
827 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
828 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
829 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
830 setOperationAction(ISD::FRINT, MVT::f32, Legal);
831 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
832 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
833 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
834 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
835 setOperationAction(ISD::FRINT, MVT::f64, Legal);
836 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
838 // FIXME: Do we need to handle scalar-to-vector here?
839 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
841 // Can turn SHL into an integer multiply.
842 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
843 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
845 // i8 and i16 vectors are custom , because the source register and source
846 // source memory operand types are not the same width. f32 vectors are
847 // custom since the immediate controlling the insert encodes additional
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
850 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
851 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
852 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
854 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
855 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
856 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
857 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
859 if (Subtarget->is64Bit()) {
860 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
861 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
865 if (Subtarget->hasSSE42()) {
866 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
869 if (!UseSoftFloat && Subtarget->hasAVX()) {
870 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
871 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
872 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
873 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
874 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
876 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
877 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
878 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
879 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
880 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
881 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
882 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
883 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
884 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
885 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
886 setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
887 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
888 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
889 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
890 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
892 // Operations to consider commented out -v16i16 v32i8
893 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
894 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
895 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
896 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
897 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
898 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
899 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
900 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
901 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
902 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
903 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
904 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
905 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
906 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
908 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
909 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
910 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
911 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
913 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
914 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
915 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
916 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
917 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
919 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
920 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
921 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
922 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
923 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
924 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
927 // Not sure we want to do this since there are no 256-bit integer
930 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
931 // This includes 256-bit vectors
932 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
933 EVT VT = (MVT::SimpleValueType)i;
935 // Do not attempt to custom lower non-power-of-2 vectors
936 if (!isPowerOf2_32(VT.getVectorNumElements()))
939 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
940 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
941 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
944 if (Subtarget->is64Bit()) {
945 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
946 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
951 // Not sure we want to do this since there are no 256-bit integer
954 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
955 // Including 256-bit vectors
956 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
957 EVT VT = (MVT::SimpleValueType)i;
959 if (!VT.is256BitVector()) {
962 setOperationAction(ISD::AND, VT, Promote);
963 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
964 setOperationAction(ISD::OR, VT, Promote);
965 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
966 setOperationAction(ISD::XOR, VT, Promote);
967 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
968 setOperationAction(ISD::LOAD, VT, Promote);
969 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
970 setOperationAction(ISD::SELECT, VT, Promote);
971 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
974 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
978 // We want to custom lower some of our intrinsics.
979 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
981 // Add/Sub/Mul with overflow operations are custom lowered.
982 setOperationAction(ISD::SADDO, MVT::i32, Custom);
983 setOperationAction(ISD::UADDO, MVT::i32, Custom);
984 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
985 setOperationAction(ISD::USUBO, MVT::i32, Custom);
986 setOperationAction(ISD::SMULO, MVT::i32, Custom);
988 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
989 // handle type legalization for these operations here.
991 // FIXME: We really should do custom legalization for addition and
992 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
993 // than generic legalization for 64-bit multiplication-with-overflow, though.
994 if (Subtarget->is64Bit()) {
995 setOperationAction(ISD::SADDO, MVT::i64, Custom);
996 setOperationAction(ISD::UADDO, MVT::i64, Custom);
997 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
998 setOperationAction(ISD::USUBO, MVT::i64, Custom);
999 setOperationAction(ISD::SMULO, MVT::i64, Custom);
1002 if (!Subtarget->is64Bit()) {
1003 // These libcalls are not available in 32-bit.
1004 setLibcallName(RTLIB::SHL_I128, 0);
1005 setLibcallName(RTLIB::SRL_I128, 0);
1006 setLibcallName(RTLIB::SRA_I128, 0);
1009 // We have target-specific dag combine patterns for the following nodes:
1010 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1011 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1012 setTargetDAGCombine(ISD::BUILD_VECTOR);
1013 setTargetDAGCombine(ISD::SELECT);
1014 setTargetDAGCombine(ISD::SHL);
1015 setTargetDAGCombine(ISD::SRA);
1016 setTargetDAGCombine(ISD::SRL);
1017 setTargetDAGCombine(ISD::OR);
1018 setTargetDAGCombine(ISD::STORE);
1019 setTargetDAGCombine(ISD::ZERO_EXTEND);
1020 if (Subtarget->is64Bit())
1021 setTargetDAGCombine(ISD::MUL);
1023 computeRegisterProperties();
1025 // FIXME: These should be based on subtarget info. Plus, the values should
1026 // be smaller when we are in optimizing for size mode.
1027 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1028 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1029 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1030 setPrefLoopAlignment(16);
1031 benefitFromCodePlacementOpt = true;
1035 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1040 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1041 /// the desired ByVal argument alignment.
1042 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1045 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1046 if (VTy->getBitWidth() == 128)
1048 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1049 unsigned EltAlign = 0;
1050 getMaxByValAlign(ATy->getElementType(), EltAlign);
1051 if (EltAlign > MaxAlign)
1052 MaxAlign = EltAlign;
1053 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1054 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1055 unsigned EltAlign = 0;
1056 getMaxByValAlign(STy->getElementType(i), EltAlign);
1057 if (EltAlign > MaxAlign)
1058 MaxAlign = EltAlign;
1066 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1067 /// function arguments in the caller parameter area. For X86, aggregates
1068 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1069 /// are at 4-byte boundaries.
1070 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1071 if (Subtarget->is64Bit()) {
1072 // Max of 8 and alignment of type.
1073 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1080 if (Subtarget->hasSSE1())
1081 getMaxByValAlign(Ty, Align);
1085 /// getOptimalMemOpType - Returns the target specific optimal type for load
1086 /// and store operations as a result of memset, memcpy, and memmove
1087 /// lowering. If DstAlign is zero that means it's safe to destination
1088 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1089 /// means there isn't a need to check it against alignment requirement,
1090 /// probably because the source does not need to be loaded. If
1091 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1092 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1093 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1094 /// constant so it does not need to be loaded.
1095 /// It returns EVT::Other if the type should be determined using generic
1096 /// target-independent logic.
1098 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1099 unsigned DstAlign, unsigned SrcAlign,
1100 bool NonScalarIntSafe,
1102 MachineFunction &MF) const {
1103 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1104 // linux. This is because the stack realignment code can't handle certain
1105 // cases like PR2962. This should be removed when PR2962 is fixed.
1106 const Function *F = MF.getFunction();
1107 if (NonScalarIntSafe &&
1108 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1110 (Subtarget->isUnalignedMemAccessFast() ||
1111 ((DstAlign == 0 || DstAlign >= 16) &&
1112 (SrcAlign == 0 || SrcAlign >= 16))) &&
1113 Subtarget->getStackAlignment() >= 16) {
1114 if (Subtarget->hasSSE2())
1116 if (Subtarget->hasSSE1())
1118 } else if (!MemcpyStrSrc && Size >= 8 &&
1119 !Subtarget->is64Bit() &&
1120 Subtarget->getStackAlignment() >= 8 &&
1121 Subtarget->hasSSE2()) {
1122 // Do not use f64 to lower memcpy if source is string constant. It's
1123 // better to use i32 to avoid the loads.
1127 if (Subtarget->is64Bit() && Size >= 8)
1132 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1133 /// current function. The returned value is a member of the
1134 /// MachineJumpTableInfo::JTEntryKind enum.
1135 unsigned X86TargetLowering::getJumpTableEncoding() const {
1136 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1138 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1139 Subtarget->isPICStyleGOT())
1140 return MachineJumpTableInfo::EK_Custom32;
1142 // Otherwise, use the normal jump table encoding heuristics.
1143 return TargetLowering::getJumpTableEncoding();
1146 /// getPICBaseSymbol - Return the X86-32 PIC base.
1148 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1149 MCContext &Ctx) const {
1150 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1151 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1152 Twine(MF->getFunctionNumber())+"$pb");
1157 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1158 const MachineBasicBlock *MBB,
1159 unsigned uid,MCContext &Ctx) const{
1160 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1161 Subtarget->isPICStyleGOT());
1162 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1164 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1165 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1168 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1170 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1171 SelectionDAG &DAG) const {
1172 if (!Subtarget->is64Bit())
1173 // This doesn't have DebugLoc associated with it, but is not really the
1174 // same as a Register.
1175 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1179 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1180 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1182 const MCExpr *X86TargetLowering::
1183 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1184 MCContext &Ctx) const {
1185 // X86-64 uses RIP relative addressing based on the jump table label.
1186 if (Subtarget->isPICStyleRIPRel())
1187 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1189 // Otherwise, the reference is relative to the PIC base.
1190 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1193 /// getFunctionAlignment - Return the Log2 alignment of this function.
1194 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1195 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1198 std::pair<const TargetRegisterClass*, uint8_t>
1199 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1200 const TargetRegisterClass *RRC = 0;
1202 switch (VT.getSimpleVT().SimpleTy) {
1204 return TargetLowering::findRepresentativeClass(VT);
1205 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1206 RRC = (Subtarget->is64Bit()
1207 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1209 case MVT::v8i8: case MVT::v4i16:
1210 case MVT::v2i32: case MVT::v1i64:
1211 RRC = X86::VR64RegisterClass;
1213 case MVT::f32: case MVT::f64:
1214 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1215 case MVT::v4f32: case MVT::v2f64:
1216 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1218 RRC = X86::VR128RegisterClass;
1221 return std::make_pair(RRC, Cost);
1225 X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
1226 MachineFunction &MF) const {
1227 unsigned FPDiff = RegInfo->hasFP(MF) ? 1 : 0;
1228 switch (RC->getID()) {
1231 case X86::GR32RegClassID:
1233 case X86::GR64RegClassID:
1235 case X86::VR128RegClassID:
1236 return Subtarget->is64Bit() ? 10 : 4;
1237 case X86::VR64RegClassID:
1242 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1243 unsigned &Offset) const {
1244 if (!Subtarget->isTargetLinux())
1247 if (Subtarget->is64Bit()) {
1248 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1250 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1263 //===----------------------------------------------------------------------===//
1264 // Return Value Calling Convention Implementation
1265 //===----------------------------------------------------------------------===//
1267 #include "X86GenCallingConv.inc"
1270 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1271 const SmallVectorImpl<ISD::OutputArg> &Outs,
1272 LLVMContext &Context) const {
1273 SmallVector<CCValAssign, 16> RVLocs;
1274 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1276 return CCInfo.CheckReturn(Outs, RetCC_X86);
1280 X86TargetLowering::LowerReturn(SDValue Chain,
1281 CallingConv::ID CallConv, bool isVarArg,
1282 const SmallVectorImpl<ISD::OutputArg> &Outs,
1283 const SmallVectorImpl<SDValue> &OutVals,
1284 DebugLoc dl, SelectionDAG &DAG) const {
1285 MachineFunction &MF = DAG.getMachineFunction();
1286 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1288 SmallVector<CCValAssign, 16> RVLocs;
1289 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1290 RVLocs, *DAG.getContext());
1291 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1293 // Add the regs to the liveout set for the function.
1294 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1295 for (unsigned i = 0; i != RVLocs.size(); ++i)
1296 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1297 MRI.addLiveOut(RVLocs[i].getLocReg());
1301 SmallVector<SDValue, 6> RetOps;
1302 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1303 // Operand #1 = Bytes To Pop
1304 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1307 // Copy the result values into the output registers.
1308 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1309 CCValAssign &VA = RVLocs[i];
1310 assert(VA.isRegLoc() && "Can only return in registers!");
1311 SDValue ValToCopy = OutVals[i];
1312 EVT ValVT = ValToCopy.getValueType();
1314 // If this is x86-64, and we disabled SSE, we can't return FP values
1315 if ((ValVT == MVT::f32 || ValVT == MVT::f64) &&
1316 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1317 report_fatal_error("SSE register return with SSE disabled");
1319 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1320 // llvm-gcc has never done it right and no one has noticed, so this
1321 // should be OK for now.
1322 if (ValVT == MVT::f64 &&
1323 (Subtarget->is64Bit() && !Subtarget->hasSSE2())) {
1324 report_fatal_error("SSE2 register return with SSE2 disabled");
1327 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1328 // the RET instruction and handled by the FP Stackifier.
1329 if (VA.getLocReg() == X86::ST0 ||
1330 VA.getLocReg() == X86::ST1) {
1331 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1332 // change the value to the FP stack register class.
1333 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1334 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1335 RetOps.push_back(ValToCopy);
1336 // Don't emit a copytoreg.
1340 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1341 // which is returned in RAX / RDX.
1342 if (Subtarget->is64Bit()) {
1343 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1344 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1345 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1346 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1349 // If we don't have SSE2 available, convert to v4f32 so the generated
1350 // register is legal.
1351 if (!Subtarget->hasSSE2())
1352 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32,ValToCopy);
1357 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1358 Flag = Chain.getValue(1);
1361 // The x86-64 ABI for returning structs by value requires that we copy
1362 // the sret argument into %rax for the return. We saved the argument into
1363 // a virtual register in the entry block, so now we copy the value out
1365 if (Subtarget->is64Bit() &&
1366 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1367 MachineFunction &MF = DAG.getMachineFunction();
1368 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1369 unsigned Reg = FuncInfo->getSRetReturnReg();
1371 "SRetReturnReg should have been set in LowerFormalArguments().");
1372 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1374 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1375 Flag = Chain.getValue(1);
1377 // RAX now acts like a return value.
1378 MRI.addLiveOut(X86::RAX);
1381 RetOps[0] = Chain; // Update chain.
1383 // Add the flag if we have it.
1385 RetOps.push_back(Flag);
1387 return DAG.getNode(X86ISD::RET_FLAG, dl,
1388 MVT::Other, &RetOps[0], RetOps.size());
1391 /// LowerCallResult - Lower the result values of a call into the
1392 /// appropriate copies out of appropriate physical registers.
1395 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1396 CallingConv::ID CallConv, bool isVarArg,
1397 const SmallVectorImpl<ISD::InputArg> &Ins,
1398 DebugLoc dl, SelectionDAG &DAG,
1399 SmallVectorImpl<SDValue> &InVals) const {
1401 // Assign locations to each value returned by this call.
1402 SmallVector<CCValAssign, 16> RVLocs;
1403 bool Is64Bit = Subtarget->is64Bit();
1404 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1405 RVLocs, *DAG.getContext());
1406 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1408 // Copy all of the result registers out of their specified physreg.
1409 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1410 CCValAssign &VA = RVLocs[i];
1411 EVT CopyVT = VA.getValVT();
1413 // If this is x86-64, and we disabled SSE, we can't return FP values
1414 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1415 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1416 report_fatal_error("SSE register return with SSE disabled");
1421 // If this is a call to a function that returns an fp value on the floating
1422 // point stack, we must guarantee the the value is popped from the stack, so
1423 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1424 // if the return value is not used. We use the FpGET_ST0 instructions
1426 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1427 // If we prefer to use the value in xmm registers, copy it out as f80 and
1428 // use a truncate to move it from fp stack reg to xmm reg.
1429 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1430 bool isST0 = VA.getLocReg() == X86::ST0;
1432 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1433 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1434 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1435 SDValue Ops[] = { Chain, InFlag };
1436 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
1438 Val = Chain.getValue(0);
1440 // Round the f80 to the right size, which also moves it to the appropriate
1442 if (CopyVT != VA.getValVT())
1443 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1444 // This truncation won't change the value.
1445 DAG.getIntPtrConstant(1));
1446 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1447 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1448 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1449 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1450 MVT::v2i64, InFlag).getValue(1);
1451 Val = Chain.getValue(0);
1452 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1453 Val, DAG.getConstant(0, MVT::i64));
1455 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1456 MVT::i64, InFlag).getValue(1);
1457 Val = Chain.getValue(0);
1459 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1461 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1462 CopyVT, InFlag).getValue(1);
1463 Val = Chain.getValue(0);
1465 InFlag = Chain.getValue(2);
1466 InVals.push_back(Val);
1473 //===----------------------------------------------------------------------===//
1474 // C & StdCall & Fast Calling Convention implementation
1475 //===----------------------------------------------------------------------===//
1476 // StdCall calling convention seems to be standard for many Windows' API
1477 // routines and around. It differs from C calling convention just a little:
1478 // callee should clean up the stack, not caller. Symbols should be also
1479 // decorated in some fancy way :) It doesn't support any vector arguments.
1480 // For info on fast calling convention see Fast Calling Convention (tail call)
1481 // implementation LowerX86_32FastCCCallTo.
1483 /// CallIsStructReturn - Determines whether a call uses struct return
1485 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1489 return Outs[0].Flags.isSRet();
1492 /// ArgsAreStructReturn - Determines whether a function uses struct
1493 /// return semantics.
1495 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1499 return Ins[0].Flags.isSRet();
1502 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1503 /// given CallingConvention value.
1504 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1505 if (Subtarget->is64Bit()) {
1506 if (CC == CallingConv::GHC)
1507 return CC_X86_64_GHC;
1508 else if (Subtarget->isTargetWin64())
1509 return CC_X86_Win64_C;
1514 if (CC == CallingConv::X86_FastCall)
1515 return CC_X86_32_FastCall;
1516 else if (CC == CallingConv::X86_ThisCall)
1517 return CC_X86_32_ThisCall;
1518 else if (CC == CallingConv::Fast)
1519 return CC_X86_32_FastCC;
1520 else if (CC == CallingConv::GHC)
1521 return CC_X86_32_GHC;
1526 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1527 /// by "Src" to address "Dst" with size and alignment information specified by
1528 /// the specific parameter attribute. The copy will be passed as a byval
1529 /// function parameter.
1531 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1532 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1534 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1535 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1536 /*isVolatile*/false, /*AlwaysInline=*/true,
1540 /// IsTailCallConvention - Return true if the calling convention is one that
1541 /// supports tail call optimization.
1542 static bool IsTailCallConvention(CallingConv::ID CC) {
1543 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1546 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1547 /// a tailcall target by changing its ABI.
1548 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1549 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1553 X86TargetLowering::LowerMemArgument(SDValue Chain,
1554 CallingConv::ID CallConv,
1555 const SmallVectorImpl<ISD::InputArg> &Ins,
1556 DebugLoc dl, SelectionDAG &DAG,
1557 const CCValAssign &VA,
1558 MachineFrameInfo *MFI,
1560 // Create the nodes corresponding to a load from this parameter slot.
1561 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1562 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1563 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1566 // If value is passed by pointer we have address passed instead of the value
1568 if (VA.getLocInfo() == CCValAssign::Indirect)
1569 ValVT = VA.getLocVT();
1571 ValVT = VA.getValVT();
1573 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1574 // changed with more analysis.
1575 // In case of tail call optimization mark all arguments mutable. Since they
1576 // could be overwritten by lowering of arguments in case of a tail call.
1577 if (Flags.isByVal()) {
1578 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1579 VA.getLocMemOffset(), isImmutable);
1580 return DAG.getFrameIndex(FI, getPointerTy());
1582 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1583 VA.getLocMemOffset(), isImmutable);
1584 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1585 return DAG.getLoad(ValVT, dl, Chain, FIN,
1586 PseudoSourceValue::getFixedStack(FI), 0,
1592 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1593 CallingConv::ID CallConv,
1595 const SmallVectorImpl<ISD::InputArg> &Ins,
1598 SmallVectorImpl<SDValue> &InVals)
1600 MachineFunction &MF = DAG.getMachineFunction();
1601 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1603 const Function* Fn = MF.getFunction();
1604 if (Fn->hasExternalLinkage() &&
1605 Subtarget->isTargetCygMing() &&
1606 Fn->getName() == "main")
1607 FuncInfo->setForceFramePointer(true);
1609 MachineFrameInfo *MFI = MF.getFrameInfo();
1610 bool Is64Bit = Subtarget->is64Bit();
1611 bool IsWin64 = Subtarget->isTargetWin64();
1613 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1614 "Var args not supported with calling convention fastcc or ghc");
1616 // Assign locations to all of the incoming arguments.
1617 SmallVector<CCValAssign, 16> ArgLocs;
1618 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1619 ArgLocs, *DAG.getContext());
1620 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1622 unsigned LastVal = ~0U;
1624 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1625 CCValAssign &VA = ArgLocs[i];
1626 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1628 assert(VA.getValNo() != LastVal &&
1629 "Don't support value assigned to multiple locs yet");
1630 LastVal = VA.getValNo();
1632 if (VA.isRegLoc()) {
1633 EVT RegVT = VA.getLocVT();
1634 TargetRegisterClass *RC = NULL;
1635 if (RegVT == MVT::i32)
1636 RC = X86::GR32RegisterClass;
1637 else if (Is64Bit && RegVT == MVT::i64)
1638 RC = X86::GR64RegisterClass;
1639 else if (RegVT == MVT::f32)
1640 RC = X86::FR32RegisterClass;
1641 else if (RegVT == MVT::f64)
1642 RC = X86::FR64RegisterClass;
1643 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1644 RC = X86::VR256RegisterClass;
1645 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1646 RC = X86::VR128RegisterClass;
1647 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1648 RC = X86::VR64RegisterClass;
1650 llvm_unreachable("Unknown argument type!");
1652 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1653 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1655 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1656 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1658 if (VA.getLocInfo() == CCValAssign::SExt)
1659 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1660 DAG.getValueType(VA.getValVT()));
1661 else if (VA.getLocInfo() == CCValAssign::ZExt)
1662 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1663 DAG.getValueType(VA.getValVT()));
1664 else if (VA.getLocInfo() == CCValAssign::BCvt)
1665 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1667 if (VA.isExtInLoc()) {
1668 // Handle MMX values passed in XMM regs.
1669 if (RegVT.isVector()) {
1670 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1671 ArgValue, DAG.getConstant(0, MVT::i64));
1672 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1674 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1677 assert(VA.isMemLoc());
1678 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1681 // If value is passed via pointer - do a load.
1682 if (VA.getLocInfo() == CCValAssign::Indirect)
1683 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1686 InVals.push_back(ArgValue);
1689 // The x86-64 ABI for returning structs by value requires that we copy
1690 // the sret argument into %rax for the return. Save the argument into
1691 // a virtual register so that we can access it from the return points.
1692 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1693 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1694 unsigned Reg = FuncInfo->getSRetReturnReg();
1696 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1697 FuncInfo->setSRetReturnReg(Reg);
1699 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1700 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1703 unsigned StackSize = CCInfo.getNextStackOffset();
1704 // Align stack specially for tail calls.
1705 if (FuncIsMadeTailCallSafe(CallConv))
1706 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1708 // If the function takes variable number of arguments, make a frame index for
1709 // the start of the first vararg value... for expansion of llvm.va_start.
1711 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1712 CallConv != CallingConv::X86_ThisCall)) {
1713 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1716 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1718 // FIXME: We should really autogenerate these arrays
1719 static const unsigned GPR64ArgRegsWin64[] = {
1720 X86::RCX, X86::RDX, X86::R8, X86::R9
1722 static const unsigned XMMArgRegsWin64[] = {
1723 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1725 static const unsigned GPR64ArgRegs64Bit[] = {
1726 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1728 static const unsigned XMMArgRegs64Bit[] = {
1729 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1730 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1732 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1735 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1736 GPR64ArgRegs = GPR64ArgRegsWin64;
1737 XMMArgRegs = XMMArgRegsWin64;
1739 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1740 GPR64ArgRegs = GPR64ArgRegs64Bit;
1741 XMMArgRegs = XMMArgRegs64Bit;
1743 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1745 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1748 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1749 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1750 "SSE register cannot be used when SSE is disabled!");
1751 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1752 "SSE register cannot be used when SSE is disabled!");
1753 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1754 // Kernel mode asks for SSE to be disabled, so don't push them
1756 TotalNumXMMRegs = 0;
1758 // For X86-64, if there are vararg parameters that are passed via
1759 // registers, then we must store them to their spots on the stack so they
1760 // may be loaded by deferencing the result of va_next.
1761 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1762 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1763 FuncInfo->setRegSaveFrameIndex(
1764 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1767 // Store the integer parameter registers.
1768 SmallVector<SDValue, 8> MemOps;
1769 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1771 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1772 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1773 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1774 DAG.getIntPtrConstant(Offset));
1775 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1776 X86::GR64RegisterClass);
1777 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1779 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1780 PseudoSourceValue::getFixedStack(
1781 FuncInfo->getRegSaveFrameIndex()),
1782 Offset, false, false, 0);
1783 MemOps.push_back(Store);
1787 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1788 // Now store the XMM (fp + vector) parameter registers.
1789 SmallVector<SDValue, 11> SaveXMMOps;
1790 SaveXMMOps.push_back(Chain);
1792 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1793 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1794 SaveXMMOps.push_back(ALVal);
1796 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1797 FuncInfo->getRegSaveFrameIndex()));
1798 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1799 FuncInfo->getVarArgsFPOffset()));
1801 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1802 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1803 X86::VR128RegisterClass);
1804 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1805 SaveXMMOps.push_back(Val);
1807 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1809 &SaveXMMOps[0], SaveXMMOps.size()));
1812 if (!MemOps.empty())
1813 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1814 &MemOps[0], MemOps.size());
1818 // Some CCs need callee pop.
1819 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1820 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1822 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1823 // If this is an sret function, the return should pop the hidden pointer.
1824 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1825 FuncInfo->setBytesToPopOnReturn(4);
1829 // RegSaveFrameIndex is X86-64 only.
1830 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1831 if (CallConv == CallingConv::X86_FastCall ||
1832 CallConv == CallingConv::X86_ThisCall)
1833 // fastcc functions can't have varargs.
1834 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1841 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1842 SDValue StackPtr, SDValue Arg,
1843 DebugLoc dl, SelectionDAG &DAG,
1844 const CCValAssign &VA,
1845 ISD::ArgFlagsTy Flags) const {
1846 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1847 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1848 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1849 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1850 if (Flags.isByVal()) {
1851 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1853 return DAG.getStore(Chain, dl, Arg, PtrOff,
1854 PseudoSourceValue::getStack(), LocMemOffset,
1858 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1859 /// optimization is performed and it is required.
1861 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1862 SDValue &OutRetAddr, SDValue Chain,
1863 bool IsTailCall, bool Is64Bit,
1864 int FPDiff, DebugLoc dl) const {
1865 // Adjust the Return address stack slot.
1866 EVT VT = getPointerTy();
1867 OutRetAddr = getReturnAddressFrameIndex(DAG);
1869 // Load the "old" Return address.
1870 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1871 return SDValue(OutRetAddr.getNode(), 1);
1874 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1875 /// optimization is performed and it is required (FPDiff!=0).
1877 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1878 SDValue Chain, SDValue RetAddrFrIdx,
1879 bool Is64Bit, int FPDiff, DebugLoc dl) {
1880 // Store the return address to the appropriate stack slot.
1881 if (!FPDiff) return Chain;
1882 // Calculate the new stack slot for the return address.
1883 int SlotSize = Is64Bit ? 8 : 4;
1884 int NewReturnAddrFI =
1885 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1886 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1887 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1888 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1889 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1895 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1896 CallingConv::ID CallConv, bool isVarArg,
1898 const SmallVectorImpl<ISD::OutputArg> &Outs,
1899 const SmallVectorImpl<SDValue> &OutVals,
1900 const SmallVectorImpl<ISD::InputArg> &Ins,
1901 DebugLoc dl, SelectionDAG &DAG,
1902 SmallVectorImpl<SDValue> &InVals) const {
1903 MachineFunction &MF = DAG.getMachineFunction();
1904 bool Is64Bit = Subtarget->is64Bit();
1905 bool IsStructRet = CallIsStructReturn(Outs);
1906 bool IsSibcall = false;
1909 // Check if it's really possible to do a tail call.
1910 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1911 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1912 Outs, OutVals, Ins, DAG);
1914 // Sibcalls are automatically detected tailcalls which do not require
1916 if (!GuaranteedTailCallOpt && isTailCall)
1923 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1924 "Var args not supported with calling convention fastcc or ghc");
1926 // Analyze operands of the call, assigning locations to each operand.
1927 SmallVector<CCValAssign, 16> ArgLocs;
1928 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1929 ArgLocs, *DAG.getContext());
1930 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1932 // Get a count of how many bytes are to be pushed on the stack.
1933 unsigned NumBytes = CCInfo.getNextStackOffset();
1935 // This is a sibcall. The memory operands are available in caller's
1936 // own caller's stack.
1938 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1939 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1942 if (isTailCall && !IsSibcall) {
1943 // Lower arguments at fp - stackoffset + fpdiff.
1944 unsigned NumBytesCallerPushed =
1945 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1946 FPDiff = NumBytesCallerPushed - NumBytes;
1948 // Set the delta of movement of the returnaddr stackslot.
1949 // But only set if delta is greater than previous delta.
1950 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1951 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1955 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1957 SDValue RetAddrFrIdx;
1958 // Load return adress for tail calls.
1959 if (isTailCall && FPDiff)
1960 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1961 Is64Bit, FPDiff, dl);
1963 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1964 SmallVector<SDValue, 8> MemOpChains;
1967 // Walk the register/memloc assignments, inserting copies/loads. In the case
1968 // of tail call optimization arguments are handle later.
1969 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1970 CCValAssign &VA = ArgLocs[i];
1971 EVT RegVT = VA.getLocVT();
1972 SDValue Arg = OutVals[i];
1973 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1974 bool isByVal = Flags.isByVal();
1976 // Promote the value if needed.
1977 switch (VA.getLocInfo()) {
1978 default: llvm_unreachable("Unknown loc info!");
1979 case CCValAssign::Full: break;
1980 case CCValAssign::SExt:
1981 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1983 case CCValAssign::ZExt:
1984 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1986 case CCValAssign::AExt:
1987 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1988 // Special case: passing MMX values in XMM registers.
1989 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1990 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1991 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1993 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1995 case CCValAssign::BCvt:
1996 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1998 case CCValAssign::Indirect: {
1999 // Store the argument.
2000 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2001 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2002 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2003 PseudoSourceValue::getFixedStack(FI), 0,
2010 if (VA.isRegLoc()) {
2011 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2012 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2013 assert(VA.isMemLoc());
2014 if (StackPtr.getNode() == 0)
2015 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2016 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2017 dl, DAG, VA, Flags));
2021 if (!MemOpChains.empty())
2022 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2023 &MemOpChains[0], MemOpChains.size());
2025 // Build a sequence of copy-to-reg nodes chained together with token chain
2026 // and flag operands which copy the outgoing args into registers.
2028 // Tail call byval lowering might overwrite argument registers so in case of
2029 // tail call optimization the copies to registers are lowered later.
2031 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2032 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2033 RegsToPass[i].second, InFlag);
2034 InFlag = Chain.getValue(1);
2037 if (Subtarget->isPICStyleGOT()) {
2038 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2041 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2042 DAG.getNode(X86ISD::GlobalBaseReg,
2043 DebugLoc(), getPointerTy()),
2045 InFlag = Chain.getValue(1);
2047 // If we are tail calling and generating PIC/GOT style code load the
2048 // address of the callee into ECX. The value in ecx is used as target of
2049 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2050 // for tail calls on PIC/GOT architectures. Normally we would just put the
2051 // address of GOT into ebx and then call target@PLT. But for tail calls
2052 // ebx would be restored (since ebx is callee saved) before jumping to the
2055 // Note: The actual moving to ECX is done further down.
2056 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2057 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2058 !G->getGlobal()->hasProtectedVisibility())
2059 Callee = LowerGlobalAddress(Callee, DAG);
2060 else if (isa<ExternalSymbolSDNode>(Callee))
2061 Callee = LowerExternalSymbol(Callee, DAG);
2065 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
2066 // From AMD64 ABI document:
2067 // For calls that may call functions that use varargs or stdargs
2068 // (prototype-less calls or calls to functions containing ellipsis (...) in
2069 // the declaration) %al is used as hidden argument to specify the number
2070 // of SSE registers used. The contents of %al do not need to match exactly
2071 // the number of registers, but must be an ubound on the number of SSE
2072 // registers used and is in the range 0 - 8 inclusive.
2074 // Count the number of XMM registers allocated.
2075 static const unsigned XMMArgRegs[] = {
2076 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2077 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2079 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2080 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2081 && "SSE registers cannot be used when SSE is disabled");
2083 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2084 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2085 InFlag = Chain.getValue(1);
2089 // For tail calls lower the arguments to the 'real' stack slot.
2091 // Force all the incoming stack arguments to be loaded from the stack
2092 // before any new outgoing arguments are stored to the stack, because the
2093 // outgoing stack slots may alias the incoming argument stack slots, and
2094 // the alias isn't otherwise explicit. This is slightly more conservative
2095 // than necessary, because it means that each store effectively depends
2096 // on every argument instead of just those arguments it would clobber.
2097 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2099 SmallVector<SDValue, 8> MemOpChains2;
2102 // Do not flag preceeding copytoreg stuff together with the following stuff.
2104 if (GuaranteedTailCallOpt) {
2105 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2106 CCValAssign &VA = ArgLocs[i];
2109 assert(VA.isMemLoc());
2110 SDValue Arg = OutVals[i];
2111 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2112 // Create frame index.
2113 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2114 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2115 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2116 FIN = DAG.getFrameIndex(FI, getPointerTy());
2118 if (Flags.isByVal()) {
2119 // Copy relative to framepointer.
2120 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2121 if (StackPtr.getNode() == 0)
2122 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2124 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2126 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2130 // Store relative to framepointer.
2131 MemOpChains2.push_back(
2132 DAG.getStore(ArgChain, dl, Arg, FIN,
2133 PseudoSourceValue::getFixedStack(FI), 0,
2139 if (!MemOpChains2.empty())
2140 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2141 &MemOpChains2[0], MemOpChains2.size());
2143 // Copy arguments to their registers.
2144 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2145 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2146 RegsToPass[i].second, InFlag);
2147 InFlag = Chain.getValue(1);
2151 // Store the return address to the appropriate stack slot.
2152 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2156 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2157 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2158 // In the 64-bit large code model, we have to make all calls
2159 // through a register, since the call instruction's 32-bit
2160 // pc-relative offset may not be large enough to hold the whole
2162 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2163 // If the callee is a GlobalAddress node (quite common, every direct call
2164 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2167 // We should use extra load for direct calls to dllimported functions in
2169 const GlobalValue *GV = G->getGlobal();
2170 if (!GV->hasDLLImportLinkage()) {
2171 unsigned char OpFlags = 0;
2173 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2174 // external symbols most go through the PLT in PIC mode. If the symbol
2175 // has hidden or protected visibility, or if it is static or local, then
2176 // we don't need to use the PLT - we can directly call it.
2177 if (Subtarget->isTargetELF() &&
2178 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2179 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2180 OpFlags = X86II::MO_PLT;
2181 } else if (Subtarget->isPICStyleStubAny() &&
2182 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2183 Subtarget->getDarwinVers() < 9) {
2184 // PC-relative references to external symbols should go through $stub,
2185 // unless we're building with the leopard linker or later, which
2186 // automatically synthesizes these stubs.
2187 OpFlags = X86II::MO_DARWIN_STUB;
2190 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2191 G->getOffset(), OpFlags);
2193 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2194 unsigned char OpFlags = 0;
2196 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2197 // symbols should go through the PLT.
2198 if (Subtarget->isTargetELF() &&
2199 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2200 OpFlags = X86II::MO_PLT;
2201 } else if (Subtarget->isPICStyleStubAny() &&
2202 Subtarget->getDarwinVers() < 9) {
2203 // PC-relative references to external symbols should go through $stub,
2204 // unless we're building with the leopard linker or later, which
2205 // automatically synthesizes these stubs.
2206 OpFlags = X86II::MO_DARWIN_STUB;
2209 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2213 // Returns a chain & a flag for retval copy to use.
2214 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2215 SmallVector<SDValue, 8> Ops;
2217 if (!IsSibcall && isTailCall) {
2218 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2219 DAG.getIntPtrConstant(0, true), InFlag);
2220 InFlag = Chain.getValue(1);
2223 Ops.push_back(Chain);
2224 Ops.push_back(Callee);
2227 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2229 // Add argument registers to the end of the list so that they are known live
2231 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2232 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2233 RegsToPass[i].second.getValueType()));
2235 // Add an implicit use GOT pointer in EBX.
2236 if (!isTailCall && Subtarget->isPICStyleGOT())
2237 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2239 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2240 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64())
2241 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2243 if (InFlag.getNode())
2244 Ops.push_back(InFlag);
2248 //// If this is the first return lowered for this function, add the regs
2249 //// to the liveout set for the function.
2250 // This isn't right, although it's probably harmless on x86; liveouts
2251 // should be computed from returns not tail calls. Consider a void
2252 // function making a tail call to a function returning int.
2253 return DAG.getNode(X86ISD::TC_RETURN, dl,
2254 NodeTys, &Ops[0], Ops.size());
2257 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2258 InFlag = Chain.getValue(1);
2260 // Create the CALLSEQ_END node.
2261 unsigned NumBytesForCalleeToPush;
2262 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2263 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2264 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2265 // If this is a call to a struct-return function, the callee
2266 // pops the hidden struct pointer, so we have to push it back.
2267 // This is common for Darwin/X86, Linux & Mingw32 targets.
2268 NumBytesForCalleeToPush = 4;
2270 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2272 // Returns a flag for retval copy to use.
2274 Chain = DAG.getCALLSEQ_END(Chain,
2275 DAG.getIntPtrConstant(NumBytes, true),
2276 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2279 InFlag = Chain.getValue(1);
2282 // Handle result values, copying them out of physregs into vregs that we
2284 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2285 Ins, dl, DAG, InVals);
2289 //===----------------------------------------------------------------------===//
2290 // Fast Calling Convention (tail call) implementation
2291 //===----------------------------------------------------------------------===//
2293 // Like std call, callee cleans arguments, convention except that ECX is
2294 // reserved for storing the tail called function address. Only 2 registers are
2295 // free for argument passing (inreg). Tail call optimization is performed
2297 // * tailcallopt is enabled
2298 // * caller/callee are fastcc
2299 // On X86_64 architecture with GOT-style position independent code only local
2300 // (within module) calls are supported at the moment.
2301 // To keep the stack aligned according to platform abi the function
2302 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2303 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2304 // If a tail called function callee has more arguments than the caller the
2305 // caller needs to make sure that there is room to move the RETADDR to. This is
2306 // achieved by reserving an area the size of the argument delta right after the
2307 // original REtADDR, but before the saved framepointer or the spilled registers
2308 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2320 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2321 /// for a 16 byte align requirement.
2323 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2324 SelectionDAG& DAG) const {
2325 MachineFunction &MF = DAG.getMachineFunction();
2326 const TargetMachine &TM = MF.getTarget();
2327 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2328 unsigned StackAlignment = TFI.getStackAlignment();
2329 uint64_t AlignMask = StackAlignment - 1;
2330 int64_t Offset = StackSize;
2331 uint64_t SlotSize = TD->getPointerSize();
2332 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2333 // Number smaller than 12 so just add the difference.
2334 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2336 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2337 Offset = ((~AlignMask) & Offset) + StackAlignment +
2338 (StackAlignment-SlotSize);
2343 /// MatchingStackOffset - Return true if the given stack call argument is
2344 /// already available in the same position (relatively) of the caller's
2345 /// incoming argument stack.
2347 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2348 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2349 const X86InstrInfo *TII) {
2350 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2352 if (Arg.getOpcode() == ISD::CopyFromReg) {
2353 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2354 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2356 MachineInstr *Def = MRI->getVRegDef(VR);
2359 if (!Flags.isByVal()) {
2360 if (!TII->isLoadFromStackSlot(Def, FI))
2363 unsigned Opcode = Def->getOpcode();
2364 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2365 Def->getOperand(1).isFI()) {
2366 FI = Def->getOperand(1).getIndex();
2367 Bytes = Flags.getByValSize();
2371 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2372 if (Flags.isByVal())
2373 // ByVal argument is passed in as a pointer but it's now being
2374 // dereferenced. e.g.
2375 // define @foo(%struct.X* %A) {
2376 // tail call @bar(%struct.X* byval %A)
2379 SDValue Ptr = Ld->getBasePtr();
2380 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2383 FI = FINode->getIndex();
2387 assert(FI != INT_MAX);
2388 if (!MFI->isFixedObjectIndex(FI))
2390 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2393 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2394 /// for tail call optimization. Targets which want to do tail call
2395 /// optimization should implement this function.
2397 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2398 CallingConv::ID CalleeCC,
2400 bool isCalleeStructRet,
2401 bool isCallerStructRet,
2402 const SmallVectorImpl<ISD::OutputArg> &Outs,
2403 const SmallVectorImpl<SDValue> &OutVals,
2404 const SmallVectorImpl<ISD::InputArg> &Ins,
2405 SelectionDAG& DAG) const {
2406 if (!IsTailCallConvention(CalleeCC) &&
2407 CalleeCC != CallingConv::C)
2410 // If -tailcallopt is specified, make fastcc functions tail-callable.
2411 const MachineFunction &MF = DAG.getMachineFunction();
2412 const Function *CallerF = DAG.getMachineFunction().getFunction();
2413 CallingConv::ID CallerCC = CallerF->getCallingConv();
2414 bool CCMatch = CallerCC == CalleeCC;
2416 if (GuaranteedTailCallOpt) {
2417 if (IsTailCallConvention(CalleeCC) && CCMatch)
2422 // Look for obvious safe cases to perform tail call optimization that do not
2423 // require ABI changes. This is what gcc calls sibcall.
2425 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2426 // emit a special epilogue.
2427 if (RegInfo->needsStackRealignment(MF))
2430 // Do not sibcall optimize vararg calls unless the call site is not passing
2432 if (isVarArg && !Outs.empty())
2435 // Also avoid sibcall optimization if either caller or callee uses struct
2436 // return semantics.
2437 if (isCalleeStructRet || isCallerStructRet)
2440 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2441 // Therefore if it's not used by the call it is not safe to optimize this into
2443 bool Unused = false;
2444 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2451 SmallVector<CCValAssign, 16> RVLocs;
2452 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2453 RVLocs, *DAG.getContext());
2454 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2455 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2456 CCValAssign &VA = RVLocs[i];
2457 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2462 // If the calling conventions do not match, then we'd better make sure the
2463 // results are returned in the same way as what the caller expects.
2465 SmallVector<CCValAssign, 16> RVLocs1;
2466 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2467 RVLocs1, *DAG.getContext());
2468 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2470 SmallVector<CCValAssign, 16> RVLocs2;
2471 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2472 RVLocs2, *DAG.getContext());
2473 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2475 if (RVLocs1.size() != RVLocs2.size())
2477 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2478 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2480 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2482 if (RVLocs1[i].isRegLoc()) {
2483 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2486 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2492 // If the callee takes no arguments then go on to check the results of the
2494 if (!Outs.empty()) {
2495 // Check if stack adjustment is needed. For now, do not do this if any
2496 // argument is passed on the stack.
2497 SmallVector<CCValAssign, 16> ArgLocs;
2498 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2499 ArgLocs, *DAG.getContext());
2500 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2501 if (CCInfo.getNextStackOffset()) {
2502 MachineFunction &MF = DAG.getMachineFunction();
2503 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2505 if (Subtarget->isTargetWin64())
2506 // Win64 ABI has additional complications.
2509 // Check if the arguments are already laid out in the right way as
2510 // the caller's fixed stack objects.
2511 MachineFrameInfo *MFI = MF.getFrameInfo();
2512 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2513 const X86InstrInfo *TII =
2514 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2515 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2516 CCValAssign &VA = ArgLocs[i];
2517 SDValue Arg = OutVals[i];
2518 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2519 if (VA.getLocInfo() == CCValAssign::Indirect)
2521 if (!VA.isRegLoc()) {
2522 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2529 // If the tailcall address may be in a register, then make sure it's
2530 // possible to register allocate for it. In 32-bit, the call address can
2531 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2532 // callee-saved registers are restored. These happen to be the same
2533 // registers used to pass 'inreg' arguments so watch out for those.
2534 if (!Subtarget->is64Bit() &&
2535 !isa<GlobalAddressSDNode>(Callee) &&
2536 !isa<ExternalSymbolSDNode>(Callee)) {
2537 unsigned NumInRegs = 0;
2538 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2539 CCValAssign &VA = ArgLocs[i];
2542 unsigned Reg = VA.getLocReg();
2545 case X86::EAX: case X86::EDX: case X86::ECX:
2546 if (++NumInRegs == 3)
2558 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2559 return X86::createFastISel(funcInfo);
2563 //===----------------------------------------------------------------------===//
2564 // Other Lowering Hooks
2565 //===----------------------------------------------------------------------===//
2567 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2568 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2570 default: llvm_unreachable("Unknown x86 shuffle node");
2571 case X86ISD::PSHUFD:
2572 case X86ISD::PSHUFHW:
2573 case X86ISD::PSHUFLW:
2574 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2580 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2581 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2583 default: llvm_unreachable("Unknown x86 shuffle node");
2584 case X86ISD::SHUFPD:
2585 case X86ISD::SHUFPS:
2586 return DAG.getNode(Opc, dl, VT, V1, V2,
2587 DAG.getConstant(TargetMask, MVT::i8));
2592 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2593 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2595 default: llvm_unreachable("Unknown x86 shuffle node");
2596 case X86ISD::MOVLHPS:
2597 case X86ISD::PUNPCKLDQ:
2598 return DAG.getNode(Opc, dl, VT, V1, V2);
2603 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2604 MachineFunction &MF = DAG.getMachineFunction();
2605 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2606 int ReturnAddrIndex = FuncInfo->getRAIndex();
2608 if (ReturnAddrIndex == 0) {
2609 // Set up a frame object for the return address.
2610 uint64_t SlotSize = TD->getPointerSize();
2611 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2613 FuncInfo->setRAIndex(ReturnAddrIndex);
2616 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2620 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2621 bool hasSymbolicDisplacement) {
2622 // Offset should fit into 32 bit immediate field.
2623 if (!isInt<32>(Offset))
2626 // If we don't have a symbolic displacement - we don't have any extra
2628 if (!hasSymbolicDisplacement)
2631 // FIXME: Some tweaks might be needed for medium code model.
2632 if (M != CodeModel::Small && M != CodeModel::Kernel)
2635 // For small code model we assume that latest object is 16MB before end of 31
2636 // bits boundary. We may also accept pretty large negative constants knowing
2637 // that all objects are in the positive half of address space.
2638 if (M == CodeModel::Small && Offset < 16*1024*1024)
2641 // For kernel code model we know that all object resist in the negative half
2642 // of 32bits address space. We may not accept negative offsets, since they may
2643 // be just off and we may accept pretty large positive ones.
2644 if (M == CodeModel::Kernel && Offset > 0)
2650 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2651 /// specific condition code, returning the condition code and the LHS/RHS of the
2652 /// comparison to make.
2653 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2654 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2656 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2657 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2658 // X > -1 -> X == 0, jump !sign.
2659 RHS = DAG.getConstant(0, RHS.getValueType());
2660 return X86::COND_NS;
2661 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2662 // X < 0 -> X == 0, jump on sign.
2664 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2666 RHS = DAG.getConstant(0, RHS.getValueType());
2667 return X86::COND_LE;
2671 switch (SetCCOpcode) {
2672 default: llvm_unreachable("Invalid integer condition!");
2673 case ISD::SETEQ: return X86::COND_E;
2674 case ISD::SETGT: return X86::COND_G;
2675 case ISD::SETGE: return X86::COND_GE;
2676 case ISD::SETLT: return X86::COND_L;
2677 case ISD::SETLE: return X86::COND_LE;
2678 case ISD::SETNE: return X86::COND_NE;
2679 case ISD::SETULT: return X86::COND_B;
2680 case ISD::SETUGT: return X86::COND_A;
2681 case ISD::SETULE: return X86::COND_BE;
2682 case ISD::SETUGE: return X86::COND_AE;
2686 // First determine if it is required or is profitable to flip the operands.
2688 // If LHS is a foldable load, but RHS is not, flip the condition.
2689 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2690 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2691 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2692 std::swap(LHS, RHS);
2695 switch (SetCCOpcode) {
2701 std::swap(LHS, RHS);
2705 // On a floating point condition, the flags are set as follows:
2707 // 0 | 0 | 0 | X > Y
2708 // 0 | 0 | 1 | X < Y
2709 // 1 | 0 | 0 | X == Y
2710 // 1 | 1 | 1 | unordered
2711 switch (SetCCOpcode) {
2712 default: llvm_unreachable("Condcode should be pre-legalized away");
2714 case ISD::SETEQ: return X86::COND_E;
2715 case ISD::SETOLT: // flipped
2717 case ISD::SETGT: return X86::COND_A;
2718 case ISD::SETOLE: // flipped
2720 case ISD::SETGE: return X86::COND_AE;
2721 case ISD::SETUGT: // flipped
2723 case ISD::SETLT: return X86::COND_B;
2724 case ISD::SETUGE: // flipped
2726 case ISD::SETLE: return X86::COND_BE;
2728 case ISD::SETNE: return X86::COND_NE;
2729 case ISD::SETUO: return X86::COND_P;
2730 case ISD::SETO: return X86::COND_NP;
2732 case ISD::SETUNE: return X86::COND_INVALID;
2736 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2737 /// code. Current x86 isa includes the following FP cmov instructions:
2738 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2739 static bool hasFPCMov(unsigned X86CC) {
2755 /// isFPImmLegal - Returns true if the target can instruction select the
2756 /// specified FP immediate natively. If false, the legalizer will
2757 /// materialize the FP immediate as a load from a constant pool.
2758 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2759 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2760 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2766 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2767 /// the specified range (L, H].
2768 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2769 return (Val < 0) || (Val >= Low && Val < Hi);
2772 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2773 /// specified value.
2774 static bool isUndefOrEqual(int Val, int CmpVal) {
2775 if (Val < 0 || Val == CmpVal)
2780 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2781 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2782 /// the second operand.
2783 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2784 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2785 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2786 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2787 return (Mask[0] < 2 && Mask[1] < 2);
2791 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2792 SmallVector<int, 8> M;
2794 return ::isPSHUFDMask(M, N->getValueType(0));
2797 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2798 /// is suitable for input to PSHUFHW.
2799 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2800 if (VT != MVT::v8i16)
2803 // Lower quadword copied in order or undef.
2804 for (int i = 0; i != 4; ++i)
2805 if (Mask[i] >= 0 && Mask[i] != i)
2808 // Upper quadword shuffled.
2809 for (int i = 4; i != 8; ++i)
2810 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2816 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2817 SmallVector<int, 8> M;
2819 return ::isPSHUFHWMask(M, N->getValueType(0));
2822 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2823 /// is suitable for input to PSHUFLW.
2824 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2825 if (VT != MVT::v8i16)
2828 // Upper quadword copied in order.
2829 for (int i = 4; i != 8; ++i)
2830 if (Mask[i] >= 0 && Mask[i] != i)
2833 // Lower quadword shuffled.
2834 for (int i = 0; i != 4; ++i)
2841 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2842 SmallVector<int, 8> M;
2844 return ::isPSHUFLWMask(M, N->getValueType(0));
2847 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2848 /// is suitable for input to PALIGNR.
2849 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2851 int i, e = VT.getVectorNumElements();
2853 // Do not handle v2i64 / v2f64 shuffles with palignr.
2854 if (e < 4 || !hasSSSE3)
2857 for (i = 0; i != e; ++i)
2861 // All undef, not a palignr.
2865 // Determine if it's ok to perform a palignr with only the LHS, since we
2866 // don't have access to the actual shuffle elements to see if RHS is undef.
2867 bool Unary = Mask[i] < (int)e;
2868 bool NeedsUnary = false;
2870 int s = Mask[i] - i;
2872 // Check the rest of the elements to see if they are consecutive.
2873 for (++i; i != e; ++i) {
2878 Unary = Unary && (m < (int)e);
2879 NeedsUnary = NeedsUnary || (m < s);
2881 if (NeedsUnary && !Unary)
2883 if (Unary && m != ((s+i) & (e-1)))
2885 if (!Unary && m != (s+i))
2891 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2892 SmallVector<int, 8> M;
2894 return ::isPALIGNRMask(M, N->getValueType(0), true);
2897 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2898 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2899 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2900 int NumElems = VT.getVectorNumElements();
2901 if (NumElems != 2 && NumElems != 4)
2904 int Half = NumElems / 2;
2905 for (int i = 0; i < Half; ++i)
2906 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2908 for (int i = Half; i < NumElems; ++i)
2909 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2915 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2916 SmallVector<int, 8> M;
2918 return ::isSHUFPMask(M, N->getValueType(0));
2921 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2922 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2923 /// half elements to come from vector 1 (which would equal the dest.) and
2924 /// the upper half to come from vector 2.
2925 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2926 int NumElems = VT.getVectorNumElements();
2928 if (NumElems != 2 && NumElems != 4)
2931 int Half = NumElems / 2;
2932 for (int i = 0; i < Half; ++i)
2933 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2935 for (int i = Half; i < NumElems; ++i)
2936 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2941 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2942 SmallVector<int, 8> M;
2944 return isCommutedSHUFPMask(M, N->getValueType(0));
2947 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2948 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2949 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2950 if (N->getValueType(0).getVectorNumElements() != 4)
2953 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2954 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2955 isUndefOrEqual(N->getMaskElt(1), 7) &&
2956 isUndefOrEqual(N->getMaskElt(2), 2) &&
2957 isUndefOrEqual(N->getMaskElt(3), 3);
2960 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2961 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2963 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2964 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2969 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2970 isUndefOrEqual(N->getMaskElt(1), 3) &&
2971 isUndefOrEqual(N->getMaskElt(2), 2) &&
2972 isUndefOrEqual(N->getMaskElt(3), 3);
2975 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2976 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2977 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2978 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2980 if (NumElems != 2 && NumElems != 4)
2983 for (unsigned i = 0; i < NumElems/2; ++i)
2984 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2987 for (unsigned i = NumElems/2; i < NumElems; ++i)
2988 if (!isUndefOrEqual(N->getMaskElt(i), i))
2994 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2995 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2996 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2997 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2999 if (NumElems != 2 && NumElems != 4)
3002 for (unsigned i = 0; i < NumElems/2; ++i)
3003 if (!isUndefOrEqual(N->getMaskElt(i), i))
3006 for (unsigned i = 0; i < NumElems/2; ++i)
3007 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3013 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3014 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3015 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3016 bool V2IsSplat = false) {
3017 int NumElts = VT.getVectorNumElements();
3018 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3021 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3023 int BitI1 = Mask[i+1];
3024 if (!isUndefOrEqual(BitI, j))
3027 if (!isUndefOrEqual(BitI1, NumElts))
3030 if (!isUndefOrEqual(BitI1, j + NumElts))
3037 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3038 SmallVector<int, 8> M;
3040 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3043 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3044 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3045 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3046 bool V2IsSplat = false) {
3047 int NumElts = VT.getVectorNumElements();
3048 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3051 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3053 int BitI1 = Mask[i+1];
3054 if (!isUndefOrEqual(BitI, j + NumElts/2))
3057 if (isUndefOrEqual(BitI1, NumElts))
3060 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3067 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3068 SmallVector<int, 8> M;
3070 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3073 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3074 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3076 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3077 int NumElems = VT.getVectorNumElements();
3078 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3081 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3083 int BitI1 = Mask[i+1];
3084 if (!isUndefOrEqual(BitI, j))
3086 if (!isUndefOrEqual(BitI1, j))
3092 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3093 SmallVector<int, 8> M;
3095 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3098 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3099 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3101 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3102 int NumElems = VT.getVectorNumElements();
3103 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3106 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3108 int BitI1 = Mask[i+1];
3109 if (!isUndefOrEqual(BitI, j))
3111 if (!isUndefOrEqual(BitI1, j))
3117 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3118 SmallVector<int, 8> M;
3120 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3123 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3124 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3125 /// MOVSD, and MOVD, i.e. setting the lowest element.
3126 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3127 if (VT.getVectorElementType().getSizeInBits() < 32)
3130 int NumElts = VT.getVectorNumElements();
3132 if (!isUndefOrEqual(Mask[0], NumElts))
3135 for (int i = 1; i < NumElts; ++i)
3136 if (!isUndefOrEqual(Mask[i], i))
3142 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3143 SmallVector<int, 8> M;
3145 return ::isMOVLMask(M, N->getValueType(0));
3148 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3149 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3150 /// element of vector 2 and the other elements to come from vector 1 in order.
3151 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3152 bool V2IsSplat = false, bool V2IsUndef = false) {
3153 int NumOps = VT.getVectorNumElements();
3154 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3157 if (!isUndefOrEqual(Mask[0], 0))
3160 for (int i = 1; i < NumOps; ++i)
3161 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3162 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3163 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3169 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3170 bool V2IsUndef = false) {
3171 SmallVector<int, 8> M;
3173 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3176 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3177 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3178 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3179 if (N->getValueType(0).getVectorNumElements() != 4)
3182 // Expect 1, 1, 3, 3
3183 for (unsigned i = 0; i < 2; ++i) {
3184 int Elt = N->getMaskElt(i);
3185 if (Elt >= 0 && Elt != 1)
3190 for (unsigned i = 2; i < 4; ++i) {
3191 int Elt = N->getMaskElt(i);
3192 if (Elt >= 0 && Elt != 3)
3197 // Don't use movshdup if it can be done with a shufps.
3198 // FIXME: verify that matching u, u, 3, 3 is what we want.
3202 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3203 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3204 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3205 if (N->getValueType(0).getVectorNumElements() != 4)
3208 // Expect 0, 0, 2, 2
3209 for (unsigned i = 0; i < 2; ++i)
3210 if (N->getMaskElt(i) > 0)
3214 for (unsigned i = 2; i < 4; ++i) {
3215 int Elt = N->getMaskElt(i);
3216 if (Elt >= 0 && Elt != 2)
3221 // Don't use movsldup if it can be done with a shufps.
3225 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3226 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3227 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3228 int e = N->getValueType(0).getVectorNumElements() / 2;
3230 for (int i = 0; i < e; ++i)
3231 if (!isUndefOrEqual(N->getMaskElt(i), i))
3233 for (int i = 0; i < e; ++i)
3234 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3239 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3240 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3241 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3242 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3243 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3245 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3247 for (int i = 0; i < NumOperands; ++i) {
3248 int Val = SVOp->getMaskElt(NumOperands-i-1);
3249 if (Val < 0) Val = 0;
3250 if (Val >= NumOperands) Val -= NumOperands;
3252 if (i != NumOperands - 1)
3258 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3259 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3260 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3261 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3263 // 8 nodes, but we only care about the last 4.
3264 for (unsigned i = 7; i >= 4; --i) {
3265 int Val = SVOp->getMaskElt(i);
3274 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3275 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3276 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3277 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3279 // 8 nodes, but we only care about the first 4.
3280 for (int i = 3; i >= 0; --i) {
3281 int Val = SVOp->getMaskElt(i);
3290 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3291 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3292 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3293 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3294 EVT VVT = N->getValueType(0);
3295 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3299 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3300 Val = SVOp->getMaskElt(i);
3304 return (Val - i) * EltSize;
3307 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3309 bool X86::isZeroNode(SDValue Elt) {
3310 return ((isa<ConstantSDNode>(Elt) &&
3311 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3312 (isa<ConstantFPSDNode>(Elt) &&
3313 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3316 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3317 /// their permute mask.
3318 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3319 SelectionDAG &DAG) {
3320 EVT VT = SVOp->getValueType(0);
3321 unsigned NumElems = VT.getVectorNumElements();
3322 SmallVector<int, 8> MaskVec;
3324 for (unsigned i = 0; i != NumElems; ++i) {
3325 int idx = SVOp->getMaskElt(i);
3327 MaskVec.push_back(idx);
3328 else if (idx < (int)NumElems)
3329 MaskVec.push_back(idx + NumElems);
3331 MaskVec.push_back(idx - NumElems);
3333 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3334 SVOp->getOperand(0), &MaskVec[0]);
3337 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3338 /// the two vector operands have swapped position.
3339 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3340 unsigned NumElems = VT.getVectorNumElements();
3341 for (unsigned i = 0; i != NumElems; ++i) {
3345 else if (idx < (int)NumElems)
3346 Mask[i] = idx + NumElems;
3348 Mask[i] = idx - NumElems;
3352 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3353 /// match movhlps. The lower half elements should come from upper half of
3354 /// V1 (and in order), and the upper half elements should come from the upper
3355 /// half of V2 (and in order).
3356 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3357 if (Op->getValueType(0).getVectorNumElements() != 4)
3359 for (unsigned i = 0, e = 2; i != e; ++i)
3360 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3362 for (unsigned i = 2; i != 4; ++i)
3363 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3368 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3369 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3371 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3372 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3374 N = N->getOperand(0).getNode();
3375 if (!ISD::isNON_EXTLoad(N))
3378 *LD = cast<LoadSDNode>(N);
3382 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3383 /// match movlp{s|d}. The lower half elements should come from lower half of
3384 /// V1 (and in order), and the upper half elements should come from the upper
3385 /// half of V2 (and in order). And since V1 will become the source of the
3386 /// MOVLP, it must be either a vector load or a scalar load to vector.
3387 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3388 ShuffleVectorSDNode *Op) {
3389 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3391 // Is V2 is a vector load, don't do this transformation. We will try to use
3392 // load folding shufps op.
3393 if (ISD::isNON_EXTLoad(V2))
3396 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3398 if (NumElems != 2 && NumElems != 4)
3400 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3401 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3403 for (unsigned i = NumElems/2; i != NumElems; ++i)
3404 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3409 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3411 static bool isSplatVector(SDNode *N) {
3412 if (N->getOpcode() != ISD::BUILD_VECTOR)
3415 SDValue SplatValue = N->getOperand(0);
3416 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3417 if (N->getOperand(i) != SplatValue)
3422 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3423 /// to an zero vector.
3424 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3425 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3426 SDValue V1 = N->getOperand(0);
3427 SDValue V2 = N->getOperand(1);
3428 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3429 for (unsigned i = 0; i != NumElems; ++i) {
3430 int Idx = N->getMaskElt(i);
3431 if (Idx >= (int)NumElems) {
3432 unsigned Opc = V2.getOpcode();
3433 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3435 if (Opc != ISD::BUILD_VECTOR ||
3436 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3438 } else if (Idx >= 0) {
3439 unsigned Opc = V1.getOpcode();
3440 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3442 if (Opc != ISD::BUILD_VECTOR ||
3443 !X86::isZeroNode(V1.getOperand(Idx)))
3450 /// getZeroVector - Returns a vector of specified type with all zero elements.
3452 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3454 assert(VT.isVector() && "Expected a vector type");
3456 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted
3457 // to their dest type. This ensures they get CSE'd.
3459 if (VT.getSizeInBits() == 64) { // MMX
3460 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3461 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3462 } else if (VT.getSizeInBits() == 128) {
3463 if (HasSSE2) { // SSE2
3464 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3465 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3467 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3468 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3470 } else if (VT.getSizeInBits() == 256) { // AVX
3471 // 256-bit logic and arithmetic instructions in AVX are
3472 // all floating-point, no support for integer ops. Default
3473 // to emitting fp zeroed vectors then.
3474 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3475 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3476 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3478 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3481 /// getOnesVector - Returns a vector of specified type with all bits set.
3483 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3484 assert(VT.isVector() && "Expected a vector type");
3486 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3487 // type. This ensures they get CSE'd.
3488 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3490 if (VT.getSizeInBits() == 64) // MMX
3491 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3493 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3494 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3498 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3499 /// that point to V2 points to its first element.
3500 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3501 EVT VT = SVOp->getValueType(0);
3502 unsigned NumElems = VT.getVectorNumElements();
3504 bool Changed = false;
3505 SmallVector<int, 8> MaskVec;
3506 SVOp->getMask(MaskVec);
3508 for (unsigned i = 0; i != NumElems; ++i) {
3509 if (MaskVec[i] > (int)NumElems) {
3510 MaskVec[i] = NumElems;
3515 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3516 SVOp->getOperand(1), &MaskVec[0]);
3517 return SDValue(SVOp, 0);
3520 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3521 /// operation of specified width.
3522 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3524 unsigned NumElems = VT.getVectorNumElements();
3525 SmallVector<int, 8> Mask;
3526 Mask.push_back(NumElems);
3527 for (unsigned i = 1; i != NumElems; ++i)
3529 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3532 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3533 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3535 unsigned NumElems = VT.getVectorNumElements();
3536 SmallVector<int, 8> Mask;
3537 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3539 Mask.push_back(i + NumElems);
3541 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3544 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3545 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3547 unsigned NumElems = VT.getVectorNumElements();
3548 unsigned Half = NumElems/2;
3549 SmallVector<int, 8> Mask;
3550 for (unsigned i = 0; i != Half; ++i) {
3551 Mask.push_back(i + Half);
3552 Mask.push_back(i + NumElems + Half);
3554 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3557 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3558 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3559 if (SV->getValueType(0).getVectorNumElements() <= 4)
3560 return SDValue(SV, 0);
3562 EVT PVT = MVT::v4f32;
3563 EVT VT = SV->getValueType(0);
3564 DebugLoc dl = SV->getDebugLoc();
3565 SDValue V1 = SV->getOperand(0);
3566 int NumElems = VT.getVectorNumElements();
3567 int EltNo = SV->getSplatIndex();
3569 // unpack elements to the correct location
3570 while (NumElems > 4) {
3571 if (EltNo < NumElems/2) {
3572 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3574 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3575 EltNo -= NumElems/2;
3580 // Perform the splat.
3581 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3582 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3583 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3584 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3587 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3588 /// vector of zero or undef vector. This produces a shuffle where the low
3589 /// element of V2 is swizzled into the zero/undef vector, landing at element
3590 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3591 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3592 bool isZero, bool HasSSE2,
3593 SelectionDAG &DAG) {
3594 EVT VT = V2.getValueType();
3596 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3597 unsigned NumElems = VT.getVectorNumElements();
3598 SmallVector<int, 16> MaskVec;
3599 for (unsigned i = 0; i != NumElems; ++i)
3600 // If this is the insertion idx, put the low elt of V2 here.
3601 MaskVec.push_back(i == Idx ? NumElems : i);
3602 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3605 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3606 /// a shuffle that is zero.
3608 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3609 bool Low, SelectionDAG &DAG) {
3610 unsigned NumZeros = 0;
3611 for (int i = 0; i < NumElems; ++i) {
3612 unsigned Index = Low ? i : NumElems-i-1;
3613 int Idx = SVOp->getMaskElt(Index);
3618 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3619 if (Elt.getNode() && X86::isZeroNode(Elt))
3627 /// isVectorShift - Returns true if the shuffle can be implemented as a
3628 /// logical left or right shift of a vector.
3629 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3630 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3631 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3632 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3635 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3638 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3642 bool SeenV1 = false;
3643 bool SeenV2 = false;
3644 for (unsigned i = NumZeros; i < NumElems; ++i) {
3645 unsigned Val = isLeft ? (i - NumZeros) : i;
3646 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3649 unsigned Idx = (unsigned) Idx_;
3659 if (SeenV1 && SeenV2)
3662 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3668 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3670 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3671 unsigned NumNonZero, unsigned NumZero,
3673 const TargetLowering &TLI) {
3677 DebugLoc dl = Op.getDebugLoc();
3680 for (unsigned i = 0; i < 16; ++i) {
3681 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3682 if (ThisIsNonZero && First) {
3684 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3686 V = DAG.getUNDEF(MVT::v8i16);
3691 SDValue ThisElt(0, 0), LastElt(0, 0);
3692 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3693 if (LastIsNonZero) {
3694 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3695 MVT::i16, Op.getOperand(i-1));
3697 if (ThisIsNonZero) {
3698 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3699 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3700 ThisElt, DAG.getConstant(8, MVT::i8));
3702 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3706 if (ThisElt.getNode())
3707 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3708 DAG.getIntPtrConstant(i/2));
3712 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3715 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3717 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3718 unsigned NumNonZero, unsigned NumZero,
3720 const TargetLowering &TLI) {
3724 DebugLoc dl = Op.getDebugLoc();
3727 for (unsigned i = 0; i < 8; ++i) {
3728 bool isNonZero = (NonZeros & (1 << i)) != 0;
3732 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3734 V = DAG.getUNDEF(MVT::v8i16);
3737 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3738 MVT::v8i16, V, Op.getOperand(i),
3739 DAG.getIntPtrConstant(i));
3746 /// getVShift - Return a vector logical shift node.
3748 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3749 unsigned NumBits, SelectionDAG &DAG,
3750 const TargetLowering &TLI, DebugLoc dl) {
3751 bool isMMX = VT.getSizeInBits() == 64;
3752 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3753 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3754 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3755 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3756 DAG.getNode(Opc, dl, ShVT, SrcOp,
3757 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3761 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3762 SelectionDAG &DAG) const {
3764 // Check if the scalar load can be widened into a vector load. And if
3765 // the address is "base + cst" see if the cst can be "absorbed" into
3766 // the shuffle mask.
3767 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3768 SDValue Ptr = LD->getBasePtr();
3769 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3771 EVT PVT = LD->getValueType(0);
3772 if (PVT != MVT::i32 && PVT != MVT::f32)
3777 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3778 FI = FINode->getIndex();
3780 } else if (Ptr.getOpcode() == ISD::ADD &&
3781 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3782 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3783 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3784 Offset = Ptr.getConstantOperandVal(1);
3785 Ptr = Ptr.getOperand(0);
3790 SDValue Chain = LD->getChain();
3791 // Make sure the stack object alignment is at least 16.
3792 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3793 if (DAG.InferPtrAlignment(Ptr) < 16) {
3794 if (MFI->isFixedObjectIndex(FI)) {
3795 // Can't change the alignment. FIXME: It's possible to compute
3796 // the exact stack offset and reference FI + adjust offset instead.
3797 // If someone *really* cares about this. That's the way to implement it.
3800 MFI->setObjectAlignment(FI, 16);
3804 // (Offset % 16) must be multiple of 4. Then address is then
3805 // Ptr + (Offset & ~15).
3808 if ((Offset % 16) & 3)
3810 int64_t StartOffset = Offset & ~15;
3812 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3813 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3815 int EltNo = (Offset - StartOffset) >> 2;
3816 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3817 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3818 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3820 // Canonicalize it to a v4i32 shuffle.
3821 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3822 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3823 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3824 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3830 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3831 /// vector of type 'VT', see if the elements can be replaced by a single large
3832 /// load which has the same value as a build_vector whose operands are 'elts'.
3834 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3836 /// FIXME: we'd also like to handle the case where the last elements are zero
3837 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3838 /// There's even a handy isZeroNode for that purpose.
3839 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3840 DebugLoc &dl, SelectionDAG &DAG) {
3841 EVT EltVT = VT.getVectorElementType();
3842 unsigned NumElems = Elts.size();
3844 LoadSDNode *LDBase = NULL;
3845 unsigned LastLoadedElt = -1U;
3847 // For each element in the initializer, see if we've found a load or an undef.
3848 // If we don't find an initial load element, or later load elements are
3849 // non-consecutive, bail out.
3850 for (unsigned i = 0; i < NumElems; ++i) {
3851 SDValue Elt = Elts[i];
3853 if (!Elt.getNode() ||
3854 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3857 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3859 LDBase = cast<LoadSDNode>(Elt.getNode());
3863 if (Elt.getOpcode() == ISD::UNDEF)
3866 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3867 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3872 // If we have found an entire vector of loads and undefs, then return a large
3873 // load of the entire vector width starting at the base pointer. If we found
3874 // consecutive loads for the low half, generate a vzext_load node.
3875 if (LastLoadedElt == NumElems - 1) {
3876 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3877 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3878 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3879 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3880 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3881 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3882 LDBase->isVolatile(), LDBase->isNonTemporal(),
3883 LDBase->getAlignment());
3884 } else if (NumElems == 4 && LastLoadedElt == 1) {
3885 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3886 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3887 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3888 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3894 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3895 DebugLoc dl = Op.getDebugLoc();
3896 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1 and
3897 // all one's are handled with pcmpeqd. In AVX, zero's are handled with
3898 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
3899 // is present, so AllOnes is ignored.
3900 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
3901 (Op.getValueType().getSizeInBits() != 256 &&
3902 ISD::isBuildVectorAllOnes(Op.getNode()))) {
3903 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3904 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3905 // eliminated on x86-32 hosts.
3906 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3909 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3910 return getOnesVector(Op.getValueType(), DAG, dl);
3911 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3914 EVT VT = Op.getValueType();
3915 EVT ExtVT = VT.getVectorElementType();
3916 unsigned EVTBits = ExtVT.getSizeInBits();
3918 unsigned NumElems = Op.getNumOperands();
3919 unsigned NumZero = 0;
3920 unsigned NumNonZero = 0;
3921 unsigned NonZeros = 0;
3922 bool IsAllConstants = true;
3923 SmallSet<SDValue, 8> Values;
3924 for (unsigned i = 0; i < NumElems; ++i) {
3925 SDValue Elt = Op.getOperand(i);
3926 if (Elt.getOpcode() == ISD::UNDEF)
3929 if (Elt.getOpcode() != ISD::Constant &&
3930 Elt.getOpcode() != ISD::ConstantFP)
3931 IsAllConstants = false;
3932 if (X86::isZeroNode(Elt))
3935 NonZeros |= (1 << i);
3940 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3941 if (NumNonZero == 0)
3942 return DAG.getUNDEF(VT);
3944 // Special case for single non-zero, non-undef, element.
3945 if (NumNonZero == 1) {
3946 unsigned Idx = CountTrailingZeros_32(NonZeros);
3947 SDValue Item = Op.getOperand(Idx);
3949 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3950 // the value are obviously zero, truncate the value to i32 and do the
3951 // insertion that way. Only do this if the value is non-constant or if the
3952 // value is a constant being inserted into element 0. It is cheaper to do
3953 // a constant pool load than it is to do a movd + shuffle.
3954 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3955 (!IsAllConstants || Idx == 0)) {
3956 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3957 // Handle MMX and SSE both.
3958 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3959 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3961 // Truncate the value (which may itself be a constant) to i32, and
3962 // convert it to a vector with movd (S2V+shuffle to zero extend).
3963 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3964 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3965 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3966 Subtarget->hasSSE2(), DAG);
3968 // Now we have our 32-bit value zero extended in the low element of
3969 // a vector. If Idx != 0, swizzle it into place.
3971 SmallVector<int, 4> Mask;
3972 Mask.push_back(Idx);
3973 for (unsigned i = 1; i != VecElts; ++i)
3975 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3976 DAG.getUNDEF(Item.getValueType()),
3979 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3983 // If we have a constant or non-constant insertion into the low element of
3984 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3985 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3986 // depending on what the source datatype is.
3989 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3990 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3991 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3992 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3993 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3994 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3996 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3997 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3998 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3999 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4000 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4001 Subtarget->hasSSE2(), DAG);
4002 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
4006 // Is it a vector logical left shift?
4007 if (NumElems == 2 && Idx == 1 &&
4008 X86::isZeroNode(Op.getOperand(0)) &&
4009 !X86::isZeroNode(Op.getOperand(1))) {
4010 unsigned NumBits = VT.getSizeInBits();
4011 return getVShift(true, VT,
4012 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4013 VT, Op.getOperand(1)),
4014 NumBits/2, DAG, *this, dl);
4017 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4020 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4021 // is a non-constant being inserted into an element other than the low one,
4022 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4023 // movd/movss) to move this into the low element, then shuffle it into
4025 if (EVTBits == 32) {
4026 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4028 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4029 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4030 Subtarget->hasSSE2(), DAG);
4031 SmallVector<int, 8> MaskVec;
4032 for (unsigned i = 0; i < NumElems; i++)
4033 MaskVec.push_back(i == Idx ? 0 : 1);
4034 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4038 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4039 if (Values.size() == 1) {
4040 if (EVTBits == 32) {
4041 // Instead of a shuffle like this:
4042 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4043 // Check if it's possible to issue this instead.
4044 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4045 unsigned Idx = CountTrailingZeros_32(NonZeros);
4046 SDValue Item = Op.getOperand(Idx);
4047 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4048 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4053 // A vector full of immediates; various special cases are already
4054 // handled, so this is best done with a single constant-pool load.
4058 // Let legalizer expand 2-wide build_vectors.
4059 if (EVTBits == 64) {
4060 if (NumNonZero == 1) {
4061 // One half is zero or undef.
4062 unsigned Idx = CountTrailingZeros_32(NonZeros);
4063 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4064 Op.getOperand(Idx));
4065 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4066 Subtarget->hasSSE2(), DAG);
4071 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4072 if (EVTBits == 8 && NumElems == 16) {
4073 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4075 if (V.getNode()) return V;
4078 if (EVTBits == 16 && NumElems == 8) {
4079 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4081 if (V.getNode()) return V;
4084 // If element VT is == 32 bits, turn it into a number of shuffles.
4085 SmallVector<SDValue, 8> V;
4087 if (NumElems == 4 && NumZero > 0) {
4088 for (unsigned i = 0; i < 4; ++i) {
4089 bool isZero = !(NonZeros & (1 << i));
4091 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4093 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4096 for (unsigned i = 0; i < 2; ++i) {
4097 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4100 V[i] = V[i*2]; // Must be a zero vector.
4103 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4106 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4109 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4114 SmallVector<int, 8> MaskVec;
4115 bool Reverse = (NonZeros & 0x3) == 2;
4116 for (unsigned i = 0; i < 2; ++i)
4117 MaskVec.push_back(Reverse ? 1-i : i);
4118 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4119 for (unsigned i = 0; i < 2; ++i)
4120 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4121 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4124 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4125 // Check for a build vector of consecutive loads.
4126 for (unsigned i = 0; i < NumElems; ++i)
4127 V[i] = Op.getOperand(i);
4129 // Check for elements which are consecutive loads.
4130 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4134 // For SSE 4.1, use inserts into undef.
4135 if (getSubtarget()->hasSSE41()) {
4136 V[0] = DAG.getUNDEF(VT);
4137 for (unsigned i = 0; i < NumElems; ++i)
4138 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4139 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
4140 Op.getOperand(i), DAG.getIntPtrConstant(i));
4144 // Otherwise, expand into a number of unpckl*
4146 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4147 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4148 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4149 for (unsigned i = 0; i < NumElems; ++i)
4150 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4152 while (NumElems != 0) {
4153 for (unsigned i = 0; i < NumElems; ++i)
4154 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4163 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4164 // We support concatenate two MMX registers and place them in a MMX
4165 // register. This is better than doing a stack convert.
4166 DebugLoc dl = Op.getDebugLoc();
4167 EVT ResVT = Op.getValueType();
4168 assert(Op.getNumOperands() == 2);
4169 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4170 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4172 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4173 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4174 InVec = Op.getOperand(1);
4175 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4176 unsigned NumElts = ResVT.getVectorNumElements();
4177 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4178 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4179 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4181 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4182 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4183 Mask[0] = 0; Mask[1] = 2;
4184 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4186 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4189 // v8i16 shuffles - Prefer shuffles in the following order:
4190 // 1. [all] pshuflw, pshufhw, optional move
4191 // 2. [ssse3] 1 x pshufb
4192 // 3. [ssse3] 2 x pshufb + 1 x por
4193 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4195 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4196 SelectionDAG &DAG) const {
4197 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4198 SDValue V1 = SVOp->getOperand(0);
4199 SDValue V2 = SVOp->getOperand(1);
4200 DebugLoc dl = SVOp->getDebugLoc();
4201 SmallVector<int, 8> MaskVals;
4203 // Determine if more than 1 of the words in each of the low and high quadwords
4204 // of the result come from the same quadword of one of the two inputs. Undef
4205 // mask values count as coming from any quadword, for better codegen.
4206 SmallVector<unsigned, 4> LoQuad(4);
4207 SmallVector<unsigned, 4> HiQuad(4);
4208 BitVector InputQuads(4);
4209 for (unsigned i = 0; i < 8; ++i) {
4210 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4211 int EltIdx = SVOp->getMaskElt(i);
4212 MaskVals.push_back(EltIdx);
4221 InputQuads.set(EltIdx / 4);
4224 int BestLoQuad = -1;
4225 unsigned MaxQuad = 1;
4226 for (unsigned i = 0; i < 4; ++i) {
4227 if (LoQuad[i] > MaxQuad) {
4229 MaxQuad = LoQuad[i];
4233 int BestHiQuad = -1;
4235 for (unsigned i = 0; i < 4; ++i) {
4236 if (HiQuad[i] > MaxQuad) {
4238 MaxQuad = HiQuad[i];
4242 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4243 // of the two input vectors, shuffle them into one input vector so only a
4244 // single pshufb instruction is necessary. If There are more than 2 input
4245 // quads, disable the next transformation since it does not help SSSE3.
4246 bool V1Used = InputQuads[0] || InputQuads[1];
4247 bool V2Used = InputQuads[2] || InputQuads[3];
4248 if (Subtarget->hasSSSE3()) {
4249 if (InputQuads.count() == 2 && V1Used && V2Used) {
4250 BestLoQuad = InputQuads.find_first();
4251 BestHiQuad = InputQuads.find_next(BestLoQuad);
4253 if (InputQuads.count() > 2) {
4259 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4260 // the shuffle mask. If a quad is scored as -1, that means that it contains
4261 // words from all 4 input quadwords.
4263 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4264 SmallVector<int, 8> MaskV;
4265 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4266 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4267 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4268 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4269 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4270 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4272 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4273 // source words for the shuffle, to aid later transformations.
4274 bool AllWordsInNewV = true;
4275 bool InOrder[2] = { true, true };
4276 for (unsigned i = 0; i != 8; ++i) {
4277 int idx = MaskVals[i];
4279 InOrder[i/4] = false;
4280 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4282 AllWordsInNewV = false;
4286 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4287 if (AllWordsInNewV) {
4288 for (int i = 0; i != 8; ++i) {
4289 int idx = MaskVals[i];
4292 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4293 if ((idx != i) && idx < 4)
4295 if ((idx != i) && idx > 3)
4304 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4305 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4306 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4307 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4308 unsigned TargetMask = 0;
4309 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4310 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4311 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4312 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4313 V1 = NewV.getOperand(0);
4314 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4318 // If we have SSSE3, and all words of the result are from 1 input vector,
4319 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4320 // is present, fall back to case 4.
4321 if (Subtarget->hasSSSE3()) {
4322 SmallVector<SDValue,16> pshufbMask;
4324 // If we have elements from both input vectors, set the high bit of the
4325 // shuffle mask element to zero out elements that come from V2 in the V1
4326 // mask, and elements that come from V1 in the V2 mask, so that the two
4327 // results can be OR'd together.
4328 bool TwoInputs = V1Used && V2Used;
4329 for (unsigned i = 0; i != 8; ++i) {
4330 int EltIdx = MaskVals[i] * 2;
4331 if (TwoInputs && (EltIdx >= 16)) {
4332 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4333 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4336 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4337 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4339 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4340 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4341 DAG.getNode(ISD::BUILD_VECTOR, dl,
4342 MVT::v16i8, &pshufbMask[0], 16));
4344 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4346 // Calculate the shuffle mask for the second input, shuffle it, and
4347 // OR it with the first shuffled input.
4349 for (unsigned i = 0; i != 8; ++i) {
4350 int EltIdx = MaskVals[i] * 2;
4352 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4353 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4356 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4357 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4359 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4360 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4361 DAG.getNode(ISD::BUILD_VECTOR, dl,
4362 MVT::v16i8, &pshufbMask[0], 16));
4363 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4364 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4367 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4368 // and update MaskVals with new element order.
4369 BitVector InOrder(8);
4370 if (BestLoQuad >= 0) {
4371 SmallVector<int, 8> MaskV;
4372 for (int i = 0; i != 4; ++i) {
4373 int idx = MaskVals[i];
4375 MaskV.push_back(-1);
4377 } else if ((idx / 4) == BestLoQuad) {
4378 MaskV.push_back(idx & 3);
4381 MaskV.push_back(-1);
4384 for (unsigned i = 4; i != 8; ++i)
4386 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4389 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4390 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4392 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4396 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4397 // and update MaskVals with the new element order.
4398 if (BestHiQuad >= 0) {
4399 SmallVector<int, 8> MaskV;
4400 for (unsigned i = 0; i != 4; ++i)
4402 for (unsigned i = 4; i != 8; ++i) {
4403 int idx = MaskVals[i];
4405 MaskV.push_back(-1);
4407 } else if ((idx / 4) == BestHiQuad) {
4408 MaskV.push_back((idx & 3) + 4);
4411 MaskV.push_back(-1);
4414 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4417 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4418 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
4420 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
4424 // In case BestHi & BestLo were both -1, which means each quadword has a word
4425 // from each of the four input quadwords, calculate the InOrder bitvector now
4426 // before falling through to the insert/extract cleanup.
4427 if (BestLoQuad == -1 && BestHiQuad == -1) {
4429 for (int i = 0; i != 8; ++i)
4430 if (MaskVals[i] < 0 || MaskVals[i] == i)
4434 // The other elements are put in the right place using pextrw and pinsrw.
4435 for (unsigned i = 0; i != 8; ++i) {
4438 int EltIdx = MaskVals[i];
4441 SDValue ExtOp = (EltIdx < 8)
4442 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4443 DAG.getIntPtrConstant(EltIdx))
4444 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4445 DAG.getIntPtrConstant(EltIdx - 8));
4446 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4447 DAG.getIntPtrConstant(i));
4452 // v16i8 shuffles - Prefer shuffles in the following order:
4453 // 1. [ssse3] 1 x pshufb
4454 // 2. [ssse3] 2 x pshufb + 1 x por
4455 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4457 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4459 const X86TargetLowering &TLI) {
4460 SDValue V1 = SVOp->getOperand(0);
4461 SDValue V2 = SVOp->getOperand(1);
4462 DebugLoc dl = SVOp->getDebugLoc();
4463 SmallVector<int, 16> MaskVals;
4464 SVOp->getMask(MaskVals);
4466 // If we have SSSE3, case 1 is generated when all result bytes come from
4467 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4468 // present, fall back to case 3.
4469 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4472 for (unsigned i = 0; i < 16; ++i) {
4473 int EltIdx = MaskVals[i];
4482 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4483 if (TLI.getSubtarget()->hasSSSE3()) {
4484 SmallVector<SDValue,16> pshufbMask;
4486 // If all result elements are from one input vector, then only translate
4487 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4489 // Otherwise, we have elements from both input vectors, and must zero out
4490 // elements that come from V2 in the first mask, and V1 in the second mask
4491 // so that we can OR them together.
4492 bool TwoInputs = !(V1Only || V2Only);
4493 for (unsigned i = 0; i != 16; ++i) {
4494 int EltIdx = MaskVals[i];
4495 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4496 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4499 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4501 // If all the elements are from V2, assign it to V1 and return after
4502 // building the first pshufb.
4505 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4506 DAG.getNode(ISD::BUILD_VECTOR, dl,
4507 MVT::v16i8, &pshufbMask[0], 16));
4511 // Calculate the shuffle mask for the second input, shuffle it, and
4512 // OR it with the first shuffled input.
4514 for (unsigned i = 0; i != 16; ++i) {
4515 int EltIdx = MaskVals[i];
4517 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4520 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4522 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4523 DAG.getNode(ISD::BUILD_VECTOR, dl,
4524 MVT::v16i8, &pshufbMask[0], 16));
4525 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4528 // No SSSE3 - Calculate in place words and then fix all out of place words
4529 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4530 // the 16 different words that comprise the two doublequadword input vectors.
4531 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4532 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4533 SDValue NewV = V2Only ? V2 : V1;
4534 for (int i = 0; i != 8; ++i) {
4535 int Elt0 = MaskVals[i*2];
4536 int Elt1 = MaskVals[i*2+1];
4538 // This word of the result is all undef, skip it.
4539 if (Elt0 < 0 && Elt1 < 0)
4542 // This word of the result is already in the correct place, skip it.
4543 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4545 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4548 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4549 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4552 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4553 // using a single extract together, load it and store it.
4554 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4555 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4556 DAG.getIntPtrConstant(Elt1 / 2));
4557 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4558 DAG.getIntPtrConstant(i));
4562 // If Elt1 is defined, extract it from the appropriate source. If the
4563 // source byte is not also odd, shift the extracted word left 8 bits
4564 // otherwise clear the bottom 8 bits if we need to do an or.
4566 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4567 DAG.getIntPtrConstant(Elt1 / 2));
4568 if ((Elt1 & 1) == 0)
4569 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4570 DAG.getConstant(8, TLI.getShiftAmountTy()));
4572 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4573 DAG.getConstant(0xFF00, MVT::i16));
4575 // If Elt0 is defined, extract it from the appropriate source. If the
4576 // source byte is not also even, shift the extracted word right 8 bits. If
4577 // Elt1 was also defined, OR the extracted values together before
4578 // inserting them in the result.
4580 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4581 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4582 if ((Elt0 & 1) != 0)
4583 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4584 DAG.getConstant(8, TLI.getShiftAmountTy()));
4586 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4587 DAG.getConstant(0x00FF, MVT::i16));
4588 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4591 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4592 DAG.getIntPtrConstant(i));
4594 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4597 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4598 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4599 /// done when every pair / quad of shuffle mask elements point to elements in
4600 /// the right sequence. e.g.
4601 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4603 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4605 const TargetLowering &TLI, DebugLoc dl) {
4606 EVT VT = SVOp->getValueType(0);
4607 SDValue V1 = SVOp->getOperand(0);
4608 SDValue V2 = SVOp->getOperand(1);
4609 unsigned NumElems = VT.getVectorNumElements();
4610 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4611 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4613 switch (VT.getSimpleVT().SimpleTy) {
4614 default: assert(false && "Unexpected!");
4615 case MVT::v4f32: NewVT = MVT::v2f64; break;
4616 case MVT::v4i32: NewVT = MVT::v2i64; break;
4617 case MVT::v8i16: NewVT = MVT::v4i32; break;
4618 case MVT::v16i8: NewVT = MVT::v4i32; break;
4621 if (NewWidth == 2) {
4627 int Scale = NumElems / NewWidth;
4628 SmallVector<int, 8> MaskVec;
4629 for (unsigned i = 0; i < NumElems; i += Scale) {
4631 for (int j = 0; j < Scale; ++j) {
4632 int EltIdx = SVOp->getMaskElt(i+j);
4636 StartIdx = EltIdx - (EltIdx % Scale);
4637 if (EltIdx != StartIdx + j)
4641 MaskVec.push_back(-1);
4643 MaskVec.push_back(StartIdx / Scale);
4646 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4647 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4648 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4651 /// getVZextMovL - Return a zero-extending vector move low node.
4653 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4654 SDValue SrcOp, SelectionDAG &DAG,
4655 const X86Subtarget *Subtarget, DebugLoc dl) {
4656 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4657 LoadSDNode *LD = NULL;
4658 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4659 LD = dyn_cast<LoadSDNode>(SrcOp);
4661 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4663 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4664 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4665 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4666 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4667 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4669 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4670 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4671 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4672 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4680 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4681 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4682 DAG.getNode(ISD::BIT_CONVERT, dl,
4686 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4689 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4690 SDValue V1 = SVOp->getOperand(0);
4691 SDValue V2 = SVOp->getOperand(1);
4692 DebugLoc dl = SVOp->getDebugLoc();
4693 EVT VT = SVOp->getValueType(0);
4695 SmallVector<std::pair<int, int>, 8> Locs;
4697 SmallVector<int, 8> Mask1(4U, -1);
4698 SmallVector<int, 8> PermMask;
4699 SVOp->getMask(PermMask);
4703 for (unsigned i = 0; i != 4; ++i) {
4704 int Idx = PermMask[i];
4706 Locs[i] = std::make_pair(-1, -1);
4708 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4710 Locs[i] = std::make_pair(0, NumLo);
4714 Locs[i] = std::make_pair(1, NumHi);
4716 Mask1[2+NumHi] = Idx;
4722 if (NumLo <= 2 && NumHi <= 2) {
4723 // If no more than two elements come from either vector. This can be
4724 // implemented with two shuffles. First shuffle gather the elements.
4725 // The second shuffle, which takes the first shuffle as both of its
4726 // vector operands, put the elements into the right order.
4727 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4729 SmallVector<int, 8> Mask2(4U, -1);
4731 for (unsigned i = 0; i != 4; ++i) {
4732 if (Locs[i].first == -1)
4735 unsigned Idx = (i < 2) ? 0 : 4;
4736 Idx += Locs[i].first * 2 + Locs[i].second;
4741 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4742 } else if (NumLo == 3 || NumHi == 3) {
4743 // Otherwise, we must have three elements from one vector, call it X, and
4744 // one element from the other, call it Y. First, use a shufps to build an
4745 // intermediate vector with the one element from Y and the element from X
4746 // that will be in the same half in the final destination (the indexes don't
4747 // matter). Then, use a shufps to build the final vector, taking the half
4748 // containing the element from Y from the intermediate, and the other half
4751 // Normalize it so the 3 elements come from V1.
4752 CommuteVectorShuffleMask(PermMask, VT);
4756 // Find the element from V2.
4758 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4759 int Val = PermMask[HiIndex];
4766 Mask1[0] = PermMask[HiIndex];
4768 Mask1[2] = PermMask[HiIndex^1];
4770 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4773 Mask1[0] = PermMask[0];
4774 Mask1[1] = PermMask[1];
4775 Mask1[2] = HiIndex & 1 ? 6 : 4;
4776 Mask1[3] = HiIndex & 1 ? 4 : 6;
4777 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4779 Mask1[0] = HiIndex & 1 ? 2 : 0;
4780 Mask1[1] = HiIndex & 1 ? 0 : 2;
4781 Mask1[2] = PermMask[2];
4782 Mask1[3] = PermMask[3];
4787 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4791 // Break it into (shuffle shuffle_hi, shuffle_lo).
4793 SmallVector<int,8> LoMask(4U, -1);
4794 SmallVector<int,8> HiMask(4U, -1);
4796 SmallVector<int,8> *MaskPtr = &LoMask;
4797 unsigned MaskIdx = 0;
4800 for (unsigned i = 0; i != 4; ++i) {
4807 int Idx = PermMask[i];
4809 Locs[i] = std::make_pair(-1, -1);
4810 } else if (Idx < 4) {
4811 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4812 (*MaskPtr)[LoIdx] = Idx;
4815 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4816 (*MaskPtr)[HiIdx] = Idx;
4821 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4822 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4823 SmallVector<int, 8> MaskOps;
4824 for (unsigned i = 0; i != 4; ++i) {
4825 if (Locs[i].first == -1) {
4826 MaskOps.push_back(-1);
4828 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4829 MaskOps.push_back(Idx);
4832 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4836 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4837 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4838 SDValue V1 = Op.getOperand(0);
4839 SDValue V2 = Op.getOperand(1);
4840 EVT VT = Op.getValueType();
4841 DebugLoc dl = Op.getDebugLoc();
4842 unsigned NumElems = VT.getVectorNumElements();
4843 bool isMMX = VT.getSizeInBits() == 64;
4844 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4845 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4846 bool V1IsSplat = false;
4847 bool V2IsSplat = false;
4848 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
4849 MachineFunction &MF = DAG.getMachineFunction();
4850 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
4852 if (isZeroShuffle(SVOp))
4853 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4855 // Promote splats to v4f32.
4856 if (SVOp->isSplat()) {
4857 if (isMMX || NumElems < 4)
4859 return PromoteSplat(SVOp, DAG);
4862 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4864 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4865 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4866 if (NewOp.getNode())
4867 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4868 LowerVECTOR_SHUFFLE(NewOp, DAG));
4869 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4870 // FIXME: Figure out a cleaner way to do this.
4871 // Try to make use of movq to zero out the top part.
4872 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4873 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4874 if (NewOp.getNode()) {
4875 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4876 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4877 DAG, Subtarget, dl);
4879 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4880 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4881 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4882 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4883 DAG, Subtarget, dl);
4887 if (X86::isPSHUFDMask(SVOp)) {
4888 // The actual implementation will match the mask in the if above and then
4889 // during isel it can match several different instructions, not only pshufd
4890 // as its name says, sad but true, emulate the behavior for now...
4891 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
4892 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
4894 if (OptForSize && HasSSE2 && X86::isUNPCKL_v_undef_Mask(SVOp) &&
4896 return getTargetShuffleNode(X86ISD::PUNPCKLDQ, dl, VT, V1, V1, DAG);
4898 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
4900 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
4901 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
4903 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
4904 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
4907 if (VT == MVT::v4f32)
4908 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
4912 // Check if this can be converted into a logical shift.
4913 bool isLeft = false;
4916 bool isShift = getSubtarget()->hasSSE2() &&
4917 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4918 if (isShift && ShVal.hasOneUse()) {
4919 // If the shifted value has multiple uses, it may be cheaper to use
4920 // v_set0 + movlhps or movhlps, etc.
4921 EVT EltVT = VT.getVectorElementType();
4922 ShAmt *= EltVT.getSizeInBits();
4923 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4926 if (X86::isMOVLMask(SVOp)) {
4929 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4930 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4935 // FIXME: fold these into legal mask.
4936 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4937 X86::isMOVSLDUPMask(SVOp) ||
4938 X86::isMOVHLPSMask(SVOp) ||
4939 X86::isMOVLHPSMask(SVOp) ||
4940 X86::isMOVLPMask(SVOp)))
4943 if (ShouldXformToMOVHLPS(SVOp) ||
4944 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4945 return CommuteVectorShuffle(SVOp, DAG);
4948 // No better options. Use a vshl / vsrl.
4949 EVT EltVT = VT.getVectorElementType();
4950 ShAmt *= EltVT.getSizeInBits();
4951 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4954 bool Commuted = false;
4955 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4956 // 1,1,1,1 -> v8i16 though.
4957 V1IsSplat = isSplatVector(V1.getNode());
4958 V2IsSplat = isSplatVector(V2.getNode());
4960 // Canonicalize the splat or undef, if present, to be on the RHS.
4961 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4962 Op = CommuteVectorShuffle(SVOp, DAG);
4963 SVOp = cast<ShuffleVectorSDNode>(Op);
4964 V1 = SVOp->getOperand(0);
4965 V2 = SVOp->getOperand(1);
4966 std::swap(V1IsSplat, V2IsSplat);
4967 std::swap(V1IsUndef, V2IsUndef);
4971 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4972 // Shuffling low element of v1 into undef, just return v1.
4975 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4976 // the instruction selector will not match, so get a canonical MOVL with
4977 // swapped operands to undo the commute.
4978 return getMOVL(DAG, dl, VT, V2, V1);
4981 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4982 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4983 X86::isUNPCKLMask(SVOp) ||
4984 X86::isUNPCKHMask(SVOp))
4988 // Normalize mask so all entries that point to V2 points to its first
4989 // element then try to match unpck{h|l} again. If match, return a
4990 // new vector_shuffle with the corrected mask.
4991 SDValue NewMask = NormalizeMask(SVOp, DAG);
4992 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4993 if (NSVOp != SVOp) {
4994 if (X86::isUNPCKLMask(NSVOp, true)) {
4996 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5003 // Commute is back and try unpck* again.
5004 // FIXME: this seems wrong.
5005 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5006 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5007 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
5008 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
5009 X86::isUNPCKLMask(NewSVOp) ||
5010 X86::isUNPCKHMask(NewSVOp))
5014 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
5016 // Normalize the node to match x86 shuffle ops if needed
5017 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5018 return CommuteVectorShuffle(SVOp, DAG);
5020 // Check for legal shuffle and return?
5021 SmallVector<int, 16> PermMask;
5022 SVOp->getMask(PermMask);
5023 if (isShuffleMaskLegal(PermMask, VT))
5026 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5027 if (VT == MVT::v8i16) {
5028 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5029 if (NewOp.getNode())
5033 if (VT == MVT::v16i8) {
5034 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5035 if (NewOp.getNode())
5039 // Handle all 4 wide cases with a number of shuffles except for MMX.
5040 if (NumElems == 4 && !isMMX)
5041 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5047 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5048 SelectionDAG &DAG) const {
5049 EVT VT = Op.getValueType();
5050 DebugLoc dl = Op.getDebugLoc();
5051 if (VT.getSizeInBits() == 8) {
5052 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
5053 Op.getOperand(0), Op.getOperand(1));
5054 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5055 DAG.getValueType(VT));
5056 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5057 } else if (VT.getSizeInBits() == 16) {
5058 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5059 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
5061 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5062 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5063 DAG.getNode(ISD::BIT_CONVERT, dl,
5067 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
5068 Op.getOperand(0), Op.getOperand(1));
5069 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5070 DAG.getValueType(VT));
5071 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5072 } else if (VT == MVT::f32) {
5073 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
5074 // the result back to FR32 register. It's only worth matching if the
5075 // result has a single use which is a store or a bitcast to i32. And in
5076 // the case of a store, it's not worth it if the index is a constant 0,
5077 // because a MOVSSmr can be used instead, which is smaller and faster.
5078 if (!Op.hasOneUse())
5080 SDNode *User = *Op.getNode()->use_begin();
5081 if ((User->getOpcode() != ISD::STORE ||
5082 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5083 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5084 (User->getOpcode() != ISD::BIT_CONVERT ||
5085 User->getValueType(0) != MVT::i32))
5087 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5088 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
5091 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
5092 } else if (VT == MVT::i32) {
5093 // ExtractPS works with constant index.
5094 if (isa<ConstantSDNode>(Op.getOperand(1)))
5102 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5103 SelectionDAG &DAG) const {
5104 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5107 if (Subtarget->hasSSE41()) {
5108 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5113 EVT VT = Op.getValueType();
5114 DebugLoc dl = Op.getDebugLoc();
5115 // TODO: handle v16i8.
5116 if (VT.getSizeInBits() == 16) {
5117 SDValue Vec = Op.getOperand(0);
5118 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5120 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5121 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5122 DAG.getNode(ISD::BIT_CONVERT, dl,
5125 // Transform it so it match pextrw which produces a 32-bit result.
5126 EVT EltVT = MVT::i32;
5127 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5128 Op.getOperand(0), Op.getOperand(1));
5129 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5130 DAG.getValueType(VT));
5131 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5132 } else if (VT.getSizeInBits() == 32) {
5133 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5137 // SHUFPS the element to the lowest double word, then movss.
5138 int Mask[4] = { Idx, -1, -1, -1 };
5139 EVT VVT = Op.getOperand(0).getValueType();
5140 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5141 DAG.getUNDEF(VVT), Mask);
5142 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5143 DAG.getIntPtrConstant(0));
5144 } else if (VT.getSizeInBits() == 64) {
5145 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5146 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5147 // to match extract_elt for f64.
5148 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5152 // UNPCKHPD the element to the lowest double word, then movsd.
5153 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5154 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5155 int Mask[2] = { 1, -1 };
5156 EVT VVT = Op.getOperand(0).getValueType();
5157 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5158 DAG.getUNDEF(VVT), Mask);
5159 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5160 DAG.getIntPtrConstant(0));
5167 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5168 SelectionDAG &DAG) const {
5169 EVT VT = Op.getValueType();
5170 EVT EltVT = VT.getVectorElementType();
5171 DebugLoc dl = Op.getDebugLoc();
5173 SDValue N0 = Op.getOperand(0);
5174 SDValue N1 = Op.getOperand(1);
5175 SDValue N2 = Op.getOperand(2);
5177 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5178 isa<ConstantSDNode>(N2)) {
5180 if (VT == MVT::v8i16)
5181 Opc = X86ISD::PINSRW;
5182 else if (VT == MVT::v4i16)
5183 Opc = X86ISD::MMX_PINSRW;
5184 else if (VT == MVT::v16i8)
5185 Opc = X86ISD::PINSRB;
5187 Opc = X86ISD::PINSRB;
5189 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5191 if (N1.getValueType() != MVT::i32)
5192 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5193 if (N2.getValueType() != MVT::i32)
5194 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5195 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5196 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5197 // Bits [7:6] of the constant are the source select. This will always be
5198 // zero here. The DAG Combiner may combine an extract_elt index into these
5199 // bits. For example (insert (extract, 3), 2) could be matched by putting
5200 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5201 // Bits [5:4] of the constant are the destination select. This is the
5202 // value of the incoming immediate.
5203 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5204 // combine either bitwise AND or insert of float 0.0 to set these bits.
5205 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5206 // Create this as a scalar to vector..
5207 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5208 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5209 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5210 // PINSR* works with constant index.
5217 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5218 EVT VT = Op.getValueType();
5219 EVT EltVT = VT.getVectorElementType();
5221 if (Subtarget->hasSSE41())
5222 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5224 if (EltVT == MVT::i8)
5227 DebugLoc dl = Op.getDebugLoc();
5228 SDValue N0 = Op.getOperand(0);
5229 SDValue N1 = Op.getOperand(1);
5230 SDValue N2 = Op.getOperand(2);
5232 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5233 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5234 // as its second argument.
5235 if (N1.getValueType() != MVT::i32)
5236 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5237 if (N2.getValueType() != MVT::i32)
5238 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5239 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5240 dl, VT, N0, N1, N2);
5246 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5247 DebugLoc dl = Op.getDebugLoc();
5249 if (Op.getValueType() == MVT::v1i64 &&
5250 Op.getOperand(0).getValueType() == MVT::i64)
5251 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5253 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5254 EVT VT = MVT::v2i32;
5255 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5262 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5263 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5266 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5267 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5268 // one of the above mentioned nodes. It has to be wrapped because otherwise
5269 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5270 // be used to form addressing mode. These wrapped nodes will be selected
5273 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5274 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5276 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5278 unsigned char OpFlag = 0;
5279 unsigned WrapperKind = X86ISD::Wrapper;
5280 CodeModel::Model M = getTargetMachine().getCodeModel();
5282 if (Subtarget->isPICStyleRIPRel() &&
5283 (M == CodeModel::Small || M == CodeModel::Kernel))
5284 WrapperKind = X86ISD::WrapperRIP;
5285 else if (Subtarget->isPICStyleGOT())
5286 OpFlag = X86II::MO_GOTOFF;
5287 else if (Subtarget->isPICStyleStubPIC())
5288 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5290 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5292 CP->getOffset(), OpFlag);
5293 DebugLoc DL = CP->getDebugLoc();
5294 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5295 // With PIC, the address is actually $g + Offset.
5297 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5298 DAG.getNode(X86ISD::GlobalBaseReg,
5299 DebugLoc(), getPointerTy()),
5306 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5307 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5309 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5311 unsigned char OpFlag = 0;
5312 unsigned WrapperKind = X86ISD::Wrapper;
5313 CodeModel::Model M = getTargetMachine().getCodeModel();
5315 if (Subtarget->isPICStyleRIPRel() &&
5316 (M == CodeModel::Small || M == CodeModel::Kernel))
5317 WrapperKind = X86ISD::WrapperRIP;
5318 else if (Subtarget->isPICStyleGOT())
5319 OpFlag = X86II::MO_GOTOFF;
5320 else if (Subtarget->isPICStyleStubPIC())
5321 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5323 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5325 DebugLoc DL = JT->getDebugLoc();
5326 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5328 // With PIC, the address is actually $g + Offset.
5330 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5331 DAG.getNode(X86ISD::GlobalBaseReg,
5332 DebugLoc(), getPointerTy()),
5340 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5341 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5343 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5345 unsigned char OpFlag = 0;
5346 unsigned WrapperKind = X86ISD::Wrapper;
5347 CodeModel::Model M = getTargetMachine().getCodeModel();
5349 if (Subtarget->isPICStyleRIPRel() &&
5350 (M == CodeModel::Small || M == CodeModel::Kernel))
5351 WrapperKind = X86ISD::WrapperRIP;
5352 else if (Subtarget->isPICStyleGOT())
5353 OpFlag = X86II::MO_GOTOFF;
5354 else if (Subtarget->isPICStyleStubPIC())
5355 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5357 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5359 DebugLoc DL = Op.getDebugLoc();
5360 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5363 // With PIC, the address is actually $g + Offset.
5364 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5365 !Subtarget->is64Bit()) {
5366 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5367 DAG.getNode(X86ISD::GlobalBaseReg,
5368 DebugLoc(), getPointerTy()),
5376 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5377 // Create the TargetBlockAddressAddress node.
5378 unsigned char OpFlags =
5379 Subtarget->ClassifyBlockAddressReference();
5380 CodeModel::Model M = getTargetMachine().getCodeModel();
5381 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5382 DebugLoc dl = Op.getDebugLoc();
5383 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5384 /*isTarget=*/true, OpFlags);
5386 if (Subtarget->isPICStyleRIPRel() &&
5387 (M == CodeModel::Small || M == CodeModel::Kernel))
5388 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5390 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5392 // With PIC, the address is actually $g + Offset.
5393 if (isGlobalRelativeToPICBase(OpFlags)) {
5394 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5395 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5403 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5405 SelectionDAG &DAG) const {
5406 // Create the TargetGlobalAddress node, folding in the constant
5407 // offset if it is legal.
5408 unsigned char OpFlags =
5409 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5410 CodeModel::Model M = getTargetMachine().getCodeModel();
5412 if (OpFlags == X86II::MO_NO_FLAG &&
5413 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5414 // A direct static reference to a global.
5415 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5418 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5421 if (Subtarget->isPICStyleRIPRel() &&
5422 (M == CodeModel::Small || M == CodeModel::Kernel))
5423 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5425 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5427 // With PIC, the address is actually $g + Offset.
5428 if (isGlobalRelativeToPICBase(OpFlags)) {
5429 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5430 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5434 // For globals that require a load from a stub to get the address, emit the
5436 if (isGlobalStubReference(OpFlags))
5437 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5438 PseudoSourceValue::getGOT(), 0, false, false, 0);
5440 // If there was a non-zero offset that we didn't fold, create an explicit
5443 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5444 DAG.getConstant(Offset, getPointerTy()));
5450 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5451 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5452 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5453 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5457 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5458 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5459 unsigned char OperandFlags) {
5460 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5461 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5462 DebugLoc dl = GA->getDebugLoc();
5463 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5464 GA->getValueType(0),
5468 SDValue Ops[] = { Chain, TGA, *InFlag };
5469 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5471 SDValue Ops[] = { Chain, TGA };
5472 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5475 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5476 MFI->setAdjustsStack(true);
5478 SDValue Flag = Chain.getValue(1);
5479 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5482 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5484 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5487 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5488 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5489 DAG.getNode(X86ISD::GlobalBaseReg,
5490 DebugLoc(), PtrVT), InFlag);
5491 InFlag = Chain.getValue(1);
5493 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5496 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5498 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5500 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5501 X86::RAX, X86II::MO_TLSGD);
5504 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5505 // "local exec" model.
5506 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5507 const EVT PtrVT, TLSModel::Model model,
5509 DebugLoc dl = GA->getDebugLoc();
5510 // Get the Thread Pointer
5511 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5513 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5516 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5517 NULL, 0, false, false, 0);
5519 unsigned char OperandFlags = 0;
5520 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5522 unsigned WrapperKind = X86ISD::Wrapper;
5523 if (model == TLSModel::LocalExec) {
5524 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5525 } else if (is64Bit) {
5526 assert(model == TLSModel::InitialExec);
5527 OperandFlags = X86II::MO_GOTTPOFF;
5528 WrapperKind = X86ISD::WrapperRIP;
5530 assert(model == TLSModel::InitialExec);
5531 OperandFlags = X86II::MO_INDNTPOFF;
5534 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5536 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5537 GA->getValueType(0),
5538 GA->getOffset(), OperandFlags);
5539 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5541 if (model == TLSModel::InitialExec)
5542 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5543 PseudoSourceValue::getGOT(), 0, false, false, 0);
5545 // The address of the thread local variable is the add of the thread
5546 // pointer with the offset of the variable.
5547 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5551 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5553 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5554 const GlobalValue *GV = GA->getGlobal();
5556 if (Subtarget->isTargetELF()) {
5557 // TODO: implement the "local dynamic" model
5558 // TODO: implement the "initial exec"model for pic executables
5560 // If GV is an alias then use the aliasee for determining
5561 // thread-localness.
5562 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5563 GV = GA->resolveAliasedGlobal(false);
5565 TLSModel::Model model
5566 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5569 case TLSModel::GeneralDynamic:
5570 case TLSModel::LocalDynamic: // not implemented
5571 if (Subtarget->is64Bit())
5572 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5573 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5575 case TLSModel::InitialExec:
5576 case TLSModel::LocalExec:
5577 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5578 Subtarget->is64Bit());
5580 } else if (Subtarget->isTargetDarwin()) {
5581 // Darwin only has one model of TLS. Lower to that.
5582 unsigned char OpFlag = 0;
5583 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5584 X86ISD::WrapperRIP : X86ISD::Wrapper;
5586 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5588 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5589 !Subtarget->is64Bit();
5591 OpFlag = X86II::MO_TLVP_PIC_BASE;
5593 OpFlag = X86II::MO_TLVP;
5594 DebugLoc DL = Op.getDebugLoc();
5595 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
5597 GA->getOffset(), OpFlag);
5598 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5600 // With PIC32, the address is actually $g + Offset.
5602 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5603 DAG.getNode(X86ISD::GlobalBaseReg,
5604 DebugLoc(), getPointerTy()),
5607 // Lowering the machine isd will make sure everything is in the right
5609 SDValue Args[] = { Offset };
5610 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5612 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5613 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5614 MFI->setAdjustsStack(true);
5616 // And our return value (tls address) is in the standard call return value
5618 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5619 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5623 "TLS not implemented for this target.");
5625 llvm_unreachable("Unreachable");
5630 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5631 /// take a 2 x i32 value to shift plus a shift amount.
5632 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5633 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5634 EVT VT = Op.getValueType();
5635 unsigned VTBits = VT.getSizeInBits();
5636 DebugLoc dl = Op.getDebugLoc();
5637 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5638 SDValue ShOpLo = Op.getOperand(0);
5639 SDValue ShOpHi = Op.getOperand(1);
5640 SDValue ShAmt = Op.getOperand(2);
5641 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5642 DAG.getConstant(VTBits - 1, MVT::i8))
5643 : DAG.getConstant(0, VT);
5646 if (Op.getOpcode() == ISD::SHL_PARTS) {
5647 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5648 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5650 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5651 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5654 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5655 DAG.getConstant(VTBits, MVT::i8));
5656 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5657 AndNode, DAG.getConstant(0, MVT::i8));
5660 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5661 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5662 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5664 if (Op.getOpcode() == ISD::SHL_PARTS) {
5665 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5666 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5668 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5669 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5672 SDValue Ops[2] = { Lo, Hi };
5673 return DAG.getMergeValues(Ops, 2, dl);
5676 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5677 SelectionDAG &DAG) const {
5678 EVT SrcVT = Op.getOperand(0).getValueType();
5680 if (SrcVT.isVector()) {
5681 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5687 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5688 "Unknown SINT_TO_FP to lower!");
5690 // These are really Legal; return the operand so the caller accepts it as
5692 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5694 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5695 Subtarget->is64Bit()) {
5699 DebugLoc dl = Op.getDebugLoc();
5700 unsigned Size = SrcVT.getSizeInBits()/8;
5701 MachineFunction &MF = DAG.getMachineFunction();
5702 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5703 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5704 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5706 PseudoSourceValue::getFixedStack(SSFI), 0,
5708 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5711 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5713 SelectionDAG &DAG) const {
5715 DebugLoc dl = Op.getDebugLoc();
5717 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5719 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5721 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5722 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5723 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5724 Tys, Ops, array_lengthof(Ops));
5727 Chain = Result.getValue(1);
5728 SDValue InFlag = Result.getValue(2);
5730 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5731 // shouldn't be necessary except that RFP cannot be live across
5732 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5733 MachineFunction &MF = DAG.getMachineFunction();
5734 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5735 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5736 Tys = DAG.getVTList(MVT::Other);
5738 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5740 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5741 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5742 PseudoSourceValue::getFixedStack(SSFI), 0,
5749 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5750 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5751 SelectionDAG &DAG) const {
5752 // This algorithm is not obvious. Here it is in C code, more or less:
5754 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5755 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5756 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5758 // Copy ints to xmm registers.
5759 __m128i xh = _mm_cvtsi32_si128( hi );
5760 __m128i xl = _mm_cvtsi32_si128( lo );
5762 // Combine into low half of a single xmm register.
5763 __m128i x = _mm_unpacklo_epi32( xh, xl );
5767 // Merge in appropriate exponents to give the integer bits the right
5769 x = _mm_unpacklo_epi32( x, exp );
5771 // Subtract away the biases to deal with the IEEE-754 double precision
5773 d = _mm_sub_pd( (__m128d) x, bias );
5775 // All conversions up to here are exact. The correctly rounded result is
5776 // calculated using the current rounding mode using the following
5778 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5779 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5780 // store doesn't really need to be here (except
5781 // maybe to zero the other double)
5786 DebugLoc dl = Op.getDebugLoc();
5787 LLVMContext *Context = DAG.getContext();
5789 // Build some magic constants.
5790 std::vector<Constant*> CV0;
5791 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5792 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5793 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5794 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5795 Constant *C0 = ConstantVector::get(CV0);
5796 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5798 std::vector<Constant*> CV1;
5800 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5802 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5803 Constant *C1 = ConstantVector::get(CV1);
5804 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5806 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5807 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5809 DAG.getIntPtrConstant(1)));
5810 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5811 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5813 DAG.getIntPtrConstant(0)));
5814 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5815 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5816 PseudoSourceValue::getConstantPool(), 0,
5818 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5819 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5820 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5821 PseudoSourceValue::getConstantPool(), 0,
5823 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5825 // Add the halves; easiest way is to swap them into another reg first.
5826 int ShufMask[2] = { 1, -1 };
5827 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5828 DAG.getUNDEF(MVT::v2f64), ShufMask);
5829 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5830 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5831 DAG.getIntPtrConstant(0));
5834 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5835 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5836 SelectionDAG &DAG) const {
5837 DebugLoc dl = Op.getDebugLoc();
5838 // FP constant to bias correct the final result.
5839 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5842 // Load the 32-bit value into an XMM register.
5843 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5844 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5846 DAG.getIntPtrConstant(0)));
5848 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5849 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5850 DAG.getIntPtrConstant(0));
5852 // Or the load with the bias.
5853 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5854 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5855 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5857 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5858 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5859 MVT::v2f64, Bias)));
5860 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5861 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5862 DAG.getIntPtrConstant(0));
5864 // Subtract the bias.
5865 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5867 // Handle final rounding.
5868 EVT DestVT = Op.getValueType();
5870 if (DestVT.bitsLT(MVT::f64)) {
5871 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5872 DAG.getIntPtrConstant(0));
5873 } else if (DestVT.bitsGT(MVT::f64)) {
5874 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5877 // Handle final rounding.
5881 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5882 SelectionDAG &DAG) const {
5883 SDValue N0 = Op.getOperand(0);
5884 DebugLoc dl = Op.getDebugLoc();
5886 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5887 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5888 // the optimization here.
5889 if (DAG.SignBitIsZero(N0))
5890 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5892 EVT SrcVT = N0.getValueType();
5893 EVT DstVT = Op.getValueType();
5894 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5895 return LowerUINT_TO_FP_i64(Op, DAG);
5896 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5897 return LowerUINT_TO_FP_i32(Op, DAG);
5899 // Make a 64-bit buffer, and use it to build an FILD.
5900 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5901 if (SrcVT == MVT::i32) {
5902 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5903 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5904 getPointerTy(), StackSlot, WordOff);
5905 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5906 StackSlot, NULL, 0, false, false, 0);
5907 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5908 OffsetSlot, NULL, 0, false, false, 0);
5909 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5913 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5914 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5915 StackSlot, NULL, 0, false, false, 0);
5916 // For i64 source, we need to add the appropriate power of 2 if the input
5917 // was negative. This is the same as the optimization in
5918 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5919 // we must be careful to do the computation in x87 extended precision, not
5920 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5921 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5922 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5923 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5925 APInt FF(32, 0x5F800000ULL);
5927 // Check whether the sign bit is set.
5928 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5929 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5932 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5933 SDValue FudgePtr = DAG.getConstantPool(
5934 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5937 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5938 SDValue Zero = DAG.getIntPtrConstant(0);
5939 SDValue Four = DAG.getIntPtrConstant(4);
5940 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5942 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5944 // Load the value out, extending it from f32 to f80.
5945 // FIXME: Avoid the extend by constructing the right constant pool?
5946 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
5947 FudgePtr, PseudoSourceValue::getConstantPool(),
5948 0, MVT::f32, false, false, 4);
5949 // Extend everything to 80 bits to force it to be done on x87.
5950 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5951 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5954 std::pair<SDValue,SDValue> X86TargetLowering::
5955 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5956 DebugLoc dl = Op.getDebugLoc();
5958 EVT DstTy = Op.getValueType();
5961 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5965 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5966 DstTy.getSimpleVT() >= MVT::i16 &&
5967 "Unknown FP_TO_SINT to lower!");
5969 // These are really Legal.
5970 if (DstTy == MVT::i32 &&
5971 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5972 return std::make_pair(SDValue(), SDValue());
5973 if (Subtarget->is64Bit() &&
5974 DstTy == MVT::i64 &&
5975 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5976 return std::make_pair(SDValue(), SDValue());
5978 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5980 MachineFunction &MF = DAG.getMachineFunction();
5981 unsigned MemSize = DstTy.getSizeInBits()/8;
5982 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5983 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5986 switch (DstTy.getSimpleVT().SimpleTy) {
5987 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5988 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5989 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5990 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5993 SDValue Chain = DAG.getEntryNode();
5994 SDValue Value = Op.getOperand(0);
5995 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5996 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5997 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5998 PseudoSourceValue::getFixedStack(SSFI), 0,
6000 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
6002 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
6004 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
6005 Chain = Value.getValue(1);
6006 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6007 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6010 // Build the FP_TO_INT*_IN_MEM
6011 SDValue Ops[] = { Chain, Value, StackSlot };
6012 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
6014 return std::make_pair(FIST, StackSlot);
6017 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
6018 SelectionDAG &DAG) const {
6019 if (Op.getValueType().isVector()) {
6020 if (Op.getValueType() == MVT::v2i32 &&
6021 Op.getOperand(0).getValueType() == MVT::v2f64) {
6027 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
6028 SDValue FIST = Vals.first, StackSlot = Vals.second;
6029 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
6030 if (FIST.getNode() == 0) return Op;
6033 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6034 FIST, StackSlot, NULL, 0, false, false, 0);
6037 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
6038 SelectionDAG &DAG) const {
6039 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
6040 SDValue FIST = Vals.first, StackSlot = Vals.second;
6041 assert(FIST.getNode() && "Unexpected failure");
6044 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6045 FIST, StackSlot, NULL, 0, false, false, 0);
6048 SDValue X86TargetLowering::LowerFABS(SDValue Op,
6049 SelectionDAG &DAG) const {
6050 LLVMContext *Context = DAG.getContext();
6051 DebugLoc dl = Op.getDebugLoc();
6052 EVT VT = Op.getValueType();
6055 EltVT = VT.getVectorElementType();
6056 std::vector<Constant*> CV;
6057 if (EltVT == MVT::f64) {
6058 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
6062 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
6068 Constant *C = ConstantVector::get(CV);
6069 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6070 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6071 PseudoSourceValue::getConstantPool(), 0,
6073 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
6076 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
6077 LLVMContext *Context = DAG.getContext();
6078 DebugLoc dl = Op.getDebugLoc();
6079 EVT VT = Op.getValueType();
6082 EltVT = VT.getVectorElementType();
6083 std::vector<Constant*> CV;
6084 if (EltVT == MVT::f64) {
6085 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
6089 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
6095 Constant *C = ConstantVector::get(CV);
6096 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6097 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6098 PseudoSourceValue::getConstantPool(), 0,
6100 if (VT.isVector()) {
6101 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6102 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6103 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6105 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
6107 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6111 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6112 LLVMContext *Context = DAG.getContext();
6113 SDValue Op0 = Op.getOperand(0);
6114 SDValue Op1 = Op.getOperand(1);
6115 DebugLoc dl = Op.getDebugLoc();
6116 EVT VT = Op.getValueType();
6117 EVT SrcVT = Op1.getValueType();
6119 // If second operand is smaller, extend it first.
6120 if (SrcVT.bitsLT(VT)) {
6121 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6124 // And if it is bigger, shrink it first.
6125 if (SrcVT.bitsGT(VT)) {
6126 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6130 // At this point the operands and the result should have the same
6131 // type, and that won't be f80 since that is not custom lowered.
6133 // First get the sign bit of second operand.
6134 std::vector<Constant*> CV;
6135 if (SrcVT == MVT::f64) {
6136 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6137 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6139 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6140 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6141 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6142 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6144 Constant *C = ConstantVector::get(CV);
6145 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6146 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6147 PseudoSourceValue::getConstantPool(), 0,
6149 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6151 // Shift sign bit right or left if the two operands have different types.
6152 if (SrcVT.bitsGT(VT)) {
6153 // Op0 is MVT::f32, Op1 is MVT::f64.
6154 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6155 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6156 DAG.getConstant(32, MVT::i32));
6157 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
6158 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6159 DAG.getIntPtrConstant(0));
6162 // Clear first operand sign bit.
6164 if (VT == MVT::f64) {
6165 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6166 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6168 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6169 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6170 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6171 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6173 C = ConstantVector::get(CV);
6174 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6175 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6176 PseudoSourceValue::getConstantPool(), 0,
6178 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6180 // Or the value with the sign bit.
6181 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6184 /// Emit nodes that will be selected as "test Op0,Op0", or something
6186 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6187 SelectionDAG &DAG) const {
6188 DebugLoc dl = Op.getDebugLoc();
6190 // CF and OF aren't always set the way we want. Determine which
6191 // of these we need.
6192 bool NeedCF = false;
6193 bool NeedOF = false;
6196 case X86::COND_A: case X86::COND_AE:
6197 case X86::COND_B: case X86::COND_BE:
6200 case X86::COND_G: case X86::COND_GE:
6201 case X86::COND_L: case X86::COND_LE:
6202 case X86::COND_O: case X86::COND_NO:
6207 // See if we can use the EFLAGS value from the operand instead of
6208 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6209 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6210 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6211 // Emit a CMP with 0, which is the TEST pattern.
6212 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6213 DAG.getConstant(0, Op.getValueType()));
6215 unsigned Opcode = 0;
6216 unsigned NumOperands = 0;
6217 switch (Op.getNode()->getOpcode()) {
6219 // Due to an isel shortcoming, be conservative if this add is likely to be
6220 // selected as part of a load-modify-store instruction. When the root node
6221 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6222 // uses of other nodes in the match, such as the ADD in this case. This
6223 // leads to the ADD being left around and reselected, with the result being
6224 // two adds in the output. Alas, even if none our users are stores, that
6225 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6226 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6227 // climbing the DAG back to the root, and it doesn't seem to be worth the
6229 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6230 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6231 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6234 if (ConstantSDNode *C =
6235 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6236 // An add of one will be selected as an INC.
6237 if (C->getAPIntValue() == 1) {
6238 Opcode = X86ISD::INC;
6243 // An add of negative one (subtract of one) will be selected as a DEC.
6244 if (C->getAPIntValue().isAllOnesValue()) {
6245 Opcode = X86ISD::DEC;
6251 // Otherwise use a regular EFLAGS-setting add.
6252 Opcode = X86ISD::ADD;
6256 // If the primary and result isn't used, don't bother using X86ISD::AND,
6257 // because a TEST instruction will be better.
6258 bool NonFlagUse = false;
6259 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6260 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6262 unsigned UOpNo = UI.getOperandNo();
6263 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6264 // Look pass truncate.
6265 UOpNo = User->use_begin().getOperandNo();
6266 User = *User->use_begin();
6269 if (User->getOpcode() != ISD::BRCOND &&
6270 User->getOpcode() != ISD::SETCC &&
6271 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6284 // Due to the ISEL shortcoming noted above, be conservative if this op is
6285 // likely to be selected as part of a load-modify-store instruction.
6286 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6287 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6288 if (UI->getOpcode() == ISD::STORE)
6291 // Otherwise use a regular EFLAGS-setting instruction.
6292 switch (Op.getNode()->getOpcode()) {
6293 default: llvm_unreachable("unexpected operator!");
6294 case ISD::SUB: Opcode = X86ISD::SUB; break;
6295 case ISD::OR: Opcode = X86ISD::OR; break;
6296 case ISD::XOR: Opcode = X86ISD::XOR; break;
6297 case ISD::AND: Opcode = X86ISD::AND; break;
6309 return SDValue(Op.getNode(), 1);
6316 // Emit a CMP with 0, which is the TEST pattern.
6317 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6318 DAG.getConstant(0, Op.getValueType()));
6320 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6321 SmallVector<SDValue, 4> Ops;
6322 for (unsigned i = 0; i != NumOperands; ++i)
6323 Ops.push_back(Op.getOperand(i));
6325 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6326 DAG.ReplaceAllUsesWith(Op, New);
6327 return SDValue(New.getNode(), 1);
6330 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6332 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6333 SelectionDAG &DAG) const {
6334 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6335 if (C->getAPIntValue() == 0)
6336 return EmitTest(Op0, X86CC, DAG);
6338 DebugLoc dl = Op0.getDebugLoc();
6339 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6342 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6343 /// if it's possible.
6344 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6345 DebugLoc dl, SelectionDAG &DAG) const {
6346 SDValue Op0 = And.getOperand(0);
6347 SDValue Op1 = And.getOperand(1);
6348 if (Op0.getOpcode() == ISD::TRUNCATE)
6349 Op0 = Op0.getOperand(0);
6350 if (Op1.getOpcode() == ISD::TRUNCATE)
6351 Op1 = Op1.getOperand(0);
6354 if (Op1.getOpcode() == ISD::SHL)
6355 std::swap(Op0, Op1);
6356 if (Op0.getOpcode() == ISD::SHL) {
6357 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6358 if (And00C->getZExtValue() == 1) {
6359 // If we looked past a truncate, check that it's only truncating away
6361 unsigned BitWidth = Op0.getValueSizeInBits();
6362 unsigned AndBitWidth = And.getValueSizeInBits();
6363 if (BitWidth > AndBitWidth) {
6364 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6365 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6366 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6370 RHS = Op0.getOperand(1);
6372 } else if (Op1.getOpcode() == ISD::Constant) {
6373 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6374 SDValue AndLHS = Op0;
6375 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6376 LHS = AndLHS.getOperand(0);
6377 RHS = AndLHS.getOperand(1);
6381 if (LHS.getNode()) {
6382 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6383 // instruction. Since the shift amount is in-range-or-undefined, we know
6384 // that doing a bittest on the i32 value is ok. We extend to i32 because
6385 // the encoding for the i16 version is larger than the i32 version.
6386 // Also promote i16 to i32 for performance / code size reason.
6387 if (LHS.getValueType() == MVT::i8 ||
6388 LHS.getValueType() == MVT::i16)
6389 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6391 // If the operand types disagree, extend the shift amount to match. Since
6392 // BT ignores high bits (like shifts) we can use anyextend.
6393 if (LHS.getValueType() != RHS.getValueType())
6394 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6396 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6397 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6398 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6399 DAG.getConstant(Cond, MVT::i8), BT);
6405 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6406 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6407 SDValue Op0 = Op.getOperand(0);
6408 SDValue Op1 = Op.getOperand(1);
6409 DebugLoc dl = Op.getDebugLoc();
6410 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6412 // Optimize to BT if possible.
6413 // Lower (X & (1 << N)) == 0 to BT(X, N).
6414 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6415 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6416 if (Op0.getOpcode() == ISD::AND &&
6418 Op1.getOpcode() == ISD::Constant &&
6419 cast<ConstantSDNode>(Op1)->isNullValue() &&
6420 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6421 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6422 if (NewSetCC.getNode())
6426 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6427 if (Op0.getOpcode() == X86ISD::SETCC &&
6428 Op1.getOpcode() == ISD::Constant &&
6429 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6430 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6431 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6432 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6433 bool Invert = (CC == ISD::SETNE) ^
6434 cast<ConstantSDNode>(Op1)->isNullValue();
6436 CCode = X86::GetOppositeBranchCondition(CCode);
6437 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6438 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6441 bool isFP = Op1.getValueType().isFloatingPoint();
6442 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6443 if (X86CC == X86::COND_INVALID)
6446 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6448 // Use sbb x, x to materialize carry bit into a GPR.
6449 if (X86CC == X86::COND_B)
6450 return DAG.getNode(ISD::AND, dl, MVT::i8,
6451 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6452 DAG.getConstant(X86CC, MVT::i8), Cond),
6453 DAG.getConstant(1, MVT::i8));
6455 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6456 DAG.getConstant(X86CC, MVT::i8), Cond);
6459 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6461 SDValue Op0 = Op.getOperand(0);
6462 SDValue Op1 = Op.getOperand(1);
6463 SDValue CC = Op.getOperand(2);
6464 EVT VT = Op.getValueType();
6465 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6466 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6467 DebugLoc dl = Op.getDebugLoc();
6471 EVT VT0 = Op0.getValueType();
6472 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6473 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6476 switch (SetCCOpcode) {
6479 case ISD::SETEQ: SSECC = 0; break;
6481 case ISD::SETGT: Swap = true; // Fallthrough
6483 case ISD::SETOLT: SSECC = 1; break;
6485 case ISD::SETGE: Swap = true; // Fallthrough
6487 case ISD::SETOLE: SSECC = 2; break;
6488 case ISD::SETUO: SSECC = 3; break;
6490 case ISD::SETNE: SSECC = 4; break;
6491 case ISD::SETULE: Swap = true;
6492 case ISD::SETUGE: SSECC = 5; break;
6493 case ISD::SETULT: Swap = true;
6494 case ISD::SETUGT: SSECC = 6; break;
6495 case ISD::SETO: SSECC = 7; break;
6498 std::swap(Op0, Op1);
6500 // In the two special cases we can't handle, emit two comparisons.
6502 if (SetCCOpcode == ISD::SETUEQ) {
6504 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6505 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6506 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6508 else if (SetCCOpcode == ISD::SETONE) {
6510 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6511 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6512 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6514 llvm_unreachable("Illegal FP comparison");
6516 // Handle all other FP comparisons here.
6517 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6520 // We are handling one of the integer comparisons here. Since SSE only has
6521 // GT and EQ comparisons for integer, swapping operands and multiple
6522 // operations may be required for some comparisons.
6523 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6524 bool Swap = false, Invert = false, FlipSigns = false;
6526 switch (VT.getSimpleVT().SimpleTy) {
6529 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6531 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6533 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6534 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6537 switch (SetCCOpcode) {
6539 case ISD::SETNE: Invert = true;
6540 case ISD::SETEQ: Opc = EQOpc; break;
6541 case ISD::SETLT: Swap = true;
6542 case ISD::SETGT: Opc = GTOpc; break;
6543 case ISD::SETGE: Swap = true;
6544 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6545 case ISD::SETULT: Swap = true;
6546 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6547 case ISD::SETUGE: Swap = true;
6548 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6551 std::swap(Op0, Op1);
6553 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6554 // bits of the inputs before performing those operations.
6556 EVT EltVT = VT.getVectorElementType();
6557 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6559 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6560 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6562 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6563 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6566 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6568 // If the logical-not of the result is required, perform that now.
6570 Result = DAG.getNOT(dl, Result, VT);
6575 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6576 static bool isX86LogicalCmp(SDValue Op) {
6577 unsigned Opc = Op.getNode()->getOpcode();
6578 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6580 if (Op.getResNo() == 1 &&
6581 (Opc == X86ISD::ADD ||
6582 Opc == X86ISD::SUB ||
6583 Opc == X86ISD::SMUL ||
6584 Opc == X86ISD::UMUL ||
6585 Opc == X86ISD::INC ||
6586 Opc == X86ISD::DEC ||
6587 Opc == X86ISD::OR ||
6588 Opc == X86ISD::XOR ||
6589 Opc == X86ISD::AND))
6595 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6596 bool addTest = true;
6597 SDValue Cond = Op.getOperand(0);
6598 DebugLoc dl = Op.getDebugLoc();
6601 if (Cond.getOpcode() == ISD::SETCC) {
6602 SDValue NewCond = LowerSETCC(Cond, DAG);
6603 if (NewCond.getNode())
6607 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6608 SDValue Op1 = Op.getOperand(1);
6609 SDValue Op2 = Op.getOperand(2);
6610 if (Cond.getOpcode() == X86ISD::SETCC &&
6611 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6612 SDValue Cmp = Cond.getOperand(1);
6613 if (Cmp.getOpcode() == X86ISD::CMP) {
6614 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6615 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6616 ConstantSDNode *RHSC =
6617 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6618 if (N1C && N1C->isAllOnesValue() &&
6619 N2C && N2C->isNullValue() &&
6620 RHSC && RHSC->isNullValue()) {
6621 SDValue CmpOp0 = Cmp.getOperand(0);
6622 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6623 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6624 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6625 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6630 // Look pass (and (setcc_carry (cmp ...)), 1).
6631 if (Cond.getOpcode() == ISD::AND &&
6632 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6633 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6634 if (C && C->getAPIntValue() == 1)
6635 Cond = Cond.getOperand(0);
6638 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6639 // setting operand in place of the X86ISD::SETCC.
6640 if (Cond.getOpcode() == X86ISD::SETCC ||
6641 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6642 CC = Cond.getOperand(0);
6644 SDValue Cmp = Cond.getOperand(1);
6645 unsigned Opc = Cmp.getOpcode();
6646 EVT VT = Op.getValueType();
6648 bool IllegalFPCMov = false;
6649 if (VT.isFloatingPoint() && !VT.isVector() &&
6650 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6651 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6653 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6654 Opc == X86ISD::BT) { // FIXME
6661 // Look pass the truncate.
6662 if (Cond.getOpcode() == ISD::TRUNCATE)
6663 Cond = Cond.getOperand(0);
6665 // We know the result of AND is compared against zero. Try to match
6667 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6668 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6669 if (NewSetCC.getNode()) {
6670 CC = NewSetCC.getOperand(0);
6671 Cond = NewSetCC.getOperand(1);
6678 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6679 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6682 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6683 // condition is true.
6684 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6685 SDValue Ops[] = { Op2, Op1, CC, Cond };
6686 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6689 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6690 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6691 // from the AND / OR.
6692 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6693 Opc = Op.getOpcode();
6694 if (Opc != ISD::OR && Opc != ISD::AND)
6696 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6697 Op.getOperand(0).hasOneUse() &&
6698 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6699 Op.getOperand(1).hasOneUse());
6702 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6703 // 1 and that the SETCC node has a single use.
6704 static bool isXor1OfSetCC(SDValue Op) {
6705 if (Op.getOpcode() != ISD::XOR)
6707 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6708 if (N1C && N1C->getAPIntValue() == 1) {
6709 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6710 Op.getOperand(0).hasOneUse();
6715 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6716 bool addTest = true;
6717 SDValue Chain = Op.getOperand(0);
6718 SDValue Cond = Op.getOperand(1);
6719 SDValue Dest = Op.getOperand(2);
6720 DebugLoc dl = Op.getDebugLoc();
6723 if (Cond.getOpcode() == ISD::SETCC) {
6724 SDValue NewCond = LowerSETCC(Cond, DAG);
6725 if (NewCond.getNode())
6729 // FIXME: LowerXALUO doesn't handle these!!
6730 else if (Cond.getOpcode() == X86ISD::ADD ||
6731 Cond.getOpcode() == X86ISD::SUB ||
6732 Cond.getOpcode() == X86ISD::SMUL ||
6733 Cond.getOpcode() == X86ISD::UMUL)
6734 Cond = LowerXALUO(Cond, DAG);
6737 // Look pass (and (setcc_carry (cmp ...)), 1).
6738 if (Cond.getOpcode() == ISD::AND &&
6739 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6740 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6741 if (C && C->getAPIntValue() == 1)
6742 Cond = Cond.getOperand(0);
6745 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6746 // setting operand in place of the X86ISD::SETCC.
6747 if (Cond.getOpcode() == X86ISD::SETCC ||
6748 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6749 CC = Cond.getOperand(0);
6751 SDValue Cmp = Cond.getOperand(1);
6752 unsigned Opc = Cmp.getOpcode();
6753 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6754 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6758 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6762 // These can only come from an arithmetic instruction with overflow,
6763 // e.g. SADDO, UADDO.
6764 Cond = Cond.getNode()->getOperand(1);
6771 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6772 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6773 if (CondOpc == ISD::OR) {
6774 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6775 // two branches instead of an explicit OR instruction with a
6777 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6778 isX86LogicalCmp(Cmp)) {
6779 CC = Cond.getOperand(0).getOperand(0);
6780 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6781 Chain, Dest, CC, Cmp);
6782 CC = Cond.getOperand(1).getOperand(0);
6786 } else { // ISD::AND
6787 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6788 // two branches instead of an explicit AND instruction with a
6789 // separate test. However, we only do this if this block doesn't
6790 // have a fall-through edge, because this requires an explicit
6791 // jmp when the condition is false.
6792 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6793 isX86LogicalCmp(Cmp) &&
6794 Op.getNode()->hasOneUse()) {
6795 X86::CondCode CCode =
6796 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6797 CCode = X86::GetOppositeBranchCondition(CCode);
6798 CC = DAG.getConstant(CCode, MVT::i8);
6799 SDNode *User = *Op.getNode()->use_begin();
6800 // Look for an unconditional branch following this conditional branch.
6801 // We need this because we need to reverse the successors in order
6802 // to implement FCMP_OEQ.
6803 if (User->getOpcode() == ISD::BR) {
6804 SDValue FalseBB = User->getOperand(1);
6806 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6807 assert(NewBR == User);
6811 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6812 Chain, Dest, CC, Cmp);
6813 X86::CondCode CCode =
6814 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6815 CCode = X86::GetOppositeBranchCondition(CCode);
6816 CC = DAG.getConstant(CCode, MVT::i8);
6822 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6823 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6824 // It should be transformed during dag combiner except when the condition
6825 // is set by a arithmetics with overflow node.
6826 X86::CondCode CCode =
6827 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6828 CCode = X86::GetOppositeBranchCondition(CCode);
6829 CC = DAG.getConstant(CCode, MVT::i8);
6830 Cond = Cond.getOperand(0).getOperand(1);
6836 // Look pass the truncate.
6837 if (Cond.getOpcode() == ISD::TRUNCATE)
6838 Cond = Cond.getOperand(0);
6840 // We know the result of AND is compared against zero. Try to match
6842 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6843 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6844 if (NewSetCC.getNode()) {
6845 CC = NewSetCC.getOperand(0);
6846 Cond = NewSetCC.getOperand(1);
6853 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6854 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6856 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6857 Chain, Dest, CC, Cond);
6861 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6862 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6863 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6864 // that the guard pages used by the OS virtual memory manager are allocated in
6865 // correct sequence.
6867 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6868 SelectionDAG &DAG) const {
6869 assert(Subtarget->isTargetCygMing() &&
6870 "This should be used only on Cygwin/Mingw targets");
6871 DebugLoc dl = Op.getDebugLoc();
6874 SDValue Chain = Op.getOperand(0);
6875 SDValue Size = Op.getOperand(1);
6876 // FIXME: Ensure alignment here
6880 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6882 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6883 Flag = Chain.getValue(1);
6885 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6887 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6888 Flag = Chain.getValue(1);
6890 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6892 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6893 return DAG.getMergeValues(Ops1, 2, dl);
6896 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6897 MachineFunction &MF = DAG.getMachineFunction();
6898 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6900 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6901 DebugLoc dl = Op.getDebugLoc();
6903 if (!Subtarget->is64Bit()) {
6904 // vastart just stores the address of the VarArgsFrameIndex slot into the
6905 // memory location argument.
6906 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6908 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6913 // gp_offset (0 - 6 * 8)
6914 // fp_offset (48 - 48 + 8 * 16)
6915 // overflow_arg_area (point to parameters coming in memory).
6917 SmallVector<SDValue, 8> MemOps;
6918 SDValue FIN = Op.getOperand(1);
6920 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6921 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6923 FIN, SV, 0, false, false, 0);
6924 MemOps.push_back(Store);
6927 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6928 FIN, DAG.getIntPtrConstant(4));
6929 Store = DAG.getStore(Op.getOperand(0), dl,
6930 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6932 FIN, SV, 4, false, false, 0);
6933 MemOps.push_back(Store);
6935 // Store ptr to overflow_arg_area
6936 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6937 FIN, DAG.getIntPtrConstant(4));
6938 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6940 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 8,
6942 MemOps.push_back(Store);
6944 // Store ptr to reg_save_area.
6945 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6946 FIN, DAG.getIntPtrConstant(8));
6947 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6949 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 16,
6951 MemOps.push_back(Store);
6952 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6953 &MemOps[0], MemOps.size());
6956 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6957 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6958 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6960 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6964 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6965 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6966 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6967 SDValue Chain = Op.getOperand(0);
6968 SDValue DstPtr = Op.getOperand(1);
6969 SDValue SrcPtr = Op.getOperand(2);
6970 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6971 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6972 DebugLoc dl = Op.getDebugLoc();
6974 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6975 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6976 false, DstSV, 0, SrcSV, 0);
6980 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6981 DebugLoc dl = Op.getDebugLoc();
6982 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6984 default: return SDValue(); // Don't custom lower most intrinsics.
6985 // Comparison intrinsics.
6986 case Intrinsic::x86_sse_comieq_ss:
6987 case Intrinsic::x86_sse_comilt_ss:
6988 case Intrinsic::x86_sse_comile_ss:
6989 case Intrinsic::x86_sse_comigt_ss:
6990 case Intrinsic::x86_sse_comige_ss:
6991 case Intrinsic::x86_sse_comineq_ss:
6992 case Intrinsic::x86_sse_ucomieq_ss:
6993 case Intrinsic::x86_sse_ucomilt_ss:
6994 case Intrinsic::x86_sse_ucomile_ss:
6995 case Intrinsic::x86_sse_ucomigt_ss:
6996 case Intrinsic::x86_sse_ucomige_ss:
6997 case Intrinsic::x86_sse_ucomineq_ss:
6998 case Intrinsic::x86_sse2_comieq_sd:
6999 case Intrinsic::x86_sse2_comilt_sd:
7000 case Intrinsic::x86_sse2_comile_sd:
7001 case Intrinsic::x86_sse2_comigt_sd:
7002 case Intrinsic::x86_sse2_comige_sd:
7003 case Intrinsic::x86_sse2_comineq_sd:
7004 case Intrinsic::x86_sse2_ucomieq_sd:
7005 case Intrinsic::x86_sse2_ucomilt_sd:
7006 case Intrinsic::x86_sse2_ucomile_sd:
7007 case Intrinsic::x86_sse2_ucomigt_sd:
7008 case Intrinsic::x86_sse2_ucomige_sd:
7009 case Intrinsic::x86_sse2_ucomineq_sd: {
7011 ISD::CondCode CC = ISD::SETCC_INVALID;
7014 case Intrinsic::x86_sse_comieq_ss:
7015 case Intrinsic::x86_sse2_comieq_sd:
7019 case Intrinsic::x86_sse_comilt_ss:
7020 case Intrinsic::x86_sse2_comilt_sd:
7024 case Intrinsic::x86_sse_comile_ss:
7025 case Intrinsic::x86_sse2_comile_sd:
7029 case Intrinsic::x86_sse_comigt_ss:
7030 case Intrinsic::x86_sse2_comigt_sd:
7034 case Intrinsic::x86_sse_comige_ss:
7035 case Intrinsic::x86_sse2_comige_sd:
7039 case Intrinsic::x86_sse_comineq_ss:
7040 case Intrinsic::x86_sse2_comineq_sd:
7044 case Intrinsic::x86_sse_ucomieq_ss:
7045 case Intrinsic::x86_sse2_ucomieq_sd:
7046 Opc = X86ISD::UCOMI;
7049 case Intrinsic::x86_sse_ucomilt_ss:
7050 case Intrinsic::x86_sse2_ucomilt_sd:
7051 Opc = X86ISD::UCOMI;
7054 case Intrinsic::x86_sse_ucomile_ss:
7055 case Intrinsic::x86_sse2_ucomile_sd:
7056 Opc = X86ISD::UCOMI;
7059 case Intrinsic::x86_sse_ucomigt_ss:
7060 case Intrinsic::x86_sse2_ucomigt_sd:
7061 Opc = X86ISD::UCOMI;
7064 case Intrinsic::x86_sse_ucomige_ss:
7065 case Intrinsic::x86_sse2_ucomige_sd:
7066 Opc = X86ISD::UCOMI;
7069 case Intrinsic::x86_sse_ucomineq_ss:
7070 case Intrinsic::x86_sse2_ucomineq_sd:
7071 Opc = X86ISD::UCOMI;
7076 SDValue LHS = Op.getOperand(1);
7077 SDValue RHS = Op.getOperand(2);
7078 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
7079 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
7080 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
7081 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7082 DAG.getConstant(X86CC, MVT::i8), Cond);
7083 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7085 // ptest and testp intrinsics. The intrinsic these come from are designed to
7086 // return an integer value, not just an instruction so lower it to the ptest
7087 // or testp pattern and a setcc for the result.
7088 case Intrinsic::x86_sse41_ptestz:
7089 case Intrinsic::x86_sse41_ptestc:
7090 case Intrinsic::x86_sse41_ptestnzc:
7091 case Intrinsic::x86_avx_ptestz_256:
7092 case Intrinsic::x86_avx_ptestc_256:
7093 case Intrinsic::x86_avx_ptestnzc_256:
7094 case Intrinsic::x86_avx_vtestz_ps:
7095 case Intrinsic::x86_avx_vtestc_ps:
7096 case Intrinsic::x86_avx_vtestnzc_ps:
7097 case Intrinsic::x86_avx_vtestz_pd:
7098 case Intrinsic::x86_avx_vtestc_pd:
7099 case Intrinsic::x86_avx_vtestnzc_pd:
7100 case Intrinsic::x86_avx_vtestz_ps_256:
7101 case Intrinsic::x86_avx_vtestc_ps_256:
7102 case Intrinsic::x86_avx_vtestnzc_ps_256:
7103 case Intrinsic::x86_avx_vtestz_pd_256:
7104 case Intrinsic::x86_avx_vtestc_pd_256:
7105 case Intrinsic::x86_avx_vtestnzc_pd_256: {
7106 bool IsTestPacked = false;
7109 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
7110 case Intrinsic::x86_avx_vtestz_ps:
7111 case Intrinsic::x86_avx_vtestz_pd:
7112 case Intrinsic::x86_avx_vtestz_ps_256:
7113 case Intrinsic::x86_avx_vtestz_pd_256:
7114 IsTestPacked = true; // Fallthrough
7115 case Intrinsic::x86_sse41_ptestz:
7116 case Intrinsic::x86_avx_ptestz_256:
7118 X86CC = X86::COND_E;
7120 case Intrinsic::x86_avx_vtestc_ps:
7121 case Intrinsic::x86_avx_vtestc_pd:
7122 case Intrinsic::x86_avx_vtestc_ps_256:
7123 case Intrinsic::x86_avx_vtestc_pd_256:
7124 IsTestPacked = true; // Fallthrough
7125 case Intrinsic::x86_sse41_ptestc:
7126 case Intrinsic::x86_avx_ptestc_256:
7128 X86CC = X86::COND_B;
7130 case Intrinsic::x86_avx_vtestnzc_ps:
7131 case Intrinsic::x86_avx_vtestnzc_pd:
7132 case Intrinsic::x86_avx_vtestnzc_ps_256:
7133 case Intrinsic::x86_avx_vtestnzc_pd_256:
7134 IsTestPacked = true; // Fallthrough
7135 case Intrinsic::x86_sse41_ptestnzc:
7136 case Intrinsic::x86_avx_ptestnzc_256:
7138 X86CC = X86::COND_A;
7142 SDValue LHS = Op.getOperand(1);
7143 SDValue RHS = Op.getOperand(2);
7144 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
7145 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
7146 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7147 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7148 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7151 // Fix vector shift instructions where the last operand is a non-immediate
7153 case Intrinsic::x86_sse2_pslli_w:
7154 case Intrinsic::x86_sse2_pslli_d:
7155 case Intrinsic::x86_sse2_pslli_q:
7156 case Intrinsic::x86_sse2_psrli_w:
7157 case Intrinsic::x86_sse2_psrli_d:
7158 case Intrinsic::x86_sse2_psrli_q:
7159 case Intrinsic::x86_sse2_psrai_w:
7160 case Intrinsic::x86_sse2_psrai_d:
7161 case Intrinsic::x86_mmx_pslli_w:
7162 case Intrinsic::x86_mmx_pslli_d:
7163 case Intrinsic::x86_mmx_pslli_q:
7164 case Intrinsic::x86_mmx_psrli_w:
7165 case Intrinsic::x86_mmx_psrli_d:
7166 case Intrinsic::x86_mmx_psrli_q:
7167 case Intrinsic::x86_mmx_psrai_w:
7168 case Intrinsic::x86_mmx_psrai_d: {
7169 SDValue ShAmt = Op.getOperand(2);
7170 if (isa<ConstantSDNode>(ShAmt))
7173 unsigned NewIntNo = 0;
7174 EVT ShAmtVT = MVT::v4i32;
7176 case Intrinsic::x86_sse2_pslli_w:
7177 NewIntNo = Intrinsic::x86_sse2_psll_w;
7179 case Intrinsic::x86_sse2_pslli_d:
7180 NewIntNo = Intrinsic::x86_sse2_psll_d;
7182 case Intrinsic::x86_sse2_pslli_q:
7183 NewIntNo = Intrinsic::x86_sse2_psll_q;
7185 case Intrinsic::x86_sse2_psrli_w:
7186 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7188 case Intrinsic::x86_sse2_psrli_d:
7189 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7191 case Intrinsic::x86_sse2_psrli_q:
7192 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7194 case Intrinsic::x86_sse2_psrai_w:
7195 NewIntNo = Intrinsic::x86_sse2_psra_w;
7197 case Intrinsic::x86_sse2_psrai_d:
7198 NewIntNo = Intrinsic::x86_sse2_psra_d;
7201 ShAmtVT = MVT::v2i32;
7203 case Intrinsic::x86_mmx_pslli_w:
7204 NewIntNo = Intrinsic::x86_mmx_psll_w;
7206 case Intrinsic::x86_mmx_pslli_d:
7207 NewIntNo = Intrinsic::x86_mmx_psll_d;
7209 case Intrinsic::x86_mmx_pslli_q:
7210 NewIntNo = Intrinsic::x86_mmx_psll_q;
7212 case Intrinsic::x86_mmx_psrli_w:
7213 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7215 case Intrinsic::x86_mmx_psrli_d:
7216 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7218 case Intrinsic::x86_mmx_psrli_q:
7219 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7221 case Intrinsic::x86_mmx_psrai_w:
7222 NewIntNo = Intrinsic::x86_mmx_psra_w;
7224 case Intrinsic::x86_mmx_psrai_d:
7225 NewIntNo = Intrinsic::x86_mmx_psra_d;
7227 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7233 // The vector shift intrinsics with scalars uses 32b shift amounts but
7234 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7238 ShOps[1] = DAG.getConstant(0, MVT::i32);
7239 if (ShAmtVT == MVT::v4i32) {
7240 ShOps[2] = DAG.getUNDEF(MVT::i32);
7241 ShOps[3] = DAG.getUNDEF(MVT::i32);
7242 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7244 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7247 EVT VT = Op.getValueType();
7248 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7249 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7250 DAG.getConstant(NewIntNo, MVT::i32),
7251 Op.getOperand(1), ShAmt);
7256 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7257 SelectionDAG &DAG) const {
7258 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7259 MFI->setReturnAddressIsTaken(true);
7261 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7262 DebugLoc dl = Op.getDebugLoc();
7265 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7267 DAG.getConstant(TD->getPointerSize(),
7268 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7269 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7270 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7272 NULL, 0, false, false, 0);
7275 // Just load the return address.
7276 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7277 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7278 RetAddrFI, NULL, 0, false, false, 0);
7281 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7282 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7283 MFI->setFrameAddressIsTaken(true);
7285 EVT VT = Op.getValueType();
7286 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7287 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7288 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7289 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7291 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7296 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7297 SelectionDAG &DAG) const {
7298 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7301 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7302 MachineFunction &MF = DAG.getMachineFunction();
7303 SDValue Chain = Op.getOperand(0);
7304 SDValue Offset = Op.getOperand(1);
7305 SDValue Handler = Op.getOperand(2);
7306 DebugLoc dl = Op.getDebugLoc();
7308 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
7309 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7311 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7313 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
7314 DAG.getIntPtrConstant(TD->getPointerSize()));
7315 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7316 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7317 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7318 MF.getRegInfo().addLiveOut(StoreAddrReg);
7320 return DAG.getNode(X86ISD::EH_RETURN, dl,
7322 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7325 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7326 SelectionDAG &DAG) const {
7327 SDValue Root = Op.getOperand(0);
7328 SDValue Trmp = Op.getOperand(1); // trampoline
7329 SDValue FPtr = Op.getOperand(2); // nested function
7330 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7331 DebugLoc dl = Op.getDebugLoc();
7333 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7335 if (Subtarget->is64Bit()) {
7336 SDValue OutChains[6];
7338 // Large code-model.
7339 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7340 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7342 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7343 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7345 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7347 // Load the pointer to the nested function into R11.
7348 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7349 SDValue Addr = Trmp;
7350 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7351 Addr, TrmpAddr, 0, false, false, 0);
7353 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7354 DAG.getConstant(2, MVT::i64));
7355 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7358 // Load the 'nest' parameter value into R10.
7359 // R10 is specified in X86CallingConv.td
7360 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7361 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7362 DAG.getConstant(10, MVT::i64));
7363 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7364 Addr, TrmpAddr, 10, false, false, 0);
7366 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7367 DAG.getConstant(12, MVT::i64));
7368 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7371 // Jump to the nested function.
7372 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7373 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7374 DAG.getConstant(20, MVT::i64));
7375 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7376 Addr, TrmpAddr, 20, false, false, 0);
7378 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7379 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7380 DAG.getConstant(22, MVT::i64));
7381 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7382 TrmpAddr, 22, false, false, 0);
7385 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7386 return DAG.getMergeValues(Ops, 2, dl);
7388 const Function *Func =
7389 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7390 CallingConv::ID CC = Func->getCallingConv();
7395 llvm_unreachable("Unsupported calling convention");
7396 case CallingConv::C:
7397 case CallingConv::X86_StdCall: {
7398 // Pass 'nest' parameter in ECX.
7399 // Must be kept in sync with X86CallingConv.td
7402 // Check that ECX wasn't needed by an 'inreg' parameter.
7403 const FunctionType *FTy = Func->getFunctionType();
7404 const AttrListPtr &Attrs = Func->getAttributes();
7406 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7407 unsigned InRegCount = 0;
7410 for (FunctionType::param_iterator I = FTy->param_begin(),
7411 E = FTy->param_end(); I != E; ++I, ++Idx)
7412 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7413 // FIXME: should only count parameters that are lowered to integers.
7414 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7416 if (InRegCount > 2) {
7417 report_fatal_error("Nest register in use - reduce number of inreg"
7423 case CallingConv::X86_FastCall:
7424 case CallingConv::X86_ThisCall:
7425 case CallingConv::Fast:
7426 // Pass 'nest' parameter in EAX.
7427 // Must be kept in sync with X86CallingConv.td
7432 SDValue OutChains[4];
7435 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7436 DAG.getConstant(10, MVT::i32));
7437 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7439 // This is storing the opcode for MOV32ri.
7440 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7441 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7442 OutChains[0] = DAG.getStore(Root, dl,
7443 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7444 Trmp, TrmpAddr, 0, false, false, 0);
7446 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7447 DAG.getConstant(1, MVT::i32));
7448 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7451 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7452 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7453 DAG.getConstant(5, MVT::i32));
7454 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7455 TrmpAddr, 5, false, false, 1);
7457 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7458 DAG.getConstant(6, MVT::i32));
7459 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7463 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7464 return DAG.getMergeValues(Ops, 2, dl);
7468 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7469 SelectionDAG &DAG) const {
7471 The rounding mode is in bits 11:10 of FPSR, and has the following
7478 FLT_ROUNDS, on the other hand, expects the following:
7485 To perform the conversion, we do:
7486 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7489 MachineFunction &MF = DAG.getMachineFunction();
7490 const TargetMachine &TM = MF.getTarget();
7491 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7492 unsigned StackAlignment = TFI.getStackAlignment();
7493 EVT VT = Op.getValueType();
7494 DebugLoc dl = Op.getDebugLoc();
7496 // Save FP Control Word to stack slot
7497 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7498 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7500 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7501 DAG.getEntryNode(), StackSlot);
7503 // Load FP Control Word from stack slot
7504 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7507 // Transform as necessary
7509 DAG.getNode(ISD::SRL, dl, MVT::i16,
7510 DAG.getNode(ISD::AND, dl, MVT::i16,
7511 CWD, DAG.getConstant(0x800, MVT::i16)),
7512 DAG.getConstant(11, MVT::i8));
7514 DAG.getNode(ISD::SRL, dl, MVT::i16,
7515 DAG.getNode(ISD::AND, dl, MVT::i16,
7516 CWD, DAG.getConstant(0x400, MVT::i16)),
7517 DAG.getConstant(9, MVT::i8));
7520 DAG.getNode(ISD::AND, dl, MVT::i16,
7521 DAG.getNode(ISD::ADD, dl, MVT::i16,
7522 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7523 DAG.getConstant(1, MVT::i16)),
7524 DAG.getConstant(3, MVT::i16));
7527 return DAG.getNode((VT.getSizeInBits() < 16 ?
7528 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7531 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7532 EVT VT = Op.getValueType();
7534 unsigned NumBits = VT.getSizeInBits();
7535 DebugLoc dl = Op.getDebugLoc();
7537 Op = Op.getOperand(0);
7538 if (VT == MVT::i8) {
7539 // Zero extend to i32 since there is not an i8 bsr.
7541 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7544 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7545 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7546 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7548 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7551 DAG.getConstant(NumBits+NumBits-1, OpVT),
7552 DAG.getConstant(X86::COND_E, MVT::i8),
7555 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7557 // Finally xor with NumBits-1.
7558 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7561 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7565 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7566 EVT VT = Op.getValueType();
7568 unsigned NumBits = VT.getSizeInBits();
7569 DebugLoc dl = Op.getDebugLoc();
7571 Op = Op.getOperand(0);
7572 if (VT == MVT::i8) {
7574 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7577 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7578 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7579 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7581 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7584 DAG.getConstant(NumBits, OpVT),
7585 DAG.getConstant(X86::COND_E, MVT::i8),
7588 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7591 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7595 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7596 EVT VT = Op.getValueType();
7597 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7598 DebugLoc dl = Op.getDebugLoc();
7600 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7601 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7602 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7603 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7604 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7606 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7607 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7608 // return AloBlo + AloBhi + AhiBlo;
7610 SDValue A = Op.getOperand(0);
7611 SDValue B = Op.getOperand(1);
7613 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7614 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7615 A, DAG.getConstant(32, MVT::i32));
7616 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7617 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7618 B, DAG.getConstant(32, MVT::i32));
7619 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7620 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7622 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7623 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7625 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7626 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7628 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7629 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7630 AloBhi, DAG.getConstant(32, MVT::i32));
7631 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7632 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7633 AhiBlo, DAG.getConstant(32, MVT::i32));
7634 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7635 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7639 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
7640 EVT VT = Op.getValueType();
7641 DebugLoc dl = Op.getDebugLoc();
7642 SDValue R = Op.getOperand(0);
7644 LLVMContext *Context = DAG.getContext();
7646 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
7648 if (VT == MVT::v4i32) {
7649 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7650 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
7651 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
7653 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
7655 std::vector<Constant*> CV(4, CI);
7656 Constant *C = ConstantVector::get(CV);
7657 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7658 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7659 PseudoSourceValue::getConstantPool(), 0,
7662 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
7663 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, Op);
7664 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
7665 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
7667 if (VT == MVT::v16i8) {
7669 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7670 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
7671 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
7673 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
7674 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
7676 std::vector<Constant*> CVM1(16, CM1);
7677 std::vector<Constant*> CVM2(16, CM2);
7678 Constant *C = ConstantVector::get(CVM1);
7679 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7680 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7681 PseudoSourceValue::getConstantPool(), 0,
7684 // r = pblendv(r, psllw(r & (char16)15, 4), a);
7685 M = DAG.getNode(ISD::AND, dl, VT, R, M);
7686 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7687 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
7688 DAG.getConstant(4, MVT::i32));
7689 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7690 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
7693 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
7695 C = ConstantVector::get(CVM2);
7696 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7697 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7698 PseudoSourceValue::getConstantPool(), 0, false, false, 16);
7700 // r = pblendv(r, psllw(r & (char16)63, 2), a);
7701 M = DAG.getNode(ISD::AND, dl, VT, R, M);
7702 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7703 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
7704 DAG.getConstant(2, MVT::i32));
7705 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7706 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
7709 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
7711 // return pblendv(r, r+r, a);
7712 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7713 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
7714 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
7720 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7721 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7722 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7723 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7724 // has only one use.
7725 SDNode *N = Op.getNode();
7726 SDValue LHS = N->getOperand(0);
7727 SDValue RHS = N->getOperand(1);
7728 unsigned BaseOp = 0;
7730 DebugLoc dl = Op.getDebugLoc();
7732 switch (Op.getOpcode()) {
7733 default: llvm_unreachable("Unknown ovf instruction!");
7735 // A subtract of one will be selected as a INC. Note that INC doesn't
7736 // set CF, so we can't do this for UADDO.
7737 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7738 if (C->getAPIntValue() == 1) {
7739 BaseOp = X86ISD::INC;
7743 BaseOp = X86ISD::ADD;
7747 BaseOp = X86ISD::ADD;
7751 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7752 // set CF, so we can't do this for USUBO.
7753 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7754 if (C->getAPIntValue() == 1) {
7755 BaseOp = X86ISD::DEC;
7759 BaseOp = X86ISD::SUB;
7763 BaseOp = X86ISD::SUB;
7767 BaseOp = X86ISD::SMUL;
7771 BaseOp = X86ISD::UMUL;
7776 // Also sets EFLAGS.
7777 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7778 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7781 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7782 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7784 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7788 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
7789 DebugLoc dl = Op.getDebugLoc();
7791 if (!Subtarget->hasSSE2()) {
7792 SDValue Chain = Op.getOperand(0);
7793 SDValue Zero = DAG.getConstant(0,
7794 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7796 DAG.getRegister(X86::ESP, MVT::i32), // Base
7797 DAG.getTargetConstant(1, MVT::i8), // Scale
7798 DAG.getRegister(0, MVT::i32), // Index
7799 DAG.getTargetConstant(0, MVT::i32), // Disp
7800 DAG.getRegister(0, MVT::i32), // Segment.
7805 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
7806 array_lengthof(Ops));
7807 return SDValue(Res, 0);
7810 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
7812 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
7814 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7815 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
7816 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
7817 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
7819 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
7820 if (!Op1 && !Op2 && !Op3 && Op4)
7821 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
7823 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
7824 if (Op1 && !Op2 && !Op3 && !Op4)
7825 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
7827 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
7829 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
7832 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7833 EVT T = Op.getValueType();
7834 DebugLoc dl = Op.getDebugLoc();
7837 switch(T.getSimpleVT().SimpleTy) {
7839 assert(false && "Invalid value type!");
7840 case MVT::i8: Reg = X86::AL; size = 1; break;
7841 case MVT::i16: Reg = X86::AX; size = 2; break;
7842 case MVT::i32: Reg = X86::EAX; size = 4; break;
7844 assert(Subtarget->is64Bit() && "Node not type legal!");
7845 Reg = X86::RAX; size = 8;
7848 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7849 Op.getOperand(2), SDValue());
7850 SDValue Ops[] = { cpIn.getValue(0),
7853 DAG.getTargetConstant(size, MVT::i8),
7855 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7856 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7858 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7862 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7863 SelectionDAG &DAG) const {
7864 assert(Subtarget->is64Bit() && "Result not type legalized?");
7865 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7866 SDValue TheChain = Op.getOperand(0);
7867 DebugLoc dl = Op.getDebugLoc();
7868 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7869 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7870 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7872 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7873 DAG.getConstant(32, MVT::i8));
7875 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7878 return DAG.getMergeValues(Ops, 2, dl);
7881 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7882 SelectionDAG &DAG) const {
7883 EVT SrcVT = Op.getOperand(0).getValueType();
7884 EVT DstVT = Op.getValueType();
7885 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7886 Subtarget->hasMMX() && !DisableMMX) &&
7887 "Unexpected custom BIT_CONVERT");
7888 assert((DstVT == MVT::i64 ||
7889 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7890 "Unexpected custom BIT_CONVERT");
7891 // i64 <=> MMX conversions are Legal.
7892 if (SrcVT==MVT::i64 && DstVT.isVector())
7894 if (DstVT==MVT::i64 && SrcVT.isVector())
7896 // MMX <=> MMX conversions are Legal.
7897 if (SrcVT.isVector() && DstVT.isVector())
7899 // All other conversions need to be expanded.
7902 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7903 SDNode *Node = Op.getNode();
7904 DebugLoc dl = Node->getDebugLoc();
7905 EVT T = Node->getValueType(0);
7906 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7907 DAG.getConstant(0, T), Node->getOperand(2));
7908 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7909 cast<AtomicSDNode>(Node)->getMemoryVT(),
7910 Node->getOperand(0),
7911 Node->getOperand(1), negOp,
7912 cast<AtomicSDNode>(Node)->getSrcValue(),
7913 cast<AtomicSDNode>(Node)->getAlignment());
7916 /// LowerOperation - Provide custom lowering hooks for some operations.
7918 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7919 switch (Op.getOpcode()) {
7920 default: llvm_unreachable("Should not custom lower this!");
7921 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
7922 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7923 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7924 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7925 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7926 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7927 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7928 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7929 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7930 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7931 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7932 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7933 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7934 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7935 case ISD::SHL_PARTS:
7936 case ISD::SRA_PARTS:
7937 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7938 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7939 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7940 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7941 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7942 case ISD::FABS: return LowerFABS(Op, DAG);
7943 case ISD::FNEG: return LowerFNEG(Op, DAG);
7944 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7945 case ISD::SETCC: return LowerSETCC(Op, DAG);
7946 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7947 case ISD::SELECT: return LowerSELECT(Op, DAG);
7948 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7949 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7950 case ISD::VASTART: return LowerVASTART(Op, DAG);
7951 case ISD::VAARG: return LowerVAARG(Op, DAG);
7952 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7953 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7954 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7955 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7956 case ISD::FRAME_TO_ARGS_OFFSET:
7957 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7958 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7959 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7960 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7961 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7962 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7963 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7964 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7965 case ISD::SHL: return LowerSHL(Op, DAG);
7971 case ISD::UMULO: return LowerXALUO(Op, DAG);
7972 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7973 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7977 void X86TargetLowering::
7978 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7979 SelectionDAG &DAG, unsigned NewOp) const {
7980 EVT T = Node->getValueType(0);
7981 DebugLoc dl = Node->getDebugLoc();
7982 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7984 SDValue Chain = Node->getOperand(0);
7985 SDValue In1 = Node->getOperand(1);
7986 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7987 Node->getOperand(2), DAG.getIntPtrConstant(0));
7988 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7989 Node->getOperand(2), DAG.getIntPtrConstant(1));
7990 SDValue Ops[] = { Chain, In1, In2L, In2H };
7991 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7993 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7994 cast<MemSDNode>(Node)->getMemOperand());
7995 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7996 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7997 Results.push_back(Result.getValue(2));
8000 /// ReplaceNodeResults - Replace a node with an illegal result type
8001 /// with a new node built out of custom code.
8002 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
8003 SmallVectorImpl<SDValue>&Results,
8004 SelectionDAG &DAG) const {
8005 DebugLoc dl = N->getDebugLoc();
8006 switch (N->getOpcode()) {
8008 assert(false && "Do not know how to custom type legalize this operation!");
8010 case ISD::FP_TO_SINT: {
8011 std::pair<SDValue,SDValue> Vals =
8012 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
8013 SDValue FIST = Vals.first, StackSlot = Vals.second;
8014 if (FIST.getNode() != 0) {
8015 EVT VT = N->getValueType(0);
8016 // Return a load from the stack slot.
8017 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
8022 case ISD::READCYCLECOUNTER: {
8023 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8024 SDValue TheChain = N->getOperand(0);
8025 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8026 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
8028 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
8030 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
8031 SDValue Ops[] = { eax, edx };
8032 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
8033 Results.push_back(edx.getValue(1));
8036 case ISD::ATOMIC_CMP_SWAP: {
8037 EVT T = N->getValueType(0);
8038 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
8039 SDValue cpInL, cpInH;
8040 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8041 DAG.getConstant(0, MVT::i32));
8042 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8043 DAG.getConstant(1, MVT::i32));
8044 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
8045 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
8047 SDValue swapInL, swapInH;
8048 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8049 DAG.getConstant(0, MVT::i32));
8050 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8051 DAG.getConstant(1, MVT::i32));
8052 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
8054 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
8055 swapInL.getValue(1));
8056 SDValue Ops[] = { swapInH.getValue(0),
8058 swapInH.getValue(1) };
8059 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8060 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
8061 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
8062 MVT::i32, Result.getValue(1));
8063 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
8064 MVT::i32, cpOutL.getValue(2));
8065 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
8066 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8067 Results.push_back(cpOutH.getValue(1));
8070 case ISD::ATOMIC_LOAD_ADD:
8071 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
8073 case ISD::ATOMIC_LOAD_AND:
8074 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
8076 case ISD::ATOMIC_LOAD_NAND:
8077 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
8079 case ISD::ATOMIC_LOAD_OR:
8080 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
8082 case ISD::ATOMIC_LOAD_SUB:
8083 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
8085 case ISD::ATOMIC_LOAD_XOR:
8086 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
8088 case ISD::ATOMIC_SWAP:
8089 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
8094 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
8096 default: return NULL;
8097 case X86ISD::BSF: return "X86ISD::BSF";
8098 case X86ISD::BSR: return "X86ISD::BSR";
8099 case X86ISD::SHLD: return "X86ISD::SHLD";
8100 case X86ISD::SHRD: return "X86ISD::SHRD";
8101 case X86ISD::FAND: return "X86ISD::FAND";
8102 case X86ISD::FOR: return "X86ISD::FOR";
8103 case X86ISD::FXOR: return "X86ISD::FXOR";
8104 case X86ISD::FSRL: return "X86ISD::FSRL";
8105 case X86ISD::FILD: return "X86ISD::FILD";
8106 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
8107 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
8108 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
8109 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
8110 case X86ISD::FLD: return "X86ISD::FLD";
8111 case X86ISD::FST: return "X86ISD::FST";
8112 case X86ISD::CALL: return "X86ISD::CALL";
8113 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
8114 case X86ISD::BT: return "X86ISD::BT";
8115 case X86ISD::CMP: return "X86ISD::CMP";
8116 case X86ISD::COMI: return "X86ISD::COMI";
8117 case X86ISD::UCOMI: return "X86ISD::UCOMI";
8118 case X86ISD::SETCC: return "X86ISD::SETCC";
8119 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
8120 case X86ISD::CMOV: return "X86ISD::CMOV";
8121 case X86ISD::BRCOND: return "X86ISD::BRCOND";
8122 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
8123 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
8124 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
8125 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
8126 case X86ISD::Wrapper: return "X86ISD::Wrapper";
8127 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
8128 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
8129 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
8130 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
8131 case X86ISD::PINSRB: return "X86ISD::PINSRB";
8132 case X86ISD::PINSRW: return "X86ISD::PINSRW";
8133 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
8134 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
8135 case X86ISD::FMAX: return "X86ISD::FMAX";
8136 case X86ISD::FMIN: return "X86ISD::FMIN";
8137 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
8138 case X86ISD::FRCP: return "X86ISD::FRCP";
8139 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
8140 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
8141 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
8142 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
8143 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
8144 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
8145 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
8146 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
8147 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
8148 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
8149 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
8150 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
8151 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
8152 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
8153 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
8154 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
8155 case X86ISD::VSHL: return "X86ISD::VSHL";
8156 case X86ISD::VSRL: return "X86ISD::VSRL";
8157 case X86ISD::CMPPD: return "X86ISD::CMPPD";
8158 case X86ISD::CMPPS: return "X86ISD::CMPPS";
8159 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
8160 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
8161 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
8162 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
8163 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
8164 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
8165 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
8166 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
8167 case X86ISD::ADD: return "X86ISD::ADD";
8168 case X86ISD::SUB: return "X86ISD::SUB";
8169 case X86ISD::SMUL: return "X86ISD::SMUL";
8170 case X86ISD::UMUL: return "X86ISD::UMUL";
8171 case X86ISD::INC: return "X86ISD::INC";
8172 case X86ISD::DEC: return "X86ISD::DEC";
8173 case X86ISD::OR: return "X86ISD::OR";
8174 case X86ISD::XOR: return "X86ISD::XOR";
8175 case X86ISD::AND: return "X86ISD::AND";
8176 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
8177 case X86ISD::PTEST: return "X86ISD::PTEST";
8178 case X86ISD::TESTP: return "X86ISD::TESTP";
8179 case X86ISD::PALIGN: return "X86ISD::PALIGN";
8180 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
8181 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
8182 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
8183 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
8184 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
8185 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
8186 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
8187 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
8188 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
8189 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
8190 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
8191 case X86ISD::MOVHPS: return "X86ISD::MOVHPS";
8192 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
8193 case X86ISD::MOVHPD: return "X86ISD::MOVHPD";
8194 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
8195 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
8196 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
8197 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
8198 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
8199 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
8200 case X86ISD::MOVSD: return "X86ISD::MOVSD";
8201 case X86ISD::MOVSS: return "X86ISD::MOVSS";
8202 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
8203 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
8204 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
8205 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
8206 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
8207 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
8208 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
8209 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
8210 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
8211 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
8212 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
8213 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
8214 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
8215 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
8219 // isLegalAddressingMode - Return true if the addressing mode represented
8220 // by AM is legal for this target, for a load/store of the specified type.
8221 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
8222 const Type *Ty) const {
8223 // X86 supports extremely general addressing modes.
8224 CodeModel::Model M = getTargetMachine().getCodeModel();
8225 Reloc::Model R = getTargetMachine().getRelocationModel();
8227 // X86 allows a sign-extended 32-bit immediate field as a displacement.
8228 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
8233 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
8235 // If a reference to this global requires an extra load, we can't fold it.
8236 if (isGlobalStubReference(GVFlags))
8239 // If BaseGV requires a register for the PIC base, we cannot also have a
8240 // BaseReg specified.
8241 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
8244 // If lower 4G is not available, then we must use rip-relative addressing.
8245 if ((M != CodeModel::Small || R != Reloc::Static) &&
8246 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
8256 // These scales always work.
8261 // These scales are formed with basereg+scalereg. Only accept if there is
8266 default: // Other stuff never works.
8274 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
8275 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
8277 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
8278 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
8279 if (NumBits1 <= NumBits2)
8284 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
8285 if (!VT1.isInteger() || !VT2.isInteger())
8287 unsigned NumBits1 = VT1.getSizeInBits();
8288 unsigned NumBits2 = VT2.getSizeInBits();
8289 if (NumBits1 <= NumBits2)
8294 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
8295 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8296 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
8299 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
8300 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8301 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
8304 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
8305 // i16 instructions are longer (0x66 prefix) and potentially slower.
8306 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
8309 /// isShuffleMaskLegal - Targets can use this to indicate that they only
8310 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
8311 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
8312 /// are assumed to be legal.
8314 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
8316 // Very little shuffling can be done for 64-bit vectors right now.
8317 if (VT.getSizeInBits() == 64)
8318 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8320 // FIXME: pshufb, blends, shifts.
8321 return (VT.getVectorNumElements() == 2 ||
8322 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8323 isMOVLMask(M, VT) ||
8324 isSHUFPMask(M, VT) ||
8325 isPSHUFDMask(M, VT) ||
8326 isPSHUFHWMask(M, VT) ||
8327 isPSHUFLWMask(M, VT) ||
8328 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8329 isUNPCKLMask(M, VT) ||
8330 isUNPCKHMask(M, VT) ||
8331 isUNPCKL_v_undef_Mask(M, VT) ||
8332 isUNPCKH_v_undef_Mask(M, VT));
8336 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8338 unsigned NumElts = VT.getVectorNumElements();
8339 // FIXME: This collection of masks seems suspect.
8342 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8343 return (isMOVLMask(Mask, VT) ||
8344 isCommutedMOVLMask(Mask, VT, true) ||
8345 isSHUFPMask(Mask, VT) ||
8346 isCommutedSHUFPMask(Mask, VT));
8351 //===----------------------------------------------------------------------===//
8352 // X86 Scheduler Hooks
8353 //===----------------------------------------------------------------------===//
8355 // private utility function
8357 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8358 MachineBasicBlock *MBB,
8365 TargetRegisterClass *RC,
8366 bool invSrc) const {
8367 // For the atomic bitwise operator, we generate
8370 // ld t1 = [bitinstr.addr]
8371 // op t2 = t1, [bitinstr.val]
8373 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8375 // fallthrough -->nextMBB
8376 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8377 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8378 MachineFunction::iterator MBBIter = MBB;
8381 /// First build the CFG
8382 MachineFunction *F = MBB->getParent();
8383 MachineBasicBlock *thisMBB = MBB;
8384 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8385 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8386 F->insert(MBBIter, newMBB);
8387 F->insert(MBBIter, nextMBB);
8389 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8390 nextMBB->splice(nextMBB->begin(), thisMBB,
8391 llvm::next(MachineBasicBlock::iterator(bInstr)),
8393 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8395 // Update thisMBB to fall through to newMBB
8396 thisMBB->addSuccessor(newMBB);
8398 // newMBB jumps to itself and fall through to nextMBB
8399 newMBB->addSuccessor(nextMBB);
8400 newMBB->addSuccessor(newMBB);
8402 // Insert instructions into newMBB based on incoming instruction
8403 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8404 "unexpected number of operands");
8405 DebugLoc dl = bInstr->getDebugLoc();
8406 MachineOperand& destOper = bInstr->getOperand(0);
8407 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8408 int numArgs = bInstr->getNumOperands() - 1;
8409 for (int i=0; i < numArgs; ++i)
8410 argOpers[i] = &bInstr->getOperand(i+1);
8412 // x86 address has 4 operands: base, index, scale, and displacement
8413 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8414 int valArgIndx = lastAddrIndx + 1;
8416 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8417 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8418 for (int i=0; i <= lastAddrIndx; ++i)
8419 (*MIB).addOperand(*argOpers[i]);
8421 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8423 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8428 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8429 assert((argOpers[valArgIndx]->isReg() ||
8430 argOpers[valArgIndx]->isImm()) &&
8432 if (argOpers[valArgIndx]->isReg())
8433 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8435 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8437 (*MIB).addOperand(*argOpers[valArgIndx]);
8439 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
8442 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8443 for (int i=0; i <= lastAddrIndx; ++i)
8444 (*MIB).addOperand(*argOpers[i]);
8446 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8447 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8448 bInstr->memoperands_end());
8450 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8454 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8456 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8460 // private utility function: 64 bit atomics on 32 bit host.
8462 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8463 MachineBasicBlock *MBB,
8468 bool invSrc) const {
8469 // For the atomic bitwise operator, we generate
8470 // thisMBB (instructions are in pairs, except cmpxchg8b)
8471 // ld t1,t2 = [bitinstr.addr]
8473 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8474 // op t5, t6 <- out1, out2, [bitinstr.val]
8475 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8476 // mov ECX, EBX <- t5, t6
8477 // mov EAX, EDX <- t1, t2
8478 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8479 // mov t3, t4 <- EAX, EDX
8481 // result in out1, out2
8482 // fallthrough -->nextMBB
8484 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8485 const unsigned LoadOpc = X86::MOV32rm;
8486 const unsigned NotOpc = X86::NOT32r;
8487 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8488 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8489 MachineFunction::iterator MBBIter = MBB;
8492 /// First build the CFG
8493 MachineFunction *F = MBB->getParent();
8494 MachineBasicBlock *thisMBB = MBB;
8495 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8496 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8497 F->insert(MBBIter, newMBB);
8498 F->insert(MBBIter, nextMBB);
8500 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8501 nextMBB->splice(nextMBB->begin(), thisMBB,
8502 llvm::next(MachineBasicBlock::iterator(bInstr)),
8504 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8506 // Update thisMBB to fall through to newMBB
8507 thisMBB->addSuccessor(newMBB);
8509 // newMBB jumps to itself and fall through to nextMBB
8510 newMBB->addSuccessor(nextMBB);
8511 newMBB->addSuccessor(newMBB);
8513 DebugLoc dl = bInstr->getDebugLoc();
8514 // Insert instructions into newMBB based on incoming instruction
8515 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8516 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
8517 "unexpected number of operands");
8518 MachineOperand& dest1Oper = bInstr->getOperand(0);
8519 MachineOperand& dest2Oper = bInstr->getOperand(1);
8520 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8521 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
8522 argOpers[i] = &bInstr->getOperand(i+2);
8524 // We use some of the operands multiple times, so conservatively just
8525 // clear any kill flags that might be present.
8526 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8527 argOpers[i]->setIsKill(false);
8530 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8531 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8533 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8534 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8535 for (int i=0; i <= lastAddrIndx; ++i)
8536 (*MIB).addOperand(*argOpers[i]);
8537 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8538 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8539 // add 4 to displacement.
8540 for (int i=0; i <= lastAddrIndx-2; ++i)
8541 (*MIB).addOperand(*argOpers[i]);
8542 MachineOperand newOp3 = *(argOpers[3]);
8544 newOp3.setImm(newOp3.getImm()+4);
8546 newOp3.setOffset(newOp3.getOffset()+4);
8547 (*MIB).addOperand(newOp3);
8548 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8550 // t3/4 are defined later, at the bottom of the loop
8551 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8552 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8553 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8554 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8555 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8556 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8558 // The subsequent operations should be using the destination registers of
8559 //the PHI instructions.
8561 t1 = F->getRegInfo().createVirtualRegister(RC);
8562 t2 = F->getRegInfo().createVirtualRegister(RC);
8563 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8564 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8566 t1 = dest1Oper.getReg();
8567 t2 = dest2Oper.getReg();
8570 int valArgIndx = lastAddrIndx + 1;
8571 assert((argOpers[valArgIndx]->isReg() ||
8572 argOpers[valArgIndx]->isImm()) &&
8574 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8575 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8576 if (argOpers[valArgIndx]->isReg())
8577 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8579 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8580 if (regOpcL != X86::MOV32rr)
8582 (*MIB).addOperand(*argOpers[valArgIndx]);
8583 assert(argOpers[valArgIndx + 1]->isReg() ==
8584 argOpers[valArgIndx]->isReg());
8585 assert(argOpers[valArgIndx + 1]->isImm() ==
8586 argOpers[valArgIndx]->isImm());
8587 if (argOpers[valArgIndx + 1]->isReg())
8588 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8590 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8591 if (regOpcH != X86::MOV32rr)
8593 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8595 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8597 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
8600 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
8602 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
8605 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8606 for (int i=0; i <= lastAddrIndx; ++i)
8607 (*MIB).addOperand(*argOpers[i]);
8609 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8610 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8611 bInstr->memoperands_end());
8613 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
8614 MIB.addReg(X86::EAX);
8615 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
8616 MIB.addReg(X86::EDX);
8619 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8621 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8625 // private utility function
8627 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8628 MachineBasicBlock *MBB,
8629 unsigned cmovOpc) const {
8630 // For the atomic min/max operator, we generate
8633 // ld t1 = [min/max.addr]
8634 // mov t2 = [min/max.val]
8636 // cmov[cond] t2 = t1
8638 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8640 // fallthrough -->nextMBB
8642 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8643 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8644 MachineFunction::iterator MBBIter = MBB;
8647 /// First build the CFG
8648 MachineFunction *F = MBB->getParent();
8649 MachineBasicBlock *thisMBB = MBB;
8650 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8651 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8652 F->insert(MBBIter, newMBB);
8653 F->insert(MBBIter, nextMBB);
8655 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8656 nextMBB->splice(nextMBB->begin(), thisMBB,
8657 llvm::next(MachineBasicBlock::iterator(mInstr)),
8659 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8661 // Update thisMBB to fall through to newMBB
8662 thisMBB->addSuccessor(newMBB);
8664 // newMBB jumps to newMBB and fall through to nextMBB
8665 newMBB->addSuccessor(nextMBB);
8666 newMBB->addSuccessor(newMBB);
8668 DebugLoc dl = mInstr->getDebugLoc();
8669 // Insert instructions into newMBB based on incoming instruction
8670 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8671 "unexpected number of operands");
8672 MachineOperand& destOper = mInstr->getOperand(0);
8673 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8674 int numArgs = mInstr->getNumOperands() - 1;
8675 for (int i=0; i < numArgs; ++i)
8676 argOpers[i] = &mInstr->getOperand(i+1);
8678 // x86 address has 4 operands: base, index, scale, and displacement
8679 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8680 int valArgIndx = lastAddrIndx + 1;
8682 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8683 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8684 for (int i=0; i <= lastAddrIndx; ++i)
8685 (*MIB).addOperand(*argOpers[i]);
8687 // We only support register and immediate values
8688 assert((argOpers[valArgIndx]->isReg() ||
8689 argOpers[valArgIndx]->isImm()) &&
8692 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8693 if (argOpers[valArgIndx]->isReg())
8694 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
8696 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8697 (*MIB).addOperand(*argOpers[valArgIndx]);
8699 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8702 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8707 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8708 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8712 // Cmp and exchange if none has modified the memory location
8713 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8714 for (int i=0; i <= lastAddrIndx; ++i)
8715 (*MIB).addOperand(*argOpers[i]);
8717 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8718 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8719 mInstr->memoperands_end());
8721 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8722 MIB.addReg(X86::EAX);
8725 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8727 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
8731 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8732 // or XMM0_V32I8 in AVX all of this code can be replaced with that
8735 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8736 unsigned numArgs, bool memArg) const {
8738 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
8739 "Target must have SSE4.2 or AVX features enabled");
8741 DebugLoc dl = MI->getDebugLoc();
8742 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8746 if (!Subtarget->hasAVX()) {
8748 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8750 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8753 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
8755 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
8758 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8760 for (unsigned i = 0; i < numArgs; ++i) {
8761 MachineOperand &Op = MI->getOperand(i+1);
8763 if (!(Op.isReg() && Op.isImplicit()))
8767 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8770 MI->eraseFromParent();
8776 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8778 MachineBasicBlock *MBB) const {
8779 // Emit code to save XMM registers to the stack. The ABI says that the
8780 // number of registers to save is given in %al, so it's theoretically
8781 // possible to do an indirect jump trick to avoid saving all of them,
8782 // however this code takes a simpler approach and just executes all
8783 // of the stores if %al is non-zero. It's less code, and it's probably
8784 // easier on the hardware branch predictor, and stores aren't all that
8785 // expensive anyway.
8787 // Create the new basic blocks. One block contains all the XMM stores,
8788 // and one block is the final destination regardless of whether any
8789 // stores were performed.
8790 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8791 MachineFunction *F = MBB->getParent();
8792 MachineFunction::iterator MBBIter = MBB;
8794 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8795 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8796 F->insert(MBBIter, XMMSaveMBB);
8797 F->insert(MBBIter, EndMBB);
8799 // Transfer the remainder of MBB and its successor edges to EndMBB.
8800 EndMBB->splice(EndMBB->begin(), MBB,
8801 llvm::next(MachineBasicBlock::iterator(MI)),
8803 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
8805 // The original block will now fall through to the XMM save block.
8806 MBB->addSuccessor(XMMSaveMBB);
8807 // The XMMSaveMBB will fall through to the end block.
8808 XMMSaveMBB->addSuccessor(EndMBB);
8810 // Now add the instructions.
8811 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8812 DebugLoc DL = MI->getDebugLoc();
8814 unsigned CountReg = MI->getOperand(0).getReg();
8815 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8816 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8818 if (!Subtarget->isTargetWin64()) {
8819 // If %al is 0, branch around the XMM save block.
8820 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8821 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8822 MBB->addSuccessor(EndMBB);
8825 // In the XMM save block, save all the XMM argument registers.
8826 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8827 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8828 MachineMemOperand *MMO =
8829 F->getMachineMemOperand(
8830 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8831 MachineMemOperand::MOStore, Offset,
8832 /*Size=*/16, /*Align=*/16);
8833 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8834 .addFrameIndex(RegSaveFrameIndex)
8835 .addImm(/*Scale=*/1)
8836 .addReg(/*IndexReg=*/0)
8837 .addImm(/*Disp=*/Offset)
8838 .addReg(/*Segment=*/0)
8839 .addReg(MI->getOperand(i).getReg())
8840 .addMemOperand(MMO);
8843 MI->eraseFromParent(); // The pseudo instruction is gone now.
8849 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8850 MachineBasicBlock *BB) const {
8851 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8852 DebugLoc DL = MI->getDebugLoc();
8854 // To "insert" a SELECT_CC instruction, we actually have to insert the
8855 // diamond control-flow pattern. The incoming instruction knows the
8856 // destination vreg to set, the condition code register to branch on, the
8857 // true/false values to select between, and a branch opcode to use.
8858 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8859 MachineFunction::iterator It = BB;
8865 // cmpTY ccX, r1, r2
8867 // fallthrough --> copy0MBB
8868 MachineBasicBlock *thisMBB = BB;
8869 MachineFunction *F = BB->getParent();
8870 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8871 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8872 F->insert(It, copy0MBB);
8873 F->insert(It, sinkMBB);
8875 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8876 // live into the sink and copy blocks.
8877 const MachineFunction *MF = BB->getParent();
8878 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8879 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8881 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
8882 const MachineOperand &MO = MI->getOperand(I);
8883 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
8884 unsigned Reg = MO.getReg();
8885 if (Reg != X86::EFLAGS) continue;
8886 copy0MBB->addLiveIn(Reg);
8887 sinkMBB->addLiveIn(Reg);
8890 // Transfer the remainder of BB and its successor edges to sinkMBB.
8891 sinkMBB->splice(sinkMBB->begin(), BB,
8892 llvm::next(MachineBasicBlock::iterator(MI)),
8894 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
8896 // Add the true and fallthrough blocks as its successors.
8897 BB->addSuccessor(copy0MBB);
8898 BB->addSuccessor(sinkMBB);
8900 // Create the conditional branch instruction.
8902 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8903 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8906 // %FalseValue = ...
8907 // # fallthrough to sinkMBB
8908 copy0MBB->addSuccessor(sinkMBB);
8911 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8913 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
8914 TII->get(X86::PHI), MI->getOperand(0).getReg())
8915 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8916 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8918 MI->eraseFromParent(); // The pseudo instruction is gone now.
8923 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8924 MachineBasicBlock *BB) const {
8925 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8926 DebugLoc DL = MI->getDebugLoc();
8928 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8929 // non-trivial part is impdef of ESP.
8930 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8933 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
8934 .addExternalSymbol("_alloca")
8935 .addReg(X86::EAX, RegState::Implicit)
8936 .addReg(X86::ESP, RegState::Implicit)
8937 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8938 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
8939 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
8941 MI->eraseFromParent(); // The pseudo instruction is gone now.
8946 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8947 MachineBasicBlock *BB) const {
8948 // This is pretty easy. We're taking the value that we received from
8949 // our load from the relocation, sticking it in either RDI (x86-64)
8950 // or EAX and doing an indirect call. The return value will then
8951 // be in the normal return register.
8952 const X86InstrInfo *TII
8953 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8954 DebugLoc DL = MI->getDebugLoc();
8955 MachineFunction *F = BB->getParent();
8956 bool IsWin64 = Subtarget->isTargetWin64();
8958 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8960 if (Subtarget->is64Bit()) {
8961 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8962 TII->get(X86::MOV64rm), X86::RDI)
8964 .addImm(0).addReg(0)
8965 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8966 MI->getOperand(3).getTargetFlags())
8968 MIB = BuildMI(*BB, MI, DL, TII->get(IsWin64 ? X86::WINCALL64m : X86::CALL64m));
8969 addDirectMem(MIB, X86::RDI);
8970 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8971 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8972 TII->get(X86::MOV32rm), X86::EAX)
8974 .addImm(0).addReg(0)
8975 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8976 MI->getOperand(3).getTargetFlags())
8978 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8979 addDirectMem(MIB, X86::EAX);
8981 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8982 TII->get(X86::MOV32rm), X86::EAX)
8983 .addReg(TII->getGlobalBaseReg(F))
8984 .addImm(0).addReg(0)
8985 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8986 MI->getOperand(3).getTargetFlags())
8988 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8989 addDirectMem(MIB, X86::EAX);
8992 MI->eraseFromParent(); // The pseudo instruction is gone now.
8997 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8998 MachineBasicBlock *BB) const {
8999 switch (MI->getOpcode()) {
9000 default: assert(false && "Unexpected instr type to insert");
9001 case X86::MINGW_ALLOCA:
9002 return EmitLoweredMingwAlloca(MI, BB);
9003 case X86::TLSCall_32:
9004 case X86::TLSCall_64:
9005 return EmitLoweredTLSCall(MI, BB);
9007 case X86::CMOV_V1I64:
9008 case X86::CMOV_FR32:
9009 case X86::CMOV_FR64:
9010 case X86::CMOV_V4F32:
9011 case X86::CMOV_V2F64:
9012 case X86::CMOV_V2I64:
9013 case X86::CMOV_GR16:
9014 case X86::CMOV_GR32:
9015 case X86::CMOV_RFP32:
9016 case X86::CMOV_RFP64:
9017 case X86::CMOV_RFP80:
9018 return EmitLoweredSelect(MI, BB);
9020 case X86::FP32_TO_INT16_IN_MEM:
9021 case X86::FP32_TO_INT32_IN_MEM:
9022 case X86::FP32_TO_INT64_IN_MEM:
9023 case X86::FP64_TO_INT16_IN_MEM:
9024 case X86::FP64_TO_INT32_IN_MEM:
9025 case X86::FP64_TO_INT64_IN_MEM:
9026 case X86::FP80_TO_INT16_IN_MEM:
9027 case X86::FP80_TO_INT32_IN_MEM:
9028 case X86::FP80_TO_INT64_IN_MEM: {
9029 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9030 DebugLoc DL = MI->getDebugLoc();
9032 // Change the floating point control register to use "round towards zero"
9033 // mode when truncating to an integer value.
9034 MachineFunction *F = BB->getParent();
9035 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
9036 addFrameReference(BuildMI(*BB, MI, DL,
9037 TII->get(X86::FNSTCW16m)), CWFrameIdx);
9039 // Load the old value of the high byte of the control word...
9041 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
9042 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
9045 // Set the high part to be round to zero...
9046 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
9049 // Reload the modified control word now...
9050 addFrameReference(BuildMI(*BB, MI, DL,
9051 TII->get(X86::FLDCW16m)), CWFrameIdx);
9053 // Restore the memory image of control word to original value
9054 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
9057 // Get the X86 opcode to use.
9059 switch (MI->getOpcode()) {
9060 default: llvm_unreachable("illegal opcode!");
9061 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
9062 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
9063 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
9064 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
9065 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
9066 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
9067 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
9068 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
9069 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
9073 MachineOperand &Op = MI->getOperand(0);
9075 AM.BaseType = X86AddressMode::RegBase;
9076 AM.Base.Reg = Op.getReg();
9078 AM.BaseType = X86AddressMode::FrameIndexBase;
9079 AM.Base.FrameIndex = Op.getIndex();
9081 Op = MI->getOperand(1);
9083 AM.Scale = Op.getImm();
9084 Op = MI->getOperand(2);
9086 AM.IndexReg = Op.getImm();
9087 Op = MI->getOperand(3);
9088 if (Op.isGlobal()) {
9089 AM.GV = Op.getGlobal();
9091 AM.Disp = Op.getImm();
9093 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
9094 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
9096 // Reload the original control word now.
9097 addFrameReference(BuildMI(*BB, MI, DL,
9098 TII->get(X86::FLDCW16m)), CWFrameIdx);
9100 MI->eraseFromParent(); // The pseudo instruction is gone now.
9103 // String/text processing lowering.
9104 case X86::PCMPISTRM128REG:
9105 case X86::VPCMPISTRM128REG:
9106 return EmitPCMP(MI, BB, 3, false /* in-mem */);
9107 case X86::PCMPISTRM128MEM:
9108 case X86::VPCMPISTRM128MEM:
9109 return EmitPCMP(MI, BB, 3, true /* in-mem */);
9110 case X86::PCMPESTRM128REG:
9111 case X86::VPCMPESTRM128REG:
9112 return EmitPCMP(MI, BB, 5, false /* in mem */);
9113 case X86::PCMPESTRM128MEM:
9114 case X86::VPCMPESTRM128MEM:
9115 return EmitPCMP(MI, BB, 5, true /* in mem */);
9118 case X86::ATOMAND32:
9119 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9120 X86::AND32ri, X86::MOV32rm,
9122 X86::NOT32r, X86::EAX,
9123 X86::GR32RegisterClass);
9125 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
9126 X86::OR32ri, X86::MOV32rm,
9128 X86::NOT32r, X86::EAX,
9129 X86::GR32RegisterClass);
9130 case X86::ATOMXOR32:
9131 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
9132 X86::XOR32ri, X86::MOV32rm,
9134 X86::NOT32r, X86::EAX,
9135 X86::GR32RegisterClass);
9136 case X86::ATOMNAND32:
9137 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9138 X86::AND32ri, X86::MOV32rm,
9140 X86::NOT32r, X86::EAX,
9141 X86::GR32RegisterClass, true);
9142 case X86::ATOMMIN32:
9143 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
9144 case X86::ATOMMAX32:
9145 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
9146 case X86::ATOMUMIN32:
9147 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
9148 case X86::ATOMUMAX32:
9149 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
9151 case X86::ATOMAND16:
9152 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9153 X86::AND16ri, X86::MOV16rm,
9155 X86::NOT16r, X86::AX,
9156 X86::GR16RegisterClass);
9158 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
9159 X86::OR16ri, X86::MOV16rm,
9161 X86::NOT16r, X86::AX,
9162 X86::GR16RegisterClass);
9163 case X86::ATOMXOR16:
9164 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
9165 X86::XOR16ri, X86::MOV16rm,
9167 X86::NOT16r, X86::AX,
9168 X86::GR16RegisterClass);
9169 case X86::ATOMNAND16:
9170 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9171 X86::AND16ri, X86::MOV16rm,
9173 X86::NOT16r, X86::AX,
9174 X86::GR16RegisterClass, true);
9175 case X86::ATOMMIN16:
9176 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
9177 case X86::ATOMMAX16:
9178 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
9179 case X86::ATOMUMIN16:
9180 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
9181 case X86::ATOMUMAX16:
9182 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
9185 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9186 X86::AND8ri, X86::MOV8rm,
9188 X86::NOT8r, X86::AL,
9189 X86::GR8RegisterClass);
9191 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
9192 X86::OR8ri, X86::MOV8rm,
9194 X86::NOT8r, X86::AL,
9195 X86::GR8RegisterClass);
9197 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
9198 X86::XOR8ri, X86::MOV8rm,
9200 X86::NOT8r, X86::AL,
9201 X86::GR8RegisterClass);
9202 case X86::ATOMNAND8:
9203 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9204 X86::AND8ri, X86::MOV8rm,
9206 X86::NOT8r, X86::AL,
9207 X86::GR8RegisterClass, true);
9208 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
9209 // This group is for 64-bit host.
9210 case X86::ATOMAND64:
9211 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9212 X86::AND64ri32, X86::MOV64rm,
9214 X86::NOT64r, X86::RAX,
9215 X86::GR64RegisterClass);
9217 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
9218 X86::OR64ri32, X86::MOV64rm,
9220 X86::NOT64r, X86::RAX,
9221 X86::GR64RegisterClass);
9222 case X86::ATOMXOR64:
9223 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
9224 X86::XOR64ri32, X86::MOV64rm,
9226 X86::NOT64r, X86::RAX,
9227 X86::GR64RegisterClass);
9228 case X86::ATOMNAND64:
9229 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9230 X86::AND64ri32, X86::MOV64rm,
9232 X86::NOT64r, X86::RAX,
9233 X86::GR64RegisterClass, true);
9234 case X86::ATOMMIN64:
9235 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
9236 case X86::ATOMMAX64:
9237 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
9238 case X86::ATOMUMIN64:
9239 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
9240 case X86::ATOMUMAX64:
9241 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
9243 // This group does 64-bit operations on a 32-bit host.
9244 case X86::ATOMAND6432:
9245 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9246 X86::AND32rr, X86::AND32rr,
9247 X86::AND32ri, X86::AND32ri,
9249 case X86::ATOMOR6432:
9250 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9251 X86::OR32rr, X86::OR32rr,
9252 X86::OR32ri, X86::OR32ri,
9254 case X86::ATOMXOR6432:
9255 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9256 X86::XOR32rr, X86::XOR32rr,
9257 X86::XOR32ri, X86::XOR32ri,
9259 case X86::ATOMNAND6432:
9260 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9261 X86::AND32rr, X86::AND32rr,
9262 X86::AND32ri, X86::AND32ri,
9264 case X86::ATOMADD6432:
9265 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9266 X86::ADD32rr, X86::ADC32rr,
9267 X86::ADD32ri, X86::ADC32ri,
9269 case X86::ATOMSUB6432:
9270 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9271 X86::SUB32rr, X86::SBB32rr,
9272 X86::SUB32ri, X86::SBB32ri,
9274 case X86::ATOMSWAP6432:
9275 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9276 X86::MOV32rr, X86::MOV32rr,
9277 X86::MOV32ri, X86::MOV32ri,
9279 case X86::VASTART_SAVE_XMM_REGS:
9280 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
9284 //===----------------------------------------------------------------------===//
9285 // X86 Optimization Hooks
9286 //===----------------------------------------------------------------------===//
9288 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9292 const SelectionDAG &DAG,
9293 unsigned Depth) const {
9294 unsigned Opc = Op.getOpcode();
9295 assert((Opc >= ISD::BUILTIN_OP_END ||
9296 Opc == ISD::INTRINSIC_WO_CHAIN ||
9297 Opc == ISD::INTRINSIC_W_CHAIN ||
9298 Opc == ISD::INTRINSIC_VOID) &&
9299 "Should use MaskedValueIsZero if you don't know whether Op"
9300 " is a target node!");
9302 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
9314 // These nodes' second result is a boolean.
9315 if (Op.getResNo() == 0)
9319 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
9320 Mask.getBitWidth() - 1);
9325 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
9326 /// node is a GlobalAddress + offset.
9327 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
9328 const GlobalValue* &GA,
9329 int64_t &Offset) const {
9330 if (N->getOpcode() == X86ISD::Wrapper) {
9331 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
9332 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
9333 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
9337 return TargetLowering::isGAPlusOffset(N, GA, Offset);
9340 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
9341 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
9342 /// if the load addresses are consecutive, non-overlapping, and in the right
9344 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
9345 const TargetLowering &TLI) {
9346 DebugLoc dl = N->getDebugLoc();
9347 EVT VT = N->getValueType(0);
9348 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9350 if (VT.getSizeInBits() != 128)
9353 SmallVector<SDValue, 16> Elts;
9354 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
9355 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
9357 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
9360 /// PerformShuffleCombine - Detect vector gather/scatter index generation
9361 /// and convert it from being a bunch of shuffles and extracts to a simple
9362 /// store and scalar loads to extract the elements.
9363 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
9364 const TargetLowering &TLI) {
9365 SDValue InputVector = N->getOperand(0);
9367 // Only operate on vectors of 4 elements, where the alternative shuffling
9368 // gets to be more expensive.
9369 if (InputVector.getValueType() != MVT::v4i32)
9372 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
9373 // single use which is a sign-extend or zero-extend, and all elements are
9375 SmallVector<SDNode *, 4> Uses;
9376 unsigned ExtractedElements = 0;
9377 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
9378 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
9379 if (UI.getUse().getResNo() != InputVector.getResNo())
9382 SDNode *Extract = *UI;
9383 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9386 if (Extract->getValueType(0) != MVT::i32)
9388 if (!Extract->hasOneUse())
9390 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9391 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9393 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9396 // Record which element was extracted.
9397 ExtractedElements |=
9398 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9400 Uses.push_back(Extract);
9403 // If not all the elements were used, this may not be worthwhile.
9404 if (ExtractedElements != 15)
9407 // Ok, we've now decided to do the transformation.
9408 DebugLoc dl = InputVector.getDebugLoc();
9410 // Store the value to a temporary stack slot.
9411 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9412 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL,
9413 0, false, false, 0);
9415 // Replace each use (extract) with a load of the appropriate element.
9416 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9417 UE = Uses.end(); UI != UE; ++UI) {
9418 SDNode *Extract = *UI;
9420 // Compute the element's address.
9421 SDValue Idx = Extract->getOperand(1);
9423 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9424 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9425 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9427 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
9428 OffsetVal, StackPtr);
9431 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
9432 ScalarAddr, NULL, 0, false, false, 0);
9434 // Replace the exact with the load.
9435 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9438 // The replacement was made in place; don't return anything.
9442 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9443 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9444 const X86Subtarget *Subtarget) {
9445 DebugLoc DL = N->getDebugLoc();
9446 SDValue Cond = N->getOperand(0);
9447 // Get the LHS/RHS of the select.
9448 SDValue LHS = N->getOperand(1);
9449 SDValue RHS = N->getOperand(2);
9451 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9452 // instructions match the semantics of the common C idiom x<y?x:y but not
9453 // x<=y?x:y, because of how they handle negative zero (which can be
9454 // ignored in unsafe-math mode).
9455 if (Subtarget->hasSSE2() &&
9456 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9457 Cond.getOpcode() == ISD::SETCC) {
9458 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9460 unsigned Opcode = 0;
9461 // Check for x CC y ? x : y.
9462 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9463 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9467 // Converting this to a min would handle NaNs incorrectly, and swapping
9468 // the operands would cause it to handle comparisons between positive
9469 // and negative zero incorrectly.
9470 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9471 if (!UnsafeFPMath &&
9472 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9474 std::swap(LHS, RHS);
9476 Opcode = X86ISD::FMIN;
9479 // Converting this to a min would handle comparisons between positive
9480 // and negative zero incorrectly.
9481 if (!UnsafeFPMath &&
9482 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9484 Opcode = X86ISD::FMIN;
9487 // Converting this to a min would handle both negative zeros and NaNs
9488 // incorrectly, but we can swap the operands to fix both.
9489 std::swap(LHS, RHS);
9493 Opcode = X86ISD::FMIN;
9497 // Converting this to a max would handle comparisons between positive
9498 // and negative zero incorrectly.
9499 if (!UnsafeFPMath &&
9500 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9502 Opcode = X86ISD::FMAX;
9505 // Converting this to a max would handle NaNs incorrectly, and swapping
9506 // the operands would cause it to handle comparisons between positive
9507 // and negative zero incorrectly.
9508 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9509 if (!UnsafeFPMath &&
9510 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9512 std::swap(LHS, RHS);
9514 Opcode = X86ISD::FMAX;
9517 // Converting this to a max would handle both negative zeros and NaNs
9518 // incorrectly, but we can swap the operands to fix both.
9519 std::swap(LHS, RHS);
9523 Opcode = X86ISD::FMAX;
9526 // Check for x CC y ? y : x -- a min/max with reversed arms.
9527 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9528 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9532 // Converting this to a min would handle comparisons between positive
9533 // and negative zero incorrectly, and swapping the operands would
9534 // cause it to handle NaNs incorrectly.
9535 if (!UnsafeFPMath &&
9536 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9537 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9539 std::swap(LHS, RHS);
9541 Opcode = X86ISD::FMIN;
9544 // Converting this to a min would handle NaNs incorrectly.
9545 if (!UnsafeFPMath &&
9546 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9548 Opcode = X86ISD::FMIN;
9551 // Converting this to a min would handle both negative zeros and NaNs
9552 // incorrectly, but we can swap the operands to fix both.
9553 std::swap(LHS, RHS);
9557 Opcode = X86ISD::FMIN;
9561 // Converting this to a max would handle NaNs incorrectly.
9562 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9564 Opcode = X86ISD::FMAX;
9567 // Converting this to a max would handle comparisons between positive
9568 // and negative zero incorrectly, and swapping the operands would
9569 // cause it to handle NaNs incorrectly.
9570 if (!UnsafeFPMath &&
9571 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9572 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9574 std::swap(LHS, RHS);
9576 Opcode = X86ISD::FMAX;
9579 // Converting this to a max would handle both negative zeros and NaNs
9580 // incorrectly, but we can swap the operands to fix both.
9581 std::swap(LHS, RHS);
9585 Opcode = X86ISD::FMAX;
9591 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9594 // If this is a select between two integer constants, try to do some
9596 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9597 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9598 // Don't do this for crazy integer types.
9599 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9600 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9601 // so that TrueC (the true value) is larger than FalseC.
9602 bool NeedsCondInvert = false;
9604 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9605 // Efficiently invertible.
9606 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9607 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9608 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9609 NeedsCondInvert = true;
9610 std::swap(TrueC, FalseC);
9613 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9614 if (FalseC->getAPIntValue() == 0 &&
9615 TrueC->getAPIntValue().isPowerOf2()) {
9616 if (NeedsCondInvert) // Invert the condition if needed.
9617 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9618 DAG.getConstant(1, Cond.getValueType()));
9620 // Zero extend the condition if needed.
9621 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9623 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9624 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9625 DAG.getConstant(ShAmt, MVT::i8));
9628 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9629 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9630 if (NeedsCondInvert) // Invert the condition if needed.
9631 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9632 DAG.getConstant(1, Cond.getValueType()));
9634 // Zero extend the condition if needed.
9635 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9636 FalseC->getValueType(0), Cond);
9637 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9638 SDValue(FalseC, 0));
9641 // Optimize cases that will turn into an LEA instruction. This requires
9642 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9643 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9644 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9645 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9647 bool isFastMultiplier = false;
9649 switch ((unsigned char)Diff) {
9651 case 1: // result = add base, cond
9652 case 2: // result = lea base( , cond*2)
9653 case 3: // result = lea base(cond, cond*2)
9654 case 4: // result = lea base( , cond*4)
9655 case 5: // result = lea base(cond, cond*4)
9656 case 8: // result = lea base( , cond*8)
9657 case 9: // result = lea base(cond, cond*8)
9658 isFastMultiplier = true;
9663 if (isFastMultiplier) {
9664 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9665 if (NeedsCondInvert) // Invert the condition if needed.
9666 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9667 DAG.getConstant(1, Cond.getValueType()));
9669 // Zero extend the condition if needed.
9670 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9672 // Scale the condition by the difference.
9674 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9675 DAG.getConstant(Diff, Cond.getValueType()));
9677 // Add the base if non-zero.
9678 if (FalseC->getAPIntValue() != 0)
9679 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9680 SDValue(FalseC, 0));
9690 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9691 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9692 TargetLowering::DAGCombinerInfo &DCI) {
9693 DebugLoc DL = N->getDebugLoc();
9695 // If the flag operand isn't dead, don't touch this CMOV.
9696 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9699 // If this is a select between two integer constants, try to do some
9700 // optimizations. Note that the operands are ordered the opposite of SELECT
9702 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9703 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9704 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9705 // larger than FalseC (the false value).
9706 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9708 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9709 CC = X86::GetOppositeBranchCondition(CC);
9710 std::swap(TrueC, FalseC);
9713 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9714 // This is efficient for any integer data type (including i8/i16) and
9716 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9717 SDValue Cond = N->getOperand(3);
9718 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9719 DAG.getConstant(CC, MVT::i8), Cond);
9721 // Zero extend the condition if needed.
9722 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9724 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9725 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9726 DAG.getConstant(ShAmt, MVT::i8));
9727 if (N->getNumValues() == 2) // Dead flag value?
9728 return DCI.CombineTo(N, Cond, SDValue());
9732 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9733 // for any integer data type, including i8/i16.
9734 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9735 SDValue Cond = N->getOperand(3);
9736 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9737 DAG.getConstant(CC, MVT::i8), Cond);
9739 // Zero extend the condition if needed.
9740 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9741 FalseC->getValueType(0), Cond);
9742 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9743 SDValue(FalseC, 0));
9745 if (N->getNumValues() == 2) // Dead flag value?
9746 return DCI.CombineTo(N, Cond, SDValue());
9750 // Optimize cases that will turn into an LEA instruction. This requires
9751 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9752 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9753 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9754 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9756 bool isFastMultiplier = false;
9758 switch ((unsigned char)Diff) {
9760 case 1: // result = add base, cond
9761 case 2: // result = lea base( , cond*2)
9762 case 3: // result = lea base(cond, cond*2)
9763 case 4: // result = lea base( , cond*4)
9764 case 5: // result = lea base(cond, cond*4)
9765 case 8: // result = lea base( , cond*8)
9766 case 9: // result = lea base(cond, cond*8)
9767 isFastMultiplier = true;
9772 if (isFastMultiplier) {
9773 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9774 SDValue Cond = N->getOperand(3);
9775 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9776 DAG.getConstant(CC, MVT::i8), Cond);
9777 // Zero extend the condition if needed.
9778 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9780 // Scale the condition by the difference.
9782 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9783 DAG.getConstant(Diff, Cond.getValueType()));
9785 // Add the base if non-zero.
9786 if (FalseC->getAPIntValue() != 0)
9787 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9788 SDValue(FalseC, 0));
9789 if (N->getNumValues() == 2) // Dead flag value?
9790 return DCI.CombineTo(N, Cond, SDValue());
9800 /// PerformMulCombine - Optimize a single multiply with constant into two
9801 /// in order to implement it with two cheaper instructions, e.g.
9802 /// LEA + SHL, LEA + LEA.
9803 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9804 TargetLowering::DAGCombinerInfo &DCI) {
9805 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9808 EVT VT = N->getValueType(0);
9812 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9815 uint64_t MulAmt = C->getZExtValue();
9816 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9819 uint64_t MulAmt1 = 0;
9820 uint64_t MulAmt2 = 0;
9821 if ((MulAmt % 9) == 0) {
9823 MulAmt2 = MulAmt / 9;
9824 } else if ((MulAmt % 5) == 0) {
9826 MulAmt2 = MulAmt / 5;
9827 } else if ((MulAmt % 3) == 0) {
9829 MulAmt2 = MulAmt / 3;
9832 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9833 DebugLoc DL = N->getDebugLoc();
9835 if (isPowerOf2_64(MulAmt2) &&
9836 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9837 // If second multiplifer is pow2, issue it first. We want the multiply by
9838 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9840 std::swap(MulAmt1, MulAmt2);
9843 if (isPowerOf2_64(MulAmt1))
9844 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9845 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9847 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9848 DAG.getConstant(MulAmt1, VT));
9850 if (isPowerOf2_64(MulAmt2))
9851 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9852 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9854 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9855 DAG.getConstant(MulAmt2, VT));
9857 // Do not add new nodes to DAG combiner worklist.
9858 DCI.CombineTo(N, NewMul, false);
9863 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9864 SDValue N0 = N->getOperand(0);
9865 SDValue N1 = N->getOperand(1);
9866 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9867 EVT VT = N0.getValueType();
9869 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9870 // since the result of setcc_c is all zero's or all ones.
9871 if (N1C && N0.getOpcode() == ISD::AND &&
9872 N0.getOperand(1).getOpcode() == ISD::Constant) {
9873 SDValue N00 = N0.getOperand(0);
9874 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9875 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9876 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9877 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9878 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9879 APInt ShAmt = N1C->getAPIntValue();
9880 Mask = Mask.shl(ShAmt);
9882 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9883 N00, DAG.getConstant(Mask, VT));
9890 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9892 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9893 const X86Subtarget *Subtarget) {
9894 EVT VT = N->getValueType(0);
9895 if (!VT.isVector() && VT.isInteger() &&
9896 N->getOpcode() == ISD::SHL)
9897 return PerformSHLCombine(N, DAG);
9899 // On X86 with SSE2 support, we can transform this to a vector shift if
9900 // all elements are shifted by the same amount. We can't do this in legalize
9901 // because the a constant vector is typically transformed to a constant pool
9902 // so we have no knowledge of the shift amount.
9903 if (!Subtarget->hasSSE2())
9906 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9909 SDValue ShAmtOp = N->getOperand(1);
9910 EVT EltVT = VT.getVectorElementType();
9911 DebugLoc DL = N->getDebugLoc();
9912 SDValue BaseShAmt = SDValue();
9913 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9914 unsigned NumElts = VT.getVectorNumElements();
9916 for (; i != NumElts; ++i) {
9917 SDValue Arg = ShAmtOp.getOperand(i);
9918 if (Arg.getOpcode() == ISD::UNDEF) continue;
9922 for (; i != NumElts; ++i) {
9923 SDValue Arg = ShAmtOp.getOperand(i);
9924 if (Arg.getOpcode() == ISD::UNDEF) continue;
9925 if (Arg != BaseShAmt) {
9929 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9930 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9931 SDValue InVec = ShAmtOp.getOperand(0);
9932 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9933 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9935 for (; i != NumElts; ++i) {
9936 SDValue Arg = InVec.getOperand(i);
9937 if (Arg.getOpcode() == ISD::UNDEF) continue;
9941 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9942 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9943 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9944 if (C->getZExtValue() == SplatIdx)
9945 BaseShAmt = InVec.getOperand(1);
9948 if (BaseShAmt.getNode() == 0)
9949 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9950 DAG.getIntPtrConstant(0));
9954 // The shift amount is an i32.
9955 if (EltVT.bitsGT(MVT::i32))
9956 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9957 else if (EltVT.bitsLT(MVT::i32))
9958 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9960 // The shift amount is identical so we can do a vector shift.
9961 SDValue ValOp = N->getOperand(0);
9962 switch (N->getOpcode()) {
9964 llvm_unreachable("Unknown shift opcode!");
9967 if (VT == MVT::v2i64)
9968 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9969 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9971 if (VT == MVT::v4i32)
9972 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9973 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9975 if (VT == MVT::v8i16)
9976 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9977 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9981 if (VT == MVT::v4i32)
9982 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9983 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9985 if (VT == MVT::v8i16)
9986 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9987 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9991 if (VT == MVT::v2i64)
9992 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9993 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9995 if (VT == MVT::v4i32)
9996 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9997 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9999 if (VT == MVT::v8i16)
10000 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10001 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
10008 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
10009 TargetLowering::DAGCombinerInfo &DCI,
10010 const X86Subtarget *Subtarget) {
10011 if (DCI.isBeforeLegalizeOps())
10014 EVT VT = N->getValueType(0);
10015 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
10018 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
10019 SDValue N0 = N->getOperand(0);
10020 SDValue N1 = N->getOperand(1);
10021 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
10023 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
10025 if (!N0.hasOneUse() || !N1.hasOneUse())
10028 SDValue ShAmt0 = N0.getOperand(1);
10029 if (ShAmt0.getValueType() != MVT::i8)
10031 SDValue ShAmt1 = N1.getOperand(1);
10032 if (ShAmt1.getValueType() != MVT::i8)
10034 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
10035 ShAmt0 = ShAmt0.getOperand(0);
10036 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
10037 ShAmt1 = ShAmt1.getOperand(0);
10039 DebugLoc DL = N->getDebugLoc();
10040 unsigned Opc = X86ISD::SHLD;
10041 SDValue Op0 = N0.getOperand(0);
10042 SDValue Op1 = N1.getOperand(0);
10043 if (ShAmt0.getOpcode() == ISD::SUB) {
10044 Opc = X86ISD::SHRD;
10045 std::swap(Op0, Op1);
10046 std::swap(ShAmt0, ShAmt1);
10049 unsigned Bits = VT.getSizeInBits();
10050 if (ShAmt1.getOpcode() == ISD::SUB) {
10051 SDValue Sum = ShAmt1.getOperand(0);
10052 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
10053 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
10054 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
10055 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
10056 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
10057 return DAG.getNode(Opc, DL, VT,
10059 DAG.getNode(ISD::TRUNCATE, DL,
10062 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
10063 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
10065 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
10066 return DAG.getNode(Opc, DL, VT,
10067 N0.getOperand(0), N1.getOperand(0),
10068 DAG.getNode(ISD::TRUNCATE, DL,
10075 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
10076 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
10077 const X86Subtarget *Subtarget) {
10078 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
10079 // the FP state in cases where an emms may be missing.
10080 // A preferable solution to the general problem is to figure out the right
10081 // places to insert EMMS. This qualifies as a quick hack.
10083 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
10084 StoreSDNode *St = cast<StoreSDNode>(N);
10085 EVT VT = St->getValue().getValueType();
10086 if (VT.getSizeInBits() != 64)
10089 const Function *F = DAG.getMachineFunction().getFunction();
10090 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
10091 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
10092 && Subtarget->hasSSE2();
10093 if ((VT.isVector() ||
10094 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
10095 isa<LoadSDNode>(St->getValue()) &&
10096 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
10097 St->getChain().hasOneUse() && !St->isVolatile()) {
10098 SDNode* LdVal = St->getValue().getNode();
10099 LoadSDNode *Ld = 0;
10100 int TokenFactorIndex = -1;
10101 SmallVector<SDValue, 8> Ops;
10102 SDNode* ChainVal = St->getChain().getNode();
10103 // Must be a store of a load. We currently handle two cases: the load
10104 // is a direct child, and it's under an intervening TokenFactor. It is
10105 // possible to dig deeper under nested TokenFactors.
10106 if (ChainVal == LdVal)
10107 Ld = cast<LoadSDNode>(St->getChain());
10108 else if (St->getValue().hasOneUse() &&
10109 ChainVal->getOpcode() == ISD::TokenFactor) {
10110 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
10111 if (ChainVal->getOperand(i).getNode() == LdVal) {
10112 TokenFactorIndex = i;
10113 Ld = cast<LoadSDNode>(St->getValue());
10115 Ops.push_back(ChainVal->getOperand(i));
10119 if (!Ld || !ISD::isNormalLoad(Ld))
10122 // If this is not the MMX case, i.e. we are just turning i64 load/store
10123 // into f64 load/store, avoid the transformation if there are multiple
10124 // uses of the loaded value.
10125 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
10128 DebugLoc LdDL = Ld->getDebugLoc();
10129 DebugLoc StDL = N->getDebugLoc();
10130 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
10131 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
10133 if (Subtarget->is64Bit() || F64IsLegal) {
10134 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
10135 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
10136 Ld->getBasePtr(), Ld->getSrcValue(),
10137 Ld->getSrcValueOffset(), Ld->isVolatile(),
10138 Ld->isNonTemporal(), Ld->getAlignment());
10139 SDValue NewChain = NewLd.getValue(1);
10140 if (TokenFactorIndex != -1) {
10141 Ops.push_back(NewChain);
10142 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10145 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
10146 St->getSrcValue(), St->getSrcValueOffset(),
10147 St->isVolatile(), St->isNonTemporal(),
10148 St->getAlignment());
10151 // Otherwise, lower to two pairs of 32-bit loads / stores.
10152 SDValue LoAddr = Ld->getBasePtr();
10153 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
10154 DAG.getConstant(4, MVT::i32));
10156 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
10157 Ld->getSrcValue(), Ld->getSrcValueOffset(),
10158 Ld->isVolatile(), Ld->isNonTemporal(),
10159 Ld->getAlignment());
10160 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
10161 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
10162 Ld->isVolatile(), Ld->isNonTemporal(),
10163 MinAlign(Ld->getAlignment(), 4));
10165 SDValue NewChain = LoLd.getValue(1);
10166 if (TokenFactorIndex != -1) {
10167 Ops.push_back(LoLd);
10168 Ops.push_back(HiLd);
10169 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10173 LoAddr = St->getBasePtr();
10174 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
10175 DAG.getConstant(4, MVT::i32));
10177 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
10178 St->getSrcValue(), St->getSrcValueOffset(),
10179 St->isVolatile(), St->isNonTemporal(),
10180 St->getAlignment());
10181 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
10183 St->getSrcValueOffset() + 4,
10185 St->isNonTemporal(),
10186 MinAlign(St->getAlignment(), 4));
10187 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
10192 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
10193 /// X86ISD::FXOR nodes.
10194 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
10195 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
10196 // F[X]OR(0.0, x) -> x
10197 // F[X]OR(x, 0.0) -> x
10198 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10199 if (C->getValueAPF().isPosZero())
10200 return N->getOperand(1);
10201 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10202 if (C->getValueAPF().isPosZero())
10203 return N->getOperand(0);
10207 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
10208 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
10209 // FAND(0.0, x) -> 0.0
10210 // FAND(x, 0.0) -> 0.0
10211 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10212 if (C->getValueAPF().isPosZero())
10213 return N->getOperand(0);
10214 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10215 if (C->getValueAPF().isPosZero())
10216 return N->getOperand(1);
10220 static SDValue PerformBTCombine(SDNode *N,
10222 TargetLowering::DAGCombinerInfo &DCI) {
10223 // BT ignores high bits in the bit index operand.
10224 SDValue Op1 = N->getOperand(1);
10225 if (Op1.hasOneUse()) {
10226 unsigned BitWidth = Op1.getValueSizeInBits();
10227 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
10228 APInt KnownZero, KnownOne;
10229 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
10230 !DCI.isBeforeLegalizeOps());
10231 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10232 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
10233 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
10234 DCI.CommitTargetLoweringOpt(TLO);
10239 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
10240 SDValue Op = N->getOperand(0);
10241 if (Op.getOpcode() == ISD::BIT_CONVERT)
10242 Op = Op.getOperand(0);
10243 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
10244 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
10245 VT.getVectorElementType().getSizeInBits() ==
10246 OpVT.getVectorElementType().getSizeInBits()) {
10247 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
10252 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
10253 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
10254 // (and (i32 x86isd::setcc_carry), 1)
10255 // This eliminates the zext. This transformation is necessary because
10256 // ISD::SETCC is always legalized to i8.
10257 DebugLoc dl = N->getDebugLoc();
10258 SDValue N0 = N->getOperand(0);
10259 EVT VT = N->getValueType(0);
10260 if (N0.getOpcode() == ISD::AND &&
10262 N0.getOperand(0).hasOneUse()) {
10263 SDValue N00 = N0.getOperand(0);
10264 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
10266 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
10267 if (!C || C->getZExtValue() != 1)
10269 return DAG.getNode(ISD::AND, dl, VT,
10270 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
10271 N00.getOperand(0), N00.getOperand(1)),
10272 DAG.getConstant(1, VT));
10278 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
10279 DAGCombinerInfo &DCI) const {
10280 SelectionDAG &DAG = DCI.DAG;
10281 switch (N->getOpcode()) {
10283 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
10284 case ISD::EXTRACT_VECTOR_ELT:
10285 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
10286 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
10287 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
10288 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
10291 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
10292 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
10293 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
10295 case X86ISD::FOR: return PerformFORCombine(N, DAG);
10296 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
10297 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
10298 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
10299 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
10305 /// isTypeDesirableForOp - Return true if the target has native support for
10306 /// the specified value type and it is 'desirable' to use the type for the
10307 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
10308 /// instruction encodings are longer and some i16 instructions are slow.
10309 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
10310 if (!isTypeLegal(VT))
10312 if (VT != MVT::i16)
10319 case ISD::SIGN_EXTEND:
10320 case ISD::ZERO_EXTEND:
10321 case ISD::ANY_EXTEND:
10334 static bool MayFoldLoad(SDValue Op) {
10335 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
10338 static bool MayFoldIntoStore(SDValue Op) {
10339 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
10342 /// IsDesirableToPromoteOp - This method query the target whether it is
10343 /// beneficial for dag combiner to promote the specified node. If true, it
10344 /// should return the desired promotion type by reference.
10345 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
10346 EVT VT = Op.getValueType();
10347 if (VT != MVT::i16)
10350 bool Promote = false;
10351 bool Commute = false;
10352 switch (Op.getOpcode()) {
10355 LoadSDNode *LD = cast<LoadSDNode>(Op);
10356 // If the non-extending load has a single use and it's not live out, then it
10357 // might be folded.
10358 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
10359 Op.hasOneUse()*/) {
10360 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
10361 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
10362 // The only case where we'd want to promote LOAD (rather then it being
10363 // promoted as an operand is when it's only use is liveout.
10364 if (UI->getOpcode() != ISD::CopyToReg)
10371 case ISD::SIGN_EXTEND:
10372 case ISD::ZERO_EXTEND:
10373 case ISD::ANY_EXTEND:
10378 SDValue N0 = Op.getOperand(0);
10379 // Look out for (store (shl (load), x)).
10380 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10393 SDValue N0 = Op.getOperand(0);
10394 SDValue N1 = Op.getOperand(1);
10395 if (!Commute && MayFoldLoad(N1))
10397 // Avoid disabling potential load folding opportunities.
10398 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10400 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10410 //===----------------------------------------------------------------------===//
10411 // X86 Inline Assembly Support
10412 //===----------------------------------------------------------------------===//
10414 static bool LowerToBSwap(CallInst *CI) {
10415 // FIXME: this should verify that we are targetting a 486 or better. If not,
10416 // we will turn this bswap into something that will be lowered to logical ops
10417 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10418 // so don't worry about this.
10420 // Verify this is a simple bswap.
10421 if (CI->getNumArgOperands() != 1 ||
10422 CI->getType() != CI->getArgOperand(0)->getType() ||
10423 !CI->getType()->isIntegerTy())
10426 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10427 if (!Ty || Ty->getBitWidth() % 16 != 0)
10430 // Okay, we can do this xform, do so now.
10431 const Type *Tys[] = { Ty };
10432 Module *M = CI->getParent()->getParent()->getParent();
10433 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10435 Value *Op = CI->getArgOperand(0);
10436 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10438 CI->replaceAllUsesWith(Op);
10439 CI->eraseFromParent();
10443 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10444 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10445 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10447 std::string AsmStr = IA->getAsmString();
10449 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10450 SmallVector<StringRef, 4> AsmPieces;
10451 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10453 switch (AsmPieces.size()) {
10454 default: return false;
10456 AsmStr = AsmPieces[0];
10458 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10461 if (AsmPieces.size() == 2 &&
10462 (AsmPieces[0] == "bswap" ||
10463 AsmPieces[0] == "bswapq" ||
10464 AsmPieces[0] == "bswapl") &&
10465 (AsmPieces[1] == "$0" ||
10466 AsmPieces[1] == "${0:q}")) {
10467 // No need to check constraints, nothing other than the equivalent of
10468 // "=r,0" would be valid here.
10469 return LowerToBSwap(CI);
10471 // rorw $$8, ${0:w} --> llvm.bswap.i16
10472 if (CI->getType()->isIntegerTy(16) &&
10473 AsmPieces.size() == 3 &&
10474 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10475 AsmPieces[1] == "$$8," &&
10476 AsmPieces[2] == "${0:w}" &&
10477 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10479 const std::string &Constraints = IA->getConstraintString();
10480 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10481 std::sort(AsmPieces.begin(), AsmPieces.end());
10482 if (AsmPieces.size() == 4 &&
10483 AsmPieces[0] == "~{cc}" &&
10484 AsmPieces[1] == "~{dirflag}" &&
10485 AsmPieces[2] == "~{flags}" &&
10486 AsmPieces[3] == "~{fpsr}") {
10487 return LowerToBSwap(CI);
10492 if (CI->getType()->isIntegerTy(64) &&
10493 Constraints.size() >= 2 &&
10494 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10495 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10496 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10497 SmallVector<StringRef, 4> Words;
10498 SplitString(AsmPieces[0], Words, " \t");
10499 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10501 SplitString(AsmPieces[1], Words, " \t");
10502 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10504 SplitString(AsmPieces[2], Words, " \t,");
10505 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10506 Words[2] == "%edx") {
10507 return LowerToBSwap(CI);
10519 /// getConstraintType - Given a constraint letter, return the type of
10520 /// constraint it is for this target.
10521 X86TargetLowering::ConstraintType
10522 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10523 if (Constraint.size() == 1) {
10524 switch (Constraint[0]) {
10536 return C_RegisterClass;
10544 return TargetLowering::getConstraintType(Constraint);
10547 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10548 /// with another that has more specific requirements based on the type of the
10549 /// corresponding operand.
10550 const char *X86TargetLowering::
10551 LowerXConstraint(EVT ConstraintVT) const {
10552 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10553 // 'f' like normal targets.
10554 if (ConstraintVT.isFloatingPoint()) {
10555 if (Subtarget->hasSSE2())
10557 if (Subtarget->hasSSE1())
10561 return TargetLowering::LowerXConstraint(ConstraintVT);
10564 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10565 /// vector. If it is invalid, don't add anything to Ops.
10566 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10568 std::vector<SDValue>&Ops,
10569 SelectionDAG &DAG) const {
10570 SDValue Result(0, 0);
10572 switch (Constraint) {
10575 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10576 if (C->getZExtValue() <= 31) {
10577 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10583 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10584 if (C->getZExtValue() <= 63) {
10585 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10591 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10592 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10593 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10599 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10600 if (C->getZExtValue() <= 255) {
10601 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10607 // 32-bit signed value
10608 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10609 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10610 C->getSExtValue())) {
10611 // Widen to 64 bits here to get it sign extended.
10612 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10615 // FIXME gcc accepts some relocatable values here too, but only in certain
10616 // memory models; it's complicated.
10621 // 32-bit unsigned value
10622 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10623 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10624 C->getZExtValue())) {
10625 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10629 // FIXME gcc accepts some relocatable values here too, but only in certain
10630 // memory models; it's complicated.
10634 // Literal immediates are always ok.
10635 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10636 // Widen to 64 bits here to get it sign extended.
10637 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10641 // In any sort of PIC mode addresses need to be computed at runtime by
10642 // adding in a register or some sort of table lookup. These can't
10643 // be used as immediates.
10644 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
10647 // If we are in non-pic codegen mode, we allow the address of a global (with
10648 // an optional displacement) to be used with 'i'.
10649 GlobalAddressSDNode *GA = 0;
10650 int64_t Offset = 0;
10652 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10654 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10655 Offset += GA->getOffset();
10657 } else if (Op.getOpcode() == ISD::ADD) {
10658 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10659 Offset += C->getZExtValue();
10660 Op = Op.getOperand(0);
10663 } else if (Op.getOpcode() == ISD::SUB) {
10664 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10665 Offset += -C->getZExtValue();
10666 Op = Op.getOperand(0);
10671 // Otherwise, this isn't something we can handle, reject it.
10675 const GlobalValue *GV = GA->getGlobal();
10676 // If we require an extra load to get this address, as in PIC mode, we
10677 // can't accept it.
10678 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10679 getTargetMachine())))
10682 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
10683 GA->getValueType(0), Offset);
10688 if (Result.getNode()) {
10689 Ops.push_back(Result);
10692 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10695 std::vector<unsigned> X86TargetLowering::
10696 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10698 if (Constraint.size() == 1) {
10699 // FIXME: not handling fp-stack yet!
10700 switch (Constraint[0]) { // GCC X86 Constraint Letters
10701 default: break; // Unknown constraint letter
10702 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10703 if (Subtarget->is64Bit()) {
10704 if (VT == MVT::i32)
10705 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10706 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10707 X86::R10D,X86::R11D,X86::R12D,
10708 X86::R13D,X86::R14D,X86::R15D,
10709 X86::EBP, X86::ESP, 0);
10710 else if (VT == MVT::i16)
10711 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10712 X86::SI, X86::DI, X86::R8W,X86::R9W,
10713 X86::R10W,X86::R11W,X86::R12W,
10714 X86::R13W,X86::R14W,X86::R15W,
10715 X86::BP, X86::SP, 0);
10716 else if (VT == MVT::i8)
10717 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10718 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10719 X86::R10B,X86::R11B,X86::R12B,
10720 X86::R13B,X86::R14B,X86::R15B,
10721 X86::BPL, X86::SPL, 0);
10723 else if (VT == MVT::i64)
10724 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10725 X86::RSI, X86::RDI, X86::R8, X86::R9,
10726 X86::R10, X86::R11, X86::R12,
10727 X86::R13, X86::R14, X86::R15,
10728 X86::RBP, X86::RSP, 0);
10732 // 32-bit fallthrough
10733 case 'Q': // Q_REGS
10734 if (VT == MVT::i32)
10735 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10736 else if (VT == MVT::i16)
10737 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10738 else if (VT == MVT::i8)
10739 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10740 else if (VT == MVT::i64)
10741 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10746 return std::vector<unsigned>();
10749 std::pair<unsigned, const TargetRegisterClass*>
10750 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10752 // First, see if this is a constraint that directly corresponds to an LLVM
10754 if (Constraint.size() == 1) {
10755 // GCC Constraint Letters
10756 switch (Constraint[0]) {
10758 case 'r': // GENERAL_REGS
10759 case 'l': // INDEX_REGS
10761 return std::make_pair(0U, X86::GR8RegisterClass);
10762 if (VT == MVT::i16)
10763 return std::make_pair(0U, X86::GR16RegisterClass);
10764 if (VT == MVT::i32 || !Subtarget->is64Bit())
10765 return std::make_pair(0U, X86::GR32RegisterClass);
10766 return std::make_pair(0U, X86::GR64RegisterClass);
10767 case 'R': // LEGACY_REGS
10769 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10770 if (VT == MVT::i16)
10771 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10772 if (VT == MVT::i32 || !Subtarget->is64Bit())
10773 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10774 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10775 case 'f': // FP Stack registers.
10776 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10777 // value to the correct fpstack register class.
10778 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10779 return std::make_pair(0U, X86::RFP32RegisterClass);
10780 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10781 return std::make_pair(0U, X86::RFP64RegisterClass);
10782 return std::make_pair(0U, X86::RFP80RegisterClass);
10783 case 'y': // MMX_REGS if MMX allowed.
10784 if (!Subtarget->hasMMX()) break;
10785 return std::make_pair(0U, X86::VR64RegisterClass);
10786 case 'Y': // SSE_REGS if SSE2 allowed
10787 if (!Subtarget->hasSSE2()) break;
10789 case 'x': // SSE_REGS if SSE1 allowed
10790 if (!Subtarget->hasSSE1()) break;
10792 switch (VT.getSimpleVT().SimpleTy) {
10794 // Scalar SSE types.
10797 return std::make_pair(0U, X86::FR32RegisterClass);
10800 return std::make_pair(0U, X86::FR64RegisterClass);
10808 return std::make_pair(0U, X86::VR128RegisterClass);
10814 // Use the default implementation in TargetLowering to convert the register
10815 // constraint into a member of a register class.
10816 std::pair<unsigned, const TargetRegisterClass*> Res;
10817 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10819 // Not found as a standard register?
10820 if (Res.second == 0) {
10821 // Map st(0) -> st(7) -> ST0
10822 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10823 tolower(Constraint[1]) == 's' &&
10824 tolower(Constraint[2]) == 't' &&
10825 Constraint[3] == '(' &&
10826 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10827 Constraint[5] == ')' &&
10828 Constraint[6] == '}') {
10830 Res.first = X86::ST0+Constraint[4]-'0';
10831 Res.second = X86::RFP80RegisterClass;
10835 // GCC allows "st(0)" to be called just plain "st".
10836 if (StringRef("{st}").equals_lower(Constraint)) {
10837 Res.first = X86::ST0;
10838 Res.second = X86::RFP80RegisterClass;
10843 if (StringRef("{flags}").equals_lower(Constraint)) {
10844 Res.first = X86::EFLAGS;
10845 Res.second = X86::CCRRegisterClass;
10849 // 'A' means EAX + EDX.
10850 if (Constraint == "A") {
10851 Res.first = X86::EAX;
10852 Res.second = X86::GR32_ADRegisterClass;
10858 // Otherwise, check to see if this is a register class of the wrong value
10859 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10860 // turn into {ax},{dx}.
10861 if (Res.second->hasType(VT))
10862 return Res; // Correct type already, nothing to do.
10864 // All of the single-register GCC register classes map their values onto
10865 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10866 // really want an 8-bit or 32-bit register, map to the appropriate register
10867 // class and return the appropriate register.
10868 if (Res.second == X86::GR16RegisterClass) {
10869 if (VT == MVT::i8) {
10870 unsigned DestReg = 0;
10871 switch (Res.first) {
10873 case X86::AX: DestReg = X86::AL; break;
10874 case X86::DX: DestReg = X86::DL; break;
10875 case X86::CX: DestReg = X86::CL; break;
10876 case X86::BX: DestReg = X86::BL; break;
10879 Res.first = DestReg;
10880 Res.second = X86::GR8RegisterClass;
10882 } else if (VT == MVT::i32) {
10883 unsigned DestReg = 0;
10884 switch (Res.first) {
10886 case X86::AX: DestReg = X86::EAX; break;
10887 case X86::DX: DestReg = X86::EDX; break;
10888 case X86::CX: DestReg = X86::ECX; break;
10889 case X86::BX: DestReg = X86::EBX; break;
10890 case X86::SI: DestReg = X86::ESI; break;
10891 case X86::DI: DestReg = X86::EDI; break;
10892 case X86::BP: DestReg = X86::EBP; break;
10893 case X86::SP: DestReg = X86::ESP; break;
10896 Res.first = DestReg;
10897 Res.second = X86::GR32RegisterClass;
10899 } else if (VT == MVT::i64) {
10900 unsigned DestReg = 0;
10901 switch (Res.first) {
10903 case X86::AX: DestReg = X86::RAX; break;
10904 case X86::DX: DestReg = X86::RDX; break;
10905 case X86::CX: DestReg = X86::RCX; break;
10906 case X86::BX: DestReg = X86::RBX; break;
10907 case X86::SI: DestReg = X86::RSI; break;
10908 case X86::DI: DestReg = X86::RDI; break;
10909 case X86::BP: DestReg = X86::RBP; break;
10910 case X86::SP: DestReg = X86::RSP; break;
10913 Res.first = DestReg;
10914 Res.second = X86::GR64RegisterClass;
10917 } else if (Res.second == X86::FR32RegisterClass ||
10918 Res.second == X86::FR64RegisterClass ||
10919 Res.second == X86::VR128RegisterClass) {
10920 // Handle references to XMM physical registers that got mapped into the
10921 // wrong class. This can happen with constraints like {xmm0} where the
10922 // target independent register mapper will just pick the first match it can
10923 // find, ignoring the required type.
10924 if (VT == MVT::f32)
10925 Res.second = X86::FR32RegisterClass;
10926 else if (VT == MVT::f64)
10927 Res.second = X86::FR64RegisterClass;
10928 else if (X86::VR128RegisterClass->hasType(VT))
10929 Res.second = X86::VR128RegisterClass;