1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalAlias.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/LLVMContext.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/Dwarf.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/raw_ostream.h"
53 using namespace dwarf;
55 STATISTIC(NumTailCalls, "Number of tail calls");
58 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
60 // Forward declarations.
61 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
64 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
66 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
68 if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
69 if (is64Bit) return new X8664_MachoTargetObjectFile();
70 return new TargetLoweringObjectFileMachO();
71 } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
72 if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
73 return new X8632_ELFTargetObjectFile(TM);
74 } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
75 return new TargetLoweringObjectFileCOFF();
77 llvm_unreachable("unknown subtarget type");
80 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
81 : TargetLowering(TM, createTLOF(TM)) {
82 Subtarget = &TM.getSubtarget<X86Subtarget>();
83 X86ScalarSSEf64 = Subtarget->hasSSE2();
84 X86ScalarSSEf32 = Subtarget->hasSSE1();
85 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
87 RegInfo = TM.getRegisterInfo();
90 // Set up the TargetLowering object.
92 // X86 is weird, it always uses i8 for shift amounts and setcc results.
93 setShiftAmountType(MVT::i8);
94 setBooleanContents(ZeroOrOneBooleanContent);
95 setSchedulingPreference(Sched::RegPressure);
96 setStackPointerRegisterToSaveRestore(X86StackPtr);
98 if (Subtarget->isTargetDarwin()) {
99 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
100 setUseUnderscoreSetJmp(false);
101 setUseUnderscoreLongJmp(false);
102 } else if (Subtarget->isTargetMingw()) {
103 // MS runtime is weird: it exports _setjmp, but longjmp!
104 setUseUnderscoreSetJmp(true);
105 setUseUnderscoreLongJmp(false);
107 setUseUnderscoreSetJmp(true);
108 setUseUnderscoreLongJmp(true);
111 // Set up the register classes.
112 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
113 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
114 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
115 if (Subtarget->is64Bit())
116 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
118 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
120 // We don't accept any truncstore of integer registers.
121 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
122 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
123 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
124 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
125 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
128 // SETOEQ and SETUNE require checking two conditions.
129 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
130 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
131 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
132 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
133 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
136 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
138 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
139 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
140 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
142 if (Subtarget->is64Bit()) {
143 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
144 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
145 } else if (!UseSoftFloat) {
146 // We have an algorithm for SSE2->double, and we turn this into a
147 // 64-bit FILD followed by conditional FADD for other targets.
148 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
149 // We have an algorithm for SSE2, and we turn this into a 64-bit
150 // FILD for other targets.
151 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
154 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
156 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
157 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
160 // SSE has no i16 to fp conversion, only i32
161 if (X86ScalarSSEf32) {
162 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
163 // f32 and f64 cases are Legal, f80 case is not
164 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
166 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
167 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
170 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
171 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
174 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
175 // are Legal, f80 is custom lowered.
176 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
179 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
181 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
182 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
184 if (X86ScalarSSEf32) {
185 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
186 // f32 and f64 cases are Legal, f80 case is not
187 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
189 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
190 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
193 // Handle FP_TO_UINT by promoting the destination to a larger signed
195 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
196 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
197 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
199 if (Subtarget->is64Bit()) {
200 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
201 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
202 } else if (!UseSoftFloat) {
203 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
204 // Expand FP_TO_UINT into a select.
205 // FIXME: We would like to use a Custom expander here eventually to do
206 // the optimal thing for SSE vs. the default expansion in the legalizer.
207 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
209 // With SSE3 we can use fisttpll to convert to a signed i64; without
210 // SSE, we're stuck with a fistpll.
211 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
214 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
215 if (!X86ScalarSSEf64) {
216 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
217 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
218 if (Subtarget->is64Bit()) {
219 setOperationAction(ISD::BIT_CONVERT , MVT::f64 , Expand);
220 // Without SSE, i64->f64 goes through memory; i64->MMX is Legal.
221 if (Subtarget->hasMMX() && !DisableMMX)
222 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Custom);
224 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Expand);
228 // Scalar integer divide and remainder are lowered to use operations that
229 // produce two results, to match the available instructions. This exposes
230 // the two-result form to trivial CSE, which is able to combine x/y and x%y
231 // into a single instruction.
233 // Scalar integer multiply-high is also lowered to use two-result
234 // operations, to match the available instructions. However, plain multiply
235 // (low) operations are left as Legal, as there are single-result
236 // instructions for this in x86. Using the two-result multiply instructions
237 // when both high and low results are needed must be arranged by dagcombine.
238 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
239 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
240 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
241 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
242 setOperationAction(ISD::SREM , MVT::i8 , Expand);
243 setOperationAction(ISD::UREM , MVT::i8 , Expand);
244 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
245 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
246 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
247 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
248 setOperationAction(ISD::SREM , MVT::i16 , Expand);
249 setOperationAction(ISD::UREM , MVT::i16 , Expand);
250 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
251 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
252 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
253 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
254 setOperationAction(ISD::SREM , MVT::i32 , Expand);
255 setOperationAction(ISD::UREM , MVT::i32 , Expand);
256 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
257 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
258 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
259 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
260 setOperationAction(ISD::SREM , MVT::i64 , Expand);
261 setOperationAction(ISD::UREM , MVT::i64 , Expand);
263 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
264 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
265 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
266 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
267 if (Subtarget->is64Bit())
268 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
269 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
270 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
272 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
273 setOperationAction(ISD::FREM , MVT::f32 , Expand);
274 setOperationAction(ISD::FREM , MVT::f64 , Expand);
275 setOperationAction(ISD::FREM , MVT::f80 , Expand);
276 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
278 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
279 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
280 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
281 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
282 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
283 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
287 if (Subtarget->is64Bit()) {
288 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
293 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
294 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
296 // These should be promoted to a larger select which is supported.
297 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
298 // X86 wants to expand cmov itself.
299 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
300 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
301 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
302 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
303 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
304 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
305 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
306 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
307 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
308 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
309 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
310 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
311 if (Subtarget->is64Bit()) {
312 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
313 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
315 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
318 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
319 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
320 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
321 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
322 if (Subtarget->is64Bit())
323 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
324 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
325 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
326 if (Subtarget->is64Bit()) {
327 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
328 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
329 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
330 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
331 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
333 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
334 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
335 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
336 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
337 if (Subtarget->is64Bit()) {
338 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
339 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
340 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
343 if (Subtarget->hasSSE1())
344 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
346 if (!Subtarget->hasSSE2())
347 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
348 // On X86 and X86-64, atomic operations are lowered to locked instructions.
349 // Locked instructions, in turn, have implicit fence semantics (all memory
350 // operations are flushed before issuing the locked instruction, and they
351 // are not buffered), so we can fold away the common pattern of
352 // fence-atomic-fence.
353 setShouldFoldAtomicFences(true);
355 // Expand certain atomics
356 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
361 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
366 if (!Subtarget->is64Bit()) {
367 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
376 // FIXME - use subtarget debug flags
377 if (!Subtarget->isTargetDarwin() &&
378 !Subtarget->isTargetELF() &&
379 !Subtarget->isTargetCygMing()) {
380 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
383 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
384 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
385 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
386 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
387 if (Subtarget->is64Bit()) {
388 setExceptionPointerRegister(X86::RAX);
389 setExceptionSelectorRegister(X86::RDX);
391 setExceptionPointerRegister(X86::EAX);
392 setExceptionSelectorRegister(X86::EDX);
394 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
395 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
397 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
399 setOperationAction(ISD::TRAP, MVT::Other, Legal);
401 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
402 setOperationAction(ISD::VASTART , MVT::Other, Custom);
403 setOperationAction(ISD::VAEND , MVT::Other, Expand);
404 if (Subtarget->is64Bit()) {
405 setOperationAction(ISD::VAARG , MVT::Other, Custom);
406 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
408 setOperationAction(ISD::VAARG , MVT::Other, Expand);
409 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
412 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
413 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
414 if (Subtarget->is64Bit())
415 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
416 if (Subtarget->isTargetCygMing())
417 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
419 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
421 if (!UseSoftFloat && X86ScalarSSEf64) {
422 // f32 and f64 use SSE.
423 // Set up the FP register classes.
424 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
425 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
427 // Use ANDPD to simulate FABS.
428 setOperationAction(ISD::FABS , MVT::f64, Custom);
429 setOperationAction(ISD::FABS , MVT::f32, Custom);
431 // Use XORP to simulate FNEG.
432 setOperationAction(ISD::FNEG , MVT::f64, Custom);
433 setOperationAction(ISD::FNEG , MVT::f32, Custom);
435 // Use ANDPD and ORPD to simulate FCOPYSIGN.
436 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
437 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
439 // We don't support sin/cos/fmod
440 setOperationAction(ISD::FSIN , MVT::f64, Expand);
441 setOperationAction(ISD::FCOS , MVT::f64, Expand);
442 setOperationAction(ISD::FSIN , MVT::f32, Expand);
443 setOperationAction(ISD::FCOS , MVT::f32, Expand);
445 // Expand FP immediates into loads from the stack, except for the special
447 addLegalFPImmediate(APFloat(+0.0)); // xorpd
448 addLegalFPImmediate(APFloat(+0.0f)); // xorps
449 } else if (!UseSoftFloat && X86ScalarSSEf32) {
450 // Use SSE for f32, x87 for f64.
451 // Set up the FP register classes.
452 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
453 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
455 // Use ANDPS to simulate FABS.
456 setOperationAction(ISD::FABS , MVT::f32, Custom);
458 // Use XORP to simulate FNEG.
459 setOperationAction(ISD::FNEG , MVT::f32, Custom);
461 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
463 // Use ANDPS and ORPS to simulate FCOPYSIGN.
464 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
465 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
467 // We don't support sin/cos/fmod
468 setOperationAction(ISD::FSIN , MVT::f32, Expand);
469 setOperationAction(ISD::FCOS , MVT::f32, Expand);
471 // Special cases we handle for FP constants.
472 addLegalFPImmediate(APFloat(+0.0f)); // xorps
473 addLegalFPImmediate(APFloat(+0.0)); // FLD0
474 addLegalFPImmediate(APFloat(+1.0)); // FLD1
475 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
476 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
479 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
480 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
482 } else if (!UseSoftFloat) {
483 // f32 and f64 in x87.
484 // Set up the FP register classes.
485 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
486 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
488 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
489 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
490 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
491 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
494 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
495 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
497 addLegalFPImmediate(APFloat(+0.0)); // FLD0
498 addLegalFPImmediate(APFloat(+1.0)); // FLD1
499 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
500 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
501 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
502 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
503 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
504 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
507 // Long double always uses X87.
509 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
510 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
511 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
514 APFloat TmpFlt(+0.0);
515 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
517 addLegalFPImmediate(TmpFlt); // FLD0
519 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
520 APFloat TmpFlt2(+1.0);
521 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
523 addLegalFPImmediate(TmpFlt2); // FLD1
524 TmpFlt2.changeSign();
525 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
529 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
530 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
534 // Always use a library call for pow.
535 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
536 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
537 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
539 setOperationAction(ISD::FLOG, MVT::f80, Expand);
540 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
541 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
542 setOperationAction(ISD::FEXP, MVT::f80, Expand);
543 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
545 // First set operation action for all vector types to either promote
546 // (for widening) or expand (for scalarization). Then we will selectively
547 // turn on ones that can be effectively codegen'd.
548 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
549 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
550 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
565 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
566 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
599 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
603 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
604 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
605 setTruncStoreAction((MVT::SimpleValueType)VT,
606 (MVT::SimpleValueType)InnerVT, Expand);
607 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
608 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
609 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
612 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
613 // with -msoft-float, disable use of MMX as well.
614 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
615 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
616 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
617 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
619 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
621 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
622 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
623 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
624 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
626 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
627 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
628 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
629 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
631 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
632 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
634 setOperationAction(ISD::AND, MVT::v8i8, Promote);
635 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
636 setOperationAction(ISD::AND, MVT::v4i16, Promote);
637 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
638 setOperationAction(ISD::AND, MVT::v2i32, Promote);
639 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
640 setOperationAction(ISD::AND, MVT::v1i64, Legal);
642 setOperationAction(ISD::OR, MVT::v8i8, Promote);
643 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
644 setOperationAction(ISD::OR, MVT::v4i16, Promote);
645 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
646 setOperationAction(ISD::OR, MVT::v2i32, Promote);
647 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
648 setOperationAction(ISD::OR, MVT::v1i64, Legal);
650 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
651 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
652 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
653 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
654 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
655 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
656 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
658 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
659 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
660 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
661 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
662 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
663 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
664 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
666 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
667 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
669 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
671 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
672 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
674 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
676 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
677 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
678 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
680 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
682 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
683 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
684 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
685 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
686 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
690 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
691 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
692 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
693 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
694 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
698 if (!UseSoftFloat && Subtarget->hasSSE1()) {
699 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
701 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
702 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
703 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
704 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
705 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
706 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
707 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
708 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
709 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
710 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
711 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
712 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
715 if (!UseSoftFloat && Subtarget->hasSSE2()) {
716 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
718 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
719 // registers cannot be used even for integer operations.
720 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
721 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
722 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
723 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
725 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
726 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
727 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
728 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
729 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
730 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
731 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
732 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
733 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
734 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
735 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
736 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
737 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
738 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
739 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
740 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
742 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
743 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
744 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
745 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
747 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
748 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
749 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
750 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
751 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
753 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
754 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
755 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
756 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
757 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
759 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
760 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
761 EVT VT = (MVT::SimpleValueType)i;
762 // Do not attempt to custom lower non-power-of-2 vectors
763 if (!isPowerOf2_32(VT.getVectorNumElements()))
765 // Do not attempt to custom lower non-128-bit vectors
766 if (!VT.is128BitVector())
768 setOperationAction(ISD::BUILD_VECTOR,
769 VT.getSimpleVT().SimpleTy, Custom);
770 setOperationAction(ISD::VECTOR_SHUFFLE,
771 VT.getSimpleVT().SimpleTy, Custom);
772 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
773 VT.getSimpleVT().SimpleTy, Custom);
776 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
777 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
778 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
779 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
780 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
781 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
783 if (Subtarget->is64Bit()) {
784 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
785 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
788 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
789 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
790 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
793 // Do not attempt to promote non-128-bit vectors
794 if (!VT.is128BitVector())
797 setOperationAction(ISD::AND, SVT, Promote);
798 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
799 setOperationAction(ISD::OR, SVT, Promote);
800 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
801 setOperationAction(ISD::XOR, SVT, Promote);
802 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
803 setOperationAction(ISD::LOAD, SVT, Promote);
804 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
805 setOperationAction(ISD::SELECT, SVT, Promote);
806 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
809 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
811 // Custom lower v2i64 and v2f64 selects.
812 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
813 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
814 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
815 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
817 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
818 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
819 if (!DisableMMX && Subtarget->hasMMX()) {
820 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
821 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
825 if (Subtarget->hasSSE41()) {
826 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
827 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
828 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
829 setOperationAction(ISD::FRINT, MVT::f32, Legal);
830 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
831 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
832 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
833 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
834 setOperationAction(ISD::FRINT, MVT::f64, Legal);
835 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
837 // FIXME: Do we need to handle scalar-to-vector here?
838 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
840 // i8 and i16 vectors are custom , because the source register and source
841 // source memory operand types are not the same width. f32 vectors are
842 // custom since the immediate controlling the insert encodes additional
844 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
849 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
851 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
852 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
854 if (Subtarget->is64Bit()) {
855 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
856 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
860 if (Subtarget->hasSSE42()) {
861 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
864 if (!UseSoftFloat && Subtarget->hasAVX()) {
865 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
866 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
867 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
868 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
870 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
871 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
872 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
873 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
874 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
875 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
876 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
877 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
878 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
879 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
880 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
881 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
882 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
883 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
884 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
886 // Operations to consider commented out -v16i16 v32i8
887 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
888 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
889 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
890 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
891 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
892 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
893 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
894 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
895 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
896 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
897 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
898 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
899 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
900 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
902 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
903 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
904 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
905 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
907 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
908 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
909 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
910 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
913 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
914 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
915 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
916 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
917 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
918 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
921 // Not sure we want to do this since there are no 256-bit integer
924 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
925 // This includes 256-bit vectors
926 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
927 EVT VT = (MVT::SimpleValueType)i;
929 // Do not attempt to custom lower non-power-of-2 vectors
930 if (!isPowerOf2_32(VT.getVectorNumElements()))
933 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
934 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
935 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
938 if (Subtarget->is64Bit()) {
939 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
940 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
945 // Not sure we want to do this since there are no 256-bit integer
948 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
949 // Including 256-bit vectors
950 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
951 EVT VT = (MVT::SimpleValueType)i;
953 if (!VT.is256BitVector()) {
956 setOperationAction(ISD::AND, VT, Promote);
957 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
958 setOperationAction(ISD::OR, VT, Promote);
959 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
960 setOperationAction(ISD::XOR, VT, Promote);
961 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
962 setOperationAction(ISD::LOAD, VT, Promote);
963 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
964 setOperationAction(ISD::SELECT, VT, Promote);
965 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
968 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
972 // We want to custom lower some of our intrinsics.
973 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
975 // Add/Sub/Mul with overflow operations are custom lowered.
976 setOperationAction(ISD::SADDO, MVT::i32, Custom);
977 setOperationAction(ISD::UADDO, MVT::i32, Custom);
978 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
979 setOperationAction(ISD::USUBO, MVT::i32, Custom);
980 setOperationAction(ISD::SMULO, MVT::i32, Custom);
982 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
983 // handle type legalization for these operations here.
985 // FIXME: We really should do custom legalization for addition and
986 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
987 // than generic legalization for 64-bit multiplication-with-overflow, though.
988 if (Subtarget->is64Bit()) {
989 setOperationAction(ISD::SADDO, MVT::i64, Custom);
990 setOperationAction(ISD::UADDO, MVT::i64, Custom);
991 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
992 setOperationAction(ISD::USUBO, MVT::i64, Custom);
993 setOperationAction(ISD::SMULO, MVT::i64, Custom);
996 if (!Subtarget->is64Bit()) {
997 // These libcalls are not available in 32-bit.
998 setLibcallName(RTLIB::SHL_I128, 0);
999 setLibcallName(RTLIB::SRL_I128, 0);
1000 setLibcallName(RTLIB::SRA_I128, 0);
1003 // We have target-specific dag combine patterns for the following nodes:
1004 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1005 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1006 setTargetDAGCombine(ISD::BUILD_VECTOR);
1007 setTargetDAGCombine(ISD::SELECT);
1008 setTargetDAGCombine(ISD::SHL);
1009 setTargetDAGCombine(ISD::SRA);
1010 setTargetDAGCombine(ISD::SRL);
1011 setTargetDAGCombine(ISD::OR);
1012 setTargetDAGCombine(ISD::STORE);
1013 setTargetDAGCombine(ISD::ZERO_EXTEND);
1014 if (Subtarget->is64Bit())
1015 setTargetDAGCombine(ISD::MUL);
1017 computeRegisterProperties();
1019 // FIXME: These should be based on subtarget info. Plus, the values should
1020 // be smaller when we are in optimizing for size mode.
1021 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1022 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1023 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1024 setPrefLoopAlignment(16);
1025 benefitFromCodePlacementOpt = true;
1029 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1034 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1035 /// the desired ByVal argument alignment.
1036 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1039 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1040 if (VTy->getBitWidth() == 128)
1042 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1043 unsigned EltAlign = 0;
1044 getMaxByValAlign(ATy->getElementType(), EltAlign);
1045 if (EltAlign > MaxAlign)
1046 MaxAlign = EltAlign;
1047 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1048 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1049 unsigned EltAlign = 0;
1050 getMaxByValAlign(STy->getElementType(i), EltAlign);
1051 if (EltAlign > MaxAlign)
1052 MaxAlign = EltAlign;
1060 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1061 /// function arguments in the caller parameter area. For X86, aggregates
1062 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1063 /// are at 4-byte boundaries.
1064 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1065 if (Subtarget->is64Bit()) {
1066 // Max of 8 and alignment of type.
1067 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1074 if (Subtarget->hasSSE1())
1075 getMaxByValAlign(Ty, Align);
1079 /// getOptimalMemOpType - Returns the target specific optimal type for load
1080 /// and store operations as a result of memset, memcpy, and memmove
1081 /// lowering. If DstAlign is zero that means it's safe to destination
1082 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1083 /// means there isn't a need to check it against alignment requirement,
1084 /// probably because the source does not need to be loaded. If
1085 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1086 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1087 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1088 /// constant so it does not need to be loaded.
1089 /// It returns EVT::Other if the type should be determined using generic
1090 /// target-independent logic.
1092 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1093 unsigned DstAlign, unsigned SrcAlign,
1094 bool NonScalarIntSafe,
1096 MachineFunction &MF) const {
1097 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1098 // linux. This is because the stack realignment code can't handle certain
1099 // cases like PR2962. This should be removed when PR2962 is fixed.
1100 const Function *F = MF.getFunction();
1101 if (NonScalarIntSafe &&
1102 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1104 (Subtarget->isUnalignedMemAccessFast() ||
1105 ((DstAlign == 0 || DstAlign >= 16) &&
1106 (SrcAlign == 0 || SrcAlign >= 16))) &&
1107 Subtarget->getStackAlignment() >= 16) {
1108 if (Subtarget->hasSSE2())
1110 if (Subtarget->hasSSE1())
1112 } else if (!MemcpyStrSrc && Size >= 8 &&
1113 !Subtarget->is64Bit() &&
1114 Subtarget->getStackAlignment() >= 8 &&
1115 Subtarget->hasSSE2()) {
1116 // Do not use f64 to lower memcpy if source is string constant. It's
1117 // better to use i32 to avoid the loads.
1121 if (Subtarget->is64Bit() && Size >= 8)
1126 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1127 /// current function. The returned value is a member of the
1128 /// MachineJumpTableInfo::JTEntryKind enum.
1129 unsigned X86TargetLowering::getJumpTableEncoding() const {
1130 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1132 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1133 Subtarget->isPICStyleGOT())
1134 return MachineJumpTableInfo::EK_Custom32;
1136 // Otherwise, use the normal jump table encoding heuristics.
1137 return TargetLowering::getJumpTableEncoding();
1140 /// getPICBaseSymbol - Return the X86-32 PIC base.
1142 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1143 MCContext &Ctx) const {
1144 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1145 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1146 Twine(MF->getFunctionNumber())+"$pb");
1151 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1152 const MachineBasicBlock *MBB,
1153 unsigned uid,MCContext &Ctx) const{
1154 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1155 Subtarget->isPICStyleGOT());
1156 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1158 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1159 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1162 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1164 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1165 SelectionDAG &DAG) const {
1166 if (!Subtarget->is64Bit())
1167 // This doesn't have DebugLoc associated with it, but is not really the
1168 // same as a Register.
1169 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1173 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1174 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1176 const MCExpr *X86TargetLowering::
1177 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1178 MCContext &Ctx) const {
1179 // X86-64 uses RIP relative addressing based on the jump table label.
1180 if (Subtarget->isPICStyleRIPRel())
1181 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1183 // Otherwise, the reference is relative to the PIC base.
1184 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1187 /// getFunctionAlignment - Return the Log2 alignment of this function.
1188 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1189 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1192 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1193 unsigned &Offset) const {
1194 if (!Subtarget->isTargetLinux())
1197 if (Subtarget->is64Bit()) {
1198 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1200 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1213 //===----------------------------------------------------------------------===//
1214 // Return Value Calling Convention Implementation
1215 //===----------------------------------------------------------------------===//
1217 #include "X86GenCallingConv.inc"
1220 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1221 const SmallVectorImpl<EVT> &OutTys,
1222 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1223 LLVMContext &Context) const {
1224 SmallVector<CCValAssign, 16> RVLocs;
1225 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1227 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1231 X86TargetLowering::LowerReturn(SDValue Chain,
1232 CallingConv::ID CallConv, bool isVarArg,
1233 const SmallVectorImpl<ISD::OutputArg> &Outs,
1234 const SmallVectorImpl<SDValue> &OutVals,
1235 DebugLoc dl, SelectionDAG &DAG) const {
1236 MachineFunction &MF = DAG.getMachineFunction();
1237 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1239 SmallVector<CCValAssign, 16> RVLocs;
1240 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1241 RVLocs, *DAG.getContext());
1242 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1244 // Add the regs to the liveout set for the function.
1245 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1246 for (unsigned i = 0; i != RVLocs.size(); ++i)
1247 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1248 MRI.addLiveOut(RVLocs[i].getLocReg());
1252 SmallVector<SDValue, 6> RetOps;
1253 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1254 // Operand #1 = Bytes To Pop
1255 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1258 // Copy the result values into the output registers.
1259 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1260 CCValAssign &VA = RVLocs[i];
1261 assert(VA.isRegLoc() && "Can only return in registers!");
1262 SDValue ValToCopy = OutVals[i];
1264 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1265 // the RET instruction and handled by the FP Stackifier.
1266 if (VA.getLocReg() == X86::ST0 ||
1267 VA.getLocReg() == X86::ST1) {
1268 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1269 // change the value to the FP stack register class.
1270 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1271 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1272 RetOps.push_back(ValToCopy);
1273 // Don't emit a copytoreg.
1277 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1278 // which is returned in RAX / RDX.
1279 if (Subtarget->is64Bit()) {
1280 EVT ValVT = ValToCopy.getValueType();
1281 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1282 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1283 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1284 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1288 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1289 Flag = Chain.getValue(1);
1292 // The x86-64 ABI for returning structs by value requires that we copy
1293 // the sret argument into %rax for the return. We saved the argument into
1294 // a virtual register in the entry block, so now we copy the value out
1296 if (Subtarget->is64Bit() &&
1297 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1298 MachineFunction &MF = DAG.getMachineFunction();
1299 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1300 unsigned Reg = FuncInfo->getSRetReturnReg();
1302 "SRetReturnReg should have been set in LowerFormalArguments().");
1303 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1305 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1306 Flag = Chain.getValue(1);
1308 // RAX now acts like a return value.
1309 MRI.addLiveOut(X86::RAX);
1312 RetOps[0] = Chain; // Update chain.
1314 // Add the flag if we have it.
1316 RetOps.push_back(Flag);
1318 return DAG.getNode(X86ISD::RET_FLAG, dl,
1319 MVT::Other, &RetOps[0], RetOps.size());
1322 /// LowerCallResult - Lower the result values of a call into the
1323 /// appropriate copies out of appropriate physical registers.
1326 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1327 CallingConv::ID CallConv, bool isVarArg,
1328 const SmallVectorImpl<ISD::InputArg> &Ins,
1329 DebugLoc dl, SelectionDAG &DAG,
1330 SmallVectorImpl<SDValue> &InVals) const {
1332 // Assign locations to each value returned by this call.
1333 SmallVector<CCValAssign, 16> RVLocs;
1334 bool Is64Bit = Subtarget->is64Bit();
1335 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1336 RVLocs, *DAG.getContext());
1337 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1339 // Copy all of the result registers out of their specified physreg.
1340 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1341 CCValAssign &VA = RVLocs[i];
1342 EVT CopyVT = VA.getValVT();
1344 // If this is x86-64, and we disabled SSE, we can't return FP values
1345 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1346 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1347 report_fatal_error("SSE register return with SSE disabled");
1352 // If this is a call to a function that returns an fp value on the floating
1353 // point stack, we must guarantee the the value is popped from the stack, so
1354 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1355 // if the return value is not used. We use the FpGET_ST0 instructions
1357 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1358 // If we prefer to use the value in xmm registers, copy it out as f80 and
1359 // use a truncate to move it from fp stack reg to xmm reg.
1360 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1361 bool isST0 = VA.getLocReg() == X86::ST0;
1363 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1364 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1365 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1366 SDValue Ops[] = { Chain, InFlag };
1367 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
1369 Val = Chain.getValue(0);
1371 // Round the f80 to the right size, which also moves it to the appropriate
1373 if (CopyVT != VA.getValVT())
1374 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1375 // This truncation won't change the value.
1376 DAG.getIntPtrConstant(1));
1377 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1378 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1379 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1380 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1381 MVT::v2i64, InFlag).getValue(1);
1382 Val = Chain.getValue(0);
1383 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1384 Val, DAG.getConstant(0, MVT::i64));
1386 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1387 MVT::i64, InFlag).getValue(1);
1388 Val = Chain.getValue(0);
1390 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1392 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1393 CopyVT, InFlag).getValue(1);
1394 Val = Chain.getValue(0);
1396 InFlag = Chain.getValue(2);
1397 InVals.push_back(Val);
1404 //===----------------------------------------------------------------------===//
1405 // C & StdCall & Fast Calling Convention implementation
1406 //===----------------------------------------------------------------------===//
1407 // StdCall calling convention seems to be standard for many Windows' API
1408 // routines and around. It differs from C calling convention just a little:
1409 // callee should clean up the stack, not caller. Symbols should be also
1410 // decorated in some fancy way :) It doesn't support any vector arguments.
1411 // For info on fast calling convention see Fast Calling Convention (tail call)
1412 // implementation LowerX86_32FastCCCallTo.
1414 /// CallIsStructReturn - Determines whether a call uses struct return
1416 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1420 return Outs[0].Flags.isSRet();
1423 /// ArgsAreStructReturn - Determines whether a function uses struct
1424 /// return semantics.
1426 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1430 return Ins[0].Flags.isSRet();
1433 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1434 /// given CallingConvention value.
1435 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1436 if (Subtarget->is64Bit()) {
1437 if (CC == CallingConv::GHC)
1438 return CC_X86_64_GHC;
1439 else if (Subtarget->isTargetWin64())
1440 return CC_X86_Win64_C;
1445 if (CC == CallingConv::X86_FastCall)
1446 return CC_X86_32_FastCall;
1447 else if (CC == CallingConv::X86_ThisCall)
1448 return CC_X86_32_ThisCall;
1449 else if (CC == CallingConv::Fast)
1450 return CC_X86_32_FastCC;
1451 else if (CC == CallingConv::GHC)
1452 return CC_X86_32_GHC;
1457 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1458 /// by "Src" to address "Dst" with size and alignment information specified by
1459 /// the specific parameter attribute. The copy will be passed as a byval
1460 /// function parameter.
1462 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1463 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1465 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1466 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1467 /*isVolatile*/false, /*AlwaysInline=*/true,
1471 /// IsTailCallConvention - Return true if the calling convention is one that
1472 /// supports tail call optimization.
1473 static bool IsTailCallConvention(CallingConv::ID CC) {
1474 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1477 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1478 /// a tailcall target by changing its ABI.
1479 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1480 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1484 X86TargetLowering::LowerMemArgument(SDValue Chain,
1485 CallingConv::ID CallConv,
1486 const SmallVectorImpl<ISD::InputArg> &Ins,
1487 DebugLoc dl, SelectionDAG &DAG,
1488 const CCValAssign &VA,
1489 MachineFrameInfo *MFI,
1491 // Create the nodes corresponding to a load from this parameter slot.
1492 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1493 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1494 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1497 // If value is passed by pointer we have address passed instead of the value
1499 if (VA.getLocInfo() == CCValAssign::Indirect)
1500 ValVT = VA.getLocVT();
1502 ValVT = VA.getValVT();
1504 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1505 // changed with more analysis.
1506 // In case of tail call optimization mark all arguments mutable. Since they
1507 // could be overwritten by lowering of arguments in case of a tail call.
1508 if (Flags.isByVal()) {
1509 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1510 VA.getLocMemOffset(), isImmutable);
1511 return DAG.getFrameIndex(FI, getPointerTy());
1513 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1514 VA.getLocMemOffset(), isImmutable);
1515 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1516 return DAG.getLoad(ValVT, dl, Chain, FIN,
1517 PseudoSourceValue::getFixedStack(FI), 0,
1523 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1524 CallingConv::ID CallConv,
1526 const SmallVectorImpl<ISD::InputArg> &Ins,
1529 SmallVectorImpl<SDValue> &InVals)
1531 MachineFunction &MF = DAG.getMachineFunction();
1532 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1534 const Function* Fn = MF.getFunction();
1535 if (Fn->hasExternalLinkage() &&
1536 Subtarget->isTargetCygMing() &&
1537 Fn->getName() == "main")
1538 FuncInfo->setForceFramePointer(true);
1540 MachineFrameInfo *MFI = MF.getFrameInfo();
1541 bool Is64Bit = Subtarget->is64Bit();
1542 bool IsWin64 = Subtarget->isTargetWin64();
1544 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1545 "Var args not supported with calling convention fastcc or ghc");
1547 // Assign locations to all of the incoming arguments.
1548 SmallVector<CCValAssign, 16> ArgLocs;
1549 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1550 ArgLocs, *DAG.getContext());
1551 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1553 unsigned LastVal = ~0U;
1555 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1556 CCValAssign &VA = ArgLocs[i];
1557 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1559 assert(VA.getValNo() != LastVal &&
1560 "Don't support value assigned to multiple locs yet");
1561 LastVal = VA.getValNo();
1563 if (VA.isRegLoc()) {
1564 EVT RegVT = VA.getLocVT();
1565 TargetRegisterClass *RC = NULL;
1566 if (RegVT == MVT::i32)
1567 RC = X86::GR32RegisterClass;
1568 else if (Is64Bit && RegVT == MVT::i64)
1569 RC = X86::GR64RegisterClass;
1570 else if (RegVT == MVT::f32)
1571 RC = X86::FR32RegisterClass;
1572 else if (RegVT == MVT::f64)
1573 RC = X86::FR64RegisterClass;
1574 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1575 RC = X86::VR128RegisterClass;
1576 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1577 RC = X86::VR64RegisterClass;
1579 llvm_unreachable("Unknown argument type!");
1581 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1582 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1584 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1585 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1587 if (VA.getLocInfo() == CCValAssign::SExt)
1588 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1589 DAG.getValueType(VA.getValVT()));
1590 else if (VA.getLocInfo() == CCValAssign::ZExt)
1591 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1592 DAG.getValueType(VA.getValVT()));
1593 else if (VA.getLocInfo() == CCValAssign::BCvt)
1594 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1596 if (VA.isExtInLoc()) {
1597 // Handle MMX values passed in XMM regs.
1598 if (RegVT.isVector()) {
1599 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1600 ArgValue, DAG.getConstant(0, MVT::i64));
1601 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1603 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1606 assert(VA.isMemLoc());
1607 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1610 // If value is passed via pointer - do a load.
1611 if (VA.getLocInfo() == CCValAssign::Indirect)
1612 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1615 InVals.push_back(ArgValue);
1618 // The x86-64 ABI for returning structs by value requires that we copy
1619 // the sret argument into %rax for the return. Save the argument into
1620 // a virtual register so that we can access it from the return points.
1621 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1622 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1623 unsigned Reg = FuncInfo->getSRetReturnReg();
1625 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1626 FuncInfo->setSRetReturnReg(Reg);
1628 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1629 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1632 unsigned StackSize = CCInfo.getNextStackOffset();
1633 // Align stack specially for tail calls.
1634 if (FuncIsMadeTailCallSafe(CallConv))
1635 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1637 // If the function takes variable number of arguments, make a frame index for
1638 // the start of the first vararg value... for expansion of llvm.va_start.
1640 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1641 CallConv != CallingConv::X86_ThisCall)) {
1642 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1645 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1647 // FIXME: We should really autogenerate these arrays
1648 static const unsigned GPR64ArgRegsWin64[] = {
1649 X86::RCX, X86::RDX, X86::R8, X86::R9
1651 static const unsigned XMMArgRegsWin64[] = {
1652 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1654 static const unsigned GPR64ArgRegs64Bit[] = {
1655 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1657 static const unsigned XMMArgRegs64Bit[] = {
1658 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1659 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1661 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1664 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1665 GPR64ArgRegs = GPR64ArgRegsWin64;
1666 XMMArgRegs = XMMArgRegsWin64;
1668 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1669 GPR64ArgRegs = GPR64ArgRegs64Bit;
1670 XMMArgRegs = XMMArgRegs64Bit;
1672 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1674 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1677 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1678 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1679 "SSE register cannot be used when SSE is disabled!");
1680 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1681 "SSE register cannot be used when SSE is disabled!");
1682 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1683 // Kernel mode asks for SSE to be disabled, so don't push them
1685 TotalNumXMMRegs = 0;
1687 // For X86-64, if there are vararg parameters that are passed via
1688 // registers, then we must store them to their spots on the stack so they
1689 // may be loaded by deferencing the result of va_next.
1690 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1691 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1692 FuncInfo->setRegSaveFrameIndex(
1693 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1696 // Store the integer parameter registers.
1697 SmallVector<SDValue, 8> MemOps;
1698 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1700 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1701 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1702 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1703 DAG.getIntPtrConstant(Offset));
1704 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1705 X86::GR64RegisterClass);
1706 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1708 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1709 PseudoSourceValue::getFixedStack(
1710 FuncInfo->getRegSaveFrameIndex()),
1711 Offset, false, false, 0);
1712 MemOps.push_back(Store);
1716 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1717 // Now store the XMM (fp + vector) parameter registers.
1718 SmallVector<SDValue, 11> SaveXMMOps;
1719 SaveXMMOps.push_back(Chain);
1721 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1722 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1723 SaveXMMOps.push_back(ALVal);
1725 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1726 FuncInfo->getRegSaveFrameIndex()));
1727 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1728 FuncInfo->getVarArgsFPOffset()));
1730 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1731 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1732 X86::VR128RegisterClass);
1733 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1734 SaveXMMOps.push_back(Val);
1736 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1738 &SaveXMMOps[0], SaveXMMOps.size()));
1741 if (!MemOps.empty())
1742 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1743 &MemOps[0], MemOps.size());
1747 // Some CCs need callee pop.
1748 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1749 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1751 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1752 // If this is an sret function, the return should pop the hidden pointer.
1753 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1754 FuncInfo->setBytesToPopOnReturn(4);
1758 // RegSaveFrameIndex is X86-64 only.
1759 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1760 if (CallConv == CallingConv::X86_FastCall ||
1761 CallConv == CallingConv::X86_ThisCall)
1762 // fastcc functions can't have varargs.
1763 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1770 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1771 SDValue StackPtr, SDValue Arg,
1772 DebugLoc dl, SelectionDAG &DAG,
1773 const CCValAssign &VA,
1774 ISD::ArgFlagsTy Flags) const {
1775 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1776 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1777 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1778 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1779 if (Flags.isByVal()) {
1780 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1782 return DAG.getStore(Chain, dl, Arg, PtrOff,
1783 PseudoSourceValue::getStack(), LocMemOffset,
1787 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1788 /// optimization is performed and it is required.
1790 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1791 SDValue &OutRetAddr, SDValue Chain,
1792 bool IsTailCall, bool Is64Bit,
1793 int FPDiff, DebugLoc dl) const {
1794 // Adjust the Return address stack slot.
1795 EVT VT = getPointerTy();
1796 OutRetAddr = getReturnAddressFrameIndex(DAG);
1798 // Load the "old" Return address.
1799 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1800 return SDValue(OutRetAddr.getNode(), 1);
1803 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1804 /// optimization is performed and it is required (FPDiff!=0).
1806 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1807 SDValue Chain, SDValue RetAddrFrIdx,
1808 bool Is64Bit, int FPDiff, DebugLoc dl) {
1809 // Store the return address to the appropriate stack slot.
1810 if (!FPDiff) return Chain;
1811 // Calculate the new stack slot for the return address.
1812 int SlotSize = Is64Bit ? 8 : 4;
1813 int NewReturnAddrFI =
1814 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1815 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1816 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1817 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1818 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1824 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1825 CallingConv::ID CallConv, bool isVarArg,
1827 const SmallVectorImpl<ISD::OutputArg> &Outs,
1828 const SmallVectorImpl<SDValue> &OutVals,
1829 const SmallVectorImpl<ISD::InputArg> &Ins,
1830 DebugLoc dl, SelectionDAG &DAG,
1831 SmallVectorImpl<SDValue> &InVals) const {
1832 MachineFunction &MF = DAG.getMachineFunction();
1833 bool Is64Bit = Subtarget->is64Bit();
1834 bool IsStructRet = CallIsStructReturn(Outs);
1835 bool IsSibcall = false;
1838 // Check if it's really possible to do a tail call.
1839 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1840 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1841 Outs, OutVals, Ins, DAG);
1843 // Sibcalls are automatically detected tailcalls which do not require
1845 if (!GuaranteedTailCallOpt && isTailCall)
1852 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1853 "Var args not supported with calling convention fastcc or ghc");
1855 // Analyze operands of the call, assigning locations to each operand.
1856 SmallVector<CCValAssign, 16> ArgLocs;
1857 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1858 ArgLocs, *DAG.getContext());
1859 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1861 // Get a count of how many bytes are to be pushed on the stack.
1862 unsigned NumBytes = CCInfo.getNextStackOffset();
1864 // This is a sibcall. The memory operands are available in caller's
1865 // own caller's stack.
1867 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1868 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1871 if (isTailCall && !IsSibcall) {
1872 // Lower arguments at fp - stackoffset + fpdiff.
1873 unsigned NumBytesCallerPushed =
1874 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1875 FPDiff = NumBytesCallerPushed - NumBytes;
1877 // Set the delta of movement of the returnaddr stackslot.
1878 // But only set if delta is greater than previous delta.
1879 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1880 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1884 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1886 SDValue RetAddrFrIdx;
1887 // Load return adress for tail calls.
1888 if (isTailCall && FPDiff)
1889 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1890 Is64Bit, FPDiff, dl);
1892 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1893 SmallVector<SDValue, 8> MemOpChains;
1896 // Walk the register/memloc assignments, inserting copies/loads. In the case
1897 // of tail call optimization arguments are handle later.
1898 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1899 CCValAssign &VA = ArgLocs[i];
1900 EVT RegVT = VA.getLocVT();
1901 SDValue Arg = OutVals[i];
1902 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1903 bool isByVal = Flags.isByVal();
1905 // Promote the value if needed.
1906 switch (VA.getLocInfo()) {
1907 default: llvm_unreachable("Unknown loc info!");
1908 case CCValAssign::Full: break;
1909 case CCValAssign::SExt:
1910 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1912 case CCValAssign::ZExt:
1913 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1915 case CCValAssign::AExt:
1916 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1917 // Special case: passing MMX values in XMM registers.
1918 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1919 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1920 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1922 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1924 case CCValAssign::BCvt:
1925 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1927 case CCValAssign::Indirect: {
1928 // Store the argument.
1929 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1930 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1931 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1932 PseudoSourceValue::getFixedStack(FI), 0,
1939 if (VA.isRegLoc()) {
1940 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1941 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1942 assert(VA.isMemLoc());
1943 if (StackPtr.getNode() == 0)
1944 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1945 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1946 dl, DAG, VA, Flags));
1950 if (!MemOpChains.empty())
1951 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1952 &MemOpChains[0], MemOpChains.size());
1954 // Build a sequence of copy-to-reg nodes chained together with token chain
1955 // and flag operands which copy the outgoing args into registers.
1957 // Tail call byval lowering might overwrite argument registers so in case of
1958 // tail call optimization the copies to registers are lowered later.
1960 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1961 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1962 RegsToPass[i].second, InFlag);
1963 InFlag = Chain.getValue(1);
1966 if (Subtarget->isPICStyleGOT()) {
1967 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1970 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1971 DAG.getNode(X86ISD::GlobalBaseReg,
1972 DebugLoc(), getPointerTy()),
1974 InFlag = Chain.getValue(1);
1976 // If we are tail calling and generating PIC/GOT style code load the
1977 // address of the callee into ECX. The value in ecx is used as target of
1978 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1979 // for tail calls on PIC/GOT architectures. Normally we would just put the
1980 // address of GOT into ebx and then call target@PLT. But for tail calls
1981 // ebx would be restored (since ebx is callee saved) before jumping to the
1984 // Note: The actual moving to ECX is done further down.
1985 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1986 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1987 !G->getGlobal()->hasProtectedVisibility())
1988 Callee = LowerGlobalAddress(Callee, DAG);
1989 else if (isa<ExternalSymbolSDNode>(Callee))
1990 Callee = LowerExternalSymbol(Callee, DAG);
1994 if (Is64Bit && isVarArg) {
1995 // From AMD64 ABI document:
1996 // For calls that may call functions that use varargs or stdargs
1997 // (prototype-less calls or calls to functions containing ellipsis (...) in
1998 // the declaration) %al is used as hidden argument to specify the number
1999 // of SSE registers used. The contents of %al do not need to match exactly
2000 // the number of registers, but must be an ubound on the number of SSE
2001 // registers used and is in the range 0 - 8 inclusive.
2003 // FIXME: Verify this on Win64
2004 // Count the number of XMM registers allocated.
2005 static const unsigned XMMArgRegs[] = {
2006 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2007 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2009 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2010 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2011 && "SSE registers cannot be used when SSE is disabled");
2013 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2014 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2015 InFlag = Chain.getValue(1);
2019 // For tail calls lower the arguments to the 'real' stack slot.
2021 // Force all the incoming stack arguments to be loaded from the stack
2022 // before any new outgoing arguments are stored to the stack, because the
2023 // outgoing stack slots may alias the incoming argument stack slots, and
2024 // the alias isn't otherwise explicit. This is slightly more conservative
2025 // than necessary, because it means that each store effectively depends
2026 // on every argument instead of just those arguments it would clobber.
2027 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2029 SmallVector<SDValue, 8> MemOpChains2;
2032 // Do not flag preceeding copytoreg stuff together with the following stuff.
2034 if (GuaranteedTailCallOpt) {
2035 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2036 CCValAssign &VA = ArgLocs[i];
2039 assert(VA.isMemLoc());
2040 SDValue Arg = OutVals[i];
2041 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2042 // Create frame index.
2043 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2044 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2045 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2046 FIN = DAG.getFrameIndex(FI, getPointerTy());
2048 if (Flags.isByVal()) {
2049 // Copy relative to framepointer.
2050 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2051 if (StackPtr.getNode() == 0)
2052 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2054 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2056 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2060 // Store relative to framepointer.
2061 MemOpChains2.push_back(
2062 DAG.getStore(ArgChain, dl, Arg, FIN,
2063 PseudoSourceValue::getFixedStack(FI), 0,
2069 if (!MemOpChains2.empty())
2070 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2071 &MemOpChains2[0], MemOpChains2.size());
2073 // Copy arguments to their registers.
2074 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2075 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2076 RegsToPass[i].second, InFlag);
2077 InFlag = Chain.getValue(1);
2081 // Store the return address to the appropriate stack slot.
2082 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2086 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2087 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2088 // In the 64-bit large code model, we have to make all calls
2089 // through a register, since the call instruction's 32-bit
2090 // pc-relative offset may not be large enough to hold the whole
2092 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2093 // If the callee is a GlobalAddress node (quite common, every direct call
2094 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2097 // We should use extra load for direct calls to dllimported functions in
2099 const GlobalValue *GV = G->getGlobal();
2100 if (!GV->hasDLLImportLinkage()) {
2101 unsigned char OpFlags = 0;
2103 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2104 // external symbols most go through the PLT in PIC mode. If the symbol
2105 // has hidden or protected visibility, or if it is static or local, then
2106 // we don't need to use the PLT - we can directly call it.
2107 if (Subtarget->isTargetELF() &&
2108 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2109 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2110 OpFlags = X86II::MO_PLT;
2111 } else if (Subtarget->isPICStyleStubAny() &&
2112 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2113 Subtarget->getDarwinVers() < 9) {
2114 // PC-relative references to external symbols should go through $stub,
2115 // unless we're building with the leopard linker or later, which
2116 // automatically synthesizes these stubs.
2117 OpFlags = X86II::MO_DARWIN_STUB;
2120 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2121 G->getOffset(), OpFlags);
2123 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2124 unsigned char OpFlags = 0;
2126 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2127 // symbols should go through the PLT.
2128 if (Subtarget->isTargetELF() &&
2129 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2130 OpFlags = X86II::MO_PLT;
2131 } else if (Subtarget->isPICStyleStubAny() &&
2132 Subtarget->getDarwinVers() < 9) {
2133 // PC-relative references to external symbols should go through $stub,
2134 // unless we're building with the leopard linker or later, which
2135 // automatically synthesizes these stubs.
2136 OpFlags = X86II::MO_DARWIN_STUB;
2139 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2143 // Returns a chain & a flag for retval copy to use.
2144 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2145 SmallVector<SDValue, 8> Ops;
2147 if (!IsSibcall && isTailCall) {
2148 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2149 DAG.getIntPtrConstant(0, true), InFlag);
2150 InFlag = Chain.getValue(1);
2153 Ops.push_back(Chain);
2154 Ops.push_back(Callee);
2157 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2159 // Add argument registers to the end of the list so that they are known live
2161 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2162 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2163 RegsToPass[i].second.getValueType()));
2165 // Add an implicit use GOT pointer in EBX.
2166 if (!isTailCall && Subtarget->isPICStyleGOT())
2167 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2169 // Add an implicit use of AL for x86 vararg functions.
2170 if (Is64Bit && isVarArg)
2171 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2173 if (InFlag.getNode())
2174 Ops.push_back(InFlag);
2178 //// If this is the first return lowered for this function, add the regs
2179 //// to the liveout set for the function.
2180 // This isn't right, although it's probably harmless on x86; liveouts
2181 // should be computed from returns not tail calls. Consider a void
2182 // function making a tail call to a function returning int.
2183 return DAG.getNode(X86ISD::TC_RETURN, dl,
2184 NodeTys, &Ops[0], Ops.size());
2187 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2188 InFlag = Chain.getValue(1);
2190 // Create the CALLSEQ_END node.
2191 unsigned NumBytesForCalleeToPush;
2192 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2193 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2194 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2195 // If this is a call to a struct-return function, the callee
2196 // pops the hidden struct pointer, so we have to push it back.
2197 // This is common for Darwin/X86, Linux & Mingw32 targets.
2198 NumBytesForCalleeToPush = 4;
2200 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2202 // Returns a flag for retval copy to use.
2204 Chain = DAG.getCALLSEQ_END(Chain,
2205 DAG.getIntPtrConstant(NumBytes, true),
2206 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2209 InFlag = Chain.getValue(1);
2212 // Handle result values, copying them out of physregs into vregs that we
2214 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2215 Ins, dl, DAG, InVals);
2219 //===----------------------------------------------------------------------===//
2220 // Fast Calling Convention (tail call) implementation
2221 //===----------------------------------------------------------------------===//
2223 // Like std call, callee cleans arguments, convention except that ECX is
2224 // reserved for storing the tail called function address. Only 2 registers are
2225 // free for argument passing (inreg). Tail call optimization is performed
2227 // * tailcallopt is enabled
2228 // * caller/callee are fastcc
2229 // On X86_64 architecture with GOT-style position independent code only local
2230 // (within module) calls are supported at the moment.
2231 // To keep the stack aligned according to platform abi the function
2232 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2233 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2234 // If a tail called function callee has more arguments than the caller the
2235 // caller needs to make sure that there is room to move the RETADDR to. This is
2236 // achieved by reserving an area the size of the argument delta right after the
2237 // original REtADDR, but before the saved framepointer or the spilled registers
2238 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2250 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2251 /// for a 16 byte align requirement.
2253 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2254 SelectionDAG& DAG) const {
2255 MachineFunction &MF = DAG.getMachineFunction();
2256 const TargetMachine &TM = MF.getTarget();
2257 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2258 unsigned StackAlignment = TFI.getStackAlignment();
2259 uint64_t AlignMask = StackAlignment - 1;
2260 int64_t Offset = StackSize;
2261 uint64_t SlotSize = TD->getPointerSize();
2262 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2263 // Number smaller than 12 so just add the difference.
2264 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2266 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2267 Offset = ((~AlignMask) & Offset) + StackAlignment +
2268 (StackAlignment-SlotSize);
2273 /// MatchingStackOffset - Return true if the given stack call argument is
2274 /// already available in the same position (relatively) of the caller's
2275 /// incoming argument stack.
2277 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2278 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2279 const X86InstrInfo *TII) {
2280 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2282 if (Arg.getOpcode() == ISD::CopyFromReg) {
2283 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2284 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2286 MachineInstr *Def = MRI->getVRegDef(VR);
2289 if (!Flags.isByVal()) {
2290 if (!TII->isLoadFromStackSlot(Def, FI))
2293 unsigned Opcode = Def->getOpcode();
2294 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2295 Def->getOperand(1).isFI()) {
2296 FI = Def->getOperand(1).getIndex();
2297 Bytes = Flags.getByValSize();
2301 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2302 if (Flags.isByVal())
2303 // ByVal argument is passed in as a pointer but it's now being
2304 // dereferenced. e.g.
2305 // define @foo(%struct.X* %A) {
2306 // tail call @bar(%struct.X* byval %A)
2309 SDValue Ptr = Ld->getBasePtr();
2310 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2313 FI = FINode->getIndex();
2317 assert(FI != INT_MAX);
2318 if (!MFI->isFixedObjectIndex(FI))
2320 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2323 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2324 /// for tail call optimization. Targets which want to do tail call
2325 /// optimization should implement this function.
2327 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2328 CallingConv::ID CalleeCC,
2330 bool isCalleeStructRet,
2331 bool isCallerStructRet,
2332 const SmallVectorImpl<ISD::OutputArg> &Outs,
2333 const SmallVectorImpl<SDValue> &OutVals,
2334 const SmallVectorImpl<ISD::InputArg> &Ins,
2335 SelectionDAG& DAG) const {
2336 if (!IsTailCallConvention(CalleeCC) &&
2337 CalleeCC != CallingConv::C)
2340 // If -tailcallopt is specified, make fastcc functions tail-callable.
2341 const MachineFunction &MF = DAG.getMachineFunction();
2342 const Function *CallerF = DAG.getMachineFunction().getFunction();
2343 CallingConv::ID CallerCC = CallerF->getCallingConv();
2344 bool CCMatch = CallerCC == CalleeCC;
2346 if (GuaranteedTailCallOpt) {
2347 if (IsTailCallConvention(CalleeCC) && CCMatch)
2352 // Look for obvious safe cases to perform tail call optimization that do not
2353 // require ABI changes. This is what gcc calls sibcall.
2355 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2356 // emit a special epilogue.
2357 if (RegInfo->needsStackRealignment(MF))
2360 // Do not sibcall optimize vararg calls unless the call site is not passing any
2362 if (isVarArg && !Outs.empty())
2365 // Also avoid sibcall optimization if either caller or callee uses struct
2366 // return semantics.
2367 if (isCalleeStructRet || isCallerStructRet)
2370 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2371 // Therefore if it's not used by the call it is not safe to optimize this into
2373 bool Unused = false;
2374 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2381 SmallVector<CCValAssign, 16> RVLocs;
2382 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2383 RVLocs, *DAG.getContext());
2384 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2385 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2386 CCValAssign &VA = RVLocs[i];
2387 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2392 // If the calling conventions do not match, then we'd better make sure the
2393 // results are returned in the same way as what the caller expects.
2395 SmallVector<CCValAssign, 16> RVLocs1;
2396 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2397 RVLocs1, *DAG.getContext());
2398 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2400 SmallVector<CCValAssign, 16> RVLocs2;
2401 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2402 RVLocs2, *DAG.getContext());
2403 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2405 if (RVLocs1.size() != RVLocs2.size())
2407 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2408 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2410 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2412 if (RVLocs1[i].isRegLoc()) {
2413 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2416 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2422 // If the callee takes no arguments then go on to check the results of the
2424 if (!Outs.empty()) {
2425 // Check if stack adjustment is needed. For now, do not do this if any
2426 // argument is passed on the stack.
2427 SmallVector<CCValAssign, 16> ArgLocs;
2428 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2429 ArgLocs, *DAG.getContext());
2430 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2431 if (CCInfo.getNextStackOffset()) {
2432 MachineFunction &MF = DAG.getMachineFunction();
2433 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2435 if (Subtarget->isTargetWin64())
2436 // Win64 ABI has additional complications.
2439 // Check if the arguments are already laid out in the right way as
2440 // the caller's fixed stack objects.
2441 MachineFrameInfo *MFI = MF.getFrameInfo();
2442 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2443 const X86InstrInfo *TII =
2444 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2445 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2446 CCValAssign &VA = ArgLocs[i];
2447 SDValue Arg = OutVals[i];
2448 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2449 if (VA.getLocInfo() == CCValAssign::Indirect)
2451 if (!VA.isRegLoc()) {
2452 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2459 // If the tailcall address may be in a register, then make sure it's
2460 // possible to register allocate for it. In 32-bit, the call address can
2461 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2462 // callee-saved registers are restored. In 64-bit, it's RAX, RCX, RDX, RSI,
2463 // RDI, R8, R9, R11.
2464 if (!isa<GlobalAddressSDNode>(Callee) &&
2465 !isa<ExternalSymbolSDNode>(Callee)) {
2466 unsigned Limit = Subtarget->is64Bit() ? 8 : 3;
2467 unsigned NumInRegs = 0;
2468 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2469 CCValAssign &VA = ArgLocs[i];
2470 if (VA.isRegLoc()) {
2471 if (++NumInRegs == Limit)
2482 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2483 return X86::createFastISel(funcInfo);
2487 //===----------------------------------------------------------------------===//
2488 // Other Lowering Hooks
2489 //===----------------------------------------------------------------------===//
2492 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2493 MachineFunction &MF = DAG.getMachineFunction();
2494 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2495 int ReturnAddrIndex = FuncInfo->getRAIndex();
2497 if (ReturnAddrIndex == 0) {
2498 // Set up a frame object for the return address.
2499 uint64_t SlotSize = TD->getPointerSize();
2500 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2502 FuncInfo->setRAIndex(ReturnAddrIndex);
2505 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2509 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2510 bool hasSymbolicDisplacement) {
2511 // Offset should fit into 32 bit immediate field.
2512 if (!isInt<32>(Offset))
2515 // If we don't have a symbolic displacement - we don't have any extra
2517 if (!hasSymbolicDisplacement)
2520 // FIXME: Some tweaks might be needed for medium code model.
2521 if (M != CodeModel::Small && M != CodeModel::Kernel)
2524 // For small code model we assume that latest object is 16MB before end of 31
2525 // bits boundary. We may also accept pretty large negative constants knowing
2526 // that all objects are in the positive half of address space.
2527 if (M == CodeModel::Small && Offset < 16*1024*1024)
2530 // For kernel code model we know that all object resist in the negative half
2531 // of 32bits address space. We may not accept negative offsets, since they may
2532 // be just off and we may accept pretty large positive ones.
2533 if (M == CodeModel::Kernel && Offset > 0)
2539 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2540 /// specific condition code, returning the condition code and the LHS/RHS of the
2541 /// comparison to make.
2542 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2543 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2545 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2546 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2547 // X > -1 -> X == 0, jump !sign.
2548 RHS = DAG.getConstant(0, RHS.getValueType());
2549 return X86::COND_NS;
2550 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2551 // X < 0 -> X == 0, jump on sign.
2553 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2555 RHS = DAG.getConstant(0, RHS.getValueType());
2556 return X86::COND_LE;
2560 switch (SetCCOpcode) {
2561 default: llvm_unreachable("Invalid integer condition!");
2562 case ISD::SETEQ: return X86::COND_E;
2563 case ISD::SETGT: return X86::COND_G;
2564 case ISD::SETGE: return X86::COND_GE;
2565 case ISD::SETLT: return X86::COND_L;
2566 case ISD::SETLE: return X86::COND_LE;
2567 case ISD::SETNE: return X86::COND_NE;
2568 case ISD::SETULT: return X86::COND_B;
2569 case ISD::SETUGT: return X86::COND_A;
2570 case ISD::SETULE: return X86::COND_BE;
2571 case ISD::SETUGE: return X86::COND_AE;
2575 // First determine if it is required or is profitable to flip the operands.
2577 // If LHS is a foldable load, but RHS is not, flip the condition.
2578 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2579 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2580 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2581 std::swap(LHS, RHS);
2584 switch (SetCCOpcode) {
2590 std::swap(LHS, RHS);
2594 // On a floating point condition, the flags are set as follows:
2596 // 0 | 0 | 0 | X > Y
2597 // 0 | 0 | 1 | X < Y
2598 // 1 | 0 | 0 | X == Y
2599 // 1 | 1 | 1 | unordered
2600 switch (SetCCOpcode) {
2601 default: llvm_unreachable("Condcode should be pre-legalized away");
2603 case ISD::SETEQ: return X86::COND_E;
2604 case ISD::SETOLT: // flipped
2606 case ISD::SETGT: return X86::COND_A;
2607 case ISD::SETOLE: // flipped
2609 case ISD::SETGE: return X86::COND_AE;
2610 case ISD::SETUGT: // flipped
2612 case ISD::SETLT: return X86::COND_B;
2613 case ISD::SETUGE: // flipped
2615 case ISD::SETLE: return X86::COND_BE;
2617 case ISD::SETNE: return X86::COND_NE;
2618 case ISD::SETUO: return X86::COND_P;
2619 case ISD::SETO: return X86::COND_NP;
2621 case ISD::SETUNE: return X86::COND_INVALID;
2625 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2626 /// code. Current x86 isa includes the following FP cmov instructions:
2627 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2628 static bool hasFPCMov(unsigned X86CC) {
2644 /// isFPImmLegal - Returns true if the target can instruction select the
2645 /// specified FP immediate natively. If false, the legalizer will
2646 /// materialize the FP immediate as a load from a constant pool.
2647 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2648 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2649 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2655 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2656 /// the specified range (L, H].
2657 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2658 return (Val < 0) || (Val >= Low && Val < Hi);
2661 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2662 /// specified value.
2663 static bool isUndefOrEqual(int Val, int CmpVal) {
2664 if (Val < 0 || Val == CmpVal)
2669 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2670 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2671 /// the second operand.
2672 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2673 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2674 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2675 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2676 return (Mask[0] < 2 && Mask[1] < 2);
2680 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2681 SmallVector<int, 8> M;
2683 return ::isPSHUFDMask(M, N->getValueType(0));
2686 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2687 /// is suitable for input to PSHUFHW.
2688 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2689 if (VT != MVT::v8i16)
2692 // Lower quadword copied in order or undef.
2693 for (int i = 0; i != 4; ++i)
2694 if (Mask[i] >= 0 && Mask[i] != i)
2697 // Upper quadword shuffled.
2698 for (int i = 4; i != 8; ++i)
2699 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2705 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2706 SmallVector<int, 8> M;
2708 return ::isPSHUFHWMask(M, N->getValueType(0));
2711 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2712 /// is suitable for input to PSHUFLW.
2713 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2714 if (VT != MVT::v8i16)
2717 // Upper quadword copied in order.
2718 for (int i = 4; i != 8; ++i)
2719 if (Mask[i] >= 0 && Mask[i] != i)
2722 // Lower quadword shuffled.
2723 for (int i = 0; i != 4; ++i)
2730 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2731 SmallVector<int, 8> M;
2733 return ::isPSHUFLWMask(M, N->getValueType(0));
2736 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2737 /// is suitable for input to PALIGNR.
2738 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2740 int i, e = VT.getVectorNumElements();
2742 // Do not handle v2i64 / v2f64 shuffles with palignr.
2743 if (e < 4 || !hasSSSE3)
2746 for (i = 0; i != e; ++i)
2750 // All undef, not a palignr.
2754 // Determine if it's ok to perform a palignr with only the LHS, since we
2755 // don't have access to the actual shuffle elements to see if RHS is undef.
2756 bool Unary = Mask[i] < (int)e;
2757 bool NeedsUnary = false;
2759 int s = Mask[i] - i;
2761 // Check the rest of the elements to see if they are consecutive.
2762 for (++i; i != e; ++i) {
2767 Unary = Unary && (m < (int)e);
2768 NeedsUnary = NeedsUnary || (m < s);
2770 if (NeedsUnary && !Unary)
2772 if (Unary && m != ((s+i) & (e-1)))
2774 if (!Unary && m != (s+i))
2780 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2781 SmallVector<int, 8> M;
2783 return ::isPALIGNRMask(M, N->getValueType(0), true);
2786 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2787 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2788 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2789 int NumElems = VT.getVectorNumElements();
2790 if (NumElems != 2 && NumElems != 4)
2793 int Half = NumElems / 2;
2794 for (int i = 0; i < Half; ++i)
2795 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2797 for (int i = Half; i < NumElems; ++i)
2798 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2804 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2805 SmallVector<int, 8> M;
2807 return ::isSHUFPMask(M, N->getValueType(0));
2810 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2811 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2812 /// half elements to come from vector 1 (which would equal the dest.) and
2813 /// the upper half to come from vector 2.
2814 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2815 int NumElems = VT.getVectorNumElements();
2817 if (NumElems != 2 && NumElems != 4)
2820 int Half = NumElems / 2;
2821 for (int i = 0; i < Half; ++i)
2822 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2824 for (int i = Half; i < NumElems; ++i)
2825 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2830 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2831 SmallVector<int, 8> M;
2833 return isCommutedSHUFPMask(M, N->getValueType(0));
2836 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2837 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2838 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2839 if (N->getValueType(0).getVectorNumElements() != 4)
2842 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2843 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2844 isUndefOrEqual(N->getMaskElt(1), 7) &&
2845 isUndefOrEqual(N->getMaskElt(2), 2) &&
2846 isUndefOrEqual(N->getMaskElt(3), 3);
2849 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2850 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2852 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2853 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2858 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2859 isUndefOrEqual(N->getMaskElt(1), 3) &&
2860 isUndefOrEqual(N->getMaskElt(2), 2) &&
2861 isUndefOrEqual(N->getMaskElt(3), 3);
2864 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2865 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2866 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2867 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2869 if (NumElems != 2 && NumElems != 4)
2872 for (unsigned i = 0; i < NumElems/2; ++i)
2873 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2876 for (unsigned i = NumElems/2; i < NumElems; ++i)
2877 if (!isUndefOrEqual(N->getMaskElt(i), i))
2883 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2884 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2885 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2886 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2888 if (NumElems != 2 && NumElems != 4)
2891 for (unsigned i = 0; i < NumElems/2; ++i)
2892 if (!isUndefOrEqual(N->getMaskElt(i), i))
2895 for (unsigned i = 0; i < NumElems/2; ++i)
2896 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2902 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2903 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2904 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2905 bool V2IsSplat = false) {
2906 int NumElts = VT.getVectorNumElements();
2907 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2910 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2912 int BitI1 = Mask[i+1];
2913 if (!isUndefOrEqual(BitI, j))
2916 if (!isUndefOrEqual(BitI1, NumElts))
2919 if (!isUndefOrEqual(BitI1, j + NumElts))
2926 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2927 SmallVector<int, 8> M;
2929 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2932 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2933 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2934 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2935 bool V2IsSplat = false) {
2936 int NumElts = VT.getVectorNumElements();
2937 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2940 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2942 int BitI1 = Mask[i+1];
2943 if (!isUndefOrEqual(BitI, j + NumElts/2))
2946 if (isUndefOrEqual(BitI1, NumElts))
2949 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2956 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2957 SmallVector<int, 8> M;
2959 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2962 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2963 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2965 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2966 int NumElems = VT.getVectorNumElements();
2967 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2970 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2972 int BitI1 = Mask[i+1];
2973 if (!isUndefOrEqual(BitI, j))
2975 if (!isUndefOrEqual(BitI1, j))
2981 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2982 SmallVector<int, 8> M;
2984 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2987 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2988 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2990 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2991 int NumElems = VT.getVectorNumElements();
2992 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2995 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2997 int BitI1 = Mask[i+1];
2998 if (!isUndefOrEqual(BitI, j))
3000 if (!isUndefOrEqual(BitI1, j))
3006 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3007 SmallVector<int, 8> M;
3009 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3012 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3013 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3014 /// MOVSD, and MOVD, i.e. setting the lowest element.
3015 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3016 if (VT.getVectorElementType().getSizeInBits() < 32)
3019 int NumElts = VT.getVectorNumElements();
3021 if (!isUndefOrEqual(Mask[0], NumElts))
3024 for (int i = 1; i < NumElts; ++i)
3025 if (!isUndefOrEqual(Mask[i], i))
3031 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3032 SmallVector<int, 8> M;
3034 return ::isMOVLMask(M, N->getValueType(0));
3037 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3038 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3039 /// element of vector 2 and the other elements to come from vector 1 in order.
3040 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3041 bool V2IsSplat = false, bool V2IsUndef = false) {
3042 int NumOps = VT.getVectorNumElements();
3043 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3046 if (!isUndefOrEqual(Mask[0], 0))
3049 for (int i = 1; i < NumOps; ++i)
3050 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3051 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3052 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3058 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3059 bool V2IsUndef = false) {
3060 SmallVector<int, 8> M;
3062 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3065 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3066 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3067 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3068 if (N->getValueType(0).getVectorNumElements() != 4)
3071 // Expect 1, 1, 3, 3
3072 for (unsigned i = 0; i < 2; ++i) {
3073 int Elt = N->getMaskElt(i);
3074 if (Elt >= 0 && Elt != 1)
3079 for (unsigned i = 2; i < 4; ++i) {
3080 int Elt = N->getMaskElt(i);
3081 if (Elt >= 0 && Elt != 3)
3086 // Don't use movshdup if it can be done with a shufps.
3087 // FIXME: verify that matching u, u, 3, 3 is what we want.
3091 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3092 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3093 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3094 if (N->getValueType(0).getVectorNumElements() != 4)
3097 // Expect 0, 0, 2, 2
3098 for (unsigned i = 0; i < 2; ++i)
3099 if (N->getMaskElt(i) > 0)
3103 for (unsigned i = 2; i < 4; ++i) {
3104 int Elt = N->getMaskElt(i);
3105 if (Elt >= 0 && Elt != 2)
3110 // Don't use movsldup if it can be done with a shufps.
3114 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3115 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3116 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3117 int e = N->getValueType(0).getVectorNumElements() / 2;
3119 for (int i = 0; i < e; ++i)
3120 if (!isUndefOrEqual(N->getMaskElt(i), i))
3122 for (int i = 0; i < e; ++i)
3123 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3128 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3129 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3130 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3131 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3132 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3134 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3136 for (int i = 0; i < NumOperands; ++i) {
3137 int Val = SVOp->getMaskElt(NumOperands-i-1);
3138 if (Val < 0) Val = 0;
3139 if (Val >= NumOperands) Val -= NumOperands;
3141 if (i != NumOperands - 1)
3147 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3148 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3149 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3150 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3152 // 8 nodes, but we only care about the last 4.
3153 for (unsigned i = 7; i >= 4; --i) {
3154 int Val = SVOp->getMaskElt(i);
3163 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3164 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3165 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3166 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3168 // 8 nodes, but we only care about the first 4.
3169 for (int i = 3; i >= 0; --i) {
3170 int Val = SVOp->getMaskElt(i);
3179 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3180 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3181 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3182 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3183 EVT VVT = N->getValueType(0);
3184 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3188 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3189 Val = SVOp->getMaskElt(i);
3193 return (Val - i) * EltSize;
3196 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3198 bool X86::isZeroNode(SDValue Elt) {
3199 return ((isa<ConstantSDNode>(Elt) &&
3200 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3201 (isa<ConstantFPSDNode>(Elt) &&
3202 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3205 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3206 /// their permute mask.
3207 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3208 SelectionDAG &DAG) {
3209 EVT VT = SVOp->getValueType(0);
3210 unsigned NumElems = VT.getVectorNumElements();
3211 SmallVector<int, 8> MaskVec;
3213 for (unsigned i = 0; i != NumElems; ++i) {
3214 int idx = SVOp->getMaskElt(i);
3216 MaskVec.push_back(idx);
3217 else if (idx < (int)NumElems)
3218 MaskVec.push_back(idx + NumElems);
3220 MaskVec.push_back(idx - NumElems);
3222 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3223 SVOp->getOperand(0), &MaskVec[0]);
3226 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3227 /// the two vector operands have swapped position.
3228 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3229 unsigned NumElems = VT.getVectorNumElements();
3230 for (unsigned i = 0; i != NumElems; ++i) {
3234 else if (idx < (int)NumElems)
3235 Mask[i] = idx + NumElems;
3237 Mask[i] = idx - NumElems;
3241 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3242 /// match movhlps. The lower half elements should come from upper half of
3243 /// V1 (and in order), and the upper half elements should come from the upper
3244 /// half of V2 (and in order).
3245 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3246 if (Op->getValueType(0).getVectorNumElements() != 4)
3248 for (unsigned i = 0, e = 2; i != e; ++i)
3249 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3251 for (unsigned i = 2; i != 4; ++i)
3252 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3257 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3258 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3260 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3261 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3263 N = N->getOperand(0).getNode();
3264 if (!ISD::isNON_EXTLoad(N))
3267 *LD = cast<LoadSDNode>(N);
3271 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3272 /// match movlp{s|d}. The lower half elements should come from lower half of
3273 /// V1 (and in order), and the upper half elements should come from the upper
3274 /// half of V2 (and in order). And since V1 will become the source of the
3275 /// MOVLP, it must be either a vector load or a scalar load to vector.
3276 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3277 ShuffleVectorSDNode *Op) {
3278 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3280 // Is V2 is a vector load, don't do this transformation. We will try to use
3281 // load folding shufps op.
3282 if (ISD::isNON_EXTLoad(V2))
3285 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3287 if (NumElems != 2 && NumElems != 4)
3289 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3290 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3292 for (unsigned i = NumElems/2; i != NumElems; ++i)
3293 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3298 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3300 static bool isSplatVector(SDNode *N) {
3301 if (N->getOpcode() != ISD::BUILD_VECTOR)
3304 SDValue SplatValue = N->getOperand(0);
3305 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3306 if (N->getOperand(i) != SplatValue)
3311 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3312 /// to an zero vector.
3313 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3314 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3315 SDValue V1 = N->getOperand(0);
3316 SDValue V2 = N->getOperand(1);
3317 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3318 for (unsigned i = 0; i != NumElems; ++i) {
3319 int Idx = N->getMaskElt(i);
3320 if (Idx >= (int)NumElems) {
3321 unsigned Opc = V2.getOpcode();
3322 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3324 if (Opc != ISD::BUILD_VECTOR ||
3325 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3327 } else if (Idx >= 0) {
3328 unsigned Opc = V1.getOpcode();
3329 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3331 if (Opc != ISD::BUILD_VECTOR ||
3332 !X86::isZeroNode(V1.getOperand(Idx)))
3339 /// getZeroVector - Returns a vector of specified type with all zero elements.
3341 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3343 assert(VT.isVector() && "Expected a vector type");
3345 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3346 // type. This ensures they get CSE'd.
3348 if (VT.getSizeInBits() == 64) { // MMX
3349 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3350 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3351 } else if (HasSSE2) { // SSE2
3352 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3353 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3355 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3356 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3358 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3361 /// getOnesVector - Returns a vector of specified type with all bits set.
3363 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3364 assert(VT.isVector() && "Expected a vector type");
3366 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3367 // type. This ensures they get CSE'd.
3368 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3370 if (VT.getSizeInBits() == 64) // MMX
3371 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3373 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3374 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3378 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3379 /// that point to V2 points to its first element.
3380 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3381 EVT VT = SVOp->getValueType(0);
3382 unsigned NumElems = VT.getVectorNumElements();
3384 bool Changed = false;
3385 SmallVector<int, 8> MaskVec;
3386 SVOp->getMask(MaskVec);
3388 for (unsigned i = 0; i != NumElems; ++i) {
3389 if (MaskVec[i] > (int)NumElems) {
3390 MaskVec[i] = NumElems;
3395 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3396 SVOp->getOperand(1), &MaskVec[0]);
3397 return SDValue(SVOp, 0);
3400 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3401 /// operation of specified width.
3402 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3404 unsigned NumElems = VT.getVectorNumElements();
3405 SmallVector<int, 8> Mask;
3406 Mask.push_back(NumElems);
3407 for (unsigned i = 1; i != NumElems; ++i)
3409 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3412 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3413 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3415 unsigned NumElems = VT.getVectorNumElements();
3416 SmallVector<int, 8> Mask;
3417 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3419 Mask.push_back(i + NumElems);
3421 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3424 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3425 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3427 unsigned NumElems = VT.getVectorNumElements();
3428 unsigned Half = NumElems/2;
3429 SmallVector<int, 8> Mask;
3430 for (unsigned i = 0; i != Half; ++i) {
3431 Mask.push_back(i + Half);
3432 Mask.push_back(i + NumElems + Half);
3434 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3437 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3438 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3440 if (SV->getValueType(0).getVectorNumElements() <= 4)
3441 return SDValue(SV, 0);
3443 EVT PVT = MVT::v4f32;
3444 EVT VT = SV->getValueType(0);
3445 DebugLoc dl = SV->getDebugLoc();
3446 SDValue V1 = SV->getOperand(0);
3447 int NumElems = VT.getVectorNumElements();
3448 int EltNo = SV->getSplatIndex();
3450 // unpack elements to the correct location
3451 while (NumElems > 4) {
3452 if (EltNo < NumElems/2) {
3453 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3455 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3456 EltNo -= NumElems/2;
3461 // Perform the splat.
3462 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3463 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3464 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3465 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3468 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3469 /// vector of zero or undef vector. This produces a shuffle where the low
3470 /// element of V2 is swizzled into the zero/undef vector, landing at element
3471 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3472 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3473 bool isZero, bool HasSSE2,
3474 SelectionDAG &DAG) {
3475 EVT VT = V2.getValueType();
3477 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3478 unsigned NumElems = VT.getVectorNumElements();
3479 SmallVector<int, 16> MaskVec;
3480 for (unsigned i = 0; i != NumElems; ++i)
3481 // If this is the insertion idx, put the low elt of V2 here.
3482 MaskVec.push_back(i == Idx ? NumElems : i);
3483 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3486 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3487 /// a shuffle that is zero.
3489 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3490 bool Low, SelectionDAG &DAG) {
3491 unsigned NumZeros = 0;
3492 for (int i = 0; i < NumElems; ++i) {
3493 unsigned Index = Low ? i : NumElems-i-1;
3494 int Idx = SVOp->getMaskElt(Index);
3499 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3500 if (Elt.getNode() && X86::isZeroNode(Elt))
3508 /// isVectorShift - Returns true if the shuffle can be implemented as a
3509 /// logical left or right shift of a vector.
3510 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3511 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3512 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3513 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3516 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3519 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3523 bool SeenV1 = false;
3524 bool SeenV2 = false;
3525 for (unsigned i = NumZeros; i < NumElems; ++i) {
3526 unsigned Val = isLeft ? (i - NumZeros) : i;
3527 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3530 unsigned Idx = (unsigned) Idx_;
3540 if (SeenV1 && SeenV2)
3543 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3549 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3551 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3552 unsigned NumNonZero, unsigned NumZero,
3554 const TargetLowering &TLI) {
3558 DebugLoc dl = Op.getDebugLoc();
3561 for (unsigned i = 0; i < 16; ++i) {
3562 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3563 if (ThisIsNonZero && First) {
3565 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3567 V = DAG.getUNDEF(MVT::v8i16);
3572 SDValue ThisElt(0, 0), LastElt(0, 0);
3573 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3574 if (LastIsNonZero) {
3575 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3576 MVT::i16, Op.getOperand(i-1));
3578 if (ThisIsNonZero) {
3579 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3580 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3581 ThisElt, DAG.getConstant(8, MVT::i8));
3583 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3587 if (ThisElt.getNode())
3588 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3589 DAG.getIntPtrConstant(i/2));
3593 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3596 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3598 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3599 unsigned NumNonZero, unsigned NumZero,
3601 const TargetLowering &TLI) {
3605 DebugLoc dl = Op.getDebugLoc();
3608 for (unsigned i = 0; i < 8; ++i) {
3609 bool isNonZero = (NonZeros & (1 << i)) != 0;
3613 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3615 V = DAG.getUNDEF(MVT::v8i16);
3618 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3619 MVT::v8i16, V, Op.getOperand(i),
3620 DAG.getIntPtrConstant(i));
3627 /// getVShift - Return a vector logical shift node.
3629 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3630 unsigned NumBits, SelectionDAG &DAG,
3631 const TargetLowering &TLI, DebugLoc dl) {
3632 bool isMMX = VT.getSizeInBits() == 64;
3633 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3634 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3635 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3636 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3637 DAG.getNode(Opc, dl, ShVT, SrcOp,
3638 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3642 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3643 SelectionDAG &DAG) const {
3645 // Check if the scalar load can be widened into a vector load. And if
3646 // the address is "base + cst" see if the cst can be "absorbed" into
3647 // the shuffle mask.
3648 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3649 SDValue Ptr = LD->getBasePtr();
3650 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3652 EVT PVT = LD->getValueType(0);
3653 if (PVT != MVT::i32 && PVT != MVT::f32)
3658 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3659 FI = FINode->getIndex();
3661 } else if (Ptr.getOpcode() == ISD::ADD &&
3662 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3663 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3664 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3665 Offset = Ptr.getConstantOperandVal(1);
3666 Ptr = Ptr.getOperand(0);
3671 SDValue Chain = LD->getChain();
3672 // Make sure the stack object alignment is at least 16.
3673 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3674 if (DAG.InferPtrAlignment(Ptr) < 16) {
3675 if (MFI->isFixedObjectIndex(FI)) {
3676 // Can't change the alignment. FIXME: It's possible to compute
3677 // the exact stack offset and reference FI + adjust offset instead.
3678 // If someone *really* cares about this. That's the way to implement it.
3681 MFI->setObjectAlignment(FI, 16);
3685 // (Offset % 16) must be multiple of 4. Then address is then
3686 // Ptr + (Offset & ~15).
3689 if ((Offset % 16) & 3)
3691 int64_t StartOffset = Offset & ~15;
3693 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3694 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3696 int EltNo = (Offset - StartOffset) >> 2;
3697 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3698 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3699 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3701 // Canonicalize it to a v4i32 shuffle.
3702 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3703 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3704 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3705 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3711 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3712 /// vector of type 'VT', see if the elements can be replaced by a single large
3713 /// load which has the same value as a build_vector whose operands are 'elts'.
3715 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3717 /// FIXME: we'd also like to handle the case where the last elements are zero
3718 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3719 /// There's even a handy isZeroNode for that purpose.
3720 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3721 DebugLoc &dl, SelectionDAG &DAG) {
3722 EVT EltVT = VT.getVectorElementType();
3723 unsigned NumElems = Elts.size();
3725 LoadSDNode *LDBase = NULL;
3726 unsigned LastLoadedElt = -1U;
3728 // For each element in the initializer, see if we've found a load or an undef.
3729 // If we don't find an initial load element, or later load elements are
3730 // non-consecutive, bail out.
3731 for (unsigned i = 0; i < NumElems; ++i) {
3732 SDValue Elt = Elts[i];
3734 if (!Elt.getNode() ||
3735 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3738 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3740 LDBase = cast<LoadSDNode>(Elt.getNode());
3744 if (Elt.getOpcode() == ISD::UNDEF)
3747 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3748 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3753 // If we have found an entire vector of loads and undefs, then return a large
3754 // load of the entire vector width starting at the base pointer. If we found
3755 // consecutive loads for the low half, generate a vzext_load node.
3756 if (LastLoadedElt == NumElems - 1) {
3757 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3758 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3759 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3760 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3761 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3762 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3763 LDBase->isVolatile(), LDBase->isNonTemporal(),
3764 LDBase->getAlignment());
3765 } else if (NumElems == 4 && LastLoadedElt == 1) {
3766 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3767 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3768 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3769 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3775 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3776 DebugLoc dl = Op.getDebugLoc();
3777 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3778 if (ISD::isBuildVectorAllZeros(Op.getNode())
3779 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3780 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3781 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3782 // eliminated on x86-32 hosts.
3783 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3786 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3787 return getOnesVector(Op.getValueType(), DAG, dl);
3788 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3791 EVT VT = Op.getValueType();
3792 EVT ExtVT = VT.getVectorElementType();
3793 unsigned EVTBits = ExtVT.getSizeInBits();
3795 unsigned NumElems = Op.getNumOperands();
3796 unsigned NumZero = 0;
3797 unsigned NumNonZero = 0;
3798 unsigned NonZeros = 0;
3799 bool IsAllConstants = true;
3800 SmallSet<SDValue, 8> Values;
3801 for (unsigned i = 0; i < NumElems; ++i) {
3802 SDValue Elt = Op.getOperand(i);
3803 if (Elt.getOpcode() == ISD::UNDEF)
3806 if (Elt.getOpcode() != ISD::Constant &&
3807 Elt.getOpcode() != ISD::ConstantFP)
3808 IsAllConstants = false;
3809 if (X86::isZeroNode(Elt))
3812 NonZeros |= (1 << i);
3817 if (NumNonZero == 0) {
3818 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3819 return DAG.getUNDEF(VT);
3822 // Special case for single non-zero, non-undef, element.
3823 if (NumNonZero == 1) {
3824 unsigned Idx = CountTrailingZeros_32(NonZeros);
3825 SDValue Item = Op.getOperand(Idx);
3827 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3828 // the value are obviously zero, truncate the value to i32 and do the
3829 // insertion that way. Only do this if the value is non-constant or if the
3830 // value is a constant being inserted into element 0. It is cheaper to do
3831 // a constant pool load than it is to do a movd + shuffle.
3832 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3833 (!IsAllConstants || Idx == 0)) {
3834 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3835 // Handle MMX and SSE both.
3836 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3837 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3839 // Truncate the value (which may itself be a constant) to i32, and
3840 // convert it to a vector with movd (S2V+shuffle to zero extend).
3841 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3842 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3843 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3844 Subtarget->hasSSE2(), DAG);
3846 // Now we have our 32-bit value zero extended in the low element of
3847 // a vector. If Idx != 0, swizzle it into place.
3849 SmallVector<int, 4> Mask;
3850 Mask.push_back(Idx);
3851 for (unsigned i = 1; i != VecElts; ++i)
3853 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3854 DAG.getUNDEF(Item.getValueType()),
3857 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3861 // If we have a constant or non-constant insertion into the low element of
3862 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3863 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3864 // depending on what the source datatype is.
3867 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3868 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3869 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3870 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3871 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3872 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3874 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3875 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3876 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3877 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3878 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3879 Subtarget->hasSSE2(), DAG);
3880 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3884 // Is it a vector logical left shift?
3885 if (NumElems == 2 && Idx == 1 &&
3886 X86::isZeroNode(Op.getOperand(0)) &&
3887 !X86::isZeroNode(Op.getOperand(1))) {
3888 unsigned NumBits = VT.getSizeInBits();
3889 return getVShift(true, VT,
3890 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3891 VT, Op.getOperand(1)),
3892 NumBits/2, DAG, *this, dl);
3895 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3898 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3899 // is a non-constant being inserted into an element other than the low one,
3900 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3901 // movd/movss) to move this into the low element, then shuffle it into
3903 if (EVTBits == 32) {
3904 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3906 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3907 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3908 Subtarget->hasSSE2(), DAG);
3909 SmallVector<int, 8> MaskVec;
3910 for (unsigned i = 0; i < NumElems; i++)
3911 MaskVec.push_back(i == Idx ? 0 : 1);
3912 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3916 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3917 if (Values.size() == 1) {
3918 if (EVTBits == 32) {
3919 // Instead of a shuffle like this:
3920 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3921 // Check if it's possible to issue this instead.
3922 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3923 unsigned Idx = CountTrailingZeros_32(NonZeros);
3924 SDValue Item = Op.getOperand(Idx);
3925 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3926 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3931 // A vector full of immediates; various special cases are already
3932 // handled, so this is best done with a single constant-pool load.
3936 // Let legalizer expand 2-wide build_vectors.
3937 if (EVTBits == 64) {
3938 if (NumNonZero == 1) {
3939 // One half is zero or undef.
3940 unsigned Idx = CountTrailingZeros_32(NonZeros);
3941 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3942 Op.getOperand(Idx));
3943 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3944 Subtarget->hasSSE2(), DAG);
3949 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3950 if (EVTBits == 8 && NumElems == 16) {
3951 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3953 if (V.getNode()) return V;
3956 if (EVTBits == 16 && NumElems == 8) {
3957 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3959 if (V.getNode()) return V;
3962 // If element VT is == 32 bits, turn it into a number of shuffles.
3963 SmallVector<SDValue, 8> V;
3965 if (NumElems == 4 && NumZero > 0) {
3966 for (unsigned i = 0; i < 4; ++i) {
3967 bool isZero = !(NonZeros & (1 << i));
3969 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3971 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3974 for (unsigned i = 0; i < 2; ++i) {
3975 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3978 V[i] = V[i*2]; // Must be a zero vector.
3981 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3984 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3987 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3992 SmallVector<int, 8> MaskVec;
3993 bool Reverse = (NonZeros & 0x3) == 2;
3994 for (unsigned i = 0; i < 2; ++i)
3995 MaskVec.push_back(Reverse ? 1-i : i);
3996 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3997 for (unsigned i = 0; i < 2; ++i)
3998 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3999 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4002 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4003 // Check for a build vector of consecutive loads.
4004 for (unsigned i = 0; i < NumElems; ++i)
4005 V[i] = Op.getOperand(i);
4007 // Check for elements which are consecutive loads.
4008 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4012 // For SSE 4.1, use inserts into undef.
4013 if (getSubtarget()->hasSSE41()) {
4014 V[0] = DAG.getUNDEF(VT);
4015 for (unsigned i = 0; i < NumElems; ++i)
4016 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4017 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
4018 Op.getOperand(i), DAG.getIntPtrConstant(i));
4022 // Otherwise, expand into a number of unpckl*
4024 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4025 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4026 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4027 for (unsigned i = 0; i < NumElems; ++i)
4028 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4030 while (NumElems != 0) {
4031 for (unsigned i = 0; i < NumElems; ++i)
4032 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4041 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4042 // We support concatenate two MMX registers and place them in a MMX
4043 // register. This is better than doing a stack convert.
4044 DebugLoc dl = Op.getDebugLoc();
4045 EVT ResVT = Op.getValueType();
4046 assert(Op.getNumOperands() == 2);
4047 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4048 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4050 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4051 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4052 InVec = Op.getOperand(1);
4053 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4054 unsigned NumElts = ResVT.getVectorNumElements();
4055 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4056 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4057 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4059 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4060 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4061 Mask[0] = 0; Mask[1] = 2;
4062 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4064 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4067 // v8i16 shuffles - Prefer shuffles in the following order:
4068 // 1. [all] pshuflw, pshufhw, optional move
4069 // 2. [ssse3] 1 x pshufb
4070 // 3. [ssse3] 2 x pshufb + 1 x por
4071 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4073 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4075 const X86TargetLowering &TLI) {
4076 SDValue V1 = SVOp->getOperand(0);
4077 SDValue V2 = SVOp->getOperand(1);
4078 DebugLoc dl = SVOp->getDebugLoc();
4079 SmallVector<int, 8> MaskVals;
4081 // Determine if more than 1 of the words in each of the low and high quadwords
4082 // of the result come from the same quadword of one of the two inputs. Undef
4083 // mask values count as coming from any quadword, for better codegen.
4084 SmallVector<unsigned, 4> LoQuad(4);
4085 SmallVector<unsigned, 4> HiQuad(4);
4086 BitVector InputQuads(4);
4087 for (unsigned i = 0; i < 8; ++i) {
4088 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4089 int EltIdx = SVOp->getMaskElt(i);
4090 MaskVals.push_back(EltIdx);
4099 InputQuads.set(EltIdx / 4);
4102 int BestLoQuad = -1;
4103 unsigned MaxQuad = 1;
4104 for (unsigned i = 0; i < 4; ++i) {
4105 if (LoQuad[i] > MaxQuad) {
4107 MaxQuad = LoQuad[i];
4111 int BestHiQuad = -1;
4113 for (unsigned i = 0; i < 4; ++i) {
4114 if (HiQuad[i] > MaxQuad) {
4116 MaxQuad = HiQuad[i];
4120 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4121 // of the two input vectors, shuffle them into one input vector so only a
4122 // single pshufb instruction is necessary. If There are more than 2 input
4123 // quads, disable the next transformation since it does not help SSSE3.
4124 bool V1Used = InputQuads[0] || InputQuads[1];
4125 bool V2Used = InputQuads[2] || InputQuads[3];
4126 if (TLI.getSubtarget()->hasSSSE3()) {
4127 if (InputQuads.count() == 2 && V1Used && V2Used) {
4128 BestLoQuad = InputQuads.find_first();
4129 BestHiQuad = InputQuads.find_next(BestLoQuad);
4131 if (InputQuads.count() > 2) {
4137 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4138 // the shuffle mask. If a quad is scored as -1, that means that it contains
4139 // words from all 4 input quadwords.
4141 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4142 SmallVector<int, 8> MaskV;
4143 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4144 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4145 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4146 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4147 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4148 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4150 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4151 // source words for the shuffle, to aid later transformations.
4152 bool AllWordsInNewV = true;
4153 bool InOrder[2] = { true, true };
4154 for (unsigned i = 0; i != 8; ++i) {
4155 int idx = MaskVals[i];
4157 InOrder[i/4] = false;
4158 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4160 AllWordsInNewV = false;
4164 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4165 if (AllWordsInNewV) {
4166 for (int i = 0; i != 8; ++i) {
4167 int idx = MaskVals[i];
4170 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4171 if ((idx != i) && idx < 4)
4173 if ((idx != i) && idx > 3)
4182 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4183 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4184 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4185 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4186 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4190 // If we have SSSE3, and all words of the result are from 1 input vector,
4191 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4192 // is present, fall back to case 4.
4193 if (TLI.getSubtarget()->hasSSSE3()) {
4194 SmallVector<SDValue,16> pshufbMask;
4196 // If we have elements from both input vectors, set the high bit of the
4197 // shuffle mask element to zero out elements that come from V2 in the V1
4198 // mask, and elements that come from V1 in the V2 mask, so that the two
4199 // results can be OR'd together.
4200 bool TwoInputs = V1Used && V2Used;
4201 for (unsigned i = 0; i != 8; ++i) {
4202 int EltIdx = MaskVals[i] * 2;
4203 if (TwoInputs && (EltIdx >= 16)) {
4204 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4205 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4208 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4209 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4211 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4212 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4213 DAG.getNode(ISD::BUILD_VECTOR, dl,
4214 MVT::v16i8, &pshufbMask[0], 16));
4216 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4218 // Calculate the shuffle mask for the second input, shuffle it, and
4219 // OR it with the first shuffled input.
4221 for (unsigned i = 0; i != 8; ++i) {
4222 int EltIdx = MaskVals[i] * 2;
4224 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4225 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4228 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4229 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4231 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4232 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4233 DAG.getNode(ISD::BUILD_VECTOR, dl,
4234 MVT::v16i8, &pshufbMask[0], 16));
4235 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4236 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4239 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4240 // and update MaskVals with new element order.
4241 BitVector InOrder(8);
4242 if (BestLoQuad >= 0) {
4243 SmallVector<int, 8> MaskV;
4244 for (int i = 0; i != 4; ++i) {
4245 int idx = MaskVals[i];
4247 MaskV.push_back(-1);
4249 } else if ((idx / 4) == BestLoQuad) {
4250 MaskV.push_back(idx & 3);
4253 MaskV.push_back(-1);
4256 for (unsigned i = 4; i != 8; ++i)
4258 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4262 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4263 // and update MaskVals with the new element order.
4264 if (BestHiQuad >= 0) {
4265 SmallVector<int, 8> MaskV;
4266 for (unsigned i = 0; i != 4; ++i)
4268 for (unsigned i = 4; i != 8; ++i) {
4269 int idx = MaskVals[i];
4271 MaskV.push_back(-1);
4273 } else if ((idx / 4) == BestHiQuad) {
4274 MaskV.push_back((idx & 3) + 4);
4277 MaskV.push_back(-1);
4280 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4284 // In case BestHi & BestLo were both -1, which means each quadword has a word
4285 // from each of the four input quadwords, calculate the InOrder bitvector now
4286 // before falling through to the insert/extract cleanup.
4287 if (BestLoQuad == -1 && BestHiQuad == -1) {
4289 for (int i = 0; i != 8; ++i)
4290 if (MaskVals[i] < 0 || MaskVals[i] == i)
4294 // The other elements are put in the right place using pextrw and pinsrw.
4295 for (unsigned i = 0; i != 8; ++i) {
4298 int EltIdx = MaskVals[i];
4301 SDValue ExtOp = (EltIdx < 8)
4302 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4303 DAG.getIntPtrConstant(EltIdx))
4304 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4305 DAG.getIntPtrConstant(EltIdx - 8));
4306 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4307 DAG.getIntPtrConstant(i));
4312 // v16i8 shuffles - Prefer shuffles in the following order:
4313 // 1. [ssse3] 1 x pshufb
4314 // 2. [ssse3] 2 x pshufb + 1 x por
4315 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4317 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4319 const X86TargetLowering &TLI) {
4320 SDValue V1 = SVOp->getOperand(0);
4321 SDValue V2 = SVOp->getOperand(1);
4322 DebugLoc dl = SVOp->getDebugLoc();
4323 SmallVector<int, 16> MaskVals;
4324 SVOp->getMask(MaskVals);
4326 // If we have SSSE3, case 1 is generated when all result bytes come from
4327 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4328 // present, fall back to case 3.
4329 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4332 for (unsigned i = 0; i < 16; ++i) {
4333 int EltIdx = MaskVals[i];
4342 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4343 if (TLI.getSubtarget()->hasSSSE3()) {
4344 SmallVector<SDValue,16> pshufbMask;
4346 // If all result elements are from one input vector, then only translate
4347 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4349 // Otherwise, we have elements from both input vectors, and must zero out
4350 // elements that come from V2 in the first mask, and V1 in the second mask
4351 // so that we can OR them together.
4352 bool TwoInputs = !(V1Only || V2Only);
4353 for (unsigned i = 0; i != 16; ++i) {
4354 int EltIdx = MaskVals[i];
4355 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4356 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4359 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4361 // If all the elements are from V2, assign it to V1 and return after
4362 // building the first pshufb.
4365 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4366 DAG.getNode(ISD::BUILD_VECTOR, dl,
4367 MVT::v16i8, &pshufbMask[0], 16));
4371 // Calculate the shuffle mask for the second input, shuffle it, and
4372 // OR it with the first shuffled input.
4374 for (unsigned i = 0; i != 16; ++i) {
4375 int EltIdx = MaskVals[i];
4377 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4380 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4382 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4383 DAG.getNode(ISD::BUILD_VECTOR, dl,
4384 MVT::v16i8, &pshufbMask[0], 16));
4385 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4388 // No SSSE3 - Calculate in place words and then fix all out of place words
4389 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4390 // the 16 different words that comprise the two doublequadword input vectors.
4391 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4392 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4393 SDValue NewV = V2Only ? V2 : V1;
4394 for (int i = 0; i != 8; ++i) {
4395 int Elt0 = MaskVals[i*2];
4396 int Elt1 = MaskVals[i*2+1];
4398 // This word of the result is all undef, skip it.
4399 if (Elt0 < 0 && Elt1 < 0)
4402 // This word of the result is already in the correct place, skip it.
4403 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4405 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4408 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4409 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4412 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4413 // using a single extract together, load it and store it.
4414 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4415 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4416 DAG.getIntPtrConstant(Elt1 / 2));
4417 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4418 DAG.getIntPtrConstant(i));
4422 // If Elt1 is defined, extract it from the appropriate source. If the
4423 // source byte is not also odd, shift the extracted word left 8 bits
4424 // otherwise clear the bottom 8 bits if we need to do an or.
4426 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4427 DAG.getIntPtrConstant(Elt1 / 2));
4428 if ((Elt1 & 1) == 0)
4429 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4430 DAG.getConstant(8, TLI.getShiftAmountTy()));
4432 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4433 DAG.getConstant(0xFF00, MVT::i16));
4435 // If Elt0 is defined, extract it from the appropriate source. If the
4436 // source byte is not also even, shift the extracted word right 8 bits. If
4437 // Elt1 was also defined, OR the extracted values together before
4438 // inserting them in the result.
4440 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4441 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4442 if ((Elt0 & 1) != 0)
4443 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4444 DAG.getConstant(8, TLI.getShiftAmountTy()));
4446 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4447 DAG.getConstant(0x00FF, MVT::i16));
4448 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4451 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4452 DAG.getIntPtrConstant(i));
4454 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4457 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4458 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4459 /// done when every pair / quad of shuffle mask elements point to elements in
4460 /// the right sequence. e.g.
4461 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4463 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4465 const TargetLowering &TLI, DebugLoc dl) {
4466 EVT VT = SVOp->getValueType(0);
4467 SDValue V1 = SVOp->getOperand(0);
4468 SDValue V2 = SVOp->getOperand(1);
4469 unsigned NumElems = VT.getVectorNumElements();
4470 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4471 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4473 switch (VT.getSimpleVT().SimpleTy) {
4474 default: assert(false && "Unexpected!");
4475 case MVT::v4f32: NewVT = MVT::v2f64; break;
4476 case MVT::v4i32: NewVT = MVT::v2i64; break;
4477 case MVT::v8i16: NewVT = MVT::v4i32; break;
4478 case MVT::v16i8: NewVT = MVT::v4i32; break;
4481 if (NewWidth == 2) {
4487 int Scale = NumElems / NewWidth;
4488 SmallVector<int, 8> MaskVec;
4489 for (unsigned i = 0; i < NumElems; i += Scale) {
4491 for (int j = 0; j < Scale; ++j) {
4492 int EltIdx = SVOp->getMaskElt(i+j);
4496 StartIdx = EltIdx - (EltIdx % Scale);
4497 if (EltIdx != StartIdx + j)
4501 MaskVec.push_back(-1);
4503 MaskVec.push_back(StartIdx / Scale);
4506 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4507 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4508 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4511 /// getVZextMovL - Return a zero-extending vector move low node.
4513 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4514 SDValue SrcOp, SelectionDAG &DAG,
4515 const X86Subtarget *Subtarget, DebugLoc dl) {
4516 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4517 LoadSDNode *LD = NULL;
4518 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4519 LD = dyn_cast<LoadSDNode>(SrcOp);
4521 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4523 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4524 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4525 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4526 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4527 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4529 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4530 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4531 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4532 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4540 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4541 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4542 DAG.getNode(ISD::BIT_CONVERT, dl,
4546 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4549 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4550 SDValue V1 = SVOp->getOperand(0);
4551 SDValue V2 = SVOp->getOperand(1);
4552 DebugLoc dl = SVOp->getDebugLoc();
4553 EVT VT = SVOp->getValueType(0);
4555 SmallVector<std::pair<int, int>, 8> Locs;
4557 SmallVector<int, 8> Mask1(4U, -1);
4558 SmallVector<int, 8> PermMask;
4559 SVOp->getMask(PermMask);
4563 for (unsigned i = 0; i != 4; ++i) {
4564 int Idx = PermMask[i];
4566 Locs[i] = std::make_pair(-1, -1);
4568 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4570 Locs[i] = std::make_pair(0, NumLo);
4574 Locs[i] = std::make_pair(1, NumHi);
4576 Mask1[2+NumHi] = Idx;
4582 if (NumLo <= 2 && NumHi <= 2) {
4583 // If no more than two elements come from either vector. This can be
4584 // implemented with two shuffles. First shuffle gather the elements.
4585 // The second shuffle, which takes the first shuffle as both of its
4586 // vector operands, put the elements into the right order.
4587 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4589 SmallVector<int, 8> Mask2(4U, -1);
4591 for (unsigned i = 0; i != 4; ++i) {
4592 if (Locs[i].first == -1)
4595 unsigned Idx = (i < 2) ? 0 : 4;
4596 Idx += Locs[i].first * 2 + Locs[i].second;
4601 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4602 } else if (NumLo == 3 || NumHi == 3) {
4603 // Otherwise, we must have three elements from one vector, call it X, and
4604 // one element from the other, call it Y. First, use a shufps to build an
4605 // intermediate vector with the one element from Y and the element from X
4606 // that will be in the same half in the final destination (the indexes don't
4607 // matter). Then, use a shufps to build the final vector, taking the half
4608 // containing the element from Y from the intermediate, and the other half
4611 // Normalize it so the 3 elements come from V1.
4612 CommuteVectorShuffleMask(PermMask, VT);
4616 // Find the element from V2.
4618 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4619 int Val = PermMask[HiIndex];
4626 Mask1[0] = PermMask[HiIndex];
4628 Mask1[2] = PermMask[HiIndex^1];
4630 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4633 Mask1[0] = PermMask[0];
4634 Mask1[1] = PermMask[1];
4635 Mask1[2] = HiIndex & 1 ? 6 : 4;
4636 Mask1[3] = HiIndex & 1 ? 4 : 6;
4637 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4639 Mask1[0] = HiIndex & 1 ? 2 : 0;
4640 Mask1[1] = HiIndex & 1 ? 0 : 2;
4641 Mask1[2] = PermMask[2];
4642 Mask1[3] = PermMask[3];
4647 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4651 // Break it into (shuffle shuffle_hi, shuffle_lo).
4653 SmallVector<int,8> LoMask(4U, -1);
4654 SmallVector<int,8> HiMask(4U, -1);
4656 SmallVector<int,8> *MaskPtr = &LoMask;
4657 unsigned MaskIdx = 0;
4660 for (unsigned i = 0; i != 4; ++i) {
4667 int Idx = PermMask[i];
4669 Locs[i] = std::make_pair(-1, -1);
4670 } else if (Idx < 4) {
4671 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4672 (*MaskPtr)[LoIdx] = Idx;
4675 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4676 (*MaskPtr)[HiIdx] = Idx;
4681 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4682 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4683 SmallVector<int, 8> MaskOps;
4684 for (unsigned i = 0; i != 4; ++i) {
4685 if (Locs[i].first == -1) {
4686 MaskOps.push_back(-1);
4688 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4689 MaskOps.push_back(Idx);
4692 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4696 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4697 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4698 SDValue V1 = Op.getOperand(0);
4699 SDValue V2 = Op.getOperand(1);
4700 EVT VT = Op.getValueType();
4701 DebugLoc dl = Op.getDebugLoc();
4702 unsigned NumElems = VT.getVectorNumElements();
4703 bool isMMX = VT.getSizeInBits() == 64;
4704 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4705 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4706 bool V1IsSplat = false;
4707 bool V2IsSplat = false;
4709 if (isZeroShuffle(SVOp))
4710 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4712 // Promote splats to v4f32.
4713 if (SVOp->isSplat()) {
4714 if (isMMX || NumElems < 4)
4716 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4719 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4721 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4722 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4723 if (NewOp.getNode())
4724 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4725 LowerVECTOR_SHUFFLE(NewOp, DAG));
4726 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4727 // FIXME: Figure out a cleaner way to do this.
4728 // Try to make use of movq to zero out the top part.
4729 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4730 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4731 if (NewOp.getNode()) {
4732 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4733 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4734 DAG, Subtarget, dl);
4736 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4737 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4738 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4739 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4740 DAG, Subtarget, dl);
4744 if (X86::isPSHUFDMask(SVOp))
4747 // Check if this can be converted into a logical shift.
4748 bool isLeft = false;
4751 bool isShift = getSubtarget()->hasSSE2() &&
4752 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4753 if (isShift && ShVal.hasOneUse()) {
4754 // If the shifted value has multiple uses, it may be cheaper to use
4755 // v_set0 + movlhps or movhlps, etc.
4756 EVT EltVT = VT.getVectorElementType();
4757 ShAmt *= EltVT.getSizeInBits();
4758 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4761 if (X86::isMOVLMask(SVOp)) {
4764 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4765 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4770 // FIXME: fold these into legal mask.
4771 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4772 X86::isMOVSLDUPMask(SVOp) ||
4773 X86::isMOVHLPSMask(SVOp) ||
4774 X86::isMOVLHPSMask(SVOp) ||
4775 X86::isMOVLPMask(SVOp)))
4778 if (ShouldXformToMOVHLPS(SVOp) ||
4779 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4780 return CommuteVectorShuffle(SVOp, DAG);
4783 // No better options. Use a vshl / vsrl.
4784 EVT EltVT = VT.getVectorElementType();
4785 ShAmt *= EltVT.getSizeInBits();
4786 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4789 bool Commuted = false;
4790 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4791 // 1,1,1,1 -> v8i16 though.
4792 V1IsSplat = isSplatVector(V1.getNode());
4793 V2IsSplat = isSplatVector(V2.getNode());
4795 // Canonicalize the splat or undef, if present, to be on the RHS.
4796 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4797 Op = CommuteVectorShuffle(SVOp, DAG);
4798 SVOp = cast<ShuffleVectorSDNode>(Op);
4799 V1 = SVOp->getOperand(0);
4800 V2 = SVOp->getOperand(1);
4801 std::swap(V1IsSplat, V2IsSplat);
4802 std::swap(V1IsUndef, V2IsUndef);
4806 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4807 // Shuffling low element of v1 into undef, just return v1.
4810 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4811 // the instruction selector will not match, so get a canonical MOVL with
4812 // swapped operands to undo the commute.
4813 return getMOVL(DAG, dl, VT, V2, V1);
4816 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4817 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4818 X86::isUNPCKLMask(SVOp) ||
4819 X86::isUNPCKHMask(SVOp))
4823 // Normalize mask so all entries that point to V2 points to its first
4824 // element then try to match unpck{h|l} again. If match, return a
4825 // new vector_shuffle with the corrected mask.
4826 SDValue NewMask = NormalizeMask(SVOp, DAG);
4827 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4828 if (NSVOp != SVOp) {
4829 if (X86::isUNPCKLMask(NSVOp, true)) {
4831 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4838 // Commute is back and try unpck* again.
4839 // FIXME: this seems wrong.
4840 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4841 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4842 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4843 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4844 X86::isUNPCKLMask(NewSVOp) ||
4845 X86::isUNPCKHMask(NewSVOp))
4849 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4851 // Normalize the node to match x86 shuffle ops if needed
4852 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4853 return CommuteVectorShuffle(SVOp, DAG);
4855 // Check for legal shuffle and return?
4856 SmallVector<int, 16> PermMask;
4857 SVOp->getMask(PermMask);
4858 if (isShuffleMaskLegal(PermMask, VT))
4861 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4862 if (VT == MVT::v8i16) {
4863 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4864 if (NewOp.getNode())
4868 if (VT == MVT::v16i8) {
4869 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4870 if (NewOp.getNode())
4874 // Handle all 4 wide cases with a number of shuffles except for MMX.
4875 if (NumElems == 4 && !isMMX)
4876 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4882 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4883 SelectionDAG &DAG) const {
4884 EVT VT = Op.getValueType();
4885 DebugLoc dl = Op.getDebugLoc();
4886 if (VT.getSizeInBits() == 8) {
4887 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4888 Op.getOperand(0), Op.getOperand(1));
4889 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4890 DAG.getValueType(VT));
4891 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4892 } else if (VT.getSizeInBits() == 16) {
4893 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4894 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4896 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4897 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4898 DAG.getNode(ISD::BIT_CONVERT, dl,
4902 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4903 Op.getOperand(0), Op.getOperand(1));
4904 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4905 DAG.getValueType(VT));
4906 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4907 } else if (VT == MVT::f32) {
4908 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4909 // the result back to FR32 register. It's only worth matching if the
4910 // result has a single use which is a store or a bitcast to i32. And in
4911 // the case of a store, it's not worth it if the index is a constant 0,
4912 // because a MOVSSmr can be used instead, which is smaller and faster.
4913 if (!Op.hasOneUse())
4915 SDNode *User = *Op.getNode()->use_begin();
4916 if ((User->getOpcode() != ISD::STORE ||
4917 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4918 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4919 (User->getOpcode() != ISD::BIT_CONVERT ||
4920 User->getValueType(0) != MVT::i32))
4922 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4923 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4926 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4927 } else if (VT == MVT::i32) {
4928 // ExtractPS works with constant index.
4929 if (isa<ConstantSDNode>(Op.getOperand(1)))
4937 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4938 SelectionDAG &DAG) const {
4939 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4942 if (Subtarget->hasSSE41()) {
4943 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4948 EVT VT = Op.getValueType();
4949 DebugLoc dl = Op.getDebugLoc();
4950 // TODO: handle v16i8.
4951 if (VT.getSizeInBits() == 16) {
4952 SDValue Vec = Op.getOperand(0);
4953 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4955 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4956 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4957 DAG.getNode(ISD::BIT_CONVERT, dl,
4960 // Transform it so it match pextrw which produces a 32-bit result.
4961 EVT EltVT = MVT::i32;
4962 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4963 Op.getOperand(0), Op.getOperand(1));
4964 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4965 DAG.getValueType(VT));
4966 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4967 } else if (VT.getSizeInBits() == 32) {
4968 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4972 // SHUFPS the element to the lowest double word, then movss.
4973 int Mask[4] = { Idx, -1, -1, -1 };
4974 EVT VVT = Op.getOperand(0).getValueType();
4975 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4976 DAG.getUNDEF(VVT), Mask);
4977 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4978 DAG.getIntPtrConstant(0));
4979 } else if (VT.getSizeInBits() == 64) {
4980 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4981 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4982 // to match extract_elt for f64.
4983 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4987 // UNPCKHPD the element to the lowest double word, then movsd.
4988 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4989 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4990 int Mask[2] = { 1, -1 };
4991 EVT VVT = Op.getOperand(0).getValueType();
4992 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4993 DAG.getUNDEF(VVT), Mask);
4994 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4995 DAG.getIntPtrConstant(0));
5002 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5003 SelectionDAG &DAG) const {
5004 EVT VT = Op.getValueType();
5005 EVT EltVT = VT.getVectorElementType();
5006 DebugLoc dl = Op.getDebugLoc();
5008 SDValue N0 = Op.getOperand(0);
5009 SDValue N1 = Op.getOperand(1);
5010 SDValue N2 = Op.getOperand(2);
5012 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5013 isa<ConstantSDNode>(N2)) {
5015 if (VT == MVT::v8i16)
5016 Opc = X86ISD::PINSRW;
5017 else if (VT == MVT::v4i16)
5018 Opc = X86ISD::MMX_PINSRW;
5019 else if (VT == MVT::v16i8)
5020 Opc = X86ISD::PINSRB;
5022 Opc = X86ISD::PINSRB;
5024 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5026 if (N1.getValueType() != MVT::i32)
5027 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5028 if (N2.getValueType() != MVT::i32)
5029 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5030 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5031 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5032 // Bits [7:6] of the constant are the source select. This will always be
5033 // zero here. The DAG Combiner may combine an extract_elt index into these
5034 // bits. For example (insert (extract, 3), 2) could be matched by putting
5035 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5036 // Bits [5:4] of the constant are the destination select. This is the
5037 // value of the incoming immediate.
5038 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5039 // combine either bitwise AND or insert of float 0.0 to set these bits.
5040 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5041 // Create this as a scalar to vector..
5042 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5043 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5044 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5045 // PINSR* works with constant index.
5052 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5053 EVT VT = Op.getValueType();
5054 EVT EltVT = VT.getVectorElementType();
5056 if (Subtarget->hasSSE41())
5057 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5059 if (EltVT == MVT::i8)
5062 DebugLoc dl = Op.getDebugLoc();
5063 SDValue N0 = Op.getOperand(0);
5064 SDValue N1 = Op.getOperand(1);
5065 SDValue N2 = Op.getOperand(2);
5067 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5068 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5069 // as its second argument.
5070 if (N1.getValueType() != MVT::i32)
5071 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5072 if (N2.getValueType() != MVT::i32)
5073 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5074 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5075 dl, VT, N0, N1, N2);
5081 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5082 DebugLoc dl = Op.getDebugLoc();
5084 if (Op.getValueType() == MVT::v1i64 &&
5085 Op.getOperand(0).getValueType() == MVT::i64)
5086 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5088 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5089 EVT VT = MVT::v2i32;
5090 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5097 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5098 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5101 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5102 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5103 // one of the above mentioned nodes. It has to be wrapped because otherwise
5104 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5105 // be used to form addressing mode. These wrapped nodes will be selected
5108 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5109 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5111 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5113 unsigned char OpFlag = 0;
5114 unsigned WrapperKind = X86ISD::Wrapper;
5115 CodeModel::Model M = getTargetMachine().getCodeModel();
5117 if (Subtarget->isPICStyleRIPRel() &&
5118 (M == CodeModel::Small || M == CodeModel::Kernel))
5119 WrapperKind = X86ISD::WrapperRIP;
5120 else if (Subtarget->isPICStyleGOT())
5121 OpFlag = X86II::MO_GOTOFF;
5122 else if (Subtarget->isPICStyleStubPIC())
5123 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5125 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5127 CP->getOffset(), OpFlag);
5128 DebugLoc DL = CP->getDebugLoc();
5129 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5130 // With PIC, the address is actually $g + Offset.
5132 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5133 DAG.getNode(X86ISD::GlobalBaseReg,
5134 DebugLoc(), getPointerTy()),
5141 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5142 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5144 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5146 unsigned char OpFlag = 0;
5147 unsigned WrapperKind = X86ISD::Wrapper;
5148 CodeModel::Model M = getTargetMachine().getCodeModel();
5150 if (Subtarget->isPICStyleRIPRel() &&
5151 (M == CodeModel::Small || M == CodeModel::Kernel))
5152 WrapperKind = X86ISD::WrapperRIP;
5153 else if (Subtarget->isPICStyleGOT())
5154 OpFlag = X86II::MO_GOTOFF;
5155 else if (Subtarget->isPICStyleStubPIC())
5156 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5158 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5160 DebugLoc DL = JT->getDebugLoc();
5161 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5163 // With PIC, the address is actually $g + Offset.
5165 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5166 DAG.getNode(X86ISD::GlobalBaseReg,
5167 DebugLoc(), getPointerTy()),
5175 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5176 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5178 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5180 unsigned char OpFlag = 0;
5181 unsigned WrapperKind = X86ISD::Wrapper;
5182 CodeModel::Model M = getTargetMachine().getCodeModel();
5184 if (Subtarget->isPICStyleRIPRel() &&
5185 (M == CodeModel::Small || M == CodeModel::Kernel))
5186 WrapperKind = X86ISD::WrapperRIP;
5187 else if (Subtarget->isPICStyleGOT())
5188 OpFlag = X86II::MO_GOTOFF;
5189 else if (Subtarget->isPICStyleStubPIC())
5190 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5192 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5194 DebugLoc DL = Op.getDebugLoc();
5195 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5198 // With PIC, the address is actually $g + Offset.
5199 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5200 !Subtarget->is64Bit()) {
5201 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5202 DAG.getNode(X86ISD::GlobalBaseReg,
5203 DebugLoc(), getPointerTy()),
5211 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5212 // Create the TargetBlockAddressAddress node.
5213 unsigned char OpFlags =
5214 Subtarget->ClassifyBlockAddressReference();
5215 CodeModel::Model M = getTargetMachine().getCodeModel();
5216 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5217 DebugLoc dl = Op.getDebugLoc();
5218 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5219 /*isTarget=*/true, OpFlags);
5221 if (Subtarget->isPICStyleRIPRel() &&
5222 (M == CodeModel::Small || M == CodeModel::Kernel))
5223 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5225 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5227 // With PIC, the address is actually $g + Offset.
5228 if (isGlobalRelativeToPICBase(OpFlags)) {
5229 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5230 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5238 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5240 SelectionDAG &DAG) const {
5241 // Create the TargetGlobalAddress node, folding in the constant
5242 // offset if it is legal.
5243 unsigned char OpFlags =
5244 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5245 CodeModel::Model M = getTargetMachine().getCodeModel();
5247 if (OpFlags == X86II::MO_NO_FLAG &&
5248 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5249 // A direct static reference to a global.
5250 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5253 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5256 if (Subtarget->isPICStyleRIPRel() &&
5257 (M == CodeModel::Small || M == CodeModel::Kernel))
5258 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5260 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5262 // With PIC, the address is actually $g + Offset.
5263 if (isGlobalRelativeToPICBase(OpFlags)) {
5264 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5265 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5269 // For globals that require a load from a stub to get the address, emit the
5271 if (isGlobalStubReference(OpFlags))
5272 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5273 PseudoSourceValue::getGOT(), 0, false, false, 0);
5275 // If there was a non-zero offset that we didn't fold, create an explicit
5278 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5279 DAG.getConstant(Offset, getPointerTy()));
5285 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5286 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5287 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5288 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5292 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5293 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5294 unsigned char OperandFlags) {
5295 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5296 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5297 DebugLoc dl = GA->getDebugLoc();
5298 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5299 GA->getValueType(0),
5303 SDValue Ops[] = { Chain, TGA, *InFlag };
5304 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5306 SDValue Ops[] = { Chain, TGA };
5307 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5310 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5311 MFI->setAdjustsStack(true);
5313 SDValue Flag = Chain.getValue(1);
5314 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5317 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5319 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5322 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5323 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5324 DAG.getNode(X86ISD::GlobalBaseReg,
5325 DebugLoc(), PtrVT), InFlag);
5326 InFlag = Chain.getValue(1);
5328 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5331 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5333 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5335 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5336 X86::RAX, X86II::MO_TLSGD);
5339 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5340 // "local exec" model.
5341 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5342 const EVT PtrVT, TLSModel::Model model,
5344 DebugLoc dl = GA->getDebugLoc();
5345 // Get the Thread Pointer
5346 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5348 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5351 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5352 NULL, 0, false, false, 0);
5354 unsigned char OperandFlags = 0;
5355 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5357 unsigned WrapperKind = X86ISD::Wrapper;
5358 if (model == TLSModel::LocalExec) {
5359 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5360 } else if (is64Bit) {
5361 assert(model == TLSModel::InitialExec);
5362 OperandFlags = X86II::MO_GOTTPOFF;
5363 WrapperKind = X86ISD::WrapperRIP;
5365 assert(model == TLSModel::InitialExec);
5366 OperandFlags = X86II::MO_INDNTPOFF;
5369 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5371 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5372 GA->getValueType(0),
5373 GA->getOffset(), OperandFlags);
5374 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5376 if (model == TLSModel::InitialExec)
5377 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5378 PseudoSourceValue::getGOT(), 0, false, false, 0);
5380 // The address of the thread local variable is the add of the thread
5381 // pointer with the offset of the variable.
5382 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5386 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5388 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5389 const GlobalValue *GV = GA->getGlobal();
5391 if (Subtarget->isTargetELF()) {
5392 // TODO: implement the "local dynamic" model
5393 // TODO: implement the "initial exec"model for pic executables
5395 // If GV is an alias then use the aliasee for determining
5396 // thread-localness.
5397 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5398 GV = GA->resolveAliasedGlobal(false);
5400 TLSModel::Model model
5401 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5404 case TLSModel::GeneralDynamic:
5405 case TLSModel::LocalDynamic: // not implemented
5406 if (Subtarget->is64Bit())
5407 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5408 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5410 case TLSModel::InitialExec:
5411 case TLSModel::LocalExec:
5412 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5413 Subtarget->is64Bit());
5415 } else if (Subtarget->isTargetDarwin()) {
5416 // Darwin only has one model of TLS. Lower to that.
5417 unsigned char OpFlag = 0;
5418 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5419 X86ISD::WrapperRIP : X86ISD::Wrapper;
5421 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5423 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5424 !Subtarget->is64Bit();
5426 OpFlag = X86II::MO_TLVP_PIC_BASE;
5428 OpFlag = X86II::MO_TLVP;
5429 DebugLoc DL = Op.getDebugLoc();
5430 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
5432 GA->getOffset(), OpFlag);
5433 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5435 // With PIC32, the address is actually $g + Offset.
5437 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5438 DAG.getNode(X86ISD::GlobalBaseReg,
5439 DebugLoc(), getPointerTy()),
5442 // Lowering the machine isd will make sure everything is in the right
5444 SDValue Args[] = { Offset };
5445 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5447 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5448 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5449 MFI->setAdjustsStack(true);
5451 // And our return value (tls address) is in the standard call return value
5453 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5454 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5458 "TLS not implemented for this target.");
5460 llvm_unreachable("Unreachable");
5465 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5466 /// take a 2 x i32 value to shift plus a shift amount.
5467 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5468 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5469 EVT VT = Op.getValueType();
5470 unsigned VTBits = VT.getSizeInBits();
5471 DebugLoc dl = Op.getDebugLoc();
5472 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5473 SDValue ShOpLo = Op.getOperand(0);
5474 SDValue ShOpHi = Op.getOperand(1);
5475 SDValue ShAmt = Op.getOperand(2);
5476 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5477 DAG.getConstant(VTBits - 1, MVT::i8))
5478 : DAG.getConstant(0, VT);
5481 if (Op.getOpcode() == ISD::SHL_PARTS) {
5482 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5483 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5485 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5486 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5489 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5490 DAG.getConstant(VTBits, MVT::i8));
5491 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5492 AndNode, DAG.getConstant(0, MVT::i8));
5495 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5496 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5497 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5499 if (Op.getOpcode() == ISD::SHL_PARTS) {
5500 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5501 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5503 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5504 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5507 SDValue Ops[2] = { Lo, Hi };
5508 return DAG.getMergeValues(Ops, 2, dl);
5511 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5512 SelectionDAG &DAG) const {
5513 EVT SrcVT = Op.getOperand(0).getValueType();
5515 if (SrcVT.isVector()) {
5516 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5522 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5523 "Unknown SINT_TO_FP to lower!");
5525 // These are really Legal; return the operand so the caller accepts it as
5527 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5529 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5530 Subtarget->is64Bit()) {
5534 DebugLoc dl = Op.getDebugLoc();
5535 unsigned Size = SrcVT.getSizeInBits()/8;
5536 MachineFunction &MF = DAG.getMachineFunction();
5537 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5538 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5539 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5541 PseudoSourceValue::getFixedStack(SSFI), 0,
5543 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5546 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5548 SelectionDAG &DAG) const {
5550 DebugLoc dl = Op.getDebugLoc();
5552 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5554 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5556 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5557 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5558 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5559 Tys, Ops, array_lengthof(Ops));
5562 Chain = Result.getValue(1);
5563 SDValue InFlag = Result.getValue(2);
5565 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5566 // shouldn't be necessary except that RFP cannot be live across
5567 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5568 MachineFunction &MF = DAG.getMachineFunction();
5569 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5570 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5571 Tys = DAG.getVTList(MVT::Other);
5573 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5575 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5576 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5577 PseudoSourceValue::getFixedStack(SSFI), 0,
5584 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5585 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5586 SelectionDAG &DAG) const {
5587 // This algorithm is not obvious. Here it is in C code, more or less:
5589 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5590 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5591 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5593 // Copy ints to xmm registers.
5594 __m128i xh = _mm_cvtsi32_si128( hi );
5595 __m128i xl = _mm_cvtsi32_si128( lo );
5597 // Combine into low half of a single xmm register.
5598 __m128i x = _mm_unpacklo_epi32( xh, xl );
5602 // Merge in appropriate exponents to give the integer bits the right
5604 x = _mm_unpacklo_epi32( x, exp );
5606 // Subtract away the biases to deal with the IEEE-754 double precision
5608 d = _mm_sub_pd( (__m128d) x, bias );
5610 // All conversions up to here are exact. The correctly rounded result is
5611 // calculated using the current rounding mode using the following
5613 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5614 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5615 // store doesn't really need to be here (except
5616 // maybe to zero the other double)
5621 DebugLoc dl = Op.getDebugLoc();
5622 LLVMContext *Context = DAG.getContext();
5624 // Build some magic constants.
5625 std::vector<Constant*> CV0;
5626 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5627 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5628 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5629 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5630 Constant *C0 = ConstantVector::get(CV0);
5631 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5633 std::vector<Constant*> CV1;
5635 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5637 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5638 Constant *C1 = ConstantVector::get(CV1);
5639 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5641 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5642 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5644 DAG.getIntPtrConstant(1)));
5645 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5646 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5648 DAG.getIntPtrConstant(0)));
5649 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5650 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5651 PseudoSourceValue::getConstantPool(), 0,
5653 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5654 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5655 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5656 PseudoSourceValue::getConstantPool(), 0,
5658 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5660 // Add the halves; easiest way is to swap them into another reg first.
5661 int ShufMask[2] = { 1, -1 };
5662 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5663 DAG.getUNDEF(MVT::v2f64), ShufMask);
5664 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5665 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5666 DAG.getIntPtrConstant(0));
5669 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5670 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5671 SelectionDAG &DAG) const {
5672 DebugLoc dl = Op.getDebugLoc();
5673 // FP constant to bias correct the final result.
5674 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5677 // Load the 32-bit value into an XMM register.
5678 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5679 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5681 DAG.getIntPtrConstant(0)));
5683 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5684 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5685 DAG.getIntPtrConstant(0));
5687 // Or the load with the bias.
5688 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5689 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5690 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5692 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5693 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5694 MVT::v2f64, Bias)));
5695 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5696 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5697 DAG.getIntPtrConstant(0));
5699 // Subtract the bias.
5700 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5702 // Handle final rounding.
5703 EVT DestVT = Op.getValueType();
5705 if (DestVT.bitsLT(MVT::f64)) {
5706 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5707 DAG.getIntPtrConstant(0));
5708 } else if (DestVT.bitsGT(MVT::f64)) {
5709 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5712 // Handle final rounding.
5716 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5717 SelectionDAG &DAG) const {
5718 SDValue N0 = Op.getOperand(0);
5719 DebugLoc dl = Op.getDebugLoc();
5721 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5722 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5723 // the optimization here.
5724 if (DAG.SignBitIsZero(N0))
5725 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5727 EVT SrcVT = N0.getValueType();
5728 EVT DstVT = Op.getValueType();
5729 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5730 return LowerUINT_TO_FP_i64(Op, DAG);
5731 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5732 return LowerUINT_TO_FP_i32(Op, DAG);
5734 // Make a 64-bit buffer, and use it to build an FILD.
5735 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5736 if (SrcVT == MVT::i32) {
5737 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5738 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5739 getPointerTy(), StackSlot, WordOff);
5740 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5741 StackSlot, NULL, 0, false, false, 0);
5742 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5743 OffsetSlot, NULL, 0, false, false, 0);
5744 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5748 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5749 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5750 StackSlot, NULL, 0, false, false, 0);
5751 // For i64 source, we need to add the appropriate power of 2 if the input
5752 // was negative. This is the same as the optimization in
5753 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5754 // we must be careful to do the computation in x87 extended precision, not
5755 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5756 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5757 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5758 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5760 APInt FF(32, 0x5F800000ULL);
5762 // Check whether the sign bit is set.
5763 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5764 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5767 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5768 SDValue FudgePtr = DAG.getConstantPool(
5769 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5772 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5773 SDValue Zero = DAG.getIntPtrConstant(0);
5774 SDValue Four = DAG.getIntPtrConstant(4);
5775 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5777 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5779 // Load the value out, extending it from f32 to f80.
5780 // FIXME: Avoid the extend by constructing the right constant pool?
5781 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
5782 FudgePtr, PseudoSourceValue::getConstantPool(),
5783 0, MVT::f32, false, false, 4);
5784 // Extend everything to 80 bits to force it to be done on x87.
5785 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5786 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5789 std::pair<SDValue,SDValue> X86TargetLowering::
5790 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5791 DebugLoc dl = Op.getDebugLoc();
5793 EVT DstTy = Op.getValueType();
5796 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5800 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5801 DstTy.getSimpleVT() >= MVT::i16 &&
5802 "Unknown FP_TO_SINT to lower!");
5804 // These are really Legal.
5805 if (DstTy == MVT::i32 &&
5806 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5807 return std::make_pair(SDValue(), SDValue());
5808 if (Subtarget->is64Bit() &&
5809 DstTy == MVT::i64 &&
5810 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5811 return std::make_pair(SDValue(), SDValue());
5813 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5815 MachineFunction &MF = DAG.getMachineFunction();
5816 unsigned MemSize = DstTy.getSizeInBits()/8;
5817 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5818 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5821 switch (DstTy.getSimpleVT().SimpleTy) {
5822 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5823 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5824 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5825 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5828 SDValue Chain = DAG.getEntryNode();
5829 SDValue Value = Op.getOperand(0);
5830 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5831 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5832 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5833 PseudoSourceValue::getFixedStack(SSFI), 0,
5835 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5837 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5839 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5840 Chain = Value.getValue(1);
5841 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5842 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5845 // Build the FP_TO_INT*_IN_MEM
5846 SDValue Ops[] = { Chain, Value, StackSlot };
5847 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5849 return std::make_pair(FIST, StackSlot);
5852 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5853 SelectionDAG &DAG) const {
5854 if (Op.getValueType().isVector()) {
5855 if (Op.getValueType() == MVT::v2i32 &&
5856 Op.getOperand(0).getValueType() == MVT::v2f64) {
5862 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5863 SDValue FIST = Vals.first, StackSlot = Vals.second;
5864 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5865 if (FIST.getNode() == 0) return Op;
5868 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5869 FIST, StackSlot, NULL, 0, false, false, 0);
5872 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5873 SelectionDAG &DAG) const {
5874 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5875 SDValue FIST = Vals.first, StackSlot = Vals.second;
5876 assert(FIST.getNode() && "Unexpected failure");
5879 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5880 FIST, StackSlot, NULL, 0, false, false, 0);
5883 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5884 SelectionDAG &DAG) const {
5885 LLVMContext *Context = DAG.getContext();
5886 DebugLoc dl = Op.getDebugLoc();
5887 EVT VT = Op.getValueType();
5890 EltVT = VT.getVectorElementType();
5891 std::vector<Constant*> CV;
5892 if (EltVT == MVT::f64) {
5893 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5897 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5903 Constant *C = ConstantVector::get(CV);
5904 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5905 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5906 PseudoSourceValue::getConstantPool(), 0,
5908 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5911 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5912 LLVMContext *Context = DAG.getContext();
5913 DebugLoc dl = Op.getDebugLoc();
5914 EVT VT = Op.getValueType();
5917 EltVT = VT.getVectorElementType();
5918 std::vector<Constant*> CV;
5919 if (EltVT == MVT::f64) {
5920 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5924 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5930 Constant *C = ConstantVector::get(CV);
5931 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5932 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5933 PseudoSourceValue::getConstantPool(), 0,
5935 if (VT.isVector()) {
5936 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5937 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5938 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5940 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5942 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5946 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5947 LLVMContext *Context = DAG.getContext();
5948 SDValue Op0 = Op.getOperand(0);
5949 SDValue Op1 = Op.getOperand(1);
5950 DebugLoc dl = Op.getDebugLoc();
5951 EVT VT = Op.getValueType();
5952 EVT SrcVT = Op1.getValueType();
5954 // If second operand is smaller, extend it first.
5955 if (SrcVT.bitsLT(VT)) {
5956 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5959 // And if it is bigger, shrink it first.
5960 if (SrcVT.bitsGT(VT)) {
5961 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5965 // At this point the operands and the result should have the same
5966 // type, and that won't be f80 since that is not custom lowered.
5968 // First get the sign bit of second operand.
5969 std::vector<Constant*> CV;
5970 if (SrcVT == MVT::f64) {
5971 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5972 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5974 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5975 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5976 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5977 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5979 Constant *C = ConstantVector::get(CV);
5980 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5981 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5982 PseudoSourceValue::getConstantPool(), 0,
5984 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5986 // Shift sign bit right or left if the two operands have different types.
5987 if (SrcVT.bitsGT(VT)) {
5988 // Op0 is MVT::f32, Op1 is MVT::f64.
5989 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5990 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5991 DAG.getConstant(32, MVT::i32));
5992 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5993 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5994 DAG.getIntPtrConstant(0));
5997 // Clear first operand sign bit.
5999 if (VT == MVT::f64) {
6000 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6001 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6003 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6004 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6005 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6006 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6008 C = ConstantVector::get(CV);
6009 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6010 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6011 PseudoSourceValue::getConstantPool(), 0,
6013 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6015 // Or the value with the sign bit.
6016 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6019 /// Emit nodes that will be selected as "test Op0,Op0", or something
6021 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6022 SelectionDAG &DAG) const {
6023 DebugLoc dl = Op.getDebugLoc();
6025 // CF and OF aren't always set the way we want. Determine which
6026 // of these we need.
6027 bool NeedCF = false;
6028 bool NeedOF = false;
6031 case X86::COND_A: case X86::COND_AE:
6032 case X86::COND_B: case X86::COND_BE:
6035 case X86::COND_G: case X86::COND_GE:
6036 case X86::COND_L: case X86::COND_LE:
6037 case X86::COND_O: case X86::COND_NO:
6042 // See if we can use the EFLAGS value from the operand instead of
6043 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6044 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6045 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6046 // Emit a CMP with 0, which is the TEST pattern.
6047 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6048 DAG.getConstant(0, Op.getValueType()));
6050 unsigned Opcode = 0;
6051 unsigned NumOperands = 0;
6052 switch (Op.getNode()->getOpcode()) {
6054 // Due to an isel shortcoming, be conservative if this add is likely to be
6055 // selected as part of a load-modify-store instruction. When the root node
6056 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6057 // uses of other nodes in the match, such as the ADD in this case. This
6058 // leads to the ADD being left around and reselected, with the result being
6059 // two adds in the output. Alas, even if none our users are stores, that
6060 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6061 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6062 // climbing the DAG back to the root, and it doesn't seem to be worth the
6064 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6065 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6066 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6069 if (ConstantSDNode *C =
6070 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6071 // An add of one will be selected as an INC.
6072 if (C->getAPIntValue() == 1) {
6073 Opcode = X86ISD::INC;
6078 // An add of negative one (subtract of one) will be selected as a DEC.
6079 if (C->getAPIntValue().isAllOnesValue()) {
6080 Opcode = X86ISD::DEC;
6086 // Otherwise use a regular EFLAGS-setting add.
6087 Opcode = X86ISD::ADD;
6091 // If the primary and result isn't used, don't bother using X86ISD::AND,
6092 // because a TEST instruction will be better.
6093 bool NonFlagUse = false;
6094 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6095 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6097 unsigned UOpNo = UI.getOperandNo();
6098 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6099 // Look pass truncate.
6100 UOpNo = User->use_begin().getOperandNo();
6101 User = *User->use_begin();
6104 if (User->getOpcode() != ISD::BRCOND &&
6105 User->getOpcode() != ISD::SETCC &&
6106 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6119 // Due to the ISEL shortcoming noted above, be conservative if this op is
6120 // likely to be selected as part of a load-modify-store instruction.
6121 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6122 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6123 if (UI->getOpcode() == ISD::STORE)
6126 // Otherwise use a regular EFLAGS-setting instruction.
6127 switch (Op.getNode()->getOpcode()) {
6128 default: llvm_unreachable("unexpected operator!");
6129 case ISD::SUB: Opcode = X86ISD::SUB; break;
6130 case ISD::OR: Opcode = X86ISD::OR; break;
6131 case ISD::XOR: Opcode = X86ISD::XOR; break;
6132 case ISD::AND: Opcode = X86ISD::AND; break;
6144 return SDValue(Op.getNode(), 1);
6151 // Emit a CMP with 0, which is the TEST pattern.
6152 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6153 DAG.getConstant(0, Op.getValueType()));
6155 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6156 SmallVector<SDValue, 4> Ops;
6157 for (unsigned i = 0; i != NumOperands; ++i)
6158 Ops.push_back(Op.getOperand(i));
6160 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6161 DAG.ReplaceAllUsesWith(Op, New);
6162 return SDValue(New.getNode(), 1);
6165 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6167 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6168 SelectionDAG &DAG) const {
6169 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6170 if (C->getAPIntValue() == 0)
6171 return EmitTest(Op0, X86CC, DAG);
6173 DebugLoc dl = Op0.getDebugLoc();
6174 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6177 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6178 /// if it's possible.
6179 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6180 DebugLoc dl, SelectionDAG &DAG) const {
6181 SDValue Op0 = And.getOperand(0);
6182 SDValue Op1 = And.getOperand(1);
6183 if (Op0.getOpcode() == ISD::TRUNCATE)
6184 Op0 = Op0.getOperand(0);
6185 if (Op1.getOpcode() == ISD::TRUNCATE)
6186 Op1 = Op1.getOperand(0);
6189 if (Op1.getOpcode() == ISD::SHL)
6190 std::swap(Op0, Op1);
6191 if (Op0.getOpcode() == ISD::SHL) {
6192 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6193 if (And00C->getZExtValue() == 1) {
6194 // If we looked past a truncate, check that it's only truncating away
6196 unsigned BitWidth = Op0.getValueSizeInBits();
6197 unsigned AndBitWidth = And.getValueSizeInBits();
6198 if (BitWidth > AndBitWidth) {
6199 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6200 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6201 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6205 RHS = Op0.getOperand(1);
6207 } else if (Op1.getOpcode() == ISD::Constant) {
6208 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6209 SDValue AndLHS = Op0;
6210 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6211 LHS = AndLHS.getOperand(0);
6212 RHS = AndLHS.getOperand(1);
6216 if (LHS.getNode()) {
6217 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6218 // instruction. Since the shift amount is in-range-or-undefined, we know
6219 // that doing a bittest on the i32 value is ok. We extend to i32 because
6220 // the encoding for the i16 version is larger than the i32 version.
6221 // Also promote i16 to i32 for performance / code size reason.
6222 if (LHS.getValueType() == MVT::i8 ||
6223 LHS.getValueType() == MVT::i16)
6224 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6226 // If the operand types disagree, extend the shift amount to match. Since
6227 // BT ignores high bits (like shifts) we can use anyextend.
6228 if (LHS.getValueType() != RHS.getValueType())
6229 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6231 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6232 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6233 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6234 DAG.getConstant(Cond, MVT::i8), BT);
6240 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6241 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6242 SDValue Op0 = Op.getOperand(0);
6243 SDValue Op1 = Op.getOperand(1);
6244 DebugLoc dl = Op.getDebugLoc();
6245 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6247 // Optimize to BT if possible.
6248 // Lower (X & (1 << N)) == 0 to BT(X, N).
6249 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6250 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6251 if (Op0.getOpcode() == ISD::AND &&
6253 Op1.getOpcode() == ISD::Constant &&
6254 cast<ConstantSDNode>(Op1)->isNullValue() &&
6255 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6256 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6257 if (NewSetCC.getNode())
6261 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6262 if (Op0.getOpcode() == X86ISD::SETCC &&
6263 Op1.getOpcode() == ISD::Constant &&
6264 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6265 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6266 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6267 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6268 bool Invert = (CC == ISD::SETNE) ^
6269 cast<ConstantSDNode>(Op1)->isNullValue();
6271 CCode = X86::GetOppositeBranchCondition(CCode);
6272 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6273 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6276 bool isFP = Op1.getValueType().isFloatingPoint();
6277 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6278 if (X86CC == X86::COND_INVALID)
6281 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6283 // Use sbb x, x to materialize carry bit into a GPR.
6284 if (X86CC == X86::COND_B)
6285 return DAG.getNode(ISD::AND, dl, MVT::i8,
6286 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6287 DAG.getConstant(X86CC, MVT::i8), Cond),
6288 DAG.getConstant(1, MVT::i8));
6290 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6291 DAG.getConstant(X86CC, MVT::i8), Cond);
6294 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6296 SDValue Op0 = Op.getOperand(0);
6297 SDValue Op1 = Op.getOperand(1);
6298 SDValue CC = Op.getOperand(2);
6299 EVT VT = Op.getValueType();
6300 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6301 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6302 DebugLoc dl = Op.getDebugLoc();
6306 EVT VT0 = Op0.getValueType();
6307 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6308 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6311 switch (SetCCOpcode) {
6314 case ISD::SETEQ: SSECC = 0; break;
6316 case ISD::SETGT: Swap = true; // Fallthrough
6318 case ISD::SETOLT: SSECC = 1; break;
6320 case ISD::SETGE: Swap = true; // Fallthrough
6322 case ISD::SETOLE: SSECC = 2; break;
6323 case ISD::SETUO: SSECC = 3; break;
6325 case ISD::SETNE: SSECC = 4; break;
6326 case ISD::SETULE: Swap = true;
6327 case ISD::SETUGE: SSECC = 5; break;
6328 case ISD::SETULT: Swap = true;
6329 case ISD::SETUGT: SSECC = 6; break;
6330 case ISD::SETO: SSECC = 7; break;
6333 std::swap(Op0, Op1);
6335 // In the two special cases we can't handle, emit two comparisons.
6337 if (SetCCOpcode == ISD::SETUEQ) {
6339 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6340 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6341 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6343 else if (SetCCOpcode == ISD::SETONE) {
6345 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6346 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6347 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6349 llvm_unreachable("Illegal FP comparison");
6351 // Handle all other FP comparisons here.
6352 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6355 // We are handling one of the integer comparisons here. Since SSE only has
6356 // GT and EQ comparisons for integer, swapping operands and multiple
6357 // operations may be required for some comparisons.
6358 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6359 bool Swap = false, Invert = false, FlipSigns = false;
6361 switch (VT.getSimpleVT().SimpleTy) {
6364 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6366 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6368 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6369 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6372 switch (SetCCOpcode) {
6374 case ISD::SETNE: Invert = true;
6375 case ISD::SETEQ: Opc = EQOpc; break;
6376 case ISD::SETLT: Swap = true;
6377 case ISD::SETGT: Opc = GTOpc; break;
6378 case ISD::SETGE: Swap = true;
6379 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6380 case ISD::SETULT: Swap = true;
6381 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6382 case ISD::SETUGE: Swap = true;
6383 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6386 std::swap(Op0, Op1);
6388 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6389 // bits of the inputs before performing those operations.
6391 EVT EltVT = VT.getVectorElementType();
6392 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6394 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6395 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6397 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6398 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6401 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6403 // If the logical-not of the result is required, perform that now.
6405 Result = DAG.getNOT(dl, Result, VT);
6410 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6411 static bool isX86LogicalCmp(SDValue Op) {
6412 unsigned Opc = Op.getNode()->getOpcode();
6413 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6415 if (Op.getResNo() == 1 &&
6416 (Opc == X86ISD::ADD ||
6417 Opc == X86ISD::SUB ||
6418 Opc == X86ISD::SMUL ||
6419 Opc == X86ISD::UMUL ||
6420 Opc == X86ISD::INC ||
6421 Opc == X86ISD::DEC ||
6422 Opc == X86ISD::OR ||
6423 Opc == X86ISD::XOR ||
6424 Opc == X86ISD::AND))
6430 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6431 bool addTest = true;
6432 SDValue Cond = Op.getOperand(0);
6433 DebugLoc dl = Op.getDebugLoc();
6436 if (Cond.getOpcode() == ISD::SETCC) {
6437 SDValue NewCond = LowerSETCC(Cond, DAG);
6438 if (NewCond.getNode())
6442 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6443 SDValue Op1 = Op.getOperand(1);
6444 SDValue Op2 = Op.getOperand(2);
6445 if (Cond.getOpcode() == X86ISD::SETCC &&
6446 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6447 SDValue Cmp = Cond.getOperand(1);
6448 if (Cmp.getOpcode() == X86ISD::CMP) {
6449 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6450 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6451 ConstantSDNode *RHSC =
6452 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6453 if (N1C && N1C->isAllOnesValue() &&
6454 N2C && N2C->isNullValue() &&
6455 RHSC && RHSC->isNullValue()) {
6456 SDValue CmpOp0 = Cmp.getOperand(0);
6457 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6458 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6459 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6460 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6465 // Look pass (and (setcc_carry (cmp ...)), 1).
6466 if (Cond.getOpcode() == ISD::AND &&
6467 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6468 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6469 if (C && C->getAPIntValue() == 1)
6470 Cond = Cond.getOperand(0);
6473 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6474 // setting operand in place of the X86ISD::SETCC.
6475 if (Cond.getOpcode() == X86ISD::SETCC ||
6476 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6477 CC = Cond.getOperand(0);
6479 SDValue Cmp = Cond.getOperand(1);
6480 unsigned Opc = Cmp.getOpcode();
6481 EVT VT = Op.getValueType();
6483 bool IllegalFPCMov = false;
6484 if (VT.isFloatingPoint() && !VT.isVector() &&
6485 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6486 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6488 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6489 Opc == X86ISD::BT) { // FIXME
6496 // Look pass the truncate.
6497 if (Cond.getOpcode() == ISD::TRUNCATE)
6498 Cond = Cond.getOperand(0);
6500 // We know the result of AND is compared against zero. Try to match
6502 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6503 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6504 if (NewSetCC.getNode()) {
6505 CC = NewSetCC.getOperand(0);
6506 Cond = NewSetCC.getOperand(1);
6513 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6514 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6517 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6518 // condition is true.
6519 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6520 SDValue Ops[] = { Op2, Op1, CC, Cond };
6521 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6524 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6525 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6526 // from the AND / OR.
6527 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6528 Opc = Op.getOpcode();
6529 if (Opc != ISD::OR && Opc != ISD::AND)
6531 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6532 Op.getOperand(0).hasOneUse() &&
6533 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6534 Op.getOperand(1).hasOneUse());
6537 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6538 // 1 and that the SETCC node has a single use.
6539 static bool isXor1OfSetCC(SDValue Op) {
6540 if (Op.getOpcode() != ISD::XOR)
6542 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6543 if (N1C && N1C->getAPIntValue() == 1) {
6544 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6545 Op.getOperand(0).hasOneUse();
6550 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6551 bool addTest = true;
6552 SDValue Chain = Op.getOperand(0);
6553 SDValue Cond = Op.getOperand(1);
6554 SDValue Dest = Op.getOperand(2);
6555 DebugLoc dl = Op.getDebugLoc();
6558 if (Cond.getOpcode() == ISD::SETCC) {
6559 SDValue NewCond = LowerSETCC(Cond, DAG);
6560 if (NewCond.getNode())
6564 // FIXME: LowerXALUO doesn't handle these!!
6565 else if (Cond.getOpcode() == X86ISD::ADD ||
6566 Cond.getOpcode() == X86ISD::SUB ||
6567 Cond.getOpcode() == X86ISD::SMUL ||
6568 Cond.getOpcode() == X86ISD::UMUL)
6569 Cond = LowerXALUO(Cond, DAG);
6572 // Look pass (and (setcc_carry (cmp ...)), 1).
6573 if (Cond.getOpcode() == ISD::AND &&
6574 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6575 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6576 if (C && C->getAPIntValue() == 1)
6577 Cond = Cond.getOperand(0);
6580 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6581 // setting operand in place of the X86ISD::SETCC.
6582 if (Cond.getOpcode() == X86ISD::SETCC ||
6583 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6584 CC = Cond.getOperand(0);
6586 SDValue Cmp = Cond.getOperand(1);
6587 unsigned Opc = Cmp.getOpcode();
6588 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6589 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6593 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6597 // These can only come from an arithmetic instruction with overflow,
6598 // e.g. SADDO, UADDO.
6599 Cond = Cond.getNode()->getOperand(1);
6606 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6607 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6608 if (CondOpc == ISD::OR) {
6609 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6610 // two branches instead of an explicit OR instruction with a
6612 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6613 isX86LogicalCmp(Cmp)) {
6614 CC = Cond.getOperand(0).getOperand(0);
6615 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6616 Chain, Dest, CC, Cmp);
6617 CC = Cond.getOperand(1).getOperand(0);
6621 } else { // ISD::AND
6622 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6623 // two branches instead of an explicit AND instruction with a
6624 // separate test. However, we only do this if this block doesn't
6625 // have a fall-through edge, because this requires an explicit
6626 // jmp when the condition is false.
6627 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6628 isX86LogicalCmp(Cmp) &&
6629 Op.getNode()->hasOneUse()) {
6630 X86::CondCode CCode =
6631 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6632 CCode = X86::GetOppositeBranchCondition(CCode);
6633 CC = DAG.getConstant(CCode, MVT::i8);
6634 SDNode *User = *Op.getNode()->use_begin();
6635 // Look for an unconditional branch following this conditional branch.
6636 // We need this because we need to reverse the successors in order
6637 // to implement FCMP_OEQ.
6638 if (User->getOpcode() == ISD::BR) {
6639 SDValue FalseBB = User->getOperand(1);
6641 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6642 assert(NewBR == User);
6646 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6647 Chain, Dest, CC, Cmp);
6648 X86::CondCode CCode =
6649 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6650 CCode = X86::GetOppositeBranchCondition(CCode);
6651 CC = DAG.getConstant(CCode, MVT::i8);
6657 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6658 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6659 // It should be transformed during dag combiner except when the condition
6660 // is set by a arithmetics with overflow node.
6661 X86::CondCode CCode =
6662 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6663 CCode = X86::GetOppositeBranchCondition(CCode);
6664 CC = DAG.getConstant(CCode, MVT::i8);
6665 Cond = Cond.getOperand(0).getOperand(1);
6671 // Look pass the truncate.
6672 if (Cond.getOpcode() == ISD::TRUNCATE)
6673 Cond = Cond.getOperand(0);
6675 // We know the result of AND is compared against zero. Try to match
6677 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6678 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6679 if (NewSetCC.getNode()) {
6680 CC = NewSetCC.getOperand(0);
6681 Cond = NewSetCC.getOperand(1);
6688 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6689 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6691 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6692 Chain, Dest, CC, Cond);
6696 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6697 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6698 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6699 // that the guard pages used by the OS virtual memory manager are allocated in
6700 // correct sequence.
6702 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6703 SelectionDAG &DAG) const {
6704 assert(Subtarget->isTargetCygMing() &&
6705 "This should be used only on Cygwin/Mingw targets");
6706 DebugLoc dl = Op.getDebugLoc();
6709 SDValue Chain = Op.getOperand(0);
6710 SDValue Size = Op.getOperand(1);
6711 // FIXME: Ensure alignment here
6715 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6717 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6718 Flag = Chain.getValue(1);
6720 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6722 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6723 Flag = Chain.getValue(1);
6725 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6727 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6728 return DAG.getMergeValues(Ops1, 2, dl);
6731 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6732 MachineFunction &MF = DAG.getMachineFunction();
6733 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6735 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6736 DebugLoc dl = Op.getDebugLoc();
6738 if (!Subtarget->is64Bit()) {
6739 // vastart just stores the address of the VarArgsFrameIndex slot into the
6740 // memory location argument.
6741 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6743 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6748 // gp_offset (0 - 6 * 8)
6749 // fp_offset (48 - 48 + 8 * 16)
6750 // overflow_arg_area (point to parameters coming in memory).
6752 SmallVector<SDValue, 8> MemOps;
6753 SDValue FIN = Op.getOperand(1);
6755 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6756 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6758 FIN, SV, 0, false, false, 0);
6759 MemOps.push_back(Store);
6762 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6763 FIN, DAG.getIntPtrConstant(4));
6764 Store = DAG.getStore(Op.getOperand(0), dl,
6765 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6767 FIN, SV, 4, false, false, 0);
6768 MemOps.push_back(Store);
6770 // Store ptr to overflow_arg_area
6771 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6772 FIN, DAG.getIntPtrConstant(4));
6773 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6775 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 8,
6777 MemOps.push_back(Store);
6779 // Store ptr to reg_save_area.
6780 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6781 FIN, DAG.getIntPtrConstant(8));
6782 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6784 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 16,
6786 MemOps.push_back(Store);
6787 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6788 &MemOps[0], MemOps.size());
6791 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6792 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6793 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6795 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6799 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6800 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6801 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6802 SDValue Chain = Op.getOperand(0);
6803 SDValue DstPtr = Op.getOperand(1);
6804 SDValue SrcPtr = Op.getOperand(2);
6805 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6806 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6807 DebugLoc dl = Op.getDebugLoc();
6809 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6810 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6811 false, DstSV, 0, SrcSV, 0);
6815 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6816 DebugLoc dl = Op.getDebugLoc();
6817 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6819 default: return SDValue(); // Don't custom lower most intrinsics.
6820 // Comparison intrinsics.
6821 case Intrinsic::x86_sse_comieq_ss:
6822 case Intrinsic::x86_sse_comilt_ss:
6823 case Intrinsic::x86_sse_comile_ss:
6824 case Intrinsic::x86_sse_comigt_ss:
6825 case Intrinsic::x86_sse_comige_ss:
6826 case Intrinsic::x86_sse_comineq_ss:
6827 case Intrinsic::x86_sse_ucomieq_ss:
6828 case Intrinsic::x86_sse_ucomilt_ss:
6829 case Intrinsic::x86_sse_ucomile_ss:
6830 case Intrinsic::x86_sse_ucomigt_ss:
6831 case Intrinsic::x86_sse_ucomige_ss:
6832 case Intrinsic::x86_sse_ucomineq_ss:
6833 case Intrinsic::x86_sse2_comieq_sd:
6834 case Intrinsic::x86_sse2_comilt_sd:
6835 case Intrinsic::x86_sse2_comile_sd:
6836 case Intrinsic::x86_sse2_comigt_sd:
6837 case Intrinsic::x86_sse2_comige_sd:
6838 case Intrinsic::x86_sse2_comineq_sd:
6839 case Intrinsic::x86_sse2_ucomieq_sd:
6840 case Intrinsic::x86_sse2_ucomilt_sd:
6841 case Intrinsic::x86_sse2_ucomile_sd:
6842 case Intrinsic::x86_sse2_ucomigt_sd:
6843 case Intrinsic::x86_sse2_ucomige_sd:
6844 case Intrinsic::x86_sse2_ucomineq_sd: {
6846 ISD::CondCode CC = ISD::SETCC_INVALID;
6849 case Intrinsic::x86_sse_comieq_ss:
6850 case Intrinsic::x86_sse2_comieq_sd:
6854 case Intrinsic::x86_sse_comilt_ss:
6855 case Intrinsic::x86_sse2_comilt_sd:
6859 case Intrinsic::x86_sse_comile_ss:
6860 case Intrinsic::x86_sse2_comile_sd:
6864 case Intrinsic::x86_sse_comigt_ss:
6865 case Intrinsic::x86_sse2_comigt_sd:
6869 case Intrinsic::x86_sse_comige_ss:
6870 case Intrinsic::x86_sse2_comige_sd:
6874 case Intrinsic::x86_sse_comineq_ss:
6875 case Intrinsic::x86_sse2_comineq_sd:
6879 case Intrinsic::x86_sse_ucomieq_ss:
6880 case Intrinsic::x86_sse2_ucomieq_sd:
6881 Opc = X86ISD::UCOMI;
6884 case Intrinsic::x86_sse_ucomilt_ss:
6885 case Intrinsic::x86_sse2_ucomilt_sd:
6886 Opc = X86ISD::UCOMI;
6889 case Intrinsic::x86_sse_ucomile_ss:
6890 case Intrinsic::x86_sse2_ucomile_sd:
6891 Opc = X86ISD::UCOMI;
6894 case Intrinsic::x86_sse_ucomigt_ss:
6895 case Intrinsic::x86_sse2_ucomigt_sd:
6896 Opc = X86ISD::UCOMI;
6899 case Intrinsic::x86_sse_ucomige_ss:
6900 case Intrinsic::x86_sse2_ucomige_sd:
6901 Opc = X86ISD::UCOMI;
6904 case Intrinsic::x86_sse_ucomineq_ss:
6905 case Intrinsic::x86_sse2_ucomineq_sd:
6906 Opc = X86ISD::UCOMI;
6911 SDValue LHS = Op.getOperand(1);
6912 SDValue RHS = Op.getOperand(2);
6913 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6914 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6915 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6916 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6917 DAG.getConstant(X86CC, MVT::i8), Cond);
6918 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6920 // ptest intrinsics. The intrinsic these come from are designed to return
6921 // an integer value, not just an instruction so lower it to the ptest
6922 // pattern and a setcc for the result.
6923 case Intrinsic::x86_sse41_ptestz:
6924 case Intrinsic::x86_sse41_ptestc:
6925 case Intrinsic::x86_sse41_ptestnzc:{
6928 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6929 case Intrinsic::x86_sse41_ptestz:
6931 X86CC = X86::COND_E;
6933 case Intrinsic::x86_sse41_ptestc:
6935 X86CC = X86::COND_B;
6937 case Intrinsic::x86_sse41_ptestnzc:
6939 X86CC = X86::COND_A;
6943 SDValue LHS = Op.getOperand(1);
6944 SDValue RHS = Op.getOperand(2);
6945 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6946 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6947 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6948 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6951 // Fix vector shift instructions where the last operand is a non-immediate
6953 case Intrinsic::x86_sse2_pslli_w:
6954 case Intrinsic::x86_sse2_pslli_d:
6955 case Intrinsic::x86_sse2_pslli_q:
6956 case Intrinsic::x86_sse2_psrli_w:
6957 case Intrinsic::x86_sse2_psrli_d:
6958 case Intrinsic::x86_sse2_psrli_q:
6959 case Intrinsic::x86_sse2_psrai_w:
6960 case Intrinsic::x86_sse2_psrai_d:
6961 case Intrinsic::x86_mmx_pslli_w:
6962 case Intrinsic::x86_mmx_pslli_d:
6963 case Intrinsic::x86_mmx_pslli_q:
6964 case Intrinsic::x86_mmx_psrli_w:
6965 case Intrinsic::x86_mmx_psrli_d:
6966 case Intrinsic::x86_mmx_psrli_q:
6967 case Intrinsic::x86_mmx_psrai_w:
6968 case Intrinsic::x86_mmx_psrai_d: {
6969 SDValue ShAmt = Op.getOperand(2);
6970 if (isa<ConstantSDNode>(ShAmt))
6973 unsigned NewIntNo = 0;
6974 EVT ShAmtVT = MVT::v4i32;
6976 case Intrinsic::x86_sse2_pslli_w:
6977 NewIntNo = Intrinsic::x86_sse2_psll_w;
6979 case Intrinsic::x86_sse2_pslli_d:
6980 NewIntNo = Intrinsic::x86_sse2_psll_d;
6982 case Intrinsic::x86_sse2_pslli_q:
6983 NewIntNo = Intrinsic::x86_sse2_psll_q;
6985 case Intrinsic::x86_sse2_psrli_w:
6986 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6988 case Intrinsic::x86_sse2_psrli_d:
6989 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6991 case Intrinsic::x86_sse2_psrli_q:
6992 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6994 case Intrinsic::x86_sse2_psrai_w:
6995 NewIntNo = Intrinsic::x86_sse2_psra_w;
6997 case Intrinsic::x86_sse2_psrai_d:
6998 NewIntNo = Intrinsic::x86_sse2_psra_d;
7001 ShAmtVT = MVT::v2i32;
7003 case Intrinsic::x86_mmx_pslli_w:
7004 NewIntNo = Intrinsic::x86_mmx_psll_w;
7006 case Intrinsic::x86_mmx_pslli_d:
7007 NewIntNo = Intrinsic::x86_mmx_psll_d;
7009 case Intrinsic::x86_mmx_pslli_q:
7010 NewIntNo = Intrinsic::x86_mmx_psll_q;
7012 case Intrinsic::x86_mmx_psrli_w:
7013 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7015 case Intrinsic::x86_mmx_psrli_d:
7016 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7018 case Intrinsic::x86_mmx_psrli_q:
7019 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7021 case Intrinsic::x86_mmx_psrai_w:
7022 NewIntNo = Intrinsic::x86_mmx_psra_w;
7024 case Intrinsic::x86_mmx_psrai_d:
7025 NewIntNo = Intrinsic::x86_mmx_psra_d;
7027 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7033 // The vector shift intrinsics with scalars uses 32b shift amounts but
7034 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7038 ShOps[1] = DAG.getConstant(0, MVT::i32);
7039 if (ShAmtVT == MVT::v4i32) {
7040 ShOps[2] = DAG.getUNDEF(MVT::i32);
7041 ShOps[3] = DAG.getUNDEF(MVT::i32);
7042 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7044 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7047 EVT VT = Op.getValueType();
7048 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7049 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7050 DAG.getConstant(NewIntNo, MVT::i32),
7051 Op.getOperand(1), ShAmt);
7056 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7057 SelectionDAG &DAG) const {
7058 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7059 MFI->setReturnAddressIsTaken(true);
7061 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7062 DebugLoc dl = Op.getDebugLoc();
7065 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7067 DAG.getConstant(TD->getPointerSize(),
7068 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7069 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7070 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7072 NULL, 0, false, false, 0);
7075 // Just load the return address.
7076 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7077 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7078 RetAddrFI, NULL, 0, false, false, 0);
7081 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7082 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7083 MFI->setFrameAddressIsTaken(true);
7085 EVT VT = Op.getValueType();
7086 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7087 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7088 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7089 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7091 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7096 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7097 SelectionDAG &DAG) const {
7098 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7101 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7102 MachineFunction &MF = DAG.getMachineFunction();
7103 SDValue Chain = Op.getOperand(0);
7104 SDValue Offset = Op.getOperand(1);
7105 SDValue Handler = Op.getOperand(2);
7106 DebugLoc dl = Op.getDebugLoc();
7108 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7110 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7112 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7113 DAG.getIntPtrConstant(-TD->getPointerSize()));
7114 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7115 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7116 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7117 MF.getRegInfo().addLiveOut(StoreAddrReg);
7119 return DAG.getNode(X86ISD::EH_RETURN, dl,
7121 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7124 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7125 SelectionDAG &DAG) const {
7126 SDValue Root = Op.getOperand(0);
7127 SDValue Trmp = Op.getOperand(1); // trampoline
7128 SDValue FPtr = Op.getOperand(2); // nested function
7129 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7130 DebugLoc dl = Op.getDebugLoc();
7132 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7134 if (Subtarget->is64Bit()) {
7135 SDValue OutChains[6];
7137 // Large code-model.
7138 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7139 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7141 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7142 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7144 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7146 // Load the pointer to the nested function into R11.
7147 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7148 SDValue Addr = Trmp;
7149 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7150 Addr, TrmpAddr, 0, false, false, 0);
7152 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7153 DAG.getConstant(2, MVT::i64));
7154 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7157 // Load the 'nest' parameter value into R10.
7158 // R10 is specified in X86CallingConv.td
7159 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7160 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7161 DAG.getConstant(10, MVT::i64));
7162 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7163 Addr, TrmpAddr, 10, false, false, 0);
7165 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7166 DAG.getConstant(12, MVT::i64));
7167 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7170 // Jump to the nested function.
7171 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7172 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7173 DAG.getConstant(20, MVT::i64));
7174 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7175 Addr, TrmpAddr, 20, false, false, 0);
7177 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7178 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7179 DAG.getConstant(22, MVT::i64));
7180 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7181 TrmpAddr, 22, false, false, 0);
7184 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7185 return DAG.getMergeValues(Ops, 2, dl);
7187 const Function *Func =
7188 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7189 CallingConv::ID CC = Func->getCallingConv();
7194 llvm_unreachable("Unsupported calling convention");
7195 case CallingConv::C:
7196 case CallingConv::X86_StdCall: {
7197 // Pass 'nest' parameter in ECX.
7198 // Must be kept in sync with X86CallingConv.td
7201 // Check that ECX wasn't needed by an 'inreg' parameter.
7202 const FunctionType *FTy = Func->getFunctionType();
7203 const AttrListPtr &Attrs = Func->getAttributes();
7205 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7206 unsigned InRegCount = 0;
7209 for (FunctionType::param_iterator I = FTy->param_begin(),
7210 E = FTy->param_end(); I != E; ++I, ++Idx)
7211 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7212 // FIXME: should only count parameters that are lowered to integers.
7213 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7215 if (InRegCount > 2) {
7216 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7221 case CallingConv::X86_FastCall:
7222 case CallingConv::X86_ThisCall:
7223 case CallingConv::Fast:
7224 // Pass 'nest' parameter in EAX.
7225 // Must be kept in sync with X86CallingConv.td
7230 SDValue OutChains[4];
7233 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7234 DAG.getConstant(10, MVT::i32));
7235 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7237 // This is storing the opcode for MOV32ri.
7238 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7239 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7240 OutChains[0] = DAG.getStore(Root, dl,
7241 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7242 Trmp, TrmpAddr, 0, false, false, 0);
7244 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7245 DAG.getConstant(1, MVT::i32));
7246 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7249 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7250 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7251 DAG.getConstant(5, MVT::i32));
7252 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7253 TrmpAddr, 5, false, false, 1);
7255 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7256 DAG.getConstant(6, MVT::i32));
7257 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7261 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7262 return DAG.getMergeValues(Ops, 2, dl);
7266 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7267 SelectionDAG &DAG) const {
7269 The rounding mode is in bits 11:10 of FPSR, and has the following
7276 FLT_ROUNDS, on the other hand, expects the following:
7283 To perform the conversion, we do:
7284 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7287 MachineFunction &MF = DAG.getMachineFunction();
7288 const TargetMachine &TM = MF.getTarget();
7289 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7290 unsigned StackAlignment = TFI.getStackAlignment();
7291 EVT VT = Op.getValueType();
7292 DebugLoc dl = Op.getDebugLoc();
7294 // Save FP Control Word to stack slot
7295 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7296 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7298 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7299 DAG.getEntryNode(), StackSlot);
7301 // Load FP Control Word from stack slot
7302 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7305 // Transform as necessary
7307 DAG.getNode(ISD::SRL, dl, MVT::i16,
7308 DAG.getNode(ISD::AND, dl, MVT::i16,
7309 CWD, DAG.getConstant(0x800, MVT::i16)),
7310 DAG.getConstant(11, MVT::i8));
7312 DAG.getNode(ISD::SRL, dl, MVT::i16,
7313 DAG.getNode(ISD::AND, dl, MVT::i16,
7314 CWD, DAG.getConstant(0x400, MVT::i16)),
7315 DAG.getConstant(9, MVT::i8));
7318 DAG.getNode(ISD::AND, dl, MVT::i16,
7319 DAG.getNode(ISD::ADD, dl, MVT::i16,
7320 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7321 DAG.getConstant(1, MVT::i16)),
7322 DAG.getConstant(3, MVT::i16));
7325 return DAG.getNode((VT.getSizeInBits() < 16 ?
7326 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7329 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7330 EVT VT = Op.getValueType();
7332 unsigned NumBits = VT.getSizeInBits();
7333 DebugLoc dl = Op.getDebugLoc();
7335 Op = Op.getOperand(0);
7336 if (VT == MVT::i8) {
7337 // Zero extend to i32 since there is not an i8 bsr.
7339 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7342 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7343 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7344 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7346 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7349 DAG.getConstant(NumBits+NumBits-1, OpVT),
7350 DAG.getConstant(X86::COND_E, MVT::i8),
7353 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7355 // Finally xor with NumBits-1.
7356 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7359 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7363 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7364 EVT VT = Op.getValueType();
7366 unsigned NumBits = VT.getSizeInBits();
7367 DebugLoc dl = Op.getDebugLoc();
7369 Op = Op.getOperand(0);
7370 if (VT == MVT::i8) {
7372 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7375 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7376 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7377 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7379 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7382 DAG.getConstant(NumBits, OpVT),
7383 DAG.getConstant(X86::COND_E, MVT::i8),
7386 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7389 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7393 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7394 EVT VT = Op.getValueType();
7395 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7396 DebugLoc dl = Op.getDebugLoc();
7398 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7399 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7400 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7401 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7402 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7404 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7405 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7406 // return AloBlo + AloBhi + AhiBlo;
7408 SDValue A = Op.getOperand(0);
7409 SDValue B = Op.getOperand(1);
7411 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7412 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7413 A, DAG.getConstant(32, MVT::i32));
7414 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7415 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7416 B, DAG.getConstant(32, MVT::i32));
7417 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7418 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7420 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7421 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7423 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7424 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7426 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7427 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7428 AloBhi, DAG.getConstant(32, MVT::i32));
7429 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7430 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7431 AhiBlo, DAG.getConstant(32, MVT::i32));
7432 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7433 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7438 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7439 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7440 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7441 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7442 // has only one use.
7443 SDNode *N = Op.getNode();
7444 SDValue LHS = N->getOperand(0);
7445 SDValue RHS = N->getOperand(1);
7446 unsigned BaseOp = 0;
7448 DebugLoc dl = Op.getDebugLoc();
7450 switch (Op.getOpcode()) {
7451 default: llvm_unreachable("Unknown ovf instruction!");
7453 // A subtract of one will be selected as a INC. Note that INC doesn't
7454 // set CF, so we can't do this for UADDO.
7455 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7456 if (C->getAPIntValue() == 1) {
7457 BaseOp = X86ISD::INC;
7461 BaseOp = X86ISD::ADD;
7465 BaseOp = X86ISD::ADD;
7469 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7470 // set CF, so we can't do this for USUBO.
7471 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7472 if (C->getAPIntValue() == 1) {
7473 BaseOp = X86ISD::DEC;
7477 BaseOp = X86ISD::SUB;
7481 BaseOp = X86ISD::SUB;
7485 BaseOp = X86ISD::SMUL;
7489 BaseOp = X86ISD::UMUL;
7494 // Also sets EFLAGS.
7495 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7496 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7499 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7500 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7502 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7506 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7507 EVT T = Op.getValueType();
7508 DebugLoc dl = Op.getDebugLoc();
7511 switch(T.getSimpleVT().SimpleTy) {
7513 assert(false && "Invalid value type!");
7514 case MVT::i8: Reg = X86::AL; size = 1; break;
7515 case MVT::i16: Reg = X86::AX; size = 2; break;
7516 case MVT::i32: Reg = X86::EAX; size = 4; break;
7518 assert(Subtarget->is64Bit() && "Node not type legal!");
7519 Reg = X86::RAX; size = 8;
7522 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7523 Op.getOperand(2), SDValue());
7524 SDValue Ops[] = { cpIn.getValue(0),
7527 DAG.getTargetConstant(size, MVT::i8),
7529 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7530 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7532 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7536 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7537 SelectionDAG &DAG) const {
7538 assert(Subtarget->is64Bit() && "Result not type legalized?");
7539 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7540 SDValue TheChain = Op.getOperand(0);
7541 DebugLoc dl = Op.getDebugLoc();
7542 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7543 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7544 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7546 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7547 DAG.getConstant(32, MVT::i8));
7549 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7552 return DAG.getMergeValues(Ops, 2, dl);
7555 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7556 SelectionDAG &DAG) const {
7557 EVT SrcVT = Op.getOperand(0).getValueType();
7558 EVT DstVT = Op.getValueType();
7559 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7560 Subtarget->hasMMX() && !DisableMMX) &&
7561 "Unexpected custom BIT_CONVERT");
7562 assert((DstVT == MVT::i64 ||
7563 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7564 "Unexpected custom BIT_CONVERT");
7565 // i64 <=> MMX conversions are Legal.
7566 if (SrcVT==MVT::i64 && DstVT.isVector())
7568 if (DstVT==MVT::i64 && SrcVT.isVector())
7570 // MMX <=> MMX conversions are Legal.
7571 if (SrcVT.isVector() && DstVT.isVector())
7573 // All other conversions need to be expanded.
7576 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7577 SDNode *Node = Op.getNode();
7578 DebugLoc dl = Node->getDebugLoc();
7579 EVT T = Node->getValueType(0);
7580 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7581 DAG.getConstant(0, T), Node->getOperand(2));
7582 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7583 cast<AtomicSDNode>(Node)->getMemoryVT(),
7584 Node->getOperand(0),
7585 Node->getOperand(1), negOp,
7586 cast<AtomicSDNode>(Node)->getSrcValue(),
7587 cast<AtomicSDNode>(Node)->getAlignment());
7590 /// LowerOperation - Provide custom lowering hooks for some operations.
7592 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7593 switch (Op.getOpcode()) {
7594 default: llvm_unreachable("Should not custom lower this!");
7595 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7596 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7597 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7598 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7599 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7600 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7601 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7602 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7603 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7604 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7605 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7606 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7607 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7608 case ISD::SHL_PARTS:
7609 case ISD::SRA_PARTS:
7610 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7611 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7612 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7613 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7614 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7615 case ISD::FABS: return LowerFABS(Op, DAG);
7616 case ISD::FNEG: return LowerFNEG(Op, DAG);
7617 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7618 case ISD::SETCC: return LowerSETCC(Op, DAG);
7619 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7620 case ISD::SELECT: return LowerSELECT(Op, DAG);
7621 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7622 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7623 case ISD::VASTART: return LowerVASTART(Op, DAG);
7624 case ISD::VAARG: return LowerVAARG(Op, DAG);
7625 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7626 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7627 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7628 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7629 case ISD::FRAME_TO_ARGS_OFFSET:
7630 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7631 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7632 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7633 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7634 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7635 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7636 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7637 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7643 case ISD::UMULO: return LowerXALUO(Op, DAG);
7644 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7645 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7649 void X86TargetLowering::
7650 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7651 SelectionDAG &DAG, unsigned NewOp) const {
7652 EVT T = Node->getValueType(0);
7653 DebugLoc dl = Node->getDebugLoc();
7654 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7656 SDValue Chain = Node->getOperand(0);
7657 SDValue In1 = Node->getOperand(1);
7658 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7659 Node->getOperand(2), DAG.getIntPtrConstant(0));
7660 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7661 Node->getOperand(2), DAG.getIntPtrConstant(1));
7662 SDValue Ops[] = { Chain, In1, In2L, In2H };
7663 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7665 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7666 cast<MemSDNode>(Node)->getMemOperand());
7667 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7668 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7669 Results.push_back(Result.getValue(2));
7672 /// ReplaceNodeResults - Replace a node with an illegal result type
7673 /// with a new node built out of custom code.
7674 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7675 SmallVectorImpl<SDValue>&Results,
7676 SelectionDAG &DAG) const {
7677 DebugLoc dl = N->getDebugLoc();
7678 switch (N->getOpcode()) {
7680 assert(false && "Do not know how to custom type legalize this operation!");
7682 case ISD::FP_TO_SINT: {
7683 std::pair<SDValue,SDValue> Vals =
7684 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7685 SDValue FIST = Vals.first, StackSlot = Vals.second;
7686 if (FIST.getNode() != 0) {
7687 EVT VT = N->getValueType(0);
7688 // Return a load from the stack slot.
7689 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7694 case ISD::READCYCLECOUNTER: {
7695 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7696 SDValue TheChain = N->getOperand(0);
7697 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7698 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7700 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7702 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7703 SDValue Ops[] = { eax, edx };
7704 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7705 Results.push_back(edx.getValue(1));
7708 case ISD::ATOMIC_CMP_SWAP: {
7709 EVT T = N->getValueType(0);
7710 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7711 SDValue cpInL, cpInH;
7712 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7713 DAG.getConstant(0, MVT::i32));
7714 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7715 DAG.getConstant(1, MVT::i32));
7716 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7717 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7719 SDValue swapInL, swapInH;
7720 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7721 DAG.getConstant(0, MVT::i32));
7722 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7723 DAG.getConstant(1, MVT::i32));
7724 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7726 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7727 swapInL.getValue(1));
7728 SDValue Ops[] = { swapInH.getValue(0),
7730 swapInH.getValue(1) };
7731 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7732 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7733 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7734 MVT::i32, Result.getValue(1));
7735 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7736 MVT::i32, cpOutL.getValue(2));
7737 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7738 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7739 Results.push_back(cpOutH.getValue(1));
7742 case ISD::ATOMIC_LOAD_ADD:
7743 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7745 case ISD::ATOMIC_LOAD_AND:
7746 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7748 case ISD::ATOMIC_LOAD_NAND:
7749 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7751 case ISD::ATOMIC_LOAD_OR:
7752 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7754 case ISD::ATOMIC_LOAD_SUB:
7755 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7757 case ISD::ATOMIC_LOAD_XOR:
7758 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7760 case ISD::ATOMIC_SWAP:
7761 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7766 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7768 default: return NULL;
7769 case X86ISD::BSF: return "X86ISD::BSF";
7770 case X86ISD::BSR: return "X86ISD::BSR";
7771 case X86ISD::SHLD: return "X86ISD::SHLD";
7772 case X86ISD::SHRD: return "X86ISD::SHRD";
7773 case X86ISD::FAND: return "X86ISD::FAND";
7774 case X86ISD::FOR: return "X86ISD::FOR";
7775 case X86ISD::FXOR: return "X86ISD::FXOR";
7776 case X86ISD::FSRL: return "X86ISD::FSRL";
7777 case X86ISD::FILD: return "X86ISD::FILD";
7778 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7779 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7780 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7781 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7782 case X86ISD::FLD: return "X86ISD::FLD";
7783 case X86ISD::FST: return "X86ISD::FST";
7784 case X86ISD::CALL: return "X86ISD::CALL";
7785 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7786 case X86ISD::BT: return "X86ISD::BT";
7787 case X86ISD::CMP: return "X86ISD::CMP";
7788 case X86ISD::COMI: return "X86ISD::COMI";
7789 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7790 case X86ISD::SETCC: return "X86ISD::SETCC";
7791 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7792 case X86ISD::CMOV: return "X86ISD::CMOV";
7793 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7794 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7795 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7796 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7797 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7798 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7799 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7800 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7801 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7802 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7803 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7804 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7805 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7806 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7807 case X86ISD::FMAX: return "X86ISD::FMAX";
7808 case X86ISD::FMIN: return "X86ISD::FMIN";
7809 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7810 case X86ISD::FRCP: return "X86ISD::FRCP";
7811 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7812 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
7813 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7814 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7815 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7816 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7817 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7818 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7819 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7820 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7821 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7822 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7823 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7824 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7825 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7826 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7827 case X86ISD::VSHL: return "X86ISD::VSHL";
7828 case X86ISD::VSRL: return "X86ISD::VSRL";
7829 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7830 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7831 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7832 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7833 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7834 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7835 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7836 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7837 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7838 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7839 case X86ISD::ADD: return "X86ISD::ADD";
7840 case X86ISD::SUB: return "X86ISD::SUB";
7841 case X86ISD::SMUL: return "X86ISD::SMUL";
7842 case X86ISD::UMUL: return "X86ISD::UMUL";
7843 case X86ISD::INC: return "X86ISD::INC";
7844 case X86ISD::DEC: return "X86ISD::DEC";
7845 case X86ISD::OR: return "X86ISD::OR";
7846 case X86ISD::XOR: return "X86ISD::XOR";
7847 case X86ISD::AND: return "X86ISD::AND";
7848 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7849 case X86ISD::PTEST: return "X86ISD::PTEST";
7850 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7851 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7855 // isLegalAddressingMode - Return true if the addressing mode represented
7856 // by AM is legal for this target, for a load/store of the specified type.
7857 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7858 const Type *Ty) const {
7859 // X86 supports extremely general addressing modes.
7860 CodeModel::Model M = getTargetMachine().getCodeModel();
7862 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7863 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7868 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7870 // If a reference to this global requires an extra load, we can't fold it.
7871 if (isGlobalStubReference(GVFlags))
7874 // If BaseGV requires a register for the PIC base, we cannot also have a
7875 // BaseReg specified.
7876 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7879 // If lower 4G is not available, then we must use rip-relative addressing.
7880 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7890 // These scales always work.
7895 // These scales are formed with basereg+scalereg. Only accept if there is
7900 default: // Other stuff never works.
7908 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7909 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7911 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7912 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7913 if (NumBits1 <= NumBits2)
7918 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7919 if (!VT1.isInteger() || !VT2.isInteger())
7921 unsigned NumBits1 = VT1.getSizeInBits();
7922 unsigned NumBits2 = VT2.getSizeInBits();
7923 if (NumBits1 <= NumBits2)
7928 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7929 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7930 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7933 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7934 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7935 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7938 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7939 // i16 instructions are longer (0x66 prefix) and potentially slower.
7940 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7943 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7944 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7945 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7946 /// are assumed to be legal.
7948 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7950 // Very little shuffling can be done for 64-bit vectors right now.
7951 if (VT.getSizeInBits() == 64)
7952 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
7954 // FIXME: pshufb, blends, shifts.
7955 return (VT.getVectorNumElements() == 2 ||
7956 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7957 isMOVLMask(M, VT) ||
7958 isSHUFPMask(M, VT) ||
7959 isPSHUFDMask(M, VT) ||
7960 isPSHUFHWMask(M, VT) ||
7961 isPSHUFLWMask(M, VT) ||
7962 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7963 isUNPCKLMask(M, VT) ||
7964 isUNPCKHMask(M, VT) ||
7965 isUNPCKL_v_undef_Mask(M, VT) ||
7966 isUNPCKH_v_undef_Mask(M, VT));
7970 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7972 unsigned NumElts = VT.getVectorNumElements();
7973 // FIXME: This collection of masks seems suspect.
7976 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7977 return (isMOVLMask(Mask, VT) ||
7978 isCommutedMOVLMask(Mask, VT, true) ||
7979 isSHUFPMask(Mask, VT) ||
7980 isCommutedSHUFPMask(Mask, VT));
7985 //===----------------------------------------------------------------------===//
7986 // X86 Scheduler Hooks
7987 //===----------------------------------------------------------------------===//
7989 // private utility function
7991 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7992 MachineBasicBlock *MBB,
8000 TargetRegisterClass *RC,
8001 bool invSrc) const {
8002 // For the atomic bitwise operator, we generate
8005 // ld t1 = [bitinstr.addr]
8006 // op t2 = t1, [bitinstr.val]
8008 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8010 // fallthrough -->nextMBB
8011 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8012 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8013 MachineFunction::iterator MBBIter = MBB;
8016 /// First build the CFG
8017 MachineFunction *F = MBB->getParent();
8018 MachineBasicBlock *thisMBB = MBB;
8019 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8020 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8021 F->insert(MBBIter, newMBB);
8022 F->insert(MBBIter, nextMBB);
8024 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8025 nextMBB->splice(nextMBB->begin(), thisMBB,
8026 llvm::next(MachineBasicBlock::iterator(bInstr)),
8028 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8030 // Update thisMBB to fall through to newMBB
8031 thisMBB->addSuccessor(newMBB);
8033 // newMBB jumps to itself and fall through to nextMBB
8034 newMBB->addSuccessor(nextMBB);
8035 newMBB->addSuccessor(newMBB);
8037 // Insert instructions into newMBB based on incoming instruction
8038 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8039 "unexpected number of operands");
8040 DebugLoc dl = bInstr->getDebugLoc();
8041 MachineOperand& destOper = bInstr->getOperand(0);
8042 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8043 int numArgs = bInstr->getNumOperands() - 1;
8044 for (int i=0; i < numArgs; ++i)
8045 argOpers[i] = &bInstr->getOperand(i+1);
8047 // x86 address has 4 operands: base, index, scale, and displacement
8048 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8049 int valArgIndx = lastAddrIndx + 1;
8051 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8052 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8053 for (int i=0; i <= lastAddrIndx; ++i)
8054 (*MIB).addOperand(*argOpers[i]);
8056 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8058 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8063 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8064 assert((argOpers[valArgIndx]->isReg() ||
8065 argOpers[valArgIndx]->isImm()) &&
8067 if (argOpers[valArgIndx]->isReg())
8068 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8070 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8072 (*MIB).addOperand(*argOpers[valArgIndx]);
8074 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8077 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8078 for (int i=0; i <= lastAddrIndx; ++i)
8079 (*MIB).addOperand(*argOpers[i]);
8081 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8082 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8083 bInstr->memoperands_end());
8085 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8089 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8091 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8095 // private utility function: 64 bit atomics on 32 bit host.
8097 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8098 MachineBasicBlock *MBB,
8103 bool invSrc) const {
8104 // For the atomic bitwise operator, we generate
8105 // thisMBB (instructions are in pairs, except cmpxchg8b)
8106 // ld t1,t2 = [bitinstr.addr]
8108 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8109 // op t5, t6 <- out1, out2, [bitinstr.val]
8110 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8111 // mov ECX, EBX <- t5, t6
8112 // mov EAX, EDX <- t1, t2
8113 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8114 // mov t3, t4 <- EAX, EDX
8116 // result in out1, out2
8117 // fallthrough -->nextMBB
8119 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8120 const unsigned LoadOpc = X86::MOV32rm;
8121 const unsigned copyOpc = X86::MOV32rr;
8122 const unsigned NotOpc = X86::NOT32r;
8123 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8124 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8125 MachineFunction::iterator MBBIter = MBB;
8128 /// First build the CFG
8129 MachineFunction *F = MBB->getParent();
8130 MachineBasicBlock *thisMBB = MBB;
8131 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8132 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8133 F->insert(MBBIter, newMBB);
8134 F->insert(MBBIter, nextMBB);
8136 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8137 nextMBB->splice(nextMBB->begin(), thisMBB,
8138 llvm::next(MachineBasicBlock::iterator(bInstr)),
8140 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8142 // Update thisMBB to fall through to newMBB
8143 thisMBB->addSuccessor(newMBB);
8145 // newMBB jumps to itself and fall through to nextMBB
8146 newMBB->addSuccessor(nextMBB);
8147 newMBB->addSuccessor(newMBB);
8149 DebugLoc dl = bInstr->getDebugLoc();
8150 // Insert instructions into newMBB based on incoming instruction
8151 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8152 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
8153 "unexpected number of operands");
8154 MachineOperand& dest1Oper = bInstr->getOperand(0);
8155 MachineOperand& dest2Oper = bInstr->getOperand(1);
8156 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8157 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
8158 argOpers[i] = &bInstr->getOperand(i+2);
8160 // We use some of the operands multiple times, so conservatively just
8161 // clear any kill flags that might be present.
8162 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8163 argOpers[i]->setIsKill(false);
8166 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8167 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8169 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8170 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8171 for (int i=0; i <= lastAddrIndx; ++i)
8172 (*MIB).addOperand(*argOpers[i]);
8173 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8174 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8175 // add 4 to displacement.
8176 for (int i=0; i <= lastAddrIndx-2; ++i)
8177 (*MIB).addOperand(*argOpers[i]);
8178 MachineOperand newOp3 = *(argOpers[3]);
8180 newOp3.setImm(newOp3.getImm()+4);
8182 newOp3.setOffset(newOp3.getOffset()+4);
8183 (*MIB).addOperand(newOp3);
8184 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8186 // t3/4 are defined later, at the bottom of the loop
8187 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8188 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8189 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8190 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8191 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8192 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8194 // The subsequent operations should be using the destination registers of
8195 //the PHI instructions.
8197 t1 = F->getRegInfo().createVirtualRegister(RC);
8198 t2 = F->getRegInfo().createVirtualRegister(RC);
8199 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8200 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8202 t1 = dest1Oper.getReg();
8203 t2 = dest2Oper.getReg();
8206 int valArgIndx = lastAddrIndx + 1;
8207 assert((argOpers[valArgIndx]->isReg() ||
8208 argOpers[valArgIndx]->isImm()) &&
8210 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8211 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8212 if (argOpers[valArgIndx]->isReg())
8213 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8215 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8216 if (regOpcL != X86::MOV32rr)
8218 (*MIB).addOperand(*argOpers[valArgIndx]);
8219 assert(argOpers[valArgIndx + 1]->isReg() ==
8220 argOpers[valArgIndx]->isReg());
8221 assert(argOpers[valArgIndx + 1]->isImm() ==
8222 argOpers[valArgIndx]->isImm());
8223 if (argOpers[valArgIndx + 1]->isReg())
8224 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8226 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8227 if (regOpcH != X86::MOV32rr)
8229 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8231 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8233 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8236 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8238 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8241 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8242 for (int i=0; i <= lastAddrIndx; ++i)
8243 (*MIB).addOperand(*argOpers[i]);
8245 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8246 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8247 bInstr->memoperands_end());
8249 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8250 MIB.addReg(X86::EAX);
8251 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8252 MIB.addReg(X86::EDX);
8255 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8257 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8261 // private utility function
8263 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8264 MachineBasicBlock *MBB,
8265 unsigned cmovOpc) const {
8266 // For the atomic min/max operator, we generate
8269 // ld t1 = [min/max.addr]
8270 // mov t2 = [min/max.val]
8272 // cmov[cond] t2 = t1
8274 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8276 // fallthrough -->nextMBB
8278 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8279 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8280 MachineFunction::iterator MBBIter = MBB;
8283 /// First build the CFG
8284 MachineFunction *F = MBB->getParent();
8285 MachineBasicBlock *thisMBB = MBB;
8286 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8287 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8288 F->insert(MBBIter, newMBB);
8289 F->insert(MBBIter, nextMBB);
8291 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8292 nextMBB->splice(nextMBB->begin(), thisMBB,
8293 llvm::next(MachineBasicBlock::iterator(mInstr)),
8295 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8297 // Update thisMBB to fall through to newMBB
8298 thisMBB->addSuccessor(newMBB);
8300 // newMBB jumps to newMBB and fall through to nextMBB
8301 newMBB->addSuccessor(nextMBB);
8302 newMBB->addSuccessor(newMBB);
8304 DebugLoc dl = mInstr->getDebugLoc();
8305 // Insert instructions into newMBB based on incoming instruction
8306 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8307 "unexpected number of operands");
8308 MachineOperand& destOper = mInstr->getOperand(0);
8309 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8310 int numArgs = mInstr->getNumOperands() - 1;
8311 for (int i=0; i < numArgs; ++i)
8312 argOpers[i] = &mInstr->getOperand(i+1);
8314 // x86 address has 4 operands: base, index, scale, and displacement
8315 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8316 int valArgIndx = lastAddrIndx + 1;
8318 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8319 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8320 for (int i=0; i <= lastAddrIndx; ++i)
8321 (*MIB).addOperand(*argOpers[i]);
8323 // We only support register and immediate values
8324 assert((argOpers[valArgIndx]->isReg() ||
8325 argOpers[valArgIndx]->isImm()) &&
8328 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8329 if (argOpers[valArgIndx]->isReg())
8330 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8332 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8333 (*MIB).addOperand(*argOpers[valArgIndx]);
8335 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8338 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8343 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8344 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8348 // Cmp and exchange if none has modified the memory location
8349 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8350 for (int i=0; i <= lastAddrIndx; ++i)
8351 (*MIB).addOperand(*argOpers[i]);
8353 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8354 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8355 mInstr->memoperands_end());
8357 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8358 MIB.addReg(X86::EAX);
8361 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8363 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
8367 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8368 // all of this code can be replaced with that in the .td file.
8370 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8371 unsigned numArgs, bool memArg) const {
8373 DebugLoc dl = MI->getDebugLoc();
8374 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8378 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8380 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8382 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8384 for (unsigned i = 0; i < numArgs; ++i) {
8385 MachineOperand &Op = MI->getOperand(i+1);
8387 if (!(Op.isReg() && Op.isImplicit()))
8391 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8394 MI->eraseFromParent();
8400 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8402 MachineBasicBlock *MBB) const {
8403 // Emit code to save XMM registers to the stack. The ABI says that the
8404 // number of registers to save is given in %al, so it's theoretically
8405 // possible to do an indirect jump trick to avoid saving all of them,
8406 // however this code takes a simpler approach and just executes all
8407 // of the stores if %al is non-zero. It's less code, and it's probably
8408 // easier on the hardware branch predictor, and stores aren't all that
8409 // expensive anyway.
8411 // Create the new basic blocks. One block contains all the XMM stores,
8412 // and one block is the final destination regardless of whether any
8413 // stores were performed.
8414 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8415 MachineFunction *F = MBB->getParent();
8416 MachineFunction::iterator MBBIter = MBB;
8418 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8419 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8420 F->insert(MBBIter, XMMSaveMBB);
8421 F->insert(MBBIter, EndMBB);
8423 // Transfer the remainder of MBB and its successor edges to EndMBB.
8424 EndMBB->splice(EndMBB->begin(), MBB,
8425 llvm::next(MachineBasicBlock::iterator(MI)),
8427 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
8429 // The original block will now fall through to the XMM save block.
8430 MBB->addSuccessor(XMMSaveMBB);
8431 // The XMMSaveMBB will fall through to the end block.
8432 XMMSaveMBB->addSuccessor(EndMBB);
8434 // Now add the instructions.
8435 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8436 DebugLoc DL = MI->getDebugLoc();
8438 unsigned CountReg = MI->getOperand(0).getReg();
8439 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8440 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8442 if (!Subtarget->isTargetWin64()) {
8443 // If %al is 0, branch around the XMM save block.
8444 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8445 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8446 MBB->addSuccessor(EndMBB);
8449 // In the XMM save block, save all the XMM argument registers.
8450 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8451 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8452 MachineMemOperand *MMO =
8453 F->getMachineMemOperand(
8454 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8455 MachineMemOperand::MOStore, Offset,
8456 /*Size=*/16, /*Align=*/16);
8457 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8458 .addFrameIndex(RegSaveFrameIndex)
8459 .addImm(/*Scale=*/1)
8460 .addReg(/*IndexReg=*/0)
8461 .addImm(/*Disp=*/Offset)
8462 .addReg(/*Segment=*/0)
8463 .addReg(MI->getOperand(i).getReg())
8464 .addMemOperand(MMO);
8467 MI->eraseFromParent(); // The pseudo instruction is gone now.
8473 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8474 MachineBasicBlock *BB) const {
8475 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8476 DebugLoc DL = MI->getDebugLoc();
8478 // To "insert" a SELECT_CC instruction, we actually have to insert the
8479 // diamond control-flow pattern. The incoming instruction knows the
8480 // destination vreg to set, the condition code register to branch on, the
8481 // true/false values to select between, and a branch opcode to use.
8482 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8483 MachineFunction::iterator It = BB;
8489 // cmpTY ccX, r1, r2
8491 // fallthrough --> copy0MBB
8492 MachineBasicBlock *thisMBB = BB;
8493 MachineFunction *F = BB->getParent();
8494 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8495 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8496 F->insert(It, copy0MBB);
8497 F->insert(It, sinkMBB);
8499 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8500 // live into the sink and copy blocks.
8501 const MachineFunction *MF = BB->getParent();
8502 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8503 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8505 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
8506 const MachineOperand &MO = MI->getOperand(I);
8507 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
8508 unsigned Reg = MO.getReg();
8509 if (Reg != X86::EFLAGS) continue;
8510 copy0MBB->addLiveIn(Reg);
8511 sinkMBB->addLiveIn(Reg);
8514 // Transfer the remainder of BB and its successor edges to sinkMBB.
8515 sinkMBB->splice(sinkMBB->begin(), BB,
8516 llvm::next(MachineBasicBlock::iterator(MI)),
8518 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
8520 // Add the true and fallthrough blocks as its successors.
8521 BB->addSuccessor(copy0MBB);
8522 BB->addSuccessor(sinkMBB);
8524 // Create the conditional branch instruction.
8526 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8527 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8530 // %FalseValue = ...
8531 // # fallthrough to sinkMBB
8532 copy0MBB->addSuccessor(sinkMBB);
8535 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8537 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
8538 TII->get(X86::PHI), MI->getOperand(0).getReg())
8539 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8540 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8542 MI->eraseFromParent(); // The pseudo instruction is gone now.
8547 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8548 MachineBasicBlock *BB) const {
8549 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8550 DebugLoc DL = MI->getDebugLoc();
8552 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8553 // non-trivial part is impdef of ESP.
8554 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8557 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
8558 .addExternalSymbol("_alloca")
8559 .addReg(X86::EAX, RegState::Implicit)
8560 .addReg(X86::ESP, RegState::Implicit)
8561 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8562 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8564 MI->eraseFromParent(); // The pseudo instruction is gone now.
8569 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8570 MachineBasicBlock *BB) const {
8571 // This is pretty easy. We're taking the value that we received from
8572 // our load from the relocation, sticking it in either RDI (x86-64)
8573 // or EAX and doing an indirect call. The return value will then
8574 // be in the normal return register.
8575 const X86InstrInfo *TII
8576 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8577 DebugLoc DL = MI->getDebugLoc();
8578 MachineFunction *F = BB->getParent();
8580 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8582 if (Subtarget->is64Bit()) {
8583 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8584 TII->get(X86::MOV64rm), X86::RDI)
8586 .addImm(0).addReg(0)
8587 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8588 MI->getOperand(3).getTargetFlags())
8590 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
8591 addDirectMem(MIB, X86::RDI);
8592 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8593 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8594 TII->get(X86::MOV32rm), X86::EAX)
8596 .addImm(0).addReg(0)
8597 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8598 MI->getOperand(3).getTargetFlags())
8600 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8601 addDirectMem(MIB, X86::EAX);
8603 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8604 TII->get(X86::MOV32rm), X86::EAX)
8605 .addReg(TII->getGlobalBaseReg(F))
8606 .addImm(0).addReg(0)
8607 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8608 MI->getOperand(3).getTargetFlags())
8610 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8611 addDirectMem(MIB, X86::EAX);
8614 MI->eraseFromParent(); // The pseudo instruction is gone now.
8619 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8620 MachineBasicBlock *BB) const {
8621 switch (MI->getOpcode()) {
8622 default: assert(false && "Unexpected instr type to insert");
8623 case X86::MINGW_ALLOCA:
8624 return EmitLoweredMingwAlloca(MI, BB);
8625 case X86::TLSCall_32:
8626 case X86::TLSCall_64:
8627 return EmitLoweredTLSCall(MI, BB);
8629 case X86::CMOV_V1I64:
8630 case X86::CMOV_FR32:
8631 case X86::CMOV_FR64:
8632 case X86::CMOV_V4F32:
8633 case X86::CMOV_V2F64:
8634 case X86::CMOV_V2I64:
8635 case X86::CMOV_GR16:
8636 case X86::CMOV_GR32:
8637 case X86::CMOV_RFP32:
8638 case X86::CMOV_RFP64:
8639 case X86::CMOV_RFP80:
8640 return EmitLoweredSelect(MI, BB);
8642 case X86::FP32_TO_INT16_IN_MEM:
8643 case X86::FP32_TO_INT32_IN_MEM:
8644 case X86::FP32_TO_INT64_IN_MEM:
8645 case X86::FP64_TO_INT16_IN_MEM:
8646 case X86::FP64_TO_INT32_IN_MEM:
8647 case X86::FP64_TO_INT64_IN_MEM:
8648 case X86::FP80_TO_INT16_IN_MEM:
8649 case X86::FP80_TO_INT32_IN_MEM:
8650 case X86::FP80_TO_INT64_IN_MEM: {
8651 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8652 DebugLoc DL = MI->getDebugLoc();
8654 // Change the floating point control register to use "round towards zero"
8655 // mode when truncating to an integer value.
8656 MachineFunction *F = BB->getParent();
8657 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8658 addFrameReference(BuildMI(*BB, MI, DL,
8659 TII->get(X86::FNSTCW16m)), CWFrameIdx);
8661 // Load the old value of the high byte of the control word...
8663 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8664 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
8667 // Set the high part to be round to zero...
8668 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8671 // Reload the modified control word now...
8672 addFrameReference(BuildMI(*BB, MI, DL,
8673 TII->get(X86::FLDCW16m)), CWFrameIdx);
8675 // Restore the memory image of control word to original value
8676 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8679 // Get the X86 opcode to use.
8681 switch (MI->getOpcode()) {
8682 default: llvm_unreachable("illegal opcode!");
8683 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8684 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8685 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8686 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8687 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8688 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8689 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8690 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8691 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8695 MachineOperand &Op = MI->getOperand(0);
8697 AM.BaseType = X86AddressMode::RegBase;
8698 AM.Base.Reg = Op.getReg();
8700 AM.BaseType = X86AddressMode::FrameIndexBase;
8701 AM.Base.FrameIndex = Op.getIndex();
8703 Op = MI->getOperand(1);
8705 AM.Scale = Op.getImm();
8706 Op = MI->getOperand(2);
8708 AM.IndexReg = Op.getImm();
8709 Op = MI->getOperand(3);
8710 if (Op.isGlobal()) {
8711 AM.GV = Op.getGlobal();
8713 AM.Disp = Op.getImm();
8715 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
8716 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
8718 // Reload the original control word now.
8719 addFrameReference(BuildMI(*BB, MI, DL,
8720 TII->get(X86::FLDCW16m)), CWFrameIdx);
8722 MI->eraseFromParent(); // The pseudo instruction is gone now.
8725 // String/text processing lowering.
8726 case X86::PCMPISTRM128REG:
8727 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8728 case X86::PCMPISTRM128MEM:
8729 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8730 case X86::PCMPESTRM128REG:
8731 return EmitPCMP(MI, BB, 5, false /* in mem */);
8732 case X86::PCMPESTRM128MEM:
8733 return EmitPCMP(MI, BB, 5, true /* in mem */);
8736 case X86::ATOMAND32:
8737 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8738 X86::AND32ri, X86::MOV32rm,
8739 X86::LCMPXCHG32, X86::MOV32rr,
8740 X86::NOT32r, X86::EAX,
8741 X86::GR32RegisterClass);
8743 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8744 X86::OR32ri, X86::MOV32rm,
8745 X86::LCMPXCHG32, X86::MOV32rr,
8746 X86::NOT32r, X86::EAX,
8747 X86::GR32RegisterClass);
8748 case X86::ATOMXOR32:
8749 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8750 X86::XOR32ri, X86::MOV32rm,
8751 X86::LCMPXCHG32, X86::MOV32rr,
8752 X86::NOT32r, X86::EAX,
8753 X86::GR32RegisterClass);
8754 case X86::ATOMNAND32:
8755 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8756 X86::AND32ri, X86::MOV32rm,
8757 X86::LCMPXCHG32, X86::MOV32rr,
8758 X86::NOT32r, X86::EAX,
8759 X86::GR32RegisterClass, true);
8760 case X86::ATOMMIN32:
8761 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8762 case X86::ATOMMAX32:
8763 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8764 case X86::ATOMUMIN32:
8765 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8766 case X86::ATOMUMAX32:
8767 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8769 case X86::ATOMAND16:
8770 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8771 X86::AND16ri, X86::MOV16rm,
8772 X86::LCMPXCHG16, X86::MOV16rr,
8773 X86::NOT16r, X86::AX,
8774 X86::GR16RegisterClass);
8776 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8777 X86::OR16ri, X86::MOV16rm,
8778 X86::LCMPXCHG16, X86::MOV16rr,
8779 X86::NOT16r, X86::AX,
8780 X86::GR16RegisterClass);
8781 case X86::ATOMXOR16:
8782 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8783 X86::XOR16ri, X86::MOV16rm,
8784 X86::LCMPXCHG16, X86::MOV16rr,
8785 X86::NOT16r, X86::AX,
8786 X86::GR16RegisterClass);
8787 case X86::ATOMNAND16:
8788 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8789 X86::AND16ri, X86::MOV16rm,
8790 X86::LCMPXCHG16, X86::MOV16rr,
8791 X86::NOT16r, X86::AX,
8792 X86::GR16RegisterClass, true);
8793 case X86::ATOMMIN16:
8794 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8795 case X86::ATOMMAX16:
8796 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8797 case X86::ATOMUMIN16:
8798 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8799 case X86::ATOMUMAX16:
8800 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8803 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8804 X86::AND8ri, X86::MOV8rm,
8805 X86::LCMPXCHG8, X86::MOV8rr,
8806 X86::NOT8r, X86::AL,
8807 X86::GR8RegisterClass);
8809 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8810 X86::OR8ri, X86::MOV8rm,
8811 X86::LCMPXCHG8, X86::MOV8rr,
8812 X86::NOT8r, X86::AL,
8813 X86::GR8RegisterClass);
8815 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8816 X86::XOR8ri, X86::MOV8rm,
8817 X86::LCMPXCHG8, X86::MOV8rr,
8818 X86::NOT8r, X86::AL,
8819 X86::GR8RegisterClass);
8820 case X86::ATOMNAND8:
8821 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8822 X86::AND8ri, X86::MOV8rm,
8823 X86::LCMPXCHG8, X86::MOV8rr,
8824 X86::NOT8r, X86::AL,
8825 X86::GR8RegisterClass, true);
8826 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8827 // This group is for 64-bit host.
8828 case X86::ATOMAND64:
8829 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8830 X86::AND64ri32, X86::MOV64rm,
8831 X86::LCMPXCHG64, X86::MOV64rr,
8832 X86::NOT64r, X86::RAX,
8833 X86::GR64RegisterClass);
8835 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8836 X86::OR64ri32, X86::MOV64rm,
8837 X86::LCMPXCHG64, X86::MOV64rr,
8838 X86::NOT64r, X86::RAX,
8839 X86::GR64RegisterClass);
8840 case X86::ATOMXOR64:
8841 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8842 X86::XOR64ri32, X86::MOV64rm,
8843 X86::LCMPXCHG64, X86::MOV64rr,
8844 X86::NOT64r, X86::RAX,
8845 X86::GR64RegisterClass);
8846 case X86::ATOMNAND64:
8847 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8848 X86::AND64ri32, X86::MOV64rm,
8849 X86::LCMPXCHG64, X86::MOV64rr,
8850 X86::NOT64r, X86::RAX,
8851 X86::GR64RegisterClass, true);
8852 case X86::ATOMMIN64:
8853 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8854 case X86::ATOMMAX64:
8855 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8856 case X86::ATOMUMIN64:
8857 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8858 case X86::ATOMUMAX64:
8859 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8861 // This group does 64-bit operations on a 32-bit host.
8862 case X86::ATOMAND6432:
8863 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8864 X86::AND32rr, X86::AND32rr,
8865 X86::AND32ri, X86::AND32ri,
8867 case X86::ATOMOR6432:
8868 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8869 X86::OR32rr, X86::OR32rr,
8870 X86::OR32ri, X86::OR32ri,
8872 case X86::ATOMXOR6432:
8873 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8874 X86::XOR32rr, X86::XOR32rr,
8875 X86::XOR32ri, X86::XOR32ri,
8877 case X86::ATOMNAND6432:
8878 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8879 X86::AND32rr, X86::AND32rr,
8880 X86::AND32ri, X86::AND32ri,
8882 case X86::ATOMADD6432:
8883 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8884 X86::ADD32rr, X86::ADC32rr,
8885 X86::ADD32ri, X86::ADC32ri,
8887 case X86::ATOMSUB6432:
8888 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8889 X86::SUB32rr, X86::SBB32rr,
8890 X86::SUB32ri, X86::SBB32ri,
8892 case X86::ATOMSWAP6432:
8893 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8894 X86::MOV32rr, X86::MOV32rr,
8895 X86::MOV32ri, X86::MOV32ri,
8897 case X86::VASTART_SAVE_XMM_REGS:
8898 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8902 //===----------------------------------------------------------------------===//
8903 // X86 Optimization Hooks
8904 //===----------------------------------------------------------------------===//
8906 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8910 const SelectionDAG &DAG,
8911 unsigned Depth) const {
8912 unsigned Opc = Op.getOpcode();
8913 assert((Opc >= ISD::BUILTIN_OP_END ||
8914 Opc == ISD::INTRINSIC_WO_CHAIN ||
8915 Opc == ISD::INTRINSIC_W_CHAIN ||
8916 Opc == ISD::INTRINSIC_VOID) &&
8917 "Should use MaskedValueIsZero if you don't know whether Op"
8918 " is a target node!");
8920 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8932 // These nodes' second result is a boolean.
8933 if (Op.getResNo() == 0)
8937 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8938 Mask.getBitWidth() - 1);
8943 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8944 /// node is a GlobalAddress + offset.
8945 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8946 const GlobalValue* &GA,
8947 int64_t &Offset) const {
8948 if (N->getOpcode() == X86ISD::Wrapper) {
8949 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8950 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8951 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8955 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8958 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8959 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8960 /// if the load addresses are consecutive, non-overlapping, and in the right
8962 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8963 const TargetLowering &TLI) {
8964 DebugLoc dl = N->getDebugLoc();
8965 EVT VT = N->getValueType(0);
8966 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8968 if (VT.getSizeInBits() != 128)
8971 SmallVector<SDValue, 16> Elts;
8972 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8973 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8975 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8978 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8979 /// and convert it from being a bunch of shuffles and extracts to a simple
8980 /// store and scalar loads to extract the elements.
8981 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8982 const TargetLowering &TLI) {
8983 SDValue InputVector = N->getOperand(0);
8985 // Only operate on vectors of 4 elements, where the alternative shuffling
8986 // gets to be more expensive.
8987 if (InputVector.getValueType() != MVT::v4i32)
8990 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8991 // single use which is a sign-extend or zero-extend, and all elements are
8993 SmallVector<SDNode *, 4> Uses;
8994 unsigned ExtractedElements = 0;
8995 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8996 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8997 if (UI.getUse().getResNo() != InputVector.getResNo())
9000 SDNode *Extract = *UI;
9001 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9004 if (Extract->getValueType(0) != MVT::i32)
9006 if (!Extract->hasOneUse())
9008 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9009 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9011 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9014 // Record which element was extracted.
9015 ExtractedElements |=
9016 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9018 Uses.push_back(Extract);
9021 // If not all the elements were used, this may not be worthwhile.
9022 if (ExtractedElements != 15)
9025 // Ok, we've now decided to do the transformation.
9026 DebugLoc dl = InputVector.getDebugLoc();
9028 // Store the value to a temporary stack slot.
9029 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9030 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
9033 // Replace each use (extract) with a load of the appropriate element.
9034 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9035 UE = Uses.end(); UI != UE; ++UI) {
9036 SDNode *Extract = *UI;
9038 // Compute the element's address.
9039 SDValue Idx = Extract->getOperand(1);
9041 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9042 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9043 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9045 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
9048 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
9049 NULL, 0, false, false, 0);
9051 // Replace the exact with the load.
9052 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9055 // The replacement was made in place; don't return anything.
9059 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9060 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9061 const X86Subtarget *Subtarget) {
9062 DebugLoc DL = N->getDebugLoc();
9063 SDValue Cond = N->getOperand(0);
9064 // Get the LHS/RHS of the select.
9065 SDValue LHS = N->getOperand(1);
9066 SDValue RHS = N->getOperand(2);
9068 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9069 // instructions match the semantics of the common C idiom x<y?x:y but not
9070 // x<=y?x:y, because of how they handle negative zero (which can be
9071 // ignored in unsafe-math mode).
9072 if (Subtarget->hasSSE2() &&
9073 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9074 Cond.getOpcode() == ISD::SETCC) {
9075 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9077 unsigned Opcode = 0;
9078 // Check for x CC y ? x : y.
9079 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9080 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9084 // Converting this to a min would handle NaNs incorrectly, and swapping
9085 // the operands would cause it to handle comparisons between positive
9086 // and negative zero incorrectly.
9087 if (!FiniteOnlyFPMath() &&
9088 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9089 if (!UnsafeFPMath &&
9090 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9092 std::swap(LHS, RHS);
9094 Opcode = X86ISD::FMIN;
9097 // Converting this to a min would handle comparisons between positive
9098 // and negative zero incorrectly.
9099 if (!UnsafeFPMath &&
9100 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9102 Opcode = X86ISD::FMIN;
9105 // Converting this to a min would handle both negative zeros and NaNs
9106 // incorrectly, but we can swap the operands to fix both.
9107 std::swap(LHS, RHS);
9111 Opcode = X86ISD::FMIN;
9115 // Converting this to a max would handle comparisons between positive
9116 // and negative zero incorrectly.
9117 if (!UnsafeFPMath &&
9118 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9120 Opcode = X86ISD::FMAX;
9123 // Converting this to a max would handle NaNs incorrectly, and swapping
9124 // the operands would cause it to handle comparisons between positive
9125 // and negative zero incorrectly.
9126 if (!FiniteOnlyFPMath() &&
9127 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9128 if (!UnsafeFPMath &&
9129 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9131 std::swap(LHS, RHS);
9133 Opcode = X86ISD::FMAX;
9136 // Converting this to a max would handle both negative zeros and NaNs
9137 // incorrectly, but we can swap the operands to fix both.
9138 std::swap(LHS, RHS);
9142 Opcode = X86ISD::FMAX;
9145 // Check for x CC y ? y : x -- a min/max with reversed arms.
9146 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9147 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9151 // Converting this to a min would handle comparisons between positive
9152 // and negative zero incorrectly, and swapping the operands would
9153 // cause it to handle NaNs incorrectly.
9154 if (!UnsafeFPMath &&
9155 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9156 if (!FiniteOnlyFPMath() &&
9157 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9159 std::swap(LHS, RHS);
9161 Opcode = X86ISD::FMIN;
9164 // Converting this to a min would handle NaNs incorrectly.
9165 if (!UnsafeFPMath &&
9166 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9168 Opcode = X86ISD::FMIN;
9171 // Converting this to a min would handle both negative zeros and NaNs
9172 // incorrectly, but we can swap the operands to fix both.
9173 std::swap(LHS, RHS);
9177 Opcode = X86ISD::FMIN;
9181 // Converting this to a max would handle NaNs incorrectly.
9182 if (!FiniteOnlyFPMath() &&
9183 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9185 Opcode = X86ISD::FMAX;
9188 // Converting this to a max would handle comparisons between positive
9189 // and negative zero incorrectly, and swapping the operands would
9190 // cause it to handle NaNs incorrectly.
9191 if (!UnsafeFPMath &&
9192 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9193 if (!FiniteOnlyFPMath() &&
9194 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9196 std::swap(LHS, RHS);
9198 Opcode = X86ISD::FMAX;
9201 // Converting this to a max would handle both negative zeros and NaNs
9202 // incorrectly, but we can swap the operands to fix both.
9203 std::swap(LHS, RHS);
9207 Opcode = X86ISD::FMAX;
9213 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9216 // If this is a select between two integer constants, try to do some
9218 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9219 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9220 // Don't do this for crazy integer types.
9221 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9222 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9223 // so that TrueC (the true value) is larger than FalseC.
9224 bool NeedsCondInvert = false;
9226 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9227 // Efficiently invertible.
9228 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9229 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9230 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9231 NeedsCondInvert = true;
9232 std::swap(TrueC, FalseC);
9235 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9236 if (FalseC->getAPIntValue() == 0 &&
9237 TrueC->getAPIntValue().isPowerOf2()) {
9238 if (NeedsCondInvert) // Invert the condition if needed.
9239 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9240 DAG.getConstant(1, Cond.getValueType()));
9242 // Zero extend the condition if needed.
9243 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9245 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9246 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9247 DAG.getConstant(ShAmt, MVT::i8));
9250 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9251 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9252 if (NeedsCondInvert) // Invert the condition if needed.
9253 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9254 DAG.getConstant(1, Cond.getValueType()));
9256 // Zero extend the condition if needed.
9257 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9258 FalseC->getValueType(0), Cond);
9259 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9260 SDValue(FalseC, 0));
9263 // Optimize cases that will turn into an LEA instruction. This requires
9264 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9265 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9266 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9267 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9269 bool isFastMultiplier = false;
9271 switch ((unsigned char)Diff) {
9273 case 1: // result = add base, cond
9274 case 2: // result = lea base( , cond*2)
9275 case 3: // result = lea base(cond, cond*2)
9276 case 4: // result = lea base( , cond*4)
9277 case 5: // result = lea base(cond, cond*4)
9278 case 8: // result = lea base( , cond*8)
9279 case 9: // result = lea base(cond, cond*8)
9280 isFastMultiplier = true;
9285 if (isFastMultiplier) {
9286 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9287 if (NeedsCondInvert) // Invert the condition if needed.
9288 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9289 DAG.getConstant(1, Cond.getValueType()));
9291 // Zero extend the condition if needed.
9292 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9294 // Scale the condition by the difference.
9296 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9297 DAG.getConstant(Diff, Cond.getValueType()));
9299 // Add the base if non-zero.
9300 if (FalseC->getAPIntValue() != 0)
9301 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9302 SDValue(FalseC, 0));
9312 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9313 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9314 TargetLowering::DAGCombinerInfo &DCI) {
9315 DebugLoc DL = N->getDebugLoc();
9317 // If the flag operand isn't dead, don't touch this CMOV.
9318 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9321 // If this is a select between two integer constants, try to do some
9322 // optimizations. Note that the operands are ordered the opposite of SELECT
9324 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9325 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9326 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9327 // larger than FalseC (the false value).
9328 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9330 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9331 CC = X86::GetOppositeBranchCondition(CC);
9332 std::swap(TrueC, FalseC);
9335 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9336 // This is efficient for any integer data type (including i8/i16) and
9338 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9339 SDValue Cond = N->getOperand(3);
9340 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9341 DAG.getConstant(CC, MVT::i8), Cond);
9343 // Zero extend the condition if needed.
9344 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9346 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9347 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9348 DAG.getConstant(ShAmt, MVT::i8));
9349 if (N->getNumValues() == 2) // Dead flag value?
9350 return DCI.CombineTo(N, Cond, SDValue());
9354 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9355 // for any integer data type, including i8/i16.
9356 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9357 SDValue Cond = N->getOperand(3);
9358 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9359 DAG.getConstant(CC, MVT::i8), Cond);
9361 // Zero extend the condition if needed.
9362 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9363 FalseC->getValueType(0), Cond);
9364 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9365 SDValue(FalseC, 0));
9367 if (N->getNumValues() == 2) // Dead flag value?
9368 return DCI.CombineTo(N, Cond, SDValue());
9372 // Optimize cases that will turn into an LEA instruction. This requires
9373 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9374 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9375 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9376 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9378 bool isFastMultiplier = false;
9380 switch ((unsigned char)Diff) {
9382 case 1: // result = add base, cond
9383 case 2: // result = lea base( , cond*2)
9384 case 3: // result = lea base(cond, cond*2)
9385 case 4: // result = lea base( , cond*4)
9386 case 5: // result = lea base(cond, cond*4)
9387 case 8: // result = lea base( , cond*8)
9388 case 9: // result = lea base(cond, cond*8)
9389 isFastMultiplier = true;
9394 if (isFastMultiplier) {
9395 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9396 SDValue Cond = N->getOperand(3);
9397 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9398 DAG.getConstant(CC, MVT::i8), Cond);
9399 // Zero extend the condition if needed.
9400 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9402 // Scale the condition by the difference.
9404 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9405 DAG.getConstant(Diff, Cond.getValueType()));
9407 // Add the base if non-zero.
9408 if (FalseC->getAPIntValue() != 0)
9409 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9410 SDValue(FalseC, 0));
9411 if (N->getNumValues() == 2) // Dead flag value?
9412 return DCI.CombineTo(N, Cond, SDValue());
9422 /// PerformMulCombine - Optimize a single multiply with constant into two
9423 /// in order to implement it with two cheaper instructions, e.g.
9424 /// LEA + SHL, LEA + LEA.
9425 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9426 TargetLowering::DAGCombinerInfo &DCI) {
9427 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9430 EVT VT = N->getValueType(0);
9434 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9437 uint64_t MulAmt = C->getZExtValue();
9438 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9441 uint64_t MulAmt1 = 0;
9442 uint64_t MulAmt2 = 0;
9443 if ((MulAmt % 9) == 0) {
9445 MulAmt2 = MulAmt / 9;
9446 } else if ((MulAmt % 5) == 0) {
9448 MulAmt2 = MulAmt / 5;
9449 } else if ((MulAmt % 3) == 0) {
9451 MulAmt2 = MulAmt / 3;
9454 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9455 DebugLoc DL = N->getDebugLoc();
9457 if (isPowerOf2_64(MulAmt2) &&
9458 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9459 // If second multiplifer is pow2, issue it first. We want the multiply by
9460 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9462 std::swap(MulAmt1, MulAmt2);
9465 if (isPowerOf2_64(MulAmt1))
9466 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9467 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9469 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9470 DAG.getConstant(MulAmt1, VT));
9472 if (isPowerOf2_64(MulAmt2))
9473 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9474 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9476 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9477 DAG.getConstant(MulAmt2, VT));
9479 // Do not add new nodes to DAG combiner worklist.
9480 DCI.CombineTo(N, NewMul, false);
9485 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9486 SDValue N0 = N->getOperand(0);
9487 SDValue N1 = N->getOperand(1);
9488 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9489 EVT VT = N0.getValueType();
9491 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9492 // since the result of setcc_c is all zero's or all ones.
9493 if (N1C && N0.getOpcode() == ISD::AND &&
9494 N0.getOperand(1).getOpcode() == ISD::Constant) {
9495 SDValue N00 = N0.getOperand(0);
9496 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9497 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9498 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9499 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9500 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9501 APInt ShAmt = N1C->getAPIntValue();
9502 Mask = Mask.shl(ShAmt);
9504 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9505 N00, DAG.getConstant(Mask, VT));
9512 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9514 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9515 const X86Subtarget *Subtarget) {
9516 EVT VT = N->getValueType(0);
9517 if (!VT.isVector() && VT.isInteger() &&
9518 N->getOpcode() == ISD::SHL)
9519 return PerformSHLCombine(N, DAG);
9521 // On X86 with SSE2 support, we can transform this to a vector shift if
9522 // all elements are shifted by the same amount. We can't do this in legalize
9523 // because the a constant vector is typically transformed to a constant pool
9524 // so we have no knowledge of the shift amount.
9525 if (!Subtarget->hasSSE2())
9528 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9531 SDValue ShAmtOp = N->getOperand(1);
9532 EVT EltVT = VT.getVectorElementType();
9533 DebugLoc DL = N->getDebugLoc();
9534 SDValue BaseShAmt = SDValue();
9535 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9536 unsigned NumElts = VT.getVectorNumElements();
9538 for (; i != NumElts; ++i) {
9539 SDValue Arg = ShAmtOp.getOperand(i);
9540 if (Arg.getOpcode() == ISD::UNDEF) continue;
9544 for (; i != NumElts; ++i) {
9545 SDValue Arg = ShAmtOp.getOperand(i);
9546 if (Arg.getOpcode() == ISD::UNDEF) continue;
9547 if (Arg != BaseShAmt) {
9551 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9552 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9553 SDValue InVec = ShAmtOp.getOperand(0);
9554 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9555 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9557 for (; i != NumElts; ++i) {
9558 SDValue Arg = InVec.getOperand(i);
9559 if (Arg.getOpcode() == ISD::UNDEF) continue;
9563 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9564 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9565 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9566 if (C->getZExtValue() == SplatIdx)
9567 BaseShAmt = InVec.getOperand(1);
9570 if (BaseShAmt.getNode() == 0)
9571 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9572 DAG.getIntPtrConstant(0));
9576 // The shift amount is an i32.
9577 if (EltVT.bitsGT(MVT::i32))
9578 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9579 else if (EltVT.bitsLT(MVT::i32))
9580 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9582 // The shift amount is identical so we can do a vector shift.
9583 SDValue ValOp = N->getOperand(0);
9584 switch (N->getOpcode()) {
9586 llvm_unreachable("Unknown shift opcode!");
9589 if (VT == MVT::v2i64)
9590 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9591 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9593 if (VT == MVT::v4i32)
9594 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9595 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9597 if (VT == MVT::v8i16)
9598 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9599 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9603 if (VT == MVT::v4i32)
9604 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9605 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9607 if (VT == MVT::v8i16)
9608 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9609 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9613 if (VT == MVT::v2i64)
9614 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9615 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9617 if (VT == MVT::v4i32)
9618 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9619 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9621 if (VT == MVT::v8i16)
9622 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9623 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9630 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9631 TargetLowering::DAGCombinerInfo &DCI,
9632 const X86Subtarget *Subtarget) {
9633 if (DCI.isBeforeLegalizeOps())
9636 EVT VT = N->getValueType(0);
9637 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9640 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9641 SDValue N0 = N->getOperand(0);
9642 SDValue N1 = N->getOperand(1);
9643 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9645 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9647 if (!N0.hasOneUse() || !N1.hasOneUse())
9650 SDValue ShAmt0 = N0.getOperand(1);
9651 if (ShAmt0.getValueType() != MVT::i8)
9653 SDValue ShAmt1 = N1.getOperand(1);
9654 if (ShAmt1.getValueType() != MVT::i8)
9656 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9657 ShAmt0 = ShAmt0.getOperand(0);
9658 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9659 ShAmt1 = ShAmt1.getOperand(0);
9661 DebugLoc DL = N->getDebugLoc();
9662 unsigned Opc = X86ISD::SHLD;
9663 SDValue Op0 = N0.getOperand(0);
9664 SDValue Op1 = N1.getOperand(0);
9665 if (ShAmt0.getOpcode() == ISD::SUB) {
9667 std::swap(Op0, Op1);
9668 std::swap(ShAmt0, ShAmt1);
9671 unsigned Bits = VT.getSizeInBits();
9672 if (ShAmt1.getOpcode() == ISD::SUB) {
9673 SDValue Sum = ShAmt1.getOperand(0);
9674 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9675 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
9676 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
9677 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
9678 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
9679 return DAG.getNode(Opc, DL, VT,
9681 DAG.getNode(ISD::TRUNCATE, DL,
9684 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9685 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9687 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9688 return DAG.getNode(Opc, DL, VT,
9689 N0.getOperand(0), N1.getOperand(0),
9690 DAG.getNode(ISD::TRUNCATE, DL,
9697 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9698 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9699 const X86Subtarget *Subtarget) {
9700 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9701 // the FP state in cases where an emms may be missing.
9702 // A preferable solution to the general problem is to figure out the right
9703 // places to insert EMMS. This qualifies as a quick hack.
9705 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9706 StoreSDNode *St = cast<StoreSDNode>(N);
9707 EVT VT = St->getValue().getValueType();
9708 if (VT.getSizeInBits() != 64)
9711 const Function *F = DAG.getMachineFunction().getFunction();
9712 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9713 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9714 && Subtarget->hasSSE2();
9715 if ((VT.isVector() ||
9716 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9717 isa<LoadSDNode>(St->getValue()) &&
9718 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9719 St->getChain().hasOneUse() && !St->isVolatile()) {
9720 SDNode* LdVal = St->getValue().getNode();
9722 int TokenFactorIndex = -1;
9723 SmallVector<SDValue, 8> Ops;
9724 SDNode* ChainVal = St->getChain().getNode();
9725 // Must be a store of a load. We currently handle two cases: the load
9726 // is a direct child, and it's under an intervening TokenFactor. It is
9727 // possible to dig deeper under nested TokenFactors.
9728 if (ChainVal == LdVal)
9729 Ld = cast<LoadSDNode>(St->getChain());
9730 else if (St->getValue().hasOneUse() &&
9731 ChainVal->getOpcode() == ISD::TokenFactor) {
9732 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9733 if (ChainVal->getOperand(i).getNode() == LdVal) {
9734 TokenFactorIndex = i;
9735 Ld = cast<LoadSDNode>(St->getValue());
9737 Ops.push_back(ChainVal->getOperand(i));
9741 if (!Ld || !ISD::isNormalLoad(Ld))
9744 // If this is not the MMX case, i.e. we are just turning i64 load/store
9745 // into f64 load/store, avoid the transformation if there are multiple
9746 // uses of the loaded value.
9747 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9750 DebugLoc LdDL = Ld->getDebugLoc();
9751 DebugLoc StDL = N->getDebugLoc();
9752 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9753 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9755 if (Subtarget->is64Bit() || F64IsLegal) {
9756 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9757 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9758 Ld->getBasePtr(), Ld->getSrcValue(),
9759 Ld->getSrcValueOffset(), Ld->isVolatile(),
9760 Ld->isNonTemporal(), Ld->getAlignment());
9761 SDValue NewChain = NewLd.getValue(1);
9762 if (TokenFactorIndex != -1) {
9763 Ops.push_back(NewChain);
9764 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9767 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9768 St->getSrcValue(), St->getSrcValueOffset(),
9769 St->isVolatile(), St->isNonTemporal(),
9770 St->getAlignment());
9773 // Otherwise, lower to two pairs of 32-bit loads / stores.
9774 SDValue LoAddr = Ld->getBasePtr();
9775 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9776 DAG.getConstant(4, MVT::i32));
9778 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9779 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9780 Ld->isVolatile(), Ld->isNonTemporal(),
9781 Ld->getAlignment());
9782 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9783 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9784 Ld->isVolatile(), Ld->isNonTemporal(),
9785 MinAlign(Ld->getAlignment(), 4));
9787 SDValue NewChain = LoLd.getValue(1);
9788 if (TokenFactorIndex != -1) {
9789 Ops.push_back(LoLd);
9790 Ops.push_back(HiLd);
9791 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9795 LoAddr = St->getBasePtr();
9796 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9797 DAG.getConstant(4, MVT::i32));
9799 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9800 St->getSrcValue(), St->getSrcValueOffset(),
9801 St->isVolatile(), St->isNonTemporal(),
9802 St->getAlignment());
9803 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9805 St->getSrcValueOffset() + 4,
9807 St->isNonTemporal(),
9808 MinAlign(St->getAlignment(), 4));
9809 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9814 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9815 /// X86ISD::FXOR nodes.
9816 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9817 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9818 // F[X]OR(0.0, x) -> x
9819 // F[X]OR(x, 0.0) -> x
9820 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9821 if (C->getValueAPF().isPosZero())
9822 return N->getOperand(1);
9823 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9824 if (C->getValueAPF().isPosZero())
9825 return N->getOperand(0);
9829 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9830 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9831 // FAND(0.0, x) -> 0.0
9832 // FAND(x, 0.0) -> 0.0
9833 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9834 if (C->getValueAPF().isPosZero())
9835 return N->getOperand(0);
9836 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9837 if (C->getValueAPF().isPosZero())
9838 return N->getOperand(1);
9842 static SDValue PerformBTCombine(SDNode *N,
9844 TargetLowering::DAGCombinerInfo &DCI) {
9845 // BT ignores high bits in the bit index operand.
9846 SDValue Op1 = N->getOperand(1);
9847 if (Op1.hasOneUse()) {
9848 unsigned BitWidth = Op1.getValueSizeInBits();
9849 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9850 APInt KnownZero, KnownOne;
9851 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9852 !DCI.isBeforeLegalizeOps());
9853 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9854 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9855 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9856 DCI.CommitTargetLoweringOpt(TLO);
9861 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9862 SDValue Op = N->getOperand(0);
9863 if (Op.getOpcode() == ISD::BIT_CONVERT)
9864 Op = Op.getOperand(0);
9865 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9866 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9867 VT.getVectorElementType().getSizeInBits() ==
9868 OpVT.getVectorElementType().getSizeInBits()) {
9869 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9874 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9875 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9876 // (and (i32 x86isd::setcc_carry), 1)
9877 // This eliminates the zext. This transformation is necessary because
9878 // ISD::SETCC is always legalized to i8.
9879 DebugLoc dl = N->getDebugLoc();
9880 SDValue N0 = N->getOperand(0);
9881 EVT VT = N->getValueType(0);
9882 if (N0.getOpcode() == ISD::AND &&
9884 N0.getOperand(0).hasOneUse()) {
9885 SDValue N00 = N0.getOperand(0);
9886 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9888 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9889 if (!C || C->getZExtValue() != 1)
9891 return DAG.getNode(ISD::AND, dl, VT,
9892 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9893 N00.getOperand(0), N00.getOperand(1)),
9894 DAG.getConstant(1, VT));
9900 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9901 DAGCombinerInfo &DCI) const {
9902 SelectionDAG &DAG = DCI.DAG;
9903 switch (N->getOpcode()) {
9905 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9906 case ISD::EXTRACT_VECTOR_ELT:
9907 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9908 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9909 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9910 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9913 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9914 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
9915 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9917 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9918 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9919 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9920 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9921 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9927 /// isTypeDesirableForOp - Return true if the target has native support for
9928 /// the specified value type and it is 'desirable' to use the type for the
9929 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9930 /// instruction encodings are longer and some i16 instructions are slow.
9931 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9932 if (!isTypeLegal(VT))
9941 case ISD::SIGN_EXTEND:
9942 case ISD::ZERO_EXTEND:
9943 case ISD::ANY_EXTEND:
9956 static bool MayFoldLoad(SDValue Op) {
9957 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
9960 static bool MayFoldIntoStore(SDValue Op) {
9961 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
9964 /// IsDesirableToPromoteOp - This method query the target whether it is
9965 /// beneficial for dag combiner to promote the specified node. If true, it
9966 /// should return the desired promotion type by reference.
9967 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
9968 EVT VT = Op.getValueType();
9972 bool Promote = false;
9973 bool Commute = false;
9974 switch (Op.getOpcode()) {
9977 LoadSDNode *LD = cast<LoadSDNode>(Op);
9978 // If the non-extending load has a single use and it's not live out, then it
9980 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
9982 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9983 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
9984 // The only case where we'd want to promote LOAD (rather then it being
9985 // promoted as an operand is when it's only use is liveout.
9986 if (UI->getOpcode() != ISD::CopyToReg)
9993 case ISD::SIGN_EXTEND:
9994 case ISD::ZERO_EXTEND:
9995 case ISD::ANY_EXTEND:
10000 SDValue N0 = Op.getOperand(0);
10001 // Look out for (store (shl (load), x)).
10002 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10015 SDValue N0 = Op.getOperand(0);
10016 SDValue N1 = Op.getOperand(1);
10017 if (!Commute && MayFoldLoad(N1))
10019 // Avoid disabling potential load folding opportunities.
10020 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10022 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10032 //===----------------------------------------------------------------------===//
10033 // X86 Inline Assembly Support
10034 //===----------------------------------------------------------------------===//
10036 static bool LowerToBSwap(CallInst *CI) {
10037 // FIXME: this should verify that we are targetting a 486 or better. If not,
10038 // we will turn this bswap into something that will be lowered to logical ops
10039 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10040 // so don't worry about this.
10042 // Verify this is a simple bswap.
10043 if (CI->getNumArgOperands() != 1 ||
10044 CI->getType() != CI->getArgOperand(0)->getType() ||
10045 !CI->getType()->isIntegerTy())
10048 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10049 if (!Ty || Ty->getBitWidth() % 16 != 0)
10052 // Okay, we can do this xform, do so now.
10053 const Type *Tys[] = { Ty };
10054 Module *M = CI->getParent()->getParent()->getParent();
10055 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10057 Value *Op = CI->getArgOperand(0);
10058 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10060 CI->replaceAllUsesWith(Op);
10061 CI->eraseFromParent();
10065 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10066 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10067 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10069 std::string AsmStr = IA->getAsmString();
10071 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10072 SmallVector<StringRef, 4> AsmPieces;
10073 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10075 switch (AsmPieces.size()) {
10076 default: return false;
10078 AsmStr = AsmPieces[0];
10080 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10083 if (AsmPieces.size() == 2 &&
10084 (AsmPieces[0] == "bswap" ||
10085 AsmPieces[0] == "bswapq" ||
10086 AsmPieces[0] == "bswapl") &&
10087 (AsmPieces[1] == "$0" ||
10088 AsmPieces[1] == "${0:q}")) {
10089 // No need to check constraints, nothing other than the equivalent of
10090 // "=r,0" would be valid here.
10091 return LowerToBSwap(CI);
10093 // rorw $$8, ${0:w} --> llvm.bswap.i16
10094 if (CI->getType()->isIntegerTy(16) &&
10095 AsmPieces.size() == 3 &&
10096 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10097 AsmPieces[1] == "$$8," &&
10098 AsmPieces[2] == "${0:w}" &&
10099 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10101 const std::string &Constraints = IA->getConstraintString();
10102 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10103 std::sort(AsmPieces.begin(), AsmPieces.end());
10104 if (AsmPieces.size() == 4 &&
10105 AsmPieces[0] == "~{cc}" &&
10106 AsmPieces[1] == "~{dirflag}" &&
10107 AsmPieces[2] == "~{flags}" &&
10108 AsmPieces[3] == "~{fpsr}") {
10109 return LowerToBSwap(CI);
10114 if (CI->getType()->isIntegerTy(64) &&
10115 Constraints.size() >= 2 &&
10116 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10117 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10118 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10119 SmallVector<StringRef, 4> Words;
10120 SplitString(AsmPieces[0], Words, " \t");
10121 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10123 SplitString(AsmPieces[1], Words, " \t");
10124 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10126 SplitString(AsmPieces[2], Words, " \t,");
10127 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10128 Words[2] == "%edx") {
10129 return LowerToBSwap(CI);
10141 /// getConstraintType - Given a constraint letter, return the type of
10142 /// constraint it is for this target.
10143 X86TargetLowering::ConstraintType
10144 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10145 if (Constraint.size() == 1) {
10146 switch (Constraint[0]) {
10158 return C_RegisterClass;
10166 return TargetLowering::getConstraintType(Constraint);
10169 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10170 /// with another that has more specific requirements based on the type of the
10171 /// corresponding operand.
10172 const char *X86TargetLowering::
10173 LowerXConstraint(EVT ConstraintVT) const {
10174 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10175 // 'f' like normal targets.
10176 if (ConstraintVT.isFloatingPoint()) {
10177 if (Subtarget->hasSSE2())
10179 if (Subtarget->hasSSE1())
10183 return TargetLowering::LowerXConstraint(ConstraintVT);
10186 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10187 /// vector. If it is invalid, don't add anything to Ops.
10188 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10190 std::vector<SDValue>&Ops,
10191 SelectionDAG &DAG) const {
10192 SDValue Result(0, 0);
10194 switch (Constraint) {
10197 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10198 if (C->getZExtValue() <= 31) {
10199 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10205 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10206 if (C->getZExtValue() <= 63) {
10207 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10213 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10214 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10215 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10221 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10222 if (C->getZExtValue() <= 255) {
10223 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10229 // 32-bit signed value
10230 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10231 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10232 C->getSExtValue())) {
10233 // Widen to 64 bits here to get it sign extended.
10234 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10237 // FIXME gcc accepts some relocatable values here too, but only in certain
10238 // memory models; it's complicated.
10243 // 32-bit unsigned value
10244 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10245 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10246 C->getZExtValue())) {
10247 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10251 // FIXME gcc accepts some relocatable values here too, but only in certain
10252 // memory models; it's complicated.
10256 // Literal immediates are always ok.
10257 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10258 // Widen to 64 bits here to get it sign extended.
10259 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10263 // In any sort of PIC mode addresses need to be computed at runtime by
10264 // adding in a register or some sort of table lookup. These can't
10265 // be used as immediates.
10266 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
10269 // If we are in non-pic codegen mode, we allow the address of a global (with
10270 // an optional displacement) to be used with 'i'.
10271 GlobalAddressSDNode *GA = 0;
10272 int64_t Offset = 0;
10274 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10276 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10277 Offset += GA->getOffset();
10279 } else if (Op.getOpcode() == ISD::ADD) {
10280 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10281 Offset += C->getZExtValue();
10282 Op = Op.getOperand(0);
10285 } else if (Op.getOpcode() == ISD::SUB) {
10286 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10287 Offset += -C->getZExtValue();
10288 Op = Op.getOperand(0);
10293 // Otherwise, this isn't something we can handle, reject it.
10297 const GlobalValue *GV = GA->getGlobal();
10298 // If we require an extra load to get this address, as in PIC mode, we
10299 // can't accept it.
10300 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10301 getTargetMachine())))
10304 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
10305 GA->getValueType(0), Offset);
10310 if (Result.getNode()) {
10311 Ops.push_back(Result);
10314 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10317 std::vector<unsigned> X86TargetLowering::
10318 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10320 if (Constraint.size() == 1) {
10321 // FIXME: not handling fp-stack yet!
10322 switch (Constraint[0]) { // GCC X86 Constraint Letters
10323 default: break; // Unknown constraint letter
10324 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10325 if (Subtarget->is64Bit()) {
10326 if (VT == MVT::i32)
10327 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10328 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10329 X86::R10D,X86::R11D,X86::R12D,
10330 X86::R13D,X86::R14D,X86::R15D,
10331 X86::EBP, X86::ESP, 0);
10332 else if (VT == MVT::i16)
10333 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10334 X86::SI, X86::DI, X86::R8W,X86::R9W,
10335 X86::R10W,X86::R11W,X86::R12W,
10336 X86::R13W,X86::R14W,X86::R15W,
10337 X86::BP, X86::SP, 0);
10338 else if (VT == MVT::i8)
10339 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10340 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10341 X86::R10B,X86::R11B,X86::R12B,
10342 X86::R13B,X86::R14B,X86::R15B,
10343 X86::BPL, X86::SPL, 0);
10345 else if (VT == MVT::i64)
10346 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10347 X86::RSI, X86::RDI, X86::R8, X86::R9,
10348 X86::R10, X86::R11, X86::R12,
10349 X86::R13, X86::R14, X86::R15,
10350 X86::RBP, X86::RSP, 0);
10354 // 32-bit fallthrough
10355 case 'Q': // Q_REGS
10356 if (VT == MVT::i32)
10357 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10358 else if (VT == MVT::i16)
10359 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10360 else if (VT == MVT::i8)
10361 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10362 else if (VT == MVT::i64)
10363 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10368 return std::vector<unsigned>();
10371 std::pair<unsigned, const TargetRegisterClass*>
10372 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10374 // First, see if this is a constraint that directly corresponds to an LLVM
10376 if (Constraint.size() == 1) {
10377 // GCC Constraint Letters
10378 switch (Constraint[0]) {
10380 case 'r': // GENERAL_REGS
10381 case 'l': // INDEX_REGS
10383 return std::make_pair(0U, X86::GR8RegisterClass);
10384 if (VT == MVT::i16)
10385 return std::make_pair(0U, X86::GR16RegisterClass);
10386 if (VT == MVT::i32 || !Subtarget->is64Bit())
10387 return std::make_pair(0U, X86::GR32RegisterClass);
10388 return std::make_pair(0U, X86::GR64RegisterClass);
10389 case 'R': // LEGACY_REGS
10391 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10392 if (VT == MVT::i16)
10393 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10394 if (VT == MVT::i32 || !Subtarget->is64Bit())
10395 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10396 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10397 case 'f': // FP Stack registers.
10398 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10399 // value to the correct fpstack register class.
10400 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10401 return std::make_pair(0U, X86::RFP32RegisterClass);
10402 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10403 return std::make_pair(0U, X86::RFP64RegisterClass);
10404 return std::make_pair(0U, X86::RFP80RegisterClass);
10405 case 'y': // MMX_REGS if MMX allowed.
10406 if (!Subtarget->hasMMX()) break;
10407 return std::make_pair(0U, X86::VR64RegisterClass);
10408 case 'Y': // SSE_REGS if SSE2 allowed
10409 if (!Subtarget->hasSSE2()) break;
10411 case 'x': // SSE_REGS if SSE1 allowed
10412 if (!Subtarget->hasSSE1()) break;
10414 switch (VT.getSimpleVT().SimpleTy) {
10416 // Scalar SSE types.
10419 return std::make_pair(0U, X86::FR32RegisterClass);
10422 return std::make_pair(0U, X86::FR64RegisterClass);
10430 return std::make_pair(0U, X86::VR128RegisterClass);
10436 // Use the default implementation in TargetLowering to convert the register
10437 // constraint into a member of a register class.
10438 std::pair<unsigned, const TargetRegisterClass*> Res;
10439 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10441 // Not found as a standard register?
10442 if (Res.second == 0) {
10443 // Map st(0) -> st(7) -> ST0
10444 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10445 tolower(Constraint[1]) == 's' &&
10446 tolower(Constraint[2]) == 't' &&
10447 Constraint[3] == '(' &&
10448 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10449 Constraint[5] == ')' &&
10450 Constraint[6] == '}') {
10452 Res.first = X86::ST0+Constraint[4]-'0';
10453 Res.second = X86::RFP80RegisterClass;
10457 // GCC allows "st(0)" to be called just plain "st".
10458 if (StringRef("{st}").equals_lower(Constraint)) {
10459 Res.first = X86::ST0;
10460 Res.second = X86::RFP80RegisterClass;
10465 if (StringRef("{flags}").equals_lower(Constraint)) {
10466 Res.first = X86::EFLAGS;
10467 Res.second = X86::CCRRegisterClass;
10471 // 'A' means EAX + EDX.
10472 if (Constraint == "A") {
10473 Res.first = X86::EAX;
10474 Res.second = X86::GR32_ADRegisterClass;
10480 // Otherwise, check to see if this is a register class of the wrong value
10481 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10482 // turn into {ax},{dx}.
10483 if (Res.second->hasType(VT))
10484 return Res; // Correct type already, nothing to do.
10486 // All of the single-register GCC register classes map their values onto
10487 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10488 // really want an 8-bit or 32-bit register, map to the appropriate register
10489 // class and return the appropriate register.
10490 if (Res.second == X86::GR16RegisterClass) {
10491 if (VT == MVT::i8) {
10492 unsigned DestReg = 0;
10493 switch (Res.first) {
10495 case X86::AX: DestReg = X86::AL; break;
10496 case X86::DX: DestReg = X86::DL; break;
10497 case X86::CX: DestReg = X86::CL; break;
10498 case X86::BX: DestReg = X86::BL; break;
10501 Res.first = DestReg;
10502 Res.second = X86::GR8RegisterClass;
10504 } else if (VT == MVT::i32) {
10505 unsigned DestReg = 0;
10506 switch (Res.first) {
10508 case X86::AX: DestReg = X86::EAX; break;
10509 case X86::DX: DestReg = X86::EDX; break;
10510 case X86::CX: DestReg = X86::ECX; break;
10511 case X86::BX: DestReg = X86::EBX; break;
10512 case X86::SI: DestReg = X86::ESI; break;
10513 case X86::DI: DestReg = X86::EDI; break;
10514 case X86::BP: DestReg = X86::EBP; break;
10515 case X86::SP: DestReg = X86::ESP; break;
10518 Res.first = DestReg;
10519 Res.second = X86::GR32RegisterClass;
10521 } else if (VT == MVT::i64) {
10522 unsigned DestReg = 0;
10523 switch (Res.first) {
10525 case X86::AX: DestReg = X86::RAX; break;
10526 case X86::DX: DestReg = X86::RDX; break;
10527 case X86::CX: DestReg = X86::RCX; break;
10528 case X86::BX: DestReg = X86::RBX; break;
10529 case X86::SI: DestReg = X86::RSI; break;
10530 case X86::DI: DestReg = X86::RDI; break;
10531 case X86::BP: DestReg = X86::RBP; break;
10532 case X86::SP: DestReg = X86::RSP; break;
10535 Res.first = DestReg;
10536 Res.second = X86::GR64RegisterClass;
10539 } else if (Res.second == X86::FR32RegisterClass ||
10540 Res.second == X86::FR64RegisterClass ||
10541 Res.second == X86::VR128RegisterClass) {
10542 // Handle references to XMM physical registers that got mapped into the
10543 // wrong class. This can happen with constraints like {xmm0} where the
10544 // target independent register mapper will just pick the first match it can
10545 // find, ignoring the required type.
10546 if (VT == MVT::f32)
10547 Res.second = X86::FR32RegisterClass;
10548 else if (VT == MVT::f64)
10549 Res.second = X86::FR64RegisterClass;
10550 else if (X86::VR128RegisterClass->hasType(VT))
10551 Res.second = X86::VR128RegisterClass;