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::SHUFPS:
2596 case X86ISD::MOVLHPS:
2597 case X86ISD::MOVLHPD:
2598 case X86ISD::MOVHLPS:
2599 case X86ISD::MOVLPS:
2600 case X86ISD::MOVLPD:
2601 case X86ISD::MOVSHDUP:
2602 case X86ISD::MOVSLDUP:
2605 case X86ISD::UNPCKLPS:
2606 case X86ISD::UNPCKLPD:
2607 case X86ISD::PUNPCKLWD:
2608 case X86ISD::PUNPCKLBW:
2609 case X86ISD::PUNPCKLDQ:
2610 case X86ISD::PUNPCKLQDQ:
2611 case X86ISD::UNPCKHPS:
2612 case X86ISD::UNPCKHPD:
2613 case X86ISD::PUNPCKHWD:
2614 case X86ISD::PUNPCKHBW:
2615 case X86ISD::PUNPCKHDQ:
2616 case X86ISD::PUNPCKHQDQ:
2622 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2623 SDValue V1, SelectionDAG &DAG) {
2625 default: llvm_unreachable("Unknown x86 shuffle node");
2626 case X86ISD::MOVSHDUP:
2627 case X86ISD::MOVSLDUP:
2628 return DAG.getNode(Opc, dl, VT, V1);
2634 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2635 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2637 default: llvm_unreachable("Unknown x86 shuffle node");
2638 case X86ISD::PSHUFD:
2639 case X86ISD::PSHUFHW:
2640 case X86ISD::PSHUFLW:
2641 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2647 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2648 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2650 default: llvm_unreachable("Unknown x86 shuffle node");
2651 case X86ISD::SHUFPD:
2652 case X86ISD::SHUFPS:
2653 return DAG.getNode(Opc, dl, VT, V1, V2,
2654 DAG.getConstant(TargetMask, MVT::i8));
2659 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2660 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2662 default: llvm_unreachable("Unknown x86 shuffle node");
2663 case X86ISD::MOVLHPS:
2664 case X86ISD::MOVLHPD:
2665 case X86ISD::MOVHLPS:
2666 case X86ISD::MOVLPS:
2667 case X86ISD::MOVLPD:
2670 case X86ISD::UNPCKLPS:
2671 case X86ISD::UNPCKLPD:
2672 case X86ISD::PUNPCKLWD:
2673 case X86ISD::PUNPCKLBW:
2674 case X86ISD::PUNPCKLDQ:
2675 case X86ISD::PUNPCKLQDQ:
2676 case X86ISD::UNPCKHPS:
2677 case X86ISD::UNPCKHPD:
2678 case X86ISD::PUNPCKHWD:
2679 case X86ISD::PUNPCKHBW:
2680 case X86ISD::PUNPCKHDQ:
2681 case X86ISD::PUNPCKHQDQ:
2682 return DAG.getNode(Opc, dl, VT, V1, V2);
2687 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2688 MachineFunction &MF = DAG.getMachineFunction();
2689 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2690 int ReturnAddrIndex = FuncInfo->getRAIndex();
2692 if (ReturnAddrIndex == 0) {
2693 // Set up a frame object for the return address.
2694 uint64_t SlotSize = TD->getPointerSize();
2695 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2697 FuncInfo->setRAIndex(ReturnAddrIndex);
2700 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2704 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2705 bool hasSymbolicDisplacement) {
2706 // Offset should fit into 32 bit immediate field.
2707 if (!isInt<32>(Offset))
2710 // If we don't have a symbolic displacement - we don't have any extra
2712 if (!hasSymbolicDisplacement)
2715 // FIXME: Some tweaks might be needed for medium code model.
2716 if (M != CodeModel::Small && M != CodeModel::Kernel)
2719 // For small code model we assume that latest object is 16MB before end of 31
2720 // bits boundary. We may also accept pretty large negative constants knowing
2721 // that all objects are in the positive half of address space.
2722 if (M == CodeModel::Small && Offset < 16*1024*1024)
2725 // For kernel code model we know that all object resist in the negative half
2726 // of 32bits address space. We may not accept negative offsets, since they may
2727 // be just off and we may accept pretty large positive ones.
2728 if (M == CodeModel::Kernel && Offset > 0)
2734 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2735 /// specific condition code, returning the condition code and the LHS/RHS of the
2736 /// comparison to make.
2737 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2738 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2740 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2741 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2742 // X > -1 -> X == 0, jump !sign.
2743 RHS = DAG.getConstant(0, RHS.getValueType());
2744 return X86::COND_NS;
2745 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2746 // X < 0 -> X == 0, jump on sign.
2748 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2750 RHS = DAG.getConstant(0, RHS.getValueType());
2751 return X86::COND_LE;
2755 switch (SetCCOpcode) {
2756 default: llvm_unreachable("Invalid integer condition!");
2757 case ISD::SETEQ: return X86::COND_E;
2758 case ISD::SETGT: return X86::COND_G;
2759 case ISD::SETGE: return X86::COND_GE;
2760 case ISD::SETLT: return X86::COND_L;
2761 case ISD::SETLE: return X86::COND_LE;
2762 case ISD::SETNE: return X86::COND_NE;
2763 case ISD::SETULT: return X86::COND_B;
2764 case ISD::SETUGT: return X86::COND_A;
2765 case ISD::SETULE: return X86::COND_BE;
2766 case ISD::SETUGE: return X86::COND_AE;
2770 // First determine if it is required or is profitable to flip the operands.
2772 // If LHS is a foldable load, but RHS is not, flip the condition.
2773 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2774 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2775 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2776 std::swap(LHS, RHS);
2779 switch (SetCCOpcode) {
2785 std::swap(LHS, RHS);
2789 // On a floating point condition, the flags are set as follows:
2791 // 0 | 0 | 0 | X > Y
2792 // 0 | 0 | 1 | X < Y
2793 // 1 | 0 | 0 | X == Y
2794 // 1 | 1 | 1 | unordered
2795 switch (SetCCOpcode) {
2796 default: llvm_unreachable("Condcode should be pre-legalized away");
2798 case ISD::SETEQ: return X86::COND_E;
2799 case ISD::SETOLT: // flipped
2801 case ISD::SETGT: return X86::COND_A;
2802 case ISD::SETOLE: // flipped
2804 case ISD::SETGE: return X86::COND_AE;
2805 case ISD::SETUGT: // flipped
2807 case ISD::SETLT: return X86::COND_B;
2808 case ISD::SETUGE: // flipped
2810 case ISD::SETLE: return X86::COND_BE;
2812 case ISD::SETNE: return X86::COND_NE;
2813 case ISD::SETUO: return X86::COND_P;
2814 case ISD::SETO: return X86::COND_NP;
2816 case ISD::SETUNE: return X86::COND_INVALID;
2820 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2821 /// code. Current x86 isa includes the following FP cmov instructions:
2822 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2823 static bool hasFPCMov(unsigned X86CC) {
2839 /// isFPImmLegal - Returns true if the target can instruction select the
2840 /// specified FP immediate natively. If false, the legalizer will
2841 /// materialize the FP immediate as a load from a constant pool.
2842 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2843 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2844 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2850 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2851 /// the specified range (L, H].
2852 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2853 return (Val < 0) || (Val >= Low && Val < Hi);
2856 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2857 /// specified value.
2858 static bool isUndefOrEqual(int Val, int CmpVal) {
2859 if (Val < 0 || Val == CmpVal)
2864 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2865 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2866 /// the second operand.
2867 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2868 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2869 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2870 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2871 return (Mask[0] < 2 && Mask[1] < 2);
2875 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2876 SmallVector<int, 8> M;
2878 return ::isPSHUFDMask(M, N->getValueType(0));
2881 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2882 /// is suitable for input to PSHUFHW.
2883 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2884 if (VT != MVT::v8i16)
2887 // Lower quadword copied in order or undef.
2888 for (int i = 0; i != 4; ++i)
2889 if (Mask[i] >= 0 && Mask[i] != i)
2892 // Upper quadword shuffled.
2893 for (int i = 4; i != 8; ++i)
2894 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2900 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2901 SmallVector<int, 8> M;
2903 return ::isPSHUFHWMask(M, N->getValueType(0));
2906 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2907 /// is suitable for input to PSHUFLW.
2908 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2909 if (VT != MVT::v8i16)
2912 // Upper quadword copied in order.
2913 for (int i = 4; i != 8; ++i)
2914 if (Mask[i] >= 0 && Mask[i] != i)
2917 // Lower quadword shuffled.
2918 for (int i = 0; i != 4; ++i)
2925 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2926 SmallVector<int, 8> M;
2928 return ::isPSHUFLWMask(M, N->getValueType(0));
2931 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2932 /// is suitable for input to PALIGNR.
2933 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2935 int i, e = VT.getVectorNumElements();
2937 // Do not handle v2i64 / v2f64 shuffles with palignr.
2938 if (e < 4 || !hasSSSE3)
2941 for (i = 0; i != e; ++i)
2945 // All undef, not a palignr.
2949 // Determine if it's ok to perform a palignr with only the LHS, since we
2950 // don't have access to the actual shuffle elements to see if RHS is undef.
2951 bool Unary = Mask[i] < (int)e;
2952 bool NeedsUnary = false;
2954 int s = Mask[i] - i;
2956 // Check the rest of the elements to see if they are consecutive.
2957 for (++i; i != e; ++i) {
2962 Unary = Unary && (m < (int)e);
2963 NeedsUnary = NeedsUnary || (m < s);
2965 if (NeedsUnary && !Unary)
2967 if (Unary && m != ((s+i) & (e-1)))
2969 if (!Unary && m != (s+i))
2975 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2976 SmallVector<int, 8> M;
2978 return ::isPALIGNRMask(M, N->getValueType(0), true);
2981 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2982 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2983 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2984 int NumElems = VT.getVectorNumElements();
2985 if (NumElems != 2 && NumElems != 4)
2988 int Half = NumElems / 2;
2989 for (int i = 0; i < Half; ++i)
2990 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2992 for (int i = Half; i < NumElems; ++i)
2993 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2999 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3000 SmallVector<int, 8> M;
3002 return ::isSHUFPMask(M, N->getValueType(0));
3005 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3006 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3007 /// half elements to come from vector 1 (which would equal the dest.) and
3008 /// the upper half to come from vector 2.
3009 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3010 int NumElems = VT.getVectorNumElements();
3012 if (NumElems != 2 && NumElems != 4)
3015 int Half = NumElems / 2;
3016 for (int i = 0; i < Half; ++i)
3017 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3019 for (int i = Half; i < NumElems; ++i)
3020 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3025 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3026 SmallVector<int, 8> M;
3028 return isCommutedSHUFPMask(M, N->getValueType(0));
3031 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3032 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3033 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3034 if (N->getValueType(0).getVectorNumElements() != 4)
3037 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3038 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3039 isUndefOrEqual(N->getMaskElt(1), 7) &&
3040 isUndefOrEqual(N->getMaskElt(2), 2) &&
3041 isUndefOrEqual(N->getMaskElt(3), 3);
3044 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3045 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3047 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3048 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3053 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3054 isUndefOrEqual(N->getMaskElt(1), 3) &&
3055 isUndefOrEqual(N->getMaskElt(2), 2) &&
3056 isUndefOrEqual(N->getMaskElt(3), 3);
3059 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3060 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3061 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3062 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3064 if (NumElems != 2 && NumElems != 4)
3067 for (unsigned i = 0; i < NumElems/2; ++i)
3068 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3071 for (unsigned i = NumElems/2; i < NumElems; ++i)
3072 if (!isUndefOrEqual(N->getMaskElt(i), i))
3078 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3079 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3080 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3081 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3083 if (NumElems != 2 && NumElems != 4)
3086 for (unsigned i = 0; i < NumElems/2; ++i)
3087 if (!isUndefOrEqual(N->getMaskElt(i), i))
3090 for (unsigned i = 0; i < NumElems/2; ++i)
3091 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3097 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3098 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3099 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3100 bool V2IsSplat = false) {
3101 int NumElts = VT.getVectorNumElements();
3102 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3105 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3107 int BitI1 = Mask[i+1];
3108 if (!isUndefOrEqual(BitI, j))
3111 if (!isUndefOrEqual(BitI1, NumElts))
3114 if (!isUndefOrEqual(BitI1, j + NumElts))
3121 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3122 SmallVector<int, 8> M;
3124 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3127 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3128 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3129 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3130 bool V2IsSplat = false) {
3131 int NumElts = VT.getVectorNumElements();
3132 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3135 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3137 int BitI1 = Mask[i+1];
3138 if (!isUndefOrEqual(BitI, j + NumElts/2))
3141 if (isUndefOrEqual(BitI1, NumElts))
3144 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3151 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3152 SmallVector<int, 8> M;
3154 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3157 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3158 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3160 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3161 int NumElems = VT.getVectorNumElements();
3162 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3165 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3167 int BitI1 = Mask[i+1];
3168 if (!isUndefOrEqual(BitI, j))
3170 if (!isUndefOrEqual(BitI1, j))
3176 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3177 SmallVector<int, 8> M;
3179 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3182 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3183 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3185 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3186 int NumElems = VT.getVectorNumElements();
3187 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3190 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3192 int BitI1 = Mask[i+1];
3193 if (!isUndefOrEqual(BitI, j))
3195 if (!isUndefOrEqual(BitI1, j))
3201 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3202 SmallVector<int, 8> M;
3204 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3207 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3208 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3209 /// MOVSD, and MOVD, i.e. setting the lowest element.
3210 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3211 if (VT.getVectorElementType().getSizeInBits() < 32)
3214 int NumElts = VT.getVectorNumElements();
3216 if (!isUndefOrEqual(Mask[0], NumElts))
3219 for (int i = 1; i < NumElts; ++i)
3220 if (!isUndefOrEqual(Mask[i], i))
3226 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3227 SmallVector<int, 8> M;
3229 return ::isMOVLMask(M, N->getValueType(0));
3232 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3233 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3234 /// element of vector 2 and the other elements to come from vector 1 in order.
3235 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3236 bool V2IsSplat = false, bool V2IsUndef = false) {
3237 int NumOps = VT.getVectorNumElements();
3238 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3241 if (!isUndefOrEqual(Mask[0], 0))
3244 for (int i = 1; i < NumOps; ++i)
3245 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3246 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3247 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3253 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3254 bool V2IsUndef = false) {
3255 SmallVector<int, 8> M;
3257 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3260 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3261 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3262 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3263 if (N->getValueType(0).getVectorNumElements() != 4)
3266 // Expect 1, 1, 3, 3
3267 for (unsigned i = 0; i < 2; ++i) {
3268 int Elt = N->getMaskElt(i);
3269 if (Elt >= 0 && Elt != 1)
3274 for (unsigned i = 2; i < 4; ++i) {
3275 int Elt = N->getMaskElt(i);
3276 if (Elt >= 0 && Elt != 3)
3281 // Don't use movshdup if it can be done with a shufps.
3282 // FIXME: verify that matching u, u, 3, 3 is what we want.
3286 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3287 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3288 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3289 if (N->getValueType(0).getVectorNumElements() != 4)
3292 // Expect 0, 0, 2, 2
3293 for (unsigned i = 0; i < 2; ++i)
3294 if (N->getMaskElt(i) > 0)
3298 for (unsigned i = 2; i < 4; ++i) {
3299 int Elt = N->getMaskElt(i);
3300 if (Elt >= 0 && Elt != 2)
3305 // Don't use movsldup if it can be done with a shufps.
3309 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3310 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3311 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3312 int e = N->getValueType(0).getVectorNumElements() / 2;
3314 for (int i = 0; i < e; ++i)
3315 if (!isUndefOrEqual(N->getMaskElt(i), i))
3317 for (int i = 0; i < e; ++i)
3318 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3323 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3324 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3325 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3326 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3327 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3329 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3331 for (int i = 0; i < NumOperands; ++i) {
3332 int Val = SVOp->getMaskElt(NumOperands-i-1);
3333 if (Val < 0) Val = 0;
3334 if (Val >= NumOperands) Val -= NumOperands;
3336 if (i != NumOperands - 1)
3342 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3343 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3344 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3345 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3347 // 8 nodes, but we only care about the last 4.
3348 for (unsigned i = 7; i >= 4; --i) {
3349 int Val = SVOp->getMaskElt(i);
3358 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3359 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3360 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3361 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3363 // 8 nodes, but we only care about the first 4.
3364 for (int i = 3; i >= 0; --i) {
3365 int Val = SVOp->getMaskElt(i);
3374 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3375 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3376 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3377 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3378 EVT VVT = N->getValueType(0);
3379 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3383 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3384 Val = SVOp->getMaskElt(i);
3388 return (Val - i) * EltSize;
3391 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3393 bool X86::isZeroNode(SDValue Elt) {
3394 return ((isa<ConstantSDNode>(Elt) &&
3395 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3396 (isa<ConstantFPSDNode>(Elt) &&
3397 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3400 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3401 /// their permute mask.
3402 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3403 SelectionDAG &DAG) {
3404 EVT VT = SVOp->getValueType(0);
3405 unsigned NumElems = VT.getVectorNumElements();
3406 SmallVector<int, 8> MaskVec;
3408 for (unsigned i = 0; i != NumElems; ++i) {
3409 int idx = SVOp->getMaskElt(i);
3411 MaskVec.push_back(idx);
3412 else if (idx < (int)NumElems)
3413 MaskVec.push_back(idx + NumElems);
3415 MaskVec.push_back(idx - NumElems);
3417 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3418 SVOp->getOperand(0), &MaskVec[0]);
3421 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3422 /// the two vector operands have swapped position.
3423 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3424 unsigned NumElems = VT.getVectorNumElements();
3425 for (unsigned i = 0; i != NumElems; ++i) {
3429 else if (idx < (int)NumElems)
3430 Mask[i] = idx + NumElems;
3432 Mask[i] = idx - NumElems;
3436 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3437 /// match movhlps. The lower half elements should come from upper half of
3438 /// V1 (and in order), and the upper half elements should come from the upper
3439 /// half of V2 (and in order).
3440 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3441 if (Op->getValueType(0).getVectorNumElements() != 4)
3443 for (unsigned i = 0, e = 2; i != e; ++i)
3444 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3446 for (unsigned i = 2; i != 4; ++i)
3447 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3452 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3453 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3455 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3456 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3458 N = N->getOperand(0).getNode();
3459 if (!ISD::isNON_EXTLoad(N))
3462 *LD = cast<LoadSDNode>(N);
3466 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3467 /// match movlp{s|d}. The lower half elements should come from lower half of
3468 /// V1 (and in order), and the upper half elements should come from the upper
3469 /// half of V2 (and in order). And since V1 will become the source of the
3470 /// MOVLP, it must be either a vector load or a scalar load to vector.
3471 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3472 ShuffleVectorSDNode *Op) {
3473 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3475 // Is V2 is a vector load, don't do this transformation. We will try to use
3476 // load folding shufps op.
3477 if (ISD::isNON_EXTLoad(V2))
3480 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3482 if (NumElems != 2 && NumElems != 4)
3484 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3485 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3487 for (unsigned i = NumElems/2; i != NumElems; ++i)
3488 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3493 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3495 static bool isSplatVector(SDNode *N) {
3496 if (N->getOpcode() != ISD::BUILD_VECTOR)
3499 SDValue SplatValue = N->getOperand(0);
3500 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3501 if (N->getOperand(i) != SplatValue)
3506 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3507 /// to an zero vector.
3508 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3509 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3510 SDValue V1 = N->getOperand(0);
3511 SDValue V2 = N->getOperand(1);
3512 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3513 for (unsigned i = 0; i != NumElems; ++i) {
3514 int Idx = N->getMaskElt(i);
3515 if (Idx >= (int)NumElems) {
3516 unsigned Opc = V2.getOpcode();
3517 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3519 if (Opc != ISD::BUILD_VECTOR ||
3520 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3522 } else if (Idx >= 0) {
3523 unsigned Opc = V1.getOpcode();
3524 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3526 if (Opc != ISD::BUILD_VECTOR ||
3527 !X86::isZeroNode(V1.getOperand(Idx)))
3534 /// getZeroVector - Returns a vector of specified type with all zero elements.
3536 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3538 assert(VT.isVector() && "Expected a vector type");
3540 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted
3541 // to their dest type. This ensures they get CSE'd.
3543 if (VT.getSizeInBits() == 64) { // MMX
3544 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3545 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3546 } else if (VT.getSizeInBits() == 128) {
3547 if (HasSSE2) { // SSE2
3548 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3549 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3551 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3552 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3554 } else if (VT.getSizeInBits() == 256) { // AVX
3555 // 256-bit logic and arithmetic instructions in AVX are
3556 // all floating-point, no support for integer ops. Default
3557 // to emitting fp zeroed vectors then.
3558 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3559 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3560 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3562 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3565 /// getOnesVector - Returns a vector of specified type with all bits set.
3567 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3568 assert(VT.isVector() && "Expected a vector type");
3570 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3571 // type. This ensures they get CSE'd.
3572 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3574 if (VT.getSizeInBits() == 64) // MMX
3575 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3577 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3578 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3582 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3583 /// that point to V2 points to its first element.
3584 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3585 EVT VT = SVOp->getValueType(0);
3586 unsigned NumElems = VT.getVectorNumElements();
3588 bool Changed = false;
3589 SmallVector<int, 8> MaskVec;
3590 SVOp->getMask(MaskVec);
3592 for (unsigned i = 0; i != NumElems; ++i) {
3593 if (MaskVec[i] > (int)NumElems) {
3594 MaskVec[i] = NumElems;
3599 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3600 SVOp->getOperand(1), &MaskVec[0]);
3601 return SDValue(SVOp, 0);
3604 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3605 /// operation of specified width.
3606 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3608 unsigned NumElems = VT.getVectorNumElements();
3609 SmallVector<int, 8> Mask;
3610 Mask.push_back(NumElems);
3611 for (unsigned i = 1; i != NumElems; ++i)
3613 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3616 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3617 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3619 unsigned NumElems = VT.getVectorNumElements();
3620 SmallVector<int, 8> Mask;
3621 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3623 Mask.push_back(i + NumElems);
3625 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3628 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3629 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3631 unsigned NumElems = VT.getVectorNumElements();
3632 unsigned Half = NumElems/2;
3633 SmallVector<int, 8> Mask;
3634 for (unsigned i = 0; i != Half; ++i) {
3635 Mask.push_back(i + Half);
3636 Mask.push_back(i + NumElems + Half);
3638 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3641 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3642 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3643 if (SV->getValueType(0).getVectorNumElements() <= 4)
3644 return SDValue(SV, 0);
3646 EVT PVT = MVT::v4f32;
3647 EVT VT = SV->getValueType(0);
3648 DebugLoc dl = SV->getDebugLoc();
3649 SDValue V1 = SV->getOperand(0);
3650 int NumElems = VT.getVectorNumElements();
3651 int EltNo = SV->getSplatIndex();
3653 // unpack elements to the correct location
3654 while (NumElems > 4) {
3655 if (EltNo < NumElems/2) {
3656 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3658 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3659 EltNo -= NumElems/2;
3664 // Perform the splat.
3665 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3666 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3667 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3668 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3671 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3672 /// vector of zero or undef vector. This produces a shuffle where the low
3673 /// element of V2 is swizzled into the zero/undef vector, landing at element
3674 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3675 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3676 bool isZero, bool HasSSE2,
3677 SelectionDAG &DAG) {
3678 EVT VT = V2.getValueType();
3680 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3681 unsigned NumElems = VT.getVectorNumElements();
3682 SmallVector<int, 16> MaskVec;
3683 for (unsigned i = 0; i != NumElems; ++i)
3684 // If this is the insertion idx, put the low elt of V2 here.
3685 MaskVec.push_back(i == Idx ? NumElems : i);
3686 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3689 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
3690 /// element of the result of the vector shuffle.
3691 SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
3694 return SDValue(); // Limit search depth.
3696 SDValue V = SDValue(N, 0);
3697 EVT VT = V.getValueType();
3698 unsigned Opcode = V.getOpcode();
3700 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
3701 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
3702 Index = SV->getMaskElt(Index);
3705 return DAG.getUNDEF(VT.getVectorElementType());
3707 int NumElems = VT.getVectorNumElements();
3708 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
3709 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
3712 // Recurse into target specific vector shuffles to find scalars.
3713 if (isTargetShuffle(Opcode)) {
3714 int NumElems = VT.getVectorNumElements();
3715 SmallVector<unsigned, 16> ShuffleMask;
3719 case X86ISD::SHUFPS:
3720 case X86ISD::SHUFPD:
3721 ImmN = N->getOperand(N->getNumOperands()-1);
3722 DecodeSHUFPSMask(NumElems,
3723 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3726 case X86ISD::PUNPCKHBW:
3727 case X86ISD::PUNPCKHWD:
3728 case X86ISD::PUNPCKHDQ:
3729 case X86ISD::PUNPCKHQDQ:
3730 DecodePUNPCKHMask(NumElems, ShuffleMask);
3732 case X86ISD::UNPCKHPS:
3733 case X86ISD::UNPCKHPD:
3734 DecodeUNPCKHPMask(NumElems, ShuffleMask);
3736 case X86ISD::PUNPCKLBW:
3737 case X86ISD::PUNPCKLWD:
3738 case X86ISD::PUNPCKLDQ:
3739 case X86ISD::PUNPCKLQDQ:
3740 DecodePUNPCKLMask(NumElems, ShuffleMask);
3742 case X86ISD::UNPCKLPS:
3743 case X86ISD::UNPCKLPD:
3744 DecodeUNPCKLPMask(NumElems, ShuffleMask);
3746 case X86ISD::MOVHLPS:
3747 DecodeMOVHLPSMask(NumElems, ShuffleMask);
3749 case X86ISD::MOVLHPS:
3750 DecodeMOVLHPSMask(NumElems, ShuffleMask);
3752 case X86ISD::PSHUFD:
3753 ImmN = N->getOperand(N->getNumOperands()-1);
3754 DecodePSHUFMask(NumElems,
3755 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3758 case X86ISD::PSHUFHW:
3759 ImmN = N->getOperand(N->getNumOperands()-1);
3760 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3763 case X86ISD::PSHUFLW:
3764 ImmN = N->getOperand(N->getNumOperands()-1);
3765 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3769 case X86ISD::MOVSD: {
3770 // The index 0 always comes from the first element of the second source,
3771 // this is why MOVSS and MOVSD are used in the first place. The other
3772 // elements come from the other positions of the first source vector.
3773 unsigned OpNum = (Index == 0) ? 1 : 0;
3774 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
3778 assert("not implemented for target shuffle node");
3782 Index = ShuffleMask[Index];
3784 return DAG.getUNDEF(VT.getVectorElementType());
3786 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
3787 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
3791 // Actual nodes that may contain scalar elements
3792 if (Opcode == ISD::BIT_CONVERT) {
3793 V = V.getOperand(0);
3794 EVT SrcVT = V.getValueType();
3795 unsigned NumElems = VT.getVectorNumElements();
3797 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
3801 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
3802 return (Index == 0) ? V.getOperand(0)
3803 : DAG.getUNDEF(VT.getVectorElementType());
3805 if (V.getOpcode() == ISD::BUILD_VECTOR)
3806 return V.getOperand(Index);
3811 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
3812 /// shuffle operation which come from a consecutively from a zero. The
3813 /// search can start in two diferent directions, from left or right.
3815 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
3816 bool ZerosFromLeft, SelectionDAG &DAG) {
3819 while (i < NumElems) {
3820 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
3821 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
3822 if (!(Elt.getNode() &&
3823 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
3831 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
3832 /// MaskE correspond consecutively to elements from one of the vector operands,
3833 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
3835 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
3836 int OpIdx, int NumElems, unsigned &OpNum) {
3837 bool SeenV1 = false;
3838 bool SeenV2 = false;
3840 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
3841 int Idx = SVOp->getMaskElt(i);
3842 // Ignore undef indicies
3851 // Only accept consecutive elements from the same vector
3852 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
3856 OpNum = SeenV1 ? 0 : 1;
3860 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
3861 /// logical left shift of a vector.
3862 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3863 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3864 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3865 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3866 false /* check zeros from right */, DAG);
3872 // Considering the elements in the mask that are not consecutive zeros,
3873 // check if they consecutively come from only one of the source vectors.
3875 // V1 = {X, A, B, C} 0
3877 // vector_shuffle V1, V2 <1, 2, 3, X>
3879 if (!isShuffleMaskConsecutive(SVOp,
3880 0, // Mask Start Index
3881 NumElems-NumZeros-1, // Mask End Index
3882 NumZeros, // Where to start looking in the src vector
3883 NumElems, // Number of elements in vector
3884 OpSrc)) // Which source operand ?
3889 ShVal = SVOp->getOperand(OpSrc);
3893 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
3894 /// logical left shift of a vector.
3895 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3896 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3897 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3898 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3899 true /* check zeros from left */, DAG);
3905 // Considering the elements in the mask that are not consecutive zeros,
3906 // check if they consecutively come from only one of the source vectors.
3908 // 0 { A, B, X, X } = V2
3910 // vector_shuffle V1, V2 <X, X, 4, 5>
3912 if (!isShuffleMaskConsecutive(SVOp,
3913 NumZeros, // Mask Start Index
3914 NumElems-1, // Mask End Index
3915 0, // Where to start looking in the src vector
3916 NumElems, // Number of elements in vector
3917 OpSrc)) // Which source operand ?
3922 ShVal = SVOp->getOperand(OpSrc);
3926 /// isVectorShift - Returns true if the shuffle can be implemented as a
3927 /// logical left or right shift of a vector.
3928 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3929 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3930 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
3931 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
3937 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3939 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3940 unsigned NumNonZero, unsigned NumZero,
3942 const TargetLowering &TLI) {
3946 DebugLoc dl = Op.getDebugLoc();
3949 for (unsigned i = 0; i < 16; ++i) {
3950 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3951 if (ThisIsNonZero && First) {
3953 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3955 V = DAG.getUNDEF(MVT::v8i16);
3960 SDValue ThisElt(0, 0), LastElt(0, 0);
3961 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3962 if (LastIsNonZero) {
3963 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3964 MVT::i16, Op.getOperand(i-1));
3966 if (ThisIsNonZero) {
3967 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3968 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3969 ThisElt, DAG.getConstant(8, MVT::i8));
3971 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3975 if (ThisElt.getNode())
3976 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3977 DAG.getIntPtrConstant(i/2));
3981 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3984 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3986 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3987 unsigned NumNonZero, unsigned NumZero,
3989 const TargetLowering &TLI) {
3993 DebugLoc dl = Op.getDebugLoc();
3996 for (unsigned i = 0; i < 8; ++i) {
3997 bool isNonZero = (NonZeros & (1 << i)) != 0;
4001 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4003 V = DAG.getUNDEF(MVT::v8i16);
4006 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4007 MVT::v8i16, V, Op.getOperand(i),
4008 DAG.getIntPtrConstant(i));
4015 /// getVShift - Return a vector logical shift node.
4017 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4018 unsigned NumBits, SelectionDAG &DAG,
4019 const TargetLowering &TLI, DebugLoc dl) {
4020 bool isMMX = VT.getSizeInBits() == 64;
4021 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
4022 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4023 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
4024 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4025 DAG.getNode(Opc, dl, ShVT, SrcOp,
4026 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
4030 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4031 SelectionDAG &DAG) const {
4033 // Check if the scalar load can be widened into a vector load. And if
4034 // the address is "base + cst" see if the cst can be "absorbed" into
4035 // the shuffle mask.
4036 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4037 SDValue Ptr = LD->getBasePtr();
4038 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4040 EVT PVT = LD->getValueType(0);
4041 if (PVT != MVT::i32 && PVT != MVT::f32)
4046 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4047 FI = FINode->getIndex();
4049 } else if (Ptr.getOpcode() == ISD::ADD &&
4050 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
4051 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4052 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4053 Offset = Ptr.getConstantOperandVal(1);
4054 Ptr = Ptr.getOperand(0);
4059 SDValue Chain = LD->getChain();
4060 // Make sure the stack object alignment is at least 16.
4061 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4062 if (DAG.InferPtrAlignment(Ptr) < 16) {
4063 if (MFI->isFixedObjectIndex(FI)) {
4064 // Can't change the alignment. FIXME: It's possible to compute
4065 // the exact stack offset and reference FI + adjust offset instead.
4066 // If someone *really* cares about this. That's the way to implement it.
4069 MFI->setObjectAlignment(FI, 16);
4073 // (Offset % 16) must be multiple of 4. Then address is then
4074 // Ptr + (Offset & ~15).
4077 if ((Offset % 16) & 3)
4079 int64_t StartOffset = Offset & ~15;
4081 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4082 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4084 int EltNo = (Offset - StartOffset) >> 2;
4085 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4086 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4087 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
4089 // Canonicalize it to a v4i32 shuffle.
4090 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
4091 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4092 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4093 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
4099 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4100 /// vector of type 'VT', see if the elements can be replaced by a single large
4101 /// load which has the same value as a build_vector whose operands are 'elts'.
4103 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4105 /// FIXME: we'd also like to handle the case where the last elements are zero
4106 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4107 /// There's even a handy isZeroNode for that purpose.
4108 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4109 DebugLoc &dl, SelectionDAG &DAG) {
4110 EVT EltVT = VT.getVectorElementType();
4111 unsigned NumElems = Elts.size();
4113 LoadSDNode *LDBase = NULL;
4114 unsigned LastLoadedElt = -1U;
4116 // For each element in the initializer, see if we've found a load or an undef.
4117 // If we don't find an initial load element, or later load elements are
4118 // non-consecutive, bail out.
4119 for (unsigned i = 0; i < NumElems; ++i) {
4120 SDValue Elt = Elts[i];
4122 if (!Elt.getNode() ||
4123 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4126 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4128 LDBase = cast<LoadSDNode>(Elt.getNode());
4132 if (Elt.getOpcode() == ISD::UNDEF)
4135 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4136 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4141 // If we have found an entire vector of loads and undefs, then return a large
4142 // load of the entire vector width starting at the base pointer. If we found
4143 // consecutive loads for the low half, generate a vzext_load node.
4144 if (LastLoadedElt == NumElems - 1) {
4145 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4146 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
4147 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
4148 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4149 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
4150 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
4151 LDBase->isVolatile(), LDBase->isNonTemporal(),
4152 LDBase->getAlignment());
4153 } else if (NumElems == 4 && LastLoadedElt == 1) {
4154 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4155 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4156 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
4157 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
4163 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4164 DebugLoc dl = Op.getDebugLoc();
4165 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
4166 // All one's are handled with pcmpeqd. In AVX, zero's are handled with
4167 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
4168 // is present, so AllOnes is ignored.
4169 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4170 (Op.getValueType().getSizeInBits() != 256 &&
4171 ISD::isBuildVectorAllOnes(Op.getNode()))) {
4172 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
4173 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4174 // eliminated on x86-32 hosts.
4175 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
4178 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4179 return getOnesVector(Op.getValueType(), DAG, dl);
4180 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4183 EVT VT = Op.getValueType();
4184 EVT ExtVT = VT.getVectorElementType();
4185 unsigned EVTBits = ExtVT.getSizeInBits();
4187 unsigned NumElems = Op.getNumOperands();
4188 unsigned NumZero = 0;
4189 unsigned NumNonZero = 0;
4190 unsigned NonZeros = 0;
4191 bool IsAllConstants = true;
4192 SmallSet<SDValue, 8> Values;
4193 for (unsigned i = 0; i < NumElems; ++i) {
4194 SDValue Elt = Op.getOperand(i);
4195 if (Elt.getOpcode() == ISD::UNDEF)
4198 if (Elt.getOpcode() != ISD::Constant &&
4199 Elt.getOpcode() != ISD::ConstantFP)
4200 IsAllConstants = false;
4201 if (X86::isZeroNode(Elt))
4204 NonZeros |= (1 << i);
4209 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4210 if (NumNonZero == 0)
4211 return DAG.getUNDEF(VT);
4213 // Special case for single non-zero, non-undef, element.
4214 if (NumNonZero == 1) {
4215 unsigned Idx = CountTrailingZeros_32(NonZeros);
4216 SDValue Item = Op.getOperand(Idx);
4218 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4219 // the value are obviously zero, truncate the value to i32 and do the
4220 // insertion that way. Only do this if the value is non-constant or if the
4221 // value is a constant being inserted into element 0. It is cheaper to do
4222 // a constant pool load than it is to do a movd + shuffle.
4223 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4224 (!IsAllConstants || Idx == 0)) {
4225 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4226 // Handle MMX and SSE both.
4227 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
4228 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
4230 // Truncate the value (which may itself be a constant) to i32, and
4231 // convert it to a vector with movd (S2V+shuffle to zero extend).
4232 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4233 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4234 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4235 Subtarget->hasSSE2(), DAG);
4237 // Now we have our 32-bit value zero extended in the low element of
4238 // a vector. If Idx != 0, swizzle it into place.
4240 SmallVector<int, 4> Mask;
4241 Mask.push_back(Idx);
4242 for (unsigned i = 1; i != VecElts; ++i)
4244 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4245 DAG.getUNDEF(Item.getValueType()),
4248 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
4252 // If we have a constant or non-constant insertion into the low element of
4253 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4254 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4255 // depending on what the source datatype is.
4258 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4259 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4260 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4261 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4262 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4263 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4265 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4266 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4267 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
4268 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4269 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4270 Subtarget->hasSSE2(), DAG);
4271 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
4275 // Is it a vector logical left shift?
4276 if (NumElems == 2 && Idx == 1 &&
4277 X86::isZeroNode(Op.getOperand(0)) &&
4278 !X86::isZeroNode(Op.getOperand(1))) {
4279 unsigned NumBits = VT.getSizeInBits();
4280 return getVShift(true, VT,
4281 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4282 VT, Op.getOperand(1)),
4283 NumBits/2, DAG, *this, dl);
4286 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4289 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4290 // is a non-constant being inserted into an element other than the low one,
4291 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4292 // movd/movss) to move this into the low element, then shuffle it into
4294 if (EVTBits == 32) {
4295 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4297 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4298 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4299 Subtarget->hasSSE2(), DAG);
4300 SmallVector<int, 8> MaskVec;
4301 for (unsigned i = 0; i < NumElems; i++)
4302 MaskVec.push_back(i == Idx ? 0 : 1);
4303 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4307 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4308 if (Values.size() == 1) {
4309 if (EVTBits == 32) {
4310 // Instead of a shuffle like this:
4311 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4312 // Check if it's possible to issue this instead.
4313 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4314 unsigned Idx = CountTrailingZeros_32(NonZeros);
4315 SDValue Item = Op.getOperand(Idx);
4316 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4317 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4322 // A vector full of immediates; various special cases are already
4323 // handled, so this is best done with a single constant-pool load.
4327 // Let legalizer expand 2-wide build_vectors.
4328 if (EVTBits == 64) {
4329 if (NumNonZero == 1) {
4330 // One half is zero or undef.
4331 unsigned Idx = CountTrailingZeros_32(NonZeros);
4332 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4333 Op.getOperand(Idx));
4334 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4335 Subtarget->hasSSE2(), DAG);
4340 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4341 if (EVTBits == 8 && NumElems == 16) {
4342 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4344 if (V.getNode()) return V;
4347 if (EVTBits == 16 && NumElems == 8) {
4348 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4350 if (V.getNode()) return V;
4353 // If element VT is == 32 bits, turn it into a number of shuffles.
4354 SmallVector<SDValue, 8> V;
4356 if (NumElems == 4 && NumZero > 0) {
4357 for (unsigned i = 0; i < 4; ++i) {
4358 bool isZero = !(NonZeros & (1 << i));
4360 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4362 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4365 for (unsigned i = 0; i < 2; ++i) {
4366 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4369 V[i] = V[i*2]; // Must be a zero vector.
4372 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4375 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4378 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4383 SmallVector<int, 8> MaskVec;
4384 bool Reverse = (NonZeros & 0x3) == 2;
4385 for (unsigned i = 0; i < 2; ++i)
4386 MaskVec.push_back(Reverse ? 1-i : i);
4387 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4388 for (unsigned i = 0; i < 2; ++i)
4389 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4390 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4393 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4394 // Check for a build vector of consecutive loads.
4395 for (unsigned i = 0; i < NumElems; ++i)
4396 V[i] = Op.getOperand(i);
4398 // Check for elements which are consecutive loads.
4399 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4403 // For SSE 4.1, use insertps to put the high elements into the low element.
4404 if (getSubtarget()->hasSSE41()) {
4406 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4407 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4409 Result = DAG.getUNDEF(VT);
4411 for (unsigned i = 1; i < NumElems; ++i) {
4412 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4413 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4414 Op.getOperand(i), DAG.getIntPtrConstant(i));
4419 // Otherwise, expand into a number of unpckl*, start by extending each of
4420 // our (non-undef) elements to the full vector width with the element in the
4421 // bottom slot of the vector (which generates no code for SSE).
4422 for (unsigned i = 0; i < NumElems; ++i) {
4423 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4424 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4426 V[i] = DAG.getUNDEF(VT);
4429 // Next, we iteratively mix elements, e.g. for v4f32:
4430 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4431 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4432 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4433 unsigned EltStride = NumElems >> 1;
4434 while (EltStride != 0) {
4435 for (unsigned i = 0; i < EltStride; ++i) {
4436 // If V[i+EltStride] is undef and this is the first round of mixing,
4437 // then it is safe to just drop this shuffle: V[i] is already in the
4438 // right place, the one element (since it's the first round) being
4439 // inserted as undef can be dropped. This isn't safe for successive
4440 // rounds because they will permute elements within both vectors.
4441 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4442 EltStride == NumElems/2)
4445 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4455 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4456 // We support concatenate two MMX registers and place them in a MMX
4457 // register. This is better than doing a stack convert.
4458 DebugLoc dl = Op.getDebugLoc();
4459 EVT ResVT = Op.getValueType();
4460 assert(Op.getNumOperands() == 2);
4461 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4462 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4464 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4465 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4466 InVec = Op.getOperand(1);
4467 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4468 unsigned NumElts = ResVT.getVectorNumElements();
4469 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4470 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4471 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4473 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4474 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4475 Mask[0] = 0; Mask[1] = 2;
4476 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4478 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4481 // v8i16 shuffles - Prefer shuffles in the following order:
4482 // 1. [all] pshuflw, pshufhw, optional move
4483 // 2. [ssse3] 1 x pshufb
4484 // 3. [ssse3] 2 x pshufb + 1 x por
4485 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4487 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4488 SelectionDAG &DAG) const {
4489 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4490 SDValue V1 = SVOp->getOperand(0);
4491 SDValue V2 = SVOp->getOperand(1);
4492 DebugLoc dl = SVOp->getDebugLoc();
4493 SmallVector<int, 8> MaskVals;
4495 // Determine if more than 1 of the words in each of the low and high quadwords
4496 // of the result come from the same quadword of one of the two inputs. Undef
4497 // mask values count as coming from any quadword, for better codegen.
4498 SmallVector<unsigned, 4> LoQuad(4);
4499 SmallVector<unsigned, 4> HiQuad(4);
4500 BitVector InputQuads(4);
4501 for (unsigned i = 0; i < 8; ++i) {
4502 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4503 int EltIdx = SVOp->getMaskElt(i);
4504 MaskVals.push_back(EltIdx);
4513 InputQuads.set(EltIdx / 4);
4516 int BestLoQuad = -1;
4517 unsigned MaxQuad = 1;
4518 for (unsigned i = 0; i < 4; ++i) {
4519 if (LoQuad[i] > MaxQuad) {
4521 MaxQuad = LoQuad[i];
4525 int BestHiQuad = -1;
4527 for (unsigned i = 0; i < 4; ++i) {
4528 if (HiQuad[i] > MaxQuad) {
4530 MaxQuad = HiQuad[i];
4534 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4535 // of the two input vectors, shuffle them into one input vector so only a
4536 // single pshufb instruction is necessary. If There are more than 2 input
4537 // quads, disable the next transformation since it does not help SSSE3.
4538 bool V1Used = InputQuads[0] || InputQuads[1];
4539 bool V2Used = InputQuads[2] || InputQuads[3];
4540 if (Subtarget->hasSSSE3()) {
4541 if (InputQuads.count() == 2 && V1Used && V2Used) {
4542 BestLoQuad = InputQuads.find_first();
4543 BestHiQuad = InputQuads.find_next(BestLoQuad);
4545 if (InputQuads.count() > 2) {
4551 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4552 // the shuffle mask. If a quad is scored as -1, that means that it contains
4553 // words from all 4 input quadwords.
4555 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4556 SmallVector<int, 8> MaskV;
4557 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4558 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4559 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4560 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4561 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4562 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4564 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4565 // source words for the shuffle, to aid later transformations.
4566 bool AllWordsInNewV = true;
4567 bool InOrder[2] = { true, true };
4568 for (unsigned i = 0; i != 8; ++i) {
4569 int idx = MaskVals[i];
4571 InOrder[i/4] = false;
4572 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4574 AllWordsInNewV = false;
4578 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4579 if (AllWordsInNewV) {
4580 for (int i = 0; i != 8; ++i) {
4581 int idx = MaskVals[i];
4584 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4585 if ((idx != i) && idx < 4)
4587 if ((idx != i) && idx > 3)
4596 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4597 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4598 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4599 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4600 unsigned TargetMask = 0;
4601 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4602 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4603 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4604 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4605 V1 = NewV.getOperand(0);
4606 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4610 // If we have SSSE3, and all words of the result are from 1 input vector,
4611 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4612 // is present, fall back to case 4.
4613 if (Subtarget->hasSSSE3()) {
4614 SmallVector<SDValue,16> pshufbMask;
4616 // If we have elements from both input vectors, set the high bit of the
4617 // shuffle mask element to zero out elements that come from V2 in the V1
4618 // mask, and elements that come from V1 in the V2 mask, so that the two
4619 // results can be OR'd together.
4620 bool TwoInputs = V1Used && V2Used;
4621 for (unsigned i = 0; i != 8; ++i) {
4622 int EltIdx = MaskVals[i] * 2;
4623 if (TwoInputs && (EltIdx >= 16)) {
4624 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4625 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4628 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4629 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4631 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4632 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4633 DAG.getNode(ISD::BUILD_VECTOR, dl,
4634 MVT::v16i8, &pshufbMask[0], 16));
4636 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4638 // Calculate the shuffle mask for the second input, shuffle it, and
4639 // OR it with the first shuffled input.
4641 for (unsigned i = 0; i != 8; ++i) {
4642 int EltIdx = MaskVals[i] * 2;
4644 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4645 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4648 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4649 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4651 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4652 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4653 DAG.getNode(ISD::BUILD_VECTOR, dl,
4654 MVT::v16i8, &pshufbMask[0], 16));
4655 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4656 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4659 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4660 // and update MaskVals with new element order.
4661 BitVector InOrder(8);
4662 if (BestLoQuad >= 0) {
4663 SmallVector<int, 8> MaskV;
4664 for (int i = 0; i != 4; ++i) {
4665 int idx = MaskVals[i];
4667 MaskV.push_back(-1);
4669 } else if ((idx / 4) == BestLoQuad) {
4670 MaskV.push_back(idx & 3);
4673 MaskV.push_back(-1);
4676 for (unsigned i = 4; i != 8; ++i)
4678 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4681 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4682 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4684 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4688 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4689 // and update MaskVals with the new element order.
4690 if (BestHiQuad >= 0) {
4691 SmallVector<int, 8> MaskV;
4692 for (unsigned i = 0; i != 4; ++i)
4694 for (unsigned i = 4; i != 8; ++i) {
4695 int idx = MaskVals[i];
4697 MaskV.push_back(-1);
4699 } else if ((idx / 4) == BestHiQuad) {
4700 MaskV.push_back((idx & 3) + 4);
4703 MaskV.push_back(-1);
4706 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4709 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4710 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
4712 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
4716 // In case BestHi & BestLo were both -1, which means each quadword has a word
4717 // from each of the four input quadwords, calculate the InOrder bitvector now
4718 // before falling through to the insert/extract cleanup.
4719 if (BestLoQuad == -1 && BestHiQuad == -1) {
4721 for (int i = 0; i != 8; ++i)
4722 if (MaskVals[i] < 0 || MaskVals[i] == i)
4726 // The other elements are put in the right place using pextrw and pinsrw.
4727 for (unsigned i = 0; i != 8; ++i) {
4730 int EltIdx = MaskVals[i];
4733 SDValue ExtOp = (EltIdx < 8)
4734 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4735 DAG.getIntPtrConstant(EltIdx))
4736 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4737 DAG.getIntPtrConstant(EltIdx - 8));
4738 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4739 DAG.getIntPtrConstant(i));
4744 // v16i8 shuffles - Prefer shuffles in the following order:
4745 // 1. [ssse3] 1 x pshufb
4746 // 2. [ssse3] 2 x pshufb + 1 x por
4747 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4749 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4751 const X86TargetLowering &TLI) {
4752 SDValue V1 = SVOp->getOperand(0);
4753 SDValue V2 = SVOp->getOperand(1);
4754 DebugLoc dl = SVOp->getDebugLoc();
4755 SmallVector<int, 16> MaskVals;
4756 SVOp->getMask(MaskVals);
4758 // If we have SSSE3, case 1 is generated when all result bytes come from
4759 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4760 // present, fall back to case 3.
4761 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4764 for (unsigned i = 0; i < 16; ++i) {
4765 int EltIdx = MaskVals[i];
4774 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4775 if (TLI.getSubtarget()->hasSSSE3()) {
4776 SmallVector<SDValue,16> pshufbMask;
4778 // If all result elements are from one input vector, then only translate
4779 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4781 // Otherwise, we have elements from both input vectors, and must zero out
4782 // elements that come from V2 in the first mask, and V1 in the second mask
4783 // so that we can OR them together.
4784 bool TwoInputs = !(V1Only || V2Only);
4785 for (unsigned i = 0; i != 16; ++i) {
4786 int EltIdx = MaskVals[i];
4787 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4788 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4791 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4793 // If all the elements are from V2, assign it to V1 and return after
4794 // building the first pshufb.
4797 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4798 DAG.getNode(ISD::BUILD_VECTOR, dl,
4799 MVT::v16i8, &pshufbMask[0], 16));
4803 // Calculate the shuffle mask for the second input, shuffle it, and
4804 // OR it with the first shuffled input.
4806 for (unsigned i = 0; i != 16; ++i) {
4807 int EltIdx = MaskVals[i];
4809 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4812 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4814 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4815 DAG.getNode(ISD::BUILD_VECTOR, dl,
4816 MVT::v16i8, &pshufbMask[0], 16));
4817 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4820 // No SSSE3 - Calculate in place words and then fix all out of place words
4821 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4822 // the 16 different words that comprise the two doublequadword input vectors.
4823 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4824 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4825 SDValue NewV = V2Only ? V2 : V1;
4826 for (int i = 0; i != 8; ++i) {
4827 int Elt0 = MaskVals[i*2];
4828 int Elt1 = MaskVals[i*2+1];
4830 // This word of the result is all undef, skip it.
4831 if (Elt0 < 0 && Elt1 < 0)
4834 // This word of the result is already in the correct place, skip it.
4835 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4837 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4840 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4841 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4844 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4845 // using a single extract together, load it and store it.
4846 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4847 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4848 DAG.getIntPtrConstant(Elt1 / 2));
4849 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4850 DAG.getIntPtrConstant(i));
4854 // If Elt1 is defined, extract it from the appropriate source. If the
4855 // source byte is not also odd, shift the extracted word left 8 bits
4856 // otherwise clear the bottom 8 bits if we need to do an or.
4858 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4859 DAG.getIntPtrConstant(Elt1 / 2));
4860 if ((Elt1 & 1) == 0)
4861 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4862 DAG.getConstant(8, TLI.getShiftAmountTy()));
4864 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4865 DAG.getConstant(0xFF00, MVT::i16));
4867 // If Elt0 is defined, extract it from the appropriate source. If the
4868 // source byte is not also even, shift the extracted word right 8 bits. If
4869 // Elt1 was also defined, OR the extracted values together before
4870 // inserting them in the result.
4872 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4873 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4874 if ((Elt0 & 1) != 0)
4875 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4876 DAG.getConstant(8, TLI.getShiftAmountTy()));
4878 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4879 DAG.getConstant(0x00FF, MVT::i16));
4880 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4883 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4884 DAG.getIntPtrConstant(i));
4886 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4889 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4890 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4891 /// done when every pair / quad of shuffle mask elements point to elements in
4892 /// the right sequence. e.g.
4893 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4895 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4897 const TargetLowering &TLI, DebugLoc dl) {
4898 EVT VT = SVOp->getValueType(0);
4899 SDValue V1 = SVOp->getOperand(0);
4900 SDValue V2 = SVOp->getOperand(1);
4901 unsigned NumElems = VT.getVectorNumElements();
4902 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4903 EVT MaskVT = (NewWidth == 4) ? MVT::v4i16 : MVT::v2i32;
4905 switch (VT.getSimpleVT().SimpleTy) {
4906 default: assert(false && "Unexpected!");
4907 case MVT::v4f32: NewVT = MVT::v2f64; break;
4908 case MVT::v4i32: NewVT = MVT::v2i64; break;
4909 case MVT::v8i16: NewVT = MVT::v4i32; break;
4910 case MVT::v16i8: NewVT = MVT::v4i32; break;
4913 if (NewWidth == 2) {
4919 int Scale = NumElems / NewWidth;
4920 SmallVector<int, 8> MaskVec;
4921 for (unsigned i = 0; i < NumElems; i += Scale) {
4923 for (int j = 0; j < Scale; ++j) {
4924 int EltIdx = SVOp->getMaskElt(i+j);
4928 StartIdx = EltIdx - (EltIdx % Scale);
4929 if (EltIdx != StartIdx + j)
4933 MaskVec.push_back(-1);
4935 MaskVec.push_back(StartIdx / Scale);
4938 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4939 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4940 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4943 /// getVZextMovL - Return a zero-extending vector move low node.
4945 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4946 SDValue SrcOp, SelectionDAG &DAG,
4947 const X86Subtarget *Subtarget, DebugLoc dl) {
4948 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4949 LoadSDNode *LD = NULL;
4950 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4951 LD = dyn_cast<LoadSDNode>(SrcOp);
4953 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4955 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4956 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4957 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4958 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4959 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4961 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4962 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4963 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4964 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4972 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4973 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4974 DAG.getNode(ISD::BIT_CONVERT, dl,
4978 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4981 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4982 SDValue V1 = SVOp->getOperand(0);
4983 SDValue V2 = SVOp->getOperand(1);
4984 DebugLoc dl = SVOp->getDebugLoc();
4985 EVT VT = SVOp->getValueType(0);
4987 SmallVector<std::pair<int, int>, 8> Locs;
4989 SmallVector<int, 8> Mask1(4U, -1);
4990 SmallVector<int, 8> PermMask;
4991 SVOp->getMask(PermMask);
4995 for (unsigned i = 0; i != 4; ++i) {
4996 int Idx = PermMask[i];
4998 Locs[i] = std::make_pair(-1, -1);
5000 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5002 Locs[i] = std::make_pair(0, NumLo);
5006 Locs[i] = std::make_pair(1, NumHi);
5008 Mask1[2+NumHi] = Idx;
5014 if (NumLo <= 2 && NumHi <= 2) {
5015 // If no more than two elements come from either vector. This can be
5016 // implemented with two shuffles. First shuffle gather the elements.
5017 // The second shuffle, which takes the first shuffle as both of its
5018 // vector operands, put the elements into the right order.
5019 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5021 SmallVector<int, 8> Mask2(4U, -1);
5023 for (unsigned i = 0; i != 4; ++i) {
5024 if (Locs[i].first == -1)
5027 unsigned Idx = (i < 2) ? 0 : 4;
5028 Idx += Locs[i].first * 2 + Locs[i].second;
5033 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5034 } else if (NumLo == 3 || NumHi == 3) {
5035 // Otherwise, we must have three elements from one vector, call it X, and
5036 // one element from the other, call it Y. First, use a shufps to build an
5037 // intermediate vector with the one element from Y and the element from X
5038 // that will be in the same half in the final destination (the indexes don't
5039 // matter). Then, use a shufps to build the final vector, taking the half
5040 // containing the element from Y from the intermediate, and the other half
5043 // Normalize it so the 3 elements come from V1.
5044 CommuteVectorShuffleMask(PermMask, VT);
5048 // Find the element from V2.
5050 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5051 int Val = PermMask[HiIndex];
5058 Mask1[0] = PermMask[HiIndex];
5060 Mask1[2] = PermMask[HiIndex^1];
5062 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5065 Mask1[0] = PermMask[0];
5066 Mask1[1] = PermMask[1];
5067 Mask1[2] = HiIndex & 1 ? 6 : 4;
5068 Mask1[3] = HiIndex & 1 ? 4 : 6;
5069 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5071 Mask1[0] = HiIndex & 1 ? 2 : 0;
5072 Mask1[1] = HiIndex & 1 ? 0 : 2;
5073 Mask1[2] = PermMask[2];
5074 Mask1[3] = PermMask[3];
5079 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5083 // Break it into (shuffle shuffle_hi, shuffle_lo).
5085 SmallVector<int,8> LoMask(4U, -1);
5086 SmallVector<int,8> HiMask(4U, -1);
5088 SmallVector<int,8> *MaskPtr = &LoMask;
5089 unsigned MaskIdx = 0;
5092 for (unsigned i = 0; i != 4; ++i) {
5099 int Idx = PermMask[i];
5101 Locs[i] = std::make_pair(-1, -1);
5102 } else if (Idx < 4) {
5103 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5104 (*MaskPtr)[LoIdx] = Idx;
5107 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5108 (*MaskPtr)[HiIdx] = Idx;
5113 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5114 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5115 SmallVector<int, 8> MaskOps;
5116 for (unsigned i = 0; i != 4; ++i) {
5117 if (Locs[i].first == -1) {
5118 MaskOps.push_back(-1);
5120 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5121 MaskOps.push_back(Idx);
5124 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5127 static bool MayFoldVectorLoad(SDValue V) {
5128 if (V.hasOneUse() && V.getOpcode() == ISD::BIT_CONVERT)
5129 V = V.getOperand(0);
5130 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5131 V = V.getOperand(0);
5138 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5140 SDValue V1 = Op.getOperand(0);
5141 SDValue V2 = Op.getOperand(1);
5142 EVT VT = Op.getValueType();
5144 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5146 if (HasSSE2 && VT == MVT::v2f64)
5147 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5150 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5154 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5155 SDValue V1 = Op.getOperand(0);
5156 SDValue V2 = Op.getOperand(1);
5157 EVT VT = Op.getValueType();
5159 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5160 "unsupported shuffle type");
5162 if (V2.getOpcode() == ISD::UNDEF)
5166 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5170 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5171 SDValue V1 = Op.getOperand(0);
5172 SDValue V2 = Op.getOperand(1);
5173 EVT VT = Op.getValueType();
5174 unsigned NumElems = VT.getVectorNumElements();
5176 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5177 // operand of these instructions is only memory, so check if there's a
5178 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5180 bool CanFoldLoad = false;
5182 // Trivial case, when V2 comes from a load.
5183 if (MayFoldVectorLoad(V2))
5186 // When V1 is a load, it can be folded later into a store in isel, example:
5187 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5189 // (MOVLPSmr addr:$src1, VR128:$src2)
5190 // So, recognize this potential and also use MOVLPS or MOVLPD
5191 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5195 if (HasSSE2 && NumElems == 2)
5196 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5199 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5202 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5203 // movl and movlp will both match v2i64, but v2i64 is never matched by
5204 // movl earlier because we make it strict to avoid messing with the movlp load
5205 // folding logic (see the code above getMOVLP call). Match it here then,
5206 // this is horrible, but will stay like this until we move all shuffle
5207 // matching to x86 specific nodes. Note that for the 1st condition all
5208 // types are matched with movsd.
5209 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5210 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5212 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5215 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5217 // Invert the operand order and use SHUFPS to match it.
5218 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5219 X86::getShuffleSHUFImmediate(SVOp), DAG);
5222 static inline unsigned getUNPCKLOpcode(EVT VT) {
5223 switch(VT.getSimpleVT().SimpleTy) {
5224 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5225 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5226 case MVT::v4f32: return X86ISD::UNPCKLPS;
5227 case MVT::v2f64: return X86ISD::UNPCKLPD;
5228 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5229 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5231 llvm_unreachable("Unknow type for unpckl");
5236 static inline unsigned getUNPCKHOpcode(EVT VT) {
5237 switch(VT.getSimpleVT().SimpleTy) {
5238 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5239 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5240 case MVT::v4f32: return X86ISD::UNPCKHPS;
5241 case MVT::v2f64: return X86ISD::UNPCKHPD;
5242 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5243 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5245 llvm_unreachable("Unknow type for unpckh");
5251 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5252 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5253 SDValue V1 = Op.getOperand(0);
5254 SDValue V2 = Op.getOperand(1);
5255 EVT VT = Op.getValueType();
5256 DebugLoc dl = Op.getDebugLoc();
5257 unsigned NumElems = VT.getVectorNumElements();
5258 bool isMMX = VT.getSizeInBits() == 64;
5259 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5260 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5261 bool V1IsSplat = false;
5262 bool V2IsSplat = false;
5263 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5264 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5265 MachineFunction &MF = DAG.getMachineFunction();
5266 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5268 if (isZeroShuffle(SVOp))
5269 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5271 // Promote splats to v4f32.
5272 if (SVOp->isSplat()) {
5273 if (isMMX || NumElems < 4)
5275 return PromoteSplat(SVOp, DAG);
5278 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5280 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5281 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5282 if (NewOp.getNode())
5283 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5284 LowerVECTOR_SHUFFLE(NewOp, DAG));
5285 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5286 // FIXME: Figure out a cleaner way to do this.
5287 // Try to make use of movq to zero out the top part.
5288 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5289 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5290 if (NewOp.getNode()) {
5291 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5292 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5293 DAG, Subtarget, dl);
5295 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5296 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5297 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5298 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5299 DAG, Subtarget, dl);
5303 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
5304 // unpckh_undef). Only use pshufd if speed is more important than size.
5305 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
5306 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5307 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5308 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
5309 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5310 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5312 if (X86::isPSHUFDMask(SVOp)) {
5313 // The actual implementation will match the mask in the if above and then
5314 // during isel it can match several different instructions, not only pshufd
5315 // as its name says, sad but true, emulate the behavior for now...
5316 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5317 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5319 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5321 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5322 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5324 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5325 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5328 if (VT == MVT::v4f32)
5329 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5333 // Check if this can be converted into a logical shift.
5334 bool isLeft = false;
5337 bool isShift = getSubtarget()->hasSSE2() &&
5338 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5339 if (isShift && ShVal.hasOneUse()) {
5340 // If the shifted value has multiple uses, it may be cheaper to use
5341 // v_set0 + movlhps or movhlps, etc.
5342 EVT EltVT = VT.getVectorElementType();
5343 ShAmt *= EltVT.getSizeInBits();
5344 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5347 if (X86::isMOVLMask(SVOp)) {
5350 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5351 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5352 if (!isMMX && !X86::isMOVLPMask(SVOp)) {
5353 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5354 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5356 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5357 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5361 // FIXME: fold these into legal mask.
5363 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5364 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5366 if (X86::isMOVHLPSMask(SVOp))
5367 return getMOVHighToLow(Op, dl, DAG);
5369 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5370 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5372 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5373 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5375 if (X86::isMOVLPMask(SVOp))
5376 return getMOVLP(Op, dl, DAG, HasSSE2);
5379 if (ShouldXformToMOVHLPS(SVOp) ||
5380 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5381 return CommuteVectorShuffle(SVOp, DAG);
5384 // No better options. Use a vshl / vsrl.
5385 EVT EltVT = VT.getVectorElementType();
5386 ShAmt *= EltVT.getSizeInBits();
5387 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5390 bool Commuted = false;
5391 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5392 // 1,1,1,1 -> v8i16 though.
5393 V1IsSplat = isSplatVector(V1.getNode());
5394 V2IsSplat = isSplatVector(V2.getNode());
5396 // Canonicalize the splat or undef, if present, to be on the RHS.
5397 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5398 Op = CommuteVectorShuffle(SVOp, DAG);
5399 SVOp = cast<ShuffleVectorSDNode>(Op);
5400 V1 = SVOp->getOperand(0);
5401 V2 = SVOp->getOperand(1);
5402 std::swap(V1IsSplat, V2IsSplat);
5403 std::swap(V1IsUndef, V2IsUndef);
5407 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5408 // Shuffling low element of v1 into undef, just return v1.
5411 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5412 // the instruction selector will not match, so get a canonical MOVL with
5413 // swapped operands to undo the commute.
5414 return getMOVL(DAG, dl, VT, V2, V1);
5417 if (X86::isUNPCKLMask(SVOp))
5419 Op : getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
5421 if (X86::isUNPCKHMask(SVOp))
5423 Op : getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
5426 // Normalize mask so all entries that point to V2 points to its first
5427 // element then try to match unpck{h|l} again. If match, return a
5428 // new vector_shuffle with the corrected mask.
5429 SDValue NewMask = NormalizeMask(SVOp, DAG);
5430 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5431 if (NSVOp != SVOp) {
5432 if (X86::isUNPCKLMask(NSVOp, true)) {
5434 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5441 // Commute is back and try unpck* again.
5442 // FIXME: this seems wrong.
5443 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5444 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5446 if (X86::isUNPCKLMask(NewSVOp))
5448 NewOp : getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
5450 if (X86::isUNPCKHMask(NewSVOp))
5452 NewOp : getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
5455 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
5457 // Normalize the node to match x86 shuffle ops if needed
5458 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5459 return CommuteVectorShuffle(SVOp, DAG);
5461 // The checks below are all present in isShuffleMaskLegal, but they are
5462 // inlined here right now to enable us to directly emit target specific
5463 // nodes, and remove one by one until they don't return Op anymore.
5464 SmallVector<int, 16> M;
5467 // Very little shuffling can be done for 64-bit vectors right now.
5468 if (VT.getSizeInBits() == 64)
5469 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ? Op : SDValue();
5471 // FIXME: pshufb, blends, shifts.
5472 if (VT.getVectorNumElements() == 2 ||
5473 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()))
5476 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
5477 SVOp->getSplatIndex() == 0 && V2IsUndef) {
5478 if (VT == MVT::v2f64)
5479 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
5480 if (VT == MVT::v2i64)
5481 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
5484 if (isPSHUFHWMask(M, VT))
5485 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
5486 X86::getShufflePSHUFHWImmediate(SVOp),
5489 if (isPSHUFLWMask(M, VT))
5490 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
5491 X86::getShufflePSHUFLWImmediate(SVOp),
5494 if (isSHUFPMask(M, VT)) {
5495 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5496 if (VT == MVT::v4f32 || VT == MVT::v4i32)
5497 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
5499 if (VT == MVT::v2f64 || VT == MVT::v2i64)
5500 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
5504 if (X86::isUNPCKL_v_undef_Mask(SVOp))
5505 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5506 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5507 if (X86::isUNPCKH_v_undef_Mask(SVOp))
5508 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5509 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5511 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5512 if (VT == MVT::v8i16) {
5513 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5514 if (NewOp.getNode())
5518 if (VT == MVT::v16i8) {
5519 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5520 if (NewOp.getNode())
5524 // Handle all 4 wide cases with a number of shuffles except for MMX.
5525 if (NumElems == 4 && !isMMX)
5526 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5532 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5533 SelectionDAG &DAG) const {
5534 EVT VT = Op.getValueType();
5535 DebugLoc dl = Op.getDebugLoc();
5536 if (VT.getSizeInBits() == 8) {
5537 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
5538 Op.getOperand(0), Op.getOperand(1));
5539 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5540 DAG.getValueType(VT));
5541 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5542 } else if (VT.getSizeInBits() == 16) {
5543 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5544 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
5546 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5547 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5548 DAG.getNode(ISD::BIT_CONVERT, dl,
5552 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
5553 Op.getOperand(0), Op.getOperand(1));
5554 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5555 DAG.getValueType(VT));
5556 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5557 } else if (VT == MVT::f32) {
5558 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
5559 // the result back to FR32 register. It's only worth matching if the
5560 // result has a single use which is a store or a bitcast to i32. And in
5561 // the case of a store, it's not worth it if the index is a constant 0,
5562 // because a MOVSSmr can be used instead, which is smaller and faster.
5563 if (!Op.hasOneUse())
5565 SDNode *User = *Op.getNode()->use_begin();
5566 if ((User->getOpcode() != ISD::STORE ||
5567 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5568 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5569 (User->getOpcode() != ISD::BIT_CONVERT ||
5570 User->getValueType(0) != MVT::i32))
5572 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5573 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
5576 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
5577 } else if (VT == MVT::i32) {
5578 // ExtractPS works with constant index.
5579 if (isa<ConstantSDNode>(Op.getOperand(1)))
5587 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5588 SelectionDAG &DAG) const {
5589 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5592 if (Subtarget->hasSSE41()) {
5593 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5598 EVT VT = Op.getValueType();
5599 DebugLoc dl = Op.getDebugLoc();
5600 // TODO: handle v16i8.
5601 if (VT.getSizeInBits() == 16) {
5602 SDValue Vec = Op.getOperand(0);
5603 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5605 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5606 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5607 DAG.getNode(ISD::BIT_CONVERT, dl,
5610 // Transform it so it match pextrw which produces a 32-bit result.
5611 EVT EltVT = MVT::i32;
5612 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5613 Op.getOperand(0), Op.getOperand(1));
5614 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5615 DAG.getValueType(VT));
5616 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5617 } else if (VT.getSizeInBits() == 32) {
5618 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5622 // SHUFPS the element to the lowest double word, then movss.
5623 int Mask[4] = { Idx, -1, -1, -1 };
5624 EVT VVT = Op.getOperand(0).getValueType();
5625 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5626 DAG.getUNDEF(VVT), Mask);
5627 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5628 DAG.getIntPtrConstant(0));
5629 } else if (VT.getSizeInBits() == 64) {
5630 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5631 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5632 // to match extract_elt for f64.
5633 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5637 // UNPCKHPD the element to the lowest double word, then movsd.
5638 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5639 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5640 int Mask[2] = { 1, -1 };
5641 EVT VVT = Op.getOperand(0).getValueType();
5642 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5643 DAG.getUNDEF(VVT), Mask);
5644 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5645 DAG.getIntPtrConstant(0));
5652 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5653 SelectionDAG &DAG) const {
5654 EVT VT = Op.getValueType();
5655 EVT EltVT = VT.getVectorElementType();
5656 DebugLoc dl = Op.getDebugLoc();
5658 SDValue N0 = Op.getOperand(0);
5659 SDValue N1 = Op.getOperand(1);
5660 SDValue N2 = Op.getOperand(2);
5662 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5663 isa<ConstantSDNode>(N2)) {
5665 if (VT == MVT::v8i16)
5666 Opc = X86ISD::PINSRW;
5667 else if (VT == MVT::v4i16)
5668 Opc = X86ISD::MMX_PINSRW;
5669 else if (VT == MVT::v16i8)
5670 Opc = X86ISD::PINSRB;
5672 Opc = X86ISD::PINSRB;
5674 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5676 if (N1.getValueType() != MVT::i32)
5677 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5678 if (N2.getValueType() != MVT::i32)
5679 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5680 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5681 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5682 // Bits [7:6] of the constant are the source select. This will always be
5683 // zero here. The DAG Combiner may combine an extract_elt index into these
5684 // bits. For example (insert (extract, 3), 2) could be matched by putting
5685 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5686 // Bits [5:4] of the constant are the destination select. This is the
5687 // value of the incoming immediate.
5688 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5689 // combine either bitwise AND or insert of float 0.0 to set these bits.
5690 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5691 // Create this as a scalar to vector..
5692 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5693 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5694 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5695 // PINSR* works with constant index.
5702 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5703 EVT VT = Op.getValueType();
5704 EVT EltVT = VT.getVectorElementType();
5706 if (Subtarget->hasSSE41())
5707 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5709 if (EltVT == MVT::i8)
5712 DebugLoc dl = Op.getDebugLoc();
5713 SDValue N0 = Op.getOperand(0);
5714 SDValue N1 = Op.getOperand(1);
5715 SDValue N2 = Op.getOperand(2);
5717 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5718 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5719 // as its second argument.
5720 if (N1.getValueType() != MVT::i32)
5721 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5722 if (N2.getValueType() != MVT::i32)
5723 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5724 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5725 dl, VT, N0, N1, N2);
5731 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5732 DebugLoc dl = Op.getDebugLoc();
5734 if (Op.getValueType() == MVT::v1i64 &&
5735 Op.getOperand(0).getValueType() == MVT::i64)
5736 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5738 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5739 EVT VT = MVT::v2i32;
5740 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5747 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5748 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5751 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5752 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5753 // one of the above mentioned nodes. It has to be wrapped because otherwise
5754 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5755 // be used to form addressing mode. These wrapped nodes will be selected
5758 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5759 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5761 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5763 unsigned char OpFlag = 0;
5764 unsigned WrapperKind = X86ISD::Wrapper;
5765 CodeModel::Model M = getTargetMachine().getCodeModel();
5767 if (Subtarget->isPICStyleRIPRel() &&
5768 (M == CodeModel::Small || M == CodeModel::Kernel))
5769 WrapperKind = X86ISD::WrapperRIP;
5770 else if (Subtarget->isPICStyleGOT())
5771 OpFlag = X86II::MO_GOTOFF;
5772 else if (Subtarget->isPICStyleStubPIC())
5773 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5775 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5777 CP->getOffset(), OpFlag);
5778 DebugLoc DL = CP->getDebugLoc();
5779 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5780 // With PIC, the address is actually $g + Offset.
5782 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5783 DAG.getNode(X86ISD::GlobalBaseReg,
5784 DebugLoc(), getPointerTy()),
5791 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5792 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5794 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5796 unsigned char OpFlag = 0;
5797 unsigned WrapperKind = X86ISD::Wrapper;
5798 CodeModel::Model M = getTargetMachine().getCodeModel();
5800 if (Subtarget->isPICStyleRIPRel() &&
5801 (M == CodeModel::Small || M == CodeModel::Kernel))
5802 WrapperKind = X86ISD::WrapperRIP;
5803 else if (Subtarget->isPICStyleGOT())
5804 OpFlag = X86II::MO_GOTOFF;
5805 else if (Subtarget->isPICStyleStubPIC())
5806 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5808 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5810 DebugLoc DL = JT->getDebugLoc();
5811 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5813 // With PIC, the address is actually $g + Offset.
5815 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5816 DAG.getNode(X86ISD::GlobalBaseReg,
5817 DebugLoc(), getPointerTy()),
5825 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5826 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5828 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5830 unsigned char OpFlag = 0;
5831 unsigned WrapperKind = X86ISD::Wrapper;
5832 CodeModel::Model M = getTargetMachine().getCodeModel();
5834 if (Subtarget->isPICStyleRIPRel() &&
5835 (M == CodeModel::Small || M == CodeModel::Kernel))
5836 WrapperKind = X86ISD::WrapperRIP;
5837 else if (Subtarget->isPICStyleGOT())
5838 OpFlag = X86II::MO_GOTOFF;
5839 else if (Subtarget->isPICStyleStubPIC())
5840 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5842 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5844 DebugLoc DL = Op.getDebugLoc();
5845 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5848 // With PIC, the address is actually $g + Offset.
5849 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5850 !Subtarget->is64Bit()) {
5851 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5852 DAG.getNode(X86ISD::GlobalBaseReg,
5853 DebugLoc(), getPointerTy()),
5861 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5862 // Create the TargetBlockAddressAddress node.
5863 unsigned char OpFlags =
5864 Subtarget->ClassifyBlockAddressReference();
5865 CodeModel::Model M = getTargetMachine().getCodeModel();
5866 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5867 DebugLoc dl = Op.getDebugLoc();
5868 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5869 /*isTarget=*/true, OpFlags);
5871 if (Subtarget->isPICStyleRIPRel() &&
5872 (M == CodeModel::Small || M == CodeModel::Kernel))
5873 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5875 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5877 // With PIC, the address is actually $g + Offset.
5878 if (isGlobalRelativeToPICBase(OpFlags)) {
5879 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5880 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5888 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5890 SelectionDAG &DAG) const {
5891 // Create the TargetGlobalAddress node, folding in the constant
5892 // offset if it is legal.
5893 unsigned char OpFlags =
5894 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5895 CodeModel::Model M = getTargetMachine().getCodeModel();
5897 if (OpFlags == X86II::MO_NO_FLAG &&
5898 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5899 // A direct static reference to a global.
5900 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5903 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5906 if (Subtarget->isPICStyleRIPRel() &&
5907 (M == CodeModel::Small || M == CodeModel::Kernel))
5908 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5910 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5912 // With PIC, the address is actually $g + Offset.
5913 if (isGlobalRelativeToPICBase(OpFlags)) {
5914 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5915 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5919 // For globals that require a load from a stub to get the address, emit the
5921 if (isGlobalStubReference(OpFlags))
5922 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5923 PseudoSourceValue::getGOT(), 0, false, false, 0);
5925 // If there was a non-zero offset that we didn't fold, create an explicit
5928 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5929 DAG.getConstant(Offset, getPointerTy()));
5935 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5936 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5937 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5938 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5942 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5943 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5944 unsigned char OperandFlags) {
5945 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5946 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5947 DebugLoc dl = GA->getDebugLoc();
5948 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5949 GA->getValueType(0),
5953 SDValue Ops[] = { Chain, TGA, *InFlag };
5954 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5956 SDValue Ops[] = { Chain, TGA };
5957 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5960 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5961 MFI->setAdjustsStack(true);
5963 SDValue Flag = Chain.getValue(1);
5964 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5967 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5969 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5972 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5973 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5974 DAG.getNode(X86ISD::GlobalBaseReg,
5975 DebugLoc(), PtrVT), InFlag);
5976 InFlag = Chain.getValue(1);
5978 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5981 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5983 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5985 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5986 X86::RAX, X86II::MO_TLSGD);
5989 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5990 // "local exec" model.
5991 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5992 const EVT PtrVT, TLSModel::Model model,
5994 DebugLoc dl = GA->getDebugLoc();
5995 // Get the Thread Pointer
5996 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5998 DAG.getRegister(is64Bit? X86::FS : X86::GS,
6001 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
6002 NULL, 0, false, false, 0);
6004 unsigned char OperandFlags = 0;
6005 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6007 unsigned WrapperKind = X86ISD::Wrapper;
6008 if (model == TLSModel::LocalExec) {
6009 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6010 } else if (is64Bit) {
6011 assert(model == TLSModel::InitialExec);
6012 OperandFlags = X86II::MO_GOTTPOFF;
6013 WrapperKind = X86ISD::WrapperRIP;
6015 assert(model == TLSModel::InitialExec);
6016 OperandFlags = X86II::MO_INDNTPOFF;
6019 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6021 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6022 GA->getValueType(0),
6023 GA->getOffset(), OperandFlags);
6024 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6026 if (model == TLSModel::InitialExec)
6027 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6028 PseudoSourceValue::getGOT(), 0, false, false, 0);
6030 // The address of the thread local variable is the add of the thread
6031 // pointer with the offset of the variable.
6032 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6036 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6038 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6039 const GlobalValue *GV = GA->getGlobal();
6041 if (Subtarget->isTargetELF()) {
6042 // TODO: implement the "local dynamic" model
6043 // TODO: implement the "initial exec"model for pic executables
6045 // If GV is an alias then use the aliasee for determining
6046 // thread-localness.
6047 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6048 GV = GA->resolveAliasedGlobal(false);
6050 TLSModel::Model model
6051 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6054 case TLSModel::GeneralDynamic:
6055 case TLSModel::LocalDynamic: // not implemented
6056 if (Subtarget->is64Bit())
6057 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6058 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6060 case TLSModel::InitialExec:
6061 case TLSModel::LocalExec:
6062 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6063 Subtarget->is64Bit());
6065 } else if (Subtarget->isTargetDarwin()) {
6066 // Darwin only has one model of TLS. Lower to that.
6067 unsigned char OpFlag = 0;
6068 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6069 X86ISD::WrapperRIP : X86ISD::Wrapper;
6071 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6073 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6074 !Subtarget->is64Bit();
6076 OpFlag = X86II::MO_TLVP_PIC_BASE;
6078 OpFlag = X86II::MO_TLVP;
6079 DebugLoc DL = Op.getDebugLoc();
6080 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6082 GA->getOffset(), OpFlag);
6083 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6085 // With PIC32, the address is actually $g + Offset.
6087 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6088 DAG.getNode(X86ISD::GlobalBaseReg,
6089 DebugLoc(), getPointerTy()),
6092 // Lowering the machine isd will make sure everything is in the right
6094 SDValue Args[] = { Offset };
6095 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
6097 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6098 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6099 MFI->setAdjustsStack(true);
6101 // And our return value (tls address) is in the standard call return value
6103 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
6104 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6108 "TLS not implemented for this target.");
6110 llvm_unreachable("Unreachable");
6115 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
6116 /// take a 2 x i32 value to shift plus a shift amount.
6117 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
6118 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6119 EVT VT = Op.getValueType();
6120 unsigned VTBits = VT.getSizeInBits();
6121 DebugLoc dl = Op.getDebugLoc();
6122 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6123 SDValue ShOpLo = Op.getOperand(0);
6124 SDValue ShOpHi = Op.getOperand(1);
6125 SDValue ShAmt = Op.getOperand(2);
6126 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6127 DAG.getConstant(VTBits - 1, MVT::i8))
6128 : DAG.getConstant(0, VT);
6131 if (Op.getOpcode() == ISD::SHL_PARTS) {
6132 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6133 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6135 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6136 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6139 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6140 DAG.getConstant(VTBits, MVT::i8));
6141 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6142 AndNode, DAG.getConstant(0, MVT::i8));
6145 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6146 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6147 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6149 if (Op.getOpcode() == ISD::SHL_PARTS) {
6150 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6151 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6153 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6154 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6157 SDValue Ops[2] = { Lo, Hi };
6158 return DAG.getMergeValues(Ops, 2, dl);
6161 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6162 SelectionDAG &DAG) const {
6163 EVT SrcVT = Op.getOperand(0).getValueType();
6165 if (SrcVT.isVector()) {
6166 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
6172 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6173 "Unknown SINT_TO_FP to lower!");
6175 // These are really Legal; return the operand so the caller accepts it as
6177 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6179 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6180 Subtarget->is64Bit()) {
6184 DebugLoc dl = Op.getDebugLoc();
6185 unsigned Size = SrcVT.getSizeInBits()/8;
6186 MachineFunction &MF = DAG.getMachineFunction();
6187 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6188 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6189 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6191 PseudoSourceValue::getFixedStack(SSFI), 0,
6193 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6196 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6198 SelectionDAG &DAG) const {
6200 DebugLoc dl = Op.getDebugLoc();
6202 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6204 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
6206 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6207 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6208 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
6209 Tys, Ops, array_lengthof(Ops));
6212 Chain = Result.getValue(1);
6213 SDValue InFlag = Result.getValue(2);
6215 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6216 // shouldn't be necessary except that RFP cannot be live across
6217 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6218 MachineFunction &MF = DAG.getMachineFunction();
6219 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
6220 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6221 Tys = DAG.getVTList(MVT::Other);
6223 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6225 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
6226 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
6227 PseudoSourceValue::getFixedStack(SSFI), 0,
6234 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6235 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6236 SelectionDAG &DAG) const {
6237 // This algorithm is not obvious. Here it is in C code, more or less:
6239 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6240 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6241 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6243 // Copy ints to xmm registers.
6244 __m128i xh = _mm_cvtsi32_si128( hi );
6245 __m128i xl = _mm_cvtsi32_si128( lo );
6247 // Combine into low half of a single xmm register.
6248 __m128i x = _mm_unpacklo_epi32( xh, xl );
6252 // Merge in appropriate exponents to give the integer bits the right
6254 x = _mm_unpacklo_epi32( x, exp );
6256 // Subtract away the biases to deal with the IEEE-754 double precision
6258 d = _mm_sub_pd( (__m128d) x, bias );
6260 // All conversions up to here are exact. The correctly rounded result is
6261 // calculated using the current rounding mode using the following
6263 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6264 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6265 // store doesn't really need to be here (except
6266 // maybe to zero the other double)
6271 DebugLoc dl = Op.getDebugLoc();
6272 LLVMContext *Context = DAG.getContext();
6274 // Build some magic constants.
6275 std::vector<Constant*> CV0;
6276 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6277 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6278 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6279 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6280 Constant *C0 = ConstantVector::get(CV0);
6281 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6283 std::vector<Constant*> CV1;
6285 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6287 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6288 Constant *C1 = ConstantVector::get(CV1);
6289 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6291 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6292 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6294 DAG.getIntPtrConstant(1)));
6295 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6296 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6298 DAG.getIntPtrConstant(0)));
6299 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6300 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6301 PseudoSourceValue::getConstantPool(), 0,
6303 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6304 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
6305 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6306 PseudoSourceValue::getConstantPool(), 0,
6308 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6310 // Add the halves; easiest way is to swap them into another reg first.
6311 int ShufMask[2] = { 1, -1 };
6312 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6313 DAG.getUNDEF(MVT::v2f64), ShufMask);
6314 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6315 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6316 DAG.getIntPtrConstant(0));
6319 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6320 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6321 SelectionDAG &DAG) const {
6322 DebugLoc dl = Op.getDebugLoc();
6323 // FP constant to bias correct the final result.
6324 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6327 // Load the 32-bit value into an XMM register.
6328 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6329 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6331 DAG.getIntPtrConstant(0)));
6333 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6334 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
6335 DAG.getIntPtrConstant(0));
6337 // Or the load with the bias.
6338 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6339 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6340 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6342 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6343 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6344 MVT::v2f64, Bias)));
6345 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6346 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
6347 DAG.getIntPtrConstant(0));
6349 // Subtract the bias.
6350 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
6352 // Handle final rounding.
6353 EVT DestVT = Op.getValueType();
6355 if (DestVT.bitsLT(MVT::f64)) {
6356 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
6357 DAG.getIntPtrConstant(0));
6358 } else if (DestVT.bitsGT(MVT::f64)) {
6359 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
6362 // Handle final rounding.
6366 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
6367 SelectionDAG &DAG) const {
6368 SDValue N0 = Op.getOperand(0);
6369 DebugLoc dl = Op.getDebugLoc();
6371 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
6372 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
6373 // the optimization here.
6374 if (DAG.SignBitIsZero(N0))
6375 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
6377 EVT SrcVT = N0.getValueType();
6378 EVT DstVT = Op.getValueType();
6379 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
6380 return LowerUINT_TO_FP_i64(Op, DAG);
6381 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
6382 return LowerUINT_TO_FP_i32(Op, DAG);
6384 // Make a 64-bit buffer, and use it to build an FILD.
6385 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
6386 if (SrcVT == MVT::i32) {
6387 SDValue WordOff = DAG.getConstant(4, getPointerTy());
6388 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
6389 getPointerTy(), StackSlot, WordOff);
6390 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6391 StackSlot, NULL, 0, false, false, 0);
6392 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
6393 OffsetSlot, NULL, 0, false, false, 0);
6394 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
6398 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
6399 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6400 StackSlot, NULL, 0, false, false, 0);
6401 // For i64 source, we need to add the appropriate power of 2 if the input
6402 // was negative. This is the same as the optimization in
6403 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
6404 // we must be careful to do the computation in x87 extended precision, not
6405 // in SSE. (The generic code can't know it's OK to do this, or how to.)
6406 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
6407 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
6408 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
6410 APInt FF(32, 0x5F800000ULL);
6412 // Check whether the sign bit is set.
6413 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
6414 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
6417 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
6418 SDValue FudgePtr = DAG.getConstantPool(
6419 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
6422 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
6423 SDValue Zero = DAG.getIntPtrConstant(0);
6424 SDValue Four = DAG.getIntPtrConstant(4);
6425 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
6427 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
6429 // Load the value out, extending it from f32 to f80.
6430 // FIXME: Avoid the extend by constructing the right constant pool?
6431 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
6432 FudgePtr, PseudoSourceValue::getConstantPool(),
6433 0, MVT::f32, false, false, 4);
6434 // Extend everything to 80 bits to force it to be done on x87.
6435 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
6436 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
6439 std::pair<SDValue,SDValue> X86TargetLowering::
6440 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
6441 DebugLoc dl = Op.getDebugLoc();
6443 EVT DstTy = Op.getValueType();
6446 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
6450 assert(DstTy.getSimpleVT() <= MVT::i64 &&
6451 DstTy.getSimpleVT() >= MVT::i16 &&
6452 "Unknown FP_TO_SINT to lower!");
6454 // These are really Legal.
6455 if (DstTy == MVT::i32 &&
6456 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6457 return std::make_pair(SDValue(), SDValue());
6458 if (Subtarget->is64Bit() &&
6459 DstTy == MVT::i64 &&
6460 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6461 return std::make_pair(SDValue(), SDValue());
6463 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
6465 MachineFunction &MF = DAG.getMachineFunction();
6466 unsigned MemSize = DstTy.getSizeInBits()/8;
6467 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6468 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6471 switch (DstTy.getSimpleVT().SimpleTy) {
6472 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
6473 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
6474 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
6475 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
6478 SDValue Chain = DAG.getEntryNode();
6479 SDValue Value = Op.getOperand(0);
6480 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
6481 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
6482 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
6483 PseudoSourceValue::getFixedStack(SSFI), 0,
6485 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
6487 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
6489 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
6490 Chain = Value.getValue(1);
6491 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6492 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6495 // Build the FP_TO_INT*_IN_MEM
6496 SDValue Ops[] = { Chain, Value, StackSlot };
6497 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
6499 return std::make_pair(FIST, StackSlot);
6502 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
6503 SelectionDAG &DAG) const {
6504 if (Op.getValueType().isVector()) {
6505 if (Op.getValueType() == MVT::v2i32 &&
6506 Op.getOperand(0).getValueType() == MVT::v2f64) {
6512 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
6513 SDValue FIST = Vals.first, StackSlot = Vals.second;
6514 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
6515 if (FIST.getNode() == 0) return Op;
6518 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6519 FIST, StackSlot, NULL, 0, false, false, 0);
6522 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
6523 SelectionDAG &DAG) const {
6524 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
6525 SDValue FIST = Vals.first, StackSlot = Vals.second;
6526 assert(FIST.getNode() && "Unexpected failure");
6529 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6530 FIST, StackSlot, NULL, 0, false, false, 0);
6533 SDValue X86TargetLowering::LowerFABS(SDValue Op,
6534 SelectionDAG &DAG) const {
6535 LLVMContext *Context = DAG.getContext();
6536 DebugLoc dl = Op.getDebugLoc();
6537 EVT VT = Op.getValueType();
6540 EltVT = VT.getVectorElementType();
6541 std::vector<Constant*> CV;
6542 if (EltVT == MVT::f64) {
6543 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
6547 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
6553 Constant *C = ConstantVector::get(CV);
6554 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6555 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6556 PseudoSourceValue::getConstantPool(), 0,
6558 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
6561 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
6562 LLVMContext *Context = DAG.getContext();
6563 DebugLoc dl = Op.getDebugLoc();
6564 EVT VT = Op.getValueType();
6567 EltVT = VT.getVectorElementType();
6568 std::vector<Constant*> CV;
6569 if (EltVT == MVT::f64) {
6570 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
6574 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
6580 Constant *C = ConstantVector::get(CV);
6581 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6582 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6583 PseudoSourceValue::getConstantPool(), 0,
6585 if (VT.isVector()) {
6586 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6587 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6588 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6590 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
6592 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6596 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6597 LLVMContext *Context = DAG.getContext();
6598 SDValue Op0 = Op.getOperand(0);
6599 SDValue Op1 = Op.getOperand(1);
6600 DebugLoc dl = Op.getDebugLoc();
6601 EVT VT = Op.getValueType();
6602 EVT SrcVT = Op1.getValueType();
6604 // If second operand is smaller, extend it first.
6605 if (SrcVT.bitsLT(VT)) {
6606 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6609 // And if it is bigger, shrink it first.
6610 if (SrcVT.bitsGT(VT)) {
6611 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6615 // At this point the operands and the result should have the same
6616 // type, and that won't be f80 since that is not custom lowered.
6618 // First get the sign bit of second operand.
6619 std::vector<Constant*> CV;
6620 if (SrcVT == MVT::f64) {
6621 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6622 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6624 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6625 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6626 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6627 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6629 Constant *C = ConstantVector::get(CV);
6630 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6631 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6632 PseudoSourceValue::getConstantPool(), 0,
6634 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6636 // Shift sign bit right or left if the two operands have different types.
6637 if (SrcVT.bitsGT(VT)) {
6638 // Op0 is MVT::f32, Op1 is MVT::f64.
6639 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6640 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6641 DAG.getConstant(32, MVT::i32));
6642 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
6643 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6644 DAG.getIntPtrConstant(0));
6647 // Clear first operand sign bit.
6649 if (VT == MVT::f64) {
6650 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6651 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6653 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6654 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6655 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6656 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6658 C = ConstantVector::get(CV);
6659 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6660 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6661 PseudoSourceValue::getConstantPool(), 0,
6663 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6665 // Or the value with the sign bit.
6666 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6669 /// Emit nodes that will be selected as "test Op0,Op0", or something
6671 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6672 SelectionDAG &DAG) const {
6673 DebugLoc dl = Op.getDebugLoc();
6675 // CF and OF aren't always set the way we want. Determine which
6676 // of these we need.
6677 bool NeedCF = false;
6678 bool NeedOF = false;
6681 case X86::COND_A: case X86::COND_AE:
6682 case X86::COND_B: case X86::COND_BE:
6685 case X86::COND_G: case X86::COND_GE:
6686 case X86::COND_L: case X86::COND_LE:
6687 case X86::COND_O: case X86::COND_NO:
6692 // See if we can use the EFLAGS value from the operand instead of
6693 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6694 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6695 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6696 // Emit a CMP with 0, which is the TEST pattern.
6697 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6698 DAG.getConstant(0, Op.getValueType()));
6700 unsigned Opcode = 0;
6701 unsigned NumOperands = 0;
6702 switch (Op.getNode()->getOpcode()) {
6704 // Due to an isel shortcoming, be conservative if this add is likely to be
6705 // selected as part of a load-modify-store instruction. When the root node
6706 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6707 // uses of other nodes in the match, such as the ADD in this case. This
6708 // leads to the ADD being left around and reselected, with the result being
6709 // two adds in the output. Alas, even if none our users are stores, that
6710 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6711 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6712 // climbing the DAG back to the root, and it doesn't seem to be worth the
6714 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6715 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6716 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6719 if (ConstantSDNode *C =
6720 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6721 // An add of one will be selected as an INC.
6722 if (C->getAPIntValue() == 1) {
6723 Opcode = X86ISD::INC;
6728 // An add of negative one (subtract of one) will be selected as a DEC.
6729 if (C->getAPIntValue().isAllOnesValue()) {
6730 Opcode = X86ISD::DEC;
6736 // Otherwise use a regular EFLAGS-setting add.
6737 Opcode = X86ISD::ADD;
6741 // If the primary and result isn't used, don't bother using X86ISD::AND,
6742 // because a TEST instruction will be better.
6743 bool NonFlagUse = false;
6744 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6745 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6747 unsigned UOpNo = UI.getOperandNo();
6748 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6749 // Look pass truncate.
6750 UOpNo = User->use_begin().getOperandNo();
6751 User = *User->use_begin();
6754 if (User->getOpcode() != ISD::BRCOND &&
6755 User->getOpcode() != ISD::SETCC &&
6756 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6769 // Due to the ISEL shortcoming noted above, be conservative if this op is
6770 // likely to be selected as part of a load-modify-store instruction.
6771 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6772 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6773 if (UI->getOpcode() == ISD::STORE)
6776 // Otherwise use a regular EFLAGS-setting instruction.
6777 switch (Op.getNode()->getOpcode()) {
6778 default: llvm_unreachable("unexpected operator!");
6779 case ISD::SUB: Opcode = X86ISD::SUB; break;
6780 case ISD::OR: Opcode = X86ISD::OR; break;
6781 case ISD::XOR: Opcode = X86ISD::XOR; break;
6782 case ISD::AND: Opcode = X86ISD::AND; break;
6794 return SDValue(Op.getNode(), 1);
6801 // Emit a CMP with 0, which is the TEST pattern.
6802 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6803 DAG.getConstant(0, Op.getValueType()));
6805 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6806 SmallVector<SDValue, 4> Ops;
6807 for (unsigned i = 0; i != NumOperands; ++i)
6808 Ops.push_back(Op.getOperand(i));
6810 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6811 DAG.ReplaceAllUsesWith(Op, New);
6812 return SDValue(New.getNode(), 1);
6815 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6817 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6818 SelectionDAG &DAG) const {
6819 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6820 if (C->getAPIntValue() == 0)
6821 return EmitTest(Op0, X86CC, DAG);
6823 DebugLoc dl = Op0.getDebugLoc();
6824 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6827 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6828 /// if it's possible.
6829 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6830 DebugLoc dl, SelectionDAG &DAG) const {
6831 SDValue Op0 = And.getOperand(0);
6832 SDValue Op1 = And.getOperand(1);
6833 if (Op0.getOpcode() == ISD::TRUNCATE)
6834 Op0 = Op0.getOperand(0);
6835 if (Op1.getOpcode() == ISD::TRUNCATE)
6836 Op1 = Op1.getOperand(0);
6839 if (Op1.getOpcode() == ISD::SHL)
6840 std::swap(Op0, Op1);
6841 if (Op0.getOpcode() == ISD::SHL) {
6842 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6843 if (And00C->getZExtValue() == 1) {
6844 // If we looked past a truncate, check that it's only truncating away
6846 unsigned BitWidth = Op0.getValueSizeInBits();
6847 unsigned AndBitWidth = And.getValueSizeInBits();
6848 if (BitWidth > AndBitWidth) {
6849 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6850 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6851 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6855 RHS = Op0.getOperand(1);
6857 } else if (Op1.getOpcode() == ISD::Constant) {
6858 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6859 SDValue AndLHS = Op0;
6860 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6861 LHS = AndLHS.getOperand(0);
6862 RHS = AndLHS.getOperand(1);
6866 if (LHS.getNode()) {
6867 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6868 // instruction. Since the shift amount is in-range-or-undefined, we know
6869 // that doing a bittest on the i32 value is ok. We extend to i32 because
6870 // the encoding for the i16 version is larger than the i32 version.
6871 // Also promote i16 to i32 for performance / code size reason.
6872 if (LHS.getValueType() == MVT::i8 ||
6873 LHS.getValueType() == MVT::i16)
6874 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6876 // If the operand types disagree, extend the shift amount to match. Since
6877 // BT ignores high bits (like shifts) we can use anyextend.
6878 if (LHS.getValueType() != RHS.getValueType())
6879 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6881 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6882 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6883 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6884 DAG.getConstant(Cond, MVT::i8), BT);
6890 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6891 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6892 SDValue Op0 = Op.getOperand(0);
6893 SDValue Op1 = Op.getOperand(1);
6894 DebugLoc dl = Op.getDebugLoc();
6895 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6897 // Optimize to BT if possible.
6898 // Lower (X & (1 << N)) == 0 to BT(X, N).
6899 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6900 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6901 if (Op0.getOpcode() == ISD::AND &&
6903 Op1.getOpcode() == ISD::Constant &&
6904 cast<ConstantSDNode>(Op1)->isNullValue() &&
6905 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6906 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6907 if (NewSetCC.getNode())
6911 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6912 if (Op0.getOpcode() == X86ISD::SETCC &&
6913 Op1.getOpcode() == ISD::Constant &&
6914 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6915 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6916 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6917 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6918 bool Invert = (CC == ISD::SETNE) ^
6919 cast<ConstantSDNode>(Op1)->isNullValue();
6921 CCode = X86::GetOppositeBranchCondition(CCode);
6922 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6923 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6926 bool isFP = Op1.getValueType().isFloatingPoint();
6927 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6928 if (X86CC == X86::COND_INVALID)
6931 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6933 // Use sbb x, x to materialize carry bit into a GPR.
6934 if (X86CC == X86::COND_B)
6935 return DAG.getNode(ISD::AND, dl, MVT::i8,
6936 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6937 DAG.getConstant(X86CC, MVT::i8), Cond),
6938 DAG.getConstant(1, MVT::i8));
6940 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6941 DAG.getConstant(X86CC, MVT::i8), Cond);
6944 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6946 SDValue Op0 = Op.getOperand(0);
6947 SDValue Op1 = Op.getOperand(1);
6948 SDValue CC = Op.getOperand(2);
6949 EVT VT = Op.getValueType();
6950 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6951 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6952 DebugLoc dl = Op.getDebugLoc();
6956 EVT VT0 = Op0.getValueType();
6957 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6958 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6961 switch (SetCCOpcode) {
6964 case ISD::SETEQ: SSECC = 0; break;
6966 case ISD::SETGT: Swap = true; // Fallthrough
6968 case ISD::SETOLT: SSECC = 1; break;
6970 case ISD::SETGE: Swap = true; // Fallthrough
6972 case ISD::SETOLE: SSECC = 2; break;
6973 case ISD::SETUO: SSECC = 3; break;
6975 case ISD::SETNE: SSECC = 4; break;
6976 case ISD::SETULE: Swap = true;
6977 case ISD::SETUGE: SSECC = 5; break;
6978 case ISD::SETULT: Swap = true;
6979 case ISD::SETUGT: SSECC = 6; break;
6980 case ISD::SETO: SSECC = 7; break;
6983 std::swap(Op0, Op1);
6985 // In the two special cases we can't handle, emit two comparisons.
6987 if (SetCCOpcode == ISD::SETUEQ) {
6989 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6990 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6991 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6993 else if (SetCCOpcode == ISD::SETONE) {
6995 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6996 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6997 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6999 llvm_unreachable("Illegal FP comparison");
7001 // Handle all other FP comparisons here.
7002 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7005 // We are handling one of the integer comparisons here. Since SSE only has
7006 // GT and EQ comparisons for integer, swapping operands and multiple
7007 // operations may be required for some comparisons.
7008 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7009 bool Swap = false, Invert = false, FlipSigns = false;
7011 switch (VT.getSimpleVT().SimpleTy) {
7014 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7016 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7018 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7019 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7022 switch (SetCCOpcode) {
7024 case ISD::SETNE: Invert = true;
7025 case ISD::SETEQ: Opc = EQOpc; break;
7026 case ISD::SETLT: Swap = true;
7027 case ISD::SETGT: Opc = GTOpc; break;
7028 case ISD::SETGE: Swap = true;
7029 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7030 case ISD::SETULT: Swap = true;
7031 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7032 case ISD::SETUGE: Swap = true;
7033 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7036 std::swap(Op0, Op1);
7038 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7039 // bits of the inputs before performing those operations.
7041 EVT EltVT = VT.getVectorElementType();
7042 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7044 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7045 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7047 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7048 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7051 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7053 // If the logical-not of the result is required, perform that now.
7055 Result = DAG.getNOT(dl, Result, VT);
7060 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
7061 static bool isX86LogicalCmp(SDValue Op) {
7062 unsigned Opc = Op.getNode()->getOpcode();
7063 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
7065 if (Op.getResNo() == 1 &&
7066 (Opc == X86ISD::ADD ||
7067 Opc == X86ISD::SUB ||
7068 Opc == X86ISD::SMUL ||
7069 Opc == X86ISD::UMUL ||
7070 Opc == X86ISD::INC ||
7071 Opc == X86ISD::DEC ||
7072 Opc == X86ISD::OR ||
7073 Opc == X86ISD::XOR ||
7074 Opc == X86ISD::AND))
7080 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
7081 bool addTest = true;
7082 SDValue Cond = Op.getOperand(0);
7083 DebugLoc dl = Op.getDebugLoc();
7086 if (Cond.getOpcode() == ISD::SETCC) {
7087 SDValue NewCond = LowerSETCC(Cond, DAG);
7088 if (NewCond.getNode())
7092 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
7093 SDValue Op1 = Op.getOperand(1);
7094 SDValue Op2 = Op.getOperand(2);
7095 if (Cond.getOpcode() == X86ISD::SETCC &&
7096 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
7097 SDValue Cmp = Cond.getOperand(1);
7098 if (Cmp.getOpcode() == X86ISD::CMP) {
7099 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
7100 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
7101 ConstantSDNode *RHSC =
7102 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
7103 if (N1C && N1C->isAllOnesValue() &&
7104 N2C && N2C->isNullValue() &&
7105 RHSC && RHSC->isNullValue()) {
7106 SDValue CmpOp0 = Cmp.getOperand(0);
7107 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
7108 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7109 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
7110 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7115 // Look pass (and (setcc_carry (cmp ...)), 1).
7116 if (Cond.getOpcode() == ISD::AND &&
7117 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7118 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7119 if (C && C->getAPIntValue() == 1)
7120 Cond = Cond.getOperand(0);
7123 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7124 // setting operand in place of the X86ISD::SETCC.
7125 if (Cond.getOpcode() == X86ISD::SETCC ||
7126 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7127 CC = Cond.getOperand(0);
7129 SDValue Cmp = Cond.getOperand(1);
7130 unsigned Opc = Cmp.getOpcode();
7131 EVT VT = Op.getValueType();
7133 bool IllegalFPCMov = false;
7134 if (VT.isFloatingPoint() && !VT.isVector() &&
7135 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7136 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7138 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7139 Opc == X86ISD::BT) { // FIXME
7146 // Look pass the truncate.
7147 if (Cond.getOpcode() == ISD::TRUNCATE)
7148 Cond = Cond.getOperand(0);
7150 // We know the result of AND is compared against zero. Try to match
7152 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7153 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7154 if (NewSetCC.getNode()) {
7155 CC = NewSetCC.getOperand(0);
7156 Cond = NewSetCC.getOperand(1);
7163 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7164 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7167 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7168 // condition is true.
7169 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
7170 SDValue Ops[] = { Op2, Op1, CC, Cond };
7171 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
7174 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7175 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7176 // from the AND / OR.
7177 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7178 Opc = Op.getOpcode();
7179 if (Opc != ISD::OR && Opc != ISD::AND)
7181 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7182 Op.getOperand(0).hasOneUse() &&
7183 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7184 Op.getOperand(1).hasOneUse());
7187 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7188 // 1 and that the SETCC node has a single use.
7189 static bool isXor1OfSetCC(SDValue Op) {
7190 if (Op.getOpcode() != ISD::XOR)
7192 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7193 if (N1C && N1C->getAPIntValue() == 1) {
7194 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7195 Op.getOperand(0).hasOneUse();
7200 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7201 bool addTest = true;
7202 SDValue Chain = Op.getOperand(0);
7203 SDValue Cond = Op.getOperand(1);
7204 SDValue Dest = Op.getOperand(2);
7205 DebugLoc dl = Op.getDebugLoc();
7208 if (Cond.getOpcode() == ISD::SETCC) {
7209 SDValue NewCond = LowerSETCC(Cond, DAG);
7210 if (NewCond.getNode())
7214 // FIXME: LowerXALUO doesn't handle these!!
7215 else if (Cond.getOpcode() == X86ISD::ADD ||
7216 Cond.getOpcode() == X86ISD::SUB ||
7217 Cond.getOpcode() == X86ISD::SMUL ||
7218 Cond.getOpcode() == X86ISD::UMUL)
7219 Cond = LowerXALUO(Cond, DAG);
7222 // Look pass (and (setcc_carry (cmp ...)), 1).
7223 if (Cond.getOpcode() == ISD::AND &&
7224 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7225 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7226 if (C && C->getAPIntValue() == 1)
7227 Cond = Cond.getOperand(0);
7230 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7231 // setting operand in place of the X86ISD::SETCC.
7232 if (Cond.getOpcode() == X86ISD::SETCC ||
7233 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7234 CC = Cond.getOperand(0);
7236 SDValue Cmp = Cond.getOperand(1);
7237 unsigned Opc = Cmp.getOpcode();
7238 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7239 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7243 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7247 // These can only come from an arithmetic instruction with overflow,
7248 // e.g. SADDO, UADDO.
7249 Cond = Cond.getNode()->getOperand(1);
7256 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7257 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7258 if (CondOpc == ISD::OR) {
7259 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7260 // two branches instead of an explicit OR instruction with a
7262 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7263 isX86LogicalCmp(Cmp)) {
7264 CC = Cond.getOperand(0).getOperand(0);
7265 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7266 Chain, Dest, CC, Cmp);
7267 CC = Cond.getOperand(1).getOperand(0);
7271 } else { // ISD::AND
7272 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7273 // two branches instead of an explicit AND instruction with a
7274 // separate test. However, we only do this if this block doesn't
7275 // have a fall-through edge, because this requires an explicit
7276 // jmp when the condition is false.
7277 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7278 isX86LogicalCmp(Cmp) &&
7279 Op.getNode()->hasOneUse()) {
7280 X86::CondCode CCode =
7281 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7282 CCode = X86::GetOppositeBranchCondition(CCode);
7283 CC = DAG.getConstant(CCode, MVT::i8);
7284 SDNode *User = *Op.getNode()->use_begin();
7285 // Look for an unconditional branch following this conditional branch.
7286 // We need this because we need to reverse the successors in order
7287 // to implement FCMP_OEQ.
7288 if (User->getOpcode() == ISD::BR) {
7289 SDValue FalseBB = User->getOperand(1);
7291 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
7292 assert(NewBR == User);
7296 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7297 Chain, Dest, CC, Cmp);
7298 X86::CondCode CCode =
7299 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
7300 CCode = X86::GetOppositeBranchCondition(CCode);
7301 CC = DAG.getConstant(CCode, MVT::i8);
7307 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
7308 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
7309 // It should be transformed during dag combiner except when the condition
7310 // is set by a arithmetics with overflow node.
7311 X86::CondCode CCode =
7312 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7313 CCode = X86::GetOppositeBranchCondition(CCode);
7314 CC = DAG.getConstant(CCode, MVT::i8);
7315 Cond = Cond.getOperand(0).getOperand(1);
7321 // Look pass the truncate.
7322 if (Cond.getOpcode() == ISD::TRUNCATE)
7323 Cond = Cond.getOperand(0);
7325 // We know the result of AND is compared against zero. Try to match
7327 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7328 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7329 if (NewSetCC.getNode()) {
7330 CC = NewSetCC.getOperand(0);
7331 Cond = NewSetCC.getOperand(1);
7338 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7339 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7341 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7342 Chain, Dest, CC, Cond);
7346 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
7347 // Calls to _alloca is needed to probe the stack when allocating more than 4k
7348 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
7349 // that the guard pages used by the OS virtual memory manager are allocated in
7350 // correct sequence.
7352 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
7353 SelectionDAG &DAG) const {
7354 assert(Subtarget->isTargetCygMing() &&
7355 "This should be used only on Cygwin/Mingw targets");
7356 DebugLoc dl = Op.getDebugLoc();
7359 SDValue Chain = Op.getOperand(0);
7360 SDValue Size = Op.getOperand(1);
7361 // FIXME: Ensure alignment here
7365 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
7367 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
7368 Flag = Chain.getValue(1);
7370 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
7372 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
7373 Flag = Chain.getValue(1);
7375 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
7377 SDValue Ops1[2] = { Chain.getValue(0), Chain };
7378 return DAG.getMergeValues(Ops1, 2, dl);
7381 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
7382 MachineFunction &MF = DAG.getMachineFunction();
7383 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
7385 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7386 DebugLoc dl = Op.getDebugLoc();
7388 if (!Subtarget->is64Bit()) {
7389 // vastart just stores the address of the VarArgsFrameIndex slot into the
7390 // memory location argument.
7391 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7393 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
7398 // gp_offset (0 - 6 * 8)
7399 // fp_offset (48 - 48 + 8 * 16)
7400 // overflow_arg_area (point to parameters coming in memory).
7402 SmallVector<SDValue, 8> MemOps;
7403 SDValue FIN = Op.getOperand(1);
7405 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
7406 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
7408 FIN, SV, 0, false, false, 0);
7409 MemOps.push_back(Store);
7412 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7413 FIN, DAG.getIntPtrConstant(4));
7414 Store = DAG.getStore(Op.getOperand(0), dl,
7415 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
7417 FIN, SV, 4, false, false, 0);
7418 MemOps.push_back(Store);
7420 // Store ptr to overflow_arg_area
7421 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7422 FIN, DAG.getIntPtrConstant(4));
7423 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7425 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 8,
7427 MemOps.push_back(Store);
7429 // Store ptr to reg_save_area.
7430 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7431 FIN, DAG.getIntPtrConstant(8));
7432 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
7434 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 16,
7436 MemOps.push_back(Store);
7437 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
7438 &MemOps[0], MemOps.size());
7441 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
7442 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7443 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
7445 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
7449 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
7450 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7451 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
7452 SDValue Chain = Op.getOperand(0);
7453 SDValue DstPtr = Op.getOperand(1);
7454 SDValue SrcPtr = Op.getOperand(2);
7455 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
7456 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7457 DebugLoc dl = Op.getDebugLoc();
7459 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
7460 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
7461 false, DstSV, 0, SrcSV, 0);
7465 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
7466 DebugLoc dl = Op.getDebugLoc();
7467 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7469 default: return SDValue(); // Don't custom lower most intrinsics.
7470 // Comparison intrinsics.
7471 case Intrinsic::x86_sse_comieq_ss:
7472 case Intrinsic::x86_sse_comilt_ss:
7473 case Intrinsic::x86_sse_comile_ss:
7474 case Intrinsic::x86_sse_comigt_ss:
7475 case Intrinsic::x86_sse_comige_ss:
7476 case Intrinsic::x86_sse_comineq_ss:
7477 case Intrinsic::x86_sse_ucomieq_ss:
7478 case Intrinsic::x86_sse_ucomilt_ss:
7479 case Intrinsic::x86_sse_ucomile_ss:
7480 case Intrinsic::x86_sse_ucomigt_ss:
7481 case Intrinsic::x86_sse_ucomige_ss:
7482 case Intrinsic::x86_sse_ucomineq_ss:
7483 case Intrinsic::x86_sse2_comieq_sd:
7484 case Intrinsic::x86_sse2_comilt_sd:
7485 case Intrinsic::x86_sse2_comile_sd:
7486 case Intrinsic::x86_sse2_comigt_sd:
7487 case Intrinsic::x86_sse2_comige_sd:
7488 case Intrinsic::x86_sse2_comineq_sd:
7489 case Intrinsic::x86_sse2_ucomieq_sd:
7490 case Intrinsic::x86_sse2_ucomilt_sd:
7491 case Intrinsic::x86_sse2_ucomile_sd:
7492 case Intrinsic::x86_sse2_ucomigt_sd:
7493 case Intrinsic::x86_sse2_ucomige_sd:
7494 case Intrinsic::x86_sse2_ucomineq_sd: {
7496 ISD::CondCode CC = ISD::SETCC_INVALID;
7499 case Intrinsic::x86_sse_comieq_ss:
7500 case Intrinsic::x86_sse2_comieq_sd:
7504 case Intrinsic::x86_sse_comilt_ss:
7505 case Intrinsic::x86_sse2_comilt_sd:
7509 case Intrinsic::x86_sse_comile_ss:
7510 case Intrinsic::x86_sse2_comile_sd:
7514 case Intrinsic::x86_sse_comigt_ss:
7515 case Intrinsic::x86_sse2_comigt_sd:
7519 case Intrinsic::x86_sse_comige_ss:
7520 case Intrinsic::x86_sse2_comige_sd:
7524 case Intrinsic::x86_sse_comineq_ss:
7525 case Intrinsic::x86_sse2_comineq_sd:
7529 case Intrinsic::x86_sse_ucomieq_ss:
7530 case Intrinsic::x86_sse2_ucomieq_sd:
7531 Opc = X86ISD::UCOMI;
7534 case Intrinsic::x86_sse_ucomilt_ss:
7535 case Intrinsic::x86_sse2_ucomilt_sd:
7536 Opc = X86ISD::UCOMI;
7539 case Intrinsic::x86_sse_ucomile_ss:
7540 case Intrinsic::x86_sse2_ucomile_sd:
7541 Opc = X86ISD::UCOMI;
7544 case Intrinsic::x86_sse_ucomigt_ss:
7545 case Intrinsic::x86_sse2_ucomigt_sd:
7546 Opc = X86ISD::UCOMI;
7549 case Intrinsic::x86_sse_ucomige_ss:
7550 case Intrinsic::x86_sse2_ucomige_sd:
7551 Opc = X86ISD::UCOMI;
7554 case Intrinsic::x86_sse_ucomineq_ss:
7555 case Intrinsic::x86_sse2_ucomineq_sd:
7556 Opc = X86ISD::UCOMI;
7561 SDValue LHS = Op.getOperand(1);
7562 SDValue RHS = Op.getOperand(2);
7563 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
7564 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
7565 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
7566 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7567 DAG.getConstant(X86CC, MVT::i8), Cond);
7568 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7570 // ptest and testp intrinsics. The intrinsic these come from are designed to
7571 // return an integer value, not just an instruction so lower it to the ptest
7572 // or testp pattern and a setcc for the result.
7573 case Intrinsic::x86_sse41_ptestz:
7574 case Intrinsic::x86_sse41_ptestc:
7575 case Intrinsic::x86_sse41_ptestnzc:
7576 case Intrinsic::x86_avx_ptestz_256:
7577 case Intrinsic::x86_avx_ptestc_256:
7578 case Intrinsic::x86_avx_ptestnzc_256:
7579 case Intrinsic::x86_avx_vtestz_ps:
7580 case Intrinsic::x86_avx_vtestc_ps:
7581 case Intrinsic::x86_avx_vtestnzc_ps:
7582 case Intrinsic::x86_avx_vtestz_pd:
7583 case Intrinsic::x86_avx_vtestc_pd:
7584 case Intrinsic::x86_avx_vtestnzc_pd:
7585 case Intrinsic::x86_avx_vtestz_ps_256:
7586 case Intrinsic::x86_avx_vtestc_ps_256:
7587 case Intrinsic::x86_avx_vtestnzc_ps_256:
7588 case Intrinsic::x86_avx_vtestz_pd_256:
7589 case Intrinsic::x86_avx_vtestc_pd_256:
7590 case Intrinsic::x86_avx_vtestnzc_pd_256: {
7591 bool IsTestPacked = false;
7594 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
7595 case Intrinsic::x86_avx_vtestz_ps:
7596 case Intrinsic::x86_avx_vtestz_pd:
7597 case Intrinsic::x86_avx_vtestz_ps_256:
7598 case Intrinsic::x86_avx_vtestz_pd_256:
7599 IsTestPacked = true; // Fallthrough
7600 case Intrinsic::x86_sse41_ptestz:
7601 case Intrinsic::x86_avx_ptestz_256:
7603 X86CC = X86::COND_E;
7605 case Intrinsic::x86_avx_vtestc_ps:
7606 case Intrinsic::x86_avx_vtestc_pd:
7607 case Intrinsic::x86_avx_vtestc_ps_256:
7608 case Intrinsic::x86_avx_vtestc_pd_256:
7609 IsTestPacked = true; // Fallthrough
7610 case Intrinsic::x86_sse41_ptestc:
7611 case Intrinsic::x86_avx_ptestc_256:
7613 X86CC = X86::COND_B;
7615 case Intrinsic::x86_avx_vtestnzc_ps:
7616 case Intrinsic::x86_avx_vtestnzc_pd:
7617 case Intrinsic::x86_avx_vtestnzc_ps_256:
7618 case Intrinsic::x86_avx_vtestnzc_pd_256:
7619 IsTestPacked = true; // Fallthrough
7620 case Intrinsic::x86_sse41_ptestnzc:
7621 case Intrinsic::x86_avx_ptestnzc_256:
7623 X86CC = X86::COND_A;
7627 SDValue LHS = Op.getOperand(1);
7628 SDValue RHS = Op.getOperand(2);
7629 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
7630 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
7631 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7632 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7633 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7636 // Fix vector shift instructions where the last operand is a non-immediate
7638 case Intrinsic::x86_sse2_pslli_w:
7639 case Intrinsic::x86_sse2_pslli_d:
7640 case Intrinsic::x86_sse2_pslli_q:
7641 case Intrinsic::x86_sse2_psrli_w:
7642 case Intrinsic::x86_sse2_psrli_d:
7643 case Intrinsic::x86_sse2_psrli_q:
7644 case Intrinsic::x86_sse2_psrai_w:
7645 case Intrinsic::x86_sse2_psrai_d:
7646 case Intrinsic::x86_mmx_pslli_w:
7647 case Intrinsic::x86_mmx_pslli_d:
7648 case Intrinsic::x86_mmx_pslli_q:
7649 case Intrinsic::x86_mmx_psrli_w:
7650 case Intrinsic::x86_mmx_psrli_d:
7651 case Intrinsic::x86_mmx_psrli_q:
7652 case Intrinsic::x86_mmx_psrai_w:
7653 case Intrinsic::x86_mmx_psrai_d: {
7654 SDValue ShAmt = Op.getOperand(2);
7655 if (isa<ConstantSDNode>(ShAmt))
7658 unsigned NewIntNo = 0;
7659 EVT ShAmtVT = MVT::v4i32;
7661 case Intrinsic::x86_sse2_pslli_w:
7662 NewIntNo = Intrinsic::x86_sse2_psll_w;
7664 case Intrinsic::x86_sse2_pslli_d:
7665 NewIntNo = Intrinsic::x86_sse2_psll_d;
7667 case Intrinsic::x86_sse2_pslli_q:
7668 NewIntNo = Intrinsic::x86_sse2_psll_q;
7670 case Intrinsic::x86_sse2_psrli_w:
7671 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7673 case Intrinsic::x86_sse2_psrli_d:
7674 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7676 case Intrinsic::x86_sse2_psrli_q:
7677 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7679 case Intrinsic::x86_sse2_psrai_w:
7680 NewIntNo = Intrinsic::x86_sse2_psra_w;
7682 case Intrinsic::x86_sse2_psrai_d:
7683 NewIntNo = Intrinsic::x86_sse2_psra_d;
7686 ShAmtVT = MVT::v2i32;
7688 case Intrinsic::x86_mmx_pslli_w:
7689 NewIntNo = Intrinsic::x86_mmx_psll_w;
7691 case Intrinsic::x86_mmx_pslli_d:
7692 NewIntNo = Intrinsic::x86_mmx_psll_d;
7694 case Intrinsic::x86_mmx_pslli_q:
7695 NewIntNo = Intrinsic::x86_mmx_psll_q;
7697 case Intrinsic::x86_mmx_psrli_w:
7698 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7700 case Intrinsic::x86_mmx_psrli_d:
7701 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7703 case Intrinsic::x86_mmx_psrli_q:
7704 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7706 case Intrinsic::x86_mmx_psrai_w:
7707 NewIntNo = Intrinsic::x86_mmx_psra_w;
7709 case Intrinsic::x86_mmx_psrai_d:
7710 NewIntNo = Intrinsic::x86_mmx_psra_d;
7712 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7718 // The vector shift intrinsics with scalars uses 32b shift amounts but
7719 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7723 ShOps[1] = DAG.getConstant(0, MVT::i32);
7724 if (ShAmtVT == MVT::v4i32) {
7725 ShOps[2] = DAG.getUNDEF(MVT::i32);
7726 ShOps[3] = DAG.getUNDEF(MVT::i32);
7727 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7729 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7732 EVT VT = Op.getValueType();
7733 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7734 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7735 DAG.getConstant(NewIntNo, MVT::i32),
7736 Op.getOperand(1), ShAmt);
7741 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7742 SelectionDAG &DAG) const {
7743 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7744 MFI->setReturnAddressIsTaken(true);
7746 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7747 DebugLoc dl = Op.getDebugLoc();
7750 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7752 DAG.getConstant(TD->getPointerSize(),
7753 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7754 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7755 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7757 NULL, 0, false, false, 0);
7760 // Just load the return address.
7761 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7762 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7763 RetAddrFI, NULL, 0, false, false, 0);
7766 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7767 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7768 MFI->setFrameAddressIsTaken(true);
7770 EVT VT = Op.getValueType();
7771 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7772 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7773 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7774 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7776 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7781 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7782 SelectionDAG &DAG) const {
7783 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7786 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7787 MachineFunction &MF = DAG.getMachineFunction();
7788 SDValue Chain = Op.getOperand(0);
7789 SDValue Offset = Op.getOperand(1);
7790 SDValue Handler = Op.getOperand(2);
7791 DebugLoc dl = Op.getDebugLoc();
7793 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
7794 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7796 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7798 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
7799 DAG.getIntPtrConstant(TD->getPointerSize()));
7800 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7801 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7802 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7803 MF.getRegInfo().addLiveOut(StoreAddrReg);
7805 return DAG.getNode(X86ISD::EH_RETURN, dl,
7807 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7810 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7811 SelectionDAG &DAG) const {
7812 SDValue Root = Op.getOperand(0);
7813 SDValue Trmp = Op.getOperand(1); // trampoline
7814 SDValue FPtr = Op.getOperand(2); // nested function
7815 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7816 DebugLoc dl = Op.getDebugLoc();
7818 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7820 if (Subtarget->is64Bit()) {
7821 SDValue OutChains[6];
7823 // Large code-model.
7824 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7825 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7827 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7828 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7830 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7832 // Load the pointer to the nested function into R11.
7833 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7834 SDValue Addr = Trmp;
7835 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7836 Addr, TrmpAddr, 0, false, false, 0);
7838 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7839 DAG.getConstant(2, MVT::i64));
7840 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7843 // Load the 'nest' parameter value into R10.
7844 // R10 is specified in X86CallingConv.td
7845 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7846 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7847 DAG.getConstant(10, MVT::i64));
7848 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7849 Addr, TrmpAddr, 10, false, false, 0);
7851 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7852 DAG.getConstant(12, MVT::i64));
7853 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7856 // Jump to the nested function.
7857 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7858 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7859 DAG.getConstant(20, MVT::i64));
7860 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7861 Addr, TrmpAddr, 20, false, false, 0);
7863 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7864 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7865 DAG.getConstant(22, MVT::i64));
7866 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7867 TrmpAddr, 22, false, false, 0);
7870 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7871 return DAG.getMergeValues(Ops, 2, dl);
7873 const Function *Func =
7874 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7875 CallingConv::ID CC = Func->getCallingConv();
7880 llvm_unreachable("Unsupported calling convention");
7881 case CallingConv::C:
7882 case CallingConv::X86_StdCall: {
7883 // Pass 'nest' parameter in ECX.
7884 // Must be kept in sync with X86CallingConv.td
7887 // Check that ECX wasn't needed by an 'inreg' parameter.
7888 const FunctionType *FTy = Func->getFunctionType();
7889 const AttrListPtr &Attrs = Func->getAttributes();
7891 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7892 unsigned InRegCount = 0;
7895 for (FunctionType::param_iterator I = FTy->param_begin(),
7896 E = FTy->param_end(); I != E; ++I, ++Idx)
7897 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7898 // FIXME: should only count parameters that are lowered to integers.
7899 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7901 if (InRegCount > 2) {
7902 report_fatal_error("Nest register in use - reduce number of inreg"
7908 case CallingConv::X86_FastCall:
7909 case CallingConv::X86_ThisCall:
7910 case CallingConv::Fast:
7911 // Pass 'nest' parameter in EAX.
7912 // Must be kept in sync with X86CallingConv.td
7917 SDValue OutChains[4];
7920 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7921 DAG.getConstant(10, MVT::i32));
7922 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7924 // This is storing the opcode for MOV32ri.
7925 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7926 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7927 OutChains[0] = DAG.getStore(Root, dl,
7928 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7929 Trmp, TrmpAddr, 0, false, false, 0);
7931 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7932 DAG.getConstant(1, MVT::i32));
7933 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7936 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7937 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7938 DAG.getConstant(5, MVT::i32));
7939 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7940 TrmpAddr, 5, false, false, 1);
7942 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7943 DAG.getConstant(6, MVT::i32));
7944 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7948 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7949 return DAG.getMergeValues(Ops, 2, dl);
7953 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7954 SelectionDAG &DAG) const {
7956 The rounding mode is in bits 11:10 of FPSR, and has the following
7963 FLT_ROUNDS, on the other hand, expects the following:
7970 To perform the conversion, we do:
7971 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7974 MachineFunction &MF = DAG.getMachineFunction();
7975 const TargetMachine &TM = MF.getTarget();
7976 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7977 unsigned StackAlignment = TFI.getStackAlignment();
7978 EVT VT = Op.getValueType();
7979 DebugLoc dl = Op.getDebugLoc();
7981 // Save FP Control Word to stack slot
7982 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7983 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7985 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7986 DAG.getEntryNode(), StackSlot);
7988 // Load FP Control Word from stack slot
7989 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7992 // Transform as necessary
7994 DAG.getNode(ISD::SRL, dl, MVT::i16,
7995 DAG.getNode(ISD::AND, dl, MVT::i16,
7996 CWD, DAG.getConstant(0x800, MVT::i16)),
7997 DAG.getConstant(11, MVT::i8));
7999 DAG.getNode(ISD::SRL, dl, MVT::i16,
8000 DAG.getNode(ISD::AND, dl, MVT::i16,
8001 CWD, DAG.getConstant(0x400, MVT::i16)),
8002 DAG.getConstant(9, MVT::i8));
8005 DAG.getNode(ISD::AND, dl, MVT::i16,
8006 DAG.getNode(ISD::ADD, dl, MVT::i16,
8007 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
8008 DAG.getConstant(1, MVT::i16)),
8009 DAG.getConstant(3, MVT::i16));
8012 return DAG.getNode((VT.getSizeInBits() < 16 ?
8013 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
8016 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
8017 EVT VT = Op.getValueType();
8019 unsigned NumBits = VT.getSizeInBits();
8020 DebugLoc dl = Op.getDebugLoc();
8022 Op = Op.getOperand(0);
8023 if (VT == MVT::i8) {
8024 // Zero extend to i32 since there is not an i8 bsr.
8026 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8029 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
8030 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8031 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
8033 // If src is zero (i.e. bsr sets ZF), returns NumBits.
8036 DAG.getConstant(NumBits+NumBits-1, OpVT),
8037 DAG.getConstant(X86::COND_E, MVT::i8),
8040 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8042 // Finally xor with NumBits-1.
8043 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
8046 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8050 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
8051 EVT VT = Op.getValueType();
8053 unsigned NumBits = VT.getSizeInBits();
8054 DebugLoc dl = Op.getDebugLoc();
8056 Op = Op.getOperand(0);
8057 if (VT == MVT::i8) {
8059 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8062 // Issue a bsf (scan bits forward) which also sets EFLAGS.
8063 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8064 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
8066 // If src is zero (i.e. bsf sets ZF), returns NumBits.
8069 DAG.getConstant(NumBits, OpVT),
8070 DAG.getConstant(X86::COND_E, MVT::i8),
8073 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8076 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8080 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
8081 EVT VT = Op.getValueType();
8082 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
8083 DebugLoc dl = Op.getDebugLoc();
8085 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
8086 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
8087 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
8088 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
8089 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
8091 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
8092 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
8093 // return AloBlo + AloBhi + AhiBlo;
8095 SDValue A = Op.getOperand(0);
8096 SDValue B = Op.getOperand(1);
8098 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8099 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8100 A, DAG.getConstant(32, MVT::i32));
8101 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8102 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8103 B, DAG.getConstant(32, MVT::i32));
8104 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8105 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8107 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8108 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8110 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8111 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8113 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8114 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8115 AloBhi, DAG.getConstant(32, MVT::i32));
8116 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8117 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8118 AhiBlo, DAG.getConstant(32, MVT::i32));
8119 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8120 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8124 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
8125 EVT VT = Op.getValueType();
8126 DebugLoc dl = Op.getDebugLoc();
8127 SDValue R = Op.getOperand(0);
8129 LLVMContext *Context = DAG.getContext();
8131 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
8133 if (VT == MVT::v4i32) {
8134 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8135 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8136 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
8138 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
8140 std::vector<Constant*> CV(4, CI);
8141 Constant *C = ConstantVector::get(CV);
8142 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8143 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8144 PseudoSourceValue::getConstantPool(), 0,
8147 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
8148 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, Op);
8149 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
8150 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
8152 if (VT == MVT::v16i8) {
8154 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8155 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8156 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
8158 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
8159 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
8161 std::vector<Constant*> CVM1(16, CM1);
8162 std::vector<Constant*> CVM2(16, CM2);
8163 Constant *C = ConstantVector::get(CVM1);
8164 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8165 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8166 PseudoSourceValue::getConstantPool(), 0,
8169 // r = pblendv(r, psllw(r & (char16)15, 4), a);
8170 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8171 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8172 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8173 DAG.getConstant(4, MVT::i32));
8174 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8175 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8178 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8180 C = ConstantVector::get(CVM2);
8181 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8182 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8183 PseudoSourceValue::getConstantPool(), 0, false, false, 16);
8185 // r = pblendv(r, psllw(r & (char16)63, 2), a);
8186 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8187 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8188 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8189 DAG.getConstant(2, MVT::i32));
8190 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8191 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8194 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8196 // return pblendv(r, r+r, a);
8197 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8198 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8199 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
8205 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
8206 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
8207 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
8208 // looks for this combo and may remove the "setcc" instruction if the "setcc"
8209 // has only one use.
8210 SDNode *N = Op.getNode();
8211 SDValue LHS = N->getOperand(0);
8212 SDValue RHS = N->getOperand(1);
8213 unsigned BaseOp = 0;
8215 DebugLoc dl = Op.getDebugLoc();
8217 switch (Op.getOpcode()) {
8218 default: llvm_unreachable("Unknown ovf instruction!");
8220 // A subtract of one will be selected as a INC. Note that INC doesn't
8221 // set CF, so we can't do this for UADDO.
8222 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8223 if (C->getAPIntValue() == 1) {
8224 BaseOp = X86ISD::INC;
8228 BaseOp = X86ISD::ADD;
8232 BaseOp = X86ISD::ADD;
8236 // A subtract of one will be selected as a DEC. Note that DEC doesn't
8237 // set CF, so we can't do this for USUBO.
8238 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8239 if (C->getAPIntValue() == 1) {
8240 BaseOp = X86ISD::DEC;
8244 BaseOp = X86ISD::SUB;
8248 BaseOp = X86ISD::SUB;
8252 BaseOp = X86ISD::SMUL;
8256 BaseOp = X86ISD::UMUL;
8261 // Also sets EFLAGS.
8262 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
8263 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
8266 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
8267 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
8269 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8273 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
8274 DebugLoc dl = Op.getDebugLoc();
8276 if (!Subtarget->hasSSE2()) {
8277 SDValue Chain = Op.getOperand(0);
8278 SDValue Zero = DAG.getConstant(0,
8279 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8281 DAG.getRegister(X86::ESP, MVT::i32), // Base
8282 DAG.getTargetConstant(1, MVT::i8), // Scale
8283 DAG.getRegister(0, MVT::i32), // Index
8284 DAG.getTargetConstant(0, MVT::i32), // Disp
8285 DAG.getRegister(0, MVT::i32), // Segment.
8290 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
8291 array_lengthof(Ops));
8292 return SDValue(Res, 0);
8295 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
8297 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
8299 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
8300 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
8301 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
8302 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
8304 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
8305 if (!Op1 && !Op2 && !Op3 && Op4)
8306 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
8308 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
8309 if (Op1 && !Op2 && !Op3 && !Op4)
8310 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
8312 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
8314 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
8317 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
8318 EVT T = Op.getValueType();
8319 DebugLoc dl = Op.getDebugLoc();
8322 switch(T.getSimpleVT().SimpleTy) {
8324 assert(false && "Invalid value type!");
8325 case MVT::i8: Reg = X86::AL; size = 1; break;
8326 case MVT::i16: Reg = X86::AX; size = 2; break;
8327 case MVT::i32: Reg = X86::EAX; size = 4; break;
8329 assert(Subtarget->is64Bit() && "Node not type legal!");
8330 Reg = X86::RAX; size = 8;
8333 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
8334 Op.getOperand(2), SDValue());
8335 SDValue Ops[] = { cpIn.getValue(0),
8338 DAG.getTargetConstant(size, MVT::i8),
8340 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8341 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
8343 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
8347 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
8348 SelectionDAG &DAG) const {
8349 assert(Subtarget->is64Bit() && "Result not type legalized?");
8350 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8351 SDValue TheChain = Op.getOperand(0);
8352 DebugLoc dl = Op.getDebugLoc();
8353 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8354 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
8355 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
8357 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
8358 DAG.getConstant(32, MVT::i8));
8360 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
8363 return DAG.getMergeValues(Ops, 2, dl);
8366 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
8367 SelectionDAG &DAG) const {
8368 EVT SrcVT = Op.getOperand(0).getValueType();
8369 EVT DstVT = Op.getValueType();
8370 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
8371 Subtarget->hasMMX() && !DisableMMX) &&
8372 "Unexpected custom BIT_CONVERT");
8373 assert((DstVT == MVT::i64 ||
8374 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
8375 "Unexpected custom BIT_CONVERT");
8376 // i64 <=> MMX conversions are Legal.
8377 if (SrcVT==MVT::i64 && DstVT.isVector())
8379 if (DstVT==MVT::i64 && SrcVT.isVector())
8381 // MMX <=> MMX conversions are Legal.
8382 if (SrcVT.isVector() && DstVT.isVector())
8384 // All other conversions need to be expanded.
8387 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
8388 SDNode *Node = Op.getNode();
8389 DebugLoc dl = Node->getDebugLoc();
8390 EVT T = Node->getValueType(0);
8391 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
8392 DAG.getConstant(0, T), Node->getOperand(2));
8393 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
8394 cast<AtomicSDNode>(Node)->getMemoryVT(),
8395 Node->getOperand(0),
8396 Node->getOperand(1), negOp,
8397 cast<AtomicSDNode>(Node)->getSrcValue(),
8398 cast<AtomicSDNode>(Node)->getAlignment());
8401 /// LowerOperation - Provide custom lowering hooks for some operations.
8403 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
8404 switch (Op.getOpcode()) {
8405 default: llvm_unreachable("Should not custom lower this!");
8406 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
8407 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
8408 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
8409 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
8410 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
8411 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
8412 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
8413 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
8414 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
8415 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
8416 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
8417 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
8418 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
8419 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
8420 case ISD::SHL_PARTS:
8421 case ISD::SRA_PARTS:
8422 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
8423 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
8424 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
8425 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
8426 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
8427 case ISD::FABS: return LowerFABS(Op, DAG);
8428 case ISD::FNEG: return LowerFNEG(Op, DAG);
8429 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
8430 case ISD::SETCC: return LowerSETCC(Op, DAG);
8431 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
8432 case ISD::SELECT: return LowerSELECT(Op, DAG);
8433 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
8434 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
8435 case ISD::VASTART: return LowerVASTART(Op, DAG);
8436 case ISD::VAARG: return LowerVAARG(Op, DAG);
8437 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
8438 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
8439 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
8440 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
8441 case ISD::FRAME_TO_ARGS_OFFSET:
8442 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
8443 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
8444 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
8445 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
8446 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
8447 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
8448 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
8449 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
8450 case ISD::SHL: return LowerSHL(Op, DAG);
8456 case ISD::UMULO: return LowerXALUO(Op, DAG);
8457 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
8458 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
8462 void X86TargetLowering::
8463 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
8464 SelectionDAG &DAG, unsigned NewOp) const {
8465 EVT T = Node->getValueType(0);
8466 DebugLoc dl = Node->getDebugLoc();
8467 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
8469 SDValue Chain = Node->getOperand(0);
8470 SDValue In1 = Node->getOperand(1);
8471 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8472 Node->getOperand(2), DAG.getIntPtrConstant(0));
8473 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8474 Node->getOperand(2), DAG.getIntPtrConstant(1));
8475 SDValue Ops[] = { Chain, In1, In2L, In2H };
8476 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
8478 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
8479 cast<MemSDNode>(Node)->getMemOperand());
8480 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
8481 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8482 Results.push_back(Result.getValue(2));
8485 /// ReplaceNodeResults - Replace a node with an illegal result type
8486 /// with a new node built out of custom code.
8487 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
8488 SmallVectorImpl<SDValue>&Results,
8489 SelectionDAG &DAG) const {
8490 DebugLoc dl = N->getDebugLoc();
8491 switch (N->getOpcode()) {
8493 assert(false && "Do not know how to custom type legalize this operation!");
8495 case ISD::FP_TO_SINT: {
8496 std::pair<SDValue,SDValue> Vals =
8497 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
8498 SDValue FIST = Vals.first, StackSlot = Vals.second;
8499 if (FIST.getNode() != 0) {
8500 EVT VT = N->getValueType(0);
8501 // Return a load from the stack slot.
8502 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
8507 case ISD::READCYCLECOUNTER: {
8508 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8509 SDValue TheChain = N->getOperand(0);
8510 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8511 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
8513 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
8515 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
8516 SDValue Ops[] = { eax, edx };
8517 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
8518 Results.push_back(edx.getValue(1));
8521 case ISD::ATOMIC_CMP_SWAP: {
8522 EVT T = N->getValueType(0);
8523 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
8524 SDValue cpInL, cpInH;
8525 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8526 DAG.getConstant(0, MVT::i32));
8527 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8528 DAG.getConstant(1, MVT::i32));
8529 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
8530 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
8532 SDValue swapInL, swapInH;
8533 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8534 DAG.getConstant(0, MVT::i32));
8535 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8536 DAG.getConstant(1, MVT::i32));
8537 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
8539 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
8540 swapInL.getValue(1));
8541 SDValue Ops[] = { swapInH.getValue(0),
8543 swapInH.getValue(1) };
8544 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8545 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
8546 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
8547 MVT::i32, Result.getValue(1));
8548 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
8549 MVT::i32, cpOutL.getValue(2));
8550 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
8551 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8552 Results.push_back(cpOutH.getValue(1));
8555 case ISD::ATOMIC_LOAD_ADD:
8556 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
8558 case ISD::ATOMIC_LOAD_AND:
8559 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
8561 case ISD::ATOMIC_LOAD_NAND:
8562 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
8564 case ISD::ATOMIC_LOAD_OR:
8565 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
8567 case ISD::ATOMIC_LOAD_SUB:
8568 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
8570 case ISD::ATOMIC_LOAD_XOR:
8571 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
8573 case ISD::ATOMIC_SWAP:
8574 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
8579 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
8581 default: return NULL;
8582 case X86ISD::BSF: return "X86ISD::BSF";
8583 case X86ISD::BSR: return "X86ISD::BSR";
8584 case X86ISD::SHLD: return "X86ISD::SHLD";
8585 case X86ISD::SHRD: return "X86ISD::SHRD";
8586 case X86ISD::FAND: return "X86ISD::FAND";
8587 case X86ISD::FOR: return "X86ISD::FOR";
8588 case X86ISD::FXOR: return "X86ISD::FXOR";
8589 case X86ISD::FSRL: return "X86ISD::FSRL";
8590 case X86ISD::FILD: return "X86ISD::FILD";
8591 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
8592 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
8593 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
8594 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
8595 case X86ISD::FLD: return "X86ISD::FLD";
8596 case X86ISD::FST: return "X86ISD::FST";
8597 case X86ISD::CALL: return "X86ISD::CALL";
8598 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
8599 case X86ISD::BT: return "X86ISD::BT";
8600 case X86ISD::CMP: return "X86ISD::CMP";
8601 case X86ISD::COMI: return "X86ISD::COMI";
8602 case X86ISD::UCOMI: return "X86ISD::UCOMI";
8603 case X86ISD::SETCC: return "X86ISD::SETCC";
8604 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
8605 case X86ISD::CMOV: return "X86ISD::CMOV";
8606 case X86ISD::BRCOND: return "X86ISD::BRCOND";
8607 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
8608 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
8609 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
8610 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
8611 case X86ISD::Wrapper: return "X86ISD::Wrapper";
8612 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
8613 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
8614 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
8615 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
8616 case X86ISD::PINSRB: return "X86ISD::PINSRB";
8617 case X86ISD::PINSRW: return "X86ISD::PINSRW";
8618 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
8619 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
8620 case X86ISD::FMAX: return "X86ISD::FMAX";
8621 case X86ISD::FMIN: return "X86ISD::FMIN";
8622 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
8623 case X86ISD::FRCP: return "X86ISD::FRCP";
8624 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
8625 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
8626 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
8627 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
8628 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
8629 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
8630 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
8631 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
8632 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
8633 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
8634 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
8635 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
8636 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
8637 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
8638 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
8639 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
8640 case X86ISD::VSHL: return "X86ISD::VSHL";
8641 case X86ISD::VSRL: return "X86ISD::VSRL";
8642 case X86ISD::CMPPD: return "X86ISD::CMPPD";
8643 case X86ISD::CMPPS: return "X86ISD::CMPPS";
8644 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
8645 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
8646 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
8647 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
8648 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
8649 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
8650 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
8651 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
8652 case X86ISD::ADD: return "X86ISD::ADD";
8653 case X86ISD::SUB: return "X86ISD::SUB";
8654 case X86ISD::SMUL: return "X86ISD::SMUL";
8655 case X86ISD::UMUL: return "X86ISD::UMUL";
8656 case X86ISD::INC: return "X86ISD::INC";
8657 case X86ISD::DEC: return "X86ISD::DEC";
8658 case X86ISD::OR: return "X86ISD::OR";
8659 case X86ISD::XOR: return "X86ISD::XOR";
8660 case X86ISD::AND: return "X86ISD::AND";
8661 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
8662 case X86ISD::PTEST: return "X86ISD::PTEST";
8663 case X86ISD::TESTP: return "X86ISD::TESTP";
8664 case X86ISD::PALIGN: return "X86ISD::PALIGN";
8665 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
8666 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
8667 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
8668 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
8669 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
8670 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
8671 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
8672 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
8673 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
8674 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
8675 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
8676 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
8677 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
8678 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
8679 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
8680 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
8681 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
8682 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
8683 case X86ISD::MOVSD: return "X86ISD::MOVSD";
8684 case X86ISD::MOVSS: return "X86ISD::MOVSS";
8685 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
8686 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
8687 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
8688 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
8689 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
8690 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
8691 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
8692 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
8693 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
8694 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
8695 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
8696 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
8697 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
8698 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
8702 // isLegalAddressingMode - Return true if the addressing mode represented
8703 // by AM is legal for this target, for a load/store of the specified type.
8704 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
8705 const Type *Ty) const {
8706 // X86 supports extremely general addressing modes.
8707 CodeModel::Model M = getTargetMachine().getCodeModel();
8708 Reloc::Model R = getTargetMachine().getRelocationModel();
8710 // X86 allows a sign-extended 32-bit immediate field as a displacement.
8711 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
8716 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
8718 // If a reference to this global requires an extra load, we can't fold it.
8719 if (isGlobalStubReference(GVFlags))
8722 // If BaseGV requires a register for the PIC base, we cannot also have a
8723 // BaseReg specified.
8724 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
8727 // If lower 4G is not available, then we must use rip-relative addressing.
8728 if ((M != CodeModel::Small || R != Reloc::Static) &&
8729 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
8739 // These scales always work.
8744 // These scales are formed with basereg+scalereg. Only accept if there is
8749 default: // Other stuff never works.
8757 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
8758 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
8760 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
8761 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
8762 if (NumBits1 <= NumBits2)
8767 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
8768 if (!VT1.isInteger() || !VT2.isInteger())
8770 unsigned NumBits1 = VT1.getSizeInBits();
8771 unsigned NumBits2 = VT2.getSizeInBits();
8772 if (NumBits1 <= NumBits2)
8777 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
8778 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8779 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
8782 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
8783 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8784 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
8787 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
8788 // i16 instructions are longer (0x66 prefix) and potentially slower.
8789 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
8792 /// isShuffleMaskLegal - Targets can use this to indicate that they only
8793 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
8794 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
8795 /// are assumed to be legal.
8797 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
8799 // Very little shuffling can be done for 64-bit vectors right now.
8800 if (VT.getSizeInBits() == 64)
8801 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8803 // FIXME: pshufb, blends, shifts.
8804 return (VT.getVectorNumElements() == 2 ||
8805 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8806 isMOVLMask(M, VT) ||
8807 isSHUFPMask(M, VT) ||
8808 isPSHUFDMask(M, VT) ||
8809 isPSHUFHWMask(M, VT) ||
8810 isPSHUFLWMask(M, VT) ||
8811 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8812 isUNPCKLMask(M, VT) ||
8813 isUNPCKHMask(M, VT) ||
8814 isUNPCKL_v_undef_Mask(M, VT) ||
8815 isUNPCKH_v_undef_Mask(M, VT));
8819 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8821 unsigned NumElts = VT.getVectorNumElements();
8822 // FIXME: This collection of masks seems suspect.
8825 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8826 return (isMOVLMask(Mask, VT) ||
8827 isCommutedMOVLMask(Mask, VT, true) ||
8828 isSHUFPMask(Mask, VT) ||
8829 isCommutedSHUFPMask(Mask, VT));
8834 //===----------------------------------------------------------------------===//
8835 // X86 Scheduler Hooks
8836 //===----------------------------------------------------------------------===//
8838 // private utility function
8840 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8841 MachineBasicBlock *MBB,
8848 TargetRegisterClass *RC,
8849 bool invSrc) const {
8850 // For the atomic bitwise operator, we generate
8853 // ld t1 = [bitinstr.addr]
8854 // op t2 = t1, [bitinstr.val]
8856 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8858 // fallthrough -->nextMBB
8859 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8860 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8861 MachineFunction::iterator MBBIter = MBB;
8864 /// First build the CFG
8865 MachineFunction *F = MBB->getParent();
8866 MachineBasicBlock *thisMBB = MBB;
8867 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8868 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8869 F->insert(MBBIter, newMBB);
8870 F->insert(MBBIter, nextMBB);
8872 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8873 nextMBB->splice(nextMBB->begin(), thisMBB,
8874 llvm::next(MachineBasicBlock::iterator(bInstr)),
8876 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8878 // Update thisMBB to fall through to newMBB
8879 thisMBB->addSuccessor(newMBB);
8881 // newMBB jumps to itself and fall through to nextMBB
8882 newMBB->addSuccessor(nextMBB);
8883 newMBB->addSuccessor(newMBB);
8885 // Insert instructions into newMBB based on incoming instruction
8886 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8887 "unexpected number of operands");
8888 DebugLoc dl = bInstr->getDebugLoc();
8889 MachineOperand& destOper = bInstr->getOperand(0);
8890 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8891 int numArgs = bInstr->getNumOperands() - 1;
8892 for (int i=0; i < numArgs; ++i)
8893 argOpers[i] = &bInstr->getOperand(i+1);
8895 // x86 address has 4 operands: base, index, scale, and displacement
8896 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8897 int valArgIndx = lastAddrIndx + 1;
8899 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8900 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8901 for (int i=0; i <= lastAddrIndx; ++i)
8902 (*MIB).addOperand(*argOpers[i]);
8904 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8906 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8911 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8912 assert((argOpers[valArgIndx]->isReg() ||
8913 argOpers[valArgIndx]->isImm()) &&
8915 if (argOpers[valArgIndx]->isReg())
8916 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8918 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8920 (*MIB).addOperand(*argOpers[valArgIndx]);
8922 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
8925 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8926 for (int i=0; i <= lastAddrIndx; ++i)
8927 (*MIB).addOperand(*argOpers[i]);
8929 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8930 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8931 bInstr->memoperands_end());
8933 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8937 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8939 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8943 // private utility function: 64 bit atomics on 32 bit host.
8945 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8946 MachineBasicBlock *MBB,
8951 bool invSrc) const {
8952 // For the atomic bitwise operator, we generate
8953 // thisMBB (instructions are in pairs, except cmpxchg8b)
8954 // ld t1,t2 = [bitinstr.addr]
8956 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8957 // op t5, t6 <- out1, out2, [bitinstr.val]
8958 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8959 // mov ECX, EBX <- t5, t6
8960 // mov EAX, EDX <- t1, t2
8961 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8962 // mov t3, t4 <- EAX, EDX
8964 // result in out1, out2
8965 // fallthrough -->nextMBB
8967 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8968 const unsigned LoadOpc = X86::MOV32rm;
8969 const unsigned NotOpc = X86::NOT32r;
8970 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8971 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8972 MachineFunction::iterator MBBIter = MBB;
8975 /// First build the CFG
8976 MachineFunction *F = MBB->getParent();
8977 MachineBasicBlock *thisMBB = MBB;
8978 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8979 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8980 F->insert(MBBIter, newMBB);
8981 F->insert(MBBIter, nextMBB);
8983 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8984 nextMBB->splice(nextMBB->begin(), thisMBB,
8985 llvm::next(MachineBasicBlock::iterator(bInstr)),
8987 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8989 // Update thisMBB to fall through to newMBB
8990 thisMBB->addSuccessor(newMBB);
8992 // newMBB jumps to itself and fall through to nextMBB
8993 newMBB->addSuccessor(nextMBB);
8994 newMBB->addSuccessor(newMBB);
8996 DebugLoc dl = bInstr->getDebugLoc();
8997 // Insert instructions into newMBB based on incoming instruction
8998 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8999 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
9000 "unexpected number of operands");
9001 MachineOperand& dest1Oper = bInstr->getOperand(0);
9002 MachineOperand& dest2Oper = bInstr->getOperand(1);
9003 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9004 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
9005 argOpers[i] = &bInstr->getOperand(i+2);
9007 // We use some of the operands multiple times, so conservatively just
9008 // clear any kill flags that might be present.
9009 if (argOpers[i]->isReg() && argOpers[i]->isUse())
9010 argOpers[i]->setIsKill(false);
9013 // x86 address has 5 operands: base, index, scale, displacement, and segment.
9014 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9016 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9017 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
9018 for (int i=0; i <= lastAddrIndx; ++i)
9019 (*MIB).addOperand(*argOpers[i]);
9020 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9021 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
9022 // add 4 to displacement.
9023 for (int i=0; i <= lastAddrIndx-2; ++i)
9024 (*MIB).addOperand(*argOpers[i]);
9025 MachineOperand newOp3 = *(argOpers[3]);
9027 newOp3.setImm(newOp3.getImm()+4);
9029 newOp3.setOffset(newOp3.getOffset()+4);
9030 (*MIB).addOperand(newOp3);
9031 (*MIB).addOperand(*argOpers[lastAddrIndx]);
9033 // t3/4 are defined later, at the bottom of the loop
9034 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
9035 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
9036 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
9037 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
9038 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
9039 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
9041 // The subsequent operations should be using the destination registers of
9042 //the PHI instructions.
9044 t1 = F->getRegInfo().createVirtualRegister(RC);
9045 t2 = F->getRegInfo().createVirtualRegister(RC);
9046 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
9047 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
9049 t1 = dest1Oper.getReg();
9050 t2 = dest2Oper.getReg();
9053 int valArgIndx = lastAddrIndx + 1;
9054 assert((argOpers[valArgIndx]->isReg() ||
9055 argOpers[valArgIndx]->isImm()) &&
9057 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
9058 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
9059 if (argOpers[valArgIndx]->isReg())
9060 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
9062 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
9063 if (regOpcL != X86::MOV32rr)
9065 (*MIB).addOperand(*argOpers[valArgIndx]);
9066 assert(argOpers[valArgIndx + 1]->isReg() ==
9067 argOpers[valArgIndx]->isReg());
9068 assert(argOpers[valArgIndx + 1]->isImm() ==
9069 argOpers[valArgIndx]->isImm());
9070 if (argOpers[valArgIndx + 1]->isReg())
9071 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
9073 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
9074 if (regOpcH != X86::MOV32rr)
9076 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
9078 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9080 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
9083 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
9085 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
9088 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
9089 for (int i=0; i <= lastAddrIndx; ++i)
9090 (*MIB).addOperand(*argOpers[i]);
9092 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9093 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9094 bInstr->memoperands_end());
9096 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
9097 MIB.addReg(X86::EAX);
9098 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
9099 MIB.addReg(X86::EDX);
9102 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9104 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9108 // private utility function
9110 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
9111 MachineBasicBlock *MBB,
9112 unsigned cmovOpc) const {
9113 // For the atomic min/max operator, we generate
9116 // ld t1 = [min/max.addr]
9117 // mov t2 = [min/max.val]
9119 // cmov[cond] t2 = t1
9121 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9123 // fallthrough -->nextMBB
9125 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9126 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9127 MachineFunction::iterator MBBIter = MBB;
9130 /// First build the CFG
9131 MachineFunction *F = MBB->getParent();
9132 MachineBasicBlock *thisMBB = MBB;
9133 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9134 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9135 F->insert(MBBIter, newMBB);
9136 F->insert(MBBIter, nextMBB);
9138 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9139 nextMBB->splice(nextMBB->begin(), thisMBB,
9140 llvm::next(MachineBasicBlock::iterator(mInstr)),
9142 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9144 // Update thisMBB to fall through to newMBB
9145 thisMBB->addSuccessor(newMBB);
9147 // newMBB jumps to newMBB and fall through to nextMBB
9148 newMBB->addSuccessor(nextMBB);
9149 newMBB->addSuccessor(newMBB);
9151 DebugLoc dl = mInstr->getDebugLoc();
9152 // Insert instructions into newMBB based on incoming instruction
9153 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9154 "unexpected number of operands");
9155 MachineOperand& destOper = mInstr->getOperand(0);
9156 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9157 int numArgs = mInstr->getNumOperands() - 1;
9158 for (int i=0; i < numArgs; ++i)
9159 argOpers[i] = &mInstr->getOperand(i+1);
9161 // x86 address has 4 operands: base, index, scale, and displacement
9162 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9163 int valArgIndx = lastAddrIndx + 1;
9165 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9166 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
9167 for (int i=0; i <= lastAddrIndx; ++i)
9168 (*MIB).addOperand(*argOpers[i]);
9170 // We only support register and immediate values
9171 assert((argOpers[valArgIndx]->isReg() ||
9172 argOpers[valArgIndx]->isImm()) &&
9175 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9176 if (argOpers[valArgIndx]->isReg())
9177 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
9179 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
9180 (*MIB).addOperand(*argOpers[valArgIndx]);
9182 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9185 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
9190 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9191 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
9195 // Cmp and exchange if none has modified the memory location
9196 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
9197 for (int i=0; i <= lastAddrIndx; ++i)
9198 (*MIB).addOperand(*argOpers[i]);
9200 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9201 (*MIB).setMemRefs(mInstr->memoperands_begin(),
9202 mInstr->memoperands_end());
9204 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9205 MIB.addReg(X86::EAX);
9208 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9210 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
9214 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
9215 // or XMM0_V32I8 in AVX all of this code can be replaced with that
9218 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
9219 unsigned numArgs, bool memArg) const {
9221 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
9222 "Target must have SSE4.2 or AVX features enabled");
9224 DebugLoc dl = MI->getDebugLoc();
9225 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9229 if (!Subtarget->hasAVX()) {
9231 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
9233 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
9236 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
9238 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
9241 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
9243 for (unsigned i = 0; i < numArgs; ++i) {
9244 MachineOperand &Op = MI->getOperand(i+1);
9246 if (!(Op.isReg() && Op.isImplicit()))
9250 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
9253 MI->eraseFromParent();
9259 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
9261 MachineBasicBlock *MBB) const {
9262 // Emit code to save XMM registers to the stack. The ABI says that the
9263 // number of registers to save is given in %al, so it's theoretically
9264 // possible to do an indirect jump trick to avoid saving all of them,
9265 // however this code takes a simpler approach and just executes all
9266 // of the stores if %al is non-zero. It's less code, and it's probably
9267 // easier on the hardware branch predictor, and stores aren't all that
9268 // expensive anyway.
9270 // Create the new basic blocks. One block contains all the XMM stores,
9271 // and one block is the final destination regardless of whether any
9272 // stores were performed.
9273 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9274 MachineFunction *F = MBB->getParent();
9275 MachineFunction::iterator MBBIter = MBB;
9277 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
9278 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
9279 F->insert(MBBIter, XMMSaveMBB);
9280 F->insert(MBBIter, EndMBB);
9282 // Transfer the remainder of MBB and its successor edges to EndMBB.
9283 EndMBB->splice(EndMBB->begin(), MBB,
9284 llvm::next(MachineBasicBlock::iterator(MI)),
9286 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
9288 // The original block will now fall through to the XMM save block.
9289 MBB->addSuccessor(XMMSaveMBB);
9290 // The XMMSaveMBB will fall through to the end block.
9291 XMMSaveMBB->addSuccessor(EndMBB);
9293 // Now add the instructions.
9294 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9295 DebugLoc DL = MI->getDebugLoc();
9297 unsigned CountReg = MI->getOperand(0).getReg();
9298 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
9299 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
9301 if (!Subtarget->isTargetWin64()) {
9302 // If %al is 0, branch around the XMM save block.
9303 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
9304 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
9305 MBB->addSuccessor(EndMBB);
9308 // In the XMM save block, save all the XMM argument registers.
9309 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
9310 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
9311 MachineMemOperand *MMO =
9312 F->getMachineMemOperand(
9313 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
9314 MachineMemOperand::MOStore, Offset,
9315 /*Size=*/16, /*Align=*/16);
9316 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
9317 .addFrameIndex(RegSaveFrameIndex)
9318 .addImm(/*Scale=*/1)
9319 .addReg(/*IndexReg=*/0)
9320 .addImm(/*Disp=*/Offset)
9321 .addReg(/*Segment=*/0)
9322 .addReg(MI->getOperand(i).getReg())
9323 .addMemOperand(MMO);
9326 MI->eraseFromParent(); // The pseudo instruction is gone now.
9332 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
9333 MachineBasicBlock *BB) const {
9334 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9335 DebugLoc DL = MI->getDebugLoc();
9337 // To "insert" a SELECT_CC instruction, we actually have to insert the
9338 // diamond control-flow pattern. The incoming instruction knows the
9339 // destination vreg to set, the condition code register to branch on, the
9340 // true/false values to select between, and a branch opcode to use.
9341 const BasicBlock *LLVM_BB = BB->getBasicBlock();
9342 MachineFunction::iterator It = BB;
9348 // cmpTY ccX, r1, r2
9350 // fallthrough --> copy0MBB
9351 MachineBasicBlock *thisMBB = BB;
9352 MachineFunction *F = BB->getParent();
9353 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
9354 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
9355 F->insert(It, copy0MBB);
9356 F->insert(It, sinkMBB);
9358 // If the EFLAGS register isn't dead in the terminator, then claim that it's
9359 // live into the sink and copy blocks.
9360 const MachineFunction *MF = BB->getParent();
9361 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
9362 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
9364 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
9365 const MachineOperand &MO = MI->getOperand(I);
9366 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
9367 unsigned Reg = MO.getReg();
9368 if (Reg != X86::EFLAGS) continue;
9369 copy0MBB->addLiveIn(Reg);
9370 sinkMBB->addLiveIn(Reg);
9373 // Transfer the remainder of BB and its successor edges to sinkMBB.
9374 sinkMBB->splice(sinkMBB->begin(), BB,
9375 llvm::next(MachineBasicBlock::iterator(MI)),
9377 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
9379 // Add the true and fallthrough blocks as its successors.
9380 BB->addSuccessor(copy0MBB);
9381 BB->addSuccessor(sinkMBB);
9383 // Create the conditional branch instruction.
9385 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
9386 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
9389 // %FalseValue = ...
9390 // # fallthrough to sinkMBB
9391 copy0MBB->addSuccessor(sinkMBB);
9394 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
9396 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
9397 TII->get(X86::PHI), MI->getOperand(0).getReg())
9398 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
9399 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
9401 MI->eraseFromParent(); // The pseudo instruction is gone now.
9406 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
9407 MachineBasicBlock *BB) const {
9408 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9409 DebugLoc DL = MI->getDebugLoc();
9411 // The lowering is pretty easy: we're just emitting the call to _alloca. The
9412 // non-trivial part is impdef of ESP.
9413 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
9416 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
9417 .addExternalSymbol("_alloca")
9418 .addReg(X86::EAX, RegState::Implicit)
9419 .addReg(X86::ESP, RegState::Implicit)
9420 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
9421 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
9422 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
9424 MI->eraseFromParent(); // The pseudo instruction is gone now.
9429 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
9430 MachineBasicBlock *BB) const {
9431 // This is pretty easy. We're taking the value that we received from
9432 // our load from the relocation, sticking it in either RDI (x86-64)
9433 // or EAX and doing an indirect call. The return value will then
9434 // be in the normal return register.
9435 const X86InstrInfo *TII
9436 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
9437 DebugLoc DL = MI->getDebugLoc();
9438 MachineFunction *F = BB->getParent();
9439 bool IsWin64 = Subtarget->isTargetWin64();
9441 assert(MI->getOperand(3).isGlobal() && "This should be a global");
9443 if (Subtarget->is64Bit()) {
9444 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9445 TII->get(X86::MOV64rm), X86::RDI)
9447 .addImm(0).addReg(0)
9448 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9449 MI->getOperand(3).getTargetFlags())
9451 MIB = BuildMI(*BB, MI, DL, TII->get(IsWin64 ? X86::WINCALL64m : X86::CALL64m));
9452 addDirectMem(MIB, X86::RDI);
9453 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
9454 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9455 TII->get(X86::MOV32rm), X86::EAX)
9457 .addImm(0).addReg(0)
9458 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9459 MI->getOperand(3).getTargetFlags())
9461 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9462 addDirectMem(MIB, X86::EAX);
9464 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9465 TII->get(X86::MOV32rm), X86::EAX)
9466 .addReg(TII->getGlobalBaseReg(F))
9467 .addImm(0).addReg(0)
9468 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9469 MI->getOperand(3).getTargetFlags())
9471 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9472 addDirectMem(MIB, X86::EAX);
9475 MI->eraseFromParent(); // The pseudo instruction is gone now.
9480 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
9481 MachineBasicBlock *BB) const {
9482 switch (MI->getOpcode()) {
9483 default: assert(false && "Unexpected instr type to insert");
9484 case X86::MINGW_ALLOCA:
9485 return EmitLoweredMingwAlloca(MI, BB);
9486 case X86::TLSCall_32:
9487 case X86::TLSCall_64:
9488 return EmitLoweredTLSCall(MI, BB);
9490 case X86::CMOV_V1I64:
9491 case X86::CMOV_FR32:
9492 case X86::CMOV_FR64:
9493 case X86::CMOV_V4F32:
9494 case X86::CMOV_V2F64:
9495 case X86::CMOV_V2I64:
9496 case X86::CMOV_GR16:
9497 case X86::CMOV_GR32:
9498 case X86::CMOV_RFP32:
9499 case X86::CMOV_RFP64:
9500 case X86::CMOV_RFP80:
9501 return EmitLoweredSelect(MI, BB);
9503 case X86::FP32_TO_INT16_IN_MEM:
9504 case X86::FP32_TO_INT32_IN_MEM:
9505 case X86::FP32_TO_INT64_IN_MEM:
9506 case X86::FP64_TO_INT16_IN_MEM:
9507 case X86::FP64_TO_INT32_IN_MEM:
9508 case X86::FP64_TO_INT64_IN_MEM:
9509 case X86::FP80_TO_INT16_IN_MEM:
9510 case X86::FP80_TO_INT32_IN_MEM:
9511 case X86::FP80_TO_INT64_IN_MEM: {
9512 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9513 DebugLoc DL = MI->getDebugLoc();
9515 // Change the floating point control register to use "round towards zero"
9516 // mode when truncating to an integer value.
9517 MachineFunction *F = BB->getParent();
9518 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
9519 addFrameReference(BuildMI(*BB, MI, DL,
9520 TII->get(X86::FNSTCW16m)), CWFrameIdx);
9522 // Load the old value of the high byte of the control word...
9524 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
9525 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
9528 // Set the high part to be round to zero...
9529 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
9532 // Reload the modified control word now...
9533 addFrameReference(BuildMI(*BB, MI, DL,
9534 TII->get(X86::FLDCW16m)), CWFrameIdx);
9536 // Restore the memory image of control word to original value
9537 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
9540 // Get the X86 opcode to use.
9542 switch (MI->getOpcode()) {
9543 default: llvm_unreachable("illegal opcode!");
9544 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
9545 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
9546 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
9547 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
9548 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
9549 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
9550 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
9551 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
9552 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
9556 MachineOperand &Op = MI->getOperand(0);
9558 AM.BaseType = X86AddressMode::RegBase;
9559 AM.Base.Reg = Op.getReg();
9561 AM.BaseType = X86AddressMode::FrameIndexBase;
9562 AM.Base.FrameIndex = Op.getIndex();
9564 Op = MI->getOperand(1);
9566 AM.Scale = Op.getImm();
9567 Op = MI->getOperand(2);
9569 AM.IndexReg = Op.getImm();
9570 Op = MI->getOperand(3);
9571 if (Op.isGlobal()) {
9572 AM.GV = Op.getGlobal();
9574 AM.Disp = Op.getImm();
9576 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
9577 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
9579 // Reload the original control word now.
9580 addFrameReference(BuildMI(*BB, MI, DL,
9581 TII->get(X86::FLDCW16m)), CWFrameIdx);
9583 MI->eraseFromParent(); // The pseudo instruction is gone now.
9586 // String/text processing lowering.
9587 case X86::PCMPISTRM128REG:
9588 case X86::VPCMPISTRM128REG:
9589 return EmitPCMP(MI, BB, 3, false /* in-mem */);
9590 case X86::PCMPISTRM128MEM:
9591 case X86::VPCMPISTRM128MEM:
9592 return EmitPCMP(MI, BB, 3, true /* in-mem */);
9593 case X86::PCMPESTRM128REG:
9594 case X86::VPCMPESTRM128REG:
9595 return EmitPCMP(MI, BB, 5, false /* in mem */);
9596 case X86::PCMPESTRM128MEM:
9597 case X86::VPCMPESTRM128MEM:
9598 return EmitPCMP(MI, BB, 5, true /* in mem */);
9601 case X86::ATOMAND32:
9602 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9603 X86::AND32ri, X86::MOV32rm,
9605 X86::NOT32r, X86::EAX,
9606 X86::GR32RegisterClass);
9608 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
9609 X86::OR32ri, X86::MOV32rm,
9611 X86::NOT32r, X86::EAX,
9612 X86::GR32RegisterClass);
9613 case X86::ATOMXOR32:
9614 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
9615 X86::XOR32ri, X86::MOV32rm,
9617 X86::NOT32r, X86::EAX,
9618 X86::GR32RegisterClass);
9619 case X86::ATOMNAND32:
9620 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9621 X86::AND32ri, X86::MOV32rm,
9623 X86::NOT32r, X86::EAX,
9624 X86::GR32RegisterClass, true);
9625 case X86::ATOMMIN32:
9626 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
9627 case X86::ATOMMAX32:
9628 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
9629 case X86::ATOMUMIN32:
9630 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
9631 case X86::ATOMUMAX32:
9632 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
9634 case X86::ATOMAND16:
9635 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9636 X86::AND16ri, X86::MOV16rm,
9638 X86::NOT16r, X86::AX,
9639 X86::GR16RegisterClass);
9641 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
9642 X86::OR16ri, X86::MOV16rm,
9644 X86::NOT16r, X86::AX,
9645 X86::GR16RegisterClass);
9646 case X86::ATOMXOR16:
9647 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
9648 X86::XOR16ri, X86::MOV16rm,
9650 X86::NOT16r, X86::AX,
9651 X86::GR16RegisterClass);
9652 case X86::ATOMNAND16:
9653 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9654 X86::AND16ri, X86::MOV16rm,
9656 X86::NOT16r, X86::AX,
9657 X86::GR16RegisterClass, true);
9658 case X86::ATOMMIN16:
9659 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
9660 case X86::ATOMMAX16:
9661 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
9662 case X86::ATOMUMIN16:
9663 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
9664 case X86::ATOMUMAX16:
9665 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
9668 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9669 X86::AND8ri, X86::MOV8rm,
9671 X86::NOT8r, X86::AL,
9672 X86::GR8RegisterClass);
9674 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
9675 X86::OR8ri, X86::MOV8rm,
9677 X86::NOT8r, X86::AL,
9678 X86::GR8RegisterClass);
9680 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
9681 X86::XOR8ri, X86::MOV8rm,
9683 X86::NOT8r, X86::AL,
9684 X86::GR8RegisterClass);
9685 case X86::ATOMNAND8:
9686 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9687 X86::AND8ri, X86::MOV8rm,
9689 X86::NOT8r, X86::AL,
9690 X86::GR8RegisterClass, true);
9691 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
9692 // This group is for 64-bit host.
9693 case X86::ATOMAND64:
9694 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9695 X86::AND64ri32, X86::MOV64rm,
9697 X86::NOT64r, X86::RAX,
9698 X86::GR64RegisterClass);
9700 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
9701 X86::OR64ri32, X86::MOV64rm,
9703 X86::NOT64r, X86::RAX,
9704 X86::GR64RegisterClass);
9705 case X86::ATOMXOR64:
9706 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
9707 X86::XOR64ri32, X86::MOV64rm,
9709 X86::NOT64r, X86::RAX,
9710 X86::GR64RegisterClass);
9711 case X86::ATOMNAND64:
9712 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9713 X86::AND64ri32, X86::MOV64rm,
9715 X86::NOT64r, X86::RAX,
9716 X86::GR64RegisterClass, true);
9717 case X86::ATOMMIN64:
9718 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
9719 case X86::ATOMMAX64:
9720 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
9721 case X86::ATOMUMIN64:
9722 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
9723 case X86::ATOMUMAX64:
9724 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
9726 // This group does 64-bit operations on a 32-bit host.
9727 case X86::ATOMAND6432:
9728 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9729 X86::AND32rr, X86::AND32rr,
9730 X86::AND32ri, X86::AND32ri,
9732 case X86::ATOMOR6432:
9733 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9734 X86::OR32rr, X86::OR32rr,
9735 X86::OR32ri, X86::OR32ri,
9737 case X86::ATOMXOR6432:
9738 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9739 X86::XOR32rr, X86::XOR32rr,
9740 X86::XOR32ri, X86::XOR32ri,
9742 case X86::ATOMNAND6432:
9743 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9744 X86::AND32rr, X86::AND32rr,
9745 X86::AND32ri, X86::AND32ri,
9747 case X86::ATOMADD6432:
9748 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9749 X86::ADD32rr, X86::ADC32rr,
9750 X86::ADD32ri, X86::ADC32ri,
9752 case X86::ATOMSUB6432:
9753 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9754 X86::SUB32rr, X86::SBB32rr,
9755 X86::SUB32ri, X86::SBB32ri,
9757 case X86::ATOMSWAP6432:
9758 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9759 X86::MOV32rr, X86::MOV32rr,
9760 X86::MOV32ri, X86::MOV32ri,
9762 case X86::VASTART_SAVE_XMM_REGS:
9763 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
9767 //===----------------------------------------------------------------------===//
9768 // X86 Optimization Hooks
9769 //===----------------------------------------------------------------------===//
9771 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9775 const SelectionDAG &DAG,
9776 unsigned Depth) const {
9777 unsigned Opc = Op.getOpcode();
9778 assert((Opc >= ISD::BUILTIN_OP_END ||
9779 Opc == ISD::INTRINSIC_WO_CHAIN ||
9780 Opc == ISD::INTRINSIC_W_CHAIN ||
9781 Opc == ISD::INTRINSIC_VOID) &&
9782 "Should use MaskedValueIsZero if you don't know whether Op"
9783 " is a target node!");
9785 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
9797 // These nodes' second result is a boolean.
9798 if (Op.getResNo() == 0)
9802 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
9803 Mask.getBitWidth() - 1);
9808 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
9809 /// node is a GlobalAddress + offset.
9810 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
9811 const GlobalValue* &GA,
9812 int64_t &Offset) const {
9813 if (N->getOpcode() == X86ISD::Wrapper) {
9814 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
9815 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
9816 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
9820 return TargetLowering::isGAPlusOffset(N, GA, Offset);
9823 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
9824 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
9825 /// if the load addresses are consecutive, non-overlapping, and in the right
9827 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
9828 const TargetLowering &TLI) {
9829 DebugLoc dl = N->getDebugLoc();
9830 EVT VT = N->getValueType(0);
9832 if (VT.getSizeInBits() != 128)
9835 SmallVector<SDValue, 16> Elts;
9836 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
9837 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
9839 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
9842 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
9843 /// generation and convert it from being a bunch of shuffles and extracts
9844 /// to a simple store and scalar loads to extract the elements.
9845 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
9846 const TargetLowering &TLI) {
9847 SDValue InputVector = N->getOperand(0);
9849 // Only operate on vectors of 4 elements, where the alternative shuffling
9850 // gets to be more expensive.
9851 if (InputVector.getValueType() != MVT::v4i32)
9854 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
9855 // single use which is a sign-extend or zero-extend, and all elements are
9857 SmallVector<SDNode *, 4> Uses;
9858 unsigned ExtractedElements = 0;
9859 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
9860 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
9861 if (UI.getUse().getResNo() != InputVector.getResNo())
9864 SDNode *Extract = *UI;
9865 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9868 if (Extract->getValueType(0) != MVT::i32)
9870 if (!Extract->hasOneUse())
9872 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9873 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9875 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9878 // Record which element was extracted.
9879 ExtractedElements |=
9880 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9882 Uses.push_back(Extract);
9885 // If not all the elements were used, this may not be worthwhile.
9886 if (ExtractedElements != 15)
9889 // Ok, we've now decided to do the transformation.
9890 DebugLoc dl = InputVector.getDebugLoc();
9892 // Store the value to a temporary stack slot.
9893 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9894 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL,
9895 0, false, false, 0);
9897 // Replace each use (extract) with a load of the appropriate element.
9898 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9899 UE = Uses.end(); UI != UE; ++UI) {
9900 SDNode *Extract = *UI;
9902 // Compute the element's address.
9903 SDValue Idx = Extract->getOperand(1);
9905 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9906 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9907 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9909 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
9910 OffsetVal, StackPtr);
9913 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
9914 ScalarAddr, NULL, 0, false, false, 0);
9916 // Replace the exact with the load.
9917 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9920 // The replacement was made in place; don't return anything.
9924 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9925 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9926 const X86Subtarget *Subtarget) {
9927 DebugLoc DL = N->getDebugLoc();
9928 SDValue Cond = N->getOperand(0);
9929 // Get the LHS/RHS of the select.
9930 SDValue LHS = N->getOperand(1);
9931 SDValue RHS = N->getOperand(2);
9933 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9934 // instructions match the semantics of the common C idiom x<y?x:y but not
9935 // x<=y?x:y, because of how they handle negative zero (which can be
9936 // ignored in unsafe-math mode).
9937 if (Subtarget->hasSSE2() &&
9938 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9939 Cond.getOpcode() == ISD::SETCC) {
9940 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9942 unsigned Opcode = 0;
9943 // Check for x CC y ? x : y.
9944 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9945 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9949 // Converting this to a min would handle NaNs incorrectly, and swapping
9950 // the operands would cause it to handle comparisons between positive
9951 // and negative zero incorrectly.
9952 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9953 if (!UnsafeFPMath &&
9954 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9956 std::swap(LHS, RHS);
9958 Opcode = X86ISD::FMIN;
9961 // Converting this to a min would handle comparisons between positive
9962 // and negative zero incorrectly.
9963 if (!UnsafeFPMath &&
9964 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9966 Opcode = X86ISD::FMIN;
9969 // Converting this to a min would handle both negative zeros and NaNs
9970 // incorrectly, but we can swap the operands to fix both.
9971 std::swap(LHS, RHS);
9975 Opcode = X86ISD::FMIN;
9979 // Converting this to a max would handle comparisons between positive
9980 // and negative zero incorrectly.
9981 if (!UnsafeFPMath &&
9982 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9984 Opcode = X86ISD::FMAX;
9987 // Converting this to a max would handle NaNs incorrectly, and swapping
9988 // the operands would cause it to handle comparisons between positive
9989 // and negative zero incorrectly.
9990 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9991 if (!UnsafeFPMath &&
9992 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9994 std::swap(LHS, RHS);
9996 Opcode = X86ISD::FMAX;
9999 // Converting this to a max would handle both negative zeros and NaNs
10000 // incorrectly, but we can swap the operands to fix both.
10001 std::swap(LHS, RHS);
10005 Opcode = X86ISD::FMAX;
10008 // Check for x CC y ? y : x -- a min/max with reversed arms.
10009 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
10010 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
10014 // Converting this to a min would handle comparisons between positive
10015 // and negative zero incorrectly, and swapping the operands would
10016 // cause it to handle NaNs incorrectly.
10017 if (!UnsafeFPMath &&
10018 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
10019 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10021 std::swap(LHS, RHS);
10023 Opcode = X86ISD::FMIN;
10026 // Converting this to a min would handle NaNs incorrectly.
10027 if (!UnsafeFPMath &&
10028 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
10030 Opcode = X86ISD::FMIN;
10033 // Converting this to a min would handle both negative zeros and NaNs
10034 // incorrectly, but we can swap the operands to fix both.
10035 std::swap(LHS, RHS);
10039 Opcode = X86ISD::FMIN;
10043 // Converting this to a max would handle NaNs incorrectly.
10044 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10046 Opcode = X86ISD::FMAX;
10049 // Converting this to a max would handle comparisons between positive
10050 // and negative zero incorrectly, and swapping the operands would
10051 // cause it to handle NaNs incorrectly.
10052 if (!UnsafeFPMath &&
10053 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
10054 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10056 std::swap(LHS, RHS);
10058 Opcode = X86ISD::FMAX;
10061 // Converting this to a max would handle both negative zeros and NaNs
10062 // incorrectly, but we can swap the operands to fix both.
10063 std::swap(LHS, RHS);
10067 Opcode = X86ISD::FMAX;
10073 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
10076 // If this is a select between two integer constants, try to do some
10078 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
10079 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
10080 // Don't do this for crazy integer types.
10081 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
10082 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
10083 // so that TrueC (the true value) is larger than FalseC.
10084 bool NeedsCondInvert = false;
10086 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
10087 // Efficiently invertible.
10088 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
10089 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
10090 isa<ConstantSDNode>(Cond.getOperand(1))))) {
10091 NeedsCondInvert = true;
10092 std::swap(TrueC, FalseC);
10095 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
10096 if (FalseC->getAPIntValue() == 0 &&
10097 TrueC->getAPIntValue().isPowerOf2()) {
10098 if (NeedsCondInvert) // Invert the condition if needed.
10099 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10100 DAG.getConstant(1, Cond.getValueType()));
10102 // Zero extend the condition if needed.
10103 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
10105 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10106 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
10107 DAG.getConstant(ShAmt, MVT::i8));
10110 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
10111 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10112 if (NeedsCondInvert) // Invert the condition if needed.
10113 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10114 DAG.getConstant(1, Cond.getValueType()));
10116 // Zero extend the condition if needed.
10117 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10118 FalseC->getValueType(0), Cond);
10119 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10120 SDValue(FalseC, 0));
10123 // Optimize cases that will turn into an LEA instruction. This requires
10124 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10125 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10126 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10127 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10129 bool isFastMultiplier = false;
10131 switch ((unsigned char)Diff) {
10133 case 1: // result = add base, cond
10134 case 2: // result = lea base( , cond*2)
10135 case 3: // result = lea base(cond, cond*2)
10136 case 4: // result = lea base( , cond*4)
10137 case 5: // result = lea base(cond, cond*4)
10138 case 8: // result = lea base( , cond*8)
10139 case 9: // result = lea base(cond, cond*8)
10140 isFastMultiplier = true;
10145 if (isFastMultiplier) {
10146 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10147 if (NeedsCondInvert) // Invert the condition if needed.
10148 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10149 DAG.getConstant(1, Cond.getValueType()));
10151 // Zero extend the condition if needed.
10152 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10154 // Scale the condition by the difference.
10156 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10157 DAG.getConstant(Diff, Cond.getValueType()));
10159 // Add the base if non-zero.
10160 if (FalseC->getAPIntValue() != 0)
10161 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10162 SDValue(FalseC, 0));
10172 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
10173 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
10174 TargetLowering::DAGCombinerInfo &DCI) {
10175 DebugLoc DL = N->getDebugLoc();
10177 // If the flag operand isn't dead, don't touch this CMOV.
10178 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
10181 // If this is a select between two integer constants, try to do some
10182 // optimizations. Note that the operands are ordered the opposite of SELECT
10184 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
10185 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
10186 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
10187 // larger than FalseC (the false value).
10188 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
10190 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
10191 CC = X86::GetOppositeBranchCondition(CC);
10192 std::swap(TrueC, FalseC);
10195 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
10196 // This is efficient for any integer data type (including i8/i16) and
10198 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
10199 SDValue Cond = N->getOperand(3);
10200 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10201 DAG.getConstant(CC, MVT::i8), Cond);
10203 // Zero extend the condition if needed.
10204 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
10206 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10207 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
10208 DAG.getConstant(ShAmt, MVT::i8));
10209 if (N->getNumValues() == 2) // Dead flag value?
10210 return DCI.CombineTo(N, Cond, SDValue());
10214 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
10215 // for any integer data type, including i8/i16.
10216 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10217 SDValue Cond = N->getOperand(3);
10218 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10219 DAG.getConstant(CC, MVT::i8), Cond);
10221 // Zero extend the condition if needed.
10222 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10223 FalseC->getValueType(0), Cond);
10224 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10225 SDValue(FalseC, 0));
10227 if (N->getNumValues() == 2) // Dead flag value?
10228 return DCI.CombineTo(N, Cond, SDValue());
10232 // Optimize cases that will turn into an LEA instruction. This requires
10233 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10234 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10235 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10236 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10238 bool isFastMultiplier = false;
10240 switch ((unsigned char)Diff) {
10242 case 1: // result = add base, cond
10243 case 2: // result = lea base( , cond*2)
10244 case 3: // result = lea base(cond, cond*2)
10245 case 4: // result = lea base( , cond*4)
10246 case 5: // result = lea base(cond, cond*4)
10247 case 8: // result = lea base( , cond*8)
10248 case 9: // result = lea base(cond, cond*8)
10249 isFastMultiplier = true;
10254 if (isFastMultiplier) {
10255 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10256 SDValue Cond = N->getOperand(3);
10257 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10258 DAG.getConstant(CC, MVT::i8), Cond);
10259 // Zero extend the condition if needed.
10260 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10262 // Scale the condition by the difference.
10264 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10265 DAG.getConstant(Diff, Cond.getValueType()));
10267 // Add the base if non-zero.
10268 if (FalseC->getAPIntValue() != 0)
10269 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10270 SDValue(FalseC, 0));
10271 if (N->getNumValues() == 2) // Dead flag value?
10272 return DCI.CombineTo(N, Cond, SDValue());
10282 /// PerformMulCombine - Optimize a single multiply with constant into two
10283 /// in order to implement it with two cheaper instructions, e.g.
10284 /// LEA + SHL, LEA + LEA.
10285 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
10286 TargetLowering::DAGCombinerInfo &DCI) {
10287 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
10290 EVT VT = N->getValueType(0);
10291 if (VT != MVT::i64)
10294 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
10297 uint64_t MulAmt = C->getZExtValue();
10298 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
10301 uint64_t MulAmt1 = 0;
10302 uint64_t MulAmt2 = 0;
10303 if ((MulAmt % 9) == 0) {
10305 MulAmt2 = MulAmt / 9;
10306 } else if ((MulAmt % 5) == 0) {
10308 MulAmt2 = MulAmt / 5;
10309 } else if ((MulAmt % 3) == 0) {
10311 MulAmt2 = MulAmt / 3;
10314 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
10315 DebugLoc DL = N->getDebugLoc();
10317 if (isPowerOf2_64(MulAmt2) &&
10318 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
10319 // If second multiplifer is pow2, issue it first. We want the multiply by
10320 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
10322 std::swap(MulAmt1, MulAmt2);
10325 if (isPowerOf2_64(MulAmt1))
10326 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
10327 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
10329 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
10330 DAG.getConstant(MulAmt1, VT));
10332 if (isPowerOf2_64(MulAmt2))
10333 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
10334 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
10336 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
10337 DAG.getConstant(MulAmt2, VT));
10339 // Do not add new nodes to DAG combiner worklist.
10340 DCI.CombineTo(N, NewMul, false);
10345 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
10346 SDValue N0 = N->getOperand(0);
10347 SDValue N1 = N->getOperand(1);
10348 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
10349 EVT VT = N0.getValueType();
10351 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
10352 // since the result of setcc_c is all zero's or all ones.
10353 if (N1C && N0.getOpcode() == ISD::AND &&
10354 N0.getOperand(1).getOpcode() == ISD::Constant) {
10355 SDValue N00 = N0.getOperand(0);
10356 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
10357 ((N00.getOpcode() == ISD::ANY_EXTEND ||
10358 N00.getOpcode() == ISD::ZERO_EXTEND) &&
10359 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
10360 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
10361 APInt ShAmt = N1C->getAPIntValue();
10362 Mask = Mask.shl(ShAmt);
10364 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
10365 N00, DAG.getConstant(Mask, VT));
10372 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
10374 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
10375 const X86Subtarget *Subtarget) {
10376 EVT VT = N->getValueType(0);
10377 if (!VT.isVector() && VT.isInteger() &&
10378 N->getOpcode() == ISD::SHL)
10379 return PerformSHLCombine(N, DAG);
10381 // On X86 with SSE2 support, we can transform this to a vector shift if
10382 // all elements are shifted by the same amount. We can't do this in legalize
10383 // because the a constant vector is typically transformed to a constant pool
10384 // so we have no knowledge of the shift amount.
10385 if (!Subtarget->hasSSE2())
10388 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
10391 SDValue ShAmtOp = N->getOperand(1);
10392 EVT EltVT = VT.getVectorElementType();
10393 DebugLoc DL = N->getDebugLoc();
10394 SDValue BaseShAmt = SDValue();
10395 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
10396 unsigned NumElts = VT.getVectorNumElements();
10398 for (; i != NumElts; ++i) {
10399 SDValue Arg = ShAmtOp.getOperand(i);
10400 if (Arg.getOpcode() == ISD::UNDEF) continue;
10404 for (; i != NumElts; ++i) {
10405 SDValue Arg = ShAmtOp.getOperand(i);
10406 if (Arg.getOpcode() == ISD::UNDEF) continue;
10407 if (Arg != BaseShAmt) {
10411 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
10412 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
10413 SDValue InVec = ShAmtOp.getOperand(0);
10414 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
10415 unsigned NumElts = InVec.getValueType().getVectorNumElements();
10417 for (; i != NumElts; ++i) {
10418 SDValue Arg = InVec.getOperand(i);
10419 if (Arg.getOpcode() == ISD::UNDEF) continue;
10423 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
10424 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
10425 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
10426 if (C->getZExtValue() == SplatIdx)
10427 BaseShAmt = InVec.getOperand(1);
10430 if (BaseShAmt.getNode() == 0)
10431 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
10432 DAG.getIntPtrConstant(0));
10436 // The shift amount is an i32.
10437 if (EltVT.bitsGT(MVT::i32))
10438 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
10439 else if (EltVT.bitsLT(MVT::i32))
10440 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
10442 // The shift amount is identical so we can do a vector shift.
10443 SDValue ValOp = N->getOperand(0);
10444 switch (N->getOpcode()) {
10446 llvm_unreachable("Unknown shift opcode!");
10449 if (VT == MVT::v2i64)
10450 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10451 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
10453 if (VT == MVT::v4i32)
10454 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10455 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
10457 if (VT == MVT::v8i16)
10458 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10459 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
10463 if (VT == MVT::v4i32)
10464 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10465 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
10467 if (VT == MVT::v8i16)
10468 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10469 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
10473 if (VT == MVT::v2i64)
10474 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10475 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
10477 if (VT == MVT::v4i32)
10478 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10479 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
10481 if (VT == MVT::v8i16)
10482 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10483 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
10490 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
10491 TargetLowering::DAGCombinerInfo &DCI,
10492 const X86Subtarget *Subtarget) {
10493 if (DCI.isBeforeLegalizeOps())
10496 EVT VT = N->getValueType(0);
10497 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
10500 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
10501 SDValue N0 = N->getOperand(0);
10502 SDValue N1 = N->getOperand(1);
10503 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
10505 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
10507 if (!N0.hasOneUse() || !N1.hasOneUse())
10510 SDValue ShAmt0 = N0.getOperand(1);
10511 if (ShAmt0.getValueType() != MVT::i8)
10513 SDValue ShAmt1 = N1.getOperand(1);
10514 if (ShAmt1.getValueType() != MVT::i8)
10516 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
10517 ShAmt0 = ShAmt0.getOperand(0);
10518 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
10519 ShAmt1 = ShAmt1.getOperand(0);
10521 DebugLoc DL = N->getDebugLoc();
10522 unsigned Opc = X86ISD::SHLD;
10523 SDValue Op0 = N0.getOperand(0);
10524 SDValue Op1 = N1.getOperand(0);
10525 if (ShAmt0.getOpcode() == ISD::SUB) {
10526 Opc = X86ISD::SHRD;
10527 std::swap(Op0, Op1);
10528 std::swap(ShAmt0, ShAmt1);
10531 unsigned Bits = VT.getSizeInBits();
10532 if (ShAmt1.getOpcode() == ISD::SUB) {
10533 SDValue Sum = ShAmt1.getOperand(0);
10534 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
10535 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
10536 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
10537 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
10538 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
10539 return DAG.getNode(Opc, DL, VT,
10541 DAG.getNode(ISD::TRUNCATE, DL,
10544 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
10545 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
10547 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
10548 return DAG.getNode(Opc, DL, VT,
10549 N0.getOperand(0), N1.getOperand(0),
10550 DAG.getNode(ISD::TRUNCATE, DL,
10557 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
10558 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
10559 const X86Subtarget *Subtarget) {
10560 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
10561 // the FP state in cases where an emms may be missing.
10562 // A preferable solution to the general problem is to figure out the right
10563 // places to insert EMMS. This qualifies as a quick hack.
10565 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
10566 StoreSDNode *St = cast<StoreSDNode>(N);
10567 EVT VT = St->getValue().getValueType();
10568 if (VT.getSizeInBits() != 64)
10571 const Function *F = DAG.getMachineFunction().getFunction();
10572 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
10573 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
10574 && Subtarget->hasSSE2();
10575 if ((VT.isVector() ||
10576 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
10577 isa<LoadSDNode>(St->getValue()) &&
10578 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
10579 St->getChain().hasOneUse() && !St->isVolatile()) {
10580 SDNode* LdVal = St->getValue().getNode();
10581 LoadSDNode *Ld = 0;
10582 int TokenFactorIndex = -1;
10583 SmallVector<SDValue, 8> Ops;
10584 SDNode* ChainVal = St->getChain().getNode();
10585 // Must be a store of a load. We currently handle two cases: the load
10586 // is a direct child, and it's under an intervening TokenFactor. It is
10587 // possible to dig deeper under nested TokenFactors.
10588 if (ChainVal == LdVal)
10589 Ld = cast<LoadSDNode>(St->getChain());
10590 else if (St->getValue().hasOneUse() &&
10591 ChainVal->getOpcode() == ISD::TokenFactor) {
10592 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
10593 if (ChainVal->getOperand(i).getNode() == LdVal) {
10594 TokenFactorIndex = i;
10595 Ld = cast<LoadSDNode>(St->getValue());
10597 Ops.push_back(ChainVal->getOperand(i));
10601 if (!Ld || !ISD::isNormalLoad(Ld))
10604 // If this is not the MMX case, i.e. we are just turning i64 load/store
10605 // into f64 load/store, avoid the transformation if there are multiple
10606 // uses of the loaded value.
10607 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
10610 DebugLoc LdDL = Ld->getDebugLoc();
10611 DebugLoc StDL = N->getDebugLoc();
10612 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
10613 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
10615 if (Subtarget->is64Bit() || F64IsLegal) {
10616 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
10617 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
10618 Ld->getBasePtr(), Ld->getSrcValue(),
10619 Ld->getSrcValueOffset(), Ld->isVolatile(),
10620 Ld->isNonTemporal(), Ld->getAlignment());
10621 SDValue NewChain = NewLd.getValue(1);
10622 if (TokenFactorIndex != -1) {
10623 Ops.push_back(NewChain);
10624 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10627 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
10628 St->getSrcValue(), St->getSrcValueOffset(),
10629 St->isVolatile(), St->isNonTemporal(),
10630 St->getAlignment());
10633 // Otherwise, lower to two pairs of 32-bit loads / stores.
10634 SDValue LoAddr = Ld->getBasePtr();
10635 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
10636 DAG.getConstant(4, MVT::i32));
10638 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
10639 Ld->getSrcValue(), Ld->getSrcValueOffset(),
10640 Ld->isVolatile(), Ld->isNonTemporal(),
10641 Ld->getAlignment());
10642 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
10643 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
10644 Ld->isVolatile(), Ld->isNonTemporal(),
10645 MinAlign(Ld->getAlignment(), 4));
10647 SDValue NewChain = LoLd.getValue(1);
10648 if (TokenFactorIndex != -1) {
10649 Ops.push_back(LoLd);
10650 Ops.push_back(HiLd);
10651 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10655 LoAddr = St->getBasePtr();
10656 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
10657 DAG.getConstant(4, MVT::i32));
10659 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
10660 St->getSrcValue(), St->getSrcValueOffset(),
10661 St->isVolatile(), St->isNonTemporal(),
10662 St->getAlignment());
10663 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
10665 St->getSrcValueOffset() + 4,
10667 St->isNonTemporal(),
10668 MinAlign(St->getAlignment(), 4));
10669 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
10674 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
10675 /// X86ISD::FXOR nodes.
10676 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
10677 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
10678 // F[X]OR(0.0, x) -> x
10679 // F[X]OR(x, 0.0) -> x
10680 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10681 if (C->getValueAPF().isPosZero())
10682 return N->getOperand(1);
10683 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10684 if (C->getValueAPF().isPosZero())
10685 return N->getOperand(0);
10689 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
10690 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
10691 // FAND(0.0, x) -> 0.0
10692 // FAND(x, 0.0) -> 0.0
10693 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10694 if (C->getValueAPF().isPosZero())
10695 return N->getOperand(0);
10696 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10697 if (C->getValueAPF().isPosZero())
10698 return N->getOperand(1);
10702 static SDValue PerformBTCombine(SDNode *N,
10704 TargetLowering::DAGCombinerInfo &DCI) {
10705 // BT ignores high bits in the bit index operand.
10706 SDValue Op1 = N->getOperand(1);
10707 if (Op1.hasOneUse()) {
10708 unsigned BitWidth = Op1.getValueSizeInBits();
10709 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
10710 APInt KnownZero, KnownOne;
10711 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
10712 !DCI.isBeforeLegalizeOps());
10713 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10714 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
10715 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
10716 DCI.CommitTargetLoweringOpt(TLO);
10721 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
10722 SDValue Op = N->getOperand(0);
10723 if (Op.getOpcode() == ISD::BIT_CONVERT)
10724 Op = Op.getOperand(0);
10725 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
10726 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
10727 VT.getVectorElementType().getSizeInBits() ==
10728 OpVT.getVectorElementType().getSizeInBits()) {
10729 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
10734 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
10735 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
10736 // (and (i32 x86isd::setcc_carry), 1)
10737 // This eliminates the zext. This transformation is necessary because
10738 // ISD::SETCC is always legalized to i8.
10739 DebugLoc dl = N->getDebugLoc();
10740 SDValue N0 = N->getOperand(0);
10741 EVT VT = N->getValueType(0);
10742 if (N0.getOpcode() == ISD::AND &&
10744 N0.getOperand(0).hasOneUse()) {
10745 SDValue N00 = N0.getOperand(0);
10746 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
10748 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
10749 if (!C || C->getZExtValue() != 1)
10751 return DAG.getNode(ISD::AND, dl, VT,
10752 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
10753 N00.getOperand(0), N00.getOperand(1)),
10754 DAG.getConstant(1, VT));
10760 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
10761 DAGCombinerInfo &DCI) const {
10762 SelectionDAG &DAG = DCI.DAG;
10763 switch (N->getOpcode()) {
10765 case ISD::EXTRACT_VECTOR_ELT:
10766 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
10767 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
10768 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
10769 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
10772 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
10773 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
10774 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
10776 case X86ISD::FOR: return PerformFORCombine(N, DAG);
10777 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
10778 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
10779 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
10780 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
10781 case X86ISD::SHUFPS: // Handle all target specific shuffles
10782 case X86ISD::SHUFPD:
10783 case X86ISD::PUNPCKHBW:
10784 case X86ISD::PUNPCKHWD:
10785 case X86ISD::PUNPCKHDQ:
10786 case X86ISD::PUNPCKHQDQ:
10787 case X86ISD::UNPCKHPS:
10788 case X86ISD::UNPCKHPD:
10789 case X86ISD::PUNPCKLBW:
10790 case X86ISD::PUNPCKLWD:
10791 case X86ISD::PUNPCKLDQ:
10792 case X86ISD::PUNPCKLQDQ:
10793 case X86ISD::UNPCKLPS:
10794 case X86ISD::UNPCKLPD:
10795 case X86ISD::MOVHLPS:
10796 case X86ISD::MOVLHPS:
10797 case X86ISD::PSHUFD:
10798 case X86ISD::PSHUFHW:
10799 case X86ISD::PSHUFLW:
10800 case X86ISD::MOVSS:
10801 case X86ISD::MOVSD:
10802 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
10808 /// isTypeDesirableForOp - Return true if the target has native support for
10809 /// the specified value type and it is 'desirable' to use the type for the
10810 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
10811 /// instruction encodings are longer and some i16 instructions are slow.
10812 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
10813 if (!isTypeLegal(VT))
10815 if (VT != MVT::i16)
10822 case ISD::SIGN_EXTEND:
10823 case ISD::ZERO_EXTEND:
10824 case ISD::ANY_EXTEND:
10837 /// IsDesirableToPromoteOp - This method query the target whether it is
10838 /// beneficial for dag combiner to promote the specified node. If true, it
10839 /// should return the desired promotion type by reference.
10840 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
10841 EVT VT = Op.getValueType();
10842 if (VT != MVT::i16)
10845 bool Promote = false;
10846 bool Commute = false;
10847 switch (Op.getOpcode()) {
10850 LoadSDNode *LD = cast<LoadSDNode>(Op);
10851 // If the non-extending load has a single use and it's not live out, then it
10852 // might be folded.
10853 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
10854 Op.hasOneUse()*/) {
10855 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
10856 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
10857 // The only case where we'd want to promote LOAD (rather then it being
10858 // promoted as an operand is when it's only use is liveout.
10859 if (UI->getOpcode() != ISD::CopyToReg)
10866 case ISD::SIGN_EXTEND:
10867 case ISD::ZERO_EXTEND:
10868 case ISD::ANY_EXTEND:
10873 SDValue N0 = Op.getOperand(0);
10874 // Look out for (store (shl (load), x)).
10875 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10888 SDValue N0 = Op.getOperand(0);
10889 SDValue N1 = Op.getOperand(1);
10890 if (!Commute && MayFoldLoad(N1))
10892 // Avoid disabling potential load folding opportunities.
10893 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10895 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10905 //===----------------------------------------------------------------------===//
10906 // X86 Inline Assembly Support
10907 //===----------------------------------------------------------------------===//
10909 static bool LowerToBSwap(CallInst *CI) {
10910 // FIXME: this should verify that we are targetting a 486 or better. If not,
10911 // we will turn this bswap into something that will be lowered to logical ops
10912 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10913 // so don't worry about this.
10915 // Verify this is a simple bswap.
10916 if (CI->getNumArgOperands() != 1 ||
10917 CI->getType() != CI->getArgOperand(0)->getType() ||
10918 !CI->getType()->isIntegerTy())
10921 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10922 if (!Ty || Ty->getBitWidth() % 16 != 0)
10925 // Okay, we can do this xform, do so now.
10926 const Type *Tys[] = { Ty };
10927 Module *M = CI->getParent()->getParent()->getParent();
10928 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10930 Value *Op = CI->getArgOperand(0);
10931 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10933 CI->replaceAllUsesWith(Op);
10934 CI->eraseFromParent();
10938 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10939 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10940 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10942 std::string AsmStr = IA->getAsmString();
10944 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10945 SmallVector<StringRef, 4> AsmPieces;
10946 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10948 switch (AsmPieces.size()) {
10949 default: return false;
10951 AsmStr = AsmPieces[0];
10953 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10956 if (AsmPieces.size() == 2 &&
10957 (AsmPieces[0] == "bswap" ||
10958 AsmPieces[0] == "bswapq" ||
10959 AsmPieces[0] == "bswapl") &&
10960 (AsmPieces[1] == "$0" ||
10961 AsmPieces[1] == "${0:q}")) {
10962 // No need to check constraints, nothing other than the equivalent of
10963 // "=r,0" would be valid here.
10964 return LowerToBSwap(CI);
10966 // rorw $$8, ${0:w} --> llvm.bswap.i16
10967 if (CI->getType()->isIntegerTy(16) &&
10968 AsmPieces.size() == 3 &&
10969 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10970 AsmPieces[1] == "$$8," &&
10971 AsmPieces[2] == "${0:w}" &&
10972 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10974 const std::string &Constraints = IA->getConstraintString();
10975 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10976 std::sort(AsmPieces.begin(), AsmPieces.end());
10977 if (AsmPieces.size() == 4 &&
10978 AsmPieces[0] == "~{cc}" &&
10979 AsmPieces[1] == "~{dirflag}" &&
10980 AsmPieces[2] == "~{flags}" &&
10981 AsmPieces[3] == "~{fpsr}") {
10982 return LowerToBSwap(CI);
10987 if (CI->getType()->isIntegerTy(64) &&
10988 Constraints.size() >= 2 &&
10989 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10990 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10991 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10992 SmallVector<StringRef, 4> Words;
10993 SplitString(AsmPieces[0], Words, " \t");
10994 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10996 SplitString(AsmPieces[1], Words, " \t");
10997 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10999 SplitString(AsmPieces[2], Words, " \t,");
11000 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
11001 Words[2] == "%edx") {
11002 return LowerToBSwap(CI);
11014 /// getConstraintType - Given a constraint letter, return the type of
11015 /// constraint it is for this target.
11016 X86TargetLowering::ConstraintType
11017 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
11018 if (Constraint.size() == 1) {
11019 switch (Constraint[0]) {
11031 return C_RegisterClass;
11039 return TargetLowering::getConstraintType(Constraint);
11042 /// LowerXConstraint - try to replace an X constraint, which matches anything,
11043 /// with another that has more specific requirements based on the type of the
11044 /// corresponding operand.
11045 const char *X86TargetLowering::
11046 LowerXConstraint(EVT ConstraintVT) const {
11047 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
11048 // 'f' like normal targets.
11049 if (ConstraintVT.isFloatingPoint()) {
11050 if (Subtarget->hasSSE2())
11052 if (Subtarget->hasSSE1())
11056 return TargetLowering::LowerXConstraint(ConstraintVT);
11059 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
11060 /// vector. If it is invalid, don't add anything to Ops.
11061 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
11063 std::vector<SDValue>&Ops,
11064 SelectionDAG &DAG) const {
11065 SDValue Result(0, 0);
11067 switch (Constraint) {
11070 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11071 if (C->getZExtValue() <= 31) {
11072 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11078 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11079 if (C->getZExtValue() <= 63) {
11080 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11086 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11087 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
11088 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11094 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11095 if (C->getZExtValue() <= 255) {
11096 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11102 // 32-bit signed value
11103 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11104 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
11105 C->getSExtValue())) {
11106 // Widen to 64 bits here to get it sign extended.
11107 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
11110 // FIXME gcc accepts some relocatable values here too, but only in certain
11111 // memory models; it's complicated.
11116 // 32-bit unsigned value
11117 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11118 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
11119 C->getZExtValue())) {
11120 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11124 // FIXME gcc accepts some relocatable values here too, but only in certain
11125 // memory models; it's complicated.
11129 // Literal immediates are always ok.
11130 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
11131 // Widen to 64 bits here to get it sign extended.
11132 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
11136 // In any sort of PIC mode addresses need to be computed at runtime by
11137 // adding in a register or some sort of table lookup. These can't
11138 // be used as immediates.
11139 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
11142 // If we are in non-pic codegen mode, we allow the address of a global (with
11143 // an optional displacement) to be used with 'i'.
11144 GlobalAddressSDNode *GA = 0;
11145 int64_t Offset = 0;
11147 // Match either (GA), (GA+C), (GA+C1+C2), etc.
11149 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
11150 Offset += GA->getOffset();
11152 } else if (Op.getOpcode() == ISD::ADD) {
11153 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
11154 Offset += C->getZExtValue();
11155 Op = Op.getOperand(0);
11158 } else if (Op.getOpcode() == ISD::SUB) {
11159 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
11160 Offset += -C->getZExtValue();
11161 Op = Op.getOperand(0);
11166 // Otherwise, this isn't something we can handle, reject it.
11170 const GlobalValue *GV = GA->getGlobal();
11171 // If we require an extra load to get this address, as in PIC mode, we
11172 // can't accept it.
11173 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
11174 getTargetMachine())))
11177 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
11178 GA->getValueType(0), Offset);
11183 if (Result.getNode()) {
11184 Ops.push_back(Result);
11187 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11190 std::vector<unsigned> X86TargetLowering::
11191 getRegClassForInlineAsmConstraint(const std::string &Constraint,
11193 if (Constraint.size() == 1) {
11194 // FIXME: not handling fp-stack yet!
11195 switch (Constraint[0]) { // GCC X86 Constraint Letters
11196 default: break; // Unknown constraint letter
11197 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
11198 if (Subtarget->is64Bit()) {
11199 if (VT == MVT::i32)
11200 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
11201 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
11202 X86::R10D,X86::R11D,X86::R12D,
11203 X86::R13D,X86::R14D,X86::R15D,
11204 X86::EBP, X86::ESP, 0);
11205 else if (VT == MVT::i16)
11206 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
11207 X86::SI, X86::DI, X86::R8W,X86::R9W,
11208 X86::R10W,X86::R11W,X86::R12W,
11209 X86::R13W,X86::R14W,X86::R15W,
11210 X86::BP, X86::SP, 0);
11211 else if (VT == MVT::i8)
11212 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
11213 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
11214 X86::R10B,X86::R11B,X86::R12B,
11215 X86::R13B,X86::R14B,X86::R15B,
11216 X86::BPL, X86::SPL, 0);
11218 else if (VT == MVT::i64)
11219 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
11220 X86::RSI, X86::RDI, X86::R8, X86::R9,
11221 X86::R10, X86::R11, X86::R12,
11222 X86::R13, X86::R14, X86::R15,
11223 X86::RBP, X86::RSP, 0);
11227 // 32-bit fallthrough
11228 case 'Q': // Q_REGS
11229 if (VT == MVT::i32)
11230 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
11231 else if (VT == MVT::i16)
11232 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
11233 else if (VT == MVT::i8)
11234 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
11235 else if (VT == MVT::i64)
11236 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
11241 return std::vector<unsigned>();
11244 std::pair<unsigned, const TargetRegisterClass*>
11245 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
11247 // First, see if this is a constraint that directly corresponds to an LLVM
11249 if (Constraint.size() == 1) {
11250 // GCC Constraint Letters
11251 switch (Constraint[0]) {
11253 case 'r': // GENERAL_REGS
11254 case 'l': // INDEX_REGS
11256 return std::make_pair(0U, X86::GR8RegisterClass);
11257 if (VT == MVT::i16)
11258 return std::make_pair(0U, X86::GR16RegisterClass);
11259 if (VT == MVT::i32 || !Subtarget->is64Bit())
11260 return std::make_pair(0U, X86::GR32RegisterClass);
11261 return std::make_pair(0U, X86::GR64RegisterClass);
11262 case 'R': // LEGACY_REGS
11264 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
11265 if (VT == MVT::i16)
11266 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
11267 if (VT == MVT::i32 || !Subtarget->is64Bit())
11268 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
11269 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
11270 case 'f': // FP Stack registers.
11271 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
11272 // value to the correct fpstack register class.
11273 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
11274 return std::make_pair(0U, X86::RFP32RegisterClass);
11275 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
11276 return std::make_pair(0U, X86::RFP64RegisterClass);
11277 return std::make_pair(0U, X86::RFP80RegisterClass);
11278 case 'y': // MMX_REGS if MMX allowed.
11279 if (!Subtarget->hasMMX()) break;
11280 return std::make_pair(0U, X86::VR64RegisterClass);
11281 case 'Y': // SSE_REGS if SSE2 allowed
11282 if (!Subtarget->hasSSE2()) break;
11284 case 'x': // SSE_REGS if SSE1 allowed
11285 if (!Subtarget->hasSSE1()) break;
11287 switch (VT.getSimpleVT().SimpleTy) {
11289 // Scalar SSE types.
11292 return std::make_pair(0U, X86::FR32RegisterClass);
11295 return std::make_pair(0U, X86::FR64RegisterClass);
11303 return std::make_pair(0U, X86::VR128RegisterClass);
11309 // Use the default implementation in TargetLowering to convert the register
11310 // constraint into a member of a register class.
11311 std::pair<unsigned, const TargetRegisterClass*> Res;
11312 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
11314 // Not found as a standard register?
11315 if (Res.second == 0) {
11316 // Map st(0) -> st(7) -> ST0
11317 if (Constraint.size() == 7 && Constraint[0] == '{' &&
11318 tolower(Constraint[1]) == 's' &&
11319 tolower(Constraint[2]) == 't' &&
11320 Constraint[3] == '(' &&
11321 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
11322 Constraint[5] == ')' &&
11323 Constraint[6] == '}') {
11325 Res.first = X86::ST0+Constraint[4]-'0';
11326 Res.second = X86::RFP80RegisterClass;
11330 // GCC allows "st(0)" to be called just plain "st".
11331 if (StringRef("{st}").equals_lower(Constraint)) {
11332 Res.first = X86::ST0;
11333 Res.second = X86::RFP80RegisterClass;
11338 if (StringRef("{flags}").equals_lower(Constraint)) {
11339 Res.first = X86::EFLAGS;
11340 Res.second = X86::CCRRegisterClass;
11344 // 'A' means EAX + EDX.
11345 if (Constraint == "A") {
11346 Res.first = X86::EAX;
11347 Res.second = X86::GR32_ADRegisterClass;
11353 // Otherwise, check to see if this is a register class of the wrong value
11354 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
11355 // turn into {ax},{dx}.
11356 if (Res.second->hasType(VT))
11357 return Res; // Correct type already, nothing to do.
11359 // All of the single-register GCC register classes map their values onto
11360 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
11361 // really want an 8-bit or 32-bit register, map to the appropriate register
11362 // class and return the appropriate register.
11363 if (Res.second == X86::GR16RegisterClass) {
11364 if (VT == MVT::i8) {
11365 unsigned DestReg = 0;
11366 switch (Res.first) {
11368 case X86::AX: DestReg = X86::AL; break;
11369 case X86::DX: DestReg = X86::DL; break;
11370 case X86::CX: DestReg = X86::CL; break;
11371 case X86::BX: DestReg = X86::BL; break;
11374 Res.first = DestReg;
11375 Res.second = X86::GR8RegisterClass;
11377 } else if (VT == MVT::i32) {
11378 unsigned DestReg = 0;
11379 switch (Res.first) {
11381 case X86::AX: DestReg = X86::EAX; break;
11382 case X86::DX: DestReg = X86::EDX; break;
11383 case X86::CX: DestReg = X86::ECX; break;
11384 case X86::BX: DestReg = X86::EBX; break;
11385 case X86::SI: DestReg = X86::ESI; break;
11386 case X86::DI: DestReg = X86::EDI; break;
11387 case X86::BP: DestReg = X86::EBP; break;
11388 case X86::SP: DestReg = X86::ESP; break;
11391 Res.first = DestReg;
11392 Res.second = X86::GR32RegisterClass;
11394 } else if (VT == MVT::i64) {
11395 unsigned DestReg = 0;
11396 switch (Res.first) {
11398 case X86::AX: DestReg = X86::RAX; break;
11399 case X86::DX: DestReg = X86::RDX; break;
11400 case X86::CX: DestReg = X86::RCX; break;
11401 case X86::BX: DestReg = X86::RBX; break;
11402 case X86::SI: DestReg = X86::RSI; break;
11403 case X86::DI: DestReg = X86::RDI; break;
11404 case X86::BP: DestReg = X86::RBP; break;
11405 case X86::SP: DestReg = X86::RSP; break;
11408 Res.first = DestReg;
11409 Res.second = X86::GR64RegisterClass;
11412 } else if (Res.second == X86::FR32RegisterClass ||
11413 Res.second == X86::FR64RegisterClass ||
11414 Res.second == X86::VR128RegisterClass) {
11415 // Handle references to XMM physical registers that got mapped into the
11416 // wrong class. This can happen with constraints like {xmm0} where the
11417 // target independent register mapper will just pick the first match it can
11418 // find, ignoring the required type.
11419 if (VT == MVT::f32)
11420 Res.second = X86::FR32RegisterClass;
11421 else if (VT == MVT::f64)
11422 Res.second = X86::FR64RegisterClass;
11423 else if (X86::VR128RegisterClass->hasType(VT))
11424 Res.second = X86::VR128RegisterClass;