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<ISD::OutputArg> &Outs,
1222 LLVMContext &Context) const {
1223 SmallVector<CCValAssign, 16> RVLocs;
1224 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1226 return CCInfo.CheckReturn(Outs, RetCC_X86);
1230 X86TargetLowering::LowerReturn(SDValue Chain,
1231 CallingConv::ID CallConv, bool isVarArg,
1232 const SmallVectorImpl<ISD::OutputArg> &Outs,
1233 const SmallVectorImpl<SDValue> &OutVals,
1234 DebugLoc dl, SelectionDAG &DAG) const {
1235 MachineFunction &MF = DAG.getMachineFunction();
1236 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1238 SmallVector<CCValAssign, 16> RVLocs;
1239 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1240 RVLocs, *DAG.getContext());
1241 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1243 // Add the regs to the liveout set for the function.
1244 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1245 for (unsigned i = 0; i != RVLocs.size(); ++i)
1246 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1247 MRI.addLiveOut(RVLocs[i].getLocReg());
1251 SmallVector<SDValue, 6> RetOps;
1252 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1253 // Operand #1 = Bytes To Pop
1254 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1257 // Copy the result values into the output registers.
1258 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1259 CCValAssign &VA = RVLocs[i];
1260 assert(VA.isRegLoc() && "Can only return in registers!");
1261 SDValue ValToCopy = OutVals[i];
1263 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1264 // the RET instruction and handled by the FP Stackifier.
1265 if (VA.getLocReg() == X86::ST0 ||
1266 VA.getLocReg() == X86::ST1) {
1267 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1268 // change the value to the FP stack register class.
1269 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1270 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1271 RetOps.push_back(ValToCopy);
1272 // Don't emit a copytoreg.
1276 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1277 // which is returned in RAX / RDX.
1278 if (Subtarget->is64Bit()) {
1279 EVT ValVT = ValToCopy.getValueType();
1280 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1281 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1282 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1283 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1287 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1288 Flag = Chain.getValue(1);
1291 // The x86-64 ABI for returning structs by value requires that we copy
1292 // the sret argument into %rax for the return. We saved the argument into
1293 // a virtual register in the entry block, so now we copy the value out
1295 if (Subtarget->is64Bit() &&
1296 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1297 MachineFunction &MF = DAG.getMachineFunction();
1298 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1299 unsigned Reg = FuncInfo->getSRetReturnReg();
1301 "SRetReturnReg should have been set in LowerFormalArguments().");
1302 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1304 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1305 Flag = Chain.getValue(1);
1307 // RAX now acts like a return value.
1308 MRI.addLiveOut(X86::RAX);
1311 RetOps[0] = Chain; // Update chain.
1313 // Add the flag if we have it.
1315 RetOps.push_back(Flag);
1317 return DAG.getNode(X86ISD::RET_FLAG, dl,
1318 MVT::Other, &RetOps[0], RetOps.size());
1321 /// LowerCallResult - Lower the result values of a call into the
1322 /// appropriate copies out of appropriate physical registers.
1325 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1326 CallingConv::ID CallConv, bool isVarArg,
1327 const SmallVectorImpl<ISD::InputArg> &Ins,
1328 DebugLoc dl, SelectionDAG &DAG,
1329 SmallVectorImpl<SDValue> &InVals) const {
1331 // Assign locations to each value returned by this call.
1332 SmallVector<CCValAssign, 16> RVLocs;
1333 bool Is64Bit = Subtarget->is64Bit();
1334 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1335 RVLocs, *DAG.getContext());
1336 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1338 // Copy all of the result registers out of their specified physreg.
1339 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1340 CCValAssign &VA = RVLocs[i];
1341 EVT CopyVT = VA.getValVT();
1343 // If this is x86-64, and we disabled SSE, we can't return FP values
1344 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1345 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1346 report_fatal_error("SSE register return with SSE disabled");
1351 // If this is a call to a function that returns an fp value on the floating
1352 // point stack, we must guarantee the the value is popped from the stack, so
1353 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1354 // if the return value is not used. We use the FpGET_ST0 instructions
1356 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1357 // If we prefer to use the value in xmm registers, copy it out as f80 and
1358 // use a truncate to move it from fp stack reg to xmm reg.
1359 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1360 bool isST0 = VA.getLocReg() == X86::ST0;
1362 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1363 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1364 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1365 SDValue Ops[] = { Chain, InFlag };
1366 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
1368 Val = Chain.getValue(0);
1370 // Round the f80 to the right size, which also moves it to the appropriate
1372 if (CopyVT != VA.getValVT())
1373 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1374 // This truncation won't change the value.
1375 DAG.getIntPtrConstant(1));
1376 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1377 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1378 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1379 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1380 MVT::v2i64, InFlag).getValue(1);
1381 Val = Chain.getValue(0);
1382 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1383 Val, DAG.getConstant(0, MVT::i64));
1385 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1386 MVT::i64, InFlag).getValue(1);
1387 Val = Chain.getValue(0);
1389 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1391 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1392 CopyVT, InFlag).getValue(1);
1393 Val = Chain.getValue(0);
1395 InFlag = Chain.getValue(2);
1396 InVals.push_back(Val);
1403 //===----------------------------------------------------------------------===//
1404 // C & StdCall & Fast Calling Convention implementation
1405 //===----------------------------------------------------------------------===//
1406 // StdCall calling convention seems to be standard for many Windows' API
1407 // routines and around. It differs from C calling convention just a little:
1408 // callee should clean up the stack, not caller. Symbols should be also
1409 // decorated in some fancy way :) It doesn't support any vector arguments.
1410 // For info on fast calling convention see Fast Calling Convention (tail call)
1411 // implementation LowerX86_32FastCCCallTo.
1413 /// CallIsStructReturn - Determines whether a call uses struct return
1415 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1419 return Outs[0].Flags.isSRet();
1422 /// ArgsAreStructReturn - Determines whether a function uses struct
1423 /// return semantics.
1425 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1429 return Ins[0].Flags.isSRet();
1432 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1433 /// given CallingConvention value.
1434 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1435 if (Subtarget->is64Bit()) {
1436 if (CC == CallingConv::GHC)
1437 return CC_X86_64_GHC;
1438 else if (Subtarget->isTargetWin64())
1439 return CC_X86_Win64_C;
1444 if (CC == CallingConv::X86_FastCall)
1445 return CC_X86_32_FastCall;
1446 else if (CC == CallingConv::X86_ThisCall)
1447 return CC_X86_32_ThisCall;
1448 else if (CC == CallingConv::Fast)
1449 return CC_X86_32_FastCC;
1450 else if (CC == CallingConv::GHC)
1451 return CC_X86_32_GHC;
1456 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1457 /// by "Src" to address "Dst" with size and alignment information specified by
1458 /// the specific parameter attribute. The copy will be passed as a byval
1459 /// function parameter.
1461 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1462 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1464 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1465 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1466 /*isVolatile*/false, /*AlwaysInline=*/true,
1470 /// IsTailCallConvention - Return true if the calling convention is one that
1471 /// supports tail call optimization.
1472 static bool IsTailCallConvention(CallingConv::ID CC) {
1473 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1476 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1477 /// a tailcall target by changing its ABI.
1478 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1479 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1483 X86TargetLowering::LowerMemArgument(SDValue Chain,
1484 CallingConv::ID CallConv,
1485 const SmallVectorImpl<ISD::InputArg> &Ins,
1486 DebugLoc dl, SelectionDAG &DAG,
1487 const CCValAssign &VA,
1488 MachineFrameInfo *MFI,
1490 // Create the nodes corresponding to a load from this parameter slot.
1491 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1492 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1493 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1496 // If value is passed by pointer we have address passed instead of the value
1498 if (VA.getLocInfo() == CCValAssign::Indirect)
1499 ValVT = VA.getLocVT();
1501 ValVT = VA.getValVT();
1503 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1504 // changed with more analysis.
1505 // In case of tail call optimization mark all arguments mutable. Since they
1506 // could be overwritten by lowering of arguments in case of a tail call.
1507 if (Flags.isByVal()) {
1508 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1509 VA.getLocMemOffset(), isImmutable);
1510 return DAG.getFrameIndex(FI, getPointerTy());
1512 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1513 VA.getLocMemOffset(), isImmutable);
1514 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1515 return DAG.getLoad(ValVT, dl, Chain, FIN,
1516 PseudoSourceValue::getFixedStack(FI), 0,
1522 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1523 CallingConv::ID CallConv,
1525 const SmallVectorImpl<ISD::InputArg> &Ins,
1528 SmallVectorImpl<SDValue> &InVals)
1530 MachineFunction &MF = DAG.getMachineFunction();
1531 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1533 const Function* Fn = MF.getFunction();
1534 if (Fn->hasExternalLinkage() &&
1535 Subtarget->isTargetCygMing() &&
1536 Fn->getName() == "main")
1537 FuncInfo->setForceFramePointer(true);
1539 MachineFrameInfo *MFI = MF.getFrameInfo();
1540 bool Is64Bit = Subtarget->is64Bit();
1541 bool IsWin64 = Subtarget->isTargetWin64();
1543 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1544 "Var args not supported with calling convention fastcc or ghc");
1546 // Assign locations to all of the incoming arguments.
1547 SmallVector<CCValAssign, 16> ArgLocs;
1548 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1549 ArgLocs, *DAG.getContext());
1550 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1552 unsigned LastVal = ~0U;
1554 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1555 CCValAssign &VA = ArgLocs[i];
1556 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1558 assert(VA.getValNo() != LastVal &&
1559 "Don't support value assigned to multiple locs yet");
1560 LastVal = VA.getValNo();
1562 if (VA.isRegLoc()) {
1563 EVT RegVT = VA.getLocVT();
1564 TargetRegisterClass *RC = NULL;
1565 if (RegVT == MVT::i32)
1566 RC = X86::GR32RegisterClass;
1567 else if (Is64Bit && RegVT == MVT::i64)
1568 RC = X86::GR64RegisterClass;
1569 else if (RegVT == MVT::f32)
1570 RC = X86::FR32RegisterClass;
1571 else if (RegVT == MVT::f64)
1572 RC = X86::FR64RegisterClass;
1573 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1574 RC = X86::VR128RegisterClass;
1575 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1576 RC = X86::VR64RegisterClass;
1578 llvm_unreachable("Unknown argument type!");
1580 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1581 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1583 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1584 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1586 if (VA.getLocInfo() == CCValAssign::SExt)
1587 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1588 DAG.getValueType(VA.getValVT()));
1589 else if (VA.getLocInfo() == CCValAssign::ZExt)
1590 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1591 DAG.getValueType(VA.getValVT()));
1592 else if (VA.getLocInfo() == CCValAssign::BCvt)
1593 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1595 if (VA.isExtInLoc()) {
1596 // Handle MMX values passed in XMM regs.
1597 if (RegVT.isVector()) {
1598 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1599 ArgValue, DAG.getConstant(0, MVT::i64));
1600 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1602 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1605 assert(VA.isMemLoc());
1606 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1609 // If value is passed via pointer - do a load.
1610 if (VA.getLocInfo() == CCValAssign::Indirect)
1611 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1614 InVals.push_back(ArgValue);
1617 // The x86-64 ABI for returning structs by value requires that we copy
1618 // the sret argument into %rax for the return. Save the argument into
1619 // a virtual register so that we can access it from the return points.
1620 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1621 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1622 unsigned Reg = FuncInfo->getSRetReturnReg();
1624 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1625 FuncInfo->setSRetReturnReg(Reg);
1627 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1628 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1631 unsigned StackSize = CCInfo.getNextStackOffset();
1632 // Align stack specially for tail calls.
1633 if (FuncIsMadeTailCallSafe(CallConv))
1634 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1636 // If the function takes variable number of arguments, make a frame index for
1637 // the start of the first vararg value... for expansion of llvm.va_start.
1639 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1640 CallConv != CallingConv::X86_ThisCall)) {
1641 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1644 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1646 // FIXME: We should really autogenerate these arrays
1647 static const unsigned GPR64ArgRegsWin64[] = {
1648 X86::RCX, X86::RDX, X86::R8, X86::R9
1650 static const unsigned XMMArgRegsWin64[] = {
1651 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1653 static const unsigned GPR64ArgRegs64Bit[] = {
1654 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1656 static const unsigned XMMArgRegs64Bit[] = {
1657 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1658 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1660 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1663 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1664 GPR64ArgRegs = GPR64ArgRegsWin64;
1665 XMMArgRegs = XMMArgRegsWin64;
1667 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1668 GPR64ArgRegs = GPR64ArgRegs64Bit;
1669 XMMArgRegs = XMMArgRegs64Bit;
1671 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1673 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1676 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1677 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1678 "SSE register cannot be used when SSE is disabled!");
1679 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1680 "SSE register cannot be used when SSE is disabled!");
1681 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1682 // Kernel mode asks for SSE to be disabled, so don't push them
1684 TotalNumXMMRegs = 0;
1686 // For X86-64, if there are vararg parameters that are passed via
1687 // registers, then we must store them to their spots on the stack so they
1688 // may be loaded by deferencing the result of va_next.
1689 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1690 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1691 FuncInfo->setRegSaveFrameIndex(
1692 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1695 // Store the integer parameter registers.
1696 SmallVector<SDValue, 8> MemOps;
1697 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1699 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1700 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1701 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1702 DAG.getIntPtrConstant(Offset));
1703 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1704 X86::GR64RegisterClass);
1705 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1707 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1708 PseudoSourceValue::getFixedStack(
1709 FuncInfo->getRegSaveFrameIndex()),
1710 Offset, false, false, 0);
1711 MemOps.push_back(Store);
1715 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1716 // Now store the XMM (fp + vector) parameter registers.
1717 SmallVector<SDValue, 11> SaveXMMOps;
1718 SaveXMMOps.push_back(Chain);
1720 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1721 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1722 SaveXMMOps.push_back(ALVal);
1724 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1725 FuncInfo->getRegSaveFrameIndex()));
1726 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1727 FuncInfo->getVarArgsFPOffset()));
1729 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1730 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1731 X86::VR128RegisterClass);
1732 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1733 SaveXMMOps.push_back(Val);
1735 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1737 &SaveXMMOps[0], SaveXMMOps.size()));
1740 if (!MemOps.empty())
1741 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1742 &MemOps[0], MemOps.size());
1746 // Some CCs need callee pop.
1747 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1748 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1750 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1751 // If this is an sret function, the return should pop the hidden pointer.
1752 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1753 FuncInfo->setBytesToPopOnReturn(4);
1757 // RegSaveFrameIndex is X86-64 only.
1758 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1759 if (CallConv == CallingConv::X86_FastCall ||
1760 CallConv == CallingConv::X86_ThisCall)
1761 // fastcc functions can't have varargs.
1762 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1769 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1770 SDValue StackPtr, SDValue Arg,
1771 DebugLoc dl, SelectionDAG &DAG,
1772 const CCValAssign &VA,
1773 ISD::ArgFlagsTy Flags) const {
1774 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1775 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1776 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1777 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1778 if (Flags.isByVal()) {
1779 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1781 return DAG.getStore(Chain, dl, Arg, PtrOff,
1782 PseudoSourceValue::getStack(), LocMemOffset,
1786 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1787 /// optimization is performed and it is required.
1789 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1790 SDValue &OutRetAddr, SDValue Chain,
1791 bool IsTailCall, bool Is64Bit,
1792 int FPDiff, DebugLoc dl) const {
1793 // Adjust the Return address stack slot.
1794 EVT VT = getPointerTy();
1795 OutRetAddr = getReturnAddressFrameIndex(DAG);
1797 // Load the "old" Return address.
1798 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1799 return SDValue(OutRetAddr.getNode(), 1);
1802 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1803 /// optimization is performed and it is required (FPDiff!=0).
1805 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1806 SDValue Chain, SDValue RetAddrFrIdx,
1807 bool Is64Bit, int FPDiff, DebugLoc dl) {
1808 // Store the return address to the appropriate stack slot.
1809 if (!FPDiff) return Chain;
1810 // Calculate the new stack slot for the return address.
1811 int SlotSize = Is64Bit ? 8 : 4;
1812 int NewReturnAddrFI =
1813 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1814 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1815 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1816 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1817 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1823 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1824 CallingConv::ID CallConv, bool isVarArg,
1826 const SmallVectorImpl<ISD::OutputArg> &Outs,
1827 const SmallVectorImpl<SDValue> &OutVals,
1828 const SmallVectorImpl<ISD::InputArg> &Ins,
1829 DebugLoc dl, SelectionDAG &DAG,
1830 SmallVectorImpl<SDValue> &InVals) const {
1831 MachineFunction &MF = DAG.getMachineFunction();
1832 bool Is64Bit = Subtarget->is64Bit();
1833 bool IsStructRet = CallIsStructReturn(Outs);
1834 bool IsSibcall = false;
1837 // Check if it's really possible to do a tail call.
1838 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1839 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1840 Outs, OutVals, Ins, DAG);
1842 // Sibcalls are automatically detected tailcalls which do not require
1844 if (!GuaranteedTailCallOpt && isTailCall)
1851 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1852 "Var args not supported with calling convention fastcc or ghc");
1854 // Analyze operands of the call, assigning locations to each operand.
1855 SmallVector<CCValAssign, 16> ArgLocs;
1856 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1857 ArgLocs, *DAG.getContext());
1858 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1860 // Get a count of how many bytes are to be pushed on the stack.
1861 unsigned NumBytes = CCInfo.getNextStackOffset();
1863 // This is a sibcall. The memory operands are available in caller's
1864 // own caller's stack.
1866 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1867 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1870 if (isTailCall && !IsSibcall) {
1871 // Lower arguments at fp - stackoffset + fpdiff.
1872 unsigned NumBytesCallerPushed =
1873 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1874 FPDiff = NumBytesCallerPushed - NumBytes;
1876 // Set the delta of movement of the returnaddr stackslot.
1877 // But only set if delta is greater than previous delta.
1878 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1879 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1883 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1885 SDValue RetAddrFrIdx;
1886 // Load return adress for tail calls.
1887 if (isTailCall && FPDiff)
1888 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1889 Is64Bit, FPDiff, dl);
1891 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1892 SmallVector<SDValue, 8> MemOpChains;
1895 // Walk the register/memloc assignments, inserting copies/loads. In the case
1896 // of tail call optimization arguments are handle later.
1897 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1898 CCValAssign &VA = ArgLocs[i];
1899 EVT RegVT = VA.getLocVT();
1900 SDValue Arg = OutVals[i];
1901 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1902 bool isByVal = Flags.isByVal();
1904 // Promote the value if needed.
1905 switch (VA.getLocInfo()) {
1906 default: llvm_unreachable("Unknown loc info!");
1907 case CCValAssign::Full: break;
1908 case CCValAssign::SExt:
1909 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1911 case CCValAssign::ZExt:
1912 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1914 case CCValAssign::AExt:
1915 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1916 // Special case: passing MMX values in XMM registers.
1917 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1918 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1919 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1921 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1923 case CCValAssign::BCvt:
1924 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1926 case CCValAssign::Indirect: {
1927 // Store the argument.
1928 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1929 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1930 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1931 PseudoSourceValue::getFixedStack(FI), 0,
1938 if (VA.isRegLoc()) {
1939 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1940 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1941 assert(VA.isMemLoc());
1942 if (StackPtr.getNode() == 0)
1943 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1944 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1945 dl, DAG, VA, Flags));
1949 if (!MemOpChains.empty())
1950 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1951 &MemOpChains[0], MemOpChains.size());
1953 // Build a sequence of copy-to-reg nodes chained together with token chain
1954 // and flag operands which copy the outgoing args into registers.
1956 // Tail call byval lowering might overwrite argument registers so in case of
1957 // tail call optimization the copies to registers are lowered later.
1959 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1960 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1961 RegsToPass[i].second, InFlag);
1962 InFlag = Chain.getValue(1);
1965 if (Subtarget->isPICStyleGOT()) {
1966 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1969 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1970 DAG.getNode(X86ISD::GlobalBaseReg,
1971 DebugLoc(), getPointerTy()),
1973 InFlag = Chain.getValue(1);
1975 // If we are tail calling and generating PIC/GOT style code load the
1976 // address of the callee into ECX. The value in ecx is used as target of
1977 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1978 // for tail calls on PIC/GOT architectures. Normally we would just put the
1979 // address of GOT into ebx and then call target@PLT. But for tail calls
1980 // ebx would be restored (since ebx is callee saved) before jumping to the
1983 // Note: The actual moving to ECX is done further down.
1984 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1985 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1986 !G->getGlobal()->hasProtectedVisibility())
1987 Callee = LowerGlobalAddress(Callee, DAG);
1988 else if (isa<ExternalSymbolSDNode>(Callee))
1989 Callee = LowerExternalSymbol(Callee, DAG);
1993 if (Is64Bit && isVarArg) {
1994 // From AMD64 ABI document:
1995 // For calls that may call functions that use varargs or stdargs
1996 // (prototype-less calls or calls to functions containing ellipsis (...) in
1997 // the declaration) %al is used as hidden argument to specify the number
1998 // of SSE registers used. The contents of %al do not need to match exactly
1999 // the number of registers, but must be an ubound on the number of SSE
2000 // registers used and is in the range 0 - 8 inclusive.
2002 // FIXME: Verify this on Win64
2003 // Count the number of XMM registers allocated.
2004 static const unsigned XMMArgRegs[] = {
2005 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2006 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2008 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2009 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2010 && "SSE registers cannot be used when SSE is disabled");
2012 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2013 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2014 InFlag = Chain.getValue(1);
2018 // For tail calls lower the arguments to the 'real' stack slot.
2020 // Force all the incoming stack arguments to be loaded from the stack
2021 // before any new outgoing arguments are stored to the stack, because the
2022 // outgoing stack slots may alias the incoming argument stack slots, and
2023 // the alias isn't otherwise explicit. This is slightly more conservative
2024 // than necessary, because it means that each store effectively depends
2025 // on every argument instead of just those arguments it would clobber.
2026 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2028 SmallVector<SDValue, 8> MemOpChains2;
2031 // Do not flag preceeding copytoreg stuff together with the following stuff.
2033 if (GuaranteedTailCallOpt) {
2034 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2035 CCValAssign &VA = ArgLocs[i];
2038 assert(VA.isMemLoc());
2039 SDValue Arg = OutVals[i];
2040 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2041 // Create frame index.
2042 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2043 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2044 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2045 FIN = DAG.getFrameIndex(FI, getPointerTy());
2047 if (Flags.isByVal()) {
2048 // Copy relative to framepointer.
2049 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2050 if (StackPtr.getNode() == 0)
2051 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2053 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2055 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2059 // Store relative to framepointer.
2060 MemOpChains2.push_back(
2061 DAG.getStore(ArgChain, dl, Arg, FIN,
2062 PseudoSourceValue::getFixedStack(FI), 0,
2068 if (!MemOpChains2.empty())
2069 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2070 &MemOpChains2[0], MemOpChains2.size());
2072 // Copy arguments to their registers.
2073 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2074 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2075 RegsToPass[i].second, InFlag);
2076 InFlag = Chain.getValue(1);
2080 // Store the return address to the appropriate stack slot.
2081 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2085 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2086 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2087 // In the 64-bit large code model, we have to make all calls
2088 // through a register, since the call instruction's 32-bit
2089 // pc-relative offset may not be large enough to hold the whole
2091 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2092 // If the callee is a GlobalAddress node (quite common, every direct call
2093 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2096 // We should use extra load for direct calls to dllimported functions in
2098 const GlobalValue *GV = G->getGlobal();
2099 if (!GV->hasDLLImportLinkage()) {
2100 unsigned char OpFlags = 0;
2102 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2103 // external symbols most go through the PLT in PIC mode. If the symbol
2104 // has hidden or protected visibility, or if it is static or local, then
2105 // we don't need to use the PLT - we can directly call it.
2106 if (Subtarget->isTargetELF() &&
2107 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2108 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2109 OpFlags = X86II::MO_PLT;
2110 } else if (Subtarget->isPICStyleStubAny() &&
2111 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2112 Subtarget->getDarwinVers() < 9) {
2113 // PC-relative references to external symbols should go through $stub,
2114 // unless we're building with the leopard linker or later, which
2115 // automatically synthesizes these stubs.
2116 OpFlags = X86II::MO_DARWIN_STUB;
2119 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2120 G->getOffset(), OpFlags);
2122 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2123 unsigned char OpFlags = 0;
2125 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2126 // symbols should go through the PLT.
2127 if (Subtarget->isTargetELF() &&
2128 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2129 OpFlags = X86II::MO_PLT;
2130 } else if (Subtarget->isPICStyleStubAny() &&
2131 Subtarget->getDarwinVers() < 9) {
2132 // PC-relative references to external symbols should go through $stub,
2133 // unless we're building with the leopard linker or later, which
2134 // automatically synthesizes these stubs.
2135 OpFlags = X86II::MO_DARWIN_STUB;
2138 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2142 // Returns a chain & a flag for retval copy to use.
2143 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2144 SmallVector<SDValue, 8> Ops;
2146 if (!IsSibcall && isTailCall) {
2147 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2148 DAG.getIntPtrConstant(0, true), InFlag);
2149 InFlag = Chain.getValue(1);
2152 Ops.push_back(Chain);
2153 Ops.push_back(Callee);
2156 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2158 // Add argument registers to the end of the list so that they are known live
2160 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2161 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2162 RegsToPass[i].second.getValueType()));
2164 // Add an implicit use GOT pointer in EBX.
2165 if (!isTailCall && Subtarget->isPICStyleGOT())
2166 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2168 // Add an implicit use of AL for x86 vararg functions.
2169 if (Is64Bit && isVarArg)
2170 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2172 if (InFlag.getNode())
2173 Ops.push_back(InFlag);
2177 //// If this is the first return lowered for this function, add the regs
2178 //// to the liveout set for the function.
2179 // This isn't right, although it's probably harmless on x86; liveouts
2180 // should be computed from returns not tail calls. Consider a void
2181 // function making a tail call to a function returning int.
2182 return DAG.getNode(X86ISD::TC_RETURN, dl,
2183 NodeTys, &Ops[0], Ops.size());
2186 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2187 InFlag = Chain.getValue(1);
2189 // Create the CALLSEQ_END node.
2190 unsigned NumBytesForCalleeToPush;
2191 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2192 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2193 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2194 // If this is a call to a struct-return function, the callee
2195 // pops the hidden struct pointer, so we have to push it back.
2196 // This is common for Darwin/X86, Linux & Mingw32 targets.
2197 NumBytesForCalleeToPush = 4;
2199 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2201 // Returns a flag for retval copy to use.
2203 Chain = DAG.getCALLSEQ_END(Chain,
2204 DAG.getIntPtrConstant(NumBytes, true),
2205 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2208 InFlag = Chain.getValue(1);
2211 // Handle result values, copying them out of physregs into vregs that we
2213 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2214 Ins, dl, DAG, InVals);
2218 //===----------------------------------------------------------------------===//
2219 // Fast Calling Convention (tail call) implementation
2220 //===----------------------------------------------------------------------===//
2222 // Like std call, callee cleans arguments, convention except that ECX is
2223 // reserved for storing the tail called function address. Only 2 registers are
2224 // free for argument passing (inreg). Tail call optimization is performed
2226 // * tailcallopt is enabled
2227 // * caller/callee are fastcc
2228 // On X86_64 architecture with GOT-style position independent code only local
2229 // (within module) calls are supported at the moment.
2230 // To keep the stack aligned according to platform abi the function
2231 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2232 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2233 // If a tail called function callee has more arguments than the caller the
2234 // caller needs to make sure that there is room to move the RETADDR to. This is
2235 // achieved by reserving an area the size of the argument delta right after the
2236 // original REtADDR, but before the saved framepointer or the spilled registers
2237 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2249 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2250 /// for a 16 byte align requirement.
2252 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2253 SelectionDAG& DAG) const {
2254 MachineFunction &MF = DAG.getMachineFunction();
2255 const TargetMachine &TM = MF.getTarget();
2256 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2257 unsigned StackAlignment = TFI.getStackAlignment();
2258 uint64_t AlignMask = StackAlignment - 1;
2259 int64_t Offset = StackSize;
2260 uint64_t SlotSize = TD->getPointerSize();
2261 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2262 // Number smaller than 12 so just add the difference.
2263 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2265 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2266 Offset = ((~AlignMask) & Offset) + StackAlignment +
2267 (StackAlignment-SlotSize);
2272 /// MatchingStackOffset - Return true if the given stack call argument is
2273 /// already available in the same position (relatively) of the caller's
2274 /// incoming argument stack.
2276 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2277 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2278 const X86InstrInfo *TII) {
2279 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2281 if (Arg.getOpcode() == ISD::CopyFromReg) {
2282 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2283 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2285 MachineInstr *Def = MRI->getVRegDef(VR);
2288 if (!Flags.isByVal()) {
2289 if (!TII->isLoadFromStackSlot(Def, FI))
2292 unsigned Opcode = Def->getOpcode();
2293 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2294 Def->getOperand(1).isFI()) {
2295 FI = Def->getOperand(1).getIndex();
2296 Bytes = Flags.getByValSize();
2300 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2301 if (Flags.isByVal())
2302 // ByVal argument is passed in as a pointer but it's now being
2303 // dereferenced. e.g.
2304 // define @foo(%struct.X* %A) {
2305 // tail call @bar(%struct.X* byval %A)
2308 SDValue Ptr = Ld->getBasePtr();
2309 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2312 FI = FINode->getIndex();
2316 assert(FI != INT_MAX);
2317 if (!MFI->isFixedObjectIndex(FI))
2319 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2322 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2323 /// for tail call optimization. Targets which want to do tail call
2324 /// optimization should implement this function.
2326 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2327 CallingConv::ID CalleeCC,
2329 bool isCalleeStructRet,
2330 bool isCallerStructRet,
2331 const SmallVectorImpl<ISD::OutputArg> &Outs,
2332 const SmallVectorImpl<SDValue> &OutVals,
2333 const SmallVectorImpl<ISD::InputArg> &Ins,
2334 SelectionDAG& DAG) const {
2335 if (!IsTailCallConvention(CalleeCC) &&
2336 CalleeCC != CallingConv::C)
2339 // If -tailcallopt is specified, make fastcc functions tail-callable.
2340 const MachineFunction &MF = DAG.getMachineFunction();
2341 const Function *CallerF = DAG.getMachineFunction().getFunction();
2342 CallingConv::ID CallerCC = CallerF->getCallingConv();
2343 bool CCMatch = CallerCC == CalleeCC;
2345 if (GuaranteedTailCallOpt) {
2346 if (IsTailCallConvention(CalleeCC) && CCMatch)
2351 // Look for obvious safe cases to perform tail call optimization that do not
2352 // require ABI changes. This is what gcc calls sibcall.
2354 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2355 // emit a special epilogue.
2356 if (RegInfo->needsStackRealignment(MF))
2359 // Do not sibcall optimize vararg calls unless the call site is not passing any
2361 if (isVarArg && !Outs.empty())
2364 // Also avoid sibcall optimization if either caller or callee uses struct
2365 // return semantics.
2366 if (isCalleeStructRet || isCallerStructRet)
2369 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2370 // Therefore if it's not used by the call it is not safe to optimize this into
2372 bool Unused = false;
2373 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2380 SmallVector<CCValAssign, 16> RVLocs;
2381 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2382 RVLocs, *DAG.getContext());
2383 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2384 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2385 CCValAssign &VA = RVLocs[i];
2386 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2391 // If the calling conventions do not match, then we'd better make sure the
2392 // results are returned in the same way as what the caller expects.
2394 SmallVector<CCValAssign, 16> RVLocs1;
2395 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2396 RVLocs1, *DAG.getContext());
2397 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2399 SmallVector<CCValAssign, 16> RVLocs2;
2400 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2401 RVLocs2, *DAG.getContext());
2402 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2404 if (RVLocs1.size() != RVLocs2.size())
2406 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2407 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2409 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2411 if (RVLocs1[i].isRegLoc()) {
2412 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2415 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2421 // If the callee takes no arguments then go on to check the results of the
2423 if (!Outs.empty()) {
2424 // Check if stack adjustment is needed. For now, do not do this if any
2425 // argument is passed on the stack.
2426 SmallVector<CCValAssign, 16> ArgLocs;
2427 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2428 ArgLocs, *DAG.getContext());
2429 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2430 if (CCInfo.getNextStackOffset()) {
2431 MachineFunction &MF = DAG.getMachineFunction();
2432 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2434 if (Subtarget->isTargetWin64())
2435 // Win64 ABI has additional complications.
2438 // Check if the arguments are already laid out in the right way as
2439 // the caller's fixed stack objects.
2440 MachineFrameInfo *MFI = MF.getFrameInfo();
2441 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2442 const X86InstrInfo *TII =
2443 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2444 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2445 CCValAssign &VA = ArgLocs[i];
2446 SDValue Arg = OutVals[i];
2447 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2448 if (VA.getLocInfo() == CCValAssign::Indirect)
2450 if (!VA.isRegLoc()) {
2451 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2458 // If the tailcall address may be in a register, then make sure it's
2459 // possible to register allocate for it. In 32-bit, the call address can
2460 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2461 // callee-saved registers are restored. These happen to be the same
2462 // registers used to pass 'inreg' arguments so watch out for those.
2463 if (!Subtarget->is64Bit() &&
2464 !isa<GlobalAddressSDNode>(Callee) &&
2465 !isa<ExternalSymbolSDNode>(Callee)) {
2466 unsigned NumInRegs = 0;
2467 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2468 CCValAssign &VA = ArgLocs[i];
2471 unsigned Reg = VA.getLocReg();
2474 case X86::EAX: case X86::EDX: case X86::ECX:
2475 if (++NumInRegs == 3)
2487 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2488 return X86::createFastISel(funcInfo);
2492 //===----------------------------------------------------------------------===//
2493 // Other Lowering Hooks
2494 //===----------------------------------------------------------------------===//
2497 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2498 MachineFunction &MF = DAG.getMachineFunction();
2499 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2500 int ReturnAddrIndex = FuncInfo->getRAIndex();
2502 if (ReturnAddrIndex == 0) {
2503 // Set up a frame object for the return address.
2504 uint64_t SlotSize = TD->getPointerSize();
2505 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2507 FuncInfo->setRAIndex(ReturnAddrIndex);
2510 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2514 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2515 bool hasSymbolicDisplacement) {
2516 // Offset should fit into 32 bit immediate field.
2517 if (!isInt<32>(Offset))
2520 // If we don't have a symbolic displacement - we don't have any extra
2522 if (!hasSymbolicDisplacement)
2525 // FIXME: Some tweaks might be needed for medium code model.
2526 if (M != CodeModel::Small && M != CodeModel::Kernel)
2529 // For small code model we assume that latest object is 16MB before end of 31
2530 // bits boundary. We may also accept pretty large negative constants knowing
2531 // that all objects are in the positive half of address space.
2532 if (M == CodeModel::Small && Offset < 16*1024*1024)
2535 // For kernel code model we know that all object resist in the negative half
2536 // of 32bits address space. We may not accept negative offsets, since they may
2537 // be just off and we may accept pretty large positive ones.
2538 if (M == CodeModel::Kernel && Offset > 0)
2544 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2545 /// specific condition code, returning the condition code and the LHS/RHS of the
2546 /// comparison to make.
2547 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2548 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2550 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2551 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2552 // X > -1 -> X == 0, jump !sign.
2553 RHS = DAG.getConstant(0, RHS.getValueType());
2554 return X86::COND_NS;
2555 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2556 // X < 0 -> X == 0, jump on sign.
2558 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2560 RHS = DAG.getConstant(0, RHS.getValueType());
2561 return X86::COND_LE;
2565 switch (SetCCOpcode) {
2566 default: llvm_unreachable("Invalid integer condition!");
2567 case ISD::SETEQ: return X86::COND_E;
2568 case ISD::SETGT: return X86::COND_G;
2569 case ISD::SETGE: return X86::COND_GE;
2570 case ISD::SETLT: return X86::COND_L;
2571 case ISD::SETLE: return X86::COND_LE;
2572 case ISD::SETNE: return X86::COND_NE;
2573 case ISD::SETULT: return X86::COND_B;
2574 case ISD::SETUGT: return X86::COND_A;
2575 case ISD::SETULE: return X86::COND_BE;
2576 case ISD::SETUGE: return X86::COND_AE;
2580 // First determine if it is required or is profitable to flip the operands.
2582 // If LHS is a foldable load, but RHS is not, flip the condition.
2583 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2584 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2585 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2586 std::swap(LHS, RHS);
2589 switch (SetCCOpcode) {
2595 std::swap(LHS, RHS);
2599 // On a floating point condition, the flags are set as follows:
2601 // 0 | 0 | 0 | X > Y
2602 // 0 | 0 | 1 | X < Y
2603 // 1 | 0 | 0 | X == Y
2604 // 1 | 1 | 1 | unordered
2605 switch (SetCCOpcode) {
2606 default: llvm_unreachable("Condcode should be pre-legalized away");
2608 case ISD::SETEQ: return X86::COND_E;
2609 case ISD::SETOLT: // flipped
2611 case ISD::SETGT: return X86::COND_A;
2612 case ISD::SETOLE: // flipped
2614 case ISD::SETGE: return X86::COND_AE;
2615 case ISD::SETUGT: // flipped
2617 case ISD::SETLT: return X86::COND_B;
2618 case ISD::SETUGE: // flipped
2620 case ISD::SETLE: return X86::COND_BE;
2622 case ISD::SETNE: return X86::COND_NE;
2623 case ISD::SETUO: return X86::COND_P;
2624 case ISD::SETO: return X86::COND_NP;
2626 case ISD::SETUNE: return X86::COND_INVALID;
2630 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2631 /// code. Current x86 isa includes the following FP cmov instructions:
2632 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2633 static bool hasFPCMov(unsigned X86CC) {
2649 /// isFPImmLegal - Returns true if the target can instruction select the
2650 /// specified FP immediate natively. If false, the legalizer will
2651 /// materialize the FP immediate as a load from a constant pool.
2652 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2653 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2654 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2660 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2661 /// the specified range (L, H].
2662 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2663 return (Val < 0) || (Val >= Low && Val < Hi);
2666 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2667 /// specified value.
2668 static bool isUndefOrEqual(int Val, int CmpVal) {
2669 if (Val < 0 || Val == CmpVal)
2674 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2675 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2676 /// the second operand.
2677 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2678 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2679 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2680 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2681 return (Mask[0] < 2 && Mask[1] < 2);
2685 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2686 SmallVector<int, 8> M;
2688 return ::isPSHUFDMask(M, N->getValueType(0));
2691 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2692 /// is suitable for input to PSHUFHW.
2693 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2694 if (VT != MVT::v8i16)
2697 // Lower quadword copied in order or undef.
2698 for (int i = 0; i != 4; ++i)
2699 if (Mask[i] >= 0 && Mask[i] != i)
2702 // Upper quadword shuffled.
2703 for (int i = 4; i != 8; ++i)
2704 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2710 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2711 SmallVector<int, 8> M;
2713 return ::isPSHUFHWMask(M, N->getValueType(0));
2716 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2717 /// is suitable for input to PSHUFLW.
2718 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2719 if (VT != MVT::v8i16)
2722 // Upper quadword copied in order.
2723 for (int i = 4; i != 8; ++i)
2724 if (Mask[i] >= 0 && Mask[i] != i)
2727 // Lower quadword shuffled.
2728 for (int i = 0; i != 4; ++i)
2735 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2736 SmallVector<int, 8> M;
2738 return ::isPSHUFLWMask(M, N->getValueType(0));
2741 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2742 /// is suitable for input to PALIGNR.
2743 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2745 int i, e = VT.getVectorNumElements();
2747 // Do not handle v2i64 / v2f64 shuffles with palignr.
2748 if (e < 4 || !hasSSSE3)
2751 for (i = 0; i != e; ++i)
2755 // All undef, not a palignr.
2759 // Determine if it's ok to perform a palignr with only the LHS, since we
2760 // don't have access to the actual shuffle elements to see if RHS is undef.
2761 bool Unary = Mask[i] < (int)e;
2762 bool NeedsUnary = false;
2764 int s = Mask[i] - i;
2766 // Check the rest of the elements to see if they are consecutive.
2767 for (++i; i != e; ++i) {
2772 Unary = Unary && (m < (int)e);
2773 NeedsUnary = NeedsUnary || (m < s);
2775 if (NeedsUnary && !Unary)
2777 if (Unary && m != ((s+i) & (e-1)))
2779 if (!Unary && m != (s+i))
2785 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2786 SmallVector<int, 8> M;
2788 return ::isPALIGNRMask(M, N->getValueType(0), true);
2791 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2792 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2793 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2794 int NumElems = VT.getVectorNumElements();
2795 if (NumElems != 2 && NumElems != 4)
2798 int Half = NumElems / 2;
2799 for (int i = 0; i < Half; ++i)
2800 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2802 for (int i = Half; i < NumElems; ++i)
2803 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2809 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2810 SmallVector<int, 8> M;
2812 return ::isSHUFPMask(M, N->getValueType(0));
2815 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2816 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2817 /// half elements to come from vector 1 (which would equal the dest.) and
2818 /// the upper half to come from vector 2.
2819 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2820 int NumElems = VT.getVectorNumElements();
2822 if (NumElems != 2 && NumElems != 4)
2825 int Half = NumElems / 2;
2826 for (int i = 0; i < Half; ++i)
2827 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2829 for (int i = Half; i < NumElems; ++i)
2830 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2835 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2836 SmallVector<int, 8> M;
2838 return isCommutedSHUFPMask(M, N->getValueType(0));
2841 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2842 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2843 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2844 if (N->getValueType(0).getVectorNumElements() != 4)
2847 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2848 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2849 isUndefOrEqual(N->getMaskElt(1), 7) &&
2850 isUndefOrEqual(N->getMaskElt(2), 2) &&
2851 isUndefOrEqual(N->getMaskElt(3), 3);
2854 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2855 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2857 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2858 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2863 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2864 isUndefOrEqual(N->getMaskElt(1), 3) &&
2865 isUndefOrEqual(N->getMaskElt(2), 2) &&
2866 isUndefOrEqual(N->getMaskElt(3), 3);
2869 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2870 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2871 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2872 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2874 if (NumElems != 2 && NumElems != 4)
2877 for (unsigned i = 0; i < NumElems/2; ++i)
2878 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2881 for (unsigned i = NumElems/2; i < NumElems; ++i)
2882 if (!isUndefOrEqual(N->getMaskElt(i), i))
2888 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2889 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2890 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2891 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2893 if (NumElems != 2 && NumElems != 4)
2896 for (unsigned i = 0; i < NumElems/2; ++i)
2897 if (!isUndefOrEqual(N->getMaskElt(i), i))
2900 for (unsigned i = 0; i < NumElems/2; ++i)
2901 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2907 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2908 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2909 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2910 bool V2IsSplat = false) {
2911 int NumElts = VT.getVectorNumElements();
2912 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2915 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2917 int BitI1 = Mask[i+1];
2918 if (!isUndefOrEqual(BitI, j))
2921 if (!isUndefOrEqual(BitI1, NumElts))
2924 if (!isUndefOrEqual(BitI1, j + NumElts))
2931 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2932 SmallVector<int, 8> M;
2934 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2937 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2938 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2939 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2940 bool V2IsSplat = false) {
2941 int NumElts = VT.getVectorNumElements();
2942 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2945 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2947 int BitI1 = Mask[i+1];
2948 if (!isUndefOrEqual(BitI, j + NumElts/2))
2951 if (isUndefOrEqual(BitI1, NumElts))
2954 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2961 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2962 SmallVector<int, 8> M;
2964 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2967 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2968 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2970 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2971 int NumElems = VT.getVectorNumElements();
2972 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2975 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2977 int BitI1 = Mask[i+1];
2978 if (!isUndefOrEqual(BitI, j))
2980 if (!isUndefOrEqual(BitI1, j))
2986 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2987 SmallVector<int, 8> M;
2989 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2992 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2993 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2995 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2996 int NumElems = VT.getVectorNumElements();
2997 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3000 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3002 int BitI1 = Mask[i+1];
3003 if (!isUndefOrEqual(BitI, j))
3005 if (!isUndefOrEqual(BitI1, j))
3011 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3012 SmallVector<int, 8> M;
3014 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3017 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3018 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3019 /// MOVSD, and MOVD, i.e. setting the lowest element.
3020 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3021 if (VT.getVectorElementType().getSizeInBits() < 32)
3024 int NumElts = VT.getVectorNumElements();
3026 if (!isUndefOrEqual(Mask[0], NumElts))
3029 for (int i = 1; i < NumElts; ++i)
3030 if (!isUndefOrEqual(Mask[i], i))
3036 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3037 SmallVector<int, 8> M;
3039 return ::isMOVLMask(M, N->getValueType(0));
3042 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3043 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3044 /// element of vector 2 and the other elements to come from vector 1 in order.
3045 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3046 bool V2IsSplat = false, bool V2IsUndef = false) {
3047 int NumOps = VT.getVectorNumElements();
3048 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3051 if (!isUndefOrEqual(Mask[0], 0))
3054 for (int i = 1; i < NumOps; ++i)
3055 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3056 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3057 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3063 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3064 bool V2IsUndef = false) {
3065 SmallVector<int, 8> M;
3067 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3070 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3071 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3072 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3073 if (N->getValueType(0).getVectorNumElements() != 4)
3076 // Expect 1, 1, 3, 3
3077 for (unsigned i = 0; i < 2; ++i) {
3078 int Elt = N->getMaskElt(i);
3079 if (Elt >= 0 && Elt != 1)
3084 for (unsigned i = 2; i < 4; ++i) {
3085 int Elt = N->getMaskElt(i);
3086 if (Elt >= 0 && Elt != 3)
3091 // Don't use movshdup if it can be done with a shufps.
3092 // FIXME: verify that matching u, u, 3, 3 is what we want.
3096 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3097 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3098 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3099 if (N->getValueType(0).getVectorNumElements() != 4)
3102 // Expect 0, 0, 2, 2
3103 for (unsigned i = 0; i < 2; ++i)
3104 if (N->getMaskElt(i) > 0)
3108 for (unsigned i = 2; i < 4; ++i) {
3109 int Elt = N->getMaskElt(i);
3110 if (Elt >= 0 && Elt != 2)
3115 // Don't use movsldup if it can be done with a shufps.
3119 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3120 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3121 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3122 int e = N->getValueType(0).getVectorNumElements() / 2;
3124 for (int i = 0; i < e; ++i)
3125 if (!isUndefOrEqual(N->getMaskElt(i), i))
3127 for (int i = 0; i < e; ++i)
3128 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3133 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3134 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3135 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3136 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3137 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3139 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3141 for (int i = 0; i < NumOperands; ++i) {
3142 int Val = SVOp->getMaskElt(NumOperands-i-1);
3143 if (Val < 0) Val = 0;
3144 if (Val >= NumOperands) Val -= NumOperands;
3146 if (i != NumOperands - 1)
3152 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3153 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3154 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3155 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3157 // 8 nodes, but we only care about the last 4.
3158 for (unsigned i = 7; i >= 4; --i) {
3159 int Val = SVOp->getMaskElt(i);
3168 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3169 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3170 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3171 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3173 // 8 nodes, but we only care about the first 4.
3174 for (int i = 3; i >= 0; --i) {
3175 int Val = SVOp->getMaskElt(i);
3184 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3185 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3186 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3187 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3188 EVT VVT = N->getValueType(0);
3189 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3193 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3194 Val = SVOp->getMaskElt(i);
3198 return (Val - i) * EltSize;
3201 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3203 bool X86::isZeroNode(SDValue Elt) {
3204 return ((isa<ConstantSDNode>(Elt) &&
3205 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3206 (isa<ConstantFPSDNode>(Elt) &&
3207 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3210 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3211 /// their permute mask.
3212 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3213 SelectionDAG &DAG) {
3214 EVT VT = SVOp->getValueType(0);
3215 unsigned NumElems = VT.getVectorNumElements();
3216 SmallVector<int, 8> MaskVec;
3218 for (unsigned i = 0; i != NumElems; ++i) {
3219 int idx = SVOp->getMaskElt(i);
3221 MaskVec.push_back(idx);
3222 else if (idx < (int)NumElems)
3223 MaskVec.push_back(idx + NumElems);
3225 MaskVec.push_back(idx - NumElems);
3227 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3228 SVOp->getOperand(0), &MaskVec[0]);
3231 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3232 /// the two vector operands have swapped position.
3233 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3234 unsigned NumElems = VT.getVectorNumElements();
3235 for (unsigned i = 0; i != NumElems; ++i) {
3239 else if (idx < (int)NumElems)
3240 Mask[i] = idx + NumElems;
3242 Mask[i] = idx - NumElems;
3246 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3247 /// match movhlps. The lower half elements should come from upper half of
3248 /// V1 (and in order), and the upper half elements should come from the upper
3249 /// half of V2 (and in order).
3250 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3251 if (Op->getValueType(0).getVectorNumElements() != 4)
3253 for (unsigned i = 0, e = 2; i != e; ++i)
3254 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3256 for (unsigned i = 2; i != 4; ++i)
3257 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3262 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3263 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3265 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3266 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3268 N = N->getOperand(0).getNode();
3269 if (!ISD::isNON_EXTLoad(N))
3272 *LD = cast<LoadSDNode>(N);
3276 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3277 /// match movlp{s|d}. The lower half elements should come from lower half of
3278 /// V1 (and in order), and the upper half elements should come from the upper
3279 /// half of V2 (and in order). And since V1 will become the source of the
3280 /// MOVLP, it must be either a vector load or a scalar load to vector.
3281 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3282 ShuffleVectorSDNode *Op) {
3283 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3285 // Is V2 is a vector load, don't do this transformation. We will try to use
3286 // load folding shufps op.
3287 if (ISD::isNON_EXTLoad(V2))
3290 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3292 if (NumElems != 2 && NumElems != 4)
3294 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3295 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3297 for (unsigned i = NumElems/2; i != NumElems; ++i)
3298 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3303 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3305 static bool isSplatVector(SDNode *N) {
3306 if (N->getOpcode() != ISD::BUILD_VECTOR)
3309 SDValue SplatValue = N->getOperand(0);
3310 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3311 if (N->getOperand(i) != SplatValue)
3316 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3317 /// to an zero vector.
3318 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3319 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3320 SDValue V1 = N->getOperand(0);
3321 SDValue V2 = N->getOperand(1);
3322 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3323 for (unsigned i = 0; i != NumElems; ++i) {
3324 int Idx = N->getMaskElt(i);
3325 if (Idx >= (int)NumElems) {
3326 unsigned Opc = V2.getOpcode();
3327 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3329 if (Opc != ISD::BUILD_VECTOR ||
3330 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3332 } else if (Idx >= 0) {
3333 unsigned Opc = V1.getOpcode();
3334 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3336 if (Opc != ISD::BUILD_VECTOR ||
3337 !X86::isZeroNode(V1.getOperand(Idx)))
3344 /// getZeroVector - Returns a vector of specified type with all zero elements.
3346 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3348 assert(VT.isVector() && "Expected a vector type");
3350 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3351 // type. This ensures they get CSE'd.
3353 if (VT.getSizeInBits() == 64) { // MMX
3354 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3355 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3356 } else if (HasSSE2) { // SSE2
3357 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3358 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3360 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3361 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3363 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3366 /// getOnesVector - Returns a vector of specified type with all bits set.
3368 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3369 assert(VT.isVector() && "Expected a vector type");
3371 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3372 // type. This ensures they get CSE'd.
3373 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3375 if (VT.getSizeInBits() == 64) // MMX
3376 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3378 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3379 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3383 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3384 /// that point to V2 points to its first element.
3385 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3386 EVT VT = SVOp->getValueType(0);
3387 unsigned NumElems = VT.getVectorNumElements();
3389 bool Changed = false;
3390 SmallVector<int, 8> MaskVec;
3391 SVOp->getMask(MaskVec);
3393 for (unsigned i = 0; i != NumElems; ++i) {
3394 if (MaskVec[i] > (int)NumElems) {
3395 MaskVec[i] = NumElems;
3400 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3401 SVOp->getOperand(1), &MaskVec[0]);
3402 return SDValue(SVOp, 0);
3405 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3406 /// operation of specified width.
3407 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3409 unsigned NumElems = VT.getVectorNumElements();
3410 SmallVector<int, 8> Mask;
3411 Mask.push_back(NumElems);
3412 for (unsigned i = 1; i != NumElems; ++i)
3414 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3417 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3418 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3420 unsigned NumElems = VT.getVectorNumElements();
3421 SmallVector<int, 8> Mask;
3422 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3424 Mask.push_back(i + NumElems);
3426 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3429 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3430 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3432 unsigned NumElems = VT.getVectorNumElements();
3433 unsigned Half = NumElems/2;
3434 SmallVector<int, 8> Mask;
3435 for (unsigned i = 0; i != Half; ++i) {
3436 Mask.push_back(i + Half);
3437 Mask.push_back(i + NumElems + Half);
3439 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3442 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3443 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3445 if (SV->getValueType(0).getVectorNumElements() <= 4)
3446 return SDValue(SV, 0);
3448 EVT PVT = MVT::v4f32;
3449 EVT VT = SV->getValueType(0);
3450 DebugLoc dl = SV->getDebugLoc();
3451 SDValue V1 = SV->getOperand(0);
3452 int NumElems = VT.getVectorNumElements();
3453 int EltNo = SV->getSplatIndex();
3455 // unpack elements to the correct location
3456 while (NumElems > 4) {
3457 if (EltNo < NumElems/2) {
3458 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3460 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3461 EltNo -= NumElems/2;
3466 // Perform the splat.
3467 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3468 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3469 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3470 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3473 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3474 /// vector of zero or undef vector. This produces a shuffle where the low
3475 /// element of V2 is swizzled into the zero/undef vector, landing at element
3476 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3477 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3478 bool isZero, bool HasSSE2,
3479 SelectionDAG &DAG) {
3480 EVT VT = V2.getValueType();
3482 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3483 unsigned NumElems = VT.getVectorNumElements();
3484 SmallVector<int, 16> MaskVec;
3485 for (unsigned i = 0; i != NumElems; ++i)
3486 // If this is the insertion idx, put the low elt of V2 here.
3487 MaskVec.push_back(i == Idx ? NumElems : i);
3488 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3491 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3492 /// a shuffle that is zero.
3494 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3495 bool Low, SelectionDAG &DAG) {
3496 unsigned NumZeros = 0;
3497 for (int i = 0; i < NumElems; ++i) {
3498 unsigned Index = Low ? i : NumElems-i-1;
3499 int Idx = SVOp->getMaskElt(Index);
3504 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3505 if (Elt.getNode() && X86::isZeroNode(Elt))
3513 /// isVectorShift - Returns true if the shuffle can be implemented as a
3514 /// logical left or right shift of a vector.
3515 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3516 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3517 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3518 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3521 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3524 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3528 bool SeenV1 = false;
3529 bool SeenV2 = false;
3530 for (unsigned i = NumZeros; i < NumElems; ++i) {
3531 unsigned Val = isLeft ? (i - NumZeros) : i;
3532 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3535 unsigned Idx = (unsigned) Idx_;
3545 if (SeenV1 && SeenV2)
3548 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3554 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3556 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3557 unsigned NumNonZero, unsigned NumZero,
3559 const TargetLowering &TLI) {
3563 DebugLoc dl = Op.getDebugLoc();
3566 for (unsigned i = 0; i < 16; ++i) {
3567 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3568 if (ThisIsNonZero && First) {
3570 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3572 V = DAG.getUNDEF(MVT::v8i16);
3577 SDValue ThisElt(0, 0), LastElt(0, 0);
3578 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3579 if (LastIsNonZero) {
3580 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3581 MVT::i16, Op.getOperand(i-1));
3583 if (ThisIsNonZero) {
3584 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3585 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3586 ThisElt, DAG.getConstant(8, MVT::i8));
3588 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3592 if (ThisElt.getNode())
3593 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3594 DAG.getIntPtrConstant(i/2));
3598 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3601 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3603 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3604 unsigned NumNonZero, unsigned NumZero,
3606 const TargetLowering &TLI) {
3610 DebugLoc dl = Op.getDebugLoc();
3613 for (unsigned i = 0; i < 8; ++i) {
3614 bool isNonZero = (NonZeros & (1 << i)) != 0;
3618 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3620 V = DAG.getUNDEF(MVT::v8i16);
3623 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3624 MVT::v8i16, V, Op.getOperand(i),
3625 DAG.getIntPtrConstant(i));
3632 /// getVShift - Return a vector logical shift node.
3634 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3635 unsigned NumBits, SelectionDAG &DAG,
3636 const TargetLowering &TLI, DebugLoc dl) {
3637 bool isMMX = VT.getSizeInBits() == 64;
3638 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3639 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3640 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3641 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3642 DAG.getNode(Opc, dl, ShVT, SrcOp,
3643 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3647 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3648 SelectionDAG &DAG) const {
3650 // Check if the scalar load can be widened into a vector load. And if
3651 // the address is "base + cst" see if the cst can be "absorbed" into
3652 // the shuffle mask.
3653 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3654 SDValue Ptr = LD->getBasePtr();
3655 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3657 EVT PVT = LD->getValueType(0);
3658 if (PVT != MVT::i32 && PVT != MVT::f32)
3663 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3664 FI = FINode->getIndex();
3666 } else if (Ptr.getOpcode() == ISD::ADD &&
3667 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3668 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3669 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3670 Offset = Ptr.getConstantOperandVal(1);
3671 Ptr = Ptr.getOperand(0);
3676 SDValue Chain = LD->getChain();
3677 // Make sure the stack object alignment is at least 16.
3678 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3679 if (DAG.InferPtrAlignment(Ptr) < 16) {
3680 if (MFI->isFixedObjectIndex(FI)) {
3681 // Can't change the alignment. FIXME: It's possible to compute
3682 // the exact stack offset and reference FI + adjust offset instead.
3683 // If someone *really* cares about this. That's the way to implement it.
3686 MFI->setObjectAlignment(FI, 16);
3690 // (Offset % 16) must be multiple of 4. Then address is then
3691 // Ptr + (Offset & ~15).
3694 if ((Offset % 16) & 3)
3696 int64_t StartOffset = Offset & ~15;
3698 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3699 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3701 int EltNo = (Offset - StartOffset) >> 2;
3702 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3703 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3704 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3706 // Canonicalize it to a v4i32 shuffle.
3707 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3708 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3709 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3710 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3716 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3717 /// vector of type 'VT', see if the elements can be replaced by a single large
3718 /// load which has the same value as a build_vector whose operands are 'elts'.
3720 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3722 /// FIXME: we'd also like to handle the case where the last elements are zero
3723 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3724 /// There's even a handy isZeroNode for that purpose.
3725 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3726 DebugLoc &dl, SelectionDAG &DAG) {
3727 EVT EltVT = VT.getVectorElementType();
3728 unsigned NumElems = Elts.size();
3730 LoadSDNode *LDBase = NULL;
3731 unsigned LastLoadedElt = -1U;
3733 // For each element in the initializer, see if we've found a load or an undef.
3734 // If we don't find an initial load element, or later load elements are
3735 // non-consecutive, bail out.
3736 for (unsigned i = 0; i < NumElems; ++i) {
3737 SDValue Elt = Elts[i];
3739 if (!Elt.getNode() ||
3740 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3743 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3745 LDBase = cast<LoadSDNode>(Elt.getNode());
3749 if (Elt.getOpcode() == ISD::UNDEF)
3752 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3753 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3758 // If we have found an entire vector of loads and undefs, then return a large
3759 // load of the entire vector width starting at the base pointer. If we found
3760 // consecutive loads for the low half, generate a vzext_load node.
3761 if (LastLoadedElt == NumElems - 1) {
3762 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3763 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3764 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3765 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3766 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3767 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3768 LDBase->isVolatile(), LDBase->isNonTemporal(),
3769 LDBase->getAlignment());
3770 } else if (NumElems == 4 && LastLoadedElt == 1) {
3771 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3772 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3773 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3774 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3780 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3781 DebugLoc dl = Op.getDebugLoc();
3782 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3783 if (ISD::isBuildVectorAllZeros(Op.getNode())
3784 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3785 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3786 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3787 // eliminated on x86-32 hosts.
3788 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3791 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3792 return getOnesVector(Op.getValueType(), DAG, dl);
3793 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3796 EVT VT = Op.getValueType();
3797 EVT ExtVT = VT.getVectorElementType();
3798 unsigned EVTBits = ExtVT.getSizeInBits();
3800 unsigned NumElems = Op.getNumOperands();
3801 unsigned NumZero = 0;
3802 unsigned NumNonZero = 0;
3803 unsigned NonZeros = 0;
3804 bool IsAllConstants = true;
3805 SmallSet<SDValue, 8> Values;
3806 for (unsigned i = 0; i < NumElems; ++i) {
3807 SDValue Elt = Op.getOperand(i);
3808 if (Elt.getOpcode() == ISD::UNDEF)
3811 if (Elt.getOpcode() != ISD::Constant &&
3812 Elt.getOpcode() != ISD::ConstantFP)
3813 IsAllConstants = false;
3814 if (X86::isZeroNode(Elt))
3817 NonZeros |= (1 << i);
3822 if (NumNonZero == 0) {
3823 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3824 return DAG.getUNDEF(VT);
3827 // Special case for single non-zero, non-undef, element.
3828 if (NumNonZero == 1) {
3829 unsigned Idx = CountTrailingZeros_32(NonZeros);
3830 SDValue Item = Op.getOperand(Idx);
3832 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3833 // the value are obviously zero, truncate the value to i32 and do the
3834 // insertion that way. Only do this if the value is non-constant or if the
3835 // value is a constant being inserted into element 0. It is cheaper to do
3836 // a constant pool load than it is to do a movd + shuffle.
3837 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3838 (!IsAllConstants || Idx == 0)) {
3839 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3840 // Handle MMX and SSE both.
3841 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3842 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3844 // Truncate the value (which may itself be a constant) to i32, and
3845 // convert it to a vector with movd (S2V+shuffle to zero extend).
3846 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3847 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3848 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3849 Subtarget->hasSSE2(), DAG);
3851 // Now we have our 32-bit value zero extended in the low element of
3852 // a vector. If Idx != 0, swizzle it into place.
3854 SmallVector<int, 4> Mask;
3855 Mask.push_back(Idx);
3856 for (unsigned i = 1; i != VecElts; ++i)
3858 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3859 DAG.getUNDEF(Item.getValueType()),
3862 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3866 // If we have a constant or non-constant insertion into the low element of
3867 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3868 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3869 // depending on what the source datatype is.
3872 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3873 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3874 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3875 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3876 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3877 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3879 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3880 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3881 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3882 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3883 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3884 Subtarget->hasSSE2(), DAG);
3885 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3889 // Is it a vector logical left shift?
3890 if (NumElems == 2 && Idx == 1 &&
3891 X86::isZeroNode(Op.getOperand(0)) &&
3892 !X86::isZeroNode(Op.getOperand(1))) {
3893 unsigned NumBits = VT.getSizeInBits();
3894 return getVShift(true, VT,
3895 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3896 VT, Op.getOperand(1)),
3897 NumBits/2, DAG, *this, dl);
3900 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3903 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3904 // is a non-constant being inserted into an element other than the low one,
3905 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3906 // movd/movss) to move this into the low element, then shuffle it into
3908 if (EVTBits == 32) {
3909 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3911 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3912 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3913 Subtarget->hasSSE2(), DAG);
3914 SmallVector<int, 8> MaskVec;
3915 for (unsigned i = 0; i < NumElems; i++)
3916 MaskVec.push_back(i == Idx ? 0 : 1);
3917 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3921 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3922 if (Values.size() == 1) {
3923 if (EVTBits == 32) {
3924 // Instead of a shuffle like this:
3925 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3926 // Check if it's possible to issue this instead.
3927 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3928 unsigned Idx = CountTrailingZeros_32(NonZeros);
3929 SDValue Item = Op.getOperand(Idx);
3930 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3931 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3936 // A vector full of immediates; various special cases are already
3937 // handled, so this is best done with a single constant-pool load.
3941 // Let legalizer expand 2-wide build_vectors.
3942 if (EVTBits == 64) {
3943 if (NumNonZero == 1) {
3944 // One half is zero or undef.
3945 unsigned Idx = CountTrailingZeros_32(NonZeros);
3946 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3947 Op.getOperand(Idx));
3948 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3949 Subtarget->hasSSE2(), DAG);
3954 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3955 if (EVTBits == 8 && NumElems == 16) {
3956 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3958 if (V.getNode()) return V;
3961 if (EVTBits == 16 && NumElems == 8) {
3962 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3964 if (V.getNode()) return V;
3967 // If element VT is == 32 bits, turn it into a number of shuffles.
3968 SmallVector<SDValue, 8> V;
3970 if (NumElems == 4 && NumZero > 0) {
3971 for (unsigned i = 0; i < 4; ++i) {
3972 bool isZero = !(NonZeros & (1 << i));
3974 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3976 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3979 for (unsigned i = 0; i < 2; ++i) {
3980 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3983 V[i] = V[i*2]; // Must be a zero vector.
3986 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3989 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3992 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3997 SmallVector<int, 8> MaskVec;
3998 bool Reverse = (NonZeros & 0x3) == 2;
3999 for (unsigned i = 0; i < 2; ++i)
4000 MaskVec.push_back(Reverse ? 1-i : i);
4001 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4002 for (unsigned i = 0; i < 2; ++i)
4003 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4004 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4007 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4008 // Check for a build vector of consecutive loads.
4009 for (unsigned i = 0; i < NumElems; ++i)
4010 V[i] = Op.getOperand(i);
4012 // Check for elements which are consecutive loads.
4013 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4017 // For SSE 4.1, use inserts into undef.
4018 if (getSubtarget()->hasSSE41()) {
4019 V[0] = DAG.getUNDEF(VT);
4020 for (unsigned i = 0; i < NumElems; ++i)
4021 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4022 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
4023 Op.getOperand(i), DAG.getIntPtrConstant(i));
4027 // Otherwise, expand into a number of unpckl*
4029 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4030 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4031 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4032 for (unsigned i = 0; i < NumElems; ++i)
4033 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4035 while (NumElems != 0) {
4036 for (unsigned i = 0; i < NumElems; ++i)
4037 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4046 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4047 // We support concatenate two MMX registers and place them in a MMX
4048 // register. This is better than doing a stack convert.
4049 DebugLoc dl = Op.getDebugLoc();
4050 EVT ResVT = Op.getValueType();
4051 assert(Op.getNumOperands() == 2);
4052 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4053 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4055 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4056 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4057 InVec = Op.getOperand(1);
4058 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4059 unsigned NumElts = ResVT.getVectorNumElements();
4060 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4061 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4062 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4064 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4065 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4066 Mask[0] = 0; Mask[1] = 2;
4067 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4069 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4072 // v8i16 shuffles - Prefer shuffles in the following order:
4073 // 1. [all] pshuflw, pshufhw, optional move
4074 // 2. [ssse3] 1 x pshufb
4075 // 3. [ssse3] 2 x pshufb + 1 x por
4076 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4078 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4080 const X86TargetLowering &TLI) {
4081 SDValue V1 = SVOp->getOperand(0);
4082 SDValue V2 = SVOp->getOperand(1);
4083 DebugLoc dl = SVOp->getDebugLoc();
4084 SmallVector<int, 8> MaskVals;
4086 // Determine if more than 1 of the words in each of the low and high quadwords
4087 // of the result come from the same quadword of one of the two inputs. Undef
4088 // mask values count as coming from any quadword, for better codegen.
4089 SmallVector<unsigned, 4> LoQuad(4);
4090 SmallVector<unsigned, 4> HiQuad(4);
4091 BitVector InputQuads(4);
4092 for (unsigned i = 0; i < 8; ++i) {
4093 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4094 int EltIdx = SVOp->getMaskElt(i);
4095 MaskVals.push_back(EltIdx);
4104 InputQuads.set(EltIdx / 4);
4107 int BestLoQuad = -1;
4108 unsigned MaxQuad = 1;
4109 for (unsigned i = 0; i < 4; ++i) {
4110 if (LoQuad[i] > MaxQuad) {
4112 MaxQuad = LoQuad[i];
4116 int BestHiQuad = -1;
4118 for (unsigned i = 0; i < 4; ++i) {
4119 if (HiQuad[i] > MaxQuad) {
4121 MaxQuad = HiQuad[i];
4125 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4126 // of the two input vectors, shuffle them into one input vector so only a
4127 // single pshufb instruction is necessary. If There are more than 2 input
4128 // quads, disable the next transformation since it does not help SSSE3.
4129 bool V1Used = InputQuads[0] || InputQuads[1];
4130 bool V2Used = InputQuads[2] || InputQuads[3];
4131 if (TLI.getSubtarget()->hasSSSE3()) {
4132 if (InputQuads.count() == 2 && V1Used && V2Used) {
4133 BestLoQuad = InputQuads.find_first();
4134 BestHiQuad = InputQuads.find_next(BestLoQuad);
4136 if (InputQuads.count() > 2) {
4142 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4143 // the shuffle mask. If a quad is scored as -1, that means that it contains
4144 // words from all 4 input quadwords.
4146 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4147 SmallVector<int, 8> MaskV;
4148 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4149 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4150 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4151 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4152 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4153 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4155 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4156 // source words for the shuffle, to aid later transformations.
4157 bool AllWordsInNewV = true;
4158 bool InOrder[2] = { true, true };
4159 for (unsigned i = 0; i != 8; ++i) {
4160 int idx = MaskVals[i];
4162 InOrder[i/4] = false;
4163 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4165 AllWordsInNewV = false;
4169 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4170 if (AllWordsInNewV) {
4171 for (int i = 0; i != 8; ++i) {
4172 int idx = MaskVals[i];
4175 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4176 if ((idx != i) && idx < 4)
4178 if ((idx != i) && idx > 3)
4187 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4188 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4189 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4190 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4191 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4195 // If we have SSSE3, and all words of the result are from 1 input vector,
4196 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4197 // is present, fall back to case 4.
4198 if (TLI.getSubtarget()->hasSSSE3()) {
4199 SmallVector<SDValue,16> pshufbMask;
4201 // If we have elements from both input vectors, set the high bit of the
4202 // shuffle mask element to zero out elements that come from V2 in the V1
4203 // mask, and elements that come from V1 in the V2 mask, so that the two
4204 // results can be OR'd together.
4205 bool TwoInputs = V1Used && V2Used;
4206 for (unsigned i = 0; i != 8; ++i) {
4207 int EltIdx = MaskVals[i] * 2;
4208 if (TwoInputs && (EltIdx >= 16)) {
4209 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4210 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4213 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4214 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4216 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4217 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4218 DAG.getNode(ISD::BUILD_VECTOR, dl,
4219 MVT::v16i8, &pshufbMask[0], 16));
4221 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4223 // Calculate the shuffle mask for the second input, shuffle it, and
4224 // OR it with the first shuffled input.
4226 for (unsigned i = 0; i != 8; ++i) {
4227 int EltIdx = MaskVals[i] * 2;
4229 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4230 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4233 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4234 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4236 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4237 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4238 DAG.getNode(ISD::BUILD_VECTOR, dl,
4239 MVT::v16i8, &pshufbMask[0], 16));
4240 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4241 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4244 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4245 // and update MaskVals with new element order.
4246 BitVector InOrder(8);
4247 if (BestLoQuad >= 0) {
4248 SmallVector<int, 8> MaskV;
4249 for (int i = 0; i != 4; ++i) {
4250 int idx = MaskVals[i];
4252 MaskV.push_back(-1);
4254 } else if ((idx / 4) == BestLoQuad) {
4255 MaskV.push_back(idx & 3);
4258 MaskV.push_back(-1);
4261 for (unsigned i = 4; i != 8; ++i)
4263 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4267 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4268 // and update MaskVals with the new element order.
4269 if (BestHiQuad >= 0) {
4270 SmallVector<int, 8> MaskV;
4271 for (unsigned i = 0; i != 4; ++i)
4273 for (unsigned i = 4; i != 8; ++i) {
4274 int idx = MaskVals[i];
4276 MaskV.push_back(-1);
4278 } else if ((idx / 4) == BestHiQuad) {
4279 MaskV.push_back((idx & 3) + 4);
4282 MaskV.push_back(-1);
4285 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4289 // In case BestHi & BestLo were both -1, which means each quadword has a word
4290 // from each of the four input quadwords, calculate the InOrder bitvector now
4291 // before falling through to the insert/extract cleanup.
4292 if (BestLoQuad == -1 && BestHiQuad == -1) {
4294 for (int i = 0; i != 8; ++i)
4295 if (MaskVals[i] < 0 || MaskVals[i] == i)
4299 // The other elements are put in the right place using pextrw and pinsrw.
4300 for (unsigned i = 0; i != 8; ++i) {
4303 int EltIdx = MaskVals[i];
4306 SDValue ExtOp = (EltIdx < 8)
4307 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4308 DAG.getIntPtrConstant(EltIdx))
4309 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4310 DAG.getIntPtrConstant(EltIdx - 8));
4311 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4312 DAG.getIntPtrConstant(i));
4317 // v16i8 shuffles - Prefer shuffles in the following order:
4318 // 1. [ssse3] 1 x pshufb
4319 // 2. [ssse3] 2 x pshufb + 1 x por
4320 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4322 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4324 const X86TargetLowering &TLI) {
4325 SDValue V1 = SVOp->getOperand(0);
4326 SDValue V2 = SVOp->getOperand(1);
4327 DebugLoc dl = SVOp->getDebugLoc();
4328 SmallVector<int, 16> MaskVals;
4329 SVOp->getMask(MaskVals);
4331 // If we have SSSE3, case 1 is generated when all result bytes come from
4332 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4333 // present, fall back to case 3.
4334 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4337 for (unsigned i = 0; i < 16; ++i) {
4338 int EltIdx = MaskVals[i];
4347 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4348 if (TLI.getSubtarget()->hasSSSE3()) {
4349 SmallVector<SDValue,16> pshufbMask;
4351 // If all result elements are from one input vector, then only translate
4352 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4354 // Otherwise, we have elements from both input vectors, and must zero out
4355 // elements that come from V2 in the first mask, and V1 in the second mask
4356 // so that we can OR them together.
4357 bool TwoInputs = !(V1Only || V2Only);
4358 for (unsigned i = 0; i != 16; ++i) {
4359 int EltIdx = MaskVals[i];
4360 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4361 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4364 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4366 // If all the elements are from V2, assign it to V1 and return after
4367 // building the first pshufb.
4370 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4371 DAG.getNode(ISD::BUILD_VECTOR, dl,
4372 MVT::v16i8, &pshufbMask[0], 16));
4376 // Calculate the shuffle mask for the second input, shuffle it, and
4377 // OR it with the first shuffled input.
4379 for (unsigned i = 0; i != 16; ++i) {
4380 int EltIdx = MaskVals[i];
4382 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4385 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4387 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4388 DAG.getNode(ISD::BUILD_VECTOR, dl,
4389 MVT::v16i8, &pshufbMask[0], 16));
4390 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4393 // No SSSE3 - Calculate in place words and then fix all out of place words
4394 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4395 // the 16 different words that comprise the two doublequadword input vectors.
4396 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4397 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4398 SDValue NewV = V2Only ? V2 : V1;
4399 for (int i = 0; i != 8; ++i) {
4400 int Elt0 = MaskVals[i*2];
4401 int Elt1 = MaskVals[i*2+1];
4403 // This word of the result is all undef, skip it.
4404 if (Elt0 < 0 && Elt1 < 0)
4407 // This word of the result is already in the correct place, skip it.
4408 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4410 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4413 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4414 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4417 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4418 // using a single extract together, load it and store it.
4419 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4420 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4421 DAG.getIntPtrConstant(Elt1 / 2));
4422 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4423 DAG.getIntPtrConstant(i));
4427 // If Elt1 is defined, extract it from the appropriate source. If the
4428 // source byte is not also odd, shift the extracted word left 8 bits
4429 // otherwise clear the bottom 8 bits if we need to do an or.
4431 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4432 DAG.getIntPtrConstant(Elt1 / 2));
4433 if ((Elt1 & 1) == 0)
4434 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4435 DAG.getConstant(8, TLI.getShiftAmountTy()));
4437 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4438 DAG.getConstant(0xFF00, MVT::i16));
4440 // If Elt0 is defined, extract it from the appropriate source. If the
4441 // source byte is not also even, shift the extracted word right 8 bits. If
4442 // Elt1 was also defined, OR the extracted values together before
4443 // inserting them in the result.
4445 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4446 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4447 if ((Elt0 & 1) != 0)
4448 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4449 DAG.getConstant(8, TLI.getShiftAmountTy()));
4451 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4452 DAG.getConstant(0x00FF, MVT::i16));
4453 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4456 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4457 DAG.getIntPtrConstant(i));
4459 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4462 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4463 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4464 /// done when every pair / quad of shuffle mask elements point to elements in
4465 /// the right sequence. e.g.
4466 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4468 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4470 const TargetLowering &TLI, DebugLoc dl) {
4471 EVT VT = SVOp->getValueType(0);
4472 SDValue V1 = SVOp->getOperand(0);
4473 SDValue V2 = SVOp->getOperand(1);
4474 unsigned NumElems = VT.getVectorNumElements();
4475 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4476 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4478 switch (VT.getSimpleVT().SimpleTy) {
4479 default: assert(false && "Unexpected!");
4480 case MVT::v4f32: NewVT = MVT::v2f64; break;
4481 case MVT::v4i32: NewVT = MVT::v2i64; break;
4482 case MVT::v8i16: NewVT = MVT::v4i32; break;
4483 case MVT::v16i8: NewVT = MVT::v4i32; break;
4486 if (NewWidth == 2) {
4492 int Scale = NumElems / NewWidth;
4493 SmallVector<int, 8> MaskVec;
4494 for (unsigned i = 0; i < NumElems; i += Scale) {
4496 for (int j = 0; j < Scale; ++j) {
4497 int EltIdx = SVOp->getMaskElt(i+j);
4501 StartIdx = EltIdx - (EltIdx % Scale);
4502 if (EltIdx != StartIdx + j)
4506 MaskVec.push_back(-1);
4508 MaskVec.push_back(StartIdx / Scale);
4511 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4512 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4513 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4516 /// getVZextMovL - Return a zero-extending vector move low node.
4518 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4519 SDValue SrcOp, SelectionDAG &DAG,
4520 const X86Subtarget *Subtarget, DebugLoc dl) {
4521 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4522 LoadSDNode *LD = NULL;
4523 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4524 LD = dyn_cast<LoadSDNode>(SrcOp);
4526 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4528 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4529 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4530 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4531 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4532 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4534 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4535 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4536 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4537 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4545 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4546 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4547 DAG.getNode(ISD::BIT_CONVERT, dl,
4551 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4554 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4555 SDValue V1 = SVOp->getOperand(0);
4556 SDValue V2 = SVOp->getOperand(1);
4557 DebugLoc dl = SVOp->getDebugLoc();
4558 EVT VT = SVOp->getValueType(0);
4560 SmallVector<std::pair<int, int>, 8> Locs;
4562 SmallVector<int, 8> Mask1(4U, -1);
4563 SmallVector<int, 8> PermMask;
4564 SVOp->getMask(PermMask);
4568 for (unsigned i = 0; i != 4; ++i) {
4569 int Idx = PermMask[i];
4571 Locs[i] = std::make_pair(-1, -1);
4573 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4575 Locs[i] = std::make_pair(0, NumLo);
4579 Locs[i] = std::make_pair(1, NumHi);
4581 Mask1[2+NumHi] = Idx;
4587 if (NumLo <= 2 && NumHi <= 2) {
4588 // If no more than two elements come from either vector. This can be
4589 // implemented with two shuffles. First shuffle gather the elements.
4590 // The second shuffle, which takes the first shuffle as both of its
4591 // vector operands, put the elements into the right order.
4592 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4594 SmallVector<int, 8> Mask2(4U, -1);
4596 for (unsigned i = 0; i != 4; ++i) {
4597 if (Locs[i].first == -1)
4600 unsigned Idx = (i < 2) ? 0 : 4;
4601 Idx += Locs[i].first * 2 + Locs[i].second;
4606 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4607 } else if (NumLo == 3 || NumHi == 3) {
4608 // Otherwise, we must have three elements from one vector, call it X, and
4609 // one element from the other, call it Y. First, use a shufps to build an
4610 // intermediate vector with the one element from Y and the element from X
4611 // that will be in the same half in the final destination (the indexes don't
4612 // matter). Then, use a shufps to build the final vector, taking the half
4613 // containing the element from Y from the intermediate, and the other half
4616 // Normalize it so the 3 elements come from V1.
4617 CommuteVectorShuffleMask(PermMask, VT);
4621 // Find the element from V2.
4623 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4624 int Val = PermMask[HiIndex];
4631 Mask1[0] = PermMask[HiIndex];
4633 Mask1[2] = PermMask[HiIndex^1];
4635 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4638 Mask1[0] = PermMask[0];
4639 Mask1[1] = PermMask[1];
4640 Mask1[2] = HiIndex & 1 ? 6 : 4;
4641 Mask1[3] = HiIndex & 1 ? 4 : 6;
4642 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4644 Mask1[0] = HiIndex & 1 ? 2 : 0;
4645 Mask1[1] = HiIndex & 1 ? 0 : 2;
4646 Mask1[2] = PermMask[2];
4647 Mask1[3] = PermMask[3];
4652 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4656 // Break it into (shuffle shuffle_hi, shuffle_lo).
4658 SmallVector<int,8> LoMask(4U, -1);
4659 SmallVector<int,8> HiMask(4U, -1);
4661 SmallVector<int,8> *MaskPtr = &LoMask;
4662 unsigned MaskIdx = 0;
4665 for (unsigned i = 0; i != 4; ++i) {
4672 int Idx = PermMask[i];
4674 Locs[i] = std::make_pair(-1, -1);
4675 } else if (Idx < 4) {
4676 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4677 (*MaskPtr)[LoIdx] = Idx;
4680 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4681 (*MaskPtr)[HiIdx] = Idx;
4686 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4687 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4688 SmallVector<int, 8> MaskOps;
4689 for (unsigned i = 0; i != 4; ++i) {
4690 if (Locs[i].first == -1) {
4691 MaskOps.push_back(-1);
4693 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4694 MaskOps.push_back(Idx);
4697 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4701 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4702 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4703 SDValue V1 = Op.getOperand(0);
4704 SDValue V2 = Op.getOperand(1);
4705 EVT VT = Op.getValueType();
4706 DebugLoc dl = Op.getDebugLoc();
4707 unsigned NumElems = VT.getVectorNumElements();
4708 bool isMMX = VT.getSizeInBits() == 64;
4709 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4710 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4711 bool V1IsSplat = false;
4712 bool V2IsSplat = false;
4714 if (isZeroShuffle(SVOp))
4715 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4717 // Promote splats to v4f32.
4718 if (SVOp->isSplat()) {
4719 if (isMMX || NumElems < 4)
4721 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4724 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4726 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4727 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4728 if (NewOp.getNode())
4729 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4730 LowerVECTOR_SHUFFLE(NewOp, DAG));
4731 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4732 // FIXME: Figure out a cleaner way to do this.
4733 // Try to make use of movq to zero out the top part.
4734 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4735 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4736 if (NewOp.getNode()) {
4737 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4738 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4739 DAG, Subtarget, dl);
4741 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4742 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4743 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4744 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4745 DAG, Subtarget, dl);
4749 if (X86::isPSHUFDMask(SVOp))
4752 // Check if this can be converted into a logical shift.
4753 bool isLeft = false;
4756 bool isShift = getSubtarget()->hasSSE2() &&
4757 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4758 if (isShift && ShVal.hasOneUse()) {
4759 // If the shifted value has multiple uses, it may be cheaper to use
4760 // v_set0 + movlhps or movhlps, etc.
4761 EVT EltVT = VT.getVectorElementType();
4762 ShAmt *= EltVT.getSizeInBits();
4763 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4766 if (X86::isMOVLMask(SVOp)) {
4769 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4770 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4775 // FIXME: fold these into legal mask.
4776 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4777 X86::isMOVSLDUPMask(SVOp) ||
4778 X86::isMOVHLPSMask(SVOp) ||
4779 X86::isMOVLHPSMask(SVOp) ||
4780 X86::isMOVLPMask(SVOp)))
4783 if (ShouldXformToMOVHLPS(SVOp) ||
4784 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4785 return CommuteVectorShuffle(SVOp, DAG);
4788 // No better options. Use a vshl / vsrl.
4789 EVT EltVT = VT.getVectorElementType();
4790 ShAmt *= EltVT.getSizeInBits();
4791 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4794 bool Commuted = false;
4795 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4796 // 1,1,1,1 -> v8i16 though.
4797 V1IsSplat = isSplatVector(V1.getNode());
4798 V2IsSplat = isSplatVector(V2.getNode());
4800 // Canonicalize the splat or undef, if present, to be on the RHS.
4801 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4802 Op = CommuteVectorShuffle(SVOp, DAG);
4803 SVOp = cast<ShuffleVectorSDNode>(Op);
4804 V1 = SVOp->getOperand(0);
4805 V2 = SVOp->getOperand(1);
4806 std::swap(V1IsSplat, V2IsSplat);
4807 std::swap(V1IsUndef, V2IsUndef);
4811 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4812 // Shuffling low element of v1 into undef, just return v1.
4815 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4816 // the instruction selector will not match, so get a canonical MOVL with
4817 // swapped operands to undo the commute.
4818 return getMOVL(DAG, dl, VT, V2, V1);
4821 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4822 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4823 X86::isUNPCKLMask(SVOp) ||
4824 X86::isUNPCKHMask(SVOp))
4828 // Normalize mask so all entries that point to V2 points to its first
4829 // element then try to match unpck{h|l} again. If match, return a
4830 // new vector_shuffle with the corrected mask.
4831 SDValue NewMask = NormalizeMask(SVOp, DAG);
4832 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4833 if (NSVOp != SVOp) {
4834 if (X86::isUNPCKLMask(NSVOp, true)) {
4836 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4843 // Commute is back and try unpck* again.
4844 // FIXME: this seems wrong.
4845 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4846 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4847 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4848 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4849 X86::isUNPCKLMask(NewSVOp) ||
4850 X86::isUNPCKHMask(NewSVOp))
4854 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4856 // Normalize the node to match x86 shuffle ops if needed
4857 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4858 return CommuteVectorShuffle(SVOp, DAG);
4860 // Check for legal shuffle and return?
4861 SmallVector<int, 16> PermMask;
4862 SVOp->getMask(PermMask);
4863 if (isShuffleMaskLegal(PermMask, VT))
4866 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4867 if (VT == MVT::v8i16) {
4868 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4869 if (NewOp.getNode())
4873 if (VT == MVT::v16i8) {
4874 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4875 if (NewOp.getNode())
4879 // Handle all 4 wide cases with a number of shuffles except for MMX.
4880 if (NumElems == 4 && !isMMX)
4881 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4887 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4888 SelectionDAG &DAG) const {
4889 EVT VT = Op.getValueType();
4890 DebugLoc dl = Op.getDebugLoc();
4891 if (VT.getSizeInBits() == 8) {
4892 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4893 Op.getOperand(0), Op.getOperand(1));
4894 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4895 DAG.getValueType(VT));
4896 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4897 } else if (VT.getSizeInBits() == 16) {
4898 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4899 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4901 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4902 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4903 DAG.getNode(ISD::BIT_CONVERT, dl,
4907 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4908 Op.getOperand(0), Op.getOperand(1));
4909 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4910 DAG.getValueType(VT));
4911 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4912 } else if (VT == MVT::f32) {
4913 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4914 // the result back to FR32 register. It's only worth matching if the
4915 // result has a single use which is a store or a bitcast to i32. And in
4916 // the case of a store, it's not worth it if the index is a constant 0,
4917 // because a MOVSSmr can be used instead, which is smaller and faster.
4918 if (!Op.hasOneUse())
4920 SDNode *User = *Op.getNode()->use_begin();
4921 if ((User->getOpcode() != ISD::STORE ||
4922 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4923 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4924 (User->getOpcode() != ISD::BIT_CONVERT ||
4925 User->getValueType(0) != MVT::i32))
4927 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4928 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4931 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4932 } else if (VT == MVT::i32) {
4933 // ExtractPS works with constant index.
4934 if (isa<ConstantSDNode>(Op.getOperand(1)))
4942 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4943 SelectionDAG &DAG) const {
4944 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4947 if (Subtarget->hasSSE41()) {
4948 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4953 EVT VT = Op.getValueType();
4954 DebugLoc dl = Op.getDebugLoc();
4955 // TODO: handle v16i8.
4956 if (VT.getSizeInBits() == 16) {
4957 SDValue Vec = Op.getOperand(0);
4958 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4960 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4961 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4962 DAG.getNode(ISD::BIT_CONVERT, dl,
4965 // Transform it so it match pextrw which produces a 32-bit result.
4966 EVT EltVT = MVT::i32;
4967 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4968 Op.getOperand(0), Op.getOperand(1));
4969 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4970 DAG.getValueType(VT));
4971 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4972 } else if (VT.getSizeInBits() == 32) {
4973 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4977 // SHUFPS the element to the lowest double word, then movss.
4978 int Mask[4] = { Idx, -1, -1, -1 };
4979 EVT VVT = Op.getOperand(0).getValueType();
4980 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4981 DAG.getUNDEF(VVT), Mask);
4982 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4983 DAG.getIntPtrConstant(0));
4984 } else if (VT.getSizeInBits() == 64) {
4985 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4986 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4987 // to match extract_elt for f64.
4988 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4992 // UNPCKHPD the element to the lowest double word, then movsd.
4993 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4994 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4995 int Mask[2] = { 1, -1 };
4996 EVT VVT = Op.getOperand(0).getValueType();
4997 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4998 DAG.getUNDEF(VVT), Mask);
4999 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5000 DAG.getIntPtrConstant(0));
5007 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5008 SelectionDAG &DAG) const {
5009 EVT VT = Op.getValueType();
5010 EVT EltVT = VT.getVectorElementType();
5011 DebugLoc dl = Op.getDebugLoc();
5013 SDValue N0 = Op.getOperand(0);
5014 SDValue N1 = Op.getOperand(1);
5015 SDValue N2 = Op.getOperand(2);
5017 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5018 isa<ConstantSDNode>(N2)) {
5020 if (VT == MVT::v8i16)
5021 Opc = X86ISD::PINSRW;
5022 else if (VT == MVT::v4i16)
5023 Opc = X86ISD::MMX_PINSRW;
5024 else if (VT == MVT::v16i8)
5025 Opc = X86ISD::PINSRB;
5027 Opc = X86ISD::PINSRB;
5029 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5031 if (N1.getValueType() != MVT::i32)
5032 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5033 if (N2.getValueType() != MVT::i32)
5034 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5035 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5036 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5037 // Bits [7:6] of the constant are the source select. This will always be
5038 // zero here. The DAG Combiner may combine an extract_elt index into these
5039 // bits. For example (insert (extract, 3), 2) could be matched by putting
5040 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5041 // Bits [5:4] of the constant are the destination select. This is the
5042 // value of the incoming immediate.
5043 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5044 // combine either bitwise AND or insert of float 0.0 to set these bits.
5045 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5046 // Create this as a scalar to vector..
5047 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5048 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5049 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5050 // PINSR* works with constant index.
5057 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5058 EVT VT = Op.getValueType();
5059 EVT EltVT = VT.getVectorElementType();
5061 if (Subtarget->hasSSE41())
5062 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5064 if (EltVT == MVT::i8)
5067 DebugLoc dl = Op.getDebugLoc();
5068 SDValue N0 = Op.getOperand(0);
5069 SDValue N1 = Op.getOperand(1);
5070 SDValue N2 = Op.getOperand(2);
5072 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5073 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5074 // as its second argument.
5075 if (N1.getValueType() != MVT::i32)
5076 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5077 if (N2.getValueType() != MVT::i32)
5078 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5079 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5080 dl, VT, N0, N1, N2);
5086 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5087 DebugLoc dl = Op.getDebugLoc();
5089 if (Op.getValueType() == MVT::v1i64 &&
5090 Op.getOperand(0).getValueType() == MVT::i64)
5091 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5093 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5094 EVT VT = MVT::v2i32;
5095 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5102 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5103 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5106 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5107 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5108 // one of the above mentioned nodes. It has to be wrapped because otherwise
5109 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5110 // be used to form addressing mode. These wrapped nodes will be selected
5113 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5114 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5116 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5118 unsigned char OpFlag = 0;
5119 unsigned WrapperKind = X86ISD::Wrapper;
5120 CodeModel::Model M = getTargetMachine().getCodeModel();
5122 if (Subtarget->isPICStyleRIPRel() &&
5123 (M == CodeModel::Small || M == CodeModel::Kernel))
5124 WrapperKind = X86ISD::WrapperRIP;
5125 else if (Subtarget->isPICStyleGOT())
5126 OpFlag = X86II::MO_GOTOFF;
5127 else if (Subtarget->isPICStyleStubPIC())
5128 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5130 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5132 CP->getOffset(), OpFlag);
5133 DebugLoc DL = CP->getDebugLoc();
5134 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5135 // With PIC, the address is actually $g + Offset.
5137 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5138 DAG.getNode(X86ISD::GlobalBaseReg,
5139 DebugLoc(), getPointerTy()),
5146 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5147 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5149 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5151 unsigned char OpFlag = 0;
5152 unsigned WrapperKind = X86ISD::Wrapper;
5153 CodeModel::Model M = getTargetMachine().getCodeModel();
5155 if (Subtarget->isPICStyleRIPRel() &&
5156 (M == CodeModel::Small || M == CodeModel::Kernel))
5157 WrapperKind = X86ISD::WrapperRIP;
5158 else if (Subtarget->isPICStyleGOT())
5159 OpFlag = X86II::MO_GOTOFF;
5160 else if (Subtarget->isPICStyleStubPIC())
5161 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5163 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5165 DebugLoc DL = JT->getDebugLoc();
5166 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5168 // With PIC, the address is actually $g + Offset.
5170 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5171 DAG.getNode(X86ISD::GlobalBaseReg,
5172 DebugLoc(), getPointerTy()),
5180 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5181 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5183 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5185 unsigned char OpFlag = 0;
5186 unsigned WrapperKind = X86ISD::Wrapper;
5187 CodeModel::Model M = getTargetMachine().getCodeModel();
5189 if (Subtarget->isPICStyleRIPRel() &&
5190 (M == CodeModel::Small || M == CodeModel::Kernel))
5191 WrapperKind = X86ISD::WrapperRIP;
5192 else if (Subtarget->isPICStyleGOT())
5193 OpFlag = X86II::MO_GOTOFF;
5194 else if (Subtarget->isPICStyleStubPIC())
5195 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5197 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5199 DebugLoc DL = Op.getDebugLoc();
5200 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5203 // With PIC, the address is actually $g + Offset.
5204 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5205 !Subtarget->is64Bit()) {
5206 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5207 DAG.getNode(X86ISD::GlobalBaseReg,
5208 DebugLoc(), getPointerTy()),
5216 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5217 // Create the TargetBlockAddressAddress node.
5218 unsigned char OpFlags =
5219 Subtarget->ClassifyBlockAddressReference();
5220 CodeModel::Model M = getTargetMachine().getCodeModel();
5221 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5222 DebugLoc dl = Op.getDebugLoc();
5223 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5224 /*isTarget=*/true, OpFlags);
5226 if (Subtarget->isPICStyleRIPRel() &&
5227 (M == CodeModel::Small || M == CodeModel::Kernel))
5228 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5230 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5232 // With PIC, the address is actually $g + Offset.
5233 if (isGlobalRelativeToPICBase(OpFlags)) {
5234 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5235 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5243 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5245 SelectionDAG &DAG) const {
5246 // Create the TargetGlobalAddress node, folding in the constant
5247 // offset if it is legal.
5248 unsigned char OpFlags =
5249 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5250 CodeModel::Model M = getTargetMachine().getCodeModel();
5252 if (OpFlags == X86II::MO_NO_FLAG &&
5253 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5254 // A direct static reference to a global.
5255 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5258 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5261 if (Subtarget->isPICStyleRIPRel() &&
5262 (M == CodeModel::Small || M == CodeModel::Kernel))
5263 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5265 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5267 // With PIC, the address is actually $g + Offset.
5268 if (isGlobalRelativeToPICBase(OpFlags)) {
5269 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5270 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5274 // For globals that require a load from a stub to get the address, emit the
5276 if (isGlobalStubReference(OpFlags))
5277 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5278 PseudoSourceValue::getGOT(), 0, false, false, 0);
5280 // If there was a non-zero offset that we didn't fold, create an explicit
5283 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5284 DAG.getConstant(Offset, getPointerTy()));
5290 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5291 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5292 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5293 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5297 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5298 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5299 unsigned char OperandFlags) {
5300 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5301 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5302 DebugLoc dl = GA->getDebugLoc();
5303 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5304 GA->getValueType(0),
5308 SDValue Ops[] = { Chain, TGA, *InFlag };
5309 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5311 SDValue Ops[] = { Chain, TGA };
5312 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5315 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5316 MFI->setAdjustsStack(true);
5318 SDValue Flag = Chain.getValue(1);
5319 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5322 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5324 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5327 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5328 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5329 DAG.getNode(X86ISD::GlobalBaseReg,
5330 DebugLoc(), PtrVT), InFlag);
5331 InFlag = Chain.getValue(1);
5333 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5336 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5338 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5340 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5341 X86::RAX, X86II::MO_TLSGD);
5344 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5345 // "local exec" model.
5346 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5347 const EVT PtrVT, TLSModel::Model model,
5349 DebugLoc dl = GA->getDebugLoc();
5350 // Get the Thread Pointer
5351 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5353 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5356 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5357 NULL, 0, false, false, 0);
5359 unsigned char OperandFlags = 0;
5360 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5362 unsigned WrapperKind = X86ISD::Wrapper;
5363 if (model == TLSModel::LocalExec) {
5364 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5365 } else if (is64Bit) {
5366 assert(model == TLSModel::InitialExec);
5367 OperandFlags = X86II::MO_GOTTPOFF;
5368 WrapperKind = X86ISD::WrapperRIP;
5370 assert(model == TLSModel::InitialExec);
5371 OperandFlags = X86II::MO_INDNTPOFF;
5374 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5376 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5377 GA->getValueType(0),
5378 GA->getOffset(), OperandFlags);
5379 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5381 if (model == TLSModel::InitialExec)
5382 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5383 PseudoSourceValue::getGOT(), 0, false, false, 0);
5385 // The address of the thread local variable is the add of the thread
5386 // pointer with the offset of the variable.
5387 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5391 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5393 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5394 const GlobalValue *GV = GA->getGlobal();
5396 if (Subtarget->isTargetELF()) {
5397 // TODO: implement the "local dynamic" model
5398 // TODO: implement the "initial exec"model for pic executables
5400 // If GV is an alias then use the aliasee for determining
5401 // thread-localness.
5402 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5403 GV = GA->resolveAliasedGlobal(false);
5405 TLSModel::Model model
5406 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5409 case TLSModel::GeneralDynamic:
5410 case TLSModel::LocalDynamic: // not implemented
5411 if (Subtarget->is64Bit())
5412 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5413 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5415 case TLSModel::InitialExec:
5416 case TLSModel::LocalExec:
5417 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5418 Subtarget->is64Bit());
5420 } else if (Subtarget->isTargetDarwin()) {
5421 // Darwin only has one model of TLS. Lower to that.
5422 unsigned char OpFlag = 0;
5423 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5424 X86ISD::WrapperRIP : X86ISD::Wrapper;
5426 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5428 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5429 !Subtarget->is64Bit();
5431 OpFlag = X86II::MO_TLVP_PIC_BASE;
5433 OpFlag = X86II::MO_TLVP;
5434 DebugLoc DL = Op.getDebugLoc();
5435 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
5437 GA->getOffset(), OpFlag);
5438 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5440 // With PIC32, the address is actually $g + Offset.
5442 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5443 DAG.getNode(X86ISD::GlobalBaseReg,
5444 DebugLoc(), getPointerTy()),
5447 // Lowering the machine isd will make sure everything is in the right
5449 SDValue Args[] = { Offset };
5450 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5452 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5453 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5454 MFI->setAdjustsStack(true);
5456 // And our return value (tls address) is in the standard call return value
5458 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5459 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5463 "TLS not implemented for this target.");
5465 llvm_unreachable("Unreachable");
5470 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5471 /// take a 2 x i32 value to shift plus a shift amount.
5472 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5473 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5474 EVT VT = Op.getValueType();
5475 unsigned VTBits = VT.getSizeInBits();
5476 DebugLoc dl = Op.getDebugLoc();
5477 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5478 SDValue ShOpLo = Op.getOperand(0);
5479 SDValue ShOpHi = Op.getOperand(1);
5480 SDValue ShAmt = Op.getOperand(2);
5481 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5482 DAG.getConstant(VTBits - 1, MVT::i8))
5483 : DAG.getConstant(0, VT);
5486 if (Op.getOpcode() == ISD::SHL_PARTS) {
5487 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5488 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5490 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5491 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5494 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5495 DAG.getConstant(VTBits, MVT::i8));
5496 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5497 AndNode, DAG.getConstant(0, MVT::i8));
5500 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5501 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5502 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5504 if (Op.getOpcode() == ISD::SHL_PARTS) {
5505 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5506 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5508 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5509 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5512 SDValue Ops[2] = { Lo, Hi };
5513 return DAG.getMergeValues(Ops, 2, dl);
5516 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5517 SelectionDAG &DAG) const {
5518 EVT SrcVT = Op.getOperand(0).getValueType();
5520 if (SrcVT.isVector()) {
5521 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5527 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5528 "Unknown SINT_TO_FP to lower!");
5530 // These are really Legal; return the operand so the caller accepts it as
5532 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5534 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5535 Subtarget->is64Bit()) {
5539 DebugLoc dl = Op.getDebugLoc();
5540 unsigned Size = SrcVT.getSizeInBits()/8;
5541 MachineFunction &MF = DAG.getMachineFunction();
5542 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5543 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5544 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5546 PseudoSourceValue::getFixedStack(SSFI), 0,
5548 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5551 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5553 SelectionDAG &DAG) const {
5555 DebugLoc dl = Op.getDebugLoc();
5557 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5559 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5561 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5562 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5563 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5564 Tys, Ops, array_lengthof(Ops));
5567 Chain = Result.getValue(1);
5568 SDValue InFlag = Result.getValue(2);
5570 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5571 // shouldn't be necessary except that RFP cannot be live across
5572 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5573 MachineFunction &MF = DAG.getMachineFunction();
5574 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5575 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5576 Tys = DAG.getVTList(MVT::Other);
5578 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5580 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5581 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5582 PseudoSourceValue::getFixedStack(SSFI), 0,
5589 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5590 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5591 SelectionDAG &DAG) const {
5592 // This algorithm is not obvious. Here it is in C code, more or less:
5594 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5595 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5596 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5598 // Copy ints to xmm registers.
5599 __m128i xh = _mm_cvtsi32_si128( hi );
5600 __m128i xl = _mm_cvtsi32_si128( lo );
5602 // Combine into low half of a single xmm register.
5603 __m128i x = _mm_unpacklo_epi32( xh, xl );
5607 // Merge in appropriate exponents to give the integer bits the right
5609 x = _mm_unpacklo_epi32( x, exp );
5611 // Subtract away the biases to deal with the IEEE-754 double precision
5613 d = _mm_sub_pd( (__m128d) x, bias );
5615 // All conversions up to here are exact. The correctly rounded result is
5616 // calculated using the current rounding mode using the following
5618 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5619 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5620 // store doesn't really need to be here (except
5621 // maybe to zero the other double)
5626 DebugLoc dl = Op.getDebugLoc();
5627 LLVMContext *Context = DAG.getContext();
5629 // Build some magic constants.
5630 std::vector<Constant*> CV0;
5631 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5632 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5633 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5634 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5635 Constant *C0 = ConstantVector::get(CV0);
5636 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5638 std::vector<Constant*> CV1;
5640 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5642 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5643 Constant *C1 = ConstantVector::get(CV1);
5644 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5646 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5647 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5649 DAG.getIntPtrConstant(1)));
5650 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5651 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5653 DAG.getIntPtrConstant(0)));
5654 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5655 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5656 PseudoSourceValue::getConstantPool(), 0,
5658 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5659 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5660 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5661 PseudoSourceValue::getConstantPool(), 0,
5663 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5665 // Add the halves; easiest way is to swap them into another reg first.
5666 int ShufMask[2] = { 1, -1 };
5667 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5668 DAG.getUNDEF(MVT::v2f64), ShufMask);
5669 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5670 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5671 DAG.getIntPtrConstant(0));
5674 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5675 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5676 SelectionDAG &DAG) const {
5677 DebugLoc dl = Op.getDebugLoc();
5678 // FP constant to bias correct the final result.
5679 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5682 // Load the 32-bit value into an XMM register.
5683 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5684 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5686 DAG.getIntPtrConstant(0)));
5688 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5689 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5690 DAG.getIntPtrConstant(0));
5692 // Or the load with the bias.
5693 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5694 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5695 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5697 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5698 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5699 MVT::v2f64, Bias)));
5700 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5701 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5702 DAG.getIntPtrConstant(0));
5704 // Subtract the bias.
5705 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5707 // Handle final rounding.
5708 EVT DestVT = Op.getValueType();
5710 if (DestVT.bitsLT(MVT::f64)) {
5711 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5712 DAG.getIntPtrConstant(0));
5713 } else if (DestVT.bitsGT(MVT::f64)) {
5714 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5717 // Handle final rounding.
5721 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5722 SelectionDAG &DAG) const {
5723 SDValue N0 = Op.getOperand(0);
5724 DebugLoc dl = Op.getDebugLoc();
5726 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5727 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5728 // the optimization here.
5729 if (DAG.SignBitIsZero(N0))
5730 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5732 EVT SrcVT = N0.getValueType();
5733 EVT DstVT = Op.getValueType();
5734 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5735 return LowerUINT_TO_FP_i64(Op, DAG);
5736 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5737 return LowerUINT_TO_FP_i32(Op, DAG);
5739 // Make a 64-bit buffer, and use it to build an FILD.
5740 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5741 if (SrcVT == MVT::i32) {
5742 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5743 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5744 getPointerTy(), StackSlot, WordOff);
5745 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5746 StackSlot, NULL, 0, false, false, 0);
5747 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5748 OffsetSlot, NULL, 0, false, false, 0);
5749 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5753 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5754 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5755 StackSlot, NULL, 0, false, false, 0);
5756 // For i64 source, we need to add the appropriate power of 2 if the input
5757 // was negative. This is the same as the optimization in
5758 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5759 // we must be careful to do the computation in x87 extended precision, not
5760 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5761 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5762 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5763 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5765 APInt FF(32, 0x5F800000ULL);
5767 // Check whether the sign bit is set.
5768 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5769 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5772 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5773 SDValue FudgePtr = DAG.getConstantPool(
5774 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5777 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5778 SDValue Zero = DAG.getIntPtrConstant(0);
5779 SDValue Four = DAG.getIntPtrConstant(4);
5780 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5782 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5784 // Load the value out, extending it from f32 to f80.
5785 // FIXME: Avoid the extend by constructing the right constant pool?
5786 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
5787 FudgePtr, PseudoSourceValue::getConstantPool(),
5788 0, MVT::f32, false, false, 4);
5789 // Extend everything to 80 bits to force it to be done on x87.
5790 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5791 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5794 std::pair<SDValue,SDValue> X86TargetLowering::
5795 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5796 DebugLoc dl = Op.getDebugLoc();
5798 EVT DstTy = Op.getValueType();
5801 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5805 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5806 DstTy.getSimpleVT() >= MVT::i16 &&
5807 "Unknown FP_TO_SINT to lower!");
5809 // These are really Legal.
5810 if (DstTy == MVT::i32 &&
5811 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5812 return std::make_pair(SDValue(), SDValue());
5813 if (Subtarget->is64Bit() &&
5814 DstTy == MVT::i64 &&
5815 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5816 return std::make_pair(SDValue(), SDValue());
5818 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5820 MachineFunction &MF = DAG.getMachineFunction();
5821 unsigned MemSize = DstTy.getSizeInBits()/8;
5822 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5823 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5826 switch (DstTy.getSimpleVT().SimpleTy) {
5827 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5828 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5829 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5830 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5833 SDValue Chain = DAG.getEntryNode();
5834 SDValue Value = Op.getOperand(0);
5835 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5836 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5837 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5838 PseudoSourceValue::getFixedStack(SSFI), 0,
5840 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5842 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5844 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5845 Chain = Value.getValue(1);
5846 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5847 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5850 // Build the FP_TO_INT*_IN_MEM
5851 SDValue Ops[] = { Chain, Value, StackSlot };
5852 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5854 return std::make_pair(FIST, StackSlot);
5857 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5858 SelectionDAG &DAG) const {
5859 if (Op.getValueType().isVector()) {
5860 if (Op.getValueType() == MVT::v2i32 &&
5861 Op.getOperand(0).getValueType() == MVT::v2f64) {
5867 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5868 SDValue FIST = Vals.first, StackSlot = Vals.second;
5869 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5870 if (FIST.getNode() == 0) return Op;
5873 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5874 FIST, StackSlot, NULL, 0, false, false, 0);
5877 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5878 SelectionDAG &DAG) const {
5879 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5880 SDValue FIST = Vals.first, StackSlot = Vals.second;
5881 assert(FIST.getNode() && "Unexpected failure");
5884 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5885 FIST, StackSlot, NULL, 0, false, false, 0);
5888 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5889 SelectionDAG &DAG) const {
5890 LLVMContext *Context = DAG.getContext();
5891 DebugLoc dl = Op.getDebugLoc();
5892 EVT VT = Op.getValueType();
5895 EltVT = VT.getVectorElementType();
5896 std::vector<Constant*> CV;
5897 if (EltVT == MVT::f64) {
5898 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5902 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5908 Constant *C = ConstantVector::get(CV);
5909 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5910 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5911 PseudoSourceValue::getConstantPool(), 0,
5913 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5916 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5917 LLVMContext *Context = DAG.getContext();
5918 DebugLoc dl = Op.getDebugLoc();
5919 EVT VT = Op.getValueType();
5922 EltVT = VT.getVectorElementType();
5923 std::vector<Constant*> CV;
5924 if (EltVT == MVT::f64) {
5925 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5929 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5935 Constant *C = ConstantVector::get(CV);
5936 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5937 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5938 PseudoSourceValue::getConstantPool(), 0,
5940 if (VT.isVector()) {
5941 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5942 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5943 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5945 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5947 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5951 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5952 LLVMContext *Context = DAG.getContext();
5953 SDValue Op0 = Op.getOperand(0);
5954 SDValue Op1 = Op.getOperand(1);
5955 DebugLoc dl = Op.getDebugLoc();
5956 EVT VT = Op.getValueType();
5957 EVT SrcVT = Op1.getValueType();
5959 // If second operand is smaller, extend it first.
5960 if (SrcVT.bitsLT(VT)) {
5961 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5964 // And if it is bigger, shrink it first.
5965 if (SrcVT.bitsGT(VT)) {
5966 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5970 // At this point the operands and the result should have the same
5971 // type, and that won't be f80 since that is not custom lowered.
5973 // First get the sign bit of second operand.
5974 std::vector<Constant*> CV;
5975 if (SrcVT == MVT::f64) {
5976 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5977 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5979 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5980 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5981 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5982 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5984 Constant *C = ConstantVector::get(CV);
5985 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5986 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5987 PseudoSourceValue::getConstantPool(), 0,
5989 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5991 // Shift sign bit right or left if the two operands have different types.
5992 if (SrcVT.bitsGT(VT)) {
5993 // Op0 is MVT::f32, Op1 is MVT::f64.
5994 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5995 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5996 DAG.getConstant(32, MVT::i32));
5997 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5998 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5999 DAG.getIntPtrConstant(0));
6002 // Clear first operand sign bit.
6004 if (VT == MVT::f64) {
6005 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6006 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6008 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6009 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6010 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6011 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6013 C = ConstantVector::get(CV);
6014 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6015 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6016 PseudoSourceValue::getConstantPool(), 0,
6018 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6020 // Or the value with the sign bit.
6021 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6024 /// Emit nodes that will be selected as "test Op0,Op0", or something
6026 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6027 SelectionDAG &DAG) const {
6028 DebugLoc dl = Op.getDebugLoc();
6030 // CF and OF aren't always set the way we want. Determine which
6031 // of these we need.
6032 bool NeedCF = false;
6033 bool NeedOF = false;
6036 case X86::COND_A: case X86::COND_AE:
6037 case X86::COND_B: case X86::COND_BE:
6040 case X86::COND_G: case X86::COND_GE:
6041 case X86::COND_L: case X86::COND_LE:
6042 case X86::COND_O: case X86::COND_NO:
6047 // See if we can use the EFLAGS value from the operand instead of
6048 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6049 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6050 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6051 // Emit a CMP with 0, which is the TEST pattern.
6052 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6053 DAG.getConstant(0, Op.getValueType()));
6055 unsigned Opcode = 0;
6056 unsigned NumOperands = 0;
6057 switch (Op.getNode()->getOpcode()) {
6059 // Due to an isel shortcoming, be conservative if this add is likely to be
6060 // selected as part of a load-modify-store instruction. When the root node
6061 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6062 // uses of other nodes in the match, such as the ADD in this case. This
6063 // leads to the ADD being left around and reselected, with the result being
6064 // two adds in the output. Alas, even if none our users are stores, that
6065 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6066 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6067 // climbing the DAG back to the root, and it doesn't seem to be worth the
6069 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6070 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6071 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6074 if (ConstantSDNode *C =
6075 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6076 // An add of one will be selected as an INC.
6077 if (C->getAPIntValue() == 1) {
6078 Opcode = X86ISD::INC;
6083 // An add of negative one (subtract of one) will be selected as a DEC.
6084 if (C->getAPIntValue().isAllOnesValue()) {
6085 Opcode = X86ISD::DEC;
6091 // Otherwise use a regular EFLAGS-setting add.
6092 Opcode = X86ISD::ADD;
6096 // If the primary and result isn't used, don't bother using X86ISD::AND,
6097 // because a TEST instruction will be better.
6098 bool NonFlagUse = false;
6099 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6100 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6102 unsigned UOpNo = UI.getOperandNo();
6103 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6104 // Look pass truncate.
6105 UOpNo = User->use_begin().getOperandNo();
6106 User = *User->use_begin();
6109 if (User->getOpcode() != ISD::BRCOND &&
6110 User->getOpcode() != ISD::SETCC &&
6111 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6124 // Due to the ISEL shortcoming noted above, be conservative if this op is
6125 // likely to be selected as part of a load-modify-store instruction.
6126 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6127 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6128 if (UI->getOpcode() == ISD::STORE)
6131 // Otherwise use a regular EFLAGS-setting instruction.
6132 switch (Op.getNode()->getOpcode()) {
6133 default: llvm_unreachable("unexpected operator!");
6134 case ISD::SUB: Opcode = X86ISD::SUB; break;
6135 case ISD::OR: Opcode = X86ISD::OR; break;
6136 case ISD::XOR: Opcode = X86ISD::XOR; break;
6137 case ISD::AND: Opcode = X86ISD::AND; break;
6149 return SDValue(Op.getNode(), 1);
6156 // Emit a CMP with 0, which is the TEST pattern.
6157 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6158 DAG.getConstant(0, Op.getValueType()));
6160 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6161 SmallVector<SDValue, 4> Ops;
6162 for (unsigned i = 0; i != NumOperands; ++i)
6163 Ops.push_back(Op.getOperand(i));
6165 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6166 DAG.ReplaceAllUsesWith(Op, New);
6167 return SDValue(New.getNode(), 1);
6170 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6172 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6173 SelectionDAG &DAG) const {
6174 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6175 if (C->getAPIntValue() == 0)
6176 return EmitTest(Op0, X86CC, DAG);
6178 DebugLoc dl = Op0.getDebugLoc();
6179 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6182 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6183 /// if it's possible.
6184 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6185 DebugLoc dl, SelectionDAG &DAG) const {
6186 SDValue Op0 = And.getOperand(0);
6187 SDValue Op1 = And.getOperand(1);
6188 if (Op0.getOpcode() == ISD::TRUNCATE)
6189 Op0 = Op0.getOperand(0);
6190 if (Op1.getOpcode() == ISD::TRUNCATE)
6191 Op1 = Op1.getOperand(0);
6194 if (Op1.getOpcode() == ISD::SHL)
6195 std::swap(Op0, Op1);
6196 if (Op0.getOpcode() == ISD::SHL) {
6197 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6198 if (And00C->getZExtValue() == 1) {
6199 // If we looked past a truncate, check that it's only truncating away
6201 unsigned BitWidth = Op0.getValueSizeInBits();
6202 unsigned AndBitWidth = And.getValueSizeInBits();
6203 if (BitWidth > AndBitWidth) {
6204 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6205 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6206 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6210 RHS = Op0.getOperand(1);
6212 } else if (Op1.getOpcode() == ISD::Constant) {
6213 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6214 SDValue AndLHS = Op0;
6215 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6216 LHS = AndLHS.getOperand(0);
6217 RHS = AndLHS.getOperand(1);
6221 if (LHS.getNode()) {
6222 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6223 // instruction. Since the shift amount is in-range-or-undefined, we know
6224 // that doing a bittest on the i32 value is ok. We extend to i32 because
6225 // the encoding for the i16 version is larger than the i32 version.
6226 // Also promote i16 to i32 for performance / code size reason.
6227 if (LHS.getValueType() == MVT::i8 ||
6228 LHS.getValueType() == MVT::i16)
6229 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6231 // If the operand types disagree, extend the shift amount to match. Since
6232 // BT ignores high bits (like shifts) we can use anyextend.
6233 if (LHS.getValueType() != RHS.getValueType())
6234 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6236 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6237 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6238 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6239 DAG.getConstant(Cond, MVT::i8), BT);
6245 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6246 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6247 SDValue Op0 = Op.getOperand(0);
6248 SDValue Op1 = Op.getOperand(1);
6249 DebugLoc dl = Op.getDebugLoc();
6250 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6252 // Optimize to BT if possible.
6253 // Lower (X & (1 << N)) == 0 to BT(X, N).
6254 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6255 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6256 if (Op0.getOpcode() == ISD::AND &&
6258 Op1.getOpcode() == ISD::Constant &&
6259 cast<ConstantSDNode>(Op1)->isNullValue() &&
6260 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6261 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6262 if (NewSetCC.getNode())
6266 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6267 if (Op0.getOpcode() == X86ISD::SETCC &&
6268 Op1.getOpcode() == ISD::Constant &&
6269 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6270 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6271 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6272 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6273 bool Invert = (CC == ISD::SETNE) ^
6274 cast<ConstantSDNode>(Op1)->isNullValue();
6276 CCode = X86::GetOppositeBranchCondition(CCode);
6277 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6278 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6281 bool isFP = Op1.getValueType().isFloatingPoint();
6282 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6283 if (X86CC == X86::COND_INVALID)
6286 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6288 // Use sbb x, x to materialize carry bit into a GPR.
6289 if (X86CC == X86::COND_B)
6290 return DAG.getNode(ISD::AND, dl, MVT::i8,
6291 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6292 DAG.getConstant(X86CC, MVT::i8), Cond),
6293 DAG.getConstant(1, MVT::i8));
6295 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6296 DAG.getConstant(X86CC, MVT::i8), Cond);
6299 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6301 SDValue Op0 = Op.getOperand(0);
6302 SDValue Op1 = Op.getOperand(1);
6303 SDValue CC = Op.getOperand(2);
6304 EVT VT = Op.getValueType();
6305 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6306 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6307 DebugLoc dl = Op.getDebugLoc();
6311 EVT VT0 = Op0.getValueType();
6312 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6313 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6316 switch (SetCCOpcode) {
6319 case ISD::SETEQ: SSECC = 0; break;
6321 case ISD::SETGT: Swap = true; // Fallthrough
6323 case ISD::SETOLT: SSECC = 1; break;
6325 case ISD::SETGE: Swap = true; // Fallthrough
6327 case ISD::SETOLE: SSECC = 2; break;
6328 case ISD::SETUO: SSECC = 3; break;
6330 case ISD::SETNE: SSECC = 4; break;
6331 case ISD::SETULE: Swap = true;
6332 case ISD::SETUGE: SSECC = 5; break;
6333 case ISD::SETULT: Swap = true;
6334 case ISD::SETUGT: SSECC = 6; break;
6335 case ISD::SETO: SSECC = 7; break;
6338 std::swap(Op0, Op1);
6340 // In the two special cases we can't handle, emit two comparisons.
6342 if (SetCCOpcode == ISD::SETUEQ) {
6344 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6345 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6346 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6348 else if (SetCCOpcode == ISD::SETONE) {
6350 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6351 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6352 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6354 llvm_unreachable("Illegal FP comparison");
6356 // Handle all other FP comparisons here.
6357 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6360 // We are handling one of the integer comparisons here. Since SSE only has
6361 // GT and EQ comparisons for integer, swapping operands and multiple
6362 // operations may be required for some comparisons.
6363 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6364 bool Swap = false, Invert = false, FlipSigns = false;
6366 switch (VT.getSimpleVT().SimpleTy) {
6369 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6371 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6373 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6374 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6377 switch (SetCCOpcode) {
6379 case ISD::SETNE: Invert = true;
6380 case ISD::SETEQ: Opc = EQOpc; break;
6381 case ISD::SETLT: Swap = true;
6382 case ISD::SETGT: Opc = GTOpc; break;
6383 case ISD::SETGE: Swap = true;
6384 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6385 case ISD::SETULT: Swap = true;
6386 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6387 case ISD::SETUGE: Swap = true;
6388 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6391 std::swap(Op0, Op1);
6393 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6394 // bits of the inputs before performing those operations.
6396 EVT EltVT = VT.getVectorElementType();
6397 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6399 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6400 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6402 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6403 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6406 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6408 // If the logical-not of the result is required, perform that now.
6410 Result = DAG.getNOT(dl, Result, VT);
6415 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6416 static bool isX86LogicalCmp(SDValue Op) {
6417 unsigned Opc = Op.getNode()->getOpcode();
6418 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6420 if (Op.getResNo() == 1 &&
6421 (Opc == X86ISD::ADD ||
6422 Opc == X86ISD::SUB ||
6423 Opc == X86ISD::SMUL ||
6424 Opc == X86ISD::UMUL ||
6425 Opc == X86ISD::INC ||
6426 Opc == X86ISD::DEC ||
6427 Opc == X86ISD::OR ||
6428 Opc == X86ISD::XOR ||
6429 Opc == X86ISD::AND))
6435 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6436 bool addTest = true;
6437 SDValue Cond = Op.getOperand(0);
6438 DebugLoc dl = Op.getDebugLoc();
6441 if (Cond.getOpcode() == ISD::SETCC) {
6442 SDValue NewCond = LowerSETCC(Cond, DAG);
6443 if (NewCond.getNode())
6447 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6448 SDValue Op1 = Op.getOperand(1);
6449 SDValue Op2 = Op.getOperand(2);
6450 if (Cond.getOpcode() == X86ISD::SETCC &&
6451 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6452 SDValue Cmp = Cond.getOperand(1);
6453 if (Cmp.getOpcode() == X86ISD::CMP) {
6454 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6455 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6456 ConstantSDNode *RHSC =
6457 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6458 if (N1C && N1C->isAllOnesValue() &&
6459 N2C && N2C->isNullValue() &&
6460 RHSC && RHSC->isNullValue()) {
6461 SDValue CmpOp0 = Cmp.getOperand(0);
6462 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6463 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6464 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6465 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6470 // Look pass (and (setcc_carry (cmp ...)), 1).
6471 if (Cond.getOpcode() == ISD::AND &&
6472 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6473 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6474 if (C && C->getAPIntValue() == 1)
6475 Cond = Cond.getOperand(0);
6478 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6479 // setting operand in place of the X86ISD::SETCC.
6480 if (Cond.getOpcode() == X86ISD::SETCC ||
6481 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6482 CC = Cond.getOperand(0);
6484 SDValue Cmp = Cond.getOperand(1);
6485 unsigned Opc = Cmp.getOpcode();
6486 EVT VT = Op.getValueType();
6488 bool IllegalFPCMov = false;
6489 if (VT.isFloatingPoint() && !VT.isVector() &&
6490 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6491 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6493 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6494 Opc == X86ISD::BT) { // FIXME
6501 // Look pass the truncate.
6502 if (Cond.getOpcode() == ISD::TRUNCATE)
6503 Cond = Cond.getOperand(0);
6505 // We know the result of AND is compared against zero. Try to match
6507 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6508 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6509 if (NewSetCC.getNode()) {
6510 CC = NewSetCC.getOperand(0);
6511 Cond = NewSetCC.getOperand(1);
6518 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6519 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6522 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6523 // condition is true.
6524 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6525 SDValue Ops[] = { Op2, Op1, CC, Cond };
6526 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6529 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6530 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6531 // from the AND / OR.
6532 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6533 Opc = Op.getOpcode();
6534 if (Opc != ISD::OR && Opc != ISD::AND)
6536 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6537 Op.getOperand(0).hasOneUse() &&
6538 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6539 Op.getOperand(1).hasOneUse());
6542 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6543 // 1 and that the SETCC node has a single use.
6544 static bool isXor1OfSetCC(SDValue Op) {
6545 if (Op.getOpcode() != ISD::XOR)
6547 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6548 if (N1C && N1C->getAPIntValue() == 1) {
6549 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6550 Op.getOperand(0).hasOneUse();
6555 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6556 bool addTest = true;
6557 SDValue Chain = Op.getOperand(0);
6558 SDValue Cond = Op.getOperand(1);
6559 SDValue Dest = Op.getOperand(2);
6560 DebugLoc dl = Op.getDebugLoc();
6563 if (Cond.getOpcode() == ISD::SETCC) {
6564 SDValue NewCond = LowerSETCC(Cond, DAG);
6565 if (NewCond.getNode())
6569 // FIXME: LowerXALUO doesn't handle these!!
6570 else if (Cond.getOpcode() == X86ISD::ADD ||
6571 Cond.getOpcode() == X86ISD::SUB ||
6572 Cond.getOpcode() == X86ISD::SMUL ||
6573 Cond.getOpcode() == X86ISD::UMUL)
6574 Cond = LowerXALUO(Cond, DAG);
6577 // Look pass (and (setcc_carry (cmp ...)), 1).
6578 if (Cond.getOpcode() == ISD::AND &&
6579 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6580 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6581 if (C && C->getAPIntValue() == 1)
6582 Cond = Cond.getOperand(0);
6585 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6586 // setting operand in place of the X86ISD::SETCC.
6587 if (Cond.getOpcode() == X86ISD::SETCC ||
6588 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6589 CC = Cond.getOperand(0);
6591 SDValue Cmp = Cond.getOperand(1);
6592 unsigned Opc = Cmp.getOpcode();
6593 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6594 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6598 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6602 // These can only come from an arithmetic instruction with overflow,
6603 // e.g. SADDO, UADDO.
6604 Cond = Cond.getNode()->getOperand(1);
6611 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6612 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6613 if (CondOpc == ISD::OR) {
6614 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6615 // two branches instead of an explicit OR instruction with a
6617 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6618 isX86LogicalCmp(Cmp)) {
6619 CC = Cond.getOperand(0).getOperand(0);
6620 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6621 Chain, Dest, CC, Cmp);
6622 CC = Cond.getOperand(1).getOperand(0);
6626 } else { // ISD::AND
6627 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6628 // two branches instead of an explicit AND instruction with a
6629 // separate test. However, we only do this if this block doesn't
6630 // have a fall-through edge, because this requires an explicit
6631 // jmp when the condition is false.
6632 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6633 isX86LogicalCmp(Cmp) &&
6634 Op.getNode()->hasOneUse()) {
6635 X86::CondCode CCode =
6636 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6637 CCode = X86::GetOppositeBranchCondition(CCode);
6638 CC = DAG.getConstant(CCode, MVT::i8);
6639 SDNode *User = *Op.getNode()->use_begin();
6640 // Look for an unconditional branch following this conditional branch.
6641 // We need this because we need to reverse the successors in order
6642 // to implement FCMP_OEQ.
6643 if (User->getOpcode() == ISD::BR) {
6644 SDValue FalseBB = User->getOperand(1);
6646 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6647 assert(NewBR == User);
6651 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6652 Chain, Dest, CC, Cmp);
6653 X86::CondCode CCode =
6654 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6655 CCode = X86::GetOppositeBranchCondition(CCode);
6656 CC = DAG.getConstant(CCode, MVT::i8);
6662 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6663 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6664 // It should be transformed during dag combiner except when the condition
6665 // is set by a arithmetics with overflow node.
6666 X86::CondCode CCode =
6667 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6668 CCode = X86::GetOppositeBranchCondition(CCode);
6669 CC = DAG.getConstant(CCode, MVT::i8);
6670 Cond = Cond.getOperand(0).getOperand(1);
6676 // Look pass the truncate.
6677 if (Cond.getOpcode() == ISD::TRUNCATE)
6678 Cond = Cond.getOperand(0);
6680 // We know the result of AND is compared against zero. Try to match
6682 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6683 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6684 if (NewSetCC.getNode()) {
6685 CC = NewSetCC.getOperand(0);
6686 Cond = NewSetCC.getOperand(1);
6693 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6694 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6696 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6697 Chain, Dest, CC, Cond);
6701 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6702 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6703 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6704 // that the guard pages used by the OS virtual memory manager are allocated in
6705 // correct sequence.
6707 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6708 SelectionDAG &DAG) const {
6709 assert(Subtarget->isTargetCygMing() &&
6710 "This should be used only on Cygwin/Mingw targets");
6711 DebugLoc dl = Op.getDebugLoc();
6714 SDValue Chain = Op.getOperand(0);
6715 SDValue Size = Op.getOperand(1);
6716 // FIXME: Ensure alignment here
6720 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6722 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6723 Flag = Chain.getValue(1);
6725 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6727 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6728 Flag = Chain.getValue(1);
6730 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6732 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6733 return DAG.getMergeValues(Ops1, 2, dl);
6736 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6737 MachineFunction &MF = DAG.getMachineFunction();
6738 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6740 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6741 DebugLoc dl = Op.getDebugLoc();
6743 if (!Subtarget->is64Bit()) {
6744 // vastart just stores the address of the VarArgsFrameIndex slot into the
6745 // memory location argument.
6746 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6748 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6753 // gp_offset (0 - 6 * 8)
6754 // fp_offset (48 - 48 + 8 * 16)
6755 // overflow_arg_area (point to parameters coming in memory).
6757 SmallVector<SDValue, 8> MemOps;
6758 SDValue FIN = Op.getOperand(1);
6760 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6761 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6763 FIN, SV, 0, false, false, 0);
6764 MemOps.push_back(Store);
6767 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6768 FIN, DAG.getIntPtrConstant(4));
6769 Store = DAG.getStore(Op.getOperand(0), dl,
6770 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6772 FIN, SV, 4, false, false, 0);
6773 MemOps.push_back(Store);
6775 // Store ptr to overflow_arg_area
6776 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6777 FIN, DAG.getIntPtrConstant(4));
6778 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6780 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 8,
6782 MemOps.push_back(Store);
6784 // Store ptr to reg_save_area.
6785 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6786 FIN, DAG.getIntPtrConstant(8));
6787 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6789 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 16,
6791 MemOps.push_back(Store);
6792 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6793 &MemOps[0], MemOps.size());
6796 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6797 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6798 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6800 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6804 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6805 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6806 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6807 SDValue Chain = Op.getOperand(0);
6808 SDValue DstPtr = Op.getOperand(1);
6809 SDValue SrcPtr = Op.getOperand(2);
6810 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6811 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6812 DebugLoc dl = Op.getDebugLoc();
6814 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6815 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6816 false, DstSV, 0, SrcSV, 0);
6820 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6821 DebugLoc dl = Op.getDebugLoc();
6822 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6824 default: return SDValue(); // Don't custom lower most intrinsics.
6825 // Comparison intrinsics.
6826 case Intrinsic::x86_sse_comieq_ss:
6827 case Intrinsic::x86_sse_comilt_ss:
6828 case Intrinsic::x86_sse_comile_ss:
6829 case Intrinsic::x86_sse_comigt_ss:
6830 case Intrinsic::x86_sse_comige_ss:
6831 case Intrinsic::x86_sse_comineq_ss:
6832 case Intrinsic::x86_sse_ucomieq_ss:
6833 case Intrinsic::x86_sse_ucomilt_ss:
6834 case Intrinsic::x86_sse_ucomile_ss:
6835 case Intrinsic::x86_sse_ucomigt_ss:
6836 case Intrinsic::x86_sse_ucomige_ss:
6837 case Intrinsic::x86_sse_ucomineq_ss:
6838 case Intrinsic::x86_sse2_comieq_sd:
6839 case Intrinsic::x86_sse2_comilt_sd:
6840 case Intrinsic::x86_sse2_comile_sd:
6841 case Intrinsic::x86_sse2_comigt_sd:
6842 case Intrinsic::x86_sse2_comige_sd:
6843 case Intrinsic::x86_sse2_comineq_sd:
6844 case Intrinsic::x86_sse2_ucomieq_sd:
6845 case Intrinsic::x86_sse2_ucomilt_sd:
6846 case Intrinsic::x86_sse2_ucomile_sd:
6847 case Intrinsic::x86_sse2_ucomigt_sd:
6848 case Intrinsic::x86_sse2_ucomige_sd:
6849 case Intrinsic::x86_sse2_ucomineq_sd: {
6851 ISD::CondCode CC = ISD::SETCC_INVALID;
6854 case Intrinsic::x86_sse_comieq_ss:
6855 case Intrinsic::x86_sse2_comieq_sd:
6859 case Intrinsic::x86_sse_comilt_ss:
6860 case Intrinsic::x86_sse2_comilt_sd:
6864 case Intrinsic::x86_sse_comile_ss:
6865 case Intrinsic::x86_sse2_comile_sd:
6869 case Intrinsic::x86_sse_comigt_ss:
6870 case Intrinsic::x86_sse2_comigt_sd:
6874 case Intrinsic::x86_sse_comige_ss:
6875 case Intrinsic::x86_sse2_comige_sd:
6879 case Intrinsic::x86_sse_comineq_ss:
6880 case Intrinsic::x86_sse2_comineq_sd:
6884 case Intrinsic::x86_sse_ucomieq_ss:
6885 case Intrinsic::x86_sse2_ucomieq_sd:
6886 Opc = X86ISD::UCOMI;
6889 case Intrinsic::x86_sse_ucomilt_ss:
6890 case Intrinsic::x86_sse2_ucomilt_sd:
6891 Opc = X86ISD::UCOMI;
6894 case Intrinsic::x86_sse_ucomile_ss:
6895 case Intrinsic::x86_sse2_ucomile_sd:
6896 Opc = X86ISD::UCOMI;
6899 case Intrinsic::x86_sse_ucomigt_ss:
6900 case Intrinsic::x86_sse2_ucomigt_sd:
6901 Opc = X86ISD::UCOMI;
6904 case Intrinsic::x86_sse_ucomige_ss:
6905 case Intrinsic::x86_sse2_ucomige_sd:
6906 Opc = X86ISD::UCOMI;
6909 case Intrinsic::x86_sse_ucomineq_ss:
6910 case Intrinsic::x86_sse2_ucomineq_sd:
6911 Opc = X86ISD::UCOMI;
6916 SDValue LHS = Op.getOperand(1);
6917 SDValue RHS = Op.getOperand(2);
6918 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6919 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6920 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6921 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6922 DAG.getConstant(X86CC, MVT::i8), Cond);
6923 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6925 // ptest intrinsics. The intrinsic these come from are designed to return
6926 // an integer value, not just an instruction so lower it to the ptest
6927 // pattern and a setcc for the result.
6928 case Intrinsic::x86_sse41_ptestz:
6929 case Intrinsic::x86_sse41_ptestc:
6930 case Intrinsic::x86_sse41_ptestnzc:{
6933 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6934 case Intrinsic::x86_sse41_ptestz:
6936 X86CC = X86::COND_E;
6938 case Intrinsic::x86_sse41_ptestc:
6940 X86CC = X86::COND_B;
6942 case Intrinsic::x86_sse41_ptestnzc:
6944 X86CC = X86::COND_A;
6948 SDValue LHS = Op.getOperand(1);
6949 SDValue RHS = Op.getOperand(2);
6950 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6951 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6952 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6953 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6956 // Fix vector shift instructions where the last operand is a non-immediate
6958 case Intrinsic::x86_sse2_pslli_w:
6959 case Intrinsic::x86_sse2_pslli_d:
6960 case Intrinsic::x86_sse2_pslli_q:
6961 case Intrinsic::x86_sse2_psrli_w:
6962 case Intrinsic::x86_sse2_psrli_d:
6963 case Intrinsic::x86_sse2_psrli_q:
6964 case Intrinsic::x86_sse2_psrai_w:
6965 case Intrinsic::x86_sse2_psrai_d:
6966 case Intrinsic::x86_mmx_pslli_w:
6967 case Intrinsic::x86_mmx_pslli_d:
6968 case Intrinsic::x86_mmx_pslli_q:
6969 case Intrinsic::x86_mmx_psrli_w:
6970 case Intrinsic::x86_mmx_psrli_d:
6971 case Intrinsic::x86_mmx_psrli_q:
6972 case Intrinsic::x86_mmx_psrai_w:
6973 case Intrinsic::x86_mmx_psrai_d: {
6974 SDValue ShAmt = Op.getOperand(2);
6975 if (isa<ConstantSDNode>(ShAmt))
6978 unsigned NewIntNo = 0;
6979 EVT ShAmtVT = MVT::v4i32;
6981 case Intrinsic::x86_sse2_pslli_w:
6982 NewIntNo = Intrinsic::x86_sse2_psll_w;
6984 case Intrinsic::x86_sse2_pslli_d:
6985 NewIntNo = Intrinsic::x86_sse2_psll_d;
6987 case Intrinsic::x86_sse2_pslli_q:
6988 NewIntNo = Intrinsic::x86_sse2_psll_q;
6990 case Intrinsic::x86_sse2_psrli_w:
6991 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6993 case Intrinsic::x86_sse2_psrli_d:
6994 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6996 case Intrinsic::x86_sse2_psrli_q:
6997 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6999 case Intrinsic::x86_sse2_psrai_w:
7000 NewIntNo = Intrinsic::x86_sse2_psra_w;
7002 case Intrinsic::x86_sse2_psrai_d:
7003 NewIntNo = Intrinsic::x86_sse2_psra_d;
7006 ShAmtVT = MVT::v2i32;
7008 case Intrinsic::x86_mmx_pslli_w:
7009 NewIntNo = Intrinsic::x86_mmx_psll_w;
7011 case Intrinsic::x86_mmx_pslli_d:
7012 NewIntNo = Intrinsic::x86_mmx_psll_d;
7014 case Intrinsic::x86_mmx_pslli_q:
7015 NewIntNo = Intrinsic::x86_mmx_psll_q;
7017 case Intrinsic::x86_mmx_psrli_w:
7018 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7020 case Intrinsic::x86_mmx_psrli_d:
7021 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7023 case Intrinsic::x86_mmx_psrli_q:
7024 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7026 case Intrinsic::x86_mmx_psrai_w:
7027 NewIntNo = Intrinsic::x86_mmx_psra_w;
7029 case Intrinsic::x86_mmx_psrai_d:
7030 NewIntNo = Intrinsic::x86_mmx_psra_d;
7032 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7038 // The vector shift intrinsics with scalars uses 32b shift amounts but
7039 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7043 ShOps[1] = DAG.getConstant(0, MVT::i32);
7044 if (ShAmtVT == MVT::v4i32) {
7045 ShOps[2] = DAG.getUNDEF(MVT::i32);
7046 ShOps[3] = DAG.getUNDEF(MVT::i32);
7047 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7049 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7052 EVT VT = Op.getValueType();
7053 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7054 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7055 DAG.getConstant(NewIntNo, MVT::i32),
7056 Op.getOperand(1), ShAmt);
7061 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7062 SelectionDAG &DAG) const {
7063 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7064 MFI->setReturnAddressIsTaken(true);
7066 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7067 DebugLoc dl = Op.getDebugLoc();
7070 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7072 DAG.getConstant(TD->getPointerSize(),
7073 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7074 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7075 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7077 NULL, 0, false, false, 0);
7080 // Just load the return address.
7081 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7082 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7083 RetAddrFI, NULL, 0, false, false, 0);
7086 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7087 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7088 MFI->setFrameAddressIsTaken(true);
7090 EVT VT = Op.getValueType();
7091 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7092 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7093 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7094 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7096 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7101 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7102 SelectionDAG &DAG) const {
7103 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7106 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7107 MachineFunction &MF = DAG.getMachineFunction();
7108 SDValue Chain = Op.getOperand(0);
7109 SDValue Offset = Op.getOperand(1);
7110 SDValue Handler = Op.getOperand(2);
7111 DebugLoc dl = Op.getDebugLoc();
7113 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7115 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7117 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7118 DAG.getIntPtrConstant(-TD->getPointerSize()));
7119 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7120 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7121 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7122 MF.getRegInfo().addLiveOut(StoreAddrReg);
7124 return DAG.getNode(X86ISD::EH_RETURN, dl,
7126 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7129 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7130 SelectionDAG &DAG) const {
7131 SDValue Root = Op.getOperand(0);
7132 SDValue Trmp = Op.getOperand(1); // trampoline
7133 SDValue FPtr = Op.getOperand(2); // nested function
7134 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7135 DebugLoc dl = Op.getDebugLoc();
7137 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7139 if (Subtarget->is64Bit()) {
7140 SDValue OutChains[6];
7142 // Large code-model.
7143 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7144 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7146 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7147 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7149 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7151 // Load the pointer to the nested function into R11.
7152 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7153 SDValue Addr = Trmp;
7154 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7155 Addr, TrmpAddr, 0, false, false, 0);
7157 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7158 DAG.getConstant(2, MVT::i64));
7159 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7162 // Load the 'nest' parameter value into R10.
7163 // R10 is specified in X86CallingConv.td
7164 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7165 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7166 DAG.getConstant(10, MVT::i64));
7167 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7168 Addr, TrmpAddr, 10, false, false, 0);
7170 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7171 DAG.getConstant(12, MVT::i64));
7172 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7175 // Jump to the nested function.
7176 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7177 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7178 DAG.getConstant(20, MVT::i64));
7179 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7180 Addr, TrmpAddr, 20, false, false, 0);
7182 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7183 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7184 DAG.getConstant(22, MVT::i64));
7185 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7186 TrmpAddr, 22, false, false, 0);
7189 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7190 return DAG.getMergeValues(Ops, 2, dl);
7192 const Function *Func =
7193 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7194 CallingConv::ID CC = Func->getCallingConv();
7199 llvm_unreachable("Unsupported calling convention");
7200 case CallingConv::C:
7201 case CallingConv::X86_StdCall: {
7202 // Pass 'nest' parameter in ECX.
7203 // Must be kept in sync with X86CallingConv.td
7206 // Check that ECX wasn't needed by an 'inreg' parameter.
7207 const FunctionType *FTy = Func->getFunctionType();
7208 const AttrListPtr &Attrs = Func->getAttributes();
7210 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7211 unsigned InRegCount = 0;
7214 for (FunctionType::param_iterator I = FTy->param_begin(),
7215 E = FTy->param_end(); I != E; ++I, ++Idx)
7216 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7217 // FIXME: should only count parameters that are lowered to integers.
7218 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7220 if (InRegCount > 2) {
7221 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7226 case CallingConv::X86_FastCall:
7227 case CallingConv::X86_ThisCall:
7228 case CallingConv::Fast:
7229 // Pass 'nest' parameter in EAX.
7230 // Must be kept in sync with X86CallingConv.td
7235 SDValue OutChains[4];
7238 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7239 DAG.getConstant(10, MVT::i32));
7240 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7242 // This is storing the opcode for MOV32ri.
7243 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7244 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7245 OutChains[0] = DAG.getStore(Root, dl,
7246 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7247 Trmp, TrmpAddr, 0, false, false, 0);
7249 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7250 DAG.getConstant(1, MVT::i32));
7251 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7254 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7255 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7256 DAG.getConstant(5, MVT::i32));
7257 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7258 TrmpAddr, 5, false, false, 1);
7260 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7261 DAG.getConstant(6, MVT::i32));
7262 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7266 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7267 return DAG.getMergeValues(Ops, 2, dl);
7271 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7272 SelectionDAG &DAG) const {
7274 The rounding mode is in bits 11:10 of FPSR, and has the following
7281 FLT_ROUNDS, on the other hand, expects the following:
7288 To perform the conversion, we do:
7289 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7292 MachineFunction &MF = DAG.getMachineFunction();
7293 const TargetMachine &TM = MF.getTarget();
7294 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7295 unsigned StackAlignment = TFI.getStackAlignment();
7296 EVT VT = Op.getValueType();
7297 DebugLoc dl = Op.getDebugLoc();
7299 // Save FP Control Word to stack slot
7300 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7301 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7303 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7304 DAG.getEntryNode(), StackSlot);
7306 // Load FP Control Word from stack slot
7307 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7310 // Transform as necessary
7312 DAG.getNode(ISD::SRL, dl, MVT::i16,
7313 DAG.getNode(ISD::AND, dl, MVT::i16,
7314 CWD, DAG.getConstant(0x800, MVT::i16)),
7315 DAG.getConstant(11, MVT::i8));
7317 DAG.getNode(ISD::SRL, dl, MVT::i16,
7318 DAG.getNode(ISD::AND, dl, MVT::i16,
7319 CWD, DAG.getConstant(0x400, MVT::i16)),
7320 DAG.getConstant(9, MVT::i8));
7323 DAG.getNode(ISD::AND, dl, MVT::i16,
7324 DAG.getNode(ISD::ADD, dl, MVT::i16,
7325 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7326 DAG.getConstant(1, MVT::i16)),
7327 DAG.getConstant(3, MVT::i16));
7330 return DAG.getNode((VT.getSizeInBits() < 16 ?
7331 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7334 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7335 EVT VT = Op.getValueType();
7337 unsigned NumBits = VT.getSizeInBits();
7338 DebugLoc dl = Op.getDebugLoc();
7340 Op = Op.getOperand(0);
7341 if (VT == MVT::i8) {
7342 // Zero extend to i32 since there is not an i8 bsr.
7344 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7347 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7348 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7349 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7351 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7354 DAG.getConstant(NumBits+NumBits-1, OpVT),
7355 DAG.getConstant(X86::COND_E, MVT::i8),
7358 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7360 // Finally xor with NumBits-1.
7361 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7364 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7368 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7369 EVT VT = Op.getValueType();
7371 unsigned NumBits = VT.getSizeInBits();
7372 DebugLoc dl = Op.getDebugLoc();
7374 Op = Op.getOperand(0);
7375 if (VT == MVT::i8) {
7377 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7380 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7381 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7382 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7384 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7387 DAG.getConstant(NumBits, OpVT),
7388 DAG.getConstant(X86::COND_E, MVT::i8),
7391 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7394 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7398 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7399 EVT VT = Op.getValueType();
7400 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7401 DebugLoc dl = Op.getDebugLoc();
7403 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7404 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7405 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7406 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7407 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7409 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7410 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7411 // return AloBlo + AloBhi + AhiBlo;
7413 SDValue A = Op.getOperand(0);
7414 SDValue B = Op.getOperand(1);
7416 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7417 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7418 A, DAG.getConstant(32, MVT::i32));
7419 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7420 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7421 B, DAG.getConstant(32, MVT::i32));
7422 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7423 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7425 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7426 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7428 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7429 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7431 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7432 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7433 AloBhi, DAG.getConstant(32, MVT::i32));
7434 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7435 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7436 AhiBlo, DAG.getConstant(32, MVT::i32));
7437 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7438 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7443 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7444 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7445 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7446 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7447 // has only one use.
7448 SDNode *N = Op.getNode();
7449 SDValue LHS = N->getOperand(0);
7450 SDValue RHS = N->getOperand(1);
7451 unsigned BaseOp = 0;
7453 DebugLoc dl = Op.getDebugLoc();
7455 switch (Op.getOpcode()) {
7456 default: llvm_unreachable("Unknown ovf instruction!");
7458 // A subtract of one will be selected as a INC. Note that INC doesn't
7459 // set CF, so we can't do this for UADDO.
7460 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7461 if (C->getAPIntValue() == 1) {
7462 BaseOp = X86ISD::INC;
7466 BaseOp = X86ISD::ADD;
7470 BaseOp = X86ISD::ADD;
7474 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7475 // set CF, so we can't do this for USUBO.
7476 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7477 if (C->getAPIntValue() == 1) {
7478 BaseOp = X86ISD::DEC;
7482 BaseOp = X86ISD::SUB;
7486 BaseOp = X86ISD::SUB;
7490 BaseOp = X86ISD::SMUL;
7494 BaseOp = X86ISD::UMUL;
7499 // Also sets EFLAGS.
7500 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7501 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7504 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7505 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7507 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7511 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7512 EVT T = Op.getValueType();
7513 DebugLoc dl = Op.getDebugLoc();
7516 switch(T.getSimpleVT().SimpleTy) {
7518 assert(false && "Invalid value type!");
7519 case MVT::i8: Reg = X86::AL; size = 1; break;
7520 case MVT::i16: Reg = X86::AX; size = 2; break;
7521 case MVT::i32: Reg = X86::EAX; size = 4; break;
7523 assert(Subtarget->is64Bit() && "Node not type legal!");
7524 Reg = X86::RAX; size = 8;
7527 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7528 Op.getOperand(2), SDValue());
7529 SDValue Ops[] = { cpIn.getValue(0),
7532 DAG.getTargetConstant(size, MVT::i8),
7534 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7535 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7537 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7541 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7542 SelectionDAG &DAG) const {
7543 assert(Subtarget->is64Bit() && "Result not type legalized?");
7544 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7545 SDValue TheChain = Op.getOperand(0);
7546 DebugLoc dl = Op.getDebugLoc();
7547 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7548 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7549 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7551 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7552 DAG.getConstant(32, MVT::i8));
7554 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7557 return DAG.getMergeValues(Ops, 2, dl);
7560 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7561 SelectionDAG &DAG) const {
7562 EVT SrcVT = Op.getOperand(0).getValueType();
7563 EVT DstVT = Op.getValueType();
7564 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7565 Subtarget->hasMMX() && !DisableMMX) &&
7566 "Unexpected custom BIT_CONVERT");
7567 assert((DstVT == MVT::i64 ||
7568 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7569 "Unexpected custom BIT_CONVERT");
7570 // i64 <=> MMX conversions are Legal.
7571 if (SrcVT==MVT::i64 && DstVT.isVector())
7573 if (DstVT==MVT::i64 && SrcVT.isVector())
7575 // MMX <=> MMX conversions are Legal.
7576 if (SrcVT.isVector() && DstVT.isVector())
7578 // All other conversions need to be expanded.
7581 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7582 SDNode *Node = Op.getNode();
7583 DebugLoc dl = Node->getDebugLoc();
7584 EVT T = Node->getValueType(0);
7585 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7586 DAG.getConstant(0, T), Node->getOperand(2));
7587 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7588 cast<AtomicSDNode>(Node)->getMemoryVT(),
7589 Node->getOperand(0),
7590 Node->getOperand(1), negOp,
7591 cast<AtomicSDNode>(Node)->getSrcValue(),
7592 cast<AtomicSDNode>(Node)->getAlignment());
7595 /// LowerOperation - Provide custom lowering hooks for some operations.
7597 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7598 switch (Op.getOpcode()) {
7599 default: llvm_unreachable("Should not custom lower this!");
7600 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7601 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7602 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7603 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7604 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7605 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7606 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7607 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7608 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7609 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7610 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7611 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7612 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7613 case ISD::SHL_PARTS:
7614 case ISD::SRA_PARTS:
7615 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7616 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7617 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7618 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7619 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7620 case ISD::FABS: return LowerFABS(Op, DAG);
7621 case ISD::FNEG: return LowerFNEG(Op, DAG);
7622 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7623 case ISD::SETCC: return LowerSETCC(Op, DAG);
7624 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7625 case ISD::SELECT: return LowerSELECT(Op, DAG);
7626 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7627 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7628 case ISD::VASTART: return LowerVASTART(Op, DAG);
7629 case ISD::VAARG: return LowerVAARG(Op, DAG);
7630 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7631 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7632 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7633 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7634 case ISD::FRAME_TO_ARGS_OFFSET:
7635 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7636 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7637 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7638 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7639 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7640 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7641 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7642 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7648 case ISD::UMULO: return LowerXALUO(Op, DAG);
7649 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7650 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7654 void X86TargetLowering::
7655 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7656 SelectionDAG &DAG, unsigned NewOp) const {
7657 EVT T = Node->getValueType(0);
7658 DebugLoc dl = Node->getDebugLoc();
7659 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7661 SDValue Chain = Node->getOperand(0);
7662 SDValue In1 = Node->getOperand(1);
7663 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7664 Node->getOperand(2), DAG.getIntPtrConstant(0));
7665 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7666 Node->getOperand(2), DAG.getIntPtrConstant(1));
7667 SDValue Ops[] = { Chain, In1, In2L, In2H };
7668 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7670 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7671 cast<MemSDNode>(Node)->getMemOperand());
7672 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7673 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7674 Results.push_back(Result.getValue(2));
7677 /// ReplaceNodeResults - Replace a node with an illegal result type
7678 /// with a new node built out of custom code.
7679 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7680 SmallVectorImpl<SDValue>&Results,
7681 SelectionDAG &DAG) const {
7682 DebugLoc dl = N->getDebugLoc();
7683 switch (N->getOpcode()) {
7685 assert(false && "Do not know how to custom type legalize this operation!");
7687 case ISD::FP_TO_SINT: {
7688 std::pair<SDValue,SDValue> Vals =
7689 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7690 SDValue FIST = Vals.first, StackSlot = Vals.second;
7691 if (FIST.getNode() != 0) {
7692 EVT VT = N->getValueType(0);
7693 // Return a load from the stack slot.
7694 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7699 case ISD::READCYCLECOUNTER: {
7700 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7701 SDValue TheChain = N->getOperand(0);
7702 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7703 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7705 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7707 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7708 SDValue Ops[] = { eax, edx };
7709 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7710 Results.push_back(edx.getValue(1));
7713 case ISD::ATOMIC_CMP_SWAP: {
7714 EVT T = N->getValueType(0);
7715 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7716 SDValue cpInL, cpInH;
7717 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7718 DAG.getConstant(0, MVT::i32));
7719 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7720 DAG.getConstant(1, MVT::i32));
7721 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7722 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7724 SDValue swapInL, swapInH;
7725 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7726 DAG.getConstant(0, MVT::i32));
7727 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7728 DAG.getConstant(1, MVT::i32));
7729 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7731 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7732 swapInL.getValue(1));
7733 SDValue Ops[] = { swapInH.getValue(0),
7735 swapInH.getValue(1) };
7736 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7737 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7738 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7739 MVT::i32, Result.getValue(1));
7740 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7741 MVT::i32, cpOutL.getValue(2));
7742 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7743 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7744 Results.push_back(cpOutH.getValue(1));
7747 case ISD::ATOMIC_LOAD_ADD:
7748 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7750 case ISD::ATOMIC_LOAD_AND:
7751 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7753 case ISD::ATOMIC_LOAD_NAND:
7754 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7756 case ISD::ATOMIC_LOAD_OR:
7757 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7759 case ISD::ATOMIC_LOAD_SUB:
7760 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7762 case ISD::ATOMIC_LOAD_XOR:
7763 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7765 case ISD::ATOMIC_SWAP:
7766 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7771 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7773 default: return NULL;
7774 case X86ISD::BSF: return "X86ISD::BSF";
7775 case X86ISD::BSR: return "X86ISD::BSR";
7776 case X86ISD::SHLD: return "X86ISD::SHLD";
7777 case X86ISD::SHRD: return "X86ISD::SHRD";
7778 case X86ISD::FAND: return "X86ISD::FAND";
7779 case X86ISD::FOR: return "X86ISD::FOR";
7780 case X86ISD::FXOR: return "X86ISD::FXOR";
7781 case X86ISD::FSRL: return "X86ISD::FSRL";
7782 case X86ISD::FILD: return "X86ISD::FILD";
7783 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7784 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7785 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7786 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7787 case X86ISD::FLD: return "X86ISD::FLD";
7788 case X86ISD::FST: return "X86ISD::FST";
7789 case X86ISD::CALL: return "X86ISD::CALL";
7790 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7791 case X86ISD::BT: return "X86ISD::BT";
7792 case X86ISD::CMP: return "X86ISD::CMP";
7793 case X86ISD::COMI: return "X86ISD::COMI";
7794 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7795 case X86ISD::SETCC: return "X86ISD::SETCC";
7796 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7797 case X86ISD::CMOV: return "X86ISD::CMOV";
7798 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7799 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7800 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7801 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7802 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7803 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7804 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7805 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7806 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7807 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7808 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7809 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7810 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7811 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7812 case X86ISD::FMAX: return "X86ISD::FMAX";
7813 case X86ISD::FMIN: return "X86ISD::FMIN";
7814 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7815 case X86ISD::FRCP: return "X86ISD::FRCP";
7816 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7817 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
7818 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7819 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7820 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7821 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7822 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7823 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7824 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7825 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7826 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7827 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7828 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7829 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7830 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7831 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7832 case X86ISD::VSHL: return "X86ISD::VSHL";
7833 case X86ISD::VSRL: return "X86ISD::VSRL";
7834 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7835 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7836 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7837 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7838 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7839 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7840 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7841 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7842 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7843 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7844 case X86ISD::ADD: return "X86ISD::ADD";
7845 case X86ISD::SUB: return "X86ISD::SUB";
7846 case X86ISD::SMUL: return "X86ISD::SMUL";
7847 case X86ISD::UMUL: return "X86ISD::UMUL";
7848 case X86ISD::INC: return "X86ISD::INC";
7849 case X86ISD::DEC: return "X86ISD::DEC";
7850 case X86ISD::OR: return "X86ISD::OR";
7851 case X86ISD::XOR: return "X86ISD::XOR";
7852 case X86ISD::AND: return "X86ISD::AND";
7853 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7854 case X86ISD::PTEST: return "X86ISD::PTEST";
7855 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7856 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7860 // isLegalAddressingMode - Return true if the addressing mode represented
7861 // by AM is legal for this target, for a load/store of the specified type.
7862 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7863 const Type *Ty) const {
7864 // X86 supports extremely general addressing modes.
7865 CodeModel::Model M = getTargetMachine().getCodeModel();
7867 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7868 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7873 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7875 // If a reference to this global requires an extra load, we can't fold it.
7876 if (isGlobalStubReference(GVFlags))
7879 // If BaseGV requires a register for the PIC base, we cannot also have a
7880 // BaseReg specified.
7881 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7884 // If lower 4G is not available, then we must use rip-relative addressing.
7885 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7895 // These scales always work.
7900 // These scales are formed with basereg+scalereg. Only accept if there is
7905 default: // Other stuff never works.
7913 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7914 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7916 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7917 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7918 if (NumBits1 <= NumBits2)
7923 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7924 if (!VT1.isInteger() || !VT2.isInteger())
7926 unsigned NumBits1 = VT1.getSizeInBits();
7927 unsigned NumBits2 = VT2.getSizeInBits();
7928 if (NumBits1 <= NumBits2)
7933 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7934 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7935 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7938 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7939 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7940 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7943 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7944 // i16 instructions are longer (0x66 prefix) and potentially slower.
7945 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7948 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7949 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7950 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7951 /// are assumed to be legal.
7953 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7955 // Very little shuffling can be done for 64-bit vectors right now.
7956 if (VT.getSizeInBits() == 64)
7957 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
7959 // FIXME: pshufb, blends, shifts.
7960 return (VT.getVectorNumElements() == 2 ||
7961 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7962 isMOVLMask(M, VT) ||
7963 isSHUFPMask(M, VT) ||
7964 isPSHUFDMask(M, VT) ||
7965 isPSHUFHWMask(M, VT) ||
7966 isPSHUFLWMask(M, VT) ||
7967 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7968 isUNPCKLMask(M, VT) ||
7969 isUNPCKHMask(M, VT) ||
7970 isUNPCKL_v_undef_Mask(M, VT) ||
7971 isUNPCKH_v_undef_Mask(M, VT));
7975 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7977 unsigned NumElts = VT.getVectorNumElements();
7978 // FIXME: This collection of masks seems suspect.
7981 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7982 return (isMOVLMask(Mask, VT) ||
7983 isCommutedMOVLMask(Mask, VT, true) ||
7984 isSHUFPMask(Mask, VT) ||
7985 isCommutedSHUFPMask(Mask, VT));
7990 //===----------------------------------------------------------------------===//
7991 // X86 Scheduler Hooks
7992 //===----------------------------------------------------------------------===//
7994 // private utility function
7996 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7997 MachineBasicBlock *MBB,
8004 TargetRegisterClass *RC,
8005 bool invSrc) const {
8006 // For the atomic bitwise operator, we generate
8009 // ld t1 = [bitinstr.addr]
8010 // op t2 = t1, [bitinstr.val]
8012 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8014 // fallthrough -->nextMBB
8015 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8016 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8017 MachineFunction::iterator MBBIter = MBB;
8020 /// First build the CFG
8021 MachineFunction *F = MBB->getParent();
8022 MachineBasicBlock *thisMBB = MBB;
8023 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8024 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8025 F->insert(MBBIter, newMBB);
8026 F->insert(MBBIter, nextMBB);
8028 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8029 nextMBB->splice(nextMBB->begin(), thisMBB,
8030 llvm::next(MachineBasicBlock::iterator(bInstr)),
8032 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8034 // Update thisMBB to fall through to newMBB
8035 thisMBB->addSuccessor(newMBB);
8037 // newMBB jumps to itself and fall through to nextMBB
8038 newMBB->addSuccessor(nextMBB);
8039 newMBB->addSuccessor(newMBB);
8041 // Insert instructions into newMBB based on incoming instruction
8042 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8043 "unexpected number of operands");
8044 DebugLoc dl = bInstr->getDebugLoc();
8045 MachineOperand& destOper = bInstr->getOperand(0);
8046 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8047 int numArgs = bInstr->getNumOperands() - 1;
8048 for (int i=0; i < numArgs; ++i)
8049 argOpers[i] = &bInstr->getOperand(i+1);
8051 // x86 address has 4 operands: base, index, scale, and displacement
8052 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8053 int valArgIndx = lastAddrIndx + 1;
8055 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8056 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8057 for (int i=0; i <= lastAddrIndx; ++i)
8058 (*MIB).addOperand(*argOpers[i]);
8060 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8062 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8067 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8068 assert((argOpers[valArgIndx]->isReg() ||
8069 argOpers[valArgIndx]->isImm()) &&
8071 if (argOpers[valArgIndx]->isReg())
8072 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8074 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8076 (*MIB).addOperand(*argOpers[valArgIndx]);
8078 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
8081 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8082 for (int i=0; i <= lastAddrIndx; ++i)
8083 (*MIB).addOperand(*argOpers[i]);
8085 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8086 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8087 bInstr->memoperands_end());
8089 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8093 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8095 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8099 // private utility function: 64 bit atomics on 32 bit host.
8101 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8102 MachineBasicBlock *MBB,
8107 bool invSrc) const {
8108 // For the atomic bitwise operator, we generate
8109 // thisMBB (instructions are in pairs, except cmpxchg8b)
8110 // ld t1,t2 = [bitinstr.addr]
8112 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8113 // op t5, t6 <- out1, out2, [bitinstr.val]
8114 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8115 // mov ECX, EBX <- t5, t6
8116 // mov EAX, EDX <- t1, t2
8117 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8118 // mov t3, t4 <- EAX, EDX
8120 // result in out1, out2
8121 // fallthrough -->nextMBB
8123 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8124 const unsigned LoadOpc = X86::MOV32rm;
8125 const unsigned NotOpc = X86::NOT32r;
8126 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8127 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8128 MachineFunction::iterator MBBIter = MBB;
8131 /// First build the CFG
8132 MachineFunction *F = MBB->getParent();
8133 MachineBasicBlock *thisMBB = MBB;
8134 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8135 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8136 F->insert(MBBIter, newMBB);
8137 F->insert(MBBIter, nextMBB);
8139 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8140 nextMBB->splice(nextMBB->begin(), thisMBB,
8141 llvm::next(MachineBasicBlock::iterator(bInstr)),
8143 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8145 // Update thisMBB to fall through to newMBB
8146 thisMBB->addSuccessor(newMBB);
8148 // newMBB jumps to itself and fall through to nextMBB
8149 newMBB->addSuccessor(nextMBB);
8150 newMBB->addSuccessor(newMBB);
8152 DebugLoc dl = bInstr->getDebugLoc();
8153 // Insert instructions into newMBB based on incoming instruction
8154 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8155 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
8156 "unexpected number of operands");
8157 MachineOperand& dest1Oper = bInstr->getOperand(0);
8158 MachineOperand& dest2Oper = bInstr->getOperand(1);
8159 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8160 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
8161 argOpers[i] = &bInstr->getOperand(i+2);
8163 // We use some of the operands multiple times, so conservatively just
8164 // clear any kill flags that might be present.
8165 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8166 argOpers[i]->setIsKill(false);
8169 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8170 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8172 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8173 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8174 for (int i=0; i <= lastAddrIndx; ++i)
8175 (*MIB).addOperand(*argOpers[i]);
8176 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8177 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8178 // add 4 to displacement.
8179 for (int i=0; i <= lastAddrIndx-2; ++i)
8180 (*MIB).addOperand(*argOpers[i]);
8181 MachineOperand newOp3 = *(argOpers[3]);
8183 newOp3.setImm(newOp3.getImm()+4);
8185 newOp3.setOffset(newOp3.getOffset()+4);
8186 (*MIB).addOperand(newOp3);
8187 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8189 // t3/4 are defined later, at the bottom of the loop
8190 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8191 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8192 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8193 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8194 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8195 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8197 // The subsequent operations should be using the destination registers of
8198 //the PHI instructions.
8200 t1 = F->getRegInfo().createVirtualRegister(RC);
8201 t2 = F->getRegInfo().createVirtualRegister(RC);
8202 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8203 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8205 t1 = dest1Oper.getReg();
8206 t2 = dest2Oper.getReg();
8209 int valArgIndx = lastAddrIndx + 1;
8210 assert((argOpers[valArgIndx]->isReg() ||
8211 argOpers[valArgIndx]->isImm()) &&
8213 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8214 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8215 if (argOpers[valArgIndx]->isReg())
8216 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8218 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8219 if (regOpcL != X86::MOV32rr)
8221 (*MIB).addOperand(*argOpers[valArgIndx]);
8222 assert(argOpers[valArgIndx + 1]->isReg() ==
8223 argOpers[valArgIndx]->isReg());
8224 assert(argOpers[valArgIndx + 1]->isImm() ==
8225 argOpers[valArgIndx]->isImm());
8226 if (argOpers[valArgIndx + 1]->isReg())
8227 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8229 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8230 if (regOpcH != X86::MOV32rr)
8232 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8234 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8236 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
8239 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
8241 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
8244 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8245 for (int i=0; i <= lastAddrIndx; ++i)
8246 (*MIB).addOperand(*argOpers[i]);
8248 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8249 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8250 bInstr->memoperands_end());
8252 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
8253 MIB.addReg(X86::EAX);
8254 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
8255 MIB.addReg(X86::EDX);
8258 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8260 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8264 // private utility function
8266 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8267 MachineBasicBlock *MBB,
8268 unsigned cmovOpc) const {
8269 // For the atomic min/max operator, we generate
8272 // ld t1 = [min/max.addr]
8273 // mov t2 = [min/max.val]
8275 // cmov[cond] t2 = t1
8277 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8279 // fallthrough -->nextMBB
8281 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8282 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8283 MachineFunction::iterator MBBIter = MBB;
8286 /// First build the CFG
8287 MachineFunction *F = MBB->getParent();
8288 MachineBasicBlock *thisMBB = MBB;
8289 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8290 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8291 F->insert(MBBIter, newMBB);
8292 F->insert(MBBIter, nextMBB);
8294 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8295 nextMBB->splice(nextMBB->begin(), thisMBB,
8296 llvm::next(MachineBasicBlock::iterator(mInstr)),
8298 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8300 // Update thisMBB to fall through to newMBB
8301 thisMBB->addSuccessor(newMBB);
8303 // newMBB jumps to newMBB and fall through to nextMBB
8304 newMBB->addSuccessor(nextMBB);
8305 newMBB->addSuccessor(newMBB);
8307 DebugLoc dl = mInstr->getDebugLoc();
8308 // Insert instructions into newMBB based on incoming instruction
8309 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8310 "unexpected number of operands");
8311 MachineOperand& destOper = mInstr->getOperand(0);
8312 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8313 int numArgs = mInstr->getNumOperands() - 1;
8314 for (int i=0; i < numArgs; ++i)
8315 argOpers[i] = &mInstr->getOperand(i+1);
8317 // x86 address has 4 operands: base, index, scale, and displacement
8318 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8319 int valArgIndx = lastAddrIndx + 1;
8321 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8322 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8323 for (int i=0; i <= lastAddrIndx; ++i)
8324 (*MIB).addOperand(*argOpers[i]);
8326 // We only support register and immediate values
8327 assert((argOpers[valArgIndx]->isReg() ||
8328 argOpers[valArgIndx]->isImm()) &&
8331 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8332 if (argOpers[valArgIndx]->isReg())
8333 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
8335 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8336 (*MIB).addOperand(*argOpers[valArgIndx]);
8338 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8341 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8346 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8347 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8351 // Cmp and exchange if none has modified the memory location
8352 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8353 for (int i=0; i <= lastAddrIndx; ++i)
8354 (*MIB).addOperand(*argOpers[i]);
8356 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8357 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8358 mInstr->memoperands_end());
8360 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8361 MIB.addReg(X86::EAX);
8364 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8366 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
8370 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8371 // all of this code can be replaced with that in the .td file.
8373 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8374 unsigned numArgs, bool memArg) const {
8376 DebugLoc dl = MI->getDebugLoc();
8377 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8381 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8383 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8385 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8387 for (unsigned i = 0; i < numArgs; ++i) {
8388 MachineOperand &Op = MI->getOperand(i+1);
8390 if (!(Op.isReg() && Op.isImplicit()))
8394 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8397 MI->eraseFromParent();
8403 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8405 MachineBasicBlock *MBB) const {
8406 // Emit code to save XMM registers to the stack. The ABI says that the
8407 // number of registers to save is given in %al, so it's theoretically
8408 // possible to do an indirect jump trick to avoid saving all of them,
8409 // however this code takes a simpler approach and just executes all
8410 // of the stores if %al is non-zero. It's less code, and it's probably
8411 // easier on the hardware branch predictor, and stores aren't all that
8412 // expensive anyway.
8414 // Create the new basic blocks. One block contains all the XMM stores,
8415 // and one block is the final destination regardless of whether any
8416 // stores were performed.
8417 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8418 MachineFunction *F = MBB->getParent();
8419 MachineFunction::iterator MBBIter = MBB;
8421 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8422 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8423 F->insert(MBBIter, XMMSaveMBB);
8424 F->insert(MBBIter, EndMBB);
8426 // Transfer the remainder of MBB and its successor edges to EndMBB.
8427 EndMBB->splice(EndMBB->begin(), MBB,
8428 llvm::next(MachineBasicBlock::iterator(MI)),
8430 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
8432 // The original block will now fall through to the XMM save block.
8433 MBB->addSuccessor(XMMSaveMBB);
8434 // The XMMSaveMBB will fall through to the end block.
8435 XMMSaveMBB->addSuccessor(EndMBB);
8437 // Now add the instructions.
8438 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8439 DebugLoc DL = MI->getDebugLoc();
8441 unsigned CountReg = MI->getOperand(0).getReg();
8442 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8443 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8445 if (!Subtarget->isTargetWin64()) {
8446 // If %al is 0, branch around the XMM save block.
8447 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8448 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8449 MBB->addSuccessor(EndMBB);
8452 // In the XMM save block, save all the XMM argument registers.
8453 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8454 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8455 MachineMemOperand *MMO =
8456 F->getMachineMemOperand(
8457 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8458 MachineMemOperand::MOStore, Offset,
8459 /*Size=*/16, /*Align=*/16);
8460 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8461 .addFrameIndex(RegSaveFrameIndex)
8462 .addImm(/*Scale=*/1)
8463 .addReg(/*IndexReg=*/0)
8464 .addImm(/*Disp=*/Offset)
8465 .addReg(/*Segment=*/0)
8466 .addReg(MI->getOperand(i).getReg())
8467 .addMemOperand(MMO);
8470 MI->eraseFromParent(); // The pseudo instruction is gone now.
8476 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8477 MachineBasicBlock *BB) const {
8478 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8479 DebugLoc DL = MI->getDebugLoc();
8481 // To "insert" a SELECT_CC instruction, we actually have to insert the
8482 // diamond control-flow pattern. The incoming instruction knows the
8483 // destination vreg to set, the condition code register to branch on, the
8484 // true/false values to select between, and a branch opcode to use.
8485 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8486 MachineFunction::iterator It = BB;
8492 // cmpTY ccX, r1, r2
8494 // fallthrough --> copy0MBB
8495 MachineBasicBlock *thisMBB = BB;
8496 MachineFunction *F = BB->getParent();
8497 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8498 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8499 F->insert(It, copy0MBB);
8500 F->insert(It, sinkMBB);
8502 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8503 // live into the sink and copy blocks.
8504 const MachineFunction *MF = BB->getParent();
8505 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8506 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8508 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
8509 const MachineOperand &MO = MI->getOperand(I);
8510 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
8511 unsigned Reg = MO.getReg();
8512 if (Reg != X86::EFLAGS) continue;
8513 copy0MBB->addLiveIn(Reg);
8514 sinkMBB->addLiveIn(Reg);
8517 // Transfer the remainder of BB and its successor edges to sinkMBB.
8518 sinkMBB->splice(sinkMBB->begin(), BB,
8519 llvm::next(MachineBasicBlock::iterator(MI)),
8521 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
8523 // Add the true and fallthrough blocks as its successors.
8524 BB->addSuccessor(copy0MBB);
8525 BB->addSuccessor(sinkMBB);
8527 // Create the conditional branch instruction.
8529 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8530 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8533 // %FalseValue = ...
8534 // # fallthrough to sinkMBB
8535 copy0MBB->addSuccessor(sinkMBB);
8538 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8540 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
8541 TII->get(X86::PHI), MI->getOperand(0).getReg())
8542 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8543 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8545 MI->eraseFromParent(); // The pseudo instruction is gone now.
8550 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8551 MachineBasicBlock *BB) const {
8552 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8553 DebugLoc DL = MI->getDebugLoc();
8555 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8556 // non-trivial part is impdef of ESP.
8557 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8560 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
8561 .addExternalSymbol("_alloca")
8562 .addReg(X86::EAX, RegState::Implicit)
8563 .addReg(X86::ESP, RegState::Implicit)
8564 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8565 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8567 MI->eraseFromParent(); // The pseudo instruction is gone now.
8572 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8573 MachineBasicBlock *BB) const {
8574 // This is pretty easy. We're taking the value that we received from
8575 // our load from the relocation, sticking it in either RDI (x86-64)
8576 // or EAX and doing an indirect call. The return value will then
8577 // be in the normal return register.
8578 const X86InstrInfo *TII
8579 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8580 DebugLoc DL = MI->getDebugLoc();
8581 MachineFunction *F = BB->getParent();
8583 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8585 if (Subtarget->is64Bit()) {
8586 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8587 TII->get(X86::MOV64rm), X86::RDI)
8589 .addImm(0).addReg(0)
8590 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8591 MI->getOperand(3).getTargetFlags())
8593 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
8594 addDirectMem(MIB, X86::RDI);
8595 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8596 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8597 TII->get(X86::MOV32rm), X86::EAX)
8599 .addImm(0).addReg(0)
8600 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8601 MI->getOperand(3).getTargetFlags())
8603 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8604 addDirectMem(MIB, X86::EAX);
8606 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8607 TII->get(X86::MOV32rm), X86::EAX)
8608 .addReg(TII->getGlobalBaseReg(F))
8609 .addImm(0).addReg(0)
8610 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8611 MI->getOperand(3).getTargetFlags())
8613 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8614 addDirectMem(MIB, X86::EAX);
8617 MI->eraseFromParent(); // The pseudo instruction is gone now.
8622 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8623 MachineBasicBlock *BB) const {
8624 switch (MI->getOpcode()) {
8625 default: assert(false && "Unexpected instr type to insert");
8626 case X86::MINGW_ALLOCA:
8627 return EmitLoweredMingwAlloca(MI, BB);
8628 case X86::TLSCall_32:
8629 case X86::TLSCall_64:
8630 return EmitLoweredTLSCall(MI, BB);
8632 case X86::CMOV_V1I64:
8633 case X86::CMOV_FR32:
8634 case X86::CMOV_FR64:
8635 case X86::CMOV_V4F32:
8636 case X86::CMOV_V2F64:
8637 case X86::CMOV_V2I64:
8638 case X86::CMOV_GR16:
8639 case X86::CMOV_GR32:
8640 case X86::CMOV_RFP32:
8641 case X86::CMOV_RFP64:
8642 case X86::CMOV_RFP80:
8643 return EmitLoweredSelect(MI, BB);
8645 case X86::FP32_TO_INT16_IN_MEM:
8646 case X86::FP32_TO_INT32_IN_MEM:
8647 case X86::FP32_TO_INT64_IN_MEM:
8648 case X86::FP64_TO_INT16_IN_MEM:
8649 case X86::FP64_TO_INT32_IN_MEM:
8650 case X86::FP64_TO_INT64_IN_MEM:
8651 case X86::FP80_TO_INT16_IN_MEM:
8652 case X86::FP80_TO_INT32_IN_MEM:
8653 case X86::FP80_TO_INT64_IN_MEM: {
8654 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8655 DebugLoc DL = MI->getDebugLoc();
8657 // Change the floating point control register to use "round towards zero"
8658 // mode when truncating to an integer value.
8659 MachineFunction *F = BB->getParent();
8660 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8661 addFrameReference(BuildMI(*BB, MI, DL,
8662 TII->get(X86::FNSTCW16m)), CWFrameIdx);
8664 // Load the old value of the high byte of the control word...
8666 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8667 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
8670 // Set the high part to be round to zero...
8671 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8674 // Reload the modified control word now...
8675 addFrameReference(BuildMI(*BB, MI, DL,
8676 TII->get(X86::FLDCW16m)), CWFrameIdx);
8678 // Restore the memory image of control word to original value
8679 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8682 // Get the X86 opcode to use.
8684 switch (MI->getOpcode()) {
8685 default: llvm_unreachable("illegal opcode!");
8686 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8687 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8688 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8689 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8690 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8691 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8692 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8693 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8694 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8698 MachineOperand &Op = MI->getOperand(0);
8700 AM.BaseType = X86AddressMode::RegBase;
8701 AM.Base.Reg = Op.getReg();
8703 AM.BaseType = X86AddressMode::FrameIndexBase;
8704 AM.Base.FrameIndex = Op.getIndex();
8706 Op = MI->getOperand(1);
8708 AM.Scale = Op.getImm();
8709 Op = MI->getOperand(2);
8711 AM.IndexReg = Op.getImm();
8712 Op = MI->getOperand(3);
8713 if (Op.isGlobal()) {
8714 AM.GV = Op.getGlobal();
8716 AM.Disp = Op.getImm();
8718 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
8719 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
8721 // Reload the original control word now.
8722 addFrameReference(BuildMI(*BB, MI, DL,
8723 TII->get(X86::FLDCW16m)), CWFrameIdx);
8725 MI->eraseFromParent(); // The pseudo instruction is gone now.
8728 // String/text processing lowering.
8729 case X86::PCMPISTRM128REG:
8730 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8731 case X86::PCMPISTRM128MEM:
8732 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8733 case X86::PCMPESTRM128REG:
8734 return EmitPCMP(MI, BB, 5, false /* in mem */);
8735 case X86::PCMPESTRM128MEM:
8736 return EmitPCMP(MI, BB, 5, true /* in mem */);
8739 case X86::ATOMAND32:
8740 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8741 X86::AND32ri, X86::MOV32rm,
8743 X86::NOT32r, X86::EAX,
8744 X86::GR32RegisterClass);
8746 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8747 X86::OR32ri, X86::MOV32rm,
8749 X86::NOT32r, X86::EAX,
8750 X86::GR32RegisterClass);
8751 case X86::ATOMXOR32:
8752 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8753 X86::XOR32ri, X86::MOV32rm,
8755 X86::NOT32r, X86::EAX,
8756 X86::GR32RegisterClass);
8757 case X86::ATOMNAND32:
8758 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8759 X86::AND32ri, X86::MOV32rm,
8761 X86::NOT32r, X86::EAX,
8762 X86::GR32RegisterClass, true);
8763 case X86::ATOMMIN32:
8764 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8765 case X86::ATOMMAX32:
8766 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8767 case X86::ATOMUMIN32:
8768 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8769 case X86::ATOMUMAX32:
8770 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8772 case X86::ATOMAND16:
8773 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8774 X86::AND16ri, X86::MOV16rm,
8776 X86::NOT16r, X86::AX,
8777 X86::GR16RegisterClass);
8779 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8780 X86::OR16ri, X86::MOV16rm,
8782 X86::NOT16r, X86::AX,
8783 X86::GR16RegisterClass);
8784 case X86::ATOMXOR16:
8785 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8786 X86::XOR16ri, X86::MOV16rm,
8788 X86::NOT16r, X86::AX,
8789 X86::GR16RegisterClass);
8790 case X86::ATOMNAND16:
8791 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8792 X86::AND16ri, X86::MOV16rm,
8794 X86::NOT16r, X86::AX,
8795 X86::GR16RegisterClass, true);
8796 case X86::ATOMMIN16:
8797 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8798 case X86::ATOMMAX16:
8799 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8800 case X86::ATOMUMIN16:
8801 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8802 case X86::ATOMUMAX16:
8803 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8806 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8807 X86::AND8ri, X86::MOV8rm,
8809 X86::NOT8r, X86::AL,
8810 X86::GR8RegisterClass);
8812 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8813 X86::OR8ri, X86::MOV8rm,
8815 X86::NOT8r, X86::AL,
8816 X86::GR8RegisterClass);
8818 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8819 X86::XOR8ri, X86::MOV8rm,
8821 X86::NOT8r, X86::AL,
8822 X86::GR8RegisterClass);
8823 case X86::ATOMNAND8:
8824 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8825 X86::AND8ri, X86::MOV8rm,
8827 X86::NOT8r, X86::AL,
8828 X86::GR8RegisterClass, true);
8829 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8830 // This group is for 64-bit host.
8831 case X86::ATOMAND64:
8832 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8833 X86::AND64ri32, X86::MOV64rm,
8835 X86::NOT64r, X86::RAX,
8836 X86::GR64RegisterClass);
8838 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8839 X86::OR64ri32, X86::MOV64rm,
8841 X86::NOT64r, X86::RAX,
8842 X86::GR64RegisterClass);
8843 case X86::ATOMXOR64:
8844 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8845 X86::XOR64ri32, X86::MOV64rm,
8847 X86::NOT64r, X86::RAX,
8848 X86::GR64RegisterClass);
8849 case X86::ATOMNAND64:
8850 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8851 X86::AND64ri32, X86::MOV64rm,
8853 X86::NOT64r, X86::RAX,
8854 X86::GR64RegisterClass, true);
8855 case X86::ATOMMIN64:
8856 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8857 case X86::ATOMMAX64:
8858 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8859 case X86::ATOMUMIN64:
8860 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8861 case X86::ATOMUMAX64:
8862 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8864 // This group does 64-bit operations on a 32-bit host.
8865 case X86::ATOMAND6432:
8866 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8867 X86::AND32rr, X86::AND32rr,
8868 X86::AND32ri, X86::AND32ri,
8870 case X86::ATOMOR6432:
8871 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8872 X86::OR32rr, X86::OR32rr,
8873 X86::OR32ri, X86::OR32ri,
8875 case X86::ATOMXOR6432:
8876 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8877 X86::XOR32rr, X86::XOR32rr,
8878 X86::XOR32ri, X86::XOR32ri,
8880 case X86::ATOMNAND6432:
8881 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8882 X86::AND32rr, X86::AND32rr,
8883 X86::AND32ri, X86::AND32ri,
8885 case X86::ATOMADD6432:
8886 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8887 X86::ADD32rr, X86::ADC32rr,
8888 X86::ADD32ri, X86::ADC32ri,
8890 case X86::ATOMSUB6432:
8891 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8892 X86::SUB32rr, X86::SBB32rr,
8893 X86::SUB32ri, X86::SBB32ri,
8895 case X86::ATOMSWAP6432:
8896 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8897 X86::MOV32rr, X86::MOV32rr,
8898 X86::MOV32ri, X86::MOV32ri,
8900 case X86::VASTART_SAVE_XMM_REGS:
8901 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8905 //===----------------------------------------------------------------------===//
8906 // X86 Optimization Hooks
8907 //===----------------------------------------------------------------------===//
8909 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8913 const SelectionDAG &DAG,
8914 unsigned Depth) const {
8915 unsigned Opc = Op.getOpcode();
8916 assert((Opc >= ISD::BUILTIN_OP_END ||
8917 Opc == ISD::INTRINSIC_WO_CHAIN ||
8918 Opc == ISD::INTRINSIC_W_CHAIN ||
8919 Opc == ISD::INTRINSIC_VOID) &&
8920 "Should use MaskedValueIsZero if you don't know whether Op"
8921 " is a target node!");
8923 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8935 // These nodes' second result is a boolean.
8936 if (Op.getResNo() == 0)
8940 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8941 Mask.getBitWidth() - 1);
8946 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8947 /// node is a GlobalAddress + offset.
8948 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8949 const GlobalValue* &GA,
8950 int64_t &Offset) const {
8951 if (N->getOpcode() == X86ISD::Wrapper) {
8952 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8953 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8954 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8958 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8961 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8962 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8963 /// if the load addresses are consecutive, non-overlapping, and in the right
8965 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8966 const TargetLowering &TLI) {
8967 DebugLoc dl = N->getDebugLoc();
8968 EVT VT = N->getValueType(0);
8969 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8971 if (VT.getSizeInBits() != 128)
8974 SmallVector<SDValue, 16> Elts;
8975 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8976 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8978 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8981 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8982 /// and convert it from being a bunch of shuffles and extracts to a simple
8983 /// store and scalar loads to extract the elements.
8984 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8985 const TargetLowering &TLI) {
8986 SDValue InputVector = N->getOperand(0);
8988 // Only operate on vectors of 4 elements, where the alternative shuffling
8989 // gets to be more expensive.
8990 if (InputVector.getValueType() != MVT::v4i32)
8993 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8994 // single use which is a sign-extend or zero-extend, and all elements are
8996 SmallVector<SDNode *, 4> Uses;
8997 unsigned ExtractedElements = 0;
8998 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8999 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
9000 if (UI.getUse().getResNo() != InputVector.getResNo())
9003 SDNode *Extract = *UI;
9004 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9007 if (Extract->getValueType(0) != MVT::i32)
9009 if (!Extract->hasOneUse())
9011 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9012 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9014 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9017 // Record which element was extracted.
9018 ExtractedElements |=
9019 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9021 Uses.push_back(Extract);
9024 // If not all the elements were used, this may not be worthwhile.
9025 if (ExtractedElements != 15)
9028 // Ok, we've now decided to do the transformation.
9029 DebugLoc dl = InputVector.getDebugLoc();
9031 // Store the value to a temporary stack slot.
9032 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9033 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
9036 // Replace each use (extract) with a load of the appropriate element.
9037 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9038 UE = Uses.end(); UI != UE; ++UI) {
9039 SDNode *Extract = *UI;
9041 // Compute the element's address.
9042 SDValue Idx = Extract->getOperand(1);
9044 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9045 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9046 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9048 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
9051 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
9052 NULL, 0, false, false, 0);
9054 // Replace the exact with the load.
9055 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9058 // The replacement was made in place; don't return anything.
9062 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9063 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9064 const X86Subtarget *Subtarget) {
9065 DebugLoc DL = N->getDebugLoc();
9066 SDValue Cond = N->getOperand(0);
9067 // Get the LHS/RHS of the select.
9068 SDValue LHS = N->getOperand(1);
9069 SDValue RHS = N->getOperand(2);
9071 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9072 // instructions match the semantics of the common C idiom x<y?x:y but not
9073 // x<=y?x:y, because of how they handle negative zero (which can be
9074 // ignored in unsafe-math mode).
9075 if (Subtarget->hasSSE2() &&
9076 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9077 Cond.getOpcode() == ISD::SETCC) {
9078 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9080 unsigned Opcode = 0;
9081 // Check for x CC y ? x : y.
9082 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9083 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9087 // Converting this to a min would handle NaNs incorrectly, and swapping
9088 // the operands would cause it to handle comparisons between positive
9089 // and negative zero incorrectly.
9090 if (!FiniteOnlyFPMath() &&
9091 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9092 if (!UnsafeFPMath &&
9093 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9095 std::swap(LHS, RHS);
9097 Opcode = X86ISD::FMIN;
9100 // Converting this to a min would handle comparisons between positive
9101 // and negative zero incorrectly.
9102 if (!UnsafeFPMath &&
9103 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9105 Opcode = X86ISD::FMIN;
9108 // Converting this to a min would handle both negative zeros and NaNs
9109 // incorrectly, but we can swap the operands to fix both.
9110 std::swap(LHS, RHS);
9114 Opcode = X86ISD::FMIN;
9118 // Converting this to a max would handle comparisons between positive
9119 // and negative zero incorrectly.
9120 if (!UnsafeFPMath &&
9121 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9123 Opcode = X86ISD::FMAX;
9126 // Converting this to a max would handle NaNs incorrectly, and swapping
9127 // the operands would cause it to handle comparisons between positive
9128 // and negative zero incorrectly.
9129 if (!FiniteOnlyFPMath() &&
9130 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9131 if (!UnsafeFPMath &&
9132 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9134 std::swap(LHS, RHS);
9136 Opcode = X86ISD::FMAX;
9139 // Converting this to a max would handle both negative zeros and NaNs
9140 // incorrectly, but we can swap the operands to fix both.
9141 std::swap(LHS, RHS);
9145 Opcode = X86ISD::FMAX;
9148 // Check for x CC y ? y : x -- a min/max with reversed arms.
9149 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9150 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9154 // Converting this to a min would handle comparisons between positive
9155 // and negative zero incorrectly, and swapping the operands would
9156 // cause it to handle NaNs incorrectly.
9157 if (!UnsafeFPMath &&
9158 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9159 if (!FiniteOnlyFPMath() &&
9160 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9162 std::swap(LHS, RHS);
9164 Opcode = X86ISD::FMIN;
9167 // Converting this to a min would handle NaNs incorrectly.
9168 if (!UnsafeFPMath &&
9169 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9171 Opcode = X86ISD::FMIN;
9174 // Converting this to a min would handle both negative zeros and NaNs
9175 // incorrectly, but we can swap the operands to fix both.
9176 std::swap(LHS, RHS);
9180 Opcode = X86ISD::FMIN;
9184 // Converting this to a max would handle NaNs incorrectly.
9185 if (!FiniteOnlyFPMath() &&
9186 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9188 Opcode = X86ISD::FMAX;
9191 // Converting this to a max would handle comparisons between positive
9192 // and negative zero incorrectly, and swapping the operands would
9193 // cause it to handle NaNs incorrectly.
9194 if (!UnsafeFPMath &&
9195 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9196 if (!FiniteOnlyFPMath() &&
9197 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9199 std::swap(LHS, RHS);
9201 Opcode = X86ISD::FMAX;
9204 // Converting this to a max would handle both negative zeros and NaNs
9205 // incorrectly, but we can swap the operands to fix both.
9206 std::swap(LHS, RHS);
9210 Opcode = X86ISD::FMAX;
9216 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9219 // If this is a select between two integer constants, try to do some
9221 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9222 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9223 // Don't do this for crazy integer types.
9224 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9225 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9226 // so that TrueC (the true value) is larger than FalseC.
9227 bool NeedsCondInvert = false;
9229 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9230 // Efficiently invertible.
9231 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9232 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9233 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9234 NeedsCondInvert = true;
9235 std::swap(TrueC, FalseC);
9238 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9239 if (FalseC->getAPIntValue() == 0 &&
9240 TrueC->getAPIntValue().isPowerOf2()) {
9241 if (NeedsCondInvert) // Invert the condition if needed.
9242 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9243 DAG.getConstant(1, Cond.getValueType()));
9245 // Zero extend the condition if needed.
9246 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9248 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9249 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9250 DAG.getConstant(ShAmt, MVT::i8));
9253 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9254 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9255 if (NeedsCondInvert) // Invert the condition if needed.
9256 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9257 DAG.getConstant(1, Cond.getValueType()));
9259 // Zero extend the condition if needed.
9260 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9261 FalseC->getValueType(0), Cond);
9262 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9263 SDValue(FalseC, 0));
9266 // Optimize cases that will turn into an LEA instruction. This requires
9267 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9268 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9269 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9270 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9272 bool isFastMultiplier = false;
9274 switch ((unsigned char)Diff) {
9276 case 1: // result = add base, cond
9277 case 2: // result = lea base( , cond*2)
9278 case 3: // result = lea base(cond, cond*2)
9279 case 4: // result = lea base( , cond*4)
9280 case 5: // result = lea base(cond, cond*4)
9281 case 8: // result = lea base( , cond*8)
9282 case 9: // result = lea base(cond, cond*8)
9283 isFastMultiplier = true;
9288 if (isFastMultiplier) {
9289 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9290 if (NeedsCondInvert) // Invert the condition if needed.
9291 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9292 DAG.getConstant(1, Cond.getValueType()));
9294 // Zero extend the condition if needed.
9295 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9297 // Scale the condition by the difference.
9299 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9300 DAG.getConstant(Diff, Cond.getValueType()));
9302 // Add the base if non-zero.
9303 if (FalseC->getAPIntValue() != 0)
9304 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9305 SDValue(FalseC, 0));
9315 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9316 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9317 TargetLowering::DAGCombinerInfo &DCI) {
9318 DebugLoc DL = N->getDebugLoc();
9320 // If the flag operand isn't dead, don't touch this CMOV.
9321 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9324 // If this is a select between two integer constants, try to do some
9325 // optimizations. Note that the operands are ordered the opposite of SELECT
9327 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9328 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9329 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9330 // larger than FalseC (the false value).
9331 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9333 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9334 CC = X86::GetOppositeBranchCondition(CC);
9335 std::swap(TrueC, FalseC);
9338 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9339 // This is efficient for any integer data type (including i8/i16) and
9341 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9342 SDValue Cond = N->getOperand(3);
9343 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9344 DAG.getConstant(CC, MVT::i8), Cond);
9346 // Zero extend the condition if needed.
9347 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9349 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9350 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9351 DAG.getConstant(ShAmt, MVT::i8));
9352 if (N->getNumValues() == 2) // Dead flag value?
9353 return DCI.CombineTo(N, Cond, SDValue());
9357 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9358 // for any integer data type, including i8/i16.
9359 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9360 SDValue Cond = N->getOperand(3);
9361 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9362 DAG.getConstant(CC, MVT::i8), Cond);
9364 // Zero extend the condition if needed.
9365 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9366 FalseC->getValueType(0), Cond);
9367 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9368 SDValue(FalseC, 0));
9370 if (N->getNumValues() == 2) // Dead flag value?
9371 return DCI.CombineTo(N, Cond, SDValue());
9375 // Optimize cases that will turn into an LEA instruction. This requires
9376 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9377 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9378 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9379 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9381 bool isFastMultiplier = false;
9383 switch ((unsigned char)Diff) {
9385 case 1: // result = add base, cond
9386 case 2: // result = lea base( , cond*2)
9387 case 3: // result = lea base(cond, cond*2)
9388 case 4: // result = lea base( , cond*4)
9389 case 5: // result = lea base(cond, cond*4)
9390 case 8: // result = lea base( , cond*8)
9391 case 9: // result = lea base(cond, cond*8)
9392 isFastMultiplier = true;
9397 if (isFastMultiplier) {
9398 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9399 SDValue Cond = N->getOperand(3);
9400 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9401 DAG.getConstant(CC, MVT::i8), Cond);
9402 // Zero extend the condition if needed.
9403 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9405 // Scale the condition by the difference.
9407 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9408 DAG.getConstant(Diff, Cond.getValueType()));
9410 // Add the base if non-zero.
9411 if (FalseC->getAPIntValue() != 0)
9412 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9413 SDValue(FalseC, 0));
9414 if (N->getNumValues() == 2) // Dead flag value?
9415 return DCI.CombineTo(N, Cond, SDValue());
9425 /// PerformMulCombine - Optimize a single multiply with constant into two
9426 /// in order to implement it with two cheaper instructions, e.g.
9427 /// LEA + SHL, LEA + LEA.
9428 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9429 TargetLowering::DAGCombinerInfo &DCI) {
9430 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9433 EVT VT = N->getValueType(0);
9437 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9440 uint64_t MulAmt = C->getZExtValue();
9441 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9444 uint64_t MulAmt1 = 0;
9445 uint64_t MulAmt2 = 0;
9446 if ((MulAmt % 9) == 0) {
9448 MulAmt2 = MulAmt / 9;
9449 } else if ((MulAmt % 5) == 0) {
9451 MulAmt2 = MulAmt / 5;
9452 } else if ((MulAmt % 3) == 0) {
9454 MulAmt2 = MulAmt / 3;
9457 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9458 DebugLoc DL = N->getDebugLoc();
9460 if (isPowerOf2_64(MulAmt2) &&
9461 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9462 // If second multiplifer is pow2, issue it first. We want the multiply by
9463 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9465 std::swap(MulAmt1, MulAmt2);
9468 if (isPowerOf2_64(MulAmt1))
9469 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9470 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9472 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9473 DAG.getConstant(MulAmt1, VT));
9475 if (isPowerOf2_64(MulAmt2))
9476 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9477 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9479 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9480 DAG.getConstant(MulAmt2, VT));
9482 // Do not add new nodes to DAG combiner worklist.
9483 DCI.CombineTo(N, NewMul, false);
9488 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9489 SDValue N0 = N->getOperand(0);
9490 SDValue N1 = N->getOperand(1);
9491 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9492 EVT VT = N0.getValueType();
9494 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9495 // since the result of setcc_c is all zero's or all ones.
9496 if (N1C && N0.getOpcode() == ISD::AND &&
9497 N0.getOperand(1).getOpcode() == ISD::Constant) {
9498 SDValue N00 = N0.getOperand(0);
9499 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9500 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9501 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9502 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9503 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9504 APInt ShAmt = N1C->getAPIntValue();
9505 Mask = Mask.shl(ShAmt);
9507 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9508 N00, DAG.getConstant(Mask, VT));
9515 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9517 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9518 const X86Subtarget *Subtarget) {
9519 EVT VT = N->getValueType(0);
9520 if (!VT.isVector() && VT.isInteger() &&
9521 N->getOpcode() == ISD::SHL)
9522 return PerformSHLCombine(N, DAG);
9524 // On X86 with SSE2 support, we can transform this to a vector shift if
9525 // all elements are shifted by the same amount. We can't do this in legalize
9526 // because the a constant vector is typically transformed to a constant pool
9527 // so we have no knowledge of the shift amount.
9528 if (!Subtarget->hasSSE2())
9531 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9534 SDValue ShAmtOp = N->getOperand(1);
9535 EVT EltVT = VT.getVectorElementType();
9536 DebugLoc DL = N->getDebugLoc();
9537 SDValue BaseShAmt = SDValue();
9538 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9539 unsigned NumElts = VT.getVectorNumElements();
9541 for (; i != NumElts; ++i) {
9542 SDValue Arg = ShAmtOp.getOperand(i);
9543 if (Arg.getOpcode() == ISD::UNDEF) continue;
9547 for (; i != NumElts; ++i) {
9548 SDValue Arg = ShAmtOp.getOperand(i);
9549 if (Arg.getOpcode() == ISD::UNDEF) continue;
9550 if (Arg != BaseShAmt) {
9554 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9555 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9556 SDValue InVec = ShAmtOp.getOperand(0);
9557 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9558 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9560 for (; i != NumElts; ++i) {
9561 SDValue Arg = InVec.getOperand(i);
9562 if (Arg.getOpcode() == ISD::UNDEF) continue;
9566 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9567 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9568 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9569 if (C->getZExtValue() == SplatIdx)
9570 BaseShAmt = InVec.getOperand(1);
9573 if (BaseShAmt.getNode() == 0)
9574 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9575 DAG.getIntPtrConstant(0));
9579 // The shift amount is an i32.
9580 if (EltVT.bitsGT(MVT::i32))
9581 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9582 else if (EltVT.bitsLT(MVT::i32))
9583 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9585 // The shift amount is identical so we can do a vector shift.
9586 SDValue ValOp = N->getOperand(0);
9587 switch (N->getOpcode()) {
9589 llvm_unreachable("Unknown shift opcode!");
9592 if (VT == MVT::v2i64)
9593 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9594 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9596 if (VT == MVT::v4i32)
9597 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9598 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9600 if (VT == MVT::v8i16)
9601 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9602 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9606 if (VT == MVT::v4i32)
9607 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9608 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9610 if (VT == MVT::v8i16)
9611 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9612 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9616 if (VT == MVT::v2i64)
9617 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9618 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9620 if (VT == MVT::v4i32)
9621 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9622 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9624 if (VT == MVT::v8i16)
9625 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9626 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9633 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9634 TargetLowering::DAGCombinerInfo &DCI,
9635 const X86Subtarget *Subtarget) {
9636 if (DCI.isBeforeLegalizeOps())
9639 EVT VT = N->getValueType(0);
9640 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9643 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9644 SDValue N0 = N->getOperand(0);
9645 SDValue N1 = N->getOperand(1);
9646 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9648 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9650 if (!N0.hasOneUse() || !N1.hasOneUse())
9653 SDValue ShAmt0 = N0.getOperand(1);
9654 if (ShAmt0.getValueType() != MVT::i8)
9656 SDValue ShAmt1 = N1.getOperand(1);
9657 if (ShAmt1.getValueType() != MVT::i8)
9659 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9660 ShAmt0 = ShAmt0.getOperand(0);
9661 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9662 ShAmt1 = ShAmt1.getOperand(0);
9664 DebugLoc DL = N->getDebugLoc();
9665 unsigned Opc = X86ISD::SHLD;
9666 SDValue Op0 = N0.getOperand(0);
9667 SDValue Op1 = N1.getOperand(0);
9668 if (ShAmt0.getOpcode() == ISD::SUB) {
9670 std::swap(Op0, Op1);
9671 std::swap(ShAmt0, ShAmt1);
9674 unsigned Bits = VT.getSizeInBits();
9675 if (ShAmt1.getOpcode() == ISD::SUB) {
9676 SDValue Sum = ShAmt1.getOperand(0);
9677 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9678 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
9679 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
9680 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
9681 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
9682 return DAG.getNode(Opc, DL, VT,
9684 DAG.getNode(ISD::TRUNCATE, DL,
9687 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9688 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9690 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9691 return DAG.getNode(Opc, DL, VT,
9692 N0.getOperand(0), N1.getOperand(0),
9693 DAG.getNode(ISD::TRUNCATE, DL,
9700 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9701 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9702 const X86Subtarget *Subtarget) {
9703 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9704 // the FP state in cases where an emms may be missing.
9705 // A preferable solution to the general problem is to figure out the right
9706 // places to insert EMMS. This qualifies as a quick hack.
9708 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9709 StoreSDNode *St = cast<StoreSDNode>(N);
9710 EVT VT = St->getValue().getValueType();
9711 if (VT.getSizeInBits() != 64)
9714 const Function *F = DAG.getMachineFunction().getFunction();
9715 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9716 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9717 && Subtarget->hasSSE2();
9718 if ((VT.isVector() ||
9719 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9720 isa<LoadSDNode>(St->getValue()) &&
9721 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9722 St->getChain().hasOneUse() && !St->isVolatile()) {
9723 SDNode* LdVal = St->getValue().getNode();
9725 int TokenFactorIndex = -1;
9726 SmallVector<SDValue, 8> Ops;
9727 SDNode* ChainVal = St->getChain().getNode();
9728 // Must be a store of a load. We currently handle two cases: the load
9729 // is a direct child, and it's under an intervening TokenFactor. It is
9730 // possible to dig deeper under nested TokenFactors.
9731 if (ChainVal == LdVal)
9732 Ld = cast<LoadSDNode>(St->getChain());
9733 else if (St->getValue().hasOneUse() &&
9734 ChainVal->getOpcode() == ISD::TokenFactor) {
9735 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9736 if (ChainVal->getOperand(i).getNode() == LdVal) {
9737 TokenFactorIndex = i;
9738 Ld = cast<LoadSDNode>(St->getValue());
9740 Ops.push_back(ChainVal->getOperand(i));
9744 if (!Ld || !ISD::isNormalLoad(Ld))
9747 // If this is not the MMX case, i.e. we are just turning i64 load/store
9748 // into f64 load/store, avoid the transformation if there are multiple
9749 // uses of the loaded value.
9750 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9753 DebugLoc LdDL = Ld->getDebugLoc();
9754 DebugLoc StDL = N->getDebugLoc();
9755 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9756 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9758 if (Subtarget->is64Bit() || F64IsLegal) {
9759 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9760 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9761 Ld->getBasePtr(), Ld->getSrcValue(),
9762 Ld->getSrcValueOffset(), Ld->isVolatile(),
9763 Ld->isNonTemporal(), Ld->getAlignment());
9764 SDValue NewChain = NewLd.getValue(1);
9765 if (TokenFactorIndex != -1) {
9766 Ops.push_back(NewChain);
9767 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9770 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9771 St->getSrcValue(), St->getSrcValueOffset(),
9772 St->isVolatile(), St->isNonTemporal(),
9773 St->getAlignment());
9776 // Otherwise, lower to two pairs of 32-bit loads / stores.
9777 SDValue LoAddr = Ld->getBasePtr();
9778 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9779 DAG.getConstant(4, MVT::i32));
9781 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9782 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9783 Ld->isVolatile(), Ld->isNonTemporal(),
9784 Ld->getAlignment());
9785 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9786 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9787 Ld->isVolatile(), Ld->isNonTemporal(),
9788 MinAlign(Ld->getAlignment(), 4));
9790 SDValue NewChain = LoLd.getValue(1);
9791 if (TokenFactorIndex != -1) {
9792 Ops.push_back(LoLd);
9793 Ops.push_back(HiLd);
9794 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9798 LoAddr = St->getBasePtr();
9799 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9800 DAG.getConstant(4, MVT::i32));
9802 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9803 St->getSrcValue(), St->getSrcValueOffset(),
9804 St->isVolatile(), St->isNonTemporal(),
9805 St->getAlignment());
9806 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9808 St->getSrcValueOffset() + 4,
9810 St->isNonTemporal(),
9811 MinAlign(St->getAlignment(), 4));
9812 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9817 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9818 /// X86ISD::FXOR nodes.
9819 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9820 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9821 // F[X]OR(0.0, x) -> x
9822 // F[X]OR(x, 0.0) -> x
9823 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9824 if (C->getValueAPF().isPosZero())
9825 return N->getOperand(1);
9826 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9827 if (C->getValueAPF().isPosZero())
9828 return N->getOperand(0);
9832 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9833 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9834 // FAND(0.0, x) -> 0.0
9835 // FAND(x, 0.0) -> 0.0
9836 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9837 if (C->getValueAPF().isPosZero())
9838 return N->getOperand(0);
9839 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9840 if (C->getValueAPF().isPosZero())
9841 return N->getOperand(1);
9845 static SDValue PerformBTCombine(SDNode *N,
9847 TargetLowering::DAGCombinerInfo &DCI) {
9848 // BT ignores high bits in the bit index operand.
9849 SDValue Op1 = N->getOperand(1);
9850 if (Op1.hasOneUse()) {
9851 unsigned BitWidth = Op1.getValueSizeInBits();
9852 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9853 APInt KnownZero, KnownOne;
9854 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9855 !DCI.isBeforeLegalizeOps());
9856 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9857 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9858 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9859 DCI.CommitTargetLoweringOpt(TLO);
9864 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9865 SDValue Op = N->getOperand(0);
9866 if (Op.getOpcode() == ISD::BIT_CONVERT)
9867 Op = Op.getOperand(0);
9868 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9869 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9870 VT.getVectorElementType().getSizeInBits() ==
9871 OpVT.getVectorElementType().getSizeInBits()) {
9872 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9877 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9878 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9879 // (and (i32 x86isd::setcc_carry), 1)
9880 // This eliminates the zext. This transformation is necessary because
9881 // ISD::SETCC is always legalized to i8.
9882 DebugLoc dl = N->getDebugLoc();
9883 SDValue N0 = N->getOperand(0);
9884 EVT VT = N->getValueType(0);
9885 if (N0.getOpcode() == ISD::AND &&
9887 N0.getOperand(0).hasOneUse()) {
9888 SDValue N00 = N0.getOperand(0);
9889 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9891 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9892 if (!C || C->getZExtValue() != 1)
9894 return DAG.getNode(ISD::AND, dl, VT,
9895 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9896 N00.getOperand(0), N00.getOperand(1)),
9897 DAG.getConstant(1, VT));
9903 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9904 DAGCombinerInfo &DCI) const {
9905 SelectionDAG &DAG = DCI.DAG;
9906 switch (N->getOpcode()) {
9908 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9909 case ISD::EXTRACT_VECTOR_ELT:
9910 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9911 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9912 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9913 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9916 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9917 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
9918 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9920 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9921 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9922 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9923 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9924 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9930 /// isTypeDesirableForOp - Return true if the target has native support for
9931 /// the specified value type and it is 'desirable' to use the type for the
9932 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9933 /// instruction encodings are longer and some i16 instructions are slow.
9934 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9935 if (!isTypeLegal(VT))
9944 case ISD::SIGN_EXTEND:
9945 case ISD::ZERO_EXTEND:
9946 case ISD::ANY_EXTEND:
9959 static bool MayFoldLoad(SDValue Op) {
9960 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
9963 static bool MayFoldIntoStore(SDValue Op) {
9964 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
9967 /// IsDesirableToPromoteOp - This method query the target whether it is
9968 /// beneficial for dag combiner to promote the specified node. If true, it
9969 /// should return the desired promotion type by reference.
9970 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
9971 EVT VT = Op.getValueType();
9975 bool Promote = false;
9976 bool Commute = false;
9977 switch (Op.getOpcode()) {
9980 LoadSDNode *LD = cast<LoadSDNode>(Op);
9981 // If the non-extending load has a single use and it's not live out, then it
9983 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
9985 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9986 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
9987 // The only case where we'd want to promote LOAD (rather then it being
9988 // promoted as an operand is when it's only use is liveout.
9989 if (UI->getOpcode() != ISD::CopyToReg)
9996 case ISD::SIGN_EXTEND:
9997 case ISD::ZERO_EXTEND:
9998 case ISD::ANY_EXTEND:
10003 SDValue N0 = Op.getOperand(0);
10004 // Look out for (store (shl (load), x)).
10005 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10018 SDValue N0 = Op.getOperand(0);
10019 SDValue N1 = Op.getOperand(1);
10020 if (!Commute && MayFoldLoad(N1))
10022 // Avoid disabling potential load folding opportunities.
10023 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10025 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10035 //===----------------------------------------------------------------------===//
10036 // X86 Inline Assembly Support
10037 //===----------------------------------------------------------------------===//
10039 static bool LowerToBSwap(CallInst *CI) {
10040 // FIXME: this should verify that we are targetting a 486 or better. If not,
10041 // we will turn this bswap into something that will be lowered to logical ops
10042 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10043 // so don't worry about this.
10045 // Verify this is a simple bswap.
10046 if (CI->getNumArgOperands() != 1 ||
10047 CI->getType() != CI->getArgOperand(0)->getType() ||
10048 !CI->getType()->isIntegerTy())
10051 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10052 if (!Ty || Ty->getBitWidth() % 16 != 0)
10055 // Okay, we can do this xform, do so now.
10056 const Type *Tys[] = { Ty };
10057 Module *M = CI->getParent()->getParent()->getParent();
10058 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10060 Value *Op = CI->getArgOperand(0);
10061 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10063 CI->replaceAllUsesWith(Op);
10064 CI->eraseFromParent();
10068 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10069 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10070 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10072 std::string AsmStr = IA->getAsmString();
10074 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10075 SmallVector<StringRef, 4> AsmPieces;
10076 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10078 switch (AsmPieces.size()) {
10079 default: return false;
10081 AsmStr = AsmPieces[0];
10083 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10086 if (AsmPieces.size() == 2 &&
10087 (AsmPieces[0] == "bswap" ||
10088 AsmPieces[0] == "bswapq" ||
10089 AsmPieces[0] == "bswapl") &&
10090 (AsmPieces[1] == "$0" ||
10091 AsmPieces[1] == "${0:q}")) {
10092 // No need to check constraints, nothing other than the equivalent of
10093 // "=r,0" would be valid here.
10094 return LowerToBSwap(CI);
10096 // rorw $$8, ${0:w} --> llvm.bswap.i16
10097 if (CI->getType()->isIntegerTy(16) &&
10098 AsmPieces.size() == 3 &&
10099 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10100 AsmPieces[1] == "$$8," &&
10101 AsmPieces[2] == "${0:w}" &&
10102 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10104 const std::string &Constraints = IA->getConstraintString();
10105 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10106 std::sort(AsmPieces.begin(), AsmPieces.end());
10107 if (AsmPieces.size() == 4 &&
10108 AsmPieces[0] == "~{cc}" &&
10109 AsmPieces[1] == "~{dirflag}" &&
10110 AsmPieces[2] == "~{flags}" &&
10111 AsmPieces[3] == "~{fpsr}") {
10112 return LowerToBSwap(CI);
10117 if (CI->getType()->isIntegerTy(64) &&
10118 Constraints.size() >= 2 &&
10119 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10120 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10121 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10122 SmallVector<StringRef, 4> Words;
10123 SplitString(AsmPieces[0], Words, " \t");
10124 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10126 SplitString(AsmPieces[1], Words, " \t");
10127 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10129 SplitString(AsmPieces[2], Words, " \t,");
10130 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10131 Words[2] == "%edx") {
10132 return LowerToBSwap(CI);
10144 /// getConstraintType - Given a constraint letter, return the type of
10145 /// constraint it is for this target.
10146 X86TargetLowering::ConstraintType
10147 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10148 if (Constraint.size() == 1) {
10149 switch (Constraint[0]) {
10161 return C_RegisterClass;
10169 return TargetLowering::getConstraintType(Constraint);
10172 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10173 /// with another that has more specific requirements based on the type of the
10174 /// corresponding operand.
10175 const char *X86TargetLowering::
10176 LowerXConstraint(EVT ConstraintVT) const {
10177 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10178 // 'f' like normal targets.
10179 if (ConstraintVT.isFloatingPoint()) {
10180 if (Subtarget->hasSSE2())
10182 if (Subtarget->hasSSE1())
10186 return TargetLowering::LowerXConstraint(ConstraintVT);
10189 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10190 /// vector. If it is invalid, don't add anything to Ops.
10191 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10193 std::vector<SDValue>&Ops,
10194 SelectionDAG &DAG) const {
10195 SDValue Result(0, 0);
10197 switch (Constraint) {
10200 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10201 if (C->getZExtValue() <= 31) {
10202 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10208 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10209 if (C->getZExtValue() <= 63) {
10210 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10216 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10217 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10218 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10224 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10225 if (C->getZExtValue() <= 255) {
10226 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10232 // 32-bit signed value
10233 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10234 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10235 C->getSExtValue())) {
10236 // Widen to 64 bits here to get it sign extended.
10237 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10240 // FIXME gcc accepts some relocatable values here too, but only in certain
10241 // memory models; it's complicated.
10246 // 32-bit unsigned value
10247 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10248 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10249 C->getZExtValue())) {
10250 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10254 // FIXME gcc accepts some relocatable values here too, but only in certain
10255 // memory models; it's complicated.
10259 // Literal immediates are always ok.
10260 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10261 // Widen to 64 bits here to get it sign extended.
10262 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10266 // In any sort of PIC mode addresses need to be computed at runtime by
10267 // adding in a register or some sort of table lookup. These can't
10268 // be used as immediates.
10269 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
10272 // If we are in non-pic codegen mode, we allow the address of a global (with
10273 // an optional displacement) to be used with 'i'.
10274 GlobalAddressSDNode *GA = 0;
10275 int64_t Offset = 0;
10277 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10279 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10280 Offset += GA->getOffset();
10282 } else if (Op.getOpcode() == ISD::ADD) {
10283 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10284 Offset += C->getZExtValue();
10285 Op = Op.getOperand(0);
10288 } else if (Op.getOpcode() == ISD::SUB) {
10289 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10290 Offset += -C->getZExtValue();
10291 Op = Op.getOperand(0);
10296 // Otherwise, this isn't something we can handle, reject it.
10300 const GlobalValue *GV = GA->getGlobal();
10301 // If we require an extra load to get this address, as in PIC mode, we
10302 // can't accept it.
10303 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10304 getTargetMachine())))
10307 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
10308 GA->getValueType(0), Offset);
10313 if (Result.getNode()) {
10314 Ops.push_back(Result);
10317 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10320 std::vector<unsigned> X86TargetLowering::
10321 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10323 if (Constraint.size() == 1) {
10324 // FIXME: not handling fp-stack yet!
10325 switch (Constraint[0]) { // GCC X86 Constraint Letters
10326 default: break; // Unknown constraint letter
10327 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10328 if (Subtarget->is64Bit()) {
10329 if (VT == MVT::i32)
10330 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10331 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10332 X86::R10D,X86::R11D,X86::R12D,
10333 X86::R13D,X86::R14D,X86::R15D,
10334 X86::EBP, X86::ESP, 0);
10335 else if (VT == MVT::i16)
10336 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10337 X86::SI, X86::DI, X86::R8W,X86::R9W,
10338 X86::R10W,X86::R11W,X86::R12W,
10339 X86::R13W,X86::R14W,X86::R15W,
10340 X86::BP, X86::SP, 0);
10341 else if (VT == MVT::i8)
10342 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10343 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10344 X86::R10B,X86::R11B,X86::R12B,
10345 X86::R13B,X86::R14B,X86::R15B,
10346 X86::BPL, X86::SPL, 0);
10348 else if (VT == MVT::i64)
10349 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10350 X86::RSI, X86::RDI, X86::R8, X86::R9,
10351 X86::R10, X86::R11, X86::R12,
10352 X86::R13, X86::R14, X86::R15,
10353 X86::RBP, X86::RSP, 0);
10357 // 32-bit fallthrough
10358 case 'Q': // Q_REGS
10359 if (VT == MVT::i32)
10360 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10361 else if (VT == MVT::i16)
10362 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10363 else if (VT == MVT::i8)
10364 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10365 else if (VT == MVT::i64)
10366 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10371 return std::vector<unsigned>();
10374 std::pair<unsigned, const TargetRegisterClass*>
10375 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10377 // First, see if this is a constraint that directly corresponds to an LLVM
10379 if (Constraint.size() == 1) {
10380 // GCC Constraint Letters
10381 switch (Constraint[0]) {
10383 case 'r': // GENERAL_REGS
10384 case 'l': // INDEX_REGS
10386 return std::make_pair(0U, X86::GR8RegisterClass);
10387 if (VT == MVT::i16)
10388 return std::make_pair(0U, X86::GR16RegisterClass);
10389 if (VT == MVT::i32 || !Subtarget->is64Bit())
10390 return std::make_pair(0U, X86::GR32RegisterClass);
10391 return std::make_pair(0U, X86::GR64RegisterClass);
10392 case 'R': // LEGACY_REGS
10394 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10395 if (VT == MVT::i16)
10396 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10397 if (VT == MVT::i32 || !Subtarget->is64Bit())
10398 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10399 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10400 case 'f': // FP Stack registers.
10401 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10402 // value to the correct fpstack register class.
10403 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10404 return std::make_pair(0U, X86::RFP32RegisterClass);
10405 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10406 return std::make_pair(0U, X86::RFP64RegisterClass);
10407 return std::make_pair(0U, X86::RFP80RegisterClass);
10408 case 'y': // MMX_REGS if MMX allowed.
10409 if (!Subtarget->hasMMX()) break;
10410 return std::make_pair(0U, X86::VR64RegisterClass);
10411 case 'Y': // SSE_REGS if SSE2 allowed
10412 if (!Subtarget->hasSSE2()) break;
10414 case 'x': // SSE_REGS if SSE1 allowed
10415 if (!Subtarget->hasSSE1()) break;
10417 switch (VT.getSimpleVT().SimpleTy) {
10419 // Scalar SSE types.
10422 return std::make_pair(0U, X86::FR32RegisterClass);
10425 return std::make_pair(0U, X86::FR64RegisterClass);
10433 return std::make_pair(0U, X86::VR128RegisterClass);
10439 // Use the default implementation in TargetLowering to convert the register
10440 // constraint into a member of a register class.
10441 std::pair<unsigned, const TargetRegisterClass*> Res;
10442 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10444 // Not found as a standard register?
10445 if (Res.second == 0) {
10446 // Map st(0) -> st(7) -> ST0
10447 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10448 tolower(Constraint[1]) == 's' &&
10449 tolower(Constraint[2]) == 't' &&
10450 Constraint[3] == '(' &&
10451 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10452 Constraint[5] == ')' &&
10453 Constraint[6] == '}') {
10455 Res.first = X86::ST0+Constraint[4]-'0';
10456 Res.second = X86::RFP80RegisterClass;
10460 // GCC allows "st(0)" to be called just plain "st".
10461 if (StringRef("{st}").equals_lower(Constraint)) {
10462 Res.first = X86::ST0;
10463 Res.second = X86::RFP80RegisterClass;
10468 if (StringRef("{flags}").equals_lower(Constraint)) {
10469 Res.first = X86::EFLAGS;
10470 Res.second = X86::CCRRegisterClass;
10474 // 'A' means EAX + EDX.
10475 if (Constraint == "A") {
10476 Res.first = X86::EAX;
10477 Res.second = X86::GR32_ADRegisterClass;
10483 // Otherwise, check to see if this is a register class of the wrong value
10484 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10485 // turn into {ax},{dx}.
10486 if (Res.second->hasType(VT))
10487 return Res; // Correct type already, nothing to do.
10489 // All of the single-register GCC register classes map their values onto
10490 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10491 // really want an 8-bit or 32-bit register, map to the appropriate register
10492 // class and return the appropriate register.
10493 if (Res.second == X86::GR16RegisterClass) {
10494 if (VT == MVT::i8) {
10495 unsigned DestReg = 0;
10496 switch (Res.first) {
10498 case X86::AX: DestReg = X86::AL; break;
10499 case X86::DX: DestReg = X86::DL; break;
10500 case X86::CX: DestReg = X86::CL; break;
10501 case X86::BX: DestReg = X86::BL; break;
10504 Res.first = DestReg;
10505 Res.second = X86::GR8RegisterClass;
10507 } else if (VT == MVT::i32) {
10508 unsigned DestReg = 0;
10509 switch (Res.first) {
10511 case X86::AX: DestReg = X86::EAX; break;
10512 case X86::DX: DestReg = X86::EDX; break;
10513 case X86::CX: DestReg = X86::ECX; break;
10514 case X86::BX: DestReg = X86::EBX; break;
10515 case X86::SI: DestReg = X86::ESI; break;
10516 case X86::DI: DestReg = X86::EDI; break;
10517 case X86::BP: DestReg = X86::EBP; break;
10518 case X86::SP: DestReg = X86::ESP; break;
10521 Res.first = DestReg;
10522 Res.second = X86::GR32RegisterClass;
10524 } else if (VT == MVT::i64) {
10525 unsigned DestReg = 0;
10526 switch (Res.first) {
10528 case X86::AX: DestReg = X86::RAX; break;
10529 case X86::DX: DestReg = X86::RDX; break;
10530 case X86::CX: DestReg = X86::RCX; break;
10531 case X86::BX: DestReg = X86::RBX; break;
10532 case X86::SI: DestReg = X86::RSI; break;
10533 case X86::DI: DestReg = X86::RDI; break;
10534 case X86::BP: DestReg = X86::RBP; break;
10535 case X86::SP: DestReg = X86::RSP; break;
10538 Res.first = DestReg;
10539 Res.second = X86::GR64RegisterClass;
10542 } else if (Res.second == X86::FR32RegisterClass ||
10543 Res.second == X86::FR64RegisterClass ||
10544 Res.second == X86::VR128RegisterClass) {
10545 // Handle references to XMM physical registers that got mapped into the
10546 // wrong class. This can happen with constraints like {xmm0} where the
10547 // target independent register mapper will just pick the first match it can
10548 // find, ignoring the required type.
10549 if (VT == MVT::f32)
10550 Res.second = X86::FR32RegisterClass;
10551 else if (VT == MVT::f64)
10552 Res.second = X86::FR64RegisterClass;
10553 else if (X86::VR128RegisterClass->hasType(VT))
10554 Res.second = X86::VR128RegisterClass;