1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalAlias.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/LLVMContext.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/Dwarf.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/raw_ostream.h"
53 using namespace dwarf;
55 STATISTIC(NumTailCalls, "Number of tail calls");
58 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
60 // Forward declarations.
61 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
64 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
66 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
68 if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
69 if (is64Bit) return new X8664_MachoTargetObjectFile();
70 return new TargetLoweringObjectFileMachO();
71 } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
72 if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
73 return new X8632_ELFTargetObjectFile(TM);
74 } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
75 return new TargetLoweringObjectFileCOFF();
77 llvm_unreachable("unknown subtarget type");
80 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
81 : TargetLowering(TM, createTLOF(TM)) {
82 Subtarget = &TM.getSubtarget<X86Subtarget>();
83 X86ScalarSSEf64 = Subtarget->hasSSE2();
84 X86ScalarSSEf32 = Subtarget->hasSSE1();
85 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
87 RegInfo = TM.getRegisterInfo();
90 // Set up the TargetLowering object.
92 // X86 is weird, it always uses i8 for shift amounts and setcc results.
93 setShiftAmountType(MVT::i8);
94 setBooleanContents(ZeroOrOneBooleanContent);
95 setSchedulingPreference(Sched::RegPressure);
96 setStackPointerRegisterToSaveRestore(X86StackPtr);
98 if (Subtarget->isTargetDarwin()) {
99 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
100 setUseUnderscoreSetJmp(false);
101 setUseUnderscoreLongJmp(false);
102 } else if (Subtarget->isTargetMingw()) {
103 // MS runtime is weird: it exports _setjmp, but longjmp!
104 setUseUnderscoreSetJmp(true);
105 setUseUnderscoreLongJmp(false);
107 setUseUnderscoreSetJmp(true);
108 setUseUnderscoreLongJmp(true);
111 // Set up the register classes.
112 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
113 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
114 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
115 if (Subtarget->is64Bit())
116 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
118 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
120 // We don't accept any truncstore of integer registers.
121 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
122 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
123 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
124 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
125 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
128 // SETOEQ and SETUNE require checking two conditions.
129 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
130 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
131 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
132 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
133 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
136 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
138 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
139 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
140 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
142 if (Subtarget->is64Bit()) {
143 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
144 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
145 } else if (!UseSoftFloat) {
146 // We have an algorithm for SSE2->double, and we turn this into a
147 // 64-bit FILD followed by conditional FADD for other targets.
148 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
149 // We have an algorithm for SSE2, and we turn this into a 64-bit
150 // FILD for other targets.
151 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
154 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
156 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
157 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
160 // SSE has no i16 to fp conversion, only i32
161 if (X86ScalarSSEf32) {
162 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
163 // f32 and f64 cases are Legal, f80 case is not
164 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
166 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
167 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
170 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
171 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
174 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
175 // are Legal, f80 is custom lowered.
176 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
179 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
181 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
182 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
184 if (X86ScalarSSEf32) {
185 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
186 // f32 and f64 cases are Legal, f80 case is not
187 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
189 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
190 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
193 // Handle FP_TO_UINT by promoting the destination to a larger signed
195 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
196 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
197 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
199 if (Subtarget->is64Bit()) {
200 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
201 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
202 } else if (!UseSoftFloat) {
203 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
204 // Expand FP_TO_UINT into a select.
205 // FIXME: We would like to use a Custom expander here eventually to do
206 // the optimal thing for SSE vs. the default expansion in the legalizer.
207 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
209 // With SSE3 we can use fisttpll to convert to a signed i64; without
210 // SSE, we're stuck with a fistpll.
211 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
214 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
215 if (!X86ScalarSSEf64) {
216 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
217 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
218 if (Subtarget->is64Bit()) {
219 setOperationAction(ISD::BIT_CONVERT , MVT::f64 , Expand);
220 // Without SSE, i64->f64 goes through memory; i64->MMX is Legal.
221 if (Subtarget->hasMMX() && !DisableMMX)
222 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Custom);
224 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Expand);
228 // Scalar integer divide and remainder are lowered to use operations that
229 // produce two results, to match the available instructions. This exposes
230 // the two-result form to trivial CSE, which is able to combine x/y and x%y
231 // into a single instruction.
233 // Scalar integer multiply-high is also lowered to use two-result
234 // operations, to match the available instructions. However, plain multiply
235 // (low) operations are left as Legal, as there are single-result
236 // instructions for this in x86. Using the two-result multiply instructions
237 // when both high and low results are needed must be arranged by dagcombine.
238 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
239 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
240 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
241 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
242 setOperationAction(ISD::SREM , MVT::i8 , Expand);
243 setOperationAction(ISD::UREM , MVT::i8 , Expand);
244 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
245 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
246 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
247 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
248 setOperationAction(ISD::SREM , MVT::i16 , Expand);
249 setOperationAction(ISD::UREM , MVT::i16 , Expand);
250 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
251 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
252 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
253 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
254 setOperationAction(ISD::SREM , MVT::i32 , Expand);
255 setOperationAction(ISD::UREM , MVT::i32 , Expand);
256 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
257 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
258 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
259 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
260 setOperationAction(ISD::SREM , MVT::i64 , Expand);
261 setOperationAction(ISD::UREM , MVT::i64 , Expand);
263 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
264 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
265 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
266 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
267 if (Subtarget->is64Bit())
268 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
269 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
270 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
272 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
273 setOperationAction(ISD::FREM , MVT::f32 , Expand);
274 setOperationAction(ISD::FREM , MVT::f64 , Expand);
275 setOperationAction(ISD::FREM , MVT::f80 , Expand);
276 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
278 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
279 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
280 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
281 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
282 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
283 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
287 if (Subtarget->is64Bit()) {
288 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
293 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
294 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
296 // These should be promoted to a larger select which is supported.
297 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
298 // X86 wants to expand cmov itself.
299 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
300 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
301 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
302 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
303 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
304 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
305 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
306 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
307 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
308 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
309 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
310 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
311 if (Subtarget->is64Bit()) {
312 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
313 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
315 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
318 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
319 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
320 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
321 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
322 if (Subtarget->is64Bit())
323 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
324 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
325 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
326 if (Subtarget->is64Bit()) {
327 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
328 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
329 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
330 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
331 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
333 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
334 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
335 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
336 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
337 if (Subtarget->is64Bit()) {
338 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
339 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
340 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
343 if (Subtarget->hasSSE1())
344 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
346 if (!Subtarget->hasSSE2())
347 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
348 // On X86 and X86-64, atomic operations are lowered to locked instructions.
349 // Locked instructions, in turn, have implicit fence semantics (all memory
350 // operations are flushed before issuing the locked instruction, and they
351 // are not buffered), so we can fold away the common pattern of
352 // fence-atomic-fence.
353 setShouldFoldAtomicFences(true);
355 // Expand certain atomics
356 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
361 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
366 if (!Subtarget->is64Bit()) {
367 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
376 // FIXME - use subtarget debug flags
377 if (!Subtarget->isTargetDarwin() &&
378 !Subtarget->isTargetELF() &&
379 !Subtarget->isTargetCygMing()) {
380 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
383 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
384 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
385 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
386 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
387 if (Subtarget->is64Bit()) {
388 setExceptionPointerRegister(X86::RAX);
389 setExceptionSelectorRegister(X86::RDX);
391 setExceptionPointerRegister(X86::EAX);
392 setExceptionSelectorRegister(X86::EDX);
394 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
395 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
397 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
399 setOperationAction(ISD::TRAP, MVT::Other, Legal);
401 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
402 setOperationAction(ISD::VASTART , MVT::Other, Custom);
403 setOperationAction(ISD::VAEND , MVT::Other, Expand);
404 if (Subtarget->is64Bit()) {
405 setOperationAction(ISD::VAARG , MVT::Other, Custom);
406 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
408 setOperationAction(ISD::VAARG , MVT::Other, Expand);
409 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
412 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
413 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
414 if (Subtarget->is64Bit())
415 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
416 if (Subtarget->isTargetCygMing())
417 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
419 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
421 if (!UseSoftFloat && X86ScalarSSEf64) {
422 // f32 and f64 use SSE.
423 // Set up the FP register classes.
424 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
425 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
427 // Use ANDPD to simulate FABS.
428 setOperationAction(ISD::FABS , MVT::f64, Custom);
429 setOperationAction(ISD::FABS , MVT::f32, Custom);
431 // Use XORP to simulate FNEG.
432 setOperationAction(ISD::FNEG , MVT::f64, Custom);
433 setOperationAction(ISD::FNEG , MVT::f32, Custom);
435 // Use ANDPD and ORPD to simulate FCOPYSIGN.
436 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
437 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
439 // We don't support sin/cos/fmod
440 setOperationAction(ISD::FSIN , MVT::f64, Expand);
441 setOperationAction(ISD::FCOS , MVT::f64, Expand);
442 setOperationAction(ISD::FSIN , MVT::f32, Expand);
443 setOperationAction(ISD::FCOS , MVT::f32, Expand);
445 // Expand FP immediates into loads from the stack, except for the special
447 addLegalFPImmediate(APFloat(+0.0)); // xorpd
448 addLegalFPImmediate(APFloat(+0.0f)); // xorps
449 } else if (!UseSoftFloat && X86ScalarSSEf32) {
450 // Use SSE for f32, x87 for f64.
451 // Set up the FP register classes.
452 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
453 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
455 // Use ANDPS to simulate FABS.
456 setOperationAction(ISD::FABS , MVT::f32, Custom);
458 // Use XORP to simulate FNEG.
459 setOperationAction(ISD::FNEG , MVT::f32, Custom);
461 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
463 // Use ANDPS and ORPS to simulate FCOPYSIGN.
464 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
465 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
467 // We don't support sin/cos/fmod
468 setOperationAction(ISD::FSIN , MVT::f32, Expand);
469 setOperationAction(ISD::FCOS , MVT::f32, Expand);
471 // Special cases we handle for FP constants.
472 addLegalFPImmediate(APFloat(+0.0f)); // xorps
473 addLegalFPImmediate(APFloat(+0.0)); // FLD0
474 addLegalFPImmediate(APFloat(+1.0)); // FLD1
475 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
476 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
479 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
480 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
482 } else if (!UseSoftFloat) {
483 // f32 and f64 in x87.
484 // Set up the FP register classes.
485 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
486 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
488 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
489 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
490 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
491 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
494 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
495 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
497 addLegalFPImmediate(APFloat(+0.0)); // FLD0
498 addLegalFPImmediate(APFloat(+1.0)); // FLD1
499 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
500 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
501 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
502 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
503 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
504 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
507 // Long double always uses X87.
509 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
510 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
511 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
514 APFloat TmpFlt(+0.0);
515 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
517 addLegalFPImmediate(TmpFlt); // FLD0
519 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
520 APFloat TmpFlt2(+1.0);
521 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
523 addLegalFPImmediate(TmpFlt2); // FLD1
524 TmpFlt2.changeSign();
525 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
529 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
530 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
534 // Always use a library call for pow.
535 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
536 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
537 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
539 setOperationAction(ISD::FLOG, MVT::f80, Expand);
540 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
541 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
542 setOperationAction(ISD::FEXP, MVT::f80, Expand);
543 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
545 // First set operation action for all vector types to either promote
546 // (for widening) or expand (for scalarization). Then we will selectively
547 // turn on ones that can be effectively codegen'd.
548 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
549 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
550 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
565 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
566 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
599 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
603 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
604 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
605 setTruncStoreAction((MVT::SimpleValueType)VT,
606 (MVT::SimpleValueType)InnerVT, Expand);
607 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
608 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
609 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
612 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
613 // with -msoft-float, disable use of MMX as well.
614 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
615 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
616 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
617 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
619 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
621 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
622 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
623 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
624 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
626 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
627 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
628 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
629 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
631 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
632 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
634 setOperationAction(ISD::AND, MVT::v8i8, Promote);
635 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
636 setOperationAction(ISD::AND, MVT::v4i16, Promote);
637 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
638 setOperationAction(ISD::AND, MVT::v2i32, Promote);
639 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
640 setOperationAction(ISD::AND, MVT::v1i64, Legal);
642 setOperationAction(ISD::OR, MVT::v8i8, Promote);
643 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
644 setOperationAction(ISD::OR, MVT::v4i16, Promote);
645 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
646 setOperationAction(ISD::OR, MVT::v2i32, Promote);
647 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
648 setOperationAction(ISD::OR, MVT::v1i64, Legal);
650 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
651 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
652 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
653 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
654 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
655 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
656 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
658 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
659 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
660 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
661 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
662 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
663 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
664 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
666 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
667 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
669 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
671 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
672 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
674 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
676 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
677 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
678 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
680 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
682 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
683 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
684 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
685 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
686 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
690 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
691 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
692 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
693 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
694 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
698 if (!UseSoftFloat && Subtarget->hasSSE1()) {
699 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
701 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
702 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
703 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
704 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
705 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
706 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
707 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
708 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
709 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
710 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
711 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
712 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
715 if (!UseSoftFloat && Subtarget->hasSSE2()) {
716 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
718 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
719 // registers cannot be used even for integer operations.
720 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
721 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
722 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
723 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
725 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
726 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
727 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
728 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
729 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
730 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
731 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
732 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
733 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
734 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
735 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
736 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
737 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
738 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
739 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
740 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
742 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
743 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
744 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
745 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
747 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
748 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
749 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
750 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
751 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
753 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
754 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
755 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
756 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
757 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
759 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
760 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
761 EVT VT = (MVT::SimpleValueType)i;
762 // Do not attempt to custom lower non-power-of-2 vectors
763 if (!isPowerOf2_32(VT.getVectorNumElements()))
765 // Do not attempt to custom lower non-128-bit vectors
766 if (!VT.is128BitVector())
768 setOperationAction(ISD::BUILD_VECTOR,
769 VT.getSimpleVT().SimpleTy, Custom);
770 setOperationAction(ISD::VECTOR_SHUFFLE,
771 VT.getSimpleVT().SimpleTy, Custom);
772 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
773 VT.getSimpleVT().SimpleTy, Custom);
776 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
777 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
778 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
779 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
780 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
781 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
783 if (Subtarget->is64Bit()) {
784 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
785 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
788 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
789 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
790 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
793 // Do not attempt to promote non-128-bit vectors
794 if (!VT.is128BitVector())
797 setOperationAction(ISD::AND, SVT, Promote);
798 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
799 setOperationAction(ISD::OR, SVT, Promote);
800 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
801 setOperationAction(ISD::XOR, SVT, Promote);
802 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
803 setOperationAction(ISD::LOAD, SVT, Promote);
804 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
805 setOperationAction(ISD::SELECT, SVT, Promote);
806 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
809 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
811 // Custom lower v2i64 and v2f64 selects.
812 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
813 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
814 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
815 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
817 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
818 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
819 if (!DisableMMX && Subtarget->hasMMX()) {
820 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
821 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
825 if (Subtarget->hasSSE41()) {
826 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
827 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
828 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
829 setOperationAction(ISD::FRINT, MVT::f32, Legal);
830 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
831 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
832 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
833 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
834 setOperationAction(ISD::FRINT, MVT::f64, Legal);
835 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
837 // FIXME: Do we need to handle scalar-to-vector here?
838 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
840 // i8 and i16 vectors are custom , because the source register and source
841 // source memory operand types are not the same width. f32 vectors are
842 // custom since the immediate controlling the insert encodes additional
844 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
849 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
851 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
852 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
854 if (Subtarget->is64Bit()) {
855 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
856 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
860 if (Subtarget->hasSSE42()) {
861 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
864 if (!UseSoftFloat && Subtarget->hasAVX()) {
865 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
866 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
867 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
868 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
870 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
871 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
872 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
873 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
874 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
875 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
876 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
877 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
878 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
879 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
880 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
881 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
882 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
883 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
884 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
886 // Operations to consider commented out -v16i16 v32i8
887 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
888 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
889 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
890 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
891 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
892 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
893 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
894 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
895 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
896 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
897 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
898 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
899 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
900 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
902 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
903 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
904 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
905 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
907 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
908 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
909 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
910 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
913 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
914 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
915 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
916 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
917 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
918 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
921 // Not sure we want to do this since there are no 256-bit integer
924 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
925 // This includes 256-bit vectors
926 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
927 EVT VT = (MVT::SimpleValueType)i;
929 // Do not attempt to custom lower non-power-of-2 vectors
930 if (!isPowerOf2_32(VT.getVectorNumElements()))
933 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
934 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
935 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
938 if (Subtarget->is64Bit()) {
939 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
940 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
945 // Not sure we want to do this since there are no 256-bit integer
948 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
949 // Including 256-bit vectors
950 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
951 EVT VT = (MVT::SimpleValueType)i;
953 if (!VT.is256BitVector()) {
956 setOperationAction(ISD::AND, VT, Promote);
957 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
958 setOperationAction(ISD::OR, VT, Promote);
959 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
960 setOperationAction(ISD::XOR, VT, Promote);
961 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
962 setOperationAction(ISD::LOAD, VT, Promote);
963 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
964 setOperationAction(ISD::SELECT, VT, Promote);
965 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
968 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
972 // We want to custom lower some of our intrinsics.
973 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
975 // Add/Sub/Mul with overflow operations are custom lowered.
976 setOperationAction(ISD::SADDO, MVT::i32, Custom);
977 setOperationAction(ISD::UADDO, MVT::i32, Custom);
978 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
979 setOperationAction(ISD::USUBO, MVT::i32, Custom);
980 setOperationAction(ISD::SMULO, MVT::i32, Custom);
982 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
983 // handle type legalization for these operations here.
985 // FIXME: We really should do custom legalization for addition and
986 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
987 // than generic legalization for 64-bit multiplication-with-overflow, though.
988 if (Subtarget->is64Bit()) {
989 setOperationAction(ISD::SADDO, MVT::i64, Custom);
990 setOperationAction(ISD::UADDO, MVT::i64, Custom);
991 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
992 setOperationAction(ISD::USUBO, MVT::i64, Custom);
993 setOperationAction(ISD::SMULO, MVT::i64, Custom);
996 if (!Subtarget->is64Bit()) {
997 // These libcalls are not available in 32-bit.
998 setLibcallName(RTLIB::SHL_I128, 0);
999 setLibcallName(RTLIB::SRL_I128, 0);
1000 setLibcallName(RTLIB::SRA_I128, 0);
1003 // We have target-specific dag combine patterns for the following nodes:
1004 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1005 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1006 setTargetDAGCombine(ISD::BUILD_VECTOR);
1007 setTargetDAGCombine(ISD::SELECT);
1008 setTargetDAGCombine(ISD::SHL);
1009 setTargetDAGCombine(ISD::SRA);
1010 setTargetDAGCombine(ISD::SRL);
1011 setTargetDAGCombine(ISD::OR);
1012 setTargetDAGCombine(ISD::STORE);
1013 setTargetDAGCombine(ISD::ZERO_EXTEND);
1014 if (Subtarget->is64Bit())
1015 setTargetDAGCombine(ISD::MUL);
1017 computeRegisterProperties();
1019 // FIXME: These should be based on subtarget info. Plus, the values should
1020 // be smaller when we are in optimizing for size mode.
1021 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1022 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1023 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1024 setPrefLoopAlignment(16);
1025 benefitFromCodePlacementOpt = true;
1029 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1034 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1035 /// the desired ByVal argument alignment.
1036 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1039 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1040 if (VTy->getBitWidth() == 128)
1042 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1043 unsigned EltAlign = 0;
1044 getMaxByValAlign(ATy->getElementType(), EltAlign);
1045 if (EltAlign > MaxAlign)
1046 MaxAlign = EltAlign;
1047 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1048 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1049 unsigned EltAlign = 0;
1050 getMaxByValAlign(STy->getElementType(i), EltAlign);
1051 if (EltAlign > MaxAlign)
1052 MaxAlign = EltAlign;
1060 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1061 /// function arguments in the caller parameter area. For X86, aggregates
1062 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1063 /// are at 4-byte boundaries.
1064 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1065 if (Subtarget->is64Bit()) {
1066 // Max of 8 and alignment of type.
1067 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1074 if (Subtarget->hasSSE1())
1075 getMaxByValAlign(Ty, Align);
1079 /// getOptimalMemOpType - Returns the target specific optimal type for load
1080 /// and store operations as a result of memset, memcpy, and memmove
1081 /// lowering. If DstAlign is zero that means it's safe to destination
1082 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1083 /// means there isn't a need to check it against alignment requirement,
1084 /// probably because the source does not need to be loaded. If
1085 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1086 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1087 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1088 /// constant so it does not need to be loaded.
1089 /// It returns EVT::Other if the type should be determined using generic
1090 /// target-independent logic.
1092 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1093 unsigned DstAlign, unsigned SrcAlign,
1094 bool NonScalarIntSafe,
1096 MachineFunction &MF) const {
1097 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1098 // linux. This is because the stack realignment code can't handle certain
1099 // cases like PR2962. This should be removed when PR2962 is fixed.
1100 const Function *F = MF.getFunction();
1101 if (NonScalarIntSafe &&
1102 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1104 (Subtarget->isUnalignedMemAccessFast() ||
1105 ((DstAlign == 0 || DstAlign >= 16) &&
1106 (SrcAlign == 0 || SrcAlign >= 16))) &&
1107 Subtarget->getStackAlignment() >= 16) {
1108 if (Subtarget->hasSSE2())
1110 if (Subtarget->hasSSE1())
1112 } else if (!MemcpyStrSrc && Size >= 8 &&
1113 !Subtarget->is64Bit() &&
1114 Subtarget->getStackAlignment() >= 8 &&
1115 Subtarget->hasSSE2()) {
1116 // Do not use f64 to lower memcpy if source is string constant. It's
1117 // better to use i32 to avoid the loads.
1121 if (Subtarget->is64Bit() && Size >= 8)
1126 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1127 /// current function. The returned value is a member of the
1128 /// MachineJumpTableInfo::JTEntryKind enum.
1129 unsigned X86TargetLowering::getJumpTableEncoding() const {
1130 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1132 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1133 Subtarget->isPICStyleGOT())
1134 return MachineJumpTableInfo::EK_Custom32;
1136 // Otherwise, use the normal jump table encoding heuristics.
1137 return TargetLowering::getJumpTableEncoding();
1140 /// getPICBaseSymbol - Return the X86-32 PIC base.
1142 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1143 MCContext &Ctx) const {
1144 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1145 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1146 Twine(MF->getFunctionNumber())+"$pb");
1151 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1152 const MachineBasicBlock *MBB,
1153 unsigned uid,MCContext &Ctx) const{
1154 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1155 Subtarget->isPICStyleGOT());
1156 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1158 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1159 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1162 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1164 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1165 SelectionDAG &DAG) const {
1166 if (!Subtarget->is64Bit())
1167 // This doesn't have DebugLoc associated with it, but is not really the
1168 // same as a Register.
1169 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1173 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1174 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1176 const MCExpr *X86TargetLowering::
1177 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1178 MCContext &Ctx) const {
1179 // X86-64 uses RIP relative addressing based on the jump table label.
1180 if (Subtarget->isPICStyleRIPRel())
1181 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1183 // Otherwise, the reference is relative to the PIC base.
1184 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1187 /// getFunctionAlignment - Return the Log2 alignment of this function.
1188 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1189 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1192 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1193 unsigned &Offset) const {
1194 if (!Subtarget->isTargetLinux())
1197 if (Subtarget->is64Bit()) {
1198 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1200 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1213 //===----------------------------------------------------------------------===//
1214 // Return Value Calling Convention Implementation
1215 //===----------------------------------------------------------------------===//
1217 #include "X86GenCallingConv.inc"
1220 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1221 const SmallVectorImpl<ISD::OutputArg> &Outs,
1222 LLVMContext &Context) const {
1223 SmallVector<CCValAssign, 16> RVLocs;
1224 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1226 return CCInfo.CheckReturn(Outs, RetCC_X86);
1230 X86TargetLowering::LowerReturn(SDValue Chain,
1231 CallingConv::ID CallConv, bool isVarArg,
1232 const SmallVectorImpl<ISD::OutputArg> &Outs,
1233 const SmallVectorImpl<SDValue> &OutVals,
1234 DebugLoc dl, SelectionDAG &DAG) const {
1235 MachineFunction &MF = DAG.getMachineFunction();
1236 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1238 SmallVector<CCValAssign, 16> RVLocs;
1239 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1240 RVLocs, *DAG.getContext());
1241 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1243 // Add the regs to the liveout set for the function.
1244 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1245 for (unsigned i = 0; i != RVLocs.size(); ++i)
1246 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1247 MRI.addLiveOut(RVLocs[i].getLocReg());
1251 SmallVector<SDValue, 6> RetOps;
1252 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1253 // Operand #1 = Bytes To Pop
1254 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1257 // Copy the result values into the output registers.
1258 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1259 CCValAssign &VA = RVLocs[i];
1260 assert(VA.isRegLoc() && "Can only return in registers!");
1261 SDValue ValToCopy = OutVals[i];
1263 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1264 // the RET instruction and handled by the FP Stackifier.
1265 if (VA.getLocReg() == X86::ST0 ||
1266 VA.getLocReg() == X86::ST1) {
1267 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1268 // change the value to the FP stack register class.
1269 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1270 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1271 RetOps.push_back(ValToCopy);
1272 // Don't emit a copytoreg.
1276 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1277 // which is returned in RAX / RDX.
1278 if (Subtarget->is64Bit()) {
1279 EVT ValVT = ValToCopy.getValueType();
1280 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1281 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1282 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1283 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1287 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1288 Flag = Chain.getValue(1);
1291 // The x86-64 ABI for returning structs by value requires that we copy
1292 // the sret argument into %rax for the return. We saved the argument into
1293 // a virtual register in the entry block, so now we copy the value out
1295 if (Subtarget->is64Bit() &&
1296 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1297 MachineFunction &MF = DAG.getMachineFunction();
1298 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1299 unsigned Reg = FuncInfo->getSRetReturnReg();
1301 "SRetReturnReg should have been set in LowerFormalArguments().");
1302 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1304 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1305 Flag = Chain.getValue(1);
1307 // RAX now acts like a return value.
1308 MRI.addLiveOut(X86::RAX);
1311 RetOps[0] = Chain; // Update chain.
1313 // Add the flag if we have it.
1315 RetOps.push_back(Flag);
1317 return DAG.getNode(X86ISD::RET_FLAG, dl,
1318 MVT::Other, &RetOps[0], RetOps.size());
1321 /// LowerCallResult - Lower the result values of a call into the
1322 /// appropriate copies out of appropriate physical registers.
1325 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1326 CallingConv::ID CallConv, bool isVarArg,
1327 const SmallVectorImpl<ISD::InputArg> &Ins,
1328 DebugLoc dl, SelectionDAG &DAG,
1329 SmallVectorImpl<SDValue> &InVals) const {
1331 // Assign locations to each value returned by this call.
1332 SmallVector<CCValAssign, 16> RVLocs;
1333 bool Is64Bit = Subtarget->is64Bit();
1334 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1335 RVLocs, *DAG.getContext());
1336 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1338 // Copy all of the result registers out of their specified physreg.
1339 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1340 CCValAssign &VA = RVLocs[i];
1341 EVT CopyVT = VA.getValVT();
1343 // If this is x86-64, and we disabled SSE, we can't return FP values
1344 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1345 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1346 report_fatal_error("SSE register return with SSE disabled");
1349 // If this is a call to a function that returns an fp value on the floating
1350 // point stack, but where we prefer to use the value in xmm registers, copy
1351 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1352 if ((VA.getLocReg() == X86::ST0 ||
1353 VA.getLocReg() == X86::ST1) &&
1354 isScalarFPTypeInSSEReg(VA.getValVT())) {
1359 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1360 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1361 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1362 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1363 MVT::v2i64, InFlag).getValue(1);
1364 Val = Chain.getValue(0);
1365 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1366 Val, DAG.getConstant(0, MVT::i64));
1368 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1369 MVT::i64, InFlag).getValue(1);
1370 Val = Chain.getValue(0);
1372 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1374 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1375 CopyVT, InFlag).getValue(1);
1376 Val = Chain.getValue(0);
1378 InFlag = Chain.getValue(2);
1380 if (CopyVT != VA.getValVT()) {
1381 // Round the F80 the right size, which also moves to the appropriate xmm
1383 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1384 // This truncation won't change the value.
1385 DAG.getIntPtrConstant(1));
1388 InVals.push_back(Val);
1395 //===----------------------------------------------------------------------===//
1396 // C & StdCall & Fast Calling Convention implementation
1397 //===----------------------------------------------------------------------===//
1398 // StdCall calling convention seems to be standard for many Windows' API
1399 // routines and around. It differs from C calling convention just a little:
1400 // callee should clean up the stack, not caller. Symbols should be also
1401 // decorated in some fancy way :) It doesn't support any vector arguments.
1402 // For info on fast calling convention see Fast Calling Convention (tail call)
1403 // implementation LowerX86_32FastCCCallTo.
1405 /// CallIsStructReturn - Determines whether a call uses struct return
1407 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1411 return Outs[0].Flags.isSRet();
1414 /// ArgsAreStructReturn - Determines whether a function uses struct
1415 /// return semantics.
1417 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1421 return Ins[0].Flags.isSRet();
1424 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1425 /// given CallingConvention value.
1426 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1427 if (Subtarget->is64Bit()) {
1428 if (CC == CallingConv::GHC)
1429 return CC_X86_64_GHC;
1430 else if (Subtarget->isTargetWin64())
1431 return CC_X86_Win64_C;
1436 if (CC == CallingConv::X86_FastCall)
1437 return CC_X86_32_FastCall;
1438 else if (CC == CallingConv::X86_ThisCall)
1439 return CC_X86_32_ThisCall;
1440 else if (CC == CallingConv::Fast)
1441 return CC_X86_32_FastCC;
1442 else if (CC == CallingConv::GHC)
1443 return CC_X86_32_GHC;
1448 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1449 /// by "Src" to address "Dst" with size and alignment information specified by
1450 /// the specific parameter attribute. The copy will be passed as a byval
1451 /// function parameter.
1453 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1454 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1456 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1457 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1458 /*isVolatile*/false, /*AlwaysInline=*/true,
1462 /// IsTailCallConvention - Return true if the calling convention is one that
1463 /// supports tail call optimization.
1464 static bool IsTailCallConvention(CallingConv::ID CC) {
1465 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1468 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1469 /// a tailcall target by changing its ABI.
1470 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1471 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1475 X86TargetLowering::LowerMemArgument(SDValue Chain,
1476 CallingConv::ID CallConv,
1477 const SmallVectorImpl<ISD::InputArg> &Ins,
1478 DebugLoc dl, SelectionDAG &DAG,
1479 const CCValAssign &VA,
1480 MachineFrameInfo *MFI,
1482 // Create the nodes corresponding to a load from this parameter slot.
1483 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1484 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1485 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1488 // If value is passed by pointer we have address passed instead of the value
1490 if (VA.getLocInfo() == CCValAssign::Indirect)
1491 ValVT = VA.getLocVT();
1493 ValVT = VA.getValVT();
1495 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1496 // changed with more analysis.
1497 // In case of tail call optimization mark all arguments mutable. Since they
1498 // could be overwritten by lowering of arguments in case of a tail call.
1499 if (Flags.isByVal()) {
1500 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1501 VA.getLocMemOffset(), isImmutable);
1502 return DAG.getFrameIndex(FI, getPointerTy());
1504 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1505 VA.getLocMemOffset(), isImmutable);
1506 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1507 return DAG.getLoad(ValVT, dl, Chain, FIN,
1508 PseudoSourceValue::getFixedStack(FI), 0,
1514 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1515 CallingConv::ID CallConv,
1517 const SmallVectorImpl<ISD::InputArg> &Ins,
1520 SmallVectorImpl<SDValue> &InVals)
1522 MachineFunction &MF = DAG.getMachineFunction();
1523 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1525 const Function* Fn = MF.getFunction();
1526 if (Fn->hasExternalLinkage() &&
1527 Subtarget->isTargetCygMing() &&
1528 Fn->getName() == "main")
1529 FuncInfo->setForceFramePointer(true);
1531 MachineFrameInfo *MFI = MF.getFrameInfo();
1532 bool Is64Bit = Subtarget->is64Bit();
1533 bool IsWin64 = Subtarget->isTargetWin64();
1535 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1536 "Var args not supported with calling convention fastcc or ghc");
1538 // Assign locations to all of the incoming arguments.
1539 SmallVector<CCValAssign, 16> ArgLocs;
1540 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1541 ArgLocs, *DAG.getContext());
1542 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1544 unsigned LastVal = ~0U;
1546 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1547 CCValAssign &VA = ArgLocs[i];
1548 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1550 assert(VA.getValNo() != LastVal &&
1551 "Don't support value assigned to multiple locs yet");
1552 LastVal = VA.getValNo();
1554 if (VA.isRegLoc()) {
1555 EVT RegVT = VA.getLocVT();
1556 TargetRegisterClass *RC = NULL;
1557 if (RegVT == MVT::i32)
1558 RC = X86::GR32RegisterClass;
1559 else if (Is64Bit && RegVT == MVT::i64)
1560 RC = X86::GR64RegisterClass;
1561 else if (RegVT == MVT::f32)
1562 RC = X86::FR32RegisterClass;
1563 else if (RegVT == MVT::f64)
1564 RC = X86::FR64RegisterClass;
1565 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1566 RC = X86::VR128RegisterClass;
1567 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1568 RC = X86::VR64RegisterClass;
1570 llvm_unreachable("Unknown argument type!");
1572 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1573 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1575 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1576 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1578 if (VA.getLocInfo() == CCValAssign::SExt)
1579 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1580 DAG.getValueType(VA.getValVT()));
1581 else if (VA.getLocInfo() == CCValAssign::ZExt)
1582 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1583 DAG.getValueType(VA.getValVT()));
1584 else if (VA.getLocInfo() == CCValAssign::BCvt)
1585 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1587 if (VA.isExtInLoc()) {
1588 // Handle MMX values passed in XMM regs.
1589 if (RegVT.isVector()) {
1590 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1591 ArgValue, DAG.getConstant(0, MVT::i64));
1592 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1594 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1597 assert(VA.isMemLoc());
1598 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1601 // If value is passed via pointer - do a load.
1602 if (VA.getLocInfo() == CCValAssign::Indirect)
1603 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1606 InVals.push_back(ArgValue);
1609 // The x86-64 ABI for returning structs by value requires that we copy
1610 // the sret argument into %rax for the return. Save the argument into
1611 // a virtual register so that we can access it from the return points.
1612 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1613 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1614 unsigned Reg = FuncInfo->getSRetReturnReg();
1616 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1617 FuncInfo->setSRetReturnReg(Reg);
1619 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1620 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1623 unsigned StackSize = CCInfo.getNextStackOffset();
1624 // Align stack specially for tail calls.
1625 if (FuncIsMadeTailCallSafe(CallConv))
1626 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1628 // If the function takes variable number of arguments, make a frame index for
1629 // the start of the first vararg value... for expansion of llvm.va_start.
1631 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1632 CallConv != CallingConv::X86_ThisCall)) {
1633 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1636 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1638 // FIXME: We should really autogenerate these arrays
1639 static const unsigned GPR64ArgRegsWin64[] = {
1640 X86::RCX, X86::RDX, X86::R8, X86::R9
1642 static const unsigned XMMArgRegsWin64[] = {
1643 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1645 static const unsigned GPR64ArgRegs64Bit[] = {
1646 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1648 static const unsigned XMMArgRegs64Bit[] = {
1649 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1650 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1652 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1655 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1656 GPR64ArgRegs = GPR64ArgRegsWin64;
1657 XMMArgRegs = XMMArgRegsWin64;
1659 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1660 GPR64ArgRegs = GPR64ArgRegs64Bit;
1661 XMMArgRegs = XMMArgRegs64Bit;
1663 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1665 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1668 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1669 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1670 "SSE register cannot be used when SSE is disabled!");
1671 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1672 "SSE register cannot be used when SSE is disabled!");
1673 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1674 // Kernel mode asks for SSE to be disabled, so don't push them
1676 TotalNumXMMRegs = 0;
1678 // For X86-64, if there are vararg parameters that are passed via
1679 // registers, then we must store them to their spots on the stack so they
1680 // may be loaded by deferencing the result of va_next.
1681 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1682 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1683 FuncInfo->setRegSaveFrameIndex(
1684 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1687 // Store the integer parameter registers.
1688 SmallVector<SDValue, 8> MemOps;
1689 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1691 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1692 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1693 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1694 DAG.getIntPtrConstant(Offset));
1695 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1696 X86::GR64RegisterClass);
1697 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1699 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1700 PseudoSourceValue::getFixedStack(
1701 FuncInfo->getRegSaveFrameIndex()),
1702 Offset, false, false, 0);
1703 MemOps.push_back(Store);
1707 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1708 // Now store the XMM (fp + vector) parameter registers.
1709 SmallVector<SDValue, 11> SaveXMMOps;
1710 SaveXMMOps.push_back(Chain);
1712 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1713 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1714 SaveXMMOps.push_back(ALVal);
1716 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1717 FuncInfo->getRegSaveFrameIndex()));
1718 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1719 FuncInfo->getVarArgsFPOffset()));
1721 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1722 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1723 X86::VR128RegisterClass);
1724 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1725 SaveXMMOps.push_back(Val);
1727 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1729 &SaveXMMOps[0], SaveXMMOps.size()));
1732 if (!MemOps.empty())
1733 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1734 &MemOps[0], MemOps.size());
1738 // Some CCs need callee pop.
1739 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1740 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1742 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1743 // If this is an sret function, the return should pop the hidden pointer.
1744 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1745 FuncInfo->setBytesToPopOnReturn(4);
1749 // RegSaveFrameIndex is X86-64 only.
1750 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1751 if (CallConv == CallingConv::X86_FastCall ||
1752 CallConv == CallingConv::X86_ThisCall)
1753 // fastcc functions can't have varargs.
1754 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1761 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1762 SDValue StackPtr, SDValue Arg,
1763 DebugLoc dl, SelectionDAG &DAG,
1764 const CCValAssign &VA,
1765 ISD::ArgFlagsTy Flags) const {
1766 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1767 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1768 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1769 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1770 if (Flags.isByVal()) {
1771 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1773 return DAG.getStore(Chain, dl, Arg, PtrOff,
1774 PseudoSourceValue::getStack(), LocMemOffset,
1778 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1779 /// optimization is performed and it is required.
1781 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1782 SDValue &OutRetAddr, SDValue Chain,
1783 bool IsTailCall, bool Is64Bit,
1784 int FPDiff, DebugLoc dl) const {
1785 // Adjust the Return address stack slot.
1786 EVT VT = getPointerTy();
1787 OutRetAddr = getReturnAddressFrameIndex(DAG);
1789 // Load the "old" Return address.
1790 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1791 return SDValue(OutRetAddr.getNode(), 1);
1794 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1795 /// optimization is performed and it is required (FPDiff!=0).
1797 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1798 SDValue Chain, SDValue RetAddrFrIdx,
1799 bool Is64Bit, int FPDiff, DebugLoc dl) {
1800 // Store the return address to the appropriate stack slot.
1801 if (!FPDiff) return Chain;
1802 // Calculate the new stack slot for the return address.
1803 int SlotSize = Is64Bit ? 8 : 4;
1804 int NewReturnAddrFI =
1805 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1806 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1807 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1808 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1809 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1815 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1816 CallingConv::ID CallConv, bool isVarArg,
1818 const SmallVectorImpl<ISD::OutputArg> &Outs,
1819 const SmallVectorImpl<SDValue> &OutVals,
1820 const SmallVectorImpl<ISD::InputArg> &Ins,
1821 DebugLoc dl, SelectionDAG &DAG,
1822 SmallVectorImpl<SDValue> &InVals) const {
1823 MachineFunction &MF = DAG.getMachineFunction();
1824 bool Is64Bit = Subtarget->is64Bit();
1825 bool IsStructRet = CallIsStructReturn(Outs);
1826 bool IsSibcall = false;
1829 // Check if it's really possible to do a tail call.
1830 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1831 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1832 Outs, OutVals, Ins, DAG);
1834 // Sibcalls are automatically detected tailcalls which do not require
1836 if (!GuaranteedTailCallOpt && isTailCall)
1843 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1844 "Var args not supported with calling convention fastcc or ghc");
1846 // Analyze operands of the call, assigning locations to each operand.
1847 SmallVector<CCValAssign, 16> ArgLocs;
1848 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1849 ArgLocs, *DAG.getContext());
1850 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1852 // Get a count of how many bytes are to be pushed on the stack.
1853 unsigned NumBytes = CCInfo.getNextStackOffset();
1855 // This is a sibcall. The memory operands are available in caller's
1856 // own caller's stack.
1858 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1859 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1862 if (isTailCall && !IsSibcall) {
1863 // Lower arguments at fp - stackoffset + fpdiff.
1864 unsigned NumBytesCallerPushed =
1865 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1866 FPDiff = NumBytesCallerPushed - NumBytes;
1868 // Set the delta of movement of the returnaddr stackslot.
1869 // But only set if delta is greater than previous delta.
1870 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1871 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1875 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1877 SDValue RetAddrFrIdx;
1878 // Load return adress for tail calls.
1879 if (isTailCall && FPDiff)
1880 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1881 Is64Bit, FPDiff, dl);
1883 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1884 SmallVector<SDValue, 8> MemOpChains;
1887 // Walk the register/memloc assignments, inserting copies/loads. In the case
1888 // of tail call optimization arguments are handle later.
1889 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1890 CCValAssign &VA = ArgLocs[i];
1891 EVT RegVT = VA.getLocVT();
1892 SDValue Arg = OutVals[i];
1893 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1894 bool isByVal = Flags.isByVal();
1896 // Promote the value if needed.
1897 switch (VA.getLocInfo()) {
1898 default: llvm_unreachable("Unknown loc info!");
1899 case CCValAssign::Full: break;
1900 case CCValAssign::SExt:
1901 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1903 case CCValAssign::ZExt:
1904 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1906 case CCValAssign::AExt:
1907 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1908 // Special case: passing MMX values in XMM registers.
1909 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1910 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1911 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1913 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1915 case CCValAssign::BCvt:
1916 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1918 case CCValAssign::Indirect: {
1919 // Store the argument.
1920 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1921 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1922 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1923 PseudoSourceValue::getFixedStack(FI), 0,
1930 if (VA.isRegLoc()) {
1931 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1932 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1933 assert(VA.isMemLoc());
1934 if (StackPtr.getNode() == 0)
1935 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1936 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1937 dl, DAG, VA, Flags));
1941 if (!MemOpChains.empty())
1942 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1943 &MemOpChains[0], MemOpChains.size());
1945 // Build a sequence of copy-to-reg nodes chained together with token chain
1946 // and flag operands which copy the outgoing args into registers.
1948 // Tail call byval lowering might overwrite argument registers so in case of
1949 // tail call optimization the copies to registers are lowered later.
1951 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1952 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1953 RegsToPass[i].second, InFlag);
1954 InFlag = Chain.getValue(1);
1957 if (Subtarget->isPICStyleGOT()) {
1958 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1961 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1962 DAG.getNode(X86ISD::GlobalBaseReg,
1963 DebugLoc(), getPointerTy()),
1965 InFlag = Chain.getValue(1);
1967 // If we are tail calling and generating PIC/GOT style code load the
1968 // address of the callee into ECX. The value in ecx is used as target of
1969 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1970 // for tail calls on PIC/GOT architectures. Normally we would just put the
1971 // address of GOT into ebx and then call target@PLT. But for tail calls
1972 // ebx would be restored (since ebx is callee saved) before jumping to the
1975 // Note: The actual moving to ECX is done further down.
1976 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1977 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1978 !G->getGlobal()->hasProtectedVisibility())
1979 Callee = LowerGlobalAddress(Callee, DAG);
1980 else if (isa<ExternalSymbolSDNode>(Callee))
1981 Callee = LowerExternalSymbol(Callee, DAG);
1985 if (Is64Bit && isVarArg) {
1986 // From AMD64 ABI document:
1987 // For calls that may call functions that use varargs or stdargs
1988 // (prototype-less calls or calls to functions containing ellipsis (...) in
1989 // the declaration) %al is used as hidden argument to specify the number
1990 // of SSE registers used. The contents of %al do not need to match exactly
1991 // the number of registers, but must be an ubound on the number of SSE
1992 // registers used and is in the range 0 - 8 inclusive.
1994 // FIXME: Verify this on Win64
1995 // Count the number of XMM registers allocated.
1996 static const unsigned XMMArgRegs[] = {
1997 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1998 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2000 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2001 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2002 && "SSE registers cannot be used when SSE is disabled");
2004 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2005 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2006 InFlag = Chain.getValue(1);
2010 // For tail calls lower the arguments to the 'real' stack slot.
2012 // Force all the incoming stack arguments to be loaded from the stack
2013 // before any new outgoing arguments are stored to the stack, because the
2014 // outgoing stack slots may alias the incoming argument stack slots, and
2015 // the alias isn't otherwise explicit. This is slightly more conservative
2016 // than necessary, because it means that each store effectively depends
2017 // on every argument instead of just those arguments it would clobber.
2018 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2020 SmallVector<SDValue, 8> MemOpChains2;
2023 // Do not flag preceeding copytoreg stuff together with the following stuff.
2025 if (GuaranteedTailCallOpt) {
2026 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2027 CCValAssign &VA = ArgLocs[i];
2030 assert(VA.isMemLoc());
2031 SDValue Arg = OutVals[i];
2032 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2033 // Create frame index.
2034 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2035 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2036 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2037 FIN = DAG.getFrameIndex(FI, getPointerTy());
2039 if (Flags.isByVal()) {
2040 // Copy relative to framepointer.
2041 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2042 if (StackPtr.getNode() == 0)
2043 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2045 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2047 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2051 // Store relative to framepointer.
2052 MemOpChains2.push_back(
2053 DAG.getStore(ArgChain, dl, Arg, FIN,
2054 PseudoSourceValue::getFixedStack(FI), 0,
2060 if (!MemOpChains2.empty())
2061 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2062 &MemOpChains2[0], MemOpChains2.size());
2064 // Copy arguments to their registers.
2065 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2066 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2067 RegsToPass[i].second, InFlag);
2068 InFlag = Chain.getValue(1);
2072 // Store the return address to the appropriate stack slot.
2073 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2077 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2078 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2079 // In the 64-bit large code model, we have to make all calls
2080 // through a register, since the call instruction's 32-bit
2081 // pc-relative offset may not be large enough to hold the whole
2083 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2084 // If the callee is a GlobalAddress node (quite common, every direct call
2085 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2088 // We should use extra load for direct calls to dllimported functions in
2090 const GlobalValue *GV = G->getGlobal();
2091 if (!GV->hasDLLImportLinkage()) {
2092 unsigned char OpFlags = 0;
2094 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2095 // external symbols most go through the PLT in PIC mode. If the symbol
2096 // has hidden or protected visibility, or if it is static or local, then
2097 // we don't need to use the PLT - we can directly call it.
2098 if (Subtarget->isTargetELF() &&
2099 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2100 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2101 OpFlags = X86II::MO_PLT;
2102 } else if (Subtarget->isPICStyleStubAny() &&
2103 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2104 Subtarget->getDarwinVers() < 9) {
2105 // PC-relative references to external symbols should go through $stub,
2106 // unless we're building with the leopard linker or later, which
2107 // automatically synthesizes these stubs.
2108 OpFlags = X86II::MO_DARWIN_STUB;
2111 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2112 G->getOffset(), OpFlags);
2114 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2115 unsigned char OpFlags = 0;
2117 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2118 // symbols should go through the PLT.
2119 if (Subtarget->isTargetELF() &&
2120 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2121 OpFlags = X86II::MO_PLT;
2122 } else if (Subtarget->isPICStyleStubAny() &&
2123 Subtarget->getDarwinVers() < 9) {
2124 // PC-relative references to external symbols should go through $stub,
2125 // unless we're building with the leopard linker or later, which
2126 // automatically synthesizes these stubs.
2127 OpFlags = X86II::MO_DARWIN_STUB;
2130 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2134 // Returns a chain & a flag for retval copy to use.
2135 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2136 SmallVector<SDValue, 8> Ops;
2138 if (!IsSibcall && isTailCall) {
2139 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2140 DAG.getIntPtrConstant(0, true), InFlag);
2141 InFlag = Chain.getValue(1);
2144 Ops.push_back(Chain);
2145 Ops.push_back(Callee);
2148 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2150 // Add argument registers to the end of the list so that they are known live
2152 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2153 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2154 RegsToPass[i].second.getValueType()));
2156 // Add an implicit use GOT pointer in EBX.
2157 if (!isTailCall && Subtarget->isPICStyleGOT())
2158 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2160 // Add an implicit use of AL for x86 vararg functions.
2161 if (Is64Bit && isVarArg)
2162 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2164 if (InFlag.getNode())
2165 Ops.push_back(InFlag);
2169 //// If this is the first return lowered for this function, add the regs
2170 //// to the liveout set for the function.
2171 // This isn't right, although it's probably harmless on x86; liveouts
2172 // should be computed from returns not tail calls. Consider a void
2173 // function making a tail call to a function returning int.
2174 return DAG.getNode(X86ISD::TC_RETURN, dl,
2175 NodeTys, &Ops[0], Ops.size());
2178 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2179 InFlag = Chain.getValue(1);
2181 // Create the CALLSEQ_END node.
2182 unsigned NumBytesForCalleeToPush;
2183 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2184 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2185 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2186 // If this is a call to a struct-return function, the callee
2187 // pops the hidden struct pointer, so we have to push it back.
2188 // This is common for Darwin/X86, Linux & Mingw32 targets.
2189 NumBytesForCalleeToPush = 4;
2191 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2193 // Returns a flag for retval copy to use.
2195 Chain = DAG.getCALLSEQ_END(Chain,
2196 DAG.getIntPtrConstant(NumBytes, true),
2197 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2200 InFlag = Chain.getValue(1);
2203 // Handle result values, copying them out of physregs into vregs that we
2205 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2206 Ins, dl, DAG, InVals);
2210 //===----------------------------------------------------------------------===//
2211 // Fast Calling Convention (tail call) implementation
2212 //===----------------------------------------------------------------------===//
2214 // Like std call, callee cleans arguments, convention except that ECX is
2215 // reserved for storing the tail called function address. Only 2 registers are
2216 // free for argument passing (inreg). Tail call optimization is performed
2218 // * tailcallopt is enabled
2219 // * caller/callee are fastcc
2220 // On X86_64 architecture with GOT-style position independent code only local
2221 // (within module) calls are supported at the moment.
2222 // To keep the stack aligned according to platform abi the function
2223 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2224 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2225 // If a tail called function callee has more arguments than the caller the
2226 // caller needs to make sure that there is room to move the RETADDR to. This is
2227 // achieved by reserving an area the size of the argument delta right after the
2228 // original REtADDR, but before the saved framepointer or the spilled registers
2229 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2241 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2242 /// for a 16 byte align requirement.
2244 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2245 SelectionDAG& DAG) const {
2246 MachineFunction &MF = DAG.getMachineFunction();
2247 const TargetMachine &TM = MF.getTarget();
2248 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2249 unsigned StackAlignment = TFI.getStackAlignment();
2250 uint64_t AlignMask = StackAlignment - 1;
2251 int64_t Offset = StackSize;
2252 uint64_t SlotSize = TD->getPointerSize();
2253 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2254 // Number smaller than 12 so just add the difference.
2255 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2257 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2258 Offset = ((~AlignMask) & Offset) + StackAlignment +
2259 (StackAlignment-SlotSize);
2264 /// MatchingStackOffset - Return true if the given stack call argument is
2265 /// already available in the same position (relatively) of the caller's
2266 /// incoming argument stack.
2268 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2269 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2270 const X86InstrInfo *TII) {
2271 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2273 if (Arg.getOpcode() == ISD::CopyFromReg) {
2274 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2275 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2277 MachineInstr *Def = MRI->getVRegDef(VR);
2280 if (!Flags.isByVal()) {
2281 if (!TII->isLoadFromStackSlot(Def, FI))
2284 unsigned Opcode = Def->getOpcode();
2285 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2286 Def->getOperand(1).isFI()) {
2287 FI = Def->getOperand(1).getIndex();
2288 Bytes = Flags.getByValSize();
2292 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2293 if (Flags.isByVal())
2294 // ByVal argument is passed in as a pointer but it's now being
2295 // dereferenced. e.g.
2296 // define @foo(%struct.X* %A) {
2297 // tail call @bar(%struct.X* byval %A)
2300 SDValue Ptr = Ld->getBasePtr();
2301 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2304 FI = FINode->getIndex();
2308 assert(FI != INT_MAX);
2309 if (!MFI->isFixedObjectIndex(FI))
2311 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2314 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2315 /// for tail call optimization. Targets which want to do tail call
2316 /// optimization should implement this function.
2318 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2319 CallingConv::ID CalleeCC,
2321 bool isCalleeStructRet,
2322 bool isCallerStructRet,
2323 const SmallVectorImpl<ISD::OutputArg> &Outs,
2324 const SmallVectorImpl<SDValue> &OutVals,
2325 const SmallVectorImpl<ISD::InputArg> &Ins,
2326 SelectionDAG& DAG) const {
2327 if (!IsTailCallConvention(CalleeCC) &&
2328 CalleeCC != CallingConv::C)
2331 // If -tailcallopt is specified, make fastcc functions tail-callable.
2332 const MachineFunction &MF = DAG.getMachineFunction();
2333 const Function *CallerF = DAG.getMachineFunction().getFunction();
2334 CallingConv::ID CallerCC = CallerF->getCallingConv();
2335 bool CCMatch = CallerCC == CalleeCC;
2337 if (GuaranteedTailCallOpt) {
2338 if (IsTailCallConvention(CalleeCC) && CCMatch)
2343 // Look for obvious safe cases to perform tail call optimization that do not
2344 // require ABI changes. This is what gcc calls sibcall.
2346 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2347 // emit a special epilogue.
2348 if (RegInfo->needsStackRealignment(MF))
2351 // Do not sibcall optimize vararg calls unless the call site is not passing any
2353 if (isVarArg && !Outs.empty())
2356 // Also avoid sibcall optimization if either caller or callee uses struct
2357 // return semantics.
2358 if (isCalleeStructRet || isCallerStructRet)
2361 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2362 // Therefore if it's not used by the call it is not safe to optimize this into
2364 bool Unused = false;
2365 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2372 SmallVector<CCValAssign, 16> RVLocs;
2373 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2374 RVLocs, *DAG.getContext());
2375 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2376 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2377 CCValAssign &VA = RVLocs[i];
2378 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2383 // If the calling conventions do not match, then we'd better make sure the
2384 // results are returned in the same way as what the caller expects.
2386 SmallVector<CCValAssign, 16> RVLocs1;
2387 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2388 RVLocs1, *DAG.getContext());
2389 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2391 SmallVector<CCValAssign, 16> RVLocs2;
2392 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2393 RVLocs2, *DAG.getContext());
2394 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2396 if (RVLocs1.size() != RVLocs2.size())
2398 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2399 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2401 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2403 if (RVLocs1[i].isRegLoc()) {
2404 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2407 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2413 // If the callee takes no arguments then go on to check the results of the
2415 if (!Outs.empty()) {
2416 // Check if stack adjustment is needed. For now, do not do this if any
2417 // argument is passed on the stack.
2418 SmallVector<CCValAssign, 16> ArgLocs;
2419 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2420 ArgLocs, *DAG.getContext());
2421 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2422 if (CCInfo.getNextStackOffset()) {
2423 MachineFunction &MF = DAG.getMachineFunction();
2424 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2426 if (Subtarget->isTargetWin64())
2427 // Win64 ABI has additional complications.
2430 // Check if the arguments are already laid out in the right way as
2431 // the caller's fixed stack objects.
2432 MachineFrameInfo *MFI = MF.getFrameInfo();
2433 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2434 const X86InstrInfo *TII =
2435 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2436 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2437 CCValAssign &VA = ArgLocs[i];
2438 SDValue Arg = OutVals[i];
2439 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2440 if (VA.getLocInfo() == CCValAssign::Indirect)
2442 if (!VA.isRegLoc()) {
2443 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2450 // If the tailcall address may be in a register, then make sure it's
2451 // possible to register allocate for it. In 32-bit, the call address can
2452 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2453 // callee-saved registers are restored. In 64-bit, it's RAX, RCX, RDX, RSI,
2454 // RDI, R8, R9, R11.
2455 if (!isa<GlobalAddressSDNode>(Callee) &&
2456 !isa<ExternalSymbolSDNode>(Callee)) {
2457 unsigned Limit = Subtarget->is64Bit() ? 8 : 3;
2458 unsigned NumInRegs = 0;
2459 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2460 CCValAssign &VA = ArgLocs[i];
2461 if (VA.isRegLoc()) {
2462 if (++NumInRegs == Limit)
2473 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2474 return X86::createFastISel(funcInfo);
2478 //===----------------------------------------------------------------------===//
2479 // Other Lowering Hooks
2480 //===----------------------------------------------------------------------===//
2483 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2484 MachineFunction &MF = DAG.getMachineFunction();
2485 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2486 int ReturnAddrIndex = FuncInfo->getRAIndex();
2488 if (ReturnAddrIndex == 0) {
2489 // Set up a frame object for the return address.
2490 uint64_t SlotSize = TD->getPointerSize();
2491 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2493 FuncInfo->setRAIndex(ReturnAddrIndex);
2496 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2500 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2501 bool hasSymbolicDisplacement) {
2502 // Offset should fit into 32 bit immediate field.
2503 if (!isInt<32>(Offset))
2506 // If we don't have a symbolic displacement - we don't have any extra
2508 if (!hasSymbolicDisplacement)
2511 // FIXME: Some tweaks might be needed for medium code model.
2512 if (M != CodeModel::Small && M != CodeModel::Kernel)
2515 // For small code model we assume that latest object is 16MB before end of 31
2516 // bits boundary. We may also accept pretty large negative constants knowing
2517 // that all objects are in the positive half of address space.
2518 if (M == CodeModel::Small && Offset < 16*1024*1024)
2521 // For kernel code model we know that all object resist in the negative half
2522 // of 32bits address space. We may not accept negative offsets, since they may
2523 // be just off and we may accept pretty large positive ones.
2524 if (M == CodeModel::Kernel && Offset > 0)
2530 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2531 /// specific condition code, returning the condition code and the LHS/RHS of the
2532 /// comparison to make.
2533 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2534 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2536 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2537 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2538 // X > -1 -> X == 0, jump !sign.
2539 RHS = DAG.getConstant(0, RHS.getValueType());
2540 return X86::COND_NS;
2541 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2542 // X < 0 -> X == 0, jump on sign.
2544 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2546 RHS = DAG.getConstant(0, RHS.getValueType());
2547 return X86::COND_LE;
2551 switch (SetCCOpcode) {
2552 default: llvm_unreachable("Invalid integer condition!");
2553 case ISD::SETEQ: return X86::COND_E;
2554 case ISD::SETGT: return X86::COND_G;
2555 case ISD::SETGE: return X86::COND_GE;
2556 case ISD::SETLT: return X86::COND_L;
2557 case ISD::SETLE: return X86::COND_LE;
2558 case ISD::SETNE: return X86::COND_NE;
2559 case ISD::SETULT: return X86::COND_B;
2560 case ISD::SETUGT: return X86::COND_A;
2561 case ISD::SETULE: return X86::COND_BE;
2562 case ISD::SETUGE: return X86::COND_AE;
2566 // First determine if it is required or is profitable to flip the operands.
2568 // If LHS is a foldable load, but RHS is not, flip the condition.
2569 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2570 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2571 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2572 std::swap(LHS, RHS);
2575 switch (SetCCOpcode) {
2581 std::swap(LHS, RHS);
2585 // On a floating point condition, the flags are set as follows:
2587 // 0 | 0 | 0 | X > Y
2588 // 0 | 0 | 1 | X < Y
2589 // 1 | 0 | 0 | X == Y
2590 // 1 | 1 | 1 | unordered
2591 switch (SetCCOpcode) {
2592 default: llvm_unreachable("Condcode should be pre-legalized away");
2594 case ISD::SETEQ: return X86::COND_E;
2595 case ISD::SETOLT: // flipped
2597 case ISD::SETGT: return X86::COND_A;
2598 case ISD::SETOLE: // flipped
2600 case ISD::SETGE: return X86::COND_AE;
2601 case ISD::SETUGT: // flipped
2603 case ISD::SETLT: return X86::COND_B;
2604 case ISD::SETUGE: // flipped
2606 case ISD::SETLE: return X86::COND_BE;
2608 case ISD::SETNE: return X86::COND_NE;
2609 case ISD::SETUO: return X86::COND_P;
2610 case ISD::SETO: return X86::COND_NP;
2612 case ISD::SETUNE: return X86::COND_INVALID;
2616 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2617 /// code. Current x86 isa includes the following FP cmov instructions:
2618 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2619 static bool hasFPCMov(unsigned X86CC) {
2635 /// isFPImmLegal - Returns true if the target can instruction select the
2636 /// specified FP immediate natively. If false, the legalizer will
2637 /// materialize the FP immediate as a load from a constant pool.
2638 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2639 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2640 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2646 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2647 /// the specified range (L, H].
2648 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2649 return (Val < 0) || (Val >= Low && Val < Hi);
2652 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2653 /// specified value.
2654 static bool isUndefOrEqual(int Val, int CmpVal) {
2655 if (Val < 0 || Val == CmpVal)
2660 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2661 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2662 /// the second operand.
2663 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2664 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2665 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2666 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2667 return (Mask[0] < 2 && Mask[1] < 2);
2671 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2672 SmallVector<int, 8> M;
2674 return ::isPSHUFDMask(M, N->getValueType(0));
2677 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2678 /// is suitable for input to PSHUFHW.
2679 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2680 if (VT != MVT::v8i16)
2683 // Lower quadword copied in order or undef.
2684 for (int i = 0; i != 4; ++i)
2685 if (Mask[i] >= 0 && Mask[i] != i)
2688 // Upper quadword shuffled.
2689 for (int i = 4; i != 8; ++i)
2690 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2696 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2697 SmallVector<int, 8> M;
2699 return ::isPSHUFHWMask(M, N->getValueType(0));
2702 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2703 /// is suitable for input to PSHUFLW.
2704 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2705 if (VT != MVT::v8i16)
2708 // Upper quadword copied in order.
2709 for (int i = 4; i != 8; ++i)
2710 if (Mask[i] >= 0 && Mask[i] != i)
2713 // Lower quadword shuffled.
2714 for (int i = 0; i != 4; ++i)
2721 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2722 SmallVector<int, 8> M;
2724 return ::isPSHUFLWMask(M, N->getValueType(0));
2727 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2728 /// is suitable for input to PALIGNR.
2729 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2731 int i, e = VT.getVectorNumElements();
2733 // Do not handle v2i64 / v2f64 shuffles with palignr.
2734 if (e < 4 || !hasSSSE3)
2737 for (i = 0; i != e; ++i)
2741 // All undef, not a palignr.
2745 // Determine if it's ok to perform a palignr with only the LHS, since we
2746 // don't have access to the actual shuffle elements to see if RHS is undef.
2747 bool Unary = Mask[i] < (int)e;
2748 bool NeedsUnary = false;
2750 int s = Mask[i] - i;
2752 // Check the rest of the elements to see if they are consecutive.
2753 for (++i; i != e; ++i) {
2758 Unary = Unary && (m < (int)e);
2759 NeedsUnary = NeedsUnary || (m < s);
2761 if (NeedsUnary && !Unary)
2763 if (Unary && m != ((s+i) & (e-1)))
2765 if (!Unary && m != (s+i))
2771 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2772 SmallVector<int, 8> M;
2774 return ::isPALIGNRMask(M, N->getValueType(0), true);
2777 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2778 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2779 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2780 int NumElems = VT.getVectorNumElements();
2781 if (NumElems != 2 && NumElems != 4)
2784 int Half = NumElems / 2;
2785 for (int i = 0; i < Half; ++i)
2786 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2788 for (int i = Half; i < NumElems; ++i)
2789 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2795 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2796 SmallVector<int, 8> M;
2798 return ::isSHUFPMask(M, N->getValueType(0));
2801 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2802 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2803 /// half elements to come from vector 1 (which would equal the dest.) and
2804 /// the upper half to come from vector 2.
2805 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2806 int NumElems = VT.getVectorNumElements();
2808 if (NumElems != 2 && NumElems != 4)
2811 int Half = NumElems / 2;
2812 for (int i = 0; i < Half; ++i)
2813 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2815 for (int i = Half; i < NumElems; ++i)
2816 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2821 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2822 SmallVector<int, 8> M;
2824 return isCommutedSHUFPMask(M, N->getValueType(0));
2827 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2828 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2829 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2830 if (N->getValueType(0).getVectorNumElements() != 4)
2833 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2834 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2835 isUndefOrEqual(N->getMaskElt(1), 7) &&
2836 isUndefOrEqual(N->getMaskElt(2), 2) &&
2837 isUndefOrEqual(N->getMaskElt(3), 3);
2840 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2841 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2843 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2844 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2849 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2850 isUndefOrEqual(N->getMaskElt(1), 3) &&
2851 isUndefOrEqual(N->getMaskElt(2), 2) &&
2852 isUndefOrEqual(N->getMaskElt(3), 3);
2855 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2856 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2857 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2858 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2860 if (NumElems != 2 && NumElems != 4)
2863 for (unsigned i = 0; i < NumElems/2; ++i)
2864 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2867 for (unsigned i = NumElems/2; i < NumElems; ++i)
2868 if (!isUndefOrEqual(N->getMaskElt(i), i))
2874 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2875 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2876 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2877 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2879 if (NumElems != 2 && NumElems != 4)
2882 for (unsigned i = 0; i < NumElems/2; ++i)
2883 if (!isUndefOrEqual(N->getMaskElt(i), i))
2886 for (unsigned i = 0; i < NumElems/2; ++i)
2887 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2893 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2894 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2895 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2896 bool V2IsSplat = false) {
2897 int NumElts = VT.getVectorNumElements();
2898 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2901 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2903 int BitI1 = Mask[i+1];
2904 if (!isUndefOrEqual(BitI, j))
2907 if (!isUndefOrEqual(BitI1, NumElts))
2910 if (!isUndefOrEqual(BitI1, j + NumElts))
2917 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2918 SmallVector<int, 8> M;
2920 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2923 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2924 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2925 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2926 bool V2IsSplat = false) {
2927 int NumElts = VT.getVectorNumElements();
2928 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2931 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2933 int BitI1 = Mask[i+1];
2934 if (!isUndefOrEqual(BitI, j + NumElts/2))
2937 if (isUndefOrEqual(BitI1, NumElts))
2940 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2947 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2948 SmallVector<int, 8> M;
2950 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2953 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2954 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2956 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2957 int NumElems = VT.getVectorNumElements();
2958 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2961 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2963 int BitI1 = Mask[i+1];
2964 if (!isUndefOrEqual(BitI, j))
2966 if (!isUndefOrEqual(BitI1, j))
2972 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2973 SmallVector<int, 8> M;
2975 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2978 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2979 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2981 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2982 int NumElems = VT.getVectorNumElements();
2983 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2986 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2988 int BitI1 = Mask[i+1];
2989 if (!isUndefOrEqual(BitI, j))
2991 if (!isUndefOrEqual(BitI1, j))
2997 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2998 SmallVector<int, 8> M;
3000 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3003 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3004 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3005 /// MOVSD, and MOVD, i.e. setting the lowest element.
3006 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3007 if (VT.getVectorElementType().getSizeInBits() < 32)
3010 int NumElts = VT.getVectorNumElements();
3012 if (!isUndefOrEqual(Mask[0], NumElts))
3015 for (int i = 1; i < NumElts; ++i)
3016 if (!isUndefOrEqual(Mask[i], i))
3022 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3023 SmallVector<int, 8> M;
3025 return ::isMOVLMask(M, N->getValueType(0));
3028 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3029 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3030 /// element of vector 2 and the other elements to come from vector 1 in order.
3031 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3032 bool V2IsSplat = false, bool V2IsUndef = false) {
3033 int NumOps = VT.getVectorNumElements();
3034 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3037 if (!isUndefOrEqual(Mask[0], 0))
3040 for (int i = 1; i < NumOps; ++i)
3041 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3042 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3043 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3049 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3050 bool V2IsUndef = false) {
3051 SmallVector<int, 8> M;
3053 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3056 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3057 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3058 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3059 if (N->getValueType(0).getVectorNumElements() != 4)
3062 // Expect 1, 1, 3, 3
3063 for (unsigned i = 0; i < 2; ++i) {
3064 int Elt = N->getMaskElt(i);
3065 if (Elt >= 0 && Elt != 1)
3070 for (unsigned i = 2; i < 4; ++i) {
3071 int Elt = N->getMaskElt(i);
3072 if (Elt >= 0 && Elt != 3)
3077 // Don't use movshdup if it can be done with a shufps.
3078 // FIXME: verify that matching u, u, 3, 3 is what we want.
3082 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3083 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3084 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3085 if (N->getValueType(0).getVectorNumElements() != 4)
3088 // Expect 0, 0, 2, 2
3089 for (unsigned i = 0; i < 2; ++i)
3090 if (N->getMaskElt(i) > 0)
3094 for (unsigned i = 2; i < 4; ++i) {
3095 int Elt = N->getMaskElt(i);
3096 if (Elt >= 0 && Elt != 2)
3101 // Don't use movsldup if it can be done with a shufps.
3105 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3106 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3107 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3108 int e = N->getValueType(0).getVectorNumElements() / 2;
3110 for (int i = 0; i < e; ++i)
3111 if (!isUndefOrEqual(N->getMaskElt(i), i))
3113 for (int i = 0; i < e; ++i)
3114 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3119 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3120 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3121 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3122 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3123 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3125 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3127 for (int i = 0; i < NumOperands; ++i) {
3128 int Val = SVOp->getMaskElt(NumOperands-i-1);
3129 if (Val < 0) Val = 0;
3130 if (Val >= NumOperands) Val -= NumOperands;
3132 if (i != NumOperands - 1)
3138 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3139 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3140 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3141 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3143 // 8 nodes, but we only care about the last 4.
3144 for (unsigned i = 7; i >= 4; --i) {
3145 int Val = SVOp->getMaskElt(i);
3154 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3155 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3156 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3157 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3159 // 8 nodes, but we only care about the first 4.
3160 for (int i = 3; i >= 0; --i) {
3161 int Val = SVOp->getMaskElt(i);
3170 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3171 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3172 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3173 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3174 EVT VVT = N->getValueType(0);
3175 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3179 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3180 Val = SVOp->getMaskElt(i);
3184 return (Val - i) * EltSize;
3187 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3189 bool X86::isZeroNode(SDValue Elt) {
3190 return ((isa<ConstantSDNode>(Elt) &&
3191 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3192 (isa<ConstantFPSDNode>(Elt) &&
3193 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3196 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3197 /// their permute mask.
3198 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3199 SelectionDAG &DAG) {
3200 EVT VT = SVOp->getValueType(0);
3201 unsigned NumElems = VT.getVectorNumElements();
3202 SmallVector<int, 8> MaskVec;
3204 for (unsigned i = 0; i != NumElems; ++i) {
3205 int idx = SVOp->getMaskElt(i);
3207 MaskVec.push_back(idx);
3208 else if (idx < (int)NumElems)
3209 MaskVec.push_back(idx + NumElems);
3211 MaskVec.push_back(idx - NumElems);
3213 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3214 SVOp->getOperand(0), &MaskVec[0]);
3217 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3218 /// the two vector operands have swapped position.
3219 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3220 unsigned NumElems = VT.getVectorNumElements();
3221 for (unsigned i = 0; i != NumElems; ++i) {
3225 else if (idx < (int)NumElems)
3226 Mask[i] = idx + NumElems;
3228 Mask[i] = idx - NumElems;
3232 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3233 /// match movhlps. The lower half elements should come from upper half of
3234 /// V1 (and in order), and the upper half elements should come from the upper
3235 /// half of V2 (and in order).
3236 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3237 if (Op->getValueType(0).getVectorNumElements() != 4)
3239 for (unsigned i = 0, e = 2; i != e; ++i)
3240 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3242 for (unsigned i = 2; i != 4; ++i)
3243 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3248 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3249 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3251 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3252 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3254 N = N->getOperand(0).getNode();
3255 if (!ISD::isNON_EXTLoad(N))
3258 *LD = cast<LoadSDNode>(N);
3262 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3263 /// match movlp{s|d}. The lower half elements should come from lower half of
3264 /// V1 (and in order), and the upper half elements should come from the upper
3265 /// half of V2 (and in order). And since V1 will become the source of the
3266 /// MOVLP, it must be either a vector load or a scalar load to vector.
3267 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3268 ShuffleVectorSDNode *Op) {
3269 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3271 // Is V2 is a vector load, don't do this transformation. We will try to use
3272 // load folding shufps op.
3273 if (ISD::isNON_EXTLoad(V2))
3276 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3278 if (NumElems != 2 && NumElems != 4)
3280 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3281 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3283 for (unsigned i = NumElems/2; i != NumElems; ++i)
3284 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3289 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3291 static bool isSplatVector(SDNode *N) {
3292 if (N->getOpcode() != ISD::BUILD_VECTOR)
3295 SDValue SplatValue = N->getOperand(0);
3296 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3297 if (N->getOperand(i) != SplatValue)
3302 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3303 /// to an zero vector.
3304 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3305 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3306 SDValue V1 = N->getOperand(0);
3307 SDValue V2 = N->getOperand(1);
3308 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3309 for (unsigned i = 0; i != NumElems; ++i) {
3310 int Idx = N->getMaskElt(i);
3311 if (Idx >= (int)NumElems) {
3312 unsigned Opc = V2.getOpcode();
3313 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3315 if (Opc != ISD::BUILD_VECTOR ||
3316 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3318 } else if (Idx >= 0) {
3319 unsigned Opc = V1.getOpcode();
3320 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3322 if (Opc != ISD::BUILD_VECTOR ||
3323 !X86::isZeroNode(V1.getOperand(Idx)))
3330 /// getZeroVector - Returns a vector of specified type with all zero elements.
3332 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3334 assert(VT.isVector() && "Expected a vector type");
3336 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3337 // type. This ensures they get CSE'd.
3339 if (VT.getSizeInBits() == 64) { // MMX
3340 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3341 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3342 } else if (HasSSE2) { // SSE2
3343 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3344 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3346 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3347 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3349 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3352 /// getOnesVector - Returns a vector of specified type with all bits set.
3354 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3355 assert(VT.isVector() && "Expected a vector type");
3357 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3358 // type. This ensures they get CSE'd.
3359 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3361 if (VT.getSizeInBits() == 64) // MMX
3362 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3364 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3365 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3369 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3370 /// that point to V2 points to its first element.
3371 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3372 EVT VT = SVOp->getValueType(0);
3373 unsigned NumElems = VT.getVectorNumElements();
3375 bool Changed = false;
3376 SmallVector<int, 8> MaskVec;
3377 SVOp->getMask(MaskVec);
3379 for (unsigned i = 0; i != NumElems; ++i) {
3380 if (MaskVec[i] > (int)NumElems) {
3381 MaskVec[i] = NumElems;
3386 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3387 SVOp->getOperand(1), &MaskVec[0]);
3388 return SDValue(SVOp, 0);
3391 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3392 /// operation of specified width.
3393 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3395 unsigned NumElems = VT.getVectorNumElements();
3396 SmallVector<int, 8> Mask;
3397 Mask.push_back(NumElems);
3398 for (unsigned i = 1; i != NumElems; ++i)
3400 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3403 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3404 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3406 unsigned NumElems = VT.getVectorNumElements();
3407 SmallVector<int, 8> Mask;
3408 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3410 Mask.push_back(i + NumElems);
3412 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3415 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3416 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3418 unsigned NumElems = VT.getVectorNumElements();
3419 unsigned Half = NumElems/2;
3420 SmallVector<int, 8> Mask;
3421 for (unsigned i = 0; i != Half; ++i) {
3422 Mask.push_back(i + Half);
3423 Mask.push_back(i + NumElems + Half);
3425 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3428 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3429 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3431 if (SV->getValueType(0).getVectorNumElements() <= 4)
3432 return SDValue(SV, 0);
3434 EVT PVT = MVT::v4f32;
3435 EVT VT = SV->getValueType(0);
3436 DebugLoc dl = SV->getDebugLoc();
3437 SDValue V1 = SV->getOperand(0);
3438 int NumElems = VT.getVectorNumElements();
3439 int EltNo = SV->getSplatIndex();
3441 // unpack elements to the correct location
3442 while (NumElems > 4) {
3443 if (EltNo < NumElems/2) {
3444 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3446 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3447 EltNo -= NumElems/2;
3452 // Perform the splat.
3453 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3454 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3455 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3456 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3459 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3460 /// vector of zero or undef vector. This produces a shuffle where the low
3461 /// element of V2 is swizzled into the zero/undef vector, landing at element
3462 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3463 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3464 bool isZero, bool HasSSE2,
3465 SelectionDAG &DAG) {
3466 EVT VT = V2.getValueType();
3468 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3469 unsigned NumElems = VT.getVectorNumElements();
3470 SmallVector<int, 16> MaskVec;
3471 for (unsigned i = 0; i != NumElems; ++i)
3472 // If this is the insertion idx, put the low elt of V2 here.
3473 MaskVec.push_back(i == Idx ? NumElems : i);
3474 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3477 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3478 /// a shuffle that is zero.
3480 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3481 bool Low, SelectionDAG &DAG) {
3482 unsigned NumZeros = 0;
3483 for (int i = 0; i < NumElems; ++i) {
3484 unsigned Index = Low ? i : NumElems-i-1;
3485 int Idx = SVOp->getMaskElt(Index);
3490 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3491 if (Elt.getNode() && X86::isZeroNode(Elt))
3499 /// isVectorShift - Returns true if the shuffle can be implemented as a
3500 /// logical left or right shift of a vector.
3501 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3502 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3503 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3504 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3507 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3510 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3514 bool SeenV1 = false;
3515 bool SeenV2 = false;
3516 for (unsigned i = NumZeros; i < NumElems; ++i) {
3517 unsigned Val = isLeft ? (i - NumZeros) : i;
3518 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3521 unsigned Idx = (unsigned) Idx_;
3531 if (SeenV1 && SeenV2)
3534 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3540 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3542 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3543 unsigned NumNonZero, unsigned NumZero,
3545 const TargetLowering &TLI) {
3549 DebugLoc dl = Op.getDebugLoc();
3552 for (unsigned i = 0; i < 16; ++i) {
3553 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3554 if (ThisIsNonZero && First) {
3556 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3558 V = DAG.getUNDEF(MVT::v8i16);
3563 SDValue ThisElt(0, 0), LastElt(0, 0);
3564 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3565 if (LastIsNonZero) {
3566 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3567 MVT::i16, Op.getOperand(i-1));
3569 if (ThisIsNonZero) {
3570 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3571 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3572 ThisElt, DAG.getConstant(8, MVT::i8));
3574 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3578 if (ThisElt.getNode())
3579 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3580 DAG.getIntPtrConstant(i/2));
3584 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3587 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3589 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3590 unsigned NumNonZero, unsigned NumZero,
3592 const TargetLowering &TLI) {
3596 DebugLoc dl = Op.getDebugLoc();
3599 for (unsigned i = 0; i < 8; ++i) {
3600 bool isNonZero = (NonZeros & (1 << i)) != 0;
3604 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3606 V = DAG.getUNDEF(MVT::v8i16);
3609 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3610 MVT::v8i16, V, Op.getOperand(i),
3611 DAG.getIntPtrConstant(i));
3618 /// getVShift - Return a vector logical shift node.
3620 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3621 unsigned NumBits, SelectionDAG &DAG,
3622 const TargetLowering &TLI, DebugLoc dl) {
3623 bool isMMX = VT.getSizeInBits() == 64;
3624 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3625 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3626 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3627 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3628 DAG.getNode(Opc, dl, ShVT, SrcOp,
3629 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3633 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3634 SelectionDAG &DAG) const {
3636 // Check if the scalar load can be widened into a vector load. And if
3637 // the address is "base + cst" see if the cst can be "absorbed" into
3638 // the shuffle mask.
3639 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3640 SDValue Ptr = LD->getBasePtr();
3641 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3643 EVT PVT = LD->getValueType(0);
3644 if (PVT != MVT::i32 && PVT != MVT::f32)
3649 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3650 FI = FINode->getIndex();
3652 } else if (Ptr.getOpcode() == ISD::ADD &&
3653 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3654 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3655 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3656 Offset = Ptr.getConstantOperandVal(1);
3657 Ptr = Ptr.getOperand(0);
3662 SDValue Chain = LD->getChain();
3663 // Make sure the stack object alignment is at least 16.
3664 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3665 if (DAG.InferPtrAlignment(Ptr) < 16) {
3666 if (MFI->isFixedObjectIndex(FI)) {
3667 // Can't change the alignment. FIXME: It's possible to compute
3668 // the exact stack offset and reference FI + adjust offset instead.
3669 // If someone *really* cares about this. That's the way to implement it.
3672 MFI->setObjectAlignment(FI, 16);
3676 // (Offset % 16) must be multiple of 4. Then address is then
3677 // Ptr + (Offset & ~15).
3680 if ((Offset % 16) & 3)
3682 int64_t StartOffset = Offset & ~15;
3684 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3685 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3687 int EltNo = (Offset - StartOffset) >> 2;
3688 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3689 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3690 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3692 // Canonicalize it to a v4i32 shuffle.
3693 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3694 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3695 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3696 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3702 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3703 /// vector of type 'VT', see if the elements can be replaced by a single large
3704 /// load which has the same value as a build_vector whose operands are 'elts'.
3706 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3708 /// FIXME: we'd also like to handle the case where the last elements are zero
3709 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3710 /// There's even a handy isZeroNode for that purpose.
3711 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3712 DebugLoc &dl, SelectionDAG &DAG) {
3713 EVT EltVT = VT.getVectorElementType();
3714 unsigned NumElems = Elts.size();
3716 LoadSDNode *LDBase = NULL;
3717 unsigned LastLoadedElt = -1U;
3719 // For each element in the initializer, see if we've found a load or an undef.
3720 // If we don't find an initial load element, or later load elements are
3721 // non-consecutive, bail out.
3722 for (unsigned i = 0; i < NumElems; ++i) {
3723 SDValue Elt = Elts[i];
3725 if (!Elt.getNode() ||
3726 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3729 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3731 LDBase = cast<LoadSDNode>(Elt.getNode());
3735 if (Elt.getOpcode() == ISD::UNDEF)
3738 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3739 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3744 // If we have found an entire vector of loads and undefs, then return a large
3745 // load of the entire vector width starting at the base pointer. If we found
3746 // consecutive loads for the low half, generate a vzext_load node.
3747 if (LastLoadedElt == NumElems - 1) {
3748 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3749 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3750 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3751 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3752 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3753 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3754 LDBase->isVolatile(), LDBase->isNonTemporal(),
3755 LDBase->getAlignment());
3756 } else if (NumElems == 4 && LastLoadedElt == 1) {
3757 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3758 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3759 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3760 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3766 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3767 DebugLoc dl = Op.getDebugLoc();
3768 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3769 if (ISD::isBuildVectorAllZeros(Op.getNode())
3770 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3771 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3772 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3773 // eliminated on x86-32 hosts.
3774 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3777 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3778 return getOnesVector(Op.getValueType(), DAG, dl);
3779 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3782 EVT VT = Op.getValueType();
3783 EVT ExtVT = VT.getVectorElementType();
3784 unsigned EVTBits = ExtVT.getSizeInBits();
3786 unsigned NumElems = Op.getNumOperands();
3787 unsigned NumZero = 0;
3788 unsigned NumNonZero = 0;
3789 unsigned NonZeros = 0;
3790 bool IsAllConstants = true;
3791 SmallSet<SDValue, 8> Values;
3792 for (unsigned i = 0; i < NumElems; ++i) {
3793 SDValue Elt = Op.getOperand(i);
3794 if (Elt.getOpcode() == ISD::UNDEF)
3797 if (Elt.getOpcode() != ISD::Constant &&
3798 Elt.getOpcode() != ISD::ConstantFP)
3799 IsAllConstants = false;
3800 if (X86::isZeroNode(Elt))
3803 NonZeros |= (1 << i);
3808 if (NumNonZero == 0) {
3809 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3810 return DAG.getUNDEF(VT);
3813 // Special case for single non-zero, non-undef, element.
3814 if (NumNonZero == 1) {
3815 unsigned Idx = CountTrailingZeros_32(NonZeros);
3816 SDValue Item = Op.getOperand(Idx);
3818 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3819 // the value are obviously zero, truncate the value to i32 and do the
3820 // insertion that way. Only do this if the value is non-constant or if the
3821 // value is a constant being inserted into element 0. It is cheaper to do
3822 // a constant pool load than it is to do a movd + shuffle.
3823 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3824 (!IsAllConstants || Idx == 0)) {
3825 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3826 // Handle MMX and SSE both.
3827 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3828 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3830 // Truncate the value (which may itself be a constant) to i32, and
3831 // convert it to a vector with movd (S2V+shuffle to zero extend).
3832 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3833 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3834 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3835 Subtarget->hasSSE2(), DAG);
3837 // Now we have our 32-bit value zero extended in the low element of
3838 // a vector. If Idx != 0, swizzle it into place.
3840 SmallVector<int, 4> Mask;
3841 Mask.push_back(Idx);
3842 for (unsigned i = 1; i != VecElts; ++i)
3844 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3845 DAG.getUNDEF(Item.getValueType()),
3848 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3852 // If we have a constant or non-constant insertion into the low element of
3853 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3854 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3855 // depending on what the source datatype is.
3858 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3859 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3860 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3861 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3862 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3863 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3865 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3866 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3867 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3868 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3869 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3870 Subtarget->hasSSE2(), DAG);
3871 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3875 // Is it a vector logical left shift?
3876 if (NumElems == 2 && Idx == 1 &&
3877 X86::isZeroNode(Op.getOperand(0)) &&
3878 !X86::isZeroNode(Op.getOperand(1))) {
3879 unsigned NumBits = VT.getSizeInBits();
3880 return getVShift(true, VT,
3881 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3882 VT, Op.getOperand(1)),
3883 NumBits/2, DAG, *this, dl);
3886 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3889 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3890 // is a non-constant being inserted into an element other than the low one,
3891 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3892 // movd/movss) to move this into the low element, then shuffle it into
3894 if (EVTBits == 32) {
3895 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3897 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3898 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3899 Subtarget->hasSSE2(), DAG);
3900 SmallVector<int, 8> MaskVec;
3901 for (unsigned i = 0; i < NumElems; i++)
3902 MaskVec.push_back(i == Idx ? 0 : 1);
3903 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3907 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3908 if (Values.size() == 1) {
3909 if (EVTBits == 32) {
3910 // Instead of a shuffle like this:
3911 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3912 // Check if it's possible to issue this instead.
3913 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3914 unsigned Idx = CountTrailingZeros_32(NonZeros);
3915 SDValue Item = Op.getOperand(Idx);
3916 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3917 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3922 // A vector full of immediates; various special cases are already
3923 // handled, so this is best done with a single constant-pool load.
3927 // Let legalizer expand 2-wide build_vectors.
3928 if (EVTBits == 64) {
3929 if (NumNonZero == 1) {
3930 // One half is zero or undef.
3931 unsigned Idx = CountTrailingZeros_32(NonZeros);
3932 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3933 Op.getOperand(Idx));
3934 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3935 Subtarget->hasSSE2(), DAG);
3940 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3941 if (EVTBits == 8 && NumElems == 16) {
3942 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3944 if (V.getNode()) return V;
3947 if (EVTBits == 16 && NumElems == 8) {
3948 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3950 if (V.getNode()) return V;
3953 // If element VT is == 32 bits, turn it into a number of shuffles.
3954 SmallVector<SDValue, 8> V;
3956 if (NumElems == 4 && NumZero > 0) {
3957 for (unsigned i = 0; i < 4; ++i) {
3958 bool isZero = !(NonZeros & (1 << i));
3960 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3962 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3965 for (unsigned i = 0; i < 2; ++i) {
3966 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3969 V[i] = V[i*2]; // Must be a zero vector.
3972 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3975 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3978 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3983 SmallVector<int, 8> MaskVec;
3984 bool Reverse = (NonZeros & 0x3) == 2;
3985 for (unsigned i = 0; i < 2; ++i)
3986 MaskVec.push_back(Reverse ? 1-i : i);
3987 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3988 for (unsigned i = 0; i < 2; ++i)
3989 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3990 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3993 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
3994 // Check for a build vector of consecutive loads.
3995 for (unsigned i = 0; i < NumElems; ++i)
3996 V[i] = Op.getOperand(i);
3998 // Check for elements which are consecutive loads.
3999 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4003 // For SSE 4.1, use inserts into undef.
4004 if (getSubtarget()->hasSSE41()) {
4005 V[0] = DAG.getUNDEF(VT);
4006 for (unsigned i = 0; i < NumElems; ++i)
4007 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4008 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
4009 Op.getOperand(i), DAG.getIntPtrConstant(i));
4013 // Otherwise, expand into a number of unpckl*
4015 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4016 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4017 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4018 for (unsigned i = 0; i < NumElems; ++i)
4019 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4021 while (NumElems != 0) {
4022 for (unsigned i = 0; i < NumElems; ++i)
4023 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4032 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4033 // We support concatenate two MMX registers and place them in a MMX
4034 // register. This is better than doing a stack convert.
4035 DebugLoc dl = Op.getDebugLoc();
4036 EVT ResVT = Op.getValueType();
4037 assert(Op.getNumOperands() == 2);
4038 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4039 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4041 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4042 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4043 InVec = Op.getOperand(1);
4044 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4045 unsigned NumElts = ResVT.getVectorNumElements();
4046 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4047 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4048 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4050 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4051 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4052 Mask[0] = 0; Mask[1] = 2;
4053 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4055 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4058 // v8i16 shuffles - Prefer shuffles in the following order:
4059 // 1. [all] pshuflw, pshufhw, optional move
4060 // 2. [ssse3] 1 x pshufb
4061 // 3. [ssse3] 2 x pshufb + 1 x por
4062 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4064 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4066 const X86TargetLowering &TLI) {
4067 SDValue V1 = SVOp->getOperand(0);
4068 SDValue V2 = SVOp->getOperand(1);
4069 DebugLoc dl = SVOp->getDebugLoc();
4070 SmallVector<int, 8> MaskVals;
4072 // Determine if more than 1 of the words in each of the low and high quadwords
4073 // of the result come from the same quadword of one of the two inputs. Undef
4074 // mask values count as coming from any quadword, for better codegen.
4075 SmallVector<unsigned, 4> LoQuad(4);
4076 SmallVector<unsigned, 4> HiQuad(4);
4077 BitVector InputQuads(4);
4078 for (unsigned i = 0; i < 8; ++i) {
4079 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4080 int EltIdx = SVOp->getMaskElt(i);
4081 MaskVals.push_back(EltIdx);
4090 InputQuads.set(EltIdx / 4);
4093 int BestLoQuad = -1;
4094 unsigned MaxQuad = 1;
4095 for (unsigned i = 0; i < 4; ++i) {
4096 if (LoQuad[i] > MaxQuad) {
4098 MaxQuad = LoQuad[i];
4102 int BestHiQuad = -1;
4104 for (unsigned i = 0; i < 4; ++i) {
4105 if (HiQuad[i] > MaxQuad) {
4107 MaxQuad = HiQuad[i];
4111 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4112 // of the two input vectors, shuffle them into one input vector so only a
4113 // single pshufb instruction is necessary. If There are more than 2 input
4114 // quads, disable the next transformation since it does not help SSSE3.
4115 bool V1Used = InputQuads[0] || InputQuads[1];
4116 bool V2Used = InputQuads[2] || InputQuads[3];
4117 if (TLI.getSubtarget()->hasSSSE3()) {
4118 if (InputQuads.count() == 2 && V1Used && V2Used) {
4119 BestLoQuad = InputQuads.find_first();
4120 BestHiQuad = InputQuads.find_next(BestLoQuad);
4122 if (InputQuads.count() > 2) {
4128 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4129 // the shuffle mask. If a quad is scored as -1, that means that it contains
4130 // words from all 4 input quadwords.
4132 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4133 SmallVector<int, 8> MaskV;
4134 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4135 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4136 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4137 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4138 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4139 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4141 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4142 // source words for the shuffle, to aid later transformations.
4143 bool AllWordsInNewV = true;
4144 bool InOrder[2] = { true, true };
4145 for (unsigned i = 0; i != 8; ++i) {
4146 int idx = MaskVals[i];
4148 InOrder[i/4] = false;
4149 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4151 AllWordsInNewV = false;
4155 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4156 if (AllWordsInNewV) {
4157 for (int i = 0; i != 8; ++i) {
4158 int idx = MaskVals[i];
4161 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4162 if ((idx != i) && idx < 4)
4164 if ((idx != i) && idx > 3)
4173 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4174 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4175 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4176 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4177 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4181 // If we have SSSE3, and all words of the result are from 1 input vector,
4182 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4183 // is present, fall back to case 4.
4184 if (TLI.getSubtarget()->hasSSSE3()) {
4185 SmallVector<SDValue,16> pshufbMask;
4187 // If we have elements from both input vectors, set the high bit of the
4188 // shuffle mask element to zero out elements that come from V2 in the V1
4189 // mask, and elements that come from V1 in the V2 mask, so that the two
4190 // results can be OR'd together.
4191 bool TwoInputs = V1Used && V2Used;
4192 for (unsigned i = 0; i != 8; ++i) {
4193 int EltIdx = MaskVals[i] * 2;
4194 if (TwoInputs && (EltIdx >= 16)) {
4195 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4196 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4199 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4200 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4202 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4203 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4204 DAG.getNode(ISD::BUILD_VECTOR, dl,
4205 MVT::v16i8, &pshufbMask[0], 16));
4207 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4209 // Calculate the shuffle mask for the second input, shuffle it, and
4210 // OR it with the first shuffled input.
4212 for (unsigned i = 0; i != 8; ++i) {
4213 int EltIdx = MaskVals[i] * 2;
4215 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4216 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4219 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4220 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4222 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4223 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4224 DAG.getNode(ISD::BUILD_VECTOR, dl,
4225 MVT::v16i8, &pshufbMask[0], 16));
4226 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4227 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4230 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4231 // and update MaskVals with new element order.
4232 BitVector InOrder(8);
4233 if (BestLoQuad >= 0) {
4234 SmallVector<int, 8> MaskV;
4235 for (int i = 0; i != 4; ++i) {
4236 int idx = MaskVals[i];
4238 MaskV.push_back(-1);
4240 } else if ((idx / 4) == BestLoQuad) {
4241 MaskV.push_back(idx & 3);
4244 MaskV.push_back(-1);
4247 for (unsigned i = 4; i != 8; ++i)
4249 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4253 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4254 // and update MaskVals with the new element order.
4255 if (BestHiQuad >= 0) {
4256 SmallVector<int, 8> MaskV;
4257 for (unsigned i = 0; i != 4; ++i)
4259 for (unsigned i = 4; i != 8; ++i) {
4260 int idx = MaskVals[i];
4262 MaskV.push_back(-1);
4264 } else if ((idx / 4) == BestHiQuad) {
4265 MaskV.push_back((idx & 3) + 4);
4268 MaskV.push_back(-1);
4271 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4275 // In case BestHi & BestLo were both -1, which means each quadword has a word
4276 // from each of the four input quadwords, calculate the InOrder bitvector now
4277 // before falling through to the insert/extract cleanup.
4278 if (BestLoQuad == -1 && BestHiQuad == -1) {
4280 for (int i = 0; i != 8; ++i)
4281 if (MaskVals[i] < 0 || MaskVals[i] == i)
4285 // The other elements are put in the right place using pextrw and pinsrw.
4286 for (unsigned i = 0; i != 8; ++i) {
4289 int EltIdx = MaskVals[i];
4292 SDValue ExtOp = (EltIdx < 8)
4293 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4294 DAG.getIntPtrConstant(EltIdx))
4295 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4296 DAG.getIntPtrConstant(EltIdx - 8));
4297 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4298 DAG.getIntPtrConstant(i));
4303 // v16i8 shuffles - Prefer shuffles in the following order:
4304 // 1. [ssse3] 1 x pshufb
4305 // 2. [ssse3] 2 x pshufb + 1 x por
4306 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4308 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4310 const X86TargetLowering &TLI) {
4311 SDValue V1 = SVOp->getOperand(0);
4312 SDValue V2 = SVOp->getOperand(1);
4313 DebugLoc dl = SVOp->getDebugLoc();
4314 SmallVector<int, 16> MaskVals;
4315 SVOp->getMask(MaskVals);
4317 // If we have SSSE3, case 1 is generated when all result bytes come from
4318 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4319 // present, fall back to case 3.
4320 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4323 for (unsigned i = 0; i < 16; ++i) {
4324 int EltIdx = MaskVals[i];
4333 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4334 if (TLI.getSubtarget()->hasSSSE3()) {
4335 SmallVector<SDValue,16> pshufbMask;
4337 // If all result elements are from one input vector, then only translate
4338 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4340 // Otherwise, we have elements from both input vectors, and must zero out
4341 // elements that come from V2 in the first mask, and V1 in the second mask
4342 // so that we can OR them together.
4343 bool TwoInputs = !(V1Only || V2Only);
4344 for (unsigned i = 0; i != 16; ++i) {
4345 int EltIdx = MaskVals[i];
4346 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4347 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4350 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4352 // If all the elements are from V2, assign it to V1 and return after
4353 // building the first pshufb.
4356 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4357 DAG.getNode(ISD::BUILD_VECTOR, dl,
4358 MVT::v16i8, &pshufbMask[0], 16));
4362 // Calculate the shuffle mask for the second input, shuffle it, and
4363 // OR it with the first shuffled input.
4365 for (unsigned i = 0; i != 16; ++i) {
4366 int EltIdx = MaskVals[i];
4368 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4371 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4373 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4374 DAG.getNode(ISD::BUILD_VECTOR, dl,
4375 MVT::v16i8, &pshufbMask[0], 16));
4376 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4379 // No SSSE3 - Calculate in place words and then fix all out of place words
4380 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4381 // the 16 different words that comprise the two doublequadword input vectors.
4382 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4383 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4384 SDValue NewV = V2Only ? V2 : V1;
4385 for (int i = 0; i != 8; ++i) {
4386 int Elt0 = MaskVals[i*2];
4387 int Elt1 = MaskVals[i*2+1];
4389 // This word of the result is all undef, skip it.
4390 if (Elt0 < 0 && Elt1 < 0)
4393 // This word of the result is already in the correct place, skip it.
4394 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4396 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4399 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4400 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4403 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4404 // using a single extract together, load it and store it.
4405 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4406 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4407 DAG.getIntPtrConstant(Elt1 / 2));
4408 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4409 DAG.getIntPtrConstant(i));
4413 // If Elt1 is defined, extract it from the appropriate source. If the
4414 // source byte is not also odd, shift the extracted word left 8 bits
4415 // otherwise clear the bottom 8 bits if we need to do an or.
4417 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4418 DAG.getIntPtrConstant(Elt1 / 2));
4419 if ((Elt1 & 1) == 0)
4420 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4421 DAG.getConstant(8, TLI.getShiftAmountTy()));
4423 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4424 DAG.getConstant(0xFF00, MVT::i16));
4426 // If Elt0 is defined, extract it from the appropriate source. If the
4427 // source byte is not also even, shift the extracted word right 8 bits. If
4428 // Elt1 was also defined, OR the extracted values together before
4429 // inserting them in the result.
4431 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4432 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4433 if ((Elt0 & 1) != 0)
4434 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4435 DAG.getConstant(8, TLI.getShiftAmountTy()));
4437 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4438 DAG.getConstant(0x00FF, MVT::i16));
4439 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4442 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4443 DAG.getIntPtrConstant(i));
4445 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4448 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4449 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4450 /// done when every pair / quad of shuffle mask elements point to elements in
4451 /// the right sequence. e.g.
4452 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4454 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4456 const TargetLowering &TLI, DebugLoc dl) {
4457 EVT VT = SVOp->getValueType(0);
4458 SDValue V1 = SVOp->getOperand(0);
4459 SDValue V2 = SVOp->getOperand(1);
4460 unsigned NumElems = VT.getVectorNumElements();
4461 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4462 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4464 switch (VT.getSimpleVT().SimpleTy) {
4465 default: assert(false && "Unexpected!");
4466 case MVT::v4f32: NewVT = MVT::v2f64; break;
4467 case MVT::v4i32: NewVT = MVT::v2i64; break;
4468 case MVT::v8i16: NewVT = MVT::v4i32; break;
4469 case MVT::v16i8: NewVT = MVT::v4i32; break;
4472 if (NewWidth == 2) {
4478 int Scale = NumElems / NewWidth;
4479 SmallVector<int, 8> MaskVec;
4480 for (unsigned i = 0; i < NumElems; i += Scale) {
4482 for (int j = 0; j < Scale; ++j) {
4483 int EltIdx = SVOp->getMaskElt(i+j);
4487 StartIdx = EltIdx - (EltIdx % Scale);
4488 if (EltIdx != StartIdx + j)
4492 MaskVec.push_back(-1);
4494 MaskVec.push_back(StartIdx / Scale);
4497 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4498 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4499 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4502 /// getVZextMovL - Return a zero-extending vector move low node.
4504 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4505 SDValue SrcOp, SelectionDAG &DAG,
4506 const X86Subtarget *Subtarget, DebugLoc dl) {
4507 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4508 LoadSDNode *LD = NULL;
4509 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4510 LD = dyn_cast<LoadSDNode>(SrcOp);
4512 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4514 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4515 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4516 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4517 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4518 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4520 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4521 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4522 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4523 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4531 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4532 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4533 DAG.getNode(ISD::BIT_CONVERT, dl,
4537 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4540 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4541 SDValue V1 = SVOp->getOperand(0);
4542 SDValue V2 = SVOp->getOperand(1);
4543 DebugLoc dl = SVOp->getDebugLoc();
4544 EVT VT = SVOp->getValueType(0);
4546 SmallVector<std::pair<int, int>, 8> Locs;
4548 SmallVector<int, 8> Mask1(4U, -1);
4549 SmallVector<int, 8> PermMask;
4550 SVOp->getMask(PermMask);
4554 for (unsigned i = 0; i != 4; ++i) {
4555 int Idx = PermMask[i];
4557 Locs[i] = std::make_pair(-1, -1);
4559 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4561 Locs[i] = std::make_pair(0, NumLo);
4565 Locs[i] = std::make_pair(1, NumHi);
4567 Mask1[2+NumHi] = Idx;
4573 if (NumLo <= 2 && NumHi <= 2) {
4574 // If no more than two elements come from either vector. This can be
4575 // implemented with two shuffles. First shuffle gather the elements.
4576 // The second shuffle, which takes the first shuffle as both of its
4577 // vector operands, put the elements into the right order.
4578 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4580 SmallVector<int, 8> Mask2(4U, -1);
4582 for (unsigned i = 0; i != 4; ++i) {
4583 if (Locs[i].first == -1)
4586 unsigned Idx = (i < 2) ? 0 : 4;
4587 Idx += Locs[i].first * 2 + Locs[i].second;
4592 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4593 } else if (NumLo == 3 || NumHi == 3) {
4594 // Otherwise, we must have three elements from one vector, call it X, and
4595 // one element from the other, call it Y. First, use a shufps to build an
4596 // intermediate vector with the one element from Y and the element from X
4597 // that will be in the same half in the final destination (the indexes don't
4598 // matter). Then, use a shufps to build the final vector, taking the half
4599 // containing the element from Y from the intermediate, and the other half
4602 // Normalize it so the 3 elements come from V1.
4603 CommuteVectorShuffleMask(PermMask, VT);
4607 // Find the element from V2.
4609 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4610 int Val = PermMask[HiIndex];
4617 Mask1[0] = PermMask[HiIndex];
4619 Mask1[2] = PermMask[HiIndex^1];
4621 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4624 Mask1[0] = PermMask[0];
4625 Mask1[1] = PermMask[1];
4626 Mask1[2] = HiIndex & 1 ? 6 : 4;
4627 Mask1[3] = HiIndex & 1 ? 4 : 6;
4628 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4630 Mask1[0] = HiIndex & 1 ? 2 : 0;
4631 Mask1[1] = HiIndex & 1 ? 0 : 2;
4632 Mask1[2] = PermMask[2];
4633 Mask1[3] = PermMask[3];
4638 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4642 // Break it into (shuffle shuffle_hi, shuffle_lo).
4644 SmallVector<int,8> LoMask(4U, -1);
4645 SmallVector<int,8> HiMask(4U, -1);
4647 SmallVector<int,8> *MaskPtr = &LoMask;
4648 unsigned MaskIdx = 0;
4651 for (unsigned i = 0; i != 4; ++i) {
4658 int Idx = PermMask[i];
4660 Locs[i] = std::make_pair(-1, -1);
4661 } else if (Idx < 4) {
4662 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4663 (*MaskPtr)[LoIdx] = Idx;
4666 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4667 (*MaskPtr)[HiIdx] = Idx;
4672 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4673 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4674 SmallVector<int, 8> MaskOps;
4675 for (unsigned i = 0; i != 4; ++i) {
4676 if (Locs[i].first == -1) {
4677 MaskOps.push_back(-1);
4679 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4680 MaskOps.push_back(Idx);
4683 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4687 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4688 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4689 SDValue V1 = Op.getOperand(0);
4690 SDValue V2 = Op.getOperand(1);
4691 EVT VT = Op.getValueType();
4692 DebugLoc dl = Op.getDebugLoc();
4693 unsigned NumElems = VT.getVectorNumElements();
4694 bool isMMX = VT.getSizeInBits() == 64;
4695 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4696 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4697 bool V1IsSplat = false;
4698 bool V2IsSplat = false;
4700 if (isZeroShuffle(SVOp))
4701 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4703 // Promote splats to v4f32.
4704 if (SVOp->isSplat()) {
4705 if (isMMX || NumElems < 4)
4707 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4710 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4712 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4713 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4714 if (NewOp.getNode())
4715 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4716 LowerVECTOR_SHUFFLE(NewOp, DAG));
4717 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4718 // FIXME: Figure out a cleaner way to do this.
4719 // Try to make use of movq to zero out the top part.
4720 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4721 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4722 if (NewOp.getNode()) {
4723 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4724 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4725 DAG, Subtarget, dl);
4727 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4728 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4729 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4730 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4731 DAG, Subtarget, dl);
4735 if (X86::isPSHUFDMask(SVOp))
4738 // Check if this can be converted into a logical shift.
4739 bool isLeft = false;
4742 bool isShift = getSubtarget()->hasSSE2() &&
4743 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4744 if (isShift && ShVal.hasOneUse()) {
4745 // If the shifted value has multiple uses, it may be cheaper to use
4746 // v_set0 + movlhps or movhlps, etc.
4747 EVT EltVT = VT.getVectorElementType();
4748 ShAmt *= EltVT.getSizeInBits();
4749 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4752 if (X86::isMOVLMask(SVOp)) {
4755 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4756 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4761 // FIXME: fold these into legal mask.
4762 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4763 X86::isMOVSLDUPMask(SVOp) ||
4764 X86::isMOVHLPSMask(SVOp) ||
4765 X86::isMOVLHPSMask(SVOp) ||
4766 X86::isMOVLPMask(SVOp)))
4769 if (ShouldXformToMOVHLPS(SVOp) ||
4770 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4771 return CommuteVectorShuffle(SVOp, DAG);
4774 // No better options. Use a vshl / vsrl.
4775 EVT EltVT = VT.getVectorElementType();
4776 ShAmt *= EltVT.getSizeInBits();
4777 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4780 bool Commuted = false;
4781 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4782 // 1,1,1,1 -> v8i16 though.
4783 V1IsSplat = isSplatVector(V1.getNode());
4784 V2IsSplat = isSplatVector(V2.getNode());
4786 // Canonicalize the splat or undef, if present, to be on the RHS.
4787 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4788 Op = CommuteVectorShuffle(SVOp, DAG);
4789 SVOp = cast<ShuffleVectorSDNode>(Op);
4790 V1 = SVOp->getOperand(0);
4791 V2 = SVOp->getOperand(1);
4792 std::swap(V1IsSplat, V2IsSplat);
4793 std::swap(V1IsUndef, V2IsUndef);
4797 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4798 // Shuffling low element of v1 into undef, just return v1.
4801 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4802 // the instruction selector will not match, so get a canonical MOVL with
4803 // swapped operands to undo the commute.
4804 return getMOVL(DAG, dl, VT, V2, V1);
4807 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4808 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4809 X86::isUNPCKLMask(SVOp) ||
4810 X86::isUNPCKHMask(SVOp))
4814 // Normalize mask so all entries that point to V2 points to its first
4815 // element then try to match unpck{h|l} again. If match, return a
4816 // new vector_shuffle with the corrected mask.
4817 SDValue NewMask = NormalizeMask(SVOp, DAG);
4818 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4819 if (NSVOp != SVOp) {
4820 if (X86::isUNPCKLMask(NSVOp, true)) {
4822 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4829 // Commute is back and try unpck* again.
4830 // FIXME: this seems wrong.
4831 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4832 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4833 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4834 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4835 X86::isUNPCKLMask(NewSVOp) ||
4836 X86::isUNPCKHMask(NewSVOp))
4840 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4842 // Normalize the node to match x86 shuffle ops if needed
4843 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4844 return CommuteVectorShuffle(SVOp, DAG);
4846 // Check for legal shuffle and return?
4847 SmallVector<int, 16> PermMask;
4848 SVOp->getMask(PermMask);
4849 if (isShuffleMaskLegal(PermMask, VT))
4852 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4853 if (VT == MVT::v8i16) {
4854 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4855 if (NewOp.getNode())
4859 if (VT == MVT::v16i8) {
4860 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4861 if (NewOp.getNode())
4865 // Handle all 4 wide cases with a number of shuffles except for MMX.
4866 if (NumElems == 4 && !isMMX)
4867 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4873 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4874 SelectionDAG &DAG) const {
4875 EVT VT = Op.getValueType();
4876 DebugLoc dl = Op.getDebugLoc();
4877 if (VT.getSizeInBits() == 8) {
4878 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4879 Op.getOperand(0), Op.getOperand(1));
4880 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4881 DAG.getValueType(VT));
4882 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4883 } else if (VT.getSizeInBits() == 16) {
4884 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4885 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4887 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4888 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4889 DAG.getNode(ISD::BIT_CONVERT, dl,
4893 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4894 Op.getOperand(0), Op.getOperand(1));
4895 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4896 DAG.getValueType(VT));
4897 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4898 } else if (VT == MVT::f32) {
4899 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4900 // the result back to FR32 register. It's only worth matching if the
4901 // result has a single use which is a store or a bitcast to i32. And in
4902 // the case of a store, it's not worth it if the index is a constant 0,
4903 // because a MOVSSmr can be used instead, which is smaller and faster.
4904 if (!Op.hasOneUse())
4906 SDNode *User = *Op.getNode()->use_begin();
4907 if ((User->getOpcode() != ISD::STORE ||
4908 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4909 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4910 (User->getOpcode() != ISD::BIT_CONVERT ||
4911 User->getValueType(0) != MVT::i32))
4913 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4914 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4917 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4918 } else if (VT == MVT::i32) {
4919 // ExtractPS works with constant index.
4920 if (isa<ConstantSDNode>(Op.getOperand(1)))
4928 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4929 SelectionDAG &DAG) const {
4930 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4933 if (Subtarget->hasSSE41()) {
4934 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4939 EVT VT = Op.getValueType();
4940 DebugLoc dl = Op.getDebugLoc();
4941 // TODO: handle v16i8.
4942 if (VT.getSizeInBits() == 16) {
4943 SDValue Vec = Op.getOperand(0);
4944 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4946 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4947 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4948 DAG.getNode(ISD::BIT_CONVERT, dl,
4951 // Transform it so it match pextrw which produces a 32-bit result.
4952 EVT EltVT = MVT::i32;
4953 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4954 Op.getOperand(0), Op.getOperand(1));
4955 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4956 DAG.getValueType(VT));
4957 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4958 } else if (VT.getSizeInBits() == 32) {
4959 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4963 // SHUFPS the element to the lowest double word, then movss.
4964 int Mask[4] = { Idx, -1, -1, -1 };
4965 EVT VVT = Op.getOperand(0).getValueType();
4966 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4967 DAG.getUNDEF(VVT), Mask);
4968 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4969 DAG.getIntPtrConstant(0));
4970 } else if (VT.getSizeInBits() == 64) {
4971 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4972 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4973 // to match extract_elt for f64.
4974 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4978 // UNPCKHPD the element to the lowest double word, then movsd.
4979 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4980 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4981 int Mask[2] = { 1, -1 };
4982 EVT VVT = Op.getOperand(0).getValueType();
4983 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4984 DAG.getUNDEF(VVT), Mask);
4985 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4986 DAG.getIntPtrConstant(0));
4993 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
4994 SelectionDAG &DAG) const {
4995 EVT VT = Op.getValueType();
4996 EVT EltVT = VT.getVectorElementType();
4997 DebugLoc dl = Op.getDebugLoc();
4999 SDValue N0 = Op.getOperand(0);
5000 SDValue N1 = Op.getOperand(1);
5001 SDValue N2 = Op.getOperand(2);
5003 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5004 isa<ConstantSDNode>(N2)) {
5006 if (VT == MVT::v8i16)
5007 Opc = X86ISD::PINSRW;
5008 else if (VT == MVT::v4i16)
5009 Opc = X86ISD::MMX_PINSRW;
5010 else if (VT == MVT::v16i8)
5011 Opc = X86ISD::PINSRB;
5013 Opc = X86ISD::PINSRB;
5015 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5017 if (N1.getValueType() != MVT::i32)
5018 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5019 if (N2.getValueType() != MVT::i32)
5020 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5021 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5022 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5023 // Bits [7:6] of the constant are the source select. This will always be
5024 // zero here. The DAG Combiner may combine an extract_elt index into these
5025 // bits. For example (insert (extract, 3), 2) could be matched by putting
5026 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5027 // Bits [5:4] of the constant are the destination select. This is the
5028 // value of the incoming immediate.
5029 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5030 // combine either bitwise AND or insert of float 0.0 to set these bits.
5031 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5032 // Create this as a scalar to vector..
5033 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5034 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5035 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5036 // PINSR* works with constant index.
5043 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5044 EVT VT = Op.getValueType();
5045 EVT EltVT = VT.getVectorElementType();
5047 if (Subtarget->hasSSE41())
5048 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5050 if (EltVT == MVT::i8)
5053 DebugLoc dl = Op.getDebugLoc();
5054 SDValue N0 = Op.getOperand(0);
5055 SDValue N1 = Op.getOperand(1);
5056 SDValue N2 = Op.getOperand(2);
5058 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5059 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5060 // as its second argument.
5061 if (N1.getValueType() != MVT::i32)
5062 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5063 if (N2.getValueType() != MVT::i32)
5064 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5065 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5066 dl, VT, N0, N1, N2);
5072 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5073 DebugLoc dl = Op.getDebugLoc();
5075 if (Op.getValueType() == MVT::v1i64 &&
5076 Op.getOperand(0).getValueType() == MVT::i64)
5077 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5079 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5080 EVT VT = MVT::v2i32;
5081 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5088 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5089 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5092 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5093 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5094 // one of the above mentioned nodes. It has to be wrapped because otherwise
5095 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5096 // be used to form addressing mode. These wrapped nodes will be selected
5099 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5100 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5102 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5104 unsigned char OpFlag = 0;
5105 unsigned WrapperKind = X86ISD::Wrapper;
5106 CodeModel::Model M = getTargetMachine().getCodeModel();
5108 if (Subtarget->isPICStyleRIPRel() &&
5109 (M == CodeModel::Small || M == CodeModel::Kernel))
5110 WrapperKind = X86ISD::WrapperRIP;
5111 else if (Subtarget->isPICStyleGOT())
5112 OpFlag = X86II::MO_GOTOFF;
5113 else if (Subtarget->isPICStyleStubPIC())
5114 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5116 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5118 CP->getOffset(), OpFlag);
5119 DebugLoc DL = CP->getDebugLoc();
5120 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5121 // With PIC, the address is actually $g + Offset.
5123 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5124 DAG.getNode(X86ISD::GlobalBaseReg,
5125 DebugLoc(), getPointerTy()),
5132 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5133 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5135 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5137 unsigned char OpFlag = 0;
5138 unsigned WrapperKind = X86ISD::Wrapper;
5139 CodeModel::Model M = getTargetMachine().getCodeModel();
5141 if (Subtarget->isPICStyleRIPRel() &&
5142 (M == CodeModel::Small || M == CodeModel::Kernel))
5143 WrapperKind = X86ISD::WrapperRIP;
5144 else if (Subtarget->isPICStyleGOT())
5145 OpFlag = X86II::MO_GOTOFF;
5146 else if (Subtarget->isPICStyleStubPIC())
5147 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5149 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5151 DebugLoc DL = JT->getDebugLoc();
5152 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5154 // With PIC, the address is actually $g + Offset.
5156 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5157 DAG.getNode(X86ISD::GlobalBaseReg,
5158 DebugLoc(), getPointerTy()),
5166 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5167 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5169 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5171 unsigned char OpFlag = 0;
5172 unsigned WrapperKind = X86ISD::Wrapper;
5173 CodeModel::Model M = getTargetMachine().getCodeModel();
5175 if (Subtarget->isPICStyleRIPRel() &&
5176 (M == CodeModel::Small || M == CodeModel::Kernel))
5177 WrapperKind = X86ISD::WrapperRIP;
5178 else if (Subtarget->isPICStyleGOT())
5179 OpFlag = X86II::MO_GOTOFF;
5180 else if (Subtarget->isPICStyleStubPIC())
5181 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5183 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5185 DebugLoc DL = Op.getDebugLoc();
5186 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5189 // With PIC, the address is actually $g + Offset.
5190 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5191 !Subtarget->is64Bit()) {
5192 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5193 DAG.getNode(X86ISD::GlobalBaseReg,
5194 DebugLoc(), getPointerTy()),
5202 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5203 // Create the TargetBlockAddressAddress node.
5204 unsigned char OpFlags =
5205 Subtarget->ClassifyBlockAddressReference();
5206 CodeModel::Model M = getTargetMachine().getCodeModel();
5207 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5208 DebugLoc dl = Op.getDebugLoc();
5209 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5210 /*isTarget=*/true, OpFlags);
5212 if (Subtarget->isPICStyleRIPRel() &&
5213 (M == CodeModel::Small || M == CodeModel::Kernel))
5214 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5216 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5218 // With PIC, the address is actually $g + Offset.
5219 if (isGlobalRelativeToPICBase(OpFlags)) {
5220 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5221 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5229 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5231 SelectionDAG &DAG) const {
5232 // Create the TargetGlobalAddress node, folding in the constant
5233 // offset if it is legal.
5234 unsigned char OpFlags =
5235 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5236 CodeModel::Model M = getTargetMachine().getCodeModel();
5238 if (OpFlags == X86II::MO_NO_FLAG &&
5239 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5240 // A direct static reference to a global.
5241 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5244 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5247 if (Subtarget->isPICStyleRIPRel() &&
5248 (M == CodeModel::Small || M == CodeModel::Kernel))
5249 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5251 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5253 // With PIC, the address is actually $g + Offset.
5254 if (isGlobalRelativeToPICBase(OpFlags)) {
5255 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5256 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5260 // For globals that require a load from a stub to get the address, emit the
5262 if (isGlobalStubReference(OpFlags))
5263 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5264 PseudoSourceValue::getGOT(), 0, false, false, 0);
5266 // If there was a non-zero offset that we didn't fold, create an explicit
5269 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5270 DAG.getConstant(Offset, getPointerTy()));
5276 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5277 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5278 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5279 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5283 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5284 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5285 unsigned char OperandFlags) {
5286 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5287 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5288 DebugLoc dl = GA->getDebugLoc();
5289 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5290 GA->getValueType(0),
5294 SDValue Ops[] = { Chain, TGA, *InFlag };
5295 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5297 SDValue Ops[] = { Chain, TGA };
5298 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5301 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5302 MFI->setAdjustsStack(true);
5304 SDValue Flag = Chain.getValue(1);
5305 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5308 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5310 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5313 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5314 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5315 DAG.getNode(X86ISD::GlobalBaseReg,
5316 DebugLoc(), PtrVT), InFlag);
5317 InFlag = Chain.getValue(1);
5319 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5322 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5324 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5326 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5327 X86::RAX, X86II::MO_TLSGD);
5330 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5331 // "local exec" model.
5332 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5333 const EVT PtrVT, TLSModel::Model model,
5335 DebugLoc dl = GA->getDebugLoc();
5336 // Get the Thread Pointer
5337 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5339 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5342 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5343 NULL, 0, false, false, 0);
5345 unsigned char OperandFlags = 0;
5346 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5348 unsigned WrapperKind = X86ISD::Wrapper;
5349 if (model == TLSModel::LocalExec) {
5350 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5351 } else if (is64Bit) {
5352 assert(model == TLSModel::InitialExec);
5353 OperandFlags = X86II::MO_GOTTPOFF;
5354 WrapperKind = X86ISD::WrapperRIP;
5356 assert(model == TLSModel::InitialExec);
5357 OperandFlags = X86II::MO_INDNTPOFF;
5360 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5362 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5363 GA->getValueType(0),
5364 GA->getOffset(), OperandFlags);
5365 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5367 if (model == TLSModel::InitialExec)
5368 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5369 PseudoSourceValue::getGOT(), 0, false, false, 0);
5371 // The address of the thread local variable is the add of the thread
5372 // pointer with the offset of the variable.
5373 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5377 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5379 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5380 const GlobalValue *GV = GA->getGlobal();
5382 if (Subtarget->isTargetELF()) {
5383 // TODO: implement the "local dynamic" model
5384 // TODO: implement the "initial exec"model for pic executables
5386 // If GV is an alias then use the aliasee for determining
5387 // thread-localness.
5388 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5389 GV = GA->resolveAliasedGlobal(false);
5391 TLSModel::Model model
5392 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5395 case TLSModel::GeneralDynamic:
5396 case TLSModel::LocalDynamic: // not implemented
5397 if (Subtarget->is64Bit())
5398 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5399 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5401 case TLSModel::InitialExec:
5402 case TLSModel::LocalExec:
5403 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5404 Subtarget->is64Bit());
5406 } else if (Subtarget->isTargetDarwin()) {
5407 // Darwin only has one model of TLS. Lower to that.
5408 unsigned char OpFlag = 0;
5409 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5410 X86ISD::WrapperRIP : X86ISD::Wrapper;
5412 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5414 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5415 !Subtarget->is64Bit();
5417 OpFlag = X86II::MO_TLVP_PIC_BASE;
5419 OpFlag = X86II::MO_TLVP;
5420 DebugLoc DL = Op.getDebugLoc();
5421 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
5423 GA->getOffset(), OpFlag);
5424 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5426 // With PIC32, the address is actually $g + Offset.
5428 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5429 DAG.getNode(X86ISD::GlobalBaseReg,
5430 DebugLoc(), getPointerTy()),
5433 // Lowering the machine isd will make sure everything is in the right
5435 SDValue Args[] = { Offset };
5436 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5438 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5439 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5440 MFI->setAdjustsStack(true);
5442 // And our return value (tls address) is in the standard call return value
5444 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5445 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5449 "TLS not implemented for this target.");
5451 llvm_unreachable("Unreachable");
5456 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5457 /// take a 2 x i32 value to shift plus a shift amount.
5458 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5459 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5460 EVT VT = Op.getValueType();
5461 unsigned VTBits = VT.getSizeInBits();
5462 DebugLoc dl = Op.getDebugLoc();
5463 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5464 SDValue ShOpLo = Op.getOperand(0);
5465 SDValue ShOpHi = Op.getOperand(1);
5466 SDValue ShAmt = Op.getOperand(2);
5467 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5468 DAG.getConstant(VTBits - 1, MVT::i8))
5469 : DAG.getConstant(0, VT);
5472 if (Op.getOpcode() == ISD::SHL_PARTS) {
5473 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5474 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5476 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5477 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5480 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5481 DAG.getConstant(VTBits, MVT::i8));
5482 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5483 AndNode, DAG.getConstant(0, MVT::i8));
5486 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5487 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5488 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5490 if (Op.getOpcode() == ISD::SHL_PARTS) {
5491 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5492 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5494 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5495 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5498 SDValue Ops[2] = { Lo, Hi };
5499 return DAG.getMergeValues(Ops, 2, dl);
5502 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5503 SelectionDAG &DAG) const {
5504 EVT SrcVT = Op.getOperand(0).getValueType();
5506 if (SrcVT.isVector()) {
5507 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5513 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5514 "Unknown SINT_TO_FP to lower!");
5516 // These are really Legal; return the operand so the caller accepts it as
5518 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5520 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5521 Subtarget->is64Bit()) {
5525 DebugLoc dl = Op.getDebugLoc();
5526 unsigned Size = SrcVT.getSizeInBits()/8;
5527 MachineFunction &MF = DAG.getMachineFunction();
5528 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5529 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5530 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5532 PseudoSourceValue::getFixedStack(SSFI), 0,
5534 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5537 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5539 SelectionDAG &DAG) const {
5541 DebugLoc dl = Op.getDebugLoc();
5543 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5545 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5547 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5548 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5549 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5550 Tys, Ops, array_lengthof(Ops));
5553 Chain = Result.getValue(1);
5554 SDValue InFlag = Result.getValue(2);
5556 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5557 // shouldn't be necessary except that RFP cannot be live across
5558 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5559 MachineFunction &MF = DAG.getMachineFunction();
5560 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5561 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5562 Tys = DAG.getVTList(MVT::Other);
5564 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5566 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5567 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5568 PseudoSourceValue::getFixedStack(SSFI), 0,
5575 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5576 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5577 SelectionDAG &DAG) const {
5578 // This algorithm is not obvious. Here it is in C code, more or less:
5580 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5581 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5582 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5584 // Copy ints to xmm registers.
5585 __m128i xh = _mm_cvtsi32_si128( hi );
5586 __m128i xl = _mm_cvtsi32_si128( lo );
5588 // Combine into low half of a single xmm register.
5589 __m128i x = _mm_unpacklo_epi32( xh, xl );
5593 // Merge in appropriate exponents to give the integer bits the right
5595 x = _mm_unpacklo_epi32( x, exp );
5597 // Subtract away the biases to deal with the IEEE-754 double precision
5599 d = _mm_sub_pd( (__m128d) x, bias );
5601 // All conversions up to here are exact. The correctly rounded result is
5602 // calculated using the current rounding mode using the following
5604 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5605 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5606 // store doesn't really need to be here (except
5607 // maybe to zero the other double)
5612 DebugLoc dl = Op.getDebugLoc();
5613 LLVMContext *Context = DAG.getContext();
5615 // Build some magic constants.
5616 std::vector<Constant*> CV0;
5617 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5618 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5619 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5620 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5621 Constant *C0 = ConstantVector::get(CV0);
5622 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5624 std::vector<Constant*> CV1;
5626 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5628 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5629 Constant *C1 = ConstantVector::get(CV1);
5630 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5632 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5633 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5635 DAG.getIntPtrConstant(1)));
5636 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5637 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5639 DAG.getIntPtrConstant(0)));
5640 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5641 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5642 PseudoSourceValue::getConstantPool(), 0,
5644 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5645 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5646 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5647 PseudoSourceValue::getConstantPool(), 0,
5649 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5651 // Add the halves; easiest way is to swap them into another reg first.
5652 int ShufMask[2] = { 1, -1 };
5653 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5654 DAG.getUNDEF(MVT::v2f64), ShufMask);
5655 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5656 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5657 DAG.getIntPtrConstant(0));
5660 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5661 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5662 SelectionDAG &DAG) const {
5663 DebugLoc dl = Op.getDebugLoc();
5664 // FP constant to bias correct the final result.
5665 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5668 // Load the 32-bit value into an XMM register.
5669 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5670 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5672 DAG.getIntPtrConstant(0)));
5674 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5675 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5676 DAG.getIntPtrConstant(0));
5678 // Or the load with the bias.
5679 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5680 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5681 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5683 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5684 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5685 MVT::v2f64, Bias)));
5686 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5687 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5688 DAG.getIntPtrConstant(0));
5690 // Subtract the bias.
5691 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5693 // Handle final rounding.
5694 EVT DestVT = Op.getValueType();
5696 if (DestVT.bitsLT(MVT::f64)) {
5697 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5698 DAG.getIntPtrConstant(0));
5699 } else if (DestVT.bitsGT(MVT::f64)) {
5700 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5703 // Handle final rounding.
5707 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5708 SelectionDAG &DAG) const {
5709 SDValue N0 = Op.getOperand(0);
5710 DebugLoc dl = Op.getDebugLoc();
5712 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5713 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5714 // the optimization here.
5715 if (DAG.SignBitIsZero(N0))
5716 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5718 EVT SrcVT = N0.getValueType();
5719 EVT DstVT = Op.getValueType();
5720 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5721 return LowerUINT_TO_FP_i64(Op, DAG);
5722 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5723 return LowerUINT_TO_FP_i32(Op, DAG);
5725 // Make a 64-bit buffer, and use it to build an FILD.
5726 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5727 if (SrcVT == MVT::i32) {
5728 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5729 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5730 getPointerTy(), StackSlot, WordOff);
5731 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5732 StackSlot, NULL, 0, false, false, 0);
5733 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5734 OffsetSlot, NULL, 0, false, false, 0);
5735 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5739 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5740 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5741 StackSlot, NULL, 0, false, false, 0);
5742 // For i64 source, we need to add the appropriate power of 2 if the input
5743 // was negative. This is the same as the optimization in
5744 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5745 // we must be careful to do the computation in x87 extended precision, not
5746 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5747 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5748 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5749 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5751 APInt FF(32, 0x5F800000ULL);
5753 // Check whether the sign bit is set.
5754 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5755 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5758 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5759 SDValue FudgePtr = DAG.getConstantPool(
5760 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5763 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5764 SDValue Zero = DAG.getIntPtrConstant(0);
5765 SDValue Four = DAG.getIntPtrConstant(4);
5766 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5768 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5770 // Load the value out, extending it from f32 to f80.
5771 // FIXME: Avoid the extend by constructing the right constant pool?
5772 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
5773 FudgePtr, PseudoSourceValue::getConstantPool(),
5774 0, MVT::f32, false, false, 4);
5775 // Extend everything to 80 bits to force it to be done on x87.
5776 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5777 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5780 std::pair<SDValue,SDValue> X86TargetLowering::
5781 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5782 DebugLoc dl = Op.getDebugLoc();
5784 EVT DstTy = Op.getValueType();
5787 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5791 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5792 DstTy.getSimpleVT() >= MVT::i16 &&
5793 "Unknown FP_TO_SINT to lower!");
5795 // These are really Legal.
5796 if (DstTy == MVT::i32 &&
5797 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5798 return std::make_pair(SDValue(), SDValue());
5799 if (Subtarget->is64Bit() &&
5800 DstTy == MVT::i64 &&
5801 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5802 return std::make_pair(SDValue(), SDValue());
5804 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5806 MachineFunction &MF = DAG.getMachineFunction();
5807 unsigned MemSize = DstTy.getSizeInBits()/8;
5808 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5809 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5812 switch (DstTy.getSimpleVT().SimpleTy) {
5813 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5814 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5815 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5816 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5819 SDValue Chain = DAG.getEntryNode();
5820 SDValue Value = Op.getOperand(0);
5821 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5822 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5823 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5824 PseudoSourceValue::getFixedStack(SSFI), 0,
5826 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5828 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5830 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5831 Chain = Value.getValue(1);
5832 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5833 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5836 // Build the FP_TO_INT*_IN_MEM
5837 SDValue Ops[] = { Chain, Value, StackSlot };
5838 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5840 return std::make_pair(FIST, StackSlot);
5843 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5844 SelectionDAG &DAG) const {
5845 if (Op.getValueType().isVector()) {
5846 if (Op.getValueType() == MVT::v2i32 &&
5847 Op.getOperand(0).getValueType() == MVT::v2f64) {
5853 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5854 SDValue FIST = Vals.first, StackSlot = Vals.second;
5855 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5856 if (FIST.getNode() == 0) return Op;
5859 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5860 FIST, StackSlot, NULL, 0, false, false, 0);
5863 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5864 SelectionDAG &DAG) const {
5865 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5866 SDValue FIST = Vals.first, StackSlot = Vals.second;
5867 assert(FIST.getNode() && "Unexpected failure");
5870 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5871 FIST, StackSlot, NULL, 0, false, false, 0);
5874 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5875 SelectionDAG &DAG) const {
5876 LLVMContext *Context = DAG.getContext();
5877 DebugLoc dl = Op.getDebugLoc();
5878 EVT VT = Op.getValueType();
5881 EltVT = VT.getVectorElementType();
5882 std::vector<Constant*> CV;
5883 if (EltVT == MVT::f64) {
5884 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5888 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5894 Constant *C = ConstantVector::get(CV);
5895 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5896 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5897 PseudoSourceValue::getConstantPool(), 0,
5899 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5902 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5903 LLVMContext *Context = DAG.getContext();
5904 DebugLoc dl = Op.getDebugLoc();
5905 EVT VT = Op.getValueType();
5908 EltVT = VT.getVectorElementType();
5909 std::vector<Constant*> CV;
5910 if (EltVT == MVT::f64) {
5911 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5915 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5921 Constant *C = ConstantVector::get(CV);
5922 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5923 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5924 PseudoSourceValue::getConstantPool(), 0,
5926 if (VT.isVector()) {
5927 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5928 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5929 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5931 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5933 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5937 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5938 LLVMContext *Context = DAG.getContext();
5939 SDValue Op0 = Op.getOperand(0);
5940 SDValue Op1 = Op.getOperand(1);
5941 DebugLoc dl = Op.getDebugLoc();
5942 EVT VT = Op.getValueType();
5943 EVT SrcVT = Op1.getValueType();
5945 // If second operand is smaller, extend it first.
5946 if (SrcVT.bitsLT(VT)) {
5947 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5950 // And if it is bigger, shrink it first.
5951 if (SrcVT.bitsGT(VT)) {
5952 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5956 // At this point the operands and the result should have the same
5957 // type, and that won't be f80 since that is not custom lowered.
5959 // First get the sign bit of second operand.
5960 std::vector<Constant*> CV;
5961 if (SrcVT == MVT::f64) {
5962 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5963 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5965 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5966 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5967 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5968 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5970 Constant *C = ConstantVector::get(CV);
5971 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5972 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5973 PseudoSourceValue::getConstantPool(), 0,
5975 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5977 // Shift sign bit right or left if the two operands have different types.
5978 if (SrcVT.bitsGT(VT)) {
5979 // Op0 is MVT::f32, Op1 is MVT::f64.
5980 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5981 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5982 DAG.getConstant(32, MVT::i32));
5983 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5984 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5985 DAG.getIntPtrConstant(0));
5988 // Clear first operand sign bit.
5990 if (VT == MVT::f64) {
5991 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5992 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5994 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5995 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5996 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5997 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5999 C = ConstantVector::get(CV);
6000 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6001 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6002 PseudoSourceValue::getConstantPool(), 0,
6004 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6006 // Or the value with the sign bit.
6007 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6010 /// Emit nodes that will be selected as "test Op0,Op0", or something
6012 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6013 SelectionDAG &DAG) const {
6014 DebugLoc dl = Op.getDebugLoc();
6016 // CF and OF aren't always set the way we want. Determine which
6017 // of these we need.
6018 bool NeedCF = false;
6019 bool NeedOF = false;
6022 case X86::COND_A: case X86::COND_AE:
6023 case X86::COND_B: case X86::COND_BE:
6026 case X86::COND_G: case X86::COND_GE:
6027 case X86::COND_L: case X86::COND_LE:
6028 case X86::COND_O: case X86::COND_NO:
6033 // See if we can use the EFLAGS value from the operand instead of
6034 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6035 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6036 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6037 // Emit a CMP with 0, which is the TEST pattern.
6038 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6039 DAG.getConstant(0, Op.getValueType()));
6041 unsigned Opcode = 0;
6042 unsigned NumOperands = 0;
6043 switch (Op.getNode()->getOpcode()) {
6045 // Due to an isel shortcoming, be conservative if this add is likely to be
6046 // selected as part of a load-modify-store instruction. When the root node
6047 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6048 // uses of other nodes in the match, such as the ADD in this case. This
6049 // leads to the ADD being left around and reselected, with the result being
6050 // two adds in the output. Alas, even if none our users are stores, that
6051 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6052 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6053 // climbing the DAG back to the root, and it doesn't seem to be worth the
6055 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6056 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6057 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6060 if (ConstantSDNode *C =
6061 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6062 // An add of one will be selected as an INC.
6063 if (C->getAPIntValue() == 1) {
6064 Opcode = X86ISD::INC;
6069 // An add of negative one (subtract of one) will be selected as a DEC.
6070 if (C->getAPIntValue().isAllOnesValue()) {
6071 Opcode = X86ISD::DEC;
6077 // Otherwise use a regular EFLAGS-setting add.
6078 Opcode = X86ISD::ADD;
6082 // If the primary and result isn't used, don't bother using X86ISD::AND,
6083 // because a TEST instruction will be better.
6084 bool NonFlagUse = false;
6085 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6086 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6088 unsigned UOpNo = UI.getOperandNo();
6089 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6090 // Look pass truncate.
6091 UOpNo = User->use_begin().getOperandNo();
6092 User = *User->use_begin();
6095 if (User->getOpcode() != ISD::BRCOND &&
6096 User->getOpcode() != ISD::SETCC &&
6097 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6110 // Due to the ISEL shortcoming noted above, be conservative if this op is
6111 // likely to be selected as part of a load-modify-store instruction.
6112 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6113 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6114 if (UI->getOpcode() == ISD::STORE)
6117 // Otherwise use a regular EFLAGS-setting instruction.
6118 switch (Op.getNode()->getOpcode()) {
6119 default: llvm_unreachable("unexpected operator!");
6120 case ISD::SUB: Opcode = X86ISD::SUB; break;
6121 case ISD::OR: Opcode = X86ISD::OR; break;
6122 case ISD::XOR: Opcode = X86ISD::XOR; break;
6123 case ISD::AND: Opcode = X86ISD::AND; break;
6135 return SDValue(Op.getNode(), 1);
6142 // Emit a CMP with 0, which is the TEST pattern.
6143 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6144 DAG.getConstant(0, Op.getValueType()));
6146 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6147 SmallVector<SDValue, 4> Ops;
6148 for (unsigned i = 0; i != NumOperands; ++i)
6149 Ops.push_back(Op.getOperand(i));
6151 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6152 DAG.ReplaceAllUsesWith(Op, New);
6153 return SDValue(New.getNode(), 1);
6156 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6158 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6159 SelectionDAG &DAG) const {
6160 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6161 if (C->getAPIntValue() == 0)
6162 return EmitTest(Op0, X86CC, DAG);
6164 DebugLoc dl = Op0.getDebugLoc();
6165 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6168 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6169 /// if it's possible.
6170 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6171 DebugLoc dl, SelectionDAG &DAG) const {
6172 SDValue Op0 = And.getOperand(0);
6173 SDValue Op1 = And.getOperand(1);
6174 if (Op0.getOpcode() == ISD::TRUNCATE)
6175 Op0 = Op0.getOperand(0);
6176 if (Op1.getOpcode() == ISD::TRUNCATE)
6177 Op1 = Op1.getOperand(0);
6180 if (Op1.getOpcode() == ISD::SHL)
6181 std::swap(Op0, Op1);
6182 if (Op0.getOpcode() == ISD::SHL) {
6183 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6184 if (And00C->getZExtValue() == 1) {
6185 // If we looked past a truncate, check that it's only truncating away
6187 unsigned BitWidth = Op0.getValueSizeInBits();
6188 unsigned AndBitWidth = And.getValueSizeInBits();
6189 if (BitWidth > AndBitWidth) {
6190 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6191 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6192 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6196 RHS = Op0.getOperand(1);
6198 } else if (Op1.getOpcode() == ISD::Constant) {
6199 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6200 SDValue AndLHS = Op0;
6201 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6202 LHS = AndLHS.getOperand(0);
6203 RHS = AndLHS.getOperand(1);
6207 if (LHS.getNode()) {
6208 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6209 // instruction. Since the shift amount is in-range-or-undefined, we know
6210 // that doing a bittest on the i32 value is ok. We extend to i32 because
6211 // the encoding for the i16 version is larger than the i32 version.
6212 // Also promote i16 to i32 for performance / code size reason.
6213 if (LHS.getValueType() == MVT::i8 ||
6214 LHS.getValueType() == MVT::i16)
6215 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6217 // If the operand types disagree, extend the shift amount to match. Since
6218 // BT ignores high bits (like shifts) we can use anyextend.
6219 if (LHS.getValueType() != RHS.getValueType())
6220 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6222 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6223 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6224 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6225 DAG.getConstant(Cond, MVT::i8), BT);
6231 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6232 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6233 SDValue Op0 = Op.getOperand(0);
6234 SDValue Op1 = Op.getOperand(1);
6235 DebugLoc dl = Op.getDebugLoc();
6236 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6238 // Optimize to BT if possible.
6239 // Lower (X & (1 << N)) == 0 to BT(X, N).
6240 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6241 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6242 if (Op0.getOpcode() == ISD::AND &&
6244 Op1.getOpcode() == ISD::Constant &&
6245 cast<ConstantSDNode>(Op1)->isNullValue() &&
6246 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6247 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6248 if (NewSetCC.getNode())
6252 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6253 if (Op0.getOpcode() == X86ISD::SETCC &&
6254 Op1.getOpcode() == ISD::Constant &&
6255 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6256 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6257 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6258 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6259 bool Invert = (CC == ISD::SETNE) ^
6260 cast<ConstantSDNode>(Op1)->isNullValue();
6262 CCode = X86::GetOppositeBranchCondition(CCode);
6263 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6264 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6267 bool isFP = Op1.getValueType().isFloatingPoint();
6268 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6269 if (X86CC == X86::COND_INVALID)
6272 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6274 // Use sbb x, x to materialize carry bit into a GPR.
6275 if (X86CC == X86::COND_B)
6276 return DAG.getNode(ISD::AND, dl, MVT::i8,
6277 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6278 DAG.getConstant(X86CC, MVT::i8), Cond),
6279 DAG.getConstant(1, MVT::i8));
6281 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6282 DAG.getConstant(X86CC, MVT::i8), Cond);
6285 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6287 SDValue Op0 = Op.getOperand(0);
6288 SDValue Op1 = Op.getOperand(1);
6289 SDValue CC = Op.getOperand(2);
6290 EVT VT = Op.getValueType();
6291 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6292 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6293 DebugLoc dl = Op.getDebugLoc();
6297 EVT VT0 = Op0.getValueType();
6298 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6299 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6302 switch (SetCCOpcode) {
6305 case ISD::SETEQ: SSECC = 0; break;
6307 case ISD::SETGT: Swap = true; // Fallthrough
6309 case ISD::SETOLT: SSECC = 1; break;
6311 case ISD::SETGE: Swap = true; // Fallthrough
6313 case ISD::SETOLE: SSECC = 2; break;
6314 case ISD::SETUO: SSECC = 3; break;
6316 case ISD::SETNE: SSECC = 4; break;
6317 case ISD::SETULE: Swap = true;
6318 case ISD::SETUGE: SSECC = 5; break;
6319 case ISD::SETULT: Swap = true;
6320 case ISD::SETUGT: SSECC = 6; break;
6321 case ISD::SETO: SSECC = 7; break;
6324 std::swap(Op0, Op1);
6326 // In the two special cases we can't handle, emit two comparisons.
6328 if (SetCCOpcode == ISD::SETUEQ) {
6330 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6331 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6332 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6334 else if (SetCCOpcode == ISD::SETONE) {
6336 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6337 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6338 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6340 llvm_unreachable("Illegal FP comparison");
6342 // Handle all other FP comparisons here.
6343 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6346 // We are handling one of the integer comparisons here. Since SSE only has
6347 // GT and EQ comparisons for integer, swapping operands and multiple
6348 // operations may be required for some comparisons.
6349 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6350 bool Swap = false, Invert = false, FlipSigns = false;
6352 switch (VT.getSimpleVT().SimpleTy) {
6355 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6357 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6359 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6360 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6363 switch (SetCCOpcode) {
6365 case ISD::SETNE: Invert = true;
6366 case ISD::SETEQ: Opc = EQOpc; break;
6367 case ISD::SETLT: Swap = true;
6368 case ISD::SETGT: Opc = GTOpc; break;
6369 case ISD::SETGE: Swap = true;
6370 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6371 case ISD::SETULT: Swap = true;
6372 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6373 case ISD::SETUGE: Swap = true;
6374 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6377 std::swap(Op0, Op1);
6379 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6380 // bits of the inputs before performing those operations.
6382 EVT EltVT = VT.getVectorElementType();
6383 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6385 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6386 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6388 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6389 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6392 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6394 // If the logical-not of the result is required, perform that now.
6396 Result = DAG.getNOT(dl, Result, VT);
6401 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6402 static bool isX86LogicalCmp(SDValue Op) {
6403 unsigned Opc = Op.getNode()->getOpcode();
6404 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6406 if (Op.getResNo() == 1 &&
6407 (Opc == X86ISD::ADD ||
6408 Opc == X86ISD::SUB ||
6409 Opc == X86ISD::SMUL ||
6410 Opc == X86ISD::UMUL ||
6411 Opc == X86ISD::INC ||
6412 Opc == X86ISD::DEC ||
6413 Opc == X86ISD::OR ||
6414 Opc == X86ISD::XOR ||
6415 Opc == X86ISD::AND))
6421 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6422 bool addTest = true;
6423 SDValue Cond = Op.getOperand(0);
6424 DebugLoc dl = Op.getDebugLoc();
6427 if (Cond.getOpcode() == ISD::SETCC) {
6428 SDValue NewCond = LowerSETCC(Cond, DAG);
6429 if (NewCond.getNode())
6433 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6434 SDValue Op1 = Op.getOperand(1);
6435 SDValue Op2 = Op.getOperand(2);
6436 if (Cond.getOpcode() == X86ISD::SETCC &&
6437 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6438 SDValue Cmp = Cond.getOperand(1);
6439 if (Cmp.getOpcode() == X86ISD::CMP) {
6440 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6441 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6442 ConstantSDNode *RHSC =
6443 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6444 if (N1C && N1C->isAllOnesValue() &&
6445 N2C && N2C->isNullValue() &&
6446 RHSC && RHSC->isNullValue()) {
6447 SDValue CmpOp0 = Cmp.getOperand(0);
6448 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6449 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6450 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6451 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6456 // Look pass (and (setcc_carry (cmp ...)), 1).
6457 if (Cond.getOpcode() == ISD::AND &&
6458 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6459 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6460 if (C && C->getAPIntValue() == 1)
6461 Cond = Cond.getOperand(0);
6464 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6465 // setting operand in place of the X86ISD::SETCC.
6466 if (Cond.getOpcode() == X86ISD::SETCC ||
6467 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6468 CC = Cond.getOperand(0);
6470 SDValue Cmp = Cond.getOperand(1);
6471 unsigned Opc = Cmp.getOpcode();
6472 EVT VT = Op.getValueType();
6474 bool IllegalFPCMov = false;
6475 if (VT.isFloatingPoint() && !VT.isVector() &&
6476 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6477 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6479 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6480 Opc == X86ISD::BT) { // FIXME
6487 // Look pass the truncate.
6488 if (Cond.getOpcode() == ISD::TRUNCATE)
6489 Cond = Cond.getOperand(0);
6491 // We know the result of AND is compared against zero. Try to match
6493 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6494 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6495 if (NewSetCC.getNode()) {
6496 CC = NewSetCC.getOperand(0);
6497 Cond = NewSetCC.getOperand(1);
6504 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6505 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6508 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6509 // condition is true.
6510 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6511 SDValue Ops[] = { Op2, Op1, CC, Cond };
6512 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6515 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6516 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6517 // from the AND / OR.
6518 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6519 Opc = Op.getOpcode();
6520 if (Opc != ISD::OR && Opc != ISD::AND)
6522 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6523 Op.getOperand(0).hasOneUse() &&
6524 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6525 Op.getOperand(1).hasOneUse());
6528 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6529 // 1 and that the SETCC node has a single use.
6530 static bool isXor1OfSetCC(SDValue Op) {
6531 if (Op.getOpcode() != ISD::XOR)
6533 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6534 if (N1C && N1C->getAPIntValue() == 1) {
6535 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6536 Op.getOperand(0).hasOneUse();
6541 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6542 bool addTest = true;
6543 SDValue Chain = Op.getOperand(0);
6544 SDValue Cond = Op.getOperand(1);
6545 SDValue Dest = Op.getOperand(2);
6546 DebugLoc dl = Op.getDebugLoc();
6549 if (Cond.getOpcode() == ISD::SETCC) {
6550 SDValue NewCond = LowerSETCC(Cond, DAG);
6551 if (NewCond.getNode())
6555 // FIXME: LowerXALUO doesn't handle these!!
6556 else if (Cond.getOpcode() == X86ISD::ADD ||
6557 Cond.getOpcode() == X86ISD::SUB ||
6558 Cond.getOpcode() == X86ISD::SMUL ||
6559 Cond.getOpcode() == X86ISD::UMUL)
6560 Cond = LowerXALUO(Cond, DAG);
6563 // Look pass (and (setcc_carry (cmp ...)), 1).
6564 if (Cond.getOpcode() == ISD::AND &&
6565 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6566 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6567 if (C && C->getAPIntValue() == 1)
6568 Cond = Cond.getOperand(0);
6571 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6572 // setting operand in place of the X86ISD::SETCC.
6573 if (Cond.getOpcode() == X86ISD::SETCC ||
6574 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6575 CC = Cond.getOperand(0);
6577 SDValue Cmp = Cond.getOperand(1);
6578 unsigned Opc = Cmp.getOpcode();
6579 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6580 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6584 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6588 // These can only come from an arithmetic instruction with overflow,
6589 // e.g. SADDO, UADDO.
6590 Cond = Cond.getNode()->getOperand(1);
6597 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6598 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6599 if (CondOpc == ISD::OR) {
6600 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6601 // two branches instead of an explicit OR instruction with a
6603 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6604 isX86LogicalCmp(Cmp)) {
6605 CC = Cond.getOperand(0).getOperand(0);
6606 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6607 Chain, Dest, CC, Cmp);
6608 CC = Cond.getOperand(1).getOperand(0);
6612 } else { // ISD::AND
6613 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6614 // two branches instead of an explicit AND instruction with a
6615 // separate test. However, we only do this if this block doesn't
6616 // have a fall-through edge, because this requires an explicit
6617 // jmp when the condition is false.
6618 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6619 isX86LogicalCmp(Cmp) &&
6620 Op.getNode()->hasOneUse()) {
6621 X86::CondCode CCode =
6622 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6623 CCode = X86::GetOppositeBranchCondition(CCode);
6624 CC = DAG.getConstant(CCode, MVT::i8);
6625 SDNode *User = *Op.getNode()->use_begin();
6626 // Look for an unconditional branch following this conditional branch.
6627 // We need this because we need to reverse the successors in order
6628 // to implement FCMP_OEQ.
6629 if (User->getOpcode() == ISD::BR) {
6630 SDValue FalseBB = User->getOperand(1);
6632 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6633 assert(NewBR == User);
6637 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6638 Chain, Dest, CC, Cmp);
6639 X86::CondCode CCode =
6640 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6641 CCode = X86::GetOppositeBranchCondition(CCode);
6642 CC = DAG.getConstant(CCode, MVT::i8);
6648 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6649 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6650 // It should be transformed during dag combiner except when the condition
6651 // is set by a arithmetics with overflow node.
6652 X86::CondCode CCode =
6653 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6654 CCode = X86::GetOppositeBranchCondition(CCode);
6655 CC = DAG.getConstant(CCode, MVT::i8);
6656 Cond = Cond.getOperand(0).getOperand(1);
6662 // Look pass the truncate.
6663 if (Cond.getOpcode() == ISD::TRUNCATE)
6664 Cond = Cond.getOperand(0);
6666 // We know the result of AND is compared against zero. Try to match
6668 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6669 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6670 if (NewSetCC.getNode()) {
6671 CC = NewSetCC.getOperand(0);
6672 Cond = NewSetCC.getOperand(1);
6679 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6680 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6682 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6683 Chain, Dest, CC, Cond);
6687 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6688 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6689 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6690 // that the guard pages used by the OS virtual memory manager are allocated in
6691 // correct sequence.
6693 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6694 SelectionDAG &DAG) const {
6695 assert(Subtarget->isTargetCygMing() &&
6696 "This should be used only on Cygwin/Mingw targets");
6697 DebugLoc dl = Op.getDebugLoc();
6700 SDValue Chain = Op.getOperand(0);
6701 SDValue Size = Op.getOperand(1);
6702 // FIXME: Ensure alignment here
6706 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6708 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6709 Flag = Chain.getValue(1);
6711 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6713 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6714 Flag = Chain.getValue(1);
6716 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6718 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6719 return DAG.getMergeValues(Ops1, 2, dl);
6722 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6723 MachineFunction &MF = DAG.getMachineFunction();
6724 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6726 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6727 DebugLoc dl = Op.getDebugLoc();
6729 if (!Subtarget->is64Bit()) {
6730 // vastart just stores the address of the VarArgsFrameIndex slot into the
6731 // memory location argument.
6732 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6734 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6739 // gp_offset (0 - 6 * 8)
6740 // fp_offset (48 - 48 + 8 * 16)
6741 // overflow_arg_area (point to parameters coming in memory).
6743 SmallVector<SDValue, 8> MemOps;
6744 SDValue FIN = Op.getOperand(1);
6746 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6747 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6749 FIN, SV, 0, false, false, 0);
6750 MemOps.push_back(Store);
6753 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6754 FIN, DAG.getIntPtrConstant(4));
6755 Store = DAG.getStore(Op.getOperand(0), dl,
6756 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6758 FIN, SV, 0, false, false, 0);
6759 MemOps.push_back(Store);
6761 // Store ptr to overflow_arg_area
6762 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6763 FIN, DAG.getIntPtrConstant(4));
6764 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6766 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6768 MemOps.push_back(Store);
6770 // Store ptr to reg_save_area.
6771 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6772 FIN, DAG.getIntPtrConstant(8));
6773 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6775 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6777 MemOps.push_back(Store);
6778 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6779 &MemOps[0], MemOps.size());
6782 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6783 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6784 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6786 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6790 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6791 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6792 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6793 SDValue Chain = Op.getOperand(0);
6794 SDValue DstPtr = Op.getOperand(1);
6795 SDValue SrcPtr = Op.getOperand(2);
6796 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6797 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6798 DebugLoc dl = Op.getDebugLoc();
6800 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6801 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6802 false, DstSV, 0, SrcSV, 0);
6806 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6807 DebugLoc dl = Op.getDebugLoc();
6808 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6810 default: return SDValue(); // Don't custom lower most intrinsics.
6811 // Comparison intrinsics.
6812 case Intrinsic::x86_sse_comieq_ss:
6813 case Intrinsic::x86_sse_comilt_ss:
6814 case Intrinsic::x86_sse_comile_ss:
6815 case Intrinsic::x86_sse_comigt_ss:
6816 case Intrinsic::x86_sse_comige_ss:
6817 case Intrinsic::x86_sse_comineq_ss:
6818 case Intrinsic::x86_sse_ucomieq_ss:
6819 case Intrinsic::x86_sse_ucomilt_ss:
6820 case Intrinsic::x86_sse_ucomile_ss:
6821 case Intrinsic::x86_sse_ucomigt_ss:
6822 case Intrinsic::x86_sse_ucomige_ss:
6823 case Intrinsic::x86_sse_ucomineq_ss:
6824 case Intrinsic::x86_sse2_comieq_sd:
6825 case Intrinsic::x86_sse2_comilt_sd:
6826 case Intrinsic::x86_sse2_comile_sd:
6827 case Intrinsic::x86_sse2_comigt_sd:
6828 case Intrinsic::x86_sse2_comige_sd:
6829 case Intrinsic::x86_sse2_comineq_sd:
6830 case Intrinsic::x86_sse2_ucomieq_sd:
6831 case Intrinsic::x86_sse2_ucomilt_sd:
6832 case Intrinsic::x86_sse2_ucomile_sd:
6833 case Intrinsic::x86_sse2_ucomigt_sd:
6834 case Intrinsic::x86_sse2_ucomige_sd:
6835 case Intrinsic::x86_sse2_ucomineq_sd: {
6837 ISD::CondCode CC = ISD::SETCC_INVALID;
6840 case Intrinsic::x86_sse_comieq_ss:
6841 case Intrinsic::x86_sse2_comieq_sd:
6845 case Intrinsic::x86_sse_comilt_ss:
6846 case Intrinsic::x86_sse2_comilt_sd:
6850 case Intrinsic::x86_sse_comile_ss:
6851 case Intrinsic::x86_sse2_comile_sd:
6855 case Intrinsic::x86_sse_comigt_ss:
6856 case Intrinsic::x86_sse2_comigt_sd:
6860 case Intrinsic::x86_sse_comige_ss:
6861 case Intrinsic::x86_sse2_comige_sd:
6865 case Intrinsic::x86_sse_comineq_ss:
6866 case Intrinsic::x86_sse2_comineq_sd:
6870 case Intrinsic::x86_sse_ucomieq_ss:
6871 case Intrinsic::x86_sse2_ucomieq_sd:
6872 Opc = X86ISD::UCOMI;
6875 case Intrinsic::x86_sse_ucomilt_ss:
6876 case Intrinsic::x86_sse2_ucomilt_sd:
6877 Opc = X86ISD::UCOMI;
6880 case Intrinsic::x86_sse_ucomile_ss:
6881 case Intrinsic::x86_sse2_ucomile_sd:
6882 Opc = X86ISD::UCOMI;
6885 case Intrinsic::x86_sse_ucomigt_ss:
6886 case Intrinsic::x86_sse2_ucomigt_sd:
6887 Opc = X86ISD::UCOMI;
6890 case Intrinsic::x86_sse_ucomige_ss:
6891 case Intrinsic::x86_sse2_ucomige_sd:
6892 Opc = X86ISD::UCOMI;
6895 case Intrinsic::x86_sse_ucomineq_ss:
6896 case Intrinsic::x86_sse2_ucomineq_sd:
6897 Opc = X86ISD::UCOMI;
6902 SDValue LHS = Op.getOperand(1);
6903 SDValue RHS = Op.getOperand(2);
6904 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6905 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6906 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6907 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6908 DAG.getConstant(X86CC, MVT::i8), Cond);
6909 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6911 // ptest intrinsics. The intrinsic these come from are designed to return
6912 // an integer value, not just an instruction so lower it to the ptest
6913 // pattern and a setcc for the result.
6914 case Intrinsic::x86_sse41_ptestz:
6915 case Intrinsic::x86_sse41_ptestc:
6916 case Intrinsic::x86_sse41_ptestnzc:{
6919 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6920 case Intrinsic::x86_sse41_ptestz:
6922 X86CC = X86::COND_E;
6924 case Intrinsic::x86_sse41_ptestc:
6926 X86CC = X86::COND_B;
6928 case Intrinsic::x86_sse41_ptestnzc:
6930 X86CC = X86::COND_A;
6934 SDValue LHS = Op.getOperand(1);
6935 SDValue RHS = Op.getOperand(2);
6936 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6937 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6938 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6939 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6942 // Fix vector shift instructions where the last operand is a non-immediate
6944 case Intrinsic::x86_sse2_pslli_w:
6945 case Intrinsic::x86_sse2_pslli_d:
6946 case Intrinsic::x86_sse2_pslli_q:
6947 case Intrinsic::x86_sse2_psrli_w:
6948 case Intrinsic::x86_sse2_psrli_d:
6949 case Intrinsic::x86_sse2_psrli_q:
6950 case Intrinsic::x86_sse2_psrai_w:
6951 case Intrinsic::x86_sse2_psrai_d:
6952 case Intrinsic::x86_mmx_pslli_w:
6953 case Intrinsic::x86_mmx_pslli_d:
6954 case Intrinsic::x86_mmx_pslli_q:
6955 case Intrinsic::x86_mmx_psrli_w:
6956 case Intrinsic::x86_mmx_psrli_d:
6957 case Intrinsic::x86_mmx_psrli_q:
6958 case Intrinsic::x86_mmx_psrai_w:
6959 case Intrinsic::x86_mmx_psrai_d: {
6960 SDValue ShAmt = Op.getOperand(2);
6961 if (isa<ConstantSDNode>(ShAmt))
6964 unsigned NewIntNo = 0;
6965 EVT ShAmtVT = MVT::v4i32;
6967 case Intrinsic::x86_sse2_pslli_w:
6968 NewIntNo = Intrinsic::x86_sse2_psll_w;
6970 case Intrinsic::x86_sse2_pslli_d:
6971 NewIntNo = Intrinsic::x86_sse2_psll_d;
6973 case Intrinsic::x86_sse2_pslli_q:
6974 NewIntNo = Intrinsic::x86_sse2_psll_q;
6976 case Intrinsic::x86_sse2_psrli_w:
6977 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6979 case Intrinsic::x86_sse2_psrli_d:
6980 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6982 case Intrinsic::x86_sse2_psrli_q:
6983 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6985 case Intrinsic::x86_sse2_psrai_w:
6986 NewIntNo = Intrinsic::x86_sse2_psra_w;
6988 case Intrinsic::x86_sse2_psrai_d:
6989 NewIntNo = Intrinsic::x86_sse2_psra_d;
6992 ShAmtVT = MVT::v2i32;
6994 case Intrinsic::x86_mmx_pslli_w:
6995 NewIntNo = Intrinsic::x86_mmx_psll_w;
6997 case Intrinsic::x86_mmx_pslli_d:
6998 NewIntNo = Intrinsic::x86_mmx_psll_d;
7000 case Intrinsic::x86_mmx_pslli_q:
7001 NewIntNo = Intrinsic::x86_mmx_psll_q;
7003 case Intrinsic::x86_mmx_psrli_w:
7004 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7006 case Intrinsic::x86_mmx_psrli_d:
7007 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7009 case Intrinsic::x86_mmx_psrli_q:
7010 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7012 case Intrinsic::x86_mmx_psrai_w:
7013 NewIntNo = Intrinsic::x86_mmx_psra_w;
7015 case Intrinsic::x86_mmx_psrai_d:
7016 NewIntNo = Intrinsic::x86_mmx_psra_d;
7018 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7024 // The vector shift intrinsics with scalars uses 32b shift amounts but
7025 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7029 ShOps[1] = DAG.getConstant(0, MVT::i32);
7030 if (ShAmtVT == MVT::v4i32) {
7031 ShOps[2] = DAG.getUNDEF(MVT::i32);
7032 ShOps[3] = DAG.getUNDEF(MVT::i32);
7033 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7035 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7038 EVT VT = Op.getValueType();
7039 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7040 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7041 DAG.getConstant(NewIntNo, MVT::i32),
7042 Op.getOperand(1), ShAmt);
7047 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7048 SelectionDAG &DAG) const {
7049 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7050 MFI->setReturnAddressIsTaken(true);
7052 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7053 DebugLoc dl = Op.getDebugLoc();
7056 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7058 DAG.getConstant(TD->getPointerSize(),
7059 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7060 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7061 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7063 NULL, 0, false, false, 0);
7066 // Just load the return address.
7067 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7068 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7069 RetAddrFI, NULL, 0, false, false, 0);
7072 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7073 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7074 MFI->setFrameAddressIsTaken(true);
7076 EVT VT = Op.getValueType();
7077 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7078 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7079 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7080 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7082 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7087 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7088 SelectionDAG &DAG) const {
7089 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7092 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7093 MachineFunction &MF = DAG.getMachineFunction();
7094 SDValue Chain = Op.getOperand(0);
7095 SDValue Offset = Op.getOperand(1);
7096 SDValue Handler = Op.getOperand(2);
7097 DebugLoc dl = Op.getDebugLoc();
7099 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7101 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7103 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7104 DAG.getIntPtrConstant(-TD->getPointerSize()));
7105 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7106 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7107 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7108 MF.getRegInfo().addLiveOut(StoreAddrReg);
7110 return DAG.getNode(X86ISD::EH_RETURN, dl,
7112 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7115 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7116 SelectionDAG &DAG) const {
7117 SDValue Root = Op.getOperand(0);
7118 SDValue Trmp = Op.getOperand(1); // trampoline
7119 SDValue FPtr = Op.getOperand(2); // nested function
7120 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7121 DebugLoc dl = Op.getDebugLoc();
7123 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7125 if (Subtarget->is64Bit()) {
7126 SDValue OutChains[6];
7128 // Large code-model.
7129 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7130 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7132 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7133 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7135 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7137 // Load the pointer to the nested function into R11.
7138 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7139 SDValue Addr = Trmp;
7140 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7141 Addr, TrmpAddr, 0, false, false, 0);
7143 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7144 DAG.getConstant(2, MVT::i64));
7145 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7148 // Load the 'nest' parameter value into R10.
7149 // R10 is specified in X86CallingConv.td
7150 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7151 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7152 DAG.getConstant(10, MVT::i64));
7153 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7154 Addr, TrmpAddr, 10, false, false, 0);
7156 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7157 DAG.getConstant(12, MVT::i64));
7158 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7161 // Jump to the nested function.
7162 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7163 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7164 DAG.getConstant(20, MVT::i64));
7165 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7166 Addr, TrmpAddr, 20, false, false, 0);
7168 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7169 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7170 DAG.getConstant(22, MVT::i64));
7171 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7172 TrmpAddr, 22, false, false, 0);
7175 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7176 return DAG.getMergeValues(Ops, 2, dl);
7178 const Function *Func =
7179 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7180 CallingConv::ID CC = Func->getCallingConv();
7185 llvm_unreachable("Unsupported calling convention");
7186 case CallingConv::C:
7187 case CallingConv::X86_StdCall: {
7188 // Pass 'nest' parameter in ECX.
7189 // Must be kept in sync with X86CallingConv.td
7192 // Check that ECX wasn't needed by an 'inreg' parameter.
7193 const FunctionType *FTy = Func->getFunctionType();
7194 const AttrListPtr &Attrs = Func->getAttributes();
7196 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7197 unsigned InRegCount = 0;
7200 for (FunctionType::param_iterator I = FTy->param_begin(),
7201 E = FTy->param_end(); I != E; ++I, ++Idx)
7202 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7203 // FIXME: should only count parameters that are lowered to integers.
7204 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7206 if (InRegCount > 2) {
7207 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7212 case CallingConv::X86_FastCall:
7213 case CallingConv::X86_ThisCall:
7214 case CallingConv::Fast:
7215 // Pass 'nest' parameter in EAX.
7216 // Must be kept in sync with X86CallingConv.td
7221 SDValue OutChains[4];
7224 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7225 DAG.getConstant(10, MVT::i32));
7226 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7228 // This is storing the opcode for MOV32ri.
7229 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7230 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7231 OutChains[0] = DAG.getStore(Root, dl,
7232 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7233 Trmp, TrmpAddr, 0, false, false, 0);
7235 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7236 DAG.getConstant(1, MVT::i32));
7237 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7240 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7241 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7242 DAG.getConstant(5, MVT::i32));
7243 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7244 TrmpAddr, 5, false, false, 1);
7246 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7247 DAG.getConstant(6, MVT::i32));
7248 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7252 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7253 return DAG.getMergeValues(Ops, 2, dl);
7257 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7258 SelectionDAG &DAG) const {
7260 The rounding mode is in bits 11:10 of FPSR, and has the following
7267 FLT_ROUNDS, on the other hand, expects the following:
7274 To perform the conversion, we do:
7275 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7278 MachineFunction &MF = DAG.getMachineFunction();
7279 const TargetMachine &TM = MF.getTarget();
7280 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7281 unsigned StackAlignment = TFI.getStackAlignment();
7282 EVT VT = Op.getValueType();
7283 DebugLoc dl = Op.getDebugLoc();
7285 // Save FP Control Word to stack slot
7286 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7287 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7289 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7290 DAG.getEntryNode(), StackSlot);
7292 // Load FP Control Word from stack slot
7293 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7296 // Transform as necessary
7298 DAG.getNode(ISD::SRL, dl, MVT::i16,
7299 DAG.getNode(ISD::AND, dl, MVT::i16,
7300 CWD, DAG.getConstant(0x800, MVT::i16)),
7301 DAG.getConstant(11, MVT::i8));
7303 DAG.getNode(ISD::SRL, dl, MVT::i16,
7304 DAG.getNode(ISD::AND, dl, MVT::i16,
7305 CWD, DAG.getConstant(0x400, MVT::i16)),
7306 DAG.getConstant(9, MVT::i8));
7309 DAG.getNode(ISD::AND, dl, MVT::i16,
7310 DAG.getNode(ISD::ADD, dl, MVT::i16,
7311 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7312 DAG.getConstant(1, MVT::i16)),
7313 DAG.getConstant(3, MVT::i16));
7316 return DAG.getNode((VT.getSizeInBits() < 16 ?
7317 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7320 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7321 EVT VT = Op.getValueType();
7323 unsigned NumBits = VT.getSizeInBits();
7324 DebugLoc dl = Op.getDebugLoc();
7326 Op = Op.getOperand(0);
7327 if (VT == MVT::i8) {
7328 // Zero extend to i32 since there is not an i8 bsr.
7330 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7333 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7334 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7335 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7337 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7340 DAG.getConstant(NumBits+NumBits-1, OpVT),
7341 DAG.getConstant(X86::COND_E, MVT::i8),
7344 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7346 // Finally xor with NumBits-1.
7347 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7350 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7354 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7355 EVT VT = Op.getValueType();
7357 unsigned NumBits = VT.getSizeInBits();
7358 DebugLoc dl = Op.getDebugLoc();
7360 Op = Op.getOperand(0);
7361 if (VT == MVT::i8) {
7363 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7366 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7367 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7368 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7370 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7373 DAG.getConstant(NumBits, OpVT),
7374 DAG.getConstant(X86::COND_E, MVT::i8),
7377 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7380 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7384 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7385 EVT VT = Op.getValueType();
7386 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7387 DebugLoc dl = Op.getDebugLoc();
7389 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7390 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7391 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7392 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7393 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7395 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7396 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7397 // return AloBlo + AloBhi + AhiBlo;
7399 SDValue A = Op.getOperand(0);
7400 SDValue B = Op.getOperand(1);
7402 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7403 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7404 A, DAG.getConstant(32, MVT::i32));
7405 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7406 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7407 B, DAG.getConstant(32, MVT::i32));
7408 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7409 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7411 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7412 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7414 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7415 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7417 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7418 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7419 AloBhi, DAG.getConstant(32, MVT::i32));
7420 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7421 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7422 AhiBlo, DAG.getConstant(32, MVT::i32));
7423 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7424 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7429 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7430 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7431 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7432 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7433 // has only one use.
7434 SDNode *N = Op.getNode();
7435 SDValue LHS = N->getOperand(0);
7436 SDValue RHS = N->getOperand(1);
7437 unsigned BaseOp = 0;
7439 DebugLoc dl = Op.getDebugLoc();
7441 switch (Op.getOpcode()) {
7442 default: llvm_unreachable("Unknown ovf instruction!");
7444 // A subtract of one will be selected as a INC. Note that INC doesn't
7445 // set CF, so we can't do this for UADDO.
7446 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7447 if (C->getAPIntValue() == 1) {
7448 BaseOp = X86ISD::INC;
7452 BaseOp = X86ISD::ADD;
7456 BaseOp = X86ISD::ADD;
7460 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7461 // set CF, so we can't do this for USUBO.
7462 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7463 if (C->getAPIntValue() == 1) {
7464 BaseOp = X86ISD::DEC;
7468 BaseOp = X86ISD::SUB;
7472 BaseOp = X86ISD::SUB;
7476 BaseOp = X86ISD::SMUL;
7480 BaseOp = X86ISD::UMUL;
7485 // Also sets EFLAGS.
7486 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7487 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7490 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7491 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7493 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7497 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7498 EVT T = Op.getValueType();
7499 DebugLoc dl = Op.getDebugLoc();
7502 switch(T.getSimpleVT().SimpleTy) {
7504 assert(false && "Invalid value type!");
7505 case MVT::i8: Reg = X86::AL; size = 1; break;
7506 case MVT::i16: Reg = X86::AX; size = 2; break;
7507 case MVT::i32: Reg = X86::EAX; size = 4; break;
7509 assert(Subtarget->is64Bit() && "Node not type legal!");
7510 Reg = X86::RAX; size = 8;
7513 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7514 Op.getOperand(2), SDValue());
7515 SDValue Ops[] = { cpIn.getValue(0),
7518 DAG.getTargetConstant(size, MVT::i8),
7520 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7521 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7523 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7527 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7528 SelectionDAG &DAG) const {
7529 assert(Subtarget->is64Bit() && "Result not type legalized?");
7530 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7531 SDValue TheChain = Op.getOperand(0);
7532 DebugLoc dl = Op.getDebugLoc();
7533 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7534 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7535 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7537 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7538 DAG.getConstant(32, MVT::i8));
7540 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7543 return DAG.getMergeValues(Ops, 2, dl);
7546 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7547 SelectionDAG &DAG) const {
7548 EVT SrcVT = Op.getOperand(0).getValueType();
7549 EVT DstVT = Op.getValueType();
7550 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7551 Subtarget->hasMMX() && !DisableMMX) &&
7552 "Unexpected custom BIT_CONVERT");
7553 assert((DstVT == MVT::i64 ||
7554 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7555 "Unexpected custom BIT_CONVERT");
7556 // i64 <=> MMX conversions are Legal.
7557 if (SrcVT==MVT::i64 && DstVT.isVector())
7559 if (DstVT==MVT::i64 && SrcVT.isVector())
7561 // MMX <=> MMX conversions are Legal.
7562 if (SrcVT.isVector() && DstVT.isVector())
7564 // All other conversions need to be expanded.
7567 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7568 SDNode *Node = Op.getNode();
7569 DebugLoc dl = Node->getDebugLoc();
7570 EVT T = Node->getValueType(0);
7571 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7572 DAG.getConstant(0, T), Node->getOperand(2));
7573 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7574 cast<AtomicSDNode>(Node)->getMemoryVT(),
7575 Node->getOperand(0),
7576 Node->getOperand(1), negOp,
7577 cast<AtomicSDNode>(Node)->getSrcValue(),
7578 cast<AtomicSDNode>(Node)->getAlignment());
7581 /// LowerOperation - Provide custom lowering hooks for some operations.
7583 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7584 switch (Op.getOpcode()) {
7585 default: llvm_unreachable("Should not custom lower this!");
7586 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7587 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7588 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7589 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7590 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7591 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7592 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7593 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7594 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7595 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7596 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7597 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7598 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7599 case ISD::SHL_PARTS:
7600 case ISD::SRA_PARTS:
7601 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7602 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7603 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7604 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7605 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7606 case ISD::FABS: return LowerFABS(Op, DAG);
7607 case ISD::FNEG: return LowerFNEG(Op, DAG);
7608 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7609 case ISD::SETCC: return LowerSETCC(Op, DAG);
7610 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7611 case ISD::SELECT: return LowerSELECT(Op, DAG);
7612 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7613 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7614 case ISD::VASTART: return LowerVASTART(Op, DAG);
7615 case ISD::VAARG: return LowerVAARG(Op, DAG);
7616 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7617 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7618 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7619 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7620 case ISD::FRAME_TO_ARGS_OFFSET:
7621 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7622 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7623 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7624 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7625 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7626 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7627 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7628 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7634 case ISD::UMULO: return LowerXALUO(Op, DAG);
7635 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7636 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7640 void X86TargetLowering::
7641 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7642 SelectionDAG &DAG, unsigned NewOp) const {
7643 EVT T = Node->getValueType(0);
7644 DebugLoc dl = Node->getDebugLoc();
7645 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7647 SDValue Chain = Node->getOperand(0);
7648 SDValue In1 = Node->getOperand(1);
7649 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7650 Node->getOperand(2), DAG.getIntPtrConstant(0));
7651 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7652 Node->getOperand(2), DAG.getIntPtrConstant(1));
7653 SDValue Ops[] = { Chain, In1, In2L, In2H };
7654 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7656 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7657 cast<MemSDNode>(Node)->getMemOperand());
7658 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7659 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7660 Results.push_back(Result.getValue(2));
7663 /// ReplaceNodeResults - Replace a node with an illegal result type
7664 /// with a new node built out of custom code.
7665 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7666 SmallVectorImpl<SDValue>&Results,
7667 SelectionDAG &DAG) const {
7668 DebugLoc dl = N->getDebugLoc();
7669 switch (N->getOpcode()) {
7671 assert(false && "Do not know how to custom type legalize this operation!");
7673 case ISD::FP_TO_SINT: {
7674 std::pair<SDValue,SDValue> Vals =
7675 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7676 SDValue FIST = Vals.first, StackSlot = Vals.second;
7677 if (FIST.getNode() != 0) {
7678 EVT VT = N->getValueType(0);
7679 // Return a load from the stack slot.
7680 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7685 case ISD::READCYCLECOUNTER: {
7686 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7687 SDValue TheChain = N->getOperand(0);
7688 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7689 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7691 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7693 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7694 SDValue Ops[] = { eax, edx };
7695 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7696 Results.push_back(edx.getValue(1));
7699 case ISD::ATOMIC_CMP_SWAP: {
7700 EVT T = N->getValueType(0);
7701 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7702 SDValue cpInL, cpInH;
7703 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7704 DAG.getConstant(0, MVT::i32));
7705 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7706 DAG.getConstant(1, MVT::i32));
7707 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7708 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7710 SDValue swapInL, swapInH;
7711 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7712 DAG.getConstant(0, MVT::i32));
7713 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7714 DAG.getConstant(1, MVT::i32));
7715 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7717 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7718 swapInL.getValue(1));
7719 SDValue Ops[] = { swapInH.getValue(0),
7721 swapInH.getValue(1) };
7722 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7723 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7724 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7725 MVT::i32, Result.getValue(1));
7726 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7727 MVT::i32, cpOutL.getValue(2));
7728 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7729 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7730 Results.push_back(cpOutH.getValue(1));
7733 case ISD::ATOMIC_LOAD_ADD:
7734 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7736 case ISD::ATOMIC_LOAD_AND:
7737 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7739 case ISD::ATOMIC_LOAD_NAND:
7740 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7742 case ISD::ATOMIC_LOAD_OR:
7743 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7745 case ISD::ATOMIC_LOAD_SUB:
7746 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7748 case ISD::ATOMIC_LOAD_XOR:
7749 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7751 case ISD::ATOMIC_SWAP:
7752 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7757 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7759 default: return NULL;
7760 case X86ISD::BSF: return "X86ISD::BSF";
7761 case X86ISD::BSR: return "X86ISD::BSR";
7762 case X86ISD::SHLD: return "X86ISD::SHLD";
7763 case X86ISD::SHRD: return "X86ISD::SHRD";
7764 case X86ISD::FAND: return "X86ISD::FAND";
7765 case X86ISD::FOR: return "X86ISD::FOR";
7766 case X86ISD::FXOR: return "X86ISD::FXOR";
7767 case X86ISD::FSRL: return "X86ISD::FSRL";
7768 case X86ISD::FILD: return "X86ISD::FILD";
7769 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7770 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7771 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7772 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7773 case X86ISD::FLD: return "X86ISD::FLD";
7774 case X86ISD::FST: return "X86ISD::FST";
7775 case X86ISD::CALL: return "X86ISD::CALL";
7776 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7777 case X86ISD::BT: return "X86ISD::BT";
7778 case X86ISD::CMP: return "X86ISD::CMP";
7779 case X86ISD::COMI: return "X86ISD::COMI";
7780 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7781 case X86ISD::SETCC: return "X86ISD::SETCC";
7782 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7783 case X86ISD::CMOV: return "X86ISD::CMOV";
7784 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7785 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7786 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7787 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7788 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7789 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7790 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7791 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7792 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7793 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7794 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7795 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7796 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7797 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7798 case X86ISD::FMAX: return "X86ISD::FMAX";
7799 case X86ISD::FMIN: return "X86ISD::FMIN";
7800 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7801 case X86ISD::FRCP: return "X86ISD::FRCP";
7802 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7803 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
7804 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7805 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7806 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7807 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7808 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7809 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7810 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7811 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7812 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7813 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7814 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7815 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7816 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7817 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7818 case X86ISD::VSHL: return "X86ISD::VSHL";
7819 case X86ISD::VSRL: return "X86ISD::VSRL";
7820 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7821 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7822 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7823 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7824 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7825 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7826 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7827 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7828 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7829 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7830 case X86ISD::ADD: return "X86ISD::ADD";
7831 case X86ISD::SUB: return "X86ISD::SUB";
7832 case X86ISD::SMUL: return "X86ISD::SMUL";
7833 case X86ISD::UMUL: return "X86ISD::UMUL";
7834 case X86ISD::INC: return "X86ISD::INC";
7835 case X86ISD::DEC: return "X86ISD::DEC";
7836 case X86ISD::OR: return "X86ISD::OR";
7837 case X86ISD::XOR: return "X86ISD::XOR";
7838 case X86ISD::AND: return "X86ISD::AND";
7839 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7840 case X86ISD::PTEST: return "X86ISD::PTEST";
7841 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7842 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7846 // isLegalAddressingMode - Return true if the addressing mode represented
7847 // by AM is legal for this target, for a load/store of the specified type.
7848 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7849 const Type *Ty) const {
7850 // X86 supports extremely general addressing modes.
7851 CodeModel::Model M = getTargetMachine().getCodeModel();
7853 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7854 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7859 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7861 // If a reference to this global requires an extra load, we can't fold it.
7862 if (isGlobalStubReference(GVFlags))
7865 // If BaseGV requires a register for the PIC base, we cannot also have a
7866 // BaseReg specified.
7867 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7870 // If lower 4G is not available, then we must use rip-relative addressing.
7871 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7881 // These scales always work.
7886 // These scales are formed with basereg+scalereg. Only accept if there is
7891 default: // Other stuff never works.
7899 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7900 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7902 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7903 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7904 if (NumBits1 <= NumBits2)
7909 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7910 if (!VT1.isInteger() || !VT2.isInteger())
7912 unsigned NumBits1 = VT1.getSizeInBits();
7913 unsigned NumBits2 = VT2.getSizeInBits();
7914 if (NumBits1 <= NumBits2)
7919 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7920 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7921 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7924 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7925 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7926 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7929 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7930 // i16 instructions are longer (0x66 prefix) and potentially slower.
7931 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7934 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7935 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7936 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7937 /// are assumed to be legal.
7939 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7941 // Very little shuffling can be done for 64-bit vectors right now.
7942 if (VT.getSizeInBits() == 64)
7943 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
7945 // FIXME: pshufb, blends, shifts.
7946 return (VT.getVectorNumElements() == 2 ||
7947 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7948 isMOVLMask(M, VT) ||
7949 isSHUFPMask(M, VT) ||
7950 isPSHUFDMask(M, VT) ||
7951 isPSHUFHWMask(M, VT) ||
7952 isPSHUFLWMask(M, VT) ||
7953 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7954 isUNPCKLMask(M, VT) ||
7955 isUNPCKHMask(M, VT) ||
7956 isUNPCKL_v_undef_Mask(M, VT) ||
7957 isUNPCKH_v_undef_Mask(M, VT));
7961 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7963 unsigned NumElts = VT.getVectorNumElements();
7964 // FIXME: This collection of masks seems suspect.
7967 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7968 return (isMOVLMask(Mask, VT) ||
7969 isCommutedMOVLMask(Mask, VT, true) ||
7970 isSHUFPMask(Mask, VT) ||
7971 isCommutedSHUFPMask(Mask, VT));
7976 //===----------------------------------------------------------------------===//
7977 // X86 Scheduler Hooks
7978 //===----------------------------------------------------------------------===//
7980 // private utility function
7982 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7983 MachineBasicBlock *MBB,
7991 TargetRegisterClass *RC,
7992 bool invSrc) const {
7993 // For the atomic bitwise operator, we generate
7996 // ld t1 = [bitinstr.addr]
7997 // op t2 = t1, [bitinstr.val]
7999 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8001 // fallthrough -->nextMBB
8002 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8003 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8004 MachineFunction::iterator MBBIter = MBB;
8007 /// First build the CFG
8008 MachineFunction *F = MBB->getParent();
8009 MachineBasicBlock *thisMBB = MBB;
8010 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8011 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8012 F->insert(MBBIter, newMBB);
8013 F->insert(MBBIter, nextMBB);
8015 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8016 nextMBB->splice(nextMBB->begin(), thisMBB,
8017 llvm::next(MachineBasicBlock::iterator(bInstr)),
8019 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8021 // Update thisMBB to fall through to newMBB
8022 thisMBB->addSuccessor(newMBB);
8024 // newMBB jumps to itself and fall through to nextMBB
8025 newMBB->addSuccessor(nextMBB);
8026 newMBB->addSuccessor(newMBB);
8028 // Insert instructions into newMBB based on incoming instruction
8029 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8030 "unexpected number of operands");
8031 DebugLoc dl = bInstr->getDebugLoc();
8032 MachineOperand& destOper = bInstr->getOperand(0);
8033 MachineOperand* argOpers[2 + X86AddrNumOperands];
8034 int numArgs = bInstr->getNumOperands() - 1;
8035 for (int i=0; i < numArgs; ++i)
8036 argOpers[i] = &bInstr->getOperand(i+1);
8038 // x86 address has 4 operands: base, index, scale, and displacement
8039 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8040 int valArgIndx = lastAddrIndx + 1;
8042 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8043 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8044 for (int i=0; i <= lastAddrIndx; ++i)
8045 (*MIB).addOperand(*argOpers[i]);
8047 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8049 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8054 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8055 assert((argOpers[valArgIndx]->isReg() ||
8056 argOpers[valArgIndx]->isImm()) &&
8058 if (argOpers[valArgIndx]->isReg())
8059 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8061 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8063 (*MIB).addOperand(*argOpers[valArgIndx]);
8065 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8068 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8069 for (int i=0; i <= lastAddrIndx; ++i)
8070 (*MIB).addOperand(*argOpers[i]);
8072 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8073 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8074 bInstr->memoperands_end());
8076 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8080 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8082 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8086 // private utility function: 64 bit atomics on 32 bit host.
8088 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8089 MachineBasicBlock *MBB,
8094 bool invSrc) const {
8095 // For the atomic bitwise operator, we generate
8096 // thisMBB (instructions are in pairs, except cmpxchg8b)
8097 // ld t1,t2 = [bitinstr.addr]
8099 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8100 // op t5, t6 <- out1, out2, [bitinstr.val]
8101 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8102 // mov ECX, EBX <- t5, t6
8103 // mov EAX, EDX <- t1, t2
8104 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8105 // mov t3, t4 <- EAX, EDX
8107 // result in out1, out2
8108 // fallthrough -->nextMBB
8110 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8111 const unsigned LoadOpc = X86::MOV32rm;
8112 const unsigned copyOpc = X86::MOV32rr;
8113 const unsigned NotOpc = X86::NOT32r;
8114 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8115 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8116 MachineFunction::iterator MBBIter = MBB;
8119 /// First build the CFG
8120 MachineFunction *F = MBB->getParent();
8121 MachineBasicBlock *thisMBB = MBB;
8122 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8123 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8124 F->insert(MBBIter, newMBB);
8125 F->insert(MBBIter, nextMBB);
8127 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8128 nextMBB->splice(nextMBB->begin(), thisMBB,
8129 llvm::next(MachineBasicBlock::iterator(bInstr)),
8131 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8133 // Update thisMBB to fall through to newMBB
8134 thisMBB->addSuccessor(newMBB);
8136 // newMBB jumps to itself and fall through to nextMBB
8137 newMBB->addSuccessor(nextMBB);
8138 newMBB->addSuccessor(newMBB);
8140 DebugLoc dl = bInstr->getDebugLoc();
8141 // Insert instructions into newMBB based on incoming instruction
8142 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8143 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8144 "unexpected number of operands");
8145 MachineOperand& dest1Oper = bInstr->getOperand(0);
8146 MachineOperand& dest2Oper = bInstr->getOperand(1);
8147 MachineOperand* argOpers[2 + X86AddrNumOperands];
8148 for (int i=0; i < 2 + X86AddrNumOperands; ++i) {
8149 argOpers[i] = &bInstr->getOperand(i+2);
8151 // We use some of the operands multiple times, so conservatively just
8152 // clear any kill flags that might be present.
8153 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8154 argOpers[i]->setIsKill(false);
8157 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8158 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8160 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8161 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8162 for (int i=0; i <= lastAddrIndx; ++i)
8163 (*MIB).addOperand(*argOpers[i]);
8164 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8165 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8166 // add 4 to displacement.
8167 for (int i=0; i <= lastAddrIndx-2; ++i)
8168 (*MIB).addOperand(*argOpers[i]);
8169 MachineOperand newOp3 = *(argOpers[3]);
8171 newOp3.setImm(newOp3.getImm()+4);
8173 newOp3.setOffset(newOp3.getOffset()+4);
8174 (*MIB).addOperand(newOp3);
8175 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8177 // t3/4 are defined later, at the bottom of the loop
8178 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8179 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8180 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8181 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8182 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8183 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8185 // The subsequent operations should be using the destination registers of
8186 //the PHI instructions.
8188 t1 = F->getRegInfo().createVirtualRegister(RC);
8189 t2 = F->getRegInfo().createVirtualRegister(RC);
8190 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8191 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8193 t1 = dest1Oper.getReg();
8194 t2 = dest2Oper.getReg();
8197 int valArgIndx = lastAddrIndx + 1;
8198 assert((argOpers[valArgIndx]->isReg() ||
8199 argOpers[valArgIndx]->isImm()) &&
8201 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8202 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8203 if (argOpers[valArgIndx]->isReg())
8204 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8206 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8207 if (regOpcL != X86::MOV32rr)
8209 (*MIB).addOperand(*argOpers[valArgIndx]);
8210 assert(argOpers[valArgIndx + 1]->isReg() ==
8211 argOpers[valArgIndx]->isReg());
8212 assert(argOpers[valArgIndx + 1]->isImm() ==
8213 argOpers[valArgIndx]->isImm());
8214 if (argOpers[valArgIndx + 1]->isReg())
8215 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8217 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8218 if (regOpcH != X86::MOV32rr)
8220 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8222 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8224 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8227 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8229 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8232 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8233 for (int i=0; i <= lastAddrIndx; ++i)
8234 (*MIB).addOperand(*argOpers[i]);
8236 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8237 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8238 bInstr->memoperands_end());
8240 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8241 MIB.addReg(X86::EAX);
8242 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8243 MIB.addReg(X86::EDX);
8246 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8248 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8252 // private utility function
8254 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8255 MachineBasicBlock *MBB,
8256 unsigned cmovOpc) const {
8257 // For the atomic min/max operator, we generate
8260 // ld t1 = [min/max.addr]
8261 // mov t2 = [min/max.val]
8263 // cmov[cond] t2 = t1
8265 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8267 // fallthrough -->nextMBB
8269 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8270 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8271 MachineFunction::iterator MBBIter = MBB;
8274 /// First build the CFG
8275 MachineFunction *F = MBB->getParent();
8276 MachineBasicBlock *thisMBB = MBB;
8277 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8278 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8279 F->insert(MBBIter, newMBB);
8280 F->insert(MBBIter, nextMBB);
8282 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8283 nextMBB->splice(nextMBB->begin(), thisMBB,
8284 llvm::next(MachineBasicBlock::iterator(mInstr)),
8286 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8288 // Update thisMBB to fall through to newMBB
8289 thisMBB->addSuccessor(newMBB);
8291 // newMBB jumps to newMBB and fall through to nextMBB
8292 newMBB->addSuccessor(nextMBB);
8293 newMBB->addSuccessor(newMBB);
8295 DebugLoc dl = mInstr->getDebugLoc();
8296 // Insert instructions into newMBB based on incoming instruction
8297 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8298 "unexpected number of operands");
8299 MachineOperand& destOper = mInstr->getOperand(0);
8300 MachineOperand* argOpers[2 + X86AddrNumOperands];
8301 int numArgs = mInstr->getNumOperands() - 1;
8302 for (int i=0; i < numArgs; ++i)
8303 argOpers[i] = &mInstr->getOperand(i+1);
8305 // x86 address has 4 operands: base, index, scale, and displacement
8306 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8307 int valArgIndx = lastAddrIndx + 1;
8309 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8310 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8311 for (int i=0; i <= lastAddrIndx; ++i)
8312 (*MIB).addOperand(*argOpers[i]);
8314 // We only support register and immediate values
8315 assert((argOpers[valArgIndx]->isReg() ||
8316 argOpers[valArgIndx]->isImm()) &&
8319 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8320 if (argOpers[valArgIndx]->isReg())
8321 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8323 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8324 (*MIB).addOperand(*argOpers[valArgIndx]);
8326 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8329 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8334 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8335 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8339 // Cmp and exchange if none has modified the memory location
8340 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8341 for (int i=0; i <= lastAddrIndx; ++i)
8342 (*MIB).addOperand(*argOpers[i]);
8344 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8345 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8346 mInstr->memoperands_end());
8348 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8349 MIB.addReg(X86::EAX);
8352 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8354 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
8358 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8359 // all of this code can be replaced with that in the .td file.
8361 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8362 unsigned numArgs, bool memArg) const {
8364 DebugLoc dl = MI->getDebugLoc();
8365 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8369 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8371 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8373 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8375 for (unsigned i = 0; i < numArgs; ++i) {
8376 MachineOperand &Op = MI->getOperand(i+1);
8378 if (!(Op.isReg() && Op.isImplicit()))
8382 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8385 MI->eraseFromParent();
8391 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8393 MachineBasicBlock *MBB) const {
8394 // Emit code to save XMM registers to the stack. The ABI says that the
8395 // number of registers to save is given in %al, so it's theoretically
8396 // possible to do an indirect jump trick to avoid saving all of them,
8397 // however this code takes a simpler approach and just executes all
8398 // of the stores if %al is non-zero. It's less code, and it's probably
8399 // easier on the hardware branch predictor, and stores aren't all that
8400 // expensive anyway.
8402 // Create the new basic blocks. One block contains all the XMM stores,
8403 // and one block is the final destination regardless of whether any
8404 // stores were performed.
8405 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8406 MachineFunction *F = MBB->getParent();
8407 MachineFunction::iterator MBBIter = MBB;
8409 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8410 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8411 F->insert(MBBIter, XMMSaveMBB);
8412 F->insert(MBBIter, EndMBB);
8414 // Transfer the remainder of MBB and its successor edges to EndMBB.
8415 EndMBB->splice(EndMBB->begin(), MBB,
8416 llvm::next(MachineBasicBlock::iterator(MI)),
8418 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
8420 // The original block will now fall through to the XMM save block.
8421 MBB->addSuccessor(XMMSaveMBB);
8422 // The XMMSaveMBB will fall through to the end block.
8423 XMMSaveMBB->addSuccessor(EndMBB);
8425 // Now add the instructions.
8426 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8427 DebugLoc DL = MI->getDebugLoc();
8429 unsigned CountReg = MI->getOperand(0).getReg();
8430 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8431 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8433 if (!Subtarget->isTargetWin64()) {
8434 // If %al is 0, branch around the XMM save block.
8435 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8436 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8437 MBB->addSuccessor(EndMBB);
8440 // In the XMM save block, save all the XMM argument registers.
8441 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8442 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8443 MachineMemOperand *MMO =
8444 F->getMachineMemOperand(
8445 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8446 MachineMemOperand::MOStore, Offset,
8447 /*Size=*/16, /*Align=*/16);
8448 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8449 .addFrameIndex(RegSaveFrameIndex)
8450 .addImm(/*Scale=*/1)
8451 .addReg(/*IndexReg=*/0)
8452 .addImm(/*Disp=*/Offset)
8453 .addReg(/*Segment=*/0)
8454 .addReg(MI->getOperand(i).getReg())
8455 .addMemOperand(MMO);
8458 MI->eraseFromParent(); // The pseudo instruction is gone now.
8464 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8465 MachineBasicBlock *BB) const {
8466 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8467 DebugLoc DL = MI->getDebugLoc();
8469 // To "insert" a SELECT_CC instruction, we actually have to insert the
8470 // diamond control-flow pattern. The incoming instruction knows the
8471 // destination vreg to set, the condition code register to branch on, the
8472 // true/false values to select between, and a branch opcode to use.
8473 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8474 MachineFunction::iterator It = BB;
8480 // cmpTY ccX, r1, r2
8482 // fallthrough --> copy0MBB
8483 MachineBasicBlock *thisMBB = BB;
8484 MachineFunction *F = BB->getParent();
8485 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8486 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8487 F->insert(It, copy0MBB);
8488 F->insert(It, sinkMBB);
8490 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8491 // live into the sink and copy blocks.
8492 const MachineFunction *MF = BB->getParent();
8493 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8494 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8496 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
8497 const MachineOperand &MO = MI->getOperand(I);
8498 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
8499 unsigned Reg = MO.getReg();
8500 if (Reg != X86::EFLAGS) continue;
8501 copy0MBB->addLiveIn(Reg);
8502 sinkMBB->addLiveIn(Reg);
8505 // Transfer the remainder of BB and its successor edges to sinkMBB.
8506 sinkMBB->splice(sinkMBB->begin(), BB,
8507 llvm::next(MachineBasicBlock::iterator(MI)),
8509 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
8511 // Add the true and fallthrough blocks as its successors.
8512 BB->addSuccessor(copy0MBB);
8513 BB->addSuccessor(sinkMBB);
8515 // Create the conditional branch instruction.
8517 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8518 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8521 // %FalseValue = ...
8522 // # fallthrough to sinkMBB
8523 copy0MBB->addSuccessor(sinkMBB);
8526 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8528 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
8529 TII->get(X86::PHI), MI->getOperand(0).getReg())
8530 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8531 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8533 MI->eraseFromParent(); // The pseudo instruction is gone now.
8538 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8539 MachineBasicBlock *BB) const {
8540 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8541 DebugLoc DL = MI->getDebugLoc();
8543 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8544 // non-trivial part is impdef of ESP.
8545 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8548 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
8549 .addExternalSymbol("_alloca")
8550 .addReg(X86::EAX, RegState::Implicit)
8551 .addReg(X86::ESP, RegState::Implicit)
8552 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8553 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8555 MI->eraseFromParent(); // The pseudo instruction is gone now.
8560 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8561 MachineBasicBlock *BB) const {
8562 // This is pretty easy. We're taking the value that we received from
8563 // our load from the relocation, sticking it in either RDI (x86-64)
8564 // or EAX and doing an indirect call. The return value will then
8565 // be in the normal return register.
8566 const X86InstrInfo *TII
8567 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8568 DebugLoc DL = MI->getDebugLoc();
8569 MachineFunction *F = BB->getParent();
8571 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8573 if (Subtarget->is64Bit()) {
8574 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8575 TII->get(X86::MOV64rm), X86::RDI)
8577 .addImm(0).addReg(0)
8578 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8579 MI->getOperand(3).getTargetFlags())
8581 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
8582 addDirectMem(MIB, X86::RDI).addReg(0);
8583 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8584 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8585 TII->get(X86::MOV32rm), X86::EAX)
8587 .addImm(0).addReg(0)
8588 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8589 MI->getOperand(3).getTargetFlags())
8591 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8592 addDirectMem(MIB, X86::EAX).addReg(0);
8594 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8595 TII->get(X86::MOV32rm), X86::EAX)
8596 .addReg(TII->getGlobalBaseReg(F))
8597 .addImm(0).addReg(0)
8598 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8599 MI->getOperand(3).getTargetFlags())
8601 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8602 addDirectMem(MIB, X86::EAX).addReg(0);
8605 MI->eraseFromParent(); // The pseudo instruction is gone now.
8610 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8611 MachineBasicBlock *BB) const {
8612 switch (MI->getOpcode()) {
8613 default: assert(false && "Unexpected instr type to insert");
8614 case X86::MINGW_ALLOCA:
8615 return EmitLoweredMingwAlloca(MI, BB);
8616 case X86::TLSCall_32:
8617 case X86::TLSCall_64:
8618 return EmitLoweredTLSCall(MI, BB);
8620 case X86::CMOV_V1I64:
8621 case X86::CMOV_FR32:
8622 case X86::CMOV_FR64:
8623 case X86::CMOV_V4F32:
8624 case X86::CMOV_V2F64:
8625 case X86::CMOV_V2I64:
8626 case X86::CMOV_GR16:
8627 case X86::CMOV_GR32:
8628 case X86::CMOV_RFP32:
8629 case X86::CMOV_RFP64:
8630 case X86::CMOV_RFP80:
8631 return EmitLoweredSelect(MI, BB);
8633 case X86::FP32_TO_INT16_IN_MEM:
8634 case X86::FP32_TO_INT32_IN_MEM:
8635 case X86::FP32_TO_INT64_IN_MEM:
8636 case X86::FP64_TO_INT16_IN_MEM:
8637 case X86::FP64_TO_INT32_IN_MEM:
8638 case X86::FP64_TO_INT64_IN_MEM:
8639 case X86::FP80_TO_INT16_IN_MEM:
8640 case X86::FP80_TO_INT32_IN_MEM:
8641 case X86::FP80_TO_INT64_IN_MEM: {
8642 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8643 DebugLoc DL = MI->getDebugLoc();
8645 // Change the floating point control register to use "round towards zero"
8646 // mode when truncating to an integer value.
8647 MachineFunction *F = BB->getParent();
8648 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8649 addFrameReference(BuildMI(*BB, MI, DL,
8650 TII->get(X86::FNSTCW16m)), CWFrameIdx);
8652 // Load the old value of the high byte of the control word...
8654 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8655 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
8658 // Set the high part to be round to zero...
8659 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8662 // Reload the modified control word now...
8663 addFrameReference(BuildMI(*BB, MI, DL,
8664 TII->get(X86::FLDCW16m)), CWFrameIdx);
8666 // Restore the memory image of control word to original value
8667 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8670 // Get the X86 opcode to use.
8672 switch (MI->getOpcode()) {
8673 default: llvm_unreachable("illegal opcode!");
8674 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8675 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8676 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8677 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8678 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8679 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8680 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8681 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8682 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8686 MachineOperand &Op = MI->getOperand(0);
8688 AM.BaseType = X86AddressMode::RegBase;
8689 AM.Base.Reg = Op.getReg();
8691 AM.BaseType = X86AddressMode::FrameIndexBase;
8692 AM.Base.FrameIndex = Op.getIndex();
8694 Op = MI->getOperand(1);
8696 AM.Scale = Op.getImm();
8697 Op = MI->getOperand(2);
8699 AM.IndexReg = Op.getImm();
8700 Op = MI->getOperand(3);
8701 if (Op.isGlobal()) {
8702 AM.GV = Op.getGlobal();
8704 AM.Disp = Op.getImm();
8706 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
8707 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8709 // Reload the original control word now.
8710 addFrameReference(BuildMI(*BB, MI, DL,
8711 TII->get(X86::FLDCW16m)), CWFrameIdx);
8713 MI->eraseFromParent(); // The pseudo instruction is gone now.
8716 // String/text processing lowering.
8717 case X86::PCMPISTRM128REG:
8718 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8719 case X86::PCMPISTRM128MEM:
8720 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8721 case X86::PCMPESTRM128REG:
8722 return EmitPCMP(MI, BB, 5, false /* in mem */);
8723 case X86::PCMPESTRM128MEM:
8724 return EmitPCMP(MI, BB, 5, true /* in mem */);
8727 case X86::ATOMAND32:
8728 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8729 X86::AND32ri, X86::MOV32rm,
8730 X86::LCMPXCHG32, X86::MOV32rr,
8731 X86::NOT32r, X86::EAX,
8732 X86::GR32RegisterClass);
8734 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8735 X86::OR32ri, X86::MOV32rm,
8736 X86::LCMPXCHG32, X86::MOV32rr,
8737 X86::NOT32r, X86::EAX,
8738 X86::GR32RegisterClass);
8739 case X86::ATOMXOR32:
8740 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8741 X86::XOR32ri, X86::MOV32rm,
8742 X86::LCMPXCHG32, X86::MOV32rr,
8743 X86::NOT32r, X86::EAX,
8744 X86::GR32RegisterClass);
8745 case X86::ATOMNAND32:
8746 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8747 X86::AND32ri, X86::MOV32rm,
8748 X86::LCMPXCHG32, X86::MOV32rr,
8749 X86::NOT32r, X86::EAX,
8750 X86::GR32RegisterClass, true);
8751 case X86::ATOMMIN32:
8752 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8753 case X86::ATOMMAX32:
8754 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8755 case X86::ATOMUMIN32:
8756 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8757 case X86::ATOMUMAX32:
8758 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8760 case X86::ATOMAND16:
8761 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8762 X86::AND16ri, X86::MOV16rm,
8763 X86::LCMPXCHG16, X86::MOV16rr,
8764 X86::NOT16r, X86::AX,
8765 X86::GR16RegisterClass);
8767 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8768 X86::OR16ri, X86::MOV16rm,
8769 X86::LCMPXCHG16, X86::MOV16rr,
8770 X86::NOT16r, X86::AX,
8771 X86::GR16RegisterClass);
8772 case X86::ATOMXOR16:
8773 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8774 X86::XOR16ri, X86::MOV16rm,
8775 X86::LCMPXCHG16, X86::MOV16rr,
8776 X86::NOT16r, X86::AX,
8777 X86::GR16RegisterClass);
8778 case X86::ATOMNAND16:
8779 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8780 X86::AND16ri, X86::MOV16rm,
8781 X86::LCMPXCHG16, X86::MOV16rr,
8782 X86::NOT16r, X86::AX,
8783 X86::GR16RegisterClass, true);
8784 case X86::ATOMMIN16:
8785 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8786 case X86::ATOMMAX16:
8787 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8788 case X86::ATOMUMIN16:
8789 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8790 case X86::ATOMUMAX16:
8791 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8794 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8795 X86::AND8ri, X86::MOV8rm,
8796 X86::LCMPXCHG8, X86::MOV8rr,
8797 X86::NOT8r, X86::AL,
8798 X86::GR8RegisterClass);
8800 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8801 X86::OR8ri, X86::MOV8rm,
8802 X86::LCMPXCHG8, X86::MOV8rr,
8803 X86::NOT8r, X86::AL,
8804 X86::GR8RegisterClass);
8806 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8807 X86::XOR8ri, X86::MOV8rm,
8808 X86::LCMPXCHG8, X86::MOV8rr,
8809 X86::NOT8r, X86::AL,
8810 X86::GR8RegisterClass);
8811 case X86::ATOMNAND8:
8812 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8813 X86::AND8ri, X86::MOV8rm,
8814 X86::LCMPXCHG8, X86::MOV8rr,
8815 X86::NOT8r, X86::AL,
8816 X86::GR8RegisterClass, true);
8817 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8818 // This group is for 64-bit host.
8819 case X86::ATOMAND64:
8820 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8821 X86::AND64ri32, X86::MOV64rm,
8822 X86::LCMPXCHG64, X86::MOV64rr,
8823 X86::NOT64r, X86::RAX,
8824 X86::GR64RegisterClass);
8826 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8827 X86::OR64ri32, X86::MOV64rm,
8828 X86::LCMPXCHG64, X86::MOV64rr,
8829 X86::NOT64r, X86::RAX,
8830 X86::GR64RegisterClass);
8831 case X86::ATOMXOR64:
8832 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8833 X86::XOR64ri32, X86::MOV64rm,
8834 X86::LCMPXCHG64, X86::MOV64rr,
8835 X86::NOT64r, X86::RAX,
8836 X86::GR64RegisterClass);
8837 case X86::ATOMNAND64:
8838 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8839 X86::AND64ri32, X86::MOV64rm,
8840 X86::LCMPXCHG64, X86::MOV64rr,
8841 X86::NOT64r, X86::RAX,
8842 X86::GR64RegisterClass, true);
8843 case X86::ATOMMIN64:
8844 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8845 case X86::ATOMMAX64:
8846 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8847 case X86::ATOMUMIN64:
8848 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8849 case X86::ATOMUMAX64:
8850 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8852 // This group does 64-bit operations on a 32-bit host.
8853 case X86::ATOMAND6432:
8854 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8855 X86::AND32rr, X86::AND32rr,
8856 X86::AND32ri, X86::AND32ri,
8858 case X86::ATOMOR6432:
8859 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8860 X86::OR32rr, X86::OR32rr,
8861 X86::OR32ri, X86::OR32ri,
8863 case X86::ATOMXOR6432:
8864 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8865 X86::XOR32rr, X86::XOR32rr,
8866 X86::XOR32ri, X86::XOR32ri,
8868 case X86::ATOMNAND6432:
8869 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8870 X86::AND32rr, X86::AND32rr,
8871 X86::AND32ri, X86::AND32ri,
8873 case X86::ATOMADD6432:
8874 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8875 X86::ADD32rr, X86::ADC32rr,
8876 X86::ADD32ri, X86::ADC32ri,
8878 case X86::ATOMSUB6432:
8879 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8880 X86::SUB32rr, X86::SBB32rr,
8881 X86::SUB32ri, X86::SBB32ri,
8883 case X86::ATOMSWAP6432:
8884 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8885 X86::MOV32rr, X86::MOV32rr,
8886 X86::MOV32ri, X86::MOV32ri,
8888 case X86::VASTART_SAVE_XMM_REGS:
8889 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8893 //===----------------------------------------------------------------------===//
8894 // X86 Optimization Hooks
8895 //===----------------------------------------------------------------------===//
8897 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8901 const SelectionDAG &DAG,
8902 unsigned Depth) const {
8903 unsigned Opc = Op.getOpcode();
8904 assert((Opc >= ISD::BUILTIN_OP_END ||
8905 Opc == ISD::INTRINSIC_WO_CHAIN ||
8906 Opc == ISD::INTRINSIC_W_CHAIN ||
8907 Opc == ISD::INTRINSIC_VOID) &&
8908 "Should use MaskedValueIsZero if you don't know whether Op"
8909 " is a target node!");
8911 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8923 // These nodes' second result is a boolean.
8924 if (Op.getResNo() == 0)
8928 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8929 Mask.getBitWidth() - 1);
8934 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8935 /// node is a GlobalAddress + offset.
8936 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8937 const GlobalValue* &GA,
8938 int64_t &Offset) const {
8939 if (N->getOpcode() == X86ISD::Wrapper) {
8940 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8941 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8942 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8946 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8949 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8950 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8951 /// if the load addresses are consecutive, non-overlapping, and in the right
8953 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8954 const TargetLowering &TLI) {
8955 DebugLoc dl = N->getDebugLoc();
8956 EVT VT = N->getValueType(0);
8957 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8959 if (VT.getSizeInBits() != 128)
8962 SmallVector<SDValue, 16> Elts;
8963 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8964 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8966 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8969 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8970 /// and convert it from being a bunch of shuffles and extracts to a simple
8971 /// store and scalar loads to extract the elements.
8972 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8973 const TargetLowering &TLI) {
8974 SDValue InputVector = N->getOperand(0);
8976 // Only operate on vectors of 4 elements, where the alternative shuffling
8977 // gets to be more expensive.
8978 if (InputVector.getValueType() != MVT::v4i32)
8981 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8982 // single use which is a sign-extend or zero-extend, and all elements are
8984 SmallVector<SDNode *, 4> Uses;
8985 unsigned ExtractedElements = 0;
8986 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8987 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8988 if (UI.getUse().getResNo() != InputVector.getResNo())
8991 SDNode *Extract = *UI;
8992 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8995 if (Extract->getValueType(0) != MVT::i32)
8997 if (!Extract->hasOneUse())
8999 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9000 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9002 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9005 // Record which element was extracted.
9006 ExtractedElements |=
9007 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9009 Uses.push_back(Extract);
9012 // If not all the elements were used, this may not be worthwhile.
9013 if (ExtractedElements != 15)
9016 // Ok, we've now decided to do the transformation.
9017 DebugLoc dl = InputVector.getDebugLoc();
9019 // Store the value to a temporary stack slot.
9020 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9021 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
9024 // Replace each use (extract) with a load of the appropriate element.
9025 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9026 UE = Uses.end(); UI != UE; ++UI) {
9027 SDNode *Extract = *UI;
9029 // Compute the element's address.
9030 SDValue Idx = Extract->getOperand(1);
9032 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9033 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9034 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9036 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
9039 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
9040 NULL, 0, false, false, 0);
9042 // Replace the exact with the load.
9043 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9046 // The replacement was made in place; don't return anything.
9050 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9051 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9052 const X86Subtarget *Subtarget) {
9053 DebugLoc DL = N->getDebugLoc();
9054 SDValue Cond = N->getOperand(0);
9055 // Get the LHS/RHS of the select.
9056 SDValue LHS = N->getOperand(1);
9057 SDValue RHS = N->getOperand(2);
9059 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9060 // instructions match the semantics of the common C idiom x<y?x:y but not
9061 // x<=y?x:y, because of how they handle negative zero (which can be
9062 // ignored in unsafe-math mode).
9063 if (Subtarget->hasSSE2() &&
9064 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9065 Cond.getOpcode() == ISD::SETCC) {
9066 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9068 unsigned Opcode = 0;
9069 // Check for x CC y ? x : y.
9070 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9071 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9075 // Converting this to a min would handle NaNs incorrectly, and swapping
9076 // the operands would cause it to handle comparisons between positive
9077 // and negative zero incorrectly.
9078 if (!FiniteOnlyFPMath() &&
9079 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9080 if (!UnsafeFPMath &&
9081 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9083 std::swap(LHS, RHS);
9085 Opcode = X86ISD::FMIN;
9088 // Converting this to a min would handle comparisons between positive
9089 // and negative zero incorrectly.
9090 if (!UnsafeFPMath &&
9091 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9093 Opcode = X86ISD::FMIN;
9096 // Converting this to a min would handle both negative zeros and NaNs
9097 // incorrectly, but we can swap the operands to fix both.
9098 std::swap(LHS, RHS);
9102 Opcode = X86ISD::FMIN;
9106 // Converting this to a max would handle comparisons between positive
9107 // and negative zero incorrectly.
9108 if (!UnsafeFPMath &&
9109 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9111 Opcode = X86ISD::FMAX;
9114 // Converting this to a max would handle NaNs incorrectly, and swapping
9115 // the operands would cause it to handle comparisons between positive
9116 // and negative zero incorrectly.
9117 if (!FiniteOnlyFPMath() &&
9118 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9119 if (!UnsafeFPMath &&
9120 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9122 std::swap(LHS, RHS);
9124 Opcode = X86ISD::FMAX;
9127 // Converting this to a max would handle both negative zeros and NaNs
9128 // incorrectly, but we can swap the operands to fix both.
9129 std::swap(LHS, RHS);
9133 Opcode = X86ISD::FMAX;
9136 // Check for x CC y ? y : x -- a min/max with reversed arms.
9137 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9138 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9142 // Converting this to a min would handle comparisons between positive
9143 // and negative zero incorrectly, and swapping the operands would
9144 // cause it to handle NaNs incorrectly.
9145 if (!UnsafeFPMath &&
9146 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9147 if (!FiniteOnlyFPMath() &&
9148 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9150 std::swap(LHS, RHS);
9152 Opcode = X86ISD::FMIN;
9155 // Converting this to a min would handle NaNs incorrectly.
9156 if (!UnsafeFPMath &&
9157 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9159 Opcode = X86ISD::FMIN;
9162 // Converting this to a min would handle both negative zeros and NaNs
9163 // incorrectly, but we can swap the operands to fix both.
9164 std::swap(LHS, RHS);
9168 Opcode = X86ISD::FMIN;
9172 // Converting this to a max would handle NaNs incorrectly.
9173 if (!FiniteOnlyFPMath() &&
9174 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9176 Opcode = X86ISD::FMAX;
9179 // Converting this to a max would handle comparisons between positive
9180 // and negative zero incorrectly, and swapping the operands would
9181 // cause it to handle NaNs incorrectly.
9182 if (!UnsafeFPMath &&
9183 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9184 if (!FiniteOnlyFPMath() &&
9185 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9187 std::swap(LHS, RHS);
9189 Opcode = X86ISD::FMAX;
9192 // Converting this to a max would handle both negative zeros and NaNs
9193 // incorrectly, but we can swap the operands to fix both.
9194 std::swap(LHS, RHS);
9198 Opcode = X86ISD::FMAX;
9204 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9207 // If this is a select between two integer constants, try to do some
9209 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9210 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9211 // Don't do this for crazy integer types.
9212 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9213 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9214 // so that TrueC (the true value) is larger than FalseC.
9215 bool NeedsCondInvert = false;
9217 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9218 // Efficiently invertible.
9219 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9220 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9221 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9222 NeedsCondInvert = true;
9223 std::swap(TrueC, FalseC);
9226 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9227 if (FalseC->getAPIntValue() == 0 &&
9228 TrueC->getAPIntValue().isPowerOf2()) {
9229 if (NeedsCondInvert) // Invert the condition if needed.
9230 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9231 DAG.getConstant(1, Cond.getValueType()));
9233 // Zero extend the condition if needed.
9234 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9236 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9237 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9238 DAG.getConstant(ShAmt, MVT::i8));
9241 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9242 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9243 if (NeedsCondInvert) // Invert the condition if needed.
9244 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9245 DAG.getConstant(1, Cond.getValueType()));
9247 // Zero extend the condition if needed.
9248 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9249 FalseC->getValueType(0), Cond);
9250 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9251 SDValue(FalseC, 0));
9254 // Optimize cases that will turn into an LEA instruction. This requires
9255 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9256 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9257 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9258 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9260 bool isFastMultiplier = false;
9262 switch ((unsigned char)Diff) {
9264 case 1: // result = add base, cond
9265 case 2: // result = lea base( , cond*2)
9266 case 3: // result = lea base(cond, cond*2)
9267 case 4: // result = lea base( , cond*4)
9268 case 5: // result = lea base(cond, cond*4)
9269 case 8: // result = lea base( , cond*8)
9270 case 9: // result = lea base(cond, cond*8)
9271 isFastMultiplier = true;
9276 if (isFastMultiplier) {
9277 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9278 if (NeedsCondInvert) // Invert the condition if needed.
9279 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9280 DAG.getConstant(1, Cond.getValueType()));
9282 // Zero extend the condition if needed.
9283 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9285 // Scale the condition by the difference.
9287 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9288 DAG.getConstant(Diff, Cond.getValueType()));
9290 // Add the base if non-zero.
9291 if (FalseC->getAPIntValue() != 0)
9292 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9293 SDValue(FalseC, 0));
9303 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9304 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9305 TargetLowering::DAGCombinerInfo &DCI) {
9306 DebugLoc DL = N->getDebugLoc();
9308 // If the flag operand isn't dead, don't touch this CMOV.
9309 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9312 // If this is a select between two integer constants, try to do some
9313 // optimizations. Note that the operands are ordered the opposite of SELECT
9315 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9316 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9317 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9318 // larger than FalseC (the false value).
9319 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9321 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9322 CC = X86::GetOppositeBranchCondition(CC);
9323 std::swap(TrueC, FalseC);
9326 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9327 // This is efficient for any integer data type (including i8/i16) and
9329 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9330 SDValue Cond = N->getOperand(3);
9331 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9332 DAG.getConstant(CC, MVT::i8), Cond);
9334 // Zero extend the condition if needed.
9335 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9337 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9338 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9339 DAG.getConstant(ShAmt, MVT::i8));
9340 if (N->getNumValues() == 2) // Dead flag value?
9341 return DCI.CombineTo(N, Cond, SDValue());
9345 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9346 // for any integer data type, including i8/i16.
9347 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9348 SDValue Cond = N->getOperand(3);
9349 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9350 DAG.getConstant(CC, MVT::i8), Cond);
9352 // Zero extend the condition if needed.
9353 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9354 FalseC->getValueType(0), Cond);
9355 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9356 SDValue(FalseC, 0));
9358 if (N->getNumValues() == 2) // Dead flag value?
9359 return DCI.CombineTo(N, Cond, SDValue());
9363 // Optimize cases that will turn into an LEA instruction. This requires
9364 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9365 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9366 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9367 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9369 bool isFastMultiplier = false;
9371 switch ((unsigned char)Diff) {
9373 case 1: // result = add base, cond
9374 case 2: // result = lea base( , cond*2)
9375 case 3: // result = lea base(cond, cond*2)
9376 case 4: // result = lea base( , cond*4)
9377 case 5: // result = lea base(cond, cond*4)
9378 case 8: // result = lea base( , cond*8)
9379 case 9: // result = lea base(cond, cond*8)
9380 isFastMultiplier = true;
9385 if (isFastMultiplier) {
9386 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9387 SDValue Cond = N->getOperand(3);
9388 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9389 DAG.getConstant(CC, MVT::i8), Cond);
9390 // Zero extend the condition if needed.
9391 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9393 // Scale the condition by the difference.
9395 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9396 DAG.getConstant(Diff, Cond.getValueType()));
9398 // Add the base if non-zero.
9399 if (FalseC->getAPIntValue() != 0)
9400 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9401 SDValue(FalseC, 0));
9402 if (N->getNumValues() == 2) // Dead flag value?
9403 return DCI.CombineTo(N, Cond, SDValue());
9413 /// PerformMulCombine - Optimize a single multiply with constant into two
9414 /// in order to implement it with two cheaper instructions, e.g.
9415 /// LEA + SHL, LEA + LEA.
9416 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9417 TargetLowering::DAGCombinerInfo &DCI) {
9418 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9421 EVT VT = N->getValueType(0);
9425 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9428 uint64_t MulAmt = C->getZExtValue();
9429 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9432 uint64_t MulAmt1 = 0;
9433 uint64_t MulAmt2 = 0;
9434 if ((MulAmt % 9) == 0) {
9436 MulAmt2 = MulAmt / 9;
9437 } else if ((MulAmt % 5) == 0) {
9439 MulAmt2 = MulAmt / 5;
9440 } else if ((MulAmt % 3) == 0) {
9442 MulAmt2 = MulAmt / 3;
9445 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9446 DebugLoc DL = N->getDebugLoc();
9448 if (isPowerOf2_64(MulAmt2) &&
9449 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9450 // If second multiplifer is pow2, issue it first. We want the multiply by
9451 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9453 std::swap(MulAmt1, MulAmt2);
9456 if (isPowerOf2_64(MulAmt1))
9457 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9458 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9460 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9461 DAG.getConstant(MulAmt1, VT));
9463 if (isPowerOf2_64(MulAmt2))
9464 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9465 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9467 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9468 DAG.getConstant(MulAmt2, VT));
9470 // Do not add new nodes to DAG combiner worklist.
9471 DCI.CombineTo(N, NewMul, false);
9476 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9477 SDValue N0 = N->getOperand(0);
9478 SDValue N1 = N->getOperand(1);
9479 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9480 EVT VT = N0.getValueType();
9482 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9483 // since the result of setcc_c is all zero's or all ones.
9484 if (N1C && N0.getOpcode() == ISD::AND &&
9485 N0.getOperand(1).getOpcode() == ISD::Constant) {
9486 SDValue N00 = N0.getOperand(0);
9487 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9488 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9489 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9490 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9491 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9492 APInt ShAmt = N1C->getAPIntValue();
9493 Mask = Mask.shl(ShAmt);
9495 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9496 N00, DAG.getConstant(Mask, VT));
9503 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9505 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9506 const X86Subtarget *Subtarget) {
9507 EVT VT = N->getValueType(0);
9508 if (!VT.isVector() && VT.isInteger() &&
9509 N->getOpcode() == ISD::SHL)
9510 return PerformSHLCombine(N, DAG);
9512 // On X86 with SSE2 support, we can transform this to a vector shift if
9513 // all elements are shifted by the same amount. We can't do this in legalize
9514 // because the a constant vector is typically transformed to a constant pool
9515 // so we have no knowledge of the shift amount.
9516 if (!Subtarget->hasSSE2())
9519 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9522 SDValue ShAmtOp = N->getOperand(1);
9523 EVT EltVT = VT.getVectorElementType();
9524 DebugLoc DL = N->getDebugLoc();
9525 SDValue BaseShAmt = SDValue();
9526 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9527 unsigned NumElts = VT.getVectorNumElements();
9529 for (; i != NumElts; ++i) {
9530 SDValue Arg = ShAmtOp.getOperand(i);
9531 if (Arg.getOpcode() == ISD::UNDEF) continue;
9535 for (; i != NumElts; ++i) {
9536 SDValue Arg = ShAmtOp.getOperand(i);
9537 if (Arg.getOpcode() == ISD::UNDEF) continue;
9538 if (Arg != BaseShAmt) {
9542 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9543 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9544 SDValue InVec = ShAmtOp.getOperand(0);
9545 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9546 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9548 for (; i != NumElts; ++i) {
9549 SDValue Arg = InVec.getOperand(i);
9550 if (Arg.getOpcode() == ISD::UNDEF) continue;
9554 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9555 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9556 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9557 if (C->getZExtValue() == SplatIdx)
9558 BaseShAmt = InVec.getOperand(1);
9561 if (BaseShAmt.getNode() == 0)
9562 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9563 DAG.getIntPtrConstant(0));
9567 // The shift amount is an i32.
9568 if (EltVT.bitsGT(MVT::i32))
9569 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9570 else if (EltVT.bitsLT(MVT::i32))
9571 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9573 // The shift amount is identical so we can do a vector shift.
9574 SDValue ValOp = N->getOperand(0);
9575 switch (N->getOpcode()) {
9577 llvm_unreachable("Unknown shift opcode!");
9580 if (VT == MVT::v2i64)
9581 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9582 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9584 if (VT == MVT::v4i32)
9585 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9586 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9588 if (VT == MVT::v8i16)
9589 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9590 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9594 if (VT == MVT::v4i32)
9595 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9596 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9598 if (VT == MVT::v8i16)
9599 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9600 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9604 if (VT == MVT::v2i64)
9605 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9606 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9608 if (VT == MVT::v4i32)
9609 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9610 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9612 if (VT == MVT::v8i16)
9613 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9614 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9621 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9622 TargetLowering::DAGCombinerInfo &DCI,
9623 const X86Subtarget *Subtarget) {
9624 if (DCI.isBeforeLegalizeOps())
9627 EVT VT = N->getValueType(0);
9628 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9631 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9632 SDValue N0 = N->getOperand(0);
9633 SDValue N1 = N->getOperand(1);
9634 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9636 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9638 if (!N0.hasOneUse() || !N1.hasOneUse())
9641 SDValue ShAmt0 = N0.getOperand(1);
9642 if (ShAmt0.getValueType() != MVT::i8)
9644 SDValue ShAmt1 = N1.getOperand(1);
9645 if (ShAmt1.getValueType() != MVT::i8)
9647 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9648 ShAmt0 = ShAmt0.getOperand(0);
9649 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9650 ShAmt1 = ShAmt1.getOperand(0);
9652 DebugLoc DL = N->getDebugLoc();
9653 unsigned Opc = X86ISD::SHLD;
9654 SDValue Op0 = N0.getOperand(0);
9655 SDValue Op1 = N1.getOperand(0);
9656 if (ShAmt0.getOpcode() == ISD::SUB) {
9658 std::swap(Op0, Op1);
9659 std::swap(ShAmt0, ShAmt1);
9662 unsigned Bits = VT.getSizeInBits();
9663 if (ShAmt1.getOpcode() == ISD::SUB) {
9664 SDValue Sum = ShAmt1.getOperand(0);
9665 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9666 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
9667 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
9668 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
9669 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
9670 return DAG.getNode(Opc, DL, VT,
9672 DAG.getNode(ISD::TRUNCATE, DL,
9675 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9676 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9678 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9679 return DAG.getNode(Opc, DL, VT,
9680 N0.getOperand(0), N1.getOperand(0),
9681 DAG.getNode(ISD::TRUNCATE, DL,
9688 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9689 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9690 const X86Subtarget *Subtarget) {
9691 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9692 // the FP state in cases where an emms may be missing.
9693 // A preferable solution to the general problem is to figure out the right
9694 // places to insert EMMS. This qualifies as a quick hack.
9696 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9697 StoreSDNode *St = cast<StoreSDNode>(N);
9698 EVT VT = St->getValue().getValueType();
9699 if (VT.getSizeInBits() != 64)
9702 const Function *F = DAG.getMachineFunction().getFunction();
9703 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9704 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9705 && Subtarget->hasSSE2();
9706 if ((VT.isVector() ||
9707 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9708 isa<LoadSDNode>(St->getValue()) &&
9709 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9710 St->getChain().hasOneUse() && !St->isVolatile()) {
9711 SDNode* LdVal = St->getValue().getNode();
9713 int TokenFactorIndex = -1;
9714 SmallVector<SDValue, 8> Ops;
9715 SDNode* ChainVal = St->getChain().getNode();
9716 // Must be a store of a load. We currently handle two cases: the load
9717 // is a direct child, and it's under an intervening TokenFactor. It is
9718 // possible to dig deeper under nested TokenFactors.
9719 if (ChainVal == LdVal)
9720 Ld = cast<LoadSDNode>(St->getChain());
9721 else if (St->getValue().hasOneUse() &&
9722 ChainVal->getOpcode() == ISD::TokenFactor) {
9723 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9724 if (ChainVal->getOperand(i).getNode() == LdVal) {
9725 TokenFactorIndex = i;
9726 Ld = cast<LoadSDNode>(St->getValue());
9728 Ops.push_back(ChainVal->getOperand(i));
9732 if (!Ld || !ISD::isNormalLoad(Ld))
9735 // If this is not the MMX case, i.e. we are just turning i64 load/store
9736 // into f64 load/store, avoid the transformation if there are multiple
9737 // uses of the loaded value.
9738 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9741 DebugLoc LdDL = Ld->getDebugLoc();
9742 DebugLoc StDL = N->getDebugLoc();
9743 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9744 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9746 if (Subtarget->is64Bit() || F64IsLegal) {
9747 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9748 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9749 Ld->getBasePtr(), Ld->getSrcValue(),
9750 Ld->getSrcValueOffset(), Ld->isVolatile(),
9751 Ld->isNonTemporal(), Ld->getAlignment());
9752 SDValue NewChain = NewLd.getValue(1);
9753 if (TokenFactorIndex != -1) {
9754 Ops.push_back(NewChain);
9755 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9758 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9759 St->getSrcValue(), St->getSrcValueOffset(),
9760 St->isVolatile(), St->isNonTemporal(),
9761 St->getAlignment());
9764 // Otherwise, lower to two pairs of 32-bit loads / stores.
9765 SDValue LoAddr = Ld->getBasePtr();
9766 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9767 DAG.getConstant(4, MVT::i32));
9769 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9770 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9771 Ld->isVolatile(), Ld->isNonTemporal(),
9772 Ld->getAlignment());
9773 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9774 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9775 Ld->isVolatile(), Ld->isNonTemporal(),
9776 MinAlign(Ld->getAlignment(), 4));
9778 SDValue NewChain = LoLd.getValue(1);
9779 if (TokenFactorIndex != -1) {
9780 Ops.push_back(LoLd);
9781 Ops.push_back(HiLd);
9782 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9786 LoAddr = St->getBasePtr();
9787 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9788 DAG.getConstant(4, MVT::i32));
9790 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9791 St->getSrcValue(), St->getSrcValueOffset(),
9792 St->isVolatile(), St->isNonTemporal(),
9793 St->getAlignment());
9794 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9796 St->getSrcValueOffset() + 4,
9798 St->isNonTemporal(),
9799 MinAlign(St->getAlignment(), 4));
9800 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9805 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9806 /// X86ISD::FXOR nodes.
9807 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9808 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9809 // F[X]OR(0.0, x) -> x
9810 // F[X]OR(x, 0.0) -> x
9811 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9812 if (C->getValueAPF().isPosZero())
9813 return N->getOperand(1);
9814 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9815 if (C->getValueAPF().isPosZero())
9816 return N->getOperand(0);
9820 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9821 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9822 // FAND(0.0, x) -> 0.0
9823 // FAND(x, 0.0) -> 0.0
9824 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9825 if (C->getValueAPF().isPosZero())
9826 return N->getOperand(0);
9827 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9828 if (C->getValueAPF().isPosZero())
9829 return N->getOperand(1);
9833 static SDValue PerformBTCombine(SDNode *N,
9835 TargetLowering::DAGCombinerInfo &DCI) {
9836 // BT ignores high bits in the bit index operand.
9837 SDValue Op1 = N->getOperand(1);
9838 if (Op1.hasOneUse()) {
9839 unsigned BitWidth = Op1.getValueSizeInBits();
9840 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9841 APInt KnownZero, KnownOne;
9842 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9843 !DCI.isBeforeLegalizeOps());
9844 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9845 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9846 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9847 DCI.CommitTargetLoweringOpt(TLO);
9852 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9853 SDValue Op = N->getOperand(0);
9854 if (Op.getOpcode() == ISD::BIT_CONVERT)
9855 Op = Op.getOperand(0);
9856 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9857 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9858 VT.getVectorElementType().getSizeInBits() ==
9859 OpVT.getVectorElementType().getSizeInBits()) {
9860 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9865 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9866 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9867 // (and (i32 x86isd::setcc_carry), 1)
9868 // This eliminates the zext. This transformation is necessary because
9869 // ISD::SETCC is always legalized to i8.
9870 DebugLoc dl = N->getDebugLoc();
9871 SDValue N0 = N->getOperand(0);
9872 EVT VT = N->getValueType(0);
9873 if (N0.getOpcode() == ISD::AND &&
9875 N0.getOperand(0).hasOneUse()) {
9876 SDValue N00 = N0.getOperand(0);
9877 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9879 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9880 if (!C || C->getZExtValue() != 1)
9882 return DAG.getNode(ISD::AND, dl, VT,
9883 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9884 N00.getOperand(0), N00.getOperand(1)),
9885 DAG.getConstant(1, VT));
9891 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9892 DAGCombinerInfo &DCI) const {
9893 SelectionDAG &DAG = DCI.DAG;
9894 switch (N->getOpcode()) {
9896 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9897 case ISD::EXTRACT_VECTOR_ELT:
9898 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9899 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9900 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9901 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9904 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9905 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
9906 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9908 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9909 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9910 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9911 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9912 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9918 /// isTypeDesirableForOp - Return true if the target has native support for
9919 /// the specified value type and it is 'desirable' to use the type for the
9920 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9921 /// instruction encodings are longer and some i16 instructions are slow.
9922 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9923 if (!isTypeLegal(VT))
9932 case ISD::SIGN_EXTEND:
9933 case ISD::ZERO_EXTEND:
9934 case ISD::ANY_EXTEND:
9947 static bool MayFoldLoad(SDValue Op) {
9948 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
9951 static bool MayFoldIntoStore(SDValue Op) {
9952 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
9955 /// IsDesirableToPromoteOp - This method query the target whether it is
9956 /// beneficial for dag combiner to promote the specified node. If true, it
9957 /// should return the desired promotion type by reference.
9958 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
9959 EVT VT = Op.getValueType();
9963 bool Promote = false;
9964 bool Commute = false;
9965 switch (Op.getOpcode()) {
9968 LoadSDNode *LD = cast<LoadSDNode>(Op);
9969 // If the non-extending load has a single use and it's not live out, then it
9971 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
9973 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9974 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
9975 // The only case where we'd want to promote LOAD (rather then it being
9976 // promoted as an operand is when it's only use is liveout.
9977 if (UI->getOpcode() != ISD::CopyToReg)
9984 case ISD::SIGN_EXTEND:
9985 case ISD::ZERO_EXTEND:
9986 case ISD::ANY_EXTEND:
9991 SDValue N0 = Op.getOperand(0);
9992 // Look out for (store (shl (load), x)).
9993 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10006 SDValue N0 = Op.getOperand(0);
10007 SDValue N1 = Op.getOperand(1);
10008 if (!Commute && MayFoldLoad(N1))
10010 // Avoid disabling potential load folding opportunities.
10011 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10013 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10023 //===----------------------------------------------------------------------===//
10024 // X86 Inline Assembly Support
10025 //===----------------------------------------------------------------------===//
10027 static bool LowerToBSwap(CallInst *CI) {
10028 // FIXME: this should verify that we are targetting a 486 or better. If not,
10029 // we will turn this bswap into something that will be lowered to logical ops
10030 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10031 // so don't worry about this.
10033 // Verify this is a simple bswap.
10034 if (CI->getNumArgOperands() != 1 ||
10035 CI->getType() != CI->getArgOperand(0)->getType() ||
10036 !CI->getType()->isIntegerTy())
10039 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10040 if (!Ty || Ty->getBitWidth() % 16 != 0)
10043 // Okay, we can do this xform, do so now.
10044 const Type *Tys[] = { Ty };
10045 Module *M = CI->getParent()->getParent()->getParent();
10046 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10048 Value *Op = CI->getArgOperand(0);
10049 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10051 CI->replaceAllUsesWith(Op);
10052 CI->eraseFromParent();
10056 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10057 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10058 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10060 std::string AsmStr = IA->getAsmString();
10062 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10063 SmallVector<StringRef, 4> AsmPieces;
10064 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10066 switch (AsmPieces.size()) {
10067 default: return false;
10069 AsmStr = AsmPieces[0];
10071 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10074 if (AsmPieces.size() == 2 &&
10075 (AsmPieces[0] == "bswap" ||
10076 AsmPieces[0] == "bswapq" ||
10077 AsmPieces[0] == "bswapl") &&
10078 (AsmPieces[1] == "$0" ||
10079 AsmPieces[1] == "${0:q}")) {
10080 // No need to check constraints, nothing other than the equivalent of
10081 // "=r,0" would be valid here.
10082 return LowerToBSwap(CI);
10084 // rorw $$8, ${0:w} --> llvm.bswap.i16
10085 if (CI->getType()->isIntegerTy(16) &&
10086 AsmPieces.size() == 3 &&
10087 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10088 AsmPieces[1] == "$$8," &&
10089 AsmPieces[2] == "${0:w}" &&
10090 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10092 const std::string &Constraints = IA->getConstraintString();
10093 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10094 std::sort(AsmPieces.begin(), AsmPieces.end());
10095 if (AsmPieces.size() == 4 &&
10096 AsmPieces[0] == "~{cc}" &&
10097 AsmPieces[1] == "~{dirflag}" &&
10098 AsmPieces[2] == "~{flags}" &&
10099 AsmPieces[3] == "~{fpsr}") {
10100 return LowerToBSwap(CI);
10105 if (CI->getType()->isIntegerTy(64) &&
10106 Constraints.size() >= 2 &&
10107 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10108 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10109 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10110 SmallVector<StringRef, 4> Words;
10111 SplitString(AsmPieces[0], Words, " \t");
10112 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10114 SplitString(AsmPieces[1], Words, " \t");
10115 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10117 SplitString(AsmPieces[2], Words, " \t,");
10118 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10119 Words[2] == "%edx") {
10120 return LowerToBSwap(CI);
10132 /// getConstraintType - Given a constraint letter, return the type of
10133 /// constraint it is for this target.
10134 X86TargetLowering::ConstraintType
10135 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10136 if (Constraint.size() == 1) {
10137 switch (Constraint[0]) {
10149 return C_RegisterClass;
10157 return TargetLowering::getConstraintType(Constraint);
10160 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10161 /// with another that has more specific requirements based on the type of the
10162 /// corresponding operand.
10163 const char *X86TargetLowering::
10164 LowerXConstraint(EVT ConstraintVT) const {
10165 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10166 // 'f' like normal targets.
10167 if (ConstraintVT.isFloatingPoint()) {
10168 if (Subtarget->hasSSE2())
10170 if (Subtarget->hasSSE1())
10174 return TargetLowering::LowerXConstraint(ConstraintVT);
10177 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10178 /// vector. If it is invalid, don't add anything to Ops.
10179 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10181 std::vector<SDValue>&Ops,
10182 SelectionDAG &DAG) const {
10183 SDValue Result(0, 0);
10185 switch (Constraint) {
10188 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10189 if (C->getZExtValue() <= 31) {
10190 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10196 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10197 if (C->getZExtValue() <= 63) {
10198 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10204 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10205 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10206 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10212 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10213 if (C->getZExtValue() <= 255) {
10214 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10220 // 32-bit signed value
10221 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10222 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10223 C->getSExtValue())) {
10224 // Widen to 64 bits here to get it sign extended.
10225 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10228 // FIXME gcc accepts some relocatable values here too, but only in certain
10229 // memory models; it's complicated.
10234 // 32-bit unsigned value
10235 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10236 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10237 C->getZExtValue())) {
10238 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10242 // FIXME gcc accepts some relocatable values here too, but only in certain
10243 // memory models; it's complicated.
10247 // Literal immediates are always ok.
10248 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10249 // Widen to 64 bits here to get it sign extended.
10250 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10254 // In any sort of PIC mode addresses need to be computed at runtime by
10255 // adding in a register or some sort of table lookup. These can't
10256 // be used as immediates.
10257 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
10260 // If we are in non-pic codegen mode, we allow the address of a global (with
10261 // an optional displacement) to be used with 'i'.
10262 GlobalAddressSDNode *GA = 0;
10263 int64_t Offset = 0;
10265 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10267 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10268 Offset += GA->getOffset();
10270 } else if (Op.getOpcode() == ISD::ADD) {
10271 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10272 Offset += C->getZExtValue();
10273 Op = Op.getOperand(0);
10276 } else if (Op.getOpcode() == ISD::SUB) {
10277 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10278 Offset += -C->getZExtValue();
10279 Op = Op.getOperand(0);
10284 // Otherwise, this isn't something we can handle, reject it.
10288 const GlobalValue *GV = GA->getGlobal();
10289 // If we require an extra load to get this address, as in PIC mode, we
10290 // can't accept it.
10291 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10292 getTargetMachine())))
10295 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
10296 GA->getValueType(0), Offset);
10301 if (Result.getNode()) {
10302 Ops.push_back(Result);
10305 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10308 std::vector<unsigned> X86TargetLowering::
10309 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10311 if (Constraint.size() == 1) {
10312 // FIXME: not handling fp-stack yet!
10313 switch (Constraint[0]) { // GCC X86 Constraint Letters
10314 default: break; // Unknown constraint letter
10315 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10316 if (Subtarget->is64Bit()) {
10317 if (VT == MVT::i32)
10318 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10319 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10320 X86::R10D,X86::R11D,X86::R12D,
10321 X86::R13D,X86::R14D,X86::R15D,
10322 X86::EBP, X86::ESP, 0);
10323 else if (VT == MVT::i16)
10324 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10325 X86::SI, X86::DI, X86::R8W,X86::R9W,
10326 X86::R10W,X86::R11W,X86::R12W,
10327 X86::R13W,X86::R14W,X86::R15W,
10328 X86::BP, X86::SP, 0);
10329 else if (VT == MVT::i8)
10330 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10331 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10332 X86::R10B,X86::R11B,X86::R12B,
10333 X86::R13B,X86::R14B,X86::R15B,
10334 X86::BPL, X86::SPL, 0);
10336 else if (VT == MVT::i64)
10337 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10338 X86::RSI, X86::RDI, X86::R8, X86::R9,
10339 X86::R10, X86::R11, X86::R12,
10340 X86::R13, X86::R14, X86::R15,
10341 X86::RBP, X86::RSP, 0);
10345 // 32-bit fallthrough
10346 case 'Q': // Q_REGS
10347 if (VT == MVT::i32)
10348 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10349 else if (VT == MVT::i16)
10350 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10351 else if (VT == MVT::i8)
10352 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10353 else if (VT == MVT::i64)
10354 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10359 return std::vector<unsigned>();
10362 std::pair<unsigned, const TargetRegisterClass*>
10363 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10365 // First, see if this is a constraint that directly corresponds to an LLVM
10367 if (Constraint.size() == 1) {
10368 // GCC Constraint Letters
10369 switch (Constraint[0]) {
10371 case 'r': // GENERAL_REGS
10372 case 'l': // INDEX_REGS
10374 return std::make_pair(0U, X86::GR8RegisterClass);
10375 if (VT == MVT::i16)
10376 return std::make_pair(0U, X86::GR16RegisterClass);
10377 if (VT == MVT::i32 || !Subtarget->is64Bit())
10378 return std::make_pair(0U, X86::GR32RegisterClass);
10379 return std::make_pair(0U, X86::GR64RegisterClass);
10380 case 'R': // LEGACY_REGS
10382 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10383 if (VT == MVT::i16)
10384 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10385 if (VT == MVT::i32 || !Subtarget->is64Bit())
10386 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10387 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10388 case 'f': // FP Stack registers.
10389 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10390 // value to the correct fpstack register class.
10391 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10392 return std::make_pair(0U, X86::RFP32RegisterClass);
10393 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10394 return std::make_pair(0U, X86::RFP64RegisterClass);
10395 return std::make_pair(0U, X86::RFP80RegisterClass);
10396 case 'y': // MMX_REGS if MMX allowed.
10397 if (!Subtarget->hasMMX()) break;
10398 return std::make_pair(0U, X86::VR64RegisterClass);
10399 case 'Y': // SSE_REGS if SSE2 allowed
10400 if (!Subtarget->hasSSE2()) break;
10402 case 'x': // SSE_REGS if SSE1 allowed
10403 if (!Subtarget->hasSSE1()) break;
10405 switch (VT.getSimpleVT().SimpleTy) {
10407 // Scalar SSE types.
10410 return std::make_pair(0U, X86::FR32RegisterClass);
10413 return std::make_pair(0U, X86::FR64RegisterClass);
10421 return std::make_pair(0U, X86::VR128RegisterClass);
10427 // Use the default implementation in TargetLowering to convert the register
10428 // constraint into a member of a register class.
10429 std::pair<unsigned, const TargetRegisterClass*> Res;
10430 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10432 // Not found as a standard register?
10433 if (Res.second == 0) {
10434 // Map st(0) -> st(7) -> ST0
10435 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10436 tolower(Constraint[1]) == 's' &&
10437 tolower(Constraint[2]) == 't' &&
10438 Constraint[3] == '(' &&
10439 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10440 Constraint[5] == ')' &&
10441 Constraint[6] == '}') {
10443 Res.first = X86::ST0+Constraint[4]-'0';
10444 Res.second = X86::RFP80RegisterClass;
10448 // GCC allows "st(0)" to be called just plain "st".
10449 if (StringRef("{st}").equals_lower(Constraint)) {
10450 Res.first = X86::ST0;
10451 Res.second = X86::RFP80RegisterClass;
10456 if (StringRef("{flags}").equals_lower(Constraint)) {
10457 Res.first = X86::EFLAGS;
10458 Res.second = X86::CCRRegisterClass;
10462 // 'A' means EAX + EDX.
10463 if (Constraint == "A") {
10464 Res.first = X86::EAX;
10465 Res.second = X86::GR32_ADRegisterClass;
10471 // Otherwise, check to see if this is a register class of the wrong value
10472 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10473 // turn into {ax},{dx}.
10474 if (Res.second->hasType(VT))
10475 return Res; // Correct type already, nothing to do.
10477 // All of the single-register GCC register classes map their values onto
10478 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10479 // really want an 8-bit or 32-bit register, map to the appropriate register
10480 // class and return the appropriate register.
10481 if (Res.second == X86::GR16RegisterClass) {
10482 if (VT == MVT::i8) {
10483 unsigned DestReg = 0;
10484 switch (Res.first) {
10486 case X86::AX: DestReg = X86::AL; break;
10487 case X86::DX: DestReg = X86::DL; break;
10488 case X86::CX: DestReg = X86::CL; break;
10489 case X86::BX: DestReg = X86::BL; break;
10492 Res.first = DestReg;
10493 Res.second = X86::GR8RegisterClass;
10495 } else if (VT == MVT::i32) {
10496 unsigned DestReg = 0;
10497 switch (Res.first) {
10499 case X86::AX: DestReg = X86::EAX; break;
10500 case X86::DX: DestReg = X86::EDX; break;
10501 case X86::CX: DestReg = X86::ECX; break;
10502 case X86::BX: DestReg = X86::EBX; break;
10503 case X86::SI: DestReg = X86::ESI; break;
10504 case X86::DI: DestReg = X86::EDI; break;
10505 case X86::BP: DestReg = X86::EBP; break;
10506 case X86::SP: DestReg = X86::ESP; break;
10509 Res.first = DestReg;
10510 Res.second = X86::GR32RegisterClass;
10512 } else if (VT == MVT::i64) {
10513 unsigned DestReg = 0;
10514 switch (Res.first) {
10516 case X86::AX: DestReg = X86::RAX; break;
10517 case X86::DX: DestReg = X86::RDX; break;
10518 case X86::CX: DestReg = X86::RCX; break;
10519 case X86::BX: DestReg = X86::RBX; break;
10520 case X86::SI: DestReg = X86::RSI; break;
10521 case X86::DI: DestReg = X86::RDI; break;
10522 case X86::BP: DestReg = X86::RBP; break;
10523 case X86::SP: DestReg = X86::RSP; break;
10526 Res.first = DestReg;
10527 Res.second = X86::GR64RegisterClass;
10530 } else if (Res.second == X86::FR32RegisterClass ||
10531 Res.second == X86::FR64RegisterClass ||
10532 Res.second == X86::VR128RegisterClass) {
10533 // Handle references to XMM physical registers that got mapped into the
10534 // wrong class. This can happen with constraints like {xmm0} where the
10535 // target independent register mapper will just pick the first match it can
10536 // find, ignoring the required type.
10537 if (VT == MVT::f32)
10538 Res.second = X86::FR32RegisterClass;
10539 else if (VT == MVT::f64)
10540 Res.second = X86::FR64RegisterClass;
10541 else if (X86::VR128RegisterClass->hasType(VT))
10542 Res.second = X86::VR128RegisterClass;