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 //===----------------------------------------------------------------------===//
1193 // Return Value Calling Convention Implementation
1194 //===----------------------------------------------------------------------===//
1196 #include "X86GenCallingConv.inc"
1199 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1200 const SmallVectorImpl<EVT> &OutTys,
1201 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1202 SelectionDAG &DAG) const {
1203 SmallVector<CCValAssign, 16> RVLocs;
1204 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1205 RVLocs, *DAG.getContext());
1206 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1210 X86TargetLowering::LowerReturn(SDValue Chain,
1211 CallingConv::ID CallConv, bool isVarArg,
1212 const SmallVectorImpl<ISD::OutputArg> &Outs,
1213 DebugLoc dl, SelectionDAG &DAG) const {
1214 MachineFunction &MF = DAG.getMachineFunction();
1215 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1217 SmallVector<CCValAssign, 16> RVLocs;
1218 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1219 RVLocs, *DAG.getContext());
1220 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1222 // Add the regs to the liveout set for the function.
1223 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1224 for (unsigned i = 0; i != RVLocs.size(); ++i)
1225 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1226 MRI.addLiveOut(RVLocs[i].getLocReg());
1230 SmallVector<SDValue, 6> RetOps;
1231 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1232 // Operand #1 = Bytes To Pop
1233 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1236 // Copy the result values into the output registers.
1237 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1238 CCValAssign &VA = RVLocs[i];
1239 assert(VA.isRegLoc() && "Can only return in registers!");
1240 SDValue ValToCopy = Outs[i].Val;
1242 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1243 // the RET instruction and handled by the FP Stackifier.
1244 if (VA.getLocReg() == X86::ST0 ||
1245 VA.getLocReg() == X86::ST1) {
1246 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1247 // change the value to the FP stack register class.
1248 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1249 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1250 RetOps.push_back(ValToCopy);
1251 // Don't emit a copytoreg.
1255 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1256 // which is returned in RAX / RDX.
1257 if (Subtarget->is64Bit()) {
1258 EVT ValVT = ValToCopy.getValueType();
1259 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1260 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1261 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1262 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1266 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1267 Flag = Chain.getValue(1);
1270 // The x86-64 ABI for returning structs by value requires that we copy
1271 // the sret argument into %rax for the return. We saved the argument into
1272 // a virtual register in the entry block, so now we copy the value out
1274 if (Subtarget->is64Bit() &&
1275 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1276 MachineFunction &MF = DAG.getMachineFunction();
1277 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1278 unsigned Reg = FuncInfo->getSRetReturnReg();
1280 "SRetReturnReg should have been set in LowerFormalArguments().");
1281 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1283 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1284 Flag = Chain.getValue(1);
1286 // RAX now acts like a return value.
1287 MRI.addLiveOut(X86::RAX);
1290 RetOps[0] = Chain; // Update chain.
1292 // Add the flag if we have it.
1294 RetOps.push_back(Flag);
1296 return DAG.getNode(X86ISD::RET_FLAG, dl,
1297 MVT::Other, &RetOps[0], RetOps.size());
1300 /// LowerCallResult - Lower the result values of a call into the
1301 /// appropriate copies out of appropriate physical registers.
1304 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1305 CallingConv::ID CallConv, bool isVarArg,
1306 const SmallVectorImpl<ISD::InputArg> &Ins,
1307 DebugLoc dl, SelectionDAG &DAG,
1308 SmallVectorImpl<SDValue> &InVals) const {
1310 // Assign locations to each value returned by this call.
1311 SmallVector<CCValAssign, 16> RVLocs;
1312 bool Is64Bit = Subtarget->is64Bit();
1313 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1314 RVLocs, *DAG.getContext());
1315 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1317 // Copy all of the result registers out of their specified physreg.
1318 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1319 CCValAssign &VA = RVLocs[i];
1320 EVT CopyVT = VA.getValVT();
1322 // If this is x86-64, and we disabled SSE, we can't return FP values
1323 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1324 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1325 report_fatal_error("SSE register return with SSE disabled");
1328 // If this is a call to a function that returns an fp value on the floating
1329 // point stack, but where we prefer to use the value in xmm registers, copy
1330 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1331 if ((VA.getLocReg() == X86::ST0 ||
1332 VA.getLocReg() == X86::ST1) &&
1333 isScalarFPTypeInSSEReg(VA.getValVT())) {
1338 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1339 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1340 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1341 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1342 MVT::v2i64, InFlag).getValue(1);
1343 Val = Chain.getValue(0);
1344 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1345 Val, DAG.getConstant(0, MVT::i64));
1347 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1348 MVT::i64, InFlag).getValue(1);
1349 Val = Chain.getValue(0);
1351 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1353 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1354 CopyVT, InFlag).getValue(1);
1355 Val = Chain.getValue(0);
1357 InFlag = Chain.getValue(2);
1359 if (CopyVT != VA.getValVT()) {
1360 // Round the F80 the right size, which also moves to the appropriate xmm
1362 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1363 // This truncation won't change the value.
1364 DAG.getIntPtrConstant(1));
1367 InVals.push_back(Val);
1374 //===----------------------------------------------------------------------===//
1375 // C & StdCall & Fast Calling Convention implementation
1376 //===----------------------------------------------------------------------===//
1377 // StdCall calling convention seems to be standard for many Windows' API
1378 // routines and around. It differs from C calling convention just a little:
1379 // callee should clean up the stack, not caller. Symbols should be also
1380 // decorated in some fancy way :) It doesn't support any vector arguments.
1381 // For info on fast calling convention see Fast Calling Convention (tail call)
1382 // implementation LowerX86_32FastCCCallTo.
1384 /// CallIsStructReturn - Determines whether a call uses struct return
1386 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1390 return Outs[0].Flags.isSRet();
1393 /// ArgsAreStructReturn - Determines whether a function uses struct
1394 /// return semantics.
1396 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1400 return Ins[0].Flags.isSRet();
1403 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1404 /// given CallingConvention value.
1405 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1406 if (Subtarget->is64Bit()) {
1407 if (CC == CallingConv::GHC)
1408 return CC_X86_64_GHC;
1409 else if (Subtarget->isTargetWin64())
1410 return CC_X86_Win64_C;
1415 if (CC == CallingConv::X86_FastCall)
1416 return CC_X86_32_FastCall;
1417 else if (CC == CallingConv::X86_ThisCall)
1418 return CC_X86_32_ThisCall;
1419 else if (CC == CallingConv::Fast)
1420 return CC_X86_32_FastCC;
1421 else if (CC == CallingConv::GHC)
1422 return CC_X86_32_GHC;
1427 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1428 /// by "Src" to address "Dst" with size and alignment information specified by
1429 /// the specific parameter attribute. The copy will be passed as a byval
1430 /// function parameter.
1432 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1433 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1435 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1436 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1437 /*isVolatile*/false, /*AlwaysInline=*/true,
1441 /// IsTailCallConvention - Return true if the calling convention is one that
1442 /// supports tail call optimization.
1443 static bool IsTailCallConvention(CallingConv::ID CC) {
1444 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1447 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1448 /// a tailcall target by changing its ABI.
1449 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1450 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1454 X86TargetLowering::LowerMemArgument(SDValue Chain,
1455 CallingConv::ID CallConv,
1456 const SmallVectorImpl<ISD::InputArg> &Ins,
1457 DebugLoc dl, SelectionDAG &DAG,
1458 const CCValAssign &VA,
1459 MachineFrameInfo *MFI,
1461 // Create the nodes corresponding to a load from this parameter slot.
1462 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1463 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1464 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1467 // If value is passed by pointer we have address passed instead of the value
1469 if (VA.getLocInfo() == CCValAssign::Indirect)
1470 ValVT = VA.getLocVT();
1472 ValVT = VA.getValVT();
1474 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1475 // changed with more analysis.
1476 // In case of tail call optimization mark all arguments mutable. Since they
1477 // could be overwritten by lowering of arguments in case of a tail call.
1478 if (Flags.isByVal()) {
1479 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1480 VA.getLocMemOffset(), isImmutable);
1481 return DAG.getFrameIndex(FI, getPointerTy());
1483 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1484 VA.getLocMemOffset(), isImmutable);
1485 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1486 return DAG.getLoad(ValVT, dl, Chain, FIN,
1487 PseudoSourceValue::getFixedStack(FI), 0,
1493 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1494 CallingConv::ID CallConv,
1496 const SmallVectorImpl<ISD::InputArg> &Ins,
1499 SmallVectorImpl<SDValue> &InVals)
1501 MachineFunction &MF = DAG.getMachineFunction();
1502 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1504 const Function* Fn = MF.getFunction();
1505 if (Fn->hasExternalLinkage() &&
1506 Subtarget->isTargetCygMing() &&
1507 Fn->getName() == "main")
1508 FuncInfo->setForceFramePointer(true);
1510 MachineFrameInfo *MFI = MF.getFrameInfo();
1511 bool Is64Bit = Subtarget->is64Bit();
1512 bool IsWin64 = Subtarget->isTargetWin64();
1514 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1515 "Var args not supported with calling convention fastcc or ghc");
1517 // Assign locations to all of the incoming arguments.
1518 SmallVector<CCValAssign, 16> ArgLocs;
1519 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1520 ArgLocs, *DAG.getContext());
1521 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1523 unsigned LastVal = ~0U;
1525 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1526 CCValAssign &VA = ArgLocs[i];
1527 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1529 assert(VA.getValNo() != LastVal &&
1530 "Don't support value assigned to multiple locs yet");
1531 LastVal = VA.getValNo();
1533 if (VA.isRegLoc()) {
1534 EVT RegVT = VA.getLocVT();
1535 TargetRegisterClass *RC = NULL;
1536 if (RegVT == MVT::i32)
1537 RC = X86::GR32RegisterClass;
1538 else if (Is64Bit && RegVT == MVT::i64)
1539 RC = X86::GR64RegisterClass;
1540 else if (RegVT == MVT::f32)
1541 RC = X86::FR32RegisterClass;
1542 else if (RegVT == MVT::f64)
1543 RC = X86::FR64RegisterClass;
1544 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1545 RC = X86::VR128RegisterClass;
1546 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1547 RC = X86::VR64RegisterClass;
1549 llvm_unreachable("Unknown argument type!");
1551 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1552 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1554 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1555 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1557 if (VA.getLocInfo() == CCValAssign::SExt)
1558 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1559 DAG.getValueType(VA.getValVT()));
1560 else if (VA.getLocInfo() == CCValAssign::ZExt)
1561 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1562 DAG.getValueType(VA.getValVT()));
1563 else if (VA.getLocInfo() == CCValAssign::BCvt)
1564 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1566 if (VA.isExtInLoc()) {
1567 // Handle MMX values passed in XMM regs.
1568 if (RegVT.isVector()) {
1569 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1570 ArgValue, DAG.getConstant(0, MVT::i64));
1571 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1573 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1576 assert(VA.isMemLoc());
1577 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1580 // If value is passed via pointer - do a load.
1581 if (VA.getLocInfo() == CCValAssign::Indirect)
1582 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1585 InVals.push_back(ArgValue);
1588 // The x86-64 ABI for returning structs by value requires that we copy
1589 // the sret argument into %rax for the return. Save the argument into
1590 // a virtual register so that we can access it from the return points.
1591 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1592 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1593 unsigned Reg = FuncInfo->getSRetReturnReg();
1595 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1596 FuncInfo->setSRetReturnReg(Reg);
1598 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1599 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1602 unsigned StackSize = CCInfo.getNextStackOffset();
1603 // Align stack specially for tail calls.
1604 if (FuncIsMadeTailCallSafe(CallConv))
1605 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1607 // If the function takes variable number of arguments, make a frame index for
1608 // the start of the first vararg value... for expansion of llvm.va_start.
1610 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1611 CallConv != CallingConv::X86_ThisCall)) {
1612 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1615 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1617 // FIXME: We should really autogenerate these arrays
1618 static const unsigned GPR64ArgRegsWin64[] = {
1619 X86::RCX, X86::RDX, X86::R8, X86::R9
1621 static const unsigned XMMArgRegsWin64[] = {
1622 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1624 static const unsigned GPR64ArgRegs64Bit[] = {
1625 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1627 static const unsigned XMMArgRegs64Bit[] = {
1628 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1629 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1631 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1634 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1635 GPR64ArgRegs = GPR64ArgRegsWin64;
1636 XMMArgRegs = XMMArgRegsWin64;
1638 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1639 GPR64ArgRegs = GPR64ArgRegs64Bit;
1640 XMMArgRegs = XMMArgRegs64Bit;
1642 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1644 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1647 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1648 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1649 "SSE register cannot be used when SSE is disabled!");
1650 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1651 "SSE register cannot be used when SSE is disabled!");
1652 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1653 // Kernel mode asks for SSE to be disabled, so don't push them
1655 TotalNumXMMRegs = 0;
1657 // For X86-64, if there are vararg parameters that are passed via
1658 // registers, then we must store them to their spots on the stack so they
1659 // may be loaded by deferencing the result of va_next.
1660 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1661 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1662 FuncInfo->setRegSaveFrameIndex(
1663 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1666 // Store the integer parameter registers.
1667 SmallVector<SDValue, 8> MemOps;
1668 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1670 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1671 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1672 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1673 DAG.getIntPtrConstant(Offset));
1674 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1675 X86::GR64RegisterClass);
1676 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1678 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1679 PseudoSourceValue::getFixedStack(
1680 FuncInfo->getRegSaveFrameIndex()),
1681 Offset, false, false, 0);
1682 MemOps.push_back(Store);
1686 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1687 // Now store the XMM (fp + vector) parameter registers.
1688 SmallVector<SDValue, 11> SaveXMMOps;
1689 SaveXMMOps.push_back(Chain);
1691 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1692 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1693 SaveXMMOps.push_back(ALVal);
1695 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1696 FuncInfo->getRegSaveFrameIndex()));
1697 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1698 FuncInfo->getVarArgsFPOffset()));
1700 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1701 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1702 X86::VR128RegisterClass);
1703 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1704 SaveXMMOps.push_back(Val);
1706 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1708 &SaveXMMOps[0], SaveXMMOps.size()));
1711 if (!MemOps.empty())
1712 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1713 &MemOps[0], MemOps.size());
1717 // Some CCs need callee pop.
1718 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1719 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1721 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1722 // If this is an sret function, the return should pop the hidden pointer.
1723 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1724 FuncInfo->setBytesToPopOnReturn(4);
1728 // RegSaveFrameIndex is X86-64 only.
1729 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1730 if (CallConv == CallingConv::X86_FastCall ||
1731 CallConv == CallingConv::X86_ThisCall)
1732 // fastcc functions can't have varargs.
1733 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1740 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1741 SDValue StackPtr, SDValue Arg,
1742 DebugLoc dl, SelectionDAG &DAG,
1743 const CCValAssign &VA,
1744 ISD::ArgFlagsTy Flags) const {
1745 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1746 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1747 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1748 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1749 if (Flags.isByVal()) {
1750 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1752 return DAG.getStore(Chain, dl, Arg, PtrOff,
1753 PseudoSourceValue::getStack(), LocMemOffset,
1757 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1758 /// optimization is performed and it is required.
1760 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1761 SDValue &OutRetAddr, SDValue Chain,
1762 bool IsTailCall, bool Is64Bit,
1763 int FPDiff, DebugLoc dl) const {
1764 // Adjust the Return address stack slot.
1765 EVT VT = getPointerTy();
1766 OutRetAddr = getReturnAddressFrameIndex(DAG);
1768 // Load the "old" Return address.
1769 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1770 return SDValue(OutRetAddr.getNode(), 1);
1773 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1774 /// optimization is performed and it is required (FPDiff!=0).
1776 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1777 SDValue Chain, SDValue RetAddrFrIdx,
1778 bool Is64Bit, int FPDiff, DebugLoc dl) {
1779 // Store the return address to the appropriate stack slot.
1780 if (!FPDiff) return Chain;
1781 // Calculate the new stack slot for the return address.
1782 int SlotSize = Is64Bit ? 8 : 4;
1783 int NewReturnAddrFI =
1784 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1785 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1786 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1787 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1788 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1794 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1795 CallingConv::ID CallConv, bool isVarArg,
1797 const SmallVectorImpl<ISD::OutputArg> &Outs,
1798 const SmallVectorImpl<ISD::InputArg> &Ins,
1799 DebugLoc dl, SelectionDAG &DAG,
1800 SmallVectorImpl<SDValue> &InVals) const {
1801 MachineFunction &MF = DAG.getMachineFunction();
1802 bool Is64Bit = Subtarget->is64Bit();
1803 bool IsStructRet = CallIsStructReturn(Outs);
1804 bool IsSibcall = false;
1807 // Check if it's really possible to do a tail call.
1808 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1809 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1812 // Sibcalls are automatically detected tailcalls which do not require
1814 if (!GuaranteedTailCallOpt && isTailCall)
1821 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1822 "Var args not supported with calling convention fastcc or ghc");
1824 // Analyze operands of the call, assigning locations to each operand.
1825 SmallVector<CCValAssign, 16> ArgLocs;
1826 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1827 ArgLocs, *DAG.getContext());
1828 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1830 // Get a count of how many bytes are to be pushed on the stack.
1831 unsigned NumBytes = CCInfo.getNextStackOffset();
1833 // This is a sibcall. The memory operands are available in caller's
1834 // own caller's stack.
1836 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1837 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1840 if (isTailCall && !IsSibcall) {
1841 // Lower arguments at fp - stackoffset + fpdiff.
1842 unsigned NumBytesCallerPushed =
1843 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1844 FPDiff = NumBytesCallerPushed - NumBytes;
1846 // Set the delta of movement of the returnaddr stackslot.
1847 // But only set if delta is greater than previous delta.
1848 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1849 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1853 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1855 SDValue RetAddrFrIdx;
1856 // Load return adress for tail calls.
1857 if (isTailCall && FPDiff)
1858 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1859 Is64Bit, FPDiff, dl);
1861 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1862 SmallVector<SDValue, 8> MemOpChains;
1865 // Walk the register/memloc assignments, inserting copies/loads. In the case
1866 // of tail call optimization arguments are handle later.
1867 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1868 CCValAssign &VA = ArgLocs[i];
1869 EVT RegVT = VA.getLocVT();
1870 SDValue Arg = Outs[i].Val;
1871 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1872 bool isByVal = Flags.isByVal();
1874 // Promote the value if needed.
1875 switch (VA.getLocInfo()) {
1876 default: llvm_unreachable("Unknown loc info!");
1877 case CCValAssign::Full: break;
1878 case CCValAssign::SExt:
1879 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1881 case CCValAssign::ZExt:
1882 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1884 case CCValAssign::AExt:
1885 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1886 // Special case: passing MMX values in XMM registers.
1887 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1888 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1889 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1891 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1893 case CCValAssign::BCvt:
1894 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1896 case CCValAssign::Indirect: {
1897 // Store the argument.
1898 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1899 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1900 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1901 PseudoSourceValue::getFixedStack(FI), 0,
1908 if (VA.isRegLoc()) {
1909 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1910 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1911 assert(VA.isMemLoc());
1912 if (StackPtr.getNode() == 0)
1913 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1914 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1915 dl, DAG, VA, Flags));
1919 if (!MemOpChains.empty())
1920 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1921 &MemOpChains[0], MemOpChains.size());
1923 // Build a sequence of copy-to-reg nodes chained together with token chain
1924 // and flag operands which copy the outgoing args into registers.
1926 // Tail call byval lowering might overwrite argument registers so in case of
1927 // tail call optimization the copies to registers are lowered later.
1929 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1930 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1931 RegsToPass[i].second, InFlag);
1932 InFlag = Chain.getValue(1);
1935 if (Subtarget->isPICStyleGOT()) {
1936 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1939 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1940 DAG.getNode(X86ISD::GlobalBaseReg,
1941 DebugLoc(), getPointerTy()),
1943 InFlag = Chain.getValue(1);
1945 // If we are tail calling and generating PIC/GOT style code load the
1946 // address of the callee into ECX. The value in ecx is used as target of
1947 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1948 // for tail calls on PIC/GOT architectures. Normally we would just put the
1949 // address of GOT into ebx and then call target@PLT. But for tail calls
1950 // ebx would be restored (since ebx is callee saved) before jumping to the
1953 // Note: The actual moving to ECX is done further down.
1954 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1955 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1956 !G->getGlobal()->hasProtectedVisibility())
1957 Callee = LowerGlobalAddress(Callee, DAG);
1958 else if (isa<ExternalSymbolSDNode>(Callee))
1959 Callee = LowerExternalSymbol(Callee, DAG);
1963 if (Is64Bit && isVarArg) {
1964 // From AMD64 ABI document:
1965 // For calls that may call functions that use varargs or stdargs
1966 // (prototype-less calls or calls to functions containing ellipsis (...) in
1967 // the declaration) %al is used as hidden argument to specify the number
1968 // of SSE registers used. The contents of %al do not need to match exactly
1969 // the number of registers, but must be an ubound on the number of SSE
1970 // registers used and is in the range 0 - 8 inclusive.
1972 // FIXME: Verify this on Win64
1973 // Count the number of XMM registers allocated.
1974 static const unsigned XMMArgRegs[] = {
1975 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1976 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1978 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1979 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1980 && "SSE registers cannot be used when SSE is disabled");
1982 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1983 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1984 InFlag = Chain.getValue(1);
1988 // For tail calls lower the arguments to the 'real' stack slot.
1990 // Force all the incoming stack arguments to be loaded from the stack
1991 // before any new outgoing arguments are stored to the stack, because the
1992 // outgoing stack slots may alias the incoming argument stack slots, and
1993 // the alias isn't otherwise explicit. This is slightly more conservative
1994 // than necessary, because it means that each store effectively depends
1995 // on every argument instead of just those arguments it would clobber.
1996 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1998 SmallVector<SDValue, 8> MemOpChains2;
2001 // Do not flag preceeding copytoreg stuff together with the following stuff.
2003 if (GuaranteedTailCallOpt) {
2004 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2005 CCValAssign &VA = ArgLocs[i];
2008 assert(VA.isMemLoc());
2009 SDValue Arg = Outs[i].Val;
2010 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2011 // Create frame index.
2012 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2013 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2014 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2015 FIN = DAG.getFrameIndex(FI, getPointerTy());
2017 if (Flags.isByVal()) {
2018 // Copy relative to framepointer.
2019 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2020 if (StackPtr.getNode() == 0)
2021 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2023 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2025 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2029 // Store relative to framepointer.
2030 MemOpChains2.push_back(
2031 DAG.getStore(ArgChain, dl, Arg, FIN,
2032 PseudoSourceValue::getFixedStack(FI), 0,
2038 if (!MemOpChains2.empty())
2039 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2040 &MemOpChains2[0], MemOpChains2.size());
2042 // Copy arguments to their registers.
2043 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2044 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2045 RegsToPass[i].second, InFlag);
2046 InFlag = Chain.getValue(1);
2050 // Store the return address to the appropriate stack slot.
2051 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2055 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2056 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2057 // In the 64-bit large code model, we have to make all calls
2058 // through a register, since the call instruction's 32-bit
2059 // pc-relative offset may not be large enough to hold the whole
2061 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2062 // If the callee is a GlobalAddress node (quite common, every direct call
2063 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2066 // We should use extra load for direct calls to dllimported functions in
2068 const GlobalValue *GV = G->getGlobal();
2069 if (!GV->hasDLLImportLinkage()) {
2070 unsigned char OpFlags = 0;
2072 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2073 // external symbols most go through the PLT in PIC mode. If the symbol
2074 // has hidden or protected visibility, or if it is static or local, then
2075 // we don't need to use the PLT - we can directly call it.
2076 if (Subtarget->isTargetELF() &&
2077 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2078 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2079 OpFlags = X86II::MO_PLT;
2080 } else if (Subtarget->isPICStyleStubAny() &&
2081 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2082 Subtarget->getDarwinVers() < 9) {
2083 // PC-relative references to external symbols should go through $stub,
2084 // unless we're building with the leopard linker or later, which
2085 // automatically synthesizes these stubs.
2086 OpFlags = X86II::MO_DARWIN_STUB;
2089 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2090 G->getOffset(), OpFlags);
2092 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2093 unsigned char OpFlags = 0;
2095 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2096 // symbols should go through the PLT.
2097 if (Subtarget->isTargetELF() &&
2098 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2099 OpFlags = X86II::MO_PLT;
2100 } else if (Subtarget->isPICStyleStubAny() &&
2101 Subtarget->getDarwinVers() < 9) {
2102 // PC-relative references to external symbols should go through $stub,
2103 // unless we're building with the leopard linker or later, which
2104 // automatically synthesizes these stubs.
2105 OpFlags = X86II::MO_DARWIN_STUB;
2108 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2112 // Returns a chain & a flag for retval copy to use.
2113 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2114 SmallVector<SDValue, 8> Ops;
2116 if (!IsSibcall && isTailCall) {
2117 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2118 DAG.getIntPtrConstant(0, true), InFlag);
2119 InFlag = Chain.getValue(1);
2122 Ops.push_back(Chain);
2123 Ops.push_back(Callee);
2126 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2128 // Add argument registers to the end of the list so that they are known live
2130 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2131 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2132 RegsToPass[i].second.getValueType()));
2134 // Add an implicit use GOT pointer in EBX.
2135 if (!isTailCall && Subtarget->isPICStyleGOT())
2136 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2138 // Add an implicit use of AL for x86 vararg functions.
2139 if (Is64Bit && isVarArg)
2140 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2142 if (InFlag.getNode())
2143 Ops.push_back(InFlag);
2147 //// If this is the first return lowered for this function, add the regs
2148 //// to the liveout set for the function.
2149 // This isn't right, although it's probably harmless on x86; liveouts
2150 // should be computed from returns not tail calls. Consider a void
2151 // function making a tail call to a function returning int.
2152 return DAG.getNode(X86ISD::TC_RETURN, dl,
2153 NodeTys, &Ops[0], Ops.size());
2156 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2157 InFlag = Chain.getValue(1);
2159 // Create the CALLSEQ_END node.
2160 unsigned NumBytesForCalleeToPush;
2161 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2162 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2163 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2164 // If this is a call to a struct-return function, the callee
2165 // pops the hidden struct pointer, so we have to push it back.
2166 // This is common for Darwin/X86, Linux & Mingw32 targets.
2167 NumBytesForCalleeToPush = 4;
2169 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2171 // Returns a flag for retval copy to use.
2173 Chain = DAG.getCALLSEQ_END(Chain,
2174 DAG.getIntPtrConstant(NumBytes, true),
2175 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2178 InFlag = Chain.getValue(1);
2181 // Handle result values, copying them out of physregs into vregs that we
2183 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2184 Ins, dl, DAG, InVals);
2188 //===----------------------------------------------------------------------===//
2189 // Fast Calling Convention (tail call) implementation
2190 //===----------------------------------------------------------------------===//
2192 // Like std call, callee cleans arguments, convention except that ECX is
2193 // reserved for storing the tail called function address. Only 2 registers are
2194 // free for argument passing (inreg). Tail call optimization is performed
2196 // * tailcallopt is enabled
2197 // * caller/callee are fastcc
2198 // On X86_64 architecture with GOT-style position independent code only local
2199 // (within module) calls are supported at the moment.
2200 // To keep the stack aligned according to platform abi the function
2201 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2202 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2203 // If a tail called function callee has more arguments than the caller the
2204 // caller needs to make sure that there is room to move the RETADDR to. This is
2205 // achieved by reserving an area the size of the argument delta right after the
2206 // original REtADDR, but before the saved framepointer or the spilled registers
2207 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2219 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2220 /// for a 16 byte align requirement.
2222 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2223 SelectionDAG& DAG) const {
2224 MachineFunction &MF = DAG.getMachineFunction();
2225 const TargetMachine &TM = MF.getTarget();
2226 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2227 unsigned StackAlignment = TFI.getStackAlignment();
2228 uint64_t AlignMask = StackAlignment - 1;
2229 int64_t Offset = StackSize;
2230 uint64_t SlotSize = TD->getPointerSize();
2231 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2232 // Number smaller than 12 so just add the difference.
2233 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2235 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2236 Offset = ((~AlignMask) & Offset) + StackAlignment +
2237 (StackAlignment-SlotSize);
2242 /// MatchingStackOffset - Return true if the given stack call argument is
2243 /// already available in the same position (relatively) of the caller's
2244 /// incoming argument stack.
2246 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2247 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2248 const X86InstrInfo *TII) {
2249 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2251 if (Arg.getOpcode() == ISD::CopyFromReg) {
2252 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2253 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2255 MachineInstr *Def = MRI->getVRegDef(VR);
2258 if (!Flags.isByVal()) {
2259 if (!TII->isLoadFromStackSlot(Def, FI))
2262 unsigned Opcode = Def->getOpcode();
2263 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2264 Def->getOperand(1).isFI()) {
2265 FI = Def->getOperand(1).getIndex();
2266 Bytes = Flags.getByValSize();
2270 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2271 if (Flags.isByVal())
2272 // ByVal argument is passed in as a pointer but it's now being
2273 // dereferenced. e.g.
2274 // define @foo(%struct.X* %A) {
2275 // tail call @bar(%struct.X* byval %A)
2278 SDValue Ptr = Ld->getBasePtr();
2279 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2282 FI = FINode->getIndex();
2286 assert(FI != INT_MAX);
2287 if (!MFI->isFixedObjectIndex(FI))
2289 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2292 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2293 /// for tail call optimization. Targets which want to do tail call
2294 /// optimization should implement this function.
2296 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2297 CallingConv::ID CalleeCC,
2299 bool isCalleeStructRet,
2300 bool isCallerStructRet,
2301 const SmallVectorImpl<ISD::OutputArg> &Outs,
2302 const SmallVectorImpl<ISD::InputArg> &Ins,
2303 SelectionDAG& DAG) const {
2304 if (!IsTailCallConvention(CalleeCC) &&
2305 CalleeCC != CallingConv::C)
2308 // If -tailcallopt is specified, make fastcc functions tail-callable.
2309 const MachineFunction &MF = DAG.getMachineFunction();
2310 const Function *CallerF = DAG.getMachineFunction().getFunction();
2311 CallingConv::ID CallerCC = CallerF->getCallingConv();
2312 bool CCMatch = CallerCC == CalleeCC;
2314 if (GuaranteedTailCallOpt) {
2315 if (IsTailCallConvention(CalleeCC) && CCMatch)
2320 // Look for obvious safe cases to perform tail call optimization that do not
2321 // require ABI changes. This is what gcc calls sibcall.
2323 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2324 // emit a special epilogue.
2325 if (RegInfo->needsStackRealignment(MF))
2328 // Do not sibcall optimize vararg calls unless the call site is not passing any
2330 if (isVarArg && !Outs.empty())
2333 // Also avoid sibcall optimization if either caller or callee uses struct
2334 // return semantics.
2335 if (isCalleeStructRet || isCallerStructRet)
2338 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2339 // Therefore if it's not used by the call it is not safe to optimize this into
2341 bool Unused = false;
2342 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2349 SmallVector<CCValAssign, 16> RVLocs;
2350 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2351 RVLocs, *DAG.getContext());
2352 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2353 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2354 CCValAssign &VA = RVLocs[i];
2355 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2360 // If the calling conventions do not match, then we'd better make sure the
2361 // results are returned in the same way as what the caller expects.
2363 SmallVector<CCValAssign, 16> RVLocs1;
2364 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2365 RVLocs1, *DAG.getContext());
2366 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2368 SmallVector<CCValAssign, 16> RVLocs2;
2369 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2370 RVLocs2, *DAG.getContext());
2371 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2373 if (RVLocs1.size() != RVLocs2.size())
2375 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2376 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2378 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2380 if (RVLocs1[i].isRegLoc()) {
2381 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2384 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2390 // If the callee takes no arguments then go on to check the results of the
2392 if (!Outs.empty()) {
2393 // Check if stack adjustment is needed. For now, do not do this if any
2394 // argument is passed on the stack.
2395 SmallVector<CCValAssign, 16> ArgLocs;
2396 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2397 ArgLocs, *DAG.getContext());
2398 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2399 if (CCInfo.getNextStackOffset()) {
2400 MachineFunction &MF = DAG.getMachineFunction();
2401 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2403 if (Subtarget->isTargetWin64())
2404 // Win64 ABI has additional complications.
2407 // Check if the arguments are already laid out in the right way as
2408 // the caller's fixed stack objects.
2409 MachineFrameInfo *MFI = MF.getFrameInfo();
2410 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2411 const X86InstrInfo *TII =
2412 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2413 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2414 CCValAssign &VA = ArgLocs[i];
2415 SDValue Arg = Outs[i].Val;
2416 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2417 if (VA.getLocInfo() == CCValAssign::Indirect)
2419 if (!VA.isRegLoc()) {
2420 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2427 // If the tailcall address may be in a register, then make sure it's
2428 // possible to register allocate for it. In 32-bit, the call address can
2429 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2430 // callee-saved registers are restored. In 64-bit, it's RAX, RCX, RDX, RSI,
2431 // RDI, R8, R9, R11.
2432 if (!isa<GlobalAddressSDNode>(Callee) &&
2433 !isa<ExternalSymbolSDNode>(Callee)) {
2434 unsigned Limit = Subtarget->is64Bit() ? 8 : 3;
2435 unsigned NumInRegs = 0;
2436 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2437 CCValAssign &VA = ArgLocs[i];
2438 if (VA.isRegLoc()) {
2439 if (++NumInRegs == Limit)
2450 X86TargetLowering::createFastISel(MachineFunction &mf,
2451 DenseMap<const Value *, unsigned> &vm,
2452 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2453 DenseMap<const AllocaInst *, int> &am,
2454 std::vector<std::pair<MachineInstr*, unsigned> > &pn
2456 , SmallSet<const Instruction *, 8> &cil
2459 return X86::createFastISel(mf, vm, bm, am, pn
2467 //===----------------------------------------------------------------------===//
2468 // Other Lowering Hooks
2469 //===----------------------------------------------------------------------===//
2472 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2473 MachineFunction &MF = DAG.getMachineFunction();
2474 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2475 int ReturnAddrIndex = FuncInfo->getRAIndex();
2477 if (ReturnAddrIndex == 0) {
2478 // Set up a frame object for the return address.
2479 uint64_t SlotSize = TD->getPointerSize();
2480 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2482 FuncInfo->setRAIndex(ReturnAddrIndex);
2485 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2489 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2490 bool hasSymbolicDisplacement) {
2491 // Offset should fit into 32 bit immediate field.
2492 if (!isInt<32>(Offset))
2495 // If we don't have a symbolic displacement - we don't have any extra
2497 if (!hasSymbolicDisplacement)
2500 // FIXME: Some tweaks might be needed for medium code model.
2501 if (M != CodeModel::Small && M != CodeModel::Kernel)
2504 // For small code model we assume that latest object is 16MB before end of 31
2505 // bits boundary. We may also accept pretty large negative constants knowing
2506 // that all objects are in the positive half of address space.
2507 if (M == CodeModel::Small && Offset < 16*1024*1024)
2510 // For kernel code model we know that all object resist in the negative half
2511 // of 32bits address space. We may not accept negative offsets, since they may
2512 // be just off and we may accept pretty large positive ones.
2513 if (M == CodeModel::Kernel && Offset > 0)
2519 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2520 /// specific condition code, returning the condition code and the LHS/RHS of the
2521 /// comparison to make.
2522 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2523 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2525 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2526 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2527 // X > -1 -> X == 0, jump !sign.
2528 RHS = DAG.getConstant(0, RHS.getValueType());
2529 return X86::COND_NS;
2530 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2531 // X < 0 -> X == 0, jump on sign.
2533 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2535 RHS = DAG.getConstant(0, RHS.getValueType());
2536 return X86::COND_LE;
2540 switch (SetCCOpcode) {
2541 default: llvm_unreachable("Invalid integer condition!");
2542 case ISD::SETEQ: return X86::COND_E;
2543 case ISD::SETGT: return X86::COND_G;
2544 case ISD::SETGE: return X86::COND_GE;
2545 case ISD::SETLT: return X86::COND_L;
2546 case ISD::SETLE: return X86::COND_LE;
2547 case ISD::SETNE: return X86::COND_NE;
2548 case ISD::SETULT: return X86::COND_B;
2549 case ISD::SETUGT: return X86::COND_A;
2550 case ISD::SETULE: return X86::COND_BE;
2551 case ISD::SETUGE: return X86::COND_AE;
2555 // First determine if it is required or is profitable to flip the operands.
2557 // If LHS is a foldable load, but RHS is not, flip the condition.
2558 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2559 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2560 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2561 std::swap(LHS, RHS);
2564 switch (SetCCOpcode) {
2570 std::swap(LHS, RHS);
2574 // On a floating point condition, the flags are set as follows:
2576 // 0 | 0 | 0 | X > Y
2577 // 0 | 0 | 1 | X < Y
2578 // 1 | 0 | 0 | X == Y
2579 // 1 | 1 | 1 | unordered
2580 switch (SetCCOpcode) {
2581 default: llvm_unreachable("Condcode should be pre-legalized away");
2583 case ISD::SETEQ: return X86::COND_E;
2584 case ISD::SETOLT: // flipped
2586 case ISD::SETGT: return X86::COND_A;
2587 case ISD::SETOLE: // flipped
2589 case ISD::SETGE: return X86::COND_AE;
2590 case ISD::SETUGT: // flipped
2592 case ISD::SETLT: return X86::COND_B;
2593 case ISD::SETUGE: // flipped
2595 case ISD::SETLE: return X86::COND_BE;
2597 case ISD::SETNE: return X86::COND_NE;
2598 case ISD::SETUO: return X86::COND_P;
2599 case ISD::SETO: return X86::COND_NP;
2601 case ISD::SETUNE: return X86::COND_INVALID;
2605 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2606 /// code. Current x86 isa includes the following FP cmov instructions:
2607 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2608 static bool hasFPCMov(unsigned X86CC) {
2624 /// isFPImmLegal - Returns true if the target can instruction select the
2625 /// specified FP immediate natively. If false, the legalizer will
2626 /// materialize the FP immediate as a load from a constant pool.
2627 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2628 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2629 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2635 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2636 /// the specified range (L, H].
2637 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2638 return (Val < 0) || (Val >= Low && Val < Hi);
2641 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2642 /// specified value.
2643 static bool isUndefOrEqual(int Val, int CmpVal) {
2644 if (Val < 0 || Val == CmpVal)
2649 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2650 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2651 /// the second operand.
2652 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2653 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2654 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2655 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2656 return (Mask[0] < 2 && Mask[1] < 2);
2660 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2661 SmallVector<int, 8> M;
2663 return ::isPSHUFDMask(M, N->getValueType(0));
2666 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2667 /// is suitable for input to PSHUFHW.
2668 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2669 if (VT != MVT::v8i16)
2672 // Lower quadword copied in order or undef.
2673 for (int i = 0; i != 4; ++i)
2674 if (Mask[i] >= 0 && Mask[i] != i)
2677 // Upper quadword shuffled.
2678 for (int i = 4; i != 8; ++i)
2679 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2685 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2686 SmallVector<int, 8> M;
2688 return ::isPSHUFHWMask(M, N->getValueType(0));
2691 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2692 /// is suitable for input to PSHUFLW.
2693 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2694 if (VT != MVT::v8i16)
2697 // Upper quadword copied in order.
2698 for (int i = 4; i != 8; ++i)
2699 if (Mask[i] >= 0 && Mask[i] != i)
2702 // Lower quadword shuffled.
2703 for (int i = 0; i != 4; ++i)
2710 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2711 SmallVector<int, 8> M;
2713 return ::isPSHUFLWMask(M, N->getValueType(0));
2716 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2717 /// is suitable for input to PALIGNR.
2718 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2720 int i, e = VT.getVectorNumElements();
2722 // Do not handle v2i64 / v2f64 shuffles with palignr.
2723 if (e < 4 || !hasSSSE3)
2726 for (i = 0; i != e; ++i)
2730 // All undef, not a palignr.
2734 // Determine if it's ok to perform a palignr with only the LHS, since we
2735 // don't have access to the actual shuffle elements to see if RHS is undef.
2736 bool Unary = Mask[i] < (int)e;
2737 bool NeedsUnary = false;
2739 int s = Mask[i] - i;
2741 // Check the rest of the elements to see if they are consecutive.
2742 for (++i; i != e; ++i) {
2747 Unary = Unary && (m < (int)e);
2748 NeedsUnary = NeedsUnary || (m < s);
2750 if (NeedsUnary && !Unary)
2752 if (Unary && m != ((s+i) & (e-1)))
2754 if (!Unary && m != (s+i))
2760 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2761 SmallVector<int, 8> M;
2763 return ::isPALIGNRMask(M, N->getValueType(0), true);
2766 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2767 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2768 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2769 int NumElems = VT.getVectorNumElements();
2770 if (NumElems != 2 && NumElems != 4)
2773 int Half = NumElems / 2;
2774 for (int i = 0; i < Half; ++i)
2775 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2777 for (int i = Half; i < NumElems; ++i)
2778 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2784 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2785 SmallVector<int, 8> M;
2787 return ::isSHUFPMask(M, N->getValueType(0));
2790 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2791 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2792 /// half elements to come from vector 1 (which would equal the dest.) and
2793 /// the upper half to come from vector 2.
2794 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2795 int NumElems = VT.getVectorNumElements();
2797 if (NumElems != 2 && NumElems != 4)
2800 int Half = NumElems / 2;
2801 for (int i = 0; i < Half; ++i)
2802 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2804 for (int i = Half; i < NumElems; ++i)
2805 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2810 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2811 SmallVector<int, 8> M;
2813 return isCommutedSHUFPMask(M, N->getValueType(0));
2816 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2817 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2818 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2819 if (N->getValueType(0).getVectorNumElements() != 4)
2822 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2823 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2824 isUndefOrEqual(N->getMaskElt(1), 7) &&
2825 isUndefOrEqual(N->getMaskElt(2), 2) &&
2826 isUndefOrEqual(N->getMaskElt(3), 3);
2829 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2830 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2832 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2833 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2838 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2839 isUndefOrEqual(N->getMaskElt(1), 3) &&
2840 isUndefOrEqual(N->getMaskElt(2), 2) &&
2841 isUndefOrEqual(N->getMaskElt(3), 3);
2844 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2845 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2846 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2847 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2849 if (NumElems != 2 && NumElems != 4)
2852 for (unsigned i = 0; i < NumElems/2; ++i)
2853 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2856 for (unsigned i = NumElems/2; i < NumElems; ++i)
2857 if (!isUndefOrEqual(N->getMaskElt(i), i))
2863 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2864 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2865 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2866 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2868 if (NumElems != 2 && NumElems != 4)
2871 for (unsigned i = 0; i < NumElems/2; ++i)
2872 if (!isUndefOrEqual(N->getMaskElt(i), i))
2875 for (unsigned i = 0; i < NumElems/2; ++i)
2876 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2882 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2883 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2884 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2885 bool V2IsSplat = false) {
2886 int NumElts = VT.getVectorNumElements();
2887 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2890 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2892 int BitI1 = Mask[i+1];
2893 if (!isUndefOrEqual(BitI, j))
2896 if (!isUndefOrEqual(BitI1, NumElts))
2899 if (!isUndefOrEqual(BitI1, j + NumElts))
2906 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2907 SmallVector<int, 8> M;
2909 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2912 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2913 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2914 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2915 bool V2IsSplat = false) {
2916 int NumElts = VT.getVectorNumElements();
2917 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2920 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2922 int BitI1 = Mask[i+1];
2923 if (!isUndefOrEqual(BitI, j + NumElts/2))
2926 if (isUndefOrEqual(BitI1, NumElts))
2929 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2936 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2937 SmallVector<int, 8> M;
2939 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2942 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2943 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2945 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2946 int NumElems = VT.getVectorNumElements();
2947 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2950 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2952 int BitI1 = Mask[i+1];
2953 if (!isUndefOrEqual(BitI, j))
2955 if (!isUndefOrEqual(BitI1, j))
2961 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2962 SmallVector<int, 8> M;
2964 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2967 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2968 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2970 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2971 int NumElems = VT.getVectorNumElements();
2972 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2975 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2977 int BitI1 = Mask[i+1];
2978 if (!isUndefOrEqual(BitI, j))
2980 if (!isUndefOrEqual(BitI1, j))
2986 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2987 SmallVector<int, 8> M;
2989 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2992 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2993 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2994 /// MOVSD, and MOVD, i.e. setting the lowest element.
2995 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2996 if (VT.getVectorElementType().getSizeInBits() < 32)
2999 int NumElts = VT.getVectorNumElements();
3001 if (!isUndefOrEqual(Mask[0], NumElts))
3004 for (int i = 1; i < NumElts; ++i)
3005 if (!isUndefOrEqual(Mask[i], i))
3011 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3012 SmallVector<int, 8> M;
3014 return ::isMOVLMask(M, N->getValueType(0));
3017 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3018 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3019 /// element of vector 2 and the other elements to come from vector 1 in order.
3020 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3021 bool V2IsSplat = false, bool V2IsUndef = false) {
3022 int NumOps = VT.getVectorNumElements();
3023 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3026 if (!isUndefOrEqual(Mask[0], 0))
3029 for (int i = 1; i < NumOps; ++i)
3030 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3031 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3032 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3038 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3039 bool V2IsUndef = false) {
3040 SmallVector<int, 8> M;
3042 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3045 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3046 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3047 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3048 if (N->getValueType(0).getVectorNumElements() != 4)
3051 // Expect 1, 1, 3, 3
3052 for (unsigned i = 0; i < 2; ++i) {
3053 int Elt = N->getMaskElt(i);
3054 if (Elt >= 0 && Elt != 1)
3059 for (unsigned i = 2; i < 4; ++i) {
3060 int Elt = N->getMaskElt(i);
3061 if (Elt >= 0 && Elt != 3)
3066 // Don't use movshdup if it can be done with a shufps.
3067 // FIXME: verify that matching u, u, 3, 3 is what we want.
3071 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3072 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3073 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3074 if (N->getValueType(0).getVectorNumElements() != 4)
3077 // Expect 0, 0, 2, 2
3078 for (unsigned i = 0; i < 2; ++i)
3079 if (N->getMaskElt(i) > 0)
3083 for (unsigned i = 2; i < 4; ++i) {
3084 int Elt = N->getMaskElt(i);
3085 if (Elt >= 0 && Elt != 2)
3090 // Don't use movsldup if it can be done with a shufps.
3094 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3095 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3096 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3097 int e = N->getValueType(0).getVectorNumElements() / 2;
3099 for (int i = 0; i < e; ++i)
3100 if (!isUndefOrEqual(N->getMaskElt(i), i))
3102 for (int i = 0; i < e; ++i)
3103 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3108 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3109 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3110 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3111 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3112 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3114 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3116 for (int i = 0; i < NumOperands; ++i) {
3117 int Val = SVOp->getMaskElt(NumOperands-i-1);
3118 if (Val < 0) Val = 0;
3119 if (Val >= NumOperands) Val -= NumOperands;
3121 if (i != NumOperands - 1)
3127 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3128 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3129 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3130 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3132 // 8 nodes, but we only care about the last 4.
3133 for (unsigned i = 7; i >= 4; --i) {
3134 int Val = SVOp->getMaskElt(i);
3143 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3144 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3145 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3146 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3148 // 8 nodes, but we only care about the first 4.
3149 for (int i = 3; i >= 0; --i) {
3150 int Val = SVOp->getMaskElt(i);
3159 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3160 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3161 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3162 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3163 EVT VVT = N->getValueType(0);
3164 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3168 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3169 Val = SVOp->getMaskElt(i);
3173 return (Val - i) * EltSize;
3176 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3178 bool X86::isZeroNode(SDValue Elt) {
3179 return ((isa<ConstantSDNode>(Elt) &&
3180 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3181 (isa<ConstantFPSDNode>(Elt) &&
3182 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3185 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3186 /// their permute mask.
3187 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3188 SelectionDAG &DAG) {
3189 EVT VT = SVOp->getValueType(0);
3190 unsigned NumElems = VT.getVectorNumElements();
3191 SmallVector<int, 8> MaskVec;
3193 for (unsigned i = 0; i != NumElems; ++i) {
3194 int idx = SVOp->getMaskElt(i);
3196 MaskVec.push_back(idx);
3197 else if (idx < (int)NumElems)
3198 MaskVec.push_back(idx + NumElems);
3200 MaskVec.push_back(idx - NumElems);
3202 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3203 SVOp->getOperand(0), &MaskVec[0]);
3206 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3207 /// the two vector operands have swapped position.
3208 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3209 unsigned NumElems = VT.getVectorNumElements();
3210 for (unsigned i = 0; i != NumElems; ++i) {
3214 else if (idx < (int)NumElems)
3215 Mask[i] = idx + NumElems;
3217 Mask[i] = idx - NumElems;
3221 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3222 /// match movhlps. The lower half elements should come from upper half of
3223 /// V1 (and in order), and the upper half elements should come from the upper
3224 /// half of V2 (and in order).
3225 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3226 if (Op->getValueType(0).getVectorNumElements() != 4)
3228 for (unsigned i = 0, e = 2; i != e; ++i)
3229 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3231 for (unsigned i = 2; i != 4; ++i)
3232 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3237 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3238 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3240 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3241 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3243 N = N->getOperand(0).getNode();
3244 if (!ISD::isNON_EXTLoad(N))
3247 *LD = cast<LoadSDNode>(N);
3251 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3252 /// match movlp{s|d}. The lower half elements should come from lower half of
3253 /// V1 (and in order), and the upper half elements should come from the upper
3254 /// half of V2 (and in order). And since V1 will become the source of the
3255 /// MOVLP, it must be either a vector load or a scalar load to vector.
3256 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3257 ShuffleVectorSDNode *Op) {
3258 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3260 // Is V2 is a vector load, don't do this transformation. We will try to use
3261 // load folding shufps op.
3262 if (ISD::isNON_EXTLoad(V2))
3265 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3267 if (NumElems != 2 && NumElems != 4)
3269 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3270 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3272 for (unsigned i = NumElems/2; i != NumElems; ++i)
3273 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3278 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3280 static bool isSplatVector(SDNode *N) {
3281 if (N->getOpcode() != ISD::BUILD_VECTOR)
3284 SDValue SplatValue = N->getOperand(0);
3285 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3286 if (N->getOperand(i) != SplatValue)
3291 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3292 /// to an zero vector.
3293 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3294 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3295 SDValue V1 = N->getOperand(0);
3296 SDValue V2 = N->getOperand(1);
3297 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3298 for (unsigned i = 0; i != NumElems; ++i) {
3299 int Idx = N->getMaskElt(i);
3300 if (Idx >= (int)NumElems) {
3301 unsigned Opc = V2.getOpcode();
3302 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3304 if (Opc != ISD::BUILD_VECTOR ||
3305 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3307 } else if (Idx >= 0) {
3308 unsigned Opc = V1.getOpcode();
3309 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3311 if (Opc != ISD::BUILD_VECTOR ||
3312 !X86::isZeroNode(V1.getOperand(Idx)))
3319 /// getZeroVector - Returns a vector of specified type with all zero elements.
3321 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3323 assert(VT.isVector() && "Expected a vector type");
3325 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3326 // type. This ensures they get CSE'd.
3328 if (VT.getSizeInBits() == 64) { // MMX
3329 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3330 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3331 } else if (HasSSE2) { // SSE2
3332 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3333 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3335 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3336 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3338 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3341 /// getOnesVector - Returns a vector of specified type with all bits set.
3343 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3344 assert(VT.isVector() && "Expected a vector type");
3346 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3347 // type. This ensures they get CSE'd.
3348 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3350 if (VT.getSizeInBits() == 64) // MMX
3351 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3353 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3354 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3358 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3359 /// that point to V2 points to its first element.
3360 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3361 EVT VT = SVOp->getValueType(0);
3362 unsigned NumElems = VT.getVectorNumElements();
3364 bool Changed = false;
3365 SmallVector<int, 8> MaskVec;
3366 SVOp->getMask(MaskVec);
3368 for (unsigned i = 0; i != NumElems; ++i) {
3369 if (MaskVec[i] > (int)NumElems) {
3370 MaskVec[i] = NumElems;
3375 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3376 SVOp->getOperand(1), &MaskVec[0]);
3377 return SDValue(SVOp, 0);
3380 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3381 /// operation of specified width.
3382 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3384 unsigned NumElems = VT.getVectorNumElements();
3385 SmallVector<int, 8> Mask;
3386 Mask.push_back(NumElems);
3387 for (unsigned i = 1; i != NumElems; ++i)
3389 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3392 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3393 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3395 unsigned NumElems = VT.getVectorNumElements();
3396 SmallVector<int, 8> Mask;
3397 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3399 Mask.push_back(i + NumElems);
3401 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3404 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3405 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3407 unsigned NumElems = VT.getVectorNumElements();
3408 unsigned Half = NumElems/2;
3409 SmallVector<int, 8> Mask;
3410 for (unsigned i = 0; i != Half; ++i) {
3411 Mask.push_back(i + Half);
3412 Mask.push_back(i + NumElems + Half);
3414 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3417 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3418 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3420 if (SV->getValueType(0).getVectorNumElements() <= 4)
3421 return SDValue(SV, 0);
3423 EVT PVT = MVT::v4f32;
3424 EVT VT = SV->getValueType(0);
3425 DebugLoc dl = SV->getDebugLoc();
3426 SDValue V1 = SV->getOperand(0);
3427 int NumElems = VT.getVectorNumElements();
3428 int EltNo = SV->getSplatIndex();
3430 // unpack elements to the correct location
3431 while (NumElems > 4) {
3432 if (EltNo < NumElems/2) {
3433 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3435 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3436 EltNo -= NumElems/2;
3441 // Perform the splat.
3442 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3443 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3444 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3445 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3448 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3449 /// vector of zero or undef vector. This produces a shuffle where the low
3450 /// element of V2 is swizzled into the zero/undef vector, landing at element
3451 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3452 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3453 bool isZero, bool HasSSE2,
3454 SelectionDAG &DAG) {
3455 EVT VT = V2.getValueType();
3457 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3458 unsigned NumElems = VT.getVectorNumElements();
3459 SmallVector<int, 16> MaskVec;
3460 for (unsigned i = 0; i != NumElems; ++i)
3461 // If this is the insertion idx, put the low elt of V2 here.
3462 MaskVec.push_back(i == Idx ? NumElems : i);
3463 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3466 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3467 /// a shuffle that is zero.
3469 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3470 bool Low, SelectionDAG &DAG) {
3471 unsigned NumZeros = 0;
3472 for (int i = 0; i < NumElems; ++i) {
3473 unsigned Index = Low ? i : NumElems-i-1;
3474 int Idx = SVOp->getMaskElt(Index);
3479 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3480 if (Elt.getNode() && X86::isZeroNode(Elt))
3488 /// isVectorShift - Returns true if the shuffle can be implemented as a
3489 /// logical left or right shift of a vector.
3490 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3491 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3492 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3493 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3496 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3499 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3503 bool SeenV1 = false;
3504 bool SeenV2 = false;
3505 for (unsigned i = NumZeros; i < NumElems; ++i) {
3506 unsigned Val = isLeft ? (i - NumZeros) : i;
3507 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3510 unsigned Idx = (unsigned) Idx_;
3520 if (SeenV1 && SeenV2)
3523 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3529 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3531 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3532 unsigned NumNonZero, unsigned NumZero,
3534 const TargetLowering &TLI) {
3538 DebugLoc dl = Op.getDebugLoc();
3541 for (unsigned i = 0; i < 16; ++i) {
3542 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3543 if (ThisIsNonZero && First) {
3545 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3547 V = DAG.getUNDEF(MVT::v8i16);
3552 SDValue ThisElt(0, 0), LastElt(0, 0);
3553 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3554 if (LastIsNonZero) {
3555 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3556 MVT::i16, Op.getOperand(i-1));
3558 if (ThisIsNonZero) {
3559 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3560 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3561 ThisElt, DAG.getConstant(8, MVT::i8));
3563 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3567 if (ThisElt.getNode())
3568 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3569 DAG.getIntPtrConstant(i/2));
3573 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3576 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3578 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3579 unsigned NumNonZero, unsigned NumZero,
3581 const TargetLowering &TLI) {
3585 DebugLoc dl = Op.getDebugLoc();
3588 for (unsigned i = 0; i < 8; ++i) {
3589 bool isNonZero = (NonZeros & (1 << i)) != 0;
3593 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3595 V = DAG.getUNDEF(MVT::v8i16);
3598 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3599 MVT::v8i16, V, Op.getOperand(i),
3600 DAG.getIntPtrConstant(i));
3607 /// getVShift - Return a vector logical shift node.
3609 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3610 unsigned NumBits, SelectionDAG &DAG,
3611 const TargetLowering &TLI, DebugLoc dl) {
3612 bool isMMX = VT.getSizeInBits() == 64;
3613 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3614 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3615 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3616 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3617 DAG.getNode(Opc, dl, ShVT, SrcOp,
3618 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3622 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3623 SelectionDAG &DAG) const {
3625 // Check if the scalar load can be widened into a vector load. And if
3626 // the address is "base + cst" see if the cst can be "absorbed" into
3627 // the shuffle mask.
3628 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3629 SDValue Ptr = LD->getBasePtr();
3630 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3632 EVT PVT = LD->getValueType(0);
3633 if (PVT != MVT::i32 && PVT != MVT::f32)
3638 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3639 FI = FINode->getIndex();
3641 } else if (Ptr.getOpcode() == ISD::ADD &&
3642 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3643 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3644 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3645 Offset = Ptr.getConstantOperandVal(1);
3646 Ptr = Ptr.getOperand(0);
3651 SDValue Chain = LD->getChain();
3652 // Make sure the stack object alignment is at least 16.
3653 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3654 if (DAG.InferPtrAlignment(Ptr) < 16) {
3655 if (MFI->isFixedObjectIndex(FI)) {
3656 // Can't change the alignment. FIXME: It's possible to compute
3657 // the exact stack offset and reference FI + adjust offset instead.
3658 // If someone *really* cares about this. That's the way to implement it.
3661 MFI->setObjectAlignment(FI, 16);
3665 // (Offset % 16) must be multiple of 4. Then address is then
3666 // Ptr + (Offset & ~15).
3669 if ((Offset % 16) & 3)
3671 int64_t StartOffset = Offset & ~15;
3673 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3674 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3676 int EltNo = (Offset - StartOffset) >> 2;
3677 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3678 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3679 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3681 // Canonicalize it to a v4i32 shuffle.
3682 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3683 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3684 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3685 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3691 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3692 /// vector of type 'VT', see if the elements can be replaced by a single large
3693 /// load which has the same value as a build_vector whose operands are 'elts'.
3695 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3697 /// FIXME: we'd also like to handle the case where the last elements are zero
3698 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3699 /// There's even a handy isZeroNode for that purpose.
3700 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3701 DebugLoc &dl, SelectionDAG &DAG) {
3702 EVT EltVT = VT.getVectorElementType();
3703 unsigned NumElems = Elts.size();
3705 LoadSDNode *LDBase = NULL;
3706 unsigned LastLoadedElt = -1U;
3708 // For each element in the initializer, see if we've found a load or an undef.
3709 // If we don't find an initial load element, or later load elements are
3710 // non-consecutive, bail out.
3711 for (unsigned i = 0; i < NumElems; ++i) {
3712 SDValue Elt = Elts[i];
3714 if (!Elt.getNode() ||
3715 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3718 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3720 LDBase = cast<LoadSDNode>(Elt.getNode());
3724 if (Elt.getOpcode() == ISD::UNDEF)
3727 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3728 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3733 // If we have found an entire vector of loads and undefs, then return a large
3734 // load of the entire vector width starting at the base pointer. If we found
3735 // consecutive loads for the low half, generate a vzext_load node.
3736 if (LastLoadedElt == NumElems - 1) {
3737 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3738 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3739 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3740 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3741 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3742 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3743 LDBase->isVolatile(), LDBase->isNonTemporal(),
3744 LDBase->getAlignment());
3745 } else if (NumElems == 4 && LastLoadedElt == 1) {
3746 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3747 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3748 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3749 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3755 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3756 DebugLoc dl = Op.getDebugLoc();
3757 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3758 if (ISD::isBuildVectorAllZeros(Op.getNode())
3759 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3760 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3761 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3762 // eliminated on x86-32 hosts.
3763 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3766 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3767 return getOnesVector(Op.getValueType(), DAG, dl);
3768 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3771 EVT VT = Op.getValueType();
3772 EVT ExtVT = VT.getVectorElementType();
3773 unsigned EVTBits = ExtVT.getSizeInBits();
3775 unsigned NumElems = Op.getNumOperands();
3776 unsigned NumZero = 0;
3777 unsigned NumNonZero = 0;
3778 unsigned NonZeros = 0;
3779 bool IsAllConstants = true;
3780 SmallSet<SDValue, 8> Values;
3781 for (unsigned i = 0; i < NumElems; ++i) {
3782 SDValue Elt = Op.getOperand(i);
3783 if (Elt.getOpcode() == ISD::UNDEF)
3786 if (Elt.getOpcode() != ISD::Constant &&
3787 Elt.getOpcode() != ISD::ConstantFP)
3788 IsAllConstants = false;
3789 if (X86::isZeroNode(Elt))
3792 NonZeros |= (1 << i);
3797 if (NumNonZero == 0) {
3798 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3799 return DAG.getUNDEF(VT);
3802 // Special case for single non-zero, non-undef, element.
3803 if (NumNonZero == 1) {
3804 unsigned Idx = CountTrailingZeros_32(NonZeros);
3805 SDValue Item = Op.getOperand(Idx);
3807 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3808 // the value are obviously zero, truncate the value to i32 and do the
3809 // insertion that way. Only do this if the value is non-constant or if the
3810 // value is a constant being inserted into element 0. It is cheaper to do
3811 // a constant pool load than it is to do a movd + shuffle.
3812 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3813 (!IsAllConstants || Idx == 0)) {
3814 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3815 // Handle MMX and SSE both.
3816 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3817 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3819 // Truncate the value (which may itself be a constant) to i32, and
3820 // convert it to a vector with movd (S2V+shuffle to zero extend).
3821 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3822 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3823 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3824 Subtarget->hasSSE2(), DAG);
3826 // Now we have our 32-bit value zero extended in the low element of
3827 // a vector. If Idx != 0, swizzle it into place.
3829 SmallVector<int, 4> Mask;
3830 Mask.push_back(Idx);
3831 for (unsigned i = 1; i != VecElts; ++i)
3833 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3834 DAG.getUNDEF(Item.getValueType()),
3837 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3841 // If we have a constant or non-constant insertion into the low element of
3842 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3843 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3844 // depending on what the source datatype is.
3847 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3848 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3849 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3850 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3851 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3852 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3854 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3855 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3856 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3857 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3858 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3859 Subtarget->hasSSE2(), DAG);
3860 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3864 // Is it a vector logical left shift?
3865 if (NumElems == 2 && Idx == 1 &&
3866 X86::isZeroNode(Op.getOperand(0)) &&
3867 !X86::isZeroNode(Op.getOperand(1))) {
3868 unsigned NumBits = VT.getSizeInBits();
3869 return getVShift(true, VT,
3870 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3871 VT, Op.getOperand(1)),
3872 NumBits/2, DAG, *this, dl);
3875 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3878 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3879 // is a non-constant being inserted into an element other than the low one,
3880 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3881 // movd/movss) to move this into the low element, then shuffle it into
3883 if (EVTBits == 32) {
3884 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3886 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3887 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3888 Subtarget->hasSSE2(), DAG);
3889 SmallVector<int, 8> MaskVec;
3890 for (unsigned i = 0; i < NumElems; i++)
3891 MaskVec.push_back(i == Idx ? 0 : 1);
3892 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3896 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3897 if (Values.size() == 1) {
3898 if (EVTBits == 32) {
3899 // Instead of a shuffle like this:
3900 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3901 // Check if it's possible to issue this instead.
3902 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3903 unsigned Idx = CountTrailingZeros_32(NonZeros);
3904 SDValue Item = Op.getOperand(Idx);
3905 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3906 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3911 // A vector full of immediates; various special cases are already
3912 // handled, so this is best done with a single constant-pool load.
3916 // Let legalizer expand 2-wide build_vectors.
3917 if (EVTBits == 64) {
3918 if (NumNonZero == 1) {
3919 // One half is zero or undef.
3920 unsigned Idx = CountTrailingZeros_32(NonZeros);
3921 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3922 Op.getOperand(Idx));
3923 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3924 Subtarget->hasSSE2(), DAG);
3929 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3930 if (EVTBits == 8 && NumElems == 16) {
3931 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3933 if (V.getNode()) return V;
3936 if (EVTBits == 16 && NumElems == 8) {
3937 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3939 if (V.getNode()) return V;
3942 // If element VT is == 32 bits, turn it into a number of shuffles.
3943 SmallVector<SDValue, 8> V;
3945 if (NumElems == 4 && NumZero > 0) {
3946 for (unsigned i = 0; i < 4; ++i) {
3947 bool isZero = !(NonZeros & (1 << i));
3949 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3951 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3954 for (unsigned i = 0; i < 2; ++i) {
3955 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3958 V[i] = V[i*2]; // Must be a zero vector.
3961 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3964 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3967 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3972 SmallVector<int, 8> MaskVec;
3973 bool Reverse = (NonZeros & 0x3) == 2;
3974 for (unsigned i = 0; i < 2; ++i)
3975 MaskVec.push_back(Reverse ? 1-i : i);
3976 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3977 for (unsigned i = 0; i < 2; ++i)
3978 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3979 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3982 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
3983 // Check for a build vector of consecutive loads.
3984 for (unsigned i = 0; i < NumElems; ++i)
3985 V[i] = Op.getOperand(i);
3987 // Check for elements which are consecutive loads.
3988 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
3992 // For SSE 4.1, use inserts into undef.
3993 if (getSubtarget()->hasSSE41()) {
3994 V[0] = DAG.getUNDEF(VT);
3995 for (unsigned i = 0; i < NumElems; ++i)
3996 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3997 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3998 Op.getOperand(i), DAG.getIntPtrConstant(i));
4002 // Otherwise, expand into a number of unpckl*
4004 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4005 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4006 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4007 for (unsigned i = 0; i < NumElems; ++i)
4008 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4010 while (NumElems != 0) {
4011 for (unsigned i = 0; i < NumElems; ++i)
4012 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4021 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4022 // We support concatenate two MMX registers and place them in a MMX
4023 // register. This is better than doing a stack convert.
4024 DebugLoc dl = Op.getDebugLoc();
4025 EVT ResVT = Op.getValueType();
4026 assert(Op.getNumOperands() == 2);
4027 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4028 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4030 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4031 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4032 InVec = Op.getOperand(1);
4033 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4034 unsigned NumElts = ResVT.getVectorNumElements();
4035 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4036 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4037 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4039 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4040 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4041 Mask[0] = 0; Mask[1] = 2;
4042 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4044 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4047 // v8i16 shuffles - Prefer shuffles in the following order:
4048 // 1. [all] pshuflw, pshufhw, optional move
4049 // 2. [ssse3] 1 x pshufb
4050 // 3. [ssse3] 2 x pshufb + 1 x por
4051 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4053 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4055 const X86TargetLowering &TLI) {
4056 SDValue V1 = SVOp->getOperand(0);
4057 SDValue V2 = SVOp->getOperand(1);
4058 DebugLoc dl = SVOp->getDebugLoc();
4059 SmallVector<int, 8> MaskVals;
4061 // Determine if more than 1 of the words in each of the low and high quadwords
4062 // of the result come from the same quadword of one of the two inputs. Undef
4063 // mask values count as coming from any quadword, for better codegen.
4064 SmallVector<unsigned, 4> LoQuad(4);
4065 SmallVector<unsigned, 4> HiQuad(4);
4066 BitVector InputQuads(4);
4067 for (unsigned i = 0; i < 8; ++i) {
4068 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4069 int EltIdx = SVOp->getMaskElt(i);
4070 MaskVals.push_back(EltIdx);
4079 InputQuads.set(EltIdx / 4);
4082 int BestLoQuad = -1;
4083 unsigned MaxQuad = 1;
4084 for (unsigned i = 0; i < 4; ++i) {
4085 if (LoQuad[i] > MaxQuad) {
4087 MaxQuad = LoQuad[i];
4091 int BestHiQuad = -1;
4093 for (unsigned i = 0; i < 4; ++i) {
4094 if (HiQuad[i] > MaxQuad) {
4096 MaxQuad = HiQuad[i];
4100 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4101 // of the two input vectors, shuffle them into one input vector so only a
4102 // single pshufb instruction is necessary. If There are more than 2 input
4103 // quads, disable the next transformation since it does not help SSSE3.
4104 bool V1Used = InputQuads[0] || InputQuads[1];
4105 bool V2Used = InputQuads[2] || InputQuads[3];
4106 if (TLI.getSubtarget()->hasSSSE3()) {
4107 if (InputQuads.count() == 2 && V1Used && V2Used) {
4108 BestLoQuad = InputQuads.find_first();
4109 BestHiQuad = InputQuads.find_next(BestLoQuad);
4111 if (InputQuads.count() > 2) {
4117 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4118 // the shuffle mask. If a quad is scored as -1, that means that it contains
4119 // words from all 4 input quadwords.
4121 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4122 SmallVector<int, 8> MaskV;
4123 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4124 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4125 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4126 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4127 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4128 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4130 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4131 // source words for the shuffle, to aid later transformations.
4132 bool AllWordsInNewV = true;
4133 bool InOrder[2] = { true, true };
4134 for (unsigned i = 0; i != 8; ++i) {
4135 int idx = MaskVals[i];
4137 InOrder[i/4] = false;
4138 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4140 AllWordsInNewV = false;
4144 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4145 if (AllWordsInNewV) {
4146 for (int i = 0; i != 8; ++i) {
4147 int idx = MaskVals[i];
4150 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4151 if ((idx != i) && idx < 4)
4153 if ((idx != i) && idx > 3)
4162 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4163 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4164 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4165 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4166 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4170 // If we have SSSE3, and all words of the result are from 1 input vector,
4171 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4172 // is present, fall back to case 4.
4173 if (TLI.getSubtarget()->hasSSSE3()) {
4174 SmallVector<SDValue,16> pshufbMask;
4176 // If we have elements from both input vectors, set the high bit of the
4177 // shuffle mask element to zero out elements that come from V2 in the V1
4178 // mask, and elements that come from V1 in the V2 mask, so that the two
4179 // results can be OR'd together.
4180 bool TwoInputs = V1Used && V2Used;
4181 for (unsigned i = 0; i != 8; ++i) {
4182 int EltIdx = MaskVals[i] * 2;
4183 if (TwoInputs && (EltIdx >= 16)) {
4184 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4185 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4188 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4189 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4191 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4192 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4193 DAG.getNode(ISD::BUILD_VECTOR, dl,
4194 MVT::v16i8, &pshufbMask[0], 16));
4196 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4198 // Calculate the shuffle mask for the second input, shuffle it, and
4199 // OR it with the first shuffled input.
4201 for (unsigned i = 0; i != 8; ++i) {
4202 int EltIdx = MaskVals[i] * 2;
4204 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4205 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4208 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4209 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4211 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4212 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4213 DAG.getNode(ISD::BUILD_VECTOR, dl,
4214 MVT::v16i8, &pshufbMask[0], 16));
4215 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4216 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4219 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4220 // and update MaskVals with new element order.
4221 BitVector InOrder(8);
4222 if (BestLoQuad >= 0) {
4223 SmallVector<int, 8> MaskV;
4224 for (int i = 0; i != 4; ++i) {
4225 int idx = MaskVals[i];
4227 MaskV.push_back(-1);
4229 } else if ((idx / 4) == BestLoQuad) {
4230 MaskV.push_back(idx & 3);
4233 MaskV.push_back(-1);
4236 for (unsigned i = 4; i != 8; ++i)
4238 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4242 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4243 // and update MaskVals with the new element order.
4244 if (BestHiQuad >= 0) {
4245 SmallVector<int, 8> MaskV;
4246 for (unsigned i = 0; i != 4; ++i)
4248 for (unsigned i = 4; i != 8; ++i) {
4249 int idx = MaskVals[i];
4251 MaskV.push_back(-1);
4253 } else if ((idx / 4) == BestHiQuad) {
4254 MaskV.push_back((idx & 3) + 4);
4257 MaskV.push_back(-1);
4260 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4264 // In case BestHi & BestLo were both -1, which means each quadword has a word
4265 // from each of the four input quadwords, calculate the InOrder bitvector now
4266 // before falling through to the insert/extract cleanup.
4267 if (BestLoQuad == -1 && BestHiQuad == -1) {
4269 for (int i = 0; i != 8; ++i)
4270 if (MaskVals[i] < 0 || MaskVals[i] == i)
4274 // The other elements are put in the right place using pextrw and pinsrw.
4275 for (unsigned i = 0; i != 8; ++i) {
4278 int EltIdx = MaskVals[i];
4281 SDValue ExtOp = (EltIdx < 8)
4282 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4283 DAG.getIntPtrConstant(EltIdx))
4284 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4285 DAG.getIntPtrConstant(EltIdx - 8));
4286 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4287 DAG.getIntPtrConstant(i));
4292 // v16i8 shuffles - Prefer shuffles in the following order:
4293 // 1. [ssse3] 1 x pshufb
4294 // 2. [ssse3] 2 x pshufb + 1 x por
4295 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4297 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4299 const X86TargetLowering &TLI) {
4300 SDValue V1 = SVOp->getOperand(0);
4301 SDValue V2 = SVOp->getOperand(1);
4302 DebugLoc dl = SVOp->getDebugLoc();
4303 SmallVector<int, 16> MaskVals;
4304 SVOp->getMask(MaskVals);
4306 // If we have SSSE3, case 1 is generated when all result bytes come from
4307 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4308 // present, fall back to case 3.
4309 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4312 for (unsigned i = 0; i < 16; ++i) {
4313 int EltIdx = MaskVals[i];
4322 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4323 if (TLI.getSubtarget()->hasSSSE3()) {
4324 SmallVector<SDValue,16> pshufbMask;
4326 // If all result elements are from one input vector, then only translate
4327 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4329 // Otherwise, we have elements from both input vectors, and must zero out
4330 // elements that come from V2 in the first mask, and V1 in the second mask
4331 // so that we can OR them together.
4332 bool TwoInputs = !(V1Only || V2Only);
4333 for (unsigned i = 0; i != 16; ++i) {
4334 int EltIdx = MaskVals[i];
4335 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4336 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4339 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4341 // If all the elements are from V2, assign it to V1 and return after
4342 // building the first pshufb.
4345 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4346 DAG.getNode(ISD::BUILD_VECTOR, dl,
4347 MVT::v16i8, &pshufbMask[0], 16));
4351 // Calculate the shuffle mask for the second input, shuffle it, and
4352 // OR it with the first shuffled input.
4354 for (unsigned i = 0; i != 16; ++i) {
4355 int EltIdx = MaskVals[i];
4357 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4360 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4362 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4363 DAG.getNode(ISD::BUILD_VECTOR, dl,
4364 MVT::v16i8, &pshufbMask[0], 16));
4365 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4368 // No SSSE3 - Calculate in place words and then fix all out of place words
4369 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4370 // the 16 different words that comprise the two doublequadword input vectors.
4371 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4372 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4373 SDValue NewV = V2Only ? V2 : V1;
4374 for (int i = 0; i != 8; ++i) {
4375 int Elt0 = MaskVals[i*2];
4376 int Elt1 = MaskVals[i*2+1];
4378 // This word of the result is all undef, skip it.
4379 if (Elt0 < 0 && Elt1 < 0)
4382 // This word of the result is already in the correct place, skip it.
4383 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4385 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4388 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4389 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4392 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4393 // using a single extract together, load it and store it.
4394 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4395 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4396 DAG.getIntPtrConstant(Elt1 / 2));
4397 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4398 DAG.getIntPtrConstant(i));
4402 // If Elt1 is defined, extract it from the appropriate source. If the
4403 // source byte is not also odd, shift the extracted word left 8 bits
4404 // otherwise clear the bottom 8 bits if we need to do an or.
4406 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4407 DAG.getIntPtrConstant(Elt1 / 2));
4408 if ((Elt1 & 1) == 0)
4409 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4410 DAG.getConstant(8, TLI.getShiftAmountTy()));
4412 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4413 DAG.getConstant(0xFF00, MVT::i16));
4415 // If Elt0 is defined, extract it from the appropriate source. If the
4416 // source byte is not also even, shift the extracted word right 8 bits. If
4417 // Elt1 was also defined, OR the extracted values together before
4418 // inserting them in the result.
4420 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4421 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4422 if ((Elt0 & 1) != 0)
4423 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4424 DAG.getConstant(8, TLI.getShiftAmountTy()));
4426 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4427 DAG.getConstant(0x00FF, MVT::i16));
4428 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4431 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4432 DAG.getIntPtrConstant(i));
4434 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4437 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4438 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4439 /// done when every pair / quad of shuffle mask elements point to elements in
4440 /// the right sequence. e.g.
4441 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4443 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4445 const TargetLowering &TLI, DebugLoc dl) {
4446 EVT VT = SVOp->getValueType(0);
4447 SDValue V1 = SVOp->getOperand(0);
4448 SDValue V2 = SVOp->getOperand(1);
4449 unsigned NumElems = VT.getVectorNumElements();
4450 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4451 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4453 switch (VT.getSimpleVT().SimpleTy) {
4454 default: assert(false && "Unexpected!");
4455 case MVT::v4f32: NewVT = MVT::v2f64; break;
4456 case MVT::v4i32: NewVT = MVT::v2i64; break;
4457 case MVT::v8i16: NewVT = MVT::v4i32; break;
4458 case MVT::v16i8: NewVT = MVT::v4i32; break;
4461 if (NewWidth == 2) {
4467 int Scale = NumElems / NewWidth;
4468 SmallVector<int, 8> MaskVec;
4469 for (unsigned i = 0; i < NumElems; i += Scale) {
4471 for (int j = 0; j < Scale; ++j) {
4472 int EltIdx = SVOp->getMaskElt(i+j);
4476 StartIdx = EltIdx - (EltIdx % Scale);
4477 if (EltIdx != StartIdx + j)
4481 MaskVec.push_back(-1);
4483 MaskVec.push_back(StartIdx / Scale);
4486 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4487 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4488 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4491 /// getVZextMovL - Return a zero-extending vector move low node.
4493 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4494 SDValue SrcOp, SelectionDAG &DAG,
4495 const X86Subtarget *Subtarget, DebugLoc dl) {
4496 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4497 LoadSDNode *LD = NULL;
4498 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4499 LD = dyn_cast<LoadSDNode>(SrcOp);
4501 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4503 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4504 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4505 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4506 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4507 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4509 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4510 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4511 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4512 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4520 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4521 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4522 DAG.getNode(ISD::BIT_CONVERT, dl,
4526 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4529 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4530 SDValue V1 = SVOp->getOperand(0);
4531 SDValue V2 = SVOp->getOperand(1);
4532 DebugLoc dl = SVOp->getDebugLoc();
4533 EVT VT = SVOp->getValueType(0);
4535 SmallVector<std::pair<int, int>, 8> Locs;
4537 SmallVector<int, 8> Mask1(4U, -1);
4538 SmallVector<int, 8> PermMask;
4539 SVOp->getMask(PermMask);
4543 for (unsigned i = 0; i != 4; ++i) {
4544 int Idx = PermMask[i];
4546 Locs[i] = std::make_pair(-1, -1);
4548 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4550 Locs[i] = std::make_pair(0, NumLo);
4554 Locs[i] = std::make_pair(1, NumHi);
4556 Mask1[2+NumHi] = Idx;
4562 if (NumLo <= 2 && NumHi <= 2) {
4563 // If no more than two elements come from either vector. This can be
4564 // implemented with two shuffles. First shuffle gather the elements.
4565 // The second shuffle, which takes the first shuffle as both of its
4566 // vector operands, put the elements into the right order.
4567 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4569 SmallVector<int, 8> Mask2(4U, -1);
4571 for (unsigned i = 0; i != 4; ++i) {
4572 if (Locs[i].first == -1)
4575 unsigned Idx = (i < 2) ? 0 : 4;
4576 Idx += Locs[i].first * 2 + Locs[i].second;
4581 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4582 } else if (NumLo == 3 || NumHi == 3) {
4583 // Otherwise, we must have three elements from one vector, call it X, and
4584 // one element from the other, call it Y. First, use a shufps to build an
4585 // intermediate vector with the one element from Y and the element from X
4586 // that will be in the same half in the final destination (the indexes don't
4587 // matter). Then, use a shufps to build the final vector, taking the half
4588 // containing the element from Y from the intermediate, and the other half
4591 // Normalize it so the 3 elements come from V1.
4592 CommuteVectorShuffleMask(PermMask, VT);
4596 // Find the element from V2.
4598 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4599 int Val = PermMask[HiIndex];
4606 Mask1[0] = PermMask[HiIndex];
4608 Mask1[2] = PermMask[HiIndex^1];
4610 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4613 Mask1[0] = PermMask[0];
4614 Mask1[1] = PermMask[1];
4615 Mask1[2] = HiIndex & 1 ? 6 : 4;
4616 Mask1[3] = HiIndex & 1 ? 4 : 6;
4617 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4619 Mask1[0] = HiIndex & 1 ? 2 : 0;
4620 Mask1[1] = HiIndex & 1 ? 0 : 2;
4621 Mask1[2] = PermMask[2];
4622 Mask1[3] = PermMask[3];
4627 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4631 // Break it into (shuffle shuffle_hi, shuffle_lo).
4633 SmallVector<int,8> LoMask(4U, -1);
4634 SmallVector<int,8> HiMask(4U, -1);
4636 SmallVector<int,8> *MaskPtr = &LoMask;
4637 unsigned MaskIdx = 0;
4640 for (unsigned i = 0; i != 4; ++i) {
4647 int Idx = PermMask[i];
4649 Locs[i] = std::make_pair(-1, -1);
4650 } else if (Idx < 4) {
4651 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4652 (*MaskPtr)[LoIdx] = Idx;
4655 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4656 (*MaskPtr)[HiIdx] = Idx;
4661 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4662 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4663 SmallVector<int, 8> MaskOps;
4664 for (unsigned i = 0; i != 4; ++i) {
4665 if (Locs[i].first == -1) {
4666 MaskOps.push_back(-1);
4668 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4669 MaskOps.push_back(Idx);
4672 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4676 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4677 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4678 SDValue V1 = Op.getOperand(0);
4679 SDValue V2 = Op.getOperand(1);
4680 EVT VT = Op.getValueType();
4681 DebugLoc dl = Op.getDebugLoc();
4682 unsigned NumElems = VT.getVectorNumElements();
4683 bool isMMX = VT.getSizeInBits() == 64;
4684 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4685 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4686 bool V1IsSplat = false;
4687 bool V2IsSplat = false;
4689 if (isZeroShuffle(SVOp))
4690 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4692 // Promote splats to v4f32.
4693 if (SVOp->isSplat()) {
4694 if (isMMX || NumElems < 4)
4696 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4699 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4701 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4702 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4703 if (NewOp.getNode())
4704 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4705 LowerVECTOR_SHUFFLE(NewOp, DAG));
4706 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4707 // FIXME: Figure out a cleaner way to do this.
4708 // Try to make use of movq to zero out the top part.
4709 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4710 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4711 if (NewOp.getNode()) {
4712 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4713 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4714 DAG, Subtarget, dl);
4716 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4717 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4718 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4719 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4720 DAG, Subtarget, dl);
4724 if (X86::isPSHUFDMask(SVOp))
4727 // Check if this can be converted into a logical shift.
4728 bool isLeft = false;
4731 bool isShift = getSubtarget()->hasSSE2() &&
4732 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4733 if (isShift && ShVal.hasOneUse()) {
4734 // If the shifted value has multiple uses, it may be cheaper to use
4735 // v_set0 + movlhps or movhlps, etc.
4736 EVT EltVT = VT.getVectorElementType();
4737 ShAmt *= EltVT.getSizeInBits();
4738 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4741 if (X86::isMOVLMask(SVOp)) {
4744 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4745 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4750 // FIXME: fold these into legal mask.
4751 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4752 X86::isMOVSLDUPMask(SVOp) ||
4753 X86::isMOVHLPSMask(SVOp) ||
4754 X86::isMOVLHPSMask(SVOp) ||
4755 X86::isMOVLPMask(SVOp)))
4758 if (ShouldXformToMOVHLPS(SVOp) ||
4759 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4760 return CommuteVectorShuffle(SVOp, DAG);
4763 // No better options. Use a vshl / vsrl.
4764 EVT EltVT = VT.getVectorElementType();
4765 ShAmt *= EltVT.getSizeInBits();
4766 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4769 bool Commuted = false;
4770 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4771 // 1,1,1,1 -> v8i16 though.
4772 V1IsSplat = isSplatVector(V1.getNode());
4773 V2IsSplat = isSplatVector(V2.getNode());
4775 // Canonicalize the splat or undef, if present, to be on the RHS.
4776 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4777 Op = CommuteVectorShuffle(SVOp, DAG);
4778 SVOp = cast<ShuffleVectorSDNode>(Op);
4779 V1 = SVOp->getOperand(0);
4780 V2 = SVOp->getOperand(1);
4781 std::swap(V1IsSplat, V2IsSplat);
4782 std::swap(V1IsUndef, V2IsUndef);
4786 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4787 // Shuffling low element of v1 into undef, just return v1.
4790 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4791 // the instruction selector will not match, so get a canonical MOVL with
4792 // swapped operands to undo the commute.
4793 return getMOVL(DAG, dl, VT, V2, V1);
4796 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4797 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4798 X86::isUNPCKLMask(SVOp) ||
4799 X86::isUNPCKHMask(SVOp))
4803 // Normalize mask so all entries that point to V2 points to its first
4804 // element then try to match unpck{h|l} again. If match, return a
4805 // new vector_shuffle with the corrected mask.
4806 SDValue NewMask = NormalizeMask(SVOp, DAG);
4807 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4808 if (NSVOp != SVOp) {
4809 if (X86::isUNPCKLMask(NSVOp, true)) {
4811 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4818 // Commute is back and try unpck* again.
4819 // FIXME: this seems wrong.
4820 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4821 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4822 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4823 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4824 X86::isUNPCKLMask(NewSVOp) ||
4825 X86::isUNPCKHMask(NewSVOp))
4829 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4831 // Normalize the node to match x86 shuffle ops if needed
4832 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4833 return CommuteVectorShuffle(SVOp, DAG);
4835 // Check for legal shuffle and return?
4836 SmallVector<int, 16> PermMask;
4837 SVOp->getMask(PermMask);
4838 if (isShuffleMaskLegal(PermMask, VT))
4841 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4842 if (VT == MVT::v8i16) {
4843 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4844 if (NewOp.getNode())
4848 if (VT == MVT::v16i8) {
4849 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4850 if (NewOp.getNode())
4854 // Handle all 4 wide cases with a number of shuffles except for MMX.
4855 if (NumElems == 4 && !isMMX)
4856 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4862 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4863 SelectionDAG &DAG) const {
4864 EVT VT = Op.getValueType();
4865 DebugLoc dl = Op.getDebugLoc();
4866 if (VT.getSizeInBits() == 8) {
4867 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4868 Op.getOperand(0), Op.getOperand(1));
4869 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4870 DAG.getValueType(VT));
4871 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4872 } else if (VT.getSizeInBits() == 16) {
4873 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4874 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4876 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4877 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4878 DAG.getNode(ISD::BIT_CONVERT, dl,
4882 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4883 Op.getOperand(0), Op.getOperand(1));
4884 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4885 DAG.getValueType(VT));
4886 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4887 } else if (VT == MVT::f32) {
4888 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4889 // the result back to FR32 register. It's only worth matching if the
4890 // result has a single use which is a store or a bitcast to i32. And in
4891 // the case of a store, it's not worth it if the index is a constant 0,
4892 // because a MOVSSmr can be used instead, which is smaller and faster.
4893 if (!Op.hasOneUse())
4895 SDNode *User = *Op.getNode()->use_begin();
4896 if ((User->getOpcode() != ISD::STORE ||
4897 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4898 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4899 (User->getOpcode() != ISD::BIT_CONVERT ||
4900 User->getValueType(0) != MVT::i32))
4902 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4903 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4906 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4907 } else if (VT == MVT::i32) {
4908 // ExtractPS works with constant index.
4909 if (isa<ConstantSDNode>(Op.getOperand(1)))
4917 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4918 SelectionDAG &DAG) const {
4919 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4922 if (Subtarget->hasSSE41()) {
4923 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4928 EVT VT = Op.getValueType();
4929 DebugLoc dl = Op.getDebugLoc();
4930 // TODO: handle v16i8.
4931 if (VT.getSizeInBits() == 16) {
4932 SDValue Vec = Op.getOperand(0);
4933 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4935 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4936 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4937 DAG.getNode(ISD::BIT_CONVERT, dl,
4940 // Transform it so it match pextrw which produces a 32-bit result.
4941 EVT EltVT = MVT::i32;
4942 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4943 Op.getOperand(0), Op.getOperand(1));
4944 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4945 DAG.getValueType(VT));
4946 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4947 } else if (VT.getSizeInBits() == 32) {
4948 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4952 // SHUFPS the element to the lowest double word, then movss.
4953 int Mask[4] = { Idx, -1, -1, -1 };
4954 EVT VVT = Op.getOperand(0).getValueType();
4955 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4956 DAG.getUNDEF(VVT), Mask);
4957 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4958 DAG.getIntPtrConstant(0));
4959 } else if (VT.getSizeInBits() == 64) {
4960 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4961 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4962 // to match extract_elt for f64.
4963 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4967 // UNPCKHPD the element to the lowest double word, then movsd.
4968 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4969 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4970 int Mask[2] = { 1, -1 };
4971 EVT VVT = Op.getOperand(0).getValueType();
4972 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4973 DAG.getUNDEF(VVT), Mask);
4974 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4975 DAG.getIntPtrConstant(0));
4982 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
4983 SelectionDAG &DAG) const {
4984 EVT VT = Op.getValueType();
4985 EVT EltVT = VT.getVectorElementType();
4986 DebugLoc dl = Op.getDebugLoc();
4988 SDValue N0 = Op.getOperand(0);
4989 SDValue N1 = Op.getOperand(1);
4990 SDValue N2 = Op.getOperand(2);
4992 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4993 isa<ConstantSDNode>(N2)) {
4995 if (VT == MVT::v8i16)
4996 Opc = X86ISD::PINSRW;
4997 else if (VT == MVT::v4i16)
4998 Opc = X86ISD::MMX_PINSRW;
4999 else if (VT == MVT::v16i8)
5000 Opc = X86ISD::PINSRB;
5002 Opc = X86ISD::PINSRB;
5004 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5006 if (N1.getValueType() != MVT::i32)
5007 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5008 if (N2.getValueType() != MVT::i32)
5009 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5010 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5011 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5012 // Bits [7:6] of the constant are the source select. This will always be
5013 // zero here. The DAG Combiner may combine an extract_elt index into these
5014 // bits. For example (insert (extract, 3), 2) could be matched by putting
5015 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5016 // Bits [5:4] of the constant are the destination select. This is the
5017 // value of the incoming immediate.
5018 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5019 // combine either bitwise AND or insert of float 0.0 to set these bits.
5020 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5021 // Create this as a scalar to vector..
5022 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5023 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5024 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5025 // PINSR* works with constant index.
5032 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5033 EVT VT = Op.getValueType();
5034 EVT EltVT = VT.getVectorElementType();
5036 if (Subtarget->hasSSE41())
5037 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5039 if (EltVT == MVT::i8)
5042 DebugLoc dl = Op.getDebugLoc();
5043 SDValue N0 = Op.getOperand(0);
5044 SDValue N1 = Op.getOperand(1);
5045 SDValue N2 = Op.getOperand(2);
5047 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5048 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5049 // as its second argument.
5050 if (N1.getValueType() != MVT::i32)
5051 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5052 if (N2.getValueType() != MVT::i32)
5053 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5054 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5055 dl, VT, N0, N1, N2);
5061 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5062 DebugLoc dl = Op.getDebugLoc();
5064 if (Op.getValueType() == MVT::v1i64 &&
5065 Op.getOperand(0).getValueType() == MVT::i64)
5066 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5068 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5069 EVT VT = MVT::v2i32;
5070 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5077 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5078 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5081 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5082 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5083 // one of the above mentioned nodes. It has to be wrapped because otherwise
5084 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5085 // be used to form addressing mode. These wrapped nodes will be selected
5088 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5089 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5091 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5093 unsigned char OpFlag = 0;
5094 unsigned WrapperKind = X86ISD::Wrapper;
5095 CodeModel::Model M = getTargetMachine().getCodeModel();
5097 if (Subtarget->isPICStyleRIPRel() &&
5098 (M == CodeModel::Small || M == CodeModel::Kernel))
5099 WrapperKind = X86ISD::WrapperRIP;
5100 else if (Subtarget->isPICStyleGOT())
5101 OpFlag = X86II::MO_GOTOFF;
5102 else if (Subtarget->isPICStyleStubPIC())
5103 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5105 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5107 CP->getOffset(), OpFlag);
5108 DebugLoc DL = CP->getDebugLoc();
5109 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5110 // With PIC, the address is actually $g + Offset.
5112 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5113 DAG.getNode(X86ISD::GlobalBaseReg,
5114 DebugLoc(), getPointerTy()),
5121 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5122 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5124 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5126 unsigned char OpFlag = 0;
5127 unsigned WrapperKind = X86ISD::Wrapper;
5128 CodeModel::Model M = getTargetMachine().getCodeModel();
5130 if (Subtarget->isPICStyleRIPRel() &&
5131 (M == CodeModel::Small || M == CodeModel::Kernel))
5132 WrapperKind = X86ISD::WrapperRIP;
5133 else if (Subtarget->isPICStyleGOT())
5134 OpFlag = X86II::MO_GOTOFF;
5135 else if (Subtarget->isPICStyleStubPIC())
5136 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5138 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5140 DebugLoc DL = JT->getDebugLoc();
5141 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5143 // With PIC, the address is actually $g + Offset.
5145 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5146 DAG.getNode(X86ISD::GlobalBaseReg,
5147 DebugLoc(), getPointerTy()),
5155 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5156 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5158 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5160 unsigned char OpFlag = 0;
5161 unsigned WrapperKind = X86ISD::Wrapper;
5162 CodeModel::Model M = getTargetMachine().getCodeModel();
5164 if (Subtarget->isPICStyleRIPRel() &&
5165 (M == CodeModel::Small || M == CodeModel::Kernel))
5166 WrapperKind = X86ISD::WrapperRIP;
5167 else if (Subtarget->isPICStyleGOT())
5168 OpFlag = X86II::MO_GOTOFF;
5169 else if (Subtarget->isPICStyleStubPIC())
5170 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5172 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5174 DebugLoc DL = Op.getDebugLoc();
5175 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5178 // With PIC, the address is actually $g + Offset.
5179 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5180 !Subtarget->is64Bit()) {
5181 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5182 DAG.getNode(X86ISD::GlobalBaseReg,
5183 DebugLoc(), getPointerTy()),
5191 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5192 // Create the TargetBlockAddressAddress node.
5193 unsigned char OpFlags =
5194 Subtarget->ClassifyBlockAddressReference();
5195 CodeModel::Model M = getTargetMachine().getCodeModel();
5196 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5197 DebugLoc dl = Op.getDebugLoc();
5198 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5199 /*isTarget=*/true, OpFlags);
5201 if (Subtarget->isPICStyleRIPRel() &&
5202 (M == CodeModel::Small || M == CodeModel::Kernel))
5203 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5205 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5207 // With PIC, the address is actually $g + Offset.
5208 if (isGlobalRelativeToPICBase(OpFlags)) {
5209 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5210 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5218 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5220 SelectionDAG &DAG) const {
5221 // Create the TargetGlobalAddress node, folding in the constant
5222 // offset if it is legal.
5223 unsigned char OpFlags =
5224 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5225 CodeModel::Model M = getTargetMachine().getCodeModel();
5227 if (OpFlags == X86II::MO_NO_FLAG &&
5228 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5229 // A direct static reference to a global.
5230 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5233 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5236 if (Subtarget->isPICStyleRIPRel() &&
5237 (M == CodeModel::Small || M == CodeModel::Kernel))
5238 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5240 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5242 // With PIC, the address is actually $g + Offset.
5243 if (isGlobalRelativeToPICBase(OpFlags)) {
5244 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5245 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5249 // For globals that require a load from a stub to get the address, emit the
5251 if (isGlobalStubReference(OpFlags))
5252 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5253 PseudoSourceValue::getGOT(), 0, false, false, 0);
5255 // If there was a non-zero offset that we didn't fold, create an explicit
5258 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5259 DAG.getConstant(Offset, getPointerTy()));
5265 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5266 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5267 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5268 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5272 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5273 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5274 unsigned char OperandFlags) {
5275 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5276 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5277 DebugLoc dl = GA->getDebugLoc();
5278 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5279 GA->getValueType(0),
5283 SDValue Ops[] = { Chain, TGA, *InFlag };
5284 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5286 SDValue Ops[] = { Chain, TGA };
5287 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5290 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5291 MFI->setAdjustsStack(true);
5293 SDValue Flag = Chain.getValue(1);
5294 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5297 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5299 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5302 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5303 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5304 DAG.getNode(X86ISD::GlobalBaseReg,
5305 DebugLoc(), PtrVT), InFlag);
5306 InFlag = Chain.getValue(1);
5308 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5311 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5313 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5315 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5316 X86::RAX, X86II::MO_TLSGD);
5319 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5320 // "local exec" model.
5321 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5322 const EVT PtrVT, TLSModel::Model model,
5324 DebugLoc dl = GA->getDebugLoc();
5325 // Get the Thread Pointer
5326 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5328 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5331 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5332 NULL, 0, false, false, 0);
5334 unsigned char OperandFlags = 0;
5335 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5337 unsigned WrapperKind = X86ISD::Wrapper;
5338 if (model == TLSModel::LocalExec) {
5339 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5340 } else if (is64Bit) {
5341 assert(model == TLSModel::InitialExec);
5342 OperandFlags = X86II::MO_GOTTPOFF;
5343 WrapperKind = X86ISD::WrapperRIP;
5345 assert(model == TLSModel::InitialExec);
5346 OperandFlags = X86II::MO_INDNTPOFF;
5349 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5351 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5352 GA->getOffset(), OperandFlags);
5353 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5355 if (model == TLSModel::InitialExec)
5356 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5357 PseudoSourceValue::getGOT(), 0, false, false, 0);
5359 // The address of the thread local variable is the add of the thread
5360 // pointer with the offset of the variable.
5361 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5365 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5367 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5368 const GlobalValue *GV = GA->getGlobal();
5370 if (Subtarget->isTargetELF()) {
5371 // TODO: implement the "local dynamic" model
5372 // TODO: implement the "initial exec"model for pic executables
5374 // If GV is an alias then use the aliasee for determining
5375 // thread-localness.
5376 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5377 GV = GA->resolveAliasedGlobal(false);
5379 TLSModel::Model model
5380 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5383 case TLSModel::GeneralDynamic:
5384 case TLSModel::LocalDynamic: // not implemented
5385 if (Subtarget->is64Bit())
5386 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5387 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5389 case TLSModel::InitialExec:
5390 case TLSModel::LocalExec:
5391 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5392 Subtarget->is64Bit());
5394 } else if (Subtarget->isTargetDarwin()) {
5395 // Darwin only has one model of TLS. Lower to that.
5396 unsigned char OpFlag = 0;
5397 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5398 X86ISD::WrapperRIP : X86ISD::Wrapper;
5400 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5402 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5403 !Subtarget->is64Bit();
5405 OpFlag = X86II::MO_TLVP_PIC_BASE;
5407 OpFlag = X86II::MO_TLVP;
5409 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(),
5411 GA->getOffset(), OpFlag);
5413 DebugLoc DL = Op.getDebugLoc();
5414 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5416 // With PIC32, the address is actually $g + Offset.
5418 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5419 DAG.getNode(X86ISD::GlobalBaseReg,
5420 DebugLoc(), getPointerTy()),
5423 // Lowering the machine isd will make sure everything is in the right
5425 SDValue Args[] = { Offset };
5426 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5428 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5429 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5430 MFI->setAdjustsStack(true);
5432 // And our return value (tls address) is in the standard call return value
5434 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5435 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5439 "TLS not implemented for this target.");
5441 llvm_unreachable("Unreachable");
5446 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5447 /// take a 2 x i32 value to shift plus a shift amount.
5448 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5449 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5450 EVT VT = Op.getValueType();
5451 unsigned VTBits = VT.getSizeInBits();
5452 DebugLoc dl = Op.getDebugLoc();
5453 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5454 SDValue ShOpLo = Op.getOperand(0);
5455 SDValue ShOpHi = Op.getOperand(1);
5456 SDValue ShAmt = Op.getOperand(2);
5457 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5458 DAG.getConstant(VTBits - 1, MVT::i8))
5459 : DAG.getConstant(0, VT);
5462 if (Op.getOpcode() == ISD::SHL_PARTS) {
5463 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5464 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5466 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5467 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5470 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5471 DAG.getConstant(VTBits, MVT::i8));
5472 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5473 AndNode, DAG.getConstant(0, MVT::i8));
5476 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5477 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5478 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5480 if (Op.getOpcode() == ISD::SHL_PARTS) {
5481 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5482 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5484 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5485 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5488 SDValue Ops[2] = { Lo, Hi };
5489 return DAG.getMergeValues(Ops, 2, dl);
5492 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5493 SelectionDAG &DAG) const {
5494 EVT SrcVT = Op.getOperand(0).getValueType();
5496 if (SrcVT.isVector()) {
5497 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5503 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5504 "Unknown SINT_TO_FP to lower!");
5506 // These are really Legal; return the operand so the caller accepts it as
5508 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5510 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5511 Subtarget->is64Bit()) {
5515 DebugLoc dl = Op.getDebugLoc();
5516 unsigned Size = SrcVT.getSizeInBits()/8;
5517 MachineFunction &MF = DAG.getMachineFunction();
5518 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5519 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5520 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5522 PseudoSourceValue::getFixedStack(SSFI), 0,
5524 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5527 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5529 SelectionDAG &DAG) const {
5531 DebugLoc dl = Op.getDebugLoc();
5533 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5535 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5537 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5538 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5539 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5540 Tys, Ops, array_lengthof(Ops));
5543 Chain = Result.getValue(1);
5544 SDValue InFlag = Result.getValue(2);
5546 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5547 // shouldn't be necessary except that RFP cannot be live across
5548 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5549 MachineFunction &MF = DAG.getMachineFunction();
5550 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5551 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5552 Tys = DAG.getVTList(MVT::Other);
5554 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5556 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5557 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5558 PseudoSourceValue::getFixedStack(SSFI), 0,
5565 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5566 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5567 SelectionDAG &DAG) const {
5568 // This algorithm is not obvious. Here it is in C code, more or less:
5570 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5571 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5572 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5574 // Copy ints to xmm registers.
5575 __m128i xh = _mm_cvtsi32_si128( hi );
5576 __m128i xl = _mm_cvtsi32_si128( lo );
5578 // Combine into low half of a single xmm register.
5579 __m128i x = _mm_unpacklo_epi32( xh, xl );
5583 // Merge in appropriate exponents to give the integer bits the right
5585 x = _mm_unpacklo_epi32( x, exp );
5587 // Subtract away the biases to deal with the IEEE-754 double precision
5589 d = _mm_sub_pd( (__m128d) x, bias );
5591 // All conversions up to here are exact. The correctly rounded result is
5592 // calculated using the current rounding mode using the following
5594 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5595 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5596 // store doesn't really need to be here (except
5597 // maybe to zero the other double)
5602 DebugLoc dl = Op.getDebugLoc();
5603 LLVMContext *Context = DAG.getContext();
5605 // Build some magic constants.
5606 std::vector<Constant*> CV0;
5607 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5608 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5609 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5610 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5611 Constant *C0 = ConstantVector::get(CV0);
5612 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5614 std::vector<Constant*> CV1;
5616 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5618 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5619 Constant *C1 = ConstantVector::get(CV1);
5620 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5622 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5623 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5625 DAG.getIntPtrConstant(1)));
5626 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5627 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5629 DAG.getIntPtrConstant(0)));
5630 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5631 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5632 PseudoSourceValue::getConstantPool(), 0,
5634 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5635 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5636 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5637 PseudoSourceValue::getConstantPool(), 0,
5639 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5641 // Add the halves; easiest way is to swap them into another reg first.
5642 int ShufMask[2] = { 1, -1 };
5643 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5644 DAG.getUNDEF(MVT::v2f64), ShufMask);
5645 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5646 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5647 DAG.getIntPtrConstant(0));
5650 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5651 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5652 SelectionDAG &DAG) const {
5653 DebugLoc dl = Op.getDebugLoc();
5654 // FP constant to bias correct the final result.
5655 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5658 // Load the 32-bit value into an XMM register.
5659 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5660 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5662 DAG.getIntPtrConstant(0)));
5664 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5665 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5666 DAG.getIntPtrConstant(0));
5668 // Or the load with the bias.
5669 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5670 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5671 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5673 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5674 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5675 MVT::v2f64, Bias)));
5676 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5677 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5678 DAG.getIntPtrConstant(0));
5680 // Subtract the bias.
5681 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5683 // Handle final rounding.
5684 EVT DestVT = Op.getValueType();
5686 if (DestVT.bitsLT(MVT::f64)) {
5687 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5688 DAG.getIntPtrConstant(0));
5689 } else if (DestVT.bitsGT(MVT::f64)) {
5690 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5693 // Handle final rounding.
5697 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5698 SelectionDAG &DAG) const {
5699 SDValue N0 = Op.getOperand(0);
5700 DebugLoc dl = Op.getDebugLoc();
5702 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5703 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5704 // the optimization here.
5705 if (DAG.SignBitIsZero(N0))
5706 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5708 EVT SrcVT = N0.getValueType();
5709 EVT DstVT = Op.getValueType();
5710 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5711 return LowerUINT_TO_FP_i64(Op, DAG);
5712 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5713 return LowerUINT_TO_FP_i32(Op, DAG);
5715 // Make a 64-bit buffer, and use it to build an FILD.
5716 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5717 if (SrcVT == MVT::i32) {
5718 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5719 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5720 getPointerTy(), StackSlot, WordOff);
5721 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5722 StackSlot, NULL, 0, false, false, 0);
5723 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5724 OffsetSlot, NULL, 0, false, false, 0);
5725 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5729 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5730 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5731 StackSlot, NULL, 0, false, false, 0);
5732 // For i64 source, we need to add the appropriate power of 2 if the input
5733 // was negative. This is the same as the optimization in
5734 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5735 // we must be careful to do the computation in x87 extended precision, not
5736 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5737 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5738 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5739 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5741 APInt FF(32, 0x5F800000ULL);
5743 // Check whether the sign bit is set.
5744 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5745 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5748 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5749 SDValue FudgePtr = DAG.getConstantPool(
5750 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5753 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5754 SDValue Zero = DAG.getIntPtrConstant(0);
5755 SDValue Four = DAG.getIntPtrConstant(4);
5756 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5758 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5760 // Load the value out, extending it from f32 to f80.
5761 // FIXME: Avoid the extend by constructing the right constant pool?
5762 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
5763 FudgePtr, PseudoSourceValue::getConstantPool(),
5764 0, MVT::f32, false, false, 4);
5765 // Extend everything to 80 bits to force it to be done on x87.
5766 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5767 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5770 std::pair<SDValue,SDValue> X86TargetLowering::
5771 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5772 DebugLoc dl = Op.getDebugLoc();
5774 EVT DstTy = Op.getValueType();
5777 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5781 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5782 DstTy.getSimpleVT() >= MVT::i16 &&
5783 "Unknown FP_TO_SINT to lower!");
5785 // These are really Legal.
5786 if (DstTy == MVT::i32 &&
5787 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5788 return std::make_pair(SDValue(), SDValue());
5789 if (Subtarget->is64Bit() &&
5790 DstTy == MVT::i64 &&
5791 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5792 return std::make_pair(SDValue(), SDValue());
5794 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5796 MachineFunction &MF = DAG.getMachineFunction();
5797 unsigned MemSize = DstTy.getSizeInBits()/8;
5798 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5799 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5802 switch (DstTy.getSimpleVT().SimpleTy) {
5803 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5804 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5805 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5806 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5809 SDValue Chain = DAG.getEntryNode();
5810 SDValue Value = Op.getOperand(0);
5811 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5812 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5813 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5814 PseudoSourceValue::getFixedStack(SSFI), 0,
5816 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5818 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5820 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5821 Chain = Value.getValue(1);
5822 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5823 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5826 // Build the FP_TO_INT*_IN_MEM
5827 SDValue Ops[] = { Chain, Value, StackSlot };
5828 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5830 return std::make_pair(FIST, StackSlot);
5833 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5834 SelectionDAG &DAG) const {
5835 if (Op.getValueType().isVector()) {
5836 if (Op.getValueType() == MVT::v2i32 &&
5837 Op.getOperand(0).getValueType() == MVT::v2f64) {
5843 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5844 SDValue FIST = Vals.first, StackSlot = Vals.second;
5845 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5846 if (FIST.getNode() == 0) return Op;
5849 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5850 FIST, StackSlot, NULL, 0, false, false, 0);
5853 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5854 SelectionDAG &DAG) const {
5855 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5856 SDValue FIST = Vals.first, StackSlot = Vals.second;
5857 assert(FIST.getNode() && "Unexpected failure");
5860 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5861 FIST, StackSlot, NULL, 0, false, false, 0);
5864 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5865 SelectionDAG &DAG) const {
5866 LLVMContext *Context = DAG.getContext();
5867 DebugLoc dl = Op.getDebugLoc();
5868 EVT VT = Op.getValueType();
5871 EltVT = VT.getVectorElementType();
5872 std::vector<Constant*> CV;
5873 if (EltVT == MVT::f64) {
5874 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5878 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5884 Constant *C = ConstantVector::get(CV);
5885 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5886 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5887 PseudoSourceValue::getConstantPool(), 0,
5889 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5892 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5893 LLVMContext *Context = DAG.getContext();
5894 DebugLoc dl = Op.getDebugLoc();
5895 EVT VT = Op.getValueType();
5898 EltVT = VT.getVectorElementType();
5899 std::vector<Constant*> CV;
5900 if (EltVT == MVT::f64) {
5901 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5905 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5911 Constant *C = ConstantVector::get(CV);
5912 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5913 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5914 PseudoSourceValue::getConstantPool(), 0,
5916 if (VT.isVector()) {
5917 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5918 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5919 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5921 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5923 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5927 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5928 LLVMContext *Context = DAG.getContext();
5929 SDValue Op0 = Op.getOperand(0);
5930 SDValue Op1 = Op.getOperand(1);
5931 DebugLoc dl = Op.getDebugLoc();
5932 EVT VT = Op.getValueType();
5933 EVT SrcVT = Op1.getValueType();
5935 // If second operand is smaller, extend it first.
5936 if (SrcVT.bitsLT(VT)) {
5937 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5940 // And if it is bigger, shrink it first.
5941 if (SrcVT.bitsGT(VT)) {
5942 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5946 // At this point the operands and the result should have the same
5947 // type, and that won't be f80 since that is not custom lowered.
5949 // First get the sign bit of second operand.
5950 std::vector<Constant*> CV;
5951 if (SrcVT == MVT::f64) {
5952 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5953 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5955 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5956 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5957 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5958 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5960 Constant *C = ConstantVector::get(CV);
5961 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5962 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5963 PseudoSourceValue::getConstantPool(), 0,
5965 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5967 // Shift sign bit right or left if the two operands have different types.
5968 if (SrcVT.bitsGT(VT)) {
5969 // Op0 is MVT::f32, Op1 is MVT::f64.
5970 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5971 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5972 DAG.getConstant(32, MVT::i32));
5973 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5974 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5975 DAG.getIntPtrConstant(0));
5978 // Clear first operand sign bit.
5980 if (VT == MVT::f64) {
5981 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5982 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5984 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5985 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5986 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5987 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5989 C = ConstantVector::get(CV);
5990 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5991 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5992 PseudoSourceValue::getConstantPool(), 0,
5994 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5996 // Or the value with the sign bit.
5997 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6000 /// Emit nodes that will be selected as "test Op0,Op0", or something
6002 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6003 SelectionDAG &DAG) const {
6004 DebugLoc dl = Op.getDebugLoc();
6006 // CF and OF aren't always set the way we want. Determine which
6007 // of these we need.
6008 bool NeedCF = false;
6009 bool NeedOF = false;
6012 case X86::COND_A: case X86::COND_AE:
6013 case X86::COND_B: case X86::COND_BE:
6016 case X86::COND_G: case X86::COND_GE:
6017 case X86::COND_L: case X86::COND_LE:
6018 case X86::COND_O: case X86::COND_NO:
6023 // See if we can use the EFLAGS value from the operand instead of
6024 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6025 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6026 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6027 // Emit a CMP with 0, which is the TEST pattern.
6028 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6029 DAG.getConstant(0, Op.getValueType()));
6031 unsigned Opcode = 0;
6032 unsigned NumOperands = 0;
6033 switch (Op.getNode()->getOpcode()) {
6035 // Due to an isel shortcoming, be conservative if this add is likely to be
6036 // selected as part of a load-modify-store instruction. When the root node
6037 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6038 // uses of other nodes in the match, such as the ADD in this case. This
6039 // leads to the ADD being left around and reselected, with the result being
6040 // two adds in the output. Alas, even if none our users are stores, that
6041 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6042 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6043 // climbing the DAG back to the root, and it doesn't seem to be worth the
6045 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6046 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6047 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6050 if (ConstantSDNode *C =
6051 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6052 // An add of one will be selected as an INC.
6053 if (C->getAPIntValue() == 1) {
6054 Opcode = X86ISD::INC;
6059 // An add of negative one (subtract of one) will be selected as a DEC.
6060 if (C->getAPIntValue().isAllOnesValue()) {
6061 Opcode = X86ISD::DEC;
6067 // Otherwise use a regular EFLAGS-setting add.
6068 Opcode = X86ISD::ADD;
6072 // If the primary and result isn't used, don't bother using X86ISD::AND,
6073 // because a TEST instruction will be better.
6074 bool NonFlagUse = false;
6075 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6076 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6078 unsigned UOpNo = UI.getOperandNo();
6079 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6080 // Look pass truncate.
6081 UOpNo = User->use_begin().getOperandNo();
6082 User = *User->use_begin();
6085 if (User->getOpcode() != ISD::BRCOND &&
6086 User->getOpcode() != ISD::SETCC &&
6087 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6100 // Due to the ISEL shortcoming noted above, be conservative if this op is
6101 // likely to be selected as part of a load-modify-store instruction.
6102 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6103 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6104 if (UI->getOpcode() == ISD::STORE)
6107 // Otherwise use a regular EFLAGS-setting instruction.
6108 switch (Op.getNode()->getOpcode()) {
6109 default: llvm_unreachable("unexpected operator!");
6110 case ISD::SUB: Opcode = X86ISD::SUB; break;
6111 case ISD::OR: Opcode = X86ISD::OR; break;
6112 case ISD::XOR: Opcode = X86ISD::XOR; break;
6113 case ISD::AND: Opcode = X86ISD::AND; break;
6125 return SDValue(Op.getNode(), 1);
6132 // Emit a CMP with 0, which is the TEST pattern.
6133 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6134 DAG.getConstant(0, Op.getValueType()));
6136 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6137 SmallVector<SDValue, 4> Ops;
6138 for (unsigned i = 0; i != NumOperands; ++i)
6139 Ops.push_back(Op.getOperand(i));
6141 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6142 DAG.ReplaceAllUsesWith(Op, New);
6143 return SDValue(New.getNode(), 1);
6146 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6148 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6149 SelectionDAG &DAG) const {
6150 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6151 if (C->getAPIntValue() == 0)
6152 return EmitTest(Op0, X86CC, DAG);
6154 DebugLoc dl = Op0.getDebugLoc();
6155 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6158 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6159 /// if it's possible.
6160 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6161 DebugLoc dl, SelectionDAG &DAG) const {
6162 SDValue Op0 = And.getOperand(0);
6163 SDValue Op1 = And.getOperand(1);
6164 if (Op0.getOpcode() == ISD::TRUNCATE)
6165 Op0 = Op0.getOperand(0);
6166 if (Op1.getOpcode() == ISD::TRUNCATE)
6167 Op1 = Op1.getOperand(0);
6170 if (Op1.getOpcode() == ISD::SHL)
6171 std::swap(Op0, Op1);
6172 if (Op0.getOpcode() == ISD::SHL) {
6173 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6174 if (And00C->getZExtValue() == 1) {
6175 // If we looked past a truncate, check that it's only truncating away
6177 unsigned BitWidth = Op0.getValueSizeInBits();
6178 unsigned AndBitWidth = And.getValueSizeInBits();
6179 if (BitWidth > AndBitWidth) {
6180 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6181 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6182 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6186 RHS = Op0.getOperand(1);
6188 } else if (Op1.getOpcode() == ISD::Constant) {
6189 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6190 SDValue AndLHS = Op0;
6191 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6192 LHS = AndLHS.getOperand(0);
6193 RHS = AndLHS.getOperand(1);
6197 if (LHS.getNode()) {
6198 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6199 // instruction. Since the shift amount is in-range-or-undefined, we know
6200 // that doing a bittest on the i32 value is ok. We extend to i32 because
6201 // the encoding for the i16 version is larger than the i32 version.
6202 // Also promote i16 to i32 for performance / code size reason.
6203 if (LHS.getValueType() == MVT::i8 ||
6204 LHS.getValueType() == MVT::i16)
6205 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6207 // If the operand types disagree, extend the shift amount to match. Since
6208 // BT ignores high bits (like shifts) we can use anyextend.
6209 if (LHS.getValueType() != RHS.getValueType())
6210 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6212 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6213 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6214 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6215 DAG.getConstant(Cond, MVT::i8), BT);
6221 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6222 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6223 SDValue Op0 = Op.getOperand(0);
6224 SDValue Op1 = Op.getOperand(1);
6225 DebugLoc dl = Op.getDebugLoc();
6226 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6228 // Optimize to BT if possible.
6229 // Lower (X & (1 << N)) == 0 to BT(X, N).
6230 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6231 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6232 if (Op0.getOpcode() == ISD::AND &&
6234 Op1.getOpcode() == ISD::Constant &&
6235 cast<ConstantSDNode>(Op1)->isNullValue() &&
6236 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6237 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6238 if (NewSetCC.getNode())
6242 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6243 if (Op0.getOpcode() == X86ISD::SETCC &&
6244 Op1.getOpcode() == ISD::Constant &&
6245 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6246 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6247 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6248 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6249 bool Invert = (CC == ISD::SETNE) ^
6250 cast<ConstantSDNode>(Op1)->isNullValue();
6252 CCode = X86::GetOppositeBranchCondition(CCode);
6253 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6254 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6257 bool isFP = Op1.getValueType().isFloatingPoint();
6258 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6259 if (X86CC == X86::COND_INVALID)
6262 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6264 // Use sbb x, x to materialize carry bit into a GPR.
6265 if (X86CC == X86::COND_B)
6266 return DAG.getNode(ISD::AND, dl, MVT::i8,
6267 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6268 DAG.getConstant(X86CC, MVT::i8), Cond),
6269 DAG.getConstant(1, MVT::i8));
6271 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6272 DAG.getConstant(X86CC, MVT::i8), Cond);
6275 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6277 SDValue Op0 = Op.getOperand(0);
6278 SDValue Op1 = Op.getOperand(1);
6279 SDValue CC = Op.getOperand(2);
6280 EVT VT = Op.getValueType();
6281 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6282 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6283 DebugLoc dl = Op.getDebugLoc();
6287 EVT VT0 = Op0.getValueType();
6288 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6289 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6292 switch (SetCCOpcode) {
6295 case ISD::SETEQ: SSECC = 0; break;
6297 case ISD::SETGT: Swap = true; // Fallthrough
6299 case ISD::SETOLT: SSECC = 1; break;
6301 case ISD::SETGE: Swap = true; // Fallthrough
6303 case ISD::SETOLE: SSECC = 2; break;
6304 case ISD::SETUO: SSECC = 3; break;
6306 case ISD::SETNE: SSECC = 4; break;
6307 case ISD::SETULE: Swap = true;
6308 case ISD::SETUGE: SSECC = 5; break;
6309 case ISD::SETULT: Swap = true;
6310 case ISD::SETUGT: SSECC = 6; break;
6311 case ISD::SETO: SSECC = 7; break;
6314 std::swap(Op0, Op1);
6316 // In the two special cases we can't handle, emit two comparisons.
6318 if (SetCCOpcode == ISD::SETUEQ) {
6320 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6321 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6322 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6324 else if (SetCCOpcode == ISD::SETONE) {
6326 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6327 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6328 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6330 llvm_unreachable("Illegal FP comparison");
6332 // Handle all other FP comparisons here.
6333 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6336 // We are handling one of the integer comparisons here. Since SSE only has
6337 // GT and EQ comparisons for integer, swapping operands and multiple
6338 // operations may be required for some comparisons.
6339 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6340 bool Swap = false, Invert = false, FlipSigns = false;
6342 switch (VT.getSimpleVT().SimpleTy) {
6345 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6347 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6349 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6350 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6353 switch (SetCCOpcode) {
6355 case ISD::SETNE: Invert = true;
6356 case ISD::SETEQ: Opc = EQOpc; break;
6357 case ISD::SETLT: Swap = true;
6358 case ISD::SETGT: Opc = GTOpc; break;
6359 case ISD::SETGE: Swap = true;
6360 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6361 case ISD::SETULT: Swap = true;
6362 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6363 case ISD::SETUGE: Swap = true;
6364 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6367 std::swap(Op0, Op1);
6369 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6370 // bits of the inputs before performing those operations.
6372 EVT EltVT = VT.getVectorElementType();
6373 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6375 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6376 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6378 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6379 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6382 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6384 // If the logical-not of the result is required, perform that now.
6386 Result = DAG.getNOT(dl, Result, VT);
6391 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6392 static bool isX86LogicalCmp(SDValue Op) {
6393 unsigned Opc = Op.getNode()->getOpcode();
6394 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6396 if (Op.getResNo() == 1 &&
6397 (Opc == X86ISD::ADD ||
6398 Opc == X86ISD::SUB ||
6399 Opc == X86ISD::SMUL ||
6400 Opc == X86ISD::UMUL ||
6401 Opc == X86ISD::INC ||
6402 Opc == X86ISD::DEC ||
6403 Opc == X86ISD::OR ||
6404 Opc == X86ISD::XOR ||
6405 Opc == X86ISD::AND))
6411 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6412 bool addTest = true;
6413 SDValue Cond = Op.getOperand(0);
6414 DebugLoc dl = Op.getDebugLoc();
6417 if (Cond.getOpcode() == ISD::SETCC) {
6418 SDValue NewCond = LowerSETCC(Cond, DAG);
6419 if (NewCond.getNode())
6423 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6424 SDValue Op1 = Op.getOperand(1);
6425 SDValue Op2 = Op.getOperand(2);
6426 if (Cond.getOpcode() == X86ISD::SETCC &&
6427 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6428 SDValue Cmp = Cond.getOperand(1);
6429 if (Cmp.getOpcode() == X86ISD::CMP) {
6430 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6431 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6432 ConstantSDNode *RHSC =
6433 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6434 if (N1C && N1C->isAllOnesValue() &&
6435 N2C && N2C->isNullValue() &&
6436 RHSC && RHSC->isNullValue()) {
6437 SDValue CmpOp0 = Cmp.getOperand(0);
6438 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6439 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6440 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6441 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6446 // Look pass (and (setcc_carry (cmp ...)), 1).
6447 if (Cond.getOpcode() == ISD::AND &&
6448 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6449 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6450 if (C && C->getAPIntValue() == 1)
6451 Cond = Cond.getOperand(0);
6454 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6455 // setting operand in place of the X86ISD::SETCC.
6456 if (Cond.getOpcode() == X86ISD::SETCC ||
6457 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6458 CC = Cond.getOperand(0);
6460 SDValue Cmp = Cond.getOperand(1);
6461 unsigned Opc = Cmp.getOpcode();
6462 EVT VT = Op.getValueType();
6464 bool IllegalFPCMov = false;
6465 if (VT.isFloatingPoint() && !VT.isVector() &&
6466 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6467 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6469 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6470 Opc == X86ISD::BT) { // FIXME
6477 // Look pass the truncate.
6478 if (Cond.getOpcode() == ISD::TRUNCATE)
6479 Cond = Cond.getOperand(0);
6481 // We know the result of AND is compared against zero. Try to match
6483 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6484 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6485 if (NewSetCC.getNode()) {
6486 CC = NewSetCC.getOperand(0);
6487 Cond = NewSetCC.getOperand(1);
6494 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6495 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6498 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6499 // condition is true.
6500 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6501 SDValue Ops[] = { Op2, Op1, CC, Cond };
6502 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6505 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6506 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6507 // from the AND / OR.
6508 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6509 Opc = Op.getOpcode();
6510 if (Opc != ISD::OR && Opc != ISD::AND)
6512 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6513 Op.getOperand(0).hasOneUse() &&
6514 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6515 Op.getOperand(1).hasOneUse());
6518 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6519 // 1 and that the SETCC node has a single use.
6520 static bool isXor1OfSetCC(SDValue Op) {
6521 if (Op.getOpcode() != ISD::XOR)
6523 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6524 if (N1C && N1C->getAPIntValue() == 1) {
6525 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6526 Op.getOperand(0).hasOneUse();
6531 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6532 bool addTest = true;
6533 SDValue Chain = Op.getOperand(0);
6534 SDValue Cond = Op.getOperand(1);
6535 SDValue Dest = Op.getOperand(2);
6536 DebugLoc dl = Op.getDebugLoc();
6539 if (Cond.getOpcode() == ISD::SETCC) {
6540 SDValue NewCond = LowerSETCC(Cond, DAG);
6541 if (NewCond.getNode())
6545 // FIXME: LowerXALUO doesn't handle these!!
6546 else if (Cond.getOpcode() == X86ISD::ADD ||
6547 Cond.getOpcode() == X86ISD::SUB ||
6548 Cond.getOpcode() == X86ISD::SMUL ||
6549 Cond.getOpcode() == X86ISD::UMUL)
6550 Cond = LowerXALUO(Cond, DAG);
6553 // Look pass (and (setcc_carry (cmp ...)), 1).
6554 if (Cond.getOpcode() == ISD::AND &&
6555 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6556 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6557 if (C && C->getAPIntValue() == 1)
6558 Cond = Cond.getOperand(0);
6561 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6562 // setting operand in place of the X86ISD::SETCC.
6563 if (Cond.getOpcode() == X86ISD::SETCC ||
6564 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6565 CC = Cond.getOperand(0);
6567 SDValue Cmp = Cond.getOperand(1);
6568 unsigned Opc = Cmp.getOpcode();
6569 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6570 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6574 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6578 // These can only come from an arithmetic instruction with overflow,
6579 // e.g. SADDO, UADDO.
6580 Cond = Cond.getNode()->getOperand(1);
6587 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6588 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6589 if (CondOpc == ISD::OR) {
6590 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6591 // two branches instead of an explicit OR instruction with a
6593 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6594 isX86LogicalCmp(Cmp)) {
6595 CC = Cond.getOperand(0).getOperand(0);
6596 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6597 Chain, Dest, CC, Cmp);
6598 CC = Cond.getOperand(1).getOperand(0);
6602 } else { // ISD::AND
6603 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6604 // two branches instead of an explicit AND instruction with a
6605 // separate test. However, we only do this if this block doesn't
6606 // have a fall-through edge, because this requires an explicit
6607 // jmp when the condition is false.
6608 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6609 isX86LogicalCmp(Cmp) &&
6610 Op.getNode()->hasOneUse()) {
6611 X86::CondCode CCode =
6612 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6613 CCode = X86::GetOppositeBranchCondition(CCode);
6614 CC = DAG.getConstant(CCode, MVT::i8);
6615 SDNode *User = *Op.getNode()->use_begin();
6616 // Look for an unconditional branch following this conditional branch.
6617 // We need this because we need to reverse the successors in order
6618 // to implement FCMP_OEQ.
6619 if (User->getOpcode() == ISD::BR) {
6620 SDValue FalseBB = User->getOperand(1);
6622 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6623 assert(NewBR == User);
6627 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6628 Chain, Dest, CC, Cmp);
6629 X86::CondCode CCode =
6630 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6631 CCode = X86::GetOppositeBranchCondition(CCode);
6632 CC = DAG.getConstant(CCode, MVT::i8);
6638 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6639 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6640 // It should be transformed during dag combiner except when the condition
6641 // is set by a arithmetics with overflow node.
6642 X86::CondCode CCode =
6643 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6644 CCode = X86::GetOppositeBranchCondition(CCode);
6645 CC = DAG.getConstant(CCode, MVT::i8);
6646 Cond = Cond.getOperand(0).getOperand(1);
6652 // Look pass the truncate.
6653 if (Cond.getOpcode() == ISD::TRUNCATE)
6654 Cond = Cond.getOperand(0);
6656 // We know the result of AND is compared against zero. Try to match
6658 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6659 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6660 if (NewSetCC.getNode()) {
6661 CC = NewSetCC.getOperand(0);
6662 Cond = NewSetCC.getOperand(1);
6669 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6670 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6672 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6673 Chain, Dest, CC, Cond);
6677 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6678 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6679 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6680 // that the guard pages used by the OS virtual memory manager are allocated in
6681 // correct sequence.
6683 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6684 SelectionDAG &DAG) const {
6685 assert(Subtarget->isTargetCygMing() &&
6686 "This should be used only on Cygwin/Mingw targets");
6687 DebugLoc dl = Op.getDebugLoc();
6690 SDValue Chain = Op.getOperand(0);
6691 SDValue Size = Op.getOperand(1);
6692 // FIXME: Ensure alignment here
6696 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6698 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6699 Flag = Chain.getValue(1);
6701 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6703 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6704 Flag = Chain.getValue(1);
6706 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6708 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6709 return DAG.getMergeValues(Ops1, 2, dl);
6712 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6713 MachineFunction &MF = DAG.getMachineFunction();
6714 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6716 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6717 DebugLoc dl = Op.getDebugLoc();
6719 if (!Subtarget->is64Bit()) {
6720 // vastart just stores the address of the VarArgsFrameIndex slot into the
6721 // memory location argument.
6722 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6724 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6729 // gp_offset (0 - 6 * 8)
6730 // fp_offset (48 - 48 + 8 * 16)
6731 // overflow_arg_area (point to parameters coming in memory).
6733 SmallVector<SDValue, 8> MemOps;
6734 SDValue FIN = Op.getOperand(1);
6736 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6737 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6739 FIN, SV, 0, false, false, 0);
6740 MemOps.push_back(Store);
6743 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6744 FIN, DAG.getIntPtrConstant(4));
6745 Store = DAG.getStore(Op.getOperand(0), dl,
6746 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6748 FIN, SV, 0, false, false, 0);
6749 MemOps.push_back(Store);
6751 // Store ptr to overflow_arg_area
6752 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6753 FIN, DAG.getIntPtrConstant(4));
6754 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6756 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6758 MemOps.push_back(Store);
6760 // Store ptr to reg_save_area.
6761 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6762 FIN, DAG.getIntPtrConstant(8));
6763 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6765 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6767 MemOps.push_back(Store);
6768 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6769 &MemOps[0], MemOps.size());
6772 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6773 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6774 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6776 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6780 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6781 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6782 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6783 SDValue Chain = Op.getOperand(0);
6784 SDValue DstPtr = Op.getOperand(1);
6785 SDValue SrcPtr = Op.getOperand(2);
6786 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6787 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6788 DebugLoc dl = Op.getDebugLoc();
6790 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6791 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6792 false, DstSV, 0, SrcSV, 0);
6796 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6797 DebugLoc dl = Op.getDebugLoc();
6798 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6800 default: return SDValue(); // Don't custom lower most intrinsics.
6801 // Comparison intrinsics.
6802 case Intrinsic::x86_sse_comieq_ss:
6803 case Intrinsic::x86_sse_comilt_ss:
6804 case Intrinsic::x86_sse_comile_ss:
6805 case Intrinsic::x86_sse_comigt_ss:
6806 case Intrinsic::x86_sse_comige_ss:
6807 case Intrinsic::x86_sse_comineq_ss:
6808 case Intrinsic::x86_sse_ucomieq_ss:
6809 case Intrinsic::x86_sse_ucomilt_ss:
6810 case Intrinsic::x86_sse_ucomile_ss:
6811 case Intrinsic::x86_sse_ucomigt_ss:
6812 case Intrinsic::x86_sse_ucomige_ss:
6813 case Intrinsic::x86_sse_ucomineq_ss:
6814 case Intrinsic::x86_sse2_comieq_sd:
6815 case Intrinsic::x86_sse2_comilt_sd:
6816 case Intrinsic::x86_sse2_comile_sd:
6817 case Intrinsic::x86_sse2_comigt_sd:
6818 case Intrinsic::x86_sse2_comige_sd:
6819 case Intrinsic::x86_sse2_comineq_sd:
6820 case Intrinsic::x86_sse2_ucomieq_sd:
6821 case Intrinsic::x86_sse2_ucomilt_sd:
6822 case Intrinsic::x86_sse2_ucomile_sd:
6823 case Intrinsic::x86_sse2_ucomigt_sd:
6824 case Intrinsic::x86_sse2_ucomige_sd:
6825 case Intrinsic::x86_sse2_ucomineq_sd: {
6827 ISD::CondCode CC = ISD::SETCC_INVALID;
6830 case Intrinsic::x86_sse_comieq_ss:
6831 case Intrinsic::x86_sse2_comieq_sd:
6835 case Intrinsic::x86_sse_comilt_ss:
6836 case Intrinsic::x86_sse2_comilt_sd:
6840 case Intrinsic::x86_sse_comile_ss:
6841 case Intrinsic::x86_sse2_comile_sd:
6845 case Intrinsic::x86_sse_comigt_ss:
6846 case Intrinsic::x86_sse2_comigt_sd:
6850 case Intrinsic::x86_sse_comige_ss:
6851 case Intrinsic::x86_sse2_comige_sd:
6855 case Intrinsic::x86_sse_comineq_ss:
6856 case Intrinsic::x86_sse2_comineq_sd:
6860 case Intrinsic::x86_sse_ucomieq_ss:
6861 case Intrinsic::x86_sse2_ucomieq_sd:
6862 Opc = X86ISD::UCOMI;
6865 case Intrinsic::x86_sse_ucomilt_ss:
6866 case Intrinsic::x86_sse2_ucomilt_sd:
6867 Opc = X86ISD::UCOMI;
6870 case Intrinsic::x86_sse_ucomile_ss:
6871 case Intrinsic::x86_sse2_ucomile_sd:
6872 Opc = X86ISD::UCOMI;
6875 case Intrinsic::x86_sse_ucomigt_ss:
6876 case Intrinsic::x86_sse2_ucomigt_sd:
6877 Opc = X86ISD::UCOMI;
6880 case Intrinsic::x86_sse_ucomige_ss:
6881 case Intrinsic::x86_sse2_ucomige_sd:
6882 Opc = X86ISD::UCOMI;
6885 case Intrinsic::x86_sse_ucomineq_ss:
6886 case Intrinsic::x86_sse2_ucomineq_sd:
6887 Opc = X86ISD::UCOMI;
6892 SDValue LHS = Op.getOperand(1);
6893 SDValue RHS = Op.getOperand(2);
6894 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6895 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6896 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6897 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6898 DAG.getConstant(X86CC, MVT::i8), Cond);
6899 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6901 // ptest intrinsics. The intrinsic these come from are designed to return
6902 // an integer value, not just an instruction so lower it to the ptest
6903 // pattern and a setcc for the result.
6904 case Intrinsic::x86_sse41_ptestz:
6905 case Intrinsic::x86_sse41_ptestc:
6906 case Intrinsic::x86_sse41_ptestnzc:{
6909 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6910 case Intrinsic::x86_sse41_ptestz:
6912 X86CC = X86::COND_E;
6914 case Intrinsic::x86_sse41_ptestc:
6916 X86CC = X86::COND_B;
6918 case Intrinsic::x86_sse41_ptestnzc:
6920 X86CC = X86::COND_A;
6924 SDValue LHS = Op.getOperand(1);
6925 SDValue RHS = Op.getOperand(2);
6926 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6927 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6928 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6929 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6932 // Fix vector shift instructions where the last operand is a non-immediate
6934 case Intrinsic::x86_sse2_pslli_w:
6935 case Intrinsic::x86_sse2_pslli_d:
6936 case Intrinsic::x86_sse2_pslli_q:
6937 case Intrinsic::x86_sse2_psrli_w:
6938 case Intrinsic::x86_sse2_psrli_d:
6939 case Intrinsic::x86_sse2_psrli_q:
6940 case Intrinsic::x86_sse2_psrai_w:
6941 case Intrinsic::x86_sse2_psrai_d:
6942 case Intrinsic::x86_mmx_pslli_w:
6943 case Intrinsic::x86_mmx_pslli_d:
6944 case Intrinsic::x86_mmx_pslli_q:
6945 case Intrinsic::x86_mmx_psrli_w:
6946 case Intrinsic::x86_mmx_psrli_d:
6947 case Intrinsic::x86_mmx_psrli_q:
6948 case Intrinsic::x86_mmx_psrai_w:
6949 case Intrinsic::x86_mmx_psrai_d: {
6950 SDValue ShAmt = Op.getOperand(2);
6951 if (isa<ConstantSDNode>(ShAmt))
6954 unsigned NewIntNo = 0;
6955 EVT ShAmtVT = MVT::v4i32;
6957 case Intrinsic::x86_sse2_pslli_w:
6958 NewIntNo = Intrinsic::x86_sse2_psll_w;
6960 case Intrinsic::x86_sse2_pslli_d:
6961 NewIntNo = Intrinsic::x86_sse2_psll_d;
6963 case Intrinsic::x86_sse2_pslli_q:
6964 NewIntNo = Intrinsic::x86_sse2_psll_q;
6966 case Intrinsic::x86_sse2_psrli_w:
6967 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6969 case Intrinsic::x86_sse2_psrli_d:
6970 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6972 case Intrinsic::x86_sse2_psrli_q:
6973 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6975 case Intrinsic::x86_sse2_psrai_w:
6976 NewIntNo = Intrinsic::x86_sse2_psra_w;
6978 case Intrinsic::x86_sse2_psrai_d:
6979 NewIntNo = Intrinsic::x86_sse2_psra_d;
6982 ShAmtVT = MVT::v2i32;
6984 case Intrinsic::x86_mmx_pslli_w:
6985 NewIntNo = Intrinsic::x86_mmx_psll_w;
6987 case Intrinsic::x86_mmx_pslli_d:
6988 NewIntNo = Intrinsic::x86_mmx_psll_d;
6990 case Intrinsic::x86_mmx_pslli_q:
6991 NewIntNo = Intrinsic::x86_mmx_psll_q;
6993 case Intrinsic::x86_mmx_psrli_w:
6994 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6996 case Intrinsic::x86_mmx_psrli_d:
6997 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6999 case Intrinsic::x86_mmx_psrli_q:
7000 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7002 case Intrinsic::x86_mmx_psrai_w:
7003 NewIntNo = Intrinsic::x86_mmx_psra_w;
7005 case Intrinsic::x86_mmx_psrai_d:
7006 NewIntNo = Intrinsic::x86_mmx_psra_d;
7008 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7014 // The vector shift intrinsics with scalars uses 32b shift amounts but
7015 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7019 ShOps[1] = DAG.getConstant(0, MVT::i32);
7020 if (ShAmtVT == MVT::v4i32) {
7021 ShOps[2] = DAG.getUNDEF(MVT::i32);
7022 ShOps[3] = DAG.getUNDEF(MVT::i32);
7023 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7025 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7028 EVT VT = Op.getValueType();
7029 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7030 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7031 DAG.getConstant(NewIntNo, MVT::i32),
7032 Op.getOperand(1), ShAmt);
7037 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7038 SelectionDAG &DAG) const {
7039 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7040 MFI->setReturnAddressIsTaken(true);
7042 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7043 DebugLoc dl = Op.getDebugLoc();
7046 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7048 DAG.getConstant(TD->getPointerSize(),
7049 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7050 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7051 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7053 NULL, 0, false, false, 0);
7056 // Just load the return address.
7057 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7058 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7059 RetAddrFI, NULL, 0, false, false, 0);
7062 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7063 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7064 MFI->setFrameAddressIsTaken(true);
7066 EVT VT = Op.getValueType();
7067 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7068 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7069 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7070 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7072 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7077 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7078 SelectionDAG &DAG) const {
7079 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7082 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7083 MachineFunction &MF = DAG.getMachineFunction();
7084 SDValue Chain = Op.getOperand(0);
7085 SDValue Offset = Op.getOperand(1);
7086 SDValue Handler = Op.getOperand(2);
7087 DebugLoc dl = Op.getDebugLoc();
7089 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7091 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7093 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7094 DAG.getIntPtrConstant(-TD->getPointerSize()));
7095 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7096 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7097 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7098 MF.getRegInfo().addLiveOut(StoreAddrReg);
7100 return DAG.getNode(X86ISD::EH_RETURN, dl,
7102 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7105 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7106 SelectionDAG &DAG) const {
7107 SDValue Root = Op.getOperand(0);
7108 SDValue Trmp = Op.getOperand(1); // trampoline
7109 SDValue FPtr = Op.getOperand(2); // nested function
7110 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7111 DebugLoc dl = Op.getDebugLoc();
7113 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7115 if (Subtarget->is64Bit()) {
7116 SDValue OutChains[6];
7118 // Large code-model.
7119 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7120 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7122 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7123 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7125 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7127 // Load the pointer to the nested function into R11.
7128 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7129 SDValue Addr = Trmp;
7130 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7131 Addr, TrmpAddr, 0, false, false, 0);
7133 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7134 DAG.getConstant(2, MVT::i64));
7135 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7138 // Load the 'nest' parameter value into R10.
7139 // R10 is specified in X86CallingConv.td
7140 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7141 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7142 DAG.getConstant(10, MVT::i64));
7143 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7144 Addr, TrmpAddr, 10, false, false, 0);
7146 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7147 DAG.getConstant(12, MVT::i64));
7148 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7151 // Jump to the nested function.
7152 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7153 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7154 DAG.getConstant(20, MVT::i64));
7155 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7156 Addr, TrmpAddr, 20, false, false, 0);
7158 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7159 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7160 DAG.getConstant(22, MVT::i64));
7161 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7162 TrmpAddr, 22, false, false, 0);
7165 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7166 return DAG.getMergeValues(Ops, 2, dl);
7168 const Function *Func =
7169 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7170 CallingConv::ID CC = Func->getCallingConv();
7175 llvm_unreachable("Unsupported calling convention");
7176 case CallingConv::C:
7177 case CallingConv::X86_StdCall: {
7178 // Pass 'nest' parameter in ECX.
7179 // Must be kept in sync with X86CallingConv.td
7182 // Check that ECX wasn't needed by an 'inreg' parameter.
7183 const FunctionType *FTy = Func->getFunctionType();
7184 const AttrListPtr &Attrs = Func->getAttributes();
7186 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7187 unsigned InRegCount = 0;
7190 for (FunctionType::param_iterator I = FTy->param_begin(),
7191 E = FTy->param_end(); I != E; ++I, ++Idx)
7192 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7193 // FIXME: should only count parameters that are lowered to integers.
7194 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7196 if (InRegCount > 2) {
7197 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7202 case CallingConv::X86_FastCall:
7203 case CallingConv::X86_ThisCall:
7204 case CallingConv::Fast:
7205 // Pass 'nest' parameter in EAX.
7206 // Must be kept in sync with X86CallingConv.td
7211 SDValue OutChains[4];
7214 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7215 DAG.getConstant(10, MVT::i32));
7216 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7218 // This is storing the opcode for MOV32ri.
7219 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7220 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7221 OutChains[0] = DAG.getStore(Root, dl,
7222 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7223 Trmp, TrmpAddr, 0, false, false, 0);
7225 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7226 DAG.getConstant(1, MVT::i32));
7227 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7230 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7231 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7232 DAG.getConstant(5, MVT::i32));
7233 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7234 TrmpAddr, 5, false, false, 1);
7236 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7237 DAG.getConstant(6, MVT::i32));
7238 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7242 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7243 return DAG.getMergeValues(Ops, 2, dl);
7247 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7248 SelectionDAG &DAG) const {
7250 The rounding mode is in bits 11:10 of FPSR, and has the following
7257 FLT_ROUNDS, on the other hand, expects the following:
7264 To perform the conversion, we do:
7265 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7268 MachineFunction &MF = DAG.getMachineFunction();
7269 const TargetMachine &TM = MF.getTarget();
7270 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7271 unsigned StackAlignment = TFI.getStackAlignment();
7272 EVT VT = Op.getValueType();
7273 DebugLoc dl = Op.getDebugLoc();
7275 // Save FP Control Word to stack slot
7276 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7277 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7279 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7280 DAG.getEntryNode(), StackSlot);
7282 // Load FP Control Word from stack slot
7283 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7286 // Transform as necessary
7288 DAG.getNode(ISD::SRL, dl, MVT::i16,
7289 DAG.getNode(ISD::AND, dl, MVT::i16,
7290 CWD, DAG.getConstant(0x800, MVT::i16)),
7291 DAG.getConstant(11, MVT::i8));
7293 DAG.getNode(ISD::SRL, dl, MVT::i16,
7294 DAG.getNode(ISD::AND, dl, MVT::i16,
7295 CWD, DAG.getConstant(0x400, MVT::i16)),
7296 DAG.getConstant(9, MVT::i8));
7299 DAG.getNode(ISD::AND, dl, MVT::i16,
7300 DAG.getNode(ISD::ADD, dl, MVT::i16,
7301 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7302 DAG.getConstant(1, MVT::i16)),
7303 DAG.getConstant(3, MVT::i16));
7306 return DAG.getNode((VT.getSizeInBits() < 16 ?
7307 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7310 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7311 EVT VT = Op.getValueType();
7313 unsigned NumBits = VT.getSizeInBits();
7314 DebugLoc dl = Op.getDebugLoc();
7316 Op = Op.getOperand(0);
7317 if (VT == MVT::i8) {
7318 // Zero extend to i32 since there is not an i8 bsr.
7320 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7323 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7324 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7325 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7327 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7330 DAG.getConstant(NumBits+NumBits-1, OpVT),
7331 DAG.getConstant(X86::COND_E, MVT::i8),
7334 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7336 // Finally xor with NumBits-1.
7337 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7340 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7344 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7345 EVT VT = Op.getValueType();
7347 unsigned NumBits = VT.getSizeInBits();
7348 DebugLoc dl = Op.getDebugLoc();
7350 Op = Op.getOperand(0);
7351 if (VT == MVT::i8) {
7353 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7356 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7357 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7358 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7360 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7363 DAG.getConstant(NumBits, OpVT),
7364 DAG.getConstant(X86::COND_E, MVT::i8),
7367 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7370 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7374 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7375 EVT VT = Op.getValueType();
7376 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7377 DebugLoc dl = Op.getDebugLoc();
7379 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7380 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7381 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7382 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7383 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7385 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7386 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7387 // return AloBlo + AloBhi + AhiBlo;
7389 SDValue A = Op.getOperand(0);
7390 SDValue B = Op.getOperand(1);
7392 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7393 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7394 A, DAG.getConstant(32, MVT::i32));
7395 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7396 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7397 B, DAG.getConstant(32, MVT::i32));
7398 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7399 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7401 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7402 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7404 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7405 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7407 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7408 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7409 AloBhi, DAG.getConstant(32, MVT::i32));
7410 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7411 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7412 AhiBlo, DAG.getConstant(32, MVT::i32));
7413 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7414 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7419 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7420 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7421 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7422 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7423 // has only one use.
7424 SDNode *N = Op.getNode();
7425 SDValue LHS = N->getOperand(0);
7426 SDValue RHS = N->getOperand(1);
7427 unsigned BaseOp = 0;
7429 DebugLoc dl = Op.getDebugLoc();
7431 switch (Op.getOpcode()) {
7432 default: llvm_unreachable("Unknown ovf instruction!");
7434 // A subtract of one will be selected as a INC. Note that INC doesn't
7435 // set CF, so we can't do this for UADDO.
7436 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7437 if (C->getAPIntValue() == 1) {
7438 BaseOp = X86ISD::INC;
7442 BaseOp = X86ISD::ADD;
7446 BaseOp = X86ISD::ADD;
7450 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7451 // set CF, so we can't do this for USUBO.
7452 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7453 if (C->getAPIntValue() == 1) {
7454 BaseOp = X86ISD::DEC;
7458 BaseOp = X86ISD::SUB;
7462 BaseOp = X86ISD::SUB;
7466 BaseOp = X86ISD::SMUL;
7470 BaseOp = X86ISD::UMUL;
7475 // Also sets EFLAGS.
7476 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7477 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7480 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7481 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7483 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7487 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7488 EVT T = Op.getValueType();
7489 DebugLoc dl = Op.getDebugLoc();
7492 switch(T.getSimpleVT().SimpleTy) {
7494 assert(false && "Invalid value type!");
7495 case MVT::i8: Reg = X86::AL; size = 1; break;
7496 case MVT::i16: Reg = X86::AX; size = 2; break;
7497 case MVT::i32: Reg = X86::EAX; size = 4; break;
7499 assert(Subtarget->is64Bit() && "Node not type legal!");
7500 Reg = X86::RAX; size = 8;
7503 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7504 Op.getOperand(2), SDValue());
7505 SDValue Ops[] = { cpIn.getValue(0),
7508 DAG.getTargetConstant(size, MVT::i8),
7510 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7511 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7513 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7517 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7518 SelectionDAG &DAG) const {
7519 assert(Subtarget->is64Bit() && "Result not type legalized?");
7520 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7521 SDValue TheChain = Op.getOperand(0);
7522 DebugLoc dl = Op.getDebugLoc();
7523 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7524 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7525 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7527 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7528 DAG.getConstant(32, MVT::i8));
7530 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7533 return DAG.getMergeValues(Ops, 2, dl);
7536 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7537 SelectionDAG &DAG) const {
7538 EVT SrcVT = Op.getOperand(0).getValueType();
7539 EVT DstVT = Op.getValueType();
7540 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7541 Subtarget->hasMMX() && !DisableMMX) &&
7542 "Unexpected custom BIT_CONVERT");
7543 assert((DstVT == MVT::i64 ||
7544 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7545 "Unexpected custom BIT_CONVERT");
7546 // i64 <=> MMX conversions are Legal.
7547 if (SrcVT==MVT::i64 && DstVT.isVector())
7549 if (DstVT==MVT::i64 && SrcVT.isVector())
7551 // MMX <=> MMX conversions are Legal.
7552 if (SrcVT.isVector() && DstVT.isVector())
7554 // All other conversions need to be expanded.
7557 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7558 SDNode *Node = Op.getNode();
7559 DebugLoc dl = Node->getDebugLoc();
7560 EVT T = Node->getValueType(0);
7561 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7562 DAG.getConstant(0, T), Node->getOperand(2));
7563 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7564 cast<AtomicSDNode>(Node)->getMemoryVT(),
7565 Node->getOperand(0),
7566 Node->getOperand(1), negOp,
7567 cast<AtomicSDNode>(Node)->getSrcValue(),
7568 cast<AtomicSDNode>(Node)->getAlignment());
7571 /// LowerOperation - Provide custom lowering hooks for some operations.
7573 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7574 switch (Op.getOpcode()) {
7575 default: llvm_unreachable("Should not custom lower this!");
7576 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7577 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7578 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7579 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7580 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7581 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7582 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7583 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7584 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7585 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7586 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7587 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7588 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7589 case ISD::SHL_PARTS:
7590 case ISD::SRA_PARTS:
7591 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7592 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7593 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7594 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7595 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7596 case ISD::FABS: return LowerFABS(Op, DAG);
7597 case ISD::FNEG: return LowerFNEG(Op, DAG);
7598 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7599 case ISD::SETCC: return LowerSETCC(Op, DAG);
7600 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7601 case ISD::SELECT: return LowerSELECT(Op, DAG);
7602 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7603 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7604 case ISD::VASTART: return LowerVASTART(Op, DAG);
7605 case ISD::VAARG: return LowerVAARG(Op, DAG);
7606 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7607 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7608 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7609 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7610 case ISD::FRAME_TO_ARGS_OFFSET:
7611 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7612 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7613 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7614 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7615 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7616 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7617 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7618 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7624 case ISD::UMULO: return LowerXALUO(Op, DAG);
7625 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7626 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7630 void X86TargetLowering::
7631 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7632 SelectionDAG &DAG, unsigned NewOp) const {
7633 EVT T = Node->getValueType(0);
7634 DebugLoc dl = Node->getDebugLoc();
7635 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7637 SDValue Chain = Node->getOperand(0);
7638 SDValue In1 = Node->getOperand(1);
7639 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7640 Node->getOperand(2), DAG.getIntPtrConstant(0));
7641 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7642 Node->getOperand(2), DAG.getIntPtrConstant(1));
7643 SDValue Ops[] = { Chain, In1, In2L, In2H };
7644 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7646 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7647 cast<MemSDNode>(Node)->getMemOperand());
7648 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7649 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7650 Results.push_back(Result.getValue(2));
7653 /// ReplaceNodeResults - Replace a node with an illegal result type
7654 /// with a new node built out of custom code.
7655 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7656 SmallVectorImpl<SDValue>&Results,
7657 SelectionDAG &DAG) const {
7658 DebugLoc dl = N->getDebugLoc();
7659 switch (N->getOpcode()) {
7661 assert(false && "Do not know how to custom type legalize this operation!");
7663 case ISD::FP_TO_SINT: {
7664 std::pair<SDValue,SDValue> Vals =
7665 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7666 SDValue FIST = Vals.first, StackSlot = Vals.second;
7667 if (FIST.getNode() != 0) {
7668 EVT VT = N->getValueType(0);
7669 // Return a load from the stack slot.
7670 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7675 case ISD::READCYCLECOUNTER: {
7676 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7677 SDValue TheChain = N->getOperand(0);
7678 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7679 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7681 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7683 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7684 SDValue Ops[] = { eax, edx };
7685 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7686 Results.push_back(edx.getValue(1));
7689 case ISD::ATOMIC_CMP_SWAP: {
7690 EVT T = N->getValueType(0);
7691 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7692 SDValue cpInL, cpInH;
7693 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7694 DAG.getConstant(0, MVT::i32));
7695 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7696 DAG.getConstant(1, MVT::i32));
7697 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7698 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7700 SDValue swapInL, swapInH;
7701 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7702 DAG.getConstant(0, MVT::i32));
7703 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7704 DAG.getConstant(1, MVT::i32));
7705 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7707 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7708 swapInL.getValue(1));
7709 SDValue Ops[] = { swapInH.getValue(0),
7711 swapInH.getValue(1) };
7712 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7713 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7714 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7715 MVT::i32, Result.getValue(1));
7716 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7717 MVT::i32, cpOutL.getValue(2));
7718 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7719 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7720 Results.push_back(cpOutH.getValue(1));
7723 case ISD::ATOMIC_LOAD_ADD:
7724 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7726 case ISD::ATOMIC_LOAD_AND:
7727 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7729 case ISD::ATOMIC_LOAD_NAND:
7730 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7732 case ISD::ATOMIC_LOAD_OR:
7733 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7735 case ISD::ATOMIC_LOAD_SUB:
7736 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7738 case ISD::ATOMIC_LOAD_XOR:
7739 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7741 case ISD::ATOMIC_SWAP:
7742 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7747 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7749 default: return NULL;
7750 case X86ISD::BSF: return "X86ISD::BSF";
7751 case X86ISD::BSR: return "X86ISD::BSR";
7752 case X86ISD::SHLD: return "X86ISD::SHLD";
7753 case X86ISD::SHRD: return "X86ISD::SHRD";
7754 case X86ISD::FAND: return "X86ISD::FAND";
7755 case X86ISD::FOR: return "X86ISD::FOR";
7756 case X86ISD::FXOR: return "X86ISD::FXOR";
7757 case X86ISD::FSRL: return "X86ISD::FSRL";
7758 case X86ISD::FILD: return "X86ISD::FILD";
7759 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7760 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7761 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7762 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7763 case X86ISD::FLD: return "X86ISD::FLD";
7764 case X86ISD::FST: return "X86ISD::FST";
7765 case X86ISD::CALL: return "X86ISD::CALL";
7766 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7767 case X86ISD::BT: return "X86ISD::BT";
7768 case X86ISD::CMP: return "X86ISD::CMP";
7769 case X86ISD::COMI: return "X86ISD::COMI";
7770 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7771 case X86ISD::SETCC: return "X86ISD::SETCC";
7772 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7773 case X86ISD::CMOV: return "X86ISD::CMOV";
7774 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7775 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7776 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7777 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7778 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7779 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7780 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7781 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7782 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7783 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7784 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7785 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7786 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7787 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7788 case X86ISD::FMAX: return "X86ISD::FMAX";
7789 case X86ISD::FMIN: return "X86ISD::FMIN";
7790 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7791 case X86ISD::FRCP: return "X86ISD::FRCP";
7792 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7793 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
7794 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7795 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7796 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7797 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7798 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7799 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7800 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7801 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7802 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7803 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7804 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7805 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7806 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7807 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7808 case X86ISD::VSHL: return "X86ISD::VSHL";
7809 case X86ISD::VSRL: return "X86ISD::VSRL";
7810 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7811 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7812 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7813 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7814 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7815 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7816 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7817 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7818 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7819 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7820 case X86ISD::ADD: return "X86ISD::ADD";
7821 case X86ISD::SUB: return "X86ISD::SUB";
7822 case X86ISD::SMUL: return "X86ISD::SMUL";
7823 case X86ISD::UMUL: return "X86ISD::UMUL";
7824 case X86ISD::INC: return "X86ISD::INC";
7825 case X86ISD::DEC: return "X86ISD::DEC";
7826 case X86ISD::OR: return "X86ISD::OR";
7827 case X86ISD::XOR: return "X86ISD::XOR";
7828 case X86ISD::AND: return "X86ISD::AND";
7829 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7830 case X86ISD::PTEST: return "X86ISD::PTEST";
7831 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7832 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7836 // isLegalAddressingMode - Return true if the addressing mode represented
7837 // by AM is legal for this target, for a load/store of the specified type.
7838 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7839 const Type *Ty) const {
7840 // X86 supports extremely general addressing modes.
7841 CodeModel::Model M = getTargetMachine().getCodeModel();
7843 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7844 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7849 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7851 // If a reference to this global requires an extra load, we can't fold it.
7852 if (isGlobalStubReference(GVFlags))
7855 // If BaseGV requires a register for the PIC base, we cannot also have a
7856 // BaseReg specified.
7857 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7860 // If lower 4G is not available, then we must use rip-relative addressing.
7861 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7871 // These scales always work.
7876 // These scales are formed with basereg+scalereg. Only accept if there is
7881 default: // Other stuff never works.
7889 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7890 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7892 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7893 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7894 if (NumBits1 <= NumBits2)
7899 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7900 if (!VT1.isInteger() || !VT2.isInteger())
7902 unsigned NumBits1 = VT1.getSizeInBits();
7903 unsigned NumBits2 = VT2.getSizeInBits();
7904 if (NumBits1 <= NumBits2)
7909 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7910 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7911 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7914 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7915 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7916 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7919 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7920 // i16 instructions are longer (0x66 prefix) and potentially slower.
7921 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7924 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7925 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7926 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7927 /// are assumed to be legal.
7929 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7931 // Very little shuffling can be done for 64-bit vectors right now.
7932 if (VT.getSizeInBits() == 64)
7933 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
7935 // FIXME: pshufb, blends, shifts.
7936 return (VT.getVectorNumElements() == 2 ||
7937 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7938 isMOVLMask(M, VT) ||
7939 isSHUFPMask(M, VT) ||
7940 isPSHUFDMask(M, VT) ||
7941 isPSHUFHWMask(M, VT) ||
7942 isPSHUFLWMask(M, VT) ||
7943 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7944 isUNPCKLMask(M, VT) ||
7945 isUNPCKHMask(M, VT) ||
7946 isUNPCKL_v_undef_Mask(M, VT) ||
7947 isUNPCKH_v_undef_Mask(M, VT));
7951 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7953 unsigned NumElts = VT.getVectorNumElements();
7954 // FIXME: This collection of masks seems suspect.
7957 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7958 return (isMOVLMask(Mask, VT) ||
7959 isCommutedMOVLMask(Mask, VT, true) ||
7960 isSHUFPMask(Mask, VT) ||
7961 isCommutedSHUFPMask(Mask, VT));
7966 //===----------------------------------------------------------------------===//
7967 // X86 Scheduler Hooks
7968 //===----------------------------------------------------------------------===//
7970 // private utility function
7972 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7973 MachineBasicBlock *MBB,
7981 TargetRegisterClass *RC,
7982 bool invSrc) const {
7983 // For the atomic bitwise operator, we generate
7986 // ld t1 = [bitinstr.addr]
7987 // op t2 = t1, [bitinstr.val]
7989 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7991 // fallthrough -->nextMBB
7992 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7993 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7994 MachineFunction::iterator MBBIter = MBB;
7997 /// First build the CFG
7998 MachineFunction *F = MBB->getParent();
7999 MachineBasicBlock *thisMBB = MBB;
8000 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8001 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8002 F->insert(MBBIter, newMBB);
8003 F->insert(MBBIter, nextMBB);
8005 // Move all successors to thisMBB to nextMBB
8006 nextMBB->transferSuccessors(thisMBB);
8008 // Update thisMBB to fall through to newMBB
8009 thisMBB->addSuccessor(newMBB);
8011 // newMBB jumps to itself and fall through to nextMBB
8012 newMBB->addSuccessor(nextMBB);
8013 newMBB->addSuccessor(newMBB);
8015 // Insert instructions into newMBB based on incoming instruction
8016 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8017 "unexpected number of operands");
8018 DebugLoc dl = bInstr->getDebugLoc();
8019 MachineOperand& destOper = bInstr->getOperand(0);
8020 MachineOperand* argOpers[2 + X86AddrNumOperands];
8021 int numArgs = bInstr->getNumOperands() - 1;
8022 for (int i=0; i < numArgs; ++i)
8023 argOpers[i] = &bInstr->getOperand(i+1);
8025 // x86 address has 4 operands: base, index, scale, and displacement
8026 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8027 int valArgIndx = lastAddrIndx + 1;
8029 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8030 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8031 for (int i=0; i <= lastAddrIndx; ++i)
8032 (*MIB).addOperand(*argOpers[i]);
8034 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8036 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8041 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8042 assert((argOpers[valArgIndx]->isReg() ||
8043 argOpers[valArgIndx]->isImm()) &&
8045 if (argOpers[valArgIndx]->isReg())
8046 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8048 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8050 (*MIB).addOperand(*argOpers[valArgIndx]);
8052 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8055 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8056 for (int i=0; i <= lastAddrIndx; ++i)
8057 (*MIB).addOperand(*argOpers[i]);
8059 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8060 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8061 bInstr->memoperands_end());
8063 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8067 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8069 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8073 // private utility function: 64 bit atomics on 32 bit host.
8075 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8076 MachineBasicBlock *MBB,
8081 bool invSrc) const {
8082 // For the atomic bitwise operator, we generate
8083 // thisMBB (instructions are in pairs, except cmpxchg8b)
8084 // ld t1,t2 = [bitinstr.addr]
8086 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8087 // op t5, t6 <- out1, out2, [bitinstr.val]
8088 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8089 // mov ECX, EBX <- t5, t6
8090 // mov EAX, EDX <- t1, t2
8091 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8092 // mov t3, t4 <- EAX, EDX
8094 // result in out1, out2
8095 // fallthrough -->nextMBB
8097 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8098 const unsigned LoadOpc = X86::MOV32rm;
8099 const unsigned copyOpc = X86::MOV32rr;
8100 const unsigned NotOpc = X86::NOT32r;
8101 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8102 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8103 MachineFunction::iterator MBBIter = MBB;
8106 /// First build the CFG
8107 MachineFunction *F = MBB->getParent();
8108 MachineBasicBlock *thisMBB = MBB;
8109 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8110 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8111 F->insert(MBBIter, newMBB);
8112 F->insert(MBBIter, nextMBB);
8114 // Move all successors to thisMBB to nextMBB
8115 nextMBB->transferSuccessors(thisMBB);
8117 // Update thisMBB to fall through to newMBB
8118 thisMBB->addSuccessor(newMBB);
8120 // newMBB jumps to itself and fall through to nextMBB
8121 newMBB->addSuccessor(nextMBB);
8122 newMBB->addSuccessor(newMBB);
8124 DebugLoc dl = bInstr->getDebugLoc();
8125 // Insert instructions into newMBB based on incoming instruction
8126 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8127 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8128 "unexpected number of operands");
8129 MachineOperand& dest1Oper = bInstr->getOperand(0);
8130 MachineOperand& dest2Oper = bInstr->getOperand(1);
8131 MachineOperand* argOpers[2 + X86AddrNumOperands];
8132 for (int i=0; i < 2 + X86AddrNumOperands; ++i) {
8133 argOpers[i] = &bInstr->getOperand(i+2);
8135 // We use some of the operands multiple times, so conservatively just
8136 // clear any kill flags that might be present.
8137 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8138 argOpers[i]->setIsKill(false);
8141 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8142 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8144 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8145 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8146 for (int i=0; i <= lastAddrIndx; ++i)
8147 (*MIB).addOperand(*argOpers[i]);
8148 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8149 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8150 // add 4 to displacement.
8151 for (int i=0; i <= lastAddrIndx-2; ++i)
8152 (*MIB).addOperand(*argOpers[i]);
8153 MachineOperand newOp3 = *(argOpers[3]);
8155 newOp3.setImm(newOp3.getImm()+4);
8157 newOp3.setOffset(newOp3.getOffset()+4);
8158 (*MIB).addOperand(newOp3);
8159 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8161 // t3/4 are defined later, at the bottom of the loop
8162 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8163 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8164 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8165 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8166 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8167 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8169 // The subsequent operations should be using the destination registers of
8170 //the PHI instructions.
8172 t1 = F->getRegInfo().createVirtualRegister(RC);
8173 t2 = F->getRegInfo().createVirtualRegister(RC);
8174 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8175 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8177 t1 = dest1Oper.getReg();
8178 t2 = dest2Oper.getReg();
8181 int valArgIndx = lastAddrIndx + 1;
8182 assert((argOpers[valArgIndx]->isReg() ||
8183 argOpers[valArgIndx]->isImm()) &&
8185 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8186 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8187 if (argOpers[valArgIndx]->isReg())
8188 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8190 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8191 if (regOpcL != X86::MOV32rr)
8193 (*MIB).addOperand(*argOpers[valArgIndx]);
8194 assert(argOpers[valArgIndx + 1]->isReg() ==
8195 argOpers[valArgIndx]->isReg());
8196 assert(argOpers[valArgIndx + 1]->isImm() ==
8197 argOpers[valArgIndx]->isImm());
8198 if (argOpers[valArgIndx + 1]->isReg())
8199 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8201 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8202 if (regOpcH != X86::MOV32rr)
8204 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8206 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8208 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8211 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8213 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8216 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8217 for (int i=0; i <= lastAddrIndx; ++i)
8218 (*MIB).addOperand(*argOpers[i]);
8220 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8221 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8222 bInstr->memoperands_end());
8224 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8225 MIB.addReg(X86::EAX);
8226 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8227 MIB.addReg(X86::EDX);
8230 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8232 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8236 // private utility function
8238 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8239 MachineBasicBlock *MBB,
8240 unsigned cmovOpc) const {
8241 // For the atomic min/max operator, we generate
8244 // ld t1 = [min/max.addr]
8245 // mov t2 = [min/max.val]
8247 // cmov[cond] t2 = t1
8249 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8251 // fallthrough -->nextMBB
8253 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8254 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8255 MachineFunction::iterator MBBIter = MBB;
8258 /// First build the CFG
8259 MachineFunction *F = MBB->getParent();
8260 MachineBasicBlock *thisMBB = MBB;
8261 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8262 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8263 F->insert(MBBIter, newMBB);
8264 F->insert(MBBIter, nextMBB);
8266 // Move all successors of thisMBB to nextMBB
8267 nextMBB->transferSuccessors(thisMBB);
8269 // Update thisMBB to fall through to newMBB
8270 thisMBB->addSuccessor(newMBB);
8272 // newMBB jumps to newMBB and fall through to nextMBB
8273 newMBB->addSuccessor(nextMBB);
8274 newMBB->addSuccessor(newMBB);
8276 DebugLoc dl = mInstr->getDebugLoc();
8277 // Insert instructions into newMBB based on incoming instruction
8278 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8279 "unexpected number of operands");
8280 MachineOperand& destOper = mInstr->getOperand(0);
8281 MachineOperand* argOpers[2 + X86AddrNumOperands];
8282 int numArgs = mInstr->getNumOperands() - 1;
8283 for (int i=0; i < numArgs; ++i)
8284 argOpers[i] = &mInstr->getOperand(i+1);
8286 // x86 address has 4 operands: base, index, scale, and displacement
8287 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8288 int valArgIndx = lastAddrIndx + 1;
8290 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8291 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8292 for (int i=0; i <= lastAddrIndx; ++i)
8293 (*MIB).addOperand(*argOpers[i]);
8295 // We only support register and immediate values
8296 assert((argOpers[valArgIndx]->isReg() ||
8297 argOpers[valArgIndx]->isImm()) &&
8300 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8301 if (argOpers[valArgIndx]->isReg())
8302 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8304 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8305 (*MIB).addOperand(*argOpers[valArgIndx]);
8307 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8310 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8315 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8316 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8320 // Cmp and exchange if none has modified the memory location
8321 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8322 for (int i=0; i <= lastAddrIndx; ++i)
8323 (*MIB).addOperand(*argOpers[i]);
8325 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8326 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8327 mInstr->memoperands_end());
8329 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8330 MIB.addReg(X86::EAX);
8333 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8335 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8339 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8340 // all of this code can be replaced with that in the .td file.
8342 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8343 unsigned numArgs, bool memArg) const {
8345 MachineFunction *F = BB->getParent();
8346 DebugLoc dl = MI->getDebugLoc();
8347 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8351 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8353 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8355 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8357 for (unsigned i = 0; i < numArgs; ++i) {
8358 MachineOperand &Op = MI->getOperand(i+1);
8360 if (!(Op.isReg() && Op.isImplicit()))
8364 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8367 F->DeleteMachineInstr(MI);
8373 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8375 MachineBasicBlock *MBB) const {
8376 // Emit code to save XMM registers to the stack. The ABI says that the
8377 // number of registers to save is given in %al, so it's theoretically
8378 // possible to do an indirect jump trick to avoid saving all of them,
8379 // however this code takes a simpler approach and just executes all
8380 // of the stores if %al is non-zero. It's less code, and it's probably
8381 // easier on the hardware branch predictor, and stores aren't all that
8382 // expensive anyway.
8384 // Create the new basic blocks. One block contains all the XMM stores,
8385 // and one block is the final destination regardless of whether any
8386 // stores were performed.
8387 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8388 MachineFunction *F = MBB->getParent();
8389 MachineFunction::iterator MBBIter = MBB;
8391 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8392 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8393 F->insert(MBBIter, XMMSaveMBB);
8394 F->insert(MBBIter, EndMBB);
8397 // Move any original successors of MBB to the end block.
8398 EndMBB->transferSuccessors(MBB);
8399 // The original block will now fall through to the XMM save block.
8400 MBB->addSuccessor(XMMSaveMBB);
8401 // The XMMSaveMBB will fall through to the end block.
8402 XMMSaveMBB->addSuccessor(EndMBB);
8404 // Now add the instructions.
8405 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8406 DebugLoc DL = MI->getDebugLoc();
8408 unsigned CountReg = MI->getOperand(0).getReg();
8409 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8410 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8412 if (!Subtarget->isTargetWin64()) {
8413 // If %al is 0, branch around the XMM save block.
8414 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8415 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8416 MBB->addSuccessor(EndMBB);
8419 // In the XMM save block, save all the XMM argument registers.
8420 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8421 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8422 MachineMemOperand *MMO =
8423 F->getMachineMemOperand(
8424 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8425 MachineMemOperand::MOStore, Offset,
8426 /*Size=*/16, /*Align=*/16);
8427 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8428 .addFrameIndex(RegSaveFrameIndex)
8429 .addImm(/*Scale=*/1)
8430 .addReg(/*IndexReg=*/0)
8431 .addImm(/*Disp=*/Offset)
8432 .addReg(/*Segment=*/0)
8433 .addReg(MI->getOperand(i).getReg())
8434 .addMemOperand(MMO);
8437 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8443 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8444 MachineBasicBlock *BB) const {
8445 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8446 DebugLoc DL = MI->getDebugLoc();
8448 // To "insert" a SELECT_CC instruction, we actually have to insert the
8449 // diamond control-flow pattern. The incoming instruction knows the
8450 // destination vreg to set, the condition code register to branch on, the
8451 // true/false values to select between, and a branch opcode to use.
8452 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8453 MachineFunction::iterator It = BB;
8459 // cmpTY ccX, r1, r2
8461 // fallthrough --> copy0MBB
8462 MachineBasicBlock *thisMBB = BB;
8463 MachineFunction *F = BB->getParent();
8464 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8465 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8467 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8469 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8470 F->insert(It, copy0MBB);
8471 F->insert(It, sinkMBB);
8473 // Update machine-CFG edges by first adding all successors of the current
8474 // block to the new block which will contain the Phi node for the select.
8475 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8476 E = BB->succ_end(); I != E; ++I)
8477 sinkMBB->addSuccessor(*I);
8479 // Next, remove all successors of the current block, and add the true
8480 // and fallthrough blocks as its successors.
8481 while (!BB->succ_empty())
8482 BB->removeSuccessor(BB->succ_begin());
8484 // Add the true and fallthrough blocks as its successors.
8485 BB->addSuccessor(copy0MBB);
8486 BB->addSuccessor(sinkMBB);
8488 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8489 // live into the sink and copy blocks.
8490 const MachineFunction *MF = BB->getParent();
8491 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8492 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8493 const MachineInstr *Term = BB->getFirstTerminator();
8495 for (unsigned I = 0, E = Term->getNumOperands(); I != E; ++I) {
8496 const MachineOperand &MO = Term->getOperand(I);
8497 if (!MO.isReg() || MO.isKill() || MO.isDead()) continue;
8498 unsigned Reg = MO.getReg();
8499 if (Reg != X86::EFLAGS) continue;
8500 copy0MBB->addLiveIn(Reg);
8501 sinkMBB->addLiveIn(Reg);
8505 // %FalseValue = ...
8506 // # fallthrough to sinkMBB
8507 copy0MBB->addSuccessor(sinkMBB);
8510 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8512 BuildMI(sinkMBB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8513 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8514 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8516 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8521 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8522 MachineBasicBlock *BB) const {
8523 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8524 DebugLoc DL = MI->getDebugLoc();
8525 MachineFunction *F = BB->getParent();
8527 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8528 // non-trivial part is impdef of ESP.
8529 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8532 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8533 .addExternalSymbol("_alloca")
8534 .addReg(X86::EAX, RegState::Implicit)
8535 .addReg(X86::ESP, RegState::Implicit)
8536 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8537 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8539 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8544 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8545 MachineBasicBlock *BB) const {
8546 // This is pretty easy. We're taking the value that we received from
8547 // our load from the relocation, sticking it in either RDI (x86-64)
8548 // or EAX and doing an indirect call. The return value will then
8549 // be in the normal return register.
8550 const X86InstrInfo *TII
8551 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8552 DebugLoc DL = MI->getDebugLoc();
8553 MachineFunction *F = BB->getParent();
8555 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8557 if (Subtarget->is64Bit()) {
8558 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV64rm), X86::RDI)
8560 .addImm(0).addReg(0)
8561 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8562 MI->getOperand(3).getTargetFlags())
8564 MIB = BuildMI(BB, DL, TII->get(X86::CALL64m));
8565 addDirectMem(MIB, X86::RDI).addReg(0);
8566 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8567 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV32rm), X86::EAX)
8569 .addImm(0).addReg(0)
8570 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8571 MI->getOperand(3).getTargetFlags())
8573 MIB = BuildMI(BB, DL, TII->get(X86::CALL32m));
8574 addDirectMem(MIB, X86::EAX).addReg(0);
8576 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV32rm), X86::EAX)
8577 .addReg(TII->getGlobalBaseReg(F))
8578 .addImm(0).addReg(0)
8579 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8580 MI->getOperand(3).getTargetFlags())
8582 MIB = BuildMI(BB, DL, TII->get(X86::CALL32m));
8583 addDirectMem(MIB, X86::EAX).addReg(0);
8586 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8591 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8592 MachineBasicBlock *BB) const {
8593 switch (MI->getOpcode()) {
8594 default: assert(false && "Unexpected instr type to insert");
8595 case X86::MINGW_ALLOCA:
8596 return EmitLoweredMingwAlloca(MI, BB);
8597 case X86::TLSCall_32:
8598 case X86::TLSCall_64:
8599 return EmitLoweredTLSCall(MI, BB);
8601 case X86::CMOV_V1I64:
8602 case X86::CMOV_FR32:
8603 case X86::CMOV_FR64:
8604 case X86::CMOV_V4F32:
8605 case X86::CMOV_V2F64:
8606 case X86::CMOV_V2I64:
8607 case X86::CMOV_GR16:
8608 case X86::CMOV_GR32:
8609 case X86::CMOV_RFP32:
8610 case X86::CMOV_RFP64:
8611 case X86::CMOV_RFP80:
8612 return EmitLoweredSelect(MI, BB);
8614 case X86::FP32_TO_INT16_IN_MEM:
8615 case X86::FP32_TO_INT32_IN_MEM:
8616 case X86::FP32_TO_INT64_IN_MEM:
8617 case X86::FP64_TO_INT16_IN_MEM:
8618 case X86::FP64_TO_INT32_IN_MEM:
8619 case X86::FP64_TO_INT64_IN_MEM:
8620 case X86::FP80_TO_INT16_IN_MEM:
8621 case X86::FP80_TO_INT32_IN_MEM:
8622 case X86::FP80_TO_INT64_IN_MEM: {
8623 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8624 DebugLoc DL = MI->getDebugLoc();
8626 // Change the floating point control register to use "round towards zero"
8627 // mode when truncating to an integer value.
8628 MachineFunction *F = BB->getParent();
8629 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8630 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8632 // Load the old value of the high byte of the control word...
8634 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8635 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8638 // Set the high part to be round to zero...
8639 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8642 // Reload the modified control word now...
8643 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8645 // Restore the memory image of control word to original value
8646 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8649 // Get the X86 opcode to use.
8651 switch (MI->getOpcode()) {
8652 default: llvm_unreachable("illegal opcode!");
8653 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8654 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8655 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8656 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8657 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8658 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8659 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8660 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8661 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8665 MachineOperand &Op = MI->getOperand(0);
8667 AM.BaseType = X86AddressMode::RegBase;
8668 AM.Base.Reg = Op.getReg();
8670 AM.BaseType = X86AddressMode::FrameIndexBase;
8671 AM.Base.FrameIndex = Op.getIndex();
8673 Op = MI->getOperand(1);
8675 AM.Scale = Op.getImm();
8676 Op = MI->getOperand(2);
8678 AM.IndexReg = Op.getImm();
8679 Op = MI->getOperand(3);
8680 if (Op.isGlobal()) {
8681 AM.GV = Op.getGlobal();
8683 AM.Disp = Op.getImm();
8685 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8686 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8688 // Reload the original control word now.
8689 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8691 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8694 // String/text processing lowering.
8695 case X86::PCMPISTRM128REG:
8696 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8697 case X86::PCMPISTRM128MEM:
8698 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8699 case X86::PCMPESTRM128REG:
8700 return EmitPCMP(MI, BB, 5, false /* in mem */);
8701 case X86::PCMPESTRM128MEM:
8702 return EmitPCMP(MI, BB, 5, true /* in mem */);
8705 case X86::ATOMAND32:
8706 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8707 X86::AND32ri, X86::MOV32rm,
8708 X86::LCMPXCHG32, X86::MOV32rr,
8709 X86::NOT32r, X86::EAX,
8710 X86::GR32RegisterClass);
8712 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8713 X86::OR32ri, X86::MOV32rm,
8714 X86::LCMPXCHG32, X86::MOV32rr,
8715 X86::NOT32r, X86::EAX,
8716 X86::GR32RegisterClass);
8717 case X86::ATOMXOR32:
8718 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8719 X86::XOR32ri, X86::MOV32rm,
8720 X86::LCMPXCHG32, X86::MOV32rr,
8721 X86::NOT32r, X86::EAX,
8722 X86::GR32RegisterClass);
8723 case X86::ATOMNAND32:
8724 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8725 X86::AND32ri, X86::MOV32rm,
8726 X86::LCMPXCHG32, X86::MOV32rr,
8727 X86::NOT32r, X86::EAX,
8728 X86::GR32RegisterClass, true);
8729 case X86::ATOMMIN32:
8730 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8731 case X86::ATOMMAX32:
8732 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8733 case X86::ATOMUMIN32:
8734 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8735 case X86::ATOMUMAX32:
8736 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8738 case X86::ATOMAND16:
8739 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8740 X86::AND16ri, X86::MOV16rm,
8741 X86::LCMPXCHG16, X86::MOV16rr,
8742 X86::NOT16r, X86::AX,
8743 X86::GR16RegisterClass);
8745 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8746 X86::OR16ri, X86::MOV16rm,
8747 X86::LCMPXCHG16, X86::MOV16rr,
8748 X86::NOT16r, X86::AX,
8749 X86::GR16RegisterClass);
8750 case X86::ATOMXOR16:
8751 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8752 X86::XOR16ri, X86::MOV16rm,
8753 X86::LCMPXCHG16, X86::MOV16rr,
8754 X86::NOT16r, X86::AX,
8755 X86::GR16RegisterClass);
8756 case X86::ATOMNAND16:
8757 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8758 X86::AND16ri, X86::MOV16rm,
8759 X86::LCMPXCHG16, X86::MOV16rr,
8760 X86::NOT16r, X86::AX,
8761 X86::GR16RegisterClass, true);
8762 case X86::ATOMMIN16:
8763 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8764 case X86::ATOMMAX16:
8765 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8766 case X86::ATOMUMIN16:
8767 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8768 case X86::ATOMUMAX16:
8769 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8772 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8773 X86::AND8ri, X86::MOV8rm,
8774 X86::LCMPXCHG8, X86::MOV8rr,
8775 X86::NOT8r, X86::AL,
8776 X86::GR8RegisterClass);
8778 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8779 X86::OR8ri, X86::MOV8rm,
8780 X86::LCMPXCHG8, X86::MOV8rr,
8781 X86::NOT8r, X86::AL,
8782 X86::GR8RegisterClass);
8784 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8785 X86::XOR8ri, X86::MOV8rm,
8786 X86::LCMPXCHG8, X86::MOV8rr,
8787 X86::NOT8r, X86::AL,
8788 X86::GR8RegisterClass);
8789 case X86::ATOMNAND8:
8790 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8791 X86::AND8ri, X86::MOV8rm,
8792 X86::LCMPXCHG8, X86::MOV8rr,
8793 X86::NOT8r, X86::AL,
8794 X86::GR8RegisterClass, true);
8795 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8796 // This group is for 64-bit host.
8797 case X86::ATOMAND64:
8798 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8799 X86::AND64ri32, X86::MOV64rm,
8800 X86::LCMPXCHG64, X86::MOV64rr,
8801 X86::NOT64r, X86::RAX,
8802 X86::GR64RegisterClass);
8804 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8805 X86::OR64ri32, X86::MOV64rm,
8806 X86::LCMPXCHG64, X86::MOV64rr,
8807 X86::NOT64r, X86::RAX,
8808 X86::GR64RegisterClass);
8809 case X86::ATOMXOR64:
8810 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8811 X86::XOR64ri32, X86::MOV64rm,
8812 X86::LCMPXCHG64, X86::MOV64rr,
8813 X86::NOT64r, X86::RAX,
8814 X86::GR64RegisterClass);
8815 case X86::ATOMNAND64:
8816 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8817 X86::AND64ri32, X86::MOV64rm,
8818 X86::LCMPXCHG64, X86::MOV64rr,
8819 X86::NOT64r, X86::RAX,
8820 X86::GR64RegisterClass, true);
8821 case X86::ATOMMIN64:
8822 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8823 case X86::ATOMMAX64:
8824 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8825 case X86::ATOMUMIN64:
8826 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8827 case X86::ATOMUMAX64:
8828 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8830 // This group does 64-bit operations on a 32-bit host.
8831 case X86::ATOMAND6432:
8832 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8833 X86::AND32rr, X86::AND32rr,
8834 X86::AND32ri, X86::AND32ri,
8836 case X86::ATOMOR6432:
8837 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8838 X86::OR32rr, X86::OR32rr,
8839 X86::OR32ri, X86::OR32ri,
8841 case X86::ATOMXOR6432:
8842 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8843 X86::XOR32rr, X86::XOR32rr,
8844 X86::XOR32ri, X86::XOR32ri,
8846 case X86::ATOMNAND6432:
8847 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8848 X86::AND32rr, X86::AND32rr,
8849 X86::AND32ri, X86::AND32ri,
8851 case X86::ATOMADD6432:
8852 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8853 X86::ADD32rr, X86::ADC32rr,
8854 X86::ADD32ri, X86::ADC32ri,
8856 case X86::ATOMSUB6432:
8857 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8858 X86::SUB32rr, X86::SBB32rr,
8859 X86::SUB32ri, X86::SBB32ri,
8861 case X86::ATOMSWAP6432:
8862 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8863 X86::MOV32rr, X86::MOV32rr,
8864 X86::MOV32ri, X86::MOV32ri,
8866 case X86::VASTART_SAVE_XMM_REGS:
8867 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8871 //===----------------------------------------------------------------------===//
8872 // X86 Optimization Hooks
8873 //===----------------------------------------------------------------------===//
8875 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8879 const SelectionDAG &DAG,
8880 unsigned Depth) const {
8881 unsigned Opc = Op.getOpcode();
8882 assert((Opc >= ISD::BUILTIN_OP_END ||
8883 Opc == ISD::INTRINSIC_WO_CHAIN ||
8884 Opc == ISD::INTRINSIC_W_CHAIN ||
8885 Opc == ISD::INTRINSIC_VOID) &&
8886 "Should use MaskedValueIsZero if you don't know whether Op"
8887 " is a target node!");
8889 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8901 // These nodes' second result is a boolean.
8902 if (Op.getResNo() == 0)
8906 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8907 Mask.getBitWidth() - 1);
8912 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8913 /// node is a GlobalAddress + offset.
8914 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8915 const GlobalValue* &GA,
8916 int64_t &Offset) const {
8917 if (N->getOpcode() == X86ISD::Wrapper) {
8918 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8919 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8920 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8924 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8927 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8928 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8929 /// if the load addresses are consecutive, non-overlapping, and in the right
8931 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8932 const TargetLowering &TLI) {
8933 DebugLoc dl = N->getDebugLoc();
8934 EVT VT = N->getValueType(0);
8935 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8937 if (VT.getSizeInBits() != 128)
8940 SmallVector<SDValue, 16> Elts;
8941 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8942 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8944 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8947 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8948 /// and convert it from being a bunch of shuffles and extracts to a simple
8949 /// store and scalar loads to extract the elements.
8950 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8951 const TargetLowering &TLI) {
8952 SDValue InputVector = N->getOperand(0);
8954 // Only operate on vectors of 4 elements, where the alternative shuffling
8955 // gets to be more expensive.
8956 if (InputVector.getValueType() != MVT::v4i32)
8959 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8960 // single use which is a sign-extend or zero-extend, and all elements are
8962 SmallVector<SDNode *, 4> Uses;
8963 unsigned ExtractedElements = 0;
8964 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8965 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8966 if (UI.getUse().getResNo() != InputVector.getResNo())
8969 SDNode *Extract = *UI;
8970 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8973 if (Extract->getValueType(0) != MVT::i32)
8975 if (!Extract->hasOneUse())
8977 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
8978 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
8980 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
8983 // Record which element was extracted.
8984 ExtractedElements |=
8985 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
8987 Uses.push_back(Extract);
8990 // If not all the elements were used, this may not be worthwhile.
8991 if (ExtractedElements != 15)
8994 // Ok, we've now decided to do the transformation.
8995 DebugLoc dl = InputVector.getDebugLoc();
8997 // Store the value to a temporary stack slot.
8998 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
8999 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
9002 // Replace each use (extract) with a load of the appropriate element.
9003 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9004 UE = Uses.end(); UI != UE; ++UI) {
9005 SDNode *Extract = *UI;
9007 // Compute the element's address.
9008 SDValue Idx = Extract->getOperand(1);
9010 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9011 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9012 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9014 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
9017 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
9018 NULL, 0, false, false, 0);
9020 // Replace the exact with the load.
9021 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9024 // The replacement was made in place; don't return anything.
9028 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9029 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9030 const X86Subtarget *Subtarget) {
9031 DebugLoc DL = N->getDebugLoc();
9032 SDValue Cond = N->getOperand(0);
9033 // Get the LHS/RHS of the select.
9034 SDValue LHS = N->getOperand(1);
9035 SDValue RHS = N->getOperand(2);
9037 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9038 // instructions match the semantics of the common C idiom x<y?x:y but not
9039 // x<=y?x:y, because of how they handle negative zero (which can be
9040 // ignored in unsafe-math mode).
9041 if (Subtarget->hasSSE2() &&
9042 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9043 Cond.getOpcode() == ISD::SETCC) {
9044 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9046 unsigned Opcode = 0;
9047 // Check for x CC y ? x : y.
9048 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9049 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9053 // Converting this to a min would handle NaNs incorrectly, and swapping
9054 // the operands would cause it to handle comparisons between positive
9055 // and negative zero incorrectly.
9056 if (!FiniteOnlyFPMath() &&
9057 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9058 if (!UnsafeFPMath &&
9059 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9061 std::swap(LHS, RHS);
9063 Opcode = X86ISD::FMIN;
9066 // Converting this to a min would handle comparisons between positive
9067 // and negative zero incorrectly.
9068 if (!UnsafeFPMath &&
9069 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9071 Opcode = X86ISD::FMIN;
9074 // Converting this to a min would handle both negative zeros and NaNs
9075 // incorrectly, but we can swap the operands to fix both.
9076 std::swap(LHS, RHS);
9080 Opcode = X86ISD::FMIN;
9084 // Converting this to a max would handle comparisons between positive
9085 // and negative zero incorrectly.
9086 if (!UnsafeFPMath &&
9087 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9089 Opcode = X86ISD::FMAX;
9092 // Converting this to a max would handle NaNs incorrectly, and swapping
9093 // the operands would cause it to handle comparisons between positive
9094 // and negative zero incorrectly.
9095 if (!FiniteOnlyFPMath() &&
9096 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9097 if (!UnsafeFPMath &&
9098 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9100 std::swap(LHS, RHS);
9102 Opcode = X86ISD::FMAX;
9105 // Converting this to a max would handle both negative zeros and NaNs
9106 // incorrectly, but we can swap the operands to fix both.
9107 std::swap(LHS, RHS);
9111 Opcode = X86ISD::FMAX;
9114 // Check for x CC y ? y : x -- a min/max with reversed arms.
9115 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9116 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9120 // Converting this to a min would handle comparisons between positive
9121 // and negative zero incorrectly, and swapping the operands would
9122 // cause it to handle NaNs incorrectly.
9123 if (!UnsafeFPMath &&
9124 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9125 if (!FiniteOnlyFPMath() &&
9126 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9128 std::swap(LHS, RHS);
9130 Opcode = X86ISD::FMIN;
9133 // Converting this to a min would handle NaNs incorrectly.
9134 if (!UnsafeFPMath &&
9135 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9137 Opcode = X86ISD::FMIN;
9140 // Converting this to a min would handle both negative zeros and NaNs
9141 // incorrectly, but we can swap the operands to fix both.
9142 std::swap(LHS, RHS);
9146 Opcode = X86ISD::FMIN;
9150 // Converting this to a max would handle NaNs incorrectly.
9151 if (!FiniteOnlyFPMath() &&
9152 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9154 Opcode = X86ISD::FMAX;
9157 // Converting this to a max would handle comparisons between positive
9158 // and negative zero incorrectly, and swapping the operands would
9159 // cause it to handle NaNs incorrectly.
9160 if (!UnsafeFPMath &&
9161 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9162 if (!FiniteOnlyFPMath() &&
9163 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9165 std::swap(LHS, RHS);
9167 Opcode = X86ISD::FMAX;
9170 // Converting this to a max would handle both negative zeros and NaNs
9171 // incorrectly, but we can swap the operands to fix both.
9172 std::swap(LHS, RHS);
9176 Opcode = X86ISD::FMAX;
9182 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9185 // If this is a select between two integer constants, try to do some
9187 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9188 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9189 // Don't do this for crazy integer types.
9190 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9191 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9192 // so that TrueC (the true value) is larger than FalseC.
9193 bool NeedsCondInvert = false;
9195 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9196 // Efficiently invertible.
9197 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9198 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9199 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9200 NeedsCondInvert = true;
9201 std::swap(TrueC, FalseC);
9204 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9205 if (FalseC->getAPIntValue() == 0 &&
9206 TrueC->getAPIntValue().isPowerOf2()) {
9207 if (NeedsCondInvert) // Invert the condition if needed.
9208 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9209 DAG.getConstant(1, Cond.getValueType()));
9211 // Zero extend the condition if needed.
9212 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9214 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9215 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9216 DAG.getConstant(ShAmt, MVT::i8));
9219 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9220 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9221 if (NeedsCondInvert) // Invert the condition if needed.
9222 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9223 DAG.getConstant(1, Cond.getValueType()));
9225 // Zero extend the condition if needed.
9226 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9227 FalseC->getValueType(0), Cond);
9228 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9229 SDValue(FalseC, 0));
9232 // Optimize cases that will turn into an LEA instruction. This requires
9233 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9234 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9235 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9236 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9238 bool isFastMultiplier = false;
9240 switch ((unsigned char)Diff) {
9242 case 1: // result = add base, cond
9243 case 2: // result = lea base( , cond*2)
9244 case 3: // result = lea base(cond, cond*2)
9245 case 4: // result = lea base( , cond*4)
9246 case 5: // result = lea base(cond, cond*4)
9247 case 8: // result = lea base( , cond*8)
9248 case 9: // result = lea base(cond, cond*8)
9249 isFastMultiplier = true;
9254 if (isFastMultiplier) {
9255 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9256 if (NeedsCondInvert) // Invert the condition if needed.
9257 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9258 DAG.getConstant(1, Cond.getValueType()));
9260 // Zero extend the condition if needed.
9261 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9263 // Scale the condition by the difference.
9265 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9266 DAG.getConstant(Diff, Cond.getValueType()));
9268 // Add the base if non-zero.
9269 if (FalseC->getAPIntValue() != 0)
9270 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9271 SDValue(FalseC, 0));
9281 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9282 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9283 TargetLowering::DAGCombinerInfo &DCI) {
9284 DebugLoc DL = N->getDebugLoc();
9286 // If the flag operand isn't dead, don't touch this CMOV.
9287 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9290 // If this is a select between two integer constants, try to do some
9291 // optimizations. Note that the operands are ordered the opposite of SELECT
9293 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9294 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9295 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9296 // larger than FalseC (the false value).
9297 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9299 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9300 CC = X86::GetOppositeBranchCondition(CC);
9301 std::swap(TrueC, FalseC);
9304 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9305 // This is efficient for any integer data type (including i8/i16) and
9307 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9308 SDValue Cond = N->getOperand(3);
9309 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9310 DAG.getConstant(CC, MVT::i8), Cond);
9312 // Zero extend the condition if needed.
9313 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9315 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9316 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9317 DAG.getConstant(ShAmt, MVT::i8));
9318 if (N->getNumValues() == 2) // Dead flag value?
9319 return DCI.CombineTo(N, Cond, SDValue());
9323 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9324 // for any integer data type, including i8/i16.
9325 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9326 SDValue Cond = N->getOperand(3);
9327 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9328 DAG.getConstant(CC, MVT::i8), Cond);
9330 // Zero extend the condition if needed.
9331 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9332 FalseC->getValueType(0), Cond);
9333 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9334 SDValue(FalseC, 0));
9336 if (N->getNumValues() == 2) // Dead flag value?
9337 return DCI.CombineTo(N, Cond, SDValue());
9341 // Optimize cases that will turn into an LEA instruction. This requires
9342 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9343 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9344 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9345 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9347 bool isFastMultiplier = false;
9349 switch ((unsigned char)Diff) {
9351 case 1: // result = add base, cond
9352 case 2: // result = lea base( , cond*2)
9353 case 3: // result = lea base(cond, cond*2)
9354 case 4: // result = lea base( , cond*4)
9355 case 5: // result = lea base(cond, cond*4)
9356 case 8: // result = lea base( , cond*8)
9357 case 9: // result = lea base(cond, cond*8)
9358 isFastMultiplier = true;
9363 if (isFastMultiplier) {
9364 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9365 SDValue Cond = N->getOperand(3);
9366 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9367 DAG.getConstant(CC, MVT::i8), Cond);
9368 // Zero extend the condition if needed.
9369 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9371 // Scale the condition by the difference.
9373 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9374 DAG.getConstant(Diff, Cond.getValueType()));
9376 // Add the base if non-zero.
9377 if (FalseC->getAPIntValue() != 0)
9378 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9379 SDValue(FalseC, 0));
9380 if (N->getNumValues() == 2) // Dead flag value?
9381 return DCI.CombineTo(N, Cond, SDValue());
9391 /// PerformMulCombine - Optimize a single multiply with constant into two
9392 /// in order to implement it with two cheaper instructions, e.g.
9393 /// LEA + SHL, LEA + LEA.
9394 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9395 TargetLowering::DAGCombinerInfo &DCI) {
9396 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9399 EVT VT = N->getValueType(0);
9403 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9406 uint64_t MulAmt = C->getZExtValue();
9407 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9410 uint64_t MulAmt1 = 0;
9411 uint64_t MulAmt2 = 0;
9412 if ((MulAmt % 9) == 0) {
9414 MulAmt2 = MulAmt / 9;
9415 } else if ((MulAmt % 5) == 0) {
9417 MulAmt2 = MulAmt / 5;
9418 } else if ((MulAmt % 3) == 0) {
9420 MulAmt2 = MulAmt / 3;
9423 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9424 DebugLoc DL = N->getDebugLoc();
9426 if (isPowerOf2_64(MulAmt2) &&
9427 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9428 // If second multiplifer is pow2, issue it first. We want the multiply by
9429 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9431 std::swap(MulAmt1, MulAmt2);
9434 if (isPowerOf2_64(MulAmt1))
9435 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9436 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9438 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9439 DAG.getConstant(MulAmt1, VT));
9441 if (isPowerOf2_64(MulAmt2))
9442 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9443 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9445 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9446 DAG.getConstant(MulAmt2, VT));
9448 // Do not add new nodes to DAG combiner worklist.
9449 DCI.CombineTo(N, NewMul, false);
9454 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9455 SDValue N0 = N->getOperand(0);
9456 SDValue N1 = N->getOperand(1);
9457 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9458 EVT VT = N0.getValueType();
9460 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9461 // since the result of setcc_c is all zero's or all ones.
9462 if (N1C && N0.getOpcode() == ISD::AND &&
9463 N0.getOperand(1).getOpcode() == ISD::Constant) {
9464 SDValue N00 = N0.getOperand(0);
9465 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9466 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9467 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9468 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9469 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9470 APInt ShAmt = N1C->getAPIntValue();
9471 Mask = Mask.shl(ShAmt);
9473 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9474 N00, DAG.getConstant(Mask, VT));
9481 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9483 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9484 const X86Subtarget *Subtarget) {
9485 EVT VT = N->getValueType(0);
9486 if (!VT.isVector() && VT.isInteger() &&
9487 N->getOpcode() == ISD::SHL)
9488 return PerformSHLCombine(N, DAG);
9490 // On X86 with SSE2 support, we can transform this to a vector shift if
9491 // all elements are shifted by the same amount. We can't do this in legalize
9492 // because the a constant vector is typically transformed to a constant pool
9493 // so we have no knowledge of the shift amount.
9494 if (!Subtarget->hasSSE2())
9497 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9500 SDValue ShAmtOp = N->getOperand(1);
9501 EVT EltVT = VT.getVectorElementType();
9502 DebugLoc DL = N->getDebugLoc();
9503 SDValue BaseShAmt = SDValue();
9504 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9505 unsigned NumElts = VT.getVectorNumElements();
9507 for (; i != NumElts; ++i) {
9508 SDValue Arg = ShAmtOp.getOperand(i);
9509 if (Arg.getOpcode() == ISD::UNDEF) continue;
9513 for (; i != NumElts; ++i) {
9514 SDValue Arg = ShAmtOp.getOperand(i);
9515 if (Arg.getOpcode() == ISD::UNDEF) continue;
9516 if (Arg != BaseShAmt) {
9520 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9521 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9522 SDValue InVec = ShAmtOp.getOperand(0);
9523 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9524 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9526 for (; i != NumElts; ++i) {
9527 SDValue Arg = InVec.getOperand(i);
9528 if (Arg.getOpcode() == ISD::UNDEF) continue;
9532 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9533 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9534 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9535 if (C->getZExtValue() == SplatIdx)
9536 BaseShAmt = InVec.getOperand(1);
9539 if (BaseShAmt.getNode() == 0)
9540 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9541 DAG.getIntPtrConstant(0));
9545 // The shift amount is an i32.
9546 if (EltVT.bitsGT(MVT::i32))
9547 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9548 else if (EltVT.bitsLT(MVT::i32))
9549 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9551 // The shift amount is identical so we can do a vector shift.
9552 SDValue ValOp = N->getOperand(0);
9553 switch (N->getOpcode()) {
9555 llvm_unreachable("Unknown shift opcode!");
9558 if (VT == MVT::v2i64)
9559 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9560 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9562 if (VT == MVT::v4i32)
9563 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9564 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9566 if (VT == MVT::v8i16)
9567 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9568 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9572 if (VT == MVT::v4i32)
9573 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9574 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9576 if (VT == MVT::v8i16)
9577 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9578 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9582 if (VT == MVT::v2i64)
9583 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9584 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9586 if (VT == MVT::v4i32)
9587 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9588 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9590 if (VT == MVT::v8i16)
9591 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9592 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9599 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9600 TargetLowering::DAGCombinerInfo &DCI,
9601 const X86Subtarget *Subtarget) {
9602 if (DCI.isBeforeLegalizeOps())
9605 EVT VT = N->getValueType(0);
9606 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9609 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9610 SDValue N0 = N->getOperand(0);
9611 SDValue N1 = N->getOperand(1);
9612 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9614 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9616 if (!N0.hasOneUse() || !N1.hasOneUse())
9619 SDValue ShAmt0 = N0.getOperand(1);
9620 if (ShAmt0.getValueType() != MVT::i8)
9622 SDValue ShAmt1 = N1.getOperand(1);
9623 if (ShAmt1.getValueType() != MVT::i8)
9625 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9626 ShAmt0 = ShAmt0.getOperand(0);
9627 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9628 ShAmt1 = ShAmt1.getOperand(0);
9630 DebugLoc DL = N->getDebugLoc();
9631 unsigned Opc = X86ISD::SHLD;
9632 SDValue Op0 = N0.getOperand(0);
9633 SDValue Op1 = N1.getOperand(0);
9634 if (ShAmt0.getOpcode() == ISD::SUB) {
9636 std::swap(Op0, Op1);
9637 std::swap(ShAmt0, ShAmt1);
9640 unsigned Bits = VT.getSizeInBits();
9641 if (ShAmt1.getOpcode() == ISD::SUB) {
9642 SDValue Sum = ShAmt1.getOperand(0);
9643 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9644 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
9645 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
9646 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
9647 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
9648 return DAG.getNode(Opc, DL, VT,
9650 DAG.getNode(ISD::TRUNCATE, DL,
9653 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9654 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9656 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9657 return DAG.getNode(Opc, DL, VT,
9658 N0.getOperand(0), N1.getOperand(0),
9659 DAG.getNode(ISD::TRUNCATE, DL,
9666 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9667 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9668 const X86Subtarget *Subtarget) {
9669 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9670 // the FP state in cases where an emms may be missing.
9671 // A preferable solution to the general problem is to figure out the right
9672 // places to insert EMMS. This qualifies as a quick hack.
9674 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9675 StoreSDNode *St = cast<StoreSDNode>(N);
9676 EVT VT = St->getValue().getValueType();
9677 if (VT.getSizeInBits() != 64)
9680 const Function *F = DAG.getMachineFunction().getFunction();
9681 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9682 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9683 && Subtarget->hasSSE2();
9684 if ((VT.isVector() ||
9685 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9686 isa<LoadSDNode>(St->getValue()) &&
9687 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9688 St->getChain().hasOneUse() && !St->isVolatile()) {
9689 SDNode* LdVal = St->getValue().getNode();
9691 int TokenFactorIndex = -1;
9692 SmallVector<SDValue, 8> Ops;
9693 SDNode* ChainVal = St->getChain().getNode();
9694 // Must be a store of a load. We currently handle two cases: the load
9695 // is a direct child, and it's under an intervening TokenFactor. It is
9696 // possible to dig deeper under nested TokenFactors.
9697 if (ChainVal == LdVal)
9698 Ld = cast<LoadSDNode>(St->getChain());
9699 else if (St->getValue().hasOneUse() &&
9700 ChainVal->getOpcode() == ISD::TokenFactor) {
9701 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9702 if (ChainVal->getOperand(i).getNode() == LdVal) {
9703 TokenFactorIndex = i;
9704 Ld = cast<LoadSDNode>(St->getValue());
9706 Ops.push_back(ChainVal->getOperand(i));
9710 if (!Ld || !ISD::isNormalLoad(Ld))
9713 // If this is not the MMX case, i.e. we are just turning i64 load/store
9714 // into f64 load/store, avoid the transformation if there are multiple
9715 // uses of the loaded value.
9716 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9719 DebugLoc LdDL = Ld->getDebugLoc();
9720 DebugLoc StDL = N->getDebugLoc();
9721 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9722 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9724 if (Subtarget->is64Bit() || F64IsLegal) {
9725 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9726 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9727 Ld->getBasePtr(), Ld->getSrcValue(),
9728 Ld->getSrcValueOffset(), Ld->isVolatile(),
9729 Ld->isNonTemporal(), Ld->getAlignment());
9730 SDValue NewChain = NewLd.getValue(1);
9731 if (TokenFactorIndex != -1) {
9732 Ops.push_back(NewChain);
9733 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9736 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9737 St->getSrcValue(), St->getSrcValueOffset(),
9738 St->isVolatile(), St->isNonTemporal(),
9739 St->getAlignment());
9742 // Otherwise, lower to two pairs of 32-bit loads / stores.
9743 SDValue LoAddr = Ld->getBasePtr();
9744 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9745 DAG.getConstant(4, MVT::i32));
9747 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9748 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9749 Ld->isVolatile(), Ld->isNonTemporal(),
9750 Ld->getAlignment());
9751 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9752 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9753 Ld->isVolatile(), Ld->isNonTemporal(),
9754 MinAlign(Ld->getAlignment(), 4));
9756 SDValue NewChain = LoLd.getValue(1);
9757 if (TokenFactorIndex != -1) {
9758 Ops.push_back(LoLd);
9759 Ops.push_back(HiLd);
9760 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9764 LoAddr = St->getBasePtr();
9765 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9766 DAG.getConstant(4, MVT::i32));
9768 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9769 St->getSrcValue(), St->getSrcValueOffset(),
9770 St->isVolatile(), St->isNonTemporal(),
9771 St->getAlignment());
9772 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9774 St->getSrcValueOffset() + 4,
9776 St->isNonTemporal(),
9777 MinAlign(St->getAlignment(), 4));
9778 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9783 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9784 /// X86ISD::FXOR nodes.
9785 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9786 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9787 // F[X]OR(0.0, x) -> x
9788 // F[X]OR(x, 0.0) -> x
9789 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9790 if (C->getValueAPF().isPosZero())
9791 return N->getOperand(1);
9792 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9793 if (C->getValueAPF().isPosZero())
9794 return N->getOperand(0);
9798 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9799 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9800 // FAND(0.0, x) -> 0.0
9801 // FAND(x, 0.0) -> 0.0
9802 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9803 if (C->getValueAPF().isPosZero())
9804 return N->getOperand(0);
9805 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9806 if (C->getValueAPF().isPosZero())
9807 return N->getOperand(1);
9811 static SDValue PerformBTCombine(SDNode *N,
9813 TargetLowering::DAGCombinerInfo &DCI) {
9814 // BT ignores high bits in the bit index operand.
9815 SDValue Op1 = N->getOperand(1);
9816 if (Op1.hasOneUse()) {
9817 unsigned BitWidth = Op1.getValueSizeInBits();
9818 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9819 APInt KnownZero, KnownOne;
9820 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9821 !DCI.isBeforeLegalizeOps());
9822 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9823 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9824 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9825 DCI.CommitTargetLoweringOpt(TLO);
9830 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9831 SDValue Op = N->getOperand(0);
9832 if (Op.getOpcode() == ISD::BIT_CONVERT)
9833 Op = Op.getOperand(0);
9834 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9835 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9836 VT.getVectorElementType().getSizeInBits() ==
9837 OpVT.getVectorElementType().getSizeInBits()) {
9838 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9843 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9844 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9845 // (and (i32 x86isd::setcc_carry), 1)
9846 // This eliminates the zext. This transformation is necessary because
9847 // ISD::SETCC is always legalized to i8.
9848 DebugLoc dl = N->getDebugLoc();
9849 SDValue N0 = N->getOperand(0);
9850 EVT VT = N->getValueType(0);
9851 if (N0.getOpcode() == ISD::AND &&
9853 N0.getOperand(0).hasOneUse()) {
9854 SDValue N00 = N0.getOperand(0);
9855 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9857 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9858 if (!C || C->getZExtValue() != 1)
9860 return DAG.getNode(ISD::AND, dl, VT,
9861 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9862 N00.getOperand(0), N00.getOperand(1)),
9863 DAG.getConstant(1, VT));
9869 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9870 DAGCombinerInfo &DCI) const {
9871 SelectionDAG &DAG = DCI.DAG;
9872 switch (N->getOpcode()) {
9874 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9875 case ISD::EXTRACT_VECTOR_ELT:
9876 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9877 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9878 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9879 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9882 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9883 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
9884 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9886 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9887 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9888 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9889 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9890 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9896 /// isTypeDesirableForOp - Return true if the target has native support for
9897 /// the specified value type and it is 'desirable' to use the type for the
9898 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9899 /// instruction encodings are longer and some i16 instructions are slow.
9900 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9901 if (!isTypeLegal(VT))
9910 case ISD::SIGN_EXTEND:
9911 case ISD::ZERO_EXTEND:
9912 case ISD::ANY_EXTEND:
9925 static bool MayFoldLoad(SDValue Op) {
9926 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
9929 static bool MayFoldIntoStore(SDValue Op) {
9930 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
9933 /// IsDesirableToPromoteOp - This method query the target whether it is
9934 /// beneficial for dag combiner to promote the specified node. If true, it
9935 /// should return the desired promotion type by reference.
9936 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
9937 EVT VT = Op.getValueType();
9941 bool Promote = false;
9942 bool Commute = false;
9943 switch (Op.getOpcode()) {
9946 LoadSDNode *LD = cast<LoadSDNode>(Op);
9947 // If the non-extending load has a single use and it's not live out, then it
9949 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
9951 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9952 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
9953 // The only case where we'd want to promote LOAD (rather then it being
9954 // promoted as an operand is when it's only use is liveout.
9955 if (UI->getOpcode() != ISD::CopyToReg)
9962 case ISD::SIGN_EXTEND:
9963 case ISD::ZERO_EXTEND:
9964 case ISD::ANY_EXTEND:
9969 SDValue N0 = Op.getOperand(0);
9970 // Look out for (store (shl (load), x)).
9971 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
9984 SDValue N0 = Op.getOperand(0);
9985 SDValue N1 = Op.getOperand(1);
9986 if (!Commute && MayFoldLoad(N1))
9988 // Avoid disabling potential load folding opportunities.
9989 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
9991 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10001 //===----------------------------------------------------------------------===//
10002 // X86 Inline Assembly Support
10003 //===----------------------------------------------------------------------===//
10005 static bool LowerToBSwap(CallInst *CI) {
10006 // FIXME: this should verify that we are targetting a 486 or better. If not,
10007 // we will turn this bswap into something that will be lowered to logical ops
10008 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10009 // so don't worry about this.
10011 // Verify this is a simple bswap.
10012 if (CI->getNumArgOperands() != 1 ||
10013 CI->getType() != CI->getArgOperand(0)->getType() ||
10014 !CI->getType()->isIntegerTy())
10017 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10018 if (!Ty || Ty->getBitWidth() % 16 != 0)
10021 // Okay, we can do this xform, do so now.
10022 const Type *Tys[] = { Ty };
10023 Module *M = CI->getParent()->getParent()->getParent();
10024 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10026 Value *Op = CI->getArgOperand(0);
10027 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10029 CI->replaceAllUsesWith(Op);
10030 CI->eraseFromParent();
10034 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10035 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10036 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10038 std::string AsmStr = IA->getAsmString();
10040 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10041 SmallVector<StringRef, 4> AsmPieces;
10042 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10044 switch (AsmPieces.size()) {
10045 default: return false;
10047 AsmStr = AsmPieces[0];
10049 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10052 if (AsmPieces.size() == 2 &&
10053 (AsmPieces[0] == "bswap" ||
10054 AsmPieces[0] == "bswapq" ||
10055 AsmPieces[0] == "bswapl") &&
10056 (AsmPieces[1] == "$0" ||
10057 AsmPieces[1] == "${0:q}")) {
10058 // No need to check constraints, nothing other than the equivalent of
10059 // "=r,0" would be valid here.
10060 return LowerToBSwap(CI);
10062 // rorw $$8, ${0:w} --> llvm.bswap.i16
10063 if (CI->getType()->isIntegerTy(16) &&
10064 AsmPieces.size() == 3 &&
10065 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10066 AsmPieces[1] == "$$8," &&
10067 AsmPieces[2] == "${0:w}" &&
10068 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10070 const std::string &Constraints = IA->getConstraintString();
10071 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10072 std::sort(AsmPieces.begin(), AsmPieces.end());
10073 if (AsmPieces.size() == 4 &&
10074 AsmPieces[0] == "~{cc}" &&
10075 AsmPieces[1] == "~{dirflag}" &&
10076 AsmPieces[2] == "~{flags}" &&
10077 AsmPieces[3] == "~{fpsr}") {
10078 return LowerToBSwap(CI);
10083 if (CI->getType()->isIntegerTy(64) &&
10084 Constraints.size() >= 2 &&
10085 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10086 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10087 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10088 SmallVector<StringRef, 4> Words;
10089 SplitString(AsmPieces[0], Words, " \t");
10090 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10092 SplitString(AsmPieces[1], Words, " \t");
10093 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10095 SplitString(AsmPieces[2], Words, " \t,");
10096 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10097 Words[2] == "%edx") {
10098 return LowerToBSwap(CI);
10110 /// getConstraintType - Given a constraint letter, return the type of
10111 /// constraint it is for this target.
10112 X86TargetLowering::ConstraintType
10113 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10114 if (Constraint.size() == 1) {
10115 switch (Constraint[0]) {
10127 return C_RegisterClass;
10135 return TargetLowering::getConstraintType(Constraint);
10138 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10139 /// with another that has more specific requirements based on the type of the
10140 /// corresponding operand.
10141 const char *X86TargetLowering::
10142 LowerXConstraint(EVT ConstraintVT) const {
10143 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10144 // 'f' like normal targets.
10145 if (ConstraintVT.isFloatingPoint()) {
10146 if (Subtarget->hasSSE2())
10148 if (Subtarget->hasSSE1())
10152 return TargetLowering::LowerXConstraint(ConstraintVT);
10155 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10156 /// vector. If it is invalid, don't add anything to Ops.
10157 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10159 std::vector<SDValue>&Ops,
10160 SelectionDAG &DAG) const {
10161 SDValue Result(0, 0);
10163 switch (Constraint) {
10166 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10167 if (C->getZExtValue() <= 31) {
10168 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10174 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10175 if (C->getZExtValue() <= 63) {
10176 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10182 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10183 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10184 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10190 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10191 if (C->getZExtValue() <= 255) {
10192 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10198 // 32-bit signed value
10199 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10200 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10201 C->getSExtValue())) {
10202 // Widen to 64 bits here to get it sign extended.
10203 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10206 // FIXME gcc accepts some relocatable values here too, but only in certain
10207 // memory models; it's complicated.
10212 // 32-bit unsigned value
10213 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10214 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10215 C->getZExtValue())) {
10216 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10220 // FIXME gcc accepts some relocatable values here too, but only in certain
10221 // memory models; it's complicated.
10225 // Literal immediates are always ok.
10226 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10227 // Widen to 64 bits here to get it sign extended.
10228 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10232 // In any sort of PIC mode addresses need to be computed at runtime by
10233 // adding in a register or some sort of table lookup. These can't
10234 // be used as immediates.
10235 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC() ||
10236 Subtarget->isPICStyleRIPRel())
10239 // If we are in non-pic codegen mode, we allow the address of a global (with
10240 // an optional displacement) to be used with 'i'.
10241 GlobalAddressSDNode *GA = 0;
10242 int64_t Offset = 0;
10244 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10246 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10247 Offset += GA->getOffset();
10249 } else if (Op.getOpcode() == ISD::ADD) {
10250 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10251 Offset += C->getZExtValue();
10252 Op = Op.getOperand(0);
10255 } else if (Op.getOpcode() == ISD::SUB) {
10256 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10257 Offset += -C->getZExtValue();
10258 Op = Op.getOperand(0);
10263 // Otherwise, this isn't something we can handle, reject it.
10267 const GlobalValue *GV = GA->getGlobal();
10268 // If we require an extra load to get this address, as in PIC mode, we
10269 // can't accept it.
10270 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10271 getTargetMachine())))
10274 Result = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10279 if (Result.getNode()) {
10280 Ops.push_back(Result);
10283 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10286 std::vector<unsigned> X86TargetLowering::
10287 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10289 if (Constraint.size() == 1) {
10290 // FIXME: not handling fp-stack yet!
10291 switch (Constraint[0]) { // GCC X86 Constraint Letters
10292 default: break; // Unknown constraint letter
10293 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10294 if (Subtarget->is64Bit()) {
10295 if (VT == MVT::i32)
10296 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10297 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10298 X86::R10D,X86::R11D,X86::R12D,
10299 X86::R13D,X86::R14D,X86::R15D,
10300 X86::EBP, X86::ESP, 0);
10301 else if (VT == MVT::i16)
10302 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10303 X86::SI, X86::DI, X86::R8W,X86::R9W,
10304 X86::R10W,X86::R11W,X86::R12W,
10305 X86::R13W,X86::R14W,X86::R15W,
10306 X86::BP, X86::SP, 0);
10307 else if (VT == MVT::i8)
10308 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10309 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10310 X86::R10B,X86::R11B,X86::R12B,
10311 X86::R13B,X86::R14B,X86::R15B,
10312 X86::BPL, X86::SPL, 0);
10314 else if (VT == MVT::i64)
10315 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10316 X86::RSI, X86::RDI, X86::R8, X86::R9,
10317 X86::R10, X86::R11, X86::R12,
10318 X86::R13, X86::R14, X86::R15,
10319 X86::RBP, X86::RSP, 0);
10323 // 32-bit fallthrough
10324 case 'Q': // Q_REGS
10325 if (VT == MVT::i32)
10326 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10327 else if (VT == MVT::i16)
10328 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10329 else if (VT == MVT::i8)
10330 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10331 else if (VT == MVT::i64)
10332 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10337 return std::vector<unsigned>();
10340 std::pair<unsigned, const TargetRegisterClass*>
10341 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10343 // First, see if this is a constraint that directly corresponds to an LLVM
10345 if (Constraint.size() == 1) {
10346 // GCC Constraint Letters
10347 switch (Constraint[0]) {
10349 case 'r': // GENERAL_REGS
10350 case 'l': // INDEX_REGS
10352 return std::make_pair(0U, X86::GR8RegisterClass);
10353 if (VT == MVT::i16)
10354 return std::make_pair(0U, X86::GR16RegisterClass);
10355 if (VT == MVT::i32 || !Subtarget->is64Bit())
10356 return std::make_pair(0U, X86::GR32RegisterClass);
10357 return std::make_pair(0U, X86::GR64RegisterClass);
10358 case 'R': // LEGACY_REGS
10360 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10361 if (VT == MVT::i16)
10362 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10363 if (VT == MVT::i32 || !Subtarget->is64Bit())
10364 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10365 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10366 case 'f': // FP Stack registers.
10367 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10368 // value to the correct fpstack register class.
10369 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10370 return std::make_pair(0U, X86::RFP32RegisterClass);
10371 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10372 return std::make_pair(0U, X86::RFP64RegisterClass);
10373 return std::make_pair(0U, X86::RFP80RegisterClass);
10374 case 'y': // MMX_REGS if MMX allowed.
10375 if (!Subtarget->hasMMX()) break;
10376 return std::make_pair(0U, X86::VR64RegisterClass);
10377 case 'Y': // SSE_REGS if SSE2 allowed
10378 if (!Subtarget->hasSSE2()) break;
10380 case 'x': // SSE_REGS if SSE1 allowed
10381 if (!Subtarget->hasSSE1()) break;
10383 switch (VT.getSimpleVT().SimpleTy) {
10385 // Scalar SSE types.
10388 return std::make_pair(0U, X86::FR32RegisterClass);
10391 return std::make_pair(0U, X86::FR64RegisterClass);
10399 return std::make_pair(0U, X86::VR128RegisterClass);
10405 // Use the default implementation in TargetLowering to convert the register
10406 // constraint into a member of a register class.
10407 std::pair<unsigned, const TargetRegisterClass*> Res;
10408 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10410 // Not found as a standard register?
10411 if (Res.second == 0) {
10412 // Map st(0) -> st(7) -> ST0
10413 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10414 tolower(Constraint[1]) == 's' &&
10415 tolower(Constraint[2]) == 't' &&
10416 Constraint[3] == '(' &&
10417 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10418 Constraint[5] == ')' &&
10419 Constraint[6] == '}') {
10421 Res.first = X86::ST0+Constraint[4]-'0';
10422 Res.second = X86::RFP80RegisterClass;
10426 // GCC allows "st(0)" to be called just plain "st".
10427 if (StringRef("{st}").equals_lower(Constraint)) {
10428 Res.first = X86::ST0;
10429 Res.second = X86::RFP80RegisterClass;
10434 if (StringRef("{flags}").equals_lower(Constraint)) {
10435 Res.first = X86::EFLAGS;
10436 Res.second = X86::CCRRegisterClass;
10440 // 'A' means EAX + EDX.
10441 if (Constraint == "A") {
10442 Res.first = X86::EAX;
10443 Res.second = X86::GR32_ADRegisterClass;
10449 // Otherwise, check to see if this is a register class of the wrong value
10450 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10451 // turn into {ax},{dx}.
10452 if (Res.second->hasType(VT))
10453 return Res; // Correct type already, nothing to do.
10455 // All of the single-register GCC register classes map their values onto
10456 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10457 // really want an 8-bit or 32-bit register, map to the appropriate register
10458 // class and return the appropriate register.
10459 if (Res.second == X86::GR16RegisterClass) {
10460 if (VT == MVT::i8) {
10461 unsigned DestReg = 0;
10462 switch (Res.first) {
10464 case X86::AX: DestReg = X86::AL; break;
10465 case X86::DX: DestReg = X86::DL; break;
10466 case X86::CX: DestReg = X86::CL; break;
10467 case X86::BX: DestReg = X86::BL; break;
10470 Res.first = DestReg;
10471 Res.second = X86::GR8RegisterClass;
10473 } else if (VT == MVT::i32) {
10474 unsigned DestReg = 0;
10475 switch (Res.first) {
10477 case X86::AX: DestReg = X86::EAX; break;
10478 case X86::DX: DestReg = X86::EDX; break;
10479 case X86::CX: DestReg = X86::ECX; break;
10480 case X86::BX: DestReg = X86::EBX; break;
10481 case X86::SI: DestReg = X86::ESI; break;
10482 case X86::DI: DestReg = X86::EDI; break;
10483 case X86::BP: DestReg = X86::EBP; break;
10484 case X86::SP: DestReg = X86::ESP; break;
10487 Res.first = DestReg;
10488 Res.second = X86::GR32RegisterClass;
10490 } else if (VT == MVT::i64) {
10491 unsigned DestReg = 0;
10492 switch (Res.first) {
10494 case X86::AX: DestReg = X86::RAX; break;
10495 case X86::DX: DestReg = X86::RDX; break;
10496 case X86::CX: DestReg = X86::RCX; break;
10497 case X86::BX: DestReg = X86::RBX; break;
10498 case X86::SI: DestReg = X86::RSI; break;
10499 case X86::DI: DestReg = X86::RDI; break;
10500 case X86::BP: DestReg = X86::RBP; break;
10501 case X86::SP: DestReg = X86::RSP; break;
10504 Res.first = DestReg;
10505 Res.second = X86::GR64RegisterClass;
10508 } else if (Res.second == X86::FR32RegisterClass ||
10509 Res.second == X86::FR64RegisterClass ||
10510 Res.second == X86::VR128RegisterClass) {
10511 // Handle references to XMM physical registers that got mapped into the
10512 // wrong class. This can happen with constraints like {xmm0} where the
10513 // target independent register mapper will just pick the first match it can
10514 // find, ignoring the required type.
10515 if (VT == MVT::f32)
10516 Res.second = X86::FR32RegisterClass;
10517 else if (VT == MVT::f64)
10518 Res.second = X86::FR64RegisterClass;
10519 else if (X86::VR128RegisterClass->hasType(VT))
10520 Res.second = X86::VR128RegisterClass;