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 "X86ShuffleDecode.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineJumpTableInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/PseudoSourceValue.h"
38 #include "llvm/MC/MCAsmInfo.h"
39 #include "llvm/MC/MCContext.h"
40 #include "llvm/MC/MCExpr.h"
41 #include "llvm/MC/MCSymbol.h"
42 #include "llvm/ADT/BitVector.h"
43 #include "llvm/ADT/SmallSet.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/ADT/StringExtras.h"
46 #include "llvm/ADT/VectorExtras.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/Dwarf.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
54 using namespace dwarf;
56 STATISTIC(NumTailCalls, "Number of tail calls");
59 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
61 // Forward declarations.
62 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
65 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
67 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
69 if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
70 if (is64Bit) return new X8664_MachoTargetObjectFile();
71 return new TargetLoweringObjectFileMachO();
72 } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
73 if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
74 return new X8632_ELFTargetObjectFile(TM);
75 } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
76 return new TargetLoweringObjectFileCOFF();
78 llvm_unreachable("unknown subtarget type");
81 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
82 : TargetLowering(TM, createTLOF(TM)) {
83 Subtarget = &TM.getSubtarget<X86Subtarget>();
84 X86ScalarSSEf64 = Subtarget->hasSSE2();
85 X86ScalarSSEf32 = Subtarget->hasSSE1();
86 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
88 RegInfo = TM.getRegisterInfo();
91 // Set up the TargetLowering object.
93 // X86 is weird, it always uses i8 for shift amounts and setcc results.
94 setShiftAmountType(MVT::i8);
95 setBooleanContents(ZeroOrOneBooleanContent);
96 setSchedulingPreference(Sched::RegPressure);
97 setStackPointerRegisterToSaveRestore(X86StackPtr);
99 if (Subtarget->isTargetDarwin()) {
100 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
101 setUseUnderscoreSetJmp(false);
102 setUseUnderscoreLongJmp(false);
103 } else if (Subtarget->isTargetMingw()) {
104 // MS runtime is weird: it exports _setjmp, but longjmp!
105 setUseUnderscoreSetJmp(true);
106 setUseUnderscoreLongJmp(false);
108 setUseUnderscoreSetJmp(true);
109 setUseUnderscoreLongJmp(true);
112 // Set up the register classes.
113 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
114 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
115 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
116 if (Subtarget->is64Bit())
117 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
119 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
121 // We don't accept any truncstore of integer registers.
122 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
123 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
124 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
125 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
126 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
127 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
129 // SETOEQ and SETUNE require checking two conditions.
130 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
131 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
132 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
133 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
135 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
137 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
139 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
140 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
141 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
143 if (Subtarget->is64Bit()) {
144 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
145 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
146 } else if (!UseSoftFloat) {
147 // We have an algorithm for SSE2->double, and we turn this into a
148 // 64-bit FILD followed by conditional FADD for other targets.
149 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
150 // We have an algorithm for SSE2, and we turn this into a 64-bit
151 // FILD for other targets.
152 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
155 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
157 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
158 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
161 // SSE has no i16 to fp conversion, only i32
162 if (X86ScalarSSEf32) {
163 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
164 // f32 and f64 cases are Legal, f80 case is not
165 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
167 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
168 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
171 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
172 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
175 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
176 // are Legal, f80 is custom lowered.
177 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
178 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
180 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
182 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
183 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
185 if (X86ScalarSSEf32) {
186 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
187 // f32 and f64 cases are Legal, f80 case is not
188 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
190 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
191 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
194 // Handle FP_TO_UINT by promoting the destination to a larger signed
196 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
197 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
198 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
200 if (Subtarget->is64Bit()) {
201 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
202 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
203 } else if (!UseSoftFloat) {
204 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
205 // Expand FP_TO_UINT into a select.
206 // FIXME: We would like to use a Custom expander here eventually to do
207 // the optimal thing for SSE vs. the default expansion in the legalizer.
208 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
210 // With SSE3 we can use fisttpll to convert to a signed i64; without
211 // SSE, we're stuck with a fistpll.
212 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
215 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
216 if (!X86ScalarSSEf64) {
217 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
218 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
219 if (Subtarget->is64Bit()) {
220 setOperationAction(ISD::BIT_CONVERT , MVT::f64 , Expand);
221 // Without SSE, i64->f64 goes through memory; i64->MMX is Legal.
222 if (Subtarget->hasMMX() && !DisableMMX)
223 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Custom);
225 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Expand);
229 // Scalar integer divide and remainder are lowered to use operations that
230 // produce two results, to match the available instructions. This exposes
231 // the two-result form to trivial CSE, which is able to combine x/y and x%y
232 // into a single instruction.
234 // Scalar integer multiply-high is also lowered to use two-result
235 // operations, to match the available instructions. However, plain multiply
236 // (low) operations are left as Legal, as there are single-result
237 // instructions for this in x86. Using the two-result multiply instructions
238 // when both high and low results are needed must be arranged by dagcombine.
239 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
240 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
241 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
242 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
243 setOperationAction(ISD::SREM , MVT::i8 , Expand);
244 setOperationAction(ISD::UREM , MVT::i8 , Expand);
245 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
246 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
247 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
248 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
249 setOperationAction(ISD::SREM , MVT::i16 , Expand);
250 setOperationAction(ISD::UREM , MVT::i16 , Expand);
251 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
252 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
253 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
254 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
255 setOperationAction(ISD::SREM , MVT::i32 , Expand);
256 setOperationAction(ISD::UREM , MVT::i32 , Expand);
257 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
258 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
259 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
260 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
261 setOperationAction(ISD::SREM , MVT::i64 , Expand);
262 setOperationAction(ISD::UREM , MVT::i64 , Expand);
264 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
265 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
266 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
267 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
268 if (Subtarget->is64Bit())
269 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
270 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
272 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
273 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
274 setOperationAction(ISD::FREM , MVT::f32 , Expand);
275 setOperationAction(ISD::FREM , MVT::f64 , Expand);
276 setOperationAction(ISD::FREM , MVT::f80 , Expand);
277 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
279 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
280 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
281 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
282 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
283 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
284 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
285 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
286 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
287 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
288 if (Subtarget->is64Bit()) {
289 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
290 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
291 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
294 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
295 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
297 // These should be promoted to a larger select which is supported.
298 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
299 // X86 wants to expand cmov itself.
300 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
301 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
302 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
303 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
304 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
305 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
306 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
307 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
308 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
309 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
310 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
311 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
312 if (Subtarget->is64Bit()) {
313 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
314 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
316 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
319 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
320 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
321 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
322 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
323 if (Subtarget->is64Bit())
324 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
325 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
326 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
327 if (Subtarget->is64Bit()) {
328 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
329 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
330 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
331 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
332 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
334 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
335 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
336 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
337 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
338 if (Subtarget->is64Bit()) {
339 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
340 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
341 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
344 if (Subtarget->hasSSE1())
345 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
347 // We may not have a libcall for MEMBARRIER so we should lower this.
348 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
350 // On X86 and X86-64, atomic operations are lowered to locked instructions.
351 // Locked instructions, in turn, have implicit fence semantics (all memory
352 // operations are flushed before issuing the locked instruction, and they
353 // are not buffered), so we can fold away the common pattern of
354 // fence-atomic-fence.
355 setShouldFoldAtomicFences(true);
357 // Expand certain atomics
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
360 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
361 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
365 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
366 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
368 if (!Subtarget->is64Bit()) {
369 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
374 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
375 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
378 // FIXME - use subtarget debug flags
379 if (!Subtarget->isTargetDarwin() &&
380 !Subtarget->isTargetELF() &&
381 !Subtarget->isTargetCygMing()) {
382 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
385 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
386 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
387 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
388 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
389 if (Subtarget->is64Bit()) {
390 setExceptionPointerRegister(X86::RAX);
391 setExceptionSelectorRegister(X86::RDX);
393 setExceptionPointerRegister(X86::EAX);
394 setExceptionSelectorRegister(X86::EDX);
396 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
397 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
399 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
401 setOperationAction(ISD::TRAP, MVT::Other, Legal);
403 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
404 setOperationAction(ISD::VASTART , MVT::Other, Custom);
405 setOperationAction(ISD::VAEND , MVT::Other, Expand);
406 if (Subtarget->is64Bit()) {
407 setOperationAction(ISD::VAARG , MVT::Other, Custom);
408 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
410 setOperationAction(ISD::VAARG , MVT::Other, Expand);
411 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
414 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
415 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
416 if (Subtarget->is64Bit())
417 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
418 if (Subtarget->isTargetCygMing())
419 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
421 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
423 if (!UseSoftFloat && X86ScalarSSEf64) {
424 // f32 and f64 use SSE.
425 // Set up the FP register classes.
426 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
427 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
429 // Use ANDPD to simulate FABS.
430 setOperationAction(ISD::FABS , MVT::f64, Custom);
431 setOperationAction(ISD::FABS , MVT::f32, Custom);
433 // Use XORP to simulate FNEG.
434 setOperationAction(ISD::FNEG , MVT::f64, Custom);
435 setOperationAction(ISD::FNEG , MVT::f32, Custom);
437 // Use ANDPD and ORPD to simulate FCOPYSIGN.
438 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
439 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
441 // We don't support sin/cos/fmod
442 setOperationAction(ISD::FSIN , MVT::f64, Expand);
443 setOperationAction(ISD::FCOS , MVT::f64, Expand);
444 setOperationAction(ISD::FSIN , MVT::f32, Expand);
445 setOperationAction(ISD::FCOS , MVT::f32, Expand);
447 // Expand FP immediates into loads from the stack, except for the special
449 addLegalFPImmediate(APFloat(+0.0)); // xorpd
450 addLegalFPImmediate(APFloat(+0.0f)); // xorps
451 } else if (!UseSoftFloat && X86ScalarSSEf32) {
452 // Use SSE for f32, x87 for f64.
453 // Set up the FP register classes.
454 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
455 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
457 // Use ANDPS to simulate FABS.
458 setOperationAction(ISD::FABS , MVT::f32, Custom);
460 // Use XORP to simulate FNEG.
461 setOperationAction(ISD::FNEG , MVT::f32, Custom);
463 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
465 // Use ANDPS and ORPS to simulate FCOPYSIGN.
466 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
467 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
469 // We don't support sin/cos/fmod
470 setOperationAction(ISD::FSIN , MVT::f32, Expand);
471 setOperationAction(ISD::FCOS , MVT::f32, Expand);
473 // Special cases we handle for FP constants.
474 addLegalFPImmediate(APFloat(+0.0f)); // xorps
475 addLegalFPImmediate(APFloat(+0.0)); // FLD0
476 addLegalFPImmediate(APFloat(+1.0)); // FLD1
477 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
478 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
481 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
482 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
484 } else if (!UseSoftFloat) {
485 // f32 and f64 in x87.
486 // Set up the FP register classes.
487 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
488 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
490 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
491 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
492 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
493 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
496 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
497 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
499 addLegalFPImmediate(APFloat(+0.0)); // FLD0
500 addLegalFPImmediate(APFloat(+1.0)); // FLD1
501 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
502 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
503 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
504 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
505 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
506 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
509 // Long double always uses X87.
511 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
512 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
513 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
516 APFloat TmpFlt(+0.0);
517 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
519 addLegalFPImmediate(TmpFlt); // FLD0
521 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
522 APFloat TmpFlt2(+1.0);
523 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
525 addLegalFPImmediate(TmpFlt2); // FLD1
526 TmpFlt2.changeSign();
527 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
531 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
532 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
536 // Always use a library call for pow.
537 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
538 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
539 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
541 setOperationAction(ISD::FLOG, MVT::f80, Expand);
542 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
543 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
544 setOperationAction(ISD::FEXP, MVT::f80, Expand);
545 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
547 // First set operation action for all vector types to either promote
548 // (for widening) or expand (for scalarization). Then we will selectively
549 // turn on ones that can be effectively codegen'd.
550 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
551 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
552 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
567 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
568 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
601 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
605 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
606 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
607 setTruncStoreAction((MVT::SimpleValueType)VT,
608 (MVT::SimpleValueType)InnerVT, Expand);
609 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
610 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
611 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
614 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
615 // with -msoft-float, disable use of MMX as well.
616 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
617 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
618 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
619 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
621 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
623 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
624 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
625 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
626 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
628 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
629 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
630 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
631 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
633 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
634 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
636 setOperationAction(ISD::AND, MVT::v8i8, Promote);
637 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
638 setOperationAction(ISD::AND, MVT::v4i16, Promote);
639 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
640 setOperationAction(ISD::AND, MVT::v2i32, Promote);
641 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
642 setOperationAction(ISD::AND, MVT::v1i64, Legal);
644 setOperationAction(ISD::OR, MVT::v8i8, Promote);
645 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
646 setOperationAction(ISD::OR, MVT::v4i16, Promote);
647 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
648 setOperationAction(ISD::OR, MVT::v2i32, Promote);
649 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
650 setOperationAction(ISD::OR, MVT::v1i64, Legal);
652 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
653 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
654 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
655 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
656 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
657 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
658 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
660 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
661 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
662 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
663 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
664 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
665 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
666 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
669 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
670 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
671 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
674 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
675 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
676 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
678 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
679 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
680 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
682 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
684 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
685 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
686 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
687 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
689 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
690 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
692 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
693 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
694 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
695 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
696 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
700 if (!UseSoftFloat && Subtarget->hasSSE1()) {
701 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
703 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
704 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
705 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
706 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
707 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
708 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
709 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
710 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
711 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
712 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
713 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
714 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
717 if (!UseSoftFloat && Subtarget->hasSSE2()) {
718 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
720 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
721 // registers cannot be used even for integer operations.
722 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
723 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
724 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
725 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
727 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
728 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
729 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
730 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
731 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
732 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
733 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
734 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
735 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
736 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
737 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
738 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
739 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
740 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
741 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
742 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
744 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
745 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
746 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
747 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
749 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
750 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
751 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
752 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
753 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
755 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
756 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
757 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
759 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
761 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
762 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
763 EVT VT = (MVT::SimpleValueType)i;
764 // Do not attempt to custom lower non-power-of-2 vectors
765 if (!isPowerOf2_32(VT.getVectorNumElements()))
767 // Do not attempt to custom lower non-128-bit vectors
768 if (!VT.is128BitVector())
770 setOperationAction(ISD::BUILD_VECTOR,
771 VT.getSimpleVT().SimpleTy, Custom);
772 setOperationAction(ISD::VECTOR_SHUFFLE,
773 VT.getSimpleVT().SimpleTy, Custom);
774 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
775 VT.getSimpleVT().SimpleTy, Custom);
778 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
779 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
780 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
781 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
782 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
783 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
785 if (Subtarget->is64Bit()) {
786 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
787 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
790 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
791 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
792 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
795 // Do not attempt to promote non-128-bit vectors
796 if (!VT.is128BitVector())
799 setOperationAction(ISD::AND, SVT, Promote);
800 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
801 setOperationAction(ISD::OR, SVT, Promote);
802 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
803 setOperationAction(ISD::XOR, SVT, Promote);
804 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
805 setOperationAction(ISD::LOAD, SVT, Promote);
806 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
807 setOperationAction(ISD::SELECT, SVT, Promote);
808 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
811 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
813 // Custom lower v2i64 and v2f64 selects.
814 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
815 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
816 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
817 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
819 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
820 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
821 if (!DisableMMX && Subtarget->hasMMX()) {
822 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
823 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
827 if (Subtarget->hasSSE41()) {
828 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
829 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
830 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
831 setOperationAction(ISD::FRINT, MVT::f32, Legal);
832 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
833 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
834 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
835 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
836 setOperationAction(ISD::FRINT, MVT::f64, Legal);
837 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
839 // FIXME: Do we need to handle scalar-to-vector here?
840 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
842 // Can turn SHL into an integer multiply.
843 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
844 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
846 // i8 and i16 vectors are custom , because the source register and source
847 // source memory operand types are not the same width. f32 vectors are
848 // custom since the immediate controlling the insert encodes additional
850 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
851 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
852 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
853 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
855 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
856 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
857 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
858 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
860 if (Subtarget->is64Bit()) {
861 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
862 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
866 if (Subtarget->hasSSE42()) {
867 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
870 if (!UseSoftFloat && Subtarget->hasAVX()) {
871 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
872 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
873 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
874 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
875 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
877 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
878 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
879 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
880 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
881 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
882 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
883 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
884 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
885 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
886 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
887 setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
888 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
889 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
890 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
891 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
893 // Operations to consider commented out -v16i16 v32i8
894 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
895 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
896 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
897 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
898 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
899 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
900 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
901 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
902 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
903 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
904 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
905 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
906 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
907 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
909 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
910 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
911 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
912 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
914 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
915 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
916 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
917 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
918 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
920 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
921 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
922 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
923 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
924 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
925 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
928 // Not sure we want to do this since there are no 256-bit integer
931 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
932 // This includes 256-bit vectors
933 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
934 EVT VT = (MVT::SimpleValueType)i;
936 // Do not attempt to custom lower non-power-of-2 vectors
937 if (!isPowerOf2_32(VT.getVectorNumElements()))
940 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
941 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
942 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
945 if (Subtarget->is64Bit()) {
946 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
947 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
952 // Not sure we want to do this since there are no 256-bit integer
955 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
956 // Including 256-bit vectors
957 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
958 EVT VT = (MVT::SimpleValueType)i;
960 if (!VT.is256BitVector()) {
963 setOperationAction(ISD::AND, VT, Promote);
964 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
965 setOperationAction(ISD::OR, VT, Promote);
966 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
967 setOperationAction(ISD::XOR, VT, Promote);
968 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
969 setOperationAction(ISD::LOAD, VT, Promote);
970 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
971 setOperationAction(ISD::SELECT, VT, Promote);
972 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
975 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
979 // We want to custom lower some of our intrinsics.
980 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
982 // Add/Sub/Mul with overflow operations are custom lowered.
983 setOperationAction(ISD::SADDO, MVT::i32, Custom);
984 setOperationAction(ISD::UADDO, MVT::i32, Custom);
985 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
986 setOperationAction(ISD::USUBO, MVT::i32, Custom);
987 setOperationAction(ISD::SMULO, MVT::i32, Custom);
989 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
990 // handle type legalization for these operations here.
992 // FIXME: We really should do custom legalization for addition and
993 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
994 // than generic legalization for 64-bit multiplication-with-overflow, though.
995 if (Subtarget->is64Bit()) {
996 setOperationAction(ISD::SADDO, MVT::i64, Custom);
997 setOperationAction(ISD::UADDO, MVT::i64, Custom);
998 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
999 setOperationAction(ISD::USUBO, MVT::i64, Custom);
1000 setOperationAction(ISD::SMULO, MVT::i64, Custom);
1003 if (!Subtarget->is64Bit()) {
1004 // These libcalls are not available in 32-bit.
1005 setLibcallName(RTLIB::SHL_I128, 0);
1006 setLibcallName(RTLIB::SRL_I128, 0);
1007 setLibcallName(RTLIB::SRA_I128, 0);
1010 // We have target-specific dag combine patterns for the following nodes:
1011 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1012 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1013 setTargetDAGCombine(ISD::BUILD_VECTOR);
1014 setTargetDAGCombine(ISD::SELECT);
1015 setTargetDAGCombine(ISD::SHL);
1016 setTargetDAGCombine(ISD::SRA);
1017 setTargetDAGCombine(ISD::SRL);
1018 setTargetDAGCombine(ISD::OR);
1019 setTargetDAGCombine(ISD::STORE);
1020 setTargetDAGCombine(ISD::ZERO_EXTEND);
1021 if (Subtarget->is64Bit())
1022 setTargetDAGCombine(ISD::MUL);
1024 computeRegisterProperties();
1026 // FIXME: These should be based on subtarget info. Plus, the values should
1027 // be smaller when we are in optimizing for size mode.
1028 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1029 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1030 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1031 setPrefLoopAlignment(16);
1032 benefitFromCodePlacementOpt = true;
1036 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1041 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1042 /// the desired ByVal argument alignment.
1043 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1046 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1047 if (VTy->getBitWidth() == 128)
1049 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1050 unsigned EltAlign = 0;
1051 getMaxByValAlign(ATy->getElementType(), EltAlign);
1052 if (EltAlign > MaxAlign)
1053 MaxAlign = EltAlign;
1054 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1055 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1056 unsigned EltAlign = 0;
1057 getMaxByValAlign(STy->getElementType(i), EltAlign);
1058 if (EltAlign > MaxAlign)
1059 MaxAlign = EltAlign;
1067 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1068 /// function arguments in the caller parameter area. For X86, aggregates
1069 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1070 /// are at 4-byte boundaries.
1071 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1072 if (Subtarget->is64Bit()) {
1073 // Max of 8 and alignment of type.
1074 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1081 if (Subtarget->hasSSE1())
1082 getMaxByValAlign(Ty, Align);
1086 /// getOptimalMemOpType - Returns the target specific optimal type for load
1087 /// and store operations as a result of memset, memcpy, and memmove
1088 /// lowering. If DstAlign is zero that means it's safe to destination
1089 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1090 /// means there isn't a need to check it against alignment requirement,
1091 /// probably because the source does not need to be loaded. If
1092 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1093 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1094 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1095 /// constant so it does not need to be loaded.
1096 /// It returns EVT::Other if the type should be determined using generic
1097 /// target-independent logic.
1099 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1100 unsigned DstAlign, unsigned SrcAlign,
1101 bool NonScalarIntSafe,
1103 MachineFunction &MF) const {
1104 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1105 // linux. This is because the stack realignment code can't handle certain
1106 // cases like PR2962. This should be removed when PR2962 is fixed.
1107 const Function *F = MF.getFunction();
1108 if (NonScalarIntSafe &&
1109 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1111 (Subtarget->isUnalignedMemAccessFast() ||
1112 ((DstAlign == 0 || DstAlign >= 16) &&
1113 (SrcAlign == 0 || SrcAlign >= 16))) &&
1114 Subtarget->getStackAlignment() >= 16) {
1115 if (Subtarget->hasSSE2())
1117 if (Subtarget->hasSSE1())
1119 } else if (!MemcpyStrSrc && Size >= 8 &&
1120 !Subtarget->is64Bit() &&
1121 Subtarget->getStackAlignment() >= 8 &&
1122 Subtarget->hasSSE2()) {
1123 // Do not use f64 to lower memcpy if source is string constant. It's
1124 // better to use i32 to avoid the loads.
1128 if (Subtarget->is64Bit() && Size >= 8)
1133 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1134 /// current function. The returned value is a member of the
1135 /// MachineJumpTableInfo::JTEntryKind enum.
1136 unsigned X86TargetLowering::getJumpTableEncoding() const {
1137 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1139 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1140 Subtarget->isPICStyleGOT())
1141 return MachineJumpTableInfo::EK_Custom32;
1143 // Otherwise, use the normal jump table encoding heuristics.
1144 return TargetLowering::getJumpTableEncoding();
1147 /// getPICBaseSymbol - Return the X86-32 PIC base.
1149 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1150 MCContext &Ctx) const {
1151 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1152 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1153 Twine(MF->getFunctionNumber())+"$pb");
1158 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1159 const MachineBasicBlock *MBB,
1160 unsigned uid,MCContext &Ctx) const{
1161 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1162 Subtarget->isPICStyleGOT());
1163 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1165 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1166 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1169 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1171 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1172 SelectionDAG &DAG) const {
1173 if (!Subtarget->is64Bit())
1174 // This doesn't have DebugLoc associated with it, but is not really the
1175 // same as a Register.
1176 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1180 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1181 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1183 const MCExpr *X86TargetLowering::
1184 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1185 MCContext &Ctx) const {
1186 // X86-64 uses RIP relative addressing based on the jump table label.
1187 if (Subtarget->isPICStyleRIPRel())
1188 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1190 // Otherwise, the reference is relative to the PIC base.
1191 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1194 /// getFunctionAlignment - Return the Log2 alignment of this function.
1195 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1196 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1199 std::pair<const TargetRegisterClass*, uint8_t>
1200 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1201 const TargetRegisterClass *RRC = 0;
1203 switch (VT.getSimpleVT().SimpleTy) {
1205 return TargetLowering::findRepresentativeClass(VT);
1206 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1207 RRC = (Subtarget->is64Bit()
1208 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1210 case MVT::v8i8: case MVT::v4i16:
1211 case MVT::v2i32: case MVT::v1i64:
1212 RRC = X86::VR64RegisterClass;
1214 case MVT::f32: case MVT::f64:
1215 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1216 case MVT::v4f32: case MVT::v2f64:
1217 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1219 RRC = X86::VR128RegisterClass;
1222 return std::make_pair(RRC, Cost);
1226 X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
1227 MachineFunction &MF) const {
1228 unsigned FPDiff = RegInfo->hasFP(MF) ? 1 : 0;
1229 switch (RC->getID()) {
1232 case X86::GR32RegClassID:
1234 case X86::GR64RegClassID:
1236 case X86::VR128RegClassID:
1237 return Subtarget->is64Bit() ? 10 : 4;
1238 case X86::VR64RegClassID:
1243 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1244 unsigned &Offset) const {
1245 if (!Subtarget->isTargetLinux())
1248 if (Subtarget->is64Bit()) {
1249 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1251 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1264 //===----------------------------------------------------------------------===//
1265 // Return Value Calling Convention Implementation
1266 //===----------------------------------------------------------------------===//
1268 #include "X86GenCallingConv.inc"
1271 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1272 const SmallVectorImpl<ISD::OutputArg> &Outs,
1273 LLVMContext &Context) const {
1274 SmallVector<CCValAssign, 16> RVLocs;
1275 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1277 return CCInfo.CheckReturn(Outs, RetCC_X86);
1281 X86TargetLowering::LowerReturn(SDValue Chain,
1282 CallingConv::ID CallConv, bool isVarArg,
1283 const SmallVectorImpl<ISD::OutputArg> &Outs,
1284 const SmallVectorImpl<SDValue> &OutVals,
1285 DebugLoc dl, SelectionDAG &DAG) const {
1286 MachineFunction &MF = DAG.getMachineFunction();
1287 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1289 SmallVector<CCValAssign, 16> RVLocs;
1290 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1291 RVLocs, *DAG.getContext());
1292 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1294 // Add the regs to the liveout set for the function.
1295 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1296 for (unsigned i = 0; i != RVLocs.size(); ++i)
1297 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1298 MRI.addLiveOut(RVLocs[i].getLocReg());
1302 SmallVector<SDValue, 6> RetOps;
1303 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1304 // Operand #1 = Bytes To Pop
1305 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1308 // Copy the result values into the output registers.
1309 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1310 CCValAssign &VA = RVLocs[i];
1311 assert(VA.isRegLoc() && "Can only return in registers!");
1312 SDValue ValToCopy = OutVals[i];
1313 EVT ValVT = ValToCopy.getValueType();
1315 // If this is x86-64, and we disabled SSE, we can't return FP values
1316 if ((ValVT == MVT::f32 || ValVT == MVT::f64) &&
1317 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1318 report_fatal_error("SSE register return with SSE disabled");
1320 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1321 // llvm-gcc has never done it right and no one has noticed, so this
1322 // should be OK for now.
1323 if (ValVT == MVT::f64 &&
1324 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1325 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 if (isVarArg && Subtarget->isTargetWin64()) {
2013 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2014 // shadow reg if callee is a varargs function.
2015 unsigned ShadowReg = 0;
2016 switch (VA.getLocReg()) {
2017 case X86::XMM0: ShadowReg = X86::RCX; break;
2018 case X86::XMM1: ShadowReg = X86::RDX; break;
2019 case X86::XMM2: ShadowReg = X86::R8; break;
2020 case X86::XMM3: ShadowReg = X86::R9; break;
2023 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2025 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2026 assert(VA.isMemLoc());
2027 if (StackPtr.getNode() == 0)
2028 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2029 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2030 dl, DAG, VA, Flags));
2034 if (!MemOpChains.empty())
2035 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2036 &MemOpChains[0], MemOpChains.size());
2038 // Build a sequence of copy-to-reg nodes chained together with token chain
2039 // and flag operands which copy the outgoing args into registers.
2041 // Tail call byval lowering might overwrite argument registers so in case of
2042 // tail call optimization the copies to registers are lowered later.
2044 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2045 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2046 RegsToPass[i].second, InFlag);
2047 InFlag = Chain.getValue(1);
2050 if (Subtarget->isPICStyleGOT()) {
2051 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2054 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2055 DAG.getNode(X86ISD::GlobalBaseReg,
2056 DebugLoc(), getPointerTy()),
2058 InFlag = Chain.getValue(1);
2060 // If we are tail calling and generating PIC/GOT style code load the
2061 // address of the callee into ECX. The value in ecx is used as target of
2062 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2063 // for tail calls on PIC/GOT architectures. Normally we would just put the
2064 // address of GOT into ebx and then call target@PLT. But for tail calls
2065 // ebx would be restored (since ebx is callee saved) before jumping to the
2068 // Note: The actual moving to ECX is done further down.
2069 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2070 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2071 !G->getGlobal()->hasProtectedVisibility())
2072 Callee = LowerGlobalAddress(Callee, DAG);
2073 else if (isa<ExternalSymbolSDNode>(Callee))
2074 Callee = LowerExternalSymbol(Callee, DAG);
2078 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
2079 // From AMD64 ABI document:
2080 // For calls that may call functions that use varargs or stdargs
2081 // (prototype-less calls or calls to functions containing ellipsis (...) in
2082 // the declaration) %al is used as hidden argument to specify the number
2083 // of SSE registers used. The contents of %al do not need to match exactly
2084 // the number of registers, but must be an ubound on the number of SSE
2085 // registers used and is in the range 0 - 8 inclusive.
2087 // Count the number of XMM registers allocated.
2088 static const unsigned XMMArgRegs[] = {
2089 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2090 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2092 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2093 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2094 && "SSE registers cannot be used when SSE is disabled");
2096 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2097 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2098 InFlag = Chain.getValue(1);
2102 // For tail calls lower the arguments to the 'real' stack slot.
2104 // Force all the incoming stack arguments to be loaded from the stack
2105 // before any new outgoing arguments are stored to the stack, because the
2106 // outgoing stack slots may alias the incoming argument stack slots, and
2107 // the alias isn't otherwise explicit. This is slightly more conservative
2108 // than necessary, because it means that each store effectively depends
2109 // on every argument instead of just those arguments it would clobber.
2110 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2112 SmallVector<SDValue, 8> MemOpChains2;
2115 // Do not flag preceeding copytoreg stuff together with the following stuff.
2117 if (GuaranteedTailCallOpt) {
2118 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2119 CCValAssign &VA = ArgLocs[i];
2122 assert(VA.isMemLoc());
2123 SDValue Arg = OutVals[i];
2124 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2125 // Create frame index.
2126 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2127 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2128 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2129 FIN = DAG.getFrameIndex(FI, getPointerTy());
2131 if (Flags.isByVal()) {
2132 // Copy relative to framepointer.
2133 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2134 if (StackPtr.getNode() == 0)
2135 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2137 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2139 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2143 // Store relative to framepointer.
2144 MemOpChains2.push_back(
2145 DAG.getStore(ArgChain, dl, Arg, FIN,
2146 PseudoSourceValue::getFixedStack(FI), 0,
2152 if (!MemOpChains2.empty())
2153 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2154 &MemOpChains2[0], MemOpChains2.size());
2156 // Copy arguments to their registers.
2157 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2158 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2159 RegsToPass[i].second, InFlag);
2160 InFlag = Chain.getValue(1);
2164 // Store the return address to the appropriate stack slot.
2165 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2169 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2170 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2171 // In the 64-bit large code model, we have to make all calls
2172 // through a register, since the call instruction's 32-bit
2173 // pc-relative offset may not be large enough to hold the whole
2175 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2176 // If the callee is a GlobalAddress node (quite common, every direct call
2177 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2180 // We should use extra load for direct calls to dllimported functions in
2182 const GlobalValue *GV = G->getGlobal();
2183 if (!GV->hasDLLImportLinkage()) {
2184 unsigned char OpFlags = 0;
2186 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2187 // external symbols most go through the PLT in PIC mode. If the symbol
2188 // has hidden or protected visibility, or if it is static or local, then
2189 // we don't need to use the PLT - we can directly call it.
2190 if (Subtarget->isTargetELF() &&
2191 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2192 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2193 OpFlags = X86II::MO_PLT;
2194 } else if (Subtarget->isPICStyleStubAny() &&
2195 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2196 Subtarget->getDarwinVers() < 9) {
2197 // PC-relative references to external symbols should go through $stub,
2198 // unless we're building with the leopard linker or later, which
2199 // automatically synthesizes these stubs.
2200 OpFlags = X86II::MO_DARWIN_STUB;
2203 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2204 G->getOffset(), OpFlags);
2206 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2207 unsigned char OpFlags = 0;
2209 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2210 // symbols should go through the PLT.
2211 if (Subtarget->isTargetELF() &&
2212 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2213 OpFlags = X86II::MO_PLT;
2214 } else if (Subtarget->isPICStyleStubAny() &&
2215 Subtarget->getDarwinVers() < 9) {
2216 // PC-relative references to external symbols should go through $stub,
2217 // unless we're building with the leopard linker or later, which
2218 // automatically synthesizes these stubs.
2219 OpFlags = X86II::MO_DARWIN_STUB;
2222 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2226 // Returns a chain & a flag for retval copy to use.
2227 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2228 SmallVector<SDValue, 8> Ops;
2230 if (!IsSibcall && isTailCall) {
2231 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2232 DAG.getIntPtrConstant(0, true), InFlag);
2233 InFlag = Chain.getValue(1);
2236 Ops.push_back(Chain);
2237 Ops.push_back(Callee);
2240 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2242 // Add argument registers to the end of the list so that they are known live
2244 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2245 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2246 RegsToPass[i].second.getValueType()));
2248 // Add an implicit use GOT pointer in EBX.
2249 if (!isTailCall && Subtarget->isPICStyleGOT())
2250 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2252 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2253 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64())
2254 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2256 if (InFlag.getNode())
2257 Ops.push_back(InFlag);
2261 //// If this is the first return lowered for this function, add the regs
2262 //// to the liveout set for the function.
2263 // This isn't right, although it's probably harmless on x86; liveouts
2264 // should be computed from returns not tail calls. Consider a void
2265 // function making a tail call to a function returning int.
2266 return DAG.getNode(X86ISD::TC_RETURN, dl,
2267 NodeTys, &Ops[0], Ops.size());
2270 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2271 InFlag = Chain.getValue(1);
2273 // Create the CALLSEQ_END node.
2274 unsigned NumBytesForCalleeToPush;
2275 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2276 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2277 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2278 // If this is a call to a struct-return function, the callee
2279 // pops the hidden struct pointer, so we have to push it back.
2280 // This is common for Darwin/X86, Linux & Mingw32 targets.
2281 NumBytesForCalleeToPush = 4;
2283 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2285 // Returns a flag for retval copy to use.
2287 Chain = DAG.getCALLSEQ_END(Chain,
2288 DAG.getIntPtrConstant(NumBytes, true),
2289 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2292 InFlag = Chain.getValue(1);
2295 // Handle result values, copying them out of physregs into vregs that we
2297 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2298 Ins, dl, DAG, InVals);
2302 //===----------------------------------------------------------------------===//
2303 // Fast Calling Convention (tail call) implementation
2304 //===----------------------------------------------------------------------===//
2306 // Like std call, callee cleans arguments, convention except that ECX is
2307 // reserved for storing the tail called function address. Only 2 registers are
2308 // free for argument passing (inreg). Tail call optimization is performed
2310 // * tailcallopt is enabled
2311 // * caller/callee are fastcc
2312 // On X86_64 architecture with GOT-style position independent code only local
2313 // (within module) calls are supported at the moment.
2314 // To keep the stack aligned according to platform abi the function
2315 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2316 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2317 // If a tail called function callee has more arguments than the caller the
2318 // caller needs to make sure that there is room to move the RETADDR to. This is
2319 // achieved by reserving an area the size of the argument delta right after the
2320 // original REtADDR, but before the saved framepointer or the spilled registers
2321 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2333 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2334 /// for a 16 byte align requirement.
2336 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2337 SelectionDAG& DAG) const {
2338 MachineFunction &MF = DAG.getMachineFunction();
2339 const TargetMachine &TM = MF.getTarget();
2340 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2341 unsigned StackAlignment = TFI.getStackAlignment();
2342 uint64_t AlignMask = StackAlignment - 1;
2343 int64_t Offset = StackSize;
2344 uint64_t SlotSize = TD->getPointerSize();
2345 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2346 // Number smaller than 12 so just add the difference.
2347 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2349 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2350 Offset = ((~AlignMask) & Offset) + StackAlignment +
2351 (StackAlignment-SlotSize);
2356 /// MatchingStackOffset - Return true if the given stack call argument is
2357 /// already available in the same position (relatively) of the caller's
2358 /// incoming argument stack.
2360 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2361 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2362 const X86InstrInfo *TII) {
2363 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2365 if (Arg.getOpcode() == ISD::CopyFromReg) {
2366 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2367 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2369 MachineInstr *Def = MRI->getVRegDef(VR);
2372 if (!Flags.isByVal()) {
2373 if (!TII->isLoadFromStackSlot(Def, FI))
2376 unsigned Opcode = Def->getOpcode();
2377 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2378 Def->getOperand(1).isFI()) {
2379 FI = Def->getOperand(1).getIndex();
2380 Bytes = Flags.getByValSize();
2384 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2385 if (Flags.isByVal())
2386 // ByVal argument is passed in as a pointer but it's now being
2387 // dereferenced. e.g.
2388 // define @foo(%struct.X* %A) {
2389 // tail call @bar(%struct.X* byval %A)
2392 SDValue Ptr = Ld->getBasePtr();
2393 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2396 FI = FINode->getIndex();
2400 assert(FI != INT_MAX);
2401 if (!MFI->isFixedObjectIndex(FI))
2403 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2406 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2407 /// for tail call optimization. Targets which want to do tail call
2408 /// optimization should implement this function.
2410 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2411 CallingConv::ID CalleeCC,
2413 bool isCalleeStructRet,
2414 bool isCallerStructRet,
2415 const SmallVectorImpl<ISD::OutputArg> &Outs,
2416 const SmallVectorImpl<SDValue> &OutVals,
2417 const SmallVectorImpl<ISD::InputArg> &Ins,
2418 SelectionDAG& DAG) const {
2419 if (!IsTailCallConvention(CalleeCC) &&
2420 CalleeCC != CallingConv::C)
2423 // If -tailcallopt is specified, make fastcc functions tail-callable.
2424 const MachineFunction &MF = DAG.getMachineFunction();
2425 const Function *CallerF = DAG.getMachineFunction().getFunction();
2426 CallingConv::ID CallerCC = CallerF->getCallingConv();
2427 bool CCMatch = CallerCC == CalleeCC;
2429 if (GuaranteedTailCallOpt) {
2430 if (IsTailCallConvention(CalleeCC) && CCMatch)
2435 // Look for obvious safe cases to perform tail call optimization that do not
2436 // require ABI changes. This is what gcc calls sibcall.
2438 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2439 // emit a special epilogue.
2440 if (RegInfo->needsStackRealignment(MF))
2443 // Do not sibcall optimize vararg calls unless the call site is not passing
2445 if (isVarArg && !Outs.empty())
2448 // Also avoid sibcall optimization if either caller or callee uses struct
2449 // return semantics.
2450 if (isCalleeStructRet || isCallerStructRet)
2453 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2454 // Therefore if it's not used by the call it is not safe to optimize this into
2456 bool Unused = false;
2457 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2464 SmallVector<CCValAssign, 16> RVLocs;
2465 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2466 RVLocs, *DAG.getContext());
2467 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2468 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2469 CCValAssign &VA = RVLocs[i];
2470 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2475 // If the calling conventions do not match, then we'd better make sure the
2476 // results are returned in the same way as what the caller expects.
2478 SmallVector<CCValAssign, 16> RVLocs1;
2479 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2480 RVLocs1, *DAG.getContext());
2481 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2483 SmallVector<CCValAssign, 16> RVLocs2;
2484 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2485 RVLocs2, *DAG.getContext());
2486 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2488 if (RVLocs1.size() != RVLocs2.size())
2490 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2491 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2493 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2495 if (RVLocs1[i].isRegLoc()) {
2496 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2499 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2505 // If the callee takes no arguments then go on to check the results of the
2507 if (!Outs.empty()) {
2508 // Check if stack adjustment is needed. For now, do not do this if any
2509 // argument is passed on the stack.
2510 SmallVector<CCValAssign, 16> ArgLocs;
2511 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2512 ArgLocs, *DAG.getContext());
2513 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2514 if (CCInfo.getNextStackOffset()) {
2515 MachineFunction &MF = DAG.getMachineFunction();
2516 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2518 if (Subtarget->isTargetWin64())
2519 // Win64 ABI has additional complications.
2522 // Check if the arguments are already laid out in the right way as
2523 // the caller's fixed stack objects.
2524 MachineFrameInfo *MFI = MF.getFrameInfo();
2525 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2526 const X86InstrInfo *TII =
2527 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2528 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2529 CCValAssign &VA = ArgLocs[i];
2530 SDValue Arg = OutVals[i];
2531 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2532 if (VA.getLocInfo() == CCValAssign::Indirect)
2534 if (!VA.isRegLoc()) {
2535 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2542 // If the tailcall address may be in a register, then make sure it's
2543 // possible to register allocate for it. In 32-bit, the call address can
2544 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2545 // callee-saved registers are restored. These happen to be the same
2546 // registers used to pass 'inreg' arguments so watch out for those.
2547 if (!Subtarget->is64Bit() &&
2548 !isa<GlobalAddressSDNode>(Callee) &&
2549 !isa<ExternalSymbolSDNode>(Callee)) {
2550 unsigned NumInRegs = 0;
2551 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2552 CCValAssign &VA = ArgLocs[i];
2555 unsigned Reg = VA.getLocReg();
2558 case X86::EAX: case X86::EDX: case X86::ECX:
2559 if (++NumInRegs == 3)
2571 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2572 return X86::createFastISel(funcInfo);
2576 //===----------------------------------------------------------------------===//
2577 // Other Lowering Hooks
2578 //===----------------------------------------------------------------------===//
2580 static bool MayFoldLoad(SDValue Op) {
2581 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2584 static bool MayFoldIntoStore(SDValue Op) {
2585 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2588 static bool isTargetShuffle(unsigned Opcode) {
2590 default: return false;
2591 case X86ISD::PSHUFD:
2592 case X86ISD::PSHUFHW:
2593 case X86ISD::PSHUFLW:
2594 case X86ISD::SHUFPD:
2595 case X86ISD::PALIGN:
2596 case X86ISD::SHUFPS:
2597 case X86ISD::MOVLHPS:
2598 case X86ISD::MOVLHPD:
2599 case X86ISD::MOVHLPS:
2600 case X86ISD::MOVLPS:
2601 case X86ISD::MOVLPD:
2602 case X86ISD::MOVSHDUP:
2603 case X86ISD::MOVSLDUP:
2606 case X86ISD::UNPCKLPS:
2607 case X86ISD::UNPCKLPD:
2608 case X86ISD::PUNPCKLWD:
2609 case X86ISD::PUNPCKLBW:
2610 case X86ISD::PUNPCKLDQ:
2611 case X86ISD::PUNPCKLQDQ:
2612 case X86ISD::UNPCKHPS:
2613 case X86ISD::UNPCKHPD:
2614 case X86ISD::PUNPCKHWD:
2615 case X86ISD::PUNPCKHBW:
2616 case X86ISD::PUNPCKHDQ:
2617 case X86ISD::PUNPCKHQDQ:
2623 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2624 SDValue V1, SelectionDAG &DAG) {
2626 default: llvm_unreachable("Unknown x86 shuffle node");
2627 case X86ISD::MOVSHDUP:
2628 case X86ISD::MOVSLDUP:
2629 return DAG.getNode(Opc, dl, VT, V1);
2635 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2636 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2638 default: llvm_unreachable("Unknown x86 shuffle node");
2639 case X86ISD::PSHUFD:
2640 case X86ISD::PSHUFHW:
2641 case X86ISD::PSHUFLW:
2642 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2648 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2649 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2651 default: llvm_unreachable("Unknown x86 shuffle node");
2652 case X86ISD::PALIGN:
2653 case X86ISD::SHUFPD:
2654 case X86ISD::SHUFPS:
2655 return DAG.getNode(Opc, dl, VT, V1, V2,
2656 DAG.getConstant(TargetMask, MVT::i8));
2661 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2662 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2664 default: llvm_unreachable("Unknown x86 shuffle node");
2665 case X86ISD::MOVLHPS:
2666 case X86ISD::MOVLHPD:
2667 case X86ISD::MOVHLPS:
2668 case X86ISD::MOVLPS:
2669 case X86ISD::MOVLPD:
2672 case X86ISD::UNPCKLPS:
2673 case X86ISD::UNPCKLPD:
2674 case X86ISD::PUNPCKLWD:
2675 case X86ISD::PUNPCKLBW:
2676 case X86ISD::PUNPCKLDQ:
2677 case X86ISD::PUNPCKLQDQ:
2678 case X86ISD::UNPCKHPS:
2679 case X86ISD::UNPCKHPD:
2680 case X86ISD::PUNPCKHWD:
2681 case X86ISD::PUNPCKHBW:
2682 case X86ISD::PUNPCKHDQ:
2683 case X86ISD::PUNPCKHQDQ:
2684 return DAG.getNode(Opc, dl, VT, V1, V2);
2689 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2690 MachineFunction &MF = DAG.getMachineFunction();
2691 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2692 int ReturnAddrIndex = FuncInfo->getRAIndex();
2694 if (ReturnAddrIndex == 0) {
2695 // Set up a frame object for the return address.
2696 uint64_t SlotSize = TD->getPointerSize();
2697 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2699 FuncInfo->setRAIndex(ReturnAddrIndex);
2702 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2706 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2707 bool hasSymbolicDisplacement) {
2708 // Offset should fit into 32 bit immediate field.
2709 if (!isInt<32>(Offset))
2712 // If we don't have a symbolic displacement - we don't have any extra
2714 if (!hasSymbolicDisplacement)
2717 // FIXME: Some tweaks might be needed for medium code model.
2718 if (M != CodeModel::Small && M != CodeModel::Kernel)
2721 // For small code model we assume that latest object is 16MB before end of 31
2722 // bits boundary. We may also accept pretty large negative constants knowing
2723 // that all objects are in the positive half of address space.
2724 if (M == CodeModel::Small && Offset < 16*1024*1024)
2727 // For kernel code model we know that all object resist in the negative half
2728 // of 32bits address space. We may not accept negative offsets, since they may
2729 // be just off and we may accept pretty large positive ones.
2730 if (M == CodeModel::Kernel && Offset > 0)
2736 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2737 /// specific condition code, returning the condition code and the LHS/RHS of the
2738 /// comparison to make.
2739 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2740 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2742 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2743 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2744 // X > -1 -> X == 0, jump !sign.
2745 RHS = DAG.getConstant(0, RHS.getValueType());
2746 return X86::COND_NS;
2747 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2748 // X < 0 -> X == 0, jump on sign.
2750 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2752 RHS = DAG.getConstant(0, RHS.getValueType());
2753 return X86::COND_LE;
2757 switch (SetCCOpcode) {
2758 default: llvm_unreachable("Invalid integer condition!");
2759 case ISD::SETEQ: return X86::COND_E;
2760 case ISD::SETGT: return X86::COND_G;
2761 case ISD::SETGE: return X86::COND_GE;
2762 case ISD::SETLT: return X86::COND_L;
2763 case ISD::SETLE: return X86::COND_LE;
2764 case ISD::SETNE: return X86::COND_NE;
2765 case ISD::SETULT: return X86::COND_B;
2766 case ISD::SETUGT: return X86::COND_A;
2767 case ISD::SETULE: return X86::COND_BE;
2768 case ISD::SETUGE: return X86::COND_AE;
2772 // First determine if it is required or is profitable to flip the operands.
2774 // If LHS is a foldable load, but RHS is not, flip the condition.
2775 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2776 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2777 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2778 std::swap(LHS, RHS);
2781 switch (SetCCOpcode) {
2787 std::swap(LHS, RHS);
2791 // On a floating point condition, the flags are set as follows:
2793 // 0 | 0 | 0 | X > Y
2794 // 0 | 0 | 1 | X < Y
2795 // 1 | 0 | 0 | X == Y
2796 // 1 | 1 | 1 | unordered
2797 switch (SetCCOpcode) {
2798 default: llvm_unreachable("Condcode should be pre-legalized away");
2800 case ISD::SETEQ: return X86::COND_E;
2801 case ISD::SETOLT: // flipped
2803 case ISD::SETGT: return X86::COND_A;
2804 case ISD::SETOLE: // flipped
2806 case ISD::SETGE: return X86::COND_AE;
2807 case ISD::SETUGT: // flipped
2809 case ISD::SETLT: return X86::COND_B;
2810 case ISD::SETUGE: // flipped
2812 case ISD::SETLE: return X86::COND_BE;
2814 case ISD::SETNE: return X86::COND_NE;
2815 case ISD::SETUO: return X86::COND_P;
2816 case ISD::SETO: return X86::COND_NP;
2818 case ISD::SETUNE: return X86::COND_INVALID;
2822 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2823 /// code. Current x86 isa includes the following FP cmov instructions:
2824 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2825 static bool hasFPCMov(unsigned X86CC) {
2841 /// isFPImmLegal - Returns true if the target can instruction select the
2842 /// specified FP immediate natively. If false, the legalizer will
2843 /// materialize the FP immediate as a load from a constant pool.
2844 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2845 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2846 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2852 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2853 /// the specified range (L, H].
2854 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2855 return (Val < 0) || (Val >= Low && Val < Hi);
2858 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2859 /// specified value.
2860 static bool isUndefOrEqual(int Val, int CmpVal) {
2861 if (Val < 0 || Val == CmpVal)
2866 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2867 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2868 /// the second operand.
2869 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2870 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2871 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2872 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2873 return (Mask[0] < 2 && Mask[1] < 2);
2877 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2878 SmallVector<int, 8> M;
2880 return ::isPSHUFDMask(M, N->getValueType(0));
2883 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2884 /// is suitable for input to PSHUFHW.
2885 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2886 if (VT != MVT::v8i16)
2889 // Lower quadword copied in order or undef.
2890 for (int i = 0; i != 4; ++i)
2891 if (Mask[i] >= 0 && Mask[i] != i)
2894 // Upper quadword shuffled.
2895 for (int i = 4; i != 8; ++i)
2896 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2902 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2903 SmallVector<int, 8> M;
2905 return ::isPSHUFHWMask(M, N->getValueType(0));
2908 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2909 /// is suitable for input to PSHUFLW.
2910 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2911 if (VT != MVT::v8i16)
2914 // Upper quadword copied in order.
2915 for (int i = 4; i != 8; ++i)
2916 if (Mask[i] >= 0 && Mask[i] != i)
2919 // Lower quadword shuffled.
2920 for (int i = 0; i != 4; ++i)
2927 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2928 SmallVector<int, 8> M;
2930 return ::isPSHUFLWMask(M, N->getValueType(0));
2933 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2934 /// is suitable for input to PALIGNR.
2935 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2937 int i, e = VT.getVectorNumElements();
2939 // Do not handle v2i64 / v2f64 shuffles with palignr.
2940 if (e < 4 || !hasSSSE3)
2943 for (i = 0; i != e; ++i)
2947 // All undef, not a palignr.
2951 // Determine if it's ok to perform a palignr with only the LHS, since we
2952 // don't have access to the actual shuffle elements to see if RHS is undef.
2953 bool Unary = Mask[i] < (int)e;
2954 bool NeedsUnary = false;
2956 int s = Mask[i] - i;
2958 // Check the rest of the elements to see if they are consecutive.
2959 for (++i; i != e; ++i) {
2964 Unary = Unary && (m < (int)e);
2965 NeedsUnary = NeedsUnary || (m < s);
2967 if (NeedsUnary && !Unary)
2969 if (Unary && m != ((s+i) & (e-1)))
2971 if (!Unary && m != (s+i))
2977 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2978 SmallVector<int, 8> M;
2980 return ::isPALIGNRMask(M, N->getValueType(0), true);
2983 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2984 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2985 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2986 int NumElems = VT.getVectorNumElements();
2987 if (NumElems != 2 && NumElems != 4)
2990 int Half = NumElems / 2;
2991 for (int i = 0; i < Half; ++i)
2992 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2994 for (int i = Half; i < NumElems; ++i)
2995 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3001 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3002 SmallVector<int, 8> M;
3004 return ::isSHUFPMask(M, N->getValueType(0));
3007 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3008 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3009 /// half elements to come from vector 1 (which would equal the dest.) and
3010 /// the upper half to come from vector 2.
3011 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3012 int NumElems = VT.getVectorNumElements();
3014 if (NumElems != 2 && NumElems != 4)
3017 int Half = NumElems / 2;
3018 for (int i = 0; i < Half; ++i)
3019 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3021 for (int i = Half; i < NumElems; ++i)
3022 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3027 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3028 SmallVector<int, 8> M;
3030 return isCommutedSHUFPMask(M, N->getValueType(0));
3033 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3034 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3035 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3036 if (N->getValueType(0).getVectorNumElements() != 4)
3039 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3040 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3041 isUndefOrEqual(N->getMaskElt(1), 7) &&
3042 isUndefOrEqual(N->getMaskElt(2), 2) &&
3043 isUndefOrEqual(N->getMaskElt(3), 3);
3046 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3047 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3049 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3050 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3055 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3056 isUndefOrEqual(N->getMaskElt(1), 3) &&
3057 isUndefOrEqual(N->getMaskElt(2), 2) &&
3058 isUndefOrEqual(N->getMaskElt(3), 3);
3061 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3062 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3063 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3064 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3066 if (NumElems != 2 && NumElems != 4)
3069 for (unsigned i = 0; i < NumElems/2; ++i)
3070 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3073 for (unsigned i = NumElems/2; i < NumElems; ++i)
3074 if (!isUndefOrEqual(N->getMaskElt(i), i))
3080 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3081 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3082 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3083 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3085 if (NumElems != 2 && NumElems != 4)
3088 for (unsigned i = 0; i < NumElems/2; ++i)
3089 if (!isUndefOrEqual(N->getMaskElt(i), i))
3092 for (unsigned i = 0; i < NumElems/2; ++i)
3093 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3099 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3100 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3101 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3102 bool V2IsSplat = false) {
3103 int NumElts = VT.getVectorNumElements();
3104 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3107 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3109 int BitI1 = Mask[i+1];
3110 if (!isUndefOrEqual(BitI, j))
3113 if (!isUndefOrEqual(BitI1, NumElts))
3116 if (!isUndefOrEqual(BitI1, j + NumElts))
3123 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3124 SmallVector<int, 8> M;
3126 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3129 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3130 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3131 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3132 bool V2IsSplat = false) {
3133 int NumElts = VT.getVectorNumElements();
3134 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3137 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3139 int BitI1 = Mask[i+1];
3140 if (!isUndefOrEqual(BitI, j + NumElts/2))
3143 if (isUndefOrEqual(BitI1, NumElts))
3146 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3153 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3154 SmallVector<int, 8> M;
3156 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3159 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3160 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3162 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3163 int NumElems = VT.getVectorNumElements();
3164 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3167 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3169 int BitI1 = Mask[i+1];
3170 if (!isUndefOrEqual(BitI, j))
3172 if (!isUndefOrEqual(BitI1, j))
3178 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3179 SmallVector<int, 8> M;
3181 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3184 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3185 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3187 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3188 int NumElems = VT.getVectorNumElements();
3189 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3192 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3194 int BitI1 = Mask[i+1];
3195 if (!isUndefOrEqual(BitI, j))
3197 if (!isUndefOrEqual(BitI1, j))
3203 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3204 SmallVector<int, 8> M;
3206 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3209 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3210 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3211 /// MOVSD, and MOVD, i.e. setting the lowest element.
3212 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3213 if (VT.getVectorElementType().getSizeInBits() < 32)
3216 int NumElts = VT.getVectorNumElements();
3218 if (!isUndefOrEqual(Mask[0], NumElts))
3221 for (int i = 1; i < NumElts; ++i)
3222 if (!isUndefOrEqual(Mask[i], i))
3228 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3229 SmallVector<int, 8> M;
3231 return ::isMOVLMask(M, N->getValueType(0));
3234 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3235 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3236 /// element of vector 2 and the other elements to come from vector 1 in order.
3237 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3238 bool V2IsSplat = false, bool V2IsUndef = false) {
3239 int NumOps = VT.getVectorNumElements();
3240 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3243 if (!isUndefOrEqual(Mask[0], 0))
3246 for (int i = 1; i < NumOps; ++i)
3247 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3248 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3249 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3255 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3256 bool V2IsUndef = false) {
3257 SmallVector<int, 8> M;
3259 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3262 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3263 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3264 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3265 if (N->getValueType(0).getVectorNumElements() != 4)
3268 // Expect 1, 1, 3, 3
3269 for (unsigned i = 0; i < 2; ++i) {
3270 int Elt = N->getMaskElt(i);
3271 if (Elt >= 0 && Elt != 1)
3276 for (unsigned i = 2; i < 4; ++i) {
3277 int Elt = N->getMaskElt(i);
3278 if (Elt >= 0 && Elt != 3)
3283 // Don't use movshdup if it can be done with a shufps.
3284 // FIXME: verify that matching u, u, 3, 3 is what we want.
3288 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3289 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3290 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3291 if (N->getValueType(0).getVectorNumElements() != 4)
3294 // Expect 0, 0, 2, 2
3295 for (unsigned i = 0; i < 2; ++i)
3296 if (N->getMaskElt(i) > 0)
3300 for (unsigned i = 2; i < 4; ++i) {
3301 int Elt = N->getMaskElt(i);
3302 if (Elt >= 0 && Elt != 2)
3307 // Don't use movsldup if it can be done with a shufps.
3311 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3312 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3313 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3314 int e = N->getValueType(0).getVectorNumElements() / 2;
3316 for (int i = 0; i < e; ++i)
3317 if (!isUndefOrEqual(N->getMaskElt(i), i))
3319 for (int i = 0; i < e; ++i)
3320 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3325 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3326 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3327 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3328 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3329 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3331 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3333 for (int i = 0; i < NumOperands; ++i) {
3334 int Val = SVOp->getMaskElt(NumOperands-i-1);
3335 if (Val < 0) Val = 0;
3336 if (Val >= NumOperands) Val -= NumOperands;
3338 if (i != NumOperands - 1)
3344 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3345 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3346 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3347 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3349 // 8 nodes, but we only care about the last 4.
3350 for (unsigned i = 7; i >= 4; --i) {
3351 int Val = SVOp->getMaskElt(i);
3360 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3361 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3362 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3363 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3365 // 8 nodes, but we only care about the first 4.
3366 for (int i = 3; i >= 0; --i) {
3367 int Val = SVOp->getMaskElt(i);
3376 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3377 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3378 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3379 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3380 EVT VVT = N->getValueType(0);
3381 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3385 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3386 Val = SVOp->getMaskElt(i);
3390 return (Val - i) * EltSize;
3393 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3395 bool X86::isZeroNode(SDValue Elt) {
3396 return ((isa<ConstantSDNode>(Elt) &&
3397 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3398 (isa<ConstantFPSDNode>(Elt) &&
3399 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3402 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3403 /// their permute mask.
3404 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3405 SelectionDAG &DAG) {
3406 EVT VT = SVOp->getValueType(0);
3407 unsigned NumElems = VT.getVectorNumElements();
3408 SmallVector<int, 8> MaskVec;
3410 for (unsigned i = 0; i != NumElems; ++i) {
3411 int idx = SVOp->getMaskElt(i);
3413 MaskVec.push_back(idx);
3414 else if (idx < (int)NumElems)
3415 MaskVec.push_back(idx + NumElems);
3417 MaskVec.push_back(idx - NumElems);
3419 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3420 SVOp->getOperand(0), &MaskVec[0]);
3423 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3424 /// the two vector operands have swapped position.
3425 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3426 unsigned NumElems = VT.getVectorNumElements();
3427 for (unsigned i = 0; i != NumElems; ++i) {
3431 else if (idx < (int)NumElems)
3432 Mask[i] = idx + NumElems;
3434 Mask[i] = idx - NumElems;
3438 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3439 /// match movhlps. The lower half elements should come from upper half of
3440 /// V1 (and in order), and the upper half elements should come from the upper
3441 /// half of V2 (and in order).
3442 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3443 if (Op->getValueType(0).getVectorNumElements() != 4)
3445 for (unsigned i = 0, e = 2; i != e; ++i)
3446 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3448 for (unsigned i = 2; i != 4; ++i)
3449 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3454 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3455 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3457 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3458 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3460 N = N->getOperand(0).getNode();
3461 if (!ISD::isNON_EXTLoad(N))
3464 *LD = cast<LoadSDNode>(N);
3468 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3469 /// match movlp{s|d}. The lower half elements should come from lower half of
3470 /// V1 (and in order), and the upper half elements should come from the upper
3471 /// half of V2 (and in order). And since V1 will become the source of the
3472 /// MOVLP, it must be either a vector load or a scalar load to vector.
3473 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3474 ShuffleVectorSDNode *Op) {
3475 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3477 // Is V2 is a vector load, don't do this transformation. We will try to use
3478 // load folding shufps op.
3479 if (ISD::isNON_EXTLoad(V2))
3482 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3484 if (NumElems != 2 && NumElems != 4)
3486 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3487 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3489 for (unsigned i = NumElems/2; i != NumElems; ++i)
3490 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3495 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3497 static bool isSplatVector(SDNode *N) {
3498 if (N->getOpcode() != ISD::BUILD_VECTOR)
3501 SDValue SplatValue = N->getOperand(0);
3502 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3503 if (N->getOperand(i) != SplatValue)
3508 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3509 /// to an zero vector.
3510 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3511 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3512 SDValue V1 = N->getOperand(0);
3513 SDValue V2 = N->getOperand(1);
3514 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3515 for (unsigned i = 0; i != NumElems; ++i) {
3516 int Idx = N->getMaskElt(i);
3517 if (Idx >= (int)NumElems) {
3518 unsigned Opc = V2.getOpcode();
3519 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3521 if (Opc != ISD::BUILD_VECTOR ||
3522 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3524 } else if (Idx >= 0) {
3525 unsigned Opc = V1.getOpcode();
3526 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3528 if (Opc != ISD::BUILD_VECTOR ||
3529 !X86::isZeroNode(V1.getOperand(Idx)))
3536 /// getZeroVector - Returns a vector of specified type with all zero elements.
3538 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3540 assert(VT.isVector() && "Expected a vector type");
3542 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted
3543 // to their dest type. This ensures they get CSE'd.
3545 if (VT.getSizeInBits() == 64) { // MMX
3546 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3547 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3548 } else if (VT.getSizeInBits() == 128) {
3549 if (HasSSE2) { // SSE2
3550 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3551 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3553 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3554 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3556 } else if (VT.getSizeInBits() == 256) { // AVX
3557 // 256-bit logic and arithmetic instructions in AVX are
3558 // all floating-point, no support for integer ops. Default
3559 // to emitting fp zeroed vectors then.
3560 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3561 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3562 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3564 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3567 /// getOnesVector - Returns a vector of specified type with all bits set.
3569 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3570 assert(VT.isVector() && "Expected a vector type");
3572 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3573 // type. This ensures they get CSE'd.
3574 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3576 if (VT.getSizeInBits() == 64) // MMX
3577 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3579 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3580 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3584 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3585 /// that point to V2 points to its first element.
3586 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3587 EVT VT = SVOp->getValueType(0);
3588 unsigned NumElems = VT.getVectorNumElements();
3590 bool Changed = false;
3591 SmallVector<int, 8> MaskVec;
3592 SVOp->getMask(MaskVec);
3594 for (unsigned i = 0; i != NumElems; ++i) {
3595 if (MaskVec[i] > (int)NumElems) {
3596 MaskVec[i] = NumElems;
3601 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3602 SVOp->getOperand(1), &MaskVec[0]);
3603 return SDValue(SVOp, 0);
3606 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3607 /// operation of specified width.
3608 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3610 unsigned NumElems = VT.getVectorNumElements();
3611 SmallVector<int, 8> Mask;
3612 Mask.push_back(NumElems);
3613 for (unsigned i = 1; i != NumElems; ++i)
3615 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3618 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3619 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3621 unsigned NumElems = VT.getVectorNumElements();
3622 SmallVector<int, 8> Mask;
3623 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3625 Mask.push_back(i + NumElems);
3627 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3630 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3631 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3633 unsigned NumElems = VT.getVectorNumElements();
3634 unsigned Half = NumElems/2;
3635 SmallVector<int, 8> Mask;
3636 for (unsigned i = 0; i != Half; ++i) {
3637 Mask.push_back(i + Half);
3638 Mask.push_back(i + NumElems + Half);
3640 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3643 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3644 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3645 if (SV->getValueType(0).getVectorNumElements() <= 4)
3646 return SDValue(SV, 0);
3648 EVT PVT = MVT::v4f32;
3649 EVT VT = SV->getValueType(0);
3650 DebugLoc dl = SV->getDebugLoc();
3651 SDValue V1 = SV->getOperand(0);
3652 int NumElems = VT.getVectorNumElements();
3653 int EltNo = SV->getSplatIndex();
3655 // unpack elements to the correct location
3656 while (NumElems > 4) {
3657 if (EltNo < NumElems/2) {
3658 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3660 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3661 EltNo -= NumElems/2;
3666 // Perform the splat.
3667 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3668 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3669 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3670 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3673 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3674 /// vector of zero or undef vector. This produces a shuffle where the low
3675 /// element of V2 is swizzled into the zero/undef vector, landing at element
3676 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3677 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3678 bool isZero, bool HasSSE2,
3679 SelectionDAG &DAG) {
3680 EVT VT = V2.getValueType();
3682 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3683 unsigned NumElems = VT.getVectorNumElements();
3684 SmallVector<int, 16> MaskVec;
3685 for (unsigned i = 0; i != NumElems; ++i)
3686 // If this is the insertion idx, put the low elt of V2 here.
3687 MaskVec.push_back(i == Idx ? NumElems : i);
3688 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3691 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
3692 /// element of the result of the vector shuffle.
3693 SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
3696 return SDValue(); // Limit search depth.
3698 SDValue V = SDValue(N, 0);
3699 EVT VT = V.getValueType();
3700 unsigned Opcode = V.getOpcode();
3702 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
3703 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
3704 Index = SV->getMaskElt(Index);
3707 return DAG.getUNDEF(VT.getVectorElementType());
3709 int NumElems = VT.getVectorNumElements();
3710 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
3711 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
3714 // Recurse into target specific vector shuffles to find scalars.
3715 if (isTargetShuffle(Opcode)) {
3716 int NumElems = VT.getVectorNumElements();
3717 SmallVector<unsigned, 16> ShuffleMask;
3721 case X86ISD::SHUFPS:
3722 case X86ISD::SHUFPD:
3723 ImmN = N->getOperand(N->getNumOperands()-1);
3724 DecodeSHUFPSMask(NumElems,
3725 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3728 case X86ISD::PUNPCKHBW:
3729 case X86ISD::PUNPCKHWD:
3730 case X86ISD::PUNPCKHDQ:
3731 case X86ISD::PUNPCKHQDQ:
3732 DecodePUNPCKHMask(NumElems, ShuffleMask);
3734 case X86ISD::UNPCKHPS:
3735 case X86ISD::UNPCKHPD:
3736 DecodeUNPCKHPMask(NumElems, ShuffleMask);
3738 case X86ISD::PUNPCKLBW:
3739 case X86ISD::PUNPCKLWD:
3740 case X86ISD::PUNPCKLDQ:
3741 case X86ISD::PUNPCKLQDQ:
3742 DecodePUNPCKLMask(NumElems, ShuffleMask);
3744 case X86ISD::UNPCKLPS:
3745 case X86ISD::UNPCKLPD:
3746 DecodeUNPCKLPMask(NumElems, ShuffleMask);
3748 case X86ISD::MOVHLPS:
3749 DecodeMOVHLPSMask(NumElems, ShuffleMask);
3751 case X86ISD::MOVLHPS:
3752 DecodeMOVLHPSMask(NumElems, ShuffleMask);
3754 case X86ISD::PSHUFD:
3755 ImmN = N->getOperand(N->getNumOperands()-1);
3756 DecodePSHUFMask(NumElems,
3757 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3760 case X86ISD::PSHUFHW:
3761 ImmN = N->getOperand(N->getNumOperands()-1);
3762 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3765 case X86ISD::PSHUFLW:
3766 ImmN = N->getOperand(N->getNumOperands()-1);
3767 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3771 case X86ISD::MOVSD: {
3772 // The index 0 always comes from the first element of the second source,
3773 // this is why MOVSS and MOVSD are used in the first place. The other
3774 // elements come from the other positions of the first source vector.
3775 unsigned OpNum = (Index == 0) ? 1 : 0;
3776 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
3780 assert("not implemented for target shuffle node");
3784 Index = ShuffleMask[Index];
3786 return DAG.getUNDEF(VT.getVectorElementType());
3788 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
3789 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
3793 // Actual nodes that may contain scalar elements
3794 if (Opcode == ISD::BIT_CONVERT) {
3795 V = V.getOperand(0);
3796 EVT SrcVT = V.getValueType();
3797 unsigned NumElems = VT.getVectorNumElements();
3799 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
3803 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
3804 return (Index == 0) ? V.getOperand(0)
3805 : DAG.getUNDEF(VT.getVectorElementType());
3807 if (V.getOpcode() == ISD::BUILD_VECTOR)
3808 return V.getOperand(Index);
3813 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
3814 /// shuffle operation which come from a consecutively from a zero. The
3815 /// search can start in two diferent directions, from left or right.
3817 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
3818 bool ZerosFromLeft, SelectionDAG &DAG) {
3821 while (i < NumElems) {
3822 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
3823 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
3824 if (!(Elt.getNode() &&
3825 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
3833 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
3834 /// MaskE correspond consecutively to elements from one of the vector operands,
3835 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
3837 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
3838 int OpIdx, int NumElems, unsigned &OpNum) {
3839 bool SeenV1 = false;
3840 bool SeenV2 = false;
3842 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
3843 int Idx = SVOp->getMaskElt(i);
3844 // Ignore undef indicies
3853 // Only accept consecutive elements from the same vector
3854 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
3858 OpNum = SeenV1 ? 0 : 1;
3862 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
3863 /// logical left shift of a vector.
3864 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3865 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3866 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3867 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3868 false /* check zeros from right */, DAG);
3874 // Considering the elements in the mask that are not consecutive zeros,
3875 // check if they consecutively come from only one of the source vectors.
3877 // V1 = {X, A, B, C} 0
3879 // vector_shuffle V1, V2 <1, 2, 3, X>
3881 if (!isShuffleMaskConsecutive(SVOp,
3882 0, // Mask Start Index
3883 NumElems-NumZeros-1, // Mask End Index
3884 NumZeros, // Where to start looking in the src vector
3885 NumElems, // Number of elements in vector
3886 OpSrc)) // Which source operand ?
3891 ShVal = SVOp->getOperand(OpSrc);
3895 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
3896 /// logical left shift of a vector.
3897 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3898 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3899 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3900 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3901 true /* check zeros from left */, DAG);
3907 // Considering the elements in the mask that are not consecutive zeros,
3908 // check if they consecutively come from only one of the source vectors.
3910 // 0 { A, B, X, X } = V2
3912 // vector_shuffle V1, V2 <X, X, 4, 5>
3914 if (!isShuffleMaskConsecutive(SVOp,
3915 NumZeros, // Mask Start Index
3916 NumElems-1, // Mask End Index
3917 0, // Where to start looking in the src vector
3918 NumElems, // Number of elements in vector
3919 OpSrc)) // Which source operand ?
3924 ShVal = SVOp->getOperand(OpSrc);
3928 /// isVectorShift - Returns true if the shuffle can be implemented as a
3929 /// logical left or right shift of a vector.
3930 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3931 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3932 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
3933 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
3939 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3941 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3942 unsigned NumNonZero, unsigned NumZero,
3944 const TargetLowering &TLI) {
3948 DebugLoc dl = Op.getDebugLoc();
3951 for (unsigned i = 0; i < 16; ++i) {
3952 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3953 if (ThisIsNonZero && First) {
3955 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3957 V = DAG.getUNDEF(MVT::v8i16);
3962 SDValue ThisElt(0, 0), LastElt(0, 0);
3963 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3964 if (LastIsNonZero) {
3965 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3966 MVT::i16, Op.getOperand(i-1));
3968 if (ThisIsNonZero) {
3969 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3970 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3971 ThisElt, DAG.getConstant(8, MVT::i8));
3973 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3977 if (ThisElt.getNode())
3978 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3979 DAG.getIntPtrConstant(i/2));
3983 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3986 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3988 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3989 unsigned NumNonZero, unsigned NumZero,
3991 const TargetLowering &TLI) {
3995 DebugLoc dl = Op.getDebugLoc();
3998 for (unsigned i = 0; i < 8; ++i) {
3999 bool isNonZero = (NonZeros & (1 << i)) != 0;
4003 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4005 V = DAG.getUNDEF(MVT::v8i16);
4008 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4009 MVT::v8i16, V, Op.getOperand(i),
4010 DAG.getIntPtrConstant(i));
4017 /// getVShift - Return a vector logical shift node.
4019 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4020 unsigned NumBits, SelectionDAG &DAG,
4021 const TargetLowering &TLI, DebugLoc dl) {
4022 bool isMMX = VT.getSizeInBits() == 64;
4023 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
4024 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4025 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
4026 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4027 DAG.getNode(Opc, dl, ShVT, SrcOp,
4028 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
4032 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4033 SelectionDAG &DAG) const {
4035 // Check if the scalar load can be widened into a vector load. And if
4036 // the address is "base + cst" see if the cst can be "absorbed" into
4037 // the shuffle mask.
4038 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4039 SDValue Ptr = LD->getBasePtr();
4040 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4042 EVT PVT = LD->getValueType(0);
4043 if (PVT != MVT::i32 && PVT != MVT::f32)
4048 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4049 FI = FINode->getIndex();
4051 } else if (Ptr.getOpcode() == ISD::ADD &&
4052 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
4053 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4054 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4055 Offset = Ptr.getConstantOperandVal(1);
4056 Ptr = Ptr.getOperand(0);
4061 SDValue Chain = LD->getChain();
4062 // Make sure the stack object alignment is at least 16.
4063 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4064 if (DAG.InferPtrAlignment(Ptr) < 16) {
4065 if (MFI->isFixedObjectIndex(FI)) {
4066 // Can't change the alignment. FIXME: It's possible to compute
4067 // the exact stack offset and reference FI + adjust offset instead.
4068 // If someone *really* cares about this. That's the way to implement it.
4071 MFI->setObjectAlignment(FI, 16);
4075 // (Offset % 16) must be multiple of 4. Then address is then
4076 // Ptr + (Offset & ~15).
4079 if ((Offset % 16) & 3)
4081 int64_t StartOffset = Offset & ~15;
4083 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4084 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4086 int EltNo = (Offset - StartOffset) >> 2;
4087 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4088 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4089 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
4091 // Canonicalize it to a v4i32 shuffle.
4092 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
4093 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4094 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4095 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
4101 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4102 /// vector of type 'VT', see if the elements can be replaced by a single large
4103 /// load which has the same value as a build_vector whose operands are 'elts'.
4105 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4107 /// FIXME: we'd also like to handle the case where the last elements are zero
4108 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4109 /// There's even a handy isZeroNode for that purpose.
4110 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4111 DebugLoc &dl, SelectionDAG &DAG) {
4112 EVT EltVT = VT.getVectorElementType();
4113 unsigned NumElems = Elts.size();
4115 LoadSDNode *LDBase = NULL;
4116 unsigned LastLoadedElt = -1U;
4118 // For each element in the initializer, see if we've found a load or an undef.
4119 // If we don't find an initial load element, or later load elements are
4120 // non-consecutive, bail out.
4121 for (unsigned i = 0; i < NumElems; ++i) {
4122 SDValue Elt = Elts[i];
4124 if (!Elt.getNode() ||
4125 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4128 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4130 LDBase = cast<LoadSDNode>(Elt.getNode());
4134 if (Elt.getOpcode() == ISD::UNDEF)
4137 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4138 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4143 // If we have found an entire vector of loads and undefs, then return a large
4144 // load of the entire vector width starting at the base pointer. If we found
4145 // consecutive loads for the low half, generate a vzext_load node.
4146 if (LastLoadedElt == NumElems - 1) {
4147 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4148 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
4149 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
4150 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4151 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
4152 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
4153 LDBase->isVolatile(), LDBase->isNonTemporal(),
4154 LDBase->getAlignment());
4155 } else if (NumElems == 4 && LastLoadedElt == 1) {
4156 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4157 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4158 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
4159 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
4165 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4166 DebugLoc dl = Op.getDebugLoc();
4167 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
4168 // All one's are handled with pcmpeqd. In AVX, zero's are handled with
4169 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
4170 // is present, so AllOnes is ignored.
4171 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4172 (Op.getValueType().getSizeInBits() != 256 &&
4173 ISD::isBuildVectorAllOnes(Op.getNode()))) {
4174 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
4175 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4176 // eliminated on x86-32 hosts.
4177 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
4180 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4181 return getOnesVector(Op.getValueType(), DAG, dl);
4182 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4185 EVT VT = Op.getValueType();
4186 EVT ExtVT = VT.getVectorElementType();
4187 unsigned EVTBits = ExtVT.getSizeInBits();
4189 unsigned NumElems = Op.getNumOperands();
4190 unsigned NumZero = 0;
4191 unsigned NumNonZero = 0;
4192 unsigned NonZeros = 0;
4193 bool IsAllConstants = true;
4194 SmallSet<SDValue, 8> Values;
4195 for (unsigned i = 0; i < NumElems; ++i) {
4196 SDValue Elt = Op.getOperand(i);
4197 if (Elt.getOpcode() == ISD::UNDEF)
4200 if (Elt.getOpcode() != ISD::Constant &&
4201 Elt.getOpcode() != ISD::ConstantFP)
4202 IsAllConstants = false;
4203 if (X86::isZeroNode(Elt))
4206 NonZeros |= (1 << i);
4211 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4212 if (NumNonZero == 0)
4213 return DAG.getUNDEF(VT);
4215 // Special case for single non-zero, non-undef, element.
4216 if (NumNonZero == 1) {
4217 unsigned Idx = CountTrailingZeros_32(NonZeros);
4218 SDValue Item = Op.getOperand(Idx);
4220 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4221 // the value are obviously zero, truncate the value to i32 and do the
4222 // insertion that way. Only do this if the value is non-constant or if the
4223 // value is a constant being inserted into element 0. It is cheaper to do
4224 // a constant pool load than it is to do a movd + shuffle.
4225 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4226 (!IsAllConstants || Idx == 0)) {
4227 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4228 // Handle MMX and SSE both.
4229 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
4230 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
4232 // Truncate the value (which may itself be a constant) to i32, and
4233 // convert it to a vector with movd (S2V+shuffle to zero extend).
4234 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4235 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4236 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4237 Subtarget->hasSSE2(), DAG);
4239 // Now we have our 32-bit value zero extended in the low element of
4240 // a vector. If Idx != 0, swizzle it into place.
4242 SmallVector<int, 4> Mask;
4243 Mask.push_back(Idx);
4244 for (unsigned i = 1; i != VecElts; ++i)
4246 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4247 DAG.getUNDEF(Item.getValueType()),
4250 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
4254 // If we have a constant or non-constant insertion into the low element of
4255 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4256 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4257 // depending on what the source datatype is.
4260 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4261 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4262 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4263 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4264 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4265 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4267 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4268 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4269 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
4270 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4271 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4272 Subtarget->hasSSE2(), DAG);
4273 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
4277 // Is it a vector logical left shift?
4278 if (NumElems == 2 && Idx == 1 &&
4279 X86::isZeroNode(Op.getOperand(0)) &&
4280 !X86::isZeroNode(Op.getOperand(1))) {
4281 unsigned NumBits = VT.getSizeInBits();
4282 return getVShift(true, VT,
4283 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4284 VT, Op.getOperand(1)),
4285 NumBits/2, DAG, *this, dl);
4288 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4291 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4292 // is a non-constant being inserted into an element other than the low one,
4293 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4294 // movd/movss) to move this into the low element, then shuffle it into
4296 if (EVTBits == 32) {
4297 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4299 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4300 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4301 Subtarget->hasSSE2(), DAG);
4302 SmallVector<int, 8> MaskVec;
4303 for (unsigned i = 0; i < NumElems; i++)
4304 MaskVec.push_back(i == Idx ? 0 : 1);
4305 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4309 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4310 if (Values.size() == 1) {
4311 if (EVTBits == 32) {
4312 // Instead of a shuffle like this:
4313 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4314 // Check if it's possible to issue this instead.
4315 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4316 unsigned Idx = CountTrailingZeros_32(NonZeros);
4317 SDValue Item = Op.getOperand(Idx);
4318 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4319 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4324 // A vector full of immediates; various special cases are already
4325 // handled, so this is best done with a single constant-pool load.
4329 // Let legalizer expand 2-wide build_vectors.
4330 if (EVTBits == 64) {
4331 if (NumNonZero == 1) {
4332 // One half is zero or undef.
4333 unsigned Idx = CountTrailingZeros_32(NonZeros);
4334 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4335 Op.getOperand(Idx));
4336 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4337 Subtarget->hasSSE2(), DAG);
4342 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4343 if (EVTBits == 8 && NumElems == 16) {
4344 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4346 if (V.getNode()) return V;
4349 if (EVTBits == 16 && NumElems == 8) {
4350 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4352 if (V.getNode()) return V;
4355 // If element VT is == 32 bits, turn it into a number of shuffles.
4356 SmallVector<SDValue, 8> V;
4358 if (NumElems == 4 && NumZero > 0) {
4359 for (unsigned i = 0; i < 4; ++i) {
4360 bool isZero = !(NonZeros & (1 << i));
4362 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4364 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4367 for (unsigned i = 0; i < 2; ++i) {
4368 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4371 V[i] = V[i*2]; // Must be a zero vector.
4374 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4377 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4380 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4385 SmallVector<int, 8> MaskVec;
4386 bool Reverse = (NonZeros & 0x3) == 2;
4387 for (unsigned i = 0; i < 2; ++i)
4388 MaskVec.push_back(Reverse ? 1-i : i);
4389 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4390 for (unsigned i = 0; i < 2; ++i)
4391 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4392 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4395 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4396 // Check for a build vector of consecutive loads.
4397 for (unsigned i = 0; i < NumElems; ++i)
4398 V[i] = Op.getOperand(i);
4400 // Check for elements which are consecutive loads.
4401 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4405 // For SSE 4.1, use insertps to put the high elements into the low element.
4406 if (getSubtarget()->hasSSE41()) {
4408 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4409 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4411 Result = DAG.getUNDEF(VT);
4413 for (unsigned i = 1; i < NumElems; ++i) {
4414 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4415 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4416 Op.getOperand(i), DAG.getIntPtrConstant(i));
4421 // Otherwise, expand into a number of unpckl*, start by extending each of
4422 // our (non-undef) elements to the full vector width with the element in the
4423 // bottom slot of the vector (which generates no code for SSE).
4424 for (unsigned i = 0; i < NumElems; ++i) {
4425 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4426 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4428 V[i] = DAG.getUNDEF(VT);
4431 // Next, we iteratively mix elements, e.g. for v4f32:
4432 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4433 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4434 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4435 unsigned EltStride = NumElems >> 1;
4436 while (EltStride != 0) {
4437 for (unsigned i = 0; i < EltStride; ++i) {
4438 // If V[i+EltStride] is undef and this is the first round of mixing,
4439 // then it is safe to just drop this shuffle: V[i] is already in the
4440 // right place, the one element (since it's the first round) being
4441 // inserted as undef can be dropped. This isn't safe for successive
4442 // rounds because they will permute elements within both vectors.
4443 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4444 EltStride == NumElems/2)
4447 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4457 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4458 // We support concatenate two MMX registers and place them in a MMX
4459 // register. This is better than doing a stack convert.
4460 DebugLoc dl = Op.getDebugLoc();
4461 EVT ResVT = Op.getValueType();
4462 assert(Op.getNumOperands() == 2);
4463 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4464 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4466 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4467 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4468 InVec = Op.getOperand(1);
4469 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4470 unsigned NumElts = ResVT.getVectorNumElements();
4471 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4472 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4473 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4475 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4476 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4477 Mask[0] = 0; Mask[1] = 2;
4478 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4480 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4483 // v8i16 shuffles - Prefer shuffles in the following order:
4484 // 1. [all] pshuflw, pshufhw, optional move
4485 // 2. [ssse3] 1 x pshufb
4486 // 3. [ssse3] 2 x pshufb + 1 x por
4487 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4489 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4490 SelectionDAG &DAG) const {
4491 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4492 SDValue V1 = SVOp->getOperand(0);
4493 SDValue V2 = SVOp->getOperand(1);
4494 DebugLoc dl = SVOp->getDebugLoc();
4495 SmallVector<int, 8> MaskVals;
4497 // Determine if more than 1 of the words in each of the low and high quadwords
4498 // of the result come from the same quadword of one of the two inputs. Undef
4499 // mask values count as coming from any quadword, for better codegen.
4500 SmallVector<unsigned, 4> LoQuad(4);
4501 SmallVector<unsigned, 4> HiQuad(4);
4502 BitVector InputQuads(4);
4503 for (unsigned i = 0; i < 8; ++i) {
4504 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4505 int EltIdx = SVOp->getMaskElt(i);
4506 MaskVals.push_back(EltIdx);
4515 InputQuads.set(EltIdx / 4);
4518 int BestLoQuad = -1;
4519 unsigned MaxQuad = 1;
4520 for (unsigned i = 0; i < 4; ++i) {
4521 if (LoQuad[i] > MaxQuad) {
4523 MaxQuad = LoQuad[i];
4527 int BestHiQuad = -1;
4529 for (unsigned i = 0; i < 4; ++i) {
4530 if (HiQuad[i] > MaxQuad) {
4532 MaxQuad = HiQuad[i];
4536 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4537 // of the two input vectors, shuffle them into one input vector so only a
4538 // single pshufb instruction is necessary. If There are more than 2 input
4539 // quads, disable the next transformation since it does not help SSSE3.
4540 bool V1Used = InputQuads[0] || InputQuads[1];
4541 bool V2Used = InputQuads[2] || InputQuads[3];
4542 if (Subtarget->hasSSSE3()) {
4543 if (InputQuads.count() == 2 && V1Used && V2Used) {
4544 BestLoQuad = InputQuads.find_first();
4545 BestHiQuad = InputQuads.find_next(BestLoQuad);
4547 if (InputQuads.count() > 2) {
4553 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4554 // the shuffle mask. If a quad is scored as -1, that means that it contains
4555 // words from all 4 input quadwords.
4557 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4558 SmallVector<int, 8> MaskV;
4559 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4560 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4561 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4562 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4563 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4564 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4566 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4567 // source words for the shuffle, to aid later transformations.
4568 bool AllWordsInNewV = true;
4569 bool InOrder[2] = { true, true };
4570 for (unsigned i = 0; i != 8; ++i) {
4571 int idx = MaskVals[i];
4573 InOrder[i/4] = false;
4574 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4576 AllWordsInNewV = false;
4580 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4581 if (AllWordsInNewV) {
4582 for (int i = 0; i != 8; ++i) {
4583 int idx = MaskVals[i];
4586 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4587 if ((idx != i) && idx < 4)
4589 if ((idx != i) && idx > 3)
4598 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4599 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4600 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4601 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4602 unsigned TargetMask = 0;
4603 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4604 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4605 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4606 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4607 V1 = NewV.getOperand(0);
4608 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4612 // If we have SSSE3, and all words of the result are from 1 input vector,
4613 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4614 // is present, fall back to case 4.
4615 if (Subtarget->hasSSSE3()) {
4616 SmallVector<SDValue,16> pshufbMask;
4618 // If we have elements from both input vectors, set the high bit of the
4619 // shuffle mask element to zero out elements that come from V2 in the V1
4620 // mask, and elements that come from V1 in the V2 mask, so that the two
4621 // results can be OR'd together.
4622 bool TwoInputs = V1Used && V2Used;
4623 for (unsigned i = 0; i != 8; ++i) {
4624 int EltIdx = MaskVals[i] * 2;
4625 if (TwoInputs && (EltIdx >= 16)) {
4626 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4627 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4630 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4631 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4633 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4634 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4635 DAG.getNode(ISD::BUILD_VECTOR, dl,
4636 MVT::v16i8, &pshufbMask[0], 16));
4638 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4640 // Calculate the shuffle mask for the second input, shuffle it, and
4641 // OR it with the first shuffled input.
4643 for (unsigned i = 0; i != 8; ++i) {
4644 int EltIdx = MaskVals[i] * 2;
4646 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4647 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4650 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4651 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4653 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4654 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4655 DAG.getNode(ISD::BUILD_VECTOR, dl,
4656 MVT::v16i8, &pshufbMask[0], 16));
4657 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4658 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4661 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4662 // and update MaskVals with new element order.
4663 BitVector InOrder(8);
4664 if (BestLoQuad >= 0) {
4665 SmallVector<int, 8> MaskV;
4666 for (int i = 0; i != 4; ++i) {
4667 int idx = MaskVals[i];
4669 MaskV.push_back(-1);
4671 } else if ((idx / 4) == BestLoQuad) {
4672 MaskV.push_back(idx & 3);
4675 MaskV.push_back(-1);
4678 for (unsigned i = 4; i != 8; ++i)
4680 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4683 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4684 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4686 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4690 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4691 // and update MaskVals with the new element order.
4692 if (BestHiQuad >= 0) {
4693 SmallVector<int, 8> MaskV;
4694 for (unsigned i = 0; i != 4; ++i)
4696 for (unsigned i = 4; i != 8; ++i) {
4697 int idx = MaskVals[i];
4699 MaskV.push_back(-1);
4701 } else if ((idx / 4) == BestHiQuad) {
4702 MaskV.push_back((idx & 3) + 4);
4705 MaskV.push_back(-1);
4708 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4711 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4712 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
4714 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
4718 // In case BestHi & BestLo were both -1, which means each quadword has a word
4719 // from each of the four input quadwords, calculate the InOrder bitvector now
4720 // before falling through to the insert/extract cleanup.
4721 if (BestLoQuad == -1 && BestHiQuad == -1) {
4723 for (int i = 0; i != 8; ++i)
4724 if (MaskVals[i] < 0 || MaskVals[i] == i)
4728 // The other elements are put in the right place using pextrw and pinsrw.
4729 for (unsigned i = 0; i != 8; ++i) {
4732 int EltIdx = MaskVals[i];
4735 SDValue ExtOp = (EltIdx < 8)
4736 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4737 DAG.getIntPtrConstant(EltIdx))
4738 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4739 DAG.getIntPtrConstant(EltIdx - 8));
4740 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4741 DAG.getIntPtrConstant(i));
4746 // v16i8 shuffles - Prefer shuffles in the following order:
4747 // 1. [ssse3] 1 x pshufb
4748 // 2. [ssse3] 2 x pshufb + 1 x por
4749 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4751 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4753 const X86TargetLowering &TLI) {
4754 SDValue V1 = SVOp->getOperand(0);
4755 SDValue V2 = SVOp->getOperand(1);
4756 DebugLoc dl = SVOp->getDebugLoc();
4757 SmallVector<int, 16> MaskVals;
4758 SVOp->getMask(MaskVals);
4760 // If we have SSSE3, case 1 is generated when all result bytes come from
4761 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4762 // present, fall back to case 3.
4763 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4766 for (unsigned i = 0; i < 16; ++i) {
4767 int EltIdx = MaskVals[i];
4776 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4777 if (TLI.getSubtarget()->hasSSSE3()) {
4778 SmallVector<SDValue,16> pshufbMask;
4780 // If all result elements are from one input vector, then only translate
4781 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4783 // Otherwise, we have elements from both input vectors, and must zero out
4784 // elements that come from V2 in the first mask, and V1 in the second mask
4785 // so that we can OR them together.
4786 bool TwoInputs = !(V1Only || V2Only);
4787 for (unsigned i = 0; i != 16; ++i) {
4788 int EltIdx = MaskVals[i];
4789 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4790 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4793 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4795 // If all the elements are from V2, assign it to V1 and return after
4796 // building the first pshufb.
4799 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4800 DAG.getNode(ISD::BUILD_VECTOR, dl,
4801 MVT::v16i8, &pshufbMask[0], 16));
4805 // Calculate the shuffle mask for the second input, shuffle it, and
4806 // OR it with the first shuffled input.
4808 for (unsigned i = 0; i != 16; ++i) {
4809 int EltIdx = MaskVals[i];
4811 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4814 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4816 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4817 DAG.getNode(ISD::BUILD_VECTOR, dl,
4818 MVT::v16i8, &pshufbMask[0], 16));
4819 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4822 // No SSSE3 - Calculate in place words and then fix all out of place words
4823 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4824 // the 16 different words that comprise the two doublequadword input vectors.
4825 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4826 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4827 SDValue NewV = V2Only ? V2 : V1;
4828 for (int i = 0; i != 8; ++i) {
4829 int Elt0 = MaskVals[i*2];
4830 int Elt1 = MaskVals[i*2+1];
4832 // This word of the result is all undef, skip it.
4833 if (Elt0 < 0 && Elt1 < 0)
4836 // This word of the result is already in the correct place, skip it.
4837 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4839 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4842 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4843 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4846 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4847 // using a single extract together, load it and store it.
4848 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4849 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4850 DAG.getIntPtrConstant(Elt1 / 2));
4851 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4852 DAG.getIntPtrConstant(i));
4856 // If Elt1 is defined, extract it from the appropriate source. If the
4857 // source byte is not also odd, shift the extracted word left 8 bits
4858 // otherwise clear the bottom 8 bits if we need to do an or.
4860 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4861 DAG.getIntPtrConstant(Elt1 / 2));
4862 if ((Elt1 & 1) == 0)
4863 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4864 DAG.getConstant(8, TLI.getShiftAmountTy()));
4866 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4867 DAG.getConstant(0xFF00, MVT::i16));
4869 // If Elt0 is defined, extract it from the appropriate source. If the
4870 // source byte is not also even, shift the extracted word right 8 bits. If
4871 // Elt1 was also defined, OR the extracted values together before
4872 // inserting them in the result.
4874 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4875 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4876 if ((Elt0 & 1) != 0)
4877 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4878 DAG.getConstant(8, TLI.getShiftAmountTy()));
4880 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4881 DAG.getConstant(0x00FF, MVT::i16));
4882 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4885 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4886 DAG.getIntPtrConstant(i));
4888 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4891 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4892 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4893 /// done when every pair / quad of shuffle mask elements point to elements in
4894 /// the right sequence. e.g.
4895 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4897 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4899 const TargetLowering &TLI, DebugLoc dl) {
4900 EVT VT = SVOp->getValueType(0);
4901 SDValue V1 = SVOp->getOperand(0);
4902 SDValue V2 = SVOp->getOperand(1);
4903 unsigned NumElems = VT.getVectorNumElements();
4904 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4905 EVT MaskVT = (NewWidth == 4) ? MVT::v4i16 : MVT::v2i32;
4907 switch (VT.getSimpleVT().SimpleTy) {
4908 default: assert(false && "Unexpected!");
4909 case MVT::v4f32: NewVT = MVT::v2f64; break;
4910 case MVT::v4i32: NewVT = MVT::v2i64; break;
4911 case MVT::v8i16: NewVT = MVT::v4i32; break;
4912 case MVT::v16i8: NewVT = MVT::v4i32; break;
4915 if (NewWidth == 2) {
4921 int Scale = NumElems / NewWidth;
4922 SmallVector<int, 8> MaskVec;
4923 for (unsigned i = 0; i < NumElems; i += Scale) {
4925 for (int j = 0; j < Scale; ++j) {
4926 int EltIdx = SVOp->getMaskElt(i+j);
4930 StartIdx = EltIdx - (EltIdx % Scale);
4931 if (EltIdx != StartIdx + j)
4935 MaskVec.push_back(-1);
4937 MaskVec.push_back(StartIdx / Scale);
4940 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4941 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4942 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4945 /// getVZextMovL - Return a zero-extending vector move low node.
4947 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4948 SDValue SrcOp, SelectionDAG &DAG,
4949 const X86Subtarget *Subtarget, DebugLoc dl) {
4950 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4951 LoadSDNode *LD = NULL;
4952 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4953 LD = dyn_cast<LoadSDNode>(SrcOp);
4955 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4957 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4958 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4959 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4960 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4961 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4963 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4964 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4965 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4966 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4974 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4975 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4976 DAG.getNode(ISD::BIT_CONVERT, dl,
4980 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4983 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4984 SDValue V1 = SVOp->getOperand(0);
4985 SDValue V2 = SVOp->getOperand(1);
4986 DebugLoc dl = SVOp->getDebugLoc();
4987 EVT VT = SVOp->getValueType(0);
4989 SmallVector<std::pair<int, int>, 8> Locs;
4991 SmallVector<int, 8> Mask1(4U, -1);
4992 SmallVector<int, 8> PermMask;
4993 SVOp->getMask(PermMask);
4997 for (unsigned i = 0; i != 4; ++i) {
4998 int Idx = PermMask[i];
5000 Locs[i] = std::make_pair(-1, -1);
5002 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5004 Locs[i] = std::make_pair(0, NumLo);
5008 Locs[i] = std::make_pair(1, NumHi);
5010 Mask1[2+NumHi] = Idx;
5016 if (NumLo <= 2 && NumHi <= 2) {
5017 // If no more than two elements come from either vector. This can be
5018 // implemented with two shuffles. First shuffle gather the elements.
5019 // The second shuffle, which takes the first shuffle as both of its
5020 // vector operands, put the elements into the right order.
5021 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5023 SmallVector<int, 8> Mask2(4U, -1);
5025 for (unsigned i = 0; i != 4; ++i) {
5026 if (Locs[i].first == -1)
5029 unsigned Idx = (i < 2) ? 0 : 4;
5030 Idx += Locs[i].first * 2 + Locs[i].second;
5035 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5036 } else if (NumLo == 3 || NumHi == 3) {
5037 // Otherwise, we must have three elements from one vector, call it X, and
5038 // one element from the other, call it Y. First, use a shufps to build an
5039 // intermediate vector with the one element from Y and the element from X
5040 // that will be in the same half in the final destination (the indexes don't
5041 // matter). Then, use a shufps to build the final vector, taking the half
5042 // containing the element from Y from the intermediate, and the other half
5045 // Normalize it so the 3 elements come from V1.
5046 CommuteVectorShuffleMask(PermMask, VT);
5050 // Find the element from V2.
5052 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5053 int Val = PermMask[HiIndex];
5060 Mask1[0] = PermMask[HiIndex];
5062 Mask1[2] = PermMask[HiIndex^1];
5064 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5067 Mask1[0] = PermMask[0];
5068 Mask1[1] = PermMask[1];
5069 Mask1[2] = HiIndex & 1 ? 6 : 4;
5070 Mask1[3] = HiIndex & 1 ? 4 : 6;
5071 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5073 Mask1[0] = HiIndex & 1 ? 2 : 0;
5074 Mask1[1] = HiIndex & 1 ? 0 : 2;
5075 Mask1[2] = PermMask[2];
5076 Mask1[3] = PermMask[3];
5081 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5085 // Break it into (shuffle shuffle_hi, shuffle_lo).
5087 SmallVector<int,8> LoMask(4U, -1);
5088 SmallVector<int,8> HiMask(4U, -1);
5090 SmallVector<int,8> *MaskPtr = &LoMask;
5091 unsigned MaskIdx = 0;
5094 for (unsigned i = 0; i != 4; ++i) {
5101 int Idx = PermMask[i];
5103 Locs[i] = std::make_pair(-1, -1);
5104 } else if (Idx < 4) {
5105 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5106 (*MaskPtr)[LoIdx] = Idx;
5109 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5110 (*MaskPtr)[HiIdx] = Idx;
5115 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5116 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5117 SmallVector<int, 8> MaskOps;
5118 for (unsigned i = 0; i != 4; ++i) {
5119 if (Locs[i].first == -1) {
5120 MaskOps.push_back(-1);
5122 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5123 MaskOps.push_back(Idx);
5126 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5129 static bool MayFoldVectorLoad(SDValue V) {
5130 if (V.hasOneUse() && V.getOpcode() == ISD::BIT_CONVERT)
5131 V = V.getOperand(0);
5132 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5133 V = V.getOperand(0);
5140 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5142 SDValue V1 = Op.getOperand(0);
5143 SDValue V2 = Op.getOperand(1);
5144 EVT VT = Op.getValueType();
5146 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5148 if (HasSSE2 && VT == MVT::v2f64)
5149 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5152 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5156 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5157 SDValue V1 = Op.getOperand(0);
5158 SDValue V2 = Op.getOperand(1);
5159 EVT VT = Op.getValueType();
5161 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5162 "unsupported shuffle type");
5164 if (V2.getOpcode() == ISD::UNDEF)
5168 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5172 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5173 SDValue V1 = Op.getOperand(0);
5174 SDValue V2 = Op.getOperand(1);
5175 EVT VT = Op.getValueType();
5176 unsigned NumElems = VT.getVectorNumElements();
5178 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5179 // operand of these instructions is only memory, so check if there's a
5180 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5182 bool CanFoldLoad = false;
5184 // Trivial case, when V2 comes from a load.
5185 if (MayFoldVectorLoad(V2))
5188 // When V1 is a load, it can be folded later into a store in isel, example:
5189 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5191 // (MOVLPSmr addr:$src1, VR128:$src2)
5192 // So, recognize this potential and also use MOVLPS or MOVLPD
5193 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5197 if (HasSSE2 && NumElems == 2)
5198 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5201 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5204 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5205 // movl and movlp will both match v2i64, but v2i64 is never matched by
5206 // movl earlier because we make it strict to avoid messing with the movlp load
5207 // folding logic (see the code above getMOVLP call). Match it here then,
5208 // this is horrible, but will stay like this until we move all shuffle
5209 // matching to x86 specific nodes. Note that for the 1st condition all
5210 // types are matched with movsd.
5211 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5212 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5214 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5217 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5219 // Invert the operand order and use SHUFPS to match it.
5220 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5221 X86::getShuffleSHUFImmediate(SVOp), DAG);
5224 static inline unsigned getUNPCKLOpcode(EVT VT) {
5225 switch(VT.getSimpleVT().SimpleTy) {
5226 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5227 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5228 case MVT::v4f32: return X86ISD::UNPCKLPS;
5229 case MVT::v2f64: return X86ISD::UNPCKLPD;
5230 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5231 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5233 llvm_unreachable("Unknow type for unpckl");
5238 static inline unsigned getUNPCKHOpcode(EVT VT) {
5239 switch(VT.getSimpleVT().SimpleTy) {
5240 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5241 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5242 case MVT::v4f32: return X86ISD::UNPCKHPS;
5243 case MVT::v2f64: return X86ISD::UNPCKHPD;
5244 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5245 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5247 llvm_unreachable("Unknow type for unpckh");
5253 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5254 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5255 SDValue V1 = Op.getOperand(0);
5256 SDValue V2 = Op.getOperand(1);
5257 EVT VT = Op.getValueType();
5258 DebugLoc dl = Op.getDebugLoc();
5259 unsigned NumElems = VT.getVectorNumElements();
5260 bool isMMX = VT.getSizeInBits() == 64;
5261 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5262 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5263 bool V1IsSplat = false;
5264 bool V2IsSplat = false;
5265 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5266 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5267 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
5268 MachineFunction &MF = DAG.getMachineFunction();
5269 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5271 if (isZeroShuffle(SVOp))
5272 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5274 // Promote splats to v4f32.
5275 if (SVOp->isSplat()) {
5276 if (isMMX || NumElems < 4)
5278 return PromoteSplat(SVOp, DAG);
5281 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5283 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5284 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5285 if (NewOp.getNode())
5286 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5287 LowerVECTOR_SHUFFLE(NewOp, DAG));
5288 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5289 // FIXME: Figure out a cleaner way to do this.
5290 // Try to make use of movq to zero out the top part.
5291 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5292 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5293 if (NewOp.getNode()) {
5294 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5295 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5296 DAG, Subtarget, dl);
5298 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5299 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5300 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5301 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5302 DAG, Subtarget, dl);
5306 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
5307 // unpckh_undef). Only use pshufd if speed is more important than size.
5308 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
5309 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5310 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5311 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
5312 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5313 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5315 if (X86::isPSHUFDMask(SVOp)) {
5316 // The actual implementation will match the mask in the if above and then
5317 // during isel it can match several different instructions, not only pshufd
5318 // as its name says, sad but true, emulate the behavior for now...
5319 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5320 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5322 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5324 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5325 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5327 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5328 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5331 if (VT == MVT::v4f32)
5332 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5336 // Check if this can be converted into a logical shift.
5337 bool isLeft = false;
5340 bool isShift = getSubtarget()->hasSSE2() &&
5341 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5342 if (isShift && ShVal.hasOneUse()) {
5343 // If the shifted value has multiple uses, it may be cheaper to use
5344 // v_set0 + movlhps or movhlps, etc.
5345 EVT EltVT = VT.getVectorElementType();
5346 ShAmt *= EltVT.getSizeInBits();
5347 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5350 if (X86::isMOVLMask(SVOp)) {
5353 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5354 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5355 if (!isMMX && !X86::isMOVLPMask(SVOp)) {
5356 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5357 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5359 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5360 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5364 // FIXME: fold these into legal mask.
5366 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5367 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5369 if (X86::isMOVHLPSMask(SVOp))
5370 return getMOVHighToLow(Op, dl, DAG);
5372 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5373 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5375 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5376 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5378 if (X86::isMOVLPMask(SVOp))
5379 return getMOVLP(Op, dl, DAG, HasSSE2);
5382 if (ShouldXformToMOVHLPS(SVOp) ||
5383 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5384 return CommuteVectorShuffle(SVOp, DAG);
5387 // No better options. Use a vshl / vsrl.
5388 EVT EltVT = VT.getVectorElementType();
5389 ShAmt *= EltVT.getSizeInBits();
5390 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5393 bool Commuted = false;
5394 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5395 // 1,1,1,1 -> v8i16 though.
5396 V1IsSplat = isSplatVector(V1.getNode());
5397 V2IsSplat = isSplatVector(V2.getNode());
5399 // Canonicalize the splat or undef, if present, to be on the RHS.
5400 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5401 Op = CommuteVectorShuffle(SVOp, DAG);
5402 SVOp = cast<ShuffleVectorSDNode>(Op);
5403 V1 = SVOp->getOperand(0);
5404 V2 = SVOp->getOperand(1);
5405 std::swap(V1IsSplat, V2IsSplat);
5406 std::swap(V1IsUndef, V2IsUndef);
5410 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5411 // Shuffling low element of v1 into undef, just return v1.
5414 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5415 // the instruction selector will not match, so get a canonical MOVL with
5416 // swapped operands to undo the commute.
5417 return getMOVL(DAG, dl, VT, V2, V1);
5420 if (X86::isUNPCKLMask(SVOp))
5422 Op : getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
5424 if (X86::isUNPCKHMask(SVOp))
5426 Op : getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
5429 // Normalize mask so all entries that point to V2 points to its first
5430 // element then try to match unpck{h|l} again. If match, return a
5431 // new vector_shuffle with the corrected mask.
5432 SDValue NewMask = NormalizeMask(SVOp, DAG);
5433 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5434 if (NSVOp != SVOp) {
5435 if (X86::isUNPCKLMask(NSVOp, true)) {
5437 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5444 // Commute is back and try unpck* again.
5445 // FIXME: this seems wrong.
5446 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5447 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5449 if (X86::isUNPCKLMask(NewSVOp))
5451 NewOp : getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
5453 if (X86::isUNPCKHMask(NewSVOp))
5455 NewOp : getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
5458 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
5460 // Normalize the node to match x86 shuffle ops if needed
5461 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5462 return CommuteVectorShuffle(SVOp, DAG);
5464 // The checks below are all present in isShuffleMaskLegal, but they are
5465 // inlined here right now to enable us to directly emit target specific
5466 // nodes, and remove one by one until they don't return Op anymore.
5467 SmallVector<int, 16> M;
5470 if (isPALIGNRMask(M, VT, HasSSSE3))
5471 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
5472 X86::getShufflePALIGNRImmediate(SVOp),
5475 // Only a few shuffle masks are handled for 64-bit vectors (MMX), and
5476 // 64-bit vectors which made to this point can't be handled, they are
5481 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
5482 SVOp->getSplatIndex() == 0 && V2IsUndef) {
5483 if (VT == MVT::v2f64)
5484 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
5485 if (VT == MVT::v2i64)
5486 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
5489 if (isPSHUFHWMask(M, VT))
5490 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
5491 X86::getShufflePSHUFHWImmediate(SVOp),
5494 if (isPSHUFLWMask(M, VT))
5495 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
5496 X86::getShufflePSHUFLWImmediate(SVOp),
5499 if (isSHUFPMask(M, VT)) {
5500 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5501 if (VT == MVT::v4f32 || VT == MVT::v4i32)
5502 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
5504 if (VT == MVT::v2f64 || VT == MVT::v2i64)
5505 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
5509 if (X86::isUNPCKL_v_undef_Mask(SVOp))
5510 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5511 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5512 if (X86::isUNPCKH_v_undef_Mask(SVOp))
5513 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5514 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5516 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5517 if (VT == MVT::v8i16) {
5518 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5519 if (NewOp.getNode())
5523 if (VT == MVT::v16i8) {
5524 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5525 if (NewOp.getNode())
5529 // Handle all 4 wide cases with a number of shuffles except for MMX.
5530 if (NumElems == 4 && !isMMX)
5531 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5537 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5538 SelectionDAG &DAG) const {
5539 EVT VT = Op.getValueType();
5540 DebugLoc dl = Op.getDebugLoc();
5541 if (VT.getSizeInBits() == 8) {
5542 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
5543 Op.getOperand(0), Op.getOperand(1));
5544 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5545 DAG.getValueType(VT));
5546 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5547 } else if (VT.getSizeInBits() == 16) {
5548 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5549 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
5551 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5552 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5553 DAG.getNode(ISD::BIT_CONVERT, dl,
5557 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
5558 Op.getOperand(0), Op.getOperand(1));
5559 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5560 DAG.getValueType(VT));
5561 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5562 } else if (VT == MVT::f32) {
5563 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
5564 // the result back to FR32 register. It's only worth matching if the
5565 // result has a single use which is a store or a bitcast to i32. And in
5566 // the case of a store, it's not worth it if the index is a constant 0,
5567 // because a MOVSSmr can be used instead, which is smaller and faster.
5568 if (!Op.hasOneUse())
5570 SDNode *User = *Op.getNode()->use_begin();
5571 if ((User->getOpcode() != ISD::STORE ||
5572 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5573 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5574 (User->getOpcode() != ISD::BIT_CONVERT ||
5575 User->getValueType(0) != MVT::i32))
5577 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5578 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
5581 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
5582 } else if (VT == MVT::i32) {
5583 // ExtractPS works with constant index.
5584 if (isa<ConstantSDNode>(Op.getOperand(1)))
5592 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5593 SelectionDAG &DAG) const {
5594 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5597 if (Subtarget->hasSSE41()) {
5598 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5603 EVT VT = Op.getValueType();
5604 DebugLoc dl = Op.getDebugLoc();
5605 // TODO: handle v16i8.
5606 if (VT.getSizeInBits() == 16) {
5607 SDValue Vec = Op.getOperand(0);
5608 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5610 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5611 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5612 DAG.getNode(ISD::BIT_CONVERT, dl,
5615 // Transform it so it match pextrw which produces a 32-bit result.
5616 EVT EltVT = MVT::i32;
5617 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5618 Op.getOperand(0), Op.getOperand(1));
5619 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5620 DAG.getValueType(VT));
5621 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5622 } else if (VT.getSizeInBits() == 32) {
5623 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5627 // SHUFPS the element to the lowest double word, then movss.
5628 int Mask[4] = { Idx, -1, -1, -1 };
5629 EVT VVT = Op.getOperand(0).getValueType();
5630 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5631 DAG.getUNDEF(VVT), Mask);
5632 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5633 DAG.getIntPtrConstant(0));
5634 } else if (VT.getSizeInBits() == 64) {
5635 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5636 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5637 // to match extract_elt for f64.
5638 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5642 // UNPCKHPD the element to the lowest double word, then movsd.
5643 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5644 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5645 int Mask[2] = { 1, -1 };
5646 EVT VVT = Op.getOperand(0).getValueType();
5647 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5648 DAG.getUNDEF(VVT), Mask);
5649 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5650 DAG.getIntPtrConstant(0));
5657 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5658 SelectionDAG &DAG) const {
5659 EVT VT = Op.getValueType();
5660 EVT EltVT = VT.getVectorElementType();
5661 DebugLoc dl = Op.getDebugLoc();
5663 SDValue N0 = Op.getOperand(0);
5664 SDValue N1 = Op.getOperand(1);
5665 SDValue N2 = Op.getOperand(2);
5667 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5668 isa<ConstantSDNode>(N2)) {
5670 if (VT == MVT::v8i16)
5671 Opc = X86ISD::PINSRW;
5672 else if (VT == MVT::v4i16)
5673 Opc = X86ISD::MMX_PINSRW;
5674 else if (VT == MVT::v16i8)
5675 Opc = X86ISD::PINSRB;
5677 Opc = X86ISD::PINSRB;
5679 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5681 if (N1.getValueType() != MVT::i32)
5682 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5683 if (N2.getValueType() != MVT::i32)
5684 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5685 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5686 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5687 // Bits [7:6] of the constant are the source select. This will always be
5688 // zero here. The DAG Combiner may combine an extract_elt index into these
5689 // bits. For example (insert (extract, 3), 2) could be matched by putting
5690 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5691 // Bits [5:4] of the constant are the destination select. This is the
5692 // value of the incoming immediate.
5693 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5694 // combine either bitwise AND or insert of float 0.0 to set these bits.
5695 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5696 // Create this as a scalar to vector..
5697 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5698 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5699 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5700 // PINSR* works with constant index.
5707 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5708 EVT VT = Op.getValueType();
5709 EVT EltVT = VT.getVectorElementType();
5711 if (Subtarget->hasSSE41())
5712 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5714 if (EltVT == MVT::i8)
5717 DebugLoc dl = Op.getDebugLoc();
5718 SDValue N0 = Op.getOperand(0);
5719 SDValue N1 = Op.getOperand(1);
5720 SDValue N2 = Op.getOperand(2);
5722 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5723 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5724 // as its second argument.
5725 if (N1.getValueType() != MVT::i32)
5726 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5727 if (N2.getValueType() != MVT::i32)
5728 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5729 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5730 dl, VT, N0, N1, N2);
5736 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5737 DebugLoc dl = Op.getDebugLoc();
5739 if (Op.getValueType() == MVT::v1i64 &&
5740 Op.getOperand(0).getValueType() == MVT::i64)
5741 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5743 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5744 EVT VT = MVT::v2i32;
5745 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5752 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5753 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5756 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5757 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5758 // one of the above mentioned nodes. It has to be wrapped because otherwise
5759 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5760 // be used to form addressing mode. These wrapped nodes will be selected
5763 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5764 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5766 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5768 unsigned char OpFlag = 0;
5769 unsigned WrapperKind = X86ISD::Wrapper;
5770 CodeModel::Model M = getTargetMachine().getCodeModel();
5772 if (Subtarget->isPICStyleRIPRel() &&
5773 (M == CodeModel::Small || M == CodeModel::Kernel))
5774 WrapperKind = X86ISD::WrapperRIP;
5775 else if (Subtarget->isPICStyleGOT())
5776 OpFlag = X86II::MO_GOTOFF;
5777 else if (Subtarget->isPICStyleStubPIC())
5778 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5780 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5782 CP->getOffset(), OpFlag);
5783 DebugLoc DL = CP->getDebugLoc();
5784 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5785 // With PIC, the address is actually $g + Offset.
5787 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5788 DAG.getNode(X86ISD::GlobalBaseReg,
5789 DebugLoc(), getPointerTy()),
5796 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5797 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5799 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5801 unsigned char OpFlag = 0;
5802 unsigned WrapperKind = X86ISD::Wrapper;
5803 CodeModel::Model M = getTargetMachine().getCodeModel();
5805 if (Subtarget->isPICStyleRIPRel() &&
5806 (M == CodeModel::Small || M == CodeModel::Kernel))
5807 WrapperKind = X86ISD::WrapperRIP;
5808 else if (Subtarget->isPICStyleGOT())
5809 OpFlag = X86II::MO_GOTOFF;
5810 else if (Subtarget->isPICStyleStubPIC())
5811 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5813 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5815 DebugLoc DL = JT->getDebugLoc();
5816 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5818 // With PIC, the address is actually $g + Offset.
5820 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5821 DAG.getNode(X86ISD::GlobalBaseReg,
5822 DebugLoc(), getPointerTy()),
5830 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5831 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5833 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5835 unsigned char OpFlag = 0;
5836 unsigned WrapperKind = X86ISD::Wrapper;
5837 CodeModel::Model M = getTargetMachine().getCodeModel();
5839 if (Subtarget->isPICStyleRIPRel() &&
5840 (M == CodeModel::Small || M == CodeModel::Kernel))
5841 WrapperKind = X86ISD::WrapperRIP;
5842 else if (Subtarget->isPICStyleGOT())
5843 OpFlag = X86II::MO_GOTOFF;
5844 else if (Subtarget->isPICStyleStubPIC())
5845 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5847 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5849 DebugLoc DL = Op.getDebugLoc();
5850 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5853 // With PIC, the address is actually $g + Offset.
5854 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5855 !Subtarget->is64Bit()) {
5856 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5857 DAG.getNode(X86ISD::GlobalBaseReg,
5858 DebugLoc(), getPointerTy()),
5866 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5867 // Create the TargetBlockAddressAddress node.
5868 unsigned char OpFlags =
5869 Subtarget->ClassifyBlockAddressReference();
5870 CodeModel::Model M = getTargetMachine().getCodeModel();
5871 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5872 DebugLoc dl = Op.getDebugLoc();
5873 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5874 /*isTarget=*/true, OpFlags);
5876 if (Subtarget->isPICStyleRIPRel() &&
5877 (M == CodeModel::Small || M == CodeModel::Kernel))
5878 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5880 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5882 // With PIC, the address is actually $g + Offset.
5883 if (isGlobalRelativeToPICBase(OpFlags)) {
5884 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5885 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5893 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5895 SelectionDAG &DAG) const {
5896 // Create the TargetGlobalAddress node, folding in the constant
5897 // offset if it is legal.
5898 unsigned char OpFlags =
5899 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5900 CodeModel::Model M = getTargetMachine().getCodeModel();
5902 if (OpFlags == X86II::MO_NO_FLAG &&
5903 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5904 // A direct static reference to a global.
5905 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5908 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5911 if (Subtarget->isPICStyleRIPRel() &&
5912 (M == CodeModel::Small || M == CodeModel::Kernel))
5913 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5915 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5917 // With PIC, the address is actually $g + Offset.
5918 if (isGlobalRelativeToPICBase(OpFlags)) {
5919 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5920 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5924 // For globals that require a load from a stub to get the address, emit the
5926 if (isGlobalStubReference(OpFlags))
5927 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5928 PseudoSourceValue::getGOT(), 0, false, false, 0);
5930 // If there was a non-zero offset that we didn't fold, create an explicit
5933 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5934 DAG.getConstant(Offset, getPointerTy()));
5940 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5941 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5942 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5943 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5947 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5948 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5949 unsigned char OperandFlags) {
5950 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5951 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5952 DebugLoc dl = GA->getDebugLoc();
5953 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5954 GA->getValueType(0),
5958 SDValue Ops[] = { Chain, TGA, *InFlag };
5959 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5961 SDValue Ops[] = { Chain, TGA };
5962 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5965 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5966 MFI->setAdjustsStack(true);
5968 SDValue Flag = Chain.getValue(1);
5969 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5972 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5974 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5977 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5978 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5979 DAG.getNode(X86ISD::GlobalBaseReg,
5980 DebugLoc(), PtrVT), InFlag);
5981 InFlag = Chain.getValue(1);
5983 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5986 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5988 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5990 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5991 X86::RAX, X86II::MO_TLSGD);
5994 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5995 // "local exec" model.
5996 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5997 const EVT PtrVT, TLSModel::Model model,
5999 DebugLoc dl = GA->getDebugLoc();
6000 // Get the Thread Pointer
6001 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
6003 DAG.getRegister(is64Bit? X86::FS : X86::GS,
6006 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
6007 NULL, 0, false, false, 0);
6009 unsigned char OperandFlags = 0;
6010 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6012 unsigned WrapperKind = X86ISD::Wrapper;
6013 if (model == TLSModel::LocalExec) {
6014 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6015 } else if (is64Bit) {
6016 assert(model == TLSModel::InitialExec);
6017 OperandFlags = X86II::MO_GOTTPOFF;
6018 WrapperKind = X86ISD::WrapperRIP;
6020 assert(model == TLSModel::InitialExec);
6021 OperandFlags = X86II::MO_INDNTPOFF;
6024 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6026 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6027 GA->getValueType(0),
6028 GA->getOffset(), OperandFlags);
6029 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6031 if (model == TLSModel::InitialExec)
6032 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6033 PseudoSourceValue::getGOT(), 0, false, false, 0);
6035 // The address of the thread local variable is the add of the thread
6036 // pointer with the offset of the variable.
6037 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6041 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6043 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6044 const GlobalValue *GV = GA->getGlobal();
6046 if (Subtarget->isTargetELF()) {
6047 // TODO: implement the "local dynamic" model
6048 // TODO: implement the "initial exec"model for pic executables
6050 // If GV is an alias then use the aliasee for determining
6051 // thread-localness.
6052 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6053 GV = GA->resolveAliasedGlobal(false);
6055 TLSModel::Model model
6056 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6059 case TLSModel::GeneralDynamic:
6060 case TLSModel::LocalDynamic: // not implemented
6061 if (Subtarget->is64Bit())
6062 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6063 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6065 case TLSModel::InitialExec:
6066 case TLSModel::LocalExec:
6067 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6068 Subtarget->is64Bit());
6070 } else if (Subtarget->isTargetDarwin()) {
6071 // Darwin only has one model of TLS. Lower to that.
6072 unsigned char OpFlag = 0;
6073 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6074 X86ISD::WrapperRIP : X86ISD::Wrapper;
6076 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6078 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6079 !Subtarget->is64Bit();
6081 OpFlag = X86II::MO_TLVP_PIC_BASE;
6083 OpFlag = X86II::MO_TLVP;
6084 DebugLoc DL = Op.getDebugLoc();
6085 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6087 GA->getOffset(), OpFlag);
6088 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6090 // With PIC32, the address is actually $g + Offset.
6092 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6093 DAG.getNode(X86ISD::GlobalBaseReg,
6094 DebugLoc(), getPointerTy()),
6097 // Lowering the machine isd will make sure everything is in the right
6099 SDValue Args[] = { Offset };
6100 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
6102 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6103 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6104 MFI->setAdjustsStack(true);
6106 // And our return value (tls address) is in the standard call return value
6108 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
6109 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6113 "TLS not implemented for this target.");
6115 llvm_unreachable("Unreachable");
6120 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
6121 /// take a 2 x i32 value to shift plus a shift amount.
6122 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
6123 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6124 EVT VT = Op.getValueType();
6125 unsigned VTBits = VT.getSizeInBits();
6126 DebugLoc dl = Op.getDebugLoc();
6127 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6128 SDValue ShOpLo = Op.getOperand(0);
6129 SDValue ShOpHi = Op.getOperand(1);
6130 SDValue ShAmt = Op.getOperand(2);
6131 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6132 DAG.getConstant(VTBits - 1, MVT::i8))
6133 : DAG.getConstant(0, VT);
6136 if (Op.getOpcode() == ISD::SHL_PARTS) {
6137 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6138 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6140 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6141 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6144 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6145 DAG.getConstant(VTBits, MVT::i8));
6146 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6147 AndNode, DAG.getConstant(0, MVT::i8));
6150 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6151 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6152 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6154 if (Op.getOpcode() == ISD::SHL_PARTS) {
6155 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6156 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6158 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6159 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6162 SDValue Ops[2] = { Lo, Hi };
6163 return DAG.getMergeValues(Ops, 2, dl);
6166 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6167 SelectionDAG &DAG) const {
6168 EVT SrcVT = Op.getOperand(0).getValueType();
6170 if (SrcVT.isVector()) {
6171 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
6177 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6178 "Unknown SINT_TO_FP to lower!");
6180 // These are really Legal; return the operand so the caller accepts it as
6182 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6184 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6185 Subtarget->is64Bit()) {
6189 DebugLoc dl = Op.getDebugLoc();
6190 unsigned Size = SrcVT.getSizeInBits()/8;
6191 MachineFunction &MF = DAG.getMachineFunction();
6192 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6193 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6194 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6196 PseudoSourceValue::getFixedStack(SSFI), 0,
6198 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6201 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6203 SelectionDAG &DAG) const {
6205 DebugLoc dl = Op.getDebugLoc();
6207 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6209 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
6211 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6212 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6213 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
6214 Tys, Ops, array_lengthof(Ops));
6217 Chain = Result.getValue(1);
6218 SDValue InFlag = Result.getValue(2);
6220 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6221 // shouldn't be necessary except that RFP cannot be live across
6222 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6223 MachineFunction &MF = DAG.getMachineFunction();
6224 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
6225 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6226 Tys = DAG.getVTList(MVT::Other);
6228 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6230 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
6231 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
6232 PseudoSourceValue::getFixedStack(SSFI), 0,
6239 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6240 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6241 SelectionDAG &DAG) const {
6242 // This algorithm is not obvious. Here it is in C code, more or less:
6244 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6245 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6246 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6248 // Copy ints to xmm registers.
6249 __m128i xh = _mm_cvtsi32_si128( hi );
6250 __m128i xl = _mm_cvtsi32_si128( lo );
6252 // Combine into low half of a single xmm register.
6253 __m128i x = _mm_unpacklo_epi32( xh, xl );
6257 // Merge in appropriate exponents to give the integer bits the right
6259 x = _mm_unpacklo_epi32( x, exp );
6261 // Subtract away the biases to deal with the IEEE-754 double precision
6263 d = _mm_sub_pd( (__m128d) x, bias );
6265 // All conversions up to here are exact. The correctly rounded result is
6266 // calculated using the current rounding mode using the following
6268 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6269 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6270 // store doesn't really need to be here (except
6271 // maybe to zero the other double)
6276 DebugLoc dl = Op.getDebugLoc();
6277 LLVMContext *Context = DAG.getContext();
6279 // Build some magic constants.
6280 std::vector<Constant*> CV0;
6281 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6282 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6283 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6284 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6285 Constant *C0 = ConstantVector::get(CV0);
6286 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6288 std::vector<Constant*> CV1;
6290 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6292 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6293 Constant *C1 = ConstantVector::get(CV1);
6294 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6296 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6297 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6299 DAG.getIntPtrConstant(1)));
6300 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6301 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6303 DAG.getIntPtrConstant(0)));
6304 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6305 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6306 PseudoSourceValue::getConstantPool(), 0,
6308 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6309 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
6310 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6311 PseudoSourceValue::getConstantPool(), 0,
6313 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6315 // Add the halves; easiest way is to swap them into another reg first.
6316 int ShufMask[2] = { 1, -1 };
6317 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6318 DAG.getUNDEF(MVT::v2f64), ShufMask);
6319 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6320 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6321 DAG.getIntPtrConstant(0));
6324 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6325 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6326 SelectionDAG &DAG) const {
6327 DebugLoc dl = Op.getDebugLoc();
6328 // FP constant to bias correct the final result.
6329 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6332 // Load the 32-bit value into an XMM register.
6333 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6334 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6336 DAG.getIntPtrConstant(0)));
6338 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6339 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
6340 DAG.getIntPtrConstant(0));
6342 // Or the load with the bias.
6343 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6344 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6345 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6347 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6348 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6349 MVT::v2f64, Bias)));
6350 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6351 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
6352 DAG.getIntPtrConstant(0));
6354 // Subtract the bias.
6355 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
6357 // Handle final rounding.
6358 EVT DestVT = Op.getValueType();
6360 if (DestVT.bitsLT(MVT::f64)) {
6361 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
6362 DAG.getIntPtrConstant(0));
6363 } else if (DestVT.bitsGT(MVT::f64)) {
6364 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
6367 // Handle final rounding.
6371 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
6372 SelectionDAG &DAG) const {
6373 SDValue N0 = Op.getOperand(0);
6374 DebugLoc dl = Op.getDebugLoc();
6376 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
6377 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
6378 // the optimization here.
6379 if (DAG.SignBitIsZero(N0))
6380 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
6382 EVT SrcVT = N0.getValueType();
6383 EVT DstVT = Op.getValueType();
6384 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
6385 return LowerUINT_TO_FP_i64(Op, DAG);
6386 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
6387 return LowerUINT_TO_FP_i32(Op, DAG);
6389 // Make a 64-bit buffer, and use it to build an FILD.
6390 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
6391 if (SrcVT == MVT::i32) {
6392 SDValue WordOff = DAG.getConstant(4, getPointerTy());
6393 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
6394 getPointerTy(), StackSlot, WordOff);
6395 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6396 StackSlot, NULL, 0, false, false, 0);
6397 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
6398 OffsetSlot, NULL, 0, false, false, 0);
6399 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
6403 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
6404 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6405 StackSlot, NULL, 0, false, false, 0);
6406 // For i64 source, we need to add the appropriate power of 2 if the input
6407 // was negative. This is the same as the optimization in
6408 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
6409 // we must be careful to do the computation in x87 extended precision, not
6410 // in SSE. (The generic code can't know it's OK to do this, or how to.)
6411 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
6412 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
6413 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
6415 APInt FF(32, 0x5F800000ULL);
6417 // Check whether the sign bit is set.
6418 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
6419 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
6422 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
6423 SDValue FudgePtr = DAG.getConstantPool(
6424 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
6427 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
6428 SDValue Zero = DAG.getIntPtrConstant(0);
6429 SDValue Four = DAG.getIntPtrConstant(4);
6430 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
6432 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
6434 // Load the value out, extending it from f32 to f80.
6435 // FIXME: Avoid the extend by constructing the right constant pool?
6436 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
6437 FudgePtr, PseudoSourceValue::getConstantPool(),
6438 0, MVT::f32, false, false, 4);
6439 // Extend everything to 80 bits to force it to be done on x87.
6440 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
6441 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
6444 std::pair<SDValue,SDValue> X86TargetLowering::
6445 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
6446 DebugLoc dl = Op.getDebugLoc();
6448 EVT DstTy = Op.getValueType();
6451 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
6455 assert(DstTy.getSimpleVT() <= MVT::i64 &&
6456 DstTy.getSimpleVT() >= MVT::i16 &&
6457 "Unknown FP_TO_SINT to lower!");
6459 // These are really Legal.
6460 if (DstTy == MVT::i32 &&
6461 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6462 return std::make_pair(SDValue(), SDValue());
6463 if (Subtarget->is64Bit() &&
6464 DstTy == MVT::i64 &&
6465 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6466 return std::make_pair(SDValue(), SDValue());
6468 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
6470 MachineFunction &MF = DAG.getMachineFunction();
6471 unsigned MemSize = DstTy.getSizeInBits()/8;
6472 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6473 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6476 switch (DstTy.getSimpleVT().SimpleTy) {
6477 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
6478 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
6479 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
6480 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
6483 SDValue Chain = DAG.getEntryNode();
6484 SDValue Value = Op.getOperand(0);
6485 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
6486 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
6487 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
6488 PseudoSourceValue::getFixedStack(SSFI), 0,
6490 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
6492 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
6494 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
6495 Chain = Value.getValue(1);
6496 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6497 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6500 // Build the FP_TO_INT*_IN_MEM
6501 SDValue Ops[] = { Chain, Value, StackSlot };
6502 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
6504 return std::make_pair(FIST, StackSlot);
6507 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
6508 SelectionDAG &DAG) const {
6509 if (Op.getValueType().isVector()) {
6510 if (Op.getValueType() == MVT::v2i32 &&
6511 Op.getOperand(0).getValueType() == MVT::v2f64) {
6517 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
6518 SDValue FIST = Vals.first, StackSlot = Vals.second;
6519 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
6520 if (FIST.getNode() == 0) return Op;
6523 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6524 FIST, StackSlot, NULL, 0, false, false, 0);
6527 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
6528 SelectionDAG &DAG) const {
6529 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
6530 SDValue FIST = Vals.first, StackSlot = Vals.second;
6531 assert(FIST.getNode() && "Unexpected failure");
6534 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6535 FIST, StackSlot, NULL, 0, false, false, 0);
6538 SDValue X86TargetLowering::LowerFABS(SDValue Op,
6539 SelectionDAG &DAG) const {
6540 LLVMContext *Context = DAG.getContext();
6541 DebugLoc dl = Op.getDebugLoc();
6542 EVT VT = Op.getValueType();
6545 EltVT = VT.getVectorElementType();
6546 std::vector<Constant*> CV;
6547 if (EltVT == MVT::f64) {
6548 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
6552 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
6558 Constant *C = ConstantVector::get(CV);
6559 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6560 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6561 PseudoSourceValue::getConstantPool(), 0,
6563 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
6566 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
6567 LLVMContext *Context = DAG.getContext();
6568 DebugLoc dl = Op.getDebugLoc();
6569 EVT VT = Op.getValueType();
6572 EltVT = VT.getVectorElementType();
6573 std::vector<Constant*> CV;
6574 if (EltVT == MVT::f64) {
6575 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
6579 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
6585 Constant *C = ConstantVector::get(CV);
6586 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6587 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6588 PseudoSourceValue::getConstantPool(), 0,
6590 if (VT.isVector()) {
6591 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6592 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6593 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6595 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
6597 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6601 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6602 LLVMContext *Context = DAG.getContext();
6603 SDValue Op0 = Op.getOperand(0);
6604 SDValue Op1 = Op.getOperand(1);
6605 DebugLoc dl = Op.getDebugLoc();
6606 EVT VT = Op.getValueType();
6607 EVT SrcVT = Op1.getValueType();
6609 // If second operand is smaller, extend it first.
6610 if (SrcVT.bitsLT(VT)) {
6611 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6614 // And if it is bigger, shrink it first.
6615 if (SrcVT.bitsGT(VT)) {
6616 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6620 // At this point the operands and the result should have the same
6621 // type, and that won't be f80 since that is not custom lowered.
6623 // First get the sign bit of second operand.
6624 std::vector<Constant*> CV;
6625 if (SrcVT == MVT::f64) {
6626 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6627 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6629 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6630 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6631 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6632 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6634 Constant *C = ConstantVector::get(CV);
6635 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6636 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6637 PseudoSourceValue::getConstantPool(), 0,
6639 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6641 // Shift sign bit right or left if the two operands have different types.
6642 if (SrcVT.bitsGT(VT)) {
6643 // Op0 is MVT::f32, Op1 is MVT::f64.
6644 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6645 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6646 DAG.getConstant(32, MVT::i32));
6647 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
6648 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6649 DAG.getIntPtrConstant(0));
6652 // Clear first operand sign bit.
6654 if (VT == MVT::f64) {
6655 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6656 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6658 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6659 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6660 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6661 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6663 C = ConstantVector::get(CV);
6664 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6665 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6666 PseudoSourceValue::getConstantPool(), 0,
6668 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6670 // Or the value with the sign bit.
6671 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6674 /// Emit nodes that will be selected as "test Op0,Op0", or something
6676 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6677 SelectionDAG &DAG) const {
6678 DebugLoc dl = Op.getDebugLoc();
6680 // CF and OF aren't always set the way we want. Determine which
6681 // of these we need.
6682 bool NeedCF = false;
6683 bool NeedOF = false;
6686 case X86::COND_A: case X86::COND_AE:
6687 case X86::COND_B: case X86::COND_BE:
6690 case X86::COND_G: case X86::COND_GE:
6691 case X86::COND_L: case X86::COND_LE:
6692 case X86::COND_O: case X86::COND_NO:
6697 // See if we can use the EFLAGS value from the operand instead of
6698 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6699 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6700 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6701 // Emit a CMP with 0, which is the TEST pattern.
6702 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6703 DAG.getConstant(0, Op.getValueType()));
6705 unsigned Opcode = 0;
6706 unsigned NumOperands = 0;
6707 switch (Op.getNode()->getOpcode()) {
6709 // Due to an isel shortcoming, be conservative if this add is likely to be
6710 // selected as part of a load-modify-store instruction. When the root node
6711 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6712 // uses of other nodes in the match, such as the ADD in this case. This
6713 // leads to the ADD being left around and reselected, with the result being
6714 // two adds in the output. Alas, even if none our users are stores, that
6715 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6716 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6717 // climbing the DAG back to the root, and it doesn't seem to be worth the
6719 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6720 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6721 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6724 if (ConstantSDNode *C =
6725 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6726 // An add of one will be selected as an INC.
6727 if (C->getAPIntValue() == 1) {
6728 Opcode = X86ISD::INC;
6733 // An add of negative one (subtract of one) will be selected as a DEC.
6734 if (C->getAPIntValue().isAllOnesValue()) {
6735 Opcode = X86ISD::DEC;
6741 // Otherwise use a regular EFLAGS-setting add.
6742 Opcode = X86ISD::ADD;
6746 // If the primary and result isn't used, don't bother using X86ISD::AND,
6747 // because a TEST instruction will be better.
6748 bool NonFlagUse = false;
6749 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6750 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6752 unsigned UOpNo = UI.getOperandNo();
6753 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6754 // Look pass truncate.
6755 UOpNo = User->use_begin().getOperandNo();
6756 User = *User->use_begin();
6759 if (User->getOpcode() != ISD::BRCOND &&
6760 User->getOpcode() != ISD::SETCC &&
6761 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6774 // Due to the ISEL shortcoming noted above, be conservative if this op is
6775 // likely to be selected as part of a load-modify-store instruction.
6776 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6777 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6778 if (UI->getOpcode() == ISD::STORE)
6781 // Otherwise use a regular EFLAGS-setting instruction.
6782 switch (Op.getNode()->getOpcode()) {
6783 default: llvm_unreachable("unexpected operator!");
6784 case ISD::SUB: Opcode = X86ISD::SUB; break;
6785 case ISD::OR: Opcode = X86ISD::OR; break;
6786 case ISD::XOR: Opcode = X86ISD::XOR; break;
6787 case ISD::AND: Opcode = X86ISD::AND; break;
6799 return SDValue(Op.getNode(), 1);
6806 // Emit a CMP with 0, which is the TEST pattern.
6807 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6808 DAG.getConstant(0, Op.getValueType()));
6810 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6811 SmallVector<SDValue, 4> Ops;
6812 for (unsigned i = 0; i != NumOperands; ++i)
6813 Ops.push_back(Op.getOperand(i));
6815 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6816 DAG.ReplaceAllUsesWith(Op, New);
6817 return SDValue(New.getNode(), 1);
6820 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6822 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6823 SelectionDAG &DAG) const {
6824 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6825 if (C->getAPIntValue() == 0)
6826 return EmitTest(Op0, X86CC, DAG);
6828 DebugLoc dl = Op0.getDebugLoc();
6829 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6832 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6833 /// if it's possible.
6834 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6835 DebugLoc dl, SelectionDAG &DAG) const {
6836 SDValue Op0 = And.getOperand(0);
6837 SDValue Op1 = And.getOperand(1);
6838 if (Op0.getOpcode() == ISD::TRUNCATE)
6839 Op0 = Op0.getOperand(0);
6840 if (Op1.getOpcode() == ISD::TRUNCATE)
6841 Op1 = Op1.getOperand(0);
6844 if (Op1.getOpcode() == ISD::SHL)
6845 std::swap(Op0, Op1);
6846 if (Op0.getOpcode() == ISD::SHL) {
6847 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6848 if (And00C->getZExtValue() == 1) {
6849 // If we looked past a truncate, check that it's only truncating away
6851 unsigned BitWidth = Op0.getValueSizeInBits();
6852 unsigned AndBitWidth = And.getValueSizeInBits();
6853 if (BitWidth > AndBitWidth) {
6854 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6855 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6856 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6860 RHS = Op0.getOperand(1);
6862 } else if (Op1.getOpcode() == ISD::Constant) {
6863 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6864 SDValue AndLHS = Op0;
6865 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6866 LHS = AndLHS.getOperand(0);
6867 RHS = AndLHS.getOperand(1);
6871 if (LHS.getNode()) {
6872 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6873 // instruction. Since the shift amount is in-range-or-undefined, we know
6874 // that doing a bittest on the i32 value is ok. We extend to i32 because
6875 // the encoding for the i16 version is larger than the i32 version.
6876 // Also promote i16 to i32 for performance / code size reason.
6877 if (LHS.getValueType() == MVT::i8 ||
6878 LHS.getValueType() == MVT::i16)
6879 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6881 // If the operand types disagree, extend the shift amount to match. Since
6882 // BT ignores high bits (like shifts) we can use anyextend.
6883 if (LHS.getValueType() != RHS.getValueType())
6884 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6886 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6887 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6888 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6889 DAG.getConstant(Cond, MVT::i8), BT);
6895 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6896 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6897 SDValue Op0 = Op.getOperand(0);
6898 SDValue Op1 = Op.getOperand(1);
6899 DebugLoc dl = Op.getDebugLoc();
6900 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6902 // Optimize to BT if possible.
6903 // Lower (X & (1 << N)) == 0 to BT(X, N).
6904 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6905 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6906 if (Op0.getOpcode() == ISD::AND &&
6908 Op1.getOpcode() == ISD::Constant &&
6909 cast<ConstantSDNode>(Op1)->isNullValue() &&
6910 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6911 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6912 if (NewSetCC.getNode())
6916 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6917 if (Op0.getOpcode() == X86ISD::SETCC &&
6918 Op1.getOpcode() == ISD::Constant &&
6919 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6920 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6921 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6922 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6923 bool Invert = (CC == ISD::SETNE) ^
6924 cast<ConstantSDNode>(Op1)->isNullValue();
6926 CCode = X86::GetOppositeBranchCondition(CCode);
6927 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6928 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6931 bool isFP = Op1.getValueType().isFloatingPoint();
6932 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6933 if (X86CC == X86::COND_INVALID)
6936 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6938 // Use sbb x, x to materialize carry bit into a GPR.
6939 if (X86CC == X86::COND_B)
6940 return DAG.getNode(ISD::AND, dl, MVT::i8,
6941 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6942 DAG.getConstant(X86CC, MVT::i8), Cond),
6943 DAG.getConstant(1, MVT::i8));
6945 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6946 DAG.getConstant(X86CC, MVT::i8), Cond);
6949 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6951 SDValue Op0 = Op.getOperand(0);
6952 SDValue Op1 = Op.getOperand(1);
6953 SDValue CC = Op.getOperand(2);
6954 EVT VT = Op.getValueType();
6955 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6956 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6957 DebugLoc dl = Op.getDebugLoc();
6961 EVT VT0 = Op0.getValueType();
6962 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6963 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6966 switch (SetCCOpcode) {
6969 case ISD::SETEQ: SSECC = 0; break;
6971 case ISD::SETGT: Swap = true; // Fallthrough
6973 case ISD::SETOLT: SSECC = 1; break;
6975 case ISD::SETGE: Swap = true; // Fallthrough
6977 case ISD::SETOLE: SSECC = 2; break;
6978 case ISD::SETUO: SSECC = 3; break;
6980 case ISD::SETNE: SSECC = 4; break;
6981 case ISD::SETULE: Swap = true;
6982 case ISD::SETUGE: SSECC = 5; break;
6983 case ISD::SETULT: Swap = true;
6984 case ISD::SETUGT: SSECC = 6; break;
6985 case ISD::SETO: SSECC = 7; break;
6988 std::swap(Op0, Op1);
6990 // In the two special cases we can't handle, emit two comparisons.
6992 if (SetCCOpcode == ISD::SETUEQ) {
6994 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6995 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6996 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6998 else if (SetCCOpcode == ISD::SETONE) {
7000 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7001 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7002 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7004 llvm_unreachable("Illegal FP comparison");
7006 // Handle all other FP comparisons here.
7007 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7010 // We are handling one of the integer comparisons here. Since SSE only has
7011 // GT and EQ comparisons for integer, swapping operands and multiple
7012 // operations may be required for some comparisons.
7013 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7014 bool Swap = false, Invert = false, FlipSigns = false;
7016 switch (VT.getSimpleVT().SimpleTy) {
7019 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7021 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7023 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7024 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7027 switch (SetCCOpcode) {
7029 case ISD::SETNE: Invert = true;
7030 case ISD::SETEQ: Opc = EQOpc; break;
7031 case ISD::SETLT: Swap = true;
7032 case ISD::SETGT: Opc = GTOpc; break;
7033 case ISD::SETGE: Swap = true;
7034 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7035 case ISD::SETULT: Swap = true;
7036 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7037 case ISD::SETUGE: Swap = true;
7038 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7041 std::swap(Op0, Op1);
7043 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7044 // bits of the inputs before performing those operations.
7046 EVT EltVT = VT.getVectorElementType();
7047 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7049 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7050 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7052 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7053 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7056 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7058 // If the logical-not of the result is required, perform that now.
7060 Result = DAG.getNOT(dl, Result, VT);
7065 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
7066 static bool isX86LogicalCmp(SDValue Op) {
7067 unsigned Opc = Op.getNode()->getOpcode();
7068 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
7070 if (Op.getResNo() == 1 &&
7071 (Opc == X86ISD::ADD ||
7072 Opc == X86ISD::SUB ||
7073 Opc == X86ISD::SMUL ||
7074 Opc == X86ISD::UMUL ||
7075 Opc == X86ISD::INC ||
7076 Opc == X86ISD::DEC ||
7077 Opc == X86ISD::OR ||
7078 Opc == X86ISD::XOR ||
7079 Opc == X86ISD::AND))
7085 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
7086 bool addTest = true;
7087 SDValue Cond = Op.getOperand(0);
7088 DebugLoc dl = Op.getDebugLoc();
7091 if (Cond.getOpcode() == ISD::SETCC) {
7092 SDValue NewCond = LowerSETCC(Cond, DAG);
7093 if (NewCond.getNode())
7097 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
7098 SDValue Op1 = Op.getOperand(1);
7099 SDValue Op2 = Op.getOperand(2);
7100 if (Cond.getOpcode() == X86ISD::SETCC &&
7101 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
7102 SDValue Cmp = Cond.getOperand(1);
7103 if (Cmp.getOpcode() == X86ISD::CMP) {
7104 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
7105 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
7106 ConstantSDNode *RHSC =
7107 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
7108 if (N1C && N1C->isAllOnesValue() &&
7109 N2C && N2C->isNullValue() &&
7110 RHSC && RHSC->isNullValue()) {
7111 SDValue CmpOp0 = Cmp.getOperand(0);
7112 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
7113 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7114 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
7115 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7120 // Look pass (and (setcc_carry (cmp ...)), 1).
7121 if (Cond.getOpcode() == ISD::AND &&
7122 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7123 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7124 if (C && C->getAPIntValue() == 1)
7125 Cond = Cond.getOperand(0);
7128 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7129 // setting operand in place of the X86ISD::SETCC.
7130 if (Cond.getOpcode() == X86ISD::SETCC ||
7131 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7132 CC = Cond.getOperand(0);
7134 SDValue Cmp = Cond.getOperand(1);
7135 unsigned Opc = Cmp.getOpcode();
7136 EVT VT = Op.getValueType();
7138 bool IllegalFPCMov = false;
7139 if (VT.isFloatingPoint() && !VT.isVector() &&
7140 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7141 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7143 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7144 Opc == X86ISD::BT) { // FIXME
7151 // Look pass the truncate.
7152 if (Cond.getOpcode() == ISD::TRUNCATE)
7153 Cond = Cond.getOperand(0);
7155 // We know the result of AND is compared against zero. Try to match
7157 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7158 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7159 if (NewSetCC.getNode()) {
7160 CC = NewSetCC.getOperand(0);
7161 Cond = NewSetCC.getOperand(1);
7168 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7169 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7172 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7173 // condition is true.
7174 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
7175 SDValue Ops[] = { Op2, Op1, CC, Cond };
7176 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
7179 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7180 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7181 // from the AND / OR.
7182 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7183 Opc = Op.getOpcode();
7184 if (Opc != ISD::OR && Opc != ISD::AND)
7186 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7187 Op.getOperand(0).hasOneUse() &&
7188 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7189 Op.getOperand(1).hasOneUse());
7192 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7193 // 1 and that the SETCC node has a single use.
7194 static bool isXor1OfSetCC(SDValue Op) {
7195 if (Op.getOpcode() != ISD::XOR)
7197 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7198 if (N1C && N1C->getAPIntValue() == 1) {
7199 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7200 Op.getOperand(0).hasOneUse();
7205 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7206 bool addTest = true;
7207 SDValue Chain = Op.getOperand(0);
7208 SDValue Cond = Op.getOperand(1);
7209 SDValue Dest = Op.getOperand(2);
7210 DebugLoc dl = Op.getDebugLoc();
7213 if (Cond.getOpcode() == ISD::SETCC) {
7214 SDValue NewCond = LowerSETCC(Cond, DAG);
7215 if (NewCond.getNode())
7219 // FIXME: LowerXALUO doesn't handle these!!
7220 else if (Cond.getOpcode() == X86ISD::ADD ||
7221 Cond.getOpcode() == X86ISD::SUB ||
7222 Cond.getOpcode() == X86ISD::SMUL ||
7223 Cond.getOpcode() == X86ISD::UMUL)
7224 Cond = LowerXALUO(Cond, DAG);
7227 // Look pass (and (setcc_carry (cmp ...)), 1).
7228 if (Cond.getOpcode() == ISD::AND &&
7229 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7230 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7231 if (C && C->getAPIntValue() == 1)
7232 Cond = Cond.getOperand(0);
7235 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7236 // setting operand in place of the X86ISD::SETCC.
7237 if (Cond.getOpcode() == X86ISD::SETCC ||
7238 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7239 CC = Cond.getOperand(0);
7241 SDValue Cmp = Cond.getOperand(1);
7242 unsigned Opc = Cmp.getOpcode();
7243 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7244 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7248 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7252 // These can only come from an arithmetic instruction with overflow,
7253 // e.g. SADDO, UADDO.
7254 Cond = Cond.getNode()->getOperand(1);
7261 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7262 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7263 if (CondOpc == ISD::OR) {
7264 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7265 // two branches instead of an explicit OR instruction with a
7267 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7268 isX86LogicalCmp(Cmp)) {
7269 CC = Cond.getOperand(0).getOperand(0);
7270 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7271 Chain, Dest, CC, Cmp);
7272 CC = Cond.getOperand(1).getOperand(0);
7276 } else { // ISD::AND
7277 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7278 // two branches instead of an explicit AND instruction with a
7279 // separate test. However, we only do this if this block doesn't
7280 // have a fall-through edge, because this requires an explicit
7281 // jmp when the condition is false.
7282 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7283 isX86LogicalCmp(Cmp) &&
7284 Op.getNode()->hasOneUse()) {
7285 X86::CondCode CCode =
7286 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7287 CCode = X86::GetOppositeBranchCondition(CCode);
7288 CC = DAG.getConstant(CCode, MVT::i8);
7289 SDNode *User = *Op.getNode()->use_begin();
7290 // Look for an unconditional branch following this conditional branch.
7291 // We need this because we need to reverse the successors in order
7292 // to implement FCMP_OEQ.
7293 if (User->getOpcode() == ISD::BR) {
7294 SDValue FalseBB = User->getOperand(1);
7296 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
7297 assert(NewBR == User);
7301 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7302 Chain, Dest, CC, Cmp);
7303 X86::CondCode CCode =
7304 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
7305 CCode = X86::GetOppositeBranchCondition(CCode);
7306 CC = DAG.getConstant(CCode, MVT::i8);
7312 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
7313 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
7314 // It should be transformed during dag combiner except when the condition
7315 // is set by a arithmetics with overflow node.
7316 X86::CondCode CCode =
7317 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7318 CCode = X86::GetOppositeBranchCondition(CCode);
7319 CC = DAG.getConstant(CCode, MVT::i8);
7320 Cond = Cond.getOperand(0).getOperand(1);
7326 // Look pass the truncate.
7327 if (Cond.getOpcode() == ISD::TRUNCATE)
7328 Cond = Cond.getOperand(0);
7330 // We know the result of AND is compared against zero. Try to match
7332 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7333 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7334 if (NewSetCC.getNode()) {
7335 CC = NewSetCC.getOperand(0);
7336 Cond = NewSetCC.getOperand(1);
7343 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7344 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7346 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7347 Chain, Dest, CC, Cond);
7351 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
7352 // Calls to _alloca is needed to probe the stack when allocating more than 4k
7353 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
7354 // that the guard pages used by the OS virtual memory manager are allocated in
7355 // correct sequence.
7357 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
7358 SelectionDAG &DAG) const {
7359 assert(Subtarget->isTargetCygMing() &&
7360 "This should be used only on Cygwin/Mingw targets");
7361 DebugLoc dl = Op.getDebugLoc();
7364 SDValue Chain = Op.getOperand(0);
7365 SDValue Size = Op.getOperand(1);
7366 // FIXME: Ensure alignment here
7370 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
7372 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
7373 Flag = Chain.getValue(1);
7375 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
7377 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
7378 Flag = Chain.getValue(1);
7380 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
7382 SDValue Ops1[2] = { Chain.getValue(0), Chain };
7383 return DAG.getMergeValues(Ops1, 2, dl);
7386 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
7387 MachineFunction &MF = DAG.getMachineFunction();
7388 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
7390 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7391 DebugLoc dl = Op.getDebugLoc();
7393 if (!Subtarget->is64Bit()) {
7394 // vastart just stores the address of the VarArgsFrameIndex slot into the
7395 // memory location argument.
7396 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7398 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
7403 // gp_offset (0 - 6 * 8)
7404 // fp_offset (48 - 48 + 8 * 16)
7405 // overflow_arg_area (point to parameters coming in memory).
7407 SmallVector<SDValue, 8> MemOps;
7408 SDValue FIN = Op.getOperand(1);
7410 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
7411 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
7413 FIN, SV, 0, false, false, 0);
7414 MemOps.push_back(Store);
7417 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7418 FIN, DAG.getIntPtrConstant(4));
7419 Store = DAG.getStore(Op.getOperand(0), dl,
7420 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
7422 FIN, SV, 4, false, false, 0);
7423 MemOps.push_back(Store);
7425 // Store ptr to overflow_arg_area
7426 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7427 FIN, DAG.getIntPtrConstant(4));
7428 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7430 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 8,
7432 MemOps.push_back(Store);
7434 // Store ptr to reg_save_area.
7435 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7436 FIN, DAG.getIntPtrConstant(8));
7437 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
7439 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 16,
7441 MemOps.push_back(Store);
7442 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
7443 &MemOps[0], MemOps.size());
7446 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
7447 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7448 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
7450 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
7454 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
7455 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7456 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
7457 SDValue Chain = Op.getOperand(0);
7458 SDValue DstPtr = Op.getOperand(1);
7459 SDValue SrcPtr = Op.getOperand(2);
7460 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
7461 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7462 DebugLoc dl = Op.getDebugLoc();
7464 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
7465 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
7466 false, DstSV, 0, SrcSV, 0);
7470 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
7471 DebugLoc dl = Op.getDebugLoc();
7472 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7474 default: return SDValue(); // Don't custom lower most intrinsics.
7475 // Comparison intrinsics.
7476 case Intrinsic::x86_sse_comieq_ss:
7477 case Intrinsic::x86_sse_comilt_ss:
7478 case Intrinsic::x86_sse_comile_ss:
7479 case Intrinsic::x86_sse_comigt_ss:
7480 case Intrinsic::x86_sse_comige_ss:
7481 case Intrinsic::x86_sse_comineq_ss:
7482 case Intrinsic::x86_sse_ucomieq_ss:
7483 case Intrinsic::x86_sse_ucomilt_ss:
7484 case Intrinsic::x86_sse_ucomile_ss:
7485 case Intrinsic::x86_sse_ucomigt_ss:
7486 case Intrinsic::x86_sse_ucomige_ss:
7487 case Intrinsic::x86_sse_ucomineq_ss:
7488 case Intrinsic::x86_sse2_comieq_sd:
7489 case Intrinsic::x86_sse2_comilt_sd:
7490 case Intrinsic::x86_sse2_comile_sd:
7491 case Intrinsic::x86_sse2_comigt_sd:
7492 case Intrinsic::x86_sse2_comige_sd:
7493 case Intrinsic::x86_sse2_comineq_sd:
7494 case Intrinsic::x86_sse2_ucomieq_sd:
7495 case Intrinsic::x86_sse2_ucomilt_sd:
7496 case Intrinsic::x86_sse2_ucomile_sd:
7497 case Intrinsic::x86_sse2_ucomigt_sd:
7498 case Intrinsic::x86_sse2_ucomige_sd:
7499 case Intrinsic::x86_sse2_ucomineq_sd: {
7501 ISD::CondCode CC = ISD::SETCC_INVALID;
7504 case Intrinsic::x86_sse_comieq_ss:
7505 case Intrinsic::x86_sse2_comieq_sd:
7509 case Intrinsic::x86_sse_comilt_ss:
7510 case Intrinsic::x86_sse2_comilt_sd:
7514 case Intrinsic::x86_sse_comile_ss:
7515 case Intrinsic::x86_sse2_comile_sd:
7519 case Intrinsic::x86_sse_comigt_ss:
7520 case Intrinsic::x86_sse2_comigt_sd:
7524 case Intrinsic::x86_sse_comige_ss:
7525 case Intrinsic::x86_sse2_comige_sd:
7529 case Intrinsic::x86_sse_comineq_ss:
7530 case Intrinsic::x86_sse2_comineq_sd:
7534 case Intrinsic::x86_sse_ucomieq_ss:
7535 case Intrinsic::x86_sse2_ucomieq_sd:
7536 Opc = X86ISD::UCOMI;
7539 case Intrinsic::x86_sse_ucomilt_ss:
7540 case Intrinsic::x86_sse2_ucomilt_sd:
7541 Opc = X86ISD::UCOMI;
7544 case Intrinsic::x86_sse_ucomile_ss:
7545 case Intrinsic::x86_sse2_ucomile_sd:
7546 Opc = X86ISD::UCOMI;
7549 case Intrinsic::x86_sse_ucomigt_ss:
7550 case Intrinsic::x86_sse2_ucomigt_sd:
7551 Opc = X86ISD::UCOMI;
7554 case Intrinsic::x86_sse_ucomige_ss:
7555 case Intrinsic::x86_sse2_ucomige_sd:
7556 Opc = X86ISD::UCOMI;
7559 case Intrinsic::x86_sse_ucomineq_ss:
7560 case Intrinsic::x86_sse2_ucomineq_sd:
7561 Opc = X86ISD::UCOMI;
7566 SDValue LHS = Op.getOperand(1);
7567 SDValue RHS = Op.getOperand(2);
7568 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
7569 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
7570 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
7571 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7572 DAG.getConstant(X86CC, MVT::i8), Cond);
7573 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7575 // ptest and testp intrinsics. The intrinsic these come from are designed to
7576 // return an integer value, not just an instruction so lower it to the ptest
7577 // or testp pattern and a setcc for the result.
7578 case Intrinsic::x86_sse41_ptestz:
7579 case Intrinsic::x86_sse41_ptestc:
7580 case Intrinsic::x86_sse41_ptestnzc:
7581 case Intrinsic::x86_avx_ptestz_256:
7582 case Intrinsic::x86_avx_ptestc_256:
7583 case Intrinsic::x86_avx_ptestnzc_256:
7584 case Intrinsic::x86_avx_vtestz_ps:
7585 case Intrinsic::x86_avx_vtestc_ps:
7586 case Intrinsic::x86_avx_vtestnzc_ps:
7587 case Intrinsic::x86_avx_vtestz_pd:
7588 case Intrinsic::x86_avx_vtestc_pd:
7589 case Intrinsic::x86_avx_vtestnzc_pd:
7590 case Intrinsic::x86_avx_vtestz_ps_256:
7591 case Intrinsic::x86_avx_vtestc_ps_256:
7592 case Intrinsic::x86_avx_vtestnzc_ps_256:
7593 case Intrinsic::x86_avx_vtestz_pd_256:
7594 case Intrinsic::x86_avx_vtestc_pd_256:
7595 case Intrinsic::x86_avx_vtestnzc_pd_256: {
7596 bool IsTestPacked = false;
7599 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
7600 case Intrinsic::x86_avx_vtestz_ps:
7601 case Intrinsic::x86_avx_vtestz_pd:
7602 case Intrinsic::x86_avx_vtestz_ps_256:
7603 case Intrinsic::x86_avx_vtestz_pd_256:
7604 IsTestPacked = true; // Fallthrough
7605 case Intrinsic::x86_sse41_ptestz:
7606 case Intrinsic::x86_avx_ptestz_256:
7608 X86CC = X86::COND_E;
7610 case Intrinsic::x86_avx_vtestc_ps:
7611 case Intrinsic::x86_avx_vtestc_pd:
7612 case Intrinsic::x86_avx_vtestc_ps_256:
7613 case Intrinsic::x86_avx_vtestc_pd_256:
7614 IsTestPacked = true; // Fallthrough
7615 case Intrinsic::x86_sse41_ptestc:
7616 case Intrinsic::x86_avx_ptestc_256:
7618 X86CC = X86::COND_B;
7620 case Intrinsic::x86_avx_vtestnzc_ps:
7621 case Intrinsic::x86_avx_vtestnzc_pd:
7622 case Intrinsic::x86_avx_vtestnzc_ps_256:
7623 case Intrinsic::x86_avx_vtestnzc_pd_256:
7624 IsTestPacked = true; // Fallthrough
7625 case Intrinsic::x86_sse41_ptestnzc:
7626 case Intrinsic::x86_avx_ptestnzc_256:
7628 X86CC = X86::COND_A;
7632 SDValue LHS = Op.getOperand(1);
7633 SDValue RHS = Op.getOperand(2);
7634 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
7635 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
7636 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7637 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7638 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7641 // Fix vector shift instructions where the last operand is a non-immediate
7643 case Intrinsic::x86_sse2_pslli_w:
7644 case Intrinsic::x86_sse2_pslli_d:
7645 case Intrinsic::x86_sse2_pslli_q:
7646 case Intrinsic::x86_sse2_psrli_w:
7647 case Intrinsic::x86_sse2_psrli_d:
7648 case Intrinsic::x86_sse2_psrli_q:
7649 case Intrinsic::x86_sse2_psrai_w:
7650 case Intrinsic::x86_sse2_psrai_d:
7651 case Intrinsic::x86_mmx_pslli_w:
7652 case Intrinsic::x86_mmx_pslli_d:
7653 case Intrinsic::x86_mmx_pslli_q:
7654 case Intrinsic::x86_mmx_psrli_w:
7655 case Intrinsic::x86_mmx_psrli_d:
7656 case Intrinsic::x86_mmx_psrli_q:
7657 case Intrinsic::x86_mmx_psrai_w:
7658 case Intrinsic::x86_mmx_psrai_d: {
7659 SDValue ShAmt = Op.getOperand(2);
7660 if (isa<ConstantSDNode>(ShAmt))
7663 unsigned NewIntNo = 0;
7664 EVT ShAmtVT = MVT::v4i32;
7666 case Intrinsic::x86_sse2_pslli_w:
7667 NewIntNo = Intrinsic::x86_sse2_psll_w;
7669 case Intrinsic::x86_sse2_pslli_d:
7670 NewIntNo = Intrinsic::x86_sse2_psll_d;
7672 case Intrinsic::x86_sse2_pslli_q:
7673 NewIntNo = Intrinsic::x86_sse2_psll_q;
7675 case Intrinsic::x86_sse2_psrli_w:
7676 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7678 case Intrinsic::x86_sse2_psrli_d:
7679 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7681 case Intrinsic::x86_sse2_psrli_q:
7682 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7684 case Intrinsic::x86_sse2_psrai_w:
7685 NewIntNo = Intrinsic::x86_sse2_psra_w;
7687 case Intrinsic::x86_sse2_psrai_d:
7688 NewIntNo = Intrinsic::x86_sse2_psra_d;
7691 ShAmtVT = MVT::v2i32;
7693 case Intrinsic::x86_mmx_pslli_w:
7694 NewIntNo = Intrinsic::x86_mmx_psll_w;
7696 case Intrinsic::x86_mmx_pslli_d:
7697 NewIntNo = Intrinsic::x86_mmx_psll_d;
7699 case Intrinsic::x86_mmx_pslli_q:
7700 NewIntNo = Intrinsic::x86_mmx_psll_q;
7702 case Intrinsic::x86_mmx_psrli_w:
7703 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7705 case Intrinsic::x86_mmx_psrli_d:
7706 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7708 case Intrinsic::x86_mmx_psrli_q:
7709 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7711 case Intrinsic::x86_mmx_psrai_w:
7712 NewIntNo = Intrinsic::x86_mmx_psra_w;
7714 case Intrinsic::x86_mmx_psrai_d:
7715 NewIntNo = Intrinsic::x86_mmx_psra_d;
7717 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7723 // The vector shift intrinsics with scalars uses 32b shift amounts but
7724 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7728 ShOps[1] = DAG.getConstant(0, MVT::i32);
7729 if (ShAmtVT == MVT::v4i32) {
7730 ShOps[2] = DAG.getUNDEF(MVT::i32);
7731 ShOps[3] = DAG.getUNDEF(MVT::i32);
7732 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7734 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7737 EVT VT = Op.getValueType();
7738 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7739 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7740 DAG.getConstant(NewIntNo, MVT::i32),
7741 Op.getOperand(1), ShAmt);
7746 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7747 SelectionDAG &DAG) const {
7748 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7749 MFI->setReturnAddressIsTaken(true);
7751 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7752 DebugLoc dl = Op.getDebugLoc();
7755 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7757 DAG.getConstant(TD->getPointerSize(),
7758 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7759 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7760 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7762 NULL, 0, false, false, 0);
7765 // Just load the return address.
7766 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7767 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7768 RetAddrFI, NULL, 0, false, false, 0);
7771 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7772 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7773 MFI->setFrameAddressIsTaken(true);
7775 EVT VT = Op.getValueType();
7776 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7777 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7778 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7779 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7781 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7786 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7787 SelectionDAG &DAG) const {
7788 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7791 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7792 MachineFunction &MF = DAG.getMachineFunction();
7793 SDValue Chain = Op.getOperand(0);
7794 SDValue Offset = Op.getOperand(1);
7795 SDValue Handler = Op.getOperand(2);
7796 DebugLoc dl = Op.getDebugLoc();
7798 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
7799 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7801 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7803 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
7804 DAG.getIntPtrConstant(TD->getPointerSize()));
7805 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7806 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7807 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7808 MF.getRegInfo().addLiveOut(StoreAddrReg);
7810 return DAG.getNode(X86ISD::EH_RETURN, dl,
7812 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7815 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7816 SelectionDAG &DAG) const {
7817 SDValue Root = Op.getOperand(0);
7818 SDValue Trmp = Op.getOperand(1); // trampoline
7819 SDValue FPtr = Op.getOperand(2); // nested function
7820 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7821 DebugLoc dl = Op.getDebugLoc();
7823 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7825 if (Subtarget->is64Bit()) {
7826 SDValue OutChains[6];
7828 // Large code-model.
7829 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7830 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7832 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7833 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7835 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7837 // Load the pointer to the nested function into R11.
7838 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7839 SDValue Addr = Trmp;
7840 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7841 Addr, TrmpAddr, 0, false, false, 0);
7843 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7844 DAG.getConstant(2, MVT::i64));
7845 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7848 // Load the 'nest' parameter value into R10.
7849 // R10 is specified in X86CallingConv.td
7850 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7851 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7852 DAG.getConstant(10, MVT::i64));
7853 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7854 Addr, TrmpAddr, 10, false, false, 0);
7856 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7857 DAG.getConstant(12, MVT::i64));
7858 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7861 // Jump to the nested function.
7862 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7863 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7864 DAG.getConstant(20, MVT::i64));
7865 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7866 Addr, TrmpAddr, 20, false, false, 0);
7868 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7869 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7870 DAG.getConstant(22, MVT::i64));
7871 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7872 TrmpAddr, 22, false, false, 0);
7875 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7876 return DAG.getMergeValues(Ops, 2, dl);
7878 const Function *Func =
7879 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7880 CallingConv::ID CC = Func->getCallingConv();
7885 llvm_unreachable("Unsupported calling convention");
7886 case CallingConv::C:
7887 case CallingConv::X86_StdCall: {
7888 // Pass 'nest' parameter in ECX.
7889 // Must be kept in sync with X86CallingConv.td
7892 // Check that ECX wasn't needed by an 'inreg' parameter.
7893 const FunctionType *FTy = Func->getFunctionType();
7894 const AttrListPtr &Attrs = Func->getAttributes();
7896 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7897 unsigned InRegCount = 0;
7900 for (FunctionType::param_iterator I = FTy->param_begin(),
7901 E = FTy->param_end(); I != E; ++I, ++Idx)
7902 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7903 // FIXME: should only count parameters that are lowered to integers.
7904 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7906 if (InRegCount > 2) {
7907 report_fatal_error("Nest register in use - reduce number of inreg"
7913 case CallingConv::X86_FastCall:
7914 case CallingConv::X86_ThisCall:
7915 case CallingConv::Fast:
7916 // Pass 'nest' parameter in EAX.
7917 // Must be kept in sync with X86CallingConv.td
7922 SDValue OutChains[4];
7925 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7926 DAG.getConstant(10, MVT::i32));
7927 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7929 // This is storing the opcode for MOV32ri.
7930 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7931 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7932 OutChains[0] = DAG.getStore(Root, dl,
7933 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7934 Trmp, TrmpAddr, 0, false, false, 0);
7936 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7937 DAG.getConstant(1, MVT::i32));
7938 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7941 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7942 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7943 DAG.getConstant(5, MVT::i32));
7944 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7945 TrmpAddr, 5, false, false, 1);
7947 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7948 DAG.getConstant(6, MVT::i32));
7949 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7953 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7954 return DAG.getMergeValues(Ops, 2, dl);
7958 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7959 SelectionDAG &DAG) const {
7961 The rounding mode is in bits 11:10 of FPSR, and has the following
7968 FLT_ROUNDS, on the other hand, expects the following:
7975 To perform the conversion, we do:
7976 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7979 MachineFunction &MF = DAG.getMachineFunction();
7980 const TargetMachine &TM = MF.getTarget();
7981 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7982 unsigned StackAlignment = TFI.getStackAlignment();
7983 EVT VT = Op.getValueType();
7984 DebugLoc dl = Op.getDebugLoc();
7986 // Save FP Control Word to stack slot
7987 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7988 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7990 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7991 DAG.getEntryNode(), StackSlot);
7993 // Load FP Control Word from stack slot
7994 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7997 // Transform as necessary
7999 DAG.getNode(ISD::SRL, dl, MVT::i16,
8000 DAG.getNode(ISD::AND, dl, MVT::i16,
8001 CWD, DAG.getConstant(0x800, MVT::i16)),
8002 DAG.getConstant(11, MVT::i8));
8004 DAG.getNode(ISD::SRL, dl, MVT::i16,
8005 DAG.getNode(ISD::AND, dl, MVT::i16,
8006 CWD, DAG.getConstant(0x400, MVT::i16)),
8007 DAG.getConstant(9, MVT::i8));
8010 DAG.getNode(ISD::AND, dl, MVT::i16,
8011 DAG.getNode(ISD::ADD, dl, MVT::i16,
8012 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
8013 DAG.getConstant(1, MVT::i16)),
8014 DAG.getConstant(3, MVT::i16));
8017 return DAG.getNode((VT.getSizeInBits() < 16 ?
8018 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
8021 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
8022 EVT VT = Op.getValueType();
8024 unsigned NumBits = VT.getSizeInBits();
8025 DebugLoc dl = Op.getDebugLoc();
8027 Op = Op.getOperand(0);
8028 if (VT == MVT::i8) {
8029 // Zero extend to i32 since there is not an i8 bsr.
8031 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8034 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
8035 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8036 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
8038 // If src is zero (i.e. bsr sets ZF), returns NumBits.
8041 DAG.getConstant(NumBits+NumBits-1, OpVT),
8042 DAG.getConstant(X86::COND_E, MVT::i8),
8045 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8047 // Finally xor with NumBits-1.
8048 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
8051 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8055 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
8056 EVT VT = Op.getValueType();
8058 unsigned NumBits = VT.getSizeInBits();
8059 DebugLoc dl = Op.getDebugLoc();
8061 Op = Op.getOperand(0);
8062 if (VT == MVT::i8) {
8064 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8067 // Issue a bsf (scan bits forward) which also sets EFLAGS.
8068 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8069 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
8071 // If src is zero (i.e. bsf sets ZF), returns NumBits.
8074 DAG.getConstant(NumBits, OpVT),
8075 DAG.getConstant(X86::COND_E, MVT::i8),
8078 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8081 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8085 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
8086 EVT VT = Op.getValueType();
8087 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
8088 DebugLoc dl = Op.getDebugLoc();
8090 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
8091 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
8092 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
8093 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
8094 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
8096 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
8097 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
8098 // return AloBlo + AloBhi + AhiBlo;
8100 SDValue A = Op.getOperand(0);
8101 SDValue B = Op.getOperand(1);
8103 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8104 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8105 A, DAG.getConstant(32, MVT::i32));
8106 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8107 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8108 B, DAG.getConstant(32, MVT::i32));
8109 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8110 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8112 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8113 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8115 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8116 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8118 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8119 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8120 AloBhi, DAG.getConstant(32, MVT::i32));
8121 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8122 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8123 AhiBlo, DAG.getConstant(32, MVT::i32));
8124 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8125 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8129 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
8130 EVT VT = Op.getValueType();
8131 DebugLoc dl = Op.getDebugLoc();
8132 SDValue R = Op.getOperand(0);
8134 LLVMContext *Context = DAG.getContext();
8136 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
8138 if (VT == MVT::v4i32) {
8139 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8140 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8141 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
8143 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
8145 std::vector<Constant*> CV(4, CI);
8146 Constant *C = ConstantVector::get(CV);
8147 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8148 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8149 PseudoSourceValue::getConstantPool(), 0,
8152 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
8153 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, Op);
8154 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
8155 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
8157 if (VT == MVT::v16i8) {
8159 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8160 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8161 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
8163 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
8164 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
8166 std::vector<Constant*> CVM1(16, CM1);
8167 std::vector<Constant*> CVM2(16, CM2);
8168 Constant *C = ConstantVector::get(CVM1);
8169 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8170 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8171 PseudoSourceValue::getConstantPool(), 0,
8174 // r = pblendv(r, psllw(r & (char16)15, 4), a);
8175 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8176 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8177 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8178 DAG.getConstant(4, MVT::i32));
8179 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8180 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8183 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8185 C = ConstantVector::get(CVM2);
8186 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8187 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8188 PseudoSourceValue::getConstantPool(), 0, false, false, 16);
8190 // r = pblendv(r, psllw(r & (char16)63, 2), a);
8191 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8192 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8193 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8194 DAG.getConstant(2, MVT::i32));
8195 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8196 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8199 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8201 // return pblendv(r, r+r, a);
8202 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8203 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8204 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
8210 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
8211 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
8212 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
8213 // looks for this combo and may remove the "setcc" instruction if the "setcc"
8214 // has only one use.
8215 SDNode *N = Op.getNode();
8216 SDValue LHS = N->getOperand(0);
8217 SDValue RHS = N->getOperand(1);
8218 unsigned BaseOp = 0;
8220 DebugLoc dl = Op.getDebugLoc();
8222 switch (Op.getOpcode()) {
8223 default: llvm_unreachable("Unknown ovf instruction!");
8225 // A subtract of one will be selected as a INC. Note that INC doesn't
8226 // set CF, so we can't do this for UADDO.
8227 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8228 if (C->getAPIntValue() == 1) {
8229 BaseOp = X86ISD::INC;
8233 BaseOp = X86ISD::ADD;
8237 BaseOp = X86ISD::ADD;
8241 // A subtract of one will be selected as a DEC. Note that DEC doesn't
8242 // set CF, so we can't do this for USUBO.
8243 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8244 if (C->getAPIntValue() == 1) {
8245 BaseOp = X86ISD::DEC;
8249 BaseOp = X86ISD::SUB;
8253 BaseOp = X86ISD::SUB;
8257 BaseOp = X86ISD::SMUL;
8261 BaseOp = X86ISD::UMUL;
8266 // Also sets EFLAGS.
8267 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
8268 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
8271 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
8272 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
8274 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8278 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
8279 DebugLoc dl = Op.getDebugLoc();
8281 if (!Subtarget->hasSSE2()) {
8282 SDValue Chain = Op.getOperand(0);
8283 SDValue Zero = DAG.getConstant(0,
8284 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8286 DAG.getRegister(X86::ESP, MVT::i32), // Base
8287 DAG.getTargetConstant(1, MVT::i8), // Scale
8288 DAG.getRegister(0, MVT::i32), // Index
8289 DAG.getTargetConstant(0, MVT::i32), // Disp
8290 DAG.getRegister(0, MVT::i32), // Segment.
8295 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
8296 array_lengthof(Ops));
8297 return SDValue(Res, 0);
8300 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
8302 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
8304 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
8305 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
8306 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
8307 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
8309 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
8310 if (!Op1 && !Op2 && !Op3 && Op4)
8311 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
8313 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
8314 if (Op1 && !Op2 && !Op3 && !Op4)
8315 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
8317 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
8319 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
8322 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
8323 EVT T = Op.getValueType();
8324 DebugLoc dl = Op.getDebugLoc();
8327 switch(T.getSimpleVT().SimpleTy) {
8329 assert(false && "Invalid value type!");
8330 case MVT::i8: Reg = X86::AL; size = 1; break;
8331 case MVT::i16: Reg = X86::AX; size = 2; break;
8332 case MVT::i32: Reg = X86::EAX; size = 4; break;
8334 assert(Subtarget->is64Bit() && "Node not type legal!");
8335 Reg = X86::RAX; size = 8;
8338 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
8339 Op.getOperand(2), SDValue());
8340 SDValue Ops[] = { cpIn.getValue(0),
8343 DAG.getTargetConstant(size, MVT::i8),
8345 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8346 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
8348 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
8352 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
8353 SelectionDAG &DAG) const {
8354 assert(Subtarget->is64Bit() && "Result not type legalized?");
8355 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8356 SDValue TheChain = Op.getOperand(0);
8357 DebugLoc dl = Op.getDebugLoc();
8358 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8359 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
8360 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
8362 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
8363 DAG.getConstant(32, MVT::i8));
8365 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
8368 return DAG.getMergeValues(Ops, 2, dl);
8371 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
8372 SelectionDAG &DAG) const {
8373 EVT SrcVT = Op.getOperand(0).getValueType();
8374 EVT DstVT = Op.getValueType();
8375 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
8376 Subtarget->hasMMX() && !DisableMMX) &&
8377 "Unexpected custom BIT_CONVERT");
8378 assert((DstVT == MVT::i64 ||
8379 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
8380 "Unexpected custom BIT_CONVERT");
8381 // i64 <=> MMX conversions are Legal.
8382 if (SrcVT==MVT::i64 && DstVT.isVector())
8384 if (DstVT==MVT::i64 && SrcVT.isVector())
8386 // MMX <=> MMX conversions are Legal.
8387 if (SrcVT.isVector() && DstVT.isVector())
8389 // All other conversions need to be expanded.
8392 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
8393 SDNode *Node = Op.getNode();
8394 DebugLoc dl = Node->getDebugLoc();
8395 EVT T = Node->getValueType(0);
8396 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
8397 DAG.getConstant(0, T), Node->getOperand(2));
8398 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
8399 cast<AtomicSDNode>(Node)->getMemoryVT(),
8400 Node->getOperand(0),
8401 Node->getOperand(1), negOp,
8402 cast<AtomicSDNode>(Node)->getSrcValue(),
8403 cast<AtomicSDNode>(Node)->getAlignment());
8406 /// LowerOperation - Provide custom lowering hooks for some operations.
8408 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
8409 switch (Op.getOpcode()) {
8410 default: llvm_unreachable("Should not custom lower this!");
8411 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
8412 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
8413 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
8414 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
8415 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
8416 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
8417 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
8418 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
8419 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
8420 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
8421 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
8422 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
8423 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
8424 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
8425 case ISD::SHL_PARTS:
8426 case ISD::SRA_PARTS:
8427 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
8428 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
8429 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
8430 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
8431 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
8432 case ISD::FABS: return LowerFABS(Op, DAG);
8433 case ISD::FNEG: return LowerFNEG(Op, DAG);
8434 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
8435 case ISD::SETCC: return LowerSETCC(Op, DAG);
8436 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
8437 case ISD::SELECT: return LowerSELECT(Op, DAG);
8438 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
8439 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
8440 case ISD::VASTART: return LowerVASTART(Op, DAG);
8441 case ISD::VAARG: return LowerVAARG(Op, DAG);
8442 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
8443 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
8444 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
8445 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
8446 case ISD::FRAME_TO_ARGS_OFFSET:
8447 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
8448 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
8449 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
8450 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
8451 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
8452 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
8453 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
8454 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
8455 case ISD::SHL: return LowerSHL(Op, DAG);
8461 case ISD::UMULO: return LowerXALUO(Op, DAG);
8462 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
8463 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
8467 void X86TargetLowering::
8468 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
8469 SelectionDAG &DAG, unsigned NewOp) const {
8470 EVT T = Node->getValueType(0);
8471 DebugLoc dl = Node->getDebugLoc();
8472 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
8474 SDValue Chain = Node->getOperand(0);
8475 SDValue In1 = Node->getOperand(1);
8476 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8477 Node->getOperand(2), DAG.getIntPtrConstant(0));
8478 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8479 Node->getOperand(2), DAG.getIntPtrConstant(1));
8480 SDValue Ops[] = { Chain, In1, In2L, In2H };
8481 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
8483 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
8484 cast<MemSDNode>(Node)->getMemOperand());
8485 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
8486 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8487 Results.push_back(Result.getValue(2));
8490 /// ReplaceNodeResults - Replace a node with an illegal result type
8491 /// with a new node built out of custom code.
8492 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
8493 SmallVectorImpl<SDValue>&Results,
8494 SelectionDAG &DAG) const {
8495 DebugLoc dl = N->getDebugLoc();
8496 switch (N->getOpcode()) {
8498 assert(false && "Do not know how to custom type legalize this operation!");
8500 case ISD::FP_TO_SINT: {
8501 std::pair<SDValue,SDValue> Vals =
8502 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
8503 SDValue FIST = Vals.first, StackSlot = Vals.second;
8504 if (FIST.getNode() != 0) {
8505 EVT VT = N->getValueType(0);
8506 // Return a load from the stack slot.
8507 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
8512 case ISD::READCYCLECOUNTER: {
8513 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8514 SDValue TheChain = N->getOperand(0);
8515 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8516 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
8518 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
8520 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
8521 SDValue Ops[] = { eax, edx };
8522 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
8523 Results.push_back(edx.getValue(1));
8526 case ISD::ATOMIC_CMP_SWAP: {
8527 EVT T = N->getValueType(0);
8528 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
8529 SDValue cpInL, cpInH;
8530 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8531 DAG.getConstant(0, MVT::i32));
8532 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8533 DAG.getConstant(1, MVT::i32));
8534 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
8535 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
8537 SDValue swapInL, swapInH;
8538 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8539 DAG.getConstant(0, MVT::i32));
8540 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8541 DAG.getConstant(1, MVT::i32));
8542 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
8544 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
8545 swapInL.getValue(1));
8546 SDValue Ops[] = { swapInH.getValue(0),
8548 swapInH.getValue(1) };
8549 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8550 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
8551 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
8552 MVT::i32, Result.getValue(1));
8553 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
8554 MVT::i32, cpOutL.getValue(2));
8555 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
8556 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8557 Results.push_back(cpOutH.getValue(1));
8560 case ISD::ATOMIC_LOAD_ADD:
8561 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
8563 case ISD::ATOMIC_LOAD_AND:
8564 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
8566 case ISD::ATOMIC_LOAD_NAND:
8567 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
8569 case ISD::ATOMIC_LOAD_OR:
8570 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
8572 case ISD::ATOMIC_LOAD_SUB:
8573 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
8575 case ISD::ATOMIC_LOAD_XOR:
8576 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
8578 case ISD::ATOMIC_SWAP:
8579 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
8584 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
8586 default: return NULL;
8587 case X86ISD::BSF: return "X86ISD::BSF";
8588 case X86ISD::BSR: return "X86ISD::BSR";
8589 case X86ISD::SHLD: return "X86ISD::SHLD";
8590 case X86ISD::SHRD: return "X86ISD::SHRD";
8591 case X86ISD::FAND: return "X86ISD::FAND";
8592 case X86ISD::FOR: return "X86ISD::FOR";
8593 case X86ISD::FXOR: return "X86ISD::FXOR";
8594 case X86ISD::FSRL: return "X86ISD::FSRL";
8595 case X86ISD::FILD: return "X86ISD::FILD";
8596 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
8597 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
8598 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
8599 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
8600 case X86ISD::FLD: return "X86ISD::FLD";
8601 case X86ISD::FST: return "X86ISD::FST";
8602 case X86ISD::CALL: return "X86ISD::CALL";
8603 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
8604 case X86ISD::BT: return "X86ISD::BT";
8605 case X86ISD::CMP: return "X86ISD::CMP";
8606 case X86ISD::COMI: return "X86ISD::COMI";
8607 case X86ISD::UCOMI: return "X86ISD::UCOMI";
8608 case X86ISD::SETCC: return "X86ISD::SETCC";
8609 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
8610 case X86ISD::CMOV: return "X86ISD::CMOV";
8611 case X86ISD::BRCOND: return "X86ISD::BRCOND";
8612 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
8613 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
8614 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
8615 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
8616 case X86ISD::Wrapper: return "X86ISD::Wrapper";
8617 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
8618 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
8619 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
8620 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
8621 case X86ISD::PINSRB: return "X86ISD::PINSRB";
8622 case X86ISD::PINSRW: return "X86ISD::PINSRW";
8623 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
8624 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
8625 case X86ISD::FMAX: return "X86ISD::FMAX";
8626 case X86ISD::FMIN: return "X86ISD::FMIN";
8627 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
8628 case X86ISD::FRCP: return "X86ISD::FRCP";
8629 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
8630 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
8631 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
8632 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
8633 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
8634 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
8635 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
8636 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
8637 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
8638 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
8639 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
8640 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
8641 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
8642 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
8643 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
8644 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
8645 case X86ISD::VSHL: return "X86ISD::VSHL";
8646 case X86ISD::VSRL: return "X86ISD::VSRL";
8647 case X86ISD::CMPPD: return "X86ISD::CMPPD";
8648 case X86ISD::CMPPS: return "X86ISD::CMPPS";
8649 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
8650 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
8651 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
8652 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
8653 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
8654 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
8655 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
8656 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
8657 case X86ISD::ADD: return "X86ISD::ADD";
8658 case X86ISD::SUB: return "X86ISD::SUB";
8659 case X86ISD::SMUL: return "X86ISD::SMUL";
8660 case X86ISD::UMUL: return "X86ISD::UMUL";
8661 case X86ISD::INC: return "X86ISD::INC";
8662 case X86ISD::DEC: return "X86ISD::DEC";
8663 case X86ISD::OR: return "X86ISD::OR";
8664 case X86ISD::XOR: return "X86ISD::XOR";
8665 case X86ISD::AND: return "X86ISD::AND";
8666 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
8667 case X86ISD::PTEST: return "X86ISD::PTEST";
8668 case X86ISD::TESTP: return "X86ISD::TESTP";
8669 case X86ISD::PALIGN: return "X86ISD::PALIGN";
8670 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
8671 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
8672 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
8673 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
8674 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
8675 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
8676 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
8677 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
8678 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
8679 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
8680 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
8681 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
8682 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
8683 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
8684 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
8685 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
8686 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
8687 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
8688 case X86ISD::MOVSD: return "X86ISD::MOVSD";
8689 case X86ISD::MOVSS: return "X86ISD::MOVSS";
8690 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
8691 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
8692 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
8693 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
8694 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
8695 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
8696 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
8697 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
8698 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
8699 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
8700 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
8701 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
8702 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
8703 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
8707 // isLegalAddressingMode - Return true if the addressing mode represented
8708 // by AM is legal for this target, for a load/store of the specified type.
8709 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
8710 const Type *Ty) const {
8711 // X86 supports extremely general addressing modes.
8712 CodeModel::Model M = getTargetMachine().getCodeModel();
8713 Reloc::Model R = getTargetMachine().getRelocationModel();
8715 // X86 allows a sign-extended 32-bit immediate field as a displacement.
8716 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
8721 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
8723 // If a reference to this global requires an extra load, we can't fold it.
8724 if (isGlobalStubReference(GVFlags))
8727 // If BaseGV requires a register for the PIC base, we cannot also have a
8728 // BaseReg specified.
8729 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
8732 // If lower 4G is not available, then we must use rip-relative addressing.
8733 if ((M != CodeModel::Small || R != Reloc::Static) &&
8734 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
8744 // These scales always work.
8749 // These scales are formed with basereg+scalereg. Only accept if there is
8754 default: // Other stuff never works.
8762 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
8763 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
8765 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
8766 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
8767 if (NumBits1 <= NumBits2)
8772 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
8773 if (!VT1.isInteger() || !VT2.isInteger())
8775 unsigned NumBits1 = VT1.getSizeInBits();
8776 unsigned NumBits2 = VT2.getSizeInBits();
8777 if (NumBits1 <= NumBits2)
8782 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
8783 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8784 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
8787 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
8788 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8789 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
8792 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
8793 // i16 instructions are longer (0x66 prefix) and potentially slower.
8794 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
8797 /// isShuffleMaskLegal - Targets can use this to indicate that they only
8798 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
8799 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
8800 /// are assumed to be legal.
8802 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
8804 // Very little shuffling can be done for 64-bit vectors right now.
8805 if (VT.getSizeInBits() == 64)
8806 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8808 // FIXME: pshufb, blends, shifts.
8809 return (VT.getVectorNumElements() == 2 ||
8810 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8811 isMOVLMask(M, VT) ||
8812 isSHUFPMask(M, VT) ||
8813 isPSHUFDMask(M, VT) ||
8814 isPSHUFHWMask(M, VT) ||
8815 isPSHUFLWMask(M, VT) ||
8816 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8817 isUNPCKLMask(M, VT) ||
8818 isUNPCKHMask(M, VT) ||
8819 isUNPCKL_v_undef_Mask(M, VT) ||
8820 isUNPCKH_v_undef_Mask(M, VT));
8824 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8826 unsigned NumElts = VT.getVectorNumElements();
8827 // FIXME: This collection of masks seems suspect.
8830 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8831 return (isMOVLMask(Mask, VT) ||
8832 isCommutedMOVLMask(Mask, VT, true) ||
8833 isSHUFPMask(Mask, VT) ||
8834 isCommutedSHUFPMask(Mask, VT));
8839 //===----------------------------------------------------------------------===//
8840 // X86 Scheduler Hooks
8841 //===----------------------------------------------------------------------===//
8843 // private utility function
8845 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8846 MachineBasicBlock *MBB,
8853 TargetRegisterClass *RC,
8854 bool invSrc) const {
8855 // For the atomic bitwise operator, we generate
8858 // ld t1 = [bitinstr.addr]
8859 // op t2 = t1, [bitinstr.val]
8861 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8863 // fallthrough -->nextMBB
8864 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8865 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8866 MachineFunction::iterator MBBIter = MBB;
8869 /// First build the CFG
8870 MachineFunction *F = MBB->getParent();
8871 MachineBasicBlock *thisMBB = MBB;
8872 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8873 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8874 F->insert(MBBIter, newMBB);
8875 F->insert(MBBIter, nextMBB);
8877 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8878 nextMBB->splice(nextMBB->begin(), thisMBB,
8879 llvm::next(MachineBasicBlock::iterator(bInstr)),
8881 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8883 // Update thisMBB to fall through to newMBB
8884 thisMBB->addSuccessor(newMBB);
8886 // newMBB jumps to itself and fall through to nextMBB
8887 newMBB->addSuccessor(nextMBB);
8888 newMBB->addSuccessor(newMBB);
8890 // Insert instructions into newMBB based on incoming instruction
8891 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8892 "unexpected number of operands");
8893 DebugLoc dl = bInstr->getDebugLoc();
8894 MachineOperand& destOper = bInstr->getOperand(0);
8895 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8896 int numArgs = bInstr->getNumOperands() - 1;
8897 for (int i=0; i < numArgs; ++i)
8898 argOpers[i] = &bInstr->getOperand(i+1);
8900 // x86 address has 4 operands: base, index, scale, and displacement
8901 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8902 int valArgIndx = lastAddrIndx + 1;
8904 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8905 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8906 for (int i=0; i <= lastAddrIndx; ++i)
8907 (*MIB).addOperand(*argOpers[i]);
8909 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8911 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8916 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8917 assert((argOpers[valArgIndx]->isReg() ||
8918 argOpers[valArgIndx]->isImm()) &&
8920 if (argOpers[valArgIndx]->isReg())
8921 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8923 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8925 (*MIB).addOperand(*argOpers[valArgIndx]);
8927 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
8930 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8931 for (int i=0; i <= lastAddrIndx; ++i)
8932 (*MIB).addOperand(*argOpers[i]);
8934 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8935 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8936 bInstr->memoperands_end());
8938 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8942 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8944 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8948 // private utility function: 64 bit atomics on 32 bit host.
8950 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8951 MachineBasicBlock *MBB,
8956 bool invSrc) const {
8957 // For the atomic bitwise operator, we generate
8958 // thisMBB (instructions are in pairs, except cmpxchg8b)
8959 // ld t1,t2 = [bitinstr.addr]
8961 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8962 // op t5, t6 <- out1, out2, [bitinstr.val]
8963 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8964 // mov ECX, EBX <- t5, t6
8965 // mov EAX, EDX <- t1, t2
8966 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8967 // mov t3, t4 <- EAX, EDX
8969 // result in out1, out2
8970 // fallthrough -->nextMBB
8972 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8973 const unsigned LoadOpc = X86::MOV32rm;
8974 const unsigned NotOpc = X86::NOT32r;
8975 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8976 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8977 MachineFunction::iterator MBBIter = MBB;
8980 /// First build the CFG
8981 MachineFunction *F = MBB->getParent();
8982 MachineBasicBlock *thisMBB = MBB;
8983 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8984 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8985 F->insert(MBBIter, newMBB);
8986 F->insert(MBBIter, nextMBB);
8988 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8989 nextMBB->splice(nextMBB->begin(), thisMBB,
8990 llvm::next(MachineBasicBlock::iterator(bInstr)),
8992 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8994 // Update thisMBB to fall through to newMBB
8995 thisMBB->addSuccessor(newMBB);
8997 // newMBB jumps to itself and fall through to nextMBB
8998 newMBB->addSuccessor(nextMBB);
8999 newMBB->addSuccessor(newMBB);
9001 DebugLoc dl = bInstr->getDebugLoc();
9002 // Insert instructions into newMBB based on incoming instruction
9003 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
9004 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
9005 "unexpected number of operands");
9006 MachineOperand& dest1Oper = bInstr->getOperand(0);
9007 MachineOperand& dest2Oper = bInstr->getOperand(1);
9008 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9009 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
9010 argOpers[i] = &bInstr->getOperand(i+2);
9012 // We use some of the operands multiple times, so conservatively just
9013 // clear any kill flags that might be present.
9014 if (argOpers[i]->isReg() && argOpers[i]->isUse())
9015 argOpers[i]->setIsKill(false);
9018 // x86 address has 5 operands: base, index, scale, displacement, and segment.
9019 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9021 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9022 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
9023 for (int i=0; i <= lastAddrIndx; ++i)
9024 (*MIB).addOperand(*argOpers[i]);
9025 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9026 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
9027 // add 4 to displacement.
9028 for (int i=0; i <= lastAddrIndx-2; ++i)
9029 (*MIB).addOperand(*argOpers[i]);
9030 MachineOperand newOp3 = *(argOpers[3]);
9032 newOp3.setImm(newOp3.getImm()+4);
9034 newOp3.setOffset(newOp3.getOffset()+4);
9035 (*MIB).addOperand(newOp3);
9036 (*MIB).addOperand(*argOpers[lastAddrIndx]);
9038 // t3/4 are defined later, at the bottom of the loop
9039 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
9040 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
9041 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
9042 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
9043 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
9044 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
9046 // The subsequent operations should be using the destination registers of
9047 //the PHI instructions.
9049 t1 = F->getRegInfo().createVirtualRegister(RC);
9050 t2 = F->getRegInfo().createVirtualRegister(RC);
9051 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
9052 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
9054 t1 = dest1Oper.getReg();
9055 t2 = dest2Oper.getReg();
9058 int valArgIndx = lastAddrIndx + 1;
9059 assert((argOpers[valArgIndx]->isReg() ||
9060 argOpers[valArgIndx]->isImm()) &&
9062 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
9063 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
9064 if (argOpers[valArgIndx]->isReg())
9065 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
9067 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
9068 if (regOpcL != X86::MOV32rr)
9070 (*MIB).addOperand(*argOpers[valArgIndx]);
9071 assert(argOpers[valArgIndx + 1]->isReg() ==
9072 argOpers[valArgIndx]->isReg());
9073 assert(argOpers[valArgIndx + 1]->isImm() ==
9074 argOpers[valArgIndx]->isImm());
9075 if (argOpers[valArgIndx + 1]->isReg())
9076 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
9078 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
9079 if (regOpcH != X86::MOV32rr)
9081 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
9083 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9085 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
9088 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
9090 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
9093 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
9094 for (int i=0; i <= lastAddrIndx; ++i)
9095 (*MIB).addOperand(*argOpers[i]);
9097 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9098 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9099 bInstr->memoperands_end());
9101 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
9102 MIB.addReg(X86::EAX);
9103 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
9104 MIB.addReg(X86::EDX);
9107 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9109 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9113 // private utility function
9115 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
9116 MachineBasicBlock *MBB,
9117 unsigned cmovOpc) const {
9118 // For the atomic min/max operator, we generate
9121 // ld t1 = [min/max.addr]
9122 // mov t2 = [min/max.val]
9124 // cmov[cond] t2 = t1
9126 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9128 // fallthrough -->nextMBB
9130 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9131 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9132 MachineFunction::iterator MBBIter = MBB;
9135 /// First build the CFG
9136 MachineFunction *F = MBB->getParent();
9137 MachineBasicBlock *thisMBB = MBB;
9138 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9139 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9140 F->insert(MBBIter, newMBB);
9141 F->insert(MBBIter, nextMBB);
9143 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9144 nextMBB->splice(nextMBB->begin(), thisMBB,
9145 llvm::next(MachineBasicBlock::iterator(mInstr)),
9147 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9149 // Update thisMBB to fall through to newMBB
9150 thisMBB->addSuccessor(newMBB);
9152 // newMBB jumps to newMBB and fall through to nextMBB
9153 newMBB->addSuccessor(nextMBB);
9154 newMBB->addSuccessor(newMBB);
9156 DebugLoc dl = mInstr->getDebugLoc();
9157 // Insert instructions into newMBB based on incoming instruction
9158 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9159 "unexpected number of operands");
9160 MachineOperand& destOper = mInstr->getOperand(0);
9161 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9162 int numArgs = mInstr->getNumOperands() - 1;
9163 for (int i=0; i < numArgs; ++i)
9164 argOpers[i] = &mInstr->getOperand(i+1);
9166 // x86 address has 4 operands: base, index, scale, and displacement
9167 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9168 int valArgIndx = lastAddrIndx + 1;
9170 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9171 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
9172 for (int i=0; i <= lastAddrIndx; ++i)
9173 (*MIB).addOperand(*argOpers[i]);
9175 // We only support register and immediate values
9176 assert((argOpers[valArgIndx]->isReg() ||
9177 argOpers[valArgIndx]->isImm()) &&
9180 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9181 if (argOpers[valArgIndx]->isReg())
9182 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
9184 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
9185 (*MIB).addOperand(*argOpers[valArgIndx]);
9187 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9190 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
9195 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9196 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
9200 // Cmp and exchange if none has modified the memory location
9201 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
9202 for (int i=0; i <= lastAddrIndx; ++i)
9203 (*MIB).addOperand(*argOpers[i]);
9205 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9206 (*MIB).setMemRefs(mInstr->memoperands_begin(),
9207 mInstr->memoperands_end());
9209 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9210 MIB.addReg(X86::EAX);
9213 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9215 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
9219 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
9220 // or XMM0_V32I8 in AVX all of this code can be replaced with that
9223 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
9224 unsigned numArgs, bool memArg) const {
9226 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
9227 "Target must have SSE4.2 or AVX features enabled");
9229 DebugLoc dl = MI->getDebugLoc();
9230 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9234 if (!Subtarget->hasAVX()) {
9236 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
9238 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
9241 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
9243 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
9246 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
9248 for (unsigned i = 0; i < numArgs; ++i) {
9249 MachineOperand &Op = MI->getOperand(i+1);
9251 if (!(Op.isReg() && Op.isImplicit()))
9255 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
9258 MI->eraseFromParent();
9264 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
9266 MachineBasicBlock *MBB) const {
9267 // Emit code to save XMM registers to the stack. The ABI says that the
9268 // number of registers to save is given in %al, so it's theoretically
9269 // possible to do an indirect jump trick to avoid saving all of them,
9270 // however this code takes a simpler approach and just executes all
9271 // of the stores if %al is non-zero. It's less code, and it's probably
9272 // easier on the hardware branch predictor, and stores aren't all that
9273 // expensive anyway.
9275 // Create the new basic blocks. One block contains all the XMM stores,
9276 // and one block is the final destination regardless of whether any
9277 // stores were performed.
9278 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9279 MachineFunction *F = MBB->getParent();
9280 MachineFunction::iterator MBBIter = MBB;
9282 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
9283 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
9284 F->insert(MBBIter, XMMSaveMBB);
9285 F->insert(MBBIter, EndMBB);
9287 // Transfer the remainder of MBB and its successor edges to EndMBB.
9288 EndMBB->splice(EndMBB->begin(), MBB,
9289 llvm::next(MachineBasicBlock::iterator(MI)),
9291 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
9293 // The original block will now fall through to the XMM save block.
9294 MBB->addSuccessor(XMMSaveMBB);
9295 // The XMMSaveMBB will fall through to the end block.
9296 XMMSaveMBB->addSuccessor(EndMBB);
9298 // Now add the instructions.
9299 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9300 DebugLoc DL = MI->getDebugLoc();
9302 unsigned CountReg = MI->getOperand(0).getReg();
9303 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
9304 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
9306 if (!Subtarget->isTargetWin64()) {
9307 // If %al is 0, branch around the XMM save block.
9308 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
9309 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
9310 MBB->addSuccessor(EndMBB);
9313 // In the XMM save block, save all the XMM argument registers.
9314 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
9315 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
9316 MachineMemOperand *MMO =
9317 F->getMachineMemOperand(
9318 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
9319 MachineMemOperand::MOStore, Offset,
9320 /*Size=*/16, /*Align=*/16);
9321 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
9322 .addFrameIndex(RegSaveFrameIndex)
9323 .addImm(/*Scale=*/1)
9324 .addReg(/*IndexReg=*/0)
9325 .addImm(/*Disp=*/Offset)
9326 .addReg(/*Segment=*/0)
9327 .addReg(MI->getOperand(i).getReg())
9328 .addMemOperand(MMO);
9331 MI->eraseFromParent(); // The pseudo instruction is gone now.
9337 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
9338 MachineBasicBlock *BB) const {
9339 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9340 DebugLoc DL = MI->getDebugLoc();
9342 // To "insert" a SELECT_CC instruction, we actually have to insert the
9343 // diamond control-flow pattern. The incoming instruction knows the
9344 // destination vreg to set, the condition code register to branch on, the
9345 // true/false values to select between, and a branch opcode to use.
9346 const BasicBlock *LLVM_BB = BB->getBasicBlock();
9347 MachineFunction::iterator It = BB;
9353 // cmpTY ccX, r1, r2
9355 // fallthrough --> copy0MBB
9356 MachineBasicBlock *thisMBB = BB;
9357 MachineFunction *F = BB->getParent();
9358 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
9359 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
9360 F->insert(It, copy0MBB);
9361 F->insert(It, sinkMBB);
9363 // If the EFLAGS register isn't dead in the terminator, then claim that it's
9364 // live into the sink and copy blocks.
9365 const MachineFunction *MF = BB->getParent();
9366 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
9367 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
9369 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
9370 const MachineOperand &MO = MI->getOperand(I);
9371 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
9372 unsigned Reg = MO.getReg();
9373 if (Reg != X86::EFLAGS) continue;
9374 copy0MBB->addLiveIn(Reg);
9375 sinkMBB->addLiveIn(Reg);
9378 // Transfer the remainder of BB and its successor edges to sinkMBB.
9379 sinkMBB->splice(sinkMBB->begin(), BB,
9380 llvm::next(MachineBasicBlock::iterator(MI)),
9382 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
9384 // Add the true and fallthrough blocks as its successors.
9385 BB->addSuccessor(copy0MBB);
9386 BB->addSuccessor(sinkMBB);
9388 // Create the conditional branch instruction.
9390 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
9391 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
9394 // %FalseValue = ...
9395 // # fallthrough to sinkMBB
9396 copy0MBB->addSuccessor(sinkMBB);
9399 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
9401 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
9402 TII->get(X86::PHI), MI->getOperand(0).getReg())
9403 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
9404 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
9406 MI->eraseFromParent(); // The pseudo instruction is gone now.
9411 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
9412 MachineBasicBlock *BB) const {
9413 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9414 DebugLoc DL = MI->getDebugLoc();
9416 // The lowering is pretty easy: we're just emitting the call to _alloca. The
9417 // non-trivial part is impdef of ESP.
9418 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
9421 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
9422 .addExternalSymbol("_alloca")
9423 .addReg(X86::EAX, RegState::Implicit)
9424 .addReg(X86::ESP, RegState::Implicit)
9425 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
9426 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
9427 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
9429 MI->eraseFromParent(); // The pseudo instruction is gone now.
9434 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
9435 MachineBasicBlock *BB) const {
9436 // This is pretty easy. We're taking the value that we received from
9437 // our load from the relocation, sticking it in either RDI (x86-64)
9438 // or EAX and doing an indirect call. The return value will then
9439 // be in the normal return register.
9440 const X86InstrInfo *TII
9441 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
9442 DebugLoc DL = MI->getDebugLoc();
9443 MachineFunction *F = BB->getParent();
9444 bool IsWin64 = Subtarget->isTargetWin64();
9446 assert(MI->getOperand(3).isGlobal() && "This should be a global");
9448 if (Subtarget->is64Bit()) {
9449 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9450 TII->get(X86::MOV64rm), X86::RDI)
9452 .addImm(0).addReg(0)
9453 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9454 MI->getOperand(3).getTargetFlags())
9456 MIB = BuildMI(*BB, MI, DL, TII->get(IsWin64 ? X86::WINCALL64m : X86::CALL64m));
9457 addDirectMem(MIB, X86::RDI);
9458 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
9459 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9460 TII->get(X86::MOV32rm), X86::EAX)
9462 .addImm(0).addReg(0)
9463 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9464 MI->getOperand(3).getTargetFlags())
9466 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9467 addDirectMem(MIB, X86::EAX);
9469 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9470 TII->get(X86::MOV32rm), X86::EAX)
9471 .addReg(TII->getGlobalBaseReg(F))
9472 .addImm(0).addReg(0)
9473 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9474 MI->getOperand(3).getTargetFlags())
9476 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9477 addDirectMem(MIB, X86::EAX);
9480 MI->eraseFromParent(); // The pseudo instruction is gone now.
9485 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
9486 MachineBasicBlock *BB) const {
9487 switch (MI->getOpcode()) {
9488 default: assert(false && "Unexpected instr type to insert");
9489 case X86::MINGW_ALLOCA:
9490 return EmitLoweredMingwAlloca(MI, BB);
9491 case X86::TLSCall_32:
9492 case X86::TLSCall_64:
9493 return EmitLoweredTLSCall(MI, BB);
9495 case X86::CMOV_V1I64:
9496 case X86::CMOV_FR32:
9497 case X86::CMOV_FR64:
9498 case X86::CMOV_V4F32:
9499 case X86::CMOV_V2F64:
9500 case X86::CMOV_V2I64:
9501 case X86::CMOV_GR16:
9502 case X86::CMOV_GR32:
9503 case X86::CMOV_RFP32:
9504 case X86::CMOV_RFP64:
9505 case X86::CMOV_RFP80:
9506 return EmitLoweredSelect(MI, BB);
9508 case X86::FP32_TO_INT16_IN_MEM:
9509 case X86::FP32_TO_INT32_IN_MEM:
9510 case X86::FP32_TO_INT64_IN_MEM:
9511 case X86::FP64_TO_INT16_IN_MEM:
9512 case X86::FP64_TO_INT32_IN_MEM:
9513 case X86::FP64_TO_INT64_IN_MEM:
9514 case X86::FP80_TO_INT16_IN_MEM:
9515 case X86::FP80_TO_INT32_IN_MEM:
9516 case X86::FP80_TO_INT64_IN_MEM: {
9517 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9518 DebugLoc DL = MI->getDebugLoc();
9520 // Change the floating point control register to use "round towards zero"
9521 // mode when truncating to an integer value.
9522 MachineFunction *F = BB->getParent();
9523 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
9524 addFrameReference(BuildMI(*BB, MI, DL,
9525 TII->get(X86::FNSTCW16m)), CWFrameIdx);
9527 // Load the old value of the high byte of the control word...
9529 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
9530 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
9533 // Set the high part to be round to zero...
9534 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
9537 // Reload the modified control word now...
9538 addFrameReference(BuildMI(*BB, MI, DL,
9539 TII->get(X86::FLDCW16m)), CWFrameIdx);
9541 // Restore the memory image of control word to original value
9542 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
9545 // Get the X86 opcode to use.
9547 switch (MI->getOpcode()) {
9548 default: llvm_unreachable("illegal opcode!");
9549 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
9550 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
9551 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
9552 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
9553 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
9554 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
9555 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
9556 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
9557 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
9561 MachineOperand &Op = MI->getOperand(0);
9563 AM.BaseType = X86AddressMode::RegBase;
9564 AM.Base.Reg = Op.getReg();
9566 AM.BaseType = X86AddressMode::FrameIndexBase;
9567 AM.Base.FrameIndex = Op.getIndex();
9569 Op = MI->getOperand(1);
9571 AM.Scale = Op.getImm();
9572 Op = MI->getOperand(2);
9574 AM.IndexReg = Op.getImm();
9575 Op = MI->getOperand(3);
9576 if (Op.isGlobal()) {
9577 AM.GV = Op.getGlobal();
9579 AM.Disp = Op.getImm();
9581 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
9582 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
9584 // Reload the original control word now.
9585 addFrameReference(BuildMI(*BB, MI, DL,
9586 TII->get(X86::FLDCW16m)), CWFrameIdx);
9588 MI->eraseFromParent(); // The pseudo instruction is gone now.
9591 // String/text processing lowering.
9592 case X86::PCMPISTRM128REG:
9593 case X86::VPCMPISTRM128REG:
9594 return EmitPCMP(MI, BB, 3, false /* in-mem */);
9595 case X86::PCMPISTRM128MEM:
9596 case X86::VPCMPISTRM128MEM:
9597 return EmitPCMP(MI, BB, 3, true /* in-mem */);
9598 case X86::PCMPESTRM128REG:
9599 case X86::VPCMPESTRM128REG:
9600 return EmitPCMP(MI, BB, 5, false /* in mem */);
9601 case X86::PCMPESTRM128MEM:
9602 case X86::VPCMPESTRM128MEM:
9603 return EmitPCMP(MI, BB, 5, true /* in mem */);
9606 case X86::ATOMAND32:
9607 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9608 X86::AND32ri, X86::MOV32rm,
9610 X86::NOT32r, X86::EAX,
9611 X86::GR32RegisterClass);
9613 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
9614 X86::OR32ri, X86::MOV32rm,
9616 X86::NOT32r, X86::EAX,
9617 X86::GR32RegisterClass);
9618 case X86::ATOMXOR32:
9619 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
9620 X86::XOR32ri, X86::MOV32rm,
9622 X86::NOT32r, X86::EAX,
9623 X86::GR32RegisterClass);
9624 case X86::ATOMNAND32:
9625 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9626 X86::AND32ri, X86::MOV32rm,
9628 X86::NOT32r, X86::EAX,
9629 X86::GR32RegisterClass, true);
9630 case X86::ATOMMIN32:
9631 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
9632 case X86::ATOMMAX32:
9633 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
9634 case X86::ATOMUMIN32:
9635 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
9636 case X86::ATOMUMAX32:
9637 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
9639 case X86::ATOMAND16:
9640 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9641 X86::AND16ri, X86::MOV16rm,
9643 X86::NOT16r, X86::AX,
9644 X86::GR16RegisterClass);
9646 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
9647 X86::OR16ri, X86::MOV16rm,
9649 X86::NOT16r, X86::AX,
9650 X86::GR16RegisterClass);
9651 case X86::ATOMXOR16:
9652 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
9653 X86::XOR16ri, X86::MOV16rm,
9655 X86::NOT16r, X86::AX,
9656 X86::GR16RegisterClass);
9657 case X86::ATOMNAND16:
9658 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9659 X86::AND16ri, X86::MOV16rm,
9661 X86::NOT16r, X86::AX,
9662 X86::GR16RegisterClass, true);
9663 case X86::ATOMMIN16:
9664 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
9665 case X86::ATOMMAX16:
9666 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
9667 case X86::ATOMUMIN16:
9668 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
9669 case X86::ATOMUMAX16:
9670 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
9673 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9674 X86::AND8ri, X86::MOV8rm,
9676 X86::NOT8r, X86::AL,
9677 X86::GR8RegisterClass);
9679 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
9680 X86::OR8ri, X86::MOV8rm,
9682 X86::NOT8r, X86::AL,
9683 X86::GR8RegisterClass);
9685 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
9686 X86::XOR8ri, X86::MOV8rm,
9688 X86::NOT8r, X86::AL,
9689 X86::GR8RegisterClass);
9690 case X86::ATOMNAND8:
9691 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9692 X86::AND8ri, X86::MOV8rm,
9694 X86::NOT8r, X86::AL,
9695 X86::GR8RegisterClass, true);
9696 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
9697 // This group is for 64-bit host.
9698 case X86::ATOMAND64:
9699 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9700 X86::AND64ri32, X86::MOV64rm,
9702 X86::NOT64r, X86::RAX,
9703 X86::GR64RegisterClass);
9705 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
9706 X86::OR64ri32, X86::MOV64rm,
9708 X86::NOT64r, X86::RAX,
9709 X86::GR64RegisterClass);
9710 case X86::ATOMXOR64:
9711 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
9712 X86::XOR64ri32, X86::MOV64rm,
9714 X86::NOT64r, X86::RAX,
9715 X86::GR64RegisterClass);
9716 case X86::ATOMNAND64:
9717 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9718 X86::AND64ri32, X86::MOV64rm,
9720 X86::NOT64r, X86::RAX,
9721 X86::GR64RegisterClass, true);
9722 case X86::ATOMMIN64:
9723 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
9724 case X86::ATOMMAX64:
9725 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
9726 case X86::ATOMUMIN64:
9727 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
9728 case X86::ATOMUMAX64:
9729 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
9731 // This group does 64-bit operations on a 32-bit host.
9732 case X86::ATOMAND6432:
9733 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9734 X86::AND32rr, X86::AND32rr,
9735 X86::AND32ri, X86::AND32ri,
9737 case X86::ATOMOR6432:
9738 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9739 X86::OR32rr, X86::OR32rr,
9740 X86::OR32ri, X86::OR32ri,
9742 case X86::ATOMXOR6432:
9743 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9744 X86::XOR32rr, X86::XOR32rr,
9745 X86::XOR32ri, X86::XOR32ri,
9747 case X86::ATOMNAND6432:
9748 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9749 X86::AND32rr, X86::AND32rr,
9750 X86::AND32ri, X86::AND32ri,
9752 case X86::ATOMADD6432:
9753 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9754 X86::ADD32rr, X86::ADC32rr,
9755 X86::ADD32ri, X86::ADC32ri,
9757 case X86::ATOMSUB6432:
9758 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9759 X86::SUB32rr, X86::SBB32rr,
9760 X86::SUB32ri, X86::SBB32ri,
9762 case X86::ATOMSWAP6432:
9763 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9764 X86::MOV32rr, X86::MOV32rr,
9765 X86::MOV32ri, X86::MOV32ri,
9767 case X86::VASTART_SAVE_XMM_REGS:
9768 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
9772 //===----------------------------------------------------------------------===//
9773 // X86 Optimization Hooks
9774 //===----------------------------------------------------------------------===//
9776 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9780 const SelectionDAG &DAG,
9781 unsigned Depth) const {
9782 unsigned Opc = Op.getOpcode();
9783 assert((Opc >= ISD::BUILTIN_OP_END ||
9784 Opc == ISD::INTRINSIC_WO_CHAIN ||
9785 Opc == ISD::INTRINSIC_W_CHAIN ||
9786 Opc == ISD::INTRINSIC_VOID) &&
9787 "Should use MaskedValueIsZero if you don't know whether Op"
9788 " is a target node!");
9790 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
9802 // These nodes' second result is a boolean.
9803 if (Op.getResNo() == 0)
9807 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
9808 Mask.getBitWidth() - 1);
9813 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
9814 /// node is a GlobalAddress + offset.
9815 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
9816 const GlobalValue* &GA,
9817 int64_t &Offset) const {
9818 if (N->getOpcode() == X86ISD::Wrapper) {
9819 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
9820 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
9821 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
9825 return TargetLowering::isGAPlusOffset(N, GA, Offset);
9828 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
9829 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
9830 /// if the load addresses are consecutive, non-overlapping, and in the right
9832 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
9833 const TargetLowering &TLI) {
9834 DebugLoc dl = N->getDebugLoc();
9835 EVT VT = N->getValueType(0);
9837 if (VT.getSizeInBits() != 128)
9840 SmallVector<SDValue, 16> Elts;
9841 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
9842 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
9844 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
9847 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
9848 /// generation and convert it from being a bunch of shuffles and extracts
9849 /// to a simple store and scalar loads to extract the elements.
9850 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
9851 const TargetLowering &TLI) {
9852 SDValue InputVector = N->getOperand(0);
9854 // Only operate on vectors of 4 elements, where the alternative shuffling
9855 // gets to be more expensive.
9856 if (InputVector.getValueType() != MVT::v4i32)
9859 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
9860 // single use which is a sign-extend or zero-extend, and all elements are
9862 SmallVector<SDNode *, 4> Uses;
9863 unsigned ExtractedElements = 0;
9864 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
9865 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
9866 if (UI.getUse().getResNo() != InputVector.getResNo())
9869 SDNode *Extract = *UI;
9870 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9873 if (Extract->getValueType(0) != MVT::i32)
9875 if (!Extract->hasOneUse())
9877 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9878 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9880 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9883 // Record which element was extracted.
9884 ExtractedElements |=
9885 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9887 Uses.push_back(Extract);
9890 // If not all the elements were used, this may not be worthwhile.
9891 if (ExtractedElements != 15)
9894 // Ok, we've now decided to do the transformation.
9895 DebugLoc dl = InputVector.getDebugLoc();
9897 // Store the value to a temporary stack slot.
9898 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9899 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL,
9900 0, false, false, 0);
9902 // Replace each use (extract) with a load of the appropriate element.
9903 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9904 UE = Uses.end(); UI != UE; ++UI) {
9905 SDNode *Extract = *UI;
9907 // Compute the element's address.
9908 SDValue Idx = Extract->getOperand(1);
9910 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9911 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9912 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9914 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
9915 OffsetVal, StackPtr);
9918 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
9919 ScalarAddr, NULL, 0, false, false, 0);
9921 // Replace the exact with the load.
9922 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9925 // The replacement was made in place; don't return anything.
9929 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9930 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9931 const X86Subtarget *Subtarget) {
9932 DebugLoc DL = N->getDebugLoc();
9933 SDValue Cond = N->getOperand(0);
9934 // Get the LHS/RHS of the select.
9935 SDValue LHS = N->getOperand(1);
9936 SDValue RHS = N->getOperand(2);
9938 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9939 // instructions match the semantics of the common C idiom x<y?x:y but not
9940 // x<=y?x:y, because of how they handle negative zero (which can be
9941 // ignored in unsafe-math mode).
9942 if (Subtarget->hasSSE2() &&
9943 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9944 Cond.getOpcode() == ISD::SETCC) {
9945 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9947 unsigned Opcode = 0;
9948 // Check for x CC y ? x : y.
9949 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9950 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9954 // Converting this to a min would handle NaNs incorrectly, and swapping
9955 // the operands would cause it to handle comparisons between positive
9956 // and negative zero incorrectly.
9957 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9958 if (!UnsafeFPMath &&
9959 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9961 std::swap(LHS, RHS);
9963 Opcode = X86ISD::FMIN;
9966 // Converting this to a min would handle comparisons between positive
9967 // and negative zero incorrectly.
9968 if (!UnsafeFPMath &&
9969 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9971 Opcode = X86ISD::FMIN;
9974 // Converting this to a min would handle both negative zeros and NaNs
9975 // incorrectly, but we can swap the operands to fix both.
9976 std::swap(LHS, RHS);
9980 Opcode = X86ISD::FMIN;
9984 // Converting this to a max would handle comparisons between positive
9985 // and negative zero incorrectly.
9986 if (!UnsafeFPMath &&
9987 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9989 Opcode = X86ISD::FMAX;
9992 // Converting this to a max would handle NaNs incorrectly, and swapping
9993 // the operands would cause it to handle comparisons between positive
9994 // and negative zero incorrectly.
9995 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9996 if (!UnsafeFPMath &&
9997 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9999 std::swap(LHS, RHS);
10001 Opcode = X86ISD::FMAX;
10004 // Converting this to a max would handle both negative zeros and NaNs
10005 // incorrectly, but we can swap the operands to fix both.
10006 std::swap(LHS, RHS);
10010 Opcode = X86ISD::FMAX;
10013 // Check for x CC y ? y : x -- a min/max with reversed arms.
10014 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
10015 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
10019 // Converting this to a min would handle comparisons between positive
10020 // and negative zero incorrectly, and swapping the operands would
10021 // cause it to handle NaNs incorrectly.
10022 if (!UnsafeFPMath &&
10023 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
10024 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10026 std::swap(LHS, RHS);
10028 Opcode = X86ISD::FMIN;
10031 // Converting this to a min would handle NaNs incorrectly.
10032 if (!UnsafeFPMath &&
10033 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
10035 Opcode = X86ISD::FMIN;
10038 // Converting this to a min would handle both negative zeros and NaNs
10039 // incorrectly, but we can swap the operands to fix both.
10040 std::swap(LHS, RHS);
10044 Opcode = X86ISD::FMIN;
10048 // Converting this to a max would handle NaNs incorrectly.
10049 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10051 Opcode = X86ISD::FMAX;
10054 // Converting this to a max would handle comparisons between positive
10055 // and negative zero incorrectly, and swapping the operands would
10056 // cause it to handle NaNs incorrectly.
10057 if (!UnsafeFPMath &&
10058 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
10059 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10061 std::swap(LHS, RHS);
10063 Opcode = X86ISD::FMAX;
10066 // Converting this to a max would handle both negative zeros and NaNs
10067 // incorrectly, but we can swap the operands to fix both.
10068 std::swap(LHS, RHS);
10072 Opcode = X86ISD::FMAX;
10078 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
10081 // If this is a select between two integer constants, try to do some
10083 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
10084 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
10085 // Don't do this for crazy integer types.
10086 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
10087 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
10088 // so that TrueC (the true value) is larger than FalseC.
10089 bool NeedsCondInvert = false;
10091 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
10092 // Efficiently invertible.
10093 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
10094 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
10095 isa<ConstantSDNode>(Cond.getOperand(1))))) {
10096 NeedsCondInvert = true;
10097 std::swap(TrueC, FalseC);
10100 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
10101 if (FalseC->getAPIntValue() == 0 &&
10102 TrueC->getAPIntValue().isPowerOf2()) {
10103 if (NeedsCondInvert) // Invert the condition if needed.
10104 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10105 DAG.getConstant(1, Cond.getValueType()));
10107 // Zero extend the condition if needed.
10108 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
10110 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10111 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
10112 DAG.getConstant(ShAmt, MVT::i8));
10115 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
10116 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10117 if (NeedsCondInvert) // Invert the condition if needed.
10118 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10119 DAG.getConstant(1, Cond.getValueType()));
10121 // Zero extend the condition if needed.
10122 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10123 FalseC->getValueType(0), Cond);
10124 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10125 SDValue(FalseC, 0));
10128 // Optimize cases that will turn into an LEA instruction. This requires
10129 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10130 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10131 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10132 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10134 bool isFastMultiplier = false;
10136 switch ((unsigned char)Diff) {
10138 case 1: // result = add base, cond
10139 case 2: // result = lea base( , cond*2)
10140 case 3: // result = lea base(cond, cond*2)
10141 case 4: // result = lea base( , cond*4)
10142 case 5: // result = lea base(cond, cond*4)
10143 case 8: // result = lea base( , cond*8)
10144 case 9: // result = lea base(cond, cond*8)
10145 isFastMultiplier = true;
10150 if (isFastMultiplier) {
10151 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10152 if (NeedsCondInvert) // Invert the condition if needed.
10153 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10154 DAG.getConstant(1, Cond.getValueType()));
10156 // Zero extend the condition if needed.
10157 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10159 // Scale the condition by the difference.
10161 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10162 DAG.getConstant(Diff, Cond.getValueType()));
10164 // Add the base if non-zero.
10165 if (FalseC->getAPIntValue() != 0)
10166 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10167 SDValue(FalseC, 0));
10177 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
10178 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
10179 TargetLowering::DAGCombinerInfo &DCI) {
10180 DebugLoc DL = N->getDebugLoc();
10182 // If the flag operand isn't dead, don't touch this CMOV.
10183 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
10186 // If this is a select between two integer constants, try to do some
10187 // optimizations. Note that the operands are ordered the opposite of SELECT
10189 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
10190 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
10191 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
10192 // larger than FalseC (the false value).
10193 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
10195 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
10196 CC = X86::GetOppositeBranchCondition(CC);
10197 std::swap(TrueC, FalseC);
10200 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
10201 // This is efficient for any integer data type (including i8/i16) and
10203 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
10204 SDValue Cond = N->getOperand(3);
10205 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10206 DAG.getConstant(CC, MVT::i8), Cond);
10208 // Zero extend the condition if needed.
10209 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
10211 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10212 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
10213 DAG.getConstant(ShAmt, MVT::i8));
10214 if (N->getNumValues() == 2) // Dead flag value?
10215 return DCI.CombineTo(N, Cond, SDValue());
10219 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
10220 // for any integer data type, including i8/i16.
10221 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10222 SDValue Cond = N->getOperand(3);
10223 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10224 DAG.getConstant(CC, MVT::i8), Cond);
10226 // Zero extend the condition if needed.
10227 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10228 FalseC->getValueType(0), Cond);
10229 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10230 SDValue(FalseC, 0));
10232 if (N->getNumValues() == 2) // Dead flag value?
10233 return DCI.CombineTo(N, Cond, SDValue());
10237 // Optimize cases that will turn into an LEA instruction. This requires
10238 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10239 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10240 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10241 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10243 bool isFastMultiplier = false;
10245 switch ((unsigned char)Diff) {
10247 case 1: // result = add base, cond
10248 case 2: // result = lea base( , cond*2)
10249 case 3: // result = lea base(cond, cond*2)
10250 case 4: // result = lea base( , cond*4)
10251 case 5: // result = lea base(cond, cond*4)
10252 case 8: // result = lea base( , cond*8)
10253 case 9: // result = lea base(cond, cond*8)
10254 isFastMultiplier = true;
10259 if (isFastMultiplier) {
10260 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10261 SDValue Cond = N->getOperand(3);
10262 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10263 DAG.getConstant(CC, MVT::i8), Cond);
10264 // Zero extend the condition if needed.
10265 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10267 // Scale the condition by the difference.
10269 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10270 DAG.getConstant(Diff, Cond.getValueType()));
10272 // Add the base if non-zero.
10273 if (FalseC->getAPIntValue() != 0)
10274 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10275 SDValue(FalseC, 0));
10276 if (N->getNumValues() == 2) // Dead flag value?
10277 return DCI.CombineTo(N, Cond, SDValue());
10287 /// PerformMulCombine - Optimize a single multiply with constant into two
10288 /// in order to implement it with two cheaper instructions, e.g.
10289 /// LEA + SHL, LEA + LEA.
10290 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
10291 TargetLowering::DAGCombinerInfo &DCI) {
10292 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
10295 EVT VT = N->getValueType(0);
10296 if (VT != MVT::i64)
10299 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
10302 uint64_t MulAmt = C->getZExtValue();
10303 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
10306 uint64_t MulAmt1 = 0;
10307 uint64_t MulAmt2 = 0;
10308 if ((MulAmt % 9) == 0) {
10310 MulAmt2 = MulAmt / 9;
10311 } else if ((MulAmt % 5) == 0) {
10313 MulAmt2 = MulAmt / 5;
10314 } else if ((MulAmt % 3) == 0) {
10316 MulAmt2 = MulAmt / 3;
10319 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
10320 DebugLoc DL = N->getDebugLoc();
10322 if (isPowerOf2_64(MulAmt2) &&
10323 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
10324 // If second multiplifer is pow2, issue it first. We want the multiply by
10325 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
10327 std::swap(MulAmt1, MulAmt2);
10330 if (isPowerOf2_64(MulAmt1))
10331 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
10332 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
10334 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
10335 DAG.getConstant(MulAmt1, VT));
10337 if (isPowerOf2_64(MulAmt2))
10338 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
10339 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
10341 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
10342 DAG.getConstant(MulAmt2, VT));
10344 // Do not add new nodes to DAG combiner worklist.
10345 DCI.CombineTo(N, NewMul, false);
10350 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
10351 SDValue N0 = N->getOperand(0);
10352 SDValue N1 = N->getOperand(1);
10353 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
10354 EVT VT = N0.getValueType();
10356 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
10357 // since the result of setcc_c is all zero's or all ones.
10358 if (N1C && N0.getOpcode() == ISD::AND &&
10359 N0.getOperand(1).getOpcode() == ISD::Constant) {
10360 SDValue N00 = N0.getOperand(0);
10361 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
10362 ((N00.getOpcode() == ISD::ANY_EXTEND ||
10363 N00.getOpcode() == ISD::ZERO_EXTEND) &&
10364 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
10365 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
10366 APInt ShAmt = N1C->getAPIntValue();
10367 Mask = Mask.shl(ShAmt);
10369 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
10370 N00, DAG.getConstant(Mask, VT));
10377 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
10379 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
10380 const X86Subtarget *Subtarget) {
10381 EVT VT = N->getValueType(0);
10382 if (!VT.isVector() && VT.isInteger() &&
10383 N->getOpcode() == ISD::SHL)
10384 return PerformSHLCombine(N, DAG);
10386 // On X86 with SSE2 support, we can transform this to a vector shift if
10387 // all elements are shifted by the same amount. We can't do this in legalize
10388 // because the a constant vector is typically transformed to a constant pool
10389 // so we have no knowledge of the shift amount.
10390 if (!Subtarget->hasSSE2())
10393 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
10396 SDValue ShAmtOp = N->getOperand(1);
10397 EVT EltVT = VT.getVectorElementType();
10398 DebugLoc DL = N->getDebugLoc();
10399 SDValue BaseShAmt = SDValue();
10400 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
10401 unsigned NumElts = VT.getVectorNumElements();
10403 for (; i != NumElts; ++i) {
10404 SDValue Arg = ShAmtOp.getOperand(i);
10405 if (Arg.getOpcode() == ISD::UNDEF) continue;
10409 for (; i != NumElts; ++i) {
10410 SDValue Arg = ShAmtOp.getOperand(i);
10411 if (Arg.getOpcode() == ISD::UNDEF) continue;
10412 if (Arg != BaseShAmt) {
10416 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
10417 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
10418 SDValue InVec = ShAmtOp.getOperand(0);
10419 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
10420 unsigned NumElts = InVec.getValueType().getVectorNumElements();
10422 for (; i != NumElts; ++i) {
10423 SDValue Arg = InVec.getOperand(i);
10424 if (Arg.getOpcode() == ISD::UNDEF) continue;
10428 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
10429 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
10430 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
10431 if (C->getZExtValue() == SplatIdx)
10432 BaseShAmt = InVec.getOperand(1);
10435 if (BaseShAmt.getNode() == 0)
10436 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
10437 DAG.getIntPtrConstant(0));
10441 // The shift amount is an i32.
10442 if (EltVT.bitsGT(MVT::i32))
10443 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
10444 else if (EltVT.bitsLT(MVT::i32))
10445 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
10447 // The shift amount is identical so we can do a vector shift.
10448 SDValue ValOp = N->getOperand(0);
10449 switch (N->getOpcode()) {
10451 llvm_unreachable("Unknown shift opcode!");
10454 if (VT == MVT::v2i64)
10455 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10456 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
10458 if (VT == MVT::v4i32)
10459 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10460 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
10462 if (VT == MVT::v8i16)
10463 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10464 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
10468 if (VT == MVT::v4i32)
10469 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10470 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
10472 if (VT == MVT::v8i16)
10473 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10474 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
10478 if (VT == MVT::v2i64)
10479 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10480 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
10482 if (VT == MVT::v4i32)
10483 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10484 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
10486 if (VT == MVT::v8i16)
10487 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10488 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
10495 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
10496 TargetLowering::DAGCombinerInfo &DCI,
10497 const X86Subtarget *Subtarget) {
10498 if (DCI.isBeforeLegalizeOps())
10501 EVT VT = N->getValueType(0);
10502 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
10505 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
10506 SDValue N0 = N->getOperand(0);
10507 SDValue N1 = N->getOperand(1);
10508 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
10510 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
10512 if (!N0.hasOneUse() || !N1.hasOneUse())
10515 SDValue ShAmt0 = N0.getOperand(1);
10516 if (ShAmt0.getValueType() != MVT::i8)
10518 SDValue ShAmt1 = N1.getOperand(1);
10519 if (ShAmt1.getValueType() != MVT::i8)
10521 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
10522 ShAmt0 = ShAmt0.getOperand(0);
10523 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
10524 ShAmt1 = ShAmt1.getOperand(0);
10526 DebugLoc DL = N->getDebugLoc();
10527 unsigned Opc = X86ISD::SHLD;
10528 SDValue Op0 = N0.getOperand(0);
10529 SDValue Op1 = N1.getOperand(0);
10530 if (ShAmt0.getOpcode() == ISD::SUB) {
10531 Opc = X86ISD::SHRD;
10532 std::swap(Op0, Op1);
10533 std::swap(ShAmt0, ShAmt1);
10536 unsigned Bits = VT.getSizeInBits();
10537 if (ShAmt1.getOpcode() == ISD::SUB) {
10538 SDValue Sum = ShAmt1.getOperand(0);
10539 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
10540 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
10541 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
10542 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
10543 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
10544 return DAG.getNode(Opc, DL, VT,
10546 DAG.getNode(ISD::TRUNCATE, DL,
10549 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
10550 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
10552 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
10553 return DAG.getNode(Opc, DL, VT,
10554 N0.getOperand(0), N1.getOperand(0),
10555 DAG.getNode(ISD::TRUNCATE, DL,
10562 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
10563 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
10564 const X86Subtarget *Subtarget) {
10565 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
10566 // the FP state in cases where an emms may be missing.
10567 // A preferable solution to the general problem is to figure out the right
10568 // places to insert EMMS. This qualifies as a quick hack.
10570 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
10571 StoreSDNode *St = cast<StoreSDNode>(N);
10572 EVT VT = St->getValue().getValueType();
10573 if (VT.getSizeInBits() != 64)
10576 const Function *F = DAG.getMachineFunction().getFunction();
10577 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
10578 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
10579 && Subtarget->hasSSE2();
10580 if ((VT.isVector() ||
10581 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
10582 isa<LoadSDNode>(St->getValue()) &&
10583 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
10584 St->getChain().hasOneUse() && !St->isVolatile()) {
10585 SDNode* LdVal = St->getValue().getNode();
10586 LoadSDNode *Ld = 0;
10587 int TokenFactorIndex = -1;
10588 SmallVector<SDValue, 8> Ops;
10589 SDNode* ChainVal = St->getChain().getNode();
10590 // Must be a store of a load. We currently handle two cases: the load
10591 // is a direct child, and it's under an intervening TokenFactor. It is
10592 // possible to dig deeper under nested TokenFactors.
10593 if (ChainVal == LdVal)
10594 Ld = cast<LoadSDNode>(St->getChain());
10595 else if (St->getValue().hasOneUse() &&
10596 ChainVal->getOpcode() == ISD::TokenFactor) {
10597 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
10598 if (ChainVal->getOperand(i).getNode() == LdVal) {
10599 TokenFactorIndex = i;
10600 Ld = cast<LoadSDNode>(St->getValue());
10602 Ops.push_back(ChainVal->getOperand(i));
10606 if (!Ld || !ISD::isNormalLoad(Ld))
10609 // If this is not the MMX case, i.e. we are just turning i64 load/store
10610 // into f64 load/store, avoid the transformation if there are multiple
10611 // uses of the loaded value.
10612 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
10615 DebugLoc LdDL = Ld->getDebugLoc();
10616 DebugLoc StDL = N->getDebugLoc();
10617 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
10618 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
10620 if (Subtarget->is64Bit() || F64IsLegal) {
10621 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
10622 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
10623 Ld->getBasePtr(), Ld->getSrcValue(),
10624 Ld->getSrcValueOffset(), Ld->isVolatile(),
10625 Ld->isNonTemporal(), Ld->getAlignment());
10626 SDValue NewChain = NewLd.getValue(1);
10627 if (TokenFactorIndex != -1) {
10628 Ops.push_back(NewChain);
10629 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10632 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
10633 St->getSrcValue(), St->getSrcValueOffset(),
10634 St->isVolatile(), St->isNonTemporal(),
10635 St->getAlignment());
10638 // Otherwise, lower to two pairs of 32-bit loads / stores.
10639 SDValue LoAddr = Ld->getBasePtr();
10640 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
10641 DAG.getConstant(4, MVT::i32));
10643 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
10644 Ld->getSrcValue(), Ld->getSrcValueOffset(),
10645 Ld->isVolatile(), Ld->isNonTemporal(),
10646 Ld->getAlignment());
10647 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
10648 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
10649 Ld->isVolatile(), Ld->isNonTemporal(),
10650 MinAlign(Ld->getAlignment(), 4));
10652 SDValue NewChain = LoLd.getValue(1);
10653 if (TokenFactorIndex != -1) {
10654 Ops.push_back(LoLd);
10655 Ops.push_back(HiLd);
10656 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10660 LoAddr = St->getBasePtr();
10661 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
10662 DAG.getConstant(4, MVT::i32));
10664 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
10665 St->getSrcValue(), St->getSrcValueOffset(),
10666 St->isVolatile(), St->isNonTemporal(),
10667 St->getAlignment());
10668 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
10670 St->getSrcValueOffset() + 4,
10672 St->isNonTemporal(),
10673 MinAlign(St->getAlignment(), 4));
10674 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
10679 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
10680 /// X86ISD::FXOR nodes.
10681 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
10682 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
10683 // F[X]OR(0.0, x) -> x
10684 // F[X]OR(x, 0.0) -> x
10685 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10686 if (C->getValueAPF().isPosZero())
10687 return N->getOperand(1);
10688 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10689 if (C->getValueAPF().isPosZero())
10690 return N->getOperand(0);
10694 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
10695 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
10696 // FAND(0.0, x) -> 0.0
10697 // FAND(x, 0.0) -> 0.0
10698 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10699 if (C->getValueAPF().isPosZero())
10700 return N->getOperand(0);
10701 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10702 if (C->getValueAPF().isPosZero())
10703 return N->getOperand(1);
10707 static SDValue PerformBTCombine(SDNode *N,
10709 TargetLowering::DAGCombinerInfo &DCI) {
10710 // BT ignores high bits in the bit index operand.
10711 SDValue Op1 = N->getOperand(1);
10712 if (Op1.hasOneUse()) {
10713 unsigned BitWidth = Op1.getValueSizeInBits();
10714 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
10715 APInt KnownZero, KnownOne;
10716 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
10717 !DCI.isBeforeLegalizeOps());
10718 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10719 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
10720 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
10721 DCI.CommitTargetLoweringOpt(TLO);
10726 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
10727 SDValue Op = N->getOperand(0);
10728 if (Op.getOpcode() == ISD::BIT_CONVERT)
10729 Op = Op.getOperand(0);
10730 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
10731 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
10732 VT.getVectorElementType().getSizeInBits() ==
10733 OpVT.getVectorElementType().getSizeInBits()) {
10734 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
10739 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
10740 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
10741 // (and (i32 x86isd::setcc_carry), 1)
10742 // This eliminates the zext. This transformation is necessary because
10743 // ISD::SETCC is always legalized to i8.
10744 DebugLoc dl = N->getDebugLoc();
10745 SDValue N0 = N->getOperand(0);
10746 EVT VT = N->getValueType(0);
10747 if (N0.getOpcode() == ISD::AND &&
10749 N0.getOperand(0).hasOneUse()) {
10750 SDValue N00 = N0.getOperand(0);
10751 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
10753 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
10754 if (!C || C->getZExtValue() != 1)
10756 return DAG.getNode(ISD::AND, dl, VT,
10757 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
10758 N00.getOperand(0), N00.getOperand(1)),
10759 DAG.getConstant(1, VT));
10765 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
10766 DAGCombinerInfo &DCI) const {
10767 SelectionDAG &DAG = DCI.DAG;
10768 switch (N->getOpcode()) {
10770 case ISD::EXTRACT_VECTOR_ELT:
10771 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
10772 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
10773 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
10774 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
10777 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
10778 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
10779 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
10781 case X86ISD::FOR: return PerformFORCombine(N, DAG);
10782 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
10783 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
10784 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
10785 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
10786 case X86ISD::SHUFPS: // Handle all target specific shuffles
10787 case X86ISD::SHUFPD:
10788 case X86ISD::PALIGN:
10789 case X86ISD::PUNPCKHBW:
10790 case X86ISD::PUNPCKHWD:
10791 case X86ISD::PUNPCKHDQ:
10792 case X86ISD::PUNPCKHQDQ:
10793 case X86ISD::UNPCKHPS:
10794 case X86ISD::UNPCKHPD:
10795 case X86ISD::PUNPCKLBW:
10796 case X86ISD::PUNPCKLWD:
10797 case X86ISD::PUNPCKLDQ:
10798 case X86ISD::PUNPCKLQDQ:
10799 case X86ISD::UNPCKLPS:
10800 case X86ISD::UNPCKLPD:
10801 case X86ISD::MOVHLPS:
10802 case X86ISD::MOVLHPS:
10803 case X86ISD::PSHUFD:
10804 case X86ISD::PSHUFHW:
10805 case X86ISD::PSHUFLW:
10806 case X86ISD::MOVSS:
10807 case X86ISD::MOVSD:
10808 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
10814 /// isTypeDesirableForOp - Return true if the target has native support for
10815 /// the specified value type and it is 'desirable' to use the type for the
10816 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
10817 /// instruction encodings are longer and some i16 instructions are slow.
10818 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
10819 if (!isTypeLegal(VT))
10821 if (VT != MVT::i16)
10828 case ISD::SIGN_EXTEND:
10829 case ISD::ZERO_EXTEND:
10830 case ISD::ANY_EXTEND:
10843 /// IsDesirableToPromoteOp - This method query the target whether it is
10844 /// beneficial for dag combiner to promote the specified node. If true, it
10845 /// should return the desired promotion type by reference.
10846 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
10847 EVT VT = Op.getValueType();
10848 if (VT != MVT::i16)
10851 bool Promote = false;
10852 bool Commute = false;
10853 switch (Op.getOpcode()) {
10856 LoadSDNode *LD = cast<LoadSDNode>(Op);
10857 // If the non-extending load has a single use and it's not live out, then it
10858 // might be folded.
10859 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
10860 Op.hasOneUse()*/) {
10861 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
10862 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
10863 // The only case where we'd want to promote LOAD (rather then it being
10864 // promoted as an operand is when it's only use is liveout.
10865 if (UI->getOpcode() != ISD::CopyToReg)
10872 case ISD::SIGN_EXTEND:
10873 case ISD::ZERO_EXTEND:
10874 case ISD::ANY_EXTEND:
10879 SDValue N0 = Op.getOperand(0);
10880 // Look out for (store (shl (load), x)).
10881 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10894 SDValue N0 = Op.getOperand(0);
10895 SDValue N1 = Op.getOperand(1);
10896 if (!Commute && MayFoldLoad(N1))
10898 // Avoid disabling potential load folding opportunities.
10899 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10901 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10911 //===----------------------------------------------------------------------===//
10912 // X86 Inline Assembly Support
10913 //===----------------------------------------------------------------------===//
10915 static bool LowerToBSwap(CallInst *CI) {
10916 // FIXME: this should verify that we are targetting a 486 or better. If not,
10917 // we will turn this bswap into something that will be lowered to logical ops
10918 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10919 // so don't worry about this.
10921 // Verify this is a simple bswap.
10922 if (CI->getNumArgOperands() != 1 ||
10923 CI->getType() != CI->getArgOperand(0)->getType() ||
10924 !CI->getType()->isIntegerTy())
10927 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10928 if (!Ty || Ty->getBitWidth() % 16 != 0)
10931 // Okay, we can do this xform, do so now.
10932 const Type *Tys[] = { Ty };
10933 Module *M = CI->getParent()->getParent()->getParent();
10934 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10936 Value *Op = CI->getArgOperand(0);
10937 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10939 CI->replaceAllUsesWith(Op);
10940 CI->eraseFromParent();
10944 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10945 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10946 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10948 std::string AsmStr = IA->getAsmString();
10950 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10951 SmallVector<StringRef, 4> AsmPieces;
10952 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10954 switch (AsmPieces.size()) {
10955 default: return false;
10957 AsmStr = AsmPieces[0];
10959 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10962 if (AsmPieces.size() == 2 &&
10963 (AsmPieces[0] == "bswap" ||
10964 AsmPieces[0] == "bswapq" ||
10965 AsmPieces[0] == "bswapl") &&
10966 (AsmPieces[1] == "$0" ||
10967 AsmPieces[1] == "${0:q}")) {
10968 // No need to check constraints, nothing other than the equivalent of
10969 // "=r,0" would be valid here.
10970 return LowerToBSwap(CI);
10972 // rorw $$8, ${0:w} --> llvm.bswap.i16
10973 if (CI->getType()->isIntegerTy(16) &&
10974 AsmPieces.size() == 3 &&
10975 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10976 AsmPieces[1] == "$$8," &&
10977 AsmPieces[2] == "${0:w}" &&
10978 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10980 const std::string &Constraints = IA->getConstraintString();
10981 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10982 std::sort(AsmPieces.begin(), AsmPieces.end());
10983 if (AsmPieces.size() == 4 &&
10984 AsmPieces[0] == "~{cc}" &&
10985 AsmPieces[1] == "~{dirflag}" &&
10986 AsmPieces[2] == "~{flags}" &&
10987 AsmPieces[3] == "~{fpsr}") {
10988 return LowerToBSwap(CI);
10993 if (CI->getType()->isIntegerTy(64) &&
10994 Constraints.size() >= 2 &&
10995 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10996 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10997 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10998 SmallVector<StringRef, 4> Words;
10999 SplitString(AsmPieces[0], Words, " \t");
11000 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
11002 SplitString(AsmPieces[1], Words, " \t");
11003 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
11005 SplitString(AsmPieces[2], Words, " \t,");
11006 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
11007 Words[2] == "%edx") {
11008 return LowerToBSwap(CI);
11020 /// getConstraintType - Given a constraint letter, return the type of
11021 /// constraint it is for this target.
11022 X86TargetLowering::ConstraintType
11023 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
11024 if (Constraint.size() == 1) {
11025 switch (Constraint[0]) {
11037 return C_RegisterClass;
11045 return TargetLowering::getConstraintType(Constraint);
11048 /// LowerXConstraint - try to replace an X constraint, which matches anything,
11049 /// with another that has more specific requirements based on the type of the
11050 /// corresponding operand.
11051 const char *X86TargetLowering::
11052 LowerXConstraint(EVT ConstraintVT) const {
11053 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
11054 // 'f' like normal targets.
11055 if (ConstraintVT.isFloatingPoint()) {
11056 if (Subtarget->hasSSE2())
11058 if (Subtarget->hasSSE1())
11062 return TargetLowering::LowerXConstraint(ConstraintVT);
11065 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
11066 /// vector. If it is invalid, don't add anything to Ops.
11067 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
11069 std::vector<SDValue>&Ops,
11070 SelectionDAG &DAG) const {
11071 SDValue Result(0, 0);
11073 switch (Constraint) {
11076 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11077 if (C->getZExtValue() <= 31) {
11078 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11084 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11085 if (C->getZExtValue() <= 63) {
11086 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11092 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11093 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
11094 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11100 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11101 if (C->getZExtValue() <= 255) {
11102 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11108 // 32-bit signed value
11109 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11110 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
11111 C->getSExtValue())) {
11112 // Widen to 64 bits here to get it sign extended.
11113 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
11116 // FIXME gcc accepts some relocatable values here too, but only in certain
11117 // memory models; it's complicated.
11122 // 32-bit unsigned value
11123 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11124 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
11125 C->getZExtValue())) {
11126 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11130 // FIXME gcc accepts some relocatable values here too, but only in certain
11131 // memory models; it's complicated.
11135 // Literal immediates are always ok.
11136 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
11137 // Widen to 64 bits here to get it sign extended.
11138 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
11142 // In any sort of PIC mode addresses need to be computed at runtime by
11143 // adding in a register or some sort of table lookup. These can't
11144 // be used as immediates.
11145 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
11148 // If we are in non-pic codegen mode, we allow the address of a global (with
11149 // an optional displacement) to be used with 'i'.
11150 GlobalAddressSDNode *GA = 0;
11151 int64_t Offset = 0;
11153 // Match either (GA), (GA+C), (GA+C1+C2), etc.
11155 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
11156 Offset += GA->getOffset();
11158 } else if (Op.getOpcode() == ISD::ADD) {
11159 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
11160 Offset += C->getZExtValue();
11161 Op = Op.getOperand(0);
11164 } else if (Op.getOpcode() == ISD::SUB) {
11165 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
11166 Offset += -C->getZExtValue();
11167 Op = Op.getOperand(0);
11172 // Otherwise, this isn't something we can handle, reject it.
11176 const GlobalValue *GV = GA->getGlobal();
11177 // If we require an extra load to get this address, as in PIC mode, we
11178 // can't accept it.
11179 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
11180 getTargetMachine())))
11183 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
11184 GA->getValueType(0), Offset);
11189 if (Result.getNode()) {
11190 Ops.push_back(Result);
11193 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11196 std::vector<unsigned> X86TargetLowering::
11197 getRegClassForInlineAsmConstraint(const std::string &Constraint,
11199 if (Constraint.size() == 1) {
11200 // FIXME: not handling fp-stack yet!
11201 switch (Constraint[0]) { // GCC X86 Constraint Letters
11202 default: break; // Unknown constraint letter
11203 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
11204 if (Subtarget->is64Bit()) {
11205 if (VT == MVT::i32)
11206 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
11207 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
11208 X86::R10D,X86::R11D,X86::R12D,
11209 X86::R13D,X86::R14D,X86::R15D,
11210 X86::EBP, X86::ESP, 0);
11211 else if (VT == MVT::i16)
11212 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
11213 X86::SI, X86::DI, X86::R8W,X86::R9W,
11214 X86::R10W,X86::R11W,X86::R12W,
11215 X86::R13W,X86::R14W,X86::R15W,
11216 X86::BP, X86::SP, 0);
11217 else if (VT == MVT::i8)
11218 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
11219 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
11220 X86::R10B,X86::R11B,X86::R12B,
11221 X86::R13B,X86::R14B,X86::R15B,
11222 X86::BPL, X86::SPL, 0);
11224 else if (VT == MVT::i64)
11225 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
11226 X86::RSI, X86::RDI, X86::R8, X86::R9,
11227 X86::R10, X86::R11, X86::R12,
11228 X86::R13, X86::R14, X86::R15,
11229 X86::RBP, X86::RSP, 0);
11233 // 32-bit fallthrough
11234 case 'Q': // Q_REGS
11235 if (VT == MVT::i32)
11236 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
11237 else if (VT == MVT::i16)
11238 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
11239 else if (VT == MVT::i8)
11240 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
11241 else if (VT == MVT::i64)
11242 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
11247 return std::vector<unsigned>();
11250 std::pair<unsigned, const TargetRegisterClass*>
11251 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
11253 // First, see if this is a constraint that directly corresponds to an LLVM
11255 if (Constraint.size() == 1) {
11256 // GCC Constraint Letters
11257 switch (Constraint[0]) {
11259 case 'r': // GENERAL_REGS
11260 case 'l': // INDEX_REGS
11262 return std::make_pair(0U, X86::GR8RegisterClass);
11263 if (VT == MVT::i16)
11264 return std::make_pair(0U, X86::GR16RegisterClass);
11265 if (VT == MVT::i32 || !Subtarget->is64Bit())
11266 return std::make_pair(0U, X86::GR32RegisterClass);
11267 return std::make_pair(0U, X86::GR64RegisterClass);
11268 case 'R': // LEGACY_REGS
11270 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
11271 if (VT == MVT::i16)
11272 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
11273 if (VT == MVT::i32 || !Subtarget->is64Bit())
11274 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
11275 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
11276 case 'f': // FP Stack registers.
11277 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
11278 // value to the correct fpstack register class.
11279 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
11280 return std::make_pair(0U, X86::RFP32RegisterClass);
11281 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
11282 return std::make_pair(0U, X86::RFP64RegisterClass);
11283 return std::make_pair(0U, X86::RFP80RegisterClass);
11284 case 'y': // MMX_REGS if MMX allowed.
11285 if (!Subtarget->hasMMX()) break;
11286 return std::make_pair(0U, X86::VR64RegisterClass);
11287 case 'Y': // SSE_REGS if SSE2 allowed
11288 if (!Subtarget->hasSSE2()) break;
11290 case 'x': // SSE_REGS if SSE1 allowed
11291 if (!Subtarget->hasSSE1()) break;
11293 switch (VT.getSimpleVT().SimpleTy) {
11295 // Scalar SSE types.
11298 return std::make_pair(0U, X86::FR32RegisterClass);
11301 return std::make_pair(0U, X86::FR64RegisterClass);
11309 return std::make_pair(0U, X86::VR128RegisterClass);
11315 // Use the default implementation in TargetLowering to convert the register
11316 // constraint into a member of a register class.
11317 std::pair<unsigned, const TargetRegisterClass*> Res;
11318 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
11320 // Not found as a standard register?
11321 if (Res.second == 0) {
11322 // Map st(0) -> st(7) -> ST0
11323 if (Constraint.size() == 7 && Constraint[0] == '{' &&
11324 tolower(Constraint[1]) == 's' &&
11325 tolower(Constraint[2]) == 't' &&
11326 Constraint[3] == '(' &&
11327 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
11328 Constraint[5] == ')' &&
11329 Constraint[6] == '}') {
11331 Res.first = X86::ST0+Constraint[4]-'0';
11332 Res.second = X86::RFP80RegisterClass;
11336 // GCC allows "st(0)" to be called just plain "st".
11337 if (StringRef("{st}").equals_lower(Constraint)) {
11338 Res.first = X86::ST0;
11339 Res.second = X86::RFP80RegisterClass;
11344 if (StringRef("{flags}").equals_lower(Constraint)) {
11345 Res.first = X86::EFLAGS;
11346 Res.second = X86::CCRRegisterClass;
11350 // 'A' means EAX + EDX.
11351 if (Constraint == "A") {
11352 Res.first = X86::EAX;
11353 Res.second = X86::GR32_ADRegisterClass;
11359 // Otherwise, check to see if this is a register class of the wrong value
11360 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
11361 // turn into {ax},{dx}.
11362 if (Res.second->hasType(VT))
11363 return Res; // Correct type already, nothing to do.
11365 // All of the single-register GCC register classes map their values onto
11366 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
11367 // really want an 8-bit or 32-bit register, map to the appropriate register
11368 // class and return the appropriate register.
11369 if (Res.second == X86::GR16RegisterClass) {
11370 if (VT == MVT::i8) {
11371 unsigned DestReg = 0;
11372 switch (Res.first) {
11374 case X86::AX: DestReg = X86::AL; break;
11375 case X86::DX: DestReg = X86::DL; break;
11376 case X86::CX: DestReg = X86::CL; break;
11377 case X86::BX: DestReg = X86::BL; break;
11380 Res.first = DestReg;
11381 Res.second = X86::GR8RegisterClass;
11383 } else if (VT == MVT::i32) {
11384 unsigned DestReg = 0;
11385 switch (Res.first) {
11387 case X86::AX: DestReg = X86::EAX; break;
11388 case X86::DX: DestReg = X86::EDX; break;
11389 case X86::CX: DestReg = X86::ECX; break;
11390 case X86::BX: DestReg = X86::EBX; break;
11391 case X86::SI: DestReg = X86::ESI; break;
11392 case X86::DI: DestReg = X86::EDI; break;
11393 case X86::BP: DestReg = X86::EBP; break;
11394 case X86::SP: DestReg = X86::ESP; break;
11397 Res.first = DestReg;
11398 Res.second = X86::GR32RegisterClass;
11400 } else if (VT == MVT::i64) {
11401 unsigned DestReg = 0;
11402 switch (Res.first) {
11404 case X86::AX: DestReg = X86::RAX; break;
11405 case X86::DX: DestReg = X86::RDX; break;
11406 case X86::CX: DestReg = X86::RCX; break;
11407 case X86::BX: DestReg = X86::RBX; break;
11408 case X86::SI: DestReg = X86::RSI; break;
11409 case X86::DI: DestReg = X86::RDI; break;
11410 case X86::BP: DestReg = X86::RBP; break;
11411 case X86::SP: DestReg = X86::RSP; break;
11414 Res.first = DestReg;
11415 Res.second = X86::GR64RegisterClass;
11418 } else if (Res.second == X86::FR32RegisterClass ||
11419 Res.second == X86::FR64RegisterClass ||
11420 Res.second == X86::VR128RegisterClass) {
11421 // Handle references to XMM physical registers that got mapped into the
11422 // wrong class. This can happen with constraints like {xmm0} where the
11423 // target independent register mapper will just pick the first match it can
11424 // find, ignoring the required type.
11425 if (VT == MVT::f32)
11426 Res.second = X86::FR32RegisterClass;
11427 else if (VT == MVT::f64)
11428 Res.second = X86::FR64RegisterClass;
11429 else if (X86::VR128RegisterClass->hasType(VT))
11430 Res.second = X86::VR128RegisterClass;