1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86ShuffleDecode.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineJumpTableInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/PseudoSourceValue.h"
38 #include "llvm/MC/MCAsmInfo.h"
39 #include "llvm/MC/MCContext.h"
40 #include "llvm/MC/MCExpr.h"
41 #include "llvm/MC/MCSymbol.h"
42 #include "llvm/ADT/BitVector.h"
43 #include "llvm/ADT/SmallSet.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/ADT/StringExtras.h"
46 #include "llvm/ADT/VectorExtras.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/Dwarf.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
54 using namespace dwarf;
56 STATISTIC(NumTailCalls, "Number of tail calls");
59 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
61 // Forward declarations.
62 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
65 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
67 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
69 if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
70 if (is64Bit) return new X8664_MachoTargetObjectFile();
71 return new TargetLoweringObjectFileMachO();
72 } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
73 if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
74 return new X8632_ELFTargetObjectFile(TM);
75 } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
76 return new TargetLoweringObjectFileCOFF();
78 llvm_unreachable("unknown subtarget type");
81 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
82 : TargetLowering(TM, createTLOF(TM)) {
83 Subtarget = &TM.getSubtarget<X86Subtarget>();
84 X86ScalarSSEf64 = Subtarget->hasXMMInt();
85 X86ScalarSSEf32 = Subtarget->hasXMM();
86 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
88 RegInfo = TM.getRegisterInfo();
91 // Set up the TargetLowering object.
93 // X86 is weird, it always uses i8 for shift amounts and setcc results.
94 setShiftAmountType(MVT::i8);
95 setBooleanContents(ZeroOrOneBooleanContent);
96 setSchedulingPreference(Sched::RegPressure);
97 setStackPointerRegisterToSaveRestore(X86StackPtr);
99 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
100 // Setup Windows compiler runtime calls.
101 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
102 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
103 setLibcallName(RTLIB::FPTOUINT_F64_I64, "_ftol2");
104 setLibcallName(RTLIB::FPTOUINT_F32_I64, "_ftol2");
105 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
106 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
107 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::C);
108 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::C);
111 if (Subtarget->isTargetDarwin()) {
112 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
113 setUseUnderscoreSetJmp(false);
114 setUseUnderscoreLongJmp(false);
115 } else if (Subtarget->isTargetMingw()) {
116 // MS runtime is weird: it exports _setjmp, but longjmp!
117 setUseUnderscoreSetJmp(true);
118 setUseUnderscoreLongJmp(false);
120 setUseUnderscoreSetJmp(true);
121 setUseUnderscoreLongJmp(true);
124 // Set up the register classes.
125 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
126 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
127 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
128 if (Subtarget->is64Bit())
129 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
131 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
133 // We don't accept any truncstore of integer registers.
134 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
135 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
136 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
137 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
138 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
139 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
141 // SETOEQ and SETUNE require checking two conditions.
142 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
143 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
144 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
145 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
146 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
147 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
149 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
151 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
152 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
153 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
155 if (Subtarget->is64Bit()) {
156 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
157 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
158 } else if (!UseSoftFloat) {
159 // We have an algorithm for SSE2->double, and we turn this into a
160 // 64-bit FILD followed by conditional FADD for other targets.
161 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
162 // We have an algorithm for SSE2, and we turn this into a 64-bit
163 // FILD for other targets.
164 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
167 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
169 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
170 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
173 // SSE has no i16 to fp conversion, only i32
174 if (X86ScalarSSEf32) {
175 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
176 // f32 and f64 cases are Legal, f80 case is not
177 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
179 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
180 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
183 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
184 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
187 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
188 // are Legal, f80 is custom lowered.
189 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
190 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
192 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
194 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
195 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
197 if (X86ScalarSSEf32) {
198 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
199 // f32 and f64 cases are Legal, f80 case is not
200 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
202 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
203 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
206 // Handle FP_TO_UINT by promoting the destination to a larger signed
208 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
209 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
210 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
212 if (Subtarget->is64Bit()) {
213 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
214 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
215 } else if (!UseSoftFloat) {
216 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
217 // Expand FP_TO_UINT into a select.
218 // FIXME: We would like to use a Custom expander here eventually to do
219 // the optimal thing for SSE vs. the default expansion in the legalizer.
220 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
222 // With SSE3 we can use fisttpll to convert to a signed i64; without
223 // SSE, we're stuck with a fistpll.
224 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
227 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
228 if (!X86ScalarSSEf64) {
229 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
230 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
231 if (Subtarget->is64Bit()) {
232 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
233 // Without SSE, i64->f64 goes through memory.
234 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
238 // Scalar integer divide and remainder are lowered to use operations that
239 // produce two results, to match the available instructions. This exposes
240 // the two-result form to trivial CSE, which is able to combine x/y and x%y
241 // into a single instruction.
243 // Scalar integer multiply-high is also lowered to use two-result
244 // operations, to match the available instructions. However, plain multiply
245 // (low) operations are left as Legal, as there are single-result
246 // instructions for this in x86. Using the two-result multiply instructions
247 // when both high and low results are needed must be arranged by dagcombine.
248 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
249 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
250 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
251 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
252 setOperationAction(ISD::SREM , MVT::i8 , Expand);
253 setOperationAction(ISD::UREM , MVT::i8 , Expand);
254 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
255 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
256 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
257 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
258 setOperationAction(ISD::SREM , MVT::i16 , Expand);
259 setOperationAction(ISD::UREM , MVT::i16 , Expand);
260 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
261 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
262 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
263 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
264 setOperationAction(ISD::SREM , MVT::i32 , Expand);
265 setOperationAction(ISD::UREM , MVT::i32 , Expand);
266 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
267 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
268 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
269 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
270 setOperationAction(ISD::SREM , MVT::i64 , Expand);
271 setOperationAction(ISD::UREM , MVT::i64 , Expand);
273 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
274 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
275 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
276 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
277 if (Subtarget->is64Bit())
278 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
279 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
280 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
281 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
282 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
283 setOperationAction(ISD::FREM , MVT::f32 , Expand);
284 setOperationAction(ISD::FREM , MVT::f64 , Expand);
285 setOperationAction(ISD::FREM , MVT::f80 , Expand);
286 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
288 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
289 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
290 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
291 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
292 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
293 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
294 if (Subtarget->is64Bit()) {
295 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
296 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
299 if (Subtarget->hasPOPCNT()) {
300 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
302 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
303 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
304 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
305 if (Subtarget->is64Bit())
306 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
309 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
310 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
312 // These should be promoted to a larger select which is supported.
313 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
314 // X86 wants to expand cmov itself.
315 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
316 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
317 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
318 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
319 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
320 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
321 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
322 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
323 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
324 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
325 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
326 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
327 if (Subtarget->is64Bit()) {
328 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
329 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
331 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
334 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
335 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
336 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
337 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
338 if (Subtarget->is64Bit())
339 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
340 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
341 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
342 if (Subtarget->is64Bit()) {
343 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
344 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
345 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
346 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
347 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
349 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
350 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
351 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
352 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
353 if (Subtarget->is64Bit()) {
354 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
355 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
356 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
359 if (Subtarget->hasXMM())
360 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
362 // We may not have a libcall for MEMBARRIER so we should lower this.
363 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
365 // On X86 and X86-64, atomic operations are lowered to locked instructions.
366 // Locked instructions, in turn, have implicit fence semantics (all memory
367 // operations are flushed before issuing the locked instruction, and they
368 // are not buffered), so we can fold away the common pattern of
369 // fence-atomic-fence.
370 setShouldFoldAtomicFences(true);
372 // Expand certain atomics
373 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
374 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
375 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
376 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
378 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
379 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
380 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
381 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
383 if (!Subtarget->is64Bit()) {
384 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
385 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
386 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
387 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
388 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
389 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
390 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
393 // FIXME - use subtarget debug flags
394 if (!Subtarget->isTargetDarwin() &&
395 !Subtarget->isTargetELF() &&
396 !Subtarget->isTargetCygMing()) {
397 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
400 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
401 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
402 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
403 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
404 if (Subtarget->is64Bit()) {
405 setExceptionPointerRegister(X86::RAX);
406 setExceptionSelectorRegister(X86::RDX);
408 setExceptionPointerRegister(X86::EAX);
409 setExceptionSelectorRegister(X86::EDX);
411 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
412 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
414 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
416 setOperationAction(ISD::TRAP, MVT::Other, Legal);
418 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
419 setOperationAction(ISD::VASTART , MVT::Other, Custom);
420 setOperationAction(ISD::VAEND , MVT::Other, Expand);
421 if (Subtarget->is64Bit()) {
422 setOperationAction(ISD::VAARG , MVT::Other, Custom);
423 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
425 setOperationAction(ISD::VAARG , MVT::Other, Expand);
426 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
429 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
430 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
431 if (Subtarget->is64Bit())
432 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
433 if (Subtarget->isTargetCygMing() || Subtarget->isTargetWindows())
434 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
436 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
438 if (!UseSoftFloat && X86ScalarSSEf64) {
439 // f32 and f64 use SSE.
440 // Set up the FP register classes.
441 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
442 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
444 // Use ANDPD to simulate FABS.
445 setOperationAction(ISD::FABS , MVT::f64, Custom);
446 setOperationAction(ISD::FABS , MVT::f32, Custom);
448 // Use XORP to simulate FNEG.
449 setOperationAction(ISD::FNEG , MVT::f64, Custom);
450 setOperationAction(ISD::FNEG , MVT::f32, Custom);
452 // Use ANDPD and ORPD to simulate FCOPYSIGN.
453 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
454 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
456 // We don't support sin/cos/fmod
457 setOperationAction(ISD::FSIN , MVT::f64, Expand);
458 setOperationAction(ISD::FCOS , MVT::f64, Expand);
459 setOperationAction(ISD::FSIN , MVT::f32, Expand);
460 setOperationAction(ISD::FCOS , MVT::f32, Expand);
462 // Expand FP immediates into loads from the stack, except for the special
464 addLegalFPImmediate(APFloat(+0.0)); // xorpd
465 addLegalFPImmediate(APFloat(+0.0f)); // xorps
466 } else if (!UseSoftFloat && X86ScalarSSEf32) {
467 // Use SSE for f32, x87 for f64.
468 // Set up the FP register classes.
469 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
470 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
472 // Use ANDPS to simulate FABS.
473 setOperationAction(ISD::FABS , MVT::f32, Custom);
475 // Use XORP to simulate FNEG.
476 setOperationAction(ISD::FNEG , MVT::f32, Custom);
478 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
480 // Use ANDPS and ORPS to simulate FCOPYSIGN.
481 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
482 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
484 // We don't support sin/cos/fmod
485 setOperationAction(ISD::FSIN , MVT::f32, Expand);
486 setOperationAction(ISD::FCOS , MVT::f32, Expand);
488 // Special cases we handle for FP constants.
489 addLegalFPImmediate(APFloat(+0.0f)); // xorps
490 addLegalFPImmediate(APFloat(+0.0)); // FLD0
491 addLegalFPImmediate(APFloat(+1.0)); // FLD1
492 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
493 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
496 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
497 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
499 } else if (!UseSoftFloat) {
500 // f32 and f64 in x87.
501 // Set up the FP register classes.
502 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
503 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
505 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
506 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
507 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
508 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
511 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
512 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
514 addLegalFPImmediate(APFloat(+0.0)); // FLD0
515 addLegalFPImmediate(APFloat(+1.0)); // FLD1
516 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
517 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
518 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
519 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
520 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
521 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
524 // Long double always uses X87.
526 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
527 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
528 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
530 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
531 addLegalFPImmediate(TmpFlt); // FLD0
533 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
536 APFloat TmpFlt2(+1.0);
537 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
539 addLegalFPImmediate(TmpFlt2); // FLD1
540 TmpFlt2.changeSign();
541 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
545 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
546 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
550 // Always use a library call for pow.
551 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
552 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
553 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
555 setOperationAction(ISD::FLOG, MVT::f80, Expand);
556 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
557 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
558 setOperationAction(ISD::FEXP, MVT::f80, Expand);
559 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
561 // First set operation action for all vector types to either promote
562 // (for widening) or expand (for scalarization). Then we will selectively
563 // turn on ones that can be effectively codegen'd.
564 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
565 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
566 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
581 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
582 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
607 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
608 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
609 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
610 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
611 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
612 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
613 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
614 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
615 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
616 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
617 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
618 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
619 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
620 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
621 setTruncStoreAction((MVT::SimpleValueType)VT,
622 (MVT::SimpleValueType)InnerVT, Expand);
623 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
624 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
625 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
628 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
629 // with -msoft-float, disable use of MMX as well.
630 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
631 addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass);
632 // No operations on x86mmx supported, everything uses intrinsics.
635 // MMX-sized vectors (other than x86mmx) are expected to be expanded
636 // into smaller operations.
637 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
638 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
639 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
640 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
641 setOperationAction(ISD::AND, MVT::v8i8, Expand);
642 setOperationAction(ISD::AND, MVT::v4i16, Expand);
643 setOperationAction(ISD::AND, MVT::v2i32, Expand);
644 setOperationAction(ISD::AND, MVT::v1i64, Expand);
645 setOperationAction(ISD::OR, MVT::v8i8, Expand);
646 setOperationAction(ISD::OR, MVT::v4i16, Expand);
647 setOperationAction(ISD::OR, MVT::v2i32, Expand);
648 setOperationAction(ISD::OR, MVT::v1i64, Expand);
649 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
650 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
651 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
652 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
653 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
654 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
655 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
656 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
657 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
658 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
659 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
660 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
661 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
662 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
663 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
664 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
665 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
667 if (!UseSoftFloat && Subtarget->hasXMM()) {
668 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
670 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
671 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
672 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
673 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
674 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
675 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
676 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
677 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
678 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
679 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
680 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
681 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
684 if (!UseSoftFloat && Subtarget->hasXMMInt()) {
685 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
687 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
688 // registers cannot be used even for integer operations.
689 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
690 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
691 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
692 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
694 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
695 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
696 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
697 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
698 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
699 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
700 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
701 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
702 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
703 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
704 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
705 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
706 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
707 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
708 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
709 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
711 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
712 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
713 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
714 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
716 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
717 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
718 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
719 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
720 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
722 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
723 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
724 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
725 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
726 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
728 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
729 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
730 EVT VT = (MVT::SimpleValueType)i;
731 // Do not attempt to custom lower non-power-of-2 vectors
732 if (!isPowerOf2_32(VT.getVectorNumElements()))
734 // Do not attempt to custom lower non-128-bit vectors
735 if (!VT.is128BitVector())
737 setOperationAction(ISD::BUILD_VECTOR,
738 VT.getSimpleVT().SimpleTy, Custom);
739 setOperationAction(ISD::VECTOR_SHUFFLE,
740 VT.getSimpleVT().SimpleTy, Custom);
741 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
742 VT.getSimpleVT().SimpleTy, Custom);
745 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
746 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
747 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
748 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
749 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
750 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
752 if (Subtarget->is64Bit()) {
753 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
754 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
757 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
758 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
759 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
762 // Do not attempt to promote non-128-bit vectors
763 if (!VT.is128BitVector())
766 setOperationAction(ISD::AND, SVT, Promote);
767 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
768 setOperationAction(ISD::OR, SVT, Promote);
769 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
770 setOperationAction(ISD::XOR, SVT, Promote);
771 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
772 setOperationAction(ISD::LOAD, SVT, Promote);
773 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
774 setOperationAction(ISD::SELECT, SVT, Promote);
775 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
778 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
780 // Custom lower v2i64 and v2f64 selects.
781 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
782 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
783 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
784 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
786 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
787 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
790 if (Subtarget->hasSSE41()) {
791 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
792 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
793 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
794 setOperationAction(ISD::FRINT, MVT::f32, Legal);
795 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
796 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
797 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
798 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
799 setOperationAction(ISD::FRINT, MVT::f64, Legal);
800 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
802 // FIXME: Do we need to handle scalar-to-vector here?
803 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
805 // Can turn SHL into an integer multiply.
806 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
807 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
809 // i8 and i16 vectors are custom , because the source register and source
810 // source memory operand types are not the same width. f32 vectors are
811 // custom since the immediate controlling the insert encodes additional
813 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
814 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
815 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
816 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
818 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
819 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
820 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
821 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
823 if (Subtarget->is64Bit()) {
824 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
825 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
829 if (Subtarget->hasSSE42()) {
830 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
833 if (!UseSoftFloat && Subtarget->hasAVX()) {
834 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
835 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
836 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
837 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
838 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
840 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
841 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
842 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
843 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
844 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
845 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
846 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
847 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
848 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
849 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
850 setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
851 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
852 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
853 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
854 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
856 // Operations to consider commented out -v16i16 v32i8
857 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
858 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
859 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
860 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
861 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
862 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
863 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
864 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
865 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
866 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
867 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
868 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
869 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
870 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
872 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
873 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
874 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
875 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
877 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
878 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
879 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
880 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
881 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
883 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
884 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
885 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
886 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
887 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
888 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
891 // Not sure we want to do this since there are no 256-bit integer
894 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
895 // This includes 256-bit vectors
896 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
897 EVT VT = (MVT::SimpleValueType)i;
899 // Do not attempt to custom lower non-power-of-2 vectors
900 if (!isPowerOf2_32(VT.getVectorNumElements()))
903 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
904 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
905 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
908 if (Subtarget->is64Bit()) {
909 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
910 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
915 // Not sure we want to do this since there are no 256-bit integer
918 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
919 // Including 256-bit vectors
920 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
921 EVT VT = (MVT::SimpleValueType)i;
923 if (!VT.is256BitVector()) {
926 setOperationAction(ISD::AND, VT, Promote);
927 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
928 setOperationAction(ISD::OR, VT, Promote);
929 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
930 setOperationAction(ISD::XOR, VT, Promote);
931 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
932 setOperationAction(ISD::LOAD, VT, Promote);
933 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
934 setOperationAction(ISD::SELECT, VT, Promote);
935 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
938 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
942 // We want to custom lower some of our intrinsics.
943 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
945 // Add/Sub/Mul with overflow operations are custom lowered.
946 setOperationAction(ISD::SADDO, MVT::i32, Custom);
947 setOperationAction(ISD::UADDO, MVT::i32, Custom);
948 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
949 setOperationAction(ISD::USUBO, MVT::i32, Custom);
950 setOperationAction(ISD::SMULO, MVT::i32, Custom);
951 setOperationAction(ISD::UMULO, MVT::i32, Custom);
953 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
954 // handle type legalization for these operations here.
956 // FIXME: We really should do custom legalization for addition and
957 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
958 // than generic legalization for 64-bit multiplication-with-overflow, though.
959 if (Subtarget->is64Bit()) {
960 setOperationAction(ISD::SADDO, MVT::i64, Custom);
961 setOperationAction(ISD::UADDO, MVT::i64, Custom);
962 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
963 setOperationAction(ISD::USUBO, MVT::i64, Custom);
964 setOperationAction(ISD::SMULO, MVT::i64, Custom);
965 setOperationAction(ISD::UMULO, MVT::i64, Custom);
968 if (!Subtarget->is64Bit()) {
969 // These libcalls are not available in 32-bit.
970 setLibcallName(RTLIB::SHL_I128, 0);
971 setLibcallName(RTLIB::SRL_I128, 0);
972 setLibcallName(RTLIB::SRA_I128, 0);
975 // We have target-specific dag combine patterns for the following nodes:
976 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
977 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
978 setTargetDAGCombine(ISD::BUILD_VECTOR);
979 setTargetDAGCombine(ISD::SELECT);
980 setTargetDAGCombine(ISD::SHL);
981 setTargetDAGCombine(ISD::SRA);
982 setTargetDAGCombine(ISD::SRL);
983 setTargetDAGCombine(ISD::OR);
984 setTargetDAGCombine(ISD::AND);
985 setTargetDAGCombine(ISD::STORE);
986 setTargetDAGCombine(ISD::ZERO_EXTEND);
987 if (Subtarget->is64Bit())
988 setTargetDAGCombine(ISD::MUL);
990 computeRegisterProperties();
992 // FIXME: These should be based on subtarget info. Plus, the values should
993 // be smaller when we are in optimizing for size mode.
994 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
995 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
996 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
997 setPrefLoopAlignment(16);
998 benefitFromCodePlacementOpt = true;
1002 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1007 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1008 /// the desired ByVal argument alignment.
1009 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1012 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1013 if (VTy->getBitWidth() == 128)
1015 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1016 unsigned EltAlign = 0;
1017 getMaxByValAlign(ATy->getElementType(), EltAlign);
1018 if (EltAlign > MaxAlign)
1019 MaxAlign = EltAlign;
1020 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1021 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1022 unsigned EltAlign = 0;
1023 getMaxByValAlign(STy->getElementType(i), EltAlign);
1024 if (EltAlign > MaxAlign)
1025 MaxAlign = EltAlign;
1033 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1034 /// function arguments in the caller parameter area. For X86, aggregates
1035 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1036 /// are at 4-byte boundaries.
1037 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1038 if (Subtarget->is64Bit()) {
1039 // Max of 8 and alignment of type.
1040 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1047 if (Subtarget->hasXMM())
1048 getMaxByValAlign(Ty, Align);
1052 /// getOptimalMemOpType - Returns the target specific optimal type for load
1053 /// and store operations as a result of memset, memcpy, and memmove
1054 /// lowering. If DstAlign is zero that means it's safe to destination
1055 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1056 /// means there isn't a need to check it against alignment requirement,
1057 /// probably because the source does not need to be loaded. If
1058 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1059 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1060 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1061 /// constant so it does not need to be loaded.
1062 /// It returns EVT::Other if the type should be determined using generic
1063 /// target-independent logic.
1065 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1066 unsigned DstAlign, unsigned SrcAlign,
1067 bool NonScalarIntSafe,
1069 MachineFunction &MF) const {
1070 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1071 // linux. This is because the stack realignment code can't handle certain
1072 // cases like PR2962. This should be removed when PR2962 is fixed.
1073 const Function *F = MF.getFunction();
1074 if (NonScalarIntSafe &&
1075 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1077 (Subtarget->isUnalignedMemAccessFast() ||
1078 ((DstAlign == 0 || DstAlign >= 16) &&
1079 (SrcAlign == 0 || SrcAlign >= 16))) &&
1080 Subtarget->getStackAlignment() >= 16) {
1081 if (Subtarget->hasSSE2())
1083 if (Subtarget->hasSSE1())
1085 } else if (!MemcpyStrSrc && Size >= 8 &&
1086 !Subtarget->is64Bit() &&
1087 Subtarget->getStackAlignment() >= 8 &&
1088 Subtarget->hasXMMInt()) {
1089 // Do not use f64 to lower memcpy if source is string constant. It's
1090 // better to use i32 to avoid the loads.
1094 if (Subtarget->is64Bit() && Size >= 8)
1099 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1100 /// current function. The returned value is a member of the
1101 /// MachineJumpTableInfo::JTEntryKind enum.
1102 unsigned X86TargetLowering::getJumpTableEncoding() const {
1103 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1105 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1106 Subtarget->isPICStyleGOT())
1107 return MachineJumpTableInfo::EK_Custom32;
1109 // Otherwise, use the normal jump table encoding heuristics.
1110 return TargetLowering::getJumpTableEncoding();
1114 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1115 const MachineBasicBlock *MBB,
1116 unsigned uid,MCContext &Ctx) const{
1117 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1118 Subtarget->isPICStyleGOT());
1119 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1121 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1122 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1125 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1127 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1128 SelectionDAG &DAG) const {
1129 if (!Subtarget->is64Bit())
1130 // This doesn't have DebugLoc associated with it, but is not really the
1131 // same as a Register.
1132 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1136 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1137 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1139 const MCExpr *X86TargetLowering::
1140 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1141 MCContext &Ctx) const {
1142 // X86-64 uses RIP relative addressing based on the jump table label.
1143 if (Subtarget->isPICStyleRIPRel())
1144 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1146 // Otherwise, the reference is relative to the PIC base.
1147 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1150 /// getFunctionAlignment - Return the Log2 alignment of this function.
1151 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1152 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1155 std::pair<const TargetRegisterClass*, uint8_t>
1156 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1157 const TargetRegisterClass *RRC = 0;
1159 switch (VT.getSimpleVT().SimpleTy) {
1161 return TargetLowering::findRepresentativeClass(VT);
1162 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1163 RRC = (Subtarget->is64Bit()
1164 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1167 RRC = X86::VR64RegisterClass;
1169 case MVT::f32: case MVT::f64:
1170 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1171 case MVT::v4f32: case MVT::v2f64:
1172 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1174 RRC = X86::VR128RegisterClass;
1177 return std::make_pair(RRC, Cost);
1181 X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
1182 MachineFunction &MF) const {
1183 const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo();
1185 unsigned FPDiff = TFI->hasFP(MF) ? 1 : 0;
1186 switch (RC->getID()) {
1189 case X86::GR32RegClassID:
1191 case X86::GR64RegClassID:
1193 case X86::VR128RegClassID:
1194 return Subtarget->is64Bit() ? 10 : 4;
1195 case X86::VR64RegClassID:
1200 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1201 unsigned &Offset) const {
1202 if (!Subtarget->isTargetLinux())
1205 if (Subtarget->is64Bit()) {
1206 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1208 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1221 //===----------------------------------------------------------------------===//
1222 // Return Value Calling Convention Implementation
1223 //===----------------------------------------------------------------------===//
1225 #include "X86GenCallingConv.inc"
1228 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1229 const SmallVectorImpl<ISD::OutputArg> &Outs,
1230 LLVMContext &Context) const {
1231 SmallVector<CCValAssign, 16> RVLocs;
1232 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1234 return CCInfo.CheckReturn(Outs, RetCC_X86);
1238 X86TargetLowering::LowerReturn(SDValue Chain,
1239 CallingConv::ID CallConv, bool isVarArg,
1240 const SmallVectorImpl<ISD::OutputArg> &Outs,
1241 const SmallVectorImpl<SDValue> &OutVals,
1242 DebugLoc dl, SelectionDAG &DAG) const {
1243 MachineFunction &MF = DAG.getMachineFunction();
1244 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1246 SmallVector<CCValAssign, 16> RVLocs;
1247 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1248 RVLocs, *DAG.getContext());
1249 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1251 // Add the regs to the liveout set for the function.
1252 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1253 for (unsigned i = 0; i != RVLocs.size(); ++i)
1254 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1255 MRI.addLiveOut(RVLocs[i].getLocReg());
1259 SmallVector<SDValue, 6> RetOps;
1260 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1261 // Operand #1 = Bytes To Pop
1262 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1265 // Copy the result values into the output registers.
1266 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1267 CCValAssign &VA = RVLocs[i];
1268 assert(VA.isRegLoc() && "Can only return in registers!");
1269 SDValue ValToCopy = OutVals[i];
1270 EVT ValVT = ValToCopy.getValueType();
1272 // If this is x86-64, and we disabled SSE, we can't return FP values,
1273 // or SSE or MMX vectors.
1274 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1275 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1276 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1277 report_fatal_error("SSE register return with SSE disabled");
1279 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1280 // llvm-gcc has never done it right and no one has noticed, so this
1281 // should be OK for now.
1282 if (ValVT == MVT::f64 &&
1283 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1284 report_fatal_error("SSE2 register return with SSE2 disabled");
1286 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1287 // the RET instruction and handled by the FP Stackifier.
1288 if (VA.getLocReg() == X86::ST0 ||
1289 VA.getLocReg() == X86::ST1) {
1290 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1291 // change the value to the FP stack register class.
1292 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1293 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1294 RetOps.push_back(ValToCopy);
1295 // Don't emit a copytoreg.
1299 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1300 // which is returned in RAX / RDX.
1301 if (Subtarget->is64Bit()) {
1302 if (ValVT == MVT::x86mmx) {
1303 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1304 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1305 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1307 // If we don't have SSE2 available, convert to v4f32 so the generated
1308 // register is legal.
1309 if (!Subtarget->hasSSE2())
1310 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1315 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1316 Flag = Chain.getValue(1);
1319 // The x86-64 ABI for returning structs by value requires that we copy
1320 // the sret argument into %rax for the return. We saved the argument into
1321 // a virtual register in the entry block, so now we copy the value out
1323 if (Subtarget->is64Bit() &&
1324 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1325 MachineFunction &MF = DAG.getMachineFunction();
1326 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1327 unsigned Reg = FuncInfo->getSRetReturnReg();
1329 "SRetReturnReg should have been set in LowerFormalArguments().");
1330 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1332 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1333 Flag = Chain.getValue(1);
1335 // RAX now acts like a return value.
1336 MRI.addLiveOut(X86::RAX);
1339 RetOps[0] = Chain; // Update chain.
1341 // Add the flag if we have it.
1343 RetOps.push_back(Flag);
1345 return DAG.getNode(X86ISD::RET_FLAG, dl,
1346 MVT::Other, &RetOps[0], RetOps.size());
1349 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1350 if (N->getNumValues() != 1)
1352 if (!N->hasNUsesOfValue(1, 0))
1355 SDNode *Copy = *N->use_begin();
1356 if (Copy->getOpcode() != ISD::CopyToReg &&
1357 Copy->getOpcode() != ISD::FP_EXTEND)
1360 bool HasRet = false;
1361 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1363 if (UI->getOpcode() != X86ISD::RET_FLAG)
1371 /// LowerCallResult - Lower the result values of a call into the
1372 /// appropriate copies out of appropriate physical registers.
1375 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1376 CallingConv::ID CallConv, bool isVarArg,
1377 const SmallVectorImpl<ISD::InputArg> &Ins,
1378 DebugLoc dl, SelectionDAG &DAG,
1379 SmallVectorImpl<SDValue> &InVals) const {
1381 // Assign locations to each value returned by this call.
1382 SmallVector<CCValAssign, 16> RVLocs;
1383 bool Is64Bit = Subtarget->is64Bit();
1384 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1385 RVLocs, *DAG.getContext());
1386 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1388 // Copy all of the result registers out of their specified physreg.
1389 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1390 CCValAssign &VA = RVLocs[i];
1391 EVT CopyVT = VA.getValVT();
1393 // If this is x86-64, and we disabled SSE, we can't return FP values
1394 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1395 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1396 report_fatal_error("SSE register return with SSE disabled");
1401 // If this is a call to a function that returns an fp value on the floating
1402 // point stack, we must guarantee the the value is popped from the stack, so
1403 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1404 // if the return value is not used. We use the FpGET_ST0 instructions
1406 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1407 // If we prefer to use the value in xmm registers, copy it out as f80 and
1408 // use a truncate to move it from fp stack reg to xmm reg.
1409 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1410 bool isST0 = VA.getLocReg() == X86::ST0;
1412 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1413 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1414 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1415 SDValue Ops[] = { Chain, InFlag };
1416 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
1418 Val = Chain.getValue(0);
1420 // Round the f80 to the right size, which also moves it to the appropriate
1422 if (CopyVT != VA.getValVT())
1423 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1424 // This truncation won't change the value.
1425 DAG.getIntPtrConstant(1));
1426 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1427 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1428 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1429 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1430 MVT::v2i64, InFlag).getValue(1);
1431 Val = Chain.getValue(0);
1432 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1433 Val, DAG.getConstant(0, MVT::i64));
1435 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1436 MVT::i64, InFlag).getValue(1);
1437 Val = Chain.getValue(0);
1439 Val = DAG.getNode(ISD::BITCAST, dl, CopyVT, Val);
1441 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1442 CopyVT, InFlag).getValue(1);
1443 Val = Chain.getValue(0);
1445 InFlag = Chain.getValue(2);
1446 InVals.push_back(Val);
1453 //===----------------------------------------------------------------------===//
1454 // C & StdCall & Fast Calling Convention implementation
1455 //===----------------------------------------------------------------------===//
1456 // StdCall calling convention seems to be standard for many Windows' API
1457 // routines and around. It differs from C calling convention just a little:
1458 // callee should clean up the stack, not caller. Symbols should be also
1459 // decorated in some fancy way :) It doesn't support any vector arguments.
1460 // For info on fast calling convention see Fast Calling Convention (tail call)
1461 // implementation LowerX86_32FastCCCallTo.
1463 /// CallIsStructReturn - Determines whether a call uses struct return
1465 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1469 return Outs[0].Flags.isSRet();
1472 /// ArgsAreStructReturn - Determines whether a function uses struct
1473 /// return semantics.
1475 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1479 return Ins[0].Flags.isSRet();
1482 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1483 /// by "Src" to address "Dst" with size and alignment information specified by
1484 /// the specific parameter attribute. The copy will be passed as a byval
1485 /// function parameter.
1487 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1488 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1490 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1492 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1493 /*isVolatile*/false, /*AlwaysInline=*/true,
1494 MachinePointerInfo(), MachinePointerInfo());
1497 /// IsTailCallConvention - Return true if the calling convention is one that
1498 /// supports tail call optimization.
1499 static bool IsTailCallConvention(CallingConv::ID CC) {
1500 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1503 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1504 /// a tailcall target by changing its ABI.
1505 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1506 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1510 X86TargetLowering::LowerMemArgument(SDValue Chain,
1511 CallingConv::ID CallConv,
1512 const SmallVectorImpl<ISD::InputArg> &Ins,
1513 DebugLoc dl, SelectionDAG &DAG,
1514 const CCValAssign &VA,
1515 MachineFrameInfo *MFI,
1517 // Create the nodes corresponding to a load from this parameter slot.
1518 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1519 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1520 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1523 // If value is passed by pointer we have address passed instead of the value
1525 if (VA.getLocInfo() == CCValAssign::Indirect)
1526 ValVT = VA.getLocVT();
1528 ValVT = VA.getValVT();
1530 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1531 // changed with more analysis.
1532 // In case of tail call optimization mark all arguments mutable. Since they
1533 // could be overwritten by lowering of arguments in case of a tail call.
1534 if (Flags.isByVal()) {
1535 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1536 VA.getLocMemOffset(), isImmutable);
1537 return DAG.getFrameIndex(FI, getPointerTy());
1539 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1540 VA.getLocMemOffset(), isImmutable);
1541 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1542 return DAG.getLoad(ValVT, dl, Chain, FIN,
1543 MachinePointerInfo::getFixedStack(FI),
1549 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1550 CallingConv::ID CallConv,
1552 const SmallVectorImpl<ISD::InputArg> &Ins,
1555 SmallVectorImpl<SDValue> &InVals)
1557 MachineFunction &MF = DAG.getMachineFunction();
1558 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1560 const Function* Fn = MF.getFunction();
1561 if (Fn->hasExternalLinkage() &&
1562 Subtarget->isTargetCygMing() &&
1563 Fn->getName() == "main")
1564 FuncInfo->setForceFramePointer(true);
1566 MachineFrameInfo *MFI = MF.getFrameInfo();
1567 bool Is64Bit = Subtarget->is64Bit();
1568 bool IsWin64 = Subtarget->isTargetWin64();
1570 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1571 "Var args not supported with calling convention fastcc or ghc");
1573 // Assign locations to all of the incoming arguments.
1574 SmallVector<CCValAssign, 16> ArgLocs;
1575 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1576 ArgLocs, *DAG.getContext());
1577 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1579 unsigned LastVal = ~0U;
1581 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1582 CCValAssign &VA = ArgLocs[i];
1583 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1585 assert(VA.getValNo() != LastVal &&
1586 "Don't support value assigned to multiple locs yet");
1587 LastVal = VA.getValNo();
1589 if (VA.isRegLoc()) {
1590 EVT RegVT = VA.getLocVT();
1591 TargetRegisterClass *RC = NULL;
1592 if (RegVT == MVT::i32)
1593 RC = X86::GR32RegisterClass;
1594 else if (Is64Bit && RegVT == MVT::i64)
1595 RC = X86::GR64RegisterClass;
1596 else if (RegVT == MVT::f32)
1597 RC = X86::FR32RegisterClass;
1598 else if (RegVT == MVT::f64)
1599 RC = X86::FR64RegisterClass;
1600 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1601 RC = X86::VR256RegisterClass;
1602 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1603 RC = X86::VR128RegisterClass;
1604 else if (RegVT == MVT::x86mmx)
1605 RC = X86::VR64RegisterClass;
1607 llvm_unreachable("Unknown argument type!");
1609 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1610 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1612 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1613 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1615 if (VA.getLocInfo() == CCValAssign::SExt)
1616 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1617 DAG.getValueType(VA.getValVT()));
1618 else if (VA.getLocInfo() == CCValAssign::ZExt)
1619 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1620 DAG.getValueType(VA.getValVT()));
1621 else if (VA.getLocInfo() == CCValAssign::BCvt)
1622 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1624 if (VA.isExtInLoc()) {
1625 // Handle MMX values passed in XMM regs.
1626 if (RegVT.isVector()) {
1627 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1630 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1633 assert(VA.isMemLoc());
1634 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1637 // If value is passed via pointer - do a load.
1638 if (VA.getLocInfo() == CCValAssign::Indirect)
1639 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1640 MachinePointerInfo(), false, false, 0);
1642 InVals.push_back(ArgValue);
1645 // The x86-64 ABI for returning structs by value requires that we copy
1646 // the sret argument into %rax for the return. Save the argument into
1647 // a virtual register so that we can access it from the return points.
1648 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1649 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1650 unsigned Reg = FuncInfo->getSRetReturnReg();
1652 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1653 FuncInfo->setSRetReturnReg(Reg);
1655 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1656 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1659 unsigned StackSize = CCInfo.getNextStackOffset();
1660 // Align stack specially for tail calls.
1661 if (FuncIsMadeTailCallSafe(CallConv))
1662 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1664 // If the function takes variable number of arguments, make a frame index for
1665 // the start of the first vararg value... for expansion of llvm.va_start.
1667 if (!IsWin64 && (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1668 CallConv != CallingConv::X86_ThisCall))) {
1669 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1672 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1674 // FIXME: We should really autogenerate these arrays
1675 static const unsigned GPR64ArgRegsWin64[] = {
1676 X86::RCX, X86::RDX, X86::R8, X86::R9
1678 static const unsigned GPR64ArgRegs64Bit[] = {
1679 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1681 static const unsigned XMMArgRegs64Bit[] = {
1682 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1683 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1685 const unsigned *GPR64ArgRegs;
1686 unsigned NumXMMRegs = 0;
1689 // The XMM registers which might contain var arg parameters are shadowed
1690 // in their paired GPR. So we only need to save the GPR to their home
1692 TotalNumIntRegs = 4;
1693 GPR64ArgRegs = GPR64ArgRegsWin64;
1695 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1696 GPR64ArgRegs = GPR64ArgRegs64Bit;
1698 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1700 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1703 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1704 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1705 "SSE register cannot be used when SSE is disabled!");
1706 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1707 "SSE register cannot be used when SSE is disabled!");
1708 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1709 // Kernel mode asks for SSE to be disabled, so don't push them
1711 TotalNumXMMRegs = 0;
1714 const TargetFrameInfo &TFI = *getTargetMachine().getFrameInfo();
1715 // Get to the caller-allocated home save location. Add 8 to account
1716 // for the return address.
1717 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1718 FuncInfo->setRegSaveFrameIndex(
1719 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1720 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1722 // For X86-64, if there are vararg parameters that are passed via
1723 // registers, then we must store them to their spots on the stack so they
1724 // may be loaded by deferencing the result of va_next.
1725 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1726 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1727 FuncInfo->setRegSaveFrameIndex(
1728 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1732 // Store the integer parameter registers.
1733 SmallVector<SDValue, 8> MemOps;
1734 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1736 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1737 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1738 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1739 DAG.getIntPtrConstant(Offset));
1740 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1741 X86::GR64RegisterClass);
1742 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1744 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1745 MachinePointerInfo::getFixedStack(
1746 FuncInfo->getRegSaveFrameIndex(), Offset),
1748 MemOps.push_back(Store);
1752 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1753 // Now store the XMM (fp + vector) parameter registers.
1754 SmallVector<SDValue, 11> SaveXMMOps;
1755 SaveXMMOps.push_back(Chain);
1757 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1758 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1759 SaveXMMOps.push_back(ALVal);
1761 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1762 FuncInfo->getRegSaveFrameIndex()));
1763 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1764 FuncInfo->getVarArgsFPOffset()));
1766 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1767 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1768 X86::VR128RegisterClass);
1769 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1770 SaveXMMOps.push_back(Val);
1772 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1774 &SaveXMMOps[0], SaveXMMOps.size()));
1777 if (!MemOps.empty())
1778 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1779 &MemOps[0], MemOps.size());
1783 // Some CCs need callee pop.
1784 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1785 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1787 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1788 // If this is an sret function, the return should pop the hidden pointer.
1789 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1790 FuncInfo->setBytesToPopOnReturn(4);
1794 // RegSaveFrameIndex is X86-64 only.
1795 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1796 if (CallConv == CallingConv::X86_FastCall ||
1797 CallConv == CallingConv::X86_ThisCall)
1798 // fastcc functions can't have varargs.
1799 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1806 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1807 SDValue StackPtr, SDValue Arg,
1808 DebugLoc dl, SelectionDAG &DAG,
1809 const CCValAssign &VA,
1810 ISD::ArgFlagsTy Flags) const {
1811 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1812 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1813 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1814 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1815 if (Flags.isByVal())
1816 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1818 return DAG.getStore(Chain, dl, Arg, PtrOff,
1819 MachinePointerInfo::getStack(LocMemOffset),
1823 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1824 /// optimization is performed and it is required.
1826 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1827 SDValue &OutRetAddr, SDValue Chain,
1828 bool IsTailCall, bool Is64Bit,
1829 int FPDiff, DebugLoc dl) const {
1830 // Adjust the Return address stack slot.
1831 EVT VT = getPointerTy();
1832 OutRetAddr = getReturnAddressFrameIndex(DAG);
1834 // Load the "old" Return address.
1835 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1837 return SDValue(OutRetAddr.getNode(), 1);
1840 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1841 /// optimization is performed and it is required (FPDiff!=0).
1843 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1844 SDValue Chain, SDValue RetAddrFrIdx,
1845 bool Is64Bit, int FPDiff, DebugLoc dl) {
1846 // Store the return address to the appropriate stack slot.
1847 if (!FPDiff) return Chain;
1848 // Calculate the new stack slot for the return address.
1849 int SlotSize = Is64Bit ? 8 : 4;
1850 int NewReturnAddrFI =
1851 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1852 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1853 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1854 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1855 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1861 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1862 CallingConv::ID CallConv, bool isVarArg,
1864 const SmallVectorImpl<ISD::OutputArg> &Outs,
1865 const SmallVectorImpl<SDValue> &OutVals,
1866 const SmallVectorImpl<ISD::InputArg> &Ins,
1867 DebugLoc dl, SelectionDAG &DAG,
1868 SmallVectorImpl<SDValue> &InVals) const {
1869 MachineFunction &MF = DAG.getMachineFunction();
1870 bool Is64Bit = Subtarget->is64Bit();
1871 bool IsStructRet = CallIsStructReturn(Outs);
1872 bool IsSibcall = false;
1875 // Check if it's really possible to do a tail call.
1876 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1877 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1878 Outs, OutVals, Ins, DAG);
1880 // Sibcalls are automatically detected tailcalls which do not require
1882 if (!GuaranteedTailCallOpt && isTailCall)
1889 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1890 "Var args not supported with calling convention fastcc or ghc");
1892 // Analyze operands of the call, assigning locations to each operand.
1893 SmallVector<CCValAssign, 16> ArgLocs;
1894 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1895 ArgLocs, *DAG.getContext());
1896 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
1898 // Get a count of how many bytes are to be pushed on the stack.
1899 unsigned NumBytes = CCInfo.getNextStackOffset();
1901 // This is a sibcall. The memory operands are available in caller's
1902 // own caller's stack.
1904 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1905 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1908 if (isTailCall && !IsSibcall) {
1909 // Lower arguments at fp - stackoffset + fpdiff.
1910 unsigned NumBytesCallerPushed =
1911 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1912 FPDiff = NumBytesCallerPushed - NumBytes;
1914 // Set the delta of movement of the returnaddr stackslot.
1915 // But only set if delta is greater than previous delta.
1916 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1917 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1921 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1923 SDValue RetAddrFrIdx;
1924 // Load return adress for tail calls.
1925 if (isTailCall && FPDiff)
1926 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1927 Is64Bit, FPDiff, dl);
1929 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1930 SmallVector<SDValue, 8> MemOpChains;
1933 // Walk the register/memloc assignments, inserting copies/loads. In the case
1934 // of tail call optimization arguments are handle later.
1935 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1936 CCValAssign &VA = ArgLocs[i];
1937 EVT RegVT = VA.getLocVT();
1938 SDValue Arg = OutVals[i];
1939 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1940 bool isByVal = Flags.isByVal();
1942 // Promote the value if needed.
1943 switch (VA.getLocInfo()) {
1944 default: llvm_unreachable("Unknown loc info!");
1945 case CCValAssign::Full: break;
1946 case CCValAssign::SExt:
1947 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1949 case CCValAssign::ZExt:
1950 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1952 case CCValAssign::AExt:
1953 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1954 // Special case: passing MMX values in XMM registers.
1955 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
1956 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1957 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1959 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1961 case CCValAssign::BCvt:
1962 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
1964 case CCValAssign::Indirect: {
1965 // Store the argument.
1966 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1967 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1968 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1969 MachinePointerInfo::getFixedStack(FI),
1976 if (VA.isRegLoc()) {
1977 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1978 if (isVarArg && Subtarget->isTargetWin64()) {
1979 // Win64 ABI requires argument XMM reg to be copied to the corresponding
1980 // shadow reg if callee is a varargs function.
1981 unsigned ShadowReg = 0;
1982 switch (VA.getLocReg()) {
1983 case X86::XMM0: ShadowReg = X86::RCX; break;
1984 case X86::XMM1: ShadowReg = X86::RDX; break;
1985 case X86::XMM2: ShadowReg = X86::R8; break;
1986 case X86::XMM3: ShadowReg = X86::R9; break;
1989 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
1991 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1992 assert(VA.isMemLoc());
1993 if (StackPtr.getNode() == 0)
1994 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1995 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1996 dl, DAG, VA, Flags));
2000 if (!MemOpChains.empty())
2001 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2002 &MemOpChains[0], MemOpChains.size());
2004 // Build a sequence of copy-to-reg nodes chained together with token chain
2005 // and flag operands which copy the outgoing args into registers.
2007 // Tail call byval lowering might overwrite argument registers so in case of
2008 // tail call optimization the copies to registers are lowered later.
2010 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2011 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2012 RegsToPass[i].second, InFlag);
2013 InFlag = Chain.getValue(1);
2016 if (Subtarget->isPICStyleGOT()) {
2017 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2020 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2021 DAG.getNode(X86ISD::GlobalBaseReg,
2022 DebugLoc(), getPointerTy()),
2024 InFlag = Chain.getValue(1);
2026 // If we are tail calling and generating PIC/GOT style code load the
2027 // address of the callee into ECX. The value in ecx is used as target of
2028 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2029 // for tail calls on PIC/GOT architectures. Normally we would just put the
2030 // address of GOT into ebx and then call target@PLT. But for tail calls
2031 // ebx would be restored (since ebx is callee saved) before jumping to the
2034 // Note: The actual moving to ECX is done further down.
2035 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2036 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2037 !G->getGlobal()->hasProtectedVisibility())
2038 Callee = LowerGlobalAddress(Callee, DAG);
2039 else if (isa<ExternalSymbolSDNode>(Callee))
2040 Callee = LowerExternalSymbol(Callee, DAG);
2044 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
2045 // From AMD64 ABI document:
2046 // For calls that may call functions that use varargs or stdargs
2047 // (prototype-less calls or calls to functions containing ellipsis (...) in
2048 // the declaration) %al is used as hidden argument to specify the number
2049 // of SSE registers used. The contents of %al do not need to match exactly
2050 // the number of registers, but must be an ubound on the number of SSE
2051 // registers used and is in the range 0 - 8 inclusive.
2053 // Count the number of XMM registers allocated.
2054 static const unsigned XMMArgRegs[] = {
2055 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2056 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2058 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2059 assert((Subtarget->hasXMM() || !NumXMMRegs)
2060 && "SSE registers cannot be used when SSE is disabled");
2062 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2063 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2064 InFlag = Chain.getValue(1);
2068 // For tail calls lower the arguments to the 'real' stack slot.
2070 // Force all the incoming stack arguments to be loaded from the stack
2071 // before any new outgoing arguments are stored to the stack, because the
2072 // outgoing stack slots may alias the incoming argument stack slots, and
2073 // the alias isn't otherwise explicit. This is slightly more conservative
2074 // than necessary, because it means that each store effectively depends
2075 // on every argument instead of just those arguments it would clobber.
2076 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2078 SmallVector<SDValue, 8> MemOpChains2;
2081 // Do not flag preceeding copytoreg stuff together with the following stuff.
2083 if (GuaranteedTailCallOpt) {
2084 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2085 CCValAssign &VA = ArgLocs[i];
2088 assert(VA.isMemLoc());
2089 SDValue Arg = OutVals[i];
2090 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2091 // Create frame index.
2092 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2093 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2094 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2095 FIN = DAG.getFrameIndex(FI, getPointerTy());
2097 if (Flags.isByVal()) {
2098 // Copy relative to framepointer.
2099 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2100 if (StackPtr.getNode() == 0)
2101 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2103 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2105 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2109 // Store relative to framepointer.
2110 MemOpChains2.push_back(
2111 DAG.getStore(ArgChain, dl, Arg, FIN,
2112 MachinePointerInfo::getFixedStack(FI),
2118 if (!MemOpChains2.empty())
2119 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2120 &MemOpChains2[0], MemOpChains2.size());
2122 // Copy arguments to their registers.
2123 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2124 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2125 RegsToPass[i].second, InFlag);
2126 InFlag = Chain.getValue(1);
2130 // Store the return address to the appropriate stack slot.
2131 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2135 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2136 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2137 // In the 64-bit large code model, we have to make all calls
2138 // through a register, since the call instruction's 32-bit
2139 // pc-relative offset may not be large enough to hold the whole
2141 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2142 // If the callee is a GlobalAddress node (quite common, every direct call
2143 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2146 // We should use extra load for direct calls to dllimported functions in
2148 const GlobalValue *GV = G->getGlobal();
2149 if (!GV->hasDLLImportLinkage()) {
2150 unsigned char OpFlags = 0;
2152 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2153 // external symbols most go through the PLT in PIC mode. If the symbol
2154 // has hidden or protected visibility, or if it is static or local, then
2155 // we don't need to use the PLT - we can directly call it.
2156 if (Subtarget->isTargetELF() &&
2157 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2158 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2159 OpFlags = X86II::MO_PLT;
2160 } else if (Subtarget->isPICStyleStubAny() &&
2161 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2162 Subtarget->getDarwinVers() < 9) {
2163 // PC-relative references to external symbols should go through $stub,
2164 // unless we're building with the leopard linker or later, which
2165 // automatically synthesizes these stubs.
2166 OpFlags = X86II::MO_DARWIN_STUB;
2169 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2170 G->getOffset(), OpFlags);
2172 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2173 unsigned char OpFlags = 0;
2175 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2176 // external symbols should go through the PLT.
2177 if (Subtarget->isTargetELF() &&
2178 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2179 OpFlags = X86II::MO_PLT;
2180 } else if (Subtarget->isPICStyleStubAny() &&
2181 Subtarget->getDarwinVers() < 9) {
2182 // PC-relative references to external symbols should go through $stub,
2183 // unless we're building with the leopard linker or later, which
2184 // automatically synthesizes these stubs.
2185 OpFlags = X86II::MO_DARWIN_STUB;
2188 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2192 // Returns a chain & a flag for retval copy to use.
2193 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2194 SmallVector<SDValue, 8> Ops;
2196 if (!IsSibcall && isTailCall) {
2197 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2198 DAG.getIntPtrConstant(0, true), InFlag);
2199 InFlag = Chain.getValue(1);
2202 Ops.push_back(Chain);
2203 Ops.push_back(Callee);
2206 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2208 // Add argument registers to the end of the list so that they are known live
2210 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2211 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2212 RegsToPass[i].second.getValueType()));
2214 // Add an implicit use GOT pointer in EBX.
2215 if (!isTailCall && Subtarget->isPICStyleGOT())
2216 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2218 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2219 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64())
2220 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2222 if (InFlag.getNode())
2223 Ops.push_back(InFlag);
2227 //// If this is the first return lowered for this function, add the regs
2228 //// to the liveout set for the function.
2229 // This isn't right, although it's probably harmless on x86; liveouts
2230 // should be computed from returns not tail calls. Consider a void
2231 // function making a tail call to a function returning int.
2232 return DAG.getNode(X86ISD::TC_RETURN, dl,
2233 NodeTys, &Ops[0], Ops.size());
2236 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2237 InFlag = Chain.getValue(1);
2239 // Create the CALLSEQ_END node.
2240 unsigned NumBytesForCalleeToPush;
2241 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2242 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2243 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2244 // If this is a call to a struct-return function, the callee
2245 // pops the hidden struct pointer, so we have to push it back.
2246 // This is common for Darwin/X86, Linux & Mingw32 targets.
2247 NumBytesForCalleeToPush = 4;
2249 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2251 // Returns a flag for retval copy to use.
2253 Chain = DAG.getCALLSEQ_END(Chain,
2254 DAG.getIntPtrConstant(NumBytes, true),
2255 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2258 InFlag = Chain.getValue(1);
2261 // Handle result values, copying them out of physregs into vregs that we
2263 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2264 Ins, dl, DAG, InVals);
2268 //===----------------------------------------------------------------------===//
2269 // Fast Calling Convention (tail call) implementation
2270 //===----------------------------------------------------------------------===//
2272 // Like std call, callee cleans arguments, convention except that ECX is
2273 // reserved for storing the tail called function address. Only 2 registers are
2274 // free for argument passing (inreg). Tail call optimization is performed
2276 // * tailcallopt is enabled
2277 // * caller/callee are fastcc
2278 // On X86_64 architecture with GOT-style position independent code only local
2279 // (within module) calls are supported at the moment.
2280 // To keep the stack aligned according to platform abi the function
2281 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2282 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2283 // If a tail called function callee has more arguments than the caller the
2284 // caller needs to make sure that there is room to move the RETADDR to. This is
2285 // achieved by reserving an area the size of the argument delta right after the
2286 // original REtADDR, but before the saved framepointer or the spilled registers
2287 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2299 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2300 /// for a 16 byte align requirement.
2302 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2303 SelectionDAG& DAG) const {
2304 MachineFunction &MF = DAG.getMachineFunction();
2305 const TargetMachine &TM = MF.getTarget();
2306 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2307 unsigned StackAlignment = TFI.getStackAlignment();
2308 uint64_t AlignMask = StackAlignment - 1;
2309 int64_t Offset = StackSize;
2310 uint64_t SlotSize = TD->getPointerSize();
2311 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2312 // Number smaller than 12 so just add the difference.
2313 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2315 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2316 Offset = ((~AlignMask) & Offset) + StackAlignment +
2317 (StackAlignment-SlotSize);
2322 /// MatchingStackOffset - Return true if the given stack call argument is
2323 /// already available in the same position (relatively) of the caller's
2324 /// incoming argument stack.
2326 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2327 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2328 const X86InstrInfo *TII) {
2329 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2331 if (Arg.getOpcode() == ISD::CopyFromReg) {
2332 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2333 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2335 MachineInstr *Def = MRI->getVRegDef(VR);
2338 if (!Flags.isByVal()) {
2339 if (!TII->isLoadFromStackSlot(Def, FI))
2342 unsigned Opcode = Def->getOpcode();
2343 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2344 Def->getOperand(1).isFI()) {
2345 FI = Def->getOperand(1).getIndex();
2346 Bytes = Flags.getByValSize();
2350 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2351 if (Flags.isByVal())
2352 // ByVal argument is passed in as a pointer but it's now being
2353 // dereferenced. e.g.
2354 // define @foo(%struct.X* %A) {
2355 // tail call @bar(%struct.X* byval %A)
2358 SDValue Ptr = Ld->getBasePtr();
2359 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2362 FI = FINode->getIndex();
2366 assert(FI != INT_MAX);
2367 if (!MFI->isFixedObjectIndex(FI))
2369 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2372 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2373 /// for tail call optimization. Targets which want to do tail call
2374 /// optimization should implement this function.
2376 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2377 CallingConv::ID CalleeCC,
2379 bool isCalleeStructRet,
2380 bool isCallerStructRet,
2381 const SmallVectorImpl<ISD::OutputArg> &Outs,
2382 const SmallVectorImpl<SDValue> &OutVals,
2383 const SmallVectorImpl<ISD::InputArg> &Ins,
2384 SelectionDAG& DAG) const {
2385 if (!IsTailCallConvention(CalleeCC) &&
2386 CalleeCC != CallingConv::C)
2389 // If -tailcallopt is specified, make fastcc functions tail-callable.
2390 const MachineFunction &MF = DAG.getMachineFunction();
2391 const Function *CallerF = DAG.getMachineFunction().getFunction();
2392 CallingConv::ID CallerCC = CallerF->getCallingConv();
2393 bool CCMatch = CallerCC == CalleeCC;
2395 if (GuaranteedTailCallOpt) {
2396 if (IsTailCallConvention(CalleeCC) && CCMatch)
2401 // Look for obvious safe cases to perform tail call optimization that do not
2402 // require ABI changes. This is what gcc calls sibcall.
2404 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2405 // emit a special epilogue.
2406 if (RegInfo->needsStackRealignment(MF))
2409 // Do not sibcall optimize vararg calls unless the call site is not passing
2411 if (isVarArg && !Outs.empty())
2414 // Also avoid sibcall optimization if either caller or callee uses struct
2415 // return semantics.
2416 if (isCalleeStructRet || isCallerStructRet)
2419 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2420 // Therefore if it's not used by the call it is not safe to optimize this into
2422 bool Unused = false;
2423 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2430 SmallVector<CCValAssign, 16> RVLocs;
2431 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2432 RVLocs, *DAG.getContext());
2433 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2434 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2435 CCValAssign &VA = RVLocs[i];
2436 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2441 // If the calling conventions do not match, then we'd better make sure the
2442 // results are returned in the same way as what the caller expects.
2444 SmallVector<CCValAssign, 16> RVLocs1;
2445 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2446 RVLocs1, *DAG.getContext());
2447 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2449 SmallVector<CCValAssign, 16> RVLocs2;
2450 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2451 RVLocs2, *DAG.getContext());
2452 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2454 if (RVLocs1.size() != RVLocs2.size())
2456 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2457 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2459 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2461 if (RVLocs1[i].isRegLoc()) {
2462 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2465 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2471 // If the callee takes no arguments then go on to check the results of the
2473 if (!Outs.empty()) {
2474 // Check if stack adjustment is needed. For now, do not do this if any
2475 // argument is passed on the stack.
2476 SmallVector<CCValAssign, 16> ArgLocs;
2477 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2478 ArgLocs, *DAG.getContext());
2479 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2480 if (CCInfo.getNextStackOffset()) {
2481 MachineFunction &MF = DAG.getMachineFunction();
2482 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2484 if (Subtarget->isTargetWin64())
2485 // Win64 ABI has additional complications.
2488 // Check if the arguments are already laid out in the right way as
2489 // the caller's fixed stack objects.
2490 MachineFrameInfo *MFI = MF.getFrameInfo();
2491 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2492 const X86InstrInfo *TII =
2493 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2494 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2495 CCValAssign &VA = ArgLocs[i];
2496 SDValue Arg = OutVals[i];
2497 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2498 if (VA.getLocInfo() == CCValAssign::Indirect)
2500 if (!VA.isRegLoc()) {
2501 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2508 // If the tailcall address may be in a register, then make sure it's
2509 // possible to register allocate for it. In 32-bit, the call address can
2510 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2511 // callee-saved registers are restored. These happen to be the same
2512 // registers used to pass 'inreg' arguments so watch out for those.
2513 if (!Subtarget->is64Bit() &&
2514 !isa<GlobalAddressSDNode>(Callee) &&
2515 !isa<ExternalSymbolSDNode>(Callee)) {
2516 unsigned NumInRegs = 0;
2517 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2518 CCValAssign &VA = ArgLocs[i];
2521 unsigned Reg = VA.getLocReg();
2524 case X86::EAX: case X86::EDX: case X86::ECX:
2525 if (++NumInRegs == 3)
2533 // An stdcall caller is expected to clean up its arguments; the callee
2534 // isn't going to do that.
2535 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2542 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2543 return X86::createFastISel(funcInfo);
2547 //===----------------------------------------------------------------------===//
2548 // Other Lowering Hooks
2549 //===----------------------------------------------------------------------===//
2551 static bool MayFoldLoad(SDValue Op) {
2552 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2555 static bool MayFoldIntoStore(SDValue Op) {
2556 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2559 static bool isTargetShuffle(unsigned Opcode) {
2561 default: return false;
2562 case X86ISD::PSHUFD:
2563 case X86ISD::PSHUFHW:
2564 case X86ISD::PSHUFLW:
2565 case X86ISD::SHUFPD:
2566 case X86ISD::PALIGN:
2567 case X86ISD::SHUFPS:
2568 case X86ISD::MOVLHPS:
2569 case X86ISD::MOVLHPD:
2570 case X86ISD::MOVHLPS:
2571 case X86ISD::MOVLPS:
2572 case X86ISD::MOVLPD:
2573 case X86ISD::MOVSHDUP:
2574 case X86ISD::MOVSLDUP:
2575 case X86ISD::MOVDDUP:
2578 case X86ISD::UNPCKLPS:
2579 case X86ISD::UNPCKLPD:
2580 case X86ISD::PUNPCKLWD:
2581 case X86ISD::PUNPCKLBW:
2582 case X86ISD::PUNPCKLDQ:
2583 case X86ISD::PUNPCKLQDQ:
2584 case X86ISD::UNPCKHPS:
2585 case X86ISD::UNPCKHPD:
2586 case X86ISD::PUNPCKHWD:
2587 case X86ISD::PUNPCKHBW:
2588 case X86ISD::PUNPCKHDQ:
2589 case X86ISD::PUNPCKHQDQ:
2595 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2596 SDValue V1, SelectionDAG &DAG) {
2598 default: llvm_unreachable("Unknown x86 shuffle node");
2599 case X86ISD::MOVSHDUP:
2600 case X86ISD::MOVSLDUP:
2601 case X86ISD::MOVDDUP:
2602 return DAG.getNode(Opc, dl, VT, V1);
2608 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2609 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2611 default: llvm_unreachable("Unknown x86 shuffle node");
2612 case X86ISD::PSHUFD:
2613 case X86ISD::PSHUFHW:
2614 case X86ISD::PSHUFLW:
2615 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2621 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2622 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2624 default: llvm_unreachable("Unknown x86 shuffle node");
2625 case X86ISD::PALIGN:
2626 case X86ISD::SHUFPD:
2627 case X86ISD::SHUFPS:
2628 return DAG.getNode(Opc, dl, VT, V1, V2,
2629 DAG.getConstant(TargetMask, MVT::i8));
2634 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2635 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2637 default: llvm_unreachable("Unknown x86 shuffle node");
2638 case X86ISD::MOVLHPS:
2639 case X86ISD::MOVLHPD:
2640 case X86ISD::MOVHLPS:
2641 case X86ISD::MOVLPS:
2642 case X86ISD::MOVLPD:
2645 case X86ISD::UNPCKLPS:
2646 case X86ISD::UNPCKLPD:
2647 case X86ISD::PUNPCKLWD:
2648 case X86ISD::PUNPCKLBW:
2649 case X86ISD::PUNPCKLDQ:
2650 case X86ISD::PUNPCKLQDQ:
2651 case X86ISD::UNPCKHPS:
2652 case X86ISD::UNPCKHPD:
2653 case X86ISD::PUNPCKHWD:
2654 case X86ISD::PUNPCKHBW:
2655 case X86ISD::PUNPCKHDQ:
2656 case X86ISD::PUNPCKHQDQ:
2657 return DAG.getNode(Opc, dl, VT, V1, V2);
2662 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2663 MachineFunction &MF = DAG.getMachineFunction();
2664 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2665 int ReturnAddrIndex = FuncInfo->getRAIndex();
2667 if (ReturnAddrIndex == 0) {
2668 // Set up a frame object for the return address.
2669 uint64_t SlotSize = TD->getPointerSize();
2670 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2672 FuncInfo->setRAIndex(ReturnAddrIndex);
2675 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2679 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2680 bool hasSymbolicDisplacement) {
2681 // Offset should fit into 32 bit immediate field.
2682 if (!isInt<32>(Offset))
2685 // If we don't have a symbolic displacement - we don't have any extra
2687 if (!hasSymbolicDisplacement)
2690 // FIXME: Some tweaks might be needed for medium code model.
2691 if (M != CodeModel::Small && M != CodeModel::Kernel)
2694 // For small code model we assume that latest object is 16MB before end of 31
2695 // bits boundary. We may also accept pretty large negative constants knowing
2696 // that all objects are in the positive half of address space.
2697 if (M == CodeModel::Small && Offset < 16*1024*1024)
2700 // For kernel code model we know that all object resist in the negative half
2701 // of 32bits address space. We may not accept negative offsets, since they may
2702 // be just off and we may accept pretty large positive ones.
2703 if (M == CodeModel::Kernel && Offset > 0)
2709 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2710 /// specific condition code, returning the condition code and the LHS/RHS of the
2711 /// comparison to make.
2712 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2713 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2715 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2716 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2717 // X > -1 -> X == 0, jump !sign.
2718 RHS = DAG.getConstant(0, RHS.getValueType());
2719 return X86::COND_NS;
2720 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2721 // X < 0 -> X == 0, jump on sign.
2723 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2725 RHS = DAG.getConstant(0, RHS.getValueType());
2726 return X86::COND_LE;
2730 switch (SetCCOpcode) {
2731 default: llvm_unreachable("Invalid integer condition!");
2732 case ISD::SETEQ: return X86::COND_E;
2733 case ISD::SETGT: return X86::COND_G;
2734 case ISD::SETGE: return X86::COND_GE;
2735 case ISD::SETLT: return X86::COND_L;
2736 case ISD::SETLE: return X86::COND_LE;
2737 case ISD::SETNE: return X86::COND_NE;
2738 case ISD::SETULT: return X86::COND_B;
2739 case ISD::SETUGT: return X86::COND_A;
2740 case ISD::SETULE: return X86::COND_BE;
2741 case ISD::SETUGE: return X86::COND_AE;
2745 // First determine if it is required or is profitable to flip the operands.
2747 // If LHS is a foldable load, but RHS is not, flip the condition.
2748 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2749 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2750 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2751 std::swap(LHS, RHS);
2754 switch (SetCCOpcode) {
2760 std::swap(LHS, RHS);
2764 // On a floating point condition, the flags are set as follows:
2766 // 0 | 0 | 0 | X > Y
2767 // 0 | 0 | 1 | X < Y
2768 // 1 | 0 | 0 | X == Y
2769 // 1 | 1 | 1 | unordered
2770 switch (SetCCOpcode) {
2771 default: llvm_unreachable("Condcode should be pre-legalized away");
2773 case ISD::SETEQ: return X86::COND_E;
2774 case ISD::SETOLT: // flipped
2776 case ISD::SETGT: return X86::COND_A;
2777 case ISD::SETOLE: // flipped
2779 case ISD::SETGE: return X86::COND_AE;
2780 case ISD::SETUGT: // flipped
2782 case ISD::SETLT: return X86::COND_B;
2783 case ISD::SETUGE: // flipped
2785 case ISD::SETLE: return X86::COND_BE;
2787 case ISD::SETNE: return X86::COND_NE;
2788 case ISD::SETUO: return X86::COND_P;
2789 case ISD::SETO: return X86::COND_NP;
2791 case ISD::SETUNE: return X86::COND_INVALID;
2795 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2796 /// code. Current x86 isa includes the following FP cmov instructions:
2797 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2798 static bool hasFPCMov(unsigned X86CC) {
2814 /// isFPImmLegal - Returns true if the target can instruction select the
2815 /// specified FP immediate natively. If false, the legalizer will
2816 /// materialize the FP immediate as a load from a constant pool.
2817 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2818 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2819 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2825 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2826 /// the specified range (L, H].
2827 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2828 return (Val < 0) || (Val >= Low && Val < Hi);
2831 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2832 /// specified value.
2833 static bool isUndefOrEqual(int Val, int CmpVal) {
2834 if (Val < 0 || Val == CmpVal)
2839 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2840 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2841 /// the second operand.
2842 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2843 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
2844 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2845 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2846 return (Mask[0] < 2 && Mask[1] < 2);
2850 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2851 SmallVector<int, 8> M;
2853 return ::isPSHUFDMask(M, N->getValueType(0));
2856 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2857 /// is suitable for input to PSHUFHW.
2858 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2859 if (VT != MVT::v8i16)
2862 // Lower quadword copied in order or undef.
2863 for (int i = 0; i != 4; ++i)
2864 if (Mask[i] >= 0 && Mask[i] != i)
2867 // Upper quadword shuffled.
2868 for (int i = 4; i != 8; ++i)
2869 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2875 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2876 SmallVector<int, 8> M;
2878 return ::isPSHUFHWMask(M, N->getValueType(0));
2881 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2882 /// is suitable for input to PSHUFLW.
2883 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2884 if (VT != MVT::v8i16)
2887 // Upper quadword copied in order.
2888 for (int i = 4; i != 8; ++i)
2889 if (Mask[i] >= 0 && Mask[i] != i)
2892 // Lower quadword shuffled.
2893 for (int i = 0; i != 4; ++i)
2900 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2901 SmallVector<int, 8> M;
2903 return ::isPSHUFLWMask(M, N->getValueType(0));
2906 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2907 /// is suitable for input to PALIGNR.
2908 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2910 int i, e = VT.getVectorNumElements();
2912 // Do not handle v2i64 / v2f64 shuffles with palignr.
2913 if (e < 4 || !hasSSSE3)
2916 for (i = 0; i != e; ++i)
2920 // All undef, not a palignr.
2924 // Determine if it's ok to perform a palignr with only the LHS, since we
2925 // don't have access to the actual shuffle elements to see if RHS is undef.
2926 bool Unary = Mask[i] < (int)e;
2927 bool NeedsUnary = false;
2929 int s = Mask[i] - i;
2931 // Check the rest of the elements to see if they are consecutive.
2932 for (++i; i != e; ++i) {
2937 Unary = Unary && (m < (int)e);
2938 NeedsUnary = NeedsUnary || (m < s);
2940 if (NeedsUnary && !Unary)
2942 if (Unary && m != ((s+i) & (e-1)))
2944 if (!Unary && m != (s+i))
2950 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2951 SmallVector<int, 8> M;
2953 return ::isPALIGNRMask(M, N->getValueType(0), true);
2956 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2957 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2958 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2959 int NumElems = VT.getVectorNumElements();
2960 if (NumElems != 2 && NumElems != 4)
2963 int Half = NumElems / 2;
2964 for (int i = 0; i < Half; ++i)
2965 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2967 for (int i = Half; i < NumElems; ++i)
2968 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2974 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2975 SmallVector<int, 8> M;
2977 return ::isSHUFPMask(M, N->getValueType(0));
2980 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2981 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2982 /// half elements to come from vector 1 (which would equal the dest.) and
2983 /// the upper half to come from vector 2.
2984 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2985 int NumElems = VT.getVectorNumElements();
2987 if (NumElems != 2 && NumElems != 4)
2990 int Half = NumElems / 2;
2991 for (int i = 0; i < Half; ++i)
2992 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2994 for (int i = Half; i < NumElems; ++i)
2995 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3000 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3001 SmallVector<int, 8> M;
3003 return isCommutedSHUFPMask(M, N->getValueType(0));
3006 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3007 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3008 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3009 if (N->getValueType(0).getVectorNumElements() != 4)
3012 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3013 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3014 isUndefOrEqual(N->getMaskElt(1), 7) &&
3015 isUndefOrEqual(N->getMaskElt(2), 2) &&
3016 isUndefOrEqual(N->getMaskElt(3), 3);
3019 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3020 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3022 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3023 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3028 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3029 isUndefOrEqual(N->getMaskElt(1), 3) &&
3030 isUndefOrEqual(N->getMaskElt(2), 2) &&
3031 isUndefOrEqual(N->getMaskElt(3), 3);
3034 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3035 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3036 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3037 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3039 if (NumElems != 2 && NumElems != 4)
3042 for (unsigned i = 0; i < NumElems/2; ++i)
3043 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3046 for (unsigned i = NumElems/2; i < NumElems; ++i)
3047 if (!isUndefOrEqual(N->getMaskElt(i), i))
3053 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3054 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3055 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3056 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3058 if (NumElems != 2 && NumElems != 4)
3061 for (unsigned i = 0; i < NumElems/2; ++i)
3062 if (!isUndefOrEqual(N->getMaskElt(i), i))
3065 for (unsigned i = 0; i < NumElems/2; ++i)
3066 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3072 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3073 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3074 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3075 bool V2IsSplat = false) {
3076 int NumElts = VT.getVectorNumElements();
3077 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3080 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3082 int BitI1 = Mask[i+1];
3083 if (!isUndefOrEqual(BitI, j))
3086 if (!isUndefOrEqual(BitI1, NumElts))
3089 if (!isUndefOrEqual(BitI1, j + NumElts))
3096 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3097 SmallVector<int, 8> M;
3099 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3102 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3103 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3104 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3105 bool V2IsSplat = false) {
3106 int NumElts = VT.getVectorNumElements();
3107 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3110 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3112 int BitI1 = Mask[i+1];
3113 if (!isUndefOrEqual(BitI, j + NumElts/2))
3116 if (isUndefOrEqual(BitI1, NumElts))
3119 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3126 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3127 SmallVector<int, 8> M;
3129 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3132 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3133 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3135 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3136 int NumElems = VT.getVectorNumElements();
3137 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3140 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3142 int BitI1 = Mask[i+1];
3143 if (!isUndefOrEqual(BitI, j))
3145 if (!isUndefOrEqual(BitI1, j))
3151 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3152 SmallVector<int, 8> M;
3154 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3157 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3158 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3160 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3161 int NumElems = VT.getVectorNumElements();
3162 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3165 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3167 int BitI1 = Mask[i+1];
3168 if (!isUndefOrEqual(BitI, j))
3170 if (!isUndefOrEqual(BitI1, j))
3176 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3177 SmallVector<int, 8> M;
3179 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3182 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3183 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3184 /// MOVSD, and MOVD, i.e. setting the lowest element.
3185 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3186 if (VT.getVectorElementType().getSizeInBits() < 32)
3189 int NumElts = VT.getVectorNumElements();
3191 if (!isUndefOrEqual(Mask[0], NumElts))
3194 for (int i = 1; i < NumElts; ++i)
3195 if (!isUndefOrEqual(Mask[i], i))
3201 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3202 SmallVector<int, 8> M;
3204 return ::isMOVLMask(M, N->getValueType(0));
3207 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3208 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3209 /// element of vector 2 and the other elements to come from vector 1 in order.
3210 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3211 bool V2IsSplat = false, bool V2IsUndef = false) {
3212 int NumOps = VT.getVectorNumElements();
3213 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3216 if (!isUndefOrEqual(Mask[0], 0))
3219 for (int i = 1; i < NumOps; ++i)
3220 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3221 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3222 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3228 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3229 bool V2IsUndef = false) {
3230 SmallVector<int, 8> M;
3232 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3235 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3236 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3237 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3238 if (N->getValueType(0).getVectorNumElements() != 4)
3241 // Expect 1, 1, 3, 3
3242 for (unsigned i = 0; i < 2; ++i) {
3243 int Elt = N->getMaskElt(i);
3244 if (Elt >= 0 && Elt != 1)
3249 for (unsigned i = 2; i < 4; ++i) {
3250 int Elt = N->getMaskElt(i);
3251 if (Elt >= 0 && Elt != 3)
3256 // Don't use movshdup if it can be done with a shufps.
3257 // FIXME: verify that matching u, u, 3, 3 is what we want.
3261 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3262 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3263 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3264 if (N->getValueType(0).getVectorNumElements() != 4)
3267 // Expect 0, 0, 2, 2
3268 for (unsigned i = 0; i < 2; ++i)
3269 if (N->getMaskElt(i) > 0)
3273 for (unsigned i = 2; i < 4; ++i) {
3274 int Elt = N->getMaskElt(i);
3275 if (Elt >= 0 && Elt != 2)
3280 // Don't use movsldup if it can be done with a shufps.
3284 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3285 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3286 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3287 int e = N->getValueType(0).getVectorNumElements() / 2;
3289 for (int i = 0; i < e; ++i)
3290 if (!isUndefOrEqual(N->getMaskElt(i), i))
3292 for (int i = 0; i < e; ++i)
3293 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3298 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3299 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3300 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3301 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3302 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3304 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3306 for (int i = 0; i < NumOperands; ++i) {
3307 int Val = SVOp->getMaskElt(NumOperands-i-1);
3308 if (Val < 0) Val = 0;
3309 if (Val >= NumOperands) Val -= NumOperands;
3311 if (i != NumOperands - 1)
3317 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3318 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3319 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3320 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3322 // 8 nodes, but we only care about the last 4.
3323 for (unsigned i = 7; i >= 4; --i) {
3324 int Val = SVOp->getMaskElt(i);
3333 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3334 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3335 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3336 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3338 // 8 nodes, but we only care about the first 4.
3339 for (int i = 3; i >= 0; --i) {
3340 int Val = SVOp->getMaskElt(i);
3349 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3350 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3351 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3352 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3353 EVT VVT = N->getValueType(0);
3354 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3358 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3359 Val = SVOp->getMaskElt(i);
3363 return (Val - i) * EltSize;
3366 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3368 bool X86::isZeroNode(SDValue Elt) {
3369 return ((isa<ConstantSDNode>(Elt) &&
3370 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3371 (isa<ConstantFPSDNode>(Elt) &&
3372 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3375 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3376 /// their permute mask.
3377 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3378 SelectionDAG &DAG) {
3379 EVT VT = SVOp->getValueType(0);
3380 unsigned NumElems = VT.getVectorNumElements();
3381 SmallVector<int, 8> MaskVec;
3383 for (unsigned i = 0; i != NumElems; ++i) {
3384 int idx = SVOp->getMaskElt(i);
3386 MaskVec.push_back(idx);
3387 else if (idx < (int)NumElems)
3388 MaskVec.push_back(idx + NumElems);
3390 MaskVec.push_back(idx - NumElems);
3392 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3393 SVOp->getOperand(0), &MaskVec[0]);
3396 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3397 /// the two vector operands have swapped position.
3398 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3399 unsigned NumElems = VT.getVectorNumElements();
3400 for (unsigned i = 0; i != NumElems; ++i) {
3404 else if (idx < (int)NumElems)
3405 Mask[i] = idx + NumElems;
3407 Mask[i] = idx - NumElems;
3411 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3412 /// match movhlps. The lower half elements should come from upper half of
3413 /// V1 (and in order), and the upper half elements should come from the upper
3414 /// half of V2 (and in order).
3415 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3416 if (Op->getValueType(0).getVectorNumElements() != 4)
3418 for (unsigned i = 0, e = 2; i != e; ++i)
3419 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3421 for (unsigned i = 2; i != 4; ++i)
3422 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3427 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3428 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3430 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3431 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3433 N = N->getOperand(0).getNode();
3434 if (!ISD::isNON_EXTLoad(N))
3437 *LD = cast<LoadSDNode>(N);
3441 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3442 /// match movlp{s|d}. The lower half elements should come from lower half of
3443 /// V1 (and in order), and the upper half elements should come from the upper
3444 /// half of V2 (and in order). And since V1 will become the source of the
3445 /// MOVLP, it must be either a vector load or a scalar load to vector.
3446 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3447 ShuffleVectorSDNode *Op) {
3448 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3450 // Is V2 is a vector load, don't do this transformation. We will try to use
3451 // load folding shufps op.
3452 if (ISD::isNON_EXTLoad(V2))
3455 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3457 if (NumElems != 2 && NumElems != 4)
3459 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3460 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3462 for (unsigned i = NumElems/2; i != NumElems; ++i)
3463 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3468 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3470 static bool isSplatVector(SDNode *N) {
3471 if (N->getOpcode() != ISD::BUILD_VECTOR)
3474 SDValue SplatValue = N->getOperand(0);
3475 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3476 if (N->getOperand(i) != SplatValue)
3481 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3482 /// to an zero vector.
3483 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3484 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3485 SDValue V1 = N->getOperand(0);
3486 SDValue V2 = N->getOperand(1);
3487 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3488 for (unsigned i = 0; i != NumElems; ++i) {
3489 int Idx = N->getMaskElt(i);
3490 if (Idx >= (int)NumElems) {
3491 unsigned Opc = V2.getOpcode();
3492 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3494 if (Opc != ISD::BUILD_VECTOR ||
3495 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3497 } else if (Idx >= 0) {
3498 unsigned Opc = V1.getOpcode();
3499 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3501 if (Opc != ISD::BUILD_VECTOR ||
3502 !X86::isZeroNode(V1.getOperand(Idx)))
3509 /// getZeroVector - Returns a vector of specified type with all zero elements.
3511 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3513 assert(VT.isVector() && "Expected a vector type");
3515 // Always build SSE zero vectors as <4 x i32> bitcasted
3516 // to their dest type. This ensures they get CSE'd.
3518 if (VT.getSizeInBits() == 128) { // SSE
3519 if (HasSSE2) { // SSE2
3520 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3521 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3523 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3524 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3526 } else if (VT.getSizeInBits() == 256) { // AVX
3527 // 256-bit logic and arithmetic instructions in AVX are
3528 // all floating-point, no support for integer ops. Default
3529 // to emitting fp zeroed vectors then.
3530 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3531 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3532 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3534 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3537 /// getOnesVector - Returns a vector of specified type with all bits set.
3539 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3540 assert(VT.isVector() && "Expected a vector type");
3542 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3543 // type. This ensures they get CSE'd.
3544 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3546 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3547 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3551 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3552 /// that point to V2 points to its first element.
3553 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3554 EVT VT = SVOp->getValueType(0);
3555 unsigned NumElems = VT.getVectorNumElements();
3557 bool Changed = false;
3558 SmallVector<int, 8> MaskVec;
3559 SVOp->getMask(MaskVec);
3561 for (unsigned i = 0; i != NumElems; ++i) {
3562 if (MaskVec[i] > (int)NumElems) {
3563 MaskVec[i] = NumElems;
3568 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3569 SVOp->getOperand(1), &MaskVec[0]);
3570 return SDValue(SVOp, 0);
3573 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3574 /// operation of specified width.
3575 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3577 unsigned NumElems = VT.getVectorNumElements();
3578 SmallVector<int, 8> Mask;
3579 Mask.push_back(NumElems);
3580 for (unsigned i = 1; i != NumElems; ++i)
3582 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3585 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3586 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3588 unsigned NumElems = VT.getVectorNumElements();
3589 SmallVector<int, 8> Mask;
3590 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3592 Mask.push_back(i + NumElems);
3594 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3597 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3598 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3600 unsigned NumElems = VT.getVectorNumElements();
3601 unsigned Half = NumElems/2;
3602 SmallVector<int, 8> Mask;
3603 for (unsigned i = 0; i != Half; ++i) {
3604 Mask.push_back(i + Half);
3605 Mask.push_back(i + NumElems + Half);
3607 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3610 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3611 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3612 EVT PVT = MVT::v4f32;
3613 EVT VT = SV->getValueType(0);
3614 DebugLoc dl = SV->getDebugLoc();
3615 SDValue V1 = SV->getOperand(0);
3616 int NumElems = VT.getVectorNumElements();
3617 int EltNo = SV->getSplatIndex();
3619 // unpack elements to the correct location
3620 while (NumElems > 4) {
3621 if (EltNo < NumElems/2) {
3622 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3624 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3625 EltNo -= NumElems/2;
3630 // Perform the splat.
3631 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3632 V1 = DAG.getNode(ISD::BITCAST, dl, PVT, V1);
3633 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3634 return DAG.getNode(ISD::BITCAST, dl, VT, V1);
3637 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3638 /// vector of zero or undef vector. This produces a shuffle where the low
3639 /// element of V2 is swizzled into the zero/undef vector, landing at element
3640 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3641 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3642 bool isZero, bool HasSSE2,
3643 SelectionDAG &DAG) {
3644 EVT VT = V2.getValueType();
3646 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3647 unsigned NumElems = VT.getVectorNumElements();
3648 SmallVector<int, 16> MaskVec;
3649 for (unsigned i = 0; i != NumElems; ++i)
3650 // If this is the insertion idx, put the low elt of V2 here.
3651 MaskVec.push_back(i == Idx ? NumElems : i);
3652 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3655 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
3656 /// element of the result of the vector shuffle.
3657 SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
3660 return SDValue(); // Limit search depth.
3662 SDValue V = SDValue(N, 0);
3663 EVT VT = V.getValueType();
3664 unsigned Opcode = V.getOpcode();
3666 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
3667 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
3668 Index = SV->getMaskElt(Index);
3671 return DAG.getUNDEF(VT.getVectorElementType());
3673 int NumElems = VT.getVectorNumElements();
3674 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
3675 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
3678 // Recurse into target specific vector shuffles to find scalars.
3679 if (isTargetShuffle(Opcode)) {
3680 int NumElems = VT.getVectorNumElements();
3681 SmallVector<unsigned, 16> ShuffleMask;
3685 case X86ISD::SHUFPS:
3686 case X86ISD::SHUFPD:
3687 ImmN = N->getOperand(N->getNumOperands()-1);
3688 DecodeSHUFPSMask(NumElems,
3689 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3692 case X86ISD::PUNPCKHBW:
3693 case X86ISD::PUNPCKHWD:
3694 case X86ISD::PUNPCKHDQ:
3695 case X86ISD::PUNPCKHQDQ:
3696 DecodePUNPCKHMask(NumElems, ShuffleMask);
3698 case X86ISD::UNPCKHPS:
3699 case X86ISD::UNPCKHPD:
3700 DecodeUNPCKHPMask(NumElems, ShuffleMask);
3702 case X86ISD::PUNPCKLBW:
3703 case X86ISD::PUNPCKLWD:
3704 case X86ISD::PUNPCKLDQ:
3705 case X86ISD::PUNPCKLQDQ:
3706 DecodePUNPCKLMask(NumElems, ShuffleMask);
3708 case X86ISD::UNPCKLPS:
3709 case X86ISD::UNPCKLPD:
3710 DecodeUNPCKLPMask(NumElems, ShuffleMask);
3712 case X86ISD::MOVHLPS:
3713 DecodeMOVHLPSMask(NumElems, ShuffleMask);
3715 case X86ISD::MOVLHPS:
3716 DecodeMOVLHPSMask(NumElems, ShuffleMask);
3718 case X86ISD::PSHUFD:
3719 ImmN = N->getOperand(N->getNumOperands()-1);
3720 DecodePSHUFMask(NumElems,
3721 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3724 case X86ISD::PSHUFHW:
3725 ImmN = N->getOperand(N->getNumOperands()-1);
3726 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3729 case X86ISD::PSHUFLW:
3730 ImmN = N->getOperand(N->getNumOperands()-1);
3731 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3735 case X86ISD::MOVSD: {
3736 // The index 0 always comes from the first element of the second source,
3737 // this is why MOVSS and MOVSD are used in the first place. The other
3738 // elements come from the other positions of the first source vector.
3739 unsigned OpNum = (Index == 0) ? 1 : 0;
3740 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
3744 assert("not implemented for target shuffle node");
3748 Index = ShuffleMask[Index];
3750 return DAG.getUNDEF(VT.getVectorElementType());
3752 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
3753 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
3757 // Actual nodes that may contain scalar elements
3758 if (Opcode == ISD::BITCAST) {
3759 V = V.getOperand(0);
3760 EVT SrcVT = V.getValueType();
3761 unsigned NumElems = VT.getVectorNumElements();
3763 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
3767 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
3768 return (Index == 0) ? V.getOperand(0)
3769 : DAG.getUNDEF(VT.getVectorElementType());
3771 if (V.getOpcode() == ISD::BUILD_VECTOR)
3772 return V.getOperand(Index);
3777 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
3778 /// shuffle operation which come from a consecutively from a zero. The
3779 /// search can start in two diferent directions, from left or right.
3781 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
3782 bool ZerosFromLeft, SelectionDAG &DAG) {
3785 while (i < NumElems) {
3786 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
3787 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
3788 if (!(Elt.getNode() &&
3789 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
3797 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
3798 /// MaskE correspond consecutively to elements from one of the vector operands,
3799 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
3801 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
3802 int OpIdx, int NumElems, unsigned &OpNum) {
3803 bool SeenV1 = false;
3804 bool SeenV2 = false;
3806 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
3807 int Idx = SVOp->getMaskElt(i);
3808 // Ignore undef indicies
3817 // Only accept consecutive elements from the same vector
3818 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
3822 OpNum = SeenV1 ? 0 : 1;
3826 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
3827 /// logical left shift of a vector.
3828 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3829 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3830 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3831 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3832 false /* check zeros from right */, DAG);
3838 // Considering the elements in the mask that are not consecutive zeros,
3839 // check if they consecutively come from only one of the source vectors.
3841 // V1 = {X, A, B, C} 0
3843 // vector_shuffle V1, V2 <1, 2, 3, X>
3845 if (!isShuffleMaskConsecutive(SVOp,
3846 0, // Mask Start Index
3847 NumElems-NumZeros-1, // Mask End Index
3848 NumZeros, // Where to start looking in the src vector
3849 NumElems, // Number of elements in vector
3850 OpSrc)) // Which source operand ?
3855 ShVal = SVOp->getOperand(OpSrc);
3859 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
3860 /// logical left shift of a vector.
3861 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3862 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3863 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3864 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3865 true /* check zeros from left */, DAG);
3871 // Considering the elements in the mask that are not consecutive zeros,
3872 // check if they consecutively come from only one of the source vectors.
3874 // 0 { A, B, X, X } = V2
3876 // vector_shuffle V1, V2 <X, X, 4, 5>
3878 if (!isShuffleMaskConsecutive(SVOp,
3879 NumZeros, // Mask Start Index
3880 NumElems-1, // Mask End Index
3881 0, // Where to start looking in the src vector
3882 NumElems, // Number of elements in vector
3883 OpSrc)) // Which source operand ?
3888 ShVal = SVOp->getOperand(OpSrc);
3892 /// isVectorShift - Returns true if the shuffle can be implemented as a
3893 /// logical left or right shift of a vector.
3894 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3895 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3896 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
3897 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
3903 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3905 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3906 unsigned NumNonZero, unsigned NumZero,
3908 const TargetLowering &TLI) {
3912 DebugLoc dl = Op.getDebugLoc();
3915 for (unsigned i = 0; i < 16; ++i) {
3916 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3917 if (ThisIsNonZero && First) {
3919 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3921 V = DAG.getUNDEF(MVT::v8i16);
3926 SDValue ThisElt(0, 0), LastElt(0, 0);
3927 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3928 if (LastIsNonZero) {
3929 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3930 MVT::i16, Op.getOperand(i-1));
3932 if (ThisIsNonZero) {
3933 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3934 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3935 ThisElt, DAG.getConstant(8, MVT::i8));
3937 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3941 if (ThisElt.getNode())
3942 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3943 DAG.getIntPtrConstant(i/2));
3947 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
3950 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3952 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3953 unsigned NumNonZero, unsigned NumZero,
3955 const TargetLowering &TLI) {
3959 DebugLoc dl = Op.getDebugLoc();
3962 for (unsigned i = 0; i < 8; ++i) {
3963 bool isNonZero = (NonZeros & (1 << i)) != 0;
3967 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3969 V = DAG.getUNDEF(MVT::v8i16);
3972 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3973 MVT::v8i16, V, Op.getOperand(i),
3974 DAG.getIntPtrConstant(i));
3981 /// getVShift - Return a vector logical shift node.
3983 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3984 unsigned NumBits, SelectionDAG &DAG,
3985 const TargetLowering &TLI, DebugLoc dl) {
3986 EVT ShVT = MVT::v2i64;
3987 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3988 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
3989 return DAG.getNode(ISD::BITCAST, dl, VT,
3990 DAG.getNode(Opc, dl, ShVT, SrcOp,
3991 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3995 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3996 SelectionDAG &DAG) const {
3998 // Check if the scalar load can be widened into a vector load. And if
3999 // the address is "base + cst" see if the cst can be "absorbed" into
4000 // the shuffle mask.
4001 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4002 SDValue Ptr = LD->getBasePtr();
4003 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4005 EVT PVT = LD->getValueType(0);
4006 if (PVT != MVT::i32 && PVT != MVT::f32)
4011 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4012 FI = FINode->getIndex();
4014 } else if (Ptr.getOpcode() == ISD::ADD &&
4015 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
4016 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4017 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4018 Offset = Ptr.getConstantOperandVal(1);
4019 Ptr = Ptr.getOperand(0);
4024 SDValue Chain = LD->getChain();
4025 // Make sure the stack object alignment is at least 16.
4026 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4027 if (DAG.InferPtrAlignment(Ptr) < 16) {
4028 if (MFI->isFixedObjectIndex(FI)) {
4029 // Can't change the alignment. FIXME: It's possible to compute
4030 // the exact stack offset and reference FI + adjust offset instead.
4031 // If someone *really* cares about this. That's the way to implement it.
4034 MFI->setObjectAlignment(FI, 16);
4038 // (Offset % 16) must be multiple of 4. Then address is then
4039 // Ptr + (Offset & ~15).
4042 if ((Offset % 16) & 3)
4044 int64_t StartOffset = Offset & ~15;
4046 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4047 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4049 int EltNo = (Offset - StartOffset) >> 2;
4050 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4051 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4052 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,
4053 LD->getPointerInfo().getWithOffset(StartOffset),
4055 // Canonicalize it to a v4i32 shuffle.
4056 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
4057 return DAG.getNode(ISD::BITCAST, dl, VT,
4058 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4059 DAG.getUNDEF(MVT::v4i32),&Mask[0]));
4065 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4066 /// vector of type 'VT', see if the elements can be replaced by a single large
4067 /// load which has the same value as a build_vector whose operands are 'elts'.
4069 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4071 /// FIXME: we'd also like to handle the case where the last elements are zero
4072 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4073 /// There's even a handy isZeroNode for that purpose.
4074 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4075 DebugLoc &DL, SelectionDAG &DAG) {
4076 EVT EltVT = VT.getVectorElementType();
4077 unsigned NumElems = Elts.size();
4079 LoadSDNode *LDBase = NULL;
4080 unsigned LastLoadedElt = -1U;
4082 // For each element in the initializer, see if we've found a load or an undef.
4083 // If we don't find an initial load element, or later load elements are
4084 // non-consecutive, bail out.
4085 for (unsigned i = 0; i < NumElems; ++i) {
4086 SDValue Elt = Elts[i];
4088 if (!Elt.getNode() ||
4089 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4092 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4094 LDBase = cast<LoadSDNode>(Elt.getNode());
4098 if (Elt.getOpcode() == ISD::UNDEF)
4101 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4102 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4107 // If we have found an entire vector of loads and undefs, then return a large
4108 // load of the entire vector width starting at the base pointer. If we found
4109 // consecutive loads for the low half, generate a vzext_load node.
4110 if (LastLoadedElt == NumElems - 1) {
4111 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4112 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4113 LDBase->getPointerInfo(),
4114 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4115 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4116 LDBase->getPointerInfo(),
4117 LDBase->isVolatile(), LDBase->isNonTemporal(),
4118 LDBase->getAlignment());
4119 } else if (NumElems == 4 && LastLoadedElt == 1) {
4120 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4121 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4122 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4124 LDBase->getMemOperand());
4125 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4131 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4132 DebugLoc dl = Op.getDebugLoc();
4133 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
4134 // All one's are handled with pcmpeqd. In AVX, zero's are handled with
4135 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
4136 // is present, so AllOnes is ignored.
4137 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4138 (Op.getValueType().getSizeInBits() != 256 &&
4139 ISD::isBuildVectorAllOnes(Op.getNode()))) {
4140 // Canonicalize this to <4 x i32> (SSE) to
4141 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4142 // eliminated on x86-32 hosts.
4143 if (Op.getValueType() == MVT::v4i32)
4146 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4147 return getOnesVector(Op.getValueType(), DAG, dl);
4148 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4151 EVT VT = Op.getValueType();
4152 EVT ExtVT = VT.getVectorElementType();
4153 unsigned EVTBits = ExtVT.getSizeInBits();
4155 unsigned NumElems = Op.getNumOperands();
4156 unsigned NumZero = 0;
4157 unsigned NumNonZero = 0;
4158 unsigned NonZeros = 0;
4159 bool IsAllConstants = true;
4160 SmallSet<SDValue, 8> Values;
4161 for (unsigned i = 0; i < NumElems; ++i) {
4162 SDValue Elt = Op.getOperand(i);
4163 if (Elt.getOpcode() == ISD::UNDEF)
4166 if (Elt.getOpcode() != ISD::Constant &&
4167 Elt.getOpcode() != ISD::ConstantFP)
4168 IsAllConstants = false;
4169 if (X86::isZeroNode(Elt))
4172 NonZeros |= (1 << i);
4177 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4178 if (NumNonZero == 0)
4179 return DAG.getUNDEF(VT);
4181 // Special case for single non-zero, non-undef, element.
4182 if (NumNonZero == 1) {
4183 unsigned Idx = CountTrailingZeros_32(NonZeros);
4184 SDValue Item = Op.getOperand(Idx);
4186 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4187 // the value are obviously zero, truncate the value to i32 and do the
4188 // insertion that way. Only do this if the value is non-constant or if the
4189 // value is a constant being inserted into element 0. It is cheaper to do
4190 // a constant pool load than it is to do a movd + shuffle.
4191 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4192 (!IsAllConstants || Idx == 0)) {
4193 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4195 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
4196 EVT VecVT = MVT::v4i32;
4197 unsigned VecElts = 4;
4199 // Truncate the value (which may itself be a constant) to i32, and
4200 // convert it to a vector with movd (S2V+shuffle to zero extend).
4201 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4202 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4203 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4204 Subtarget->hasSSE2(), DAG);
4206 // Now we have our 32-bit value zero extended in the low element of
4207 // a vector. If Idx != 0, swizzle it into place.
4209 SmallVector<int, 4> Mask;
4210 Mask.push_back(Idx);
4211 for (unsigned i = 1; i != VecElts; ++i)
4213 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4214 DAG.getUNDEF(Item.getValueType()),
4217 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
4221 // If we have a constant or non-constant insertion into the low element of
4222 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4223 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4224 // depending on what the source datatype is.
4227 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4228 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4229 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4230 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4231 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4232 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4234 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4235 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4236 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
4237 EVT MiddleVT = MVT::v4i32;
4238 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4239 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4240 Subtarget->hasSSE2(), DAG);
4241 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
4245 // Is it a vector logical left shift?
4246 if (NumElems == 2 && Idx == 1 &&
4247 X86::isZeroNode(Op.getOperand(0)) &&
4248 !X86::isZeroNode(Op.getOperand(1))) {
4249 unsigned NumBits = VT.getSizeInBits();
4250 return getVShift(true, VT,
4251 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4252 VT, Op.getOperand(1)),
4253 NumBits/2, DAG, *this, dl);
4256 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4259 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4260 // is a non-constant being inserted into an element other than the low one,
4261 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4262 // movd/movss) to move this into the low element, then shuffle it into
4264 if (EVTBits == 32) {
4265 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4267 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4268 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4269 Subtarget->hasSSE2(), DAG);
4270 SmallVector<int, 8> MaskVec;
4271 for (unsigned i = 0; i < NumElems; i++)
4272 MaskVec.push_back(i == Idx ? 0 : 1);
4273 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4277 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4278 if (Values.size() == 1) {
4279 if (EVTBits == 32) {
4280 // Instead of a shuffle like this:
4281 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4282 // Check if it's possible to issue this instead.
4283 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4284 unsigned Idx = CountTrailingZeros_32(NonZeros);
4285 SDValue Item = Op.getOperand(Idx);
4286 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4287 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4292 // A vector full of immediates; various special cases are already
4293 // handled, so this is best done with a single constant-pool load.
4297 // Let legalizer expand 2-wide build_vectors.
4298 if (EVTBits == 64) {
4299 if (NumNonZero == 1) {
4300 // One half is zero or undef.
4301 unsigned Idx = CountTrailingZeros_32(NonZeros);
4302 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4303 Op.getOperand(Idx));
4304 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4305 Subtarget->hasSSE2(), DAG);
4310 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4311 if (EVTBits == 8 && NumElems == 16) {
4312 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4314 if (V.getNode()) return V;
4317 if (EVTBits == 16 && NumElems == 8) {
4318 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4320 if (V.getNode()) return V;
4323 // If element VT is == 32 bits, turn it into a number of shuffles.
4324 SmallVector<SDValue, 8> V;
4326 if (NumElems == 4 && NumZero > 0) {
4327 for (unsigned i = 0; i < 4; ++i) {
4328 bool isZero = !(NonZeros & (1 << i));
4330 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4332 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4335 for (unsigned i = 0; i < 2; ++i) {
4336 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4339 V[i] = V[i*2]; // Must be a zero vector.
4342 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4345 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4348 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4353 SmallVector<int, 8> MaskVec;
4354 bool Reverse = (NonZeros & 0x3) == 2;
4355 for (unsigned i = 0; i < 2; ++i)
4356 MaskVec.push_back(Reverse ? 1-i : i);
4357 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4358 for (unsigned i = 0; i < 2; ++i)
4359 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4360 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4363 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4364 // Check for a build vector of consecutive loads.
4365 for (unsigned i = 0; i < NumElems; ++i)
4366 V[i] = Op.getOperand(i);
4368 // Check for elements which are consecutive loads.
4369 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4373 // For SSE 4.1, use insertps to put the high elements into the low element.
4374 if (getSubtarget()->hasSSE41()) {
4376 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4377 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4379 Result = DAG.getUNDEF(VT);
4381 for (unsigned i = 1; i < NumElems; ++i) {
4382 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4383 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4384 Op.getOperand(i), DAG.getIntPtrConstant(i));
4389 // Otherwise, expand into a number of unpckl*, start by extending each of
4390 // our (non-undef) elements to the full vector width with the element in the
4391 // bottom slot of the vector (which generates no code for SSE).
4392 for (unsigned i = 0; i < NumElems; ++i) {
4393 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4394 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4396 V[i] = DAG.getUNDEF(VT);
4399 // Next, we iteratively mix elements, e.g. for v4f32:
4400 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4401 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4402 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4403 unsigned EltStride = NumElems >> 1;
4404 while (EltStride != 0) {
4405 for (unsigned i = 0; i < EltStride; ++i) {
4406 // If V[i+EltStride] is undef and this is the first round of mixing,
4407 // then it is safe to just drop this shuffle: V[i] is already in the
4408 // right place, the one element (since it's the first round) being
4409 // inserted as undef can be dropped. This isn't safe for successive
4410 // rounds because they will permute elements within both vectors.
4411 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4412 EltStride == NumElems/2)
4415 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4425 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4426 // We support concatenate two MMX registers and place them in a MMX
4427 // register. This is better than doing a stack convert.
4428 DebugLoc dl = Op.getDebugLoc();
4429 EVT ResVT = Op.getValueType();
4430 assert(Op.getNumOperands() == 2);
4431 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4432 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4434 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
4435 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4436 InVec = Op.getOperand(1);
4437 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4438 unsigned NumElts = ResVT.getVectorNumElements();
4439 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4440 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4441 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4443 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
4444 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4445 Mask[0] = 0; Mask[1] = 2;
4446 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4448 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4451 // v8i16 shuffles - Prefer shuffles in the following order:
4452 // 1. [all] pshuflw, pshufhw, optional move
4453 // 2. [ssse3] 1 x pshufb
4454 // 3. [ssse3] 2 x pshufb + 1 x por
4455 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4457 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4458 SelectionDAG &DAG) const {
4459 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4460 SDValue V1 = SVOp->getOperand(0);
4461 SDValue V2 = SVOp->getOperand(1);
4462 DebugLoc dl = SVOp->getDebugLoc();
4463 SmallVector<int, 8> MaskVals;
4465 // Determine if more than 1 of the words in each of the low and high quadwords
4466 // of the result come from the same quadword of one of the two inputs. Undef
4467 // mask values count as coming from any quadword, for better codegen.
4468 SmallVector<unsigned, 4> LoQuad(4);
4469 SmallVector<unsigned, 4> HiQuad(4);
4470 BitVector InputQuads(4);
4471 for (unsigned i = 0; i < 8; ++i) {
4472 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4473 int EltIdx = SVOp->getMaskElt(i);
4474 MaskVals.push_back(EltIdx);
4483 InputQuads.set(EltIdx / 4);
4486 int BestLoQuad = -1;
4487 unsigned MaxQuad = 1;
4488 for (unsigned i = 0; i < 4; ++i) {
4489 if (LoQuad[i] > MaxQuad) {
4491 MaxQuad = LoQuad[i];
4495 int BestHiQuad = -1;
4497 for (unsigned i = 0; i < 4; ++i) {
4498 if (HiQuad[i] > MaxQuad) {
4500 MaxQuad = HiQuad[i];
4504 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4505 // of the two input vectors, shuffle them into one input vector so only a
4506 // single pshufb instruction is necessary. If There are more than 2 input
4507 // quads, disable the next transformation since it does not help SSSE3.
4508 bool V1Used = InputQuads[0] || InputQuads[1];
4509 bool V2Used = InputQuads[2] || InputQuads[3];
4510 if (Subtarget->hasSSSE3()) {
4511 if (InputQuads.count() == 2 && V1Used && V2Used) {
4512 BestLoQuad = InputQuads.find_first();
4513 BestHiQuad = InputQuads.find_next(BestLoQuad);
4515 if (InputQuads.count() > 2) {
4521 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4522 // the shuffle mask. If a quad is scored as -1, that means that it contains
4523 // words from all 4 input quadwords.
4525 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4526 SmallVector<int, 8> MaskV;
4527 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4528 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4529 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4530 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
4531 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
4532 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
4534 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4535 // source words for the shuffle, to aid later transformations.
4536 bool AllWordsInNewV = true;
4537 bool InOrder[2] = { true, true };
4538 for (unsigned i = 0; i != 8; ++i) {
4539 int idx = MaskVals[i];
4541 InOrder[i/4] = false;
4542 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4544 AllWordsInNewV = false;
4548 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4549 if (AllWordsInNewV) {
4550 for (int i = 0; i != 8; ++i) {
4551 int idx = MaskVals[i];
4554 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4555 if ((idx != i) && idx < 4)
4557 if ((idx != i) && idx > 3)
4566 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4567 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4568 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4569 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4570 unsigned TargetMask = 0;
4571 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4572 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4573 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4574 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4575 V1 = NewV.getOperand(0);
4576 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4580 // If we have SSSE3, and all words of the result are from 1 input vector,
4581 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4582 // is present, fall back to case 4.
4583 if (Subtarget->hasSSSE3()) {
4584 SmallVector<SDValue,16> pshufbMask;
4586 // If we have elements from both input vectors, set the high bit of the
4587 // shuffle mask element to zero out elements that come from V2 in the V1
4588 // mask, and elements that come from V1 in the V2 mask, so that the two
4589 // results can be OR'd together.
4590 bool TwoInputs = V1Used && V2Used;
4591 for (unsigned i = 0; i != 8; ++i) {
4592 int EltIdx = MaskVals[i] * 2;
4593 if (TwoInputs && (EltIdx >= 16)) {
4594 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4595 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4598 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4599 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4601 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
4602 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4603 DAG.getNode(ISD::BUILD_VECTOR, dl,
4604 MVT::v16i8, &pshufbMask[0], 16));
4606 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
4608 // Calculate the shuffle mask for the second input, shuffle it, and
4609 // OR it with the first shuffled input.
4611 for (unsigned i = 0; i != 8; ++i) {
4612 int EltIdx = MaskVals[i] * 2;
4614 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4615 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4618 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4619 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4621 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
4622 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4623 DAG.getNode(ISD::BUILD_VECTOR, dl,
4624 MVT::v16i8, &pshufbMask[0], 16));
4625 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4626 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
4629 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4630 // and update MaskVals with new element order.
4631 BitVector InOrder(8);
4632 if (BestLoQuad >= 0) {
4633 SmallVector<int, 8> MaskV;
4634 for (int i = 0; i != 4; ++i) {
4635 int idx = MaskVals[i];
4637 MaskV.push_back(-1);
4639 } else if ((idx / 4) == BestLoQuad) {
4640 MaskV.push_back(idx & 3);
4643 MaskV.push_back(-1);
4646 for (unsigned i = 4; i != 8; ++i)
4648 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4651 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4652 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4654 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4658 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4659 // and update MaskVals with the new element order.
4660 if (BestHiQuad >= 0) {
4661 SmallVector<int, 8> MaskV;
4662 for (unsigned i = 0; i != 4; ++i)
4664 for (unsigned i = 4; i != 8; ++i) {
4665 int idx = MaskVals[i];
4667 MaskV.push_back(-1);
4669 } else if ((idx / 4) == BestHiQuad) {
4670 MaskV.push_back((idx & 3) + 4);
4673 MaskV.push_back(-1);
4676 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4679 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4680 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
4682 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
4686 // In case BestHi & BestLo were both -1, which means each quadword has a word
4687 // from each of the four input quadwords, calculate the InOrder bitvector now
4688 // before falling through to the insert/extract cleanup.
4689 if (BestLoQuad == -1 && BestHiQuad == -1) {
4691 for (int i = 0; i != 8; ++i)
4692 if (MaskVals[i] < 0 || MaskVals[i] == i)
4696 // The other elements are put in the right place using pextrw and pinsrw.
4697 for (unsigned i = 0; i != 8; ++i) {
4700 int EltIdx = MaskVals[i];
4703 SDValue ExtOp = (EltIdx < 8)
4704 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4705 DAG.getIntPtrConstant(EltIdx))
4706 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4707 DAG.getIntPtrConstant(EltIdx - 8));
4708 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4709 DAG.getIntPtrConstant(i));
4714 // v16i8 shuffles - Prefer shuffles in the following order:
4715 // 1. [ssse3] 1 x pshufb
4716 // 2. [ssse3] 2 x pshufb + 1 x por
4717 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4719 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4721 const X86TargetLowering &TLI) {
4722 SDValue V1 = SVOp->getOperand(0);
4723 SDValue V2 = SVOp->getOperand(1);
4724 DebugLoc dl = SVOp->getDebugLoc();
4725 SmallVector<int, 16> MaskVals;
4726 SVOp->getMask(MaskVals);
4728 // If we have SSSE3, case 1 is generated when all result bytes come from
4729 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4730 // present, fall back to case 3.
4731 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4734 for (unsigned i = 0; i < 16; ++i) {
4735 int EltIdx = MaskVals[i];
4744 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4745 if (TLI.getSubtarget()->hasSSSE3()) {
4746 SmallVector<SDValue,16> pshufbMask;
4748 // If all result elements are from one input vector, then only translate
4749 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4751 // Otherwise, we have elements from both input vectors, and must zero out
4752 // elements that come from V2 in the first mask, and V1 in the second mask
4753 // so that we can OR them together.
4754 bool TwoInputs = !(V1Only || V2Only);
4755 for (unsigned i = 0; i != 16; ++i) {
4756 int EltIdx = MaskVals[i];
4757 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4758 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4761 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4763 // If all the elements are from V2, assign it to V1 and return after
4764 // building the first pshufb.
4767 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4768 DAG.getNode(ISD::BUILD_VECTOR, dl,
4769 MVT::v16i8, &pshufbMask[0], 16));
4773 // Calculate the shuffle mask for the second input, shuffle it, and
4774 // OR it with the first shuffled input.
4776 for (unsigned i = 0; i != 16; ++i) {
4777 int EltIdx = MaskVals[i];
4779 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4782 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4784 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4785 DAG.getNode(ISD::BUILD_VECTOR, dl,
4786 MVT::v16i8, &pshufbMask[0], 16));
4787 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4790 // No SSSE3 - Calculate in place words and then fix all out of place words
4791 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4792 // the 16 different words that comprise the two doublequadword input vectors.
4793 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
4794 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
4795 SDValue NewV = V2Only ? V2 : V1;
4796 for (int i = 0; i != 8; ++i) {
4797 int Elt0 = MaskVals[i*2];
4798 int Elt1 = MaskVals[i*2+1];
4800 // This word of the result is all undef, skip it.
4801 if (Elt0 < 0 && Elt1 < 0)
4804 // This word of the result is already in the correct place, skip it.
4805 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4807 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4810 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4811 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4814 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4815 // using a single extract together, load it and store it.
4816 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4817 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4818 DAG.getIntPtrConstant(Elt1 / 2));
4819 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4820 DAG.getIntPtrConstant(i));
4824 // If Elt1 is defined, extract it from the appropriate source. If the
4825 // source byte is not also odd, shift the extracted word left 8 bits
4826 // otherwise clear the bottom 8 bits if we need to do an or.
4828 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4829 DAG.getIntPtrConstant(Elt1 / 2));
4830 if ((Elt1 & 1) == 0)
4831 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4832 DAG.getConstant(8, TLI.getShiftAmountTy()));
4834 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4835 DAG.getConstant(0xFF00, MVT::i16));
4837 // If Elt0 is defined, extract it from the appropriate source. If the
4838 // source byte is not also even, shift the extracted word right 8 bits. If
4839 // Elt1 was also defined, OR the extracted values together before
4840 // inserting them in the result.
4842 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4843 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4844 if ((Elt0 & 1) != 0)
4845 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4846 DAG.getConstant(8, TLI.getShiftAmountTy()));
4848 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4849 DAG.getConstant(0x00FF, MVT::i16));
4850 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4853 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4854 DAG.getIntPtrConstant(i));
4856 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
4859 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4860 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
4861 /// done when every pair / quad of shuffle mask elements point to elements in
4862 /// the right sequence. e.g.
4863 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
4865 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4866 SelectionDAG &DAG, DebugLoc dl) {
4867 EVT VT = SVOp->getValueType(0);
4868 SDValue V1 = SVOp->getOperand(0);
4869 SDValue V2 = SVOp->getOperand(1);
4870 unsigned NumElems = VT.getVectorNumElements();
4871 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4873 switch (VT.getSimpleVT().SimpleTy) {
4874 default: assert(false && "Unexpected!");
4875 case MVT::v4f32: NewVT = MVT::v2f64; break;
4876 case MVT::v4i32: NewVT = MVT::v2i64; break;
4877 case MVT::v8i16: NewVT = MVT::v4i32; break;
4878 case MVT::v16i8: NewVT = MVT::v4i32; break;
4881 int Scale = NumElems / NewWidth;
4882 SmallVector<int, 8> MaskVec;
4883 for (unsigned i = 0; i < NumElems; i += Scale) {
4885 for (int j = 0; j < Scale; ++j) {
4886 int EltIdx = SVOp->getMaskElt(i+j);
4890 StartIdx = EltIdx - (EltIdx % Scale);
4891 if (EltIdx != StartIdx + j)
4895 MaskVec.push_back(-1);
4897 MaskVec.push_back(StartIdx / Scale);
4900 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
4901 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
4902 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4905 /// getVZextMovL - Return a zero-extending vector move low node.
4907 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4908 SDValue SrcOp, SelectionDAG &DAG,
4909 const X86Subtarget *Subtarget, DebugLoc dl) {
4910 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4911 LoadSDNode *LD = NULL;
4912 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4913 LD = dyn_cast<LoadSDNode>(SrcOp);
4915 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4917 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4918 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
4919 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4920 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
4921 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4923 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4924 return DAG.getNode(ISD::BITCAST, dl, VT,
4925 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4926 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4934 return DAG.getNode(ISD::BITCAST, dl, VT,
4935 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4936 DAG.getNode(ISD::BITCAST, dl,
4940 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4943 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4944 SDValue V1 = SVOp->getOperand(0);
4945 SDValue V2 = SVOp->getOperand(1);
4946 DebugLoc dl = SVOp->getDebugLoc();
4947 EVT VT = SVOp->getValueType(0);
4949 SmallVector<std::pair<int, int>, 8> Locs;
4951 SmallVector<int, 8> Mask1(4U, -1);
4952 SmallVector<int, 8> PermMask;
4953 SVOp->getMask(PermMask);
4957 for (unsigned i = 0; i != 4; ++i) {
4958 int Idx = PermMask[i];
4960 Locs[i] = std::make_pair(-1, -1);
4962 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4964 Locs[i] = std::make_pair(0, NumLo);
4968 Locs[i] = std::make_pair(1, NumHi);
4970 Mask1[2+NumHi] = Idx;
4976 if (NumLo <= 2 && NumHi <= 2) {
4977 // If no more than two elements come from either vector. This can be
4978 // implemented with two shuffles. First shuffle gather the elements.
4979 // The second shuffle, which takes the first shuffle as both of its
4980 // vector operands, put the elements into the right order.
4981 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4983 SmallVector<int, 8> Mask2(4U, -1);
4985 for (unsigned i = 0; i != 4; ++i) {
4986 if (Locs[i].first == -1)
4989 unsigned Idx = (i < 2) ? 0 : 4;
4990 Idx += Locs[i].first * 2 + Locs[i].second;
4995 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4996 } else if (NumLo == 3 || NumHi == 3) {
4997 // Otherwise, we must have three elements from one vector, call it X, and
4998 // one element from the other, call it Y. First, use a shufps to build an
4999 // intermediate vector with the one element from Y and the element from X
5000 // that will be in the same half in the final destination (the indexes don't
5001 // matter). Then, use a shufps to build the final vector, taking the half
5002 // containing the element from Y from the intermediate, and the other half
5005 // Normalize it so the 3 elements come from V1.
5006 CommuteVectorShuffleMask(PermMask, VT);
5010 // Find the element from V2.
5012 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5013 int Val = PermMask[HiIndex];
5020 Mask1[0] = PermMask[HiIndex];
5022 Mask1[2] = PermMask[HiIndex^1];
5024 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5027 Mask1[0] = PermMask[0];
5028 Mask1[1] = PermMask[1];
5029 Mask1[2] = HiIndex & 1 ? 6 : 4;
5030 Mask1[3] = HiIndex & 1 ? 4 : 6;
5031 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5033 Mask1[0] = HiIndex & 1 ? 2 : 0;
5034 Mask1[1] = HiIndex & 1 ? 0 : 2;
5035 Mask1[2] = PermMask[2];
5036 Mask1[3] = PermMask[3];
5041 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5045 // Break it into (shuffle shuffle_hi, shuffle_lo).
5047 SmallVector<int,8> LoMask(4U, -1);
5048 SmallVector<int,8> HiMask(4U, -1);
5050 SmallVector<int,8> *MaskPtr = &LoMask;
5051 unsigned MaskIdx = 0;
5054 for (unsigned i = 0; i != 4; ++i) {
5061 int Idx = PermMask[i];
5063 Locs[i] = std::make_pair(-1, -1);
5064 } else if (Idx < 4) {
5065 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5066 (*MaskPtr)[LoIdx] = Idx;
5069 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5070 (*MaskPtr)[HiIdx] = Idx;
5075 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5076 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5077 SmallVector<int, 8> MaskOps;
5078 for (unsigned i = 0; i != 4; ++i) {
5079 if (Locs[i].first == -1) {
5080 MaskOps.push_back(-1);
5082 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5083 MaskOps.push_back(Idx);
5086 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5089 static bool MayFoldVectorLoad(SDValue V) {
5090 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5091 V = V.getOperand(0);
5092 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5093 V = V.getOperand(0);
5099 // FIXME: the version above should always be used. Since there's
5100 // a bug where several vector shuffles can't be folded because the
5101 // DAG is not updated during lowering and a node claims to have two
5102 // uses while it only has one, use this version, and let isel match
5103 // another instruction if the load really happens to have more than
5104 // one use. Remove this version after this bug get fixed.
5105 // rdar://8434668, PR8156
5106 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5107 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5108 V = V.getOperand(0);
5109 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5110 V = V.getOperand(0);
5111 if (ISD::isNormalLoad(V.getNode()))
5116 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5117 /// a vector extract, and if both can be later optimized into a single load.
5118 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5119 /// here because otherwise a target specific shuffle node is going to be
5120 /// emitted for this shuffle, and the optimization not done.
5121 /// FIXME: This is probably not the best approach, but fix the problem
5122 /// until the right path is decided.
5124 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5125 const TargetLowering &TLI) {
5126 EVT VT = V.getValueType();
5127 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5129 // Be sure that the vector shuffle is present in a pattern like this:
5130 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5134 SDNode *N = *V.getNode()->use_begin();
5135 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5138 SDValue EltNo = N->getOperand(1);
5139 if (!isa<ConstantSDNode>(EltNo))
5142 // If the bit convert changed the number of elements, it is unsafe
5143 // to examine the mask.
5144 bool HasShuffleIntoBitcast = false;
5145 if (V.getOpcode() == ISD::BITCAST) {
5146 EVT SrcVT = V.getOperand(0).getValueType();
5147 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5149 V = V.getOperand(0);
5150 HasShuffleIntoBitcast = true;
5153 // Select the input vector, guarding against out of range extract vector.
5154 unsigned NumElems = VT.getVectorNumElements();
5155 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5156 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5157 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5159 // Skip one more bit_convert if necessary
5160 if (V.getOpcode() == ISD::BITCAST)
5161 V = V.getOperand(0);
5163 if (ISD::isNormalLoad(V.getNode())) {
5164 // Is the original load suitable?
5165 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5167 // FIXME: avoid the multi-use bug that is preventing lots of
5168 // of foldings to be detected, this is still wrong of course, but
5169 // give the temporary desired behavior, and if it happens that
5170 // the load has real more uses, during isel it will not fold, and
5171 // will generate poor code.
5172 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5175 if (!HasShuffleIntoBitcast)
5178 // If there's a bitcast before the shuffle, check if the load type and
5179 // alignment is valid.
5180 unsigned Align = LN0->getAlignment();
5182 TLI.getTargetData()->getABITypeAlignment(
5183 VT.getTypeForEVT(*DAG.getContext()));
5185 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
5193 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
5194 EVT VT = Op.getValueType();
5196 // Canonizalize to v2f64.
5197 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
5198 return DAG.getNode(ISD::BITCAST, dl, VT,
5199 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
5204 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5206 SDValue V1 = Op.getOperand(0);
5207 SDValue V2 = Op.getOperand(1);
5208 EVT VT = Op.getValueType();
5210 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5212 if (HasSSE2 && VT == MVT::v2f64)
5213 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5216 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5220 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5221 SDValue V1 = Op.getOperand(0);
5222 SDValue V2 = Op.getOperand(1);
5223 EVT VT = Op.getValueType();
5225 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5226 "unsupported shuffle type");
5228 if (V2.getOpcode() == ISD::UNDEF)
5232 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5236 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5237 SDValue V1 = Op.getOperand(0);
5238 SDValue V2 = Op.getOperand(1);
5239 EVT VT = Op.getValueType();
5240 unsigned NumElems = VT.getVectorNumElements();
5242 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5243 // operand of these instructions is only memory, so check if there's a
5244 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5246 bool CanFoldLoad = false;
5248 // Trivial case, when V2 comes from a load.
5249 if (MayFoldVectorLoad(V2))
5252 // When V1 is a load, it can be folded later into a store in isel, example:
5253 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5255 // (MOVLPSmr addr:$src1, VR128:$src2)
5256 // So, recognize this potential and also use MOVLPS or MOVLPD
5257 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5261 if (HasSSE2 && NumElems == 2)
5262 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5265 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5268 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5269 // movl and movlp will both match v2i64, but v2i64 is never matched by
5270 // movl earlier because we make it strict to avoid messing with the movlp load
5271 // folding logic (see the code above getMOVLP call). Match it here then,
5272 // this is horrible, but will stay like this until we move all shuffle
5273 // matching to x86 specific nodes. Note that for the 1st condition all
5274 // types are matched with movsd.
5275 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5276 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5278 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5281 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5283 // Invert the operand order and use SHUFPS to match it.
5284 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5285 X86::getShuffleSHUFImmediate(SVOp), DAG);
5288 static inline unsigned getUNPCKLOpcode(EVT VT) {
5289 switch(VT.getSimpleVT().SimpleTy) {
5290 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5291 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5292 case MVT::v4f32: return X86ISD::UNPCKLPS;
5293 case MVT::v2f64: return X86ISD::UNPCKLPD;
5294 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5295 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5297 llvm_unreachable("Unknow type for unpckl");
5302 static inline unsigned getUNPCKHOpcode(EVT VT) {
5303 switch(VT.getSimpleVT().SimpleTy) {
5304 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5305 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5306 case MVT::v4f32: return X86ISD::UNPCKHPS;
5307 case MVT::v2f64: return X86ISD::UNPCKHPD;
5308 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5309 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5311 llvm_unreachable("Unknow type for unpckh");
5317 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
5318 const TargetLowering &TLI,
5319 const X86Subtarget *Subtarget) {
5320 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5321 EVT VT = Op.getValueType();
5322 DebugLoc dl = Op.getDebugLoc();
5323 SDValue V1 = Op.getOperand(0);
5324 SDValue V2 = Op.getOperand(1);
5326 if (isZeroShuffle(SVOp))
5327 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5329 // Handle splat operations
5330 if (SVOp->isSplat()) {
5331 // Special case, this is the only place now where it's
5332 // allowed to return a vector_shuffle operation without
5333 // using a target specific node, because *hopefully* it
5334 // will be optimized away by the dag combiner.
5335 if (VT.getVectorNumElements() <= 4 &&
5336 CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
5339 // Handle splats by matching through known masks
5340 if (VT.getVectorNumElements() <= 4)
5343 // Canonicalize all of the remaining to v4f32.
5344 return PromoteSplat(SVOp, DAG);
5347 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5349 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5350 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5351 if (NewOp.getNode())
5352 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
5353 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5354 // FIXME: Figure out a cleaner way to do this.
5355 // Try to make use of movq to zero out the top part.
5356 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5357 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5358 if (NewOp.getNode()) {
5359 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5360 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5361 DAG, Subtarget, dl);
5363 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5364 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5365 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5366 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5367 DAG, Subtarget, dl);
5374 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5375 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5376 SDValue V1 = Op.getOperand(0);
5377 SDValue V2 = Op.getOperand(1);
5378 EVT VT = Op.getValueType();
5379 DebugLoc dl = Op.getDebugLoc();
5380 unsigned NumElems = VT.getVectorNumElements();
5381 bool isMMX = VT.getSizeInBits() == 64;
5382 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5383 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5384 bool V1IsSplat = false;
5385 bool V2IsSplat = false;
5386 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5387 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5388 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
5389 MachineFunction &MF = DAG.getMachineFunction();
5390 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5392 // Shuffle operations on MMX not supported.
5396 // Vector shuffle lowering takes 3 steps:
5398 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
5399 // narrowing and commutation of operands should be handled.
5400 // 2) Matching of shuffles with known shuffle masks to x86 target specific
5402 // 3) Rewriting of unmatched masks into new generic shuffle operations,
5403 // so the shuffle can be broken into other shuffles and the legalizer can
5404 // try the lowering again.
5406 // The general ideia is that no vector_shuffle operation should be left to
5407 // be matched during isel, all of them must be converted to a target specific
5410 // Normalize the input vectors. Here splats, zeroed vectors, profitable
5411 // narrowing and commutation of operands should be handled. The actual code
5412 // doesn't include all of those, work in progress...
5413 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
5414 if (NewOp.getNode())
5417 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
5418 // unpckh_undef). Only use pshufd if speed is more important than size.
5419 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
5420 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5421 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5422 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
5423 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5424 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5426 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
5427 RelaxedMayFoldVectorLoad(V1))
5428 return getMOVDDup(Op, dl, V1, DAG);
5430 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
5431 return getMOVHighToLow(Op, dl, DAG);
5433 // Use to match splats
5434 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
5435 (VT == MVT::v2f64 || VT == MVT::v2i64))
5436 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5438 if (X86::isPSHUFDMask(SVOp)) {
5439 // The actual implementation will match the mask in the if above and then
5440 // during isel it can match several different instructions, not only pshufd
5441 // as its name says, sad but true, emulate the behavior for now...
5442 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5443 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5445 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5447 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5448 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5450 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5451 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5454 if (VT == MVT::v4f32)
5455 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5459 // Check if this can be converted into a logical shift.
5460 bool isLeft = false;
5463 bool isShift = getSubtarget()->hasSSE2() &&
5464 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5465 if (isShift && ShVal.hasOneUse()) {
5466 // If the shifted value has multiple uses, it may be cheaper to use
5467 // v_set0 + movlhps or movhlps, etc.
5468 EVT EltVT = VT.getVectorElementType();
5469 ShAmt *= EltVT.getSizeInBits();
5470 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5473 if (X86::isMOVLMask(SVOp)) {
5476 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5477 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5478 if (!X86::isMOVLPMask(SVOp)) {
5479 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5480 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5482 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5483 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5487 // FIXME: fold these into legal mask.
5488 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5489 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5491 if (X86::isMOVHLPSMask(SVOp))
5492 return getMOVHighToLow(Op, dl, DAG);
5494 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5495 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5497 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5498 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5500 if (X86::isMOVLPMask(SVOp))
5501 return getMOVLP(Op, dl, DAG, HasSSE2);
5503 if (ShouldXformToMOVHLPS(SVOp) ||
5504 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5505 return CommuteVectorShuffle(SVOp, DAG);
5508 // No better options. Use a vshl / vsrl.
5509 EVT EltVT = VT.getVectorElementType();
5510 ShAmt *= EltVT.getSizeInBits();
5511 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5514 bool Commuted = false;
5515 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5516 // 1,1,1,1 -> v8i16 though.
5517 V1IsSplat = isSplatVector(V1.getNode());
5518 V2IsSplat = isSplatVector(V2.getNode());
5520 // Canonicalize the splat or undef, if present, to be on the RHS.
5521 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5522 Op = CommuteVectorShuffle(SVOp, DAG);
5523 SVOp = cast<ShuffleVectorSDNode>(Op);
5524 V1 = SVOp->getOperand(0);
5525 V2 = SVOp->getOperand(1);
5526 std::swap(V1IsSplat, V2IsSplat);
5527 std::swap(V1IsUndef, V2IsUndef);
5531 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5532 // Shuffling low element of v1 into undef, just return v1.
5535 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5536 // the instruction selector will not match, so get a canonical MOVL with
5537 // swapped operands to undo the commute.
5538 return getMOVL(DAG, dl, VT, V2, V1);
5541 if (X86::isUNPCKLMask(SVOp))
5542 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
5544 if (X86::isUNPCKHMask(SVOp))
5545 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
5548 // Normalize mask so all entries that point to V2 points to its first
5549 // element then try to match unpck{h|l} again. If match, return a
5550 // new vector_shuffle with the corrected mask.
5551 SDValue NewMask = NormalizeMask(SVOp, DAG);
5552 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5553 if (NSVOp != SVOp) {
5554 if (X86::isUNPCKLMask(NSVOp, true)) {
5556 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5563 // Commute is back and try unpck* again.
5564 // FIXME: this seems wrong.
5565 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5566 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5568 if (X86::isUNPCKLMask(NewSVOp))
5569 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
5571 if (X86::isUNPCKHMask(NewSVOp))
5572 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
5575 // Normalize the node to match x86 shuffle ops if needed
5576 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5577 return CommuteVectorShuffle(SVOp, DAG);
5579 // The checks below are all present in isShuffleMaskLegal, but they are
5580 // inlined here right now to enable us to directly emit target specific
5581 // nodes, and remove one by one until they don't return Op anymore.
5582 SmallVector<int, 16> M;
5585 if (isPALIGNRMask(M, VT, HasSSSE3))
5586 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
5587 X86::getShufflePALIGNRImmediate(SVOp),
5590 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
5591 SVOp->getSplatIndex() == 0 && V2IsUndef) {
5592 if (VT == MVT::v2f64)
5593 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
5594 if (VT == MVT::v2i64)
5595 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
5598 if (isPSHUFHWMask(M, VT))
5599 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
5600 X86::getShufflePSHUFHWImmediate(SVOp),
5603 if (isPSHUFLWMask(M, VT))
5604 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
5605 X86::getShufflePSHUFLWImmediate(SVOp),
5608 if (isSHUFPMask(M, VT)) {
5609 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5610 if (VT == MVT::v4f32 || VT == MVT::v4i32)
5611 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
5613 if (VT == MVT::v2f64 || VT == MVT::v2i64)
5614 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
5618 if (X86::isUNPCKL_v_undef_Mask(SVOp))
5619 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5620 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5621 if (X86::isUNPCKH_v_undef_Mask(SVOp))
5622 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5623 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5625 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5626 if (VT == MVT::v8i16) {
5627 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5628 if (NewOp.getNode())
5632 if (VT == MVT::v16i8) {
5633 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5634 if (NewOp.getNode())
5638 // Handle all 4 wide cases with a number of shuffles.
5640 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5646 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5647 SelectionDAG &DAG) const {
5648 EVT VT = Op.getValueType();
5649 DebugLoc dl = Op.getDebugLoc();
5650 if (VT.getSizeInBits() == 8) {
5651 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
5652 Op.getOperand(0), Op.getOperand(1));
5653 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5654 DAG.getValueType(VT));
5655 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5656 } else if (VT.getSizeInBits() == 16) {
5657 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5658 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
5660 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5661 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5662 DAG.getNode(ISD::BITCAST, dl,
5666 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
5667 Op.getOperand(0), Op.getOperand(1));
5668 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5669 DAG.getValueType(VT));
5670 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5671 } else if (VT == MVT::f32) {
5672 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
5673 // the result back to FR32 register. It's only worth matching if the
5674 // result has a single use which is a store or a bitcast to i32. And in
5675 // the case of a store, it's not worth it if the index is a constant 0,
5676 // because a MOVSSmr can be used instead, which is smaller and faster.
5677 if (!Op.hasOneUse())
5679 SDNode *User = *Op.getNode()->use_begin();
5680 if ((User->getOpcode() != ISD::STORE ||
5681 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5682 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5683 (User->getOpcode() != ISD::BITCAST ||
5684 User->getValueType(0) != MVT::i32))
5686 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5687 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
5690 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
5691 } else if (VT == MVT::i32) {
5692 // ExtractPS works with constant index.
5693 if (isa<ConstantSDNode>(Op.getOperand(1)))
5701 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5702 SelectionDAG &DAG) const {
5703 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5706 if (Subtarget->hasSSE41()) {
5707 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5712 EVT VT = Op.getValueType();
5713 DebugLoc dl = Op.getDebugLoc();
5714 // TODO: handle v16i8.
5715 if (VT.getSizeInBits() == 16) {
5716 SDValue Vec = Op.getOperand(0);
5717 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5719 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5720 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5721 DAG.getNode(ISD::BITCAST, dl,
5724 // Transform it so it match pextrw which produces a 32-bit result.
5725 EVT EltVT = MVT::i32;
5726 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5727 Op.getOperand(0), Op.getOperand(1));
5728 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5729 DAG.getValueType(VT));
5730 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5731 } else if (VT.getSizeInBits() == 32) {
5732 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5736 // SHUFPS the element to the lowest double word, then movss.
5737 int Mask[4] = { Idx, -1, -1, -1 };
5738 EVT VVT = Op.getOperand(0).getValueType();
5739 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5740 DAG.getUNDEF(VVT), Mask);
5741 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5742 DAG.getIntPtrConstant(0));
5743 } else if (VT.getSizeInBits() == 64) {
5744 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5745 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5746 // to match extract_elt for f64.
5747 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5751 // UNPCKHPD the element to the lowest double word, then movsd.
5752 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5753 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5754 int Mask[2] = { 1, -1 };
5755 EVT VVT = Op.getOperand(0).getValueType();
5756 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5757 DAG.getUNDEF(VVT), Mask);
5758 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5759 DAG.getIntPtrConstant(0));
5766 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5767 SelectionDAG &DAG) const {
5768 EVT VT = Op.getValueType();
5769 EVT EltVT = VT.getVectorElementType();
5770 DebugLoc dl = Op.getDebugLoc();
5772 SDValue N0 = Op.getOperand(0);
5773 SDValue N1 = Op.getOperand(1);
5774 SDValue N2 = Op.getOperand(2);
5776 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5777 isa<ConstantSDNode>(N2)) {
5779 if (VT == MVT::v8i16)
5780 Opc = X86ISD::PINSRW;
5781 else if (VT == MVT::v16i8)
5782 Opc = X86ISD::PINSRB;
5784 Opc = X86ISD::PINSRB;
5786 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5788 if (N1.getValueType() != MVT::i32)
5789 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5790 if (N2.getValueType() != MVT::i32)
5791 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5792 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5793 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5794 // Bits [7:6] of the constant are the source select. This will always be
5795 // zero here. The DAG Combiner may combine an extract_elt index into these
5796 // bits. For example (insert (extract, 3), 2) could be matched by putting
5797 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5798 // Bits [5:4] of the constant are the destination select. This is the
5799 // value of the incoming immediate.
5800 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5801 // combine either bitwise AND or insert of float 0.0 to set these bits.
5802 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5803 // Create this as a scalar to vector..
5804 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5805 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5806 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5807 // PINSR* works with constant index.
5814 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5815 EVT VT = Op.getValueType();
5816 EVT EltVT = VT.getVectorElementType();
5818 if (Subtarget->hasSSE41())
5819 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5821 if (EltVT == MVT::i8)
5824 DebugLoc dl = Op.getDebugLoc();
5825 SDValue N0 = Op.getOperand(0);
5826 SDValue N1 = Op.getOperand(1);
5827 SDValue N2 = Op.getOperand(2);
5829 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5830 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5831 // as its second argument.
5832 if (N1.getValueType() != MVT::i32)
5833 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5834 if (N2.getValueType() != MVT::i32)
5835 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5836 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
5842 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5843 DebugLoc dl = Op.getDebugLoc();
5845 if (Op.getValueType() == MVT::v1i64 &&
5846 Op.getOperand(0).getValueType() == MVT::i64)
5847 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5849 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5850 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
5851 "Expected an SSE type!");
5852 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
5853 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
5856 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5857 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5858 // one of the above mentioned nodes. It has to be wrapped because otherwise
5859 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5860 // be used to form addressing mode. These wrapped nodes will be selected
5863 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5864 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5866 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5868 unsigned char OpFlag = 0;
5869 unsigned WrapperKind = X86ISD::Wrapper;
5870 CodeModel::Model M = getTargetMachine().getCodeModel();
5872 if (Subtarget->isPICStyleRIPRel() &&
5873 (M == CodeModel::Small || M == CodeModel::Kernel))
5874 WrapperKind = X86ISD::WrapperRIP;
5875 else if (Subtarget->isPICStyleGOT())
5876 OpFlag = X86II::MO_GOTOFF;
5877 else if (Subtarget->isPICStyleStubPIC())
5878 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5880 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5882 CP->getOffset(), OpFlag);
5883 DebugLoc DL = CP->getDebugLoc();
5884 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5885 // With PIC, the address is actually $g + Offset.
5887 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5888 DAG.getNode(X86ISD::GlobalBaseReg,
5889 DebugLoc(), getPointerTy()),
5896 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5897 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5899 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5901 unsigned char OpFlag = 0;
5902 unsigned WrapperKind = X86ISD::Wrapper;
5903 CodeModel::Model M = getTargetMachine().getCodeModel();
5905 if (Subtarget->isPICStyleRIPRel() &&
5906 (M == CodeModel::Small || M == CodeModel::Kernel))
5907 WrapperKind = X86ISD::WrapperRIP;
5908 else if (Subtarget->isPICStyleGOT())
5909 OpFlag = X86II::MO_GOTOFF;
5910 else if (Subtarget->isPICStyleStubPIC())
5911 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5913 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5915 DebugLoc DL = JT->getDebugLoc();
5916 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5918 // With PIC, the address is actually $g + Offset.
5920 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5921 DAG.getNode(X86ISD::GlobalBaseReg,
5922 DebugLoc(), getPointerTy()),
5929 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5930 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5932 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5934 unsigned char OpFlag = 0;
5935 unsigned WrapperKind = X86ISD::Wrapper;
5936 CodeModel::Model M = getTargetMachine().getCodeModel();
5938 if (Subtarget->isPICStyleRIPRel() &&
5939 (M == CodeModel::Small || M == CodeModel::Kernel))
5940 WrapperKind = X86ISD::WrapperRIP;
5941 else if (Subtarget->isPICStyleGOT())
5942 OpFlag = X86II::MO_GOTOFF;
5943 else if (Subtarget->isPICStyleStubPIC())
5944 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5946 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5948 DebugLoc DL = Op.getDebugLoc();
5949 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5952 // With PIC, the address is actually $g + Offset.
5953 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5954 !Subtarget->is64Bit()) {
5955 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5956 DAG.getNode(X86ISD::GlobalBaseReg,
5957 DebugLoc(), getPointerTy()),
5965 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5966 // Create the TargetBlockAddressAddress node.
5967 unsigned char OpFlags =
5968 Subtarget->ClassifyBlockAddressReference();
5969 CodeModel::Model M = getTargetMachine().getCodeModel();
5970 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5971 DebugLoc dl = Op.getDebugLoc();
5972 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5973 /*isTarget=*/true, OpFlags);
5975 if (Subtarget->isPICStyleRIPRel() &&
5976 (M == CodeModel::Small || M == CodeModel::Kernel))
5977 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5979 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5981 // With PIC, the address is actually $g + Offset.
5982 if (isGlobalRelativeToPICBase(OpFlags)) {
5983 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5984 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5992 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5994 SelectionDAG &DAG) const {
5995 // Create the TargetGlobalAddress node, folding in the constant
5996 // offset if it is legal.
5997 unsigned char OpFlags =
5998 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5999 CodeModel::Model M = getTargetMachine().getCodeModel();
6001 if (OpFlags == X86II::MO_NO_FLAG &&
6002 X86::isOffsetSuitableForCodeModel(Offset, M)) {
6003 // A direct static reference to a global.
6004 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
6007 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6010 if (Subtarget->isPICStyleRIPRel() &&
6011 (M == CodeModel::Small || M == CodeModel::Kernel))
6012 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6014 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6016 // With PIC, the address is actually $g + Offset.
6017 if (isGlobalRelativeToPICBase(OpFlags)) {
6018 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6019 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6023 // For globals that require a load from a stub to get the address, emit the
6025 if (isGlobalStubReference(OpFlags))
6026 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
6027 MachinePointerInfo::getGOT(), false, false, 0);
6029 // If there was a non-zero offset that we didn't fold, create an explicit
6032 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
6033 DAG.getConstant(Offset, getPointerTy()));
6039 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
6040 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
6041 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
6042 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
6046 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
6047 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
6048 unsigned char OperandFlags) {
6049 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6050 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6051 DebugLoc dl = GA->getDebugLoc();
6052 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6053 GA->getValueType(0),
6057 SDValue Ops[] = { Chain, TGA, *InFlag };
6058 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
6060 SDValue Ops[] = { Chain, TGA };
6061 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
6064 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
6065 MFI->setAdjustsStack(true);
6067 SDValue Flag = Chain.getValue(1);
6068 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
6071 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
6073 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6076 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
6077 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
6078 DAG.getNode(X86ISD::GlobalBaseReg,
6079 DebugLoc(), PtrVT), InFlag);
6080 InFlag = Chain.getValue(1);
6082 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
6085 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
6087 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6089 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
6090 X86::RAX, X86II::MO_TLSGD);
6093 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
6094 // "local exec" model.
6095 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6096 const EVT PtrVT, TLSModel::Model model,
6098 DebugLoc dl = GA->getDebugLoc();
6100 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
6101 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
6102 is64Bit ? 257 : 256));
6104 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
6105 DAG.getIntPtrConstant(0),
6106 MachinePointerInfo(Ptr), false, false, 0);
6108 unsigned char OperandFlags = 0;
6109 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6111 unsigned WrapperKind = X86ISD::Wrapper;
6112 if (model == TLSModel::LocalExec) {
6113 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6114 } else if (is64Bit) {
6115 assert(model == TLSModel::InitialExec);
6116 OperandFlags = X86II::MO_GOTTPOFF;
6117 WrapperKind = X86ISD::WrapperRIP;
6119 assert(model == TLSModel::InitialExec);
6120 OperandFlags = X86II::MO_INDNTPOFF;
6123 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6125 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6126 GA->getValueType(0),
6127 GA->getOffset(), OperandFlags);
6128 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6130 if (model == TLSModel::InitialExec)
6131 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6132 MachinePointerInfo::getGOT(), false, false, 0);
6134 // The address of the thread local variable is the add of the thread
6135 // pointer with the offset of the variable.
6136 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6140 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6142 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6143 const GlobalValue *GV = GA->getGlobal();
6145 if (Subtarget->isTargetELF()) {
6146 // TODO: implement the "local dynamic" model
6147 // TODO: implement the "initial exec"model for pic executables
6149 // If GV is an alias then use the aliasee for determining
6150 // thread-localness.
6151 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6152 GV = GA->resolveAliasedGlobal(false);
6154 TLSModel::Model model
6155 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6158 case TLSModel::GeneralDynamic:
6159 case TLSModel::LocalDynamic: // not implemented
6160 if (Subtarget->is64Bit())
6161 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6162 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6164 case TLSModel::InitialExec:
6165 case TLSModel::LocalExec:
6166 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6167 Subtarget->is64Bit());
6169 } else if (Subtarget->isTargetDarwin()) {
6170 // Darwin only has one model of TLS. Lower to that.
6171 unsigned char OpFlag = 0;
6172 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6173 X86ISD::WrapperRIP : X86ISD::Wrapper;
6175 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6177 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6178 !Subtarget->is64Bit();
6180 OpFlag = X86II::MO_TLVP_PIC_BASE;
6182 OpFlag = X86II::MO_TLVP;
6183 DebugLoc DL = Op.getDebugLoc();
6184 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6185 GA->getValueType(0),
6186 GA->getOffset(), OpFlag);
6187 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6189 // With PIC32, the address is actually $g + Offset.
6191 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6192 DAG.getNode(X86ISD::GlobalBaseReg,
6193 DebugLoc(), getPointerTy()),
6196 // Lowering the machine isd will make sure everything is in the right
6198 SDValue Chain = DAG.getEntryNode();
6199 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6200 SDValue Args[] = { Chain, Offset };
6201 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
6203 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6204 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6205 MFI->setAdjustsStack(true);
6207 // And our return value (tls address) is in the standard call return value
6209 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
6210 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6214 "TLS not implemented for this target.");
6216 llvm_unreachable("Unreachable");
6221 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
6222 /// take a 2 x i32 value to shift plus a shift amount.
6223 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
6224 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6225 EVT VT = Op.getValueType();
6226 unsigned VTBits = VT.getSizeInBits();
6227 DebugLoc dl = Op.getDebugLoc();
6228 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6229 SDValue ShOpLo = Op.getOperand(0);
6230 SDValue ShOpHi = Op.getOperand(1);
6231 SDValue ShAmt = Op.getOperand(2);
6232 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6233 DAG.getConstant(VTBits - 1, MVT::i8))
6234 : DAG.getConstant(0, VT);
6237 if (Op.getOpcode() == ISD::SHL_PARTS) {
6238 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6239 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6241 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6242 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6245 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6246 DAG.getConstant(VTBits, MVT::i8));
6247 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6248 AndNode, DAG.getConstant(0, MVT::i8));
6251 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6252 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6253 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6255 if (Op.getOpcode() == ISD::SHL_PARTS) {
6256 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6257 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6259 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6260 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6263 SDValue Ops[2] = { Lo, Hi };
6264 return DAG.getMergeValues(Ops, 2, dl);
6267 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6268 SelectionDAG &DAG) const {
6269 EVT SrcVT = Op.getOperand(0).getValueType();
6271 if (SrcVT.isVector())
6274 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6275 "Unknown SINT_TO_FP to lower!");
6277 // These are really Legal; return the operand so the caller accepts it as
6279 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6281 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6282 Subtarget->is64Bit()) {
6286 DebugLoc dl = Op.getDebugLoc();
6287 unsigned Size = SrcVT.getSizeInBits()/8;
6288 MachineFunction &MF = DAG.getMachineFunction();
6289 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6290 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6291 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6293 MachinePointerInfo::getFixedStack(SSFI),
6295 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6298 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6300 SelectionDAG &DAG) const {
6302 DebugLoc DL = Op.getDebugLoc();
6304 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6306 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
6308 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6310 unsigned ByteSize = SrcVT.getSizeInBits()/8;
6312 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
6313 MachineMemOperand *MMO =
6314 DAG.getMachineFunction()
6315 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6316 MachineMemOperand::MOLoad, ByteSize, ByteSize);
6318 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6319 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
6321 Tys, Ops, array_lengthof(Ops),
6325 Chain = Result.getValue(1);
6326 SDValue InFlag = Result.getValue(2);
6328 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6329 // shouldn't be necessary except that RFP cannot be live across
6330 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6331 MachineFunction &MF = DAG.getMachineFunction();
6332 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
6333 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
6334 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6335 Tys = DAG.getVTList(MVT::Other);
6337 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6339 MachineMemOperand *MMO =
6340 DAG.getMachineFunction()
6341 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6342 MachineMemOperand::MOStore, SSFISize, SSFISize);
6344 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
6345 Ops, array_lengthof(Ops),
6346 Op.getValueType(), MMO);
6347 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
6348 MachinePointerInfo::getFixedStack(SSFI),
6355 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6356 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6357 SelectionDAG &DAG) const {
6358 // This algorithm is not obvious. Here it is in C code, more or less:
6360 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6361 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6362 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6364 // Copy ints to xmm registers.
6365 __m128i xh = _mm_cvtsi32_si128( hi );
6366 __m128i xl = _mm_cvtsi32_si128( lo );
6368 // Combine into low half of a single xmm register.
6369 __m128i x = _mm_unpacklo_epi32( xh, xl );
6373 // Merge in appropriate exponents to give the integer bits the right
6375 x = _mm_unpacklo_epi32( x, exp );
6377 // Subtract away the biases to deal with the IEEE-754 double precision
6379 d = _mm_sub_pd( (__m128d) x, bias );
6381 // All conversions up to here are exact. The correctly rounded result is
6382 // calculated using the current rounding mode using the following
6384 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6385 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6386 // store doesn't really need to be here (except
6387 // maybe to zero the other double)
6392 DebugLoc dl = Op.getDebugLoc();
6393 LLVMContext *Context = DAG.getContext();
6395 // Build some magic constants.
6396 std::vector<Constant*> CV0;
6397 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6398 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6399 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6400 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6401 Constant *C0 = ConstantVector::get(CV0);
6402 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6404 std::vector<Constant*> CV1;
6406 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6408 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6409 Constant *C1 = ConstantVector::get(CV1);
6410 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6412 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6413 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6415 DAG.getIntPtrConstant(1)));
6416 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6417 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6419 DAG.getIntPtrConstant(0)));
6420 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6421 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6422 MachinePointerInfo::getConstantPool(),
6424 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6425 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
6426 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6427 MachinePointerInfo::getConstantPool(),
6429 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6431 // Add the halves; easiest way is to swap them into another reg first.
6432 int ShufMask[2] = { 1, -1 };
6433 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6434 DAG.getUNDEF(MVT::v2f64), ShufMask);
6435 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6436 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6437 DAG.getIntPtrConstant(0));
6440 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6441 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6442 SelectionDAG &DAG) const {
6443 DebugLoc dl = Op.getDebugLoc();
6444 // FP constant to bias correct the final result.
6445 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6448 // Load the 32-bit value into an XMM register.
6449 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6450 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6452 DAG.getIntPtrConstant(0)));
6454 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6455 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
6456 DAG.getIntPtrConstant(0));
6458 // Or the load with the bias.
6459 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6460 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6461 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6463 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6464 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6465 MVT::v2f64, Bias)));
6466 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6467 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
6468 DAG.getIntPtrConstant(0));
6470 // Subtract the bias.
6471 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
6473 // Handle final rounding.
6474 EVT DestVT = Op.getValueType();
6476 if (DestVT.bitsLT(MVT::f64)) {
6477 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
6478 DAG.getIntPtrConstant(0));
6479 } else if (DestVT.bitsGT(MVT::f64)) {
6480 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
6483 // Handle final rounding.
6487 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
6488 SelectionDAG &DAG) const {
6489 SDValue N0 = Op.getOperand(0);
6490 DebugLoc dl = Op.getDebugLoc();
6492 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
6493 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
6494 // the optimization here.
6495 if (DAG.SignBitIsZero(N0))
6496 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
6498 EVT SrcVT = N0.getValueType();
6499 EVT DstVT = Op.getValueType();
6500 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
6501 return LowerUINT_TO_FP_i64(Op, DAG);
6502 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
6503 return LowerUINT_TO_FP_i32(Op, DAG);
6505 // Make a 64-bit buffer, and use it to build an FILD.
6506 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
6507 if (SrcVT == MVT::i32) {
6508 SDValue WordOff = DAG.getConstant(4, getPointerTy());
6509 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
6510 getPointerTy(), StackSlot, WordOff);
6511 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6512 StackSlot, MachinePointerInfo(),
6514 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
6515 OffsetSlot, MachinePointerInfo(),
6517 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
6521 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
6522 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6523 StackSlot, MachinePointerInfo(),
6525 // For i64 source, we need to add the appropriate power of 2 if the input
6526 // was negative. This is the same as the optimization in
6527 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
6528 // we must be careful to do the computation in x87 extended precision, not
6529 // in SSE. (The generic code can't know it's OK to do this, or how to.)
6530 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
6531 MachineMemOperand *MMO =
6532 DAG.getMachineFunction()
6533 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6534 MachineMemOperand::MOLoad, 8, 8);
6536 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
6537 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
6538 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
6541 APInt FF(32, 0x5F800000ULL);
6543 // Check whether the sign bit is set.
6544 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
6545 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
6548 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
6549 SDValue FudgePtr = DAG.getConstantPool(
6550 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
6553 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
6554 SDValue Zero = DAG.getIntPtrConstant(0);
6555 SDValue Four = DAG.getIntPtrConstant(4);
6556 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
6558 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
6560 // Load the value out, extending it from f32 to f80.
6561 // FIXME: Avoid the extend by constructing the right constant pool?
6562 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
6563 FudgePtr, MachinePointerInfo::getConstantPool(),
6564 MVT::f32, false, false, 4);
6565 // Extend everything to 80 bits to force it to be done on x87.
6566 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
6567 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
6570 std::pair<SDValue,SDValue> X86TargetLowering::
6571 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
6572 DebugLoc DL = Op.getDebugLoc();
6574 EVT DstTy = Op.getValueType();
6577 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
6581 assert(DstTy.getSimpleVT() <= MVT::i64 &&
6582 DstTy.getSimpleVT() >= MVT::i16 &&
6583 "Unknown FP_TO_SINT to lower!");
6585 // These are really Legal.
6586 if (DstTy == MVT::i32 &&
6587 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6588 return std::make_pair(SDValue(), SDValue());
6589 if (Subtarget->is64Bit() &&
6590 DstTy == MVT::i64 &&
6591 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6592 return std::make_pair(SDValue(), SDValue());
6594 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
6596 MachineFunction &MF = DAG.getMachineFunction();
6597 unsigned MemSize = DstTy.getSizeInBits()/8;
6598 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6599 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6604 switch (DstTy.getSimpleVT().SimpleTy) {
6605 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
6606 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
6607 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
6608 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
6611 SDValue Chain = DAG.getEntryNode();
6612 SDValue Value = Op.getOperand(0);
6613 EVT TheVT = Op.getOperand(0).getValueType();
6614 if (isScalarFPTypeInSSEReg(TheVT)) {
6615 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
6616 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
6617 MachinePointerInfo::getFixedStack(SSFI),
6619 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
6621 Chain, StackSlot, DAG.getValueType(TheVT)
6624 MachineMemOperand *MMO =
6625 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6626 MachineMemOperand::MOLoad, MemSize, MemSize);
6627 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
6629 Chain = Value.getValue(1);
6630 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6631 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6634 MachineMemOperand *MMO =
6635 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6636 MachineMemOperand::MOStore, MemSize, MemSize);
6638 // Build the FP_TO_INT*_IN_MEM
6639 SDValue Ops[] = { Chain, Value, StackSlot };
6640 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
6641 Ops, 3, DstTy, MMO);
6643 return std::make_pair(FIST, StackSlot);
6646 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
6647 SelectionDAG &DAG) const {
6648 if (Op.getValueType().isVector())
6651 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
6652 SDValue FIST = Vals.first, StackSlot = Vals.second;
6653 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
6654 if (FIST.getNode() == 0) return Op;
6657 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6658 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
6661 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
6662 SelectionDAG &DAG) const {
6663 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
6664 SDValue FIST = Vals.first, StackSlot = Vals.second;
6665 assert(FIST.getNode() && "Unexpected failure");
6668 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6669 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
6672 SDValue X86TargetLowering::LowerFABS(SDValue Op,
6673 SelectionDAG &DAG) const {
6674 LLVMContext *Context = DAG.getContext();
6675 DebugLoc dl = Op.getDebugLoc();
6676 EVT VT = Op.getValueType();
6679 EltVT = VT.getVectorElementType();
6680 std::vector<Constant*> CV;
6681 if (EltVT == MVT::f64) {
6682 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
6686 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
6692 Constant *C = ConstantVector::get(CV);
6693 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6694 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6695 MachinePointerInfo::getConstantPool(),
6697 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
6700 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
6701 LLVMContext *Context = DAG.getContext();
6702 DebugLoc dl = Op.getDebugLoc();
6703 EVT VT = Op.getValueType();
6706 EltVT = VT.getVectorElementType();
6707 std::vector<Constant*> CV;
6708 if (EltVT == MVT::f64) {
6709 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
6713 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
6719 Constant *C = ConstantVector::get(CV);
6720 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6721 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6722 MachinePointerInfo::getConstantPool(),
6724 if (VT.isVector()) {
6725 return DAG.getNode(ISD::BITCAST, dl, VT,
6726 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6727 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6729 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
6731 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6735 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6736 LLVMContext *Context = DAG.getContext();
6737 SDValue Op0 = Op.getOperand(0);
6738 SDValue Op1 = Op.getOperand(1);
6739 DebugLoc dl = Op.getDebugLoc();
6740 EVT VT = Op.getValueType();
6741 EVT SrcVT = Op1.getValueType();
6743 // If second operand is smaller, extend it first.
6744 if (SrcVT.bitsLT(VT)) {
6745 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6748 // And if it is bigger, shrink it first.
6749 if (SrcVT.bitsGT(VT)) {
6750 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6754 // At this point the operands and the result should have the same
6755 // type, and that won't be f80 since that is not custom lowered.
6757 // First get the sign bit of second operand.
6758 std::vector<Constant*> CV;
6759 if (SrcVT == MVT::f64) {
6760 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6761 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6763 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6764 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6765 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6766 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6768 Constant *C = ConstantVector::get(CV);
6769 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6770 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6771 MachinePointerInfo::getConstantPool(),
6773 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6775 // Shift sign bit right or left if the two operands have different types.
6776 if (SrcVT.bitsGT(VT)) {
6777 // Op0 is MVT::f32, Op1 is MVT::f64.
6778 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6779 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6780 DAG.getConstant(32, MVT::i32));
6781 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
6782 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6783 DAG.getIntPtrConstant(0));
6786 // Clear first operand sign bit.
6788 if (VT == MVT::f64) {
6789 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6790 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6792 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6793 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6794 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6795 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6797 C = ConstantVector::get(CV);
6798 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6799 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6800 MachinePointerInfo::getConstantPool(),
6802 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6804 // Or the value with the sign bit.
6805 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6808 /// Emit nodes that will be selected as "test Op0,Op0", or something
6810 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6811 SelectionDAG &DAG) const {
6812 DebugLoc dl = Op.getDebugLoc();
6814 // CF and OF aren't always set the way we want. Determine which
6815 // of these we need.
6816 bool NeedCF = false;
6817 bool NeedOF = false;
6820 case X86::COND_A: case X86::COND_AE:
6821 case X86::COND_B: case X86::COND_BE:
6824 case X86::COND_G: case X86::COND_GE:
6825 case X86::COND_L: case X86::COND_LE:
6826 case X86::COND_O: case X86::COND_NO:
6831 // See if we can use the EFLAGS value from the operand instead of
6832 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6833 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6834 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6835 // Emit a CMP with 0, which is the TEST pattern.
6836 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6837 DAG.getConstant(0, Op.getValueType()));
6839 unsigned Opcode = 0;
6840 unsigned NumOperands = 0;
6841 switch (Op.getNode()->getOpcode()) {
6843 // Due to an isel shortcoming, be conservative if this add is likely to be
6844 // selected as part of a load-modify-store instruction. When the root node
6845 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6846 // uses of other nodes in the match, such as the ADD in this case. This
6847 // leads to the ADD being left around and reselected, with the result being
6848 // two adds in the output. Alas, even if none our users are stores, that
6849 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6850 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6851 // climbing the DAG back to the root, and it doesn't seem to be worth the
6853 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6854 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6855 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6858 if (ConstantSDNode *C =
6859 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6860 // An add of one will be selected as an INC.
6861 if (C->getAPIntValue() == 1) {
6862 Opcode = X86ISD::INC;
6867 // An add of negative one (subtract of one) will be selected as a DEC.
6868 if (C->getAPIntValue().isAllOnesValue()) {
6869 Opcode = X86ISD::DEC;
6875 // Otherwise use a regular EFLAGS-setting add.
6876 Opcode = X86ISD::ADD;
6880 // If the primary and result isn't used, don't bother using X86ISD::AND,
6881 // because a TEST instruction will be better.
6882 bool NonFlagUse = false;
6883 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6884 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6886 unsigned UOpNo = UI.getOperandNo();
6887 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6888 // Look pass truncate.
6889 UOpNo = User->use_begin().getOperandNo();
6890 User = *User->use_begin();
6893 if (User->getOpcode() != ISD::BRCOND &&
6894 User->getOpcode() != ISD::SETCC &&
6895 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6908 // Due to the ISEL shortcoming noted above, be conservative if this op is
6909 // likely to be selected as part of a load-modify-store instruction.
6910 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6911 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6912 if (UI->getOpcode() == ISD::STORE)
6915 // Otherwise use a regular EFLAGS-setting instruction.
6916 switch (Op.getNode()->getOpcode()) {
6917 default: llvm_unreachable("unexpected operator!");
6918 case ISD::SUB: Opcode = X86ISD::SUB; break;
6919 case ISD::OR: Opcode = X86ISD::OR; break;
6920 case ISD::XOR: Opcode = X86ISD::XOR; break;
6921 case ISD::AND: Opcode = X86ISD::AND; break;
6933 return SDValue(Op.getNode(), 1);
6940 // Emit a CMP with 0, which is the TEST pattern.
6941 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6942 DAG.getConstant(0, Op.getValueType()));
6944 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6945 SmallVector<SDValue, 4> Ops;
6946 for (unsigned i = 0; i != NumOperands; ++i)
6947 Ops.push_back(Op.getOperand(i));
6949 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6950 DAG.ReplaceAllUsesWith(Op, New);
6951 return SDValue(New.getNode(), 1);
6954 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6956 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6957 SelectionDAG &DAG) const {
6958 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6959 if (C->getAPIntValue() == 0)
6960 return EmitTest(Op0, X86CC, DAG);
6962 DebugLoc dl = Op0.getDebugLoc();
6963 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6966 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6967 /// if it's possible.
6968 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6969 DebugLoc dl, SelectionDAG &DAG) const {
6970 SDValue Op0 = And.getOperand(0);
6971 SDValue Op1 = And.getOperand(1);
6972 if (Op0.getOpcode() == ISD::TRUNCATE)
6973 Op0 = Op0.getOperand(0);
6974 if (Op1.getOpcode() == ISD::TRUNCATE)
6975 Op1 = Op1.getOperand(0);
6978 if (Op1.getOpcode() == ISD::SHL)
6979 std::swap(Op0, Op1);
6980 if (Op0.getOpcode() == ISD::SHL) {
6981 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6982 if (And00C->getZExtValue() == 1) {
6983 // If we looked past a truncate, check that it's only truncating away
6985 unsigned BitWidth = Op0.getValueSizeInBits();
6986 unsigned AndBitWidth = And.getValueSizeInBits();
6987 if (BitWidth > AndBitWidth) {
6988 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6989 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6990 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6994 RHS = Op0.getOperand(1);
6996 } else if (Op1.getOpcode() == ISD::Constant) {
6997 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6998 SDValue AndLHS = Op0;
6999 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
7000 LHS = AndLHS.getOperand(0);
7001 RHS = AndLHS.getOperand(1);
7005 if (LHS.getNode()) {
7006 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
7007 // instruction. Since the shift amount is in-range-or-undefined, we know
7008 // that doing a bittest on the i32 value is ok. We extend to i32 because
7009 // the encoding for the i16 version is larger than the i32 version.
7010 // Also promote i16 to i32 for performance / code size reason.
7011 if (LHS.getValueType() == MVT::i8 ||
7012 LHS.getValueType() == MVT::i16)
7013 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
7015 // If the operand types disagree, extend the shift amount to match. Since
7016 // BT ignores high bits (like shifts) we can use anyextend.
7017 if (LHS.getValueType() != RHS.getValueType())
7018 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
7020 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
7021 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
7022 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7023 DAG.getConstant(Cond, MVT::i8), BT);
7029 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
7030 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
7031 SDValue Op0 = Op.getOperand(0);
7032 SDValue Op1 = Op.getOperand(1);
7033 DebugLoc dl = Op.getDebugLoc();
7034 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7036 // Optimize to BT if possible.
7037 // Lower (X & (1 << N)) == 0 to BT(X, N).
7038 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
7039 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
7040 if (Op0.getOpcode() == ISD::AND &&
7042 Op1.getOpcode() == ISD::Constant &&
7043 cast<ConstantSDNode>(Op1)->isNullValue() &&
7044 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7045 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
7046 if (NewSetCC.getNode())
7050 // Look for "(setcc) == / != 1" to avoid unnecessary setcc.
7051 if (Op0.getOpcode() == X86ISD::SETCC &&
7052 Op1.getOpcode() == ISD::Constant &&
7053 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
7054 cast<ConstantSDNode>(Op1)->isNullValue()) &&
7055 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7056 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
7057 bool Invert = (CC == ISD::SETNE) ^
7058 cast<ConstantSDNode>(Op1)->isNullValue();
7060 CCode = X86::GetOppositeBranchCondition(CCode);
7061 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7062 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
7065 bool isFP = Op1.getValueType().isFloatingPoint();
7066 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
7067 if (X86CC == X86::COND_INVALID)
7070 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
7072 // Use sbb x, x to materialize carry bit into a GPR.
7073 if (X86CC == X86::COND_B)
7074 return DAG.getNode(ISD::AND, dl, MVT::i8,
7075 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
7076 DAG.getConstant(X86CC, MVT::i8), Cond),
7077 DAG.getConstant(1, MVT::i8));
7079 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7080 DAG.getConstant(X86CC, MVT::i8), Cond);
7083 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
7085 SDValue Op0 = Op.getOperand(0);
7086 SDValue Op1 = Op.getOperand(1);
7087 SDValue CC = Op.getOperand(2);
7088 EVT VT = Op.getValueType();
7089 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
7090 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
7091 DebugLoc dl = Op.getDebugLoc();
7095 EVT VT0 = Op0.getValueType();
7096 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
7097 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
7100 switch (SetCCOpcode) {
7103 case ISD::SETEQ: SSECC = 0; break;
7105 case ISD::SETGT: Swap = true; // Fallthrough
7107 case ISD::SETOLT: SSECC = 1; break;
7109 case ISD::SETGE: Swap = true; // Fallthrough
7111 case ISD::SETOLE: SSECC = 2; break;
7112 case ISD::SETUO: SSECC = 3; break;
7114 case ISD::SETNE: SSECC = 4; break;
7115 case ISD::SETULE: Swap = true;
7116 case ISD::SETUGE: SSECC = 5; break;
7117 case ISD::SETULT: Swap = true;
7118 case ISD::SETUGT: SSECC = 6; break;
7119 case ISD::SETO: SSECC = 7; break;
7122 std::swap(Op0, Op1);
7124 // In the two special cases we can't handle, emit two comparisons.
7126 if (SetCCOpcode == ISD::SETUEQ) {
7128 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
7129 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
7130 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
7132 else if (SetCCOpcode == ISD::SETONE) {
7134 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7135 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7136 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7138 llvm_unreachable("Illegal FP comparison");
7140 // Handle all other FP comparisons here.
7141 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7144 // We are handling one of the integer comparisons here. Since SSE only has
7145 // GT and EQ comparisons for integer, swapping operands and multiple
7146 // operations may be required for some comparisons.
7147 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7148 bool Swap = false, Invert = false, FlipSigns = false;
7150 switch (VT.getSimpleVT().SimpleTy) {
7152 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7153 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7154 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7155 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7158 switch (SetCCOpcode) {
7160 case ISD::SETNE: Invert = true;
7161 case ISD::SETEQ: Opc = EQOpc; break;
7162 case ISD::SETLT: Swap = true;
7163 case ISD::SETGT: Opc = GTOpc; break;
7164 case ISD::SETGE: Swap = true;
7165 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7166 case ISD::SETULT: Swap = true;
7167 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7168 case ISD::SETUGE: Swap = true;
7169 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7172 std::swap(Op0, Op1);
7174 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7175 // bits of the inputs before performing those operations.
7177 EVT EltVT = VT.getVectorElementType();
7178 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7180 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7181 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7183 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7184 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7187 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7189 // If the logical-not of the result is required, perform that now.
7191 Result = DAG.getNOT(dl, Result, VT);
7196 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
7197 static bool isX86LogicalCmp(SDValue Op) {
7198 unsigned Opc = Op.getNode()->getOpcode();
7199 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
7201 if (Op.getResNo() == 1 &&
7202 (Opc == X86ISD::ADD ||
7203 Opc == X86ISD::SUB ||
7204 Opc == X86ISD::SMUL ||
7205 Opc == X86ISD::UMUL ||
7206 Opc == X86ISD::INC ||
7207 Opc == X86ISD::DEC ||
7208 Opc == X86ISD::OR ||
7209 Opc == X86ISD::XOR ||
7210 Opc == X86ISD::AND))
7213 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
7219 static bool isZero(SDValue V) {
7220 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7221 return C && C->isNullValue();
7224 static bool isAllOnes(SDValue V) {
7225 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7226 return C && C->isAllOnesValue();
7229 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
7230 bool addTest = true;
7231 SDValue Cond = Op.getOperand(0);
7232 SDValue Op1 = Op.getOperand(1);
7233 SDValue Op2 = Op.getOperand(2);
7234 DebugLoc DL = Op.getDebugLoc();
7237 if (Cond.getOpcode() == ISD::SETCC) {
7238 SDValue NewCond = LowerSETCC(Cond, DAG);
7239 if (NewCond.getNode())
7243 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
7244 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
7245 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
7246 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
7247 if (Cond.getOpcode() == X86ISD::SETCC &&
7248 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
7249 isZero(Cond.getOperand(1).getOperand(1))) {
7250 SDValue Cmp = Cond.getOperand(1);
7252 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
7254 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
7255 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
7256 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
7258 SDValue CmpOp0 = Cmp.getOperand(0);
7259 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
7260 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7262 SDValue Res = // Res = 0 or -1.
7263 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7264 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7266 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
7267 Res = DAG.getNOT(DL, Res, Res.getValueType());
7269 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
7270 if (N2C == 0 || !N2C->isNullValue())
7271 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
7276 // Look past (and (setcc_carry (cmp ...)), 1).
7277 if (Cond.getOpcode() == ISD::AND &&
7278 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7279 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7280 if (C && C->getAPIntValue() == 1)
7281 Cond = Cond.getOperand(0);
7284 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7285 // setting operand in place of the X86ISD::SETCC.
7286 if (Cond.getOpcode() == X86ISD::SETCC ||
7287 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7288 CC = Cond.getOperand(0);
7290 SDValue Cmp = Cond.getOperand(1);
7291 unsigned Opc = Cmp.getOpcode();
7292 EVT VT = Op.getValueType();
7294 bool IllegalFPCMov = false;
7295 if (VT.isFloatingPoint() && !VT.isVector() &&
7296 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7297 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7299 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7300 Opc == X86ISD::BT) { // FIXME
7307 // Look pass the truncate.
7308 if (Cond.getOpcode() == ISD::TRUNCATE)
7309 Cond = Cond.getOperand(0);
7311 // We know the result of AND is compared against zero. Try to match
7313 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7314 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
7315 if (NewSetCC.getNode()) {
7316 CC = NewSetCC.getOperand(0);
7317 Cond = NewSetCC.getOperand(1);
7324 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7325 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7328 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7329 // condition is true.
7330 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
7331 SDValue Ops[] = { Op2, Op1, CC, Cond };
7332 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
7335 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7336 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7337 // from the AND / OR.
7338 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7339 Opc = Op.getOpcode();
7340 if (Opc != ISD::OR && Opc != ISD::AND)
7342 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7343 Op.getOperand(0).hasOneUse() &&
7344 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7345 Op.getOperand(1).hasOneUse());
7348 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7349 // 1 and that the SETCC node has a single use.
7350 static bool isXor1OfSetCC(SDValue Op) {
7351 if (Op.getOpcode() != ISD::XOR)
7353 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7354 if (N1C && N1C->getAPIntValue() == 1) {
7355 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7356 Op.getOperand(0).hasOneUse();
7361 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7362 bool addTest = true;
7363 SDValue Chain = Op.getOperand(0);
7364 SDValue Cond = Op.getOperand(1);
7365 SDValue Dest = Op.getOperand(2);
7366 DebugLoc dl = Op.getDebugLoc();
7369 if (Cond.getOpcode() == ISD::SETCC) {
7370 SDValue NewCond = LowerSETCC(Cond, DAG);
7371 if (NewCond.getNode())
7375 // FIXME: LowerXALUO doesn't handle these!!
7376 else if (Cond.getOpcode() == X86ISD::ADD ||
7377 Cond.getOpcode() == X86ISD::SUB ||
7378 Cond.getOpcode() == X86ISD::SMUL ||
7379 Cond.getOpcode() == X86ISD::UMUL)
7380 Cond = LowerXALUO(Cond, DAG);
7383 // Look pass (and (setcc_carry (cmp ...)), 1).
7384 if (Cond.getOpcode() == ISD::AND &&
7385 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7386 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7387 if (C && C->getAPIntValue() == 1)
7388 Cond = Cond.getOperand(0);
7391 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7392 // setting operand in place of the X86ISD::SETCC.
7393 if (Cond.getOpcode() == X86ISD::SETCC ||
7394 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7395 CC = Cond.getOperand(0);
7397 SDValue Cmp = Cond.getOperand(1);
7398 unsigned Opc = Cmp.getOpcode();
7399 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7400 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7404 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7408 // These can only come from an arithmetic instruction with overflow,
7409 // e.g. SADDO, UADDO.
7410 Cond = Cond.getNode()->getOperand(1);
7417 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7418 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7419 if (CondOpc == ISD::OR) {
7420 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7421 // two branches instead of an explicit OR instruction with a
7423 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7424 isX86LogicalCmp(Cmp)) {
7425 CC = Cond.getOperand(0).getOperand(0);
7426 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7427 Chain, Dest, CC, Cmp);
7428 CC = Cond.getOperand(1).getOperand(0);
7432 } else { // ISD::AND
7433 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7434 // two branches instead of an explicit AND instruction with a
7435 // separate test. However, we only do this if this block doesn't
7436 // have a fall-through edge, because this requires an explicit
7437 // jmp when the condition is false.
7438 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7439 isX86LogicalCmp(Cmp) &&
7440 Op.getNode()->hasOneUse()) {
7441 X86::CondCode CCode =
7442 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7443 CCode = X86::GetOppositeBranchCondition(CCode);
7444 CC = DAG.getConstant(CCode, MVT::i8);
7445 SDNode *User = *Op.getNode()->use_begin();
7446 // Look for an unconditional branch following this conditional branch.
7447 // We need this because we need to reverse the successors in order
7448 // to implement FCMP_OEQ.
7449 if (User->getOpcode() == ISD::BR) {
7450 SDValue FalseBB = User->getOperand(1);
7452 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
7453 assert(NewBR == User);
7457 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7458 Chain, Dest, CC, Cmp);
7459 X86::CondCode CCode =
7460 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
7461 CCode = X86::GetOppositeBranchCondition(CCode);
7462 CC = DAG.getConstant(CCode, MVT::i8);
7468 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
7469 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
7470 // It should be transformed during dag combiner except when the condition
7471 // is set by a arithmetics with overflow node.
7472 X86::CondCode CCode =
7473 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7474 CCode = X86::GetOppositeBranchCondition(CCode);
7475 CC = DAG.getConstant(CCode, MVT::i8);
7476 Cond = Cond.getOperand(0).getOperand(1);
7482 // Look pass the truncate.
7483 if (Cond.getOpcode() == ISD::TRUNCATE)
7484 Cond = Cond.getOperand(0);
7486 // We know the result of AND is compared against zero. Try to match
7488 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7489 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7490 if (NewSetCC.getNode()) {
7491 CC = NewSetCC.getOperand(0);
7492 Cond = NewSetCC.getOperand(1);
7499 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7500 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7502 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7503 Chain, Dest, CC, Cond);
7507 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
7508 // Calls to _alloca is needed to probe the stack when allocating more than 4k
7509 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
7510 // that the guard pages used by the OS virtual memory manager are allocated in
7511 // correct sequence.
7513 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
7514 SelectionDAG &DAG) const {
7515 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows()) &&
7516 "This should be used only on Windows targets");
7517 DebugLoc dl = Op.getDebugLoc();
7520 SDValue Chain = Op.getOperand(0);
7521 SDValue Size = Op.getOperand(1);
7522 // FIXME: Ensure alignment here
7526 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
7528 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
7529 Flag = Chain.getValue(1);
7531 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
7533 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
7534 Flag = Chain.getValue(1);
7536 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
7538 SDValue Ops1[2] = { Chain.getValue(0), Chain };
7539 return DAG.getMergeValues(Ops1, 2, dl);
7542 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
7543 MachineFunction &MF = DAG.getMachineFunction();
7544 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
7546 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7547 DebugLoc DL = Op.getDebugLoc();
7549 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
7550 // vastart just stores the address of the VarArgsFrameIndex slot into the
7551 // memory location argument.
7552 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7554 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
7555 MachinePointerInfo(SV), false, false, 0);
7559 // gp_offset (0 - 6 * 8)
7560 // fp_offset (48 - 48 + 8 * 16)
7561 // overflow_arg_area (point to parameters coming in memory).
7563 SmallVector<SDValue, 8> MemOps;
7564 SDValue FIN = Op.getOperand(1);
7566 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
7567 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
7569 FIN, MachinePointerInfo(SV), false, false, 0);
7570 MemOps.push_back(Store);
7573 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7574 FIN, DAG.getIntPtrConstant(4));
7575 Store = DAG.getStore(Op.getOperand(0), DL,
7576 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
7578 FIN, MachinePointerInfo(SV, 4), false, false, 0);
7579 MemOps.push_back(Store);
7581 // Store ptr to overflow_arg_area
7582 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7583 FIN, DAG.getIntPtrConstant(4));
7584 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7586 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
7587 MachinePointerInfo(SV, 8),
7589 MemOps.push_back(Store);
7591 // Store ptr to reg_save_area.
7592 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7593 FIN, DAG.getIntPtrConstant(8));
7594 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
7596 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
7597 MachinePointerInfo(SV, 16), false, false, 0);
7598 MemOps.push_back(Store);
7599 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
7600 &MemOps[0], MemOps.size());
7603 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
7604 assert(Subtarget->is64Bit() &&
7605 "LowerVAARG only handles 64-bit va_arg!");
7606 assert((Subtarget->isTargetLinux() ||
7607 Subtarget->isTargetDarwin()) &&
7608 "Unhandled target in LowerVAARG");
7609 assert(Op.getNode()->getNumOperands() == 4);
7610 SDValue Chain = Op.getOperand(0);
7611 SDValue SrcPtr = Op.getOperand(1);
7612 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7613 unsigned Align = Op.getConstantOperandVal(3);
7614 DebugLoc dl = Op.getDebugLoc();
7616 EVT ArgVT = Op.getNode()->getValueType(0);
7617 const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
7618 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
7621 // Decide which area this value should be read from.
7622 // TODO: Implement the AMD64 ABI in its entirety. This simple
7623 // selection mechanism works only for the basic types.
7624 if (ArgVT == MVT::f80) {
7625 llvm_unreachable("va_arg for f80 not yet implemented");
7626 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
7627 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
7628 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
7629 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
7631 llvm_unreachable("Unhandled argument type in LowerVAARG");
7635 // Sanity Check: Make sure using fp_offset makes sense.
7636 assert(!UseSoftFloat &&
7637 !(DAG.getMachineFunction()
7638 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
7639 Subtarget->hasXMM());
7642 // Insert VAARG_64 node into the DAG
7643 // VAARG_64 returns two values: Variable Argument Address, Chain
7644 SmallVector<SDValue, 11> InstOps;
7645 InstOps.push_back(Chain);
7646 InstOps.push_back(SrcPtr);
7647 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
7648 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
7649 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
7650 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
7651 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
7652 VTs, &InstOps[0], InstOps.size(),
7654 MachinePointerInfo(SV),
7659 Chain = VAARG.getValue(1);
7661 // Load the next argument and return it
7662 return DAG.getLoad(ArgVT, dl,
7665 MachinePointerInfo(),
7669 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
7670 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7671 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
7672 SDValue Chain = Op.getOperand(0);
7673 SDValue DstPtr = Op.getOperand(1);
7674 SDValue SrcPtr = Op.getOperand(2);
7675 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
7676 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7677 DebugLoc DL = Op.getDebugLoc();
7679 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
7680 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
7682 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
7686 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
7687 DebugLoc dl = Op.getDebugLoc();
7688 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7690 default: return SDValue(); // Don't custom lower most intrinsics.
7691 // Comparison intrinsics.
7692 case Intrinsic::x86_sse_comieq_ss:
7693 case Intrinsic::x86_sse_comilt_ss:
7694 case Intrinsic::x86_sse_comile_ss:
7695 case Intrinsic::x86_sse_comigt_ss:
7696 case Intrinsic::x86_sse_comige_ss:
7697 case Intrinsic::x86_sse_comineq_ss:
7698 case Intrinsic::x86_sse_ucomieq_ss:
7699 case Intrinsic::x86_sse_ucomilt_ss:
7700 case Intrinsic::x86_sse_ucomile_ss:
7701 case Intrinsic::x86_sse_ucomigt_ss:
7702 case Intrinsic::x86_sse_ucomige_ss:
7703 case Intrinsic::x86_sse_ucomineq_ss:
7704 case Intrinsic::x86_sse2_comieq_sd:
7705 case Intrinsic::x86_sse2_comilt_sd:
7706 case Intrinsic::x86_sse2_comile_sd:
7707 case Intrinsic::x86_sse2_comigt_sd:
7708 case Intrinsic::x86_sse2_comige_sd:
7709 case Intrinsic::x86_sse2_comineq_sd:
7710 case Intrinsic::x86_sse2_ucomieq_sd:
7711 case Intrinsic::x86_sse2_ucomilt_sd:
7712 case Intrinsic::x86_sse2_ucomile_sd:
7713 case Intrinsic::x86_sse2_ucomigt_sd:
7714 case Intrinsic::x86_sse2_ucomige_sd:
7715 case Intrinsic::x86_sse2_ucomineq_sd: {
7717 ISD::CondCode CC = ISD::SETCC_INVALID;
7720 case Intrinsic::x86_sse_comieq_ss:
7721 case Intrinsic::x86_sse2_comieq_sd:
7725 case Intrinsic::x86_sse_comilt_ss:
7726 case Intrinsic::x86_sse2_comilt_sd:
7730 case Intrinsic::x86_sse_comile_ss:
7731 case Intrinsic::x86_sse2_comile_sd:
7735 case Intrinsic::x86_sse_comigt_ss:
7736 case Intrinsic::x86_sse2_comigt_sd:
7740 case Intrinsic::x86_sse_comige_ss:
7741 case Intrinsic::x86_sse2_comige_sd:
7745 case Intrinsic::x86_sse_comineq_ss:
7746 case Intrinsic::x86_sse2_comineq_sd:
7750 case Intrinsic::x86_sse_ucomieq_ss:
7751 case Intrinsic::x86_sse2_ucomieq_sd:
7752 Opc = X86ISD::UCOMI;
7755 case Intrinsic::x86_sse_ucomilt_ss:
7756 case Intrinsic::x86_sse2_ucomilt_sd:
7757 Opc = X86ISD::UCOMI;
7760 case Intrinsic::x86_sse_ucomile_ss:
7761 case Intrinsic::x86_sse2_ucomile_sd:
7762 Opc = X86ISD::UCOMI;
7765 case Intrinsic::x86_sse_ucomigt_ss:
7766 case Intrinsic::x86_sse2_ucomigt_sd:
7767 Opc = X86ISD::UCOMI;
7770 case Intrinsic::x86_sse_ucomige_ss:
7771 case Intrinsic::x86_sse2_ucomige_sd:
7772 Opc = X86ISD::UCOMI;
7775 case Intrinsic::x86_sse_ucomineq_ss:
7776 case Intrinsic::x86_sse2_ucomineq_sd:
7777 Opc = X86ISD::UCOMI;
7782 SDValue LHS = Op.getOperand(1);
7783 SDValue RHS = Op.getOperand(2);
7784 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
7785 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
7786 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
7787 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7788 DAG.getConstant(X86CC, MVT::i8), Cond);
7789 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7791 // ptest and testp intrinsics. The intrinsic these come from are designed to
7792 // return an integer value, not just an instruction so lower it to the ptest
7793 // or testp pattern and a setcc for the result.
7794 case Intrinsic::x86_sse41_ptestz:
7795 case Intrinsic::x86_sse41_ptestc:
7796 case Intrinsic::x86_sse41_ptestnzc:
7797 case Intrinsic::x86_avx_ptestz_256:
7798 case Intrinsic::x86_avx_ptestc_256:
7799 case Intrinsic::x86_avx_ptestnzc_256:
7800 case Intrinsic::x86_avx_vtestz_ps:
7801 case Intrinsic::x86_avx_vtestc_ps:
7802 case Intrinsic::x86_avx_vtestnzc_ps:
7803 case Intrinsic::x86_avx_vtestz_pd:
7804 case Intrinsic::x86_avx_vtestc_pd:
7805 case Intrinsic::x86_avx_vtestnzc_pd:
7806 case Intrinsic::x86_avx_vtestz_ps_256:
7807 case Intrinsic::x86_avx_vtestc_ps_256:
7808 case Intrinsic::x86_avx_vtestnzc_ps_256:
7809 case Intrinsic::x86_avx_vtestz_pd_256:
7810 case Intrinsic::x86_avx_vtestc_pd_256:
7811 case Intrinsic::x86_avx_vtestnzc_pd_256: {
7812 bool IsTestPacked = false;
7815 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
7816 case Intrinsic::x86_avx_vtestz_ps:
7817 case Intrinsic::x86_avx_vtestz_pd:
7818 case Intrinsic::x86_avx_vtestz_ps_256:
7819 case Intrinsic::x86_avx_vtestz_pd_256:
7820 IsTestPacked = true; // Fallthrough
7821 case Intrinsic::x86_sse41_ptestz:
7822 case Intrinsic::x86_avx_ptestz_256:
7824 X86CC = X86::COND_E;
7826 case Intrinsic::x86_avx_vtestc_ps:
7827 case Intrinsic::x86_avx_vtestc_pd:
7828 case Intrinsic::x86_avx_vtestc_ps_256:
7829 case Intrinsic::x86_avx_vtestc_pd_256:
7830 IsTestPacked = true; // Fallthrough
7831 case Intrinsic::x86_sse41_ptestc:
7832 case Intrinsic::x86_avx_ptestc_256:
7834 X86CC = X86::COND_B;
7836 case Intrinsic::x86_avx_vtestnzc_ps:
7837 case Intrinsic::x86_avx_vtestnzc_pd:
7838 case Intrinsic::x86_avx_vtestnzc_ps_256:
7839 case Intrinsic::x86_avx_vtestnzc_pd_256:
7840 IsTestPacked = true; // Fallthrough
7841 case Intrinsic::x86_sse41_ptestnzc:
7842 case Intrinsic::x86_avx_ptestnzc_256:
7844 X86CC = X86::COND_A;
7848 SDValue LHS = Op.getOperand(1);
7849 SDValue RHS = Op.getOperand(2);
7850 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
7851 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
7852 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7853 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7854 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7857 // Fix vector shift instructions where the last operand is a non-immediate
7859 case Intrinsic::x86_sse2_pslli_w:
7860 case Intrinsic::x86_sse2_pslli_d:
7861 case Intrinsic::x86_sse2_pslli_q:
7862 case Intrinsic::x86_sse2_psrli_w:
7863 case Intrinsic::x86_sse2_psrli_d:
7864 case Intrinsic::x86_sse2_psrli_q:
7865 case Intrinsic::x86_sse2_psrai_w:
7866 case Intrinsic::x86_sse2_psrai_d:
7867 case Intrinsic::x86_mmx_pslli_w:
7868 case Intrinsic::x86_mmx_pslli_d:
7869 case Intrinsic::x86_mmx_pslli_q:
7870 case Intrinsic::x86_mmx_psrli_w:
7871 case Intrinsic::x86_mmx_psrli_d:
7872 case Intrinsic::x86_mmx_psrli_q:
7873 case Intrinsic::x86_mmx_psrai_w:
7874 case Intrinsic::x86_mmx_psrai_d: {
7875 SDValue ShAmt = Op.getOperand(2);
7876 if (isa<ConstantSDNode>(ShAmt))
7879 unsigned NewIntNo = 0;
7880 EVT ShAmtVT = MVT::v4i32;
7882 case Intrinsic::x86_sse2_pslli_w:
7883 NewIntNo = Intrinsic::x86_sse2_psll_w;
7885 case Intrinsic::x86_sse2_pslli_d:
7886 NewIntNo = Intrinsic::x86_sse2_psll_d;
7888 case Intrinsic::x86_sse2_pslli_q:
7889 NewIntNo = Intrinsic::x86_sse2_psll_q;
7891 case Intrinsic::x86_sse2_psrli_w:
7892 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7894 case Intrinsic::x86_sse2_psrli_d:
7895 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7897 case Intrinsic::x86_sse2_psrli_q:
7898 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7900 case Intrinsic::x86_sse2_psrai_w:
7901 NewIntNo = Intrinsic::x86_sse2_psra_w;
7903 case Intrinsic::x86_sse2_psrai_d:
7904 NewIntNo = Intrinsic::x86_sse2_psra_d;
7907 ShAmtVT = MVT::v2i32;
7909 case Intrinsic::x86_mmx_pslli_w:
7910 NewIntNo = Intrinsic::x86_mmx_psll_w;
7912 case Intrinsic::x86_mmx_pslli_d:
7913 NewIntNo = Intrinsic::x86_mmx_psll_d;
7915 case Intrinsic::x86_mmx_pslli_q:
7916 NewIntNo = Intrinsic::x86_mmx_psll_q;
7918 case Intrinsic::x86_mmx_psrli_w:
7919 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7921 case Intrinsic::x86_mmx_psrli_d:
7922 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7924 case Intrinsic::x86_mmx_psrli_q:
7925 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7927 case Intrinsic::x86_mmx_psrai_w:
7928 NewIntNo = Intrinsic::x86_mmx_psra_w;
7930 case Intrinsic::x86_mmx_psrai_d:
7931 NewIntNo = Intrinsic::x86_mmx_psra_d;
7933 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7939 // The vector shift intrinsics with scalars uses 32b shift amounts but
7940 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7944 ShOps[1] = DAG.getConstant(0, MVT::i32);
7945 if (ShAmtVT == MVT::v4i32) {
7946 ShOps[2] = DAG.getUNDEF(MVT::i32);
7947 ShOps[3] = DAG.getUNDEF(MVT::i32);
7948 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7950 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7951 // FIXME this must be lowered to get rid of the invalid type.
7954 EVT VT = Op.getValueType();
7955 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
7956 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7957 DAG.getConstant(NewIntNo, MVT::i32),
7958 Op.getOperand(1), ShAmt);
7963 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7964 SelectionDAG &DAG) const {
7965 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7966 MFI->setReturnAddressIsTaken(true);
7968 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7969 DebugLoc dl = Op.getDebugLoc();
7972 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7974 DAG.getConstant(TD->getPointerSize(),
7975 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7976 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7977 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7979 MachinePointerInfo(), false, false, 0);
7982 // Just load the return address.
7983 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7984 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7985 RetAddrFI, MachinePointerInfo(), false, false, 0);
7988 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7989 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7990 MFI->setFrameAddressIsTaken(true);
7992 EVT VT = Op.getValueType();
7993 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7994 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7995 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7996 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7998 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
7999 MachinePointerInfo(),
8004 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
8005 SelectionDAG &DAG) const {
8006 return DAG.getIntPtrConstant(2*TD->getPointerSize());
8009 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
8010 MachineFunction &MF = DAG.getMachineFunction();
8011 SDValue Chain = Op.getOperand(0);
8012 SDValue Offset = Op.getOperand(1);
8013 SDValue Handler = Op.getOperand(2);
8014 DebugLoc dl = Op.getDebugLoc();
8016 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
8017 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
8019 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
8021 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
8022 DAG.getIntPtrConstant(TD->getPointerSize()));
8023 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
8024 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
8026 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
8027 MF.getRegInfo().addLiveOut(StoreAddrReg);
8029 return DAG.getNode(X86ISD::EH_RETURN, dl,
8031 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
8034 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
8035 SelectionDAG &DAG) const {
8036 SDValue Root = Op.getOperand(0);
8037 SDValue Trmp = Op.getOperand(1); // trampoline
8038 SDValue FPtr = Op.getOperand(2); // nested function
8039 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
8040 DebugLoc dl = Op.getDebugLoc();
8042 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8044 if (Subtarget->is64Bit()) {
8045 SDValue OutChains[6];
8047 // Large code-model.
8048 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
8049 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
8051 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
8052 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
8054 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
8056 // Load the pointer to the nested function into R11.
8057 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
8058 SDValue Addr = Trmp;
8059 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8060 Addr, MachinePointerInfo(TrmpAddr),
8063 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8064 DAG.getConstant(2, MVT::i64));
8065 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
8066 MachinePointerInfo(TrmpAddr, 2),
8069 // Load the 'nest' parameter value into R10.
8070 // R10 is specified in X86CallingConv.td
8071 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
8072 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8073 DAG.getConstant(10, MVT::i64));
8074 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8075 Addr, MachinePointerInfo(TrmpAddr, 10),
8078 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8079 DAG.getConstant(12, MVT::i64));
8080 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
8081 MachinePointerInfo(TrmpAddr, 12),
8084 // Jump to the nested function.
8085 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
8086 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8087 DAG.getConstant(20, MVT::i64));
8088 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8089 Addr, MachinePointerInfo(TrmpAddr, 20),
8092 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
8093 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8094 DAG.getConstant(22, MVT::i64));
8095 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
8096 MachinePointerInfo(TrmpAddr, 22),
8100 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
8101 return DAG.getMergeValues(Ops, 2, dl);
8103 const Function *Func =
8104 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
8105 CallingConv::ID CC = Func->getCallingConv();
8110 llvm_unreachable("Unsupported calling convention");
8111 case CallingConv::C:
8112 case CallingConv::X86_StdCall: {
8113 // Pass 'nest' parameter in ECX.
8114 // Must be kept in sync with X86CallingConv.td
8117 // Check that ECX wasn't needed by an 'inreg' parameter.
8118 const FunctionType *FTy = Func->getFunctionType();
8119 const AttrListPtr &Attrs = Func->getAttributes();
8121 if (!Attrs.isEmpty() && !Func->isVarArg()) {
8122 unsigned InRegCount = 0;
8125 for (FunctionType::param_iterator I = FTy->param_begin(),
8126 E = FTy->param_end(); I != E; ++I, ++Idx)
8127 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
8128 // FIXME: should only count parameters that are lowered to integers.
8129 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
8131 if (InRegCount > 2) {
8132 report_fatal_error("Nest register in use - reduce number of inreg"
8138 case CallingConv::X86_FastCall:
8139 case CallingConv::X86_ThisCall:
8140 case CallingConv::Fast:
8141 // Pass 'nest' parameter in EAX.
8142 // Must be kept in sync with X86CallingConv.td
8147 SDValue OutChains[4];
8150 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8151 DAG.getConstant(10, MVT::i32));
8152 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
8154 // This is storing the opcode for MOV32ri.
8155 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
8156 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
8157 OutChains[0] = DAG.getStore(Root, dl,
8158 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
8159 Trmp, MachinePointerInfo(TrmpAddr),
8162 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8163 DAG.getConstant(1, MVT::i32));
8164 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
8165 MachinePointerInfo(TrmpAddr, 1),
8168 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
8169 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8170 DAG.getConstant(5, MVT::i32));
8171 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
8172 MachinePointerInfo(TrmpAddr, 5),
8175 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8176 DAG.getConstant(6, MVT::i32));
8177 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
8178 MachinePointerInfo(TrmpAddr, 6),
8182 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
8183 return DAG.getMergeValues(Ops, 2, dl);
8187 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
8188 SelectionDAG &DAG) const {
8190 The rounding mode is in bits 11:10 of FPSR, and has the following
8197 FLT_ROUNDS, on the other hand, expects the following:
8204 To perform the conversion, we do:
8205 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
8208 MachineFunction &MF = DAG.getMachineFunction();
8209 const TargetMachine &TM = MF.getTarget();
8210 const TargetFrameInfo &TFI = *TM.getFrameInfo();
8211 unsigned StackAlignment = TFI.getStackAlignment();
8212 EVT VT = Op.getValueType();
8213 DebugLoc DL = Op.getDebugLoc();
8215 // Save FP Control Word to stack slot
8216 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
8217 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8220 MachineMemOperand *MMO =
8221 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8222 MachineMemOperand::MOStore, 2, 2);
8224 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
8225 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
8226 DAG.getVTList(MVT::Other),
8227 Ops, 2, MVT::i16, MMO);
8229 // Load FP Control Word from stack slot
8230 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
8231 MachinePointerInfo(), false, false, 0);
8233 // Transform as necessary
8235 DAG.getNode(ISD::SRL, DL, MVT::i16,
8236 DAG.getNode(ISD::AND, DL, MVT::i16,
8237 CWD, DAG.getConstant(0x800, MVT::i16)),
8238 DAG.getConstant(11, MVT::i8));
8240 DAG.getNode(ISD::SRL, DL, MVT::i16,
8241 DAG.getNode(ISD::AND, DL, MVT::i16,
8242 CWD, DAG.getConstant(0x400, MVT::i16)),
8243 DAG.getConstant(9, MVT::i8));
8246 DAG.getNode(ISD::AND, DL, MVT::i16,
8247 DAG.getNode(ISD::ADD, DL, MVT::i16,
8248 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
8249 DAG.getConstant(1, MVT::i16)),
8250 DAG.getConstant(3, MVT::i16));
8253 return DAG.getNode((VT.getSizeInBits() < 16 ?
8254 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
8257 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
8258 EVT VT = Op.getValueType();
8260 unsigned NumBits = VT.getSizeInBits();
8261 DebugLoc dl = Op.getDebugLoc();
8263 Op = Op.getOperand(0);
8264 if (VT == MVT::i8) {
8265 // Zero extend to i32 since there is not an i8 bsr.
8267 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8270 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
8271 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8272 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
8274 // If src is zero (i.e. bsr sets ZF), returns NumBits.
8277 DAG.getConstant(NumBits+NumBits-1, OpVT),
8278 DAG.getConstant(X86::COND_E, MVT::i8),
8281 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8283 // Finally xor with NumBits-1.
8284 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
8287 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8291 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
8292 EVT VT = Op.getValueType();
8294 unsigned NumBits = VT.getSizeInBits();
8295 DebugLoc dl = Op.getDebugLoc();
8297 Op = Op.getOperand(0);
8298 if (VT == MVT::i8) {
8300 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8303 // Issue a bsf (scan bits forward) which also sets EFLAGS.
8304 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8305 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
8307 // If src is zero (i.e. bsf sets ZF), returns NumBits.
8310 DAG.getConstant(NumBits, OpVT),
8311 DAG.getConstant(X86::COND_E, MVT::i8),
8314 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8317 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8321 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
8322 EVT VT = Op.getValueType();
8323 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
8324 DebugLoc dl = Op.getDebugLoc();
8326 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
8327 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
8328 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
8329 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
8330 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
8332 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
8333 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
8334 // return AloBlo + AloBhi + AhiBlo;
8336 SDValue A = Op.getOperand(0);
8337 SDValue B = Op.getOperand(1);
8339 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8340 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8341 A, DAG.getConstant(32, MVT::i32));
8342 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8343 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8344 B, DAG.getConstant(32, MVT::i32));
8345 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8346 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8348 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8349 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8351 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8352 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8354 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8355 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8356 AloBhi, DAG.getConstant(32, MVT::i32));
8357 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8358 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8359 AhiBlo, DAG.getConstant(32, MVT::i32));
8360 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8361 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8365 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
8366 EVT VT = Op.getValueType();
8367 DebugLoc dl = Op.getDebugLoc();
8368 SDValue R = Op.getOperand(0);
8370 LLVMContext *Context = DAG.getContext();
8372 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
8374 if (VT == MVT::v4i32) {
8375 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8376 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8377 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
8379 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
8381 std::vector<Constant*> CV(4, CI);
8382 Constant *C = ConstantVector::get(CV);
8383 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8384 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8385 MachinePointerInfo::getConstantPool(),
8388 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
8389 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
8390 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
8391 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
8393 if (VT == MVT::v16i8) {
8395 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8396 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8397 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
8399 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
8400 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
8402 std::vector<Constant*> CVM1(16, CM1);
8403 std::vector<Constant*> CVM2(16, CM2);
8404 Constant *C = ConstantVector::get(CVM1);
8405 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8406 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8407 MachinePointerInfo::getConstantPool(),
8410 // r = pblendv(r, psllw(r & (char16)15, 4), a);
8411 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8412 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8413 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8414 DAG.getConstant(4, MVT::i32));
8415 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8416 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8419 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8421 C = ConstantVector::get(CVM2);
8422 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8423 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8424 MachinePointerInfo::getConstantPool(),
8427 // r = pblendv(r, psllw(r & (char16)63, 2), a);
8428 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8429 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8430 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8431 DAG.getConstant(2, MVT::i32));
8432 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8433 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8436 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8438 // return pblendv(r, r+r, a);
8439 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8440 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8441 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
8447 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
8448 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
8449 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
8450 // looks for this combo and may remove the "setcc" instruction if the "setcc"
8451 // has only one use.
8452 SDNode *N = Op.getNode();
8453 SDValue LHS = N->getOperand(0);
8454 SDValue RHS = N->getOperand(1);
8455 unsigned BaseOp = 0;
8457 DebugLoc DL = Op.getDebugLoc();
8458 switch (Op.getOpcode()) {
8459 default: llvm_unreachable("Unknown ovf instruction!");
8461 // A subtract of one will be selected as a INC. Note that INC doesn't
8462 // set CF, so we can't do this for UADDO.
8463 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8464 if (C->getAPIntValue() == 1) {
8465 BaseOp = X86ISD::INC;
8469 BaseOp = X86ISD::ADD;
8473 BaseOp = X86ISD::ADD;
8477 // A subtract of one will be selected as a DEC. Note that DEC doesn't
8478 // set CF, so we can't do this for USUBO.
8479 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8480 if (C->getAPIntValue() == 1) {
8481 BaseOp = X86ISD::DEC;
8485 BaseOp = X86ISD::SUB;
8489 BaseOp = X86ISD::SUB;
8493 BaseOp = X86ISD::SMUL;
8496 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
8497 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
8499 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
8502 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8503 DAG.getConstant(X86::COND_O, MVT::i32),
8504 SDValue(Sum.getNode(), 2));
8506 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8511 // Also sets EFLAGS.
8512 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
8513 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
8516 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
8517 DAG.getConstant(Cond, MVT::i32),
8518 SDValue(Sum.getNode(), 1));
8520 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8524 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
8525 DebugLoc dl = Op.getDebugLoc();
8527 if (!Subtarget->hasSSE2()) {
8528 SDValue Chain = Op.getOperand(0);
8529 SDValue Zero = DAG.getConstant(0,
8530 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8532 DAG.getRegister(X86::ESP, MVT::i32), // Base
8533 DAG.getTargetConstant(1, MVT::i8), // Scale
8534 DAG.getRegister(0, MVT::i32), // Index
8535 DAG.getTargetConstant(0, MVT::i32), // Disp
8536 DAG.getRegister(0, MVT::i32), // Segment.
8541 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
8542 array_lengthof(Ops));
8543 return SDValue(Res, 0);
8546 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
8548 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
8550 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
8551 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
8552 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
8553 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
8555 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
8556 if (!Op1 && !Op2 && !Op3 && Op4)
8557 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
8559 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
8560 if (Op1 && !Op2 && !Op3 && !Op4)
8561 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
8563 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
8565 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
8568 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
8569 EVT T = Op.getValueType();
8570 DebugLoc DL = Op.getDebugLoc();
8573 switch(T.getSimpleVT().SimpleTy) {
8575 assert(false && "Invalid value type!");
8576 case MVT::i8: Reg = X86::AL; size = 1; break;
8577 case MVT::i16: Reg = X86::AX; size = 2; break;
8578 case MVT::i32: Reg = X86::EAX; size = 4; break;
8580 assert(Subtarget->is64Bit() && "Node not type legal!");
8581 Reg = X86::RAX; size = 8;
8584 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
8585 Op.getOperand(2), SDValue());
8586 SDValue Ops[] = { cpIn.getValue(0),
8589 DAG.getTargetConstant(size, MVT::i8),
8591 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8592 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
8593 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
8596 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
8600 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
8601 SelectionDAG &DAG) const {
8602 assert(Subtarget->is64Bit() && "Result not type legalized?");
8603 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8604 SDValue TheChain = Op.getOperand(0);
8605 DebugLoc dl = Op.getDebugLoc();
8606 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8607 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
8608 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
8610 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
8611 DAG.getConstant(32, MVT::i8));
8613 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
8616 return DAG.getMergeValues(Ops, 2, dl);
8619 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
8620 SelectionDAG &DAG) const {
8621 EVT SrcVT = Op.getOperand(0).getValueType();
8622 EVT DstVT = Op.getValueType();
8623 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
8624 Subtarget->hasMMX() && !DisableMMX) &&
8625 "Unexpected custom BITCAST");
8626 assert((DstVT == MVT::i64 ||
8627 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
8628 "Unexpected custom BITCAST");
8629 // i64 <=> MMX conversions are Legal.
8630 if (SrcVT==MVT::i64 && DstVT.isVector())
8632 if (DstVT==MVT::i64 && SrcVT.isVector())
8634 // MMX <=> MMX conversions are Legal.
8635 if (SrcVT.isVector() && DstVT.isVector())
8637 // All other conversions need to be expanded.
8640 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
8641 SDNode *Node = Op.getNode();
8642 DebugLoc dl = Node->getDebugLoc();
8643 EVT T = Node->getValueType(0);
8644 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
8645 DAG.getConstant(0, T), Node->getOperand(2));
8646 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
8647 cast<AtomicSDNode>(Node)->getMemoryVT(),
8648 Node->getOperand(0),
8649 Node->getOperand(1), negOp,
8650 cast<AtomicSDNode>(Node)->getSrcValue(),
8651 cast<AtomicSDNode>(Node)->getAlignment());
8654 /// LowerOperation - Provide custom lowering hooks for some operations.
8656 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
8657 switch (Op.getOpcode()) {
8658 default: llvm_unreachable("Should not custom lower this!");
8659 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
8660 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
8661 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
8662 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
8663 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
8664 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
8665 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
8666 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
8667 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
8668 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
8669 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
8670 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
8671 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
8672 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
8673 case ISD::SHL_PARTS:
8674 case ISD::SRA_PARTS:
8675 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
8676 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
8677 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
8678 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
8679 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
8680 case ISD::FABS: return LowerFABS(Op, DAG);
8681 case ISD::FNEG: return LowerFNEG(Op, DAG);
8682 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
8683 case ISD::SETCC: return LowerSETCC(Op, DAG);
8684 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
8685 case ISD::SELECT: return LowerSELECT(Op, DAG);
8686 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
8687 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
8688 case ISD::VASTART: return LowerVASTART(Op, DAG);
8689 case ISD::VAARG: return LowerVAARG(Op, DAG);
8690 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
8691 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
8692 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
8693 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
8694 case ISD::FRAME_TO_ARGS_OFFSET:
8695 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
8696 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
8697 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
8698 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
8699 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
8700 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
8701 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
8702 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
8703 case ISD::SHL: return LowerSHL(Op, DAG);
8709 case ISD::UMULO: return LowerXALUO(Op, DAG);
8710 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
8711 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
8715 void X86TargetLowering::
8716 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
8717 SelectionDAG &DAG, unsigned NewOp) const {
8718 EVT T = Node->getValueType(0);
8719 DebugLoc dl = Node->getDebugLoc();
8720 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
8722 SDValue Chain = Node->getOperand(0);
8723 SDValue In1 = Node->getOperand(1);
8724 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8725 Node->getOperand(2), DAG.getIntPtrConstant(0));
8726 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8727 Node->getOperand(2), DAG.getIntPtrConstant(1));
8728 SDValue Ops[] = { Chain, In1, In2L, In2H };
8729 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
8731 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
8732 cast<MemSDNode>(Node)->getMemOperand());
8733 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
8734 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8735 Results.push_back(Result.getValue(2));
8738 /// ReplaceNodeResults - Replace a node with an illegal result type
8739 /// with a new node built out of custom code.
8740 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
8741 SmallVectorImpl<SDValue>&Results,
8742 SelectionDAG &DAG) const {
8743 DebugLoc dl = N->getDebugLoc();
8744 switch (N->getOpcode()) {
8746 assert(false && "Do not know how to custom type legalize this operation!");
8748 case ISD::FP_TO_SINT: {
8749 std::pair<SDValue,SDValue> Vals =
8750 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
8751 SDValue FIST = Vals.first, StackSlot = Vals.second;
8752 if (FIST.getNode() != 0) {
8753 EVT VT = N->getValueType(0);
8754 // Return a load from the stack slot.
8755 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
8756 MachinePointerInfo(), false, false, 0));
8760 case ISD::READCYCLECOUNTER: {
8761 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8762 SDValue TheChain = N->getOperand(0);
8763 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8764 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
8766 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
8768 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
8769 SDValue Ops[] = { eax, edx };
8770 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
8771 Results.push_back(edx.getValue(1));
8774 case ISD::ATOMIC_CMP_SWAP: {
8775 EVT T = N->getValueType(0);
8776 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
8777 SDValue cpInL, cpInH;
8778 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8779 DAG.getConstant(0, MVT::i32));
8780 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8781 DAG.getConstant(1, MVT::i32));
8782 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
8783 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
8785 SDValue swapInL, swapInH;
8786 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8787 DAG.getConstant(0, MVT::i32));
8788 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8789 DAG.getConstant(1, MVT::i32));
8790 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
8792 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
8793 swapInL.getValue(1));
8794 SDValue Ops[] = { swapInH.getValue(0),
8796 swapInH.getValue(1) };
8797 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8798 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
8799 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
8801 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
8802 MVT::i32, Result.getValue(1));
8803 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
8804 MVT::i32, cpOutL.getValue(2));
8805 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
8806 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8807 Results.push_back(cpOutH.getValue(1));
8810 case ISD::ATOMIC_LOAD_ADD:
8811 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
8813 case ISD::ATOMIC_LOAD_AND:
8814 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
8816 case ISD::ATOMIC_LOAD_NAND:
8817 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
8819 case ISD::ATOMIC_LOAD_OR:
8820 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
8822 case ISD::ATOMIC_LOAD_SUB:
8823 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
8825 case ISD::ATOMIC_LOAD_XOR:
8826 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
8828 case ISD::ATOMIC_SWAP:
8829 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
8834 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
8836 default: return NULL;
8837 case X86ISD::BSF: return "X86ISD::BSF";
8838 case X86ISD::BSR: return "X86ISD::BSR";
8839 case X86ISD::SHLD: return "X86ISD::SHLD";
8840 case X86ISD::SHRD: return "X86ISD::SHRD";
8841 case X86ISD::FAND: return "X86ISD::FAND";
8842 case X86ISD::FOR: return "X86ISD::FOR";
8843 case X86ISD::FXOR: return "X86ISD::FXOR";
8844 case X86ISD::FSRL: return "X86ISD::FSRL";
8845 case X86ISD::FILD: return "X86ISD::FILD";
8846 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
8847 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
8848 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
8849 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
8850 case X86ISD::FLD: return "X86ISD::FLD";
8851 case X86ISD::FST: return "X86ISD::FST";
8852 case X86ISD::CALL: return "X86ISD::CALL";
8853 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
8854 case X86ISD::BT: return "X86ISD::BT";
8855 case X86ISD::CMP: return "X86ISD::CMP";
8856 case X86ISD::COMI: return "X86ISD::COMI";
8857 case X86ISD::UCOMI: return "X86ISD::UCOMI";
8858 case X86ISD::SETCC: return "X86ISD::SETCC";
8859 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
8860 case X86ISD::CMOV: return "X86ISD::CMOV";
8861 case X86ISD::BRCOND: return "X86ISD::BRCOND";
8862 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
8863 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
8864 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
8865 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
8866 case X86ISD::Wrapper: return "X86ISD::Wrapper";
8867 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
8868 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
8869 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
8870 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
8871 case X86ISD::PINSRB: return "X86ISD::PINSRB";
8872 case X86ISD::PINSRW: return "X86ISD::PINSRW";
8873 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
8874 case X86ISD::PANDN: return "X86ISD::PANDN";
8875 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
8876 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
8877 case X86ISD::PSIGND: return "X86ISD::PSIGND";
8878 case X86ISD::FMAX: return "X86ISD::FMAX";
8879 case X86ISD::FMIN: return "X86ISD::FMIN";
8880 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
8881 case X86ISD::FRCP: return "X86ISD::FRCP";
8882 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
8883 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
8884 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
8885 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
8886 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
8887 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
8888 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
8889 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
8890 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
8891 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
8892 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
8893 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
8894 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
8895 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
8896 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
8897 case X86ISD::VSHL: return "X86ISD::VSHL";
8898 case X86ISD::VSRL: return "X86ISD::VSRL";
8899 case X86ISD::CMPPD: return "X86ISD::CMPPD";
8900 case X86ISD::CMPPS: return "X86ISD::CMPPS";
8901 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
8902 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
8903 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
8904 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
8905 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
8906 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
8907 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
8908 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
8909 case X86ISD::ADD: return "X86ISD::ADD";
8910 case X86ISD::SUB: return "X86ISD::SUB";
8911 case X86ISD::SMUL: return "X86ISD::SMUL";
8912 case X86ISD::UMUL: return "X86ISD::UMUL";
8913 case X86ISD::INC: return "X86ISD::INC";
8914 case X86ISD::DEC: return "X86ISD::DEC";
8915 case X86ISD::OR: return "X86ISD::OR";
8916 case X86ISD::XOR: return "X86ISD::XOR";
8917 case X86ISD::AND: return "X86ISD::AND";
8918 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
8919 case X86ISD::PTEST: return "X86ISD::PTEST";
8920 case X86ISD::TESTP: return "X86ISD::TESTP";
8921 case X86ISD::PALIGN: return "X86ISD::PALIGN";
8922 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
8923 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
8924 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
8925 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
8926 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
8927 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
8928 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
8929 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
8930 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
8931 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
8932 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
8933 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
8934 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
8935 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
8936 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
8937 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
8938 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
8939 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
8940 case X86ISD::MOVSD: return "X86ISD::MOVSD";
8941 case X86ISD::MOVSS: return "X86ISD::MOVSS";
8942 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
8943 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
8944 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
8945 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
8946 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
8947 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
8948 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
8949 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
8950 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
8951 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
8952 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
8953 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
8954 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
8955 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
8956 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
8960 // isLegalAddressingMode - Return true if the addressing mode represented
8961 // by AM is legal for this target, for a load/store of the specified type.
8962 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
8963 const Type *Ty) const {
8964 // X86 supports extremely general addressing modes.
8965 CodeModel::Model M = getTargetMachine().getCodeModel();
8966 Reloc::Model R = getTargetMachine().getRelocationModel();
8968 // X86 allows a sign-extended 32-bit immediate field as a displacement.
8969 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
8974 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
8976 // If a reference to this global requires an extra load, we can't fold it.
8977 if (isGlobalStubReference(GVFlags))
8980 // If BaseGV requires a register for the PIC base, we cannot also have a
8981 // BaseReg specified.
8982 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
8985 // If lower 4G is not available, then we must use rip-relative addressing.
8986 if ((M != CodeModel::Small || R != Reloc::Static) &&
8987 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
8997 // These scales always work.
9002 // These scales are formed with basereg+scalereg. Only accept if there is
9007 default: // Other stuff never works.
9015 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
9016 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9018 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
9019 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
9020 if (NumBits1 <= NumBits2)
9025 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
9026 if (!VT1.isInteger() || !VT2.isInteger())
9028 unsigned NumBits1 = VT1.getSizeInBits();
9029 unsigned NumBits2 = VT2.getSizeInBits();
9030 if (NumBits1 <= NumBits2)
9035 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
9036 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9037 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
9040 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
9041 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9042 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
9045 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
9046 // i16 instructions are longer (0x66 prefix) and potentially slower.
9047 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
9050 /// isShuffleMaskLegal - Targets can use this to indicate that they only
9051 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
9052 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
9053 /// are assumed to be legal.
9055 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
9057 // Very little shuffling can be done for 64-bit vectors right now.
9058 if (VT.getSizeInBits() == 64)
9059 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
9061 // FIXME: pshufb, blends, shifts.
9062 return (VT.getVectorNumElements() == 2 ||
9063 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
9064 isMOVLMask(M, VT) ||
9065 isSHUFPMask(M, VT) ||
9066 isPSHUFDMask(M, VT) ||
9067 isPSHUFHWMask(M, VT) ||
9068 isPSHUFLWMask(M, VT) ||
9069 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
9070 isUNPCKLMask(M, VT) ||
9071 isUNPCKHMask(M, VT) ||
9072 isUNPCKL_v_undef_Mask(M, VT) ||
9073 isUNPCKH_v_undef_Mask(M, VT));
9077 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
9079 unsigned NumElts = VT.getVectorNumElements();
9080 // FIXME: This collection of masks seems suspect.
9083 if (NumElts == 4 && VT.getSizeInBits() == 128) {
9084 return (isMOVLMask(Mask, VT) ||
9085 isCommutedMOVLMask(Mask, VT, true) ||
9086 isSHUFPMask(Mask, VT) ||
9087 isCommutedSHUFPMask(Mask, VT));
9092 //===----------------------------------------------------------------------===//
9093 // X86 Scheduler Hooks
9094 //===----------------------------------------------------------------------===//
9096 // private utility function
9098 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
9099 MachineBasicBlock *MBB,
9106 TargetRegisterClass *RC,
9107 bool invSrc) const {
9108 // For the atomic bitwise operator, we generate
9111 // ld t1 = [bitinstr.addr]
9112 // op t2 = t1, [bitinstr.val]
9114 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9116 // fallthrough -->nextMBB
9117 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9118 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9119 MachineFunction::iterator MBBIter = MBB;
9122 /// First build the CFG
9123 MachineFunction *F = MBB->getParent();
9124 MachineBasicBlock *thisMBB = MBB;
9125 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9126 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9127 F->insert(MBBIter, newMBB);
9128 F->insert(MBBIter, nextMBB);
9130 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9131 nextMBB->splice(nextMBB->begin(), thisMBB,
9132 llvm::next(MachineBasicBlock::iterator(bInstr)),
9134 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9136 // Update thisMBB to fall through to newMBB
9137 thisMBB->addSuccessor(newMBB);
9139 // newMBB jumps to itself and fall through to nextMBB
9140 newMBB->addSuccessor(nextMBB);
9141 newMBB->addSuccessor(newMBB);
9143 // Insert instructions into newMBB based on incoming instruction
9144 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9145 "unexpected number of operands");
9146 DebugLoc dl = bInstr->getDebugLoc();
9147 MachineOperand& destOper = bInstr->getOperand(0);
9148 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9149 int numArgs = bInstr->getNumOperands() - 1;
9150 for (int i=0; i < numArgs; ++i)
9151 argOpers[i] = &bInstr->getOperand(i+1);
9153 // x86 address has 4 operands: base, index, scale, and displacement
9154 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9155 int valArgIndx = lastAddrIndx + 1;
9157 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9158 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
9159 for (int i=0; i <= lastAddrIndx; ++i)
9160 (*MIB).addOperand(*argOpers[i]);
9162 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
9164 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
9169 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9170 assert((argOpers[valArgIndx]->isReg() ||
9171 argOpers[valArgIndx]->isImm()) &&
9173 if (argOpers[valArgIndx]->isReg())
9174 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
9176 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
9178 (*MIB).addOperand(*argOpers[valArgIndx]);
9180 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
9183 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
9184 for (int i=0; i <= lastAddrIndx; ++i)
9185 (*MIB).addOperand(*argOpers[i]);
9187 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9188 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9189 bInstr->memoperands_end());
9191 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9195 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9197 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9201 // private utility function: 64 bit atomics on 32 bit host.
9203 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
9204 MachineBasicBlock *MBB,
9209 bool invSrc) const {
9210 // For the atomic bitwise operator, we generate
9211 // thisMBB (instructions are in pairs, except cmpxchg8b)
9212 // ld t1,t2 = [bitinstr.addr]
9214 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
9215 // op t5, t6 <- out1, out2, [bitinstr.val]
9216 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
9217 // mov ECX, EBX <- t5, t6
9218 // mov EAX, EDX <- t1, t2
9219 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
9220 // mov t3, t4 <- EAX, EDX
9222 // result in out1, out2
9223 // fallthrough -->nextMBB
9225 const TargetRegisterClass *RC = X86::GR32RegisterClass;
9226 const unsigned LoadOpc = X86::MOV32rm;
9227 const unsigned NotOpc = X86::NOT32r;
9228 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9229 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9230 MachineFunction::iterator MBBIter = MBB;
9233 /// First build the CFG
9234 MachineFunction *F = MBB->getParent();
9235 MachineBasicBlock *thisMBB = MBB;
9236 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9237 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9238 F->insert(MBBIter, newMBB);
9239 F->insert(MBBIter, nextMBB);
9241 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9242 nextMBB->splice(nextMBB->begin(), thisMBB,
9243 llvm::next(MachineBasicBlock::iterator(bInstr)),
9245 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9247 // Update thisMBB to fall through to newMBB
9248 thisMBB->addSuccessor(newMBB);
9250 // newMBB jumps to itself and fall through to nextMBB
9251 newMBB->addSuccessor(nextMBB);
9252 newMBB->addSuccessor(newMBB);
9254 DebugLoc dl = bInstr->getDebugLoc();
9255 // Insert instructions into newMBB based on incoming instruction
9256 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
9257 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
9258 "unexpected number of operands");
9259 MachineOperand& dest1Oper = bInstr->getOperand(0);
9260 MachineOperand& dest2Oper = bInstr->getOperand(1);
9261 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9262 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
9263 argOpers[i] = &bInstr->getOperand(i+2);
9265 // We use some of the operands multiple times, so conservatively just
9266 // clear any kill flags that might be present.
9267 if (argOpers[i]->isReg() && argOpers[i]->isUse())
9268 argOpers[i]->setIsKill(false);
9271 // x86 address has 5 operands: base, index, scale, displacement, and segment.
9272 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9274 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9275 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
9276 for (int i=0; i <= lastAddrIndx; ++i)
9277 (*MIB).addOperand(*argOpers[i]);
9278 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9279 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
9280 // add 4 to displacement.
9281 for (int i=0; i <= lastAddrIndx-2; ++i)
9282 (*MIB).addOperand(*argOpers[i]);
9283 MachineOperand newOp3 = *(argOpers[3]);
9285 newOp3.setImm(newOp3.getImm()+4);
9287 newOp3.setOffset(newOp3.getOffset()+4);
9288 (*MIB).addOperand(newOp3);
9289 (*MIB).addOperand(*argOpers[lastAddrIndx]);
9291 // t3/4 are defined later, at the bottom of the loop
9292 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
9293 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
9294 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
9295 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
9296 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
9297 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
9299 // The subsequent operations should be using the destination registers of
9300 //the PHI instructions.
9302 t1 = F->getRegInfo().createVirtualRegister(RC);
9303 t2 = F->getRegInfo().createVirtualRegister(RC);
9304 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
9305 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
9307 t1 = dest1Oper.getReg();
9308 t2 = dest2Oper.getReg();
9311 int valArgIndx = lastAddrIndx + 1;
9312 assert((argOpers[valArgIndx]->isReg() ||
9313 argOpers[valArgIndx]->isImm()) &&
9315 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
9316 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
9317 if (argOpers[valArgIndx]->isReg())
9318 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
9320 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
9321 if (regOpcL != X86::MOV32rr)
9323 (*MIB).addOperand(*argOpers[valArgIndx]);
9324 assert(argOpers[valArgIndx + 1]->isReg() ==
9325 argOpers[valArgIndx]->isReg());
9326 assert(argOpers[valArgIndx + 1]->isImm() ==
9327 argOpers[valArgIndx]->isImm());
9328 if (argOpers[valArgIndx + 1]->isReg())
9329 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
9331 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
9332 if (regOpcH != X86::MOV32rr)
9334 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
9336 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9338 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
9341 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
9343 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
9346 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
9347 for (int i=0; i <= lastAddrIndx; ++i)
9348 (*MIB).addOperand(*argOpers[i]);
9350 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9351 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9352 bInstr->memoperands_end());
9354 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
9355 MIB.addReg(X86::EAX);
9356 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
9357 MIB.addReg(X86::EDX);
9360 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9362 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9366 // private utility function
9368 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
9369 MachineBasicBlock *MBB,
9370 unsigned cmovOpc) const {
9371 // For the atomic min/max operator, we generate
9374 // ld t1 = [min/max.addr]
9375 // mov t2 = [min/max.val]
9377 // cmov[cond] t2 = t1
9379 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9381 // fallthrough -->nextMBB
9383 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9384 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9385 MachineFunction::iterator MBBIter = MBB;
9388 /// First build the CFG
9389 MachineFunction *F = MBB->getParent();
9390 MachineBasicBlock *thisMBB = MBB;
9391 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9392 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9393 F->insert(MBBIter, newMBB);
9394 F->insert(MBBIter, nextMBB);
9396 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9397 nextMBB->splice(nextMBB->begin(), thisMBB,
9398 llvm::next(MachineBasicBlock::iterator(mInstr)),
9400 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9402 // Update thisMBB to fall through to newMBB
9403 thisMBB->addSuccessor(newMBB);
9405 // newMBB jumps to newMBB and fall through to nextMBB
9406 newMBB->addSuccessor(nextMBB);
9407 newMBB->addSuccessor(newMBB);
9409 DebugLoc dl = mInstr->getDebugLoc();
9410 // Insert instructions into newMBB based on incoming instruction
9411 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9412 "unexpected number of operands");
9413 MachineOperand& destOper = mInstr->getOperand(0);
9414 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9415 int numArgs = mInstr->getNumOperands() - 1;
9416 for (int i=0; i < numArgs; ++i)
9417 argOpers[i] = &mInstr->getOperand(i+1);
9419 // x86 address has 4 operands: base, index, scale, and displacement
9420 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9421 int valArgIndx = lastAddrIndx + 1;
9423 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9424 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
9425 for (int i=0; i <= lastAddrIndx; ++i)
9426 (*MIB).addOperand(*argOpers[i]);
9428 // We only support register and immediate values
9429 assert((argOpers[valArgIndx]->isReg() ||
9430 argOpers[valArgIndx]->isImm()) &&
9433 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9434 if (argOpers[valArgIndx]->isReg())
9435 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
9437 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
9438 (*MIB).addOperand(*argOpers[valArgIndx]);
9440 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9443 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
9448 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9449 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
9453 // Cmp and exchange if none has modified the memory location
9454 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
9455 for (int i=0; i <= lastAddrIndx; ++i)
9456 (*MIB).addOperand(*argOpers[i]);
9458 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9459 (*MIB).setMemRefs(mInstr->memoperands_begin(),
9460 mInstr->memoperands_end());
9462 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9463 MIB.addReg(X86::EAX);
9466 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9468 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
9472 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
9473 // or XMM0_V32I8 in AVX all of this code can be replaced with that
9476 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
9477 unsigned numArgs, bool memArg) const {
9478 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
9479 "Target must have SSE4.2 or AVX features enabled");
9481 DebugLoc dl = MI->getDebugLoc();
9482 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9484 if (!Subtarget->hasAVX()) {
9486 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
9488 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
9491 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
9493 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
9496 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
9497 for (unsigned i = 0; i < numArgs; ++i) {
9498 MachineOperand &Op = MI->getOperand(i+1);
9499 if (!(Op.isReg() && Op.isImplicit()))
9502 BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
9505 MI->eraseFromParent();
9510 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
9511 DebugLoc dl = MI->getDebugLoc();
9512 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9514 // Address into RAX/EAX, other two args into ECX, EDX.
9515 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
9516 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
9517 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
9518 for (int i = 0; i < X86::AddrNumOperands; ++i)
9519 MIB.addOperand(MI->getOperand(i));
9521 unsigned ValOps = X86::AddrNumOperands;
9522 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
9523 .addReg(MI->getOperand(ValOps).getReg());
9524 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
9525 .addReg(MI->getOperand(ValOps+1).getReg());
9527 // The instruction doesn't actually take any operands though.
9528 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
9530 MI->eraseFromParent(); // The pseudo is gone now.
9535 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
9536 DebugLoc dl = MI->getDebugLoc();
9537 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9539 // First arg in ECX, the second in EAX.
9540 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
9541 .addReg(MI->getOperand(0).getReg());
9542 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
9543 .addReg(MI->getOperand(1).getReg());
9545 // The instruction doesn't actually take any operands though.
9546 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
9548 MI->eraseFromParent(); // The pseudo is gone now.
9553 X86TargetLowering::EmitVAARG64WithCustomInserter(
9555 MachineBasicBlock *MBB) const {
9556 // Emit va_arg instruction on X86-64.
9558 // Operands to this pseudo-instruction:
9559 // 0 ) Output : destination address (reg)
9560 // 1-5) Input : va_list address (addr, i64mem)
9561 // 6 ) ArgSize : Size (in bytes) of vararg type
9562 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
9563 // 8 ) Align : Alignment of type
9564 // 9 ) EFLAGS (implicit-def)
9566 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
9567 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
9569 unsigned DestReg = MI->getOperand(0).getReg();
9570 MachineOperand &Base = MI->getOperand(1);
9571 MachineOperand &Scale = MI->getOperand(2);
9572 MachineOperand &Index = MI->getOperand(3);
9573 MachineOperand &Disp = MI->getOperand(4);
9574 MachineOperand &Segment = MI->getOperand(5);
9575 unsigned ArgSize = MI->getOperand(6).getImm();
9576 unsigned ArgMode = MI->getOperand(7).getImm();
9577 unsigned Align = MI->getOperand(8).getImm();
9580 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
9581 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
9582 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
9584 // Machine Information
9585 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9586 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
9587 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
9588 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
9589 DebugLoc DL = MI->getDebugLoc();
9594 // i64 overflow_area (address)
9595 // i64 reg_save_area (address)
9597 // sizeof(va_list) = 24
9598 // alignment(va_list) = 8
9600 unsigned TotalNumIntRegs = 6;
9601 unsigned TotalNumXMMRegs = 8;
9602 bool UseGPOffset = (ArgMode == 1);
9603 bool UseFPOffset = (ArgMode == 2);
9604 unsigned MaxOffset = TotalNumIntRegs * 8 +
9605 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
9607 /* Align ArgSize to a multiple of 8 */
9608 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
9609 bool NeedsAlign = (Align > 8);
9611 MachineBasicBlock *thisMBB = MBB;
9612 MachineBasicBlock *overflowMBB;
9613 MachineBasicBlock *offsetMBB;
9614 MachineBasicBlock *endMBB;
9616 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
9617 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
9618 unsigned OffsetReg = 0;
9620 if (!UseGPOffset && !UseFPOffset) {
9621 // If we only pull from the overflow region, we don't create a branch.
9622 // We don't need to alter control flow.
9623 OffsetDestReg = 0; // unused
9624 OverflowDestReg = DestReg;
9627 overflowMBB = thisMBB;
9630 // First emit code to check if gp_offset (or fp_offset) is below the bound.
9631 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
9632 // If not, pull from overflow_area. (branch to overflowMBB)
9637 // offsetMBB overflowMBB
9642 // Registers for the PHI in endMBB
9643 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
9644 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
9646 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9647 MachineFunction *MF = MBB->getParent();
9648 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
9649 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
9650 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
9652 MachineFunction::iterator MBBIter = MBB;
9655 // Insert the new basic blocks
9656 MF->insert(MBBIter, offsetMBB);
9657 MF->insert(MBBIter, overflowMBB);
9658 MF->insert(MBBIter, endMBB);
9660 // Transfer the remainder of MBB and its successor edges to endMBB.
9661 endMBB->splice(endMBB->begin(), thisMBB,
9662 llvm::next(MachineBasicBlock::iterator(MI)),
9664 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9666 // Make offsetMBB and overflowMBB successors of thisMBB
9667 thisMBB->addSuccessor(offsetMBB);
9668 thisMBB->addSuccessor(overflowMBB);
9670 // endMBB is a successor of both offsetMBB and overflowMBB
9671 offsetMBB->addSuccessor(endMBB);
9672 overflowMBB->addSuccessor(endMBB);
9674 // Load the offset value into a register
9675 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
9676 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
9680 .addDisp(Disp, UseFPOffset ? 4 : 0)
9681 .addOperand(Segment)
9682 .setMemRefs(MMOBegin, MMOEnd);
9684 // Check if there is enough room left to pull this argument.
9685 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
9687 .addImm(MaxOffset + 8 - ArgSizeA8);
9689 // Branch to "overflowMBB" if offset >= max
9690 // Fall through to "offsetMBB" otherwise
9691 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
9692 .addMBB(overflowMBB);
9695 // In offsetMBB, emit code to use the reg_save_area.
9697 assert(OffsetReg != 0);
9699 // Read the reg_save_area address.
9700 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
9701 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
9706 .addOperand(Segment)
9707 .setMemRefs(MMOBegin, MMOEnd);
9709 // Zero-extend the offset
9710 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
9711 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
9714 .addImm(X86::sub_32bit);
9716 // Add the offset to the reg_save_area to get the final address.
9717 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
9718 .addReg(OffsetReg64)
9719 .addReg(RegSaveReg);
9721 // Compute the offset for the next argument
9722 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
9723 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
9725 .addImm(UseFPOffset ? 16 : 8);
9727 // Store it back into the va_list.
9728 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
9732 .addDisp(Disp, UseFPOffset ? 4 : 0)
9733 .addOperand(Segment)
9734 .addReg(NextOffsetReg)
9735 .setMemRefs(MMOBegin, MMOEnd);
9738 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
9743 // Emit code to use overflow area
9746 // Load the overflow_area address into a register.
9747 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
9748 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
9753 .addOperand(Segment)
9754 .setMemRefs(MMOBegin, MMOEnd);
9756 // If we need to align it, do so. Otherwise, just copy the address
9757 // to OverflowDestReg.
9759 // Align the overflow address
9760 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
9761 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
9763 // aligned_addr = (addr + (align-1)) & ~(align-1)
9764 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
9765 .addReg(OverflowAddrReg)
9768 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
9770 .addImm(~(uint64_t)(Align-1));
9772 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
9773 .addReg(OverflowAddrReg);
9776 // Compute the next overflow address after this argument.
9777 // (the overflow address should be kept 8-byte aligned)
9778 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
9779 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
9780 .addReg(OverflowDestReg)
9783 // Store the new overflow address.
9784 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
9789 .addOperand(Segment)
9790 .addReg(NextAddrReg)
9791 .setMemRefs(MMOBegin, MMOEnd);
9793 // If we branched, emit the PHI to the front of endMBB.
9795 BuildMI(*endMBB, endMBB->begin(), DL,
9796 TII->get(X86::PHI), DestReg)
9797 .addReg(OffsetDestReg).addMBB(offsetMBB)
9798 .addReg(OverflowDestReg).addMBB(overflowMBB);
9801 // Erase the pseudo instruction
9802 MI->eraseFromParent();
9808 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
9810 MachineBasicBlock *MBB) const {
9811 // Emit code to save XMM registers to the stack. The ABI says that the
9812 // number of registers to save is given in %al, so it's theoretically
9813 // possible to do an indirect jump trick to avoid saving all of them,
9814 // however this code takes a simpler approach and just executes all
9815 // of the stores if %al is non-zero. It's less code, and it's probably
9816 // easier on the hardware branch predictor, and stores aren't all that
9817 // expensive anyway.
9819 // Create the new basic blocks. One block contains all the XMM stores,
9820 // and one block is the final destination regardless of whether any
9821 // stores were performed.
9822 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9823 MachineFunction *F = MBB->getParent();
9824 MachineFunction::iterator MBBIter = MBB;
9826 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
9827 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
9828 F->insert(MBBIter, XMMSaveMBB);
9829 F->insert(MBBIter, EndMBB);
9831 // Transfer the remainder of MBB and its successor edges to EndMBB.
9832 EndMBB->splice(EndMBB->begin(), MBB,
9833 llvm::next(MachineBasicBlock::iterator(MI)),
9835 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
9837 // The original block will now fall through to the XMM save block.
9838 MBB->addSuccessor(XMMSaveMBB);
9839 // The XMMSaveMBB will fall through to the end block.
9840 XMMSaveMBB->addSuccessor(EndMBB);
9842 // Now add the instructions.
9843 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9844 DebugLoc DL = MI->getDebugLoc();
9846 unsigned CountReg = MI->getOperand(0).getReg();
9847 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
9848 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
9850 if (!Subtarget->isTargetWin64()) {
9851 // If %al is 0, branch around the XMM save block.
9852 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
9853 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
9854 MBB->addSuccessor(EndMBB);
9857 // In the XMM save block, save all the XMM argument registers.
9858 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
9859 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
9860 MachineMemOperand *MMO =
9861 F->getMachineMemOperand(
9862 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
9863 MachineMemOperand::MOStore,
9864 /*Size=*/16, /*Align=*/16);
9865 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
9866 .addFrameIndex(RegSaveFrameIndex)
9867 .addImm(/*Scale=*/1)
9868 .addReg(/*IndexReg=*/0)
9869 .addImm(/*Disp=*/Offset)
9870 .addReg(/*Segment=*/0)
9871 .addReg(MI->getOperand(i).getReg())
9872 .addMemOperand(MMO);
9875 MI->eraseFromParent(); // The pseudo instruction is gone now.
9881 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
9882 MachineBasicBlock *BB) const {
9883 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9884 DebugLoc DL = MI->getDebugLoc();
9886 // To "insert" a SELECT_CC instruction, we actually have to insert the
9887 // diamond control-flow pattern. The incoming instruction knows the
9888 // destination vreg to set, the condition code register to branch on, the
9889 // true/false values to select between, and a branch opcode to use.
9890 const BasicBlock *LLVM_BB = BB->getBasicBlock();
9891 MachineFunction::iterator It = BB;
9897 // cmpTY ccX, r1, r2
9899 // fallthrough --> copy0MBB
9900 MachineBasicBlock *thisMBB = BB;
9901 MachineFunction *F = BB->getParent();
9902 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
9903 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
9904 F->insert(It, copy0MBB);
9905 F->insert(It, sinkMBB);
9907 // If the EFLAGS register isn't dead in the terminator, then claim that it's
9908 // live into the sink and copy blocks.
9909 const MachineFunction *MF = BB->getParent();
9910 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
9911 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
9913 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
9914 const MachineOperand &MO = MI->getOperand(I);
9915 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
9916 unsigned Reg = MO.getReg();
9917 if (Reg != X86::EFLAGS) continue;
9918 copy0MBB->addLiveIn(Reg);
9919 sinkMBB->addLiveIn(Reg);
9922 // Transfer the remainder of BB and its successor edges to sinkMBB.
9923 sinkMBB->splice(sinkMBB->begin(), BB,
9924 llvm::next(MachineBasicBlock::iterator(MI)),
9926 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
9928 // Add the true and fallthrough blocks as its successors.
9929 BB->addSuccessor(copy0MBB);
9930 BB->addSuccessor(sinkMBB);
9932 // Create the conditional branch instruction.
9934 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
9935 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
9938 // %FalseValue = ...
9939 // # fallthrough to sinkMBB
9940 copy0MBB->addSuccessor(sinkMBB);
9943 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
9945 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
9946 TII->get(X86::PHI), MI->getOperand(0).getReg())
9947 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
9948 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
9950 MI->eraseFromParent(); // The pseudo instruction is gone now.
9955 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
9956 MachineBasicBlock *BB) const {
9957 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9958 DebugLoc DL = MI->getDebugLoc();
9960 // The lowering is pretty easy: we're just emitting the call to _alloca. The
9961 // non-trivial part is impdef of ESP.
9962 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
9965 const char *StackProbeSymbol =
9966 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
9968 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
9969 .addExternalSymbol(StackProbeSymbol)
9970 .addReg(X86::EAX, RegState::Implicit)
9971 .addReg(X86::ESP, RegState::Implicit)
9972 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
9973 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
9974 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
9976 MI->eraseFromParent(); // The pseudo instruction is gone now.
9981 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
9982 MachineBasicBlock *BB) const {
9983 // This is pretty easy. We're taking the value that we received from
9984 // our load from the relocation, sticking it in either RDI (x86-64)
9985 // or EAX and doing an indirect call. The return value will then
9986 // be in the normal return register.
9987 const X86InstrInfo *TII
9988 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
9989 DebugLoc DL = MI->getDebugLoc();
9990 MachineFunction *F = BB->getParent();
9992 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
9993 assert(MI->getOperand(3).isGlobal() && "This should be a global");
9995 if (Subtarget->is64Bit()) {
9996 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9997 TII->get(X86::MOV64rm), X86::RDI)
9999 .addImm(0).addReg(0)
10000 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10001 MI->getOperand(3).getTargetFlags())
10003 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
10004 addDirectMem(MIB, X86::RDI);
10005 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
10006 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10007 TII->get(X86::MOV32rm), X86::EAX)
10009 .addImm(0).addReg(0)
10010 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10011 MI->getOperand(3).getTargetFlags())
10013 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10014 addDirectMem(MIB, X86::EAX);
10016 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10017 TII->get(X86::MOV32rm), X86::EAX)
10018 .addReg(TII->getGlobalBaseReg(F))
10019 .addImm(0).addReg(0)
10020 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10021 MI->getOperand(3).getTargetFlags())
10023 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10024 addDirectMem(MIB, X86::EAX);
10027 MI->eraseFromParent(); // The pseudo instruction is gone now.
10031 MachineBasicBlock *
10032 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
10033 MachineBasicBlock *BB) const {
10034 switch (MI->getOpcode()) {
10035 default: assert(false && "Unexpected instr type to insert");
10036 case X86::WIN_ALLOCA:
10037 return EmitLoweredWinAlloca(MI, BB);
10038 case X86::TLSCall_32:
10039 case X86::TLSCall_64:
10040 return EmitLoweredTLSCall(MI, BB);
10041 case X86::CMOV_GR8:
10042 case X86::CMOV_FR32:
10043 case X86::CMOV_FR64:
10044 case X86::CMOV_V4F32:
10045 case X86::CMOV_V2F64:
10046 case X86::CMOV_V2I64:
10047 case X86::CMOV_GR16:
10048 case X86::CMOV_GR32:
10049 case X86::CMOV_RFP32:
10050 case X86::CMOV_RFP64:
10051 case X86::CMOV_RFP80:
10052 return EmitLoweredSelect(MI, BB);
10054 case X86::FP32_TO_INT16_IN_MEM:
10055 case X86::FP32_TO_INT32_IN_MEM:
10056 case X86::FP32_TO_INT64_IN_MEM:
10057 case X86::FP64_TO_INT16_IN_MEM:
10058 case X86::FP64_TO_INT32_IN_MEM:
10059 case X86::FP64_TO_INT64_IN_MEM:
10060 case X86::FP80_TO_INT16_IN_MEM:
10061 case X86::FP80_TO_INT32_IN_MEM:
10062 case X86::FP80_TO_INT64_IN_MEM: {
10063 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10064 DebugLoc DL = MI->getDebugLoc();
10066 // Change the floating point control register to use "round towards zero"
10067 // mode when truncating to an integer value.
10068 MachineFunction *F = BB->getParent();
10069 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
10070 addFrameReference(BuildMI(*BB, MI, DL,
10071 TII->get(X86::FNSTCW16m)), CWFrameIdx);
10073 // Load the old value of the high byte of the control word...
10075 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
10076 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
10079 // Set the high part to be round to zero...
10080 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
10083 // Reload the modified control word now...
10084 addFrameReference(BuildMI(*BB, MI, DL,
10085 TII->get(X86::FLDCW16m)), CWFrameIdx);
10087 // Restore the memory image of control word to original value
10088 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
10091 // Get the X86 opcode to use.
10093 switch (MI->getOpcode()) {
10094 default: llvm_unreachable("illegal opcode!");
10095 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
10096 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
10097 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
10098 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
10099 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
10100 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
10101 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
10102 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
10103 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
10107 MachineOperand &Op = MI->getOperand(0);
10109 AM.BaseType = X86AddressMode::RegBase;
10110 AM.Base.Reg = Op.getReg();
10112 AM.BaseType = X86AddressMode::FrameIndexBase;
10113 AM.Base.FrameIndex = Op.getIndex();
10115 Op = MI->getOperand(1);
10117 AM.Scale = Op.getImm();
10118 Op = MI->getOperand(2);
10120 AM.IndexReg = Op.getImm();
10121 Op = MI->getOperand(3);
10122 if (Op.isGlobal()) {
10123 AM.GV = Op.getGlobal();
10125 AM.Disp = Op.getImm();
10127 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
10128 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
10130 // Reload the original control word now.
10131 addFrameReference(BuildMI(*BB, MI, DL,
10132 TII->get(X86::FLDCW16m)), CWFrameIdx);
10134 MI->eraseFromParent(); // The pseudo instruction is gone now.
10137 // String/text processing lowering.
10138 case X86::PCMPISTRM128REG:
10139 case X86::VPCMPISTRM128REG:
10140 return EmitPCMP(MI, BB, 3, false /* in-mem */);
10141 case X86::PCMPISTRM128MEM:
10142 case X86::VPCMPISTRM128MEM:
10143 return EmitPCMP(MI, BB, 3, true /* in-mem */);
10144 case X86::PCMPESTRM128REG:
10145 case X86::VPCMPESTRM128REG:
10146 return EmitPCMP(MI, BB, 5, false /* in mem */);
10147 case X86::PCMPESTRM128MEM:
10148 case X86::VPCMPESTRM128MEM:
10149 return EmitPCMP(MI, BB, 5, true /* in mem */);
10151 // Thread synchronization.
10153 return EmitMonitor(MI, BB);
10155 return EmitMwait(MI, BB);
10157 // Atomic Lowering.
10158 case X86::ATOMAND32:
10159 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10160 X86::AND32ri, X86::MOV32rm,
10162 X86::NOT32r, X86::EAX,
10163 X86::GR32RegisterClass);
10164 case X86::ATOMOR32:
10165 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
10166 X86::OR32ri, X86::MOV32rm,
10168 X86::NOT32r, X86::EAX,
10169 X86::GR32RegisterClass);
10170 case X86::ATOMXOR32:
10171 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
10172 X86::XOR32ri, X86::MOV32rm,
10174 X86::NOT32r, X86::EAX,
10175 X86::GR32RegisterClass);
10176 case X86::ATOMNAND32:
10177 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10178 X86::AND32ri, X86::MOV32rm,
10180 X86::NOT32r, X86::EAX,
10181 X86::GR32RegisterClass, true);
10182 case X86::ATOMMIN32:
10183 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
10184 case X86::ATOMMAX32:
10185 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
10186 case X86::ATOMUMIN32:
10187 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
10188 case X86::ATOMUMAX32:
10189 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
10191 case X86::ATOMAND16:
10192 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
10193 X86::AND16ri, X86::MOV16rm,
10195 X86::NOT16r, X86::AX,
10196 X86::GR16RegisterClass);
10197 case X86::ATOMOR16:
10198 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
10199 X86::OR16ri, X86::MOV16rm,
10201 X86::NOT16r, X86::AX,
10202 X86::GR16RegisterClass);
10203 case X86::ATOMXOR16:
10204 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
10205 X86::XOR16ri, X86::MOV16rm,
10207 X86::NOT16r, X86::AX,
10208 X86::GR16RegisterClass);
10209 case X86::ATOMNAND16:
10210 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
10211 X86::AND16ri, X86::MOV16rm,
10213 X86::NOT16r, X86::AX,
10214 X86::GR16RegisterClass, true);
10215 case X86::ATOMMIN16:
10216 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
10217 case X86::ATOMMAX16:
10218 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
10219 case X86::ATOMUMIN16:
10220 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
10221 case X86::ATOMUMAX16:
10222 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
10224 case X86::ATOMAND8:
10225 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
10226 X86::AND8ri, X86::MOV8rm,
10228 X86::NOT8r, X86::AL,
10229 X86::GR8RegisterClass);
10231 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
10232 X86::OR8ri, X86::MOV8rm,
10234 X86::NOT8r, X86::AL,
10235 X86::GR8RegisterClass);
10236 case X86::ATOMXOR8:
10237 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
10238 X86::XOR8ri, X86::MOV8rm,
10240 X86::NOT8r, X86::AL,
10241 X86::GR8RegisterClass);
10242 case X86::ATOMNAND8:
10243 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
10244 X86::AND8ri, X86::MOV8rm,
10246 X86::NOT8r, X86::AL,
10247 X86::GR8RegisterClass, true);
10248 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
10249 // This group is for 64-bit host.
10250 case X86::ATOMAND64:
10251 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
10252 X86::AND64ri32, X86::MOV64rm,
10254 X86::NOT64r, X86::RAX,
10255 X86::GR64RegisterClass);
10256 case X86::ATOMOR64:
10257 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
10258 X86::OR64ri32, X86::MOV64rm,
10260 X86::NOT64r, X86::RAX,
10261 X86::GR64RegisterClass);
10262 case X86::ATOMXOR64:
10263 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
10264 X86::XOR64ri32, X86::MOV64rm,
10266 X86::NOT64r, X86::RAX,
10267 X86::GR64RegisterClass);
10268 case X86::ATOMNAND64:
10269 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
10270 X86::AND64ri32, X86::MOV64rm,
10272 X86::NOT64r, X86::RAX,
10273 X86::GR64RegisterClass, true);
10274 case X86::ATOMMIN64:
10275 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
10276 case X86::ATOMMAX64:
10277 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
10278 case X86::ATOMUMIN64:
10279 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
10280 case X86::ATOMUMAX64:
10281 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
10283 // This group does 64-bit operations on a 32-bit host.
10284 case X86::ATOMAND6432:
10285 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10286 X86::AND32rr, X86::AND32rr,
10287 X86::AND32ri, X86::AND32ri,
10289 case X86::ATOMOR6432:
10290 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10291 X86::OR32rr, X86::OR32rr,
10292 X86::OR32ri, X86::OR32ri,
10294 case X86::ATOMXOR6432:
10295 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10296 X86::XOR32rr, X86::XOR32rr,
10297 X86::XOR32ri, X86::XOR32ri,
10299 case X86::ATOMNAND6432:
10300 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10301 X86::AND32rr, X86::AND32rr,
10302 X86::AND32ri, X86::AND32ri,
10304 case X86::ATOMADD6432:
10305 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10306 X86::ADD32rr, X86::ADC32rr,
10307 X86::ADD32ri, X86::ADC32ri,
10309 case X86::ATOMSUB6432:
10310 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10311 X86::SUB32rr, X86::SBB32rr,
10312 X86::SUB32ri, X86::SBB32ri,
10314 case X86::ATOMSWAP6432:
10315 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10316 X86::MOV32rr, X86::MOV32rr,
10317 X86::MOV32ri, X86::MOV32ri,
10319 case X86::VASTART_SAVE_XMM_REGS:
10320 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
10322 case X86::VAARG_64:
10323 return EmitVAARG64WithCustomInserter(MI, BB);
10327 //===----------------------------------------------------------------------===//
10328 // X86 Optimization Hooks
10329 //===----------------------------------------------------------------------===//
10331 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
10335 const SelectionDAG &DAG,
10336 unsigned Depth) const {
10337 unsigned Opc = Op.getOpcode();
10338 assert((Opc >= ISD::BUILTIN_OP_END ||
10339 Opc == ISD::INTRINSIC_WO_CHAIN ||
10340 Opc == ISD::INTRINSIC_W_CHAIN ||
10341 Opc == ISD::INTRINSIC_VOID) &&
10342 "Should use MaskedValueIsZero if you don't know whether Op"
10343 " is a target node!");
10345 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
10357 // These nodes' second result is a boolean.
10358 if (Op.getResNo() == 0)
10361 case X86ISD::SETCC:
10362 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
10363 Mask.getBitWidth() - 1);
10368 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
10369 unsigned Depth) const {
10370 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
10371 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
10372 return Op.getValueType().getScalarType().getSizeInBits();
10378 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
10379 /// node is a GlobalAddress + offset.
10380 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
10381 const GlobalValue* &GA,
10382 int64_t &Offset) const {
10383 if (N->getOpcode() == X86ISD::Wrapper) {
10384 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
10385 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
10386 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
10390 return TargetLowering::isGAPlusOffset(N, GA, Offset);
10393 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
10394 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
10395 /// if the load addresses are consecutive, non-overlapping, and in the right
10397 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
10398 const TargetLowering &TLI) {
10399 DebugLoc dl = N->getDebugLoc();
10400 EVT VT = N->getValueType(0);
10402 if (VT.getSizeInBits() != 128)
10405 SmallVector<SDValue, 16> Elts;
10406 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
10407 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
10409 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
10412 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
10413 /// generation and convert it from being a bunch of shuffles and extracts
10414 /// to a simple store and scalar loads to extract the elements.
10415 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
10416 const TargetLowering &TLI) {
10417 SDValue InputVector = N->getOperand(0);
10419 // Only operate on vectors of 4 elements, where the alternative shuffling
10420 // gets to be more expensive.
10421 if (InputVector.getValueType() != MVT::v4i32)
10424 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
10425 // single use which is a sign-extend or zero-extend, and all elements are
10427 SmallVector<SDNode *, 4> Uses;
10428 unsigned ExtractedElements = 0;
10429 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
10430 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
10431 if (UI.getUse().getResNo() != InputVector.getResNo())
10434 SDNode *Extract = *UI;
10435 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
10438 if (Extract->getValueType(0) != MVT::i32)
10440 if (!Extract->hasOneUse())
10442 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
10443 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
10445 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
10448 // Record which element was extracted.
10449 ExtractedElements |=
10450 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
10452 Uses.push_back(Extract);
10455 // If not all the elements were used, this may not be worthwhile.
10456 if (ExtractedElements != 15)
10459 // Ok, we've now decided to do the transformation.
10460 DebugLoc dl = InputVector.getDebugLoc();
10462 // Store the value to a temporary stack slot.
10463 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
10464 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
10465 MachinePointerInfo(), false, false, 0);
10467 // Replace each use (extract) with a load of the appropriate element.
10468 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
10469 UE = Uses.end(); UI != UE; ++UI) {
10470 SDNode *Extract = *UI;
10472 // Compute the element's address.
10473 SDValue Idx = Extract->getOperand(1);
10475 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
10476 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
10477 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
10479 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
10480 StackPtr, OffsetVal);
10482 // Load the scalar.
10483 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
10484 ScalarAddr, MachinePointerInfo(),
10487 // Replace the exact with the load.
10488 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
10491 // The replacement was made in place; don't return anything.
10495 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
10496 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
10497 const X86Subtarget *Subtarget) {
10498 DebugLoc DL = N->getDebugLoc();
10499 SDValue Cond = N->getOperand(0);
10500 // Get the LHS/RHS of the select.
10501 SDValue LHS = N->getOperand(1);
10502 SDValue RHS = N->getOperand(2);
10504 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
10505 // instructions match the semantics of the common C idiom x<y?x:y but not
10506 // x<=y?x:y, because of how they handle negative zero (which can be
10507 // ignored in unsafe-math mode).
10508 if (Subtarget->hasSSE2() &&
10509 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
10510 Cond.getOpcode() == ISD::SETCC) {
10511 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
10513 unsigned Opcode = 0;
10514 // Check for x CC y ? x : y.
10515 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
10516 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
10520 // Converting this to a min would handle NaNs incorrectly, and swapping
10521 // the operands would cause it to handle comparisons between positive
10522 // and negative zero incorrectly.
10523 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
10524 if (!UnsafeFPMath &&
10525 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10527 std::swap(LHS, RHS);
10529 Opcode = X86ISD::FMIN;
10532 // Converting this to a min would handle comparisons between positive
10533 // and negative zero incorrectly.
10534 if (!UnsafeFPMath &&
10535 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
10537 Opcode = X86ISD::FMIN;
10540 // Converting this to a min would handle both negative zeros and NaNs
10541 // incorrectly, but we can swap the operands to fix both.
10542 std::swap(LHS, RHS);
10546 Opcode = X86ISD::FMIN;
10550 // Converting this to a max would handle comparisons between positive
10551 // and negative zero incorrectly.
10552 if (!UnsafeFPMath &&
10553 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
10555 Opcode = X86ISD::FMAX;
10558 // Converting this to a max would handle NaNs incorrectly, and swapping
10559 // the operands would cause it to handle comparisons between positive
10560 // and negative zero incorrectly.
10561 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
10562 if (!UnsafeFPMath &&
10563 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10565 std::swap(LHS, RHS);
10567 Opcode = X86ISD::FMAX;
10570 // Converting this to a max would handle both negative zeros and NaNs
10571 // incorrectly, but we can swap the operands to fix both.
10572 std::swap(LHS, RHS);
10576 Opcode = X86ISD::FMAX;
10579 // Check for x CC y ? y : x -- a min/max with reversed arms.
10580 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
10581 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
10585 // Converting this to a min would handle comparisons between positive
10586 // and negative zero incorrectly, and swapping the operands would
10587 // cause it to handle NaNs incorrectly.
10588 if (!UnsafeFPMath &&
10589 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
10590 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10592 std::swap(LHS, RHS);
10594 Opcode = X86ISD::FMIN;
10597 // Converting this to a min would handle NaNs incorrectly.
10598 if (!UnsafeFPMath &&
10599 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
10601 Opcode = X86ISD::FMIN;
10604 // Converting this to a min would handle both negative zeros and NaNs
10605 // incorrectly, but we can swap the operands to fix both.
10606 std::swap(LHS, RHS);
10610 Opcode = X86ISD::FMIN;
10614 // Converting this to a max would handle NaNs incorrectly.
10615 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10617 Opcode = X86ISD::FMAX;
10620 // Converting this to a max would handle comparisons between positive
10621 // and negative zero incorrectly, and swapping the operands would
10622 // cause it to handle NaNs incorrectly.
10623 if (!UnsafeFPMath &&
10624 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
10625 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10627 std::swap(LHS, RHS);
10629 Opcode = X86ISD::FMAX;
10632 // Converting this to a max would handle both negative zeros and NaNs
10633 // incorrectly, but we can swap the operands to fix both.
10634 std::swap(LHS, RHS);
10638 Opcode = X86ISD::FMAX;
10644 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
10647 // If this is a select between two integer constants, try to do some
10649 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
10650 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
10651 // Don't do this for crazy integer types.
10652 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
10653 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
10654 // so that TrueC (the true value) is larger than FalseC.
10655 bool NeedsCondInvert = false;
10657 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
10658 // Efficiently invertible.
10659 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
10660 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
10661 isa<ConstantSDNode>(Cond.getOperand(1))))) {
10662 NeedsCondInvert = true;
10663 std::swap(TrueC, FalseC);
10666 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
10667 if (FalseC->getAPIntValue() == 0 &&
10668 TrueC->getAPIntValue().isPowerOf2()) {
10669 if (NeedsCondInvert) // Invert the condition if needed.
10670 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10671 DAG.getConstant(1, Cond.getValueType()));
10673 // Zero extend the condition if needed.
10674 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
10676 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10677 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
10678 DAG.getConstant(ShAmt, MVT::i8));
10681 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
10682 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10683 if (NeedsCondInvert) // Invert the condition if needed.
10684 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10685 DAG.getConstant(1, Cond.getValueType()));
10687 // Zero extend the condition if needed.
10688 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10689 FalseC->getValueType(0), Cond);
10690 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10691 SDValue(FalseC, 0));
10694 // Optimize cases that will turn into an LEA instruction. This requires
10695 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10696 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10697 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10698 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10700 bool isFastMultiplier = false;
10702 switch ((unsigned char)Diff) {
10704 case 1: // result = add base, cond
10705 case 2: // result = lea base( , cond*2)
10706 case 3: // result = lea base(cond, cond*2)
10707 case 4: // result = lea base( , cond*4)
10708 case 5: // result = lea base(cond, cond*4)
10709 case 8: // result = lea base( , cond*8)
10710 case 9: // result = lea base(cond, cond*8)
10711 isFastMultiplier = true;
10716 if (isFastMultiplier) {
10717 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10718 if (NeedsCondInvert) // Invert the condition if needed.
10719 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10720 DAG.getConstant(1, Cond.getValueType()));
10722 // Zero extend the condition if needed.
10723 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10725 // Scale the condition by the difference.
10727 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10728 DAG.getConstant(Diff, Cond.getValueType()));
10730 // Add the base if non-zero.
10731 if (FalseC->getAPIntValue() != 0)
10732 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10733 SDValue(FalseC, 0));
10743 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
10744 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
10745 TargetLowering::DAGCombinerInfo &DCI) {
10746 DebugLoc DL = N->getDebugLoc();
10748 // If the flag operand isn't dead, don't touch this CMOV.
10749 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
10752 // If this is a select between two integer constants, try to do some
10753 // optimizations. Note that the operands are ordered the opposite of SELECT
10755 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
10756 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
10757 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
10758 // larger than FalseC (the false value).
10759 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
10761 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
10762 CC = X86::GetOppositeBranchCondition(CC);
10763 std::swap(TrueC, FalseC);
10766 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
10767 // This is efficient for any integer data type (including i8/i16) and
10769 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
10770 SDValue Cond = N->getOperand(3);
10771 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10772 DAG.getConstant(CC, MVT::i8), Cond);
10774 // Zero extend the condition if needed.
10775 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
10777 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10778 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
10779 DAG.getConstant(ShAmt, MVT::i8));
10780 if (N->getNumValues() == 2) // Dead flag value?
10781 return DCI.CombineTo(N, Cond, SDValue());
10785 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
10786 // for any integer data type, including i8/i16.
10787 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10788 SDValue Cond = N->getOperand(3);
10789 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10790 DAG.getConstant(CC, MVT::i8), Cond);
10792 // Zero extend the condition if needed.
10793 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10794 FalseC->getValueType(0), Cond);
10795 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10796 SDValue(FalseC, 0));
10798 if (N->getNumValues() == 2) // Dead flag value?
10799 return DCI.CombineTo(N, Cond, SDValue());
10803 // Optimize cases that will turn into an LEA instruction. This requires
10804 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10805 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10806 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10807 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10809 bool isFastMultiplier = false;
10811 switch ((unsigned char)Diff) {
10813 case 1: // result = add base, cond
10814 case 2: // result = lea base( , cond*2)
10815 case 3: // result = lea base(cond, cond*2)
10816 case 4: // result = lea base( , cond*4)
10817 case 5: // result = lea base(cond, cond*4)
10818 case 8: // result = lea base( , cond*8)
10819 case 9: // result = lea base(cond, cond*8)
10820 isFastMultiplier = true;
10825 if (isFastMultiplier) {
10826 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10827 SDValue Cond = N->getOperand(3);
10828 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10829 DAG.getConstant(CC, MVT::i8), Cond);
10830 // Zero extend the condition if needed.
10831 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10833 // Scale the condition by the difference.
10835 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10836 DAG.getConstant(Diff, Cond.getValueType()));
10838 // Add the base if non-zero.
10839 if (FalseC->getAPIntValue() != 0)
10840 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10841 SDValue(FalseC, 0));
10842 if (N->getNumValues() == 2) // Dead flag value?
10843 return DCI.CombineTo(N, Cond, SDValue());
10853 /// PerformMulCombine - Optimize a single multiply with constant into two
10854 /// in order to implement it with two cheaper instructions, e.g.
10855 /// LEA + SHL, LEA + LEA.
10856 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
10857 TargetLowering::DAGCombinerInfo &DCI) {
10858 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
10861 EVT VT = N->getValueType(0);
10862 if (VT != MVT::i64)
10865 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
10868 uint64_t MulAmt = C->getZExtValue();
10869 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
10872 uint64_t MulAmt1 = 0;
10873 uint64_t MulAmt2 = 0;
10874 if ((MulAmt % 9) == 0) {
10876 MulAmt2 = MulAmt / 9;
10877 } else if ((MulAmt % 5) == 0) {
10879 MulAmt2 = MulAmt / 5;
10880 } else if ((MulAmt % 3) == 0) {
10882 MulAmt2 = MulAmt / 3;
10885 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
10886 DebugLoc DL = N->getDebugLoc();
10888 if (isPowerOf2_64(MulAmt2) &&
10889 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
10890 // If second multiplifer is pow2, issue it first. We want the multiply by
10891 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
10893 std::swap(MulAmt1, MulAmt2);
10896 if (isPowerOf2_64(MulAmt1))
10897 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
10898 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
10900 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
10901 DAG.getConstant(MulAmt1, VT));
10903 if (isPowerOf2_64(MulAmt2))
10904 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
10905 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
10907 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
10908 DAG.getConstant(MulAmt2, VT));
10910 // Do not add new nodes to DAG combiner worklist.
10911 DCI.CombineTo(N, NewMul, false);
10916 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
10917 SDValue N0 = N->getOperand(0);
10918 SDValue N1 = N->getOperand(1);
10919 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
10920 EVT VT = N0.getValueType();
10922 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
10923 // since the result of setcc_c is all zero's or all ones.
10924 if (N1C && N0.getOpcode() == ISD::AND &&
10925 N0.getOperand(1).getOpcode() == ISD::Constant) {
10926 SDValue N00 = N0.getOperand(0);
10927 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
10928 ((N00.getOpcode() == ISD::ANY_EXTEND ||
10929 N00.getOpcode() == ISD::ZERO_EXTEND) &&
10930 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
10931 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
10932 APInt ShAmt = N1C->getAPIntValue();
10933 Mask = Mask.shl(ShAmt);
10935 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
10936 N00, DAG.getConstant(Mask, VT));
10943 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
10945 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
10946 const X86Subtarget *Subtarget) {
10947 EVT VT = N->getValueType(0);
10948 if (!VT.isVector() && VT.isInteger() &&
10949 N->getOpcode() == ISD::SHL)
10950 return PerformSHLCombine(N, DAG);
10952 // On X86 with SSE2 support, we can transform this to a vector shift if
10953 // all elements are shifted by the same amount. We can't do this in legalize
10954 // because the a constant vector is typically transformed to a constant pool
10955 // so we have no knowledge of the shift amount.
10956 if (!Subtarget->hasSSE2())
10959 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
10962 SDValue ShAmtOp = N->getOperand(1);
10963 EVT EltVT = VT.getVectorElementType();
10964 DebugLoc DL = N->getDebugLoc();
10965 SDValue BaseShAmt = SDValue();
10966 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
10967 unsigned NumElts = VT.getVectorNumElements();
10969 for (; i != NumElts; ++i) {
10970 SDValue Arg = ShAmtOp.getOperand(i);
10971 if (Arg.getOpcode() == ISD::UNDEF) continue;
10975 for (; i != NumElts; ++i) {
10976 SDValue Arg = ShAmtOp.getOperand(i);
10977 if (Arg.getOpcode() == ISD::UNDEF) continue;
10978 if (Arg != BaseShAmt) {
10982 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
10983 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
10984 SDValue InVec = ShAmtOp.getOperand(0);
10985 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
10986 unsigned NumElts = InVec.getValueType().getVectorNumElements();
10988 for (; i != NumElts; ++i) {
10989 SDValue Arg = InVec.getOperand(i);
10990 if (Arg.getOpcode() == ISD::UNDEF) continue;
10994 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
10995 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
10996 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
10997 if (C->getZExtValue() == SplatIdx)
10998 BaseShAmt = InVec.getOperand(1);
11001 if (BaseShAmt.getNode() == 0)
11002 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
11003 DAG.getIntPtrConstant(0));
11007 // The shift amount is an i32.
11008 if (EltVT.bitsGT(MVT::i32))
11009 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
11010 else if (EltVT.bitsLT(MVT::i32))
11011 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
11013 // The shift amount is identical so we can do a vector shift.
11014 SDValue ValOp = N->getOperand(0);
11015 switch (N->getOpcode()) {
11017 llvm_unreachable("Unknown shift opcode!");
11020 if (VT == MVT::v2i64)
11021 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11022 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
11024 if (VT == MVT::v4i32)
11025 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11026 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
11028 if (VT == MVT::v8i16)
11029 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11030 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
11034 if (VT == MVT::v4i32)
11035 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11036 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
11038 if (VT == MVT::v8i16)
11039 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11040 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
11044 if (VT == MVT::v2i64)
11045 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11046 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
11048 if (VT == MVT::v4i32)
11049 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11050 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
11052 if (VT == MVT::v8i16)
11053 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11054 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
11062 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
11063 TargetLowering::DAGCombinerInfo &DCI,
11064 const X86Subtarget *Subtarget) {
11065 if (DCI.isBeforeLegalizeOps())
11068 // Want to form PANDN nodes, in the hopes of then easily combining them with
11069 // OR and AND nodes to form PBLEND/PSIGN.
11070 EVT VT = N->getValueType(0);
11071 if (VT != MVT::v2i64)
11074 SDValue N0 = N->getOperand(0);
11075 SDValue N1 = N->getOperand(1);
11076 DebugLoc DL = N->getDebugLoc();
11078 // Check LHS for vnot
11079 if (N0.getOpcode() == ISD::XOR &&
11080 ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
11081 return DAG.getNode(X86ISD::PANDN, DL, VT, N0.getOperand(0), N1);
11083 // Check RHS for vnot
11084 if (N1.getOpcode() == ISD::XOR &&
11085 ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
11086 return DAG.getNode(X86ISD::PANDN, DL, VT, N1.getOperand(0), N0);
11091 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
11092 TargetLowering::DAGCombinerInfo &DCI,
11093 const X86Subtarget *Subtarget) {
11094 if (DCI.isBeforeLegalizeOps())
11097 EVT VT = N->getValueType(0);
11098 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
11101 SDValue N0 = N->getOperand(0);
11102 SDValue N1 = N->getOperand(1);
11104 // look for psign/blend
11105 if (Subtarget->hasSSSE3()) {
11106 if (VT == MVT::v2i64) {
11107 // Canonicalize pandn to RHS
11108 if (N0.getOpcode() == X86ISD::PANDN)
11110 // or (and (m, x), (pandn m, y))
11111 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::PANDN) {
11112 SDValue Mask = N1.getOperand(0);
11113 SDValue X = N1.getOperand(1);
11115 if (N0.getOperand(0) == Mask)
11116 Y = N0.getOperand(1);
11117 if (N0.getOperand(1) == Mask)
11118 Y = N0.getOperand(0);
11120 // Check to see if the mask appeared in both the AND and PANDN and
11124 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
11125 if (Mask.getOpcode() != ISD::BITCAST ||
11126 X.getOpcode() != ISD::BITCAST ||
11127 Y.getOpcode() != ISD::BITCAST)
11130 // Look through mask bitcast.
11131 Mask = Mask.getOperand(0);
11132 EVT MaskVT = Mask.getValueType();
11134 // Validate that the Mask operand is a vector sra node. The sra node
11135 // will be an intrinsic.
11136 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
11139 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
11140 // there is no psrai.b
11141 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
11142 case Intrinsic::x86_sse2_psrai_w:
11143 case Intrinsic::x86_sse2_psrai_d:
11145 default: return SDValue();
11148 // Check that the SRA is all signbits.
11149 SDValue SraC = Mask.getOperand(2);
11150 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
11151 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
11152 if ((SraAmt + 1) != EltBits)
11155 DebugLoc DL = N->getDebugLoc();
11157 // Now we know we at least have a plendvb with the mask val. See if
11158 // we can form a psignb/w/d.
11159 // psign = x.type == y.type == mask.type && y = sub(0, x);
11160 X = X.getOperand(0);
11161 Y = Y.getOperand(0);
11162 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
11163 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
11164 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
11167 case 8: Opc = X86ISD::PSIGNB; break;
11168 case 16: Opc = X86ISD::PSIGNW; break;
11169 case 32: Opc = X86ISD::PSIGND; break;
11173 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
11174 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
11177 // PBLENDVB only available on SSE 4.1
11178 if (!Subtarget->hasSSE41())
11181 unsigned IID = Intrinsic::x86_sse41_pblendvb;
11182 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
11183 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
11184 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
11185 Mask = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, MVT::v16i8,
11186 DAG.getConstant(IID, MVT::i32), X, Y, Mask);
11187 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
11192 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
11193 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
11195 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
11197 if (!N0.hasOneUse() || !N1.hasOneUse())
11200 SDValue ShAmt0 = N0.getOperand(1);
11201 if (ShAmt0.getValueType() != MVT::i8)
11203 SDValue ShAmt1 = N1.getOperand(1);
11204 if (ShAmt1.getValueType() != MVT::i8)
11206 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
11207 ShAmt0 = ShAmt0.getOperand(0);
11208 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
11209 ShAmt1 = ShAmt1.getOperand(0);
11211 DebugLoc DL = N->getDebugLoc();
11212 unsigned Opc = X86ISD::SHLD;
11213 SDValue Op0 = N0.getOperand(0);
11214 SDValue Op1 = N1.getOperand(0);
11215 if (ShAmt0.getOpcode() == ISD::SUB) {
11216 Opc = X86ISD::SHRD;
11217 std::swap(Op0, Op1);
11218 std::swap(ShAmt0, ShAmt1);
11221 unsigned Bits = VT.getSizeInBits();
11222 if (ShAmt1.getOpcode() == ISD::SUB) {
11223 SDValue Sum = ShAmt1.getOperand(0);
11224 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
11225 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
11226 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
11227 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
11228 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
11229 return DAG.getNode(Opc, DL, VT,
11231 DAG.getNode(ISD::TRUNCATE, DL,
11234 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
11235 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
11237 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
11238 return DAG.getNode(Opc, DL, VT,
11239 N0.getOperand(0), N1.getOperand(0),
11240 DAG.getNode(ISD::TRUNCATE, DL,
11247 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
11248 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
11249 const X86Subtarget *Subtarget) {
11250 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
11251 // the FP state in cases where an emms may be missing.
11252 // A preferable solution to the general problem is to figure out the right
11253 // places to insert EMMS. This qualifies as a quick hack.
11255 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
11256 StoreSDNode *St = cast<StoreSDNode>(N);
11257 EVT VT = St->getValue().getValueType();
11258 if (VT.getSizeInBits() != 64)
11261 const Function *F = DAG.getMachineFunction().getFunction();
11262 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
11263 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
11264 && Subtarget->hasSSE2();
11265 if ((VT.isVector() ||
11266 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
11267 isa<LoadSDNode>(St->getValue()) &&
11268 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
11269 St->getChain().hasOneUse() && !St->isVolatile()) {
11270 SDNode* LdVal = St->getValue().getNode();
11271 LoadSDNode *Ld = 0;
11272 int TokenFactorIndex = -1;
11273 SmallVector<SDValue, 8> Ops;
11274 SDNode* ChainVal = St->getChain().getNode();
11275 // Must be a store of a load. We currently handle two cases: the load
11276 // is a direct child, and it's under an intervening TokenFactor. It is
11277 // possible to dig deeper under nested TokenFactors.
11278 if (ChainVal == LdVal)
11279 Ld = cast<LoadSDNode>(St->getChain());
11280 else if (St->getValue().hasOneUse() &&
11281 ChainVal->getOpcode() == ISD::TokenFactor) {
11282 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
11283 if (ChainVal->getOperand(i).getNode() == LdVal) {
11284 TokenFactorIndex = i;
11285 Ld = cast<LoadSDNode>(St->getValue());
11287 Ops.push_back(ChainVal->getOperand(i));
11291 if (!Ld || !ISD::isNormalLoad(Ld))
11294 // If this is not the MMX case, i.e. we are just turning i64 load/store
11295 // into f64 load/store, avoid the transformation if there are multiple
11296 // uses of the loaded value.
11297 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
11300 DebugLoc LdDL = Ld->getDebugLoc();
11301 DebugLoc StDL = N->getDebugLoc();
11302 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
11303 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
11305 if (Subtarget->is64Bit() || F64IsLegal) {
11306 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
11307 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
11308 Ld->getPointerInfo(), Ld->isVolatile(),
11309 Ld->isNonTemporal(), Ld->getAlignment());
11310 SDValue NewChain = NewLd.getValue(1);
11311 if (TokenFactorIndex != -1) {
11312 Ops.push_back(NewChain);
11313 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
11316 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
11317 St->getPointerInfo(),
11318 St->isVolatile(), St->isNonTemporal(),
11319 St->getAlignment());
11322 // Otherwise, lower to two pairs of 32-bit loads / stores.
11323 SDValue LoAddr = Ld->getBasePtr();
11324 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
11325 DAG.getConstant(4, MVT::i32));
11327 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
11328 Ld->getPointerInfo(),
11329 Ld->isVolatile(), Ld->isNonTemporal(),
11330 Ld->getAlignment());
11331 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
11332 Ld->getPointerInfo().getWithOffset(4),
11333 Ld->isVolatile(), Ld->isNonTemporal(),
11334 MinAlign(Ld->getAlignment(), 4));
11336 SDValue NewChain = LoLd.getValue(1);
11337 if (TokenFactorIndex != -1) {
11338 Ops.push_back(LoLd);
11339 Ops.push_back(HiLd);
11340 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
11344 LoAddr = St->getBasePtr();
11345 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
11346 DAG.getConstant(4, MVT::i32));
11348 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
11349 St->getPointerInfo(),
11350 St->isVolatile(), St->isNonTemporal(),
11351 St->getAlignment());
11352 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
11353 St->getPointerInfo().getWithOffset(4),
11355 St->isNonTemporal(),
11356 MinAlign(St->getAlignment(), 4));
11357 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
11362 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
11363 /// X86ISD::FXOR nodes.
11364 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
11365 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
11366 // F[X]OR(0.0, x) -> x
11367 // F[X]OR(x, 0.0) -> x
11368 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
11369 if (C->getValueAPF().isPosZero())
11370 return N->getOperand(1);
11371 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
11372 if (C->getValueAPF().isPosZero())
11373 return N->getOperand(0);
11377 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
11378 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
11379 // FAND(0.0, x) -> 0.0
11380 // FAND(x, 0.0) -> 0.0
11381 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
11382 if (C->getValueAPF().isPosZero())
11383 return N->getOperand(0);
11384 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
11385 if (C->getValueAPF().isPosZero())
11386 return N->getOperand(1);
11390 static SDValue PerformBTCombine(SDNode *N,
11392 TargetLowering::DAGCombinerInfo &DCI) {
11393 // BT ignores high bits in the bit index operand.
11394 SDValue Op1 = N->getOperand(1);
11395 if (Op1.hasOneUse()) {
11396 unsigned BitWidth = Op1.getValueSizeInBits();
11397 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
11398 APInt KnownZero, KnownOne;
11399 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
11400 !DCI.isBeforeLegalizeOps());
11401 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11402 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
11403 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
11404 DCI.CommitTargetLoweringOpt(TLO);
11409 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
11410 SDValue Op = N->getOperand(0);
11411 if (Op.getOpcode() == ISD::BITCAST)
11412 Op = Op.getOperand(0);
11413 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
11414 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
11415 VT.getVectorElementType().getSizeInBits() ==
11416 OpVT.getVectorElementType().getSizeInBits()) {
11417 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
11422 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
11423 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
11424 // (and (i32 x86isd::setcc_carry), 1)
11425 // This eliminates the zext. This transformation is necessary because
11426 // ISD::SETCC is always legalized to i8.
11427 DebugLoc dl = N->getDebugLoc();
11428 SDValue N0 = N->getOperand(0);
11429 EVT VT = N->getValueType(0);
11430 if (N0.getOpcode() == ISD::AND &&
11432 N0.getOperand(0).hasOneUse()) {
11433 SDValue N00 = N0.getOperand(0);
11434 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
11436 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
11437 if (!C || C->getZExtValue() != 1)
11439 return DAG.getNode(ISD::AND, dl, VT,
11440 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
11441 N00.getOperand(0), N00.getOperand(1)),
11442 DAG.getConstant(1, VT));
11448 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
11449 DAGCombinerInfo &DCI) const {
11450 SelectionDAG &DAG = DCI.DAG;
11451 switch (N->getOpcode()) {
11453 case ISD::EXTRACT_VECTOR_ELT:
11454 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
11455 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
11456 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
11457 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
11460 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
11461 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
11462 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
11463 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
11465 case X86ISD::FOR: return PerformFORCombine(N, DAG);
11466 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
11467 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
11468 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
11469 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
11470 case X86ISD::SHUFPS: // Handle all target specific shuffles
11471 case X86ISD::SHUFPD:
11472 case X86ISD::PALIGN:
11473 case X86ISD::PUNPCKHBW:
11474 case X86ISD::PUNPCKHWD:
11475 case X86ISD::PUNPCKHDQ:
11476 case X86ISD::PUNPCKHQDQ:
11477 case X86ISD::UNPCKHPS:
11478 case X86ISD::UNPCKHPD:
11479 case X86ISD::PUNPCKLBW:
11480 case X86ISD::PUNPCKLWD:
11481 case X86ISD::PUNPCKLDQ:
11482 case X86ISD::PUNPCKLQDQ:
11483 case X86ISD::UNPCKLPS:
11484 case X86ISD::UNPCKLPD:
11485 case X86ISD::MOVHLPS:
11486 case X86ISD::MOVLHPS:
11487 case X86ISD::PSHUFD:
11488 case X86ISD::PSHUFHW:
11489 case X86ISD::PSHUFLW:
11490 case X86ISD::MOVSS:
11491 case X86ISD::MOVSD:
11492 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
11498 /// isTypeDesirableForOp - Return true if the target has native support for
11499 /// the specified value type and it is 'desirable' to use the type for the
11500 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
11501 /// instruction encodings are longer and some i16 instructions are slow.
11502 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
11503 if (!isTypeLegal(VT))
11505 if (VT != MVT::i16)
11512 case ISD::SIGN_EXTEND:
11513 case ISD::ZERO_EXTEND:
11514 case ISD::ANY_EXTEND:
11527 /// IsDesirableToPromoteOp - This method query the target whether it is
11528 /// beneficial for dag combiner to promote the specified node. If true, it
11529 /// should return the desired promotion type by reference.
11530 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
11531 EVT VT = Op.getValueType();
11532 if (VT != MVT::i16)
11535 bool Promote = false;
11536 bool Commute = false;
11537 switch (Op.getOpcode()) {
11540 LoadSDNode *LD = cast<LoadSDNode>(Op);
11541 // If the non-extending load has a single use and it's not live out, then it
11542 // might be folded.
11543 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
11544 Op.hasOneUse()*/) {
11545 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
11546 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
11547 // The only case where we'd want to promote LOAD (rather then it being
11548 // promoted as an operand is when it's only use is liveout.
11549 if (UI->getOpcode() != ISD::CopyToReg)
11556 case ISD::SIGN_EXTEND:
11557 case ISD::ZERO_EXTEND:
11558 case ISD::ANY_EXTEND:
11563 SDValue N0 = Op.getOperand(0);
11564 // Look out for (store (shl (load), x)).
11565 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
11578 SDValue N0 = Op.getOperand(0);
11579 SDValue N1 = Op.getOperand(1);
11580 if (!Commute && MayFoldLoad(N1))
11582 // Avoid disabling potential load folding opportunities.
11583 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
11585 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
11595 //===----------------------------------------------------------------------===//
11596 // X86 Inline Assembly Support
11597 //===----------------------------------------------------------------------===//
11599 static bool LowerToBSwap(CallInst *CI) {
11600 // FIXME: this should verify that we are targetting a 486 or better. If not,
11601 // we will turn this bswap into something that will be lowered to logical ops
11602 // instead of emitting the bswap asm. For now, we don't support 486 or lower
11603 // so don't worry about this.
11605 // Verify this is a simple bswap.
11606 if (CI->getNumArgOperands() != 1 ||
11607 CI->getType() != CI->getArgOperand(0)->getType() ||
11608 !CI->getType()->isIntegerTy())
11611 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
11612 if (!Ty || Ty->getBitWidth() % 16 != 0)
11615 // Okay, we can do this xform, do so now.
11616 const Type *Tys[] = { Ty };
11617 Module *M = CI->getParent()->getParent()->getParent();
11618 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
11620 Value *Op = CI->getArgOperand(0);
11621 Op = CallInst::Create(Int, Op, CI->getName(), CI);
11623 CI->replaceAllUsesWith(Op);
11624 CI->eraseFromParent();
11628 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
11629 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
11630 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
11632 std::string AsmStr = IA->getAsmString();
11634 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
11635 SmallVector<StringRef, 4> AsmPieces;
11636 SplitString(AsmStr, AsmPieces, ";\n");
11638 switch (AsmPieces.size()) {
11639 default: return false;
11641 AsmStr = AsmPieces[0];
11643 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
11646 if (AsmPieces.size() == 2 &&
11647 (AsmPieces[0] == "bswap" ||
11648 AsmPieces[0] == "bswapq" ||
11649 AsmPieces[0] == "bswapl") &&
11650 (AsmPieces[1] == "$0" ||
11651 AsmPieces[1] == "${0:q}")) {
11652 // No need to check constraints, nothing other than the equivalent of
11653 // "=r,0" would be valid here.
11654 return LowerToBSwap(CI);
11656 // rorw $$8, ${0:w} --> llvm.bswap.i16
11657 if (CI->getType()->isIntegerTy(16) &&
11658 AsmPieces.size() == 3 &&
11659 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
11660 AsmPieces[1] == "$$8," &&
11661 AsmPieces[2] == "${0:w}" &&
11662 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
11664 const std::string &Constraints = IA->getConstraintString();
11665 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
11666 std::sort(AsmPieces.begin(), AsmPieces.end());
11667 if (AsmPieces.size() == 4 &&
11668 AsmPieces[0] == "~{cc}" &&
11669 AsmPieces[1] == "~{dirflag}" &&
11670 AsmPieces[2] == "~{flags}" &&
11671 AsmPieces[3] == "~{fpsr}") {
11672 return LowerToBSwap(CI);
11677 if (CI->getType()->isIntegerTy(32) &&
11678 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
11679 SmallVector<StringRef, 4> Words;
11680 SplitString(AsmPieces[0], Words, " \t,");
11681 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
11682 Words[2] == "${0:w}") {
11684 SplitString(AsmPieces[1], Words, " \t,");
11685 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
11686 Words[2] == "$0") {
11688 SplitString(AsmPieces[2], Words, " \t,");
11689 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
11690 Words[2] == "${0:w}") {
11692 const std::string &Constraints = IA->getConstraintString();
11693 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
11694 std::sort(AsmPieces.begin(), AsmPieces.end());
11695 if (AsmPieces.size() == 4 &&
11696 AsmPieces[0] == "~{cc}" &&
11697 AsmPieces[1] == "~{dirflag}" &&
11698 AsmPieces[2] == "~{flags}" &&
11699 AsmPieces[3] == "~{fpsr}") {
11700 return LowerToBSwap(CI);
11706 if (CI->getType()->isIntegerTy(64) &&
11707 Constraints.size() >= 2 &&
11708 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
11709 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
11710 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
11711 SmallVector<StringRef, 4> Words;
11712 SplitString(AsmPieces[0], Words, " \t");
11713 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
11715 SplitString(AsmPieces[1], Words, " \t");
11716 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
11718 SplitString(AsmPieces[2], Words, " \t,");
11719 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
11720 Words[2] == "%edx") {
11721 return LowerToBSwap(CI);
11733 /// getConstraintType - Given a constraint letter, return the type of
11734 /// constraint it is for this target.
11735 X86TargetLowering::ConstraintType
11736 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
11737 if (Constraint.size() == 1) {
11738 switch (Constraint[0]) {
11748 return C_RegisterClass;
11772 return TargetLowering::getConstraintType(Constraint);
11775 /// Examine constraint type and operand type and determine a weight value.
11776 /// This object must already have been set up with the operand type
11777 /// and the current alternative constraint selected.
11778 TargetLowering::ConstraintWeight
11779 X86TargetLowering::getSingleConstraintMatchWeight(
11780 AsmOperandInfo &info, const char *constraint) const {
11781 ConstraintWeight weight = CW_Invalid;
11782 Value *CallOperandVal = info.CallOperandVal;
11783 // If we don't have a value, we can't do a match,
11784 // but allow it at the lowest weight.
11785 if (CallOperandVal == NULL)
11787 const Type *type = CallOperandVal->getType();
11788 // Look at the constraint type.
11789 switch (*constraint) {
11791 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
11802 if (CallOperandVal->getType()->isIntegerTy())
11803 weight = CW_SpecificReg;
11808 if (type->isFloatingPointTy())
11809 weight = CW_SpecificReg;
11812 if (type->isX86_MMXTy() && !DisableMMX && Subtarget->hasMMX())
11813 weight = CW_SpecificReg;
11817 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
11818 weight = CW_Register;
11821 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
11822 if (C->getZExtValue() <= 31)
11823 weight = CW_Constant;
11827 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
11828 if (C->getZExtValue() <= 63)
11829 weight = CW_Constant;
11833 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
11834 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
11835 weight = CW_Constant;
11839 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
11840 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
11841 weight = CW_Constant;
11845 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
11846 if (C->getZExtValue() <= 3)
11847 weight = CW_Constant;
11851 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
11852 if (C->getZExtValue() <= 0xff)
11853 weight = CW_Constant;
11858 if (dyn_cast<ConstantFP>(CallOperandVal)) {
11859 weight = CW_Constant;
11863 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
11864 if ((C->getSExtValue() >= -0x80000000LL) &&
11865 (C->getSExtValue() <= 0x7fffffffLL))
11866 weight = CW_Constant;
11870 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
11871 if (C->getZExtValue() <= 0xffffffff)
11872 weight = CW_Constant;
11879 /// LowerXConstraint - try to replace an X constraint, which matches anything,
11880 /// with another that has more specific requirements based on the type of the
11881 /// corresponding operand.
11882 const char *X86TargetLowering::
11883 LowerXConstraint(EVT ConstraintVT) const {
11884 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
11885 // 'f' like normal targets.
11886 if (ConstraintVT.isFloatingPoint()) {
11887 if (Subtarget->hasXMMInt())
11889 if (Subtarget->hasXMM())
11893 return TargetLowering::LowerXConstraint(ConstraintVT);
11896 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
11897 /// vector. If it is invalid, don't add anything to Ops.
11898 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
11900 std::vector<SDValue>&Ops,
11901 SelectionDAG &DAG) const {
11902 SDValue Result(0, 0);
11904 switch (Constraint) {
11907 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11908 if (C->getZExtValue() <= 31) {
11909 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11915 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11916 if (C->getZExtValue() <= 63) {
11917 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11923 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11924 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
11925 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11931 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11932 if (C->getZExtValue() <= 255) {
11933 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11939 // 32-bit signed value
11940 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11941 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
11942 C->getSExtValue())) {
11943 // Widen to 64 bits here to get it sign extended.
11944 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
11947 // FIXME gcc accepts some relocatable values here too, but only in certain
11948 // memory models; it's complicated.
11953 // 32-bit unsigned value
11954 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11955 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
11956 C->getZExtValue())) {
11957 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11961 // FIXME gcc accepts some relocatable values here too, but only in certain
11962 // memory models; it's complicated.
11966 // Literal immediates are always ok.
11967 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
11968 // Widen to 64 bits here to get it sign extended.
11969 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
11973 // In any sort of PIC mode addresses need to be computed at runtime by
11974 // adding in a register or some sort of table lookup. These can't
11975 // be used as immediates.
11976 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
11979 // If we are in non-pic codegen mode, we allow the address of a global (with
11980 // an optional displacement) to be used with 'i'.
11981 GlobalAddressSDNode *GA = 0;
11982 int64_t Offset = 0;
11984 // Match either (GA), (GA+C), (GA+C1+C2), etc.
11986 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
11987 Offset += GA->getOffset();
11989 } else if (Op.getOpcode() == ISD::ADD) {
11990 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
11991 Offset += C->getZExtValue();
11992 Op = Op.getOperand(0);
11995 } else if (Op.getOpcode() == ISD::SUB) {
11996 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
11997 Offset += -C->getZExtValue();
11998 Op = Op.getOperand(0);
12003 // Otherwise, this isn't something we can handle, reject it.
12007 const GlobalValue *GV = GA->getGlobal();
12008 // If we require an extra load to get this address, as in PIC mode, we
12009 // can't accept it.
12010 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
12011 getTargetMachine())))
12014 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
12015 GA->getValueType(0), Offset);
12020 if (Result.getNode()) {
12021 Ops.push_back(Result);
12024 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
12027 std::vector<unsigned> X86TargetLowering::
12028 getRegClassForInlineAsmConstraint(const std::string &Constraint,
12030 if (Constraint.size() == 1) {
12031 // FIXME: not handling fp-stack yet!
12032 switch (Constraint[0]) { // GCC X86 Constraint Letters
12033 default: break; // Unknown constraint letter
12034 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
12035 if (Subtarget->is64Bit()) {
12036 if (VT == MVT::i32)
12037 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
12038 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
12039 X86::R10D,X86::R11D,X86::R12D,
12040 X86::R13D,X86::R14D,X86::R15D,
12041 X86::EBP, X86::ESP, 0);
12042 else if (VT == MVT::i16)
12043 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
12044 X86::SI, X86::DI, X86::R8W,X86::R9W,
12045 X86::R10W,X86::R11W,X86::R12W,
12046 X86::R13W,X86::R14W,X86::R15W,
12047 X86::BP, X86::SP, 0);
12048 else if (VT == MVT::i8)
12049 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
12050 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
12051 X86::R10B,X86::R11B,X86::R12B,
12052 X86::R13B,X86::R14B,X86::R15B,
12053 X86::BPL, X86::SPL, 0);
12055 else if (VT == MVT::i64)
12056 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
12057 X86::RSI, X86::RDI, X86::R8, X86::R9,
12058 X86::R10, X86::R11, X86::R12,
12059 X86::R13, X86::R14, X86::R15,
12060 X86::RBP, X86::RSP, 0);
12064 // 32-bit fallthrough
12065 case 'Q': // Q_REGS
12066 if (VT == MVT::i32)
12067 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
12068 else if (VT == MVT::i16)
12069 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
12070 else if (VT == MVT::i8)
12071 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
12072 else if (VT == MVT::i64)
12073 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
12078 return std::vector<unsigned>();
12081 std::pair<unsigned, const TargetRegisterClass*>
12082 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
12084 // First, see if this is a constraint that directly corresponds to an LLVM
12086 if (Constraint.size() == 1) {
12087 // GCC Constraint Letters
12088 switch (Constraint[0]) {
12090 case 'r': // GENERAL_REGS
12091 case 'l': // INDEX_REGS
12093 return std::make_pair(0U, X86::GR8RegisterClass);
12094 if (VT == MVT::i16)
12095 return std::make_pair(0U, X86::GR16RegisterClass);
12096 if (VT == MVT::i32 || !Subtarget->is64Bit())
12097 return std::make_pair(0U, X86::GR32RegisterClass);
12098 return std::make_pair(0U, X86::GR64RegisterClass);
12099 case 'R': // LEGACY_REGS
12101 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
12102 if (VT == MVT::i16)
12103 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
12104 if (VT == MVT::i32 || !Subtarget->is64Bit())
12105 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
12106 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
12107 case 'f': // FP Stack registers.
12108 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
12109 // value to the correct fpstack register class.
12110 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
12111 return std::make_pair(0U, X86::RFP32RegisterClass);
12112 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
12113 return std::make_pair(0U, X86::RFP64RegisterClass);
12114 return std::make_pair(0U, X86::RFP80RegisterClass);
12115 case 'y': // MMX_REGS if MMX allowed.
12116 if (!Subtarget->hasMMX()) break;
12117 return std::make_pair(0U, X86::VR64RegisterClass);
12118 case 'Y': // SSE_REGS if SSE2 allowed
12119 if (!Subtarget->hasXMMInt()) break;
12121 case 'x': // SSE_REGS if SSE1 allowed
12122 if (!Subtarget->hasXMM()) break;
12124 switch (VT.getSimpleVT().SimpleTy) {
12126 // Scalar SSE types.
12129 return std::make_pair(0U, X86::FR32RegisterClass);
12132 return std::make_pair(0U, X86::FR64RegisterClass);
12140 return std::make_pair(0U, X86::VR128RegisterClass);
12146 // Use the default implementation in TargetLowering to convert the register
12147 // constraint into a member of a register class.
12148 std::pair<unsigned, const TargetRegisterClass*> Res;
12149 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
12151 // Not found as a standard register?
12152 if (Res.second == 0) {
12153 // Map st(0) -> st(7) -> ST0
12154 if (Constraint.size() == 7 && Constraint[0] == '{' &&
12155 tolower(Constraint[1]) == 's' &&
12156 tolower(Constraint[2]) == 't' &&
12157 Constraint[3] == '(' &&
12158 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
12159 Constraint[5] == ')' &&
12160 Constraint[6] == '}') {
12162 Res.first = X86::ST0+Constraint[4]-'0';
12163 Res.second = X86::RFP80RegisterClass;
12167 // GCC allows "st(0)" to be called just plain "st".
12168 if (StringRef("{st}").equals_lower(Constraint)) {
12169 Res.first = X86::ST0;
12170 Res.second = X86::RFP80RegisterClass;
12175 if (StringRef("{flags}").equals_lower(Constraint)) {
12176 Res.first = X86::EFLAGS;
12177 Res.second = X86::CCRRegisterClass;
12181 // 'A' means EAX + EDX.
12182 if (Constraint == "A") {
12183 Res.first = X86::EAX;
12184 Res.second = X86::GR32_ADRegisterClass;
12190 // Otherwise, check to see if this is a register class of the wrong value
12191 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
12192 // turn into {ax},{dx}.
12193 if (Res.second->hasType(VT))
12194 return Res; // Correct type already, nothing to do.
12196 // All of the single-register GCC register classes map their values onto
12197 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
12198 // really want an 8-bit or 32-bit register, map to the appropriate register
12199 // class and return the appropriate register.
12200 if (Res.second == X86::GR16RegisterClass) {
12201 if (VT == MVT::i8) {
12202 unsigned DestReg = 0;
12203 switch (Res.first) {
12205 case X86::AX: DestReg = X86::AL; break;
12206 case X86::DX: DestReg = X86::DL; break;
12207 case X86::CX: DestReg = X86::CL; break;
12208 case X86::BX: DestReg = X86::BL; break;
12211 Res.first = DestReg;
12212 Res.second = X86::GR8RegisterClass;
12214 } else if (VT == MVT::i32) {
12215 unsigned DestReg = 0;
12216 switch (Res.first) {
12218 case X86::AX: DestReg = X86::EAX; break;
12219 case X86::DX: DestReg = X86::EDX; break;
12220 case X86::CX: DestReg = X86::ECX; break;
12221 case X86::BX: DestReg = X86::EBX; break;
12222 case X86::SI: DestReg = X86::ESI; break;
12223 case X86::DI: DestReg = X86::EDI; break;
12224 case X86::BP: DestReg = X86::EBP; break;
12225 case X86::SP: DestReg = X86::ESP; break;
12228 Res.first = DestReg;
12229 Res.second = X86::GR32RegisterClass;
12231 } else if (VT == MVT::i64) {
12232 unsigned DestReg = 0;
12233 switch (Res.first) {
12235 case X86::AX: DestReg = X86::RAX; break;
12236 case X86::DX: DestReg = X86::RDX; break;
12237 case X86::CX: DestReg = X86::RCX; break;
12238 case X86::BX: DestReg = X86::RBX; break;
12239 case X86::SI: DestReg = X86::RSI; break;
12240 case X86::DI: DestReg = X86::RDI; break;
12241 case X86::BP: DestReg = X86::RBP; break;
12242 case X86::SP: DestReg = X86::RSP; break;
12245 Res.first = DestReg;
12246 Res.second = X86::GR64RegisterClass;
12249 } else if (Res.second == X86::FR32RegisterClass ||
12250 Res.second == X86::FR64RegisterClass ||
12251 Res.second == X86::VR128RegisterClass) {
12252 // Handle references to XMM physical registers that got mapped into the
12253 // wrong class. This can happen with constraints like {xmm0} where the
12254 // target independent register mapper will just pick the first match it can
12255 // find, ignoring the required type.
12256 if (VT == MVT::f32)
12257 Res.second = X86::FR32RegisterClass;
12258 else if (VT == MVT::f64)
12259 Res.second = X86::FR64RegisterClass;
12260 else if (X86::VR128RegisterClass->hasType(VT))
12261 Res.second = X86::VR128RegisterClass;