1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalAlias.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/LLVMContext.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/Dwarf.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/raw_ostream.h"
53 using namespace dwarf;
55 STATISTIC(NumTailCalls, "Number of tail calls");
58 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
60 // Forward declarations.
61 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
64 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
66 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
68 if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
69 if (is64Bit) return new X8664_MachoTargetObjectFile();
70 return new TargetLoweringObjectFileMachO();
71 } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
72 if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
73 return new X8632_ELFTargetObjectFile(TM);
74 } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
75 return new TargetLoweringObjectFileCOFF();
77 llvm_unreachable("unknown subtarget type");
80 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
81 : TargetLowering(TM, createTLOF(TM)) {
82 Subtarget = &TM.getSubtarget<X86Subtarget>();
83 X86ScalarSSEf64 = Subtarget->hasSSE2();
84 X86ScalarSSEf32 = Subtarget->hasSSE1();
85 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
87 RegInfo = TM.getRegisterInfo();
90 // Set up the TargetLowering object.
92 // X86 is weird, it always uses i8 for shift amounts and setcc results.
93 setShiftAmountType(MVT::i8);
94 setBooleanContents(ZeroOrOneBooleanContent);
95 setSchedulingPreference(Sched::RegPressure);
96 setStackPointerRegisterToSaveRestore(X86StackPtr);
98 if (Subtarget->isTargetDarwin()) {
99 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
100 setUseUnderscoreSetJmp(false);
101 setUseUnderscoreLongJmp(false);
102 } else if (Subtarget->isTargetMingw()) {
103 // MS runtime is weird: it exports _setjmp, but longjmp!
104 setUseUnderscoreSetJmp(true);
105 setUseUnderscoreLongJmp(false);
107 setUseUnderscoreSetJmp(true);
108 setUseUnderscoreLongJmp(true);
111 // Set up the register classes.
112 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
113 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
114 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
115 if (Subtarget->is64Bit())
116 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
118 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
120 // We don't accept any truncstore of integer registers.
121 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
122 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
123 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
124 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
125 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
128 // SETOEQ and SETUNE require checking two conditions.
129 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
130 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
131 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
132 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
133 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
136 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
138 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
139 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
140 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
142 if (Subtarget->is64Bit()) {
143 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
144 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
145 } else if (!UseSoftFloat) {
146 // We have an algorithm for SSE2->double, and we turn this into a
147 // 64-bit FILD followed by conditional FADD for other targets.
148 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
149 // We have an algorithm for SSE2, and we turn this into a 64-bit
150 // FILD for other targets.
151 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
154 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
156 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
157 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
160 // SSE has no i16 to fp conversion, only i32
161 if (X86ScalarSSEf32) {
162 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
163 // f32 and f64 cases are Legal, f80 case is not
164 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
166 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
167 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
170 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
171 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
174 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
175 // are Legal, f80 is custom lowered.
176 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
179 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
181 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
182 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
184 if (X86ScalarSSEf32) {
185 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
186 // f32 and f64 cases are Legal, f80 case is not
187 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
189 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
190 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
193 // Handle FP_TO_UINT by promoting the destination to a larger signed
195 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
196 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
197 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
199 if (Subtarget->is64Bit()) {
200 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
201 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
202 } else if (!UseSoftFloat) {
203 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
204 // Expand FP_TO_UINT into a select.
205 // FIXME: We would like to use a Custom expander here eventually to do
206 // the optimal thing for SSE vs. the default expansion in the legalizer.
207 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
209 // With SSE3 we can use fisttpll to convert to a signed i64; without
210 // SSE, we're stuck with a fistpll.
211 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
214 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
215 if (!X86ScalarSSEf64) {
216 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
217 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
218 if (Subtarget->is64Bit()) {
219 setOperationAction(ISD::BIT_CONVERT , MVT::f64 , Expand);
220 // Without SSE, i64->f64 goes through memory; i64->MMX is Legal.
221 if (Subtarget->hasMMX() && !DisableMMX)
222 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Custom);
224 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Expand);
228 // Scalar integer divide and remainder are lowered to use operations that
229 // produce two results, to match the available instructions. This exposes
230 // the two-result form to trivial CSE, which is able to combine x/y and x%y
231 // into a single instruction.
233 // Scalar integer multiply-high is also lowered to use two-result
234 // operations, to match the available instructions. However, plain multiply
235 // (low) operations are left as Legal, as there are single-result
236 // instructions for this in x86. Using the two-result multiply instructions
237 // when both high and low results are needed must be arranged by dagcombine.
238 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
239 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
240 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
241 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
242 setOperationAction(ISD::SREM , MVT::i8 , Expand);
243 setOperationAction(ISD::UREM , MVT::i8 , Expand);
244 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
245 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
246 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
247 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
248 setOperationAction(ISD::SREM , MVT::i16 , Expand);
249 setOperationAction(ISD::UREM , MVT::i16 , Expand);
250 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
251 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
252 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
253 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
254 setOperationAction(ISD::SREM , MVT::i32 , Expand);
255 setOperationAction(ISD::UREM , MVT::i32 , Expand);
256 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
257 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
258 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
259 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
260 setOperationAction(ISD::SREM , MVT::i64 , Expand);
261 setOperationAction(ISD::UREM , MVT::i64 , Expand);
263 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
264 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
265 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
266 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
267 if (Subtarget->is64Bit())
268 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
269 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
270 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
272 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
273 setOperationAction(ISD::FREM , MVT::f32 , Expand);
274 setOperationAction(ISD::FREM , MVT::f64 , Expand);
275 setOperationAction(ISD::FREM , MVT::f80 , Expand);
276 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
278 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
279 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
280 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
281 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
282 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
283 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
287 if (Subtarget->is64Bit()) {
288 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
293 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
294 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
296 // These should be promoted to a larger select which is supported.
297 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
298 // X86 wants to expand cmov itself.
299 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
300 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
301 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
302 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
303 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
304 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
305 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
306 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
307 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
308 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
309 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
310 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
311 if (Subtarget->is64Bit()) {
312 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
313 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
315 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
318 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
319 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
320 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
321 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
322 if (Subtarget->is64Bit())
323 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
324 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
325 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
326 if (Subtarget->is64Bit()) {
327 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
328 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
329 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
330 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
331 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
333 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
334 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
335 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
336 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
337 if (Subtarget->is64Bit()) {
338 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
339 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
340 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
343 if (Subtarget->hasSSE1())
344 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
346 if (!Subtarget->hasSSE2())
347 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
348 // On X86 and X86-64, atomic operations are lowered to locked instructions.
349 // Locked instructions, in turn, have implicit fence semantics (all memory
350 // operations are flushed before issuing the locked instruction, and they
351 // are not buffered), so we can fold away the common pattern of
352 // fence-atomic-fence.
353 setShouldFoldAtomicFences(true);
355 // Expand certain atomics
356 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
361 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
366 if (!Subtarget->is64Bit()) {
367 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
376 // FIXME - use subtarget debug flags
377 if (!Subtarget->isTargetDarwin() &&
378 !Subtarget->isTargetELF() &&
379 !Subtarget->isTargetCygMing()) {
380 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
383 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
384 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
385 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
386 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
387 if (Subtarget->is64Bit()) {
388 setExceptionPointerRegister(X86::RAX);
389 setExceptionSelectorRegister(X86::RDX);
391 setExceptionPointerRegister(X86::EAX);
392 setExceptionSelectorRegister(X86::EDX);
394 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
395 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
397 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
399 setOperationAction(ISD::TRAP, MVT::Other, Legal);
401 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
402 setOperationAction(ISD::VASTART , MVT::Other, Custom);
403 setOperationAction(ISD::VAEND , MVT::Other, Expand);
404 if (Subtarget->is64Bit()) {
405 setOperationAction(ISD::VAARG , MVT::Other, Custom);
406 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
408 setOperationAction(ISD::VAARG , MVT::Other, Expand);
409 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
412 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
413 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
414 if (Subtarget->is64Bit())
415 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
416 if (Subtarget->isTargetCygMing())
417 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
419 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
421 if (!UseSoftFloat && X86ScalarSSEf64) {
422 // f32 and f64 use SSE.
423 // Set up the FP register classes.
424 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
425 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
427 // Use ANDPD to simulate FABS.
428 setOperationAction(ISD::FABS , MVT::f64, Custom);
429 setOperationAction(ISD::FABS , MVT::f32, Custom);
431 // Use XORP to simulate FNEG.
432 setOperationAction(ISD::FNEG , MVT::f64, Custom);
433 setOperationAction(ISD::FNEG , MVT::f32, Custom);
435 // Use ANDPD and ORPD to simulate FCOPYSIGN.
436 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
437 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
439 // We don't support sin/cos/fmod
440 setOperationAction(ISD::FSIN , MVT::f64, Expand);
441 setOperationAction(ISD::FCOS , MVT::f64, Expand);
442 setOperationAction(ISD::FSIN , MVT::f32, Expand);
443 setOperationAction(ISD::FCOS , MVT::f32, Expand);
445 // Expand FP immediates into loads from the stack, except for the special
447 addLegalFPImmediate(APFloat(+0.0)); // xorpd
448 addLegalFPImmediate(APFloat(+0.0f)); // xorps
449 } else if (!UseSoftFloat && X86ScalarSSEf32) {
450 // Use SSE for f32, x87 for f64.
451 // Set up the FP register classes.
452 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
453 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
455 // Use ANDPS to simulate FABS.
456 setOperationAction(ISD::FABS , MVT::f32, Custom);
458 // Use XORP to simulate FNEG.
459 setOperationAction(ISD::FNEG , MVT::f32, Custom);
461 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
463 // Use ANDPS and ORPS to simulate FCOPYSIGN.
464 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
465 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
467 // We don't support sin/cos/fmod
468 setOperationAction(ISD::FSIN , MVT::f32, Expand);
469 setOperationAction(ISD::FCOS , MVT::f32, Expand);
471 // Special cases we handle for FP constants.
472 addLegalFPImmediate(APFloat(+0.0f)); // xorps
473 addLegalFPImmediate(APFloat(+0.0)); // FLD0
474 addLegalFPImmediate(APFloat(+1.0)); // FLD1
475 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
476 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
479 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
480 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
482 } else if (!UseSoftFloat) {
483 // f32 and f64 in x87.
484 // Set up the FP register classes.
485 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
486 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
488 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
489 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
490 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
491 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
494 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
495 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
497 addLegalFPImmediate(APFloat(+0.0)); // FLD0
498 addLegalFPImmediate(APFloat(+1.0)); // FLD1
499 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
500 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
501 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
502 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
503 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
504 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
507 // Long double always uses X87.
509 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
510 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
511 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
514 APFloat TmpFlt(+0.0);
515 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
517 addLegalFPImmediate(TmpFlt); // FLD0
519 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
520 APFloat TmpFlt2(+1.0);
521 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
523 addLegalFPImmediate(TmpFlt2); // FLD1
524 TmpFlt2.changeSign();
525 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
529 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
530 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
534 // Always use a library call for pow.
535 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
536 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
537 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
539 setOperationAction(ISD::FLOG, MVT::f80, Expand);
540 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
541 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
542 setOperationAction(ISD::FEXP, MVT::f80, Expand);
543 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
545 // First set operation action for all vector types to either promote
546 // (for widening) or expand (for scalarization). Then we will selectively
547 // turn on ones that can be effectively codegen'd.
548 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
549 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
550 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
565 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
566 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
599 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
603 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
604 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
605 setTruncStoreAction((MVT::SimpleValueType)VT,
606 (MVT::SimpleValueType)InnerVT, Expand);
607 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
608 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
609 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
612 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
613 // with -msoft-float, disable use of MMX as well.
614 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
615 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
616 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
617 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
619 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
621 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
622 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
623 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
624 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
626 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
627 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
628 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
629 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
631 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
632 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
634 setOperationAction(ISD::AND, MVT::v8i8, Promote);
635 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
636 setOperationAction(ISD::AND, MVT::v4i16, Promote);
637 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
638 setOperationAction(ISD::AND, MVT::v2i32, Promote);
639 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
640 setOperationAction(ISD::AND, MVT::v1i64, Legal);
642 setOperationAction(ISD::OR, MVT::v8i8, Promote);
643 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
644 setOperationAction(ISD::OR, MVT::v4i16, Promote);
645 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
646 setOperationAction(ISD::OR, MVT::v2i32, Promote);
647 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
648 setOperationAction(ISD::OR, MVT::v1i64, Legal);
650 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
651 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
652 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
653 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
654 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
655 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
656 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
658 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
659 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
660 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
661 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
662 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
663 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
664 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
666 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
667 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
669 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
671 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
672 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
674 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
676 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
677 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
678 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
680 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
682 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
683 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
684 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
685 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
686 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
690 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
691 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
692 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
693 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
694 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
698 if (!UseSoftFloat && Subtarget->hasSSE1()) {
699 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
701 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
702 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
703 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
704 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
705 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
706 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
707 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
708 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
709 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
710 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
711 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
712 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
715 if (!UseSoftFloat && Subtarget->hasSSE2()) {
716 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
718 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
719 // registers cannot be used even for integer operations.
720 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
721 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
722 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
723 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
725 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
726 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
727 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
728 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
729 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
730 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
731 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
732 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
733 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
734 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
735 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
736 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
737 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
738 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
739 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
740 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
742 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
743 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
744 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
745 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
747 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
748 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
749 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
750 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
751 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
753 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
754 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
755 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
756 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
757 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
759 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
760 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
761 EVT VT = (MVT::SimpleValueType)i;
762 // Do not attempt to custom lower non-power-of-2 vectors
763 if (!isPowerOf2_32(VT.getVectorNumElements()))
765 // Do not attempt to custom lower non-128-bit vectors
766 if (!VT.is128BitVector())
768 setOperationAction(ISD::BUILD_VECTOR,
769 VT.getSimpleVT().SimpleTy, Custom);
770 setOperationAction(ISD::VECTOR_SHUFFLE,
771 VT.getSimpleVT().SimpleTy, Custom);
772 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
773 VT.getSimpleVT().SimpleTy, Custom);
776 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
777 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
778 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
779 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
780 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
781 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
783 if (Subtarget->is64Bit()) {
784 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
785 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
788 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
789 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
790 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
793 // Do not attempt to promote non-128-bit vectors
794 if (!VT.is128BitVector())
797 setOperationAction(ISD::AND, SVT, Promote);
798 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
799 setOperationAction(ISD::OR, SVT, Promote);
800 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
801 setOperationAction(ISD::XOR, SVT, Promote);
802 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
803 setOperationAction(ISD::LOAD, SVT, Promote);
804 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
805 setOperationAction(ISD::SELECT, SVT, Promote);
806 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
809 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
811 // Custom lower v2i64 and v2f64 selects.
812 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
813 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
814 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
815 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
817 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
818 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
819 if (!DisableMMX && Subtarget->hasMMX()) {
820 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
821 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
825 if (Subtarget->hasSSE41()) {
826 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
827 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
828 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
829 setOperationAction(ISD::FRINT, MVT::f32, Legal);
830 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
831 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
832 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
833 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
834 setOperationAction(ISD::FRINT, MVT::f64, Legal);
835 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
837 // FIXME: Do we need to handle scalar-to-vector here?
838 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
840 // i8 and i16 vectors are custom , because the source register and source
841 // source memory operand types are not the same width. f32 vectors are
842 // custom since the immediate controlling the insert encodes additional
844 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
849 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
851 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
852 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
854 if (Subtarget->is64Bit()) {
855 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
856 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
860 if (Subtarget->hasSSE42()) {
861 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
864 if (!UseSoftFloat && Subtarget->hasAVX()) {
865 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
866 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
867 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
868 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
870 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
871 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
872 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
873 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
874 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
875 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
876 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
877 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
878 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
879 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
880 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
881 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
882 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
883 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
884 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
886 // Operations to consider commented out -v16i16 v32i8
887 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
888 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
889 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
890 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
891 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
892 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
893 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
894 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
895 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
896 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
897 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
898 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
899 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
900 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
902 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
903 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
904 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
905 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
907 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
908 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
909 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
910 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
913 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
914 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
915 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
916 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
917 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
918 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
921 // Not sure we want to do this since there are no 256-bit integer
924 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
925 // This includes 256-bit vectors
926 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
927 EVT VT = (MVT::SimpleValueType)i;
929 // Do not attempt to custom lower non-power-of-2 vectors
930 if (!isPowerOf2_32(VT.getVectorNumElements()))
933 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
934 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
935 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
938 if (Subtarget->is64Bit()) {
939 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
940 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
945 // Not sure we want to do this since there are no 256-bit integer
948 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
949 // Including 256-bit vectors
950 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
951 EVT VT = (MVT::SimpleValueType)i;
953 if (!VT.is256BitVector()) {
956 setOperationAction(ISD::AND, VT, Promote);
957 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
958 setOperationAction(ISD::OR, VT, Promote);
959 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
960 setOperationAction(ISD::XOR, VT, Promote);
961 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
962 setOperationAction(ISD::LOAD, VT, Promote);
963 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
964 setOperationAction(ISD::SELECT, VT, Promote);
965 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
968 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
972 // We want to custom lower some of our intrinsics.
973 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
975 // Add/Sub/Mul with overflow operations are custom lowered.
976 setOperationAction(ISD::SADDO, MVT::i32, Custom);
977 setOperationAction(ISD::UADDO, MVT::i32, Custom);
978 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
979 setOperationAction(ISD::USUBO, MVT::i32, Custom);
980 setOperationAction(ISD::SMULO, MVT::i32, Custom);
982 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
983 // handle type legalization for these operations here.
985 // FIXME: We really should do custom legalization for addition and
986 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
987 // than generic legalization for 64-bit multiplication-with-overflow, though.
988 if (Subtarget->is64Bit()) {
989 setOperationAction(ISD::SADDO, MVT::i64, Custom);
990 setOperationAction(ISD::UADDO, MVT::i64, Custom);
991 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
992 setOperationAction(ISD::USUBO, MVT::i64, Custom);
993 setOperationAction(ISD::SMULO, MVT::i64, Custom);
996 if (!Subtarget->is64Bit()) {
997 // These libcalls are not available in 32-bit.
998 setLibcallName(RTLIB::SHL_I128, 0);
999 setLibcallName(RTLIB::SRL_I128, 0);
1000 setLibcallName(RTLIB::SRA_I128, 0);
1003 // We have target-specific dag combine patterns for the following nodes:
1004 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1005 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1006 setTargetDAGCombine(ISD::BUILD_VECTOR);
1007 setTargetDAGCombine(ISD::SELECT);
1008 setTargetDAGCombine(ISD::SHL);
1009 setTargetDAGCombine(ISD::SRA);
1010 setTargetDAGCombine(ISD::SRL);
1011 setTargetDAGCombine(ISD::OR);
1012 setTargetDAGCombine(ISD::STORE);
1013 setTargetDAGCombine(ISD::ZERO_EXTEND);
1014 if (Subtarget->is64Bit())
1015 setTargetDAGCombine(ISD::MUL);
1017 computeRegisterProperties();
1019 // FIXME: These should be based on subtarget info. Plus, the values should
1020 // be smaller when we are in optimizing for size mode.
1021 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1022 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1023 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1024 setPrefLoopAlignment(16);
1025 benefitFromCodePlacementOpt = true;
1029 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1034 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1035 /// the desired ByVal argument alignment.
1036 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1039 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1040 if (VTy->getBitWidth() == 128)
1042 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1043 unsigned EltAlign = 0;
1044 getMaxByValAlign(ATy->getElementType(), EltAlign);
1045 if (EltAlign > MaxAlign)
1046 MaxAlign = EltAlign;
1047 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1048 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1049 unsigned EltAlign = 0;
1050 getMaxByValAlign(STy->getElementType(i), EltAlign);
1051 if (EltAlign > MaxAlign)
1052 MaxAlign = EltAlign;
1060 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1061 /// function arguments in the caller parameter area. For X86, aggregates
1062 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1063 /// are at 4-byte boundaries.
1064 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1065 if (Subtarget->is64Bit()) {
1066 // Max of 8 and alignment of type.
1067 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1074 if (Subtarget->hasSSE1())
1075 getMaxByValAlign(Ty, Align);
1079 /// getOptimalMemOpType - Returns the target specific optimal type for load
1080 /// and store operations as a result of memset, memcpy, and memmove
1081 /// lowering. If DstAlign is zero that means it's safe to destination
1082 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1083 /// means there isn't a need to check it against alignment requirement,
1084 /// probably because the source does not need to be loaded. If
1085 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1086 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1087 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1088 /// constant so it does not need to be loaded.
1089 /// It returns EVT::Other if the type should be determined using generic
1090 /// target-independent logic.
1092 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1093 unsigned DstAlign, unsigned SrcAlign,
1094 bool NonScalarIntSafe,
1096 MachineFunction &MF) const {
1097 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1098 // linux. This is because the stack realignment code can't handle certain
1099 // cases like PR2962. This should be removed when PR2962 is fixed.
1100 const Function *F = MF.getFunction();
1101 if (NonScalarIntSafe &&
1102 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1104 (Subtarget->isUnalignedMemAccessFast() ||
1105 ((DstAlign == 0 || DstAlign >= 16) &&
1106 (SrcAlign == 0 || SrcAlign >= 16))) &&
1107 Subtarget->getStackAlignment() >= 16) {
1108 if (Subtarget->hasSSE2())
1110 if (Subtarget->hasSSE1())
1112 } else if (!MemcpyStrSrc && Size >= 8 &&
1113 !Subtarget->is64Bit() &&
1114 Subtarget->getStackAlignment() >= 8 &&
1115 Subtarget->hasSSE2()) {
1116 // Do not use f64 to lower memcpy if source is string constant. It's
1117 // better to use i32 to avoid the loads.
1121 if (Subtarget->is64Bit() && Size >= 8)
1126 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1127 /// current function. The returned value is a member of the
1128 /// MachineJumpTableInfo::JTEntryKind enum.
1129 unsigned X86TargetLowering::getJumpTableEncoding() const {
1130 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1132 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1133 Subtarget->isPICStyleGOT())
1134 return MachineJumpTableInfo::EK_Custom32;
1136 // Otherwise, use the normal jump table encoding heuristics.
1137 return TargetLowering::getJumpTableEncoding();
1140 /// getPICBaseSymbol - Return the X86-32 PIC base.
1142 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1143 MCContext &Ctx) const {
1144 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1145 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1146 Twine(MF->getFunctionNumber())+"$pb");
1151 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1152 const MachineBasicBlock *MBB,
1153 unsigned uid,MCContext &Ctx) const{
1154 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1155 Subtarget->isPICStyleGOT());
1156 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1158 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1159 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1162 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1164 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1165 SelectionDAG &DAG) const {
1166 if (!Subtarget->is64Bit())
1167 // This doesn't have DebugLoc associated with it, but is not really the
1168 // same as a Register.
1169 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1173 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1174 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1176 const MCExpr *X86TargetLowering::
1177 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1178 MCContext &Ctx) const {
1179 // X86-64 uses RIP relative addressing based on the jump table label.
1180 if (Subtarget->isPICStyleRIPRel())
1181 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1183 // Otherwise, the reference is relative to the PIC base.
1184 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1187 /// getFunctionAlignment - Return the Log2 alignment of this function.
1188 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1189 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1192 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1193 unsigned &Offset) const {
1194 if (!Subtarget->isTargetLinux())
1197 if (Subtarget->is64Bit()) {
1198 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1200 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1213 //===----------------------------------------------------------------------===//
1214 // Return Value Calling Convention Implementation
1215 //===----------------------------------------------------------------------===//
1217 #include "X86GenCallingConv.inc"
1220 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1221 const SmallVectorImpl<EVT> &OutTys,
1222 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1223 LLVMContext &Context) const {
1224 SmallVector<CCValAssign, 16> RVLocs;
1225 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1227 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1231 X86TargetLowering::LowerReturn(SDValue Chain,
1232 CallingConv::ID CallConv, bool isVarArg,
1233 const SmallVectorImpl<ISD::OutputArg> &Outs,
1234 const SmallVectorImpl<SDValue> &OutVals,
1235 DebugLoc dl, SelectionDAG &DAG) const {
1236 MachineFunction &MF = DAG.getMachineFunction();
1237 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1239 SmallVector<CCValAssign, 16> RVLocs;
1240 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1241 RVLocs, *DAG.getContext());
1242 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1244 // Add the regs to the liveout set for the function.
1245 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1246 for (unsigned i = 0; i != RVLocs.size(); ++i)
1247 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1248 MRI.addLiveOut(RVLocs[i].getLocReg());
1252 SmallVector<SDValue, 6> RetOps;
1253 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1254 // Operand #1 = Bytes To Pop
1255 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1258 // Copy the result values into the output registers.
1259 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1260 CCValAssign &VA = RVLocs[i];
1261 assert(VA.isRegLoc() && "Can only return in registers!");
1262 SDValue ValToCopy = OutVals[i];
1264 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1265 // the RET instruction and handled by the FP Stackifier.
1266 if (VA.getLocReg() == X86::ST0 ||
1267 VA.getLocReg() == X86::ST1) {
1268 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1269 // change the value to the FP stack register class.
1270 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1271 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1272 RetOps.push_back(ValToCopy);
1273 // Don't emit a copytoreg.
1277 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1278 // which is returned in RAX / RDX.
1279 if (Subtarget->is64Bit()) {
1280 EVT ValVT = ValToCopy.getValueType();
1281 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1282 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1283 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1284 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1288 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1289 Flag = Chain.getValue(1);
1292 // The x86-64 ABI for returning structs by value requires that we copy
1293 // the sret argument into %rax for the return. We saved the argument into
1294 // a virtual register in the entry block, so now we copy the value out
1296 if (Subtarget->is64Bit() &&
1297 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1298 MachineFunction &MF = DAG.getMachineFunction();
1299 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1300 unsigned Reg = FuncInfo->getSRetReturnReg();
1302 "SRetReturnReg should have been set in LowerFormalArguments().");
1303 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1305 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1306 Flag = Chain.getValue(1);
1308 // RAX now acts like a return value.
1309 MRI.addLiveOut(X86::RAX);
1312 RetOps[0] = Chain; // Update chain.
1314 // Add the flag if we have it.
1316 RetOps.push_back(Flag);
1318 return DAG.getNode(X86ISD::RET_FLAG, dl,
1319 MVT::Other, &RetOps[0], RetOps.size());
1322 /// LowerCallResult - Lower the result values of a call into the
1323 /// appropriate copies out of appropriate physical registers.
1326 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1327 CallingConv::ID CallConv, bool isVarArg,
1328 const SmallVectorImpl<ISD::InputArg> &Ins,
1329 DebugLoc dl, SelectionDAG &DAG,
1330 SmallVectorImpl<SDValue> &InVals) const {
1332 // Assign locations to each value returned by this call.
1333 SmallVector<CCValAssign, 16> RVLocs;
1334 bool Is64Bit = Subtarget->is64Bit();
1335 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1336 RVLocs, *DAG.getContext());
1337 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1339 // Copy all of the result registers out of their specified physreg.
1340 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1341 CCValAssign &VA = RVLocs[i];
1342 EVT CopyVT = VA.getValVT();
1344 // If this is x86-64, and we disabled SSE, we can't return FP values
1345 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1346 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1347 report_fatal_error("SSE register return with SSE disabled");
1350 // If this is a call to a function that returns an fp value on the floating
1351 // point stack, but where we prefer to use the value in xmm registers, copy
1352 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1353 if ((VA.getLocReg() == X86::ST0 ||
1354 VA.getLocReg() == X86::ST1) &&
1355 isScalarFPTypeInSSEReg(VA.getValVT())) {
1360 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1361 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1362 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1363 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1364 MVT::v2i64, InFlag).getValue(1);
1365 Val = Chain.getValue(0);
1366 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1367 Val, DAG.getConstant(0, MVT::i64));
1369 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1370 MVT::i64, InFlag).getValue(1);
1371 Val = Chain.getValue(0);
1373 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1375 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1376 CopyVT, InFlag).getValue(1);
1377 Val = Chain.getValue(0);
1379 InFlag = Chain.getValue(2);
1381 if (CopyVT != VA.getValVT()) {
1382 // Round the F80 the right size, which also moves to the appropriate xmm
1384 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1385 // This truncation won't change the value.
1386 DAG.getIntPtrConstant(1));
1389 InVals.push_back(Val);
1396 //===----------------------------------------------------------------------===//
1397 // C & StdCall & Fast Calling Convention implementation
1398 //===----------------------------------------------------------------------===//
1399 // StdCall calling convention seems to be standard for many Windows' API
1400 // routines and around. It differs from C calling convention just a little:
1401 // callee should clean up the stack, not caller. Symbols should be also
1402 // decorated in some fancy way :) It doesn't support any vector arguments.
1403 // For info on fast calling convention see Fast Calling Convention (tail call)
1404 // implementation LowerX86_32FastCCCallTo.
1406 /// CallIsStructReturn - Determines whether a call uses struct return
1408 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1412 return Outs[0].Flags.isSRet();
1415 /// ArgsAreStructReturn - Determines whether a function uses struct
1416 /// return semantics.
1418 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1422 return Ins[0].Flags.isSRet();
1425 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1426 /// given CallingConvention value.
1427 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1428 if (Subtarget->is64Bit()) {
1429 if (CC == CallingConv::GHC)
1430 return CC_X86_64_GHC;
1431 else if (Subtarget->isTargetWin64())
1432 return CC_X86_Win64_C;
1437 if (CC == CallingConv::X86_FastCall)
1438 return CC_X86_32_FastCall;
1439 else if (CC == CallingConv::X86_ThisCall)
1440 return CC_X86_32_ThisCall;
1441 else if (CC == CallingConv::Fast)
1442 return CC_X86_32_FastCC;
1443 else if (CC == CallingConv::GHC)
1444 return CC_X86_32_GHC;
1449 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1450 /// by "Src" to address "Dst" with size and alignment information specified by
1451 /// the specific parameter attribute. The copy will be passed as a byval
1452 /// function parameter.
1454 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1455 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1457 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1458 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1459 /*isVolatile*/false, /*AlwaysInline=*/true,
1463 /// IsTailCallConvention - Return true if the calling convention is one that
1464 /// supports tail call optimization.
1465 static bool IsTailCallConvention(CallingConv::ID CC) {
1466 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1469 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1470 /// a tailcall target by changing its ABI.
1471 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1472 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1476 X86TargetLowering::LowerMemArgument(SDValue Chain,
1477 CallingConv::ID CallConv,
1478 const SmallVectorImpl<ISD::InputArg> &Ins,
1479 DebugLoc dl, SelectionDAG &DAG,
1480 const CCValAssign &VA,
1481 MachineFrameInfo *MFI,
1483 // Create the nodes corresponding to a load from this parameter slot.
1484 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1485 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1486 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1489 // If value is passed by pointer we have address passed instead of the value
1491 if (VA.getLocInfo() == CCValAssign::Indirect)
1492 ValVT = VA.getLocVT();
1494 ValVT = VA.getValVT();
1496 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1497 // changed with more analysis.
1498 // In case of tail call optimization mark all arguments mutable. Since they
1499 // could be overwritten by lowering of arguments in case of a tail call.
1500 if (Flags.isByVal()) {
1501 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1502 VA.getLocMemOffset(), isImmutable);
1503 return DAG.getFrameIndex(FI, getPointerTy());
1505 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1506 VA.getLocMemOffset(), isImmutable);
1507 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1508 return DAG.getLoad(ValVT, dl, Chain, FIN,
1509 PseudoSourceValue::getFixedStack(FI), 0,
1515 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1516 CallingConv::ID CallConv,
1518 const SmallVectorImpl<ISD::InputArg> &Ins,
1521 SmallVectorImpl<SDValue> &InVals)
1523 MachineFunction &MF = DAG.getMachineFunction();
1524 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1526 const Function* Fn = MF.getFunction();
1527 if (Fn->hasExternalLinkage() &&
1528 Subtarget->isTargetCygMing() &&
1529 Fn->getName() == "main")
1530 FuncInfo->setForceFramePointer(true);
1532 MachineFrameInfo *MFI = MF.getFrameInfo();
1533 bool Is64Bit = Subtarget->is64Bit();
1534 bool IsWin64 = Subtarget->isTargetWin64();
1536 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1537 "Var args not supported with calling convention fastcc or ghc");
1539 // Assign locations to all of the incoming arguments.
1540 SmallVector<CCValAssign, 16> ArgLocs;
1541 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1542 ArgLocs, *DAG.getContext());
1543 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1545 unsigned LastVal = ~0U;
1547 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1548 CCValAssign &VA = ArgLocs[i];
1549 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1551 assert(VA.getValNo() != LastVal &&
1552 "Don't support value assigned to multiple locs yet");
1553 LastVal = VA.getValNo();
1555 if (VA.isRegLoc()) {
1556 EVT RegVT = VA.getLocVT();
1557 TargetRegisterClass *RC = NULL;
1558 if (RegVT == MVT::i32)
1559 RC = X86::GR32RegisterClass;
1560 else if (Is64Bit && RegVT == MVT::i64)
1561 RC = X86::GR64RegisterClass;
1562 else if (RegVT == MVT::f32)
1563 RC = X86::FR32RegisterClass;
1564 else if (RegVT == MVT::f64)
1565 RC = X86::FR64RegisterClass;
1566 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1567 RC = X86::VR128RegisterClass;
1568 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1569 RC = X86::VR64RegisterClass;
1571 llvm_unreachable("Unknown argument type!");
1573 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1574 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1576 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1577 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1579 if (VA.getLocInfo() == CCValAssign::SExt)
1580 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1581 DAG.getValueType(VA.getValVT()));
1582 else if (VA.getLocInfo() == CCValAssign::ZExt)
1583 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1584 DAG.getValueType(VA.getValVT()));
1585 else if (VA.getLocInfo() == CCValAssign::BCvt)
1586 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1588 if (VA.isExtInLoc()) {
1589 // Handle MMX values passed in XMM regs.
1590 if (RegVT.isVector()) {
1591 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1592 ArgValue, DAG.getConstant(0, MVT::i64));
1593 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1595 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1598 assert(VA.isMemLoc());
1599 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1602 // If value is passed via pointer - do a load.
1603 if (VA.getLocInfo() == CCValAssign::Indirect)
1604 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1607 InVals.push_back(ArgValue);
1610 // The x86-64 ABI for returning structs by value requires that we copy
1611 // the sret argument into %rax for the return. Save the argument into
1612 // a virtual register so that we can access it from the return points.
1613 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1614 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1615 unsigned Reg = FuncInfo->getSRetReturnReg();
1617 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1618 FuncInfo->setSRetReturnReg(Reg);
1620 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1621 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1624 unsigned StackSize = CCInfo.getNextStackOffset();
1625 // Align stack specially for tail calls.
1626 if (FuncIsMadeTailCallSafe(CallConv))
1627 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1629 // If the function takes variable number of arguments, make a frame index for
1630 // the start of the first vararg value... for expansion of llvm.va_start.
1632 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1633 CallConv != CallingConv::X86_ThisCall)) {
1634 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1637 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1639 // FIXME: We should really autogenerate these arrays
1640 static const unsigned GPR64ArgRegsWin64[] = {
1641 X86::RCX, X86::RDX, X86::R8, X86::R9
1643 static const unsigned XMMArgRegsWin64[] = {
1644 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1646 static const unsigned GPR64ArgRegs64Bit[] = {
1647 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1649 static const unsigned XMMArgRegs64Bit[] = {
1650 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1651 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1653 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1656 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1657 GPR64ArgRegs = GPR64ArgRegsWin64;
1658 XMMArgRegs = XMMArgRegsWin64;
1660 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1661 GPR64ArgRegs = GPR64ArgRegs64Bit;
1662 XMMArgRegs = XMMArgRegs64Bit;
1664 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1666 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1669 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1670 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1671 "SSE register cannot be used when SSE is disabled!");
1672 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1673 "SSE register cannot be used when SSE is disabled!");
1674 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1675 // Kernel mode asks for SSE to be disabled, so don't push them
1677 TotalNumXMMRegs = 0;
1679 // For X86-64, if there are vararg parameters that are passed via
1680 // registers, then we must store them to their spots on the stack so they
1681 // may be loaded by deferencing the result of va_next.
1682 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1683 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1684 FuncInfo->setRegSaveFrameIndex(
1685 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1688 // Store the integer parameter registers.
1689 SmallVector<SDValue, 8> MemOps;
1690 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1692 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1693 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1694 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1695 DAG.getIntPtrConstant(Offset));
1696 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1697 X86::GR64RegisterClass);
1698 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1700 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1701 PseudoSourceValue::getFixedStack(
1702 FuncInfo->getRegSaveFrameIndex()),
1703 Offset, false, false, 0);
1704 MemOps.push_back(Store);
1708 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1709 // Now store the XMM (fp + vector) parameter registers.
1710 SmallVector<SDValue, 11> SaveXMMOps;
1711 SaveXMMOps.push_back(Chain);
1713 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1714 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1715 SaveXMMOps.push_back(ALVal);
1717 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1718 FuncInfo->getRegSaveFrameIndex()));
1719 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1720 FuncInfo->getVarArgsFPOffset()));
1722 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1723 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1724 X86::VR128RegisterClass);
1725 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1726 SaveXMMOps.push_back(Val);
1728 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1730 &SaveXMMOps[0], SaveXMMOps.size()));
1733 if (!MemOps.empty())
1734 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1735 &MemOps[0], MemOps.size());
1739 // Some CCs need callee pop.
1740 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1741 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1743 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1744 // If this is an sret function, the return should pop the hidden pointer.
1745 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1746 FuncInfo->setBytesToPopOnReturn(4);
1750 // RegSaveFrameIndex is X86-64 only.
1751 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1752 if (CallConv == CallingConv::X86_FastCall ||
1753 CallConv == CallingConv::X86_ThisCall)
1754 // fastcc functions can't have varargs.
1755 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1762 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1763 SDValue StackPtr, SDValue Arg,
1764 DebugLoc dl, SelectionDAG &DAG,
1765 const CCValAssign &VA,
1766 ISD::ArgFlagsTy Flags) const {
1767 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1768 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1769 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1770 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1771 if (Flags.isByVal()) {
1772 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1774 return DAG.getStore(Chain, dl, Arg, PtrOff,
1775 PseudoSourceValue::getStack(), LocMemOffset,
1779 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1780 /// optimization is performed and it is required.
1782 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1783 SDValue &OutRetAddr, SDValue Chain,
1784 bool IsTailCall, bool Is64Bit,
1785 int FPDiff, DebugLoc dl) const {
1786 // Adjust the Return address stack slot.
1787 EVT VT = getPointerTy();
1788 OutRetAddr = getReturnAddressFrameIndex(DAG);
1790 // Load the "old" Return address.
1791 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1792 return SDValue(OutRetAddr.getNode(), 1);
1795 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1796 /// optimization is performed and it is required (FPDiff!=0).
1798 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1799 SDValue Chain, SDValue RetAddrFrIdx,
1800 bool Is64Bit, int FPDiff, DebugLoc dl) {
1801 // Store the return address to the appropriate stack slot.
1802 if (!FPDiff) return Chain;
1803 // Calculate the new stack slot for the return address.
1804 int SlotSize = Is64Bit ? 8 : 4;
1805 int NewReturnAddrFI =
1806 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1807 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1808 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1809 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1810 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1816 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1817 CallingConv::ID CallConv, bool isVarArg,
1819 const SmallVectorImpl<ISD::OutputArg> &Outs,
1820 const SmallVectorImpl<SDValue> &OutVals,
1821 const SmallVectorImpl<ISD::InputArg> &Ins,
1822 DebugLoc dl, SelectionDAG &DAG,
1823 SmallVectorImpl<SDValue> &InVals) const {
1824 MachineFunction &MF = DAG.getMachineFunction();
1825 bool Is64Bit = Subtarget->is64Bit();
1826 bool IsStructRet = CallIsStructReturn(Outs);
1827 bool IsSibcall = false;
1830 // Check if it's really possible to do a tail call.
1831 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1832 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1833 Outs, OutVals, Ins, DAG);
1835 // Sibcalls are automatically detected tailcalls which do not require
1837 if (!GuaranteedTailCallOpt && isTailCall)
1844 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1845 "Var args not supported with calling convention fastcc or ghc");
1847 // Analyze operands of the call, assigning locations to each operand.
1848 SmallVector<CCValAssign, 16> ArgLocs;
1849 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1850 ArgLocs, *DAG.getContext());
1851 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1853 // Get a count of how many bytes are to be pushed on the stack.
1854 unsigned NumBytes = CCInfo.getNextStackOffset();
1856 // This is a sibcall. The memory operands are available in caller's
1857 // own caller's stack.
1859 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1860 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1863 if (isTailCall && !IsSibcall) {
1864 // Lower arguments at fp - stackoffset + fpdiff.
1865 unsigned NumBytesCallerPushed =
1866 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1867 FPDiff = NumBytesCallerPushed - NumBytes;
1869 // Set the delta of movement of the returnaddr stackslot.
1870 // But only set if delta is greater than previous delta.
1871 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1872 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1876 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1878 SDValue RetAddrFrIdx;
1879 // Load return adress for tail calls.
1880 if (isTailCall && FPDiff)
1881 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1882 Is64Bit, FPDiff, dl);
1884 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1885 SmallVector<SDValue, 8> MemOpChains;
1888 // Walk the register/memloc assignments, inserting copies/loads. In the case
1889 // of tail call optimization arguments are handle later.
1890 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1891 CCValAssign &VA = ArgLocs[i];
1892 EVT RegVT = VA.getLocVT();
1893 SDValue Arg = OutVals[i];
1894 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1895 bool isByVal = Flags.isByVal();
1897 // Promote the value if needed.
1898 switch (VA.getLocInfo()) {
1899 default: llvm_unreachable("Unknown loc info!");
1900 case CCValAssign::Full: break;
1901 case CCValAssign::SExt:
1902 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1904 case CCValAssign::ZExt:
1905 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1907 case CCValAssign::AExt:
1908 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1909 // Special case: passing MMX values in XMM registers.
1910 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1911 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1912 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1914 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1916 case CCValAssign::BCvt:
1917 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1919 case CCValAssign::Indirect: {
1920 // Store the argument.
1921 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1922 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1923 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1924 PseudoSourceValue::getFixedStack(FI), 0,
1931 if (VA.isRegLoc()) {
1932 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1933 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1934 assert(VA.isMemLoc());
1935 if (StackPtr.getNode() == 0)
1936 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1937 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1938 dl, DAG, VA, Flags));
1942 if (!MemOpChains.empty())
1943 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1944 &MemOpChains[0], MemOpChains.size());
1946 // Build a sequence of copy-to-reg nodes chained together with token chain
1947 // and flag operands which copy the outgoing args into registers.
1949 // Tail call byval lowering might overwrite argument registers so in case of
1950 // tail call optimization the copies to registers are lowered later.
1952 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1953 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1954 RegsToPass[i].second, InFlag);
1955 InFlag = Chain.getValue(1);
1958 if (Subtarget->isPICStyleGOT()) {
1959 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1962 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1963 DAG.getNode(X86ISD::GlobalBaseReg,
1964 DebugLoc(), getPointerTy()),
1966 InFlag = Chain.getValue(1);
1968 // If we are tail calling and generating PIC/GOT style code load the
1969 // address of the callee into ECX. The value in ecx is used as target of
1970 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1971 // for tail calls on PIC/GOT architectures. Normally we would just put the
1972 // address of GOT into ebx and then call target@PLT. But for tail calls
1973 // ebx would be restored (since ebx is callee saved) before jumping to the
1976 // Note: The actual moving to ECX is done further down.
1977 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1978 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1979 !G->getGlobal()->hasProtectedVisibility())
1980 Callee = LowerGlobalAddress(Callee, DAG);
1981 else if (isa<ExternalSymbolSDNode>(Callee))
1982 Callee = LowerExternalSymbol(Callee, DAG);
1986 if (Is64Bit && isVarArg) {
1987 // From AMD64 ABI document:
1988 // For calls that may call functions that use varargs or stdargs
1989 // (prototype-less calls or calls to functions containing ellipsis (...) in
1990 // the declaration) %al is used as hidden argument to specify the number
1991 // of SSE registers used. The contents of %al do not need to match exactly
1992 // the number of registers, but must be an ubound on the number of SSE
1993 // registers used and is in the range 0 - 8 inclusive.
1995 // FIXME: Verify this on Win64
1996 // Count the number of XMM registers allocated.
1997 static const unsigned XMMArgRegs[] = {
1998 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1999 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2001 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2002 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2003 && "SSE registers cannot be used when SSE is disabled");
2005 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2006 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2007 InFlag = Chain.getValue(1);
2011 // For tail calls lower the arguments to the 'real' stack slot.
2013 // Force all the incoming stack arguments to be loaded from the stack
2014 // before any new outgoing arguments are stored to the stack, because the
2015 // outgoing stack slots may alias the incoming argument stack slots, and
2016 // the alias isn't otherwise explicit. This is slightly more conservative
2017 // than necessary, because it means that each store effectively depends
2018 // on every argument instead of just those arguments it would clobber.
2019 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2021 SmallVector<SDValue, 8> MemOpChains2;
2024 // Do not flag preceeding copytoreg stuff together with the following stuff.
2026 if (GuaranteedTailCallOpt) {
2027 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2028 CCValAssign &VA = ArgLocs[i];
2031 assert(VA.isMemLoc());
2032 SDValue Arg = OutVals[i];
2033 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2034 // Create frame index.
2035 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2036 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2037 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2038 FIN = DAG.getFrameIndex(FI, getPointerTy());
2040 if (Flags.isByVal()) {
2041 // Copy relative to framepointer.
2042 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2043 if (StackPtr.getNode() == 0)
2044 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2046 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2048 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2052 // Store relative to framepointer.
2053 MemOpChains2.push_back(
2054 DAG.getStore(ArgChain, dl, Arg, FIN,
2055 PseudoSourceValue::getFixedStack(FI), 0,
2061 if (!MemOpChains2.empty())
2062 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2063 &MemOpChains2[0], MemOpChains2.size());
2065 // Copy arguments to their registers.
2066 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2067 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2068 RegsToPass[i].second, InFlag);
2069 InFlag = Chain.getValue(1);
2073 // Store the return address to the appropriate stack slot.
2074 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2078 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2079 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2080 // In the 64-bit large code model, we have to make all calls
2081 // through a register, since the call instruction's 32-bit
2082 // pc-relative offset may not be large enough to hold the whole
2084 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2085 // If the callee is a GlobalAddress node (quite common, every direct call
2086 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2089 // We should use extra load for direct calls to dllimported functions in
2091 const GlobalValue *GV = G->getGlobal();
2092 if (!GV->hasDLLImportLinkage()) {
2093 unsigned char OpFlags = 0;
2095 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2096 // external symbols most go through the PLT in PIC mode. If the symbol
2097 // has hidden or protected visibility, or if it is static or local, then
2098 // we don't need to use the PLT - we can directly call it.
2099 if (Subtarget->isTargetELF() &&
2100 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2101 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2102 OpFlags = X86II::MO_PLT;
2103 } else if (Subtarget->isPICStyleStubAny() &&
2104 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2105 Subtarget->getDarwinVers() < 9) {
2106 // PC-relative references to external symbols should go through $stub,
2107 // unless we're building with the leopard linker or later, which
2108 // automatically synthesizes these stubs.
2109 OpFlags = X86II::MO_DARWIN_STUB;
2112 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2113 G->getOffset(), OpFlags);
2115 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2116 unsigned char OpFlags = 0;
2118 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2119 // symbols should go through the PLT.
2120 if (Subtarget->isTargetELF() &&
2121 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2122 OpFlags = X86II::MO_PLT;
2123 } else if (Subtarget->isPICStyleStubAny() &&
2124 Subtarget->getDarwinVers() < 9) {
2125 // PC-relative references to external symbols should go through $stub,
2126 // unless we're building with the leopard linker or later, which
2127 // automatically synthesizes these stubs.
2128 OpFlags = X86II::MO_DARWIN_STUB;
2131 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2135 // Returns a chain & a flag for retval copy to use.
2136 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2137 SmallVector<SDValue, 8> Ops;
2139 if (!IsSibcall && isTailCall) {
2140 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2141 DAG.getIntPtrConstant(0, true), InFlag);
2142 InFlag = Chain.getValue(1);
2145 Ops.push_back(Chain);
2146 Ops.push_back(Callee);
2149 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2151 // Add argument registers to the end of the list so that they are known live
2153 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2154 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2155 RegsToPass[i].second.getValueType()));
2157 // Add an implicit use GOT pointer in EBX.
2158 if (!isTailCall && Subtarget->isPICStyleGOT())
2159 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2161 // Add an implicit use of AL for x86 vararg functions.
2162 if (Is64Bit && isVarArg)
2163 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2165 if (InFlag.getNode())
2166 Ops.push_back(InFlag);
2170 //// If this is the first return lowered for this function, add the regs
2171 //// to the liveout set for the function.
2172 // This isn't right, although it's probably harmless on x86; liveouts
2173 // should be computed from returns not tail calls. Consider a void
2174 // function making a tail call to a function returning int.
2175 return DAG.getNode(X86ISD::TC_RETURN, dl,
2176 NodeTys, &Ops[0], Ops.size());
2179 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2180 InFlag = Chain.getValue(1);
2182 // Create the CALLSEQ_END node.
2183 unsigned NumBytesForCalleeToPush;
2184 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2185 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2186 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2187 // If this is a call to a struct-return function, the callee
2188 // pops the hidden struct pointer, so we have to push it back.
2189 // This is common for Darwin/X86, Linux & Mingw32 targets.
2190 NumBytesForCalleeToPush = 4;
2192 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2194 // Returns a flag for retval copy to use.
2196 Chain = DAG.getCALLSEQ_END(Chain,
2197 DAG.getIntPtrConstant(NumBytes, true),
2198 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2201 InFlag = Chain.getValue(1);
2204 // Handle result values, copying them out of physregs into vregs that we
2206 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2207 Ins, dl, DAG, InVals);
2211 //===----------------------------------------------------------------------===//
2212 // Fast Calling Convention (tail call) implementation
2213 //===----------------------------------------------------------------------===//
2215 // Like std call, callee cleans arguments, convention except that ECX is
2216 // reserved for storing the tail called function address. Only 2 registers are
2217 // free for argument passing (inreg). Tail call optimization is performed
2219 // * tailcallopt is enabled
2220 // * caller/callee are fastcc
2221 // On X86_64 architecture with GOT-style position independent code only local
2222 // (within module) calls are supported at the moment.
2223 // To keep the stack aligned according to platform abi the function
2224 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2225 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2226 // If a tail called function callee has more arguments than the caller the
2227 // caller needs to make sure that there is room to move the RETADDR to. This is
2228 // achieved by reserving an area the size of the argument delta right after the
2229 // original REtADDR, but before the saved framepointer or the spilled registers
2230 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2242 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2243 /// for a 16 byte align requirement.
2245 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2246 SelectionDAG& DAG) const {
2247 MachineFunction &MF = DAG.getMachineFunction();
2248 const TargetMachine &TM = MF.getTarget();
2249 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2250 unsigned StackAlignment = TFI.getStackAlignment();
2251 uint64_t AlignMask = StackAlignment - 1;
2252 int64_t Offset = StackSize;
2253 uint64_t SlotSize = TD->getPointerSize();
2254 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2255 // Number smaller than 12 so just add the difference.
2256 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2258 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2259 Offset = ((~AlignMask) & Offset) + StackAlignment +
2260 (StackAlignment-SlotSize);
2265 /// MatchingStackOffset - Return true if the given stack call argument is
2266 /// already available in the same position (relatively) of the caller's
2267 /// incoming argument stack.
2269 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2270 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2271 const X86InstrInfo *TII) {
2272 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2274 if (Arg.getOpcode() == ISD::CopyFromReg) {
2275 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2276 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2278 MachineInstr *Def = MRI->getVRegDef(VR);
2281 if (!Flags.isByVal()) {
2282 if (!TII->isLoadFromStackSlot(Def, FI))
2285 unsigned Opcode = Def->getOpcode();
2286 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2287 Def->getOperand(1).isFI()) {
2288 FI = Def->getOperand(1).getIndex();
2289 Bytes = Flags.getByValSize();
2293 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2294 if (Flags.isByVal())
2295 // ByVal argument is passed in as a pointer but it's now being
2296 // dereferenced. e.g.
2297 // define @foo(%struct.X* %A) {
2298 // tail call @bar(%struct.X* byval %A)
2301 SDValue Ptr = Ld->getBasePtr();
2302 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2305 FI = FINode->getIndex();
2309 assert(FI != INT_MAX);
2310 if (!MFI->isFixedObjectIndex(FI))
2312 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2315 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2316 /// for tail call optimization. Targets which want to do tail call
2317 /// optimization should implement this function.
2319 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2320 CallingConv::ID CalleeCC,
2322 bool isCalleeStructRet,
2323 bool isCallerStructRet,
2324 const SmallVectorImpl<ISD::OutputArg> &Outs,
2325 const SmallVectorImpl<SDValue> &OutVals,
2326 const SmallVectorImpl<ISD::InputArg> &Ins,
2327 SelectionDAG& DAG) const {
2328 if (!IsTailCallConvention(CalleeCC) &&
2329 CalleeCC != CallingConv::C)
2332 // If -tailcallopt is specified, make fastcc functions tail-callable.
2333 const MachineFunction &MF = DAG.getMachineFunction();
2334 const Function *CallerF = DAG.getMachineFunction().getFunction();
2335 CallingConv::ID CallerCC = CallerF->getCallingConv();
2336 bool CCMatch = CallerCC == CalleeCC;
2338 if (GuaranteedTailCallOpt) {
2339 if (IsTailCallConvention(CalleeCC) && CCMatch)
2344 // Look for obvious safe cases to perform tail call optimization that do not
2345 // require ABI changes. This is what gcc calls sibcall.
2347 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2348 // emit a special epilogue.
2349 if (RegInfo->needsStackRealignment(MF))
2352 // Do not sibcall optimize vararg calls unless the call site is not passing any
2354 if (isVarArg && !Outs.empty())
2357 // Also avoid sibcall optimization if either caller or callee uses struct
2358 // return semantics.
2359 if (isCalleeStructRet || isCallerStructRet)
2362 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2363 // Therefore if it's not used by the call it is not safe to optimize this into
2365 bool Unused = false;
2366 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2373 SmallVector<CCValAssign, 16> RVLocs;
2374 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2375 RVLocs, *DAG.getContext());
2376 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2377 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2378 CCValAssign &VA = RVLocs[i];
2379 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2384 // If the calling conventions do not match, then we'd better make sure the
2385 // results are returned in the same way as what the caller expects.
2387 SmallVector<CCValAssign, 16> RVLocs1;
2388 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2389 RVLocs1, *DAG.getContext());
2390 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2392 SmallVector<CCValAssign, 16> RVLocs2;
2393 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2394 RVLocs2, *DAG.getContext());
2395 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2397 if (RVLocs1.size() != RVLocs2.size())
2399 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2400 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2402 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2404 if (RVLocs1[i].isRegLoc()) {
2405 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2408 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2414 // If the callee takes no arguments then go on to check the results of the
2416 if (!Outs.empty()) {
2417 // Check if stack adjustment is needed. For now, do not do this if any
2418 // argument is passed on the stack.
2419 SmallVector<CCValAssign, 16> ArgLocs;
2420 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2421 ArgLocs, *DAG.getContext());
2422 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2423 if (CCInfo.getNextStackOffset()) {
2424 MachineFunction &MF = DAG.getMachineFunction();
2425 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2427 if (Subtarget->isTargetWin64())
2428 // Win64 ABI has additional complications.
2431 // Check if the arguments are already laid out in the right way as
2432 // the caller's fixed stack objects.
2433 MachineFrameInfo *MFI = MF.getFrameInfo();
2434 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2435 const X86InstrInfo *TII =
2436 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2437 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2438 CCValAssign &VA = ArgLocs[i];
2439 SDValue Arg = OutVals[i];
2440 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2441 if (VA.getLocInfo() == CCValAssign::Indirect)
2443 if (!VA.isRegLoc()) {
2444 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2451 // If the tailcall address may be in a register, then make sure it's
2452 // possible to register allocate for it. In 32-bit, the call address can
2453 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2454 // callee-saved registers are restored. In 64-bit, it's RAX, RCX, RDX, RSI,
2455 // RDI, R8, R9, R11.
2456 if (!isa<GlobalAddressSDNode>(Callee) &&
2457 !isa<ExternalSymbolSDNode>(Callee)) {
2458 unsigned Limit = Subtarget->is64Bit() ? 8 : 3;
2459 unsigned NumInRegs = 0;
2460 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2461 CCValAssign &VA = ArgLocs[i];
2462 if (VA.isRegLoc()) {
2463 if (++NumInRegs == Limit)
2474 X86TargetLowering::createFastISel(MachineFunction &mf,
2475 DenseMap<const Value *, unsigned> &vm,
2476 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2477 DenseMap<const AllocaInst *, int> &am,
2478 std::vector<std::pair<MachineInstr*, unsigned> > &pn
2480 , SmallSet<const Instruction *, 8> &cil
2483 return X86::createFastISel(mf, vm, bm, am, pn
2491 //===----------------------------------------------------------------------===//
2492 // Other Lowering Hooks
2493 //===----------------------------------------------------------------------===//
2496 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2497 MachineFunction &MF = DAG.getMachineFunction();
2498 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2499 int ReturnAddrIndex = FuncInfo->getRAIndex();
2501 if (ReturnAddrIndex == 0) {
2502 // Set up a frame object for the return address.
2503 uint64_t SlotSize = TD->getPointerSize();
2504 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2506 FuncInfo->setRAIndex(ReturnAddrIndex);
2509 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2513 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2514 bool hasSymbolicDisplacement) {
2515 // Offset should fit into 32 bit immediate field.
2516 if (!isInt<32>(Offset))
2519 // If we don't have a symbolic displacement - we don't have any extra
2521 if (!hasSymbolicDisplacement)
2524 // FIXME: Some tweaks might be needed for medium code model.
2525 if (M != CodeModel::Small && M != CodeModel::Kernel)
2528 // For small code model we assume that latest object is 16MB before end of 31
2529 // bits boundary. We may also accept pretty large negative constants knowing
2530 // that all objects are in the positive half of address space.
2531 if (M == CodeModel::Small && Offset < 16*1024*1024)
2534 // For kernel code model we know that all object resist in the negative half
2535 // of 32bits address space. We may not accept negative offsets, since they may
2536 // be just off and we may accept pretty large positive ones.
2537 if (M == CodeModel::Kernel && Offset > 0)
2543 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2544 /// specific condition code, returning the condition code and the LHS/RHS of the
2545 /// comparison to make.
2546 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2547 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2549 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2550 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2551 // X > -1 -> X == 0, jump !sign.
2552 RHS = DAG.getConstant(0, RHS.getValueType());
2553 return X86::COND_NS;
2554 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2555 // X < 0 -> X == 0, jump on sign.
2557 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2559 RHS = DAG.getConstant(0, RHS.getValueType());
2560 return X86::COND_LE;
2564 switch (SetCCOpcode) {
2565 default: llvm_unreachable("Invalid integer condition!");
2566 case ISD::SETEQ: return X86::COND_E;
2567 case ISD::SETGT: return X86::COND_G;
2568 case ISD::SETGE: return X86::COND_GE;
2569 case ISD::SETLT: return X86::COND_L;
2570 case ISD::SETLE: return X86::COND_LE;
2571 case ISD::SETNE: return X86::COND_NE;
2572 case ISD::SETULT: return X86::COND_B;
2573 case ISD::SETUGT: return X86::COND_A;
2574 case ISD::SETULE: return X86::COND_BE;
2575 case ISD::SETUGE: return X86::COND_AE;
2579 // First determine if it is required or is profitable to flip the operands.
2581 // If LHS is a foldable load, but RHS is not, flip the condition.
2582 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2583 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2584 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2585 std::swap(LHS, RHS);
2588 switch (SetCCOpcode) {
2594 std::swap(LHS, RHS);
2598 // On a floating point condition, the flags are set as follows:
2600 // 0 | 0 | 0 | X > Y
2601 // 0 | 0 | 1 | X < Y
2602 // 1 | 0 | 0 | X == Y
2603 // 1 | 1 | 1 | unordered
2604 switch (SetCCOpcode) {
2605 default: llvm_unreachable("Condcode should be pre-legalized away");
2607 case ISD::SETEQ: return X86::COND_E;
2608 case ISD::SETOLT: // flipped
2610 case ISD::SETGT: return X86::COND_A;
2611 case ISD::SETOLE: // flipped
2613 case ISD::SETGE: return X86::COND_AE;
2614 case ISD::SETUGT: // flipped
2616 case ISD::SETLT: return X86::COND_B;
2617 case ISD::SETUGE: // flipped
2619 case ISD::SETLE: return X86::COND_BE;
2621 case ISD::SETNE: return X86::COND_NE;
2622 case ISD::SETUO: return X86::COND_P;
2623 case ISD::SETO: return X86::COND_NP;
2625 case ISD::SETUNE: return X86::COND_INVALID;
2629 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2630 /// code. Current x86 isa includes the following FP cmov instructions:
2631 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2632 static bool hasFPCMov(unsigned X86CC) {
2648 /// isFPImmLegal - Returns true if the target can instruction select the
2649 /// specified FP immediate natively. If false, the legalizer will
2650 /// materialize the FP immediate as a load from a constant pool.
2651 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2652 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2653 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2659 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2660 /// the specified range (L, H].
2661 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2662 return (Val < 0) || (Val >= Low && Val < Hi);
2665 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2666 /// specified value.
2667 static bool isUndefOrEqual(int Val, int CmpVal) {
2668 if (Val < 0 || Val == CmpVal)
2673 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2674 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2675 /// the second operand.
2676 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2677 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2678 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2679 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2680 return (Mask[0] < 2 && Mask[1] < 2);
2684 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2685 SmallVector<int, 8> M;
2687 return ::isPSHUFDMask(M, N->getValueType(0));
2690 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2691 /// is suitable for input to PSHUFHW.
2692 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2693 if (VT != MVT::v8i16)
2696 // Lower quadword copied in order or undef.
2697 for (int i = 0; i != 4; ++i)
2698 if (Mask[i] >= 0 && Mask[i] != i)
2701 // Upper quadword shuffled.
2702 for (int i = 4; i != 8; ++i)
2703 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2709 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2710 SmallVector<int, 8> M;
2712 return ::isPSHUFHWMask(M, N->getValueType(0));
2715 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2716 /// is suitable for input to PSHUFLW.
2717 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2718 if (VT != MVT::v8i16)
2721 // Upper quadword copied in order.
2722 for (int i = 4; i != 8; ++i)
2723 if (Mask[i] >= 0 && Mask[i] != i)
2726 // Lower quadword shuffled.
2727 for (int i = 0; i != 4; ++i)
2734 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2735 SmallVector<int, 8> M;
2737 return ::isPSHUFLWMask(M, N->getValueType(0));
2740 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2741 /// is suitable for input to PALIGNR.
2742 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2744 int i, e = VT.getVectorNumElements();
2746 // Do not handle v2i64 / v2f64 shuffles with palignr.
2747 if (e < 4 || !hasSSSE3)
2750 for (i = 0; i != e; ++i)
2754 // All undef, not a palignr.
2758 // Determine if it's ok to perform a palignr with only the LHS, since we
2759 // don't have access to the actual shuffle elements to see if RHS is undef.
2760 bool Unary = Mask[i] < (int)e;
2761 bool NeedsUnary = false;
2763 int s = Mask[i] - i;
2765 // Check the rest of the elements to see if they are consecutive.
2766 for (++i; i != e; ++i) {
2771 Unary = Unary && (m < (int)e);
2772 NeedsUnary = NeedsUnary || (m < s);
2774 if (NeedsUnary && !Unary)
2776 if (Unary && m != ((s+i) & (e-1)))
2778 if (!Unary && m != (s+i))
2784 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2785 SmallVector<int, 8> M;
2787 return ::isPALIGNRMask(M, N->getValueType(0), true);
2790 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2791 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2792 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2793 int NumElems = VT.getVectorNumElements();
2794 if (NumElems != 2 && NumElems != 4)
2797 int Half = NumElems / 2;
2798 for (int i = 0; i < Half; ++i)
2799 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2801 for (int i = Half; i < NumElems; ++i)
2802 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2808 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2809 SmallVector<int, 8> M;
2811 return ::isSHUFPMask(M, N->getValueType(0));
2814 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2815 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2816 /// half elements to come from vector 1 (which would equal the dest.) and
2817 /// the upper half to come from vector 2.
2818 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2819 int NumElems = VT.getVectorNumElements();
2821 if (NumElems != 2 && NumElems != 4)
2824 int Half = NumElems / 2;
2825 for (int i = 0; i < Half; ++i)
2826 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2828 for (int i = Half; i < NumElems; ++i)
2829 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2834 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2835 SmallVector<int, 8> M;
2837 return isCommutedSHUFPMask(M, N->getValueType(0));
2840 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2841 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2842 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2843 if (N->getValueType(0).getVectorNumElements() != 4)
2846 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2847 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2848 isUndefOrEqual(N->getMaskElt(1), 7) &&
2849 isUndefOrEqual(N->getMaskElt(2), 2) &&
2850 isUndefOrEqual(N->getMaskElt(3), 3);
2853 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2854 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2856 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2857 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2862 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2863 isUndefOrEqual(N->getMaskElt(1), 3) &&
2864 isUndefOrEqual(N->getMaskElt(2), 2) &&
2865 isUndefOrEqual(N->getMaskElt(3), 3);
2868 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2869 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2870 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2871 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2873 if (NumElems != 2 && NumElems != 4)
2876 for (unsigned i = 0; i < NumElems/2; ++i)
2877 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2880 for (unsigned i = NumElems/2; i < NumElems; ++i)
2881 if (!isUndefOrEqual(N->getMaskElt(i), i))
2887 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2888 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2889 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2890 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2892 if (NumElems != 2 && NumElems != 4)
2895 for (unsigned i = 0; i < NumElems/2; ++i)
2896 if (!isUndefOrEqual(N->getMaskElt(i), i))
2899 for (unsigned i = 0; i < NumElems/2; ++i)
2900 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2906 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2907 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2908 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2909 bool V2IsSplat = false) {
2910 int NumElts = VT.getVectorNumElements();
2911 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2914 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2916 int BitI1 = Mask[i+1];
2917 if (!isUndefOrEqual(BitI, j))
2920 if (!isUndefOrEqual(BitI1, NumElts))
2923 if (!isUndefOrEqual(BitI1, j + NumElts))
2930 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2931 SmallVector<int, 8> M;
2933 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2936 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2937 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2938 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2939 bool V2IsSplat = false) {
2940 int NumElts = VT.getVectorNumElements();
2941 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2944 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2946 int BitI1 = Mask[i+1];
2947 if (!isUndefOrEqual(BitI, j + NumElts/2))
2950 if (isUndefOrEqual(BitI1, NumElts))
2953 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2960 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2961 SmallVector<int, 8> M;
2963 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2966 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2967 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2969 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2970 int NumElems = VT.getVectorNumElements();
2971 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2974 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2976 int BitI1 = Mask[i+1];
2977 if (!isUndefOrEqual(BitI, j))
2979 if (!isUndefOrEqual(BitI1, j))
2985 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2986 SmallVector<int, 8> M;
2988 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2991 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2992 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2994 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2995 int NumElems = VT.getVectorNumElements();
2996 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2999 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3001 int BitI1 = Mask[i+1];
3002 if (!isUndefOrEqual(BitI, j))
3004 if (!isUndefOrEqual(BitI1, j))
3010 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3011 SmallVector<int, 8> M;
3013 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3016 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3017 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3018 /// MOVSD, and MOVD, i.e. setting the lowest element.
3019 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3020 if (VT.getVectorElementType().getSizeInBits() < 32)
3023 int NumElts = VT.getVectorNumElements();
3025 if (!isUndefOrEqual(Mask[0], NumElts))
3028 for (int i = 1; i < NumElts; ++i)
3029 if (!isUndefOrEqual(Mask[i], i))
3035 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3036 SmallVector<int, 8> M;
3038 return ::isMOVLMask(M, N->getValueType(0));
3041 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3042 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3043 /// element of vector 2 and the other elements to come from vector 1 in order.
3044 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3045 bool V2IsSplat = false, bool V2IsUndef = false) {
3046 int NumOps = VT.getVectorNumElements();
3047 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3050 if (!isUndefOrEqual(Mask[0], 0))
3053 for (int i = 1; i < NumOps; ++i)
3054 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3055 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3056 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3062 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3063 bool V2IsUndef = false) {
3064 SmallVector<int, 8> M;
3066 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3069 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3070 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3071 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3072 if (N->getValueType(0).getVectorNumElements() != 4)
3075 // Expect 1, 1, 3, 3
3076 for (unsigned i = 0; i < 2; ++i) {
3077 int Elt = N->getMaskElt(i);
3078 if (Elt >= 0 && Elt != 1)
3083 for (unsigned i = 2; i < 4; ++i) {
3084 int Elt = N->getMaskElt(i);
3085 if (Elt >= 0 && Elt != 3)
3090 // Don't use movshdup if it can be done with a shufps.
3091 // FIXME: verify that matching u, u, 3, 3 is what we want.
3095 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3096 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3097 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3098 if (N->getValueType(0).getVectorNumElements() != 4)
3101 // Expect 0, 0, 2, 2
3102 for (unsigned i = 0; i < 2; ++i)
3103 if (N->getMaskElt(i) > 0)
3107 for (unsigned i = 2; i < 4; ++i) {
3108 int Elt = N->getMaskElt(i);
3109 if (Elt >= 0 && Elt != 2)
3114 // Don't use movsldup if it can be done with a shufps.
3118 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3119 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3120 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3121 int e = N->getValueType(0).getVectorNumElements() / 2;
3123 for (int i = 0; i < e; ++i)
3124 if (!isUndefOrEqual(N->getMaskElt(i), i))
3126 for (int i = 0; i < e; ++i)
3127 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3132 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3133 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3134 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3135 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3136 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3138 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3140 for (int i = 0; i < NumOperands; ++i) {
3141 int Val = SVOp->getMaskElt(NumOperands-i-1);
3142 if (Val < 0) Val = 0;
3143 if (Val >= NumOperands) Val -= NumOperands;
3145 if (i != NumOperands - 1)
3151 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3152 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3153 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3154 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3156 // 8 nodes, but we only care about the last 4.
3157 for (unsigned i = 7; i >= 4; --i) {
3158 int Val = SVOp->getMaskElt(i);
3167 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3168 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3169 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3170 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3172 // 8 nodes, but we only care about the first 4.
3173 for (int i = 3; i >= 0; --i) {
3174 int Val = SVOp->getMaskElt(i);
3183 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3184 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3185 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3186 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3187 EVT VVT = N->getValueType(0);
3188 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3192 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3193 Val = SVOp->getMaskElt(i);
3197 return (Val - i) * EltSize;
3200 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3202 bool X86::isZeroNode(SDValue Elt) {
3203 return ((isa<ConstantSDNode>(Elt) &&
3204 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3205 (isa<ConstantFPSDNode>(Elt) &&
3206 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3209 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3210 /// their permute mask.
3211 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3212 SelectionDAG &DAG) {
3213 EVT VT = SVOp->getValueType(0);
3214 unsigned NumElems = VT.getVectorNumElements();
3215 SmallVector<int, 8> MaskVec;
3217 for (unsigned i = 0; i != NumElems; ++i) {
3218 int idx = SVOp->getMaskElt(i);
3220 MaskVec.push_back(idx);
3221 else if (idx < (int)NumElems)
3222 MaskVec.push_back(idx + NumElems);
3224 MaskVec.push_back(idx - NumElems);
3226 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3227 SVOp->getOperand(0), &MaskVec[0]);
3230 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3231 /// the two vector operands have swapped position.
3232 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3233 unsigned NumElems = VT.getVectorNumElements();
3234 for (unsigned i = 0; i != NumElems; ++i) {
3238 else if (idx < (int)NumElems)
3239 Mask[i] = idx + NumElems;
3241 Mask[i] = idx - NumElems;
3245 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3246 /// match movhlps. The lower half elements should come from upper half of
3247 /// V1 (and in order), and the upper half elements should come from the upper
3248 /// half of V2 (and in order).
3249 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3250 if (Op->getValueType(0).getVectorNumElements() != 4)
3252 for (unsigned i = 0, e = 2; i != e; ++i)
3253 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3255 for (unsigned i = 2; i != 4; ++i)
3256 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3261 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3262 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3264 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3265 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3267 N = N->getOperand(0).getNode();
3268 if (!ISD::isNON_EXTLoad(N))
3271 *LD = cast<LoadSDNode>(N);
3275 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3276 /// match movlp{s|d}. The lower half elements should come from lower half of
3277 /// V1 (and in order), and the upper half elements should come from the upper
3278 /// half of V2 (and in order). And since V1 will become the source of the
3279 /// MOVLP, it must be either a vector load or a scalar load to vector.
3280 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3281 ShuffleVectorSDNode *Op) {
3282 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3284 // Is V2 is a vector load, don't do this transformation. We will try to use
3285 // load folding shufps op.
3286 if (ISD::isNON_EXTLoad(V2))
3289 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3291 if (NumElems != 2 && NumElems != 4)
3293 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3294 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3296 for (unsigned i = NumElems/2; i != NumElems; ++i)
3297 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3302 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3304 static bool isSplatVector(SDNode *N) {
3305 if (N->getOpcode() != ISD::BUILD_VECTOR)
3308 SDValue SplatValue = N->getOperand(0);
3309 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3310 if (N->getOperand(i) != SplatValue)
3315 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3316 /// to an zero vector.
3317 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3318 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3319 SDValue V1 = N->getOperand(0);
3320 SDValue V2 = N->getOperand(1);
3321 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3322 for (unsigned i = 0; i != NumElems; ++i) {
3323 int Idx = N->getMaskElt(i);
3324 if (Idx >= (int)NumElems) {
3325 unsigned Opc = V2.getOpcode();
3326 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3328 if (Opc != ISD::BUILD_VECTOR ||
3329 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3331 } else if (Idx >= 0) {
3332 unsigned Opc = V1.getOpcode();
3333 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3335 if (Opc != ISD::BUILD_VECTOR ||
3336 !X86::isZeroNode(V1.getOperand(Idx)))
3343 /// getZeroVector - Returns a vector of specified type with all zero elements.
3345 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3347 assert(VT.isVector() && "Expected a vector type");
3349 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3350 // type. This ensures they get CSE'd.
3352 if (VT.getSizeInBits() == 64) { // MMX
3353 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3354 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3355 } else if (HasSSE2) { // SSE2
3356 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3357 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3359 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3360 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3362 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3365 /// getOnesVector - Returns a vector of specified type with all bits set.
3367 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3368 assert(VT.isVector() && "Expected a vector type");
3370 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3371 // type. This ensures they get CSE'd.
3372 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3374 if (VT.getSizeInBits() == 64) // MMX
3375 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3377 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3378 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3382 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3383 /// that point to V2 points to its first element.
3384 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3385 EVT VT = SVOp->getValueType(0);
3386 unsigned NumElems = VT.getVectorNumElements();
3388 bool Changed = false;
3389 SmallVector<int, 8> MaskVec;
3390 SVOp->getMask(MaskVec);
3392 for (unsigned i = 0; i != NumElems; ++i) {
3393 if (MaskVec[i] > (int)NumElems) {
3394 MaskVec[i] = NumElems;
3399 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3400 SVOp->getOperand(1), &MaskVec[0]);
3401 return SDValue(SVOp, 0);
3404 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3405 /// operation of specified width.
3406 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3408 unsigned NumElems = VT.getVectorNumElements();
3409 SmallVector<int, 8> Mask;
3410 Mask.push_back(NumElems);
3411 for (unsigned i = 1; i != NumElems; ++i)
3413 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3416 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3417 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3419 unsigned NumElems = VT.getVectorNumElements();
3420 SmallVector<int, 8> Mask;
3421 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3423 Mask.push_back(i + NumElems);
3425 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3428 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3429 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3431 unsigned NumElems = VT.getVectorNumElements();
3432 unsigned Half = NumElems/2;
3433 SmallVector<int, 8> Mask;
3434 for (unsigned i = 0; i != Half; ++i) {
3435 Mask.push_back(i + Half);
3436 Mask.push_back(i + NumElems + Half);
3438 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3441 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3442 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3444 if (SV->getValueType(0).getVectorNumElements() <= 4)
3445 return SDValue(SV, 0);
3447 EVT PVT = MVT::v4f32;
3448 EVT VT = SV->getValueType(0);
3449 DebugLoc dl = SV->getDebugLoc();
3450 SDValue V1 = SV->getOperand(0);
3451 int NumElems = VT.getVectorNumElements();
3452 int EltNo = SV->getSplatIndex();
3454 // unpack elements to the correct location
3455 while (NumElems > 4) {
3456 if (EltNo < NumElems/2) {
3457 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3459 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3460 EltNo -= NumElems/2;
3465 // Perform the splat.
3466 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3467 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3468 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3469 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3472 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3473 /// vector of zero or undef vector. This produces a shuffle where the low
3474 /// element of V2 is swizzled into the zero/undef vector, landing at element
3475 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3476 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3477 bool isZero, bool HasSSE2,
3478 SelectionDAG &DAG) {
3479 EVT VT = V2.getValueType();
3481 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3482 unsigned NumElems = VT.getVectorNumElements();
3483 SmallVector<int, 16> MaskVec;
3484 for (unsigned i = 0; i != NumElems; ++i)
3485 // If this is the insertion idx, put the low elt of V2 here.
3486 MaskVec.push_back(i == Idx ? NumElems : i);
3487 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3490 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3491 /// a shuffle that is zero.
3493 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3494 bool Low, SelectionDAG &DAG) {
3495 unsigned NumZeros = 0;
3496 for (int i = 0; i < NumElems; ++i) {
3497 unsigned Index = Low ? i : NumElems-i-1;
3498 int Idx = SVOp->getMaskElt(Index);
3503 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3504 if (Elt.getNode() && X86::isZeroNode(Elt))
3512 /// isVectorShift - Returns true if the shuffle can be implemented as a
3513 /// logical left or right shift of a vector.
3514 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3515 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3516 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3517 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3520 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3523 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3527 bool SeenV1 = false;
3528 bool SeenV2 = false;
3529 for (unsigned i = NumZeros; i < NumElems; ++i) {
3530 unsigned Val = isLeft ? (i - NumZeros) : i;
3531 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3534 unsigned Idx = (unsigned) Idx_;
3544 if (SeenV1 && SeenV2)
3547 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3553 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3555 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3556 unsigned NumNonZero, unsigned NumZero,
3558 const TargetLowering &TLI) {
3562 DebugLoc dl = Op.getDebugLoc();
3565 for (unsigned i = 0; i < 16; ++i) {
3566 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3567 if (ThisIsNonZero && First) {
3569 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3571 V = DAG.getUNDEF(MVT::v8i16);
3576 SDValue ThisElt(0, 0), LastElt(0, 0);
3577 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3578 if (LastIsNonZero) {
3579 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3580 MVT::i16, Op.getOperand(i-1));
3582 if (ThisIsNonZero) {
3583 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3584 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3585 ThisElt, DAG.getConstant(8, MVT::i8));
3587 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3591 if (ThisElt.getNode())
3592 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3593 DAG.getIntPtrConstant(i/2));
3597 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3600 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3602 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3603 unsigned NumNonZero, unsigned NumZero,
3605 const TargetLowering &TLI) {
3609 DebugLoc dl = Op.getDebugLoc();
3612 for (unsigned i = 0; i < 8; ++i) {
3613 bool isNonZero = (NonZeros & (1 << i)) != 0;
3617 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3619 V = DAG.getUNDEF(MVT::v8i16);
3622 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3623 MVT::v8i16, V, Op.getOperand(i),
3624 DAG.getIntPtrConstant(i));
3631 /// getVShift - Return a vector logical shift node.
3633 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3634 unsigned NumBits, SelectionDAG &DAG,
3635 const TargetLowering &TLI, DebugLoc dl) {
3636 bool isMMX = VT.getSizeInBits() == 64;
3637 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3638 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3639 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3640 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3641 DAG.getNode(Opc, dl, ShVT, SrcOp,
3642 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3646 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3647 SelectionDAG &DAG) const {
3649 // Check if the scalar load can be widened into a vector load. And if
3650 // the address is "base + cst" see if the cst can be "absorbed" into
3651 // the shuffle mask.
3652 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3653 SDValue Ptr = LD->getBasePtr();
3654 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3656 EVT PVT = LD->getValueType(0);
3657 if (PVT != MVT::i32 && PVT != MVT::f32)
3662 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3663 FI = FINode->getIndex();
3665 } else if (Ptr.getOpcode() == ISD::ADD &&
3666 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3667 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3668 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3669 Offset = Ptr.getConstantOperandVal(1);
3670 Ptr = Ptr.getOperand(0);
3675 SDValue Chain = LD->getChain();
3676 // Make sure the stack object alignment is at least 16.
3677 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3678 if (DAG.InferPtrAlignment(Ptr) < 16) {
3679 if (MFI->isFixedObjectIndex(FI)) {
3680 // Can't change the alignment. FIXME: It's possible to compute
3681 // the exact stack offset and reference FI + adjust offset instead.
3682 // If someone *really* cares about this. That's the way to implement it.
3685 MFI->setObjectAlignment(FI, 16);
3689 // (Offset % 16) must be multiple of 4. Then address is then
3690 // Ptr + (Offset & ~15).
3693 if ((Offset % 16) & 3)
3695 int64_t StartOffset = Offset & ~15;
3697 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3698 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3700 int EltNo = (Offset - StartOffset) >> 2;
3701 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3702 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3703 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3705 // Canonicalize it to a v4i32 shuffle.
3706 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3707 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3708 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3709 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3715 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3716 /// vector of type 'VT', see if the elements can be replaced by a single large
3717 /// load which has the same value as a build_vector whose operands are 'elts'.
3719 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3721 /// FIXME: we'd also like to handle the case where the last elements are zero
3722 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3723 /// There's even a handy isZeroNode for that purpose.
3724 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3725 DebugLoc &dl, SelectionDAG &DAG) {
3726 EVT EltVT = VT.getVectorElementType();
3727 unsigned NumElems = Elts.size();
3729 LoadSDNode *LDBase = NULL;
3730 unsigned LastLoadedElt = -1U;
3732 // For each element in the initializer, see if we've found a load or an undef.
3733 // If we don't find an initial load element, or later load elements are
3734 // non-consecutive, bail out.
3735 for (unsigned i = 0; i < NumElems; ++i) {
3736 SDValue Elt = Elts[i];
3738 if (!Elt.getNode() ||
3739 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3742 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3744 LDBase = cast<LoadSDNode>(Elt.getNode());
3748 if (Elt.getOpcode() == ISD::UNDEF)
3751 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3752 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3757 // If we have found an entire vector of loads and undefs, then return a large
3758 // load of the entire vector width starting at the base pointer. If we found
3759 // consecutive loads for the low half, generate a vzext_load node.
3760 if (LastLoadedElt == NumElems - 1) {
3761 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3762 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3763 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3764 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3765 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3766 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3767 LDBase->isVolatile(), LDBase->isNonTemporal(),
3768 LDBase->getAlignment());
3769 } else if (NumElems == 4 && LastLoadedElt == 1) {
3770 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3771 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3772 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3773 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3779 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3780 DebugLoc dl = Op.getDebugLoc();
3781 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3782 if (ISD::isBuildVectorAllZeros(Op.getNode())
3783 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3784 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3785 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3786 // eliminated on x86-32 hosts.
3787 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3790 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3791 return getOnesVector(Op.getValueType(), DAG, dl);
3792 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3795 EVT VT = Op.getValueType();
3796 EVT ExtVT = VT.getVectorElementType();
3797 unsigned EVTBits = ExtVT.getSizeInBits();
3799 unsigned NumElems = Op.getNumOperands();
3800 unsigned NumZero = 0;
3801 unsigned NumNonZero = 0;
3802 unsigned NonZeros = 0;
3803 bool IsAllConstants = true;
3804 SmallSet<SDValue, 8> Values;
3805 for (unsigned i = 0; i < NumElems; ++i) {
3806 SDValue Elt = Op.getOperand(i);
3807 if (Elt.getOpcode() == ISD::UNDEF)
3810 if (Elt.getOpcode() != ISD::Constant &&
3811 Elt.getOpcode() != ISD::ConstantFP)
3812 IsAllConstants = false;
3813 if (X86::isZeroNode(Elt))
3816 NonZeros |= (1 << i);
3821 if (NumNonZero == 0) {
3822 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3823 return DAG.getUNDEF(VT);
3826 // Special case for single non-zero, non-undef, element.
3827 if (NumNonZero == 1) {
3828 unsigned Idx = CountTrailingZeros_32(NonZeros);
3829 SDValue Item = Op.getOperand(Idx);
3831 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3832 // the value are obviously zero, truncate the value to i32 and do the
3833 // insertion that way. Only do this if the value is non-constant or if the
3834 // value is a constant being inserted into element 0. It is cheaper to do
3835 // a constant pool load than it is to do a movd + shuffle.
3836 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3837 (!IsAllConstants || Idx == 0)) {
3838 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3839 // Handle MMX and SSE both.
3840 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3841 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3843 // Truncate the value (which may itself be a constant) to i32, and
3844 // convert it to a vector with movd (S2V+shuffle to zero extend).
3845 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3846 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3847 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3848 Subtarget->hasSSE2(), DAG);
3850 // Now we have our 32-bit value zero extended in the low element of
3851 // a vector. If Idx != 0, swizzle it into place.
3853 SmallVector<int, 4> Mask;
3854 Mask.push_back(Idx);
3855 for (unsigned i = 1; i != VecElts; ++i)
3857 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3858 DAG.getUNDEF(Item.getValueType()),
3861 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3865 // If we have a constant or non-constant insertion into the low element of
3866 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3867 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3868 // depending on what the source datatype is.
3871 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3872 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3873 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3874 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3875 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3876 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3878 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3879 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3880 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3881 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3882 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3883 Subtarget->hasSSE2(), DAG);
3884 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3888 // Is it a vector logical left shift?
3889 if (NumElems == 2 && Idx == 1 &&
3890 X86::isZeroNode(Op.getOperand(0)) &&
3891 !X86::isZeroNode(Op.getOperand(1))) {
3892 unsigned NumBits = VT.getSizeInBits();
3893 return getVShift(true, VT,
3894 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3895 VT, Op.getOperand(1)),
3896 NumBits/2, DAG, *this, dl);
3899 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3902 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3903 // is a non-constant being inserted into an element other than the low one,
3904 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3905 // movd/movss) to move this into the low element, then shuffle it into
3907 if (EVTBits == 32) {
3908 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3910 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3911 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3912 Subtarget->hasSSE2(), DAG);
3913 SmallVector<int, 8> MaskVec;
3914 for (unsigned i = 0; i < NumElems; i++)
3915 MaskVec.push_back(i == Idx ? 0 : 1);
3916 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3920 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3921 if (Values.size() == 1) {
3922 if (EVTBits == 32) {
3923 // Instead of a shuffle like this:
3924 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3925 // Check if it's possible to issue this instead.
3926 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3927 unsigned Idx = CountTrailingZeros_32(NonZeros);
3928 SDValue Item = Op.getOperand(Idx);
3929 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3930 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3935 // A vector full of immediates; various special cases are already
3936 // handled, so this is best done with a single constant-pool load.
3940 // Let legalizer expand 2-wide build_vectors.
3941 if (EVTBits == 64) {
3942 if (NumNonZero == 1) {
3943 // One half is zero or undef.
3944 unsigned Idx = CountTrailingZeros_32(NonZeros);
3945 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3946 Op.getOperand(Idx));
3947 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3948 Subtarget->hasSSE2(), DAG);
3953 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3954 if (EVTBits == 8 && NumElems == 16) {
3955 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3957 if (V.getNode()) return V;
3960 if (EVTBits == 16 && NumElems == 8) {
3961 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3963 if (V.getNode()) return V;
3966 // If element VT is == 32 bits, turn it into a number of shuffles.
3967 SmallVector<SDValue, 8> V;
3969 if (NumElems == 4 && NumZero > 0) {
3970 for (unsigned i = 0; i < 4; ++i) {
3971 bool isZero = !(NonZeros & (1 << i));
3973 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3975 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3978 for (unsigned i = 0; i < 2; ++i) {
3979 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3982 V[i] = V[i*2]; // Must be a zero vector.
3985 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3988 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3991 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3996 SmallVector<int, 8> MaskVec;
3997 bool Reverse = (NonZeros & 0x3) == 2;
3998 for (unsigned i = 0; i < 2; ++i)
3999 MaskVec.push_back(Reverse ? 1-i : i);
4000 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4001 for (unsigned i = 0; i < 2; ++i)
4002 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4003 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4006 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4007 // Check for a build vector of consecutive loads.
4008 for (unsigned i = 0; i < NumElems; ++i)
4009 V[i] = Op.getOperand(i);
4011 // Check for elements which are consecutive loads.
4012 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4016 // For SSE 4.1, use inserts into undef.
4017 if (getSubtarget()->hasSSE41()) {
4018 V[0] = DAG.getUNDEF(VT);
4019 for (unsigned i = 0; i < NumElems; ++i)
4020 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4021 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
4022 Op.getOperand(i), DAG.getIntPtrConstant(i));
4026 // Otherwise, expand into a number of unpckl*
4028 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4029 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4030 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4031 for (unsigned i = 0; i < NumElems; ++i)
4032 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4034 while (NumElems != 0) {
4035 for (unsigned i = 0; i < NumElems; ++i)
4036 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4045 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4046 // We support concatenate two MMX registers and place them in a MMX
4047 // register. This is better than doing a stack convert.
4048 DebugLoc dl = Op.getDebugLoc();
4049 EVT ResVT = Op.getValueType();
4050 assert(Op.getNumOperands() == 2);
4051 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4052 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4054 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4055 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4056 InVec = Op.getOperand(1);
4057 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4058 unsigned NumElts = ResVT.getVectorNumElements();
4059 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4060 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4061 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4063 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4064 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4065 Mask[0] = 0; Mask[1] = 2;
4066 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4068 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4071 // v8i16 shuffles - Prefer shuffles in the following order:
4072 // 1. [all] pshuflw, pshufhw, optional move
4073 // 2. [ssse3] 1 x pshufb
4074 // 3. [ssse3] 2 x pshufb + 1 x por
4075 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4077 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4079 const X86TargetLowering &TLI) {
4080 SDValue V1 = SVOp->getOperand(0);
4081 SDValue V2 = SVOp->getOperand(1);
4082 DebugLoc dl = SVOp->getDebugLoc();
4083 SmallVector<int, 8> MaskVals;
4085 // Determine if more than 1 of the words in each of the low and high quadwords
4086 // of the result come from the same quadword of one of the two inputs. Undef
4087 // mask values count as coming from any quadword, for better codegen.
4088 SmallVector<unsigned, 4> LoQuad(4);
4089 SmallVector<unsigned, 4> HiQuad(4);
4090 BitVector InputQuads(4);
4091 for (unsigned i = 0; i < 8; ++i) {
4092 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4093 int EltIdx = SVOp->getMaskElt(i);
4094 MaskVals.push_back(EltIdx);
4103 InputQuads.set(EltIdx / 4);
4106 int BestLoQuad = -1;
4107 unsigned MaxQuad = 1;
4108 for (unsigned i = 0; i < 4; ++i) {
4109 if (LoQuad[i] > MaxQuad) {
4111 MaxQuad = LoQuad[i];
4115 int BestHiQuad = -1;
4117 for (unsigned i = 0; i < 4; ++i) {
4118 if (HiQuad[i] > MaxQuad) {
4120 MaxQuad = HiQuad[i];
4124 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4125 // of the two input vectors, shuffle them into one input vector so only a
4126 // single pshufb instruction is necessary. If There are more than 2 input
4127 // quads, disable the next transformation since it does not help SSSE3.
4128 bool V1Used = InputQuads[0] || InputQuads[1];
4129 bool V2Used = InputQuads[2] || InputQuads[3];
4130 if (TLI.getSubtarget()->hasSSSE3()) {
4131 if (InputQuads.count() == 2 && V1Used && V2Used) {
4132 BestLoQuad = InputQuads.find_first();
4133 BestHiQuad = InputQuads.find_next(BestLoQuad);
4135 if (InputQuads.count() > 2) {
4141 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4142 // the shuffle mask. If a quad is scored as -1, that means that it contains
4143 // words from all 4 input quadwords.
4145 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4146 SmallVector<int, 8> MaskV;
4147 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4148 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4149 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4150 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4151 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4152 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4154 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4155 // source words for the shuffle, to aid later transformations.
4156 bool AllWordsInNewV = true;
4157 bool InOrder[2] = { true, true };
4158 for (unsigned i = 0; i != 8; ++i) {
4159 int idx = MaskVals[i];
4161 InOrder[i/4] = false;
4162 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4164 AllWordsInNewV = false;
4168 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4169 if (AllWordsInNewV) {
4170 for (int i = 0; i != 8; ++i) {
4171 int idx = MaskVals[i];
4174 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4175 if ((idx != i) && idx < 4)
4177 if ((idx != i) && idx > 3)
4186 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4187 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4188 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4189 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4190 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4194 // If we have SSSE3, and all words of the result are from 1 input vector,
4195 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4196 // is present, fall back to case 4.
4197 if (TLI.getSubtarget()->hasSSSE3()) {
4198 SmallVector<SDValue,16> pshufbMask;
4200 // If we have elements from both input vectors, set the high bit of the
4201 // shuffle mask element to zero out elements that come from V2 in the V1
4202 // mask, and elements that come from V1 in the V2 mask, so that the two
4203 // results can be OR'd together.
4204 bool TwoInputs = V1Used && V2Used;
4205 for (unsigned i = 0; i != 8; ++i) {
4206 int EltIdx = MaskVals[i] * 2;
4207 if (TwoInputs && (EltIdx >= 16)) {
4208 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4209 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4212 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4213 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4215 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4216 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4217 DAG.getNode(ISD::BUILD_VECTOR, dl,
4218 MVT::v16i8, &pshufbMask[0], 16));
4220 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4222 // Calculate the shuffle mask for the second input, shuffle it, and
4223 // OR it with the first shuffled input.
4225 for (unsigned i = 0; i != 8; ++i) {
4226 int EltIdx = MaskVals[i] * 2;
4228 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4229 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4232 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4233 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4235 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4236 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4237 DAG.getNode(ISD::BUILD_VECTOR, dl,
4238 MVT::v16i8, &pshufbMask[0], 16));
4239 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4240 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4243 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4244 // and update MaskVals with new element order.
4245 BitVector InOrder(8);
4246 if (BestLoQuad >= 0) {
4247 SmallVector<int, 8> MaskV;
4248 for (int i = 0; i != 4; ++i) {
4249 int idx = MaskVals[i];
4251 MaskV.push_back(-1);
4253 } else if ((idx / 4) == BestLoQuad) {
4254 MaskV.push_back(idx & 3);
4257 MaskV.push_back(-1);
4260 for (unsigned i = 4; i != 8; ++i)
4262 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4266 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4267 // and update MaskVals with the new element order.
4268 if (BestHiQuad >= 0) {
4269 SmallVector<int, 8> MaskV;
4270 for (unsigned i = 0; i != 4; ++i)
4272 for (unsigned i = 4; i != 8; ++i) {
4273 int idx = MaskVals[i];
4275 MaskV.push_back(-1);
4277 } else if ((idx / 4) == BestHiQuad) {
4278 MaskV.push_back((idx & 3) + 4);
4281 MaskV.push_back(-1);
4284 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4288 // In case BestHi & BestLo were both -1, which means each quadword has a word
4289 // from each of the four input quadwords, calculate the InOrder bitvector now
4290 // before falling through to the insert/extract cleanup.
4291 if (BestLoQuad == -1 && BestHiQuad == -1) {
4293 for (int i = 0; i != 8; ++i)
4294 if (MaskVals[i] < 0 || MaskVals[i] == i)
4298 // The other elements are put in the right place using pextrw and pinsrw.
4299 for (unsigned i = 0; i != 8; ++i) {
4302 int EltIdx = MaskVals[i];
4305 SDValue ExtOp = (EltIdx < 8)
4306 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4307 DAG.getIntPtrConstant(EltIdx))
4308 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4309 DAG.getIntPtrConstant(EltIdx - 8));
4310 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4311 DAG.getIntPtrConstant(i));
4316 // v16i8 shuffles - Prefer shuffles in the following order:
4317 // 1. [ssse3] 1 x pshufb
4318 // 2. [ssse3] 2 x pshufb + 1 x por
4319 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4321 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4323 const X86TargetLowering &TLI) {
4324 SDValue V1 = SVOp->getOperand(0);
4325 SDValue V2 = SVOp->getOperand(1);
4326 DebugLoc dl = SVOp->getDebugLoc();
4327 SmallVector<int, 16> MaskVals;
4328 SVOp->getMask(MaskVals);
4330 // If we have SSSE3, case 1 is generated when all result bytes come from
4331 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4332 // present, fall back to case 3.
4333 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4336 for (unsigned i = 0; i < 16; ++i) {
4337 int EltIdx = MaskVals[i];
4346 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4347 if (TLI.getSubtarget()->hasSSSE3()) {
4348 SmallVector<SDValue,16> pshufbMask;
4350 // If all result elements are from one input vector, then only translate
4351 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4353 // Otherwise, we have elements from both input vectors, and must zero out
4354 // elements that come from V2 in the first mask, and V1 in the second mask
4355 // so that we can OR them together.
4356 bool TwoInputs = !(V1Only || V2Only);
4357 for (unsigned i = 0; i != 16; ++i) {
4358 int EltIdx = MaskVals[i];
4359 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4360 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4363 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4365 // If all the elements are from V2, assign it to V1 and return after
4366 // building the first pshufb.
4369 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4370 DAG.getNode(ISD::BUILD_VECTOR, dl,
4371 MVT::v16i8, &pshufbMask[0], 16));
4375 // Calculate the shuffle mask for the second input, shuffle it, and
4376 // OR it with the first shuffled input.
4378 for (unsigned i = 0; i != 16; ++i) {
4379 int EltIdx = MaskVals[i];
4381 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4384 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4386 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4387 DAG.getNode(ISD::BUILD_VECTOR, dl,
4388 MVT::v16i8, &pshufbMask[0], 16));
4389 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4392 // No SSSE3 - Calculate in place words and then fix all out of place words
4393 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4394 // the 16 different words that comprise the two doublequadword input vectors.
4395 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4396 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4397 SDValue NewV = V2Only ? V2 : V1;
4398 for (int i = 0; i != 8; ++i) {
4399 int Elt0 = MaskVals[i*2];
4400 int Elt1 = MaskVals[i*2+1];
4402 // This word of the result is all undef, skip it.
4403 if (Elt0 < 0 && Elt1 < 0)
4406 // This word of the result is already in the correct place, skip it.
4407 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4409 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4412 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4413 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4416 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4417 // using a single extract together, load it and store it.
4418 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4419 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4420 DAG.getIntPtrConstant(Elt1 / 2));
4421 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4422 DAG.getIntPtrConstant(i));
4426 // If Elt1 is defined, extract it from the appropriate source. If the
4427 // source byte is not also odd, shift the extracted word left 8 bits
4428 // otherwise clear the bottom 8 bits if we need to do an or.
4430 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4431 DAG.getIntPtrConstant(Elt1 / 2));
4432 if ((Elt1 & 1) == 0)
4433 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4434 DAG.getConstant(8, TLI.getShiftAmountTy()));
4436 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4437 DAG.getConstant(0xFF00, MVT::i16));
4439 // If Elt0 is defined, extract it from the appropriate source. If the
4440 // source byte is not also even, shift the extracted word right 8 bits. If
4441 // Elt1 was also defined, OR the extracted values together before
4442 // inserting them in the result.
4444 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4445 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4446 if ((Elt0 & 1) != 0)
4447 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4448 DAG.getConstant(8, TLI.getShiftAmountTy()));
4450 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4451 DAG.getConstant(0x00FF, MVT::i16));
4452 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4455 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4456 DAG.getIntPtrConstant(i));
4458 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4461 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4462 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4463 /// done when every pair / quad of shuffle mask elements point to elements in
4464 /// the right sequence. e.g.
4465 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4467 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4469 const TargetLowering &TLI, DebugLoc dl) {
4470 EVT VT = SVOp->getValueType(0);
4471 SDValue V1 = SVOp->getOperand(0);
4472 SDValue V2 = SVOp->getOperand(1);
4473 unsigned NumElems = VT.getVectorNumElements();
4474 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4475 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4477 switch (VT.getSimpleVT().SimpleTy) {
4478 default: assert(false && "Unexpected!");
4479 case MVT::v4f32: NewVT = MVT::v2f64; break;
4480 case MVT::v4i32: NewVT = MVT::v2i64; break;
4481 case MVT::v8i16: NewVT = MVT::v4i32; break;
4482 case MVT::v16i8: NewVT = MVT::v4i32; break;
4485 if (NewWidth == 2) {
4491 int Scale = NumElems / NewWidth;
4492 SmallVector<int, 8> MaskVec;
4493 for (unsigned i = 0; i < NumElems; i += Scale) {
4495 for (int j = 0; j < Scale; ++j) {
4496 int EltIdx = SVOp->getMaskElt(i+j);
4500 StartIdx = EltIdx - (EltIdx % Scale);
4501 if (EltIdx != StartIdx + j)
4505 MaskVec.push_back(-1);
4507 MaskVec.push_back(StartIdx / Scale);
4510 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4511 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4512 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4515 /// getVZextMovL - Return a zero-extending vector move low node.
4517 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4518 SDValue SrcOp, SelectionDAG &DAG,
4519 const X86Subtarget *Subtarget, DebugLoc dl) {
4520 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4521 LoadSDNode *LD = NULL;
4522 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4523 LD = dyn_cast<LoadSDNode>(SrcOp);
4525 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4527 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4528 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4529 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4530 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4531 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4533 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4534 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4535 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4536 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4544 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4545 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4546 DAG.getNode(ISD::BIT_CONVERT, dl,
4550 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4553 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4554 SDValue V1 = SVOp->getOperand(0);
4555 SDValue V2 = SVOp->getOperand(1);
4556 DebugLoc dl = SVOp->getDebugLoc();
4557 EVT VT = SVOp->getValueType(0);
4559 SmallVector<std::pair<int, int>, 8> Locs;
4561 SmallVector<int, 8> Mask1(4U, -1);
4562 SmallVector<int, 8> PermMask;
4563 SVOp->getMask(PermMask);
4567 for (unsigned i = 0; i != 4; ++i) {
4568 int Idx = PermMask[i];
4570 Locs[i] = std::make_pair(-1, -1);
4572 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4574 Locs[i] = std::make_pair(0, NumLo);
4578 Locs[i] = std::make_pair(1, NumHi);
4580 Mask1[2+NumHi] = Idx;
4586 if (NumLo <= 2 && NumHi <= 2) {
4587 // If no more than two elements come from either vector. This can be
4588 // implemented with two shuffles. First shuffle gather the elements.
4589 // The second shuffle, which takes the first shuffle as both of its
4590 // vector operands, put the elements into the right order.
4591 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4593 SmallVector<int, 8> Mask2(4U, -1);
4595 for (unsigned i = 0; i != 4; ++i) {
4596 if (Locs[i].first == -1)
4599 unsigned Idx = (i < 2) ? 0 : 4;
4600 Idx += Locs[i].first * 2 + Locs[i].second;
4605 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4606 } else if (NumLo == 3 || NumHi == 3) {
4607 // Otherwise, we must have three elements from one vector, call it X, and
4608 // one element from the other, call it Y. First, use a shufps to build an
4609 // intermediate vector with the one element from Y and the element from X
4610 // that will be in the same half in the final destination (the indexes don't
4611 // matter). Then, use a shufps to build the final vector, taking the half
4612 // containing the element from Y from the intermediate, and the other half
4615 // Normalize it so the 3 elements come from V1.
4616 CommuteVectorShuffleMask(PermMask, VT);
4620 // Find the element from V2.
4622 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4623 int Val = PermMask[HiIndex];
4630 Mask1[0] = PermMask[HiIndex];
4632 Mask1[2] = PermMask[HiIndex^1];
4634 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4637 Mask1[0] = PermMask[0];
4638 Mask1[1] = PermMask[1];
4639 Mask1[2] = HiIndex & 1 ? 6 : 4;
4640 Mask1[3] = HiIndex & 1 ? 4 : 6;
4641 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4643 Mask1[0] = HiIndex & 1 ? 2 : 0;
4644 Mask1[1] = HiIndex & 1 ? 0 : 2;
4645 Mask1[2] = PermMask[2];
4646 Mask1[3] = PermMask[3];
4651 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4655 // Break it into (shuffle shuffle_hi, shuffle_lo).
4657 SmallVector<int,8> LoMask(4U, -1);
4658 SmallVector<int,8> HiMask(4U, -1);
4660 SmallVector<int,8> *MaskPtr = &LoMask;
4661 unsigned MaskIdx = 0;
4664 for (unsigned i = 0; i != 4; ++i) {
4671 int Idx = PermMask[i];
4673 Locs[i] = std::make_pair(-1, -1);
4674 } else if (Idx < 4) {
4675 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4676 (*MaskPtr)[LoIdx] = Idx;
4679 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4680 (*MaskPtr)[HiIdx] = Idx;
4685 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4686 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4687 SmallVector<int, 8> MaskOps;
4688 for (unsigned i = 0; i != 4; ++i) {
4689 if (Locs[i].first == -1) {
4690 MaskOps.push_back(-1);
4692 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4693 MaskOps.push_back(Idx);
4696 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4700 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4701 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4702 SDValue V1 = Op.getOperand(0);
4703 SDValue V2 = Op.getOperand(1);
4704 EVT VT = Op.getValueType();
4705 DebugLoc dl = Op.getDebugLoc();
4706 unsigned NumElems = VT.getVectorNumElements();
4707 bool isMMX = VT.getSizeInBits() == 64;
4708 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4709 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4710 bool V1IsSplat = false;
4711 bool V2IsSplat = false;
4713 if (isZeroShuffle(SVOp))
4714 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4716 // Promote splats to v4f32.
4717 if (SVOp->isSplat()) {
4718 if (isMMX || NumElems < 4)
4720 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4723 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4725 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4726 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4727 if (NewOp.getNode())
4728 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4729 LowerVECTOR_SHUFFLE(NewOp, DAG));
4730 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4731 // FIXME: Figure out a cleaner way to do this.
4732 // Try to make use of movq to zero out the top part.
4733 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4734 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4735 if (NewOp.getNode()) {
4736 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4737 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4738 DAG, Subtarget, dl);
4740 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4741 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4742 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4743 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4744 DAG, Subtarget, dl);
4748 if (X86::isPSHUFDMask(SVOp))
4751 // Check if this can be converted into a logical shift.
4752 bool isLeft = false;
4755 bool isShift = getSubtarget()->hasSSE2() &&
4756 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4757 if (isShift && ShVal.hasOneUse()) {
4758 // If the shifted value has multiple uses, it may be cheaper to use
4759 // v_set0 + movlhps or movhlps, etc.
4760 EVT EltVT = VT.getVectorElementType();
4761 ShAmt *= EltVT.getSizeInBits();
4762 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4765 if (X86::isMOVLMask(SVOp)) {
4768 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4769 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4774 // FIXME: fold these into legal mask.
4775 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4776 X86::isMOVSLDUPMask(SVOp) ||
4777 X86::isMOVHLPSMask(SVOp) ||
4778 X86::isMOVLHPSMask(SVOp) ||
4779 X86::isMOVLPMask(SVOp)))
4782 if (ShouldXformToMOVHLPS(SVOp) ||
4783 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4784 return CommuteVectorShuffle(SVOp, DAG);
4787 // No better options. Use a vshl / vsrl.
4788 EVT EltVT = VT.getVectorElementType();
4789 ShAmt *= EltVT.getSizeInBits();
4790 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4793 bool Commuted = false;
4794 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4795 // 1,1,1,1 -> v8i16 though.
4796 V1IsSplat = isSplatVector(V1.getNode());
4797 V2IsSplat = isSplatVector(V2.getNode());
4799 // Canonicalize the splat or undef, if present, to be on the RHS.
4800 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4801 Op = CommuteVectorShuffle(SVOp, DAG);
4802 SVOp = cast<ShuffleVectorSDNode>(Op);
4803 V1 = SVOp->getOperand(0);
4804 V2 = SVOp->getOperand(1);
4805 std::swap(V1IsSplat, V2IsSplat);
4806 std::swap(V1IsUndef, V2IsUndef);
4810 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4811 // Shuffling low element of v1 into undef, just return v1.
4814 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4815 // the instruction selector will not match, so get a canonical MOVL with
4816 // swapped operands to undo the commute.
4817 return getMOVL(DAG, dl, VT, V2, V1);
4820 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4821 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4822 X86::isUNPCKLMask(SVOp) ||
4823 X86::isUNPCKHMask(SVOp))
4827 // Normalize mask so all entries that point to V2 points to its first
4828 // element then try to match unpck{h|l} again. If match, return a
4829 // new vector_shuffle with the corrected mask.
4830 SDValue NewMask = NormalizeMask(SVOp, DAG);
4831 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4832 if (NSVOp != SVOp) {
4833 if (X86::isUNPCKLMask(NSVOp, true)) {
4835 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4842 // Commute is back and try unpck* again.
4843 // FIXME: this seems wrong.
4844 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4845 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4846 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4847 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4848 X86::isUNPCKLMask(NewSVOp) ||
4849 X86::isUNPCKHMask(NewSVOp))
4853 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4855 // Normalize the node to match x86 shuffle ops if needed
4856 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4857 return CommuteVectorShuffle(SVOp, DAG);
4859 // Check for legal shuffle and return?
4860 SmallVector<int, 16> PermMask;
4861 SVOp->getMask(PermMask);
4862 if (isShuffleMaskLegal(PermMask, VT))
4865 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4866 if (VT == MVT::v8i16) {
4867 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4868 if (NewOp.getNode())
4872 if (VT == MVT::v16i8) {
4873 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4874 if (NewOp.getNode())
4878 // Handle all 4 wide cases with a number of shuffles except for MMX.
4879 if (NumElems == 4 && !isMMX)
4880 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4886 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4887 SelectionDAG &DAG) const {
4888 EVT VT = Op.getValueType();
4889 DebugLoc dl = Op.getDebugLoc();
4890 if (VT.getSizeInBits() == 8) {
4891 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4892 Op.getOperand(0), Op.getOperand(1));
4893 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4894 DAG.getValueType(VT));
4895 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4896 } else if (VT.getSizeInBits() == 16) {
4897 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4898 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4900 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4901 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4902 DAG.getNode(ISD::BIT_CONVERT, dl,
4906 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4907 Op.getOperand(0), Op.getOperand(1));
4908 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4909 DAG.getValueType(VT));
4910 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4911 } else if (VT == MVT::f32) {
4912 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4913 // the result back to FR32 register. It's only worth matching if the
4914 // result has a single use which is a store or a bitcast to i32. And in
4915 // the case of a store, it's not worth it if the index is a constant 0,
4916 // because a MOVSSmr can be used instead, which is smaller and faster.
4917 if (!Op.hasOneUse())
4919 SDNode *User = *Op.getNode()->use_begin();
4920 if ((User->getOpcode() != ISD::STORE ||
4921 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4922 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4923 (User->getOpcode() != ISD::BIT_CONVERT ||
4924 User->getValueType(0) != MVT::i32))
4926 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4927 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4930 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4931 } else if (VT == MVT::i32) {
4932 // ExtractPS works with constant index.
4933 if (isa<ConstantSDNode>(Op.getOperand(1)))
4941 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4942 SelectionDAG &DAG) const {
4943 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4946 if (Subtarget->hasSSE41()) {
4947 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4952 EVT VT = Op.getValueType();
4953 DebugLoc dl = Op.getDebugLoc();
4954 // TODO: handle v16i8.
4955 if (VT.getSizeInBits() == 16) {
4956 SDValue Vec = Op.getOperand(0);
4957 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4959 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4960 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4961 DAG.getNode(ISD::BIT_CONVERT, dl,
4964 // Transform it so it match pextrw which produces a 32-bit result.
4965 EVT EltVT = MVT::i32;
4966 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4967 Op.getOperand(0), Op.getOperand(1));
4968 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4969 DAG.getValueType(VT));
4970 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4971 } else if (VT.getSizeInBits() == 32) {
4972 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4976 // SHUFPS the element to the lowest double word, then movss.
4977 int Mask[4] = { Idx, -1, -1, -1 };
4978 EVT VVT = Op.getOperand(0).getValueType();
4979 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4980 DAG.getUNDEF(VVT), Mask);
4981 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4982 DAG.getIntPtrConstant(0));
4983 } else if (VT.getSizeInBits() == 64) {
4984 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4985 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4986 // to match extract_elt for f64.
4987 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4991 // UNPCKHPD the element to the lowest double word, then movsd.
4992 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4993 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4994 int Mask[2] = { 1, -1 };
4995 EVT VVT = Op.getOperand(0).getValueType();
4996 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4997 DAG.getUNDEF(VVT), Mask);
4998 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4999 DAG.getIntPtrConstant(0));
5006 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5007 SelectionDAG &DAG) const {
5008 EVT VT = Op.getValueType();
5009 EVT EltVT = VT.getVectorElementType();
5010 DebugLoc dl = Op.getDebugLoc();
5012 SDValue N0 = Op.getOperand(0);
5013 SDValue N1 = Op.getOperand(1);
5014 SDValue N2 = Op.getOperand(2);
5016 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5017 isa<ConstantSDNode>(N2)) {
5019 if (VT == MVT::v8i16)
5020 Opc = X86ISD::PINSRW;
5021 else if (VT == MVT::v4i16)
5022 Opc = X86ISD::MMX_PINSRW;
5023 else if (VT == MVT::v16i8)
5024 Opc = X86ISD::PINSRB;
5026 Opc = X86ISD::PINSRB;
5028 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5030 if (N1.getValueType() != MVT::i32)
5031 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5032 if (N2.getValueType() != MVT::i32)
5033 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5034 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5035 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5036 // Bits [7:6] of the constant are the source select. This will always be
5037 // zero here. The DAG Combiner may combine an extract_elt index into these
5038 // bits. For example (insert (extract, 3), 2) could be matched by putting
5039 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5040 // Bits [5:4] of the constant are the destination select. This is the
5041 // value of the incoming immediate.
5042 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5043 // combine either bitwise AND or insert of float 0.0 to set these bits.
5044 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5045 // Create this as a scalar to vector..
5046 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5047 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5048 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5049 // PINSR* works with constant index.
5056 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5057 EVT VT = Op.getValueType();
5058 EVT EltVT = VT.getVectorElementType();
5060 if (Subtarget->hasSSE41())
5061 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5063 if (EltVT == MVT::i8)
5066 DebugLoc dl = Op.getDebugLoc();
5067 SDValue N0 = Op.getOperand(0);
5068 SDValue N1 = Op.getOperand(1);
5069 SDValue N2 = Op.getOperand(2);
5071 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5072 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5073 // as its second argument.
5074 if (N1.getValueType() != MVT::i32)
5075 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5076 if (N2.getValueType() != MVT::i32)
5077 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5078 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5079 dl, VT, N0, N1, N2);
5085 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5086 DebugLoc dl = Op.getDebugLoc();
5088 if (Op.getValueType() == MVT::v1i64 &&
5089 Op.getOperand(0).getValueType() == MVT::i64)
5090 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5092 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5093 EVT VT = MVT::v2i32;
5094 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5101 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5102 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5105 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5106 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5107 // one of the above mentioned nodes. It has to be wrapped because otherwise
5108 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5109 // be used to form addressing mode. These wrapped nodes will be selected
5112 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5113 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5115 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5117 unsigned char OpFlag = 0;
5118 unsigned WrapperKind = X86ISD::Wrapper;
5119 CodeModel::Model M = getTargetMachine().getCodeModel();
5121 if (Subtarget->isPICStyleRIPRel() &&
5122 (M == CodeModel::Small || M == CodeModel::Kernel))
5123 WrapperKind = X86ISD::WrapperRIP;
5124 else if (Subtarget->isPICStyleGOT())
5125 OpFlag = X86II::MO_GOTOFF;
5126 else if (Subtarget->isPICStyleStubPIC())
5127 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5129 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5131 CP->getOffset(), OpFlag);
5132 DebugLoc DL = CP->getDebugLoc();
5133 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5134 // With PIC, the address is actually $g + Offset.
5136 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5137 DAG.getNode(X86ISD::GlobalBaseReg,
5138 DebugLoc(), getPointerTy()),
5145 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5146 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5148 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5150 unsigned char OpFlag = 0;
5151 unsigned WrapperKind = X86ISD::Wrapper;
5152 CodeModel::Model M = getTargetMachine().getCodeModel();
5154 if (Subtarget->isPICStyleRIPRel() &&
5155 (M == CodeModel::Small || M == CodeModel::Kernel))
5156 WrapperKind = X86ISD::WrapperRIP;
5157 else if (Subtarget->isPICStyleGOT())
5158 OpFlag = X86II::MO_GOTOFF;
5159 else if (Subtarget->isPICStyleStubPIC())
5160 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5162 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5164 DebugLoc DL = JT->getDebugLoc();
5165 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5167 // With PIC, the address is actually $g + Offset.
5169 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5170 DAG.getNode(X86ISD::GlobalBaseReg,
5171 DebugLoc(), getPointerTy()),
5179 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5180 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5182 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5184 unsigned char OpFlag = 0;
5185 unsigned WrapperKind = X86ISD::Wrapper;
5186 CodeModel::Model M = getTargetMachine().getCodeModel();
5188 if (Subtarget->isPICStyleRIPRel() &&
5189 (M == CodeModel::Small || M == CodeModel::Kernel))
5190 WrapperKind = X86ISD::WrapperRIP;
5191 else if (Subtarget->isPICStyleGOT())
5192 OpFlag = X86II::MO_GOTOFF;
5193 else if (Subtarget->isPICStyleStubPIC())
5194 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5196 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5198 DebugLoc DL = Op.getDebugLoc();
5199 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5202 // With PIC, the address is actually $g + Offset.
5203 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5204 !Subtarget->is64Bit()) {
5205 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5206 DAG.getNode(X86ISD::GlobalBaseReg,
5207 DebugLoc(), getPointerTy()),
5215 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5216 // Create the TargetBlockAddressAddress node.
5217 unsigned char OpFlags =
5218 Subtarget->ClassifyBlockAddressReference();
5219 CodeModel::Model M = getTargetMachine().getCodeModel();
5220 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5221 DebugLoc dl = Op.getDebugLoc();
5222 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5223 /*isTarget=*/true, OpFlags);
5225 if (Subtarget->isPICStyleRIPRel() &&
5226 (M == CodeModel::Small || M == CodeModel::Kernel))
5227 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5229 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5231 // With PIC, the address is actually $g + Offset.
5232 if (isGlobalRelativeToPICBase(OpFlags)) {
5233 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5234 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5242 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5244 SelectionDAG &DAG) const {
5245 // Create the TargetGlobalAddress node, folding in the constant
5246 // offset if it is legal.
5247 unsigned char OpFlags =
5248 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5249 CodeModel::Model M = getTargetMachine().getCodeModel();
5251 if (OpFlags == X86II::MO_NO_FLAG &&
5252 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5253 // A direct static reference to a global.
5254 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5257 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5260 if (Subtarget->isPICStyleRIPRel() &&
5261 (M == CodeModel::Small || M == CodeModel::Kernel))
5262 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5264 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5266 // With PIC, the address is actually $g + Offset.
5267 if (isGlobalRelativeToPICBase(OpFlags)) {
5268 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5269 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5273 // For globals that require a load from a stub to get the address, emit the
5275 if (isGlobalStubReference(OpFlags))
5276 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5277 PseudoSourceValue::getGOT(), 0, false, false, 0);
5279 // If there was a non-zero offset that we didn't fold, create an explicit
5282 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5283 DAG.getConstant(Offset, getPointerTy()));
5289 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5290 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5291 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5292 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5296 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5297 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5298 unsigned char OperandFlags) {
5299 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5300 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5301 DebugLoc dl = GA->getDebugLoc();
5302 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5303 GA->getValueType(0),
5307 SDValue Ops[] = { Chain, TGA, *InFlag };
5308 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5310 SDValue Ops[] = { Chain, TGA };
5311 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5314 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5315 MFI->setAdjustsStack(true);
5317 SDValue Flag = Chain.getValue(1);
5318 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5321 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5323 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5326 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5327 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5328 DAG.getNode(X86ISD::GlobalBaseReg,
5329 DebugLoc(), PtrVT), InFlag);
5330 InFlag = Chain.getValue(1);
5332 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5335 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5337 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5339 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5340 X86::RAX, X86II::MO_TLSGD);
5343 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5344 // "local exec" model.
5345 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5346 const EVT PtrVT, TLSModel::Model model,
5348 DebugLoc dl = GA->getDebugLoc();
5349 // Get the Thread Pointer
5350 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5352 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5355 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5356 NULL, 0, false, false, 0);
5358 unsigned char OperandFlags = 0;
5359 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5361 unsigned WrapperKind = X86ISD::Wrapper;
5362 if (model == TLSModel::LocalExec) {
5363 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5364 } else if (is64Bit) {
5365 assert(model == TLSModel::InitialExec);
5366 OperandFlags = X86II::MO_GOTTPOFF;
5367 WrapperKind = X86ISD::WrapperRIP;
5369 assert(model == TLSModel::InitialExec);
5370 OperandFlags = X86II::MO_INDNTPOFF;
5373 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5375 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5376 GA->getValueType(0),
5377 GA->getOffset(), OperandFlags);
5378 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5380 if (model == TLSModel::InitialExec)
5381 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5382 PseudoSourceValue::getGOT(), 0, false, false, 0);
5384 // The address of the thread local variable is the add of the thread
5385 // pointer with the offset of the variable.
5386 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5390 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5392 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5393 const GlobalValue *GV = GA->getGlobal();
5395 if (Subtarget->isTargetELF()) {
5396 // TODO: implement the "local dynamic" model
5397 // TODO: implement the "initial exec"model for pic executables
5399 // If GV is an alias then use the aliasee for determining
5400 // thread-localness.
5401 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5402 GV = GA->resolveAliasedGlobal(false);
5404 TLSModel::Model model
5405 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5408 case TLSModel::GeneralDynamic:
5409 case TLSModel::LocalDynamic: // not implemented
5410 if (Subtarget->is64Bit())
5411 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5412 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5414 case TLSModel::InitialExec:
5415 case TLSModel::LocalExec:
5416 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5417 Subtarget->is64Bit());
5419 } else if (Subtarget->isTargetDarwin()) {
5420 // Darwin only has one model of TLS. Lower to that.
5421 unsigned char OpFlag = 0;
5422 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5423 X86ISD::WrapperRIP : X86ISD::Wrapper;
5425 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5427 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5428 !Subtarget->is64Bit();
5430 OpFlag = X86II::MO_TLVP_PIC_BASE;
5432 OpFlag = X86II::MO_TLVP;
5433 DebugLoc DL = Op.getDebugLoc();
5434 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
5436 GA->getOffset(), OpFlag);
5437 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5439 // With PIC32, the address is actually $g + Offset.
5441 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5442 DAG.getNode(X86ISD::GlobalBaseReg,
5443 DebugLoc(), getPointerTy()),
5446 // Lowering the machine isd will make sure everything is in the right
5448 SDValue Args[] = { Offset };
5449 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5451 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5452 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5453 MFI->setAdjustsStack(true);
5455 // And our return value (tls address) is in the standard call return value
5457 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5458 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5462 "TLS not implemented for this target.");
5464 llvm_unreachable("Unreachable");
5469 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5470 /// take a 2 x i32 value to shift plus a shift amount.
5471 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5472 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5473 EVT VT = Op.getValueType();
5474 unsigned VTBits = VT.getSizeInBits();
5475 DebugLoc dl = Op.getDebugLoc();
5476 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5477 SDValue ShOpLo = Op.getOperand(0);
5478 SDValue ShOpHi = Op.getOperand(1);
5479 SDValue ShAmt = Op.getOperand(2);
5480 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5481 DAG.getConstant(VTBits - 1, MVT::i8))
5482 : DAG.getConstant(0, VT);
5485 if (Op.getOpcode() == ISD::SHL_PARTS) {
5486 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5487 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5489 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5490 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5493 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5494 DAG.getConstant(VTBits, MVT::i8));
5495 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5496 AndNode, DAG.getConstant(0, MVT::i8));
5499 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5500 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5501 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5503 if (Op.getOpcode() == ISD::SHL_PARTS) {
5504 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5505 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5507 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5508 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5511 SDValue Ops[2] = { Lo, Hi };
5512 return DAG.getMergeValues(Ops, 2, dl);
5515 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5516 SelectionDAG &DAG) const {
5517 EVT SrcVT = Op.getOperand(0).getValueType();
5519 if (SrcVT.isVector()) {
5520 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5526 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5527 "Unknown SINT_TO_FP to lower!");
5529 // These are really Legal; return the operand so the caller accepts it as
5531 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5533 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5534 Subtarget->is64Bit()) {
5538 DebugLoc dl = Op.getDebugLoc();
5539 unsigned Size = SrcVT.getSizeInBits()/8;
5540 MachineFunction &MF = DAG.getMachineFunction();
5541 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5542 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5543 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5545 PseudoSourceValue::getFixedStack(SSFI), 0,
5547 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5550 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5552 SelectionDAG &DAG) const {
5554 DebugLoc dl = Op.getDebugLoc();
5556 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5558 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5560 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5561 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5562 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5563 Tys, Ops, array_lengthof(Ops));
5566 Chain = Result.getValue(1);
5567 SDValue InFlag = Result.getValue(2);
5569 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5570 // shouldn't be necessary except that RFP cannot be live across
5571 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5572 MachineFunction &MF = DAG.getMachineFunction();
5573 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5574 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5575 Tys = DAG.getVTList(MVT::Other);
5577 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5579 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5580 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5581 PseudoSourceValue::getFixedStack(SSFI), 0,
5588 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5589 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5590 SelectionDAG &DAG) const {
5591 // This algorithm is not obvious. Here it is in C code, more or less:
5593 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5594 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5595 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5597 // Copy ints to xmm registers.
5598 __m128i xh = _mm_cvtsi32_si128( hi );
5599 __m128i xl = _mm_cvtsi32_si128( lo );
5601 // Combine into low half of a single xmm register.
5602 __m128i x = _mm_unpacklo_epi32( xh, xl );
5606 // Merge in appropriate exponents to give the integer bits the right
5608 x = _mm_unpacklo_epi32( x, exp );
5610 // Subtract away the biases to deal with the IEEE-754 double precision
5612 d = _mm_sub_pd( (__m128d) x, bias );
5614 // All conversions up to here are exact. The correctly rounded result is
5615 // calculated using the current rounding mode using the following
5617 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5618 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5619 // store doesn't really need to be here (except
5620 // maybe to zero the other double)
5625 DebugLoc dl = Op.getDebugLoc();
5626 LLVMContext *Context = DAG.getContext();
5628 // Build some magic constants.
5629 std::vector<Constant*> CV0;
5630 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5631 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5632 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5633 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5634 Constant *C0 = ConstantVector::get(CV0);
5635 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5637 std::vector<Constant*> CV1;
5639 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5641 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5642 Constant *C1 = ConstantVector::get(CV1);
5643 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5645 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5646 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5648 DAG.getIntPtrConstant(1)));
5649 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5650 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5652 DAG.getIntPtrConstant(0)));
5653 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5654 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5655 PseudoSourceValue::getConstantPool(), 0,
5657 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5658 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5659 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5660 PseudoSourceValue::getConstantPool(), 0,
5662 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5664 // Add the halves; easiest way is to swap them into another reg first.
5665 int ShufMask[2] = { 1, -1 };
5666 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5667 DAG.getUNDEF(MVT::v2f64), ShufMask);
5668 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5669 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5670 DAG.getIntPtrConstant(0));
5673 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5674 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5675 SelectionDAG &DAG) const {
5676 DebugLoc dl = Op.getDebugLoc();
5677 // FP constant to bias correct the final result.
5678 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5681 // Load the 32-bit value into an XMM register.
5682 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5683 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5685 DAG.getIntPtrConstant(0)));
5687 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5688 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5689 DAG.getIntPtrConstant(0));
5691 // Or the load with the bias.
5692 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5693 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5694 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5696 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5697 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5698 MVT::v2f64, Bias)));
5699 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5700 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5701 DAG.getIntPtrConstant(0));
5703 // Subtract the bias.
5704 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5706 // Handle final rounding.
5707 EVT DestVT = Op.getValueType();
5709 if (DestVT.bitsLT(MVT::f64)) {
5710 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5711 DAG.getIntPtrConstant(0));
5712 } else if (DestVT.bitsGT(MVT::f64)) {
5713 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5716 // Handle final rounding.
5720 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5721 SelectionDAG &DAG) const {
5722 SDValue N0 = Op.getOperand(0);
5723 DebugLoc dl = Op.getDebugLoc();
5725 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5726 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5727 // the optimization here.
5728 if (DAG.SignBitIsZero(N0))
5729 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5731 EVT SrcVT = N0.getValueType();
5732 EVT DstVT = Op.getValueType();
5733 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5734 return LowerUINT_TO_FP_i64(Op, DAG);
5735 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5736 return LowerUINT_TO_FP_i32(Op, DAG);
5738 // Make a 64-bit buffer, and use it to build an FILD.
5739 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5740 if (SrcVT == MVT::i32) {
5741 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5742 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5743 getPointerTy(), StackSlot, WordOff);
5744 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5745 StackSlot, NULL, 0, false, false, 0);
5746 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5747 OffsetSlot, NULL, 0, false, false, 0);
5748 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5752 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5753 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5754 StackSlot, NULL, 0, false, false, 0);
5755 // For i64 source, we need to add the appropriate power of 2 if the input
5756 // was negative. This is the same as the optimization in
5757 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5758 // we must be careful to do the computation in x87 extended precision, not
5759 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5760 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5761 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5762 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5764 APInt FF(32, 0x5F800000ULL);
5766 // Check whether the sign bit is set.
5767 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5768 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5771 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5772 SDValue FudgePtr = DAG.getConstantPool(
5773 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5776 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5777 SDValue Zero = DAG.getIntPtrConstant(0);
5778 SDValue Four = DAG.getIntPtrConstant(4);
5779 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5781 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5783 // Load the value out, extending it from f32 to f80.
5784 // FIXME: Avoid the extend by constructing the right constant pool?
5785 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
5786 FudgePtr, PseudoSourceValue::getConstantPool(),
5787 0, MVT::f32, false, false, 4);
5788 // Extend everything to 80 bits to force it to be done on x87.
5789 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5790 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5793 std::pair<SDValue,SDValue> X86TargetLowering::
5794 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5795 DebugLoc dl = Op.getDebugLoc();
5797 EVT DstTy = Op.getValueType();
5800 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5804 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5805 DstTy.getSimpleVT() >= MVT::i16 &&
5806 "Unknown FP_TO_SINT to lower!");
5808 // These are really Legal.
5809 if (DstTy == MVT::i32 &&
5810 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5811 return std::make_pair(SDValue(), SDValue());
5812 if (Subtarget->is64Bit() &&
5813 DstTy == MVT::i64 &&
5814 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5815 return std::make_pair(SDValue(), SDValue());
5817 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5819 MachineFunction &MF = DAG.getMachineFunction();
5820 unsigned MemSize = DstTy.getSizeInBits()/8;
5821 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5822 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5825 switch (DstTy.getSimpleVT().SimpleTy) {
5826 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5827 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5828 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5829 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5832 SDValue Chain = DAG.getEntryNode();
5833 SDValue Value = Op.getOperand(0);
5834 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5835 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5836 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5837 PseudoSourceValue::getFixedStack(SSFI), 0,
5839 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5841 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5843 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5844 Chain = Value.getValue(1);
5845 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5846 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5849 // Build the FP_TO_INT*_IN_MEM
5850 SDValue Ops[] = { Chain, Value, StackSlot };
5851 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5853 return std::make_pair(FIST, StackSlot);
5856 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5857 SelectionDAG &DAG) const {
5858 if (Op.getValueType().isVector()) {
5859 if (Op.getValueType() == MVT::v2i32 &&
5860 Op.getOperand(0).getValueType() == MVT::v2f64) {
5866 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5867 SDValue FIST = Vals.first, StackSlot = Vals.second;
5868 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5869 if (FIST.getNode() == 0) return Op;
5872 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5873 FIST, StackSlot, NULL, 0, false, false, 0);
5876 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5877 SelectionDAG &DAG) const {
5878 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5879 SDValue FIST = Vals.first, StackSlot = Vals.second;
5880 assert(FIST.getNode() && "Unexpected failure");
5883 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5884 FIST, StackSlot, NULL, 0, false, false, 0);
5887 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5888 SelectionDAG &DAG) const {
5889 LLVMContext *Context = DAG.getContext();
5890 DebugLoc dl = Op.getDebugLoc();
5891 EVT VT = Op.getValueType();
5894 EltVT = VT.getVectorElementType();
5895 std::vector<Constant*> CV;
5896 if (EltVT == MVT::f64) {
5897 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5901 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5907 Constant *C = ConstantVector::get(CV);
5908 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5909 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5910 PseudoSourceValue::getConstantPool(), 0,
5912 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5915 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5916 LLVMContext *Context = DAG.getContext();
5917 DebugLoc dl = Op.getDebugLoc();
5918 EVT VT = Op.getValueType();
5921 EltVT = VT.getVectorElementType();
5922 std::vector<Constant*> CV;
5923 if (EltVT == MVT::f64) {
5924 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5928 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5934 Constant *C = ConstantVector::get(CV);
5935 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5936 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5937 PseudoSourceValue::getConstantPool(), 0,
5939 if (VT.isVector()) {
5940 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5941 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5942 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5944 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5946 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5950 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5951 LLVMContext *Context = DAG.getContext();
5952 SDValue Op0 = Op.getOperand(0);
5953 SDValue Op1 = Op.getOperand(1);
5954 DebugLoc dl = Op.getDebugLoc();
5955 EVT VT = Op.getValueType();
5956 EVT SrcVT = Op1.getValueType();
5958 // If second operand is smaller, extend it first.
5959 if (SrcVT.bitsLT(VT)) {
5960 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5963 // And if it is bigger, shrink it first.
5964 if (SrcVT.bitsGT(VT)) {
5965 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5969 // At this point the operands and the result should have the same
5970 // type, and that won't be f80 since that is not custom lowered.
5972 // First get the sign bit of second operand.
5973 std::vector<Constant*> CV;
5974 if (SrcVT == MVT::f64) {
5975 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5976 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5978 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5979 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5980 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5981 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5983 Constant *C = ConstantVector::get(CV);
5984 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5985 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5986 PseudoSourceValue::getConstantPool(), 0,
5988 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5990 // Shift sign bit right or left if the two operands have different types.
5991 if (SrcVT.bitsGT(VT)) {
5992 // Op0 is MVT::f32, Op1 is MVT::f64.
5993 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5994 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5995 DAG.getConstant(32, MVT::i32));
5996 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5997 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5998 DAG.getIntPtrConstant(0));
6001 // Clear first operand sign bit.
6003 if (VT == MVT::f64) {
6004 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6005 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6007 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6008 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6009 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6010 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6012 C = ConstantVector::get(CV);
6013 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6014 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6015 PseudoSourceValue::getConstantPool(), 0,
6017 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6019 // Or the value with the sign bit.
6020 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6023 /// Emit nodes that will be selected as "test Op0,Op0", or something
6025 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6026 SelectionDAG &DAG) const {
6027 DebugLoc dl = Op.getDebugLoc();
6029 // CF and OF aren't always set the way we want. Determine which
6030 // of these we need.
6031 bool NeedCF = false;
6032 bool NeedOF = false;
6035 case X86::COND_A: case X86::COND_AE:
6036 case X86::COND_B: case X86::COND_BE:
6039 case X86::COND_G: case X86::COND_GE:
6040 case X86::COND_L: case X86::COND_LE:
6041 case X86::COND_O: case X86::COND_NO:
6046 // See if we can use the EFLAGS value from the operand instead of
6047 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6048 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6049 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6050 // Emit a CMP with 0, which is the TEST pattern.
6051 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6052 DAG.getConstant(0, Op.getValueType()));
6054 unsigned Opcode = 0;
6055 unsigned NumOperands = 0;
6056 switch (Op.getNode()->getOpcode()) {
6058 // Due to an isel shortcoming, be conservative if this add is likely to be
6059 // selected as part of a load-modify-store instruction. When the root node
6060 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6061 // uses of other nodes in the match, such as the ADD in this case. This
6062 // leads to the ADD being left around and reselected, with the result being
6063 // two adds in the output. Alas, even if none our users are stores, that
6064 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6065 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6066 // climbing the DAG back to the root, and it doesn't seem to be worth the
6068 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6069 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6070 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6073 if (ConstantSDNode *C =
6074 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6075 // An add of one will be selected as an INC.
6076 if (C->getAPIntValue() == 1) {
6077 Opcode = X86ISD::INC;
6082 // An add of negative one (subtract of one) will be selected as a DEC.
6083 if (C->getAPIntValue().isAllOnesValue()) {
6084 Opcode = X86ISD::DEC;
6090 // Otherwise use a regular EFLAGS-setting add.
6091 Opcode = X86ISD::ADD;
6095 // If the primary and result isn't used, don't bother using X86ISD::AND,
6096 // because a TEST instruction will be better.
6097 bool NonFlagUse = false;
6098 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6099 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6101 unsigned UOpNo = UI.getOperandNo();
6102 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6103 // Look pass truncate.
6104 UOpNo = User->use_begin().getOperandNo();
6105 User = *User->use_begin();
6108 if (User->getOpcode() != ISD::BRCOND &&
6109 User->getOpcode() != ISD::SETCC &&
6110 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6123 // Due to the ISEL shortcoming noted above, be conservative if this op is
6124 // likely to be selected as part of a load-modify-store instruction.
6125 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6126 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6127 if (UI->getOpcode() == ISD::STORE)
6130 // Otherwise use a regular EFLAGS-setting instruction.
6131 switch (Op.getNode()->getOpcode()) {
6132 default: llvm_unreachable("unexpected operator!");
6133 case ISD::SUB: Opcode = X86ISD::SUB; break;
6134 case ISD::OR: Opcode = X86ISD::OR; break;
6135 case ISD::XOR: Opcode = X86ISD::XOR; break;
6136 case ISD::AND: Opcode = X86ISD::AND; break;
6148 return SDValue(Op.getNode(), 1);
6155 // Emit a CMP with 0, which is the TEST pattern.
6156 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6157 DAG.getConstant(0, Op.getValueType()));
6159 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6160 SmallVector<SDValue, 4> Ops;
6161 for (unsigned i = 0; i != NumOperands; ++i)
6162 Ops.push_back(Op.getOperand(i));
6164 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6165 DAG.ReplaceAllUsesWith(Op, New);
6166 return SDValue(New.getNode(), 1);
6169 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6171 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6172 SelectionDAG &DAG) const {
6173 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6174 if (C->getAPIntValue() == 0)
6175 return EmitTest(Op0, X86CC, DAG);
6177 DebugLoc dl = Op0.getDebugLoc();
6178 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6181 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6182 /// if it's possible.
6183 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6184 DebugLoc dl, SelectionDAG &DAG) const {
6185 SDValue Op0 = And.getOperand(0);
6186 SDValue Op1 = And.getOperand(1);
6187 if (Op0.getOpcode() == ISD::TRUNCATE)
6188 Op0 = Op0.getOperand(0);
6189 if (Op1.getOpcode() == ISD::TRUNCATE)
6190 Op1 = Op1.getOperand(0);
6193 if (Op1.getOpcode() == ISD::SHL)
6194 std::swap(Op0, Op1);
6195 if (Op0.getOpcode() == ISD::SHL) {
6196 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6197 if (And00C->getZExtValue() == 1) {
6198 // If we looked past a truncate, check that it's only truncating away
6200 unsigned BitWidth = Op0.getValueSizeInBits();
6201 unsigned AndBitWidth = And.getValueSizeInBits();
6202 if (BitWidth > AndBitWidth) {
6203 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6204 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6205 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6209 RHS = Op0.getOperand(1);
6211 } else if (Op1.getOpcode() == ISD::Constant) {
6212 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6213 SDValue AndLHS = Op0;
6214 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6215 LHS = AndLHS.getOperand(0);
6216 RHS = AndLHS.getOperand(1);
6220 if (LHS.getNode()) {
6221 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6222 // instruction. Since the shift amount is in-range-or-undefined, we know
6223 // that doing a bittest on the i32 value is ok. We extend to i32 because
6224 // the encoding for the i16 version is larger than the i32 version.
6225 // Also promote i16 to i32 for performance / code size reason.
6226 if (LHS.getValueType() == MVT::i8 ||
6227 LHS.getValueType() == MVT::i16)
6228 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6230 // If the operand types disagree, extend the shift amount to match. Since
6231 // BT ignores high bits (like shifts) we can use anyextend.
6232 if (LHS.getValueType() != RHS.getValueType())
6233 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6235 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6236 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6237 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6238 DAG.getConstant(Cond, MVT::i8), BT);
6244 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6245 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6246 SDValue Op0 = Op.getOperand(0);
6247 SDValue Op1 = Op.getOperand(1);
6248 DebugLoc dl = Op.getDebugLoc();
6249 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6251 // Optimize to BT if possible.
6252 // Lower (X & (1 << N)) == 0 to BT(X, N).
6253 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6254 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6255 if (Op0.getOpcode() == ISD::AND &&
6257 Op1.getOpcode() == ISD::Constant &&
6258 cast<ConstantSDNode>(Op1)->isNullValue() &&
6259 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6260 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6261 if (NewSetCC.getNode())
6265 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6266 if (Op0.getOpcode() == X86ISD::SETCC &&
6267 Op1.getOpcode() == ISD::Constant &&
6268 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6269 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6270 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6271 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6272 bool Invert = (CC == ISD::SETNE) ^
6273 cast<ConstantSDNode>(Op1)->isNullValue();
6275 CCode = X86::GetOppositeBranchCondition(CCode);
6276 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6277 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6280 bool isFP = Op1.getValueType().isFloatingPoint();
6281 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6282 if (X86CC == X86::COND_INVALID)
6285 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6287 // Use sbb x, x to materialize carry bit into a GPR.
6288 if (X86CC == X86::COND_B)
6289 return DAG.getNode(ISD::AND, dl, MVT::i8,
6290 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6291 DAG.getConstant(X86CC, MVT::i8), Cond),
6292 DAG.getConstant(1, MVT::i8));
6294 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6295 DAG.getConstant(X86CC, MVT::i8), Cond);
6298 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6300 SDValue Op0 = Op.getOperand(0);
6301 SDValue Op1 = Op.getOperand(1);
6302 SDValue CC = Op.getOperand(2);
6303 EVT VT = Op.getValueType();
6304 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6305 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6306 DebugLoc dl = Op.getDebugLoc();
6310 EVT VT0 = Op0.getValueType();
6311 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6312 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6315 switch (SetCCOpcode) {
6318 case ISD::SETEQ: SSECC = 0; break;
6320 case ISD::SETGT: Swap = true; // Fallthrough
6322 case ISD::SETOLT: SSECC = 1; break;
6324 case ISD::SETGE: Swap = true; // Fallthrough
6326 case ISD::SETOLE: SSECC = 2; break;
6327 case ISD::SETUO: SSECC = 3; break;
6329 case ISD::SETNE: SSECC = 4; break;
6330 case ISD::SETULE: Swap = true;
6331 case ISD::SETUGE: SSECC = 5; break;
6332 case ISD::SETULT: Swap = true;
6333 case ISD::SETUGT: SSECC = 6; break;
6334 case ISD::SETO: SSECC = 7; break;
6337 std::swap(Op0, Op1);
6339 // In the two special cases we can't handle, emit two comparisons.
6341 if (SetCCOpcode == ISD::SETUEQ) {
6343 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6344 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6345 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6347 else if (SetCCOpcode == ISD::SETONE) {
6349 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6350 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6351 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6353 llvm_unreachable("Illegal FP comparison");
6355 // Handle all other FP comparisons here.
6356 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6359 // We are handling one of the integer comparisons here. Since SSE only has
6360 // GT and EQ comparisons for integer, swapping operands and multiple
6361 // operations may be required for some comparisons.
6362 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6363 bool Swap = false, Invert = false, FlipSigns = false;
6365 switch (VT.getSimpleVT().SimpleTy) {
6368 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6370 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6372 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6373 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6376 switch (SetCCOpcode) {
6378 case ISD::SETNE: Invert = true;
6379 case ISD::SETEQ: Opc = EQOpc; break;
6380 case ISD::SETLT: Swap = true;
6381 case ISD::SETGT: Opc = GTOpc; break;
6382 case ISD::SETGE: Swap = true;
6383 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6384 case ISD::SETULT: Swap = true;
6385 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6386 case ISD::SETUGE: Swap = true;
6387 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6390 std::swap(Op0, Op1);
6392 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6393 // bits of the inputs before performing those operations.
6395 EVT EltVT = VT.getVectorElementType();
6396 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6398 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6399 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6401 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6402 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6405 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6407 // If the logical-not of the result is required, perform that now.
6409 Result = DAG.getNOT(dl, Result, VT);
6414 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6415 static bool isX86LogicalCmp(SDValue Op) {
6416 unsigned Opc = Op.getNode()->getOpcode();
6417 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6419 if (Op.getResNo() == 1 &&
6420 (Opc == X86ISD::ADD ||
6421 Opc == X86ISD::SUB ||
6422 Opc == X86ISD::SMUL ||
6423 Opc == X86ISD::UMUL ||
6424 Opc == X86ISD::INC ||
6425 Opc == X86ISD::DEC ||
6426 Opc == X86ISD::OR ||
6427 Opc == X86ISD::XOR ||
6428 Opc == X86ISD::AND))
6434 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6435 bool addTest = true;
6436 SDValue Cond = Op.getOperand(0);
6437 DebugLoc dl = Op.getDebugLoc();
6440 if (Cond.getOpcode() == ISD::SETCC) {
6441 SDValue NewCond = LowerSETCC(Cond, DAG);
6442 if (NewCond.getNode())
6446 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6447 SDValue Op1 = Op.getOperand(1);
6448 SDValue Op2 = Op.getOperand(2);
6449 if (Cond.getOpcode() == X86ISD::SETCC &&
6450 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6451 SDValue Cmp = Cond.getOperand(1);
6452 if (Cmp.getOpcode() == X86ISD::CMP) {
6453 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6454 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6455 ConstantSDNode *RHSC =
6456 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6457 if (N1C && N1C->isAllOnesValue() &&
6458 N2C && N2C->isNullValue() &&
6459 RHSC && RHSC->isNullValue()) {
6460 SDValue CmpOp0 = Cmp.getOperand(0);
6461 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6462 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6463 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6464 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6469 // Look pass (and (setcc_carry (cmp ...)), 1).
6470 if (Cond.getOpcode() == ISD::AND &&
6471 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6472 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6473 if (C && C->getAPIntValue() == 1)
6474 Cond = Cond.getOperand(0);
6477 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6478 // setting operand in place of the X86ISD::SETCC.
6479 if (Cond.getOpcode() == X86ISD::SETCC ||
6480 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6481 CC = Cond.getOperand(0);
6483 SDValue Cmp = Cond.getOperand(1);
6484 unsigned Opc = Cmp.getOpcode();
6485 EVT VT = Op.getValueType();
6487 bool IllegalFPCMov = false;
6488 if (VT.isFloatingPoint() && !VT.isVector() &&
6489 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6490 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6492 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6493 Opc == X86ISD::BT) { // FIXME
6500 // Look pass the truncate.
6501 if (Cond.getOpcode() == ISD::TRUNCATE)
6502 Cond = Cond.getOperand(0);
6504 // We know the result of AND is compared against zero. Try to match
6506 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6507 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6508 if (NewSetCC.getNode()) {
6509 CC = NewSetCC.getOperand(0);
6510 Cond = NewSetCC.getOperand(1);
6517 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6518 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6521 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6522 // condition is true.
6523 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6524 SDValue Ops[] = { Op2, Op1, CC, Cond };
6525 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6528 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6529 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6530 // from the AND / OR.
6531 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6532 Opc = Op.getOpcode();
6533 if (Opc != ISD::OR && Opc != ISD::AND)
6535 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6536 Op.getOperand(0).hasOneUse() &&
6537 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6538 Op.getOperand(1).hasOneUse());
6541 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6542 // 1 and that the SETCC node has a single use.
6543 static bool isXor1OfSetCC(SDValue Op) {
6544 if (Op.getOpcode() != ISD::XOR)
6546 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6547 if (N1C && N1C->getAPIntValue() == 1) {
6548 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6549 Op.getOperand(0).hasOneUse();
6554 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6555 bool addTest = true;
6556 SDValue Chain = Op.getOperand(0);
6557 SDValue Cond = Op.getOperand(1);
6558 SDValue Dest = Op.getOperand(2);
6559 DebugLoc dl = Op.getDebugLoc();
6562 if (Cond.getOpcode() == ISD::SETCC) {
6563 SDValue NewCond = LowerSETCC(Cond, DAG);
6564 if (NewCond.getNode())
6568 // FIXME: LowerXALUO doesn't handle these!!
6569 else if (Cond.getOpcode() == X86ISD::ADD ||
6570 Cond.getOpcode() == X86ISD::SUB ||
6571 Cond.getOpcode() == X86ISD::SMUL ||
6572 Cond.getOpcode() == X86ISD::UMUL)
6573 Cond = LowerXALUO(Cond, DAG);
6576 // Look pass (and (setcc_carry (cmp ...)), 1).
6577 if (Cond.getOpcode() == ISD::AND &&
6578 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6579 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6580 if (C && C->getAPIntValue() == 1)
6581 Cond = Cond.getOperand(0);
6584 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6585 // setting operand in place of the X86ISD::SETCC.
6586 if (Cond.getOpcode() == X86ISD::SETCC ||
6587 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6588 CC = Cond.getOperand(0);
6590 SDValue Cmp = Cond.getOperand(1);
6591 unsigned Opc = Cmp.getOpcode();
6592 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6593 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6597 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6601 // These can only come from an arithmetic instruction with overflow,
6602 // e.g. SADDO, UADDO.
6603 Cond = Cond.getNode()->getOperand(1);
6610 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6611 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6612 if (CondOpc == ISD::OR) {
6613 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6614 // two branches instead of an explicit OR instruction with a
6616 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6617 isX86LogicalCmp(Cmp)) {
6618 CC = Cond.getOperand(0).getOperand(0);
6619 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6620 Chain, Dest, CC, Cmp);
6621 CC = Cond.getOperand(1).getOperand(0);
6625 } else { // ISD::AND
6626 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6627 // two branches instead of an explicit AND instruction with a
6628 // separate test. However, we only do this if this block doesn't
6629 // have a fall-through edge, because this requires an explicit
6630 // jmp when the condition is false.
6631 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6632 isX86LogicalCmp(Cmp) &&
6633 Op.getNode()->hasOneUse()) {
6634 X86::CondCode CCode =
6635 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6636 CCode = X86::GetOppositeBranchCondition(CCode);
6637 CC = DAG.getConstant(CCode, MVT::i8);
6638 SDNode *User = *Op.getNode()->use_begin();
6639 // Look for an unconditional branch following this conditional branch.
6640 // We need this because we need to reverse the successors in order
6641 // to implement FCMP_OEQ.
6642 if (User->getOpcode() == ISD::BR) {
6643 SDValue FalseBB = User->getOperand(1);
6645 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6646 assert(NewBR == User);
6650 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6651 Chain, Dest, CC, Cmp);
6652 X86::CondCode CCode =
6653 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6654 CCode = X86::GetOppositeBranchCondition(CCode);
6655 CC = DAG.getConstant(CCode, MVT::i8);
6661 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6662 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6663 // It should be transformed during dag combiner except when the condition
6664 // is set by a arithmetics with overflow node.
6665 X86::CondCode CCode =
6666 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6667 CCode = X86::GetOppositeBranchCondition(CCode);
6668 CC = DAG.getConstant(CCode, MVT::i8);
6669 Cond = Cond.getOperand(0).getOperand(1);
6675 // Look pass the truncate.
6676 if (Cond.getOpcode() == ISD::TRUNCATE)
6677 Cond = Cond.getOperand(0);
6679 // We know the result of AND is compared against zero. Try to match
6681 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6682 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6683 if (NewSetCC.getNode()) {
6684 CC = NewSetCC.getOperand(0);
6685 Cond = NewSetCC.getOperand(1);
6692 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6693 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6695 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6696 Chain, Dest, CC, Cond);
6700 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6701 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6702 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6703 // that the guard pages used by the OS virtual memory manager are allocated in
6704 // correct sequence.
6706 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6707 SelectionDAG &DAG) const {
6708 assert(Subtarget->isTargetCygMing() &&
6709 "This should be used only on Cygwin/Mingw targets");
6710 DebugLoc dl = Op.getDebugLoc();
6713 SDValue Chain = Op.getOperand(0);
6714 SDValue Size = Op.getOperand(1);
6715 // FIXME: Ensure alignment here
6719 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6721 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6722 Flag = Chain.getValue(1);
6724 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6726 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6727 Flag = Chain.getValue(1);
6729 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6731 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6732 return DAG.getMergeValues(Ops1, 2, dl);
6735 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6736 MachineFunction &MF = DAG.getMachineFunction();
6737 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6739 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6740 DebugLoc dl = Op.getDebugLoc();
6742 if (!Subtarget->is64Bit()) {
6743 // vastart just stores the address of the VarArgsFrameIndex slot into the
6744 // memory location argument.
6745 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6747 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6752 // gp_offset (0 - 6 * 8)
6753 // fp_offset (48 - 48 + 8 * 16)
6754 // overflow_arg_area (point to parameters coming in memory).
6756 SmallVector<SDValue, 8> MemOps;
6757 SDValue FIN = Op.getOperand(1);
6759 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6760 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6762 FIN, SV, 0, false, false, 0);
6763 MemOps.push_back(Store);
6766 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6767 FIN, DAG.getIntPtrConstant(4));
6768 Store = DAG.getStore(Op.getOperand(0), dl,
6769 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6771 FIN, SV, 0, false, false, 0);
6772 MemOps.push_back(Store);
6774 // Store ptr to overflow_arg_area
6775 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6776 FIN, DAG.getIntPtrConstant(4));
6777 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6779 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6781 MemOps.push_back(Store);
6783 // Store ptr to reg_save_area.
6784 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6785 FIN, DAG.getIntPtrConstant(8));
6786 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6788 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6790 MemOps.push_back(Store);
6791 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6792 &MemOps[0], MemOps.size());
6795 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6796 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6797 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6799 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6803 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6804 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6805 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6806 SDValue Chain = Op.getOperand(0);
6807 SDValue DstPtr = Op.getOperand(1);
6808 SDValue SrcPtr = Op.getOperand(2);
6809 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6810 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6811 DebugLoc dl = Op.getDebugLoc();
6813 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6814 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6815 false, DstSV, 0, SrcSV, 0);
6819 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6820 DebugLoc dl = Op.getDebugLoc();
6821 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6823 default: return SDValue(); // Don't custom lower most intrinsics.
6824 // Comparison intrinsics.
6825 case Intrinsic::x86_sse_comieq_ss:
6826 case Intrinsic::x86_sse_comilt_ss:
6827 case Intrinsic::x86_sse_comile_ss:
6828 case Intrinsic::x86_sse_comigt_ss:
6829 case Intrinsic::x86_sse_comige_ss:
6830 case Intrinsic::x86_sse_comineq_ss:
6831 case Intrinsic::x86_sse_ucomieq_ss:
6832 case Intrinsic::x86_sse_ucomilt_ss:
6833 case Intrinsic::x86_sse_ucomile_ss:
6834 case Intrinsic::x86_sse_ucomigt_ss:
6835 case Intrinsic::x86_sse_ucomige_ss:
6836 case Intrinsic::x86_sse_ucomineq_ss:
6837 case Intrinsic::x86_sse2_comieq_sd:
6838 case Intrinsic::x86_sse2_comilt_sd:
6839 case Intrinsic::x86_sse2_comile_sd:
6840 case Intrinsic::x86_sse2_comigt_sd:
6841 case Intrinsic::x86_sse2_comige_sd:
6842 case Intrinsic::x86_sse2_comineq_sd:
6843 case Intrinsic::x86_sse2_ucomieq_sd:
6844 case Intrinsic::x86_sse2_ucomilt_sd:
6845 case Intrinsic::x86_sse2_ucomile_sd:
6846 case Intrinsic::x86_sse2_ucomigt_sd:
6847 case Intrinsic::x86_sse2_ucomige_sd:
6848 case Intrinsic::x86_sse2_ucomineq_sd: {
6850 ISD::CondCode CC = ISD::SETCC_INVALID;
6853 case Intrinsic::x86_sse_comieq_ss:
6854 case Intrinsic::x86_sse2_comieq_sd:
6858 case Intrinsic::x86_sse_comilt_ss:
6859 case Intrinsic::x86_sse2_comilt_sd:
6863 case Intrinsic::x86_sse_comile_ss:
6864 case Intrinsic::x86_sse2_comile_sd:
6868 case Intrinsic::x86_sse_comigt_ss:
6869 case Intrinsic::x86_sse2_comigt_sd:
6873 case Intrinsic::x86_sse_comige_ss:
6874 case Intrinsic::x86_sse2_comige_sd:
6878 case Intrinsic::x86_sse_comineq_ss:
6879 case Intrinsic::x86_sse2_comineq_sd:
6883 case Intrinsic::x86_sse_ucomieq_ss:
6884 case Intrinsic::x86_sse2_ucomieq_sd:
6885 Opc = X86ISD::UCOMI;
6888 case Intrinsic::x86_sse_ucomilt_ss:
6889 case Intrinsic::x86_sse2_ucomilt_sd:
6890 Opc = X86ISD::UCOMI;
6893 case Intrinsic::x86_sse_ucomile_ss:
6894 case Intrinsic::x86_sse2_ucomile_sd:
6895 Opc = X86ISD::UCOMI;
6898 case Intrinsic::x86_sse_ucomigt_ss:
6899 case Intrinsic::x86_sse2_ucomigt_sd:
6900 Opc = X86ISD::UCOMI;
6903 case Intrinsic::x86_sse_ucomige_ss:
6904 case Intrinsic::x86_sse2_ucomige_sd:
6905 Opc = X86ISD::UCOMI;
6908 case Intrinsic::x86_sse_ucomineq_ss:
6909 case Intrinsic::x86_sse2_ucomineq_sd:
6910 Opc = X86ISD::UCOMI;
6915 SDValue LHS = Op.getOperand(1);
6916 SDValue RHS = Op.getOperand(2);
6917 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6918 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6919 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6920 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6921 DAG.getConstant(X86CC, MVT::i8), Cond);
6922 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6924 // ptest intrinsics. The intrinsic these come from are designed to return
6925 // an integer value, not just an instruction so lower it to the ptest
6926 // pattern and a setcc for the result.
6927 case Intrinsic::x86_sse41_ptestz:
6928 case Intrinsic::x86_sse41_ptestc:
6929 case Intrinsic::x86_sse41_ptestnzc:{
6932 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6933 case Intrinsic::x86_sse41_ptestz:
6935 X86CC = X86::COND_E;
6937 case Intrinsic::x86_sse41_ptestc:
6939 X86CC = X86::COND_B;
6941 case Intrinsic::x86_sse41_ptestnzc:
6943 X86CC = X86::COND_A;
6947 SDValue LHS = Op.getOperand(1);
6948 SDValue RHS = Op.getOperand(2);
6949 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6950 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6951 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6952 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6955 // Fix vector shift instructions where the last operand is a non-immediate
6957 case Intrinsic::x86_sse2_pslli_w:
6958 case Intrinsic::x86_sse2_pslli_d:
6959 case Intrinsic::x86_sse2_pslli_q:
6960 case Intrinsic::x86_sse2_psrli_w:
6961 case Intrinsic::x86_sse2_psrli_d:
6962 case Intrinsic::x86_sse2_psrli_q:
6963 case Intrinsic::x86_sse2_psrai_w:
6964 case Intrinsic::x86_sse2_psrai_d:
6965 case Intrinsic::x86_mmx_pslli_w:
6966 case Intrinsic::x86_mmx_pslli_d:
6967 case Intrinsic::x86_mmx_pslli_q:
6968 case Intrinsic::x86_mmx_psrli_w:
6969 case Intrinsic::x86_mmx_psrli_d:
6970 case Intrinsic::x86_mmx_psrli_q:
6971 case Intrinsic::x86_mmx_psrai_w:
6972 case Intrinsic::x86_mmx_psrai_d: {
6973 SDValue ShAmt = Op.getOperand(2);
6974 if (isa<ConstantSDNode>(ShAmt))
6977 unsigned NewIntNo = 0;
6978 EVT ShAmtVT = MVT::v4i32;
6980 case Intrinsic::x86_sse2_pslli_w:
6981 NewIntNo = Intrinsic::x86_sse2_psll_w;
6983 case Intrinsic::x86_sse2_pslli_d:
6984 NewIntNo = Intrinsic::x86_sse2_psll_d;
6986 case Intrinsic::x86_sse2_pslli_q:
6987 NewIntNo = Intrinsic::x86_sse2_psll_q;
6989 case Intrinsic::x86_sse2_psrli_w:
6990 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6992 case Intrinsic::x86_sse2_psrli_d:
6993 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6995 case Intrinsic::x86_sse2_psrli_q:
6996 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6998 case Intrinsic::x86_sse2_psrai_w:
6999 NewIntNo = Intrinsic::x86_sse2_psra_w;
7001 case Intrinsic::x86_sse2_psrai_d:
7002 NewIntNo = Intrinsic::x86_sse2_psra_d;
7005 ShAmtVT = MVT::v2i32;
7007 case Intrinsic::x86_mmx_pslli_w:
7008 NewIntNo = Intrinsic::x86_mmx_psll_w;
7010 case Intrinsic::x86_mmx_pslli_d:
7011 NewIntNo = Intrinsic::x86_mmx_psll_d;
7013 case Intrinsic::x86_mmx_pslli_q:
7014 NewIntNo = Intrinsic::x86_mmx_psll_q;
7016 case Intrinsic::x86_mmx_psrli_w:
7017 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7019 case Intrinsic::x86_mmx_psrli_d:
7020 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7022 case Intrinsic::x86_mmx_psrli_q:
7023 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7025 case Intrinsic::x86_mmx_psrai_w:
7026 NewIntNo = Intrinsic::x86_mmx_psra_w;
7028 case Intrinsic::x86_mmx_psrai_d:
7029 NewIntNo = Intrinsic::x86_mmx_psra_d;
7031 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7037 // The vector shift intrinsics with scalars uses 32b shift amounts but
7038 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7042 ShOps[1] = DAG.getConstant(0, MVT::i32);
7043 if (ShAmtVT == MVT::v4i32) {
7044 ShOps[2] = DAG.getUNDEF(MVT::i32);
7045 ShOps[3] = DAG.getUNDEF(MVT::i32);
7046 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7048 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7051 EVT VT = Op.getValueType();
7052 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7053 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7054 DAG.getConstant(NewIntNo, MVT::i32),
7055 Op.getOperand(1), ShAmt);
7060 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7061 SelectionDAG &DAG) const {
7062 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7063 MFI->setReturnAddressIsTaken(true);
7065 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7066 DebugLoc dl = Op.getDebugLoc();
7069 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7071 DAG.getConstant(TD->getPointerSize(),
7072 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7073 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7074 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7076 NULL, 0, false, false, 0);
7079 // Just load the return address.
7080 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7081 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7082 RetAddrFI, NULL, 0, false, false, 0);
7085 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7086 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7087 MFI->setFrameAddressIsTaken(true);
7089 EVT VT = Op.getValueType();
7090 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7091 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7092 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7093 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7095 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7100 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7101 SelectionDAG &DAG) const {
7102 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7105 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7106 MachineFunction &MF = DAG.getMachineFunction();
7107 SDValue Chain = Op.getOperand(0);
7108 SDValue Offset = Op.getOperand(1);
7109 SDValue Handler = Op.getOperand(2);
7110 DebugLoc dl = Op.getDebugLoc();
7112 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7114 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7116 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7117 DAG.getIntPtrConstant(-TD->getPointerSize()));
7118 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7119 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7120 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7121 MF.getRegInfo().addLiveOut(StoreAddrReg);
7123 return DAG.getNode(X86ISD::EH_RETURN, dl,
7125 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7128 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7129 SelectionDAG &DAG) const {
7130 SDValue Root = Op.getOperand(0);
7131 SDValue Trmp = Op.getOperand(1); // trampoline
7132 SDValue FPtr = Op.getOperand(2); // nested function
7133 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7134 DebugLoc dl = Op.getDebugLoc();
7136 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7138 if (Subtarget->is64Bit()) {
7139 SDValue OutChains[6];
7141 // Large code-model.
7142 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7143 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7145 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7146 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7148 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7150 // Load the pointer to the nested function into R11.
7151 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7152 SDValue Addr = Trmp;
7153 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7154 Addr, TrmpAddr, 0, false, false, 0);
7156 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7157 DAG.getConstant(2, MVT::i64));
7158 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7161 // Load the 'nest' parameter value into R10.
7162 // R10 is specified in X86CallingConv.td
7163 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7164 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7165 DAG.getConstant(10, MVT::i64));
7166 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7167 Addr, TrmpAddr, 10, false, false, 0);
7169 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7170 DAG.getConstant(12, MVT::i64));
7171 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7174 // Jump to the nested function.
7175 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7176 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7177 DAG.getConstant(20, MVT::i64));
7178 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7179 Addr, TrmpAddr, 20, false, false, 0);
7181 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7182 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7183 DAG.getConstant(22, MVT::i64));
7184 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7185 TrmpAddr, 22, false, false, 0);
7188 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7189 return DAG.getMergeValues(Ops, 2, dl);
7191 const Function *Func =
7192 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7193 CallingConv::ID CC = Func->getCallingConv();
7198 llvm_unreachable("Unsupported calling convention");
7199 case CallingConv::C:
7200 case CallingConv::X86_StdCall: {
7201 // Pass 'nest' parameter in ECX.
7202 // Must be kept in sync with X86CallingConv.td
7205 // Check that ECX wasn't needed by an 'inreg' parameter.
7206 const FunctionType *FTy = Func->getFunctionType();
7207 const AttrListPtr &Attrs = Func->getAttributes();
7209 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7210 unsigned InRegCount = 0;
7213 for (FunctionType::param_iterator I = FTy->param_begin(),
7214 E = FTy->param_end(); I != E; ++I, ++Idx)
7215 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7216 // FIXME: should only count parameters that are lowered to integers.
7217 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7219 if (InRegCount > 2) {
7220 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7225 case CallingConv::X86_FastCall:
7226 case CallingConv::X86_ThisCall:
7227 case CallingConv::Fast:
7228 // Pass 'nest' parameter in EAX.
7229 // Must be kept in sync with X86CallingConv.td
7234 SDValue OutChains[4];
7237 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7238 DAG.getConstant(10, MVT::i32));
7239 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7241 // This is storing the opcode for MOV32ri.
7242 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7243 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7244 OutChains[0] = DAG.getStore(Root, dl,
7245 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7246 Trmp, TrmpAddr, 0, false, false, 0);
7248 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7249 DAG.getConstant(1, MVT::i32));
7250 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7253 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7254 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7255 DAG.getConstant(5, MVT::i32));
7256 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7257 TrmpAddr, 5, false, false, 1);
7259 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7260 DAG.getConstant(6, MVT::i32));
7261 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7265 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7266 return DAG.getMergeValues(Ops, 2, dl);
7270 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7271 SelectionDAG &DAG) const {
7273 The rounding mode is in bits 11:10 of FPSR, and has the following
7280 FLT_ROUNDS, on the other hand, expects the following:
7287 To perform the conversion, we do:
7288 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7291 MachineFunction &MF = DAG.getMachineFunction();
7292 const TargetMachine &TM = MF.getTarget();
7293 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7294 unsigned StackAlignment = TFI.getStackAlignment();
7295 EVT VT = Op.getValueType();
7296 DebugLoc dl = Op.getDebugLoc();
7298 // Save FP Control Word to stack slot
7299 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7300 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7302 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7303 DAG.getEntryNode(), StackSlot);
7305 // Load FP Control Word from stack slot
7306 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7309 // Transform as necessary
7311 DAG.getNode(ISD::SRL, dl, MVT::i16,
7312 DAG.getNode(ISD::AND, dl, MVT::i16,
7313 CWD, DAG.getConstant(0x800, MVT::i16)),
7314 DAG.getConstant(11, MVT::i8));
7316 DAG.getNode(ISD::SRL, dl, MVT::i16,
7317 DAG.getNode(ISD::AND, dl, MVT::i16,
7318 CWD, DAG.getConstant(0x400, MVT::i16)),
7319 DAG.getConstant(9, MVT::i8));
7322 DAG.getNode(ISD::AND, dl, MVT::i16,
7323 DAG.getNode(ISD::ADD, dl, MVT::i16,
7324 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7325 DAG.getConstant(1, MVT::i16)),
7326 DAG.getConstant(3, MVT::i16));
7329 return DAG.getNode((VT.getSizeInBits() < 16 ?
7330 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7333 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7334 EVT VT = Op.getValueType();
7336 unsigned NumBits = VT.getSizeInBits();
7337 DebugLoc dl = Op.getDebugLoc();
7339 Op = Op.getOperand(0);
7340 if (VT == MVT::i8) {
7341 // Zero extend to i32 since there is not an i8 bsr.
7343 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7346 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7347 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7348 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7350 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7353 DAG.getConstant(NumBits+NumBits-1, OpVT),
7354 DAG.getConstant(X86::COND_E, MVT::i8),
7357 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7359 // Finally xor with NumBits-1.
7360 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7363 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7367 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7368 EVT VT = Op.getValueType();
7370 unsigned NumBits = VT.getSizeInBits();
7371 DebugLoc dl = Op.getDebugLoc();
7373 Op = Op.getOperand(0);
7374 if (VT == MVT::i8) {
7376 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7379 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7380 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7381 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7383 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7386 DAG.getConstant(NumBits, OpVT),
7387 DAG.getConstant(X86::COND_E, MVT::i8),
7390 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7393 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7397 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7398 EVT VT = Op.getValueType();
7399 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7400 DebugLoc dl = Op.getDebugLoc();
7402 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7403 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7404 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7405 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7406 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7408 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7409 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7410 // return AloBlo + AloBhi + AhiBlo;
7412 SDValue A = Op.getOperand(0);
7413 SDValue B = Op.getOperand(1);
7415 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7416 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7417 A, DAG.getConstant(32, MVT::i32));
7418 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7419 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7420 B, DAG.getConstant(32, MVT::i32));
7421 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7422 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7424 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7425 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7427 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7428 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7430 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7431 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7432 AloBhi, DAG.getConstant(32, MVT::i32));
7433 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7434 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7435 AhiBlo, DAG.getConstant(32, MVT::i32));
7436 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7437 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7442 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7443 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7444 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7445 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7446 // has only one use.
7447 SDNode *N = Op.getNode();
7448 SDValue LHS = N->getOperand(0);
7449 SDValue RHS = N->getOperand(1);
7450 unsigned BaseOp = 0;
7452 DebugLoc dl = Op.getDebugLoc();
7454 switch (Op.getOpcode()) {
7455 default: llvm_unreachable("Unknown ovf instruction!");
7457 // A subtract of one will be selected as a INC. Note that INC doesn't
7458 // set CF, so we can't do this for UADDO.
7459 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7460 if (C->getAPIntValue() == 1) {
7461 BaseOp = X86ISD::INC;
7465 BaseOp = X86ISD::ADD;
7469 BaseOp = X86ISD::ADD;
7473 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7474 // set CF, so we can't do this for USUBO.
7475 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7476 if (C->getAPIntValue() == 1) {
7477 BaseOp = X86ISD::DEC;
7481 BaseOp = X86ISD::SUB;
7485 BaseOp = X86ISD::SUB;
7489 BaseOp = X86ISD::SMUL;
7493 BaseOp = X86ISD::UMUL;
7498 // Also sets EFLAGS.
7499 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7500 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7503 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7504 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7506 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7510 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7511 EVT T = Op.getValueType();
7512 DebugLoc dl = Op.getDebugLoc();
7515 switch(T.getSimpleVT().SimpleTy) {
7517 assert(false && "Invalid value type!");
7518 case MVT::i8: Reg = X86::AL; size = 1; break;
7519 case MVT::i16: Reg = X86::AX; size = 2; break;
7520 case MVT::i32: Reg = X86::EAX; size = 4; break;
7522 assert(Subtarget->is64Bit() && "Node not type legal!");
7523 Reg = X86::RAX; size = 8;
7526 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7527 Op.getOperand(2), SDValue());
7528 SDValue Ops[] = { cpIn.getValue(0),
7531 DAG.getTargetConstant(size, MVT::i8),
7533 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7534 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7536 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7540 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7541 SelectionDAG &DAG) const {
7542 assert(Subtarget->is64Bit() && "Result not type legalized?");
7543 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7544 SDValue TheChain = Op.getOperand(0);
7545 DebugLoc dl = Op.getDebugLoc();
7546 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7547 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7548 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7550 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7551 DAG.getConstant(32, MVT::i8));
7553 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7556 return DAG.getMergeValues(Ops, 2, dl);
7559 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7560 SelectionDAG &DAG) const {
7561 EVT SrcVT = Op.getOperand(0).getValueType();
7562 EVT DstVT = Op.getValueType();
7563 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7564 Subtarget->hasMMX() && !DisableMMX) &&
7565 "Unexpected custom BIT_CONVERT");
7566 assert((DstVT == MVT::i64 ||
7567 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7568 "Unexpected custom BIT_CONVERT");
7569 // i64 <=> MMX conversions are Legal.
7570 if (SrcVT==MVT::i64 && DstVT.isVector())
7572 if (DstVT==MVT::i64 && SrcVT.isVector())
7574 // MMX <=> MMX conversions are Legal.
7575 if (SrcVT.isVector() && DstVT.isVector())
7577 // All other conversions need to be expanded.
7580 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7581 SDNode *Node = Op.getNode();
7582 DebugLoc dl = Node->getDebugLoc();
7583 EVT T = Node->getValueType(0);
7584 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7585 DAG.getConstant(0, T), Node->getOperand(2));
7586 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7587 cast<AtomicSDNode>(Node)->getMemoryVT(),
7588 Node->getOperand(0),
7589 Node->getOperand(1), negOp,
7590 cast<AtomicSDNode>(Node)->getSrcValue(),
7591 cast<AtomicSDNode>(Node)->getAlignment());
7594 /// LowerOperation - Provide custom lowering hooks for some operations.
7596 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7597 switch (Op.getOpcode()) {
7598 default: llvm_unreachable("Should not custom lower this!");
7599 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7600 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7601 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7602 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7603 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7604 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7605 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7606 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7607 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7608 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7609 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7610 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7611 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7612 case ISD::SHL_PARTS:
7613 case ISD::SRA_PARTS:
7614 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7615 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7616 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7617 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7618 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7619 case ISD::FABS: return LowerFABS(Op, DAG);
7620 case ISD::FNEG: return LowerFNEG(Op, DAG);
7621 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7622 case ISD::SETCC: return LowerSETCC(Op, DAG);
7623 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7624 case ISD::SELECT: return LowerSELECT(Op, DAG);
7625 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7626 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7627 case ISD::VASTART: return LowerVASTART(Op, DAG);
7628 case ISD::VAARG: return LowerVAARG(Op, DAG);
7629 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7630 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7631 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7632 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7633 case ISD::FRAME_TO_ARGS_OFFSET:
7634 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7635 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7636 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7637 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7638 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7639 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7640 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7641 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7647 case ISD::UMULO: return LowerXALUO(Op, DAG);
7648 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7649 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7653 void X86TargetLowering::
7654 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7655 SelectionDAG &DAG, unsigned NewOp) const {
7656 EVT T = Node->getValueType(0);
7657 DebugLoc dl = Node->getDebugLoc();
7658 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7660 SDValue Chain = Node->getOperand(0);
7661 SDValue In1 = Node->getOperand(1);
7662 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7663 Node->getOperand(2), DAG.getIntPtrConstant(0));
7664 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7665 Node->getOperand(2), DAG.getIntPtrConstant(1));
7666 SDValue Ops[] = { Chain, In1, In2L, In2H };
7667 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7669 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7670 cast<MemSDNode>(Node)->getMemOperand());
7671 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7672 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7673 Results.push_back(Result.getValue(2));
7676 /// ReplaceNodeResults - Replace a node with an illegal result type
7677 /// with a new node built out of custom code.
7678 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7679 SmallVectorImpl<SDValue>&Results,
7680 SelectionDAG &DAG) const {
7681 DebugLoc dl = N->getDebugLoc();
7682 switch (N->getOpcode()) {
7684 assert(false && "Do not know how to custom type legalize this operation!");
7686 case ISD::FP_TO_SINT: {
7687 std::pair<SDValue,SDValue> Vals =
7688 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7689 SDValue FIST = Vals.first, StackSlot = Vals.second;
7690 if (FIST.getNode() != 0) {
7691 EVT VT = N->getValueType(0);
7692 // Return a load from the stack slot.
7693 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7698 case ISD::READCYCLECOUNTER: {
7699 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7700 SDValue TheChain = N->getOperand(0);
7701 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7702 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7704 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7706 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7707 SDValue Ops[] = { eax, edx };
7708 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7709 Results.push_back(edx.getValue(1));
7712 case ISD::ATOMIC_CMP_SWAP: {
7713 EVT T = N->getValueType(0);
7714 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7715 SDValue cpInL, cpInH;
7716 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7717 DAG.getConstant(0, MVT::i32));
7718 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7719 DAG.getConstant(1, MVT::i32));
7720 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7721 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7723 SDValue swapInL, swapInH;
7724 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7725 DAG.getConstant(0, MVT::i32));
7726 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7727 DAG.getConstant(1, MVT::i32));
7728 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7730 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7731 swapInL.getValue(1));
7732 SDValue Ops[] = { swapInH.getValue(0),
7734 swapInH.getValue(1) };
7735 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7736 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7737 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7738 MVT::i32, Result.getValue(1));
7739 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7740 MVT::i32, cpOutL.getValue(2));
7741 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7742 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7743 Results.push_back(cpOutH.getValue(1));
7746 case ISD::ATOMIC_LOAD_ADD:
7747 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7749 case ISD::ATOMIC_LOAD_AND:
7750 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7752 case ISD::ATOMIC_LOAD_NAND:
7753 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7755 case ISD::ATOMIC_LOAD_OR:
7756 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7758 case ISD::ATOMIC_LOAD_SUB:
7759 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7761 case ISD::ATOMIC_LOAD_XOR:
7762 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7764 case ISD::ATOMIC_SWAP:
7765 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7770 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7772 default: return NULL;
7773 case X86ISD::BSF: return "X86ISD::BSF";
7774 case X86ISD::BSR: return "X86ISD::BSR";
7775 case X86ISD::SHLD: return "X86ISD::SHLD";
7776 case X86ISD::SHRD: return "X86ISD::SHRD";
7777 case X86ISD::FAND: return "X86ISD::FAND";
7778 case X86ISD::FOR: return "X86ISD::FOR";
7779 case X86ISD::FXOR: return "X86ISD::FXOR";
7780 case X86ISD::FSRL: return "X86ISD::FSRL";
7781 case X86ISD::FILD: return "X86ISD::FILD";
7782 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7783 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7784 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7785 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7786 case X86ISD::FLD: return "X86ISD::FLD";
7787 case X86ISD::FST: return "X86ISD::FST";
7788 case X86ISD::CALL: return "X86ISD::CALL";
7789 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7790 case X86ISD::BT: return "X86ISD::BT";
7791 case X86ISD::CMP: return "X86ISD::CMP";
7792 case X86ISD::COMI: return "X86ISD::COMI";
7793 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7794 case X86ISD::SETCC: return "X86ISD::SETCC";
7795 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7796 case X86ISD::CMOV: return "X86ISD::CMOV";
7797 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7798 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7799 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7800 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7801 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7802 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7803 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7804 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7805 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7806 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7807 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7808 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7809 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7810 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7811 case X86ISD::FMAX: return "X86ISD::FMAX";
7812 case X86ISD::FMIN: return "X86ISD::FMIN";
7813 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7814 case X86ISD::FRCP: return "X86ISD::FRCP";
7815 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7816 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
7817 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7818 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7819 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7820 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7821 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7822 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7823 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7824 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7825 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7826 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7827 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7828 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7829 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7830 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7831 case X86ISD::VSHL: return "X86ISD::VSHL";
7832 case X86ISD::VSRL: return "X86ISD::VSRL";
7833 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7834 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7835 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7836 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7837 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7838 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7839 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7840 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7841 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7842 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7843 case X86ISD::ADD: return "X86ISD::ADD";
7844 case X86ISD::SUB: return "X86ISD::SUB";
7845 case X86ISD::SMUL: return "X86ISD::SMUL";
7846 case X86ISD::UMUL: return "X86ISD::UMUL";
7847 case X86ISD::INC: return "X86ISD::INC";
7848 case X86ISD::DEC: return "X86ISD::DEC";
7849 case X86ISD::OR: return "X86ISD::OR";
7850 case X86ISD::XOR: return "X86ISD::XOR";
7851 case X86ISD::AND: return "X86ISD::AND";
7852 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7853 case X86ISD::PTEST: return "X86ISD::PTEST";
7854 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7855 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7859 // isLegalAddressingMode - Return true if the addressing mode represented
7860 // by AM is legal for this target, for a load/store of the specified type.
7861 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7862 const Type *Ty) const {
7863 // X86 supports extremely general addressing modes.
7864 CodeModel::Model M = getTargetMachine().getCodeModel();
7866 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7867 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7872 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7874 // If a reference to this global requires an extra load, we can't fold it.
7875 if (isGlobalStubReference(GVFlags))
7878 // If BaseGV requires a register for the PIC base, we cannot also have a
7879 // BaseReg specified.
7880 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7883 // If lower 4G is not available, then we must use rip-relative addressing.
7884 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7894 // These scales always work.
7899 // These scales are formed with basereg+scalereg. Only accept if there is
7904 default: // Other stuff never works.
7912 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7913 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7915 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7916 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7917 if (NumBits1 <= NumBits2)
7922 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7923 if (!VT1.isInteger() || !VT2.isInteger())
7925 unsigned NumBits1 = VT1.getSizeInBits();
7926 unsigned NumBits2 = VT2.getSizeInBits();
7927 if (NumBits1 <= NumBits2)
7932 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7933 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7934 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7937 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7938 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7939 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7942 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7943 // i16 instructions are longer (0x66 prefix) and potentially slower.
7944 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7947 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7948 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7949 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7950 /// are assumed to be legal.
7952 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7954 // Very little shuffling can be done for 64-bit vectors right now.
7955 if (VT.getSizeInBits() == 64)
7956 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
7958 // FIXME: pshufb, blends, shifts.
7959 return (VT.getVectorNumElements() == 2 ||
7960 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7961 isMOVLMask(M, VT) ||
7962 isSHUFPMask(M, VT) ||
7963 isPSHUFDMask(M, VT) ||
7964 isPSHUFHWMask(M, VT) ||
7965 isPSHUFLWMask(M, VT) ||
7966 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7967 isUNPCKLMask(M, VT) ||
7968 isUNPCKHMask(M, VT) ||
7969 isUNPCKL_v_undef_Mask(M, VT) ||
7970 isUNPCKH_v_undef_Mask(M, VT));
7974 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7976 unsigned NumElts = VT.getVectorNumElements();
7977 // FIXME: This collection of masks seems suspect.
7980 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7981 return (isMOVLMask(Mask, VT) ||
7982 isCommutedMOVLMask(Mask, VT, true) ||
7983 isSHUFPMask(Mask, VT) ||
7984 isCommutedSHUFPMask(Mask, VT));
7989 //===----------------------------------------------------------------------===//
7990 // X86 Scheduler Hooks
7991 //===----------------------------------------------------------------------===//
7993 // private utility function
7995 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7996 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() < X86AddrNumOperands + 4 &&
8043 "unexpected number of operands");
8044 DebugLoc dl = bInstr->getDebugLoc();
8045 MachineOperand& destOper = bInstr->getOperand(0);
8046 MachineOperand* argOpers[2 + X86AddrNumOperands];
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 = X86AddrNumOperands - 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(copyOpc), 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(copyOpc), 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 copyOpc = X86::MOV32rr;
8126 const unsigned NotOpc = X86::NOT32r;
8127 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8128 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8129 MachineFunction::iterator MBBIter = MBB;
8132 /// First build the CFG
8133 MachineFunction *F = MBB->getParent();
8134 MachineBasicBlock *thisMBB = MBB;
8135 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8136 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8137 F->insert(MBBIter, newMBB);
8138 F->insert(MBBIter, nextMBB);
8140 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8141 nextMBB->splice(nextMBB->begin(), thisMBB,
8142 llvm::next(MachineBasicBlock::iterator(bInstr)),
8144 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8146 // Update thisMBB to fall through to newMBB
8147 thisMBB->addSuccessor(newMBB);
8149 // newMBB jumps to itself and fall through to nextMBB
8150 newMBB->addSuccessor(nextMBB);
8151 newMBB->addSuccessor(newMBB);
8153 DebugLoc dl = bInstr->getDebugLoc();
8154 // Insert instructions into newMBB based on incoming instruction
8155 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8156 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8157 "unexpected number of operands");
8158 MachineOperand& dest1Oper = bInstr->getOperand(0);
8159 MachineOperand& dest2Oper = bInstr->getOperand(1);
8160 MachineOperand* argOpers[2 + X86AddrNumOperands];
8161 for (int i=0; i < 2 + X86AddrNumOperands; ++i) {
8162 argOpers[i] = &bInstr->getOperand(i+2);
8164 // We use some of the operands multiple times, so conservatively just
8165 // clear any kill flags that might be present.
8166 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8167 argOpers[i]->setIsKill(false);
8170 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8171 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8173 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8174 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8175 for (int i=0; i <= lastAddrIndx; ++i)
8176 (*MIB).addOperand(*argOpers[i]);
8177 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8178 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8179 // add 4 to displacement.
8180 for (int i=0; i <= lastAddrIndx-2; ++i)
8181 (*MIB).addOperand(*argOpers[i]);
8182 MachineOperand newOp3 = *(argOpers[3]);
8184 newOp3.setImm(newOp3.getImm()+4);
8186 newOp3.setOffset(newOp3.getOffset()+4);
8187 (*MIB).addOperand(newOp3);
8188 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8190 // t3/4 are defined later, at the bottom of the loop
8191 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8192 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8193 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8194 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8195 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8196 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8198 // The subsequent operations should be using the destination registers of
8199 //the PHI instructions.
8201 t1 = F->getRegInfo().createVirtualRegister(RC);
8202 t2 = F->getRegInfo().createVirtualRegister(RC);
8203 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8204 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8206 t1 = dest1Oper.getReg();
8207 t2 = dest2Oper.getReg();
8210 int valArgIndx = lastAddrIndx + 1;
8211 assert((argOpers[valArgIndx]->isReg() ||
8212 argOpers[valArgIndx]->isImm()) &&
8214 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8215 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8216 if (argOpers[valArgIndx]->isReg())
8217 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8219 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8220 if (regOpcL != X86::MOV32rr)
8222 (*MIB).addOperand(*argOpers[valArgIndx]);
8223 assert(argOpers[valArgIndx + 1]->isReg() ==
8224 argOpers[valArgIndx]->isReg());
8225 assert(argOpers[valArgIndx + 1]->isImm() ==
8226 argOpers[valArgIndx]->isImm());
8227 if (argOpers[valArgIndx + 1]->isReg())
8228 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8230 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8231 if (regOpcH != X86::MOV32rr)
8233 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8235 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8237 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8240 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8242 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8245 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8246 for (int i=0; i <= lastAddrIndx; ++i)
8247 (*MIB).addOperand(*argOpers[i]);
8249 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8250 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8251 bInstr->memoperands_end());
8253 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8254 MIB.addReg(X86::EAX);
8255 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8256 MIB.addReg(X86::EDX);
8259 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8261 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8265 // private utility function
8267 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8268 MachineBasicBlock *MBB,
8269 unsigned cmovOpc) const {
8270 // For the atomic min/max operator, we generate
8273 // ld t1 = [min/max.addr]
8274 // mov t2 = [min/max.val]
8276 // cmov[cond] t2 = t1
8278 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8280 // fallthrough -->nextMBB
8282 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8283 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8284 MachineFunction::iterator MBBIter = MBB;
8287 /// First build the CFG
8288 MachineFunction *F = MBB->getParent();
8289 MachineBasicBlock *thisMBB = MBB;
8290 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8291 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8292 F->insert(MBBIter, newMBB);
8293 F->insert(MBBIter, nextMBB);
8295 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8296 nextMBB->splice(nextMBB->begin(), thisMBB,
8297 llvm::next(MachineBasicBlock::iterator(mInstr)),
8299 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8301 // Update thisMBB to fall through to newMBB
8302 thisMBB->addSuccessor(newMBB);
8304 // newMBB jumps to newMBB and fall through to nextMBB
8305 newMBB->addSuccessor(nextMBB);
8306 newMBB->addSuccessor(newMBB);
8308 DebugLoc dl = mInstr->getDebugLoc();
8309 // Insert instructions into newMBB based on incoming instruction
8310 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8311 "unexpected number of operands");
8312 MachineOperand& destOper = mInstr->getOperand(0);
8313 MachineOperand* argOpers[2 + X86AddrNumOperands];
8314 int numArgs = mInstr->getNumOperands() - 1;
8315 for (int i=0; i < numArgs; ++i)
8316 argOpers[i] = &mInstr->getOperand(i+1);
8318 // x86 address has 4 operands: base, index, scale, and displacement
8319 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8320 int valArgIndx = lastAddrIndx + 1;
8322 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8323 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8324 for (int i=0; i <= lastAddrIndx; ++i)
8325 (*MIB).addOperand(*argOpers[i]);
8327 // We only support register and immediate values
8328 assert((argOpers[valArgIndx]->isReg() ||
8329 argOpers[valArgIndx]->isImm()) &&
8332 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8333 if (argOpers[valArgIndx]->isReg())
8334 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8336 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8337 (*MIB).addOperand(*argOpers[valArgIndx]);
8339 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8342 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8347 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8348 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8352 // Cmp and exchange if none has modified the memory location
8353 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8354 for (int i=0; i <= lastAddrIndx; ++i)
8355 (*MIB).addOperand(*argOpers[i]);
8357 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8358 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8359 mInstr->memoperands_end());
8361 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8362 MIB.addReg(X86::EAX);
8365 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8367 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
8371 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8372 // all of this code can be replaced with that in the .td file.
8374 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8375 unsigned numArgs, bool memArg) const {
8377 DebugLoc dl = MI->getDebugLoc();
8378 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8382 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8384 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8386 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8388 for (unsigned i = 0; i < numArgs; ++i) {
8389 MachineOperand &Op = MI->getOperand(i+1);
8391 if (!(Op.isReg() && Op.isImplicit()))
8395 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8398 MI->eraseFromParent();
8404 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8406 MachineBasicBlock *MBB) const {
8407 // Emit code to save XMM registers to the stack. The ABI says that the
8408 // number of registers to save is given in %al, so it's theoretically
8409 // possible to do an indirect jump trick to avoid saving all of them,
8410 // however this code takes a simpler approach and just executes all
8411 // of the stores if %al is non-zero. It's less code, and it's probably
8412 // easier on the hardware branch predictor, and stores aren't all that
8413 // expensive anyway.
8415 // Create the new basic blocks. One block contains all the XMM stores,
8416 // and one block is the final destination regardless of whether any
8417 // stores were performed.
8418 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8419 MachineFunction *F = MBB->getParent();
8420 MachineFunction::iterator MBBIter = MBB;
8422 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8423 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8424 F->insert(MBBIter, XMMSaveMBB);
8425 F->insert(MBBIter, EndMBB);
8427 // Transfer the remainder of MBB and its successor edges to EndMBB.
8428 EndMBB->splice(EndMBB->begin(), MBB,
8429 llvm::next(MachineBasicBlock::iterator(MI)),
8431 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
8433 // The original block will now fall through to the XMM save block.
8434 MBB->addSuccessor(XMMSaveMBB);
8435 // The XMMSaveMBB will fall through to the end block.
8436 XMMSaveMBB->addSuccessor(EndMBB);
8438 // Now add the instructions.
8439 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8440 DebugLoc DL = MI->getDebugLoc();
8442 unsigned CountReg = MI->getOperand(0).getReg();
8443 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8444 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8446 if (!Subtarget->isTargetWin64()) {
8447 // If %al is 0, branch around the XMM save block.
8448 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8449 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8450 MBB->addSuccessor(EndMBB);
8453 // In the XMM save block, save all the XMM argument registers.
8454 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8455 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8456 MachineMemOperand *MMO =
8457 F->getMachineMemOperand(
8458 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8459 MachineMemOperand::MOStore, Offset,
8460 /*Size=*/16, /*Align=*/16);
8461 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8462 .addFrameIndex(RegSaveFrameIndex)
8463 .addImm(/*Scale=*/1)
8464 .addReg(/*IndexReg=*/0)
8465 .addImm(/*Disp=*/Offset)
8466 .addReg(/*Segment=*/0)
8467 .addReg(MI->getOperand(i).getReg())
8468 .addMemOperand(MMO);
8471 MI->eraseFromParent(); // The pseudo instruction is gone now.
8477 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8478 MachineBasicBlock *BB) const {
8479 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8480 DebugLoc DL = MI->getDebugLoc();
8482 // To "insert" a SELECT_CC instruction, we actually have to insert the
8483 // diamond control-flow pattern. The incoming instruction knows the
8484 // destination vreg to set, the condition code register to branch on, the
8485 // true/false values to select between, and a branch opcode to use.
8486 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8487 MachineFunction::iterator It = BB;
8493 // cmpTY ccX, r1, r2
8495 // fallthrough --> copy0MBB
8496 MachineBasicBlock *thisMBB = BB;
8497 MachineFunction *F = BB->getParent();
8498 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8499 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8500 F->insert(It, copy0MBB);
8501 F->insert(It, sinkMBB);
8503 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8504 // live into the sink and copy blocks.
8505 const MachineFunction *MF = BB->getParent();
8506 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8507 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8509 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
8510 const MachineOperand &MO = MI->getOperand(I);
8511 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
8512 unsigned Reg = MO.getReg();
8513 if (Reg != X86::EFLAGS) continue;
8514 copy0MBB->addLiveIn(Reg);
8515 sinkMBB->addLiveIn(Reg);
8518 // Transfer the remainder of BB and its successor edges to sinkMBB.
8519 sinkMBB->splice(sinkMBB->begin(), BB,
8520 llvm::next(MachineBasicBlock::iterator(MI)),
8522 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
8524 // Add the true and fallthrough blocks as its successors.
8525 BB->addSuccessor(copy0MBB);
8526 BB->addSuccessor(sinkMBB);
8528 // Create the conditional branch instruction.
8530 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8531 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8534 // %FalseValue = ...
8535 // # fallthrough to sinkMBB
8536 copy0MBB->addSuccessor(sinkMBB);
8539 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8541 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
8542 TII->get(X86::PHI), MI->getOperand(0).getReg())
8543 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8544 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8546 MI->eraseFromParent(); // The pseudo instruction is gone now.
8551 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8552 MachineBasicBlock *BB) const {
8553 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8554 DebugLoc DL = MI->getDebugLoc();
8556 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8557 // non-trivial part is impdef of ESP.
8558 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8561 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
8562 .addExternalSymbol("_alloca")
8563 .addReg(X86::EAX, RegState::Implicit)
8564 .addReg(X86::ESP, RegState::Implicit)
8565 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8566 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8568 MI->eraseFromParent(); // The pseudo instruction is gone now.
8573 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8574 MachineBasicBlock *BB) const {
8575 // This is pretty easy. We're taking the value that we received from
8576 // our load from the relocation, sticking it in either RDI (x86-64)
8577 // or EAX and doing an indirect call. The return value will then
8578 // be in the normal return register.
8579 const X86InstrInfo *TII
8580 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8581 DebugLoc DL = MI->getDebugLoc();
8582 MachineFunction *F = BB->getParent();
8584 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8586 if (Subtarget->is64Bit()) {
8587 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8588 TII->get(X86::MOV64rm), X86::RDI)
8590 .addImm(0).addReg(0)
8591 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8592 MI->getOperand(3).getTargetFlags())
8594 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
8595 addDirectMem(MIB, X86::RDI).addReg(0);
8596 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8597 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8598 TII->get(X86::MOV32rm), X86::EAX)
8600 .addImm(0).addReg(0)
8601 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8602 MI->getOperand(3).getTargetFlags())
8604 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8605 addDirectMem(MIB, X86::EAX).addReg(0);
8607 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8608 TII->get(X86::MOV32rm), X86::EAX)
8609 .addReg(TII->getGlobalBaseReg(F))
8610 .addImm(0).addReg(0)
8611 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8612 MI->getOperand(3).getTargetFlags())
8614 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8615 addDirectMem(MIB, X86::EAX).addReg(0);
8618 MI->eraseFromParent(); // The pseudo instruction is gone now.
8623 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8624 MachineBasicBlock *BB) const {
8625 switch (MI->getOpcode()) {
8626 default: assert(false && "Unexpected instr type to insert");
8627 case X86::MINGW_ALLOCA:
8628 return EmitLoweredMingwAlloca(MI, BB);
8629 case X86::TLSCall_32:
8630 case X86::TLSCall_64:
8631 return EmitLoweredTLSCall(MI, BB);
8633 case X86::CMOV_V1I64:
8634 case X86::CMOV_FR32:
8635 case X86::CMOV_FR64:
8636 case X86::CMOV_V4F32:
8637 case X86::CMOV_V2F64:
8638 case X86::CMOV_V2I64:
8639 case X86::CMOV_GR16:
8640 case X86::CMOV_GR32:
8641 case X86::CMOV_RFP32:
8642 case X86::CMOV_RFP64:
8643 case X86::CMOV_RFP80:
8644 return EmitLoweredSelect(MI, BB);
8646 case X86::FP32_TO_INT16_IN_MEM:
8647 case X86::FP32_TO_INT32_IN_MEM:
8648 case X86::FP32_TO_INT64_IN_MEM:
8649 case X86::FP64_TO_INT16_IN_MEM:
8650 case X86::FP64_TO_INT32_IN_MEM:
8651 case X86::FP64_TO_INT64_IN_MEM:
8652 case X86::FP80_TO_INT16_IN_MEM:
8653 case X86::FP80_TO_INT32_IN_MEM:
8654 case X86::FP80_TO_INT64_IN_MEM: {
8655 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8656 DebugLoc DL = MI->getDebugLoc();
8658 // Change the floating point control register to use "round towards zero"
8659 // mode when truncating to an integer value.
8660 MachineFunction *F = BB->getParent();
8661 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8662 addFrameReference(BuildMI(*BB, MI, DL,
8663 TII->get(X86::FNSTCW16m)), CWFrameIdx);
8665 // Load the old value of the high byte of the control word...
8667 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8668 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
8671 // Set the high part to be round to zero...
8672 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8675 // Reload the modified control word now...
8676 addFrameReference(BuildMI(*BB, MI, DL,
8677 TII->get(X86::FLDCW16m)), CWFrameIdx);
8679 // Restore the memory image of control word to original value
8680 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8683 // Get the X86 opcode to use.
8685 switch (MI->getOpcode()) {
8686 default: llvm_unreachable("illegal opcode!");
8687 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8688 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8689 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8690 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8691 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8692 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8693 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8694 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8695 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8699 MachineOperand &Op = MI->getOperand(0);
8701 AM.BaseType = X86AddressMode::RegBase;
8702 AM.Base.Reg = Op.getReg();
8704 AM.BaseType = X86AddressMode::FrameIndexBase;
8705 AM.Base.FrameIndex = Op.getIndex();
8707 Op = MI->getOperand(1);
8709 AM.Scale = Op.getImm();
8710 Op = MI->getOperand(2);
8712 AM.IndexReg = Op.getImm();
8713 Op = MI->getOperand(3);
8714 if (Op.isGlobal()) {
8715 AM.GV = Op.getGlobal();
8717 AM.Disp = Op.getImm();
8719 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
8720 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8722 // Reload the original control word now.
8723 addFrameReference(BuildMI(*BB, MI, DL,
8724 TII->get(X86::FLDCW16m)), CWFrameIdx);
8726 MI->eraseFromParent(); // The pseudo instruction is gone now.
8729 // String/text processing lowering.
8730 case X86::PCMPISTRM128REG:
8731 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8732 case X86::PCMPISTRM128MEM:
8733 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8734 case X86::PCMPESTRM128REG:
8735 return EmitPCMP(MI, BB, 5, false /* in mem */);
8736 case X86::PCMPESTRM128MEM:
8737 return EmitPCMP(MI, BB, 5, true /* in mem */);
8740 case X86::ATOMAND32:
8741 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8742 X86::AND32ri, X86::MOV32rm,
8743 X86::LCMPXCHG32, X86::MOV32rr,
8744 X86::NOT32r, X86::EAX,
8745 X86::GR32RegisterClass);
8747 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8748 X86::OR32ri, X86::MOV32rm,
8749 X86::LCMPXCHG32, X86::MOV32rr,
8750 X86::NOT32r, X86::EAX,
8751 X86::GR32RegisterClass);
8752 case X86::ATOMXOR32:
8753 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8754 X86::XOR32ri, X86::MOV32rm,
8755 X86::LCMPXCHG32, X86::MOV32rr,
8756 X86::NOT32r, X86::EAX,
8757 X86::GR32RegisterClass);
8758 case X86::ATOMNAND32:
8759 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8760 X86::AND32ri, X86::MOV32rm,
8761 X86::LCMPXCHG32, X86::MOV32rr,
8762 X86::NOT32r, X86::EAX,
8763 X86::GR32RegisterClass, true);
8764 case X86::ATOMMIN32:
8765 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8766 case X86::ATOMMAX32:
8767 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8768 case X86::ATOMUMIN32:
8769 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8770 case X86::ATOMUMAX32:
8771 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8773 case X86::ATOMAND16:
8774 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8775 X86::AND16ri, X86::MOV16rm,
8776 X86::LCMPXCHG16, X86::MOV16rr,
8777 X86::NOT16r, X86::AX,
8778 X86::GR16RegisterClass);
8780 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8781 X86::OR16ri, X86::MOV16rm,
8782 X86::LCMPXCHG16, X86::MOV16rr,
8783 X86::NOT16r, X86::AX,
8784 X86::GR16RegisterClass);
8785 case X86::ATOMXOR16:
8786 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8787 X86::XOR16ri, X86::MOV16rm,
8788 X86::LCMPXCHG16, X86::MOV16rr,
8789 X86::NOT16r, X86::AX,
8790 X86::GR16RegisterClass);
8791 case X86::ATOMNAND16:
8792 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8793 X86::AND16ri, X86::MOV16rm,
8794 X86::LCMPXCHG16, X86::MOV16rr,
8795 X86::NOT16r, X86::AX,
8796 X86::GR16RegisterClass, true);
8797 case X86::ATOMMIN16:
8798 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8799 case X86::ATOMMAX16:
8800 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8801 case X86::ATOMUMIN16:
8802 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8803 case X86::ATOMUMAX16:
8804 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8807 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8808 X86::AND8ri, X86::MOV8rm,
8809 X86::LCMPXCHG8, X86::MOV8rr,
8810 X86::NOT8r, X86::AL,
8811 X86::GR8RegisterClass);
8813 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8814 X86::OR8ri, X86::MOV8rm,
8815 X86::LCMPXCHG8, X86::MOV8rr,
8816 X86::NOT8r, X86::AL,
8817 X86::GR8RegisterClass);
8819 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8820 X86::XOR8ri, X86::MOV8rm,
8821 X86::LCMPXCHG8, X86::MOV8rr,
8822 X86::NOT8r, X86::AL,
8823 X86::GR8RegisterClass);
8824 case X86::ATOMNAND8:
8825 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8826 X86::AND8ri, X86::MOV8rm,
8827 X86::LCMPXCHG8, X86::MOV8rr,
8828 X86::NOT8r, X86::AL,
8829 X86::GR8RegisterClass, true);
8830 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8831 // This group is for 64-bit host.
8832 case X86::ATOMAND64:
8833 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8834 X86::AND64ri32, X86::MOV64rm,
8835 X86::LCMPXCHG64, X86::MOV64rr,
8836 X86::NOT64r, X86::RAX,
8837 X86::GR64RegisterClass);
8839 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8840 X86::OR64ri32, X86::MOV64rm,
8841 X86::LCMPXCHG64, X86::MOV64rr,
8842 X86::NOT64r, X86::RAX,
8843 X86::GR64RegisterClass);
8844 case X86::ATOMXOR64:
8845 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8846 X86::XOR64ri32, X86::MOV64rm,
8847 X86::LCMPXCHG64, X86::MOV64rr,
8848 X86::NOT64r, X86::RAX,
8849 X86::GR64RegisterClass);
8850 case X86::ATOMNAND64:
8851 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8852 X86::AND64ri32, X86::MOV64rm,
8853 X86::LCMPXCHG64, X86::MOV64rr,
8854 X86::NOT64r, X86::RAX,
8855 X86::GR64RegisterClass, true);
8856 case X86::ATOMMIN64:
8857 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8858 case X86::ATOMMAX64:
8859 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8860 case X86::ATOMUMIN64:
8861 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8862 case X86::ATOMUMAX64:
8863 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8865 // This group does 64-bit operations on a 32-bit host.
8866 case X86::ATOMAND6432:
8867 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8868 X86::AND32rr, X86::AND32rr,
8869 X86::AND32ri, X86::AND32ri,
8871 case X86::ATOMOR6432:
8872 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8873 X86::OR32rr, X86::OR32rr,
8874 X86::OR32ri, X86::OR32ri,
8876 case X86::ATOMXOR6432:
8877 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8878 X86::XOR32rr, X86::XOR32rr,
8879 X86::XOR32ri, X86::XOR32ri,
8881 case X86::ATOMNAND6432:
8882 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8883 X86::AND32rr, X86::AND32rr,
8884 X86::AND32ri, X86::AND32ri,
8886 case X86::ATOMADD6432:
8887 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8888 X86::ADD32rr, X86::ADC32rr,
8889 X86::ADD32ri, X86::ADC32ri,
8891 case X86::ATOMSUB6432:
8892 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8893 X86::SUB32rr, X86::SBB32rr,
8894 X86::SUB32ri, X86::SBB32ri,
8896 case X86::ATOMSWAP6432:
8897 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8898 X86::MOV32rr, X86::MOV32rr,
8899 X86::MOV32ri, X86::MOV32ri,
8901 case X86::VASTART_SAVE_XMM_REGS:
8902 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8906 //===----------------------------------------------------------------------===//
8907 // X86 Optimization Hooks
8908 //===----------------------------------------------------------------------===//
8910 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8914 const SelectionDAG &DAG,
8915 unsigned Depth) const {
8916 unsigned Opc = Op.getOpcode();
8917 assert((Opc >= ISD::BUILTIN_OP_END ||
8918 Opc == ISD::INTRINSIC_WO_CHAIN ||
8919 Opc == ISD::INTRINSIC_W_CHAIN ||
8920 Opc == ISD::INTRINSIC_VOID) &&
8921 "Should use MaskedValueIsZero if you don't know whether Op"
8922 " is a target node!");
8924 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8936 // These nodes' second result is a boolean.
8937 if (Op.getResNo() == 0)
8941 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8942 Mask.getBitWidth() - 1);
8947 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8948 /// node is a GlobalAddress + offset.
8949 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8950 const GlobalValue* &GA,
8951 int64_t &Offset) const {
8952 if (N->getOpcode() == X86ISD::Wrapper) {
8953 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8954 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8955 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8959 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8962 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8963 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8964 /// if the load addresses are consecutive, non-overlapping, and in the right
8966 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8967 const TargetLowering &TLI) {
8968 DebugLoc dl = N->getDebugLoc();
8969 EVT VT = N->getValueType(0);
8970 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8972 if (VT.getSizeInBits() != 128)
8975 SmallVector<SDValue, 16> Elts;
8976 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8977 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8979 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8982 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8983 /// and convert it from being a bunch of shuffles and extracts to a simple
8984 /// store and scalar loads to extract the elements.
8985 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8986 const TargetLowering &TLI) {
8987 SDValue InputVector = N->getOperand(0);
8989 // Only operate on vectors of 4 elements, where the alternative shuffling
8990 // gets to be more expensive.
8991 if (InputVector.getValueType() != MVT::v4i32)
8994 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8995 // single use which is a sign-extend or zero-extend, and all elements are
8997 SmallVector<SDNode *, 4> Uses;
8998 unsigned ExtractedElements = 0;
8999 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
9000 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
9001 if (UI.getUse().getResNo() != InputVector.getResNo())
9004 SDNode *Extract = *UI;
9005 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9008 if (Extract->getValueType(0) != MVT::i32)
9010 if (!Extract->hasOneUse())
9012 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9013 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9015 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9018 // Record which element was extracted.
9019 ExtractedElements |=
9020 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9022 Uses.push_back(Extract);
9025 // If not all the elements were used, this may not be worthwhile.
9026 if (ExtractedElements != 15)
9029 // Ok, we've now decided to do the transformation.
9030 DebugLoc dl = InputVector.getDebugLoc();
9032 // Store the value to a temporary stack slot.
9033 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9034 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
9037 // Replace each use (extract) with a load of the appropriate element.
9038 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9039 UE = Uses.end(); UI != UE; ++UI) {
9040 SDNode *Extract = *UI;
9042 // Compute the element's address.
9043 SDValue Idx = Extract->getOperand(1);
9045 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9046 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9047 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9049 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
9052 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
9053 NULL, 0, false, false, 0);
9055 // Replace the exact with the load.
9056 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9059 // The replacement was made in place; don't return anything.
9063 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9064 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9065 const X86Subtarget *Subtarget) {
9066 DebugLoc DL = N->getDebugLoc();
9067 SDValue Cond = N->getOperand(0);
9068 // Get the LHS/RHS of the select.
9069 SDValue LHS = N->getOperand(1);
9070 SDValue RHS = N->getOperand(2);
9072 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9073 // instructions match the semantics of the common C idiom x<y?x:y but not
9074 // x<=y?x:y, because of how they handle negative zero (which can be
9075 // ignored in unsafe-math mode).
9076 if (Subtarget->hasSSE2() &&
9077 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9078 Cond.getOpcode() == ISD::SETCC) {
9079 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9081 unsigned Opcode = 0;
9082 // Check for x CC y ? x : y.
9083 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9084 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9088 // Converting this to a min would handle NaNs incorrectly, and swapping
9089 // the operands would cause it to handle comparisons between positive
9090 // and negative zero incorrectly.
9091 if (!FiniteOnlyFPMath() &&
9092 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9093 if (!UnsafeFPMath &&
9094 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9096 std::swap(LHS, RHS);
9098 Opcode = X86ISD::FMIN;
9101 // Converting this to a min would handle comparisons between positive
9102 // and negative zero incorrectly.
9103 if (!UnsafeFPMath &&
9104 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9106 Opcode = X86ISD::FMIN;
9109 // Converting this to a min would handle both negative zeros and NaNs
9110 // incorrectly, but we can swap the operands to fix both.
9111 std::swap(LHS, RHS);
9115 Opcode = X86ISD::FMIN;
9119 // Converting this to a max would handle comparisons between positive
9120 // and negative zero incorrectly.
9121 if (!UnsafeFPMath &&
9122 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9124 Opcode = X86ISD::FMAX;
9127 // Converting this to a max would handle NaNs incorrectly, and swapping
9128 // the operands would cause it to handle comparisons between positive
9129 // and negative zero incorrectly.
9130 if (!FiniteOnlyFPMath() &&
9131 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9132 if (!UnsafeFPMath &&
9133 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9135 std::swap(LHS, RHS);
9137 Opcode = X86ISD::FMAX;
9140 // Converting this to a max would handle both negative zeros and NaNs
9141 // incorrectly, but we can swap the operands to fix both.
9142 std::swap(LHS, RHS);
9146 Opcode = X86ISD::FMAX;
9149 // Check for x CC y ? y : x -- a min/max with reversed arms.
9150 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9151 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9155 // Converting this to a min would handle comparisons between positive
9156 // and negative zero incorrectly, and swapping the operands would
9157 // cause it to handle NaNs incorrectly.
9158 if (!UnsafeFPMath &&
9159 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9160 if (!FiniteOnlyFPMath() &&
9161 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9163 std::swap(LHS, RHS);
9165 Opcode = X86ISD::FMIN;
9168 // Converting this to a min would handle NaNs incorrectly.
9169 if (!UnsafeFPMath &&
9170 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9172 Opcode = X86ISD::FMIN;
9175 // Converting this to a min would handle both negative zeros and NaNs
9176 // incorrectly, but we can swap the operands to fix both.
9177 std::swap(LHS, RHS);
9181 Opcode = X86ISD::FMIN;
9185 // Converting this to a max would handle NaNs incorrectly.
9186 if (!FiniteOnlyFPMath() &&
9187 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9189 Opcode = X86ISD::FMAX;
9192 // Converting this to a max would handle comparisons between positive
9193 // and negative zero incorrectly, and swapping the operands would
9194 // cause it to handle NaNs incorrectly.
9195 if (!UnsafeFPMath &&
9196 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9197 if (!FiniteOnlyFPMath() &&
9198 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9200 std::swap(LHS, RHS);
9202 Opcode = X86ISD::FMAX;
9205 // Converting this to a max would handle both negative zeros and NaNs
9206 // incorrectly, but we can swap the operands to fix both.
9207 std::swap(LHS, RHS);
9211 Opcode = X86ISD::FMAX;
9217 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9220 // If this is a select between two integer constants, try to do some
9222 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9223 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9224 // Don't do this for crazy integer types.
9225 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9226 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9227 // so that TrueC (the true value) is larger than FalseC.
9228 bool NeedsCondInvert = false;
9230 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9231 // Efficiently invertible.
9232 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9233 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9234 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9235 NeedsCondInvert = true;
9236 std::swap(TrueC, FalseC);
9239 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9240 if (FalseC->getAPIntValue() == 0 &&
9241 TrueC->getAPIntValue().isPowerOf2()) {
9242 if (NeedsCondInvert) // Invert the condition if needed.
9243 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9244 DAG.getConstant(1, Cond.getValueType()));
9246 // Zero extend the condition if needed.
9247 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9249 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9250 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9251 DAG.getConstant(ShAmt, MVT::i8));
9254 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9255 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9256 if (NeedsCondInvert) // Invert the condition if needed.
9257 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9258 DAG.getConstant(1, Cond.getValueType()));
9260 // Zero extend the condition if needed.
9261 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9262 FalseC->getValueType(0), Cond);
9263 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9264 SDValue(FalseC, 0));
9267 // Optimize cases that will turn into an LEA instruction. This requires
9268 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9269 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9270 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9271 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9273 bool isFastMultiplier = false;
9275 switch ((unsigned char)Diff) {
9277 case 1: // result = add base, cond
9278 case 2: // result = lea base( , cond*2)
9279 case 3: // result = lea base(cond, cond*2)
9280 case 4: // result = lea base( , cond*4)
9281 case 5: // result = lea base(cond, cond*4)
9282 case 8: // result = lea base( , cond*8)
9283 case 9: // result = lea base(cond, cond*8)
9284 isFastMultiplier = true;
9289 if (isFastMultiplier) {
9290 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9291 if (NeedsCondInvert) // Invert the condition if needed.
9292 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9293 DAG.getConstant(1, Cond.getValueType()));
9295 // Zero extend the condition if needed.
9296 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9298 // Scale the condition by the difference.
9300 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9301 DAG.getConstant(Diff, Cond.getValueType()));
9303 // Add the base if non-zero.
9304 if (FalseC->getAPIntValue() != 0)
9305 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9306 SDValue(FalseC, 0));
9316 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9317 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9318 TargetLowering::DAGCombinerInfo &DCI) {
9319 DebugLoc DL = N->getDebugLoc();
9321 // If the flag operand isn't dead, don't touch this CMOV.
9322 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9325 // If this is a select between two integer constants, try to do some
9326 // optimizations. Note that the operands are ordered the opposite of SELECT
9328 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9329 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9330 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9331 // larger than FalseC (the false value).
9332 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9334 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9335 CC = X86::GetOppositeBranchCondition(CC);
9336 std::swap(TrueC, FalseC);
9339 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9340 // This is efficient for any integer data type (including i8/i16) and
9342 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9343 SDValue Cond = N->getOperand(3);
9344 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9345 DAG.getConstant(CC, MVT::i8), Cond);
9347 // Zero extend the condition if needed.
9348 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9350 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9351 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9352 DAG.getConstant(ShAmt, MVT::i8));
9353 if (N->getNumValues() == 2) // Dead flag value?
9354 return DCI.CombineTo(N, Cond, SDValue());
9358 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9359 // for any integer data type, including i8/i16.
9360 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9361 SDValue Cond = N->getOperand(3);
9362 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9363 DAG.getConstant(CC, MVT::i8), Cond);
9365 // Zero extend the condition if needed.
9366 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9367 FalseC->getValueType(0), Cond);
9368 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9369 SDValue(FalseC, 0));
9371 if (N->getNumValues() == 2) // Dead flag value?
9372 return DCI.CombineTo(N, Cond, SDValue());
9376 // Optimize cases that will turn into an LEA instruction. This requires
9377 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9378 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9379 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9380 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9382 bool isFastMultiplier = false;
9384 switch ((unsigned char)Diff) {
9386 case 1: // result = add base, cond
9387 case 2: // result = lea base( , cond*2)
9388 case 3: // result = lea base(cond, cond*2)
9389 case 4: // result = lea base( , cond*4)
9390 case 5: // result = lea base(cond, cond*4)
9391 case 8: // result = lea base( , cond*8)
9392 case 9: // result = lea base(cond, cond*8)
9393 isFastMultiplier = true;
9398 if (isFastMultiplier) {
9399 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9400 SDValue Cond = N->getOperand(3);
9401 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9402 DAG.getConstant(CC, MVT::i8), Cond);
9403 // Zero extend the condition if needed.
9404 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9406 // Scale the condition by the difference.
9408 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9409 DAG.getConstant(Diff, Cond.getValueType()));
9411 // Add the base if non-zero.
9412 if (FalseC->getAPIntValue() != 0)
9413 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9414 SDValue(FalseC, 0));
9415 if (N->getNumValues() == 2) // Dead flag value?
9416 return DCI.CombineTo(N, Cond, SDValue());
9426 /// PerformMulCombine - Optimize a single multiply with constant into two
9427 /// in order to implement it with two cheaper instructions, e.g.
9428 /// LEA + SHL, LEA + LEA.
9429 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9430 TargetLowering::DAGCombinerInfo &DCI) {
9431 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9434 EVT VT = N->getValueType(0);
9438 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9441 uint64_t MulAmt = C->getZExtValue();
9442 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9445 uint64_t MulAmt1 = 0;
9446 uint64_t MulAmt2 = 0;
9447 if ((MulAmt % 9) == 0) {
9449 MulAmt2 = MulAmt / 9;
9450 } else if ((MulAmt % 5) == 0) {
9452 MulAmt2 = MulAmt / 5;
9453 } else if ((MulAmt % 3) == 0) {
9455 MulAmt2 = MulAmt / 3;
9458 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9459 DebugLoc DL = N->getDebugLoc();
9461 if (isPowerOf2_64(MulAmt2) &&
9462 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9463 // If second multiplifer is pow2, issue it first. We want the multiply by
9464 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9466 std::swap(MulAmt1, MulAmt2);
9469 if (isPowerOf2_64(MulAmt1))
9470 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9471 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9473 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9474 DAG.getConstant(MulAmt1, VT));
9476 if (isPowerOf2_64(MulAmt2))
9477 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9478 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9480 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9481 DAG.getConstant(MulAmt2, VT));
9483 // Do not add new nodes to DAG combiner worklist.
9484 DCI.CombineTo(N, NewMul, false);
9489 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9490 SDValue N0 = N->getOperand(0);
9491 SDValue N1 = N->getOperand(1);
9492 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9493 EVT VT = N0.getValueType();
9495 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9496 // since the result of setcc_c is all zero's or all ones.
9497 if (N1C && N0.getOpcode() == ISD::AND &&
9498 N0.getOperand(1).getOpcode() == ISD::Constant) {
9499 SDValue N00 = N0.getOperand(0);
9500 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9501 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9502 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9503 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9504 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9505 APInt ShAmt = N1C->getAPIntValue();
9506 Mask = Mask.shl(ShAmt);
9508 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9509 N00, DAG.getConstant(Mask, VT));
9516 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9518 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9519 const X86Subtarget *Subtarget) {
9520 EVT VT = N->getValueType(0);
9521 if (!VT.isVector() && VT.isInteger() &&
9522 N->getOpcode() == ISD::SHL)
9523 return PerformSHLCombine(N, DAG);
9525 // On X86 with SSE2 support, we can transform this to a vector shift if
9526 // all elements are shifted by the same amount. We can't do this in legalize
9527 // because the a constant vector is typically transformed to a constant pool
9528 // so we have no knowledge of the shift amount.
9529 if (!Subtarget->hasSSE2())
9532 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9535 SDValue ShAmtOp = N->getOperand(1);
9536 EVT EltVT = VT.getVectorElementType();
9537 DebugLoc DL = N->getDebugLoc();
9538 SDValue BaseShAmt = SDValue();
9539 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9540 unsigned NumElts = VT.getVectorNumElements();
9542 for (; i != NumElts; ++i) {
9543 SDValue Arg = ShAmtOp.getOperand(i);
9544 if (Arg.getOpcode() == ISD::UNDEF) continue;
9548 for (; i != NumElts; ++i) {
9549 SDValue Arg = ShAmtOp.getOperand(i);
9550 if (Arg.getOpcode() == ISD::UNDEF) continue;
9551 if (Arg != BaseShAmt) {
9555 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9556 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9557 SDValue InVec = ShAmtOp.getOperand(0);
9558 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9559 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9561 for (; i != NumElts; ++i) {
9562 SDValue Arg = InVec.getOperand(i);
9563 if (Arg.getOpcode() == ISD::UNDEF) continue;
9567 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9568 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9569 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9570 if (C->getZExtValue() == SplatIdx)
9571 BaseShAmt = InVec.getOperand(1);
9574 if (BaseShAmt.getNode() == 0)
9575 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9576 DAG.getIntPtrConstant(0));
9580 // The shift amount is an i32.
9581 if (EltVT.bitsGT(MVT::i32))
9582 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9583 else if (EltVT.bitsLT(MVT::i32))
9584 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9586 // The shift amount is identical so we can do a vector shift.
9587 SDValue ValOp = N->getOperand(0);
9588 switch (N->getOpcode()) {
9590 llvm_unreachable("Unknown shift opcode!");
9593 if (VT == MVT::v2i64)
9594 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9595 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9597 if (VT == MVT::v4i32)
9598 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9599 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9601 if (VT == MVT::v8i16)
9602 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9603 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9607 if (VT == MVT::v4i32)
9608 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9609 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9611 if (VT == MVT::v8i16)
9612 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9613 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9617 if (VT == MVT::v2i64)
9618 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9619 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9621 if (VT == MVT::v4i32)
9622 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9623 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9625 if (VT == MVT::v8i16)
9626 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9627 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9634 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9635 TargetLowering::DAGCombinerInfo &DCI,
9636 const X86Subtarget *Subtarget) {
9637 if (DCI.isBeforeLegalizeOps())
9640 EVT VT = N->getValueType(0);
9641 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9644 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9645 SDValue N0 = N->getOperand(0);
9646 SDValue N1 = N->getOperand(1);
9647 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9649 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9651 if (!N0.hasOneUse() || !N1.hasOneUse())
9654 SDValue ShAmt0 = N0.getOperand(1);
9655 if (ShAmt0.getValueType() != MVT::i8)
9657 SDValue ShAmt1 = N1.getOperand(1);
9658 if (ShAmt1.getValueType() != MVT::i8)
9660 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9661 ShAmt0 = ShAmt0.getOperand(0);
9662 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9663 ShAmt1 = ShAmt1.getOperand(0);
9665 DebugLoc DL = N->getDebugLoc();
9666 unsigned Opc = X86ISD::SHLD;
9667 SDValue Op0 = N0.getOperand(0);
9668 SDValue Op1 = N1.getOperand(0);
9669 if (ShAmt0.getOpcode() == ISD::SUB) {
9671 std::swap(Op0, Op1);
9672 std::swap(ShAmt0, ShAmt1);
9675 unsigned Bits = VT.getSizeInBits();
9676 if (ShAmt1.getOpcode() == ISD::SUB) {
9677 SDValue Sum = ShAmt1.getOperand(0);
9678 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9679 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
9680 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
9681 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
9682 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
9683 return DAG.getNode(Opc, DL, VT,
9685 DAG.getNode(ISD::TRUNCATE, DL,
9688 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9689 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9691 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9692 return DAG.getNode(Opc, DL, VT,
9693 N0.getOperand(0), N1.getOperand(0),
9694 DAG.getNode(ISD::TRUNCATE, DL,
9701 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9702 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9703 const X86Subtarget *Subtarget) {
9704 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9705 // the FP state in cases where an emms may be missing.
9706 // A preferable solution to the general problem is to figure out the right
9707 // places to insert EMMS. This qualifies as a quick hack.
9709 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9710 StoreSDNode *St = cast<StoreSDNode>(N);
9711 EVT VT = St->getValue().getValueType();
9712 if (VT.getSizeInBits() != 64)
9715 const Function *F = DAG.getMachineFunction().getFunction();
9716 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9717 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9718 && Subtarget->hasSSE2();
9719 if ((VT.isVector() ||
9720 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9721 isa<LoadSDNode>(St->getValue()) &&
9722 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9723 St->getChain().hasOneUse() && !St->isVolatile()) {
9724 SDNode* LdVal = St->getValue().getNode();
9726 int TokenFactorIndex = -1;
9727 SmallVector<SDValue, 8> Ops;
9728 SDNode* ChainVal = St->getChain().getNode();
9729 // Must be a store of a load. We currently handle two cases: the load
9730 // is a direct child, and it's under an intervening TokenFactor. It is
9731 // possible to dig deeper under nested TokenFactors.
9732 if (ChainVal == LdVal)
9733 Ld = cast<LoadSDNode>(St->getChain());
9734 else if (St->getValue().hasOneUse() &&
9735 ChainVal->getOpcode() == ISD::TokenFactor) {
9736 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9737 if (ChainVal->getOperand(i).getNode() == LdVal) {
9738 TokenFactorIndex = i;
9739 Ld = cast<LoadSDNode>(St->getValue());
9741 Ops.push_back(ChainVal->getOperand(i));
9745 if (!Ld || !ISD::isNormalLoad(Ld))
9748 // If this is not the MMX case, i.e. we are just turning i64 load/store
9749 // into f64 load/store, avoid the transformation if there are multiple
9750 // uses of the loaded value.
9751 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9754 DebugLoc LdDL = Ld->getDebugLoc();
9755 DebugLoc StDL = N->getDebugLoc();
9756 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9757 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9759 if (Subtarget->is64Bit() || F64IsLegal) {
9760 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9761 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9762 Ld->getBasePtr(), Ld->getSrcValue(),
9763 Ld->getSrcValueOffset(), Ld->isVolatile(),
9764 Ld->isNonTemporal(), Ld->getAlignment());
9765 SDValue NewChain = NewLd.getValue(1);
9766 if (TokenFactorIndex != -1) {
9767 Ops.push_back(NewChain);
9768 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9771 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9772 St->getSrcValue(), St->getSrcValueOffset(),
9773 St->isVolatile(), St->isNonTemporal(),
9774 St->getAlignment());
9777 // Otherwise, lower to two pairs of 32-bit loads / stores.
9778 SDValue LoAddr = Ld->getBasePtr();
9779 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9780 DAG.getConstant(4, MVT::i32));
9782 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9783 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9784 Ld->isVolatile(), Ld->isNonTemporal(),
9785 Ld->getAlignment());
9786 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9787 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9788 Ld->isVolatile(), Ld->isNonTemporal(),
9789 MinAlign(Ld->getAlignment(), 4));
9791 SDValue NewChain = LoLd.getValue(1);
9792 if (TokenFactorIndex != -1) {
9793 Ops.push_back(LoLd);
9794 Ops.push_back(HiLd);
9795 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9799 LoAddr = St->getBasePtr();
9800 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9801 DAG.getConstant(4, MVT::i32));
9803 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9804 St->getSrcValue(), St->getSrcValueOffset(),
9805 St->isVolatile(), St->isNonTemporal(),
9806 St->getAlignment());
9807 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9809 St->getSrcValueOffset() + 4,
9811 St->isNonTemporal(),
9812 MinAlign(St->getAlignment(), 4));
9813 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9818 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9819 /// X86ISD::FXOR nodes.
9820 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9821 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9822 // F[X]OR(0.0, x) -> x
9823 // F[X]OR(x, 0.0) -> x
9824 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9825 if (C->getValueAPF().isPosZero())
9826 return N->getOperand(1);
9827 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9828 if (C->getValueAPF().isPosZero())
9829 return N->getOperand(0);
9833 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9834 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9835 // FAND(0.0, x) -> 0.0
9836 // FAND(x, 0.0) -> 0.0
9837 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9838 if (C->getValueAPF().isPosZero())
9839 return N->getOperand(0);
9840 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9841 if (C->getValueAPF().isPosZero())
9842 return N->getOperand(1);
9846 static SDValue PerformBTCombine(SDNode *N,
9848 TargetLowering::DAGCombinerInfo &DCI) {
9849 // BT ignores high bits in the bit index operand.
9850 SDValue Op1 = N->getOperand(1);
9851 if (Op1.hasOneUse()) {
9852 unsigned BitWidth = Op1.getValueSizeInBits();
9853 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9854 APInt KnownZero, KnownOne;
9855 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9856 !DCI.isBeforeLegalizeOps());
9857 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9858 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9859 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9860 DCI.CommitTargetLoweringOpt(TLO);
9865 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9866 SDValue Op = N->getOperand(0);
9867 if (Op.getOpcode() == ISD::BIT_CONVERT)
9868 Op = Op.getOperand(0);
9869 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9870 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9871 VT.getVectorElementType().getSizeInBits() ==
9872 OpVT.getVectorElementType().getSizeInBits()) {
9873 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9878 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9879 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9880 // (and (i32 x86isd::setcc_carry), 1)
9881 // This eliminates the zext. This transformation is necessary because
9882 // ISD::SETCC is always legalized to i8.
9883 DebugLoc dl = N->getDebugLoc();
9884 SDValue N0 = N->getOperand(0);
9885 EVT VT = N->getValueType(0);
9886 if (N0.getOpcode() == ISD::AND &&
9888 N0.getOperand(0).hasOneUse()) {
9889 SDValue N00 = N0.getOperand(0);
9890 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9892 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9893 if (!C || C->getZExtValue() != 1)
9895 return DAG.getNode(ISD::AND, dl, VT,
9896 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9897 N00.getOperand(0), N00.getOperand(1)),
9898 DAG.getConstant(1, VT));
9904 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9905 DAGCombinerInfo &DCI) const {
9906 SelectionDAG &DAG = DCI.DAG;
9907 switch (N->getOpcode()) {
9909 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9910 case ISD::EXTRACT_VECTOR_ELT:
9911 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9912 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9913 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9914 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9917 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9918 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
9919 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9921 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9922 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9923 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9924 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9925 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9931 /// isTypeDesirableForOp - Return true if the target has native support for
9932 /// the specified value type and it is 'desirable' to use the type for the
9933 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9934 /// instruction encodings are longer and some i16 instructions are slow.
9935 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9936 if (!isTypeLegal(VT))
9945 case ISD::SIGN_EXTEND:
9946 case ISD::ZERO_EXTEND:
9947 case ISD::ANY_EXTEND:
9960 static bool MayFoldLoad(SDValue Op) {
9961 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
9964 static bool MayFoldIntoStore(SDValue Op) {
9965 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
9968 /// IsDesirableToPromoteOp - This method query the target whether it is
9969 /// beneficial for dag combiner to promote the specified node. If true, it
9970 /// should return the desired promotion type by reference.
9971 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
9972 EVT VT = Op.getValueType();
9976 bool Promote = false;
9977 bool Commute = false;
9978 switch (Op.getOpcode()) {
9981 LoadSDNode *LD = cast<LoadSDNode>(Op);
9982 // If the non-extending load has a single use and it's not live out, then it
9984 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
9986 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9987 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
9988 // The only case where we'd want to promote LOAD (rather then it being
9989 // promoted as an operand is when it's only use is liveout.
9990 if (UI->getOpcode() != ISD::CopyToReg)
9997 case ISD::SIGN_EXTEND:
9998 case ISD::ZERO_EXTEND:
9999 case ISD::ANY_EXTEND:
10004 SDValue N0 = Op.getOperand(0);
10005 // Look out for (store (shl (load), x)).
10006 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10019 SDValue N0 = Op.getOperand(0);
10020 SDValue N1 = Op.getOperand(1);
10021 if (!Commute && MayFoldLoad(N1))
10023 // Avoid disabling potential load folding opportunities.
10024 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10026 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10036 //===----------------------------------------------------------------------===//
10037 // X86 Inline Assembly Support
10038 //===----------------------------------------------------------------------===//
10040 static bool LowerToBSwap(CallInst *CI) {
10041 // FIXME: this should verify that we are targetting a 486 or better. If not,
10042 // we will turn this bswap into something that will be lowered to logical ops
10043 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10044 // so don't worry about this.
10046 // Verify this is a simple bswap.
10047 if (CI->getNumArgOperands() != 1 ||
10048 CI->getType() != CI->getArgOperand(0)->getType() ||
10049 !CI->getType()->isIntegerTy())
10052 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10053 if (!Ty || Ty->getBitWidth() % 16 != 0)
10056 // Okay, we can do this xform, do so now.
10057 const Type *Tys[] = { Ty };
10058 Module *M = CI->getParent()->getParent()->getParent();
10059 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10061 Value *Op = CI->getArgOperand(0);
10062 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10064 CI->replaceAllUsesWith(Op);
10065 CI->eraseFromParent();
10069 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10070 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10071 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10073 std::string AsmStr = IA->getAsmString();
10075 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10076 SmallVector<StringRef, 4> AsmPieces;
10077 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10079 switch (AsmPieces.size()) {
10080 default: return false;
10082 AsmStr = AsmPieces[0];
10084 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10087 if (AsmPieces.size() == 2 &&
10088 (AsmPieces[0] == "bswap" ||
10089 AsmPieces[0] == "bswapq" ||
10090 AsmPieces[0] == "bswapl") &&
10091 (AsmPieces[1] == "$0" ||
10092 AsmPieces[1] == "${0:q}")) {
10093 // No need to check constraints, nothing other than the equivalent of
10094 // "=r,0" would be valid here.
10095 return LowerToBSwap(CI);
10097 // rorw $$8, ${0:w} --> llvm.bswap.i16
10098 if (CI->getType()->isIntegerTy(16) &&
10099 AsmPieces.size() == 3 &&
10100 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10101 AsmPieces[1] == "$$8," &&
10102 AsmPieces[2] == "${0:w}" &&
10103 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10105 const std::string &Constraints = IA->getConstraintString();
10106 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10107 std::sort(AsmPieces.begin(), AsmPieces.end());
10108 if (AsmPieces.size() == 4 &&
10109 AsmPieces[0] == "~{cc}" &&
10110 AsmPieces[1] == "~{dirflag}" &&
10111 AsmPieces[2] == "~{flags}" &&
10112 AsmPieces[3] == "~{fpsr}") {
10113 return LowerToBSwap(CI);
10118 if (CI->getType()->isIntegerTy(64) &&
10119 Constraints.size() >= 2 &&
10120 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10121 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10122 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10123 SmallVector<StringRef, 4> Words;
10124 SplitString(AsmPieces[0], Words, " \t");
10125 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10127 SplitString(AsmPieces[1], Words, " \t");
10128 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10130 SplitString(AsmPieces[2], Words, " \t,");
10131 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10132 Words[2] == "%edx") {
10133 return LowerToBSwap(CI);
10145 /// getConstraintType - Given a constraint letter, return the type of
10146 /// constraint it is for this target.
10147 X86TargetLowering::ConstraintType
10148 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10149 if (Constraint.size() == 1) {
10150 switch (Constraint[0]) {
10162 return C_RegisterClass;
10170 return TargetLowering::getConstraintType(Constraint);
10173 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10174 /// with another that has more specific requirements based on the type of the
10175 /// corresponding operand.
10176 const char *X86TargetLowering::
10177 LowerXConstraint(EVT ConstraintVT) const {
10178 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10179 // 'f' like normal targets.
10180 if (ConstraintVT.isFloatingPoint()) {
10181 if (Subtarget->hasSSE2())
10183 if (Subtarget->hasSSE1())
10187 return TargetLowering::LowerXConstraint(ConstraintVT);
10190 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10191 /// vector. If it is invalid, don't add anything to Ops.
10192 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10194 std::vector<SDValue>&Ops,
10195 SelectionDAG &DAG) const {
10196 SDValue Result(0, 0);
10198 switch (Constraint) {
10201 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10202 if (C->getZExtValue() <= 31) {
10203 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10209 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10210 if (C->getZExtValue() <= 63) {
10211 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10217 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10218 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10219 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10225 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10226 if (C->getZExtValue() <= 255) {
10227 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10233 // 32-bit signed value
10234 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10235 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10236 C->getSExtValue())) {
10237 // Widen to 64 bits here to get it sign extended.
10238 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10241 // FIXME gcc accepts some relocatable values here too, but only in certain
10242 // memory models; it's complicated.
10247 // 32-bit unsigned value
10248 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10249 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10250 C->getZExtValue())) {
10251 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10255 // FIXME gcc accepts some relocatable values here too, but only in certain
10256 // memory models; it's complicated.
10260 // Literal immediates are always ok.
10261 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10262 // Widen to 64 bits here to get it sign extended.
10263 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10267 // In any sort of PIC mode addresses need to be computed at runtime by
10268 // adding in a register or some sort of table lookup. These can't
10269 // be used as immediates.
10270 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
10273 // If we are in non-pic codegen mode, we allow the address of a global (with
10274 // an optional displacement) to be used with 'i'.
10275 GlobalAddressSDNode *GA = 0;
10276 int64_t Offset = 0;
10278 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10280 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10281 Offset += GA->getOffset();
10283 } else if (Op.getOpcode() == ISD::ADD) {
10284 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10285 Offset += C->getZExtValue();
10286 Op = Op.getOperand(0);
10289 } else if (Op.getOpcode() == ISD::SUB) {
10290 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10291 Offset += -C->getZExtValue();
10292 Op = Op.getOperand(0);
10297 // Otherwise, this isn't something we can handle, reject it.
10301 const GlobalValue *GV = GA->getGlobal();
10302 // If we require an extra load to get this address, as in PIC mode, we
10303 // can't accept it.
10304 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10305 getTargetMachine())))
10308 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
10309 GA->getValueType(0), Offset);
10314 if (Result.getNode()) {
10315 Ops.push_back(Result);
10318 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10321 std::vector<unsigned> X86TargetLowering::
10322 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10324 if (Constraint.size() == 1) {
10325 // FIXME: not handling fp-stack yet!
10326 switch (Constraint[0]) { // GCC X86 Constraint Letters
10327 default: break; // Unknown constraint letter
10328 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10329 if (Subtarget->is64Bit()) {
10330 if (VT == MVT::i32)
10331 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10332 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10333 X86::R10D,X86::R11D,X86::R12D,
10334 X86::R13D,X86::R14D,X86::R15D,
10335 X86::EBP, X86::ESP, 0);
10336 else if (VT == MVT::i16)
10337 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10338 X86::SI, X86::DI, X86::R8W,X86::R9W,
10339 X86::R10W,X86::R11W,X86::R12W,
10340 X86::R13W,X86::R14W,X86::R15W,
10341 X86::BP, X86::SP, 0);
10342 else if (VT == MVT::i8)
10343 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10344 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10345 X86::R10B,X86::R11B,X86::R12B,
10346 X86::R13B,X86::R14B,X86::R15B,
10347 X86::BPL, X86::SPL, 0);
10349 else if (VT == MVT::i64)
10350 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10351 X86::RSI, X86::RDI, X86::R8, X86::R9,
10352 X86::R10, X86::R11, X86::R12,
10353 X86::R13, X86::R14, X86::R15,
10354 X86::RBP, X86::RSP, 0);
10358 // 32-bit fallthrough
10359 case 'Q': // Q_REGS
10360 if (VT == MVT::i32)
10361 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10362 else if (VT == MVT::i16)
10363 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10364 else if (VT == MVT::i8)
10365 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10366 else if (VT == MVT::i64)
10367 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10372 return std::vector<unsigned>();
10375 std::pair<unsigned, const TargetRegisterClass*>
10376 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10378 // First, see if this is a constraint that directly corresponds to an LLVM
10380 if (Constraint.size() == 1) {
10381 // GCC Constraint Letters
10382 switch (Constraint[0]) {
10384 case 'r': // GENERAL_REGS
10385 case 'l': // INDEX_REGS
10387 return std::make_pair(0U, X86::GR8RegisterClass);
10388 if (VT == MVT::i16)
10389 return std::make_pair(0U, X86::GR16RegisterClass);
10390 if (VT == MVT::i32 || !Subtarget->is64Bit())
10391 return std::make_pair(0U, X86::GR32RegisterClass);
10392 return std::make_pair(0U, X86::GR64RegisterClass);
10393 case 'R': // LEGACY_REGS
10395 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10396 if (VT == MVT::i16)
10397 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10398 if (VT == MVT::i32 || !Subtarget->is64Bit())
10399 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10400 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10401 case 'f': // FP Stack registers.
10402 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10403 // value to the correct fpstack register class.
10404 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10405 return std::make_pair(0U, X86::RFP32RegisterClass);
10406 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10407 return std::make_pair(0U, X86::RFP64RegisterClass);
10408 return std::make_pair(0U, X86::RFP80RegisterClass);
10409 case 'y': // MMX_REGS if MMX allowed.
10410 if (!Subtarget->hasMMX()) break;
10411 return std::make_pair(0U, X86::VR64RegisterClass);
10412 case 'Y': // SSE_REGS if SSE2 allowed
10413 if (!Subtarget->hasSSE2()) break;
10415 case 'x': // SSE_REGS if SSE1 allowed
10416 if (!Subtarget->hasSSE1()) break;
10418 switch (VT.getSimpleVT().SimpleTy) {
10420 // Scalar SSE types.
10423 return std::make_pair(0U, X86::FR32RegisterClass);
10426 return std::make_pair(0U, X86::FR64RegisterClass);
10434 return std::make_pair(0U, X86::VR128RegisterClass);
10440 // Use the default implementation in TargetLowering to convert the register
10441 // constraint into a member of a register class.
10442 std::pair<unsigned, const TargetRegisterClass*> Res;
10443 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10445 // Not found as a standard register?
10446 if (Res.second == 0) {
10447 // Map st(0) -> st(7) -> ST0
10448 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10449 tolower(Constraint[1]) == 's' &&
10450 tolower(Constraint[2]) == 't' &&
10451 Constraint[3] == '(' &&
10452 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10453 Constraint[5] == ')' &&
10454 Constraint[6] == '}') {
10456 Res.first = X86::ST0+Constraint[4]-'0';
10457 Res.second = X86::RFP80RegisterClass;
10461 // GCC allows "st(0)" to be called just plain "st".
10462 if (StringRef("{st}").equals_lower(Constraint)) {
10463 Res.first = X86::ST0;
10464 Res.second = X86::RFP80RegisterClass;
10469 if (StringRef("{flags}").equals_lower(Constraint)) {
10470 Res.first = X86::EFLAGS;
10471 Res.second = X86::CCRRegisterClass;
10475 // 'A' means EAX + EDX.
10476 if (Constraint == "A") {
10477 Res.first = X86::EAX;
10478 Res.second = X86::GR32_ADRegisterClass;
10484 // Otherwise, check to see if this is a register class of the wrong value
10485 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10486 // turn into {ax},{dx}.
10487 if (Res.second->hasType(VT))
10488 return Res; // Correct type already, nothing to do.
10490 // All of the single-register GCC register classes map their values onto
10491 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10492 // really want an 8-bit or 32-bit register, map to the appropriate register
10493 // class and return the appropriate register.
10494 if (Res.second == X86::GR16RegisterClass) {
10495 if (VT == MVT::i8) {
10496 unsigned DestReg = 0;
10497 switch (Res.first) {
10499 case X86::AX: DestReg = X86::AL; break;
10500 case X86::DX: DestReg = X86::DL; break;
10501 case X86::CX: DestReg = X86::CL; break;
10502 case X86::BX: DestReg = X86::BL; break;
10505 Res.first = DestReg;
10506 Res.second = X86::GR8RegisterClass;
10508 } else if (VT == MVT::i32) {
10509 unsigned DestReg = 0;
10510 switch (Res.first) {
10512 case X86::AX: DestReg = X86::EAX; break;
10513 case X86::DX: DestReg = X86::EDX; break;
10514 case X86::CX: DestReg = X86::ECX; break;
10515 case X86::BX: DestReg = X86::EBX; break;
10516 case X86::SI: DestReg = X86::ESI; break;
10517 case X86::DI: DestReg = X86::EDI; break;
10518 case X86::BP: DestReg = X86::EBP; break;
10519 case X86::SP: DestReg = X86::ESP; break;
10522 Res.first = DestReg;
10523 Res.second = X86::GR32RegisterClass;
10525 } else if (VT == MVT::i64) {
10526 unsigned DestReg = 0;
10527 switch (Res.first) {
10529 case X86::AX: DestReg = X86::RAX; break;
10530 case X86::DX: DestReg = X86::RDX; break;
10531 case X86::CX: DestReg = X86::RCX; break;
10532 case X86::BX: DestReg = X86::RBX; break;
10533 case X86::SI: DestReg = X86::RSI; break;
10534 case X86::DI: DestReg = X86::RDI; break;
10535 case X86::BP: DestReg = X86::RBP; break;
10536 case X86::SP: DestReg = X86::RSP; break;
10539 Res.first = DestReg;
10540 Res.second = X86::GR64RegisterClass;
10543 } else if (Res.second == X86::FR32RegisterClass ||
10544 Res.second == X86::FR64RegisterClass ||
10545 Res.second == X86::VR128RegisterClass) {
10546 // Handle references to XMM physical registers that got mapped into the
10547 // wrong class. This can happen with constraints like {xmm0} where the
10548 // target independent register mapper will just pick the first match it can
10549 // find, ignoring the required type.
10550 if (VT == MVT::f32)
10551 Res.second = X86::FR32RegisterClass;
10552 else if (VT == MVT::f64)
10553 Res.second = X86::FR64RegisterClass;
10554 else if (X86::VR128RegisterClass->hasType(VT))
10555 Res.second = X86::VR128RegisterClass;