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::x86mmx, X86::VR64RegisterClass, false);
619 // FIXME: Remove the rest of this stuff.
620 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
621 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
622 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
624 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
626 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
627 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
628 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
629 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
631 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
632 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
633 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
634 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
636 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
637 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
639 setOperationAction(ISD::AND, MVT::v8i8, Promote);
640 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
641 setOperationAction(ISD::AND, MVT::v4i16, Promote);
642 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
643 setOperationAction(ISD::AND, MVT::v2i32, Promote);
644 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
645 setOperationAction(ISD::AND, MVT::v1i64, Legal);
647 setOperationAction(ISD::OR, MVT::v8i8, Promote);
648 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
649 setOperationAction(ISD::OR, MVT::v4i16, Promote);
650 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
651 setOperationAction(ISD::OR, MVT::v2i32, Promote);
652 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
653 setOperationAction(ISD::OR, MVT::v1i64, Legal);
655 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
656 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
657 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
658 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
659 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
660 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
661 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
663 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
664 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
665 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
666 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
667 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
668 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
669 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
671 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
672 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
673 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
674 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
676 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
677 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
678 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
679 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
681 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
682 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
683 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
685 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
687 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
688 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
689 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
690 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
691 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
692 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
693 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
695 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
696 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
697 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
698 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
699 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
703 if (!UseSoftFloat && Subtarget->hasSSE1()) {
704 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
706 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
707 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
708 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
709 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
710 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
711 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
712 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
713 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
714 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
715 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
716 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
717 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
720 if (!UseSoftFloat && Subtarget->hasSSE2()) {
721 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
723 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
724 // registers cannot be used even for integer operations.
725 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
726 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
727 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
728 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
730 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
731 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
732 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
733 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
734 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
735 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
736 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
737 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
738 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
739 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
740 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
741 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
742 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
743 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
744 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
745 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
747 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
748 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
749 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
750 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
752 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
753 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
754 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
755 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
756 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
759 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
760 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
761 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
762 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
764 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
765 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
766 EVT VT = (MVT::SimpleValueType)i;
767 // Do not attempt to custom lower non-power-of-2 vectors
768 if (!isPowerOf2_32(VT.getVectorNumElements()))
770 // Do not attempt to custom lower non-128-bit vectors
771 if (!VT.is128BitVector())
773 setOperationAction(ISD::BUILD_VECTOR,
774 VT.getSimpleVT().SimpleTy, Custom);
775 setOperationAction(ISD::VECTOR_SHUFFLE,
776 VT.getSimpleVT().SimpleTy, Custom);
777 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
778 VT.getSimpleVT().SimpleTy, Custom);
781 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
782 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
783 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
784 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
788 if (Subtarget->is64Bit()) {
789 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
790 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
793 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
794 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
795 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
798 // Do not attempt to promote non-128-bit vectors
799 if (!VT.is128BitVector())
802 setOperationAction(ISD::AND, SVT, Promote);
803 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
804 setOperationAction(ISD::OR, SVT, Promote);
805 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
806 setOperationAction(ISD::XOR, SVT, Promote);
807 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
808 setOperationAction(ISD::LOAD, SVT, Promote);
809 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
810 setOperationAction(ISD::SELECT, SVT, Promote);
811 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
814 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
816 // Custom lower v2i64 and v2f64 selects.
817 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
818 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
819 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
820 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
822 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
823 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
824 if (!DisableMMX && Subtarget->hasMMX()) {
825 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
826 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
830 if (Subtarget->hasSSE41()) {
831 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
832 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
833 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
834 setOperationAction(ISD::FRINT, MVT::f32, Legal);
835 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
836 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
837 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
838 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
839 setOperationAction(ISD::FRINT, MVT::f64, Legal);
840 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
842 // FIXME: Do we need to handle scalar-to-vector here?
843 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
845 // Can turn SHL into an integer multiply.
846 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
847 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
849 // i8 and i16 vectors are custom , because the source register and source
850 // source memory operand types are not the same width. f32 vectors are
851 // custom since the immediate controlling the insert encodes additional
853 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
854 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
855 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
856 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
858 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
859 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
860 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
861 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
863 if (Subtarget->is64Bit()) {
864 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
865 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
869 if (Subtarget->hasSSE42()) {
870 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
873 if (!UseSoftFloat && Subtarget->hasAVX()) {
874 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
875 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
876 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
877 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
878 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
880 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
881 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
882 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
883 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
884 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
885 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
886 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
887 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
888 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
889 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
890 setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
891 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
892 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
893 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
894 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
896 // Operations to consider commented out -v16i16 v32i8
897 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
898 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
899 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
900 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
901 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
902 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
903 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
904 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
905 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
906 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
907 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
908 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
909 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
910 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
912 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
913 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
914 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
915 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
917 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
918 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
919 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
920 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
921 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
923 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
924 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
925 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
926 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
927 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
928 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
931 // Not sure we want to do this since there are no 256-bit integer
934 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
935 // This includes 256-bit vectors
936 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
937 EVT VT = (MVT::SimpleValueType)i;
939 // Do not attempt to custom lower non-power-of-2 vectors
940 if (!isPowerOf2_32(VT.getVectorNumElements()))
943 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
944 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
945 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
948 if (Subtarget->is64Bit()) {
949 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
950 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
955 // Not sure we want to do this since there are no 256-bit integer
958 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
959 // Including 256-bit vectors
960 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
961 EVT VT = (MVT::SimpleValueType)i;
963 if (!VT.is256BitVector()) {
966 setOperationAction(ISD::AND, VT, Promote);
967 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
968 setOperationAction(ISD::OR, VT, Promote);
969 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
970 setOperationAction(ISD::XOR, VT, Promote);
971 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
972 setOperationAction(ISD::LOAD, VT, Promote);
973 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
974 setOperationAction(ISD::SELECT, VT, Promote);
975 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
978 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
982 // We want to custom lower some of our intrinsics.
983 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
985 // Add/Sub/Mul with overflow operations are custom lowered.
986 setOperationAction(ISD::SADDO, MVT::i32, Custom);
987 setOperationAction(ISD::UADDO, MVT::i32, Custom);
988 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
989 setOperationAction(ISD::USUBO, MVT::i32, Custom);
990 setOperationAction(ISD::SMULO, MVT::i32, Custom);
992 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
993 // handle type legalization for these operations here.
995 // FIXME: We really should do custom legalization for addition and
996 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
997 // than generic legalization for 64-bit multiplication-with-overflow, though.
998 if (Subtarget->is64Bit()) {
999 setOperationAction(ISD::SADDO, MVT::i64, Custom);
1000 setOperationAction(ISD::UADDO, MVT::i64, Custom);
1001 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
1002 setOperationAction(ISD::USUBO, MVT::i64, Custom);
1003 setOperationAction(ISD::SMULO, MVT::i64, Custom);
1006 if (!Subtarget->is64Bit()) {
1007 // These libcalls are not available in 32-bit.
1008 setLibcallName(RTLIB::SHL_I128, 0);
1009 setLibcallName(RTLIB::SRL_I128, 0);
1010 setLibcallName(RTLIB::SRA_I128, 0);
1013 // We have target-specific dag combine patterns for the following nodes:
1014 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1015 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1016 setTargetDAGCombine(ISD::BUILD_VECTOR);
1017 setTargetDAGCombine(ISD::SELECT);
1018 setTargetDAGCombine(ISD::SHL);
1019 setTargetDAGCombine(ISD::SRA);
1020 setTargetDAGCombine(ISD::SRL);
1021 setTargetDAGCombine(ISD::OR);
1022 setTargetDAGCombine(ISD::STORE);
1023 setTargetDAGCombine(ISD::ZERO_EXTEND);
1024 if (Subtarget->is64Bit())
1025 setTargetDAGCombine(ISD::MUL);
1027 computeRegisterProperties();
1029 // FIXME: These should be based on subtarget info. Plus, the values should
1030 // be smaller when we are in optimizing for size mode.
1031 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1032 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1033 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1034 setPrefLoopAlignment(16);
1035 benefitFromCodePlacementOpt = true;
1039 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1044 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1045 /// the desired ByVal argument alignment.
1046 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1049 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1050 if (VTy->getBitWidth() == 128)
1052 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1053 unsigned EltAlign = 0;
1054 getMaxByValAlign(ATy->getElementType(), EltAlign);
1055 if (EltAlign > MaxAlign)
1056 MaxAlign = EltAlign;
1057 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1058 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1059 unsigned EltAlign = 0;
1060 getMaxByValAlign(STy->getElementType(i), EltAlign);
1061 if (EltAlign > MaxAlign)
1062 MaxAlign = EltAlign;
1070 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1071 /// function arguments in the caller parameter area. For X86, aggregates
1072 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1073 /// are at 4-byte boundaries.
1074 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1075 if (Subtarget->is64Bit()) {
1076 // Max of 8 and alignment of type.
1077 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1084 if (Subtarget->hasSSE1())
1085 getMaxByValAlign(Ty, Align);
1089 /// getOptimalMemOpType - Returns the target specific optimal type for load
1090 /// and store operations as a result of memset, memcpy, and memmove
1091 /// lowering. If DstAlign is zero that means it's safe to destination
1092 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1093 /// means there isn't a need to check it against alignment requirement,
1094 /// probably because the source does not need to be loaded. If
1095 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1096 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1097 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1098 /// constant so it does not need to be loaded.
1099 /// It returns EVT::Other if the type should be determined using generic
1100 /// target-independent logic.
1102 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1103 unsigned DstAlign, unsigned SrcAlign,
1104 bool NonScalarIntSafe,
1106 MachineFunction &MF) const {
1107 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1108 // linux. This is because the stack realignment code can't handle certain
1109 // cases like PR2962. This should be removed when PR2962 is fixed.
1110 const Function *F = MF.getFunction();
1111 if (NonScalarIntSafe &&
1112 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1114 (Subtarget->isUnalignedMemAccessFast() ||
1115 ((DstAlign == 0 || DstAlign >= 16) &&
1116 (SrcAlign == 0 || SrcAlign >= 16))) &&
1117 Subtarget->getStackAlignment() >= 16) {
1118 if (Subtarget->hasSSE2())
1120 if (Subtarget->hasSSE1())
1122 } else if (!MemcpyStrSrc && Size >= 8 &&
1123 !Subtarget->is64Bit() &&
1124 Subtarget->getStackAlignment() >= 8 &&
1125 Subtarget->hasSSE2()) {
1126 // Do not use f64 to lower memcpy if source is string constant. It's
1127 // better to use i32 to avoid the loads.
1131 if (Subtarget->is64Bit() && Size >= 8)
1136 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1137 /// current function. The returned value is a member of the
1138 /// MachineJumpTableInfo::JTEntryKind enum.
1139 unsigned X86TargetLowering::getJumpTableEncoding() const {
1140 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1142 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1143 Subtarget->isPICStyleGOT())
1144 return MachineJumpTableInfo::EK_Custom32;
1146 // Otherwise, use the normal jump table encoding heuristics.
1147 return TargetLowering::getJumpTableEncoding();
1150 /// getPICBaseSymbol - Return the X86-32 PIC base.
1152 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1153 MCContext &Ctx) const {
1154 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1155 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1156 Twine(MF->getFunctionNumber())+"$pb");
1161 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1162 const MachineBasicBlock *MBB,
1163 unsigned uid,MCContext &Ctx) const{
1164 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1165 Subtarget->isPICStyleGOT());
1166 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1168 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1169 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1172 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1174 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1175 SelectionDAG &DAG) const {
1176 if (!Subtarget->is64Bit())
1177 // This doesn't have DebugLoc associated with it, but is not really the
1178 // same as a Register.
1179 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1183 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1184 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1186 const MCExpr *X86TargetLowering::
1187 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1188 MCContext &Ctx) const {
1189 // X86-64 uses RIP relative addressing based on the jump table label.
1190 if (Subtarget->isPICStyleRIPRel())
1191 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1193 // Otherwise, the reference is relative to the PIC base.
1194 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1197 /// getFunctionAlignment - Return the Log2 alignment of this function.
1198 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1199 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1202 std::pair<const TargetRegisterClass*, uint8_t>
1203 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1204 const TargetRegisterClass *RRC = 0;
1206 switch (VT.getSimpleVT().SimpleTy) {
1208 return TargetLowering::findRepresentativeClass(VT);
1209 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1210 RRC = (Subtarget->is64Bit()
1211 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1213 case MVT::v8i8: case MVT::v4i16:
1214 case MVT::v2i32: case MVT::v1i64:
1215 RRC = X86::VR64RegisterClass;
1217 case MVT::f32: case MVT::f64:
1218 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1219 case MVT::v4f32: case MVT::v2f64:
1220 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1222 RRC = X86::VR128RegisterClass;
1225 return std::make_pair(RRC, Cost);
1229 X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
1230 MachineFunction &MF) const {
1231 unsigned FPDiff = RegInfo->hasFP(MF) ? 1 : 0;
1232 switch (RC->getID()) {
1235 case X86::GR32RegClassID:
1237 case X86::GR64RegClassID:
1239 case X86::VR128RegClassID:
1240 return Subtarget->is64Bit() ? 10 : 4;
1241 case X86::VR64RegClassID:
1246 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1247 unsigned &Offset) const {
1248 if (!Subtarget->isTargetLinux())
1251 if (Subtarget->is64Bit()) {
1252 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1254 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1267 //===----------------------------------------------------------------------===//
1268 // Return Value Calling Convention Implementation
1269 //===----------------------------------------------------------------------===//
1271 #include "X86GenCallingConv.inc"
1274 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1275 const SmallVectorImpl<ISD::OutputArg> &Outs,
1276 LLVMContext &Context) const {
1277 SmallVector<CCValAssign, 16> RVLocs;
1278 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1280 return CCInfo.CheckReturn(Outs, RetCC_X86);
1284 X86TargetLowering::LowerReturn(SDValue Chain,
1285 CallingConv::ID CallConv, bool isVarArg,
1286 const SmallVectorImpl<ISD::OutputArg> &Outs,
1287 const SmallVectorImpl<SDValue> &OutVals,
1288 DebugLoc dl, SelectionDAG &DAG) const {
1289 MachineFunction &MF = DAG.getMachineFunction();
1290 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1292 SmallVector<CCValAssign, 16> RVLocs;
1293 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1294 RVLocs, *DAG.getContext());
1295 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1297 // Add the regs to the liveout set for the function.
1298 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1299 for (unsigned i = 0; i != RVLocs.size(); ++i)
1300 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1301 MRI.addLiveOut(RVLocs[i].getLocReg());
1305 SmallVector<SDValue, 6> RetOps;
1306 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1307 // Operand #1 = Bytes To Pop
1308 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1311 // Copy the result values into the output registers.
1312 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1313 CCValAssign &VA = RVLocs[i];
1314 assert(VA.isRegLoc() && "Can only return in registers!");
1315 SDValue ValToCopy = OutVals[i];
1316 EVT ValVT = ValToCopy.getValueType();
1318 // If this is x86-64, and we disabled SSE, we can't return FP values
1319 if ((ValVT == MVT::f32 || ValVT == MVT::f64) &&
1320 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1321 report_fatal_error("SSE register return with SSE disabled");
1323 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1324 // llvm-gcc has never done it right and no one has noticed, so this
1325 // should be OK for now.
1326 if (ValVT == MVT::f64 &&
1327 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1328 report_fatal_error("SSE2 register return with SSE2 disabled");
1330 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1331 // the RET instruction and handled by the FP Stackifier.
1332 if (VA.getLocReg() == X86::ST0 ||
1333 VA.getLocReg() == X86::ST1) {
1334 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1335 // change the value to the FP stack register class.
1336 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1337 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1338 RetOps.push_back(ValToCopy);
1339 // Don't emit a copytoreg.
1343 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1344 // which is returned in RAX / RDX.
1345 if (Subtarget->is64Bit()) {
1346 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1347 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1348 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1349 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1352 // If we don't have SSE2 available, convert to v4f32 so the generated
1353 // register is legal.
1354 if (!Subtarget->hasSSE2())
1355 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32,ValToCopy);
1360 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1361 Flag = Chain.getValue(1);
1364 // The x86-64 ABI for returning structs by value requires that we copy
1365 // the sret argument into %rax for the return. We saved the argument into
1366 // a virtual register in the entry block, so now we copy the value out
1368 if (Subtarget->is64Bit() &&
1369 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1370 MachineFunction &MF = DAG.getMachineFunction();
1371 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1372 unsigned Reg = FuncInfo->getSRetReturnReg();
1374 "SRetReturnReg should have been set in LowerFormalArguments().");
1375 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1377 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1378 Flag = Chain.getValue(1);
1380 // RAX now acts like a return value.
1381 MRI.addLiveOut(X86::RAX);
1384 RetOps[0] = Chain; // Update chain.
1386 // Add the flag if we have it.
1388 RetOps.push_back(Flag);
1390 return DAG.getNode(X86ISD::RET_FLAG, dl,
1391 MVT::Other, &RetOps[0], RetOps.size());
1394 /// LowerCallResult - Lower the result values of a call into the
1395 /// appropriate copies out of appropriate physical registers.
1398 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1399 CallingConv::ID CallConv, bool isVarArg,
1400 const SmallVectorImpl<ISD::InputArg> &Ins,
1401 DebugLoc dl, SelectionDAG &DAG,
1402 SmallVectorImpl<SDValue> &InVals) const {
1404 // Assign locations to each value returned by this call.
1405 SmallVector<CCValAssign, 16> RVLocs;
1406 bool Is64Bit = Subtarget->is64Bit();
1407 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1408 RVLocs, *DAG.getContext());
1409 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1411 // Copy all of the result registers out of their specified physreg.
1412 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1413 CCValAssign &VA = RVLocs[i];
1414 EVT CopyVT = VA.getValVT();
1416 // If this is x86-64, and we disabled SSE, we can't return FP values
1417 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1418 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1419 report_fatal_error("SSE register return with SSE disabled");
1424 // If this is a call to a function that returns an fp value on the floating
1425 // point stack, we must guarantee the the value is popped from the stack, so
1426 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1427 // if the return value is not used. We use the FpGET_ST0 instructions
1429 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1430 // If we prefer to use the value in xmm registers, copy it out as f80 and
1431 // use a truncate to move it from fp stack reg to xmm reg.
1432 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1433 bool isST0 = VA.getLocReg() == X86::ST0;
1435 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1436 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1437 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1438 SDValue Ops[] = { Chain, InFlag };
1439 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
1441 Val = Chain.getValue(0);
1443 // Round the f80 to the right size, which also moves it to the appropriate
1445 if (CopyVT != VA.getValVT())
1446 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1447 // This truncation won't change the value.
1448 DAG.getIntPtrConstant(1));
1449 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1450 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1451 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1452 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1453 MVT::v2i64, InFlag).getValue(1);
1454 Val = Chain.getValue(0);
1455 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1456 Val, DAG.getConstant(0, MVT::i64));
1458 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1459 MVT::i64, InFlag).getValue(1);
1460 Val = Chain.getValue(0);
1462 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1464 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1465 CopyVT, InFlag).getValue(1);
1466 Val = Chain.getValue(0);
1468 InFlag = Chain.getValue(2);
1469 InVals.push_back(Val);
1476 //===----------------------------------------------------------------------===//
1477 // C & StdCall & Fast Calling Convention implementation
1478 //===----------------------------------------------------------------------===//
1479 // StdCall calling convention seems to be standard for many Windows' API
1480 // routines and around. It differs from C calling convention just a little:
1481 // callee should clean up the stack, not caller. Symbols should be also
1482 // decorated in some fancy way :) It doesn't support any vector arguments.
1483 // For info on fast calling convention see Fast Calling Convention (tail call)
1484 // implementation LowerX86_32FastCCCallTo.
1486 /// CallIsStructReturn - Determines whether a call uses struct return
1488 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1492 return Outs[0].Flags.isSRet();
1495 /// ArgsAreStructReturn - Determines whether a function uses struct
1496 /// return semantics.
1498 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1502 return Ins[0].Flags.isSRet();
1505 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1506 /// given CallingConvention value.
1507 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1508 if (Subtarget->is64Bit()) {
1509 if (CC == CallingConv::GHC)
1510 return CC_X86_64_GHC;
1511 else if (Subtarget->isTargetWin64())
1512 return CC_X86_Win64_C;
1517 if (CC == CallingConv::X86_FastCall)
1518 return CC_X86_32_FastCall;
1519 else if (CC == CallingConv::X86_ThisCall)
1520 return CC_X86_32_ThisCall;
1521 else if (CC == CallingConv::Fast)
1522 return CC_X86_32_FastCC;
1523 else if (CC == CallingConv::GHC)
1524 return CC_X86_32_GHC;
1529 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1530 /// by "Src" to address "Dst" with size and alignment information specified by
1531 /// the specific parameter attribute. The copy will be passed as a byval
1532 /// function parameter.
1534 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1535 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1537 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1539 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1540 /*isVolatile*/false, /*AlwaysInline=*/true,
1541 MachinePointerInfo(), MachinePointerInfo());
1544 /// IsTailCallConvention - Return true if the calling convention is one that
1545 /// supports tail call optimization.
1546 static bool IsTailCallConvention(CallingConv::ID CC) {
1547 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1550 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1551 /// a tailcall target by changing its ABI.
1552 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1553 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1557 X86TargetLowering::LowerMemArgument(SDValue Chain,
1558 CallingConv::ID CallConv,
1559 const SmallVectorImpl<ISD::InputArg> &Ins,
1560 DebugLoc dl, SelectionDAG &DAG,
1561 const CCValAssign &VA,
1562 MachineFrameInfo *MFI,
1564 // Create the nodes corresponding to a load from this parameter slot.
1565 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1566 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1567 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1570 // If value is passed by pointer we have address passed instead of the value
1572 if (VA.getLocInfo() == CCValAssign::Indirect)
1573 ValVT = VA.getLocVT();
1575 ValVT = VA.getValVT();
1577 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1578 // changed with more analysis.
1579 // In case of tail call optimization mark all arguments mutable. Since they
1580 // could be overwritten by lowering of arguments in case of a tail call.
1581 if (Flags.isByVal()) {
1582 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1583 VA.getLocMemOffset(), isImmutable);
1584 return DAG.getFrameIndex(FI, getPointerTy());
1586 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1587 VA.getLocMemOffset(), isImmutable);
1588 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1589 return DAG.getLoad(ValVT, dl, Chain, FIN,
1590 MachinePointerInfo::getFixedStack(FI),
1596 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1597 CallingConv::ID CallConv,
1599 const SmallVectorImpl<ISD::InputArg> &Ins,
1602 SmallVectorImpl<SDValue> &InVals)
1604 MachineFunction &MF = DAG.getMachineFunction();
1605 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1607 const Function* Fn = MF.getFunction();
1608 if (Fn->hasExternalLinkage() &&
1609 Subtarget->isTargetCygMing() &&
1610 Fn->getName() == "main")
1611 FuncInfo->setForceFramePointer(true);
1613 MachineFrameInfo *MFI = MF.getFrameInfo();
1614 bool Is64Bit = Subtarget->is64Bit();
1615 bool IsWin64 = Subtarget->isTargetWin64();
1617 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1618 "Var args not supported with calling convention fastcc or ghc");
1620 // Assign locations to all of the incoming arguments.
1621 SmallVector<CCValAssign, 16> ArgLocs;
1622 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1623 ArgLocs, *DAG.getContext());
1624 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1626 unsigned LastVal = ~0U;
1628 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1629 CCValAssign &VA = ArgLocs[i];
1630 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1632 assert(VA.getValNo() != LastVal &&
1633 "Don't support value assigned to multiple locs yet");
1634 LastVal = VA.getValNo();
1636 if (VA.isRegLoc()) {
1637 EVT RegVT = VA.getLocVT();
1638 TargetRegisterClass *RC = NULL;
1639 if (RegVT == MVT::i32)
1640 RC = X86::GR32RegisterClass;
1641 else if (Is64Bit && RegVT == MVT::i64)
1642 RC = X86::GR64RegisterClass;
1643 else if (RegVT == MVT::f32)
1644 RC = X86::FR32RegisterClass;
1645 else if (RegVT == MVT::f64)
1646 RC = X86::FR64RegisterClass;
1647 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1648 RC = X86::VR256RegisterClass;
1649 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1650 RC = X86::VR128RegisterClass;
1651 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1652 RC = X86::VR64RegisterClass;
1654 llvm_unreachable("Unknown argument type!");
1656 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1657 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1659 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1660 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1662 if (VA.getLocInfo() == CCValAssign::SExt)
1663 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1664 DAG.getValueType(VA.getValVT()));
1665 else if (VA.getLocInfo() == CCValAssign::ZExt)
1666 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1667 DAG.getValueType(VA.getValVT()));
1668 else if (VA.getLocInfo() == CCValAssign::BCvt)
1669 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1671 if (VA.isExtInLoc()) {
1672 // Handle MMX values passed in XMM regs.
1673 if (RegVT.isVector()) {
1674 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1675 ArgValue, DAG.getConstant(0, MVT::i64));
1676 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1678 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1681 assert(VA.isMemLoc());
1682 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1685 // If value is passed via pointer - do a load.
1686 if (VA.getLocInfo() == CCValAssign::Indirect)
1687 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1688 MachinePointerInfo(), false, false, 0);
1690 InVals.push_back(ArgValue);
1693 // The x86-64 ABI for returning structs by value requires that we copy
1694 // the sret argument into %rax for the return. Save the argument into
1695 // a virtual register so that we can access it from the return points.
1696 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1697 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1698 unsigned Reg = FuncInfo->getSRetReturnReg();
1700 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1701 FuncInfo->setSRetReturnReg(Reg);
1703 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1704 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1707 unsigned StackSize = CCInfo.getNextStackOffset();
1708 // Align stack specially for tail calls.
1709 if (FuncIsMadeTailCallSafe(CallConv))
1710 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1712 // If the function takes variable number of arguments, make a frame index for
1713 // the start of the first vararg value... for expansion of llvm.va_start.
1715 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1716 CallConv != CallingConv::X86_ThisCall)) {
1717 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1720 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1722 // FIXME: We should really autogenerate these arrays
1723 static const unsigned GPR64ArgRegsWin64[] = {
1724 X86::RCX, X86::RDX, X86::R8, X86::R9
1726 static const unsigned XMMArgRegsWin64[] = {
1727 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1729 static const unsigned GPR64ArgRegs64Bit[] = {
1730 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1732 static const unsigned XMMArgRegs64Bit[] = {
1733 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1734 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1736 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1739 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1740 GPR64ArgRegs = GPR64ArgRegsWin64;
1741 XMMArgRegs = XMMArgRegsWin64;
1743 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1744 GPR64ArgRegs = GPR64ArgRegs64Bit;
1745 XMMArgRegs = XMMArgRegs64Bit;
1747 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1749 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1752 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1753 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1754 "SSE register cannot be used when SSE is disabled!");
1755 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1756 "SSE register cannot be used when SSE is disabled!");
1757 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1758 // Kernel mode asks for SSE to be disabled, so don't push them
1760 TotalNumXMMRegs = 0;
1762 // For X86-64, if there are vararg parameters that are passed via
1763 // registers, then we must store them to their spots on the stack so they
1764 // may be loaded by deferencing the result of va_next.
1765 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1766 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1767 FuncInfo->setRegSaveFrameIndex(
1768 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1771 // Store the integer parameter registers.
1772 SmallVector<SDValue, 8> MemOps;
1773 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1775 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1776 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1777 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1778 DAG.getIntPtrConstant(Offset));
1779 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1780 X86::GR64RegisterClass);
1781 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1783 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1784 MachinePointerInfo::getFixedStack(
1785 FuncInfo->getRegSaveFrameIndex(), Offset),
1787 MemOps.push_back(Store);
1791 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1792 // Now store the XMM (fp + vector) parameter registers.
1793 SmallVector<SDValue, 11> SaveXMMOps;
1794 SaveXMMOps.push_back(Chain);
1796 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1797 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1798 SaveXMMOps.push_back(ALVal);
1800 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1801 FuncInfo->getRegSaveFrameIndex()));
1802 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1803 FuncInfo->getVarArgsFPOffset()));
1805 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1806 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1807 X86::VR128RegisterClass);
1808 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1809 SaveXMMOps.push_back(Val);
1811 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1813 &SaveXMMOps[0], SaveXMMOps.size()));
1816 if (!MemOps.empty())
1817 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1818 &MemOps[0], MemOps.size());
1822 // Some CCs need callee pop.
1823 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1824 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1826 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1827 // If this is an sret function, the return should pop the hidden pointer.
1828 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1829 FuncInfo->setBytesToPopOnReturn(4);
1833 // RegSaveFrameIndex is X86-64 only.
1834 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1835 if (CallConv == CallingConv::X86_FastCall ||
1836 CallConv == CallingConv::X86_ThisCall)
1837 // fastcc functions can't have varargs.
1838 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1845 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1846 SDValue StackPtr, SDValue Arg,
1847 DebugLoc dl, SelectionDAG &DAG,
1848 const CCValAssign &VA,
1849 ISD::ArgFlagsTy Flags) const {
1850 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1851 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1852 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1853 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1854 if (Flags.isByVal())
1855 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1857 return DAG.getStore(Chain, dl, Arg, PtrOff,
1858 MachinePointerInfo::getStack(LocMemOffset),
1862 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1863 /// optimization is performed and it is required.
1865 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1866 SDValue &OutRetAddr, SDValue Chain,
1867 bool IsTailCall, bool Is64Bit,
1868 int FPDiff, DebugLoc dl) const {
1869 // Adjust the Return address stack slot.
1870 EVT VT = getPointerTy();
1871 OutRetAddr = getReturnAddressFrameIndex(DAG);
1873 // Load the "old" Return address.
1874 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1876 return SDValue(OutRetAddr.getNode(), 1);
1879 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1880 /// optimization is performed and it is required (FPDiff!=0).
1882 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1883 SDValue Chain, SDValue RetAddrFrIdx,
1884 bool Is64Bit, int FPDiff, DebugLoc dl) {
1885 // Store the return address to the appropriate stack slot.
1886 if (!FPDiff) return Chain;
1887 // Calculate the new stack slot for the return address.
1888 int SlotSize = Is64Bit ? 8 : 4;
1889 int NewReturnAddrFI =
1890 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1891 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1892 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1893 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1894 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1900 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1901 CallingConv::ID CallConv, bool isVarArg,
1903 const SmallVectorImpl<ISD::OutputArg> &Outs,
1904 const SmallVectorImpl<SDValue> &OutVals,
1905 const SmallVectorImpl<ISD::InputArg> &Ins,
1906 DebugLoc dl, SelectionDAG &DAG,
1907 SmallVectorImpl<SDValue> &InVals) const {
1908 MachineFunction &MF = DAG.getMachineFunction();
1909 bool Is64Bit = Subtarget->is64Bit();
1910 bool IsStructRet = CallIsStructReturn(Outs);
1911 bool IsSibcall = false;
1914 // Check if it's really possible to do a tail call.
1915 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1916 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1917 Outs, OutVals, Ins, DAG);
1919 // Sibcalls are automatically detected tailcalls which do not require
1921 if (!GuaranteedTailCallOpt && isTailCall)
1928 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1929 "Var args not supported with calling convention fastcc or ghc");
1931 // Analyze operands of the call, assigning locations to each operand.
1932 SmallVector<CCValAssign, 16> ArgLocs;
1933 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1934 ArgLocs, *DAG.getContext());
1935 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1937 // Get a count of how many bytes are to be pushed on the stack.
1938 unsigned NumBytes = CCInfo.getNextStackOffset();
1940 // This is a sibcall. The memory operands are available in caller's
1941 // own caller's stack.
1943 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1944 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1947 if (isTailCall && !IsSibcall) {
1948 // Lower arguments at fp - stackoffset + fpdiff.
1949 unsigned NumBytesCallerPushed =
1950 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1951 FPDiff = NumBytesCallerPushed - NumBytes;
1953 // Set the delta of movement of the returnaddr stackslot.
1954 // But only set if delta is greater than previous delta.
1955 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1956 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1960 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1962 SDValue RetAddrFrIdx;
1963 // Load return adress for tail calls.
1964 if (isTailCall && FPDiff)
1965 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1966 Is64Bit, FPDiff, dl);
1968 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1969 SmallVector<SDValue, 8> MemOpChains;
1972 // Walk the register/memloc assignments, inserting copies/loads. In the case
1973 // of tail call optimization arguments are handle later.
1974 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1975 CCValAssign &VA = ArgLocs[i];
1976 EVT RegVT = VA.getLocVT();
1977 SDValue Arg = OutVals[i];
1978 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1979 bool isByVal = Flags.isByVal();
1981 // Promote the value if needed.
1982 switch (VA.getLocInfo()) {
1983 default: llvm_unreachable("Unknown loc info!");
1984 case CCValAssign::Full: break;
1985 case CCValAssign::SExt:
1986 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1988 case CCValAssign::ZExt:
1989 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1991 case CCValAssign::AExt:
1992 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1993 // Special case: passing MMX values in XMM registers.
1994 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1995 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1996 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1998 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2000 case CCValAssign::BCvt:
2001 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
2003 case CCValAssign::Indirect: {
2004 // Store the argument.
2005 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2006 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2007 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2008 MachinePointerInfo::getFixedStack(FI),
2015 if (VA.isRegLoc()) {
2016 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2017 if (isVarArg && Subtarget->isTargetWin64()) {
2018 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2019 // shadow reg if callee is a varargs function.
2020 unsigned ShadowReg = 0;
2021 switch (VA.getLocReg()) {
2022 case X86::XMM0: ShadowReg = X86::RCX; break;
2023 case X86::XMM1: ShadowReg = X86::RDX; break;
2024 case X86::XMM2: ShadowReg = X86::R8; break;
2025 case X86::XMM3: ShadowReg = X86::R9; break;
2028 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2030 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2031 assert(VA.isMemLoc());
2032 if (StackPtr.getNode() == 0)
2033 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2034 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2035 dl, DAG, VA, Flags));
2039 if (!MemOpChains.empty())
2040 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2041 &MemOpChains[0], MemOpChains.size());
2043 // Build a sequence of copy-to-reg nodes chained together with token chain
2044 // and flag operands which copy the outgoing args into registers.
2046 // Tail call byval lowering might overwrite argument registers so in case of
2047 // tail call optimization the copies to registers are lowered later.
2049 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2050 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2051 RegsToPass[i].second, InFlag);
2052 InFlag = Chain.getValue(1);
2055 if (Subtarget->isPICStyleGOT()) {
2056 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2059 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2060 DAG.getNode(X86ISD::GlobalBaseReg,
2061 DebugLoc(), getPointerTy()),
2063 InFlag = Chain.getValue(1);
2065 // If we are tail calling and generating PIC/GOT style code load the
2066 // address of the callee into ECX. The value in ecx is used as target of
2067 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2068 // for tail calls on PIC/GOT architectures. Normally we would just put the
2069 // address of GOT into ebx and then call target@PLT. But for tail calls
2070 // ebx would be restored (since ebx is callee saved) before jumping to the
2073 // Note: The actual moving to ECX is done further down.
2074 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2075 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2076 !G->getGlobal()->hasProtectedVisibility())
2077 Callee = LowerGlobalAddress(Callee, DAG);
2078 else if (isa<ExternalSymbolSDNode>(Callee))
2079 Callee = LowerExternalSymbol(Callee, DAG);
2083 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
2084 // From AMD64 ABI document:
2085 // For calls that may call functions that use varargs or stdargs
2086 // (prototype-less calls or calls to functions containing ellipsis (...) in
2087 // the declaration) %al is used as hidden argument to specify the number
2088 // of SSE registers used. The contents of %al do not need to match exactly
2089 // the number of registers, but must be an ubound on the number of SSE
2090 // registers used and is in the range 0 - 8 inclusive.
2092 // Count the number of XMM registers allocated.
2093 static const unsigned XMMArgRegs[] = {
2094 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2095 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2097 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2098 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2099 && "SSE registers cannot be used when SSE is disabled");
2101 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2102 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2103 InFlag = Chain.getValue(1);
2107 // For tail calls lower the arguments to the 'real' stack slot.
2109 // Force all the incoming stack arguments to be loaded from the stack
2110 // before any new outgoing arguments are stored to the stack, because the
2111 // outgoing stack slots may alias the incoming argument stack slots, and
2112 // the alias isn't otherwise explicit. This is slightly more conservative
2113 // than necessary, because it means that each store effectively depends
2114 // on every argument instead of just those arguments it would clobber.
2115 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2117 SmallVector<SDValue, 8> MemOpChains2;
2120 // Do not flag preceeding copytoreg stuff together with the following stuff.
2122 if (GuaranteedTailCallOpt) {
2123 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2124 CCValAssign &VA = ArgLocs[i];
2127 assert(VA.isMemLoc());
2128 SDValue Arg = OutVals[i];
2129 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2130 // Create frame index.
2131 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2132 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2133 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2134 FIN = DAG.getFrameIndex(FI, getPointerTy());
2136 if (Flags.isByVal()) {
2137 // Copy relative to framepointer.
2138 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2139 if (StackPtr.getNode() == 0)
2140 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2142 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2144 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2148 // Store relative to framepointer.
2149 MemOpChains2.push_back(
2150 DAG.getStore(ArgChain, dl, Arg, FIN,
2151 MachinePointerInfo::getFixedStack(FI),
2157 if (!MemOpChains2.empty())
2158 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2159 &MemOpChains2[0], MemOpChains2.size());
2161 // Copy arguments to their registers.
2162 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2163 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2164 RegsToPass[i].second, InFlag);
2165 InFlag = Chain.getValue(1);
2169 // Store the return address to the appropriate stack slot.
2170 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2174 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2175 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2176 // In the 64-bit large code model, we have to make all calls
2177 // through a register, since the call instruction's 32-bit
2178 // pc-relative offset may not be large enough to hold the whole
2180 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2181 // If the callee is a GlobalAddress node (quite common, every direct call
2182 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2185 // We should use extra load for direct calls to dllimported functions in
2187 const GlobalValue *GV = G->getGlobal();
2188 if (!GV->hasDLLImportLinkage()) {
2189 unsigned char OpFlags = 0;
2191 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2192 // external symbols most go through the PLT in PIC mode. If the symbol
2193 // has hidden or protected visibility, or if it is static or local, then
2194 // we don't need to use the PLT - we can directly call it.
2195 if (Subtarget->isTargetELF() &&
2196 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2197 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2198 OpFlags = X86II::MO_PLT;
2199 } else if (Subtarget->isPICStyleStubAny() &&
2200 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2201 Subtarget->getDarwinVers() < 9) {
2202 // PC-relative references to external symbols should go through $stub,
2203 // unless we're building with the leopard linker or later, which
2204 // automatically synthesizes these stubs.
2205 OpFlags = X86II::MO_DARWIN_STUB;
2208 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2209 G->getOffset(), OpFlags);
2211 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2212 unsigned char OpFlags = 0;
2214 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2215 // symbols should go through the PLT.
2216 if (Subtarget->isTargetELF() &&
2217 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2218 OpFlags = X86II::MO_PLT;
2219 } else if (Subtarget->isPICStyleStubAny() &&
2220 Subtarget->getDarwinVers() < 9) {
2221 // PC-relative references to external symbols should go through $stub,
2222 // unless we're building with the leopard linker or later, which
2223 // automatically synthesizes these stubs.
2224 OpFlags = X86II::MO_DARWIN_STUB;
2227 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2231 // Returns a chain & a flag for retval copy to use.
2232 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2233 SmallVector<SDValue, 8> Ops;
2235 if (!IsSibcall && isTailCall) {
2236 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2237 DAG.getIntPtrConstant(0, true), InFlag);
2238 InFlag = Chain.getValue(1);
2241 Ops.push_back(Chain);
2242 Ops.push_back(Callee);
2245 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2247 // Add argument registers to the end of the list so that they are known live
2249 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2250 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2251 RegsToPass[i].second.getValueType()));
2253 // Add an implicit use GOT pointer in EBX.
2254 if (!isTailCall && Subtarget->isPICStyleGOT())
2255 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2257 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2258 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64())
2259 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2261 if (InFlag.getNode())
2262 Ops.push_back(InFlag);
2266 //// If this is the first return lowered for this function, add the regs
2267 //// to the liveout set for the function.
2268 // This isn't right, although it's probably harmless on x86; liveouts
2269 // should be computed from returns not tail calls. Consider a void
2270 // function making a tail call to a function returning int.
2271 return DAG.getNode(X86ISD::TC_RETURN, dl,
2272 NodeTys, &Ops[0], Ops.size());
2275 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2276 InFlag = Chain.getValue(1);
2278 // Create the CALLSEQ_END node.
2279 unsigned NumBytesForCalleeToPush;
2280 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2281 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2282 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2283 // If this is a call to a struct-return function, the callee
2284 // pops the hidden struct pointer, so we have to push it back.
2285 // This is common for Darwin/X86, Linux & Mingw32 targets.
2286 NumBytesForCalleeToPush = 4;
2288 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2290 // Returns a flag for retval copy to use.
2292 Chain = DAG.getCALLSEQ_END(Chain,
2293 DAG.getIntPtrConstant(NumBytes, true),
2294 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2297 InFlag = Chain.getValue(1);
2300 // Handle result values, copying them out of physregs into vregs that we
2302 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2303 Ins, dl, DAG, InVals);
2307 //===----------------------------------------------------------------------===//
2308 // Fast Calling Convention (tail call) implementation
2309 //===----------------------------------------------------------------------===//
2311 // Like std call, callee cleans arguments, convention except that ECX is
2312 // reserved for storing the tail called function address. Only 2 registers are
2313 // free for argument passing (inreg). Tail call optimization is performed
2315 // * tailcallopt is enabled
2316 // * caller/callee are fastcc
2317 // On X86_64 architecture with GOT-style position independent code only local
2318 // (within module) calls are supported at the moment.
2319 // To keep the stack aligned according to platform abi the function
2320 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2321 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2322 // If a tail called function callee has more arguments than the caller the
2323 // caller needs to make sure that there is room to move the RETADDR to. This is
2324 // achieved by reserving an area the size of the argument delta right after the
2325 // original REtADDR, but before the saved framepointer or the spilled registers
2326 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2338 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2339 /// for a 16 byte align requirement.
2341 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2342 SelectionDAG& DAG) const {
2343 MachineFunction &MF = DAG.getMachineFunction();
2344 const TargetMachine &TM = MF.getTarget();
2345 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2346 unsigned StackAlignment = TFI.getStackAlignment();
2347 uint64_t AlignMask = StackAlignment - 1;
2348 int64_t Offset = StackSize;
2349 uint64_t SlotSize = TD->getPointerSize();
2350 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2351 // Number smaller than 12 so just add the difference.
2352 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2354 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2355 Offset = ((~AlignMask) & Offset) + StackAlignment +
2356 (StackAlignment-SlotSize);
2361 /// MatchingStackOffset - Return true if the given stack call argument is
2362 /// already available in the same position (relatively) of the caller's
2363 /// incoming argument stack.
2365 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2366 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2367 const X86InstrInfo *TII) {
2368 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2370 if (Arg.getOpcode() == ISD::CopyFromReg) {
2371 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2372 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2374 MachineInstr *Def = MRI->getVRegDef(VR);
2377 if (!Flags.isByVal()) {
2378 if (!TII->isLoadFromStackSlot(Def, FI))
2381 unsigned Opcode = Def->getOpcode();
2382 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2383 Def->getOperand(1).isFI()) {
2384 FI = Def->getOperand(1).getIndex();
2385 Bytes = Flags.getByValSize();
2389 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2390 if (Flags.isByVal())
2391 // ByVal argument is passed in as a pointer but it's now being
2392 // dereferenced. e.g.
2393 // define @foo(%struct.X* %A) {
2394 // tail call @bar(%struct.X* byval %A)
2397 SDValue Ptr = Ld->getBasePtr();
2398 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2401 FI = FINode->getIndex();
2405 assert(FI != INT_MAX);
2406 if (!MFI->isFixedObjectIndex(FI))
2408 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2411 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2412 /// for tail call optimization. Targets which want to do tail call
2413 /// optimization should implement this function.
2415 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2416 CallingConv::ID CalleeCC,
2418 bool isCalleeStructRet,
2419 bool isCallerStructRet,
2420 const SmallVectorImpl<ISD::OutputArg> &Outs,
2421 const SmallVectorImpl<SDValue> &OutVals,
2422 const SmallVectorImpl<ISD::InputArg> &Ins,
2423 SelectionDAG& DAG) const {
2424 if (!IsTailCallConvention(CalleeCC) &&
2425 CalleeCC != CallingConv::C)
2428 // If -tailcallopt is specified, make fastcc functions tail-callable.
2429 const MachineFunction &MF = DAG.getMachineFunction();
2430 const Function *CallerF = DAG.getMachineFunction().getFunction();
2431 CallingConv::ID CallerCC = CallerF->getCallingConv();
2432 bool CCMatch = CallerCC == CalleeCC;
2434 if (GuaranteedTailCallOpt) {
2435 if (IsTailCallConvention(CalleeCC) && CCMatch)
2440 // Look for obvious safe cases to perform tail call optimization that do not
2441 // require ABI changes. This is what gcc calls sibcall.
2443 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2444 // emit a special epilogue.
2445 if (RegInfo->needsStackRealignment(MF))
2448 // Do not sibcall optimize vararg calls unless the call site is not passing
2450 if (isVarArg && !Outs.empty())
2453 // Also avoid sibcall optimization if either caller or callee uses struct
2454 // return semantics.
2455 if (isCalleeStructRet || isCallerStructRet)
2458 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2459 // Therefore if it's not used by the call it is not safe to optimize this into
2461 bool Unused = false;
2462 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2469 SmallVector<CCValAssign, 16> RVLocs;
2470 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2471 RVLocs, *DAG.getContext());
2472 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2473 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2474 CCValAssign &VA = RVLocs[i];
2475 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2480 // If the calling conventions do not match, then we'd better make sure the
2481 // results are returned in the same way as what the caller expects.
2483 SmallVector<CCValAssign, 16> RVLocs1;
2484 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2485 RVLocs1, *DAG.getContext());
2486 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2488 SmallVector<CCValAssign, 16> RVLocs2;
2489 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2490 RVLocs2, *DAG.getContext());
2491 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2493 if (RVLocs1.size() != RVLocs2.size())
2495 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2496 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2498 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2500 if (RVLocs1[i].isRegLoc()) {
2501 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2504 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2510 // If the callee takes no arguments then go on to check the results of the
2512 if (!Outs.empty()) {
2513 // Check if stack adjustment is needed. For now, do not do this if any
2514 // argument is passed on the stack.
2515 SmallVector<CCValAssign, 16> ArgLocs;
2516 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2517 ArgLocs, *DAG.getContext());
2518 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2519 if (CCInfo.getNextStackOffset()) {
2520 MachineFunction &MF = DAG.getMachineFunction();
2521 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2523 if (Subtarget->isTargetWin64())
2524 // Win64 ABI has additional complications.
2527 // Check if the arguments are already laid out in the right way as
2528 // the caller's fixed stack objects.
2529 MachineFrameInfo *MFI = MF.getFrameInfo();
2530 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2531 const X86InstrInfo *TII =
2532 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2533 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2534 CCValAssign &VA = ArgLocs[i];
2535 SDValue Arg = OutVals[i];
2536 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2537 if (VA.getLocInfo() == CCValAssign::Indirect)
2539 if (!VA.isRegLoc()) {
2540 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2547 // If the tailcall address may be in a register, then make sure it's
2548 // possible to register allocate for it. In 32-bit, the call address can
2549 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2550 // callee-saved registers are restored. These happen to be the same
2551 // registers used to pass 'inreg' arguments so watch out for those.
2552 if (!Subtarget->is64Bit() &&
2553 !isa<GlobalAddressSDNode>(Callee) &&
2554 !isa<ExternalSymbolSDNode>(Callee)) {
2555 unsigned NumInRegs = 0;
2556 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2557 CCValAssign &VA = ArgLocs[i];
2560 unsigned Reg = VA.getLocReg();
2563 case X86::EAX: case X86::EDX: case X86::ECX:
2564 if (++NumInRegs == 3)
2576 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2577 return X86::createFastISel(funcInfo);
2581 //===----------------------------------------------------------------------===//
2582 // Other Lowering Hooks
2583 //===----------------------------------------------------------------------===//
2585 static bool MayFoldLoad(SDValue Op) {
2586 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2589 static bool MayFoldIntoStore(SDValue Op) {
2590 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2593 static bool isTargetShuffle(unsigned Opcode) {
2595 default: return false;
2596 case X86ISD::PSHUFD:
2597 case X86ISD::PSHUFHW:
2598 case X86ISD::PSHUFLW:
2599 case X86ISD::SHUFPD:
2600 case X86ISD::PALIGN:
2601 case X86ISD::SHUFPS:
2602 case X86ISD::MOVLHPS:
2603 case X86ISD::MOVLHPD:
2604 case X86ISD::MOVHLPS:
2605 case X86ISD::MOVLPS:
2606 case X86ISD::MOVLPD:
2607 case X86ISD::MOVSHDUP:
2608 case X86ISD::MOVSLDUP:
2609 case X86ISD::MOVDDUP:
2612 case X86ISD::UNPCKLPS:
2613 case X86ISD::UNPCKLPD:
2614 case X86ISD::PUNPCKLWD:
2615 case X86ISD::PUNPCKLBW:
2616 case X86ISD::PUNPCKLDQ:
2617 case X86ISD::PUNPCKLQDQ:
2618 case X86ISD::UNPCKHPS:
2619 case X86ISD::UNPCKHPD:
2620 case X86ISD::PUNPCKHWD:
2621 case X86ISD::PUNPCKHBW:
2622 case X86ISD::PUNPCKHDQ:
2623 case X86ISD::PUNPCKHQDQ:
2629 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2630 SDValue V1, SelectionDAG &DAG) {
2632 default: llvm_unreachable("Unknown x86 shuffle node");
2633 case X86ISD::MOVSHDUP:
2634 case X86ISD::MOVSLDUP:
2635 case X86ISD::MOVDDUP:
2636 return DAG.getNode(Opc, dl, VT, V1);
2642 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2643 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2645 default: llvm_unreachable("Unknown x86 shuffle node");
2646 case X86ISD::PSHUFD:
2647 case X86ISD::PSHUFHW:
2648 case X86ISD::PSHUFLW:
2649 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2655 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2656 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2658 default: llvm_unreachable("Unknown x86 shuffle node");
2659 case X86ISD::PALIGN:
2660 case X86ISD::SHUFPD:
2661 case X86ISD::SHUFPS:
2662 return DAG.getNode(Opc, dl, VT, V1, V2,
2663 DAG.getConstant(TargetMask, MVT::i8));
2668 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2669 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2671 default: llvm_unreachable("Unknown x86 shuffle node");
2672 case X86ISD::MOVLHPS:
2673 case X86ISD::MOVLHPD:
2674 case X86ISD::MOVHLPS:
2675 case X86ISD::MOVLPS:
2676 case X86ISD::MOVLPD:
2679 case X86ISD::UNPCKLPS:
2680 case X86ISD::UNPCKLPD:
2681 case X86ISD::PUNPCKLWD:
2682 case X86ISD::PUNPCKLBW:
2683 case X86ISD::PUNPCKLDQ:
2684 case X86ISD::PUNPCKLQDQ:
2685 case X86ISD::UNPCKHPS:
2686 case X86ISD::UNPCKHPD:
2687 case X86ISD::PUNPCKHWD:
2688 case X86ISD::PUNPCKHBW:
2689 case X86ISD::PUNPCKHDQ:
2690 case X86ISD::PUNPCKHQDQ:
2691 return DAG.getNode(Opc, dl, VT, V1, V2);
2696 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2697 MachineFunction &MF = DAG.getMachineFunction();
2698 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2699 int ReturnAddrIndex = FuncInfo->getRAIndex();
2701 if (ReturnAddrIndex == 0) {
2702 // Set up a frame object for the return address.
2703 uint64_t SlotSize = TD->getPointerSize();
2704 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2706 FuncInfo->setRAIndex(ReturnAddrIndex);
2709 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2713 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2714 bool hasSymbolicDisplacement) {
2715 // Offset should fit into 32 bit immediate field.
2716 if (!isInt<32>(Offset))
2719 // If we don't have a symbolic displacement - we don't have any extra
2721 if (!hasSymbolicDisplacement)
2724 // FIXME: Some tweaks might be needed for medium code model.
2725 if (M != CodeModel::Small && M != CodeModel::Kernel)
2728 // For small code model we assume that latest object is 16MB before end of 31
2729 // bits boundary. We may also accept pretty large negative constants knowing
2730 // that all objects are in the positive half of address space.
2731 if (M == CodeModel::Small && Offset < 16*1024*1024)
2734 // For kernel code model we know that all object resist in the negative half
2735 // of 32bits address space. We may not accept negative offsets, since they may
2736 // be just off and we may accept pretty large positive ones.
2737 if (M == CodeModel::Kernel && Offset > 0)
2743 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2744 /// specific condition code, returning the condition code and the LHS/RHS of the
2745 /// comparison to make.
2746 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2747 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2749 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2750 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2751 // X > -1 -> X == 0, jump !sign.
2752 RHS = DAG.getConstant(0, RHS.getValueType());
2753 return X86::COND_NS;
2754 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2755 // X < 0 -> X == 0, jump on sign.
2757 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2759 RHS = DAG.getConstant(0, RHS.getValueType());
2760 return X86::COND_LE;
2764 switch (SetCCOpcode) {
2765 default: llvm_unreachable("Invalid integer condition!");
2766 case ISD::SETEQ: return X86::COND_E;
2767 case ISD::SETGT: return X86::COND_G;
2768 case ISD::SETGE: return X86::COND_GE;
2769 case ISD::SETLT: return X86::COND_L;
2770 case ISD::SETLE: return X86::COND_LE;
2771 case ISD::SETNE: return X86::COND_NE;
2772 case ISD::SETULT: return X86::COND_B;
2773 case ISD::SETUGT: return X86::COND_A;
2774 case ISD::SETULE: return X86::COND_BE;
2775 case ISD::SETUGE: return X86::COND_AE;
2779 // First determine if it is required or is profitable to flip the operands.
2781 // If LHS is a foldable load, but RHS is not, flip the condition.
2782 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2783 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2784 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2785 std::swap(LHS, RHS);
2788 switch (SetCCOpcode) {
2794 std::swap(LHS, RHS);
2798 // On a floating point condition, the flags are set as follows:
2800 // 0 | 0 | 0 | X > Y
2801 // 0 | 0 | 1 | X < Y
2802 // 1 | 0 | 0 | X == Y
2803 // 1 | 1 | 1 | unordered
2804 switch (SetCCOpcode) {
2805 default: llvm_unreachable("Condcode should be pre-legalized away");
2807 case ISD::SETEQ: return X86::COND_E;
2808 case ISD::SETOLT: // flipped
2810 case ISD::SETGT: return X86::COND_A;
2811 case ISD::SETOLE: // flipped
2813 case ISD::SETGE: return X86::COND_AE;
2814 case ISD::SETUGT: // flipped
2816 case ISD::SETLT: return X86::COND_B;
2817 case ISD::SETUGE: // flipped
2819 case ISD::SETLE: return X86::COND_BE;
2821 case ISD::SETNE: return X86::COND_NE;
2822 case ISD::SETUO: return X86::COND_P;
2823 case ISD::SETO: return X86::COND_NP;
2825 case ISD::SETUNE: return X86::COND_INVALID;
2829 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2830 /// code. Current x86 isa includes the following FP cmov instructions:
2831 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2832 static bool hasFPCMov(unsigned X86CC) {
2848 /// isFPImmLegal - Returns true if the target can instruction select the
2849 /// specified FP immediate natively. If false, the legalizer will
2850 /// materialize the FP immediate as a load from a constant pool.
2851 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2852 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2853 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2859 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2860 /// the specified range (L, H].
2861 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2862 return (Val < 0) || (Val >= Low && Val < Hi);
2865 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2866 /// specified value.
2867 static bool isUndefOrEqual(int Val, int CmpVal) {
2868 if (Val < 0 || Val == CmpVal)
2873 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2874 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2875 /// the second operand.
2876 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2877 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2878 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2879 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2880 return (Mask[0] < 2 && Mask[1] < 2);
2884 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2885 SmallVector<int, 8> M;
2887 return ::isPSHUFDMask(M, N->getValueType(0));
2890 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2891 /// is suitable for input to PSHUFHW.
2892 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2893 if (VT != MVT::v8i16)
2896 // Lower quadword copied in order or undef.
2897 for (int i = 0; i != 4; ++i)
2898 if (Mask[i] >= 0 && Mask[i] != i)
2901 // Upper quadword shuffled.
2902 for (int i = 4; i != 8; ++i)
2903 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2909 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2910 SmallVector<int, 8> M;
2912 return ::isPSHUFHWMask(M, N->getValueType(0));
2915 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2916 /// is suitable for input to PSHUFLW.
2917 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2918 if (VT != MVT::v8i16)
2921 // Upper quadword copied in order.
2922 for (int i = 4; i != 8; ++i)
2923 if (Mask[i] >= 0 && Mask[i] != i)
2926 // Lower quadword shuffled.
2927 for (int i = 0; i != 4; ++i)
2934 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2935 SmallVector<int, 8> M;
2937 return ::isPSHUFLWMask(M, N->getValueType(0));
2940 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2941 /// is suitable for input to PALIGNR.
2942 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2944 int i, e = VT.getVectorNumElements();
2946 // Do not handle v2i64 / v2f64 shuffles with palignr.
2947 if (e < 4 || !hasSSSE3)
2950 for (i = 0; i != e; ++i)
2954 // All undef, not a palignr.
2958 // Determine if it's ok to perform a palignr with only the LHS, since we
2959 // don't have access to the actual shuffle elements to see if RHS is undef.
2960 bool Unary = Mask[i] < (int)e;
2961 bool NeedsUnary = false;
2963 int s = Mask[i] - i;
2965 // Check the rest of the elements to see if they are consecutive.
2966 for (++i; i != e; ++i) {
2971 Unary = Unary && (m < (int)e);
2972 NeedsUnary = NeedsUnary || (m < s);
2974 if (NeedsUnary && !Unary)
2976 if (Unary && m != ((s+i) & (e-1)))
2978 if (!Unary && m != (s+i))
2984 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2985 SmallVector<int, 8> M;
2987 return ::isPALIGNRMask(M, N->getValueType(0), true);
2990 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2991 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2992 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2993 int NumElems = VT.getVectorNumElements();
2994 if (NumElems != 2 && NumElems != 4)
2997 int Half = NumElems / 2;
2998 for (int i = 0; i < Half; ++i)
2999 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3001 for (int i = Half; i < NumElems; ++i)
3002 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3008 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3009 SmallVector<int, 8> M;
3011 return ::isSHUFPMask(M, N->getValueType(0));
3014 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3015 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3016 /// half elements to come from vector 1 (which would equal the dest.) and
3017 /// the upper half to come from vector 2.
3018 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3019 int NumElems = VT.getVectorNumElements();
3021 if (NumElems != 2 && NumElems != 4)
3024 int Half = NumElems / 2;
3025 for (int i = 0; i < Half; ++i)
3026 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3028 for (int i = Half; i < NumElems; ++i)
3029 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3034 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3035 SmallVector<int, 8> M;
3037 return isCommutedSHUFPMask(M, N->getValueType(0));
3040 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3041 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3042 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3043 if (N->getValueType(0).getVectorNumElements() != 4)
3046 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3047 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3048 isUndefOrEqual(N->getMaskElt(1), 7) &&
3049 isUndefOrEqual(N->getMaskElt(2), 2) &&
3050 isUndefOrEqual(N->getMaskElt(3), 3);
3053 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3054 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3056 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3057 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3062 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3063 isUndefOrEqual(N->getMaskElt(1), 3) &&
3064 isUndefOrEqual(N->getMaskElt(2), 2) &&
3065 isUndefOrEqual(N->getMaskElt(3), 3);
3068 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3069 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3070 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3071 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3073 if (NumElems != 2 && NumElems != 4)
3076 for (unsigned i = 0; i < NumElems/2; ++i)
3077 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3080 for (unsigned i = NumElems/2; i < NumElems; ++i)
3081 if (!isUndefOrEqual(N->getMaskElt(i), i))
3087 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3088 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3089 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3090 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3092 if (NumElems != 2 && NumElems != 4)
3095 for (unsigned i = 0; i < NumElems/2; ++i)
3096 if (!isUndefOrEqual(N->getMaskElt(i), i))
3099 for (unsigned i = 0; i < NumElems/2; ++i)
3100 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3106 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3107 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3108 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3109 bool V2IsSplat = false) {
3110 int NumElts = VT.getVectorNumElements();
3111 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3114 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3116 int BitI1 = Mask[i+1];
3117 if (!isUndefOrEqual(BitI, j))
3120 if (!isUndefOrEqual(BitI1, NumElts))
3123 if (!isUndefOrEqual(BitI1, j + NumElts))
3130 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3131 SmallVector<int, 8> M;
3133 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3136 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3137 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3138 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3139 bool V2IsSplat = false) {
3140 int NumElts = VT.getVectorNumElements();
3141 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3144 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3146 int BitI1 = Mask[i+1];
3147 if (!isUndefOrEqual(BitI, j + NumElts/2))
3150 if (isUndefOrEqual(BitI1, NumElts))
3153 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3160 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3161 SmallVector<int, 8> M;
3163 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3166 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3167 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3169 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3170 int NumElems = VT.getVectorNumElements();
3171 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3174 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3176 int BitI1 = Mask[i+1];
3177 if (!isUndefOrEqual(BitI, j))
3179 if (!isUndefOrEqual(BitI1, j))
3185 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3186 SmallVector<int, 8> M;
3188 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3191 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3192 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3194 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3195 int NumElems = VT.getVectorNumElements();
3196 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3199 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3201 int BitI1 = Mask[i+1];
3202 if (!isUndefOrEqual(BitI, j))
3204 if (!isUndefOrEqual(BitI1, j))
3210 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3211 SmallVector<int, 8> M;
3213 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3216 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3217 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3218 /// MOVSD, and MOVD, i.e. setting the lowest element.
3219 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3220 if (VT.getVectorElementType().getSizeInBits() < 32)
3223 int NumElts = VT.getVectorNumElements();
3225 if (!isUndefOrEqual(Mask[0], NumElts))
3228 for (int i = 1; i < NumElts; ++i)
3229 if (!isUndefOrEqual(Mask[i], i))
3235 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3236 SmallVector<int, 8> M;
3238 return ::isMOVLMask(M, N->getValueType(0));
3241 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3242 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3243 /// element of vector 2 and the other elements to come from vector 1 in order.
3244 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3245 bool V2IsSplat = false, bool V2IsUndef = false) {
3246 int NumOps = VT.getVectorNumElements();
3247 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3250 if (!isUndefOrEqual(Mask[0], 0))
3253 for (int i = 1; i < NumOps; ++i)
3254 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3255 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3256 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3262 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3263 bool V2IsUndef = false) {
3264 SmallVector<int, 8> M;
3266 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3269 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3270 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3271 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3272 if (N->getValueType(0).getVectorNumElements() != 4)
3275 // Expect 1, 1, 3, 3
3276 for (unsigned i = 0; i < 2; ++i) {
3277 int Elt = N->getMaskElt(i);
3278 if (Elt >= 0 && Elt != 1)
3283 for (unsigned i = 2; i < 4; ++i) {
3284 int Elt = N->getMaskElt(i);
3285 if (Elt >= 0 && Elt != 3)
3290 // Don't use movshdup if it can be done with a shufps.
3291 // FIXME: verify that matching u, u, 3, 3 is what we want.
3295 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3296 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3297 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3298 if (N->getValueType(0).getVectorNumElements() != 4)
3301 // Expect 0, 0, 2, 2
3302 for (unsigned i = 0; i < 2; ++i)
3303 if (N->getMaskElt(i) > 0)
3307 for (unsigned i = 2; i < 4; ++i) {
3308 int Elt = N->getMaskElt(i);
3309 if (Elt >= 0 && Elt != 2)
3314 // Don't use movsldup if it can be done with a shufps.
3318 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3319 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3320 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3321 int e = N->getValueType(0).getVectorNumElements() / 2;
3323 for (int i = 0; i < e; ++i)
3324 if (!isUndefOrEqual(N->getMaskElt(i), i))
3326 for (int i = 0; i < e; ++i)
3327 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3332 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3333 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3334 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3335 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3336 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3338 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3340 for (int i = 0; i < NumOperands; ++i) {
3341 int Val = SVOp->getMaskElt(NumOperands-i-1);
3342 if (Val < 0) Val = 0;
3343 if (Val >= NumOperands) Val -= NumOperands;
3345 if (i != NumOperands - 1)
3351 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3352 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3353 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3354 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3356 // 8 nodes, but we only care about the last 4.
3357 for (unsigned i = 7; i >= 4; --i) {
3358 int Val = SVOp->getMaskElt(i);
3367 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3368 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3369 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3370 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3372 // 8 nodes, but we only care about the first 4.
3373 for (int i = 3; i >= 0; --i) {
3374 int Val = SVOp->getMaskElt(i);
3383 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3384 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3385 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3386 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3387 EVT VVT = N->getValueType(0);
3388 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3392 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3393 Val = SVOp->getMaskElt(i);
3397 return (Val - i) * EltSize;
3400 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3402 bool X86::isZeroNode(SDValue Elt) {
3403 return ((isa<ConstantSDNode>(Elt) &&
3404 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3405 (isa<ConstantFPSDNode>(Elt) &&
3406 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3409 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3410 /// their permute mask.
3411 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3412 SelectionDAG &DAG) {
3413 EVT VT = SVOp->getValueType(0);
3414 unsigned NumElems = VT.getVectorNumElements();
3415 SmallVector<int, 8> MaskVec;
3417 for (unsigned i = 0; i != NumElems; ++i) {
3418 int idx = SVOp->getMaskElt(i);
3420 MaskVec.push_back(idx);
3421 else if (idx < (int)NumElems)
3422 MaskVec.push_back(idx + NumElems);
3424 MaskVec.push_back(idx - NumElems);
3426 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3427 SVOp->getOperand(0), &MaskVec[0]);
3430 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3431 /// the two vector operands have swapped position.
3432 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3433 unsigned NumElems = VT.getVectorNumElements();
3434 for (unsigned i = 0; i != NumElems; ++i) {
3438 else if (idx < (int)NumElems)
3439 Mask[i] = idx + NumElems;
3441 Mask[i] = idx - NumElems;
3445 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3446 /// match movhlps. The lower half elements should come from upper half of
3447 /// V1 (and in order), and the upper half elements should come from the upper
3448 /// half of V2 (and in order).
3449 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3450 if (Op->getValueType(0).getVectorNumElements() != 4)
3452 for (unsigned i = 0, e = 2; i != e; ++i)
3453 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3455 for (unsigned i = 2; i != 4; ++i)
3456 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3461 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3462 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3464 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3465 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3467 N = N->getOperand(0).getNode();
3468 if (!ISD::isNON_EXTLoad(N))
3471 *LD = cast<LoadSDNode>(N);
3475 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3476 /// match movlp{s|d}. The lower half elements should come from lower half of
3477 /// V1 (and in order), and the upper half elements should come from the upper
3478 /// half of V2 (and in order). And since V1 will become the source of the
3479 /// MOVLP, it must be either a vector load or a scalar load to vector.
3480 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3481 ShuffleVectorSDNode *Op) {
3482 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3484 // Is V2 is a vector load, don't do this transformation. We will try to use
3485 // load folding shufps op.
3486 if (ISD::isNON_EXTLoad(V2))
3489 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3491 if (NumElems != 2 && NumElems != 4)
3493 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3494 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3496 for (unsigned i = NumElems/2; i != NumElems; ++i)
3497 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3502 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3504 static bool isSplatVector(SDNode *N) {
3505 if (N->getOpcode() != ISD::BUILD_VECTOR)
3508 SDValue SplatValue = N->getOperand(0);
3509 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3510 if (N->getOperand(i) != SplatValue)
3515 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3516 /// to an zero vector.
3517 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3518 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3519 SDValue V1 = N->getOperand(0);
3520 SDValue V2 = N->getOperand(1);
3521 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3522 for (unsigned i = 0; i != NumElems; ++i) {
3523 int Idx = N->getMaskElt(i);
3524 if (Idx >= (int)NumElems) {
3525 unsigned Opc = V2.getOpcode();
3526 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3528 if (Opc != ISD::BUILD_VECTOR ||
3529 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3531 } else if (Idx >= 0) {
3532 unsigned Opc = V1.getOpcode();
3533 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3535 if (Opc != ISD::BUILD_VECTOR ||
3536 !X86::isZeroNode(V1.getOperand(Idx)))
3543 /// getZeroVector - Returns a vector of specified type with all zero elements.
3545 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3547 assert(VT.isVector() && "Expected a vector type");
3549 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted
3550 // to their dest type. This ensures they get CSE'd.
3552 if (VT.getSizeInBits() == 64) { // MMX
3553 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3554 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3555 } else if (VT.getSizeInBits() == 128) {
3556 if (HasSSE2) { // SSE2
3557 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3558 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3560 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3561 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3563 } else if (VT.getSizeInBits() == 256) { // AVX
3564 // 256-bit logic and arithmetic instructions in AVX are
3565 // all floating-point, no support for integer ops. Default
3566 // to emitting fp zeroed vectors then.
3567 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3568 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3569 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3571 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3574 /// getOnesVector - Returns a vector of specified type with all bits set.
3576 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3577 assert(VT.isVector() && "Expected a vector type");
3579 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3580 // type. This ensures they get CSE'd.
3581 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3583 if (VT.getSizeInBits() == 64) // MMX
3584 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3586 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3587 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3591 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3592 /// that point to V2 points to its first element.
3593 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3594 EVT VT = SVOp->getValueType(0);
3595 unsigned NumElems = VT.getVectorNumElements();
3597 bool Changed = false;
3598 SmallVector<int, 8> MaskVec;
3599 SVOp->getMask(MaskVec);
3601 for (unsigned i = 0; i != NumElems; ++i) {
3602 if (MaskVec[i] > (int)NumElems) {
3603 MaskVec[i] = NumElems;
3608 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3609 SVOp->getOperand(1), &MaskVec[0]);
3610 return SDValue(SVOp, 0);
3613 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3614 /// operation of specified width.
3615 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3617 unsigned NumElems = VT.getVectorNumElements();
3618 SmallVector<int, 8> Mask;
3619 Mask.push_back(NumElems);
3620 for (unsigned i = 1; i != NumElems; ++i)
3622 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3625 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3626 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3628 unsigned NumElems = VT.getVectorNumElements();
3629 SmallVector<int, 8> Mask;
3630 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3632 Mask.push_back(i + NumElems);
3634 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3637 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3638 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3640 unsigned NumElems = VT.getVectorNumElements();
3641 unsigned Half = NumElems/2;
3642 SmallVector<int, 8> Mask;
3643 for (unsigned i = 0; i != Half; ++i) {
3644 Mask.push_back(i + Half);
3645 Mask.push_back(i + NumElems + Half);
3647 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3650 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3651 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3652 EVT PVT = MVT::v4f32;
3653 EVT VT = SV->getValueType(0);
3654 DebugLoc dl = SV->getDebugLoc();
3655 SDValue V1 = SV->getOperand(0);
3656 int NumElems = VT.getVectorNumElements();
3657 int EltNo = SV->getSplatIndex();
3659 // unpack elements to the correct location
3660 while (NumElems > 4) {
3661 if (EltNo < NumElems/2) {
3662 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3664 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3665 EltNo -= NumElems/2;
3670 // Perform the splat.
3671 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3672 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3673 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3674 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3677 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3678 /// vector of zero or undef vector. This produces a shuffle where the low
3679 /// element of V2 is swizzled into the zero/undef vector, landing at element
3680 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3681 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3682 bool isZero, bool HasSSE2,
3683 SelectionDAG &DAG) {
3684 EVT VT = V2.getValueType();
3686 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3687 unsigned NumElems = VT.getVectorNumElements();
3688 SmallVector<int, 16> MaskVec;
3689 for (unsigned i = 0; i != NumElems; ++i)
3690 // If this is the insertion idx, put the low elt of V2 here.
3691 MaskVec.push_back(i == Idx ? NumElems : i);
3692 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3695 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
3696 /// element of the result of the vector shuffle.
3697 SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
3700 return SDValue(); // Limit search depth.
3702 SDValue V = SDValue(N, 0);
3703 EVT VT = V.getValueType();
3704 unsigned Opcode = V.getOpcode();
3706 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
3707 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
3708 Index = SV->getMaskElt(Index);
3711 return DAG.getUNDEF(VT.getVectorElementType());
3713 int NumElems = VT.getVectorNumElements();
3714 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
3715 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
3718 // Recurse into target specific vector shuffles to find scalars.
3719 if (isTargetShuffle(Opcode)) {
3720 int NumElems = VT.getVectorNumElements();
3721 SmallVector<unsigned, 16> ShuffleMask;
3725 case X86ISD::SHUFPS:
3726 case X86ISD::SHUFPD:
3727 ImmN = N->getOperand(N->getNumOperands()-1);
3728 DecodeSHUFPSMask(NumElems,
3729 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3732 case X86ISD::PUNPCKHBW:
3733 case X86ISD::PUNPCKHWD:
3734 case X86ISD::PUNPCKHDQ:
3735 case X86ISD::PUNPCKHQDQ:
3736 DecodePUNPCKHMask(NumElems, ShuffleMask);
3738 case X86ISD::UNPCKHPS:
3739 case X86ISD::UNPCKHPD:
3740 DecodeUNPCKHPMask(NumElems, ShuffleMask);
3742 case X86ISD::PUNPCKLBW:
3743 case X86ISD::PUNPCKLWD:
3744 case X86ISD::PUNPCKLDQ:
3745 case X86ISD::PUNPCKLQDQ:
3746 DecodePUNPCKLMask(NumElems, ShuffleMask);
3748 case X86ISD::UNPCKLPS:
3749 case X86ISD::UNPCKLPD:
3750 DecodeUNPCKLPMask(NumElems, ShuffleMask);
3752 case X86ISD::MOVHLPS:
3753 DecodeMOVHLPSMask(NumElems, ShuffleMask);
3755 case X86ISD::MOVLHPS:
3756 DecodeMOVLHPSMask(NumElems, ShuffleMask);
3758 case X86ISD::PSHUFD:
3759 ImmN = N->getOperand(N->getNumOperands()-1);
3760 DecodePSHUFMask(NumElems,
3761 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3764 case X86ISD::PSHUFHW:
3765 ImmN = N->getOperand(N->getNumOperands()-1);
3766 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3769 case X86ISD::PSHUFLW:
3770 ImmN = N->getOperand(N->getNumOperands()-1);
3771 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3775 case X86ISD::MOVSD: {
3776 // The index 0 always comes from the first element of the second source,
3777 // this is why MOVSS and MOVSD are used in the first place. The other
3778 // elements come from the other positions of the first source vector.
3779 unsigned OpNum = (Index == 0) ? 1 : 0;
3780 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
3784 assert("not implemented for target shuffle node");
3788 Index = ShuffleMask[Index];
3790 return DAG.getUNDEF(VT.getVectorElementType());
3792 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
3793 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
3797 // Actual nodes that may contain scalar elements
3798 if (Opcode == ISD::BIT_CONVERT) {
3799 V = V.getOperand(0);
3800 EVT SrcVT = V.getValueType();
3801 unsigned NumElems = VT.getVectorNumElements();
3803 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
3807 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
3808 return (Index == 0) ? V.getOperand(0)
3809 : DAG.getUNDEF(VT.getVectorElementType());
3811 if (V.getOpcode() == ISD::BUILD_VECTOR)
3812 return V.getOperand(Index);
3817 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
3818 /// shuffle operation which come from a consecutively from a zero. The
3819 /// search can start in two diferent directions, from left or right.
3821 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
3822 bool ZerosFromLeft, SelectionDAG &DAG) {
3825 while (i < NumElems) {
3826 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
3827 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
3828 if (!(Elt.getNode() &&
3829 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
3837 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
3838 /// MaskE correspond consecutively to elements from one of the vector operands,
3839 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
3841 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
3842 int OpIdx, int NumElems, unsigned &OpNum) {
3843 bool SeenV1 = false;
3844 bool SeenV2 = false;
3846 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
3847 int Idx = SVOp->getMaskElt(i);
3848 // Ignore undef indicies
3857 // Only accept consecutive elements from the same vector
3858 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
3862 OpNum = SeenV1 ? 0 : 1;
3866 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
3867 /// logical left shift of a vector.
3868 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3869 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3870 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3871 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3872 false /* check zeros from right */, DAG);
3878 // Considering the elements in the mask that are not consecutive zeros,
3879 // check if they consecutively come from only one of the source vectors.
3881 // V1 = {X, A, B, C} 0
3883 // vector_shuffle V1, V2 <1, 2, 3, X>
3885 if (!isShuffleMaskConsecutive(SVOp,
3886 0, // Mask Start Index
3887 NumElems-NumZeros-1, // Mask End Index
3888 NumZeros, // Where to start looking in the src vector
3889 NumElems, // Number of elements in vector
3890 OpSrc)) // Which source operand ?
3895 ShVal = SVOp->getOperand(OpSrc);
3899 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
3900 /// logical left shift of a vector.
3901 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3902 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3903 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3904 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3905 true /* check zeros from left */, DAG);
3911 // Considering the elements in the mask that are not consecutive zeros,
3912 // check if they consecutively come from only one of the source vectors.
3914 // 0 { A, B, X, X } = V2
3916 // vector_shuffle V1, V2 <X, X, 4, 5>
3918 if (!isShuffleMaskConsecutive(SVOp,
3919 NumZeros, // Mask Start Index
3920 NumElems-1, // Mask End Index
3921 0, // Where to start looking in the src vector
3922 NumElems, // Number of elements in vector
3923 OpSrc)) // Which source operand ?
3928 ShVal = SVOp->getOperand(OpSrc);
3932 /// isVectorShift - Returns true if the shuffle can be implemented as a
3933 /// logical left or right shift of a vector.
3934 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3935 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3936 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
3937 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
3943 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3945 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3946 unsigned NumNonZero, unsigned NumZero,
3948 const TargetLowering &TLI) {
3952 DebugLoc dl = Op.getDebugLoc();
3955 for (unsigned i = 0; i < 16; ++i) {
3956 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3957 if (ThisIsNonZero && First) {
3959 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3961 V = DAG.getUNDEF(MVT::v8i16);
3966 SDValue ThisElt(0, 0), LastElt(0, 0);
3967 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3968 if (LastIsNonZero) {
3969 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3970 MVT::i16, Op.getOperand(i-1));
3972 if (ThisIsNonZero) {
3973 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3974 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3975 ThisElt, DAG.getConstant(8, MVT::i8));
3977 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3981 if (ThisElt.getNode())
3982 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3983 DAG.getIntPtrConstant(i/2));
3987 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3990 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3992 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3993 unsigned NumNonZero, unsigned NumZero,
3995 const TargetLowering &TLI) {
3999 DebugLoc dl = Op.getDebugLoc();
4002 for (unsigned i = 0; i < 8; ++i) {
4003 bool isNonZero = (NonZeros & (1 << i)) != 0;
4007 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4009 V = DAG.getUNDEF(MVT::v8i16);
4012 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4013 MVT::v8i16, V, Op.getOperand(i),
4014 DAG.getIntPtrConstant(i));
4021 /// getVShift - Return a vector logical shift node.
4023 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4024 unsigned NumBits, SelectionDAG &DAG,
4025 const TargetLowering &TLI, DebugLoc dl) {
4026 bool isMMX = VT.getSizeInBits() == 64;
4027 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
4028 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4029 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
4030 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4031 DAG.getNode(Opc, dl, ShVT, SrcOp,
4032 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
4036 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4037 SelectionDAG &DAG) const {
4039 // Check if the scalar load can be widened into a vector load. And if
4040 // the address is "base + cst" see if the cst can be "absorbed" into
4041 // the shuffle mask.
4042 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4043 SDValue Ptr = LD->getBasePtr();
4044 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4046 EVT PVT = LD->getValueType(0);
4047 if (PVT != MVT::i32 && PVT != MVT::f32)
4052 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4053 FI = FINode->getIndex();
4055 } else if (Ptr.getOpcode() == ISD::ADD &&
4056 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
4057 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4058 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4059 Offset = Ptr.getConstantOperandVal(1);
4060 Ptr = Ptr.getOperand(0);
4065 SDValue Chain = LD->getChain();
4066 // Make sure the stack object alignment is at least 16.
4067 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4068 if (DAG.InferPtrAlignment(Ptr) < 16) {
4069 if (MFI->isFixedObjectIndex(FI)) {
4070 // Can't change the alignment. FIXME: It's possible to compute
4071 // the exact stack offset and reference FI + adjust offset instead.
4072 // If someone *really* cares about this. That's the way to implement it.
4075 MFI->setObjectAlignment(FI, 16);
4079 // (Offset % 16) must be multiple of 4. Then address is then
4080 // Ptr + (Offset & ~15).
4083 if ((Offset % 16) & 3)
4085 int64_t StartOffset = Offset & ~15;
4087 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4088 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4090 int EltNo = (Offset - StartOffset) >> 2;
4091 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4092 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4093 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,
4094 LD->getPointerInfo().getWithOffset(StartOffset),
4096 // Canonicalize it to a v4i32 shuffle.
4097 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
4098 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4099 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4100 DAG.getUNDEF(MVT::v4i32),&Mask[0]));
4106 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4107 /// vector of type 'VT', see if the elements can be replaced by a single large
4108 /// load which has the same value as a build_vector whose operands are 'elts'.
4110 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4112 /// FIXME: we'd also like to handle the case where the last elements are zero
4113 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4114 /// There's even a handy isZeroNode for that purpose.
4115 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4116 DebugLoc &dl, SelectionDAG &DAG) {
4117 EVT EltVT = VT.getVectorElementType();
4118 unsigned NumElems = Elts.size();
4120 LoadSDNode *LDBase = NULL;
4121 unsigned LastLoadedElt = -1U;
4123 // For each element in the initializer, see if we've found a load or an undef.
4124 // If we don't find an initial load element, or later load elements are
4125 // non-consecutive, bail out.
4126 for (unsigned i = 0; i < NumElems; ++i) {
4127 SDValue Elt = Elts[i];
4129 if (!Elt.getNode() ||
4130 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4133 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4135 LDBase = cast<LoadSDNode>(Elt.getNode());
4139 if (Elt.getOpcode() == ISD::UNDEF)
4142 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4143 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4148 // If we have found an entire vector of loads and undefs, then return a large
4149 // load of the entire vector width starting at the base pointer. If we found
4150 // consecutive loads for the low half, generate a vzext_load node.
4151 if (LastLoadedElt == NumElems - 1) {
4152 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4153 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
4154 LDBase->getPointerInfo(),
4155 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4156 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
4157 LDBase->getPointerInfo(),
4158 LDBase->isVolatile(), LDBase->isNonTemporal(),
4159 LDBase->getAlignment());
4160 } else if (NumElems == 4 && LastLoadedElt == 1) {
4161 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4162 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4163 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
4164 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
4170 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4171 DebugLoc dl = Op.getDebugLoc();
4172 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
4173 // All one's are handled with pcmpeqd. In AVX, zero's are handled with
4174 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
4175 // is present, so AllOnes is ignored.
4176 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4177 (Op.getValueType().getSizeInBits() != 256 &&
4178 ISD::isBuildVectorAllOnes(Op.getNode()))) {
4179 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
4180 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4181 // eliminated on x86-32 hosts.
4182 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
4185 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4186 return getOnesVector(Op.getValueType(), DAG, dl);
4187 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4190 EVT VT = Op.getValueType();
4191 EVT ExtVT = VT.getVectorElementType();
4192 unsigned EVTBits = ExtVT.getSizeInBits();
4194 unsigned NumElems = Op.getNumOperands();
4195 unsigned NumZero = 0;
4196 unsigned NumNonZero = 0;
4197 unsigned NonZeros = 0;
4198 bool IsAllConstants = true;
4199 SmallSet<SDValue, 8> Values;
4200 for (unsigned i = 0; i < NumElems; ++i) {
4201 SDValue Elt = Op.getOperand(i);
4202 if (Elt.getOpcode() == ISD::UNDEF)
4205 if (Elt.getOpcode() != ISD::Constant &&
4206 Elt.getOpcode() != ISD::ConstantFP)
4207 IsAllConstants = false;
4208 if (X86::isZeroNode(Elt))
4211 NonZeros |= (1 << i);
4216 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4217 if (NumNonZero == 0)
4218 return DAG.getUNDEF(VT);
4220 // Special case for single non-zero, non-undef, element.
4221 if (NumNonZero == 1) {
4222 unsigned Idx = CountTrailingZeros_32(NonZeros);
4223 SDValue Item = Op.getOperand(Idx);
4225 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4226 // the value are obviously zero, truncate the value to i32 and do the
4227 // insertion that way. Only do this if the value is non-constant or if the
4228 // value is a constant being inserted into element 0. It is cheaper to do
4229 // a constant pool load than it is to do a movd + shuffle.
4230 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4231 (!IsAllConstants || Idx == 0)) {
4232 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4233 // Handle MMX and SSE both.
4234 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
4235 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
4237 // Truncate the value (which may itself be a constant) to i32, and
4238 // convert it to a vector with movd (S2V+shuffle to zero extend).
4239 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4240 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4241 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4242 Subtarget->hasSSE2(), DAG);
4244 // Now we have our 32-bit value zero extended in the low element of
4245 // a vector. If Idx != 0, swizzle it into place.
4247 SmallVector<int, 4> Mask;
4248 Mask.push_back(Idx);
4249 for (unsigned i = 1; i != VecElts; ++i)
4251 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4252 DAG.getUNDEF(Item.getValueType()),
4255 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
4259 // If we have a constant or non-constant insertion into the low element of
4260 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4261 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4262 // depending on what the source datatype is.
4265 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4266 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4267 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4268 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4269 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4270 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4272 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4273 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4274 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
4275 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4276 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4277 Subtarget->hasSSE2(), DAG);
4278 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
4282 // Is it a vector logical left shift?
4283 if (NumElems == 2 && Idx == 1 &&
4284 X86::isZeroNode(Op.getOperand(0)) &&
4285 !X86::isZeroNode(Op.getOperand(1))) {
4286 unsigned NumBits = VT.getSizeInBits();
4287 return getVShift(true, VT,
4288 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4289 VT, Op.getOperand(1)),
4290 NumBits/2, DAG, *this, dl);
4293 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4296 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4297 // is a non-constant being inserted into an element other than the low one,
4298 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4299 // movd/movss) to move this into the low element, then shuffle it into
4301 if (EVTBits == 32) {
4302 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4304 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4305 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4306 Subtarget->hasSSE2(), DAG);
4307 SmallVector<int, 8> MaskVec;
4308 for (unsigned i = 0; i < NumElems; i++)
4309 MaskVec.push_back(i == Idx ? 0 : 1);
4310 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4314 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4315 if (Values.size() == 1) {
4316 if (EVTBits == 32) {
4317 // Instead of a shuffle like this:
4318 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4319 // Check if it's possible to issue this instead.
4320 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4321 unsigned Idx = CountTrailingZeros_32(NonZeros);
4322 SDValue Item = Op.getOperand(Idx);
4323 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4324 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4329 // A vector full of immediates; various special cases are already
4330 // handled, so this is best done with a single constant-pool load.
4334 // Let legalizer expand 2-wide build_vectors.
4335 if (EVTBits == 64) {
4336 if (NumNonZero == 1) {
4337 // One half is zero or undef.
4338 unsigned Idx = CountTrailingZeros_32(NonZeros);
4339 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4340 Op.getOperand(Idx));
4341 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4342 Subtarget->hasSSE2(), DAG);
4347 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4348 if (EVTBits == 8 && NumElems == 16) {
4349 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4351 if (V.getNode()) return V;
4354 if (EVTBits == 16 && NumElems == 8) {
4355 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4357 if (V.getNode()) return V;
4360 // If element VT is == 32 bits, turn it into a number of shuffles.
4361 SmallVector<SDValue, 8> V;
4363 if (NumElems == 4 && NumZero > 0) {
4364 for (unsigned i = 0; i < 4; ++i) {
4365 bool isZero = !(NonZeros & (1 << i));
4367 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4369 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4372 for (unsigned i = 0; i < 2; ++i) {
4373 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4376 V[i] = V[i*2]; // Must be a zero vector.
4379 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4382 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4385 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4390 SmallVector<int, 8> MaskVec;
4391 bool Reverse = (NonZeros & 0x3) == 2;
4392 for (unsigned i = 0; i < 2; ++i)
4393 MaskVec.push_back(Reverse ? 1-i : i);
4394 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4395 for (unsigned i = 0; i < 2; ++i)
4396 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4397 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4400 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4401 // Check for a build vector of consecutive loads.
4402 for (unsigned i = 0; i < NumElems; ++i)
4403 V[i] = Op.getOperand(i);
4405 // Check for elements which are consecutive loads.
4406 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4410 // For SSE 4.1, use insertps to put the high elements into the low element.
4411 if (getSubtarget()->hasSSE41()) {
4413 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4414 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4416 Result = DAG.getUNDEF(VT);
4418 for (unsigned i = 1; i < NumElems; ++i) {
4419 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4420 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4421 Op.getOperand(i), DAG.getIntPtrConstant(i));
4426 // Otherwise, expand into a number of unpckl*, start by extending each of
4427 // our (non-undef) elements to the full vector width with the element in the
4428 // bottom slot of the vector (which generates no code for SSE).
4429 for (unsigned i = 0; i < NumElems; ++i) {
4430 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4431 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4433 V[i] = DAG.getUNDEF(VT);
4436 // Next, we iteratively mix elements, e.g. for v4f32:
4437 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4438 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4439 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4440 unsigned EltStride = NumElems >> 1;
4441 while (EltStride != 0) {
4442 for (unsigned i = 0; i < EltStride; ++i) {
4443 // If V[i+EltStride] is undef and this is the first round of mixing,
4444 // then it is safe to just drop this shuffle: V[i] is already in the
4445 // right place, the one element (since it's the first round) being
4446 // inserted as undef can be dropped. This isn't safe for successive
4447 // rounds because they will permute elements within both vectors.
4448 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4449 EltStride == NumElems/2)
4452 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4462 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4463 // We support concatenate two MMX registers and place them in a MMX
4464 // register. This is better than doing a stack convert.
4465 DebugLoc dl = Op.getDebugLoc();
4466 EVT ResVT = Op.getValueType();
4467 assert(Op.getNumOperands() == 2);
4468 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4469 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4471 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4472 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4473 InVec = Op.getOperand(1);
4474 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4475 unsigned NumElts = ResVT.getVectorNumElements();
4476 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4477 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4478 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4480 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4481 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4482 Mask[0] = 0; Mask[1] = 2;
4483 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4485 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4488 // v8i16 shuffles - Prefer shuffles in the following order:
4489 // 1. [all] pshuflw, pshufhw, optional move
4490 // 2. [ssse3] 1 x pshufb
4491 // 3. [ssse3] 2 x pshufb + 1 x por
4492 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4494 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4495 SelectionDAG &DAG) const {
4496 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4497 SDValue V1 = SVOp->getOperand(0);
4498 SDValue V2 = SVOp->getOperand(1);
4499 DebugLoc dl = SVOp->getDebugLoc();
4500 SmallVector<int, 8> MaskVals;
4502 // Determine if more than 1 of the words in each of the low and high quadwords
4503 // of the result come from the same quadword of one of the two inputs. Undef
4504 // mask values count as coming from any quadword, for better codegen.
4505 SmallVector<unsigned, 4> LoQuad(4);
4506 SmallVector<unsigned, 4> HiQuad(4);
4507 BitVector InputQuads(4);
4508 for (unsigned i = 0; i < 8; ++i) {
4509 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4510 int EltIdx = SVOp->getMaskElt(i);
4511 MaskVals.push_back(EltIdx);
4520 InputQuads.set(EltIdx / 4);
4523 int BestLoQuad = -1;
4524 unsigned MaxQuad = 1;
4525 for (unsigned i = 0; i < 4; ++i) {
4526 if (LoQuad[i] > MaxQuad) {
4528 MaxQuad = LoQuad[i];
4532 int BestHiQuad = -1;
4534 for (unsigned i = 0; i < 4; ++i) {
4535 if (HiQuad[i] > MaxQuad) {
4537 MaxQuad = HiQuad[i];
4541 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4542 // of the two input vectors, shuffle them into one input vector so only a
4543 // single pshufb instruction is necessary. If There are more than 2 input
4544 // quads, disable the next transformation since it does not help SSSE3.
4545 bool V1Used = InputQuads[0] || InputQuads[1];
4546 bool V2Used = InputQuads[2] || InputQuads[3];
4547 if (Subtarget->hasSSSE3()) {
4548 if (InputQuads.count() == 2 && V1Used && V2Used) {
4549 BestLoQuad = InputQuads.find_first();
4550 BestHiQuad = InputQuads.find_next(BestLoQuad);
4552 if (InputQuads.count() > 2) {
4558 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4559 // the shuffle mask. If a quad is scored as -1, that means that it contains
4560 // words from all 4 input quadwords.
4562 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4563 SmallVector<int, 8> MaskV;
4564 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4565 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4566 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4567 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4568 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4569 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4571 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4572 // source words for the shuffle, to aid later transformations.
4573 bool AllWordsInNewV = true;
4574 bool InOrder[2] = { true, true };
4575 for (unsigned i = 0; i != 8; ++i) {
4576 int idx = MaskVals[i];
4578 InOrder[i/4] = false;
4579 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4581 AllWordsInNewV = false;
4585 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4586 if (AllWordsInNewV) {
4587 for (int i = 0; i != 8; ++i) {
4588 int idx = MaskVals[i];
4591 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4592 if ((idx != i) && idx < 4)
4594 if ((idx != i) && idx > 3)
4603 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4604 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4605 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4606 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4607 unsigned TargetMask = 0;
4608 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4609 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4610 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4611 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4612 V1 = NewV.getOperand(0);
4613 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4617 // If we have SSSE3, and all words of the result are from 1 input vector,
4618 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4619 // is present, fall back to case 4.
4620 if (Subtarget->hasSSSE3()) {
4621 SmallVector<SDValue,16> pshufbMask;
4623 // If we have elements from both input vectors, set the high bit of the
4624 // shuffle mask element to zero out elements that come from V2 in the V1
4625 // mask, and elements that come from V1 in the V2 mask, so that the two
4626 // results can be OR'd together.
4627 bool TwoInputs = V1Used && V2Used;
4628 for (unsigned i = 0; i != 8; ++i) {
4629 int EltIdx = MaskVals[i] * 2;
4630 if (TwoInputs && (EltIdx >= 16)) {
4631 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4632 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4635 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4636 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4638 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4639 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4640 DAG.getNode(ISD::BUILD_VECTOR, dl,
4641 MVT::v16i8, &pshufbMask[0], 16));
4643 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4645 // Calculate the shuffle mask for the second input, shuffle it, and
4646 // OR it with the first shuffled input.
4648 for (unsigned i = 0; i != 8; ++i) {
4649 int EltIdx = MaskVals[i] * 2;
4651 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4652 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4655 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4656 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4658 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4659 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4660 DAG.getNode(ISD::BUILD_VECTOR, dl,
4661 MVT::v16i8, &pshufbMask[0], 16));
4662 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4663 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4666 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4667 // and update MaskVals with new element order.
4668 BitVector InOrder(8);
4669 if (BestLoQuad >= 0) {
4670 SmallVector<int, 8> MaskV;
4671 for (int i = 0; i != 4; ++i) {
4672 int idx = MaskVals[i];
4674 MaskV.push_back(-1);
4676 } else if ((idx / 4) == BestLoQuad) {
4677 MaskV.push_back(idx & 3);
4680 MaskV.push_back(-1);
4683 for (unsigned i = 4; i != 8; ++i)
4685 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4688 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4689 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4691 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4695 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4696 // and update MaskVals with the new element order.
4697 if (BestHiQuad >= 0) {
4698 SmallVector<int, 8> MaskV;
4699 for (unsigned i = 0; i != 4; ++i)
4701 for (unsigned i = 4; i != 8; ++i) {
4702 int idx = MaskVals[i];
4704 MaskV.push_back(-1);
4706 } else if ((idx / 4) == BestHiQuad) {
4707 MaskV.push_back((idx & 3) + 4);
4710 MaskV.push_back(-1);
4713 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4716 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4717 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
4719 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
4723 // In case BestHi & BestLo were both -1, which means each quadword has a word
4724 // from each of the four input quadwords, calculate the InOrder bitvector now
4725 // before falling through to the insert/extract cleanup.
4726 if (BestLoQuad == -1 && BestHiQuad == -1) {
4728 for (int i = 0; i != 8; ++i)
4729 if (MaskVals[i] < 0 || MaskVals[i] == i)
4733 // The other elements are put in the right place using pextrw and pinsrw.
4734 for (unsigned i = 0; i != 8; ++i) {
4737 int EltIdx = MaskVals[i];
4740 SDValue ExtOp = (EltIdx < 8)
4741 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4742 DAG.getIntPtrConstant(EltIdx))
4743 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4744 DAG.getIntPtrConstant(EltIdx - 8));
4745 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4746 DAG.getIntPtrConstant(i));
4751 // v16i8 shuffles - Prefer shuffles in the following order:
4752 // 1. [ssse3] 1 x pshufb
4753 // 2. [ssse3] 2 x pshufb + 1 x por
4754 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4756 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4758 const X86TargetLowering &TLI) {
4759 SDValue V1 = SVOp->getOperand(0);
4760 SDValue V2 = SVOp->getOperand(1);
4761 DebugLoc dl = SVOp->getDebugLoc();
4762 SmallVector<int, 16> MaskVals;
4763 SVOp->getMask(MaskVals);
4765 // If we have SSSE3, case 1 is generated when all result bytes come from
4766 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4767 // present, fall back to case 3.
4768 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4771 for (unsigned i = 0; i < 16; ++i) {
4772 int EltIdx = MaskVals[i];
4781 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4782 if (TLI.getSubtarget()->hasSSSE3()) {
4783 SmallVector<SDValue,16> pshufbMask;
4785 // If all result elements are from one input vector, then only translate
4786 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4788 // Otherwise, we have elements from both input vectors, and must zero out
4789 // elements that come from V2 in the first mask, and V1 in the second mask
4790 // so that we can OR them together.
4791 bool TwoInputs = !(V1Only || V2Only);
4792 for (unsigned i = 0; i != 16; ++i) {
4793 int EltIdx = MaskVals[i];
4794 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4795 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4798 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4800 // If all the elements are from V2, assign it to V1 and return after
4801 // building the first pshufb.
4804 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4805 DAG.getNode(ISD::BUILD_VECTOR, dl,
4806 MVT::v16i8, &pshufbMask[0], 16));
4810 // Calculate the shuffle mask for the second input, shuffle it, and
4811 // OR it with the first shuffled input.
4813 for (unsigned i = 0; i != 16; ++i) {
4814 int EltIdx = MaskVals[i];
4816 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4819 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4821 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4822 DAG.getNode(ISD::BUILD_VECTOR, dl,
4823 MVT::v16i8, &pshufbMask[0], 16));
4824 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4827 // No SSSE3 - Calculate in place words and then fix all out of place words
4828 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4829 // the 16 different words that comprise the two doublequadword input vectors.
4830 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4831 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4832 SDValue NewV = V2Only ? V2 : V1;
4833 for (int i = 0; i != 8; ++i) {
4834 int Elt0 = MaskVals[i*2];
4835 int Elt1 = MaskVals[i*2+1];
4837 // This word of the result is all undef, skip it.
4838 if (Elt0 < 0 && Elt1 < 0)
4841 // This word of the result is already in the correct place, skip it.
4842 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4844 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4847 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4848 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4851 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4852 // using a single extract together, load it and store it.
4853 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4854 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4855 DAG.getIntPtrConstant(Elt1 / 2));
4856 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4857 DAG.getIntPtrConstant(i));
4861 // If Elt1 is defined, extract it from the appropriate source. If the
4862 // source byte is not also odd, shift the extracted word left 8 bits
4863 // otherwise clear the bottom 8 bits if we need to do an or.
4865 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4866 DAG.getIntPtrConstant(Elt1 / 2));
4867 if ((Elt1 & 1) == 0)
4868 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4869 DAG.getConstant(8, TLI.getShiftAmountTy()));
4871 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4872 DAG.getConstant(0xFF00, MVT::i16));
4874 // If Elt0 is defined, extract it from the appropriate source. If the
4875 // source byte is not also even, shift the extracted word right 8 bits. If
4876 // Elt1 was also defined, OR the extracted values together before
4877 // inserting them in the result.
4879 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4880 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4881 if ((Elt0 & 1) != 0)
4882 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4883 DAG.getConstant(8, TLI.getShiftAmountTy()));
4885 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4886 DAG.getConstant(0x00FF, MVT::i16));
4887 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4890 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4891 DAG.getIntPtrConstant(i));
4893 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4896 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4897 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
4898 /// done when every pair / quad of shuffle mask elements point to elements in
4899 /// the right sequence. e.g.
4900 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
4902 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4903 SelectionDAG &DAG, DebugLoc dl) {
4904 EVT VT = SVOp->getValueType(0);
4905 SDValue V1 = SVOp->getOperand(0);
4906 SDValue V2 = SVOp->getOperand(1);
4907 unsigned NumElems = VT.getVectorNumElements();
4908 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4910 switch (VT.getSimpleVT().SimpleTy) {
4911 default: assert(false && "Unexpected!");
4912 case MVT::v4f32: NewVT = MVT::v2f64; break;
4913 case MVT::v4i32: NewVT = MVT::v2i64; break;
4914 case MVT::v8i16: NewVT = MVT::v4i32; break;
4915 case MVT::v16i8: NewVT = MVT::v4i32; break;
4918 int Scale = NumElems / NewWidth;
4919 SmallVector<int, 8> MaskVec;
4920 for (unsigned i = 0; i < NumElems; i += Scale) {
4922 for (int j = 0; j < Scale; ++j) {
4923 int EltIdx = SVOp->getMaskElt(i+j);
4927 StartIdx = EltIdx - (EltIdx % Scale);
4928 if (EltIdx != StartIdx + j)
4932 MaskVec.push_back(-1);
4934 MaskVec.push_back(StartIdx / Scale);
4937 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4938 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4939 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4942 /// getVZextMovL - Return a zero-extending vector move low node.
4944 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4945 SDValue SrcOp, SelectionDAG &DAG,
4946 const X86Subtarget *Subtarget, DebugLoc dl) {
4947 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4948 LoadSDNode *LD = NULL;
4949 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4950 LD = dyn_cast<LoadSDNode>(SrcOp);
4952 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4954 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4955 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4956 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4957 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4958 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4960 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4961 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4962 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4963 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4971 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4972 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4973 DAG.getNode(ISD::BIT_CONVERT, dl,
4977 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4980 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4981 SDValue V1 = SVOp->getOperand(0);
4982 SDValue V2 = SVOp->getOperand(1);
4983 DebugLoc dl = SVOp->getDebugLoc();
4984 EVT VT = SVOp->getValueType(0);
4986 SmallVector<std::pair<int, int>, 8> Locs;
4988 SmallVector<int, 8> Mask1(4U, -1);
4989 SmallVector<int, 8> PermMask;
4990 SVOp->getMask(PermMask);
4994 for (unsigned i = 0; i != 4; ++i) {
4995 int Idx = PermMask[i];
4997 Locs[i] = std::make_pair(-1, -1);
4999 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5001 Locs[i] = std::make_pair(0, NumLo);
5005 Locs[i] = std::make_pair(1, NumHi);
5007 Mask1[2+NumHi] = Idx;
5013 if (NumLo <= 2 && NumHi <= 2) {
5014 // If no more than two elements come from either vector. This can be
5015 // implemented with two shuffles. First shuffle gather the elements.
5016 // The second shuffle, which takes the first shuffle as both of its
5017 // vector operands, put the elements into the right order.
5018 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5020 SmallVector<int, 8> Mask2(4U, -1);
5022 for (unsigned i = 0; i != 4; ++i) {
5023 if (Locs[i].first == -1)
5026 unsigned Idx = (i < 2) ? 0 : 4;
5027 Idx += Locs[i].first * 2 + Locs[i].second;
5032 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5033 } else if (NumLo == 3 || NumHi == 3) {
5034 // Otherwise, we must have three elements from one vector, call it X, and
5035 // one element from the other, call it Y. First, use a shufps to build an
5036 // intermediate vector with the one element from Y and the element from X
5037 // that will be in the same half in the final destination (the indexes don't
5038 // matter). Then, use a shufps to build the final vector, taking the half
5039 // containing the element from Y from the intermediate, and the other half
5042 // Normalize it so the 3 elements come from V1.
5043 CommuteVectorShuffleMask(PermMask, VT);
5047 // Find the element from V2.
5049 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5050 int Val = PermMask[HiIndex];
5057 Mask1[0] = PermMask[HiIndex];
5059 Mask1[2] = PermMask[HiIndex^1];
5061 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5064 Mask1[0] = PermMask[0];
5065 Mask1[1] = PermMask[1];
5066 Mask1[2] = HiIndex & 1 ? 6 : 4;
5067 Mask1[3] = HiIndex & 1 ? 4 : 6;
5068 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5070 Mask1[0] = HiIndex & 1 ? 2 : 0;
5071 Mask1[1] = HiIndex & 1 ? 0 : 2;
5072 Mask1[2] = PermMask[2];
5073 Mask1[3] = PermMask[3];
5078 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5082 // Break it into (shuffle shuffle_hi, shuffle_lo).
5084 SmallVector<int,8> LoMask(4U, -1);
5085 SmallVector<int,8> HiMask(4U, -1);
5087 SmallVector<int,8> *MaskPtr = &LoMask;
5088 unsigned MaskIdx = 0;
5091 for (unsigned i = 0; i != 4; ++i) {
5098 int Idx = PermMask[i];
5100 Locs[i] = std::make_pair(-1, -1);
5101 } else if (Idx < 4) {
5102 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5103 (*MaskPtr)[LoIdx] = Idx;
5106 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5107 (*MaskPtr)[HiIdx] = Idx;
5112 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5113 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5114 SmallVector<int, 8> MaskOps;
5115 for (unsigned i = 0; i != 4; ++i) {
5116 if (Locs[i].first == -1) {
5117 MaskOps.push_back(-1);
5119 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5120 MaskOps.push_back(Idx);
5123 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5126 static bool MayFoldVectorLoad(SDValue V) {
5127 if (V.hasOneUse() && V.getOpcode() == ISD::BIT_CONVERT)
5128 V = V.getOperand(0);
5129 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5130 V = V.getOperand(0);
5136 // FIXME: the version above should always be used. Since there's
5137 // a bug where several vector shuffles can't be folded because the
5138 // DAG is not updated during lowering and a node claims to have two
5139 // uses while it only has one, use this version, and let isel match
5140 // another instruction if the load really happens to have more than
5141 // one use. Remove this version after this bug get fixed.
5142 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5143 if (V.hasOneUse() && V.getOpcode() == ISD::BIT_CONVERT)
5144 V = V.getOperand(0);
5145 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5146 V = V.getOperand(0);
5147 if (ISD::isNormalLoad(V.getNode()))
5152 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5153 /// a vector extract, and if both can be later optimized into a single load.
5154 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5155 /// here because otherwise a target specific shuffle node is going to be
5156 /// emitted for this shuffle, and the optimization not done.
5157 /// FIXME: This is probably not the best approach, but fix the problem
5158 /// until the right path is decided.
5160 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5161 const TargetLowering &TLI) {
5162 EVT VT = V.getValueType();
5163 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5165 // Be sure that the vector shuffle is present in a pattern like this:
5166 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5170 SDNode *N = *V.getNode()->use_begin();
5171 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5174 SDValue EltNo = N->getOperand(1);
5175 if (!isa<ConstantSDNode>(EltNo))
5178 // If the bit convert changed the number of elements, it is unsafe
5179 // to examine the mask.
5180 bool HasShuffleIntoBitcast = false;
5181 if (V.getOpcode() == ISD::BIT_CONVERT) {
5182 EVT SrcVT = V.getOperand(0).getValueType();
5183 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5185 V = V.getOperand(0);
5186 HasShuffleIntoBitcast = true;
5189 // Select the input vector, guarding against out of range extract vector.
5190 unsigned NumElems = VT.getVectorNumElements();
5191 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5192 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5193 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5195 // Skip one more bit_convert if necessary
5196 if (V.getOpcode() == ISD::BIT_CONVERT)
5197 V = V.getOperand(0);
5199 if (ISD::isNormalLoad(V.getNode())) {
5200 // Is the original load suitable?
5201 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5203 // FIXME: avoid the multi-use bug that is preventing lots of
5204 // of foldings to be detected, this is still wrong of course, but
5205 // give the temporary desired behavior, and if it happens that
5206 // the load has real more uses, during isel it will not fold, and
5207 // will generate poor code.
5208 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5211 if (!HasShuffleIntoBitcast)
5214 // If there's a bitcast before the shuffle, check if the load type and
5215 // alignment is valid.
5216 unsigned Align = LN0->getAlignment();
5218 TLI.getTargetData()->getABITypeAlignment(
5219 VT.getTypeForEVT(*DAG.getContext()));
5221 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
5229 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5231 SDValue V1 = Op.getOperand(0);
5232 SDValue V2 = Op.getOperand(1);
5233 EVT VT = Op.getValueType();
5235 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5237 if (HasSSE2 && VT == MVT::v2f64)
5238 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5241 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5245 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5246 SDValue V1 = Op.getOperand(0);
5247 SDValue V2 = Op.getOperand(1);
5248 EVT VT = Op.getValueType();
5250 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5251 "unsupported shuffle type");
5253 if (V2.getOpcode() == ISD::UNDEF)
5257 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5261 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5262 SDValue V1 = Op.getOperand(0);
5263 SDValue V2 = Op.getOperand(1);
5264 EVT VT = Op.getValueType();
5265 unsigned NumElems = VT.getVectorNumElements();
5267 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5268 // operand of these instructions is only memory, so check if there's a
5269 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5271 bool CanFoldLoad = false;
5273 // Trivial case, when V2 comes from a load.
5274 if (MayFoldVectorLoad(V2))
5277 // When V1 is a load, it can be folded later into a store in isel, example:
5278 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5280 // (MOVLPSmr addr:$src1, VR128:$src2)
5281 // So, recognize this potential and also use MOVLPS or MOVLPD
5282 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5286 if (HasSSE2 && NumElems == 2)
5287 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5290 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5293 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5294 // movl and movlp will both match v2i64, but v2i64 is never matched by
5295 // movl earlier because we make it strict to avoid messing with the movlp load
5296 // folding logic (see the code above getMOVLP call). Match it here then,
5297 // this is horrible, but will stay like this until we move all shuffle
5298 // matching to x86 specific nodes. Note that for the 1st condition all
5299 // types are matched with movsd.
5300 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5301 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5303 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5306 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5308 // Invert the operand order and use SHUFPS to match it.
5309 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5310 X86::getShuffleSHUFImmediate(SVOp), DAG);
5313 static inline unsigned getUNPCKLOpcode(EVT VT) {
5314 switch(VT.getSimpleVT().SimpleTy) {
5315 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5316 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5317 case MVT::v4f32: return X86ISD::UNPCKLPS;
5318 case MVT::v2f64: return X86ISD::UNPCKLPD;
5319 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5320 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5322 llvm_unreachable("Unknow type for unpckl");
5327 static inline unsigned getUNPCKHOpcode(EVT VT) {
5328 switch(VT.getSimpleVT().SimpleTy) {
5329 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5330 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5331 case MVT::v4f32: return X86ISD::UNPCKHPS;
5332 case MVT::v2f64: return X86ISD::UNPCKHPD;
5333 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5334 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5336 llvm_unreachable("Unknow type for unpckh");
5342 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
5343 const TargetLowering &TLI,
5344 const X86Subtarget *Subtarget) {
5345 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5346 EVT VT = Op.getValueType();
5347 DebugLoc dl = Op.getDebugLoc();
5348 SDValue V1 = Op.getOperand(0);
5349 SDValue V2 = Op.getOperand(1);
5351 if (isZeroShuffle(SVOp))
5352 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5354 // Handle splat operations
5355 if (SVOp->isSplat()) {
5356 // Special case, this is the only place now where it's
5357 // allowed to return a vector_shuffle operation without
5358 // using a target specific node, because *hopefully* it
5359 // will be optimized away by the dag combiner.
5360 if (VT.getVectorNumElements() <= 4 &&
5361 CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
5364 // Handle splats by matching through known masks
5365 if (VT.getVectorNumElements() <= 4)
5368 // Canonize all of the remaining to v4f32.
5369 return PromoteSplat(SVOp, DAG);
5372 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5374 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5375 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5376 if (NewOp.getNode())
5377 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, NewOp);
5378 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5379 // FIXME: Figure out a cleaner way to do this.
5380 // Try to make use of movq to zero out the top part.
5381 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5382 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5383 if (NewOp.getNode()) {
5384 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5385 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5386 DAG, Subtarget, dl);
5388 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5389 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5390 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5391 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5392 DAG, Subtarget, dl);
5399 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5400 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5401 SDValue V1 = Op.getOperand(0);
5402 SDValue V2 = Op.getOperand(1);
5403 EVT VT = Op.getValueType();
5404 DebugLoc dl = Op.getDebugLoc();
5405 unsigned NumElems = VT.getVectorNumElements();
5406 bool isMMX = VT.getSizeInBits() == 64;
5407 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5408 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5409 bool V1IsSplat = false;
5410 bool V2IsSplat = false;
5411 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5412 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5413 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
5414 MachineFunction &MF = DAG.getMachineFunction();
5415 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5417 // FIXME: this is somehow handled during isel by MMX pattern fragments. Remove
5418 // the check or come up with another solution when all MMX move to intrinsics,
5419 // but don't allow this to be considered legal, we don't want vector_shuffle
5420 // operations to be matched during isel anymore.
5421 if (isMMX && SVOp->isSplat())
5424 // Vector shuffle lowering takes 3 steps:
5426 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
5427 // narrowing and commutation of operands should be handled.
5428 // 2) Matching of shuffles with known shuffle masks to x86 target specific
5430 // 3) Rewriting of unmatched masks into new generic shuffle operations,
5431 // so the shuffle can be broken into other shuffles and the legalizer can
5432 // try the lowering again.
5434 // The general ideia is that no vector_shuffle operation should be left to
5435 // be matched during isel, all of them must be converted to a target specific
5438 // Normalize the input vectors. Here splats, zeroed vectors, profitable
5439 // narrowing and commutation of operands should be handled. The actual code
5440 // doesn't include all of those, work in progress...
5441 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
5442 if (NewOp.getNode())
5445 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
5446 // unpckh_undef). Only use pshufd if speed is more important than size.
5447 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
5448 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5449 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5450 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
5451 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5452 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5454 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
5455 RelaxedMayFoldVectorLoad(V1) && !isMMX)
5456 return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
5458 if (!isMMX && X86::isMOVHLPS_v_undef_Mask(SVOp))
5459 return getMOVHighToLow(Op, dl, DAG);
5461 // Use to match splats
5462 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
5463 (VT == MVT::v2f64 || VT == MVT::v2i64))
5464 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5466 if (X86::isPSHUFDMask(SVOp)) {
5467 // The actual implementation will match the mask in the if above and then
5468 // during isel it can match several different instructions, not only pshufd
5469 // as its name says, sad but true, emulate the behavior for now...
5470 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5471 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5473 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5475 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5476 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5478 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5479 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5482 if (VT == MVT::v4f32)
5483 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5487 // Check if this can be converted into a logical shift.
5488 bool isLeft = false;
5491 bool isShift = getSubtarget()->hasSSE2() &&
5492 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5493 if (isShift && ShVal.hasOneUse()) {
5494 // If the shifted value has multiple uses, it may be cheaper to use
5495 // v_set0 + movlhps or movhlps, etc.
5496 EVT EltVT = VT.getVectorElementType();
5497 ShAmt *= EltVT.getSizeInBits();
5498 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5501 if (X86::isMOVLMask(SVOp)) {
5504 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5505 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5506 if (!isMMX && !X86::isMOVLPMask(SVOp)) {
5507 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5508 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5510 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5511 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5515 // FIXME: fold these into legal mask.
5517 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5518 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5520 if (X86::isMOVHLPSMask(SVOp))
5521 return getMOVHighToLow(Op, dl, DAG);
5523 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5524 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5526 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5527 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5529 if (X86::isMOVLPMask(SVOp))
5530 return getMOVLP(Op, dl, DAG, HasSSE2);
5533 if (ShouldXformToMOVHLPS(SVOp) ||
5534 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5535 return CommuteVectorShuffle(SVOp, DAG);
5538 // No better options. Use a vshl / vsrl.
5539 EVT EltVT = VT.getVectorElementType();
5540 ShAmt *= EltVT.getSizeInBits();
5541 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5544 bool Commuted = false;
5545 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5546 // 1,1,1,1 -> v8i16 though.
5547 V1IsSplat = isSplatVector(V1.getNode());
5548 V2IsSplat = isSplatVector(V2.getNode());
5550 // Canonicalize the splat or undef, if present, to be on the RHS.
5551 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5552 Op = CommuteVectorShuffle(SVOp, DAG);
5553 SVOp = cast<ShuffleVectorSDNode>(Op);
5554 V1 = SVOp->getOperand(0);
5555 V2 = SVOp->getOperand(1);
5556 std::swap(V1IsSplat, V2IsSplat);
5557 std::swap(V1IsUndef, V2IsUndef);
5561 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5562 // Shuffling low element of v1 into undef, just return v1.
5565 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5566 // the instruction selector will not match, so get a canonical MOVL with
5567 // swapped operands to undo the commute.
5568 return getMOVL(DAG, dl, VT, V2, V1);
5571 if (X86::isUNPCKLMask(SVOp))
5573 Op : getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
5575 if (X86::isUNPCKHMask(SVOp))
5577 Op : getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
5580 // Normalize mask so all entries that point to V2 points to its first
5581 // element then try to match unpck{h|l} again. If match, return a
5582 // new vector_shuffle with the corrected mask.
5583 SDValue NewMask = NormalizeMask(SVOp, DAG);
5584 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5585 if (NSVOp != SVOp) {
5586 if (X86::isUNPCKLMask(NSVOp, true)) {
5588 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5595 // Commute is back and try unpck* again.
5596 // FIXME: this seems wrong.
5597 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5598 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5600 if (X86::isUNPCKLMask(NewSVOp))
5602 NewOp : getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
5604 if (X86::isUNPCKHMask(NewSVOp))
5606 NewOp : getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
5609 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
5611 // Normalize the node to match x86 shuffle ops if needed
5612 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5613 return CommuteVectorShuffle(SVOp, DAG);
5615 // The checks below are all present in isShuffleMaskLegal, but they are
5616 // inlined here right now to enable us to directly emit target specific
5617 // nodes, and remove one by one until they don't return Op anymore.
5618 SmallVector<int, 16> M;
5621 if (isPALIGNRMask(M, VT, HasSSSE3))
5622 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
5623 X86::getShufflePALIGNRImmediate(SVOp),
5626 // Only a few shuffle masks are handled for 64-bit vectors (MMX), and
5627 // 64-bit vectors which made to this point can't be handled, they are
5632 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
5633 SVOp->getSplatIndex() == 0 && V2IsUndef) {
5634 if (VT == MVT::v2f64)
5635 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
5636 if (VT == MVT::v2i64)
5637 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
5640 if (isPSHUFHWMask(M, VT))
5641 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
5642 X86::getShufflePSHUFHWImmediate(SVOp),
5645 if (isPSHUFLWMask(M, VT))
5646 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
5647 X86::getShufflePSHUFLWImmediate(SVOp),
5650 if (isSHUFPMask(M, VT)) {
5651 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5652 if (VT == MVT::v4f32 || VT == MVT::v4i32)
5653 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
5655 if (VT == MVT::v2f64 || VT == MVT::v2i64)
5656 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
5660 if (X86::isUNPCKL_v_undef_Mask(SVOp))
5661 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5662 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5663 if (X86::isUNPCKH_v_undef_Mask(SVOp))
5664 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5665 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5667 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5668 if (VT == MVT::v8i16) {
5669 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5670 if (NewOp.getNode())
5674 if (VT == MVT::v16i8) {
5675 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5676 if (NewOp.getNode())
5680 // Handle all 4 wide cases with a number of shuffles except for MMX.
5681 if (NumElems == 4 && !isMMX)
5682 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5688 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5689 SelectionDAG &DAG) const {
5690 EVT VT = Op.getValueType();
5691 DebugLoc dl = Op.getDebugLoc();
5692 if (VT.getSizeInBits() == 8) {
5693 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
5694 Op.getOperand(0), Op.getOperand(1));
5695 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5696 DAG.getValueType(VT));
5697 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5698 } else if (VT.getSizeInBits() == 16) {
5699 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5700 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
5702 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5703 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5704 DAG.getNode(ISD::BIT_CONVERT, dl,
5708 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
5709 Op.getOperand(0), Op.getOperand(1));
5710 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5711 DAG.getValueType(VT));
5712 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5713 } else if (VT == MVT::f32) {
5714 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
5715 // the result back to FR32 register. It's only worth matching if the
5716 // result has a single use which is a store or a bitcast to i32. And in
5717 // the case of a store, it's not worth it if the index is a constant 0,
5718 // because a MOVSSmr can be used instead, which is smaller and faster.
5719 if (!Op.hasOneUse())
5721 SDNode *User = *Op.getNode()->use_begin();
5722 if ((User->getOpcode() != ISD::STORE ||
5723 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5724 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5725 (User->getOpcode() != ISD::BIT_CONVERT ||
5726 User->getValueType(0) != MVT::i32))
5728 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5729 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
5732 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
5733 } else if (VT == MVT::i32) {
5734 // ExtractPS works with constant index.
5735 if (isa<ConstantSDNode>(Op.getOperand(1)))
5743 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5744 SelectionDAG &DAG) const {
5745 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5748 if (Subtarget->hasSSE41()) {
5749 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5754 EVT VT = Op.getValueType();
5755 DebugLoc dl = Op.getDebugLoc();
5756 // TODO: handle v16i8.
5757 if (VT.getSizeInBits() == 16) {
5758 SDValue Vec = Op.getOperand(0);
5759 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5761 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5762 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5763 DAG.getNode(ISD::BIT_CONVERT, dl,
5766 // Transform it so it match pextrw which produces a 32-bit result.
5767 EVT EltVT = MVT::i32;
5768 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5769 Op.getOperand(0), Op.getOperand(1));
5770 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5771 DAG.getValueType(VT));
5772 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5773 } else if (VT.getSizeInBits() == 32) {
5774 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5778 // SHUFPS the element to the lowest double word, then movss.
5779 int Mask[4] = { Idx, -1, -1, -1 };
5780 EVT VVT = Op.getOperand(0).getValueType();
5781 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5782 DAG.getUNDEF(VVT), Mask);
5783 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5784 DAG.getIntPtrConstant(0));
5785 } else if (VT.getSizeInBits() == 64) {
5786 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5787 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5788 // to match extract_elt for f64.
5789 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5793 // UNPCKHPD the element to the lowest double word, then movsd.
5794 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5795 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5796 int Mask[2] = { 1, -1 };
5797 EVT VVT = Op.getOperand(0).getValueType();
5798 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5799 DAG.getUNDEF(VVT), Mask);
5800 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5801 DAG.getIntPtrConstant(0));
5808 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5809 SelectionDAG &DAG) const {
5810 EVT VT = Op.getValueType();
5811 EVT EltVT = VT.getVectorElementType();
5812 DebugLoc dl = Op.getDebugLoc();
5814 SDValue N0 = Op.getOperand(0);
5815 SDValue N1 = Op.getOperand(1);
5816 SDValue N2 = Op.getOperand(2);
5818 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5819 isa<ConstantSDNode>(N2)) {
5821 if (VT == MVT::v8i16)
5822 Opc = X86ISD::PINSRW;
5823 else if (VT == MVT::v4i16)
5824 Opc = X86ISD::MMX_PINSRW;
5825 else if (VT == MVT::v16i8)
5826 Opc = X86ISD::PINSRB;
5828 Opc = X86ISD::PINSRB;
5830 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5832 if (N1.getValueType() != MVT::i32)
5833 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5834 if (N2.getValueType() != MVT::i32)
5835 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5836 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5837 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5838 // Bits [7:6] of the constant are the source select. This will always be
5839 // zero here. The DAG Combiner may combine an extract_elt index into these
5840 // bits. For example (insert (extract, 3), 2) could be matched by putting
5841 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5842 // Bits [5:4] of the constant are the destination select. This is the
5843 // value of the incoming immediate.
5844 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5845 // combine either bitwise AND or insert of float 0.0 to set these bits.
5846 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5847 // Create this as a scalar to vector..
5848 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5849 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5850 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5851 // PINSR* works with constant index.
5858 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5859 EVT VT = Op.getValueType();
5860 EVT EltVT = VT.getVectorElementType();
5862 if (Subtarget->hasSSE41())
5863 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5865 if (EltVT == MVT::i8)
5868 DebugLoc dl = Op.getDebugLoc();
5869 SDValue N0 = Op.getOperand(0);
5870 SDValue N1 = Op.getOperand(1);
5871 SDValue N2 = Op.getOperand(2);
5873 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5874 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5875 // as its second argument.
5876 if (N1.getValueType() != MVT::i32)
5877 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5878 if (N2.getValueType() != MVT::i32)
5879 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5880 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5881 dl, VT, N0, N1, N2);
5887 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5888 DebugLoc dl = Op.getDebugLoc();
5890 if (Op.getValueType() == MVT::v1i64 &&
5891 Op.getOperand(0).getValueType() == MVT::i64)
5892 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5894 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5895 EVT VT = MVT::v2i32;
5896 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5903 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5904 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5907 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5908 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5909 // one of the above mentioned nodes. It has to be wrapped because otherwise
5910 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5911 // be used to form addressing mode. These wrapped nodes will be selected
5914 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5915 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5917 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5919 unsigned char OpFlag = 0;
5920 unsigned WrapperKind = X86ISD::Wrapper;
5921 CodeModel::Model M = getTargetMachine().getCodeModel();
5923 if (Subtarget->isPICStyleRIPRel() &&
5924 (M == CodeModel::Small || M == CodeModel::Kernel))
5925 WrapperKind = X86ISD::WrapperRIP;
5926 else if (Subtarget->isPICStyleGOT())
5927 OpFlag = X86II::MO_GOTOFF;
5928 else if (Subtarget->isPICStyleStubPIC())
5929 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5931 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5933 CP->getOffset(), OpFlag);
5934 DebugLoc DL = CP->getDebugLoc();
5935 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5936 // With PIC, the address is actually $g + Offset.
5938 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5939 DAG.getNode(X86ISD::GlobalBaseReg,
5940 DebugLoc(), getPointerTy()),
5947 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5948 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5950 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5952 unsigned char OpFlag = 0;
5953 unsigned WrapperKind = X86ISD::Wrapper;
5954 CodeModel::Model M = getTargetMachine().getCodeModel();
5956 if (Subtarget->isPICStyleRIPRel() &&
5957 (M == CodeModel::Small || M == CodeModel::Kernel))
5958 WrapperKind = X86ISD::WrapperRIP;
5959 else if (Subtarget->isPICStyleGOT())
5960 OpFlag = X86II::MO_GOTOFF;
5961 else if (Subtarget->isPICStyleStubPIC())
5962 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5964 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5966 DebugLoc DL = JT->getDebugLoc();
5967 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5969 // With PIC, the address is actually $g + Offset.
5971 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5972 DAG.getNode(X86ISD::GlobalBaseReg,
5973 DebugLoc(), getPointerTy()),
5981 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5982 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5984 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5986 unsigned char OpFlag = 0;
5987 unsigned WrapperKind = X86ISD::Wrapper;
5988 CodeModel::Model M = getTargetMachine().getCodeModel();
5990 if (Subtarget->isPICStyleRIPRel() &&
5991 (M == CodeModel::Small || M == CodeModel::Kernel))
5992 WrapperKind = X86ISD::WrapperRIP;
5993 else if (Subtarget->isPICStyleGOT())
5994 OpFlag = X86II::MO_GOTOFF;
5995 else if (Subtarget->isPICStyleStubPIC())
5996 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5998 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
6000 DebugLoc DL = Op.getDebugLoc();
6001 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6004 // With PIC, the address is actually $g + Offset.
6005 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
6006 !Subtarget->is64Bit()) {
6007 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6008 DAG.getNode(X86ISD::GlobalBaseReg,
6009 DebugLoc(), getPointerTy()),
6017 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
6018 // Create the TargetBlockAddressAddress node.
6019 unsigned char OpFlags =
6020 Subtarget->ClassifyBlockAddressReference();
6021 CodeModel::Model M = getTargetMachine().getCodeModel();
6022 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
6023 DebugLoc dl = Op.getDebugLoc();
6024 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
6025 /*isTarget=*/true, OpFlags);
6027 if (Subtarget->isPICStyleRIPRel() &&
6028 (M == CodeModel::Small || M == CodeModel::Kernel))
6029 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6031 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6033 // With PIC, the address is actually $g + Offset.
6034 if (isGlobalRelativeToPICBase(OpFlags)) {
6035 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6036 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6044 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
6046 SelectionDAG &DAG) const {
6047 // Create the TargetGlobalAddress node, folding in the constant
6048 // offset if it is legal.
6049 unsigned char OpFlags =
6050 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
6051 CodeModel::Model M = getTargetMachine().getCodeModel();
6053 if (OpFlags == X86II::MO_NO_FLAG &&
6054 X86::isOffsetSuitableForCodeModel(Offset, M)) {
6055 // A direct static reference to a global.
6056 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
6059 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6062 if (Subtarget->isPICStyleRIPRel() &&
6063 (M == CodeModel::Small || M == CodeModel::Kernel))
6064 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6066 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6068 // With PIC, the address is actually $g + Offset.
6069 if (isGlobalRelativeToPICBase(OpFlags)) {
6070 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6071 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6075 // For globals that require a load from a stub to get the address, emit the
6077 if (isGlobalStubReference(OpFlags))
6078 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
6079 MachinePointerInfo::getGOT(), false, false, 0);
6081 // If there was a non-zero offset that we didn't fold, create an explicit
6084 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
6085 DAG.getConstant(Offset, getPointerTy()));
6091 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
6092 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
6093 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
6094 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
6098 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
6099 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
6100 unsigned char OperandFlags) {
6101 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6102 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6103 DebugLoc dl = GA->getDebugLoc();
6104 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6105 GA->getValueType(0),
6109 SDValue Ops[] = { Chain, TGA, *InFlag };
6110 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
6112 SDValue Ops[] = { Chain, TGA };
6113 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
6116 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
6117 MFI->setAdjustsStack(true);
6119 SDValue Flag = Chain.getValue(1);
6120 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
6123 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
6125 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6128 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
6129 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
6130 DAG.getNode(X86ISD::GlobalBaseReg,
6131 DebugLoc(), PtrVT), InFlag);
6132 InFlag = Chain.getValue(1);
6134 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
6137 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
6139 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6141 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
6142 X86::RAX, X86II::MO_TLSGD);
6145 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
6146 // "local exec" model.
6147 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6148 const EVT PtrVT, TLSModel::Model model,
6150 DebugLoc dl = GA->getDebugLoc();
6151 // Get the Thread Pointer
6152 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
6154 DAG.getRegister(is64Bit? X86::FS : X86::GS,
6157 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
6158 MachinePointerInfo(), false, false, 0);
6160 unsigned char OperandFlags = 0;
6161 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6163 unsigned WrapperKind = X86ISD::Wrapper;
6164 if (model == TLSModel::LocalExec) {
6165 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6166 } else if (is64Bit) {
6167 assert(model == TLSModel::InitialExec);
6168 OperandFlags = X86II::MO_GOTTPOFF;
6169 WrapperKind = X86ISD::WrapperRIP;
6171 assert(model == TLSModel::InitialExec);
6172 OperandFlags = X86II::MO_INDNTPOFF;
6175 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6177 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6178 GA->getValueType(0),
6179 GA->getOffset(), OperandFlags);
6180 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6182 if (model == TLSModel::InitialExec)
6183 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6184 MachinePointerInfo::getGOT(), false, false, 0);
6186 // The address of the thread local variable is the add of the thread
6187 // pointer with the offset of the variable.
6188 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6192 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6194 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6195 const GlobalValue *GV = GA->getGlobal();
6197 if (Subtarget->isTargetELF()) {
6198 // TODO: implement the "local dynamic" model
6199 // TODO: implement the "initial exec"model for pic executables
6201 // If GV is an alias then use the aliasee for determining
6202 // thread-localness.
6203 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6204 GV = GA->resolveAliasedGlobal(false);
6206 TLSModel::Model model
6207 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6210 case TLSModel::GeneralDynamic:
6211 case TLSModel::LocalDynamic: // not implemented
6212 if (Subtarget->is64Bit())
6213 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6214 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6216 case TLSModel::InitialExec:
6217 case TLSModel::LocalExec:
6218 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6219 Subtarget->is64Bit());
6221 } else if (Subtarget->isTargetDarwin()) {
6222 // Darwin only has one model of TLS. Lower to that.
6223 unsigned char OpFlag = 0;
6224 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6225 X86ISD::WrapperRIP : X86ISD::Wrapper;
6227 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6229 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6230 !Subtarget->is64Bit();
6232 OpFlag = X86II::MO_TLVP_PIC_BASE;
6234 OpFlag = X86II::MO_TLVP;
6235 DebugLoc DL = Op.getDebugLoc();
6236 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6238 GA->getOffset(), OpFlag);
6239 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6241 // With PIC32, the address is actually $g + Offset.
6243 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6244 DAG.getNode(X86ISD::GlobalBaseReg,
6245 DebugLoc(), getPointerTy()),
6248 // Lowering the machine isd will make sure everything is in the right
6250 SDValue Args[] = { Offset };
6251 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
6253 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6254 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6255 MFI->setAdjustsStack(true);
6257 // And our return value (tls address) is in the standard call return value
6259 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
6260 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6264 "TLS not implemented for this target.");
6266 llvm_unreachable("Unreachable");
6271 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
6272 /// take a 2 x i32 value to shift plus a shift amount.
6273 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
6274 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6275 EVT VT = Op.getValueType();
6276 unsigned VTBits = VT.getSizeInBits();
6277 DebugLoc dl = Op.getDebugLoc();
6278 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6279 SDValue ShOpLo = Op.getOperand(0);
6280 SDValue ShOpHi = Op.getOperand(1);
6281 SDValue ShAmt = Op.getOperand(2);
6282 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6283 DAG.getConstant(VTBits - 1, MVT::i8))
6284 : DAG.getConstant(0, VT);
6287 if (Op.getOpcode() == ISD::SHL_PARTS) {
6288 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6289 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6291 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6292 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6295 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6296 DAG.getConstant(VTBits, MVT::i8));
6297 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6298 AndNode, DAG.getConstant(0, MVT::i8));
6301 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6302 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6303 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6305 if (Op.getOpcode() == ISD::SHL_PARTS) {
6306 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6307 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6309 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6310 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6313 SDValue Ops[2] = { Lo, Hi };
6314 return DAG.getMergeValues(Ops, 2, dl);
6317 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6318 SelectionDAG &DAG) const {
6319 EVT SrcVT = Op.getOperand(0).getValueType();
6321 if (SrcVT.isVector()) {
6322 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
6328 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6329 "Unknown SINT_TO_FP to lower!");
6331 // These are really Legal; return the operand so the caller accepts it as
6333 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6335 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6336 Subtarget->is64Bit()) {
6340 DebugLoc dl = Op.getDebugLoc();
6341 unsigned Size = SrcVT.getSizeInBits()/8;
6342 MachineFunction &MF = DAG.getMachineFunction();
6343 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6344 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6345 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6347 MachinePointerInfo::getFixedStack(SSFI),
6349 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6352 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6354 SelectionDAG &DAG) const {
6356 DebugLoc dl = Op.getDebugLoc();
6358 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6360 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
6362 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6363 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6364 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
6365 Tys, Ops, array_lengthof(Ops));
6368 Chain = Result.getValue(1);
6369 SDValue InFlag = Result.getValue(2);
6371 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6372 // shouldn't be necessary except that RFP cannot be live across
6373 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6374 MachineFunction &MF = DAG.getMachineFunction();
6375 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
6376 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6377 Tys = DAG.getVTList(MVT::Other);
6379 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6381 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
6382 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
6383 MachinePointerInfo::getFixedStack(SSFI),
6390 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6391 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6392 SelectionDAG &DAG) const {
6393 // This algorithm is not obvious. Here it is in C code, more or less:
6395 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6396 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6397 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6399 // Copy ints to xmm registers.
6400 __m128i xh = _mm_cvtsi32_si128( hi );
6401 __m128i xl = _mm_cvtsi32_si128( lo );
6403 // Combine into low half of a single xmm register.
6404 __m128i x = _mm_unpacklo_epi32( xh, xl );
6408 // Merge in appropriate exponents to give the integer bits the right
6410 x = _mm_unpacklo_epi32( x, exp );
6412 // Subtract away the biases to deal with the IEEE-754 double precision
6414 d = _mm_sub_pd( (__m128d) x, bias );
6416 // All conversions up to here are exact. The correctly rounded result is
6417 // calculated using the current rounding mode using the following
6419 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6420 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6421 // store doesn't really need to be here (except
6422 // maybe to zero the other double)
6427 DebugLoc dl = Op.getDebugLoc();
6428 LLVMContext *Context = DAG.getContext();
6430 // Build some magic constants.
6431 std::vector<Constant*> CV0;
6432 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6433 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6434 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6435 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6436 Constant *C0 = ConstantVector::get(CV0);
6437 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6439 std::vector<Constant*> CV1;
6441 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6443 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6444 Constant *C1 = ConstantVector::get(CV1);
6445 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6447 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6448 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6450 DAG.getIntPtrConstant(1)));
6451 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6452 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6454 DAG.getIntPtrConstant(0)));
6455 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6456 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6457 MachinePointerInfo::getConstantPool(),
6459 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6460 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
6461 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6462 MachinePointerInfo::getConstantPool(),
6464 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6466 // Add the halves; easiest way is to swap them into another reg first.
6467 int ShufMask[2] = { 1, -1 };
6468 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6469 DAG.getUNDEF(MVT::v2f64), ShufMask);
6470 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6471 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6472 DAG.getIntPtrConstant(0));
6475 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6476 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6477 SelectionDAG &DAG) const {
6478 DebugLoc dl = Op.getDebugLoc();
6479 // FP constant to bias correct the final result.
6480 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6483 // Load the 32-bit value into an XMM register.
6484 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6485 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6487 DAG.getIntPtrConstant(0)));
6489 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6490 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
6491 DAG.getIntPtrConstant(0));
6493 // Or the load with the bias.
6494 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6495 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6496 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6498 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6499 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6500 MVT::v2f64, Bias)));
6501 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6502 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
6503 DAG.getIntPtrConstant(0));
6505 // Subtract the bias.
6506 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
6508 // Handle final rounding.
6509 EVT DestVT = Op.getValueType();
6511 if (DestVT.bitsLT(MVT::f64)) {
6512 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
6513 DAG.getIntPtrConstant(0));
6514 } else if (DestVT.bitsGT(MVT::f64)) {
6515 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
6518 // Handle final rounding.
6522 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
6523 SelectionDAG &DAG) const {
6524 SDValue N0 = Op.getOperand(0);
6525 DebugLoc dl = Op.getDebugLoc();
6527 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
6528 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
6529 // the optimization here.
6530 if (DAG.SignBitIsZero(N0))
6531 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
6533 EVT SrcVT = N0.getValueType();
6534 EVT DstVT = Op.getValueType();
6535 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
6536 return LowerUINT_TO_FP_i64(Op, DAG);
6537 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
6538 return LowerUINT_TO_FP_i32(Op, DAG);
6540 // Make a 64-bit buffer, and use it to build an FILD.
6541 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
6542 if (SrcVT == MVT::i32) {
6543 SDValue WordOff = DAG.getConstant(4, getPointerTy());
6544 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
6545 getPointerTy(), StackSlot, WordOff);
6546 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6547 StackSlot, MachinePointerInfo(),
6549 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
6550 OffsetSlot, MachinePointerInfo(),
6552 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
6556 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
6557 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6558 StackSlot, MachinePointerInfo(),
6560 // For i64 source, we need to add the appropriate power of 2 if the input
6561 // was negative. This is the same as the optimization in
6562 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
6563 // we must be careful to do the computation in x87 extended precision, not
6564 // in SSE. (The generic code can't know it's OK to do this, or how to.)
6565 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
6566 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
6567 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
6569 APInt FF(32, 0x5F800000ULL);
6571 // Check whether the sign bit is set.
6572 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
6573 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
6576 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
6577 SDValue FudgePtr = DAG.getConstantPool(
6578 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
6581 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
6582 SDValue Zero = DAG.getIntPtrConstant(0);
6583 SDValue Four = DAG.getIntPtrConstant(4);
6584 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
6586 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
6588 // Load the value out, extending it from f32 to f80.
6589 // FIXME: Avoid the extend by constructing the right constant pool?
6590 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
6591 FudgePtr, MachinePointerInfo::getConstantPool(),
6592 MVT::f32, false, false, 4);
6593 // Extend everything to 80 bits to force it to be done on x87.
6594 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
6595 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
6598 std::pair<SDValue,SDValue> X86TargetLowering::
6599 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
6600 DebugLoc dl = Op.getDebugLoc();
6602 EVT DstTy = Op.getValueType();
6605 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
6609 assert(DstTy.getSimpleVT() <= MVT::i64 &&
6610 DstTy.getSimpleVT() >= MVT::i16 &&
6611 "Unknown FP_TO_SINT to lower!");
6613 // These are really Legal.
6614 if (DstTy == MVT::i32 &&
6615 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6616 return std::make_pair(SDValue(), SDValue());
6617 if (Subtarget->is64Bit() &&
6618 DstTy == MVT::i64 &&
6619 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6620 return std::make_pair(SDValue(), SDValue());
6622 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
6624 MachineFunction &MF = DAG.getMachineFunction();
6625 unsigned MemSize = DstTy.getSizeInBits()/8;
6626 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6627 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6630 switch (DstTy.getSimpleVT().SimpleTy) {
6631 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
6632 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
6633 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
6634 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
6637 SDValue Chain = DAG.getEntryNode();
6638 SDValue Value = Op.getOperand(0);
6639 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
6640 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
6641 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
6642 MachinePointerInfo::getFixedStack(SSFI),
6644 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
6646 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
6648 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
6649 Chain = Value.getValue(1);
6650 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6651 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6654 // Build the FP_TO_INT*_IN_MEM
6655 SDValue Ops[] = { Chain, Value, StackSlot };
6656 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
6658 return std::make_pair(FIST, StackSlot);
6661 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
6662 SelectionDAG &DAG) const {
6663 if (Op.getValueType().isVector()) {
6664 if (Op.getValueType() == MVT::v2i32 &&
6665 Op.getOperand(0).getValueType() == MVT::v2f64) {
6671 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
6672 SDValue FIST = Vals.first, StackSlot = Vals.second;
6673 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
6674 if (FIST.getNode() == 0) return Op;
6677 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6678 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
6681 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
6682 SelectionDAG &DAG) const {
6683 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
6684 SDValue FIST = Vals.first, StackSlot = Vals.second;
6685 assert(FIST.getNode() && "Unexpected failure");
6688 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6689 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
6692 SDValue X86TargetLowering::LowerFABS(SDValue Op,
6693 SelectionDAG &DAG) const {
6694 LLVMContext *Context = DAG.getContext();
6695 DebugLoc dl = Op.getDebugLoc();
6696 EVT VT = Op.getValueType();
6699 EltVT = VT.getVectorElementType();
6700 std::vector<Constant*> CV;
6701 if (EltVT == MVT::f64) {
6702 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
6706 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
6712 Constant *C = ConstantVector::get(CV);
6713 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6714 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6715 MachinePointerInfo::getConstantPool(),
6717 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
6720 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
6721 LLVMContext *Context = DAG.getContext();
6722 DebugLoc dl = Op.getDebugLoc();
6723 EVT VT = Op.getValueType();
6726 EltVT = VT.getVectorElementType();
6727 std::vector<Constant*> CV;
6728 if (EltVT == MVT::f64) {
6729 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
6733 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
6739 Constant *C = ConstantVector::get(CV);
6740 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6741 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6742 MachinePointerInfo::getConstantPool(),
6744 if (VT.isVector()) {
6745 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6746 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6747 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6749 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
6751 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6755 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6756 LLVMContext *Context = DAG.getContext();
6757 SDValue Op0 = Op.getOperand(0);
6758 SDValue Op1 = Op.getOperand(1);
6759 DebugLoc dl = Op.getDebugLoc();
6760 EVT VT = Op.getValueType();
6761 EVT SrcVT = Op1.getValueType();
6763 // If second operand is smaller, extend it first.
6764 if (SrcVT.bitsLT(VT)) {
6765 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6768 // And if it is bigger, shrink it first.
6769 if (SrcVT.bitsGT(VT)) {
6770 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6774 // At this point the operands and the result should have the same
6775 // type, and that won't be f80 since that is not custom lowered.
6777 // First get the sign bit of second operand.
6778 std::vector<Constant*> CV;
6779 if (SrcVT == MVT::f64) {
6780 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6781 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6783 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6784 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6785 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6786 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6788 Constant *C = ConstantVector::get(CV);
6789 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6790 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6791 MachinePointerInfo::getConstantPool(),
6793 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6795 // Shift sign bit right or left if the two operands have different types.
6796 if (SrcVT.bitsGT(VT)) {
6797 // Op0 is MVT::f32, Op1 is MVT::f64.
6798 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6799 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6800 DAG.getConstant(32, MVT::i32));
6801 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
6802 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6803 DAG.getIntPtrConstant(0));
6806 // Clear first operand sign bit.
6808 if (VT == MVT::f64) {
6809 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6810 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6812 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6813 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6814 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6815 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6817 C = ConstantVector::get(CV);
6818 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6819 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6820 MachinePointerInfo::getConstantPool(),
6822 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6824 // Or the value with the sign bit.
6825 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6828 /// Emit nodes that will be selected as "test Op0,Op0", or something
6830 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6831 SelectionDAG &DAG) const {
6832 DebugLoc dl = Op.getDebugLoc();
6834 // CF and OF aren't always set the way we want. Determine which
6835 // of these we need.
6836 bool NeedCF = false;
6837 bool NeedOF = false;
6840 case X86::COND_A: case X86::COND_AE:
6841 case X86::COND_B: case X86::COND_BE:
6844 case X86::COND_G: case X86::COND_GE:
6845 case X86::COND_L: case X86::COND_LE:
6846 case X86::COND_O: case X86::COND_NO:
6851 // See if we can use the EFLAGS value from the operand instead of
6852 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6853 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6854 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6855 // Emit a CMP with 0, which is the TEST pattern.
6856 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6857 DAG.getConstant(0, Op.getValueType()));
6859 unsigned Opcode = 0;
6860 unsigned NumOperands = 0;
6861 switch (Op.getNode()->getOpcode()) {
6863 // Due to an isel shortcoming, be conservative if this add is likely to be
6864 // selected as part of a load-modify-store instruction. When the root node
6865 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6866 // uses of other nodes in the match, such as the ADD in this case. This
6867 // leads to the ADD being left around and reselected, with the result being
6868 // two adds in the output. Alas, even if none our users are stores, that
6869 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6870 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6871 // climbing the DAG back to the root, and it doesn't seem to be worth the
6873 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6874 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6875 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6878 if (ConstantSDNode *C =
6879 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6880 // An add of one will be selected as an INC.
6881 if (C->getAPIntValue() == 1) {
6882 Opcode = X86ISD::INC;
6887 // An add of negative one (subtract of one) will be selected as a DEC.
6888 if (C->getAPIntValue().isAllOnesValue()) {
6889 Opcode = X86ISD::DEC;
6895 // Otherwise use a regular EFLAGS-setting add.
6896 Opcode = X86ISD::ADD;
6900 // If the primary and result isn't used, don't bother using X86ISD::AND,
6901 // because a TEST instruction will be better.
6902 bool NonFlagUse = false;
6903 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6904 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6906 unsigned UOpNo = UI.getOperandNo();
6907 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6908 // Look pass truncate.
6909 UOpNo = User->use_begin().getOperandNo();
6910 User = *User->use_begin();
6913 if (User->getOpcode() != ISD::BRCOND &&
6914 User->getOpcode() != ISD::SETCC &&
6915 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6928 // Due to the ISEL shortcoming noted above, be conservative if this op is
6929 // likely to be selected as part of a load-modify-store instruction.
6930 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6931 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6932 if (UI->getOpcode() == ISD::STORE)
6935 // Otherwise use a regular EFLAGS-setting instruction.
6936 switch (Op.getNode()->getOpcode()) {
6937 default: llvm_unreachable("unexpected operator!");
6938 case ISD::SUB: Opcode = X86ISD::SUB; break;
6939 case ISD::OR: Opcode = X86ISD::OR; break;
6940 case ISD::XOR: Opcode = X86ISD::XOR; break;
6941 case ISD::AND: Opcode = X86ISD::AND; break;
6953 return SDValue(Op.getNode(), 1);
6960 // Emit a CMP with 0, which is the TEST pattern.
6961 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6962 DAG.getConstant(0, Op.getValueType()));
6964 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6965 SmallVector<SDValue, 4> Ops;
6966 for (unsigned i = 0; i != NumOperands; ++i)
6967 Ops.push_back(Op.getOperand(i));
6969 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6970 DAG.ReplaceAllUsesWith(Op, New);
6971 return SDValue(New.getNode(), 1);
6974 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6976 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6977 SelectionDAG &DAG) const {
6978 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6979 if (C->getAPIntValue() == 0)
6980 return EmitTest(Op0, X86CC, DAG);
6982 DebugLoc dl = Op0.getDebugLoc();
6983 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6986 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6987 /// if it's possible.
6988 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6989 DebugLoc dl, SelectionDAG &DAG) const {
6990 SDValue Op0 = And.getOperand(0);
6991 SDValue Op1 = And.getOperand(1);
6992 if (Op0.getOpcode() == ISD::TRUNCATE)
6993 Op0 = Op0.getOperand(0);
6994 if (Op1.getOpcode() == ISD::TRUNCATE)
6995 Op1 = Op1.getOperand(0);
6998 if (Op1.getOpcode() == ISD::SHL)
6999 std::swap(Op0, Op1);
7000 if (Op0.getOpcode() == ISD::SHL) {
7001 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
7002 if (And00C->getZExtValue() == 1) {
7003 // If we looked past a truncate, check that it's only truncating away
7005 unsigned BitWidth = Op0.getValueSizeInBits();
7006 unsigned AndBitWidth = And.getValueSizeInBits();
7007 if (BitWidth > AndBitWidth) {
7008 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
7009 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
7010 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
7014 RHS = Op0.getOperand(1);
7016 } else if (Op1.getOpcode() == ISD::Constant) {
7017 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
7018 SDValue AndLHS = Op0;
7019 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
7020 LHS = AndLHS.getOperand(0);
7021 RHS = AndLHS.getOperand(1);
7025 if (LHS.getNode()) {
7026 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
7027 // instruction. Since the shift amount is in-range-or-undefined, we know
7028 // that doing a bittest on the i32 value is ok. We extend to i32 because
7029 // the encoding for the i16 version is larger than the i32 version.
7030 // Also promote i16 to i32 for performance / code size reason.
7031 if (LHS.getValueType() == MVT::i8 ||
7032 LHS.getValueType() == MVT::i16)
7033 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
7035 // If the operand types disagree, extend the shift amount to match. Since
7036 // BT ignores high bits (like shifts) we can use anyextend.
7037 if (LHS.getValueType() != RHS.getValueType())
7038 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
7040 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
7041 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
7042 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7043 DAG.getConstant(Cond, MVT::i8), BT);
7049 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
7050 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
7051 SDValue Op0 = Op.getOperand(0);
7052 SDValue Op1 = Op.getOperand(1);
7053 DebugLoc dl = Op.getDebugLoc();
7054 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7056 // Optimize to BT if possible.
7057 // Lower (X & (1 << N)) == 0 to BT(X, N).
7058 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
7059 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
7060 if (Op0.getOpcode() == ISD::AND &&
7062 Op1.getOpcode() == ISD::Constant &&
7063 cast<ConstantSDNode>(Op1)->isNullValue() &&
7064 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7065 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
7066 if (NewSetCC.getNode())
7070 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
7071 if (Op0.getOpcode() == X86ISD::SETCC &&
7072 Op1.getOpcode() == ISD::Constant &&
7073 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
7074 cast<ConstantSDNode>(Op1)->isNullValue()) &&
7075 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7076 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
7077 bool Invert = (CC == ISD::SETNE) ^
7078 cast<ConstantSDNode>(Op1)->isNullValue();
7080 CCode = X86::GetOppositeBranchCondition(CCode);
7081 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7082 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
7085 bool isFP = Op1.getValueType().isFloatingPoint();
7086 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
7087 if (X86CC == X86::COND_INVALID)
7090 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
7092 // Use sbb x, x to materialize carry bit into a GPR.
7093 if (X86CC == X86::COND_B)
7094 return DAG.getNode(ISD::AND, dl, MVT::i8,
7095 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
7096 DAG.getConstant(X86CC, MVT::i8), Cond),
7097 DAG.getConstant(1, MVT::i8));
7099 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7100 DAG.getConstant(X86CC, MVT::i8), Cond);
7103 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
7105 SDValue Op0 = Op.getOperand(0);
7106 SDValue Op1 = Op.getOperand(1);
7107 SDValue CC = Op.getOperand(2);
7108 EVT VT = Op.getValueType();
7109 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
7110 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
7111 DebugLoc dl = Op.getDebugLoc();
7115 EVT VT0 = Op0.getValueType();
7116 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
7117 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
7120 switch (SetCCOpcode) {
7123 case ISD::SETEQ: SSECC = 0; break;
7125 case ISD::SETGT: Swap = true; // Fallthrough
7127 case ISD::SETOLT: SSECC = 1; break;
7129 case ISD::SETGE: Swap = true; // Fallthrough
7131 case ISD::SETOLE: SSECC = 2; break;
7132 case ISD::SETUO: SSECC = 3; break;
7134 case ISD::SETNE: SSECC = 4; break;
7135 case ISD::SETULE: Swap = true;
7136 case ISD::SETUGE: SSECC = 5; break;
7137 case ISD::SETULT: Swap = true;
7138 case ISD::SETUGT: SSECC = 6; break;
7139 case ISD::SETO: SSECC = 7; break;
7142 std::swap(Op0, Op1);
7144 // In the two special cases we can't handle, emit two comparisons.
7146 if (SetCCOpcode == ISD::SETUEQ) {
7148 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
7149 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
7150 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
7152 else if (SetCCOpcode == ISD::SETONE) {
7154 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7155 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7156 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7158 llvm_unreachable("Illegal FP comparison");
7160 // Handle all other FP comparisons here.
7161 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7164 // We are handling one of the integer comparisons here. Since SSE only has
7165 // GT and EQ comparisons for integer, swapping operands and multiple
7166 // operations may be required for some comparisons.
7167 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7168 bool Swap = false, Invert = false, FlipSigns = false;
7170 switch (VT.getSimpleVT().SimpleTy) {
7173 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7175 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7177 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7178 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7181 switch (SetCCOpcode) {
7183 case ISD::SETNE: Invert = true;
7184 case ISD::SETEQ: Opc = EQOpc; break;
7185 case ISD::SETLT: Swap = true;
7186 case ISD::SETGT: Opc = GTOpc; break;
7187 case ISD::SETGE: Swap = true;
7188 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7189 case ISD::SETULT: Swap = true;
7190 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7191 case ISD::SETUGE: Swap = true;
7192 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7195 std::swap(Op0, Op1);
7197 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7198 // bits of the inputs before performing those operations.
7200 EVT EltVT = VT.getVectorElementType();
7201 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7203 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7204 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7206 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7207 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7210 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7212 // If the logical-not of the result is required, perform that now.
7214 Result = DAG.getNOT(dl, Result, VT);
7219 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
7220 static bool isX86LogicalCmp(SDValue Op) {
7221 unsigned Opc = Op.getNode()->getOpcode();
7222 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
7224 if (Op.getResNo() == 1 &&
7225 (Opc == X86ISD::ADD ||
7226 Opc == X86ISD::SUB ||
7227 Opc == X86ISD::SMUL ||
7228 Opc == X86ISD::UMUL ||
7229 Opc == X86ISD::INC ||
7230 Opc == X86ISD::DEC ||
7231 Opc == X86ISD::OR ||
7232 Opc == X86ISD::XOR ||
7233 Opc == X86ISD::AND))
7239 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
7240 bool addTest = true;
7241 SDValue Cond = Op.getOperand(0);
7242 DebugLoc dl = Op.getDebugLoc();
7245 if (Cond.getOpcode() == ISD::SETCC) {
7246 SDValue NewCond = LowerSETCC(Cond, DAG);
7247 if (NewCond.getNode())
7251 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
7252 SDValue Op1 = Op.getOperand(1);
7253 SDValue Op2 = Op.getOperand(2);
7254 if (Cond.getOpcode() == X86ISD::SETCC &&
7255 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
7256 SDValue Cmp = Cond.getOperand(1);
7257 if (Cmp.getOpcode() == X86ISD::CMP) {
7258 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
7259 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
7260 ConstantSDNode *RHSC =
7261 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
7262 if (N1C && N1C->isAllOnesValue() &&
7263 N2C && N2C->isNullValue() &&
7264 RHSC && RHSC->isNullValue()) {
7265 SDValue CmpOp0 = Cmp.getOperand(0);
7266 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
7267 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7268 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
7269 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7274 // Look pass (and (setcc_carry (cmp ...)), 1).
7275 if (Cond.getOpcode() == ISD::AND &&
7276 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7277 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7278 if (C && C->getAPIntValue() == 1)
7279 Cond = Cond.getOperand(0);
7282 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7283 // setting operand in place of the X86ISD::SETCC.
7284 if (Cond.getOpcode() == X86ISD::SETCC ||
7285 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7286 CC = Cond.getOperand(0);
7288 SDValue Cmp = Cond.getOperand(1);
7289 unsigned Opc = Cmp.getOpcode();
7290 EVT VT = Op.getValueType();
7292 bool IllegalFPCMov = false;
7293 if (VT.isFloatingPoint() && !VT.isVector() &&
7294 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7295 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7297 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7298 Opc == X86ISD::BT) { // FIXME
7305 // Look pass the truncate.
7306 if (Cond.getOpcode() == ISD::TRUNCATE)
7307 Cond = Cond.getOperand(0);
7309 // We know the result of AND is compared against zero. Try to match
7311 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7312 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7313 if (NewSetCC.getNode()) {
7314 CC = NewSetCC.getOperand(0);
7315 Cond = NewSetCC.getOperand(1);
7322 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7323 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7326 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7327 // condition is true.
7328 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
7329 SDValue Ops[] = { Op2, Op1, CC, Cond };
7330 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
7333 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7334 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7335 // from the AND / OR.
7336 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7337 Opc = Op.getOpcode();
7338 if (Opc != ISD::OR && Opc != ISD::AND)
7340 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7341 Op.getOperand(0).hasOneUse() &&
7342 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7343 Op.getOperand(1).hasOneUse());
7346 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7347 // 1 and that the SETCC node has a single use.
7348 static bool isXor1OfSetCC(SDValue Op) {
7349 if (Op.getOpcode() != ISD::XOR)
7351 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7352 if (N1C && N1C->getAPIntValue() == 1) {
7353 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7354 Op.getOperand(0).hasOneUse();
7359 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7360 bool addTest = true;
7361 SDValue Chain = Op.getOperand(0);
7362 SDValue Cond = Op.getOperand(1);
7363 SDValue Dest = Op.getOperand(2);
7364 DebugLoc dl = Op.getDebugLoc();
7367 if (Cond.getOpcode() == ISD::SETCC) {
7368 SDValue NewCond = LowerSETCC(Cond, DAG);
7369 if (NewCond.getNode())
7373 // FIXME: LowerXALUO doesn't handle these!!
7374 else if (Cond.getOpcode() == X86ISD::ADD ||
7375 Cond.getOpcode() == X86ISD::SUB ||
7376 Cond.getOpcode() == X86ISD::SMUL ||
7377 Cond.getOpcode() == X86ISD::UMUL)
7378 Cond = LowerXALUO(Cond, DAG);
7381 // Look pass (and (setcc_carry (cmp ...)), 1).
7382 if (Cond.getOpcode() == ISD::AND &&
7383 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7384 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7385 if (C && C->getAPIntValue() == 1)
7386 Cond = Cond.getOperand(0);
7389 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7390 // setting operand in place of the X86ISD::SETCC.
7391 if (Cond.getOpcode() == X86ISD::SETCC ||
7392 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7393 CC = Cond.getOperand(0);
7395 SDValue Cmp = Cond.getOperand(1);
7396 unsigned Opc = Cmp.getOpcode();
7397 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7398 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7402 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7406 // These can only come from an arithmetic instruction with overflow,
7407 // e.g. SADDO, UADDO.
7408 Cond = Cond.getNode()->getOperand(1);
7415 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7416 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7417 if (CondOpc == ISD::OR) {
7418 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7419 // two branches instead of an explicit OR instruction with a
7421 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7422 isX86LogicalCmp(Cmp)) {
7423 CC = Cond.getOperand(0).getOperand(0);
7424 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7425 Chain, Dest, CC, Cmp);
7426 CC = Cond.getOperand(1).getOperand(0);
7430 } else { // ISD::AND
7431 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7432 // two branches instead of an explicit AND instruction with a
7433 // separate test. However, we only do this if this block doesn't
7434 // have a fall-through edge, because this requires an explicit
7435 // jmp when the condition is false.
7436 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7437 isX86LogicalCmp(Cmp) &&
7438 Op.getNode()->hasOneUse()) {
7439 X86::CondCode CCode =
7440 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7441 CCode = X86::GetOppositeBranchCondition(CCode);
7442 CC = DAG.getConstant(CCode, MVT::i8);
7443 SDNode *User = *Op.getNode()->use_begin();
7444 // Look for an unconditional branch following this conditional branch.
7445 // We need this because we need to reverse the successors in order
7446 // to implement FCMP_OEQ.
7447 if (User->getOpcode() == ISD::BR) {
7448 SDValue FalseBB = User->getOperand(1);
7450 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
7451 assert(NewBR == User);
7455 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7456 Chain, Dest, CC, Cmp);
7457 X86::CondCode CCode =
7458 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
7459 CCode = X86::GetOppositeBranchCondition(CCode);
7460 CC = DAG.getConstant(CCode, MVT::i8);
7466 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
7467 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
7468 // It should be transformed during dag combiner except when the condition
7469 // is set by a arithmetics with overflow node.
7470 X86::CondCode CCode =
7471 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7472 CCode = X86::GetOppositeBranchCondition(CCode);
7473 CC = DAG.getConstant(CCode, MVT::i8);
7474 Cond = Cond.getOperand(0).getOperand(1);
7480 // Look pass the truncate.
7481 if (Cond.getOpcode() == ISD::TRUNCATE)
7482 Cond = Cond.getOperand(0);
7484 // We know the result of AND is compared against zero. Try to match
7486 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7487 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7488 if (NewSetCC.getNode()) {
7489 CC = NewSetCC.getOperand(0);
7490 Cond = NewSetCC.getOperand(1);
7497 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7498 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7500 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7501 Chain, Dest, CC, Cond);
7505 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
7506 // Calls to _alloca is needed to probe the stack when allocating more than 4k
7507 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
7508 // that the guard pages used by the OS virtual memory manager are allocated in
7509 // correct sequence.
7511 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
7512 SelectionDAG &DAG) const {
7513 assert(Subtarget->isTargetCygMing() &&
7514 "This should be used only on Cygwin/Mingw targets");
7515 DebugLoc dl = Op.getDebugLoc();
7518 SDValue Chain = Op.getOperand(0);
7519 SDValue Size = Op.getOperand(1);
7520 // FIXME: Ensure alignment here
7524 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
7526 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
7527 Flag = Chain.getValue(1);
7529 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
7531 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
7532 Flag = Chain.getValue(1);
7534 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
7536 SDValue Ops1[2] = { Chain.getValue(0), Chain };
7537 return DAG.getMergeValues(Ops1, 2, dl);
7540 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
7541 MachineFunction &MF = DAG.getMachineFunction();
7542 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
7544 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7545 DebugLoc DL = Op.getDebugLoc();
7547 if (!Subtarget->is64Bit()) {
7548 // vastart just stores the address of the VarArgsFrameIndex slot into the
7549 // memory location argument.
7550 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7552 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
7553 MachinePointerInfo(SV), false, false, 0);
7557 // gp_offset (0 - 6 * 8)
7558 // fp_offset (48 - 48 + 8 * 16)
7559 // overflow_arg_area (point to parameters coming in memory).
7561 SmallVector<SDValue, 8> MemOps;
7562 SDValue FIN = Op.getOperand(1);
7564 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
7565 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
7567 FIN, MachinePointerInfo(SV), false, false, 0);
7568 MemOps.push_back(Store);
7571 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7572 FIN, DAG.getIntPtrConstant(4));
7573 Store = DAG.getStore(Op.getOperand(0), DL,
7574 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
7576 FIN, MachinePointerInfo(SV, 4), false, false, 0);
7577 MemOps.push_back(Store);
7579 // Store ptr to overflow_arg_area
7580 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7581 FIN, DAG.getIntPtrConstant(4));
7582 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7584 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
7585 MachinePointerInfo(SV, 8),
7587 MemOps.push_back(Store);
7589 // Store ptr to reg_save_area.
7590 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7591 FIN, DAG.getIntPtrConstant(8));
7592 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
7594 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
7595 MachinePointerInfo(SV, 16), false, false, 0);
7596 MemOps.push_back(Store);
7597 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
7598 &MemOps[0], MemOps.size());
7601 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
7602 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7603 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
7605 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
7609 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
7610 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7611 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
7612 SDValue Chain = Op.getOperand(0);
7613 SDValue DstPtr = Op.getOperand(1);
7614 SDValue SrcPtr = Op.getOperand(2);
7615 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
7616 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7617 DebugLoc DL = Op.getDebugLoc();
7619 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
7620 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
7622 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
7626 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
7627 DebugLoc dl = Op.getDebugLoc();
7628 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7630 default: return SDValue(); // Don't custom lower most intrinsics.
7631 // Comparison intrinsics.
7632 case Intrinsic::x86_sse_comieq_ss:
7633 case Intrinsic::x86_sse_comilt_ss:
7634 case Intrinsic::x86_sse_comile_ss:
7635 case Intrinsic::x86_sse_comigt_ss:
7636 case Intrinsic::x86_sse_comige_ss:
7637 case Intrinsic::x86_sse_comineq_ss:
7638 case Intrinsic::x86_sse_ucomieq_ss:
7639 case Intrinsic::x86_sse_ucomilt_ss:
7640 case Intrinsic::x86_sse_ucomile_ss:
7641 case Intrinsic::x86_sse_ucomigt_ss:
7642 case Intrinsic::x86_sse_ucomige_ss:
7643 case Intrinsic::x86_sse_ucomineq_ss:
7644 case Intrinsic::x86_sse2_comieq_sd:
7645 case Intrinsic::x86_sse2_comilt_sd:
7646 case Intrinsic::x86_sse2_comile_sd:
7647 case Intrinsic::x86_sse2_comigt_sd:
7648 case Intrinsic::x86_sse2_comige_sd:
7649 case Intrinsic::x86_sse2_comineq_sd:
7650 case Intrinsic::x86_sse2_ucomieq_sd:
7651 case Intrinsic::x86_sse2_ucomilt_sd:
7652 case Intrinsic::x86_sse2_ucomile_sd:
7653 case Intrinsic::x86_sse2_ucomigt_sd:
7654 case Intrinsic::x86_sse2_ucomige_sd:
7655 case Intrinsic::x86_sse2_ucomineq_sd: {
7657 ISD::CondCode CC = ISD::SETCC_INVALID;
7660 case Intrinsic::x86_sse_comieq_ss:
7661 case Intrinsic::x86_sse2_comieq_sd:
7665 case Intrinsic::x86_sse_comilt_ss:
7666 case Intrinsic::x86_sse2_comilt_sd:
7670 case Intrinsic::x86_sse_comile_ss:
7671 case Intrinsic::x86_sse2_comile_sd:
7675 case Intrinsic::x86_sse_comigt_ss:
7676 case Intrinsic::x86_sse2_comigt_sd:
7680 case Intrinsic::x86_sse_comige_ss:
7681 case Intrinsic::x86_sse2_comige_sd:
7685 case Intrinsic::x86_sse_comineq_ss:
7686 case Intrinsic::x86_sse2_comineq_sd:
7690 case Intrinsic::x86_sse_ucomieq_ss:
7691 case Intrinsic::x86_sse2_ucomieq_sd:
7692 Opc = X86ISD::UCOMI;
7695 case Intrinsic::x86_sse_ucomilt_ss:
7696 case Intrinsic::x86_sse2_ucomilt_sd:
7697 Opc = X86ISD::UCOMI;
7700 case Intrinsic::x86_sse_ucomile_ss:
7701 case Intrinsic::x86_sse2_ucomile_sd:
7702 Opc = X86ISD::UCOMI;
7705 case Intrinsic::x86_sse_ucomigt_ss:
7706 case Intrinsic::x86_sse2_ucomigt_sd:
7707 Opc = X86ISD::UCOMI;
7710 case Intrinsic::x86_sse_ucomige_ss:
7711 case Intrinsic::x86_sse2_ucomige_sd:
7712 Opc = X86ISD::UCOMI;
7715 case Intrinsic::x86_sse_ucomineq_ss:
7716 case Intrinsic::x86_sse2_ucomineq_sd:
7717 Opc = X86ISD::UCOMI;
7722 SDValue LHS = Op.getOperand(1);
7723 SDValue RHS = Op.getOperand(2);
7724 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
7725 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
7726 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
7727 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7728 DAG.getConstant(X86CC, MVT::i8), Cond);
7729 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7731 // ptest and testp intrinsics. The intrinsic these come from are designed to
7732 // return an integer value, not just an instruction so lower it to the ptest
7733 // or testp pattern and a setcc for the result.
7734 case Intrinsic::x86_sse41_ptestz:
7735 case Intrinsic::x86_sse41_ptestc:
7736 case Intrinsic::x86_sse41_ptestnzc:
7737 case Intrinsic::x86_avx_ptestz_256:
7738 case Intrinsic::x86_avx_ptestc_256:
7739 case Intrinsic::x86_avx_ptestnzc_256:
7740 case Intrinsic::x86_avx_vtestz_ps:
7741 case Intrinsic::x86_avx_vtestc_ps:
7742 case Intrinsic::x86_avx_vtestnzc_ps:
7743 case Intrinsic::x86_avx_vtestz_pd:
7744 case Intrinsic::x86_avx_vtestc_pd:
7745 case Intrinsic::x86_avx_vtestnzc_pd:
7746 case Intrinsic::x86_avx_vtestz_ps_256:
7747 case Intrinsic::x86_avx_vtestc_ps_256:
7748 case Intrinsic::x86_avx_vtestnzc_ps_256:
7749 case Intrinsic::x86_avx_vtestz_pd_256:
7750 case Intrinsic::x86_avx_vtestc_pd_256:
7751 case Intrinsic::x86_avx_vtestnzc_pd_256: {
7752 bool IsTestPacked = false;
7755 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
7756 case Intrinsic::x86_avx_vtestz_ps:
7757 case Intrinsic::x86_avx_vtestz_pd:
7758 case Intrinsic::x86_avx_vtestz_ps_256:
7759 case Intrinsic::x86_avx_vtestz_pd_256:
7760 IsTestPacked = true; // Fallthrough
7761 case Intrinsic::x86_sse41_ptestz:
7762 case Intrinsic::x86_avx_ptestz_256:
7764 X86CC = X86::COND_E;
7766 case Intrinsic::x86_avx_vtestc_ps:
7767 case Intrinsic::x86_avx_vtestc_pd:
7768 case Intrinsic::x86_avx_vtestc_ps_256:
7769 case Intrinsic::x86_avx_vtestc_pd_256:
7770 IsTestPacked = true; // Fallthrough
7771 case Intrinsic::x86_sse41_ptestc:
7772 case Intrinsic::x86_avx_ptestc_256:
7774 X86CC = X86::COND_B;
7776 case Intrinsic::x86_avx_vtestnzc_ps:
7777 case Intrinsic::x86_avx_vtestnzc_pd:
7778 case Intrinsic::x86_avx_vtestnzc_ps_256:
7779 case Intrinsic::x86_avx_vtestnzc_pd_256:
7780 IsTestPacked = true; // Fallthrough
7781 case Intrinsic::x86_sse41_ptestnzc:
7782 case Intrinsic::x86_avx_ptestnzc_256:
7784 X86CC = X86::COND_A;
7788 SDValue LHS = Op.getOperand(1);
7789 SDValue RHS = Op.getOperand(2);
7790 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
7791 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
7792 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7793 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7794 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7797 // Fix vector shift instructions where the last operand is a non-immediate
7799 case Intrinsic::x86_sse2_pslli_w:
7800 case Intrinsic::x86_sse2_pslli_d:
7801 case Intrinsic::x86_sse2_pslli_q:
7802 case Intrinsic::x86_sse2_psrli_w:
7803 case Intrinsic::x86_sse2_psrli_d:
7804 case Intrinsic::x86_sse2_psrli_q:
7805 case Intrinsic::x86_sse2_psrai_w:
7806 case Intrinsic::x86_sse2_psrai_d:
7807 case Intrinsic::x86_mmx_pslli_w:
7808 case Intrinsic::x86_mmx_pslli_d:
7809 case Intrinsic::x86_mmx_pslli_q:
7810 case Intrinsic::x86_mmx_psrli_w:
7811 case Intrinsic::x86_mmx_psrli_d:
7812 case Intrinsic::x86_mmx_psrli_q:
7813 case Intrinsic::x86_mmx_psrai_w:
7814 case Intrinsic::x86_mmx_psrai_d: {
7815 SDValue ShAmt = Op.getOperand(2);
7816 if (isa<ConstantSDNode>(ShAmt))
7819 unsigned NewIntNo = 0;
7820 EVT ShAmtVT = MVT::v4i32;
7822 case Intrinsic::x86_sse2_pslli_w:
7823 NewIntNo = Intrinsic::x86_sse2_psll_w;
7825 case Intrinsic::x86_sse2_pslli_d:
7826 NewIntNo = Intrinsic::x86_sse2_psll_d;
7828 case Intrinsic::x86_sse2_pslli_q:
7829 NewIntNo = Intrinsic::x86_sse2_psll_q;
7831 case Intrinsic::x86_sse2_psrli_w:
7832 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7834 case Intrinsic::x86_sse2_psrli_d:
7835 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7837 case Intrinsic::x86_sse2_psrli_q:
7838 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7840 case Intrinsic::x86_sse2_psrai_w:
7841 NewIntNo = Intrinsic::x86_sse2_psra_w;
7843 case Intrinsic::x86_sse2_psrai_d:
7844 NewIntNo = Intrinsic::x86_sse2_psra_d;
7847 ShAmtVT = MVT::v2i32;
7849 case Intrinsic::x86_mmx_pslli_w:
7850 NewIntNo = Intrinsic::x86_mmx_psll_w;
7852 case Intrinsic::x86_mmx_pslli_d:
7853 NewIntNo = Intrinsic::x86_mmx_psll_d;
7855 case Intrinsic::x86_mmx_pslli_q:
7856 NewIntNo = Intrinsic::x86_mmx_psll_q;
7858 case Intrinsic::x86_mmx_psrli_w:
7859 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7861 case Intrinsic::x86_mmx_psrli_d:
7862 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7864 case Intrinsic::x86_mmx_psrli_q:
7865 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7867 case Intrinsic::x86_mmx_psrai_w:
7868 NewIntNo = Intrinsic::x86_mmx_psra_w;
7870 case Intrinsic::x86_mmx_psrai_d:
7871 NewIntNo = Intrinsic::x86_mmx_psra_d;
7873 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7879 // The vector shift intrinsics with scalars uses 32b shift amounts but
7880 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7884 ShOps[1] = DAG.getConstant(0, MVT::i32);
7885 if (ShAmtVT == MVT::v4i32) {
7886 ShOps[2] = DAG.getUNDEF(MVT::i32);
7887 ShOps[3] = DAG.getUNDEF(MVT::i32);
7888 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7890 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7893 EVT VT = Op.getValueType();
7894 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7895 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7896 DAG.getConstant(NewIntNo, MVT::i32),
7897 Op.getOperand(1), ShAmt);
7902 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7903 SelectionDAG &DAG) const {
7904 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7905 MFI->setReturnAddressIsTaken(true);
7907 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7908 DebugLoc dl = Op.getDebugLoc();
7911 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7913 DAG.getConstant(TD->getPointerSize(),
7914 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7915 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7916 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7918 MachinePointerInfo(), false, false, 0);
7921 // Just load the return address.
7922 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7923 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7924 RetAddrFI, MachinePointerInfo(), false, false, 0);
7927 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7928 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7929 MFI->setFrameAddressIsTaken(true);
7931 EVT VT = Op.getValueType();
7932 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7933 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7934 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7935 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7937 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
7938 MachinePointerInfo(),
7943 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7944 SelectionDAG &DAG) const {
7945 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7948 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7949 MachineFunction &MF = DAG.getMachineFunction();
7950 SDValue Chain = Op.getOperand(0);
7951 SDValue Offset = Op.getOperand(1);
7952 SDValue Handler = Op.getOperand(2);
7953 DebugLoc dl = Op.getDebugLoc();
7955 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
7956 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7958 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7960 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
7961 DAG.getIntPtrConstant(TD->getPointerSize()));
7962 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7963 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
7965 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7966 MF.getRegInfo().addLiveOut(StoreAddrReg);
7968 return DAG.getNode(X86ISD::EH_RETURN, dl,
7970 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7973 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7974 SelectionDAG &DAG) const {
7975 SDValue Root = Op.getOperand(0);
7976 SDValue Trmp = Op.getOperand(1); // trampoline
7977 SDValue FPtr = Op.getOperand(2); // nested function
7978 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7979 DebugLoc dl = Op.getDebugLoc();
7981 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7983 if (Subtarget->is64Bit()) {
7984 SDValue OutChains[6];
7986 // Large code-model.
7987 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7988 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7990 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7991 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7993 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7995 // Load the pointer to the nested function into R11.
7996 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7997 SDValue Addr = Trmp;
7998 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7999 Addr, MachinePointerInfo(TrmpAddr),
8002 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8003 DAG.getConstant(2, MVT::i64));
8004 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
8005 MachinePointerInfo(TrmpAddr, 2),
8008 // Load the 'nest' parameter value into R10.
8009 // R10 is specified in X86CallingConv.td
8010 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
8011 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8012 DAG.getConstant(10, MVT::i64));
8013 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8014 Addr, MachinePointerInfo(TrmpAddr, 10),
8017 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8018 DAG.getConstant(12, MVT::i64));
8019 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
8020 MachinePointerInfo(TrmpAddr, 12),
8023 // Jump to the nested function.
8024 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
8025 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8026 DAG.getConstant(20, MVT::i64));
8027 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8028 Addr, MachinePointerInfo(TrmpAddr, 20),
8031 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
8032 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8033 DAG.getConstant(22, MVT::i64));
8034 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
8035 MachinePointerInfo(TrmpAddr, 22),
8039 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
8040 return DAG.getMergeValues(Ops, 2, dl);
8042 const Function *Func =
8043 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
8044 CallingConv::ID CC = Func->getCallingConv();
8049 llvm_unreachable("Unsupported calling convention");
8050 case CallingConv::C:
8051 case CallingConv::X86_StdCall: {
8052 // Pass 'nest' parameter in ECX.
8053 // Must be kept in sync with X86CallingConv.td
8056 // Check that ECX wasn't needed by an 'inreg' parameter.
8057 const FunctionType *FTy = Func->getFunctionType();
8058 const AttrListPtr &Attrs = Func->getAttributes();
8060 if (!Attrs.isEmpty() && !Func->isVarArg()) {
8061 unsigned InRegCount = 0;
8064 for (FunctionType::param_iterator I = FTy->param_begin(),
8065 E = FTy->param_end(); I != E; ++I, ++Idx)
8066 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
8067 // FIXME: should only count parameters that are lowered to integers.
8068 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
8070 if (InRegCount > 2) {
8071 report_fatal_error("Nest register in use - reduce number of inreg"
8077 case CallingConv::X86_FastCall:
8078 case CallingConv::X86_ThisCall:
8079 case CallingConv::Fast:
8080 // Pass 'nest' parameter in EAX.
8081 // Must be kept in sync with X86CallingConv.td
8086 SDValue OutChains[4];
8089 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8090 DAG.getConstant(10, MVT::i32));
8091 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
8093 // This is storing the opcode for MOV32ri.
8094 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
8095 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
8096 OutChains[0] = DAG.getStore(Root, dl,
8097 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
8098 Trmp, MachinePointerInfo(TrmpAddr),
8101 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8102 DAG.getConstant(1, MVT::i32));
8103 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
8104 MachinePointerInfo(TrmpAddr, 1),
8107 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
8108 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8109 DAG.getConstant(5, MVT::i32));
8110 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
8111 MachinePointerInfo(TrmpAddr, 5),
8114 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8115 DAG.getConstant(6, MVT::i32));
8116 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
8117 MachinePointerInfo(TrmpAddr, 6),
8121 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
8122 return DAG.getMergeValues(Ops, 2, dl);
8126 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
8127 SelectionDAG &DAG) const {
8129 The rounding mode is in bits 11:10 of FPSR, and has the following
8136 FLT_ROUNDS, on the other hand, expects the following:
8143 To perform the conversion, we do:
8144 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
8147 MachineFunction &MF = DAG.getMachineFunction();
8148 const TargetMachine &TM = MF.getTarget();
8149 const TargetFrameInfo &TFI = *TM.getFrameInfo();
8150 unsigned StackAlignment = TFI.getStackAlignment();
8151 EVT VT = Op.getValueType();
8152 DebugLoc dl = Op.getDebugLoc();
8154 // Save FP Control Word to stack slot
8155 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
8156 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8158 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
8159 DAG.getEntryNode(), StackSlot);
8161 // Load FP Control Word from stack slot
8162 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot,
8163 MachinePointerInfo(), false, false, 0);
8165 // Transform as necessary
8167 DAG.getNode(ISD::SRL, dl, MVT::i16,
8168 DAG.getNode(ISD::AND, dl, MVT::i16,
8169 CWD, DAG.getConstant(0x800, MVT::i16)),
8170 DAG.getConstant(11, MVT::i8));
8172 DAG.getNode(ISD::SRL, dl, MVT::i16,
8173 DAG.getNode(ISD::AND, dl, MVT::i16,
8174 CWD, DAG.getConstant(0x400, MVT::i16)),
8175 DAG.getConstant(9, MVT::i8));
8178 DAG.getNode(ISD::AND, dl, MVT::i16,
8179 DAG.getNode(ISD::ADD, dl, MVT::i16,
8180 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
8181 DAG.getConstant(1, MVT::i16)),
8182 DAG.getConstant(3, MVT::i16));
8185 return DAG.getNode((VT.getSizeInBits() < 16 ?
8186 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
8189 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
8190 EVT VT = Op.getValueType();
8192 unsigned NumBits = VT.getSizeInBits();
8193 DebugLoc dl = Op.getDebugLoc();
8195 Op = Op.getOperand(0);
8196 if (VT == MVT::i8) {
8197 // Zero extend to i32 since there is not an i8 bsr.
8199 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8202 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
8203 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8204 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
8206 // If src is zero (i.e. bsr sets ZF), returns NumBits.
8209 DAG.getConstant(NumBits+NumBits-1, OpVT),
8210 DAG.getConstant(X86::COND_E, MVT::i8),
8213 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8215 // Finally xor with NumBits-1.
8216 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
8219 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8223 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
8224 EVT VT = Op.getValueType();
8226 unsigned NumBits = VT.getSizeInBits();
8227 DebugLoc dl = Op.getDebugLoc();
8229 Op = Op.getOperand(0);
8230 if (VT == MVT::i8) {
8232 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8235 // Issue a bsf (scan bits forward) which also sets EFLAGS.
8236 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8237 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
8239 // If src is zero (i.e. bsf sets ZF), returns NumBits.
8242 DAG.getConstant(NumBits, OpVT),
8243 DAG.getConstant(X86::COND_E, MVT::i8),
8246 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8249 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8253 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
8254 EVT VT = Op.getValueType();
8255 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
8256 DebugLoc dl = Op.getDebugLoc();
8258 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
8259 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
8260 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
8261 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
8262 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
8264 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
8265 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
8266 // return AloBlo + AloBhi + AhiBlo;
8268 SDValue A = Op.getOperand(0);
8269 SDValue B = Op.getOperand(1);
8271 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8272 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8273 A, DAG.getConstant(32, MVT::i32));
8274 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8275 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8276 B, DAG.getConstant(32, MVT::i32));
8277 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8278 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8280 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8281 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8283 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8284 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8286 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8287 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8288 AloBhi, DAG.getConstant(32, MVT::i32));
8289 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8290 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8291 AhiBlo, DAG.getConstant(32, MVT::i32));
8292 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8293 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8297 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
8298 EVT VT = Op.getValueType();
8299 DebugLoc dl = Op.getDebugLoc();
8300 SDValue R = Op.getOperand(0);
8302 LLVMContext *Context = DAG.getContext();
8304 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
8306 if (VT == MVT::v4i32) {
8307 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8308 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8309 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
8311 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
8313 std::vector<Constant*> CV(4, CI);
8314 Constant *C = ConstantVector::get(CV);
8315 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8316 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8317 MachinePointerInfo::getConstantPool(),
8320 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
8321 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, Op);
8322 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
8323 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
8325 if (VT == MVT::v16i8) {
8327 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8328 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8329 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
8331 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
8332 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
8334 std::vector<Constant*> CVM1(16, CM1);
8335 std::vector<Constant*> CVM2(16, CM2);
8336 Constant *C = ConstantVector::get(CVM1);
8337 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8338 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8339 MachinePointerInfo::getConstantPool(),
8342 // r = pblendv(r, psllw(r & (char16)15, 4), a);
8343 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8344 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8345 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8346 DAG.getConstant(4, MVT::i32));
8347 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8348 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8351 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8353 C = ConstantVector::get(CVM2);
8354 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8355 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8356 MachinePointerInfo::getConstantPool(),
8359 // r = pblendv(r, psllw(r & (char16)63, 2), a);
8360 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8361 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8362 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8363 DAG.getConstant(2, MVT::i32));
8364 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8365 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8368 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8370 // return pblendv(r, r+r, a);
8371 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8372 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8373 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
8379 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
8380 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
8381 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
8382 // looks for this combo and may remove the "setcc" instruction if the "setcc"
8383 // has only one use.
8384 SDNode *N = Op.getNode();
8385 SDValue LHS = N->getOperand(0);
8386 SDValue RHS = N->getOperand(1);
8387 unsigned BaseOp = 0;
8389 DebugLoc dl = Op.getDebugLoc();
8391 switch (Op.getOpcode()) {
8392 default: llvm_unreachable("Unknown ovf instruction!");
8394 // A subtract of one will be selected as a INC. Note that INC doesn't
8395 // set CF, so we can't do this for UADDO.
8396 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8397 if (C->getAPIntValue() == 1) {
8398 BaseOp = X86ISD::INC;
8402 BaseOp = X86ISD::ADD;
8406 BaseOp = X86ISD::ADD;
8410 // A subtract of one will be selected as a DEC. Note that DEC doesn't
8411 // set CF, so we can't do this for USUBO.
8412 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8413 if (C->getAPIntValue() == 1) {
8414 BaseOp = X86ISD::DEC;
8418 BaseOp = X86ISD::SUB;
8422 BaseOp = X86ISD::SUB;
8426 BaseOp = X86ISD::SMUL;
8430 BaseOp = X86ISD::UMUL;
8435 // Also sets EFLAGS.
8436 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
8437 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
8440 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
8441 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
8443 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8447 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
8448 DebugLoc dl = Op.getDebugLoc();
8450 if (!Subtarget->hasSSE2()) {
8451 SDValue Chain = Op.getOperand(0);
8452 SDValue Zero = DAG.getConstant(0,
8453 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8455 DAG.getRegister(X86::ESP, MVT::i32), // Base
8456 DAG.getTargetConstant(1, MVT::i8), // Scale
8457 DAG.getRegister(0, MVT::i32), // Index
8458 DAG.getTargetConstant(0, MVT::i32), // Disp
8459 DAG.getRegister(0, MVT::i32), // Segment.
8464 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
8465 array_lengthof(Ops));
8466 return SDValue(Res, 0);
8469 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
8471 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
8473 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
8474 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
8475 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
8476 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
8478 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
8479 if (!Op1 && !Op2 && !Op3 && Op4)
8480 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
8482 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
8483 if (Op1 && !Op2 && !Op3 && !Op4)
8484 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
8486 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
8488 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
8491 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
8492 EVT T = Op.getValueType();
8493 DebugLoc dl = Op.getDebugLoc();
8496 switch(T.getSimpleVT().SimpleTy) {
8498 assert(false && "Invalid value type!");
8499 case MVT::i8: Reg = X86::AL; size = 1; break;
8500 case MVT::i16: Reg = X86::AX; size = 2; break;
8501 case MVT::i32: Reg = X86::EAX; size = 4; break;
8503 assert(Subtarget->is64Bit() && "Node not type legal!");
8504 Reg = X86::RAX; size = 8;
8507 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
8508 Op.getOperand(2), SDValue());
8509 SDValue Ops[] = { cpIn.getValue(0),
8512 DAG.getTargetConstant(size, MVT::i8),
8514 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8515 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
8517 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
8521 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
8522 SelectionDAG &DAG) const {
8523 assert(Subtarget->is64Bit() && "Result not type legalized?");
8524 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8525 SDValue TheChain = Op.getOperand(0);
8526 DebugLoc dl = Op.getDebugLoc();
8527 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8528 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
8529 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
8531 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
8532 DAG.getConstant(32, MVT::i8));
8534 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
8537 return DAG.getMergeValues(Ops, 2, dl);
8540 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
8541 SelectionDAG &DAG) const {
8542 EVT SrcVT = Op.getOperand(0).getValueType();
8543 EVT DstVT = Op.getValueType();
8544 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
8545 Subtarget->hasMMX() && !DisableMMX) &&
8546 "Unexpected custom BIT_CONVERT");
8547 assert((DstVT == MVT::i64 ||
8548 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
8549 "Unexpected custom BIT_CONVERT");
8550 // i64 <=> MMX conversions are Legal.
8551 if (SrcVT==MVT::i64 && DstVT.isVector())
8553 if (DstVT==MVT::i64 && SrcVT.isVector())
8555 // MMX <=> MMX conversions are Legal.
8556 if (SrcVT.isVector() && DstVT.isVector())
8558 // All other conversions need to be expanded.
8561 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
8562 SDNode *Node = Op.getNode();
8563 DebugLoc dl = Node->getDebugLoc();
8564 EVT T = Node->getValueType(0);
8565 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
8566 DAG.getConstant(0, T), Node->getOperand(2));
8567 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
8568 cast<AtomicSDNode>(Node)->getMemoryVT(),
8569 Node->getOperand(0),
8570 Node->getOperand(1), negOp,
8571 cast<AtomicSDNode>(Node)->getSrcValue(),
8572 cast<AtomicSDNode>(Node)->getAlignment());
8575 /// LowerOperation - Provide custom lowering hooks for some operations.
8577 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
8578 switch (Op.getOpcode()) {
8579 default: llvm_unreachable("Should not custom lower this!");
8580 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
8581 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
8582 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
8583 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
8584 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
8585 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
8586 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
8587 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
8588 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
8589 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
8590 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
8591 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
8592 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
8593 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
8594 case ISD::SHL_PARTS:
8595 case ISD::SRA_PARTS:
8596 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
8597 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
8598 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
8599 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
8600 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
8601 case ISD::FABS: return LowerFABS(Op, DAG);
8602 case ISD::FNEG: return LowerFNEG(Op, DAG);
8603 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
8604 case ISD::SETCC: return LowerSETCC(Op, DAG);
8605 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
8606 case ISD::SELECT: return LowerSELECT(Op, DAG);
8607 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
8608 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
8609 case ISD::VASTART: return LowerVASTART(Op, DAG);
8610 case ISD::VAARG: return LowerVAARG(Op, DAG);
8611 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
8612 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
8613 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
8614 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
8615 case ISD::FRAME_TO_ARGS_OFFSET:
8616 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
8617 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
8618 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
8619 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
8620 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
8621 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
8622 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
8623 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
8624 case ISD::SHL: return LowerSHL(Op, DAG);
8630 case ISD::UMULO: return LowerXALUO(Op, DAG);
8631 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
8632 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
8636 void X86TargetLowering::
8637 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
8638 SelectionDAG &DAG, unsigned NewOp) const {
8639 EVT T = Node->getValueType(0);
8640 DebugLoc dl = Node->getDebugLoc();
8641 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
8643 SDValue Chain = Node->getOperand(0);
8644 SDValue In1 = Node->getOperand(1);
8645 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8646 Node->getOperand(2), DAG.getIntPtrConstant(0));
8647 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8648 Node->getOperand(2), DAG.getIntPtrConstant(1));
8649 SDValue Ops[] = { Chain, In1, In2L, In2H };
8650 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
8652 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
8653 cast<MemSDNode>(Node)->getMemOperand());
8654 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
8655 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8656 Results.push_back(Result.getValue(2));
8659 /// ReplaceNodeResults - Replace a node with an illegal result type
8660 /// with a new node built out of custom code.
8661 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
8662 SmallVectorImpl<SDValue>&Results,
8663 SelectionDAG &DAG) const {
8664 DebugLoc dl = N->getDebugLoc();
8665 switch (N->getOpcode()) {
8667 assert(false && "Do not know how to custom type legalize this operation!");
8669 case ISD::FP_TO_SINT: {
8670 std::pair<SDValue,SDValue> Vals =
8671 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
8672 SDValue FIST = Vals.first, StackSlot = Vals.second;
8673 if (FIST.getNode() != 0) {
8674 EVT VT = N->getValueType(0);
8675 // Return a load from the stack slot.
8676 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
8677 MachinePointerInfo(), false, false, 0));
8681 case ISD::READCYCLECOUNTER: {
8682 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8683 SDValue TheChain = N->getOperand(0);
8684 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8685 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
8687 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
8689 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
8690 SDValue Ops[] = { eax, edx };
8691 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
8692 Results.push_back(edx.getValue(1));
8695 case ISD::ATOMIC_CMP_SWAP: {
8696 EVT T = N->getValueType(0);
8697 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
8698 SDValue cpInL, cpInH;
8699 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8700 DAG.getConstant(0, MVT::i32));
8701 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8702 DAG.getConstant(1, MVT::i32));
8703 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
8704 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
8706 SDValue swapInL, swapInH;
8707 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8708 DAG.getConstant(0, MVT::i32));
8709 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8710 DAG.getConstant(1, MVT::i32));
8711 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
8713 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
8714 swapInL.getValue(1));
8715 SDValue Ops[] = { swapInH.getValue(0),
8717 swapInH.getValue(1) };
8718 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8719 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
8720 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
8721 MVT::i32, Result.getValue(1));
8722 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
8723 MVT::i32, cpOutL.getValue(2));
8724 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
8725 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8726 Results.push_back(cpOutH.getValue(1));
8729 case ISD::ATOMIC_LOAD_ADD:
8730 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
8732 case ISD::ATOMIC_LOAD_AND:
8733 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
8735 case ISD::ATOMIC_LOAD_NAND:
8736 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
8738 case ISD::ATOMIC_LOAD_OR:
8739 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
8741 case ISD::ATOMIC_LOAD_SUB:
8742 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
8744 case ISD::ATOMIC_LOAD_XOR:
8745 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
8747 case ISD::ATOMIC_SWAP:
8748 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
8753 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
8755 default: return NULL;
8756 case X86ISD::BSF: return "X86ISD::BSF";
8757 case X86ISD::BSR: return "X86ISD::BSR";
8758 case X86ISD::SHLD: return "X86ISD::SHLD";
8759 case X86ISD::SHRD: return "X86ISD::SHRD";
8760 case X86ISD::FAND: return "X86ISD::FAND";
8761 case X86ISD::FOR: return "X86ISD::FOR";
8762 case X86ISD::FXOR: return "X86ISD::FXOR";
8763 case X86ISD::FSRL: return "X86ISD::FSRL";
8764 case X86ISD::FILD: return "X86ISD::FILD";
8765 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
8766 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
8767 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
8768 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
8769 case X86ISD::FLD: return "X86ISD::FLD";
8770 case X86ISD::FST: return "X86ISD::FST";
8771 case X86ISD::CALL: return "X86ISD::CALL";
8772 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
8773 case X86ISD::BT: return "X86ISD::BT";
8774 case X86ISD::CMP: return "X86ISD::CMP";
8775 case X86ISD::COMI: return "X86ISD::COMI";
8776 case X86ISD::UCOMI: return "X86ISD::UCOMI";
8777 case X86ISD::SETCC: return "X86ISD::SETCC";
8778 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
8779 case X86ISD::CMOV: return "X86ISD::CMOV";
8780 case X86ISD::BRCOND: return "X86ISD::BRCOND";
8781 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
8782 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
8783 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
8784 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
8785 case X86ISD::Wrapper: return "X86ISD::Wrapper";
8786 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
8787 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
8788 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
8789 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
8790 case X86ISD::PINSRB: return "X86ISD::PINSRB";
8791 case X86ISD::PINSRW: return "X86ISD::PINSRW";
8792 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
8793 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
8794 case X86ISD::FMAX: return "X86ISD::FMAX";
8795 case X86ISD::FMIN: return "X86ISD::FMIN";
8796 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
8797 case X86ISD::FRCP: return "X86ISD::FRCP";
8798 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
8799 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
8800 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
8801 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
8802 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
8803 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
8804 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
8805 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
8806 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
8807 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
8808 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
8809 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
8810 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
8811 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
8812 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
8813 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
8814 case X86ISD::VSHL: return "X86ISD::VSHL";
8815 case X86ISD::VSRL: return "X86ISD::VSRL";
8816 case X86ISD::CMPPD: return "X86ISD::CMPPD";
8817 case X86ISD::CMPPS: return "X86ISD::CMPPS";
8818 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
8819 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
8820 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
8821 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
8822 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
8823 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
8824 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
8825 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
8826 case X86ISD::ADD: return "X86ISD::ADD";
8827 case X86ISD::SUB: return "X86ISD::SUB";
8828 case X86ISD::SMUL: return "X86ISD::SMUL";
8829 case X86ISD::UMUL: return "X86ISD::UMUL";
8830 case X86ISD::INC: return "X86ISD::INC";
8831 case X86ISD::DEC: return "X86ISD::DEC";
8832 case X86ISD::OR: return "X86ISD::OR";
8833 case X86ISD::XOR: return "X86ISD::XOR";
8834 case X86ISD::AND: return "X86ISD::AND";
8835 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
8836 case X86ISD::PTEST: return "X86ISD::PTEST";
8837 case X86ISD::TESTP: return "X86ISD::TESTP";
8838 case X86ISD::PALIGN: return "X86ISD::PALIGN";
8839 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
8840 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
8841 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
8842 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
8843 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
8844 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
8845 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
8846 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
8847 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
8848 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
8849 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
8850 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
8851 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
8852 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
8853 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
8854 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
8855 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
8856 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
8857 case X86ISD::MOVSD: return "X86ISD::MOVSD";
8858 case X86ISD::MOVSS: return "X86ISD::MOVSS";
8859 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
8860 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
8861 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
8862 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
8863 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
8864 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
8865 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
8866 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
8867 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
8868 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
8869 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
8870 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
8871 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
8872 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
8876 // isLegalAddressingMode - Return true if the addressing mode represented
8877 // by AM is legal for this target, for a load/store of the specified type.
8878 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
8879 const Type *Ty) const {
8880 // X86 supports extremely general addressing modes.
8881 CodeModel::Model M = getTargetMachine().getCodeModel();
8882 Reloc::Model R = getTargetMachine().getRelocationModel();
8884 // X86 allows a sign-extended 32-bit immediate field as a displacement.
8885 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
8890 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
8892 // If a reference to this global requires an extra load, we can't fold it.
8893 if (isGlobalStubReference(GVFlags))
8896 // If BaseGV requires a register for the PIC base, we cannot also have a
8897 // BaseReg specified.
8898 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
8901 // If lower 4G is not available, then we must use rip-relative addressing.
8902 if ((M != CodeModel::Small || R != Reloc::Static) &&
8903 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
8913 // These scales always work.
8918 // These scales are formed with basereg+scalereg. Only accept if there is
8923 default: // Other stuff never works.
8931 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
8932 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
8934 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
8935 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
8936 if (NumBits1 <= NumBits2)
8941 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
8942 if (!VT1.isInteger() || !VT2.isInteger())
8944 unsigned NumBits1 = VT1.getSizeInBits();
8945 unsigned NumBits2 = VT2.getSizeInBits();
8946 if (NumBits1 <= NumBits2)
8951 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
8952 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8953 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
8956 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
8957 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8958 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
8961 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
8962 // i16 instructions are longer (0x66 prefix) and potentially slower.
8963 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
8966 /// isShuffleMaskLegal - Targets can use this to indicate that they only
8967 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
8968 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
8969 /// are assumed to be legal.
8971 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
8973 // Very little shuffling can be done for 64-bit vectors right now.
8974 if (VT.getSizeInBits() == 64)
8975 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8977 // FIXME: pshufb, blends, shifts.
8978 return (VT.getVectorNumElements() == 2 ||
8979 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8980 isMOVLMask(M, VT) ||
8981 isSHUFPMask(M, VT) ||
8982 isPSHUFDMask(M, VT) ||
8983 isPSHUFHWMask(M, VT) ||
8984 isPSHUFLWMask(M, VT) ||
8985 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8986 isUNPCKLMask(M, VT) ||
8987 isUNPCKHMask(M, VT) ||
8988 isUNPCKL_v_undef_Mask(M, VT) ||
8989 isUNPCKH_v_undef_Mask(M, VT));
8993 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8995 unsigned NumElts = VT.getVectorNumElements();
8996 // FIXME: This collection of masks seems suspect.
8999 if (NumElts == 4 && VT.getSizeInBits() == 128) {
9000 return (isMOVLMask(Mask, VT) ||
9001 isCommutedMOVLMask(Mask, VT, true) ||
9002 isSHUFPMask(Mask, VT) ||
9003 isCommutedSHUFPMask(Mask, VT));
9008 //===----------------------------------------------------------------------===//
9009 // X86 Scheduler Hooks
9010 //===----------------------------------------------------------------------===//
9012 // private utility function
9014 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
9015 MachineBasicBlock *MBB,
9022 TargetRegisterClass *RC,
9023 bool invSrc) const {
9024 // For the atomic bitwise operator, we generate
9027 // ld t1 = [bitinstr.addr]
9028 // op t2 = t1, [bitinstr.val]
9030 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9032 // fallthrough -->nextMBB
9033 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9034 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9035 MachineFunction::iterator MBBIter = MBB;
9038 /// First build the CFG
9039 MachineFunction *F = MBB->getParent();
9040 MachineBasicBlock *thisMBB = MBB;
9041 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9042 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9043 F->insert(MBBIter, newMBB);
9044 F->insert(MBBIter, nextMBB);
9046 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9047 nextMBB->splice(nextMBB->begin(), thisMBB,
9048 llvm::next(MachineBasicBlock::iterator(bInstr)),
9050 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9052 // Update thisMBB to fall through to newMBB
9053 thisMBB->addSuccessor(newMBB);
9055 // newMBB jumps to itself and fall through to nextMBB
9056 newMBB->addSuccessor(nextMBB);
9057 newMBB->addSuccessor(newMBB);
9059 // Insert instructions into newMBB based on incoming instruction
9060 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9061 "unexpected number of operands");
9062 DebugLoc dl = bInstr->getDebugLoc();
9063 MachineOperand& destOper = bInstr->getOperand(0);
9064 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9065 int numArgs = bInstr->getNumOperands() - 1;
9066 for (int i=0; i < numArgs; ++i)
9067 argOpers[i] = &bInstr->getOperand(i+1);
9069 // x86 address has 4 operands: base, index, scale, and displacement
9070 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9071 int valArgIndx = lastAddrIndx + 1;
9073 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9074 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
9075 for (int i=0; i <= lastAddrIndx; ++i)
9076 (*MIB).addOperand(*argOpers[i]);
9078 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
9080 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
9085 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9086 assert((argOpers[valArgIndx]->isReg() ||
9087 argOpers[valArgIndx]->isImm()) &&
9089 if (argOpers[valArgIndx]->isReg())
9090 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
9092 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
9094 (*MIB).addOperand(*argOpers[valArgIndx]);
9096 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
9099 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
9100 for (int i=0; i <= lastAddrIndx; ++i)
9101 (*MIB).addOperand(*argOpers[i]);
9103 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9104 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9105 bInstr->memoperands_end());
9107 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9111 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9113 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9117 // private utility function: 64 bit atomics on 32 bit host.
9119 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
9120 MachineBasicBlock *MBB,
9125 bool invSrc) const {
9126 // For the atomic bitwise operator, we generate
9127 // thisMBB (instructions are in pairs, except cmpxchg8b)
9128 // ld t1,t2 = [bitinstr.addr]
9130 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
9131 // op t5, t6 <- out1, out2, [bitinstr.val]
9132 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
9133 // mov ECX, EBX <- t5, t6
9134 // mov EAX, EDX <- t1, t2
9135 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
9136 // mov t3, t4 <- EAX, EDX
9138 // result in out1, out2
9139 // fallthrough -->nextMBB
9141 const TargetRegisterClass *RC = X86::GR32RegisterClass;
9142 const unsigned LoadOpc = X86::MOV32rm;
9143 const unsigned NotOpc = X86::NOT32r;
9144 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9145 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9146 MachineFunction::iterator MBBIter = MBB;
9149 /// First build the CFG
9150 MachineFunction *F = MBB->getParent();
9151 MachineBasicBlock *thisMBB = MBB;
9152 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9153 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9154 F->insert(MBBIter, newMBB);
9155 F->insert(MBBIter, nextMBB);
9157 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9158 nextMBB->splice(nextMBB->begin(), thisMBB,
9159 llvm::next(MachineBasicBlock::iterator(bInstr)),
9161 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9163 // Update thisMBB to fall through to newMBB
9164 thisMBB->addSuccessor(newMBB);
9166 // newMBB jumps to itself and fall through to nextMBB
9167 newMBB->addSuccessor(nextMBB);
9168 newMBB->addSuccessor(newMBB);
9170 DebugLoc dl = bInstr->getDebugLoc();
9171 // Insert instructions into newMBB based on incoming instruction
9172 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
9173 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
9174 "unexpected number of operands");
9175 MachineOperand& dest1Oper = bInstr->getOperand(0);
9176 MachineOperand& dest2Oper = bInstr->getOperand(1);
9177 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9178 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
9179 argOpers[i] = &bInstr->getOperand(i+2);
9181 // We use some of the operands multiple times, so conservatively just
9182 // clear any kill flags that might be present.
9183 if (argOpers[i]->isReg() && argOpers[i]->isUse())
9184 argOpers[i]->setIsKill(false);
9187 // x86 address has 5 operands: base, index, scale, displacement, and segment.
9188 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9190 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9191 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
9192 for (int i=0; i <= lastAddrIndx; ++i)
9193 (*MIB).addOperand(*argOpers[i]);
9194 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9195 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
9196 // add 4 to displacement.
9197 for (int i=0; i <= lastAddrIndx-2; ++i)
9198 (*MIB).addOperand(*argOpers[i]);
9199 MachineOperand newOp3 = *(argOpers[3]);
9201 newOp3.setImm(newOp3.getImm()+4);
9203 newOp3.setOffset(newOp3.getOffset()+4);
9204 (*MIB).addOperand(newOp3);
9205 (*MIB).addOperand(*argOpers[lastAddrIndx]);
9207 // t3/4 are defined later, at the bottom of the loop
9208 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
9209 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
9210 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
9211 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
9212 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
9213 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
9215 // The subsequent operations should be using the destination registers of
9216 //the PHI instructions.
9218 t1 = F->getRegInfo().createVirtualRegister(RC);
9219 t2 = F->getRegInfo().createVirtualRegister(RC);
9220 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
9221 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
9223 t1 = dest1Oper.getReg();
9224 t2 = dest2Oper.getReg();
9227 int valArgIndx = lastAddrIndx + 1;
9228 assert((argOpers[valArgIndx]->isReg() ||
9229 argOpers[valArgIndx]->isImm()) &&
9231 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
9232 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
9233 if (argOpers[valArgIndx]->isReg())
9234 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
9236 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
9237 if (regOpcL != X86::MOV32rr)
9239 (*MIB).addOperand(*argOpers[valArgIndx]);
9240 assert(argOpers[valArgIndx + 1]->isReg() ==
9241 argOpers[valArgIndx]->isReg());
9242 assert(argOpers[valArgIndx + 1]->isImm() ==
9243 argOpers[valArgIndx]->isImm());
9244 if (argOpers[valArgIndx + 1]->isReg())
9245 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
9247 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
9248 if (regOpcH != X86::MOV32rr)
9250 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
9252 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9254 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
9257 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
9259 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
9262 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
9263 for (int i=0; i <= lastAddrIndx; ++i)
9264 (*MIB).addOperand(*argOpers[i]);
9266 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9267 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9268 bInstr->memoperands_end());
9270 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
9271 MIB.addReg(X86::EAX);
9272 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
9273 MIB.addReg(X86::EDX);
9276 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9278 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9282 // private utility function
9284 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
9285 MachineBasicBlock *MBB,
9286 unsigned cmovOpc) const {
9287 // For the atomic min/max operator, we generate
9290 // ld t1 = [min/max.addr]
9291 // mov t2 = [min/max.val]
9293 // cmov[cond] t2 = t1
9295 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9297 // fallthrough -->nextMBB
9299 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9300 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9301 MachineFunction::iterator MBBIter = MBB;
9304 /// First build the CFG
9305 MachineFunction *F = MBB->getParent();
9306 MachineBasicBlock *thisMBB = MBB;
9307 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9308 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9309 F->insert(MBBIter, newMBB);
9310 F->insert(MBBIter, nextMBB);
9312 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9313 nextMBB->splice(nextMBB->begin(), thisMBB,
9314 llvm::next(MachineBasicBlock::iterator(mInstr)),
9316 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9318 // Update thisMBB to fall through to newMBB
9319 thisMBB->addSuccessor(newMBB);
9321 // newMBB jumps to newMBB and fall through to nextMBB
9322 newMBB->addSuccessor(nextMBB);
9323 newMBB->addSuccessor(newMBB);
9325 DebugLoc dl = mInstr->getDebugLoc();
9326 // Insert instructions into newMBB based on incoming instruction
9327 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9328 "unexpected number of operands");
9329 MachineOperand& destOper = mInstr->getOperand(0);
9330 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9331 int numArgs = mInstr->getNumOperands() - 1;
9332 for (int i=0; i < numArgs; ++i)
9333 argOpers[i] = &mInstr->getOperand(i+1);
9335 // x86 address has 4 operands: base, index, scale, and displacement
9336 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9337 int valArgIndx = lastAddrIndx + 1;
9339 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9340 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
9341 for (int i=0; i <= lastAddrIndx; ++i)
9342 (*MIB).addOperand(*argOpers[i]);
9344 // We only support register and immediate values
9345 assert((argOpers[valArgIndx]->isReg() ||
9346 argOpers[valArgIndx]->isImm()) &&
9349 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9350 if (argOpers[valArgIndx]->isReg())
9351 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
9353 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
9354 (*MIB).addOperand(*argOpers[valArgIndx]);
9356 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9359 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
9364 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9365 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
9369 // Cmp and exchange if none has modified the memory location
9370 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
9371 for (int i=0; i <= lastAddrIndx; ++i)
9372 (*MIB).addOperand(*argOpers[i]);
9374 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9375 (*MIB).setMemRefs(mInstr->memoperands_begin(),
9376 mInstr->memoperands_end());
9378 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9379 MIB.addReg(X86::EAX);
9382 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9384 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
9388 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
9389 // or XMM0_V32I8 in AVX all of this code can be replaced with that
9392 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
9393 unsigned numArgs, bool memArg) const {
9395 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
9396 "Target must have SSE4.2 or AVX features enabled");
9398 DebugLoc dl = MI->getDebugLoc();
9399 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9403 if (!Subtarget->hasAVX()) {
9405 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
9407 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
9410 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
9412 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
9415 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
9417 for (unsigned i = 0; i < numArgs; ++i) {
9418 MachineOperand &Op = MI->getOperand(i+1);
9420 if (!(Op.isReg() && Op.isImplicit()))
9424 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
9427 MI->eraseFromParent();
9433 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
9435 MachineBasicBlock *MBB) const {
9436 // Emit code to save XMM registers to the stack. The ABI says that the
9437 // number of registers to save is given in %al, so it's theoretically
9438 // possible to do an indirect jump trick to avoid saving all of them,
9439 // however this code takes a simpler approach and just executes all
9440 // of the stores if %al is non-zero. It's less code, and it's probably
9441 // easier on the hardware branch predictor, and stores aren't all that
9442 // expensive anyway.
9444 // Create the new basic blocks. One block contains all the XMM stores,
9445 // and one block is the final destination regardless of whether any
9446 // stores were performed.
9447 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9448 MachineFunction *F = MBB->getParent();
9449 MachineFunction::iterator MBBIter = MBB;
9451 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
9452 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
9453 F->insert(MBBIter, XMMSaveMBB);
9454 F->insert(MBBIter, EndMBB);
9456 // Transfer the remainder of MBB and its successor edges to EndMBB.
9457 EndMBB->splice(EndMBB->begin(), MBB,
9458 llvm::next(MachineBasicBlock::iterator(MI)),
9460 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
9462 // The original block will now fall through to the XMM save block.
9463 MBB->addSuccessor(XMMSaveMBB);
9464 // The XMMSaveMBB will fall through to the end block.
9465 XMMSaveMBB->addSuccessor(EndMBB);
9467 // Now add the instructions.
9468 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9469 DebugLoc DL = MI->getDebugLoc();
9471 unsigned CountReg = MI->getOperand(0).getReg();
9472 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
9473 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
9475 if (!Subtarget->isTargetWin64()) {
9476 // If %al is 0, branch around the XMM save block.
9477 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
9478 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
9479 MBB->addSuccessor(EndMBB);
9482 // In the XMM save block, save all the XMM argument registers.
9483 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
9484 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
9485 MachineMemOperand *MMO =
9486 F->getMachineMemOperand(
9487 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
9488 MachineMemOperand::MOStore,
9489 /*Size=*/16, /*Align=*/16);
9490 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
9491 .addFrameIndex(RegSaveFrameIndex)
9492 .addImm(/*Scale=*/1)
9493 .addReg(/*IndexReg=*/0)
9494 .addImm(/*Disp=*/Offset)
9495 .addReg(/*Segment=*/0)
9496 .addReg(MI->getOperand(i).getReg())
9497 .addMemOperand(MMO);
9500 MI->eraseFromParent(); // The pseudo instruction is gone now.
9506 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
9507 MachineBasicBlock *BB) const {
9508 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9509 DebugLoc DL = MI->getDebugLoc();
9511 // To "insert" a SELECT_CC instruction, we actually have to insert the
9512 // diamond control-flow pattern. The incoming instruction knows the
9513 // destination vreg to set, the condition code register to branch on, the
9514 // true/false values to select between, and a branch opcode to use.
9515 const BasicBlock *LLVM_BB = BB->getBasicBlock();
9516 MachineFunction::iterator It = BB;
9522 // cmpTY ccX, r1, r2
9524 // fallthrough --> copy0MBB
9525 MachineBasicBlock *thisMBB = BB;
9526 MachineFunction *F = BB->getParent();
9527 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
9528 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
9529 F->insert(It, copy0MBB);
9530 F->insert(It, sinkMBB);
9532 // If the EFLAGS register isn't dead in the terminator, then claim that it's
9533 // live into the sink and copy blocks.
9534 const MachineFunction *MF = BB->getParent();
9535 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
9536 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
9538 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
9539 const MachineOperand &MO = MI->getOperand(I);
9540 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
9541 unsigned Reg = MO.getReg();
9542 if (Reg != X86::EFLAGS) continue;
9543 copy0MBB->addLiveIn(Reg);
9544 sinkMBB->addLiveIn(Reg);
9547 // Transfer the remainder of BB and its successor edges to sinkMBB.
9548 sinkMBB->splice(sinkMBB->begin(), BB,
9549 llvm::next(MachineBasicBlock::iterator(MI)),
9551 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
9553 // Add the true and fallthrough blocks as its successors.
9554 BB->addSuccessor(copy0MBB);
9555 BB->addSuccessor(sinkMBB);
9557 // Create the conditional branch instruction.
9559 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
9560 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
9563 // %FalseValue = ...
9564 // # fallthrough to sinkMBB
9565 copy0MBB->addSuccessor(sinkMBB);
9568 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
9570 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
9571 TII->get(X86::PHI), MI->getOperand(0).getReg())
9572 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
9573 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
9575 MI->eraseFromParent(); // The pseudo instruction is gone now.
9580 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
9581 MachineBasicBlock *BB) const {
9582 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9583 DebugLoc DL = MI->getDebugLoc();
9585 // The lowering is pretty easy: we're just emitting the call to _alloca. The
9586 // non-trivial part is impdef of ESP.
9587 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
9590 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
9591 .addExternalSymbol("_alloca")
9592 .addReg(X86::EAX, RegState::Implicit)
9593 .addReg(X86::ESP, RegState::Implicit)
9594 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
9595 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
9596 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
9598 MI->eraseFromParent(); // The pseudo instruction is gone now.
9603 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
9604 MachineBasicBlock *BB) const {
9605 // This is pretty easy. We're taking the value that we received from
9606 // our load from the relocation, sticking it in either RDI (x86-64)
9607 // or EAX and doing an indirect call. The return value will then
9608 // be in the normal return register.
9609 const X86InstrInfo *TII
9610 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
9611 DebugLoc DL = MI->getDebugLoc();
9612 MachineFunction *F = BB->getParent();
9613 bool IsWin64 = Subtarget->isTargetWin64();
9615 assert(MI->getOperand(3).isGlobal() && "This should be a global");
9617 if (Subtarget->is64Bit()) {
9618 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9619 TII->get(X86::MOV64rm), X86::RDI)
9621 .addImm(0).addReg(0)
9622 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9623 MI->getOperand(3).getTargetFlags())
9625 MIB = BuildMI(*BB, MI, DL, TII->get(IsWin64 ? X86::WINCALL64m : X86::CALL64m));
9626 addDirectMem(MIB, X86::RDI);
9627 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
9628 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9629 TII->get(X86::MOV32rm), X86::EAX)
9631 .addImm(0).addReg(0)
9632 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9633 MI->getOperand(3).getTargetFlags())
9635 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9636 addDirectMem(MIB, X86::EAX);
9638 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9639 TII->get(X86::MOV32rm), X86::EAX)
9640 .addReg(TII->getGlobalBaseReg(F))
9641 .addImm(0).addReg(0)
9642 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9643 MI->getOperand(3).getTargetFlags())
9645 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9646 addDirectMem(MIB, X86::EAX);
9649 MI->eraseFromParent(); // The pseudo instruction is gone now.
9654 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
9655 MachineBasicBlock *BB) const {
9656 switch (MI->getOpcode()) {
9657 default: assert(false && "Unexpected instr type to insert");
9658 case X86::MINGW_ALLOCA:
9659 return EmitLoweredMingwAlloca(MI, BB);
9660 case X86::TLSCall_32:
9661 case X86::TLSCall_64:
9662 return EmitLoweredTLSCall(MI, BB);
9664 case X86::CMOV_V1I64:
9665 case X86::CMOV_FR32:
9666 case X86::CMOV_FR64:
9667 case X86::CMOV_V4F32:
9668 case X86::CMOV_V2F64:
9669 case X86::CMOV_V2I64:
9670 case X86::CMOV_GR16:
9671 case X86::CMOV_GR32:
9672 case X86::CMOV_RFP32:
9673 case X86::CMOV_RFP64:
9674 case X86::CMOV_RFP80:
9675 return EmitLoweredSelect(MI, BB);
9677 case X86::FP32_TO_INT16_IN_MEM:
9678 case X86::FP32_TO_INT32_IN_MEM:
9679 case X86::FP32_TO_INT64_IN_MEM:
9680 case X86::FP64_TO_INT16_IN_MEM:
9681 case X86::FP64_TO_INT32_IN_MEM:
9682 case X86::FP64_TO_INT64_IN_MEM:
9683 case X86::FP80_TO_INT16_IN_MEM:
9684 case X86::FP80_TO_INT32_IN_MEM:
9685 case X86::FP80_TO_INT64_IN_MEM: {
9686 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9687 DebugLoc DL = MI->getDebugLoc();
9689 // Change the floating point control register to use "round towards zero"
9690 // mode when truncating to an integer value.
9691 MachineFunction *F = BB->getParent();
9692 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
9693 addFrameReference(BuildMI(*BB, MI, DL,
9694 TII->get(X86::FNSTCW16m)), CWFrameIdx);
9696 // Load the old value of the high byte of the control word...
9698 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
9699 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
9702 // Set the high part to be round to zero...
9703 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
9706 // Reload the modified control word now...
9707 addFrameReference(BuildMI(*BB, MI, DL,
9708 TII->get(X86::FLDCW16m)), CWFrameIdx);
9710 // Restore the memory image of control word to original value
9711 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
9714 // Get the X86 opcode to use.
9716 switch (MI->getOpcode()) {
9717 default: llvm_unreachable("illegal opcode!");
9718 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
9719 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
9720 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
9721 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
9722 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
9723 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
9724 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
9725 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
9726 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
9730 MachineOperand &Op = MI->getOperand(0);
9732 AM.BaseType = X86AddressMode::RegBase;
9733 AM.Base.Reg = Op.getReg();
9735 AM.BaseType = X86AddressMode::FrameIndexBase;
9736 AM.Base.FrameIndex = Op.getIndex();
9738 Op = MI->getOperand(1);
9740 AM.Scale = Op.getImm();
9741 Op = MI->getOperand(2);
9743 AM.IndexReg = Op.getImm();
9744 Op = MI->getOperand(3);
9745 if (Op.isGlobal()) {
9746 AM.GV = Op.getGlobal();
9748 AM.Disp = Op.getImm();
9750 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
9751 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
9753 // Reload the original control word now.
9754 addFrameReference(BuildMI(*BB, MI, DL,
9755 TII->get(X86::FLDCW16m)), CWFrameIdx);
9757 MI->eraseFromParent(); // The pseudo instruction is gone now.
9760 // String/text processing lowering.
9761 case X86::PCMPISTRM128REG:
9762 case X86::VPCMPISTRM128REG:
9763 return EmitPCMP(MI, BB, 3, false /* in-mem */);
9764 case X86::PCMPISTRM128MEM:
9765 case X86::VPCMPISTRM128MEM:
9766 return EmitPCMP(MI, BB, 3, true /* in-mem */);
9767 case X86::PCMPESTRM128REG:
9768 case X86::VPCMPESTRM128REG:
9769 return EmitPCMP(MI, BB, 5, false /* in mem */);
9770 case X86::PCMPESTRM128MEM:
9771 case X86::VPCMPESTRM128MEM:
9772 return EmitPCMP(MI, BB, 5, true /* in mem */);
9775 case X86::ATOMAND32:
9776 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9777 X86::AND32ri, X86::MOV32rm,
9779 X86::NOT32r, X86::EAX,
9780 X86::GR32RegisterClass);
9782 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
9783 X86::OR32ri, X86::MOV32rm,
9785 X86::NOT32r, X86::EAX,
9786 X86::GR32RegisterClass);
9787 case X86::ATOMXOR32:
9788 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
9789 X86::XOR32ri, X86::MOV32rm,
9791 X86::NOT32r, X86::EAX,
9792 X86::GR32RegisterClass);
9793 case X86::ATOMNAND32:
9794 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9795 X86::AND32ri, X86::MOV32rm,
9797 X86::NOT32r, X86::EAX,
9798 X86::GR32RegisterClass, true);
9799 case X86::ATOMMIN32:
9800 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
9801 case X86::ATOMMAX32:
9802 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
9803 case X86::ATOMUMIN32:
9804 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
9805 case X86::ATOMUMAX32:
9806 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
9808 case X86::ATOMAND16:
9809 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9810 X86::AND16ri, X86::MOV16rm,
9812 X86::NOT16r, X86::AX,
9813 X86::GR16RegisterClass);
9815 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
9816 X86::OR16ri, X86::MOV16rm,
9818 X86::NOT16r, X86::AX,
9819 X86::GR16RegisterClass);
9820 case X86::ATOMXOR16:
9821 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
9822 X86::XOR16ri, X86::MOV16rm,
9824 X86::NOT16r, X86::AX,
9825 X86::GR16RegisterClass);
9826 case X86::ATOMNAND16:
9827 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9828 X86::AND16ri, X86::MOV16rm,
9830 X86::NOT16r, X86::AX,
9831 X86::GR16RegisterClass, true);
9832 case X86::ATOMMIN16:
9833 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
9834 case X86::ATOMMAX16:
9835 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
9836 case X86::ATOMUMIN16:
9837 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
9838 case X86::ATOMUMAX16:
9839 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
9842 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9843 X86::AND8ri, X86::MOV8rm,
9845 X86::NOT8r, X86::AL,
9846 X86::GR8RegisterClass);
9848 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
9849 X86::OR8ri, X86::MOV8rm,
9851 X86::NOT8r, X86::AL,
9852 X86::GR8RegisterClass);
9854 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
9855 X86::XOR8ri, X86::MOV8rm,
9857 X86::NOT8r, X86::AL,
9858 X86::GR8RegisterClass);
9859 case X86::ATOMNAND8:
9860 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9861 X86::AND8ri, X86::MOV8rm,
9863 X86::NOT8r, X86::AL,
9864 X86::GR8RegisterClass, true);
9865 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
9866 // This group is for 64-bit host.
9867 case X86::ATOMAND64:
9868 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9869 X86::AND64ri32, X86::MOV64rm,
9871 X86::NOT64r, X86::RAX,
9872 X86::GR64RegisterClass);
9874 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
9875 X86::OR64ri32, X86::MOV64rm,
9877 X86::NOT64r, X86::RAX,
9878 X86::GR64RegisterClass);
9879 case X86::ATOMXOR64:
9880 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
9881 X86::XOR64ri32, X86::MOV64rm,
9883 X86::NOT64r, X86::RAX,
9884 X86::GR64RegisterClass);
9885 case X86::ATOMNAND64:
9886 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9887 X86::AND64ri32, X86::MOV64rm,
9889 X86::NOT64r, X86::RAX,
9890 X86::GR64RegisterClass, true);
9891 case X86::ATOMMIN64:
9892 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
9893 case X86::ATOMMAX64:
9894 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
9895 case X86::ATOMUMIN64:
9896 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
9897 case X86::ATOMUMAX64:
9898 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
9900 // This group does 64-bit operations on a 32-bit host.
9901 case X86::ATOMAND6432:
9902 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9903 X86::AND32rr, X86::AND32rr,
9904 X86::AND32ri, X86::AND32ri,
9906 case X86::ATOMOR6432:
9907 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9908 X86::OR32rr, X86::OR32rr,
9909 X86::OR32ri, X86::OR32ri,
9911 case X86::ATOMXOR6432:
9912 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9913 X86::XOR32rr, X86::XOR32rr,
9914 X86::XOR32ri, X86::XOR32ri,
9916 case X86::ATOMNAND6432:
9917 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9918 X86::AND32rr, X86::AND32rr,
9919 X86::AND32ri, X86::AND32ri,
9921 case X86::ATOMADD6432:
9922 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9923 X86::ADD32rr, X86::ADC32rr,
9924 X86::ADD32ri, X86::ADC32ri,
9926 case X86::ATOMSUB6432:
9927 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9928 X86::SUB32rr, X86::SBB32rr,
9929 X86::SUB32ri, X86::SBB32ri,
9931 case X86::ATOMSWAP6432:
9932 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9933 X86::MOV32rr, X86::MOV32rr,
9934 X86::MOV32ri, X86::MOV32ri,
9936 case X86::VASTART_SAVE_XMM_REGS:
9937 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
9941 //===----------------------------------------------------------------------===//
9942 // X86 Optimization Hooks
9943 //===----------------------------------------------------------------------===//
9945 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9949 const SelectionDAG &DAG,
9950 unsigned Depth) const {
9951 unsigned Opc = Op.getOpcode();
9952 assert((Opc >= ISD::BUILTIN_OP_END ||
9953 Opc == ISD::INTRINSIC_WO_CHAIN ||
9954 Opc == ISD::INTRINSIC_W_CHAIN ||
9955 Opc == ISD::INTRINSIC_VOID) &&
9956 "Should use MaskedValueIsZero if you don't know whether Op"
9957 " is a target node!");
9959 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
9971 // These nodes' second result is a boolean.
9972 if (Op.getResNo() == 0)
9976 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
9977 Mask.getBitWidth() - 1);
9982 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
9983 unsigned Depth) const {
9984 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
9985 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
9986 return Op.getValueType().getScalarType().getSizeInBits();
9992 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
9993 /// node is a GlobalAddress + offset.
9994 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
9995 const GlobalValue* &GA,
9996 int64_t &Offset) const {
9997 if (N->getOpcode() == X86ISD::Wrapper) {
9998 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
9999 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
10000 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
10004 return TargetLowering::isGAPlusOffset(N, GA, Offset);
10007 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
10008 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
10009 /// if the load addresses are consecutive, non-overlapping, and in the right
10011 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
10012 const TargetLowering &TLI) {
10013 DebugLoc dl = N->getDebugLoc();
10014 EVT VT = N->getValueType(0);
10016 if (VT.getSizeInBits() != 128)
10019 SmallVector<SDValue, 16> Elts;
10020 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
10021 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
10023 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
10026 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
10027 /// generation and convert it from being a bunch of shuffles and extracts
10028 /// to a simple store and scalar loads to extract the elements.
10029 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
10030 const TargetLowering &TLI) {
10031 SDValue InputVector = N->getOperand(0);
10033 // Only operate on vectors of 4 elements, where the alternative shuffling
10034 // gets to be more expensive.
10035 if (InputVector.getValueType() != MVT::v4i32)
10038 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
10039 // single use which is a sign-extend or zero-extend, and all elements are
10041 SmallVector<SDNode *, 4> Uses;
10042 unsigned ExtractedElements = 0;
10043 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
10044 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
10045 if (UI.getUse().getResNo() != InputVector.getResNo())
10048 SDNode *Extract = *UI;
10049 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
10052 if (Extract->getValueType(0) != MVT::i32)
10054 if (!Extract->hasOneUse())
10056 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
10057 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
10059 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
10062 // Record which element was extracted.
10063 ExtractedElements |=
10064 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
10066 Uses.push_back(Extract);
10069 // If not all the elements were used, this may not be worthwhile.
10070 if (ExtractedElements != 15)
10073 // Ok, we've now decided to do the transformation.
10074 DebugLoc dl = InputVector.getDebugLoc();
10076 // Store the value to a temporary stack slot.
10077 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
10078 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
10079 MachinePointerInfo(), false, false, 0);
10081 // Replace each use (extract) with a load of the appropriate element.
10082 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
10083 UE = Uses.end(); UI != UE; ++UI) {
10084 SDNode *Extract = *UI;
10086 // Compute the element's address.
10087 SDValue Idx = Extract->getOperand(1);
10089 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
10090 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
10091 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
10093 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
10094 StackPtr, OffsetVal);
10096 // Load the scalar.
10097 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
10098 ScalarAddr, MachinePointerInfo(),
10101 // Replace the exact with the load.
10102 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
10105 // The replacement was made in place; don't return anything.
10109 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
10110 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
10111 const X86Subtarget *Subtarget) {
10112 DebugLoc DL = N->getDebugLoc();
10113 SDValue Cond = N->getOperand(0);
10114 // Get the LHS/RHS of the select.
10115 SDValue LHS = N->getOperand(1);
10116 SDValue RHS = N->getOperand(2);
10118 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
10119 // instructions match the semantics of the common C idiom x<y?x:y but not
10120 // x<=y?x:y, because of how they handle negative zero (which can be
10121 // ignored in unsafe-math mode).
10122 if (Subtarget->hasSSE2() &&
10123 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
10124 Cond.getOpcode() == ISD::SETCC) {
10125 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
10127 unsigned Opcode = 0;
10128 // Check for x CC y ? x : y.
10129 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
10130 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
10134 // Converting this to a min would handle NaNs incorrectly, and swapping
10135 // the operands would cause it to handle comparisons between positive
10136 // and negative zero incorrectly.
10137 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
10138 if (!UnsafeFPMath &&
10139 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10141 std::swap(LHS, RHS);
10143 Opcode = X86ISD::FMIN;
10146 // Converting this to a min would handle comparisons between positive
10147 // and negative zero incorrectly.
10148 if (!UnsafeFPMath &&
10149 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
10151 Opcode = X86ISD::FMIN;
10154 // Converting this to a min would handle both negative zeros and NaNs
10155 // incorrectly, but we can swap the operands to fix both.
10156 std::swap(LHS, RHS);
10160 Opcode = X86ISD::FMIN;
10164 // Converting this to a max would handle comparisons between positive
10165 // and negative zero incorrectly.
10166 if (!UnsafeFPMath &&
10167 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
10169 Opcode = X86ISD::FMAX;
10172 // Converting this to a max would handle NaNs incorrectly, and swapping
10173 // the operands would cause it to handle comparisons between positive
10174 // and negative zero incorrectly.
10175 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
10176 if (!UnsafeFPMath &&
10177 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10179 std::swap(LHS, RHS);
10181 Opcode = X86ISD::FMAX;
10184 // Converting this to a max would handle both negative zeros and NaNs
10185 // incorrectly, but we can swap the operands to fix both.
10186 std::swap(LHS, RHS);
10190 Opcode = X86ISD::FMAX;
10193 // Check for x CC y ? y : x -- a min/max with reversed arms.
10194 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
10195 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
10199 // Converting this to a min would handle comparisons between positive
10200 // and negative zero incorrectly, and swapping the operands would
10201 // cause it to handle NaNs incorrectly.
10202 if (!UnsafeFPMath &&
10203 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
10204 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10206 std::swap(LHS, RHS);
10208 Opcode = X86ISD::FMIN;
10211 // Converting this to a min would handle NaNs incorrectly.
10212 if (!UnsafeFPMath &&
10213 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
10215 Opcode = X86ISD::FMIN;
10218 // Converting this to a min would handle both negative zeros and NaNs
10219 // incorrectly, but we can swap the operands to fix both.
10220 std::swap(LHS, RHS);
10224 Opcode = X86ISD::FMIN;
10228 // Converting this to a max would handle NaNs incorrectly.
10229 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10231 Opcode = X86ISD::FMAX;
10234 // Converting this to a max would handle comparisons between positive
10235 // and negative zero incorrectly, and swapping the operands would
10236 // cause it to handle NaNs incorrectly.
10237 if (!UnsafeFPMath &&
10238 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
10239 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10241 std::swap(LHS, RHS);
10243 Opcode = X86ISD::FMAX;
10246 // Converting this to a max would handle both negative zeros and NaNs
10247 // incorrectly, but we can swap the operands to fix both.
10248 std::swap(LHS, RHS);
10252 Opcode = X86ISD::FMAX;
10258 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
10261 // If this is a select between two integer constants, try to do some
10263 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
10264 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
10265 // Don't do this for crazy integer types.
10266 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
10267 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
10268 // so that TrueC (the true value) is larger than FalseC.
10269 bool NeedsCondInvert = false;
10271 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
10272 // Efficiently invertible.
10273 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
10274 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
10275 isa<ConstantSDNode>(Cond.getOperand(1))))) {
10276 NeedsCondInvert = true;
10277 std::swap(TrueC, FalseC);
10280 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
10281 if (FalseC->getAPIntValue() == 0 &&
10282 TrueC->getAPIntValue().isPowerOf2()) {
10283 if (NeedsCondInvert) // Invert the condition if needed.
10284 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10285 DAG.getConstant(1, Cond.getValueType()));
10287 // Zero extend the condition if needed.
10288 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
10290 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10291 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
10292 DAG.getConstant(ShAmt, MVT::i8));
10295 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
10296 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10297 if (NeedsCondInvert) // Invert the condition if needed.
10298 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10299 DAG.getConstant(1, Cond.getValueType()));
10301 // Zero extend the condition if needed.
10302 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10303 FalseC->getValueType(0), Cond);
10304 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10305 SDValue(FalseC, 0));
10308 // Optimize cases that will turn into an LEA instruction. This requires
10309 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10310 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10311 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10312 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10314 bool isFastMultiplier = false;
10316 switch ((unsigned char)Diff) {
10318 case 1: // result = add base, cond
10319 case 2: // result = lea base( , cond*2)
10320 case 3: // result = lea base(cond, cond*2)
10321 case 4: // result = lea base( , cond*4)
10322 case 5: // result = lea base(cond, cond*4)
10323 case 8: // result = lea base( , cond*8)
10324 case 9: // result = lea base(cond, cond*8)
10325 isFastMultiplier = true;
10330 if (isFastMultiplier) {
10331 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10332 if (NeedsCondInvert) // Invert the condition if needed.
10333 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10334 DAG.getConstant(1, Cond.getValueType()));
10336 // Zero extend the condition if needed.
10337 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10339 // Scale the condition by the difference.
10341 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10342 DAG.getConstant(Diff, Cond.getValueType()));
10344 // Add the base if non-zero.
10345 if (FalseC->getAPIntValue() != 0)
10346 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10347 SDValue(FalseC, 0));
10357 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
10358 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
10359 TargetLowering::DAGCombinerInfo &DCI) {
10360 DebugLoc DL = N->getDebugLoc();
10362 // If the flag operand isn't dead, don't touch this CMOV.
10363 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
10366 // If this is a select between two integer constants, try to do some
10367 // optimizations. Note that the operands are ordered the opposite of SELECT
10369 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
10370 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
10371 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
10372 // larger than FalseC (the false value).
10373 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
10375 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
10376 CC = X86::GetOppositeBranchCondition(CC);
10377 std::swap(TrueC, FalseC);
10380 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
10381 // This is efficient for any integer data type (including i8/i16) and
10383 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
10384 SDValue Cond = N->getOperand(3);
10385 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10386 DAG.getConstant(CC, MVT::i8), Cond);
10388 // Zero extend the condition if needed.
10389 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
10391 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10392 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
10393 DAG.getConstant(ShAmt, MVT::i8));
10394 if (N->getNumValues() == 2) // Dead flag value?
10395 return DCI.CombineTo(N, Cond, SDValue());
10399 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
10400 // for any integer data type, including i8/i16.
10401 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10402 SDValue Cond = N->getOperand(3);
10403 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10404 DAG.getConstant(CC, MVT::i8), Cond);
10406 // Zero extend the condition if needed.
10407 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10408 FalseC->getValueType(0), Cond);
10409 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10410 SDValue(FalseC, 0));
10412 if (N->getNumValues() == 2) // Dead flag value?
10413 return DCI.CombineTo(N, Cond, SDValue());
10417 // Optimize cases that will turn into an LEA instruction. This requires
10418 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10419 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10420 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10421 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10423 bool isFastMultiplier = false;
10425 switch ((unsigned char)Diff) {
10427 case 1: // result = add base, cond
10428 case 2: // result = lea base( , cond*2)
10429 case 3: // result = lea base(cond, cond*2)
10430 case 4: // result = lea base( , cond*4)
10431 case 5: // result = lea base(cond, cond*4)
10432 case 8: // result = lea base( , cond*8)
10433 case 9: // result = lea base(cond, cond*8)
10434 isFastMultiplier = true;
10439 if (isFastMultiplier) {
10440 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10441 SDValue Cond = N->getOperand(3);
10442 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10443 DAG.getConstant(CC, MVT::i8), Cond);
10444 // Zero extend the condition if needed.
10445 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10447 // Scale the condition by the difference.
10449 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10450 DAG.getConstant(Diff, Cond.getValueType()));
10452 // Add the base if non-zero.
10453 if (FalseC->getAPIntValue() != 0)
10454 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10455 SDValue(FalseC, 0));
10456 if (N->getNumValues() == 2) // Dead flag value?
10457 return DCI.CombineTo(N, Cond, SDValue());
10467 /// PerformMulCombine - Optimize a single multiply with constant into two
10468 /// in order to implement it with two cheaper instructions, e.g.
10469 /// LEA + SHL, LEA + LEA.
10470 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
10471 TargetLowering::DAGCombinerInfo &DCI) {
10472 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
10475 EVT VT = N->getValueType(0);
10476 if (VT != MVT::i64)
10479 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
10482 uint64_t MulAmt = C->getZExtValue();
10483 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
10486 uint64_t MulAmt1 = 0;
10487 uint64_t MulAmt2 = 0;
10488 if ((MulAmt % 9) == 0) {
10490 MulAmt2 = MulAmt / 9;
10491 } else if ((MulAmt % 5) == 0) {
10493 MulAmt2 = MulAmt / 5;
10494 } else if ((MulAmt % 3) == 0) {
10496 MulAmt2 = MulAmt / 3;
10499 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
10500 DebugLoc DL = N->getDebugLoc();
10502 if (isPowerOf2_64(MulAmt2) &&
10503 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
10504 // If second multiplifer is pow2, issue it first. We want the multiply by
10505 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
10507 std::swap(MulAmt1, MulAmt2);
10510 if (isPowerOf2_64(MulAmt1))
10511 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
10512 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
10514 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
10515 DAG.getConstant(MulAmt1, VT));
10517 if (isPowerOf2_64(MulAmt2))
10518 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
10519 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
10521 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
10522 DAG.getConstant(MulAmt2, VT));
10524 // Do not add new nodes to DAG combiner worklist.
10525 DCI.CombineTo(N, NewMul, false);
10530 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
10531 SDValue N0 = N->getOperand(0);
10532 SDValue N1 = N->getOperand(1);
10533 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
10534 EVT VT = N0.getValueType();
10536 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
10537 // since the result of setcc_c is all zero's or all ones.
10538 if (N1C && N0.getOpcode() == ISD::AND &&
10539 N0.getOperand(1).getOpcode() == ISD::Constant) {
10540 SDValue N00 = N0.getOperand(0);
10541 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
10542 ((N00.getOpcode() == ISD::ANY_EXTEND ||
10543 N00.getOpcode() == ISD::ZERO_EXTEND) &&
10544 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
10545 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
10546 APInt ShAmt = N1C->getAPIntValue();
10547 Mask = Mask.shl(ShAmt);
10549 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
10550 N00, DAG.getConstant(Mask, VT));
10557 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
10559 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
10560 const X86Subtarget *Subtarget) {
10561 EVT VT = N->getValueType(0);
10562 if (!VT.isVector() && VT.isInteger() &&
10563 N->getOpcode() == ISD::SHL)
10564 return PerformSHLCombine(N, DAG);
10566 // On X86 with SSE2 support, we can transform this to a vector shift if
10567 // all elements are shifted by the same amount. We can't do this in legalize
10568 // because the a constant vector is typically transformed to a constant pool
10569 // so we have no knowledge of the shift amount.
10570 if (!Subtarget->hasSSE2())
10573 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
10576 SDValue ShAmtOp = N->getOperand(1);
10577 EVT EltVT = VT.getVectorElementType();
10578 DebugLoc DL = N->getDebugLoc();
10579 SDValue BaseShAmt = SDValue();
10580 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
10581 unsigned NumElts = VT.getVectorNumElements();
10583 for (; i != NumElts; ++i) {
10584 SDValue Arg = ShAmtOp.getOperand(i);
10585 if (Arg.getOpcode() == ISD::UNDEF) continue;
10589 for (; i != NumElts; ++i) {
10590 SDValue Arg = ShAmtOp.getOperand(i);
10591 if (Arg.getOpcode() == ISD::UNDEF) continue;
10592 if (Arg != BaseShAmt) {
10596 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
10597 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
10598 SDValue InVec = ShAmtOp.getOperand(0);
10599 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
10600 unsigned NumElts = InVec.getValueType().getVectorNumElements();
10602 for (; i != NumElts; ++i) {
10603 SDValue Arg = InVec.getOperand(i);
10604 if (Arg.getOpcode() == ISD::UNDEF) continue;
10608 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
10609 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
10610 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
10611 if (C->getZExtValue() == SplatIdx)
10612 BaseShAmt = InVec.getOperand(1);
10615 if (BaseShAmt.getNode() == 0)
10616 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
10617 DAG.getIntPtrConstant(0));
10621 // The shift amount is an i32.
10622 if (EltVT.bitsGT(MVT::i32))
10623 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
10624 else if (EltVT.bitsLT(MVT::i32))
10625 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
10627 // The shift amount is identical so we can do a vector shift.
10628 SDValue ValOp = N->getOperand(0);
10629 switch (N->getOpcode()) {
10631 llvm_unreachable("Unknown shift opcode!");
10634 if (VT == MVT::v2i64)
10635 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10636 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
10638 if (VT == MVT::v4i32)
10639 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10640 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
10642 if (VT == MVT::v8i16)
10643 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10644 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
10648 if (VT == MVT::v4i32)
10649 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10650 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
10652 if (VT == MVT::v8i16)
10653 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10654 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
10658 if (VT == MVT::v2i64)
10659 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10660 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
10662 if (VT == MVT::v4i32)
10663 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10664 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
10666 if (VT == MVT::v8i16)
10667 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10668 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
10675 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
10676 TargetLowering::DAGCombinerInfo &DCI,
10677 const X86Subtarget *Subtarget) {
10678 if (DCI.isBeforeLegalizeOps())
10681 EVT VT = N->getValueType(0);
10682 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
10685 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
10686 SDValue N0 = N->getOperand(0);
10687 SDValue N1 = N->getOperand(1);
10688 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
10690 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
10692 if (!N0.hasOneUse() || !N1.hasOneUse())
10695 SDValue ShAmt0 = N0.getOperand(1);
10696 if (ShAmt0.getValueType() != MVT::i8)
10698 SDValue ShAmt1 = N1.getOperand(1);
10699 if (ShAmt1.getValueType() != MVT::i8)
10701 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
10702 ShAmt0 = ShAmt0.getOperand(0);
10703 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
10704 ShAmt1 = ShAmt1.getOperand(0);
10706 DebugLoc DL = N->getDebugLoc();
10707 unsigned Opc = X86ISD::SHLD;
10708 SDValue Op0 = N0.getOperand(0);
10709 SDValue Op1 = N1.getOperand(0);
10710 if (ShAmt0.getOpcode() == ISD::SUB) {
10711 Opc = X86ISD::SHRD;
10712 std::swap(Op0, Op1);
10713 std::swap(ShAmt0, ShAmt1);
10716 unsigned Bits = VT.getSizeInBits();
10717 if (ShAmt1.getOpcode() == ISD::SUB) {
10718 SDValue Sum = ShAmt1.getOperand(0);
10719 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
10720 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
10721 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
10722 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
10723 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
10724 return DAG.getNode(Opc, DL, VT,
10726 DAG.getNode(ISD::TRUNCATE, DL,
10729 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
10730 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
10732 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
10733 return DAG.getNode(Opc, DL, VT,
10734 N0.getOperand(0), N1.getOperand(0),
10735 DAG.getNode(ISD::TRUNCATE, DL,
10742 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
10743 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
10744 const X86Subtarget *Subtarget) {
10745 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
10746 // the FP state in cases where an emms may be missing.
10747 // A preferable solution to the general problem is to figure out the right
10748 // places to insert EMMS. This qualifies as a quick hack.
10750 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
10751 StoreSDNode *St = cast<StoreSDNode>(N);
10752 EVT VT = St->getValue().getValueType();
10753 if (VT.getSizeInBits() != 64)
10756 const Function *F = DAG.getMachineFunction().getFunction();
10757 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
10758 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
10759 && Subtarget->hasSSE2();
10760 if ((VT.isVector() ||
10761 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
10762 isa<LoadSDNode>(St->getValue()) &&
10763 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
10764 St->getChain().hasOneUse() && !St->isVolatile()) {
10765 SDNode* LdVal = St->getValue().getNode();
10766 LoadSDNode *Ld = 0;
10767 int TokenFactorIndex = -1;
10768 SmallVector<SDValue, 8> Ops;
10769 SDNode* ChainVal = St->getChain().getNode();
10770 // Must be a store of a load. We currently handle two cases: the load
10771 // is a direct child, and it's under an intervening TokenFactor. It is
10772 // possible to dig deeper under nested TokenFactors.
10773 if (ChainVal == LdVal)
10774 Ld = cast<LoadSDNode>(St->getChain());
10775 else if (St->getValue().hasOneUse() &&
10776 ChainVal->getOpcode() == ISD::TokenFactor) {
10777 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
10778 if (ChainVal->getOperand(i).getNode() == LdVal) {
10779 TokenFactorIndex = i;
10780 Ld = cast<LoadSDNode>(St->getValue());
10782 Ops.push_back(ChainVal->getOperand(i));
10786 if (!Ld || !ISD::isNormalLoad(Ld))
10789 // If this is not the MMX case, i.e. we are just turning i64 load/store
10790 // into f64 load/store, avoid the transformation if there are multiple
10791 // uses of the loaded value.
10792 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
10795 DebugLoc LdDL = Ld->getDebugLoc();
10796 DebugLoc StDL = N->getDebugLoc();
10797 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
10798 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
10800 if (Subtarget->is64Bit() || F64IsLegal) {
10801 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
10802 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
10803 Ld->getPointerInfo(), Ld->isVolatile(),
10804 Ld->isNonTemporal(), Ld->getAlignment());
10805 SDValue NewChain = NewLd.getValue(1);
10806 if (TokenFactorIndex != -1) {
10807 Ops.push_back(NewChain);
10808 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10811 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
10812 St->getPointerInfo(),
10813 St->isVolatile(), St->isNonTemporal(),
10814 St->getAlignment());
10817 // Otherwise, lower to two pairs of 32-bit loads / stores.
10818 SDValue LoAddr = Ld->getBasePtr();
10819 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
10820 DAG.getConstant(4, MVT::i32));
10822 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
10823 Ld->getPointerInfo(),
10824 Ld->isVolatile(), Ld->isNonTemporal(),
10825 Ld->getAlignment());
10826 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
10827 Ld->getPointerInfo().getWithOffset(4),
10828 Ld->isVolatile(), Ld->isNonTemporal(),
10829 MinAlign(Ld->getAlignment(), 4));
10831 SDValue NewChain = LoLd.getValue(1);
10832 if (TokenFactorIndex != -1) {
10833 Ops.push_back(LoLd);
10834 Ops.push_back(HiLd);
10835 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10839 LoAddr = St->getBasePtr();
10840 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
10841 DAG.getConstant(4, MVT::i32));
10843 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
10844 St->getPointerInfo(),
10845 St->isVolatile(), St->isNonTemporal(),
10846 St->getAlignment());
10847 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
10848 St->getPointerInfo().getWithOffset(4),
10850 St->isNonTemporal(),
10851 MinAlign(St->getAlignment(), 4));
10852 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
10857 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
10858 /// X86ISD::FXOR nodes.
10859 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
10860 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
10861 // F[X]OR(0.0, x) -> x
10862 // F[X]OR(x, 0.0) -> x
10863 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10864 if (C->getValueAPF().isPosZero())
10865 return N->getOperand(1);
10866 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10867 if (C->getValueAPF().isPosZero())
10868 return N->getOperand(0);
10872 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
10873 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
10874 // FAND(0.0, x) -> 0.0
10875 // FAND(x, 0.0) -> 0.0
10876 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10877 if (C->getValueAPF().isPosZero())
10878 return N->getOperand(0);
10879 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10880 if (C->getValueAPF().isPosZero())
10881 return N->getOperand(1);
10885 static SDValue PerformBTCombine(SDNode *N,
10887 TargetLowering::DAGCombinerInfo &DCI) {
10888 // BT ignores high bits in the bit index operand.
10889 SDValue Op1 = N->getOperand(1);
10890 if (Op1.hasOneUse()) {
10891 unsigned BitWidth = Op1.getValueSizeInBits();
10892 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
10893 APInt KnownZero, KnownOne;
10894 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
10895 !DCI.isBeforeLegalizeOps());
10896 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10897 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
10898 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
10899 DCI.CommitTargetLoweringOpt(TLO);
10904 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
10905 SDValue Op = N->getOperand(0);
10906 if (Op.getOpcode() == ISD::BIT_CONVERT)
10907 Op = Op.getOperand(0);
10908 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
10909 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
10910 VT.getVectorElementType().getSizeInBits() ==
10911 OpVT.getVectorElementType().getSizeInBits()) {
10912 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
10917 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
10918 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
10919 // (and (i32 x86isd::setcc_carry), 1)
10920 // This eliminates the zext. This transformation is necessary because
10921 // ISD::SETCC is always legalized to i8.
10922 DebugLoc dl = N->getDebugLoc();
10923 SDValue N0 = N->getOperand(0);
10924 EVT VT = N->getValueType(0);
10925 if (N0.getOpcode() == ISD::AND &&
10927 N0.getOperand(0).hasOneUse()) {
10928 SDValue N00 = N0.getOperand(0);
10929 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
10931 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
10932 if (!C || C->getZExtValue() != 1)
10934 return DAG.getNode(ISD::AND, dl, VT,
10935 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
10936 N00.getOperand(0), N00.getOperand(1)),
10937 DAG.getConstant(1, VT));
10943 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
10944 DAGCombinerInfo &DCI) const {
10945 SelectionDAG &DAG = DCI.DAG;
10946 switch (N->getOpcode()) {
10948 case ISD::EXTRACT_VECTOR_ELT:
10949 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
10950 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
10951 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
10952 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
10955 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
10956 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
10957 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
10959 case X86ISD::FOR: return PerformFORCombine(N, DAG);
10960 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
10961 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
10962 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
10963 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
10964 case X86ISD::SHUFPS: // Handle all target specific shuffles
10965 case X86ISD::SHUFPD:
10966 case X86ISD::PALIGN:
10967 case X86ISD::PUNPCKHBW:
10968 case X86ISD::PUNPCKHWD:
10969 case X86ISD::PUNPCKHDQ:
10970 case X86ISD::PUNPCKHQDQ:
10971 case X86ISD::UNPCKHPS:
10972 case X86ISD::UNPCKHPD:
10973 case X86ISD::PUNPCKLBW:
10974 case X86ISD::PUNPCKLWD:
10975 case X86ISD::PUNPCKLDQ:
10976 case X86ISD::PUNPCKLQDQ:
10977 case X86ISD::UNPCKLPS:
10978 case X86ISD::UNPCKLPD:
10979 case X86ISD::MOVHLPS:
10980 case X86ISD::MOVLHPS:
10981 case X86ISD::PSHUFD:
10982 case X86ISD::PSHUFHW:
10983 case X86ISD::PSHUFLW:
10984 case X86ISD::MOVSS:
10985 case X86ISD::MOVSD:
10986 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
10992 /// isTypeDesirableForOp - Return true if the target has native support for
10993 /// the specified value type and it is 'desirable' to use the type for the
10994 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
10995 /// instruction encodings are longer and some i16 instructions are slow.
10996 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
10997 if (!isTypeLegal(VT))
10999 if (VT != MVT::i16)
11006 case ISD::SIGN_EXTEND:
11007 case ISD::ZERO_EXTEND:
11008 case ISD::ANY_EXTEND:
11021 /// IsDesirableToPromoteOp - This method query the target whether it is
11022 /// beneficial for dag combiner to promote the specified node. If true, it
11023 /// should return the desired promotion type by reference.
11024 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
11025 EVT VT = Op.getValueType();
11026 if (VT != MVT::i16)
11029 bool Promote = false;
11030 bool Commute = false;
11031 switch (Op.getOpcode()) {
11034 LoadSDNode *LD = cast<LoadSDNode>(Op);
11035 // If the non-extending load has a single use and it's not live out, then it
11036 // might be folded.
11037 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
11038 Op.hasOneUse()*/) {
11039 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
11040 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
11041 // The only case where we'd want to promote LOAD (rather then it being
11042 // promoted as an operand is when it's only use is liveout.
11043 if (UI->getOpcode() != ISD::CopyToReg)
11050 case ISD::SIGN_EXTEND:
11051 case ISD::ZERO_EXTEND:
11052 case ISD::ANY_EXTEND:
11057 SDValue N0 = Op.getOperand(0);
11058 // Look out for (store (shl (load), x)).
11059 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
11072 SDValue N0 = Op.getOperand(0);
11073 SDValue N1 = Op.getOperand(1);
11074 if (!Commute && MayFoldLoad(N1))
11076 // Avoid disabling potential load folding opportunities.
11077 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
11079 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
11089 //===----------------------------------------------------------------------===//
11090 // X86 Inline Assembly Support
11091 //===----------------------------------------------------------------------===//
11093 static bool LowerToBSwap(CallInst *CI) {
11094 // FIXME: this should verify that we are targetting a 486 or better. If not,
11095 // we will turn this bswap into something that will be lowered to logical ops
11096 // instead of emitting the bswap asm. For now, we don't support 486 or lower
11097 // so don't worry about this.
11099 // Verify this is a simple bswap.
11100 if (CI->getNumArgOperands() != 1 ||
11101 CI->getType() != CI->getArgOperand(0)->getType() ||
11102 !CI->getType()->isIntegerTy())
11105 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
11106 if (!Ty || Ty->getBitWidth() % 16 != 0)
11109 // Okay, we can do this xform, do so now.
11110 const Type *Tys[] = { Ty };
11111 Module *M = CI->getParent()->getParent()->getParent();
11112 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
11114 Value *Op = CI->getArgOperand(0);
11115 Op = CallInst::Create(Int, Op, CI->getName(), CI);
11117 CI->replaceAllUsesWith(Op);
11118 CI->eraseFromParent();
11122 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
11123 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
11124 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
11126 std::string AsmStr = IA->getAsmString();
11128 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
11129 SmallVector<StringRef, 4> AsmPieces;
11130 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
11132 switch (AsmPieces.size()) {
11133 default: return false;
11135 AsmStr = AsmPieces[0];
11137 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
11140 if (AsmPieces.size() == 2 &&
11141 (AsmPieces[0] == "bswap" ||
11142 AsmPieces[0] == "bswapq" ||
11143 AsmPieces[0] == "bswapl") &&
11144 (AsmPieces[1] == "$0" ||
11145 AsmPieces[1] == "${0:q}")) {
11146 // No need to check constraints, nothing other than the equivalent of
11147 // "=r,0" would be valid here.
11148 return LowerToBSwap(CI);
11150 // rorw $$8, ${0:w} --> llvm.bswap.i16
11151 if (CI->getType()->isIntegerTy(16) &&
11152 AsmPieces.size() == 3 &&
11153 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
11154 AsmPieces[1] == "$$8," &&
11155 AsmPieces[2] == "${0:w}" &&
11156 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
11158 const std::string &Constraints = IA->getConstraintString();
11159 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
11160 std::sort(AsmPieces.begin(), AsmPieces.end());
11161 if (AsmPieces.size() == 4 &&
11162 AsmPieces[0] == "~{cc}" &&
11163 AsmPieces[1] == "~{dirflag}" &&
11164 AsmPieces[2] == "~{flags}" &&
11165 AsmPieces[3] == "~{fpsr}") {
11166 return LowerToBSwap(CI);
11171 if (CI->getType()->isIntegerTy(64) &&
11172 Constraints.size() >= 2 &&
11173 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
11174 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
11175 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
11176 SmallVector<StringRef, 4> Words;
11177 SplitString(AsmPieces[0], Words, " \t");
11178 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
11180 SplitString(AsmPieces[1], Words, " \t");
11181 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
11183 SplitString(AsmPieces[2], Words, " \t,");
11184 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
11185 Words[2] == "%edx") {
11186 return LowerToBSwap(CI);
11198 /// getConstraintType - Given a constraint letter, return the type of
11199 /// constraint it is for this target.
11200 X86TargetLowering::ConstraintType
11201 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
11202 if (Constraint.size() == 1) {
11203 switch (Constraint[0]) {
11215 return C_RegisterClass;
11223 return TargetLowering::getConstraintType(Constraint);
11226 /// Examine constraint type and operand type and determine a weight value,
11227 /// where: -1 = invalid match, and 0 = so-so match to 3 = good match.
11228 /// This object must already have been set up with the operand type
11229 /// and the current alternative constraint selected.
11230 int X86TargetLowering::getSingleConstraintMatchWeight(
11231 AsmOperandInfo &info, const char *constraint) const {
11233 Value *CallOperandVal = info.CallOperandVal;
11234 // If we don't have a value, we can't do a match,
11235 // but allow it at the lowest weight.
11236 if (CallOperandVal == NULL)
11238 // Look at the constraint type.
11239 switch (*constraint) {
11241 return TargetLowering::getSingleConstraintMatchWeight(info, constraint);
11244 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
11245 if (C->getZExtValue() <= 31)
11254 /// LowerXConstraint - try to replace an X constraint, which matches anything,
11255 /// with another that has more specific requirements based on the type of the
11256 /// corresponding operand.
11257 const char *X86TargetLowering::
11258 LowerXConstraint(EVT ConstraintVT) const {
11259 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
11260 // 'f' like normal targets.
11261 if (ConstraintVT.isFloatingPoint()) {
11262 if (Subtarget->hasSSE2())
11264 if (Subtarget->hasSSE1())
11268 return TargetLowering::LowerXConstraint(ConstraintVT);
11271 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
11272 /// vector. If it is invalid, don't add anything to Ops.
11273 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
11275 std::vector<SDValue>&Ops,
11276 SelectionDAG &DAG) const {
11277 SDValue Result(0, 0);
11279 switch (Constraint) {
11282 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11283 if (C->getZExtValue() <= 31) {
11284 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11290 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11291 if (C->getZExtValue() <= 63) {
11292 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11298 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11299 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
11300 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11306 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11307 if (C->getZExtValue() <= 255) {
11308 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11314 // 32-bit signed value
11315 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11316 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
11317 C->getSExtValue())) {
11318 // Widen to 64 bits here to get it sign extended.
11319 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
11322 // FIXME gcc accepts some relocatable values here too, but only in certain
11323 // memory models; it's complicated.
11328 // 32-bit unsigned value
11329 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11330 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
11331 C->getZExtValue())) {
11332 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11336 // FIXME gcc accepts some relocatable values here too, but only in certain
11337 // memory models; it's complicated.
11341 // Literal immediates are always ok.
11342 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
11343 // Widen to 64 bits here to get it sign extended.
11344 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
11348 // In any sort of PIC mode addresses need to be computed at runtime by
11349 // adding in a register or some sort of table lookup. These can't
11350 // be used as immediates.
11351 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
11354 // If we are in non-pic codegen mode, we allow the address of a global (with
11355 // an optional displacement) to be used with 'i'.
11356 GlobalAddressSDNode *GA = 0;
11357 int64_t Offset = 0;
11359 // Match either (GA), (GA+C), (GA+C1+C2), etc.
11361 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
11362 Offset += GA->getOffset();
11364 } else if (Op.getOpcode() == ISD::ADD) {
11365 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
11366 Offset += C->getZExtValue();
11367 Op = Op.getOperand(0);
11370 } else if (Op.getOpcode() == ISD::SUB) {
11371 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
11372 Offset += -C->getZExtValue();
11373 Op = Op.getOperand(0);
11378 // Otherwise, this isn't something we can handle, reject it.
11382 const GlobalValue *GV = GA->getGlobal();
11383 // If we require an extra load to get this address, as in PIC mode, we
11384 // can't accept it.
11385 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
11386 getTargetMachine())))
11389 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
11390 GA->getValueType(0), Offset);
11395 if (Result.getNode()) {
11396 Ops.push_back(Result);
11399 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11402 std::vector<unsigned> X86TargetLowering::
11403 getRegClassForInlineAsmConstraint(const std::string &Constraint,
11405 if (Constraint.size() == 1) {
11406 // FIXME: not handling fp-stack yet!
11407 switch (Constraint[0]) { // GCC X86 Constraint Letters
11408 default: break; // Unknown constraint letter
11409 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
11410 if (Subtarget->is64Bit()) {
11411 if (VT == MVT::i32)
11412 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
11413 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
11414 X86::R10D,X86::R11D,X86::R12D,
11415 X86::R13D,X86::R14D,X86::R15D,
11416 X86::EBP, X86::ESP, 0);
11417 else if (VT == MVT::i16)
11418 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
11419 X86::SI, X86::DI, X86::R8W,X86::R9W,
11420 X86::R10W,X86::R11W,X86::R12W,
11421 X86::R13W,X86::R14W,X86::R15W,
11422 X86::BP, X86::SP, 0);
11423 else if (VT == MVT::i8)
11424 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
11425 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
11426 X86::R10B,X86::R11B,X86::R12B,
11427 X86::R13B,X86::R14B,X86::R15B,
11428 X86::BPL, X86::SPL, 0);
11430 else if (VT == MVT::i64)
11431 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
11432 X86::RSI, X86::RDI, X86::R8, X86::R9,
11433 X86::R10, X86::R11, X86::R12,
11434 X86::R13, X86::R14, X86::R15,
11435 X86::RBP, X86::RSP, 0);
11439 // 32-bit fallthrough
11440 case 'Q': // Q_REGS
11441 if (VT == MVT::i32)
11442 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
11443 else if (VT == MVT::i16)
11444 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
11445 else if (VT == MVT::i8)
11446 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
11447 else if (VT == MVT::i64)
11448 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
11453 return std::vector<unsigned>();
11456 std::pair<unsigned, const TargetRegisterClass*>
11457 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
11459 // First, see if this is a constraint that directly corresponds to an LLVM
11461 if (Constraint.size() == 1) {
11462 // GCC Constraint Letters
11463 switch (Constraint[0]) {
11465 case 'r': // GENERAL_REGS
11466 case 'l': // INDEX_REGS
11468 return std::make_pair(0U, X86::GR8RegisterClass);
11469 if (VT == MVT::i16)
11470 return std::make_pair(0U, X86::GR16RegisterClass);
11471 if (VT == MVT::i32 || !Subtarget->is64Bit())
11472 return std::make_pair(0U, X86::GR32RegisterClass);
11473 return std::make_pair(0U, X86::GR64RegisterClass);
11474 case 'R': // LEGACY_REGS
11476 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
11477 if (VT == MVT::i16)
11478 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
11479 if (VT == MVT::i32 || !Subtarget->is64Bit())
11480 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
11481 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
11482 case 'f': // FP Stack registers.
11483 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
11484 // value to the correct fpstack register class.
11485 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
11486 return std::make_pair(0U, X86::RFP32RegisterClass);
11487 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
11488 return std::make_pair(0U, X86::RFP64RegisterClass);
11489 return std::make_pair(0U, X86::RFP80RegisterClass);
11490 case 'y': // MMX_REGS if MMX allowed.
11491 if (!Subtarget->hasMMX()) break;
11492 return std::make_pair(0U, X86::VR64RegisterClass);
11493 case 'Y': // SSE_REGS if SSE2 allowed
11494 if (!Subtarget->hasSSE2()) break;
11496 case 'x': // SSE_REGS if SSE1 allowed
11497 if (!Subtarget->hasSSE1()) break;
11499 switch (VT.getSimpleVT().SimpleTy) {
11501 // Scalar SSE types.
11504 return std::make_pair(0U, X86::FR32RegisterClass);
11507 return std::make_pair(0U, X86::FR64RegisterClass);
11515 return std::make_pair(0U, X86::VR128RegisterClass);
11521 // Use the default implementation in TargetLowering to convert the register
11522 // constraint into a member of a register class.
11523 std::pair<unsigned, const TargetRegisterClass*> Res;
11524 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
11526 // Not found as a standard register?
11527 if (Res.second == 0) {
11528 // Map st(0) -> st(7) -> ST0
11529 if (Constraint.size() == 7 && Constraint[0] == '{' &&
11530 tolower(Constraint[1]) == 's' &&
11531 tolower(Constraint[2]) == 't' &&
11532 Constraint[3] == '(' &&
11533 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
11534 Constraint[5] == ')' &&
11535 Constraint[6] == '}') {
11537 Res.first = X86::ST0+Constraint[4]-'0';
11538 Res.second = X86::RFP80RegisterClass;
11542 // GCC allows "st(0)" to be called just plain "st".
11543 if (StringRef("{st}").equals_lower(Constraint)) {
11544 Res.first = X86::ST0;
11545 Res.second = X86::RFP80RegisterClass;
11550 if (StringRef("{flags}").equals_lower(Constraint)) {
11551 Res.first = X86::EFLAGS;
11552 Res.second = X86::CCRRegisterClass;
11556 // 'A' means EAX + EDX.
11557 if (Constraint == "A") {
11558 Res.first = X86::EAX;
11559 Res.second = X86::GR32_ADRegisterClass;
11565 // Otherwise, check to see if this is a register class of the wrong value
11566 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
11567 // turn into {ax},{dx}.
11568 if (Res.second->hasType(VT))
11569 return Res; // Correct type already, nothing to do.
11571 // All of the single-register GCC register classes map their values onto
11572 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
11573 // really want an 8-bit or 32-bit register, map to the appropriate register
11574 // class and return the appropriate register.
11575 if (Res.second == X86::GR16RegisterClass) {
11576 if (VT == MVT::i8) {
11577 unsigned DestReg = 0;
11578 switch (Res.first) {
11580 case X86::AX: DestReg = X86::AL; break;
11581 case X86::DX: DestReg = X86::DL; break;
11582 case X86::CX: DestReg = X86::CL; break;
11583 case X86::BX: DestReg = X86::BL; break;
11586 Res.first = DestReg;
11587 Res.second = X86::GR8RegisterClass;
11589 } else if (VT == MVT::i32) {
11590 unsigned DestReg = 0;
11591 switch (Res.first) {
11593 case X86::AX: DestReg = X86::EAX; break;
11594 case X86::DX: DestReg = X86::EDX; break;
11595 case X86::CX: DestReg = X86::ECX; break;
11596 case X86::BX: DestReg = X86::EBX; break;
11597 case X86::SI: DestReg = X86::ESI; break;
11598 case X86::DI: DestReg = X86::EDI; break;
11599 case X86::BP: DestReg = X86::EBP; break;
11600 case X86::SP: DestReg = X86::ESP; break;
11603 Res.first = DestReg;
11604 Res.second = X86::GR32RegisterClass;
11606 } else if (VT == MVT::i64) {
11607 unsigned DestReg = 0;
11608 switch (Res.first) {
11610 case X86::AX: DestReg = X86::RAX; break;
11611 case X86::DX: DestReg = X86::RDX; break;
11612 case X86::CX: DestReg = X86::RCX; break;
11613 case X86::BX: DestReg = X86::RBX; break;
11614 case X86::SI: DestReg = X86::RSI; break;
11615 case X86::DI: DestReg = X86::RDI; break;
11616 case X86::BP: DestReg = X86::RBP; break;
11617 case X86::SP: DestReg = X86::RSP; break;
11620 Res.first = DestReg;
11621 Res.second = X86::GR64RegisterClass;
11624 } else if (Res.second == X86::FR32RegisterClass ||
11625 Res.second == X86::FR64RegisterClass ||
11626 Res.second == X86::VR128RegisterClass) {
11627 // Handle references to XMM physical registers that got mapped into the
11628 // wrong class. This can happen with constraints like {xmm0} where the
11629 // target independent register mapper will just pick the first match it can
11630 // find, ignoring the required type.
11631 if (VT == MVT::f32)
11632 Res.second = X86::FR32RegisterClass;
11633 else if (VT == MVT::f64)
11634 Res.second = X86::FR64RegisterClass;
11635 else if (X86::VR128RegisterClass->hasType(VT))
11636 Res.second = X86::VR128RegisterClass;