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 // Disable16Bit - 16-bit operations typically have a larger encoding than
61 // corresponding 32-bit instructions, and 16-bit code is slow on some
62 // processors. This is an experimental flag to disable 16-bit operations
63 // (which forces them to be Legalized to 32-bit operations).
65 Disable16Bit("disable-16bit", cl::Hidden,
66 cl::desc("Disable use of 16-bit instructions"));
68 Promote16Bit("promote-16bit", cl::Hidden,
69 cl::desc("Promote 16-bit instructions"));
71 // Forward declarations.
72 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
75 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
76 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
77 default: llvm_unreachable("unknown subtarget type");
78 case X86Subtarget::isDarwin:
79 if (TM.getSubtarget<X86Subtarget>().is64Bit())
80 return new X8664_MachoTargetObjectFile();
81 return new TargetLoweringObjectFileMachO();
82 case X86Subtarget::isELF:
83 if (TM.getSubtarget<X86Subtarget>().is64Bit())
84 return new X8664_ELFTargetObjectFile(TM);
85 return new X8632_ELFTargetObjectFile(TM);
86 case X86Subtarget::isMingw:
87 case X86Subtarget::isCygwin:
88 case X86Subtarget::isWindows:
89 return new TargetLoweringObjectFileCOFF();
93 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
94 : TargetLowering(TM, createTLOF(TM)) {
95 Subtarget = &TM.getSubtarget<X86Subtarget>();
96 X86ScalarSSEf64 = Subtarget->hasSSE2();
97 X86ScalarSSEf32 = Subtarget->hasSSE1();
98 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
100 RegInfo = TM.getRegisterInfo();
101 TD = getTargetData();
103 // Set up the TargetLowering object.
105 // X86 is weird, it always uses i8 for shift amounts and setcc results.
106 setShiftAmountType(MVT::i8);
107 setBooleanContents(ZeroOrOneBooleanContent);
108 setSchedulingPreference(SchedulingForRegPressure);
109 setStackPointerRegisterToSaveRestore(X86StackPtr);
111 if (Subtarget->isTargetDarwin()) {
112 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
113 setUseUnderscoreSetJmp(false);
114 setUseUnderscoreLongJmp(false);
115 } else if (Subtarget->isTargetMingw()) {
116 // MS runtime is weird: it exports _setjmp, but longjmp!
117 setUseUnderscoreSetJmp(true);
118 setUseUnderscoreLongJmp(false);
120 setUseUnderscoreSetJmp(true);
121 setUseUnderscoreLongJmp(true);
124 // Set up the register classes.
125 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
127 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
128 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
129 if (Subtarget->is64Bit())
130 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
132 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
134 // We don't accept any truncstore of integer registers.
135 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
137 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
138 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
140 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
141 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
142 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
144 // SETOEQ and SETUNE require checking two conditions.
145 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
146 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
147 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
148 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
149 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
150 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
152 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
154 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
155 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
156 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
158 if (Subtarget->is64Bit()) {
159 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
160 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
161 } else if (!UseSoftFloat) {
162 if (X86ScalarSSEf64) {
163 // We have an impenetrably clever algorithm for ui64->double only.
164 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
166 // We have an algorithm for SSE2, and we turn this into a 64-bit
167 // FILD for other targets.
168 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
171 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
173 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
174 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
177 // SSE has no i16 to fp conversion, only i32
178 if (X86ScalarSSEf32) {
179 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
180 // f32 and f64 cases are Legal, f80 case is not
181 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
183 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
184 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
187 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
188 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
191 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
192 // are Legal, f80 is custom lowered.
193 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
194 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
196 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
198 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
199 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
201 if (X86ScalarSSEf32) {
202 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
203 // f32 and f64 cases are Legal, f80 case is not
204 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
206 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
207 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
210 // Handle FP_TO_UINT by promoting the destination to a larger signed
212 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
213 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
214 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
216 if (Subtarget->is64Bit()) {
217 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
218 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
219 } else if (!UseSoftFloat) {
220 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
221 // Expand FP_TO_UINT into a select.
222 // FIXME: We would like to use a Custom expander here eventually to do
223 // the optimal thing for SSE vs. the default expansion in the legalizer.
224 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
226 // With SSE3 we can use fisttpll to convert to a signed i64; without
227 // SSE, we're stuck with a fistpll.
228 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
231 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
232 if (!X86ScalarSSEf64) {
233 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
234 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
237 // Scalar integer divide and remainder are lowered to use operations that
238 // produce two results, to match the available instructions. This exposes
239 // the two-result form to trivial CSE, which is able to combine x/y and x%y
240 // into a single instruction.
242 // Scalar integer multiply-high is also lowered to use two-result
243 // operations, to match the available instructions. However, plain multiply
244 // (low) operations are left as Legal, as there are single-result
245 // instructions for this in x86. Using the two-result multiply instructions
246 // when both high and low results are needed must be arranged by dagcombine.
247 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
248 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
249 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
250 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
251 setOperationAction(ISD::SREM , MVT::i8 , Expand);
252 setOperationAction(ISD::UREM , MVT::i8 , Expand);
253 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
254 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
255 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
256 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
257 setOperationAction(ISD::SREM , MVT::i16 , Expand);
258 setOperationAction(ISD::UREM , MVT::i16 , Expand);
259 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
260 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
261 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
262 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
263 setOperationAction(ISD::SREM , MVT::i32 , Expand);
264 setOperationAction(ISD::UREM , MVT::i32 , Expand);
265 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
266 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
267 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
268 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
269 setOperationAction(ISD::SREM , MVT::i64 , Expand);
270 setOperationAction(ISD::UREM , MVT::i64 , Expand);
272 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
273 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
274 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
275 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
276 if (Subtarget->is64Bit())
277 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
278 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
279 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
280 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
281 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
282 setOperationAction(ISD::FREM , MVT::f32 , Expand);
283 setOperationAction(ISD::FREM , MVT::f64 , Expand);
284 setOperationAction(ISD::FREM , MVT::f80 , Expand);
285 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
287 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
288 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
289 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
290 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
292 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
293 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
295 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
296 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
298 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
299 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
300 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
301 if (Subtarget->is64Bit()) {
302 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
303 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
304 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
307 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
308 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
310 // These should be promoted to a larger select which is supported.
311 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
312 // X86 wants to expand cmov itself.
313 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
315 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
317 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
318 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
319 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
320 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
321 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
322 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
324 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
326 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
327 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
328 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
329 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
330 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
331 if (Subtarget->is64Bit()) {
332 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
333 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
335 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
338 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
339 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
340 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
341 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
342 if (Subtarget->is64Bit())
343 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
344 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
345 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
346 if (Subtarget->is64Bit()) {
347 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
348 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
349 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
350 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
351 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
353 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
354 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
355 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
356 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
357 if (Subtarget->is64Bit()) {
358 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
359 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
360 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
363 if (Subtarget->hasSSE1())
364 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
366 if (!Subtarget->hasSSE2())
367 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
369 // Expand certain atomics
370 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
371 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
372 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
373 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
375 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
376 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
377 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
378 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
380 if (!Subtarget->is64Bit()) {
381 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
382 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
383 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
384 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
385 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
386 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
387 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
390 // FIXME - use subtarget debug flags
391 if (!Subtarget->isTargetDarwin() &&
392 !Subtarget->isTargetELF() &&
393 !Subtarget->isTargetCygMing()) {
394 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
397 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
398 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
399 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
400 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
401 if (Subtarget->is64Bit()) {
402 setExceptionPointerRegister(X86::RAX);
403 setExceptionSelectorRegister(X86::RDX);
405 setExceptionPointerRegister(X86::EAX);
406 setExceptionSelectorRegister(X86::EDX);
408 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
409 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
411 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
413 setOperationAction(ISD::TRAP, MVT::Other, Legal);
415 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
416 setOperationAction(ISD::VASTART , MVT::Other, Custom);
417 setOperationAction(ISD::VAEND , MVT::Other, Expand);
418 if (Subtarget->is64Bit()) {
419 setOperationAction(ISD::VAARG , MVT::Other, Custom);
420 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
422 setOperationAction(ISD::VAARG , MVT::Other, Expand);
423 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
426 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
427 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
428 if (Subtarget->is64Bit())
429 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
430 if (Subtarget->isTargetCygMing())
431 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
433 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
435 if (!UseSoftFloat && X86ScalarSSEf64) {
436 // f32 and f64 use SSE.
437 // Set up the FP register classes.
438 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
439 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
441 // Use ANDPD to simulate FABS.
442 setOperationAction(ISD::FABS , MVT::f64, Custom);
443 setOperationAction(ISD::FABS , MVT::f32, Custom);
445 // Use XORP to simulate FNEG.
446 setOperationAction(ISD::FNEG , MVT::f64, Custom);
447 setOperationAction(ISD::FNEG , MVT::f32, Custom);
449 // Use ANDPD and ORPD to simulate FCOPYSIGN.
450 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
451 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
453 // We don't support sin/cos/fmod
454 setOperationAction(ISD::FSIN , MVT::f64, Expand);
455 setOperationAction(ISD::FCOS , MVT::f64, Expand);
456 setOperationAction(ISD::FSIN , MVT::f32, Expand);
457 setOperationAction(ISD::FCOS , MVT::f32, Expand);
459 // Expand FP immediates into loads from the stack, except for the special
461 addLegalFPImmediate(APFloat(+0.0)); // xorpd
462 addLegalFPImmediate(APFloat(+0.0f)); // xorps
463 } else if (!UseSoftFloat && X86ScalarSSEf32) {
464 // Use SSE for f32, x87 for f64.
465 // Set up the FP register classes.
466 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
467 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
469 // Use ANDPS to simulate FABS.
470 setOperationAction(ISD::FABS , MVT::f32, Custom);
472 // Use XORP to simulate FNEG.
473 setOperationAction(ISD::FNEG , MVT::f32, Custom);
475 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
477 // Use ANDPS and ORPS to simulate FCOPYSIGN.
478 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
479 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
481 // We don't support sin/cos/fmod
482 setOperationAction(ISD::FSIN , MVT::f32, Expand);
483 setOperationAction(ISD::FCOS , MVT::f32, Expand);
485 // Special cases we handle for FP constants.
486 addLegalFPImmediate(APFloat(+0.0f)); // xorps
487 addLegalFPImmediate(APFloat(+0.0)); // FLD0
488 addLegalFPImmediate(APFloat(+1.0)); // FLD1
489 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
490 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
493 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
494 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
496 } else if (!UseSoftFloat) {
497 // f32 and f64 in x87.
498 // Set up the FP register classes.
499 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
500 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
502 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
503 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
504 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
505 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
508 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
509 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
511 addLegalFPImmediate(APFloat(+0.0)); // FLD0
512 addLegalFPImmediate(APFloat(+1.0)); // FLD1
513 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
514 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
515 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
516 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
517 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
518 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
521 // Long double always uses X87.
523 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
524 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
525 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
528 APFloat TmpFlt(+0.0);
529 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
531 addLegalFPImmediate(TmpFlt); // FLD0
533 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
534 APFloat TmpFlt2(+1.0);
535 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
537 addLegalFPImmediate(TmpFlt2); // FLD1
538 TmpFlt2.changeSign();
539 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
543 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
544 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
548 // Always use a library call for pow.
549 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
550 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
551 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
553 setOperationAction(ISD::FLOG, MVT::f80, Expand);
554 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
555 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
556 setOperationAction(ISD::FEXP, MVT::f80, Expand);
557 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
559 // First set operation action for all vector types to either promote
560 // (for widening) or expand (for scalarization). Then we will selectively
561 // turn on ones that can be effectively codegen'd.
562 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
563 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
564 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
579 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
580 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
607 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
608 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
609 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
610 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
611 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
612 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
613 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
614 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
615 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
616 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
617 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
618 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
619 setTruncStoreAction((MVT::SimpleValueType)VT,
620 (MVT::SimpleValueType)InnerVT, Expand);
621 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
622 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
623 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
626 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
627 // with -msoft-float, disable use of MMX as well.
628 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
629 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
630 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
631 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
632 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
633 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
635 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
636 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
637 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
638 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
640 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
641 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
642 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
643 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
645 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
646 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
648 setOperationAction(ISD::AND, MVT::v8i8, Promote);
649 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
650 setOperationAction(ISD::AND, MVT::v4i16, Promote);
651 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
652 setOperationAction(ISD::AND, MVT::v2i32, Promote);
653 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
654 setOperationAction(ISD::AND, MVT::v1i64, Legal);
656 setOperationAction(ISD::OR, MVT::v8i8, Promote);
657 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
658 setOperationAction(ISD::OR, MVT::v4i16, Promote);
659 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
660 setOperationAction(ISD::OR, MVT::v2i32, Promote);
661 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
662 setOperationAction(ISD::OR, MVT::v1i64, Legal);
664 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
665 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
666 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
667 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
668 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
669 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
670 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
672 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
673 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
674 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
675 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
676 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
677 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
678 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
679 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
680 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
682 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
683 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
684 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
685 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
686 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
688 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
689 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
690 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
691 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
693 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
694 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
695 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
696 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
698 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
700 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
701 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
702 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
703 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
704 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
705 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
706 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
709 if (!UseSoftFloat && Subtarget->hasSSE1()) {
710 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
712 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
713 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
714 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
715 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
716 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
717 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
718 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
719 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
720 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
721 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
722 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
723 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
726 if (!UseSoftFloat && Subtarget->hasSSE2()) {
727 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
729 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
730 // registers cannot be used even for integer operations.
731 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
732 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
733 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
734 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
736 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
737 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
738 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
739 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
740 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
741 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
742 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
743 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
744 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
745 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
746 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
747 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
748 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
749 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
750 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
751 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
753 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
754 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
755 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
756 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
758 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
759 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
760 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
761 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
762 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
764 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
765 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
766 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
767 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
768 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
770 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
771 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
772 EVT VT = (MVT::SimpleValueType)i;
773 // Do not attempt to custom lower non-power-of-2 vectors
774 if (!isPowerOf2_32(VT.getVectorNumElements()))
776 // Do not attempt to custom lower non-128-bit vectors
777 if (!VT.is128BitVector())
779 setOperationAction(ISD::BUILD_VECTOR,
780 VT.getSimpleVT().SimpleTy, Custom);
781 setOperationAction(ISD::VECTOR_SHUFFLE,
782 VT.getSimpleVT().SimpleTy, Custom);
783 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
784 VT.getSimpleVT().SimpleTy, Custom);
787 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
788 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
789 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
790 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
791 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
792 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
794 if (Subtarget->is64Bit()) {
795 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
796 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
799 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
800 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
801 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
804 // Do not attempt to promote non-128-bit vectors
805 if (!VT.is128BitVector()) {
809 setOperationAction(ISD::AND, SVT, Promote);
810 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
811 setOperationAction(ISD::OR, SVT, Promote);
812 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
813 setOperationAction(ISD::XOR, SVT, Promote);
814 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
815 setOperationAction(ISD::LOAD, SVT, Promote);
816 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
817 setOperationAction(ISD::SELECT, SVT, Promote);
818 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
821 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
823 // Custom lower v2i64 and v2f64 selects.
824 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
825 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
826 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
827 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
829 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
830 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
831 if (!DisableMMX && Subtarget->hasMMX()) {
832 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
833 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
837 if (Subtarget->hasSSE41()) {
838 // FIXME: Do we need to handle scalar-to-vector here?
839 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
841 // i8 and i16 vectors are custom , because the source register and source
842 // source memory operand types are not the same width. f32 vectors are
843 // custom since the immediate controlling the insert encodes additional
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
848 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
851 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
852 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
853 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
855 if (Subtarget->is64Bit()) {
856 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
857 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
861 if (Subtarget->hasSSE42()) {
862 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
865 if (!UseSoftFloat && Subtarget->hasAVX()) {
866 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
867 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
868 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
869 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
871 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
872 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
873 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
874 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
875 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
876 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
877 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
878 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
879 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
880 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
881 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
882 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
883 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
884 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
885 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
887 // Operations to consider commented out -v16i16 v32i8
888 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
889 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
890 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
891 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
892 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
893 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
894 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
895 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
896 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
897 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
898 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
899 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
900 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
901 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
903 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
904 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
905 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
906 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
908 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
909 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
910 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
912 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
914 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
915 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
916 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
917 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
918 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
919 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
922 // Not sure we want to do this since there are no 256-bit integer
925 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
926 // This includes 256-bit vectors
927 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
928 EVT VT = (MVT::SimpleValueType)i;
930 // Do not attempt to custom lower non-power-of-2 vectors
931 if (!isPowerOf2_32(VT.getVectorNumElements()))
934 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
935 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
936 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
939 if (Subtarget->is64Bit()) {
940 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
941 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
946 // Not sure we want to do this since there are no 256-bit integer
949 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
950 // Including 256-bit vectors
951 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
952 EVT VT = (MVT::SimpleValueType)i;
954 if (!VT.is256BitVector()) {
957 setOperationAction(ISD::AND, VT, Promote);
958 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
959 setOperationAction(ISD::OR, VT, Promote);
960 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
961 setOperationAction(ISD::XOR, VT, Promote);
962 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
963 setOperationAction(ISD::LOAD, VT, Promote);
964 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
965 setOperationAction(ISD::SELECT, VT, Promote);
966 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
969 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
973 // We want to custom lower some of our intrinsics.
974 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
976 // Add/Sub/Mul with overflow operations are custom lowered.
977 setOperationAction(ISD::SADDO, MVT::i32, Custom);
978 setOperationAction(ISD::SADDO, MVT::i64, Custom);
979 setOperationAction(ISD::UADDO, MVT::i32, Custom);
980 setOperationAction(ISD::UADDO, MVT::i64, Custom);
981 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
982 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
983 setOperationAction(ISD::USUBO, MVT::i32, Custom);
984 setOperationAction(ISD::USUBO, MVT::i64, Custom);
985 setOperationAction(ISD::SMULO, MVT::i32, Custom);
986 setOperationAction(ISD::SMULO, MVT::i64, Custom);
988 if (!Subtarget->is64Bit()) {
989 // These libcalls are not available in 32-bit.
990 setLibcallName(RTLIB::SHL_I128, 0);
991 setLibcallName(RTLIB::SRL_I128, 0);
992 setLibcallName(RTLIB::SRA_I128, 0);
995 // We have target-specific dag combine patterns for the following nodes:
996 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
997 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
998 setTargetDAGCombine(ISD::BUILD_VECTOR);
999 setTargetDAGCombine(ISD::SELECT);
1000 setTargetDAGCombine(ISD::SHL);
1001 setTargetDAGCombine(ISD::SRA);
1002 setTargetDAGCombine(ISD::SRL);
1003 setTargetDAGCombine(ISD::OR);
1004 setTargetDAGCombine(ISD::STORE);
1005 setTargetDAGCombine(ISD::MEMBARRIER);
1006 setTargetDAGCombine(ISD::ZERO_EXTEND);
1007 if (Subtarget->is64Bit())
1008 setTargetDAGCombine(ISD::MUL);
1010 computeRegisterProperties();
1012 // FIXME: These should be based on subtarget info. Plus, the values should
1013 // be smaller when we are in optimizing for size mode.
1014 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1015 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1016 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1017 setPrefLoopAlignment(16);
1018 benefitFromCodePlacementOpt = true;
1022 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1027 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1028 /// the desired ByVal argument alignment.
1029 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1032 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1033 if (VTy->getBitWidth() == 128)
1035 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1036 unsigned EltAlign = 0;
1037 getMaxByValAlign(ATy->getElementType(), EltAlign);
1038 if (EltAlign > MaxAlign)
1039 MaxAlign = EltAlign;
1040 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1041 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1042 unsigned EltAlign = 0;
1043 getMaxByValAlign(STy->getElementType(i), EltAlign);
1044 if (EltAlign > MaxAlign)
1045 MaxAlign = EltAlign;
1053 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1054 /// function arguments in the caller parameter area. For X86, aggregates
1055 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1056 /// are at 4-byte boundaries.
1057 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1058 if (Subtarget->is64Bit()) {
1059 // Max of 8 and alignment of type.
1060 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1067 if (Subtarget->hasSSE1())
1068 getMaxByValAlign(Ty, Align);
1072 /// getOptimalMemOpType - Returns the target specific optimal type for load
1073 /// and store operations as a result of memset, memcpy, and memmove
1074 /// lowering. If DstAlign is zero that means it's safe to destination
1075 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1076 /// means there isn't a need to check it against alignment requirement,
1077 /// probably because the source does not need to be loaded. If
1078 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1079 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1080 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1081 /// constant so it does not need to be loaded.
1082 /// It returns EVT::Other if SelectionDAG should be responsible for
1083 /// determining the type.
1085 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1086 unsigned DstAlign, unsigned SrcAlign,
1087 bool NonScalarIntSafe,
1089 SelectionDAG &DAG) const {
1090 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1091 // linux. This is because the stack realignment code can't handle certain
1092 // cases like PR2962. This should be removed when PR2962 is fixed.
1093 const Function *F = DAG.getMachineFunction().getFunction();
1094 if (NonScalarIntSafe &&
1095 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1097 (Subtarget->isUnalignedMemAccessFast() ||
1098 ((DstAlign == 0 || DstAlign >= 16) &&
1099 (SrcAlign == 0 || SrcAlign >= 16))) &&
1100 Subtarget->getStackAlignment() >= 16) {
1101 if (Subtarget->hasSSE2())
1103 if (Subtarget->hasSSE1())
1105 } else if (!MemcpyStrSrc && Size >= 8 &&
1106 !Subtarget->is64Bit() &&
1107 Subtarget->getStackAlignment() >= 8 &&
1108 Subtarget->hasSSE2()) {
1109 // Do not use f64 to lower memcpy if source is string constant. It's
1110 // better to use i32 to avoid the loads.
1114 if (Subtarget->is64Bit() && Size >= 8)
1119 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1120 /// current function. The returned value is a member of the
1121 /// MachineJumpTableInfo::JTEntryKind enum.
1122 unsigned X86TargetLowering::getJumpTableEncoding() const {
1123 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1125 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1126 Subtarget->isPICStyleGOT())
1127 return MachineJumpTableInfo::EK_Custom32;
1129 // Otherwise, use the normal jump table encoding heuristics.
1130 return TargetLowering::getJumpTableEncoding();
1133 /// getPICBaseSymbol - Return the X86-32 PIC base.
1135 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1136 MCContext &Ctx) const {
1137 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1138 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1139 Twine(MF->getFunctionNumber())+"$pb");
1144 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1145 const MachineBasicBlock *MBB,
1146 unsigned uid,MCContext &Ctx) const{
1147 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1148 Subtarget->isPICStyleGOT());
1149 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1151 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1152 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1155 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1157 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1158 SelectionDAG &DAG) const {
1159 if (!Subtarget->is64Bit())
1160 // This doesn't have DebugLoc associated with it, but is not really the
1161 // same as a Register.
1162 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1166 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1167 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1169 const MCExpr *X86TargetLowering::
1170 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1171 MCContext &Ctx) const {
1172 // X86-64 uses RIP relative addressing based on the jump table label.
1173 if (Subtarget->isPICStyleRIPRel())
1174 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1176 // Otherwise, the reference is relative to the PIC base.
1177 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1180 /// getFunctionAlignment - Return the Log2 alignment of this function.
1181 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1182 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1185 //===----------------------------------------------------------------------===//
1186 // Return Value Calling Convention Implementation
1187 //===----------------------------------------------------------------------===//
1189 #include "X86GenCallingConv.inc"
1192 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1193 const SmallVectorImpl<EVT> &OutTys,
1194 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1195 SelectionDAG &DAG) {
1196 SmallVector<CCValAssign, 16> RVLocs;
1197 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1198 RVLocs, *DAG.getContext());
1199 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1203 X86TargetLowering::LowerReturn(SDValue Chain,
1204 CallingConv::ID CallConv, bool isVarArg,
1205 const SmallVectorImpl<ISD::OutputArg> &Outs,
1206 DebugLoc dl, SelectionDAG &DAG) {
1208 SmallVector<CCValAssign, 16> RVLocs;
1209 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1210 RVLocs, *DAG.getContext());
1211 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1213 // Add the regs to the liveout set for the function.
1214 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1215 for (unsigned i = 0; i != RVLocs.size(); ++i)
1216 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1217 MRI.addLiveOut(RVLocs[i].getLocReg());
1221 SmallVector<SDValue, 6> RetOps;
1222 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1223 // Operand #1 = Bytes To Pop
1224 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1226 // Copy the result values into the output registers.
1227 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1228 CCValAssign &VA = RVLocs[i];
1229 assert(VA.isRegLoc() && "Can only return in registers!");
1230 SDValue ValToCopy = Outs[i].Val;
1232 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1233 // the RET instruction and handled by the FP Stackifier.
1234 if (VA.getLocReg() == X86::ST0 ||
1235 VA.getLocReg() == X86::ST1) {
1236 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1237 // change the value to the FP stack register class.
1238 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1239 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1240 RetOps.push_back(ValToCopy);
1241 // Don't emit a copytoreg.
1245 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1246 // which is returned in RAX / RDX.
1247 if (Subtarget->is64Bit()) {
1248 EVT ValVT = ValToCopy.getValueType();
1249 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1250 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1251 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1252 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1256 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1257 Flag = Chain.getValue(1);
1260 // The x86-64 ABI for returning structs by value requires that we copy
1261 // the sret argument into %rax for the return. We saved the argument into
1262 // a virtual register in the entry block, so now we copy the value out
1264 if (Subtarget->is64Bit() &&
1265 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1266 MachineFunction &MF = DAG.getMachineFunction();
1267 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1268 unsigned Reg = FuncInfo->getSRetReturnReg();
1270 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1271 FuncInfo->setSRetReturnReg(Reg);
1273 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1275 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1276 Flag = Chain.getValue(1);
1278 // RAX now acts like a return value.
1279 MRI.addLiveOut(X86::RAX);
1282 RetOps[0] = Chain; // Update chain.
1284 // Add the flag if we have it.
1286 RetOps.push_back(Flag);
1288 return DAG.getNode(X86ISD::RET_FLAG, dl,
1289 MVT::Other, &RetOps[0], RetOps.size());
1292 /// LowerCallResult - Lower the result values of a call into the
1293 /// appropriate copies out of appropriate physical registers.
1296 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1297 CallingConv::ID CallConv, bool isVarArg,
1298 const SmallVectorImpl<ISD::InputArg> &Ins,
1299 DebugLoc dl, SelectionDAG &DAG,
1300 SmallVectorImpl<SDValue> &InVals) {
1302 // Assign locations to each value returned by this call.
1303 SmallVector<CCValAssign, 16> RVLocs;
1304 bool Is64Bit = Subtarget->is64Bit();
1305 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1306 RVLocs, *DAG.getContext());
1307 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1309 // Copy all of the result registers out of their specified physreg.
1310 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1311 CCValAssign &VA = RVLocs[i];
1312 EVT CopyVT = VA.getValVT();
1314 // If this is x86-64, and we disabled SSE, we can't return FP values
1315 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1316 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1317 report_fatal_error("SSE register return with SSE disabled");
1320 // If this is a call to a function that returns an fp value on the floating
1321 // point stack, but where we prefer to use the value in xmm registers, copy
1322 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1323 if ((VA.getLocReg() == X86::ST0 ||
1324 VA.getLocReg() == X86::ST1) &&
1325 isScalarFPTypeInSSEReg(VA.getValVT())) {
1330 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1331 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1332 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1333 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1334 MVT::v2i64, InFlag).getValue(1);
1335 Val = Chain.getValue(0);
1336 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1337 Val, DAG.getConstant(0, MVT::i64));
1339 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1340 MVT::i64, InFlag).getValue(1);
1341 Val = Chain.getValue(0);
1343 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1345 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1346 CopyVT, InFlag).getValue(1);
1347 Val = Chain.getValue(0);
1349 InFlag = Chain.getValue(2);
1351 if (CopyVT != VA.getValVT()) {
1352 // Round the F80 the right size, which also moves to the appropriate xmm
1354 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1355 // This truncation won't change the value.
1356 DAG.getIntPtrConstant(1));
1359 InVals.push_back(Val);
1366 //===----------------------------------------------------------------------===//
1367 // C & StdCall & Fast Calling Convention implementation
1368 //===----------------------------------------------------------------------===//
1369 // StdCall calling convention seems to be standard for many Windows' API
1370 // routines and around. It differs from C calling convention just a little:
1371 // callee should clean up the stack, not caller. Symbols should be also
1372 // decorated in some fancy way :) It doesn't support any vector arguments.
1373 // For info on fast calling convention see Fast Calling Convention (tail call)
1374 // implementation LowerX86_32FastCCCallTo.
1376 /// CallIsStructReturn - Determines whether a call uses struct return
1378 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1382 return Outs[0].Flags.isSRet();
1385 /// ArgsAreStructReturn - Determines whether a function uses struct
1386 /// return semantics.
1388 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1392 return Ins[0].Flags.isSRet();
1395 /// IsCalleePop - Determines whether the callee is required to pop its
1396 /// own arguments. Callee pop is necessary to support tail calls.
1397 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1401 switch (CallingConv) {
1404 case CallingConv::X86_StdCall:
1405 return !Subtarget->is64Bit();
1406 case CallingConv::X86_FastCall:
1407 return !Subtarget->is64Bit();
1408 case CallingConv::Fast:
1409 return GuaranteedTailCallOpt;
1410 case CallingConv::GHC:
1411 return GuaranteedTailCallOpt;
1415 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1416 /// given CallingConvention value.
1417 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1418 if (Subtarget->is64Bit()) {
1419 if (CC == CallingConv::GHC)
1420 return CC_X86_64_GHC;
1421 else if (Subtarget->isTargetWin64())
1422 return CC_X86_Win64_C;
1427 if (CC == CallingConv::X86_FastCall)
1428 return CC_X86_32_FastCall;
1429 else if (CC == CallingConv::Fast)
1430 return CC_X86_32_FastCC;
1431 else if (CC == CallingConv::GHC)
1432 return CC_X86_32_GHC;
1437 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1438 /// by "Src" to address "Dst" with size and alignment information specified by
1439 /// the specific parameter attribute. The copy will be passed as a byval
1440 /// function parameter.
1442 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1443 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1445 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1446 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1447 /*isVolatile*/false, /*AlwaysInline=*/true,
1451 /// IsTailCallConvention - Return true if the calling convention is one that
1452 /// supports tail call optimization.
1453 static bool IsTailCallConvention(CallingConv::ID CC) {
1454 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1457 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1458 /// a tailcall target by changing its ABI.
1459 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1460 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1464 X86TargetLowering::LowerMemArgument(SDValue Chain,
1465 CallingConv::ID CallConv,
1466 const SmallVectorImpl<ISD::InputArg> &Ins,
1467 DebugLoc dl, SelectionDAG &DAG,
1468 const CCValAssign &VA,
1469 MachineFrameInfo *MFI,
1471 // Create the nodes corresponding to a load from this parameter slot.
1472 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1473 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1474 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1477 // If value is passed by pointer we have address passed instead of the value
1479 if (VA.getLocInfo() == CCValAssign::Indirect)
1480 ValVT = VA.getLocVT();
1482 ValVT = VA.getValVT();
1484 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1485 // changed with more analysis.
1486 // In case of tail call optimization mark all arguments mutable. Since they
1487 // could be overwritten by lowering of arguments in case of a tail call.
1488 if (Flags.isByVal()) {
1489 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1490 VA.getLocMemOffset(), isImmutable, false);
1491 return DAG.getFrameIndex(FI, getPointerTy());
1493 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1494 VA.getLocMemOffset(), isImmutable, false);
1495 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1496 return DAG.getLoad(ValVT, dl, Chain, FIN,
1497 PseudoSourceValue::getFixedStack(FI), 0,
1503 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1504 CallingConv::ID CallConv,
1506 const SmallVectorImpl<ISD::InputArg> &Ins,
1509 SmallVectorImpl<SDValue> &InVals) {
1510 MachineFunction &MF = DAG.getMachineFunction();
1511 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1513 const Function* Fn = MF.getFunction();
1514 if (Fn->hasExternalLinkage() &&
1515 Subtarget->isTargetCygMing() &&
1516 Fn->getName() == "main")
1517 FuncInfo->setForceFramePointer(true);
1519 MachineFrameInfo *MFI = MF.getFrameInfo();
1520 bool Is64Bit = Subtarget->is64Bit();
1521 bool IsWin64 = Subtarget->isTargetWin64();
1523 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1524 "Var args not supported with calling convention fastcc or ghc");
1526 // Assign locations to all of the incoming arguments.
1527 SmallVector<CCValAssign, 16> ArgLocs;
1528 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1529 ArgLocs, *DAG.getContext());
1530 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1532 unsigned LastVal = ~0U;
1534 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1535 CCValAssign &VA = ArgLocs[i];
1536 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1538 assert(VA.getValNo() != LastVal &&
1539 "Don't support value assigned to multiple locs yet");
1540 LastVal = VA.getValNo();
1542 if (VA.isRegLoc()) {
1543 EVT RegVT = VA.getLocVT();
1544 TargetRegisterClass *RC = NULL;
1545 if (RegVT == MVT::i32)
1546 RC = X86::GR32RegisterClass;
1547 else if (Is64Bit && RegVT == MVT::i64)
1548 RC = X86::GR64RegisterClass;
1549 else if (RegVT == MVT::f32)
1550 RC = X86::FR32RegisterClass;
1551 else if (RegVT == MVT::f64)
1552 RC = X86::FR64RegisterClass;
1553 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1554 RC = X86::VR128RegisterClass;
1555 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1556 RC = X86::VR64RegisterClass;
1558 llvm_unreachable("Unknown argument type!");
1560 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1561 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1563 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1564 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1566 if (VA.getLocInfo() == CCValAssign::SExt)
1567 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1568 DAG.getValueType(VA.getValVT()));
1569 else if (VA.getLocInfo() == CCValAssign::ZExt)
1570 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1571 DAG.getValueType(VA.getValVT()));
1572 else if (VA.getLocInfo() == CCValAssign::BCvt)
1573 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1575 if (VA.isExtInLoc()) {
1576 // Handle MMX values passed in XMM regs.
1577 if (RegVT.isVector()) {
1578 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1579 ArgValue, DAG.getConstant(0, MVT::i64));
1580 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1582 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1585 assert(VA.isMemLoc());
1586 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1589 // If value is passed via pointer - do a load.
1590 if (VA.getLocInfo() == CCValAssign::Indirect)
1591 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1594 InVals.push_back(ArgValue);
1597 // The x86-64 ABI for returning structs by value requires that we copy
1598 // the sret argument into %rax for the return. Save the argument into
1599 // a virtual register so that we can access it from the return points.
1600 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1601 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1602 unsigned Reg = FuncInfo->getSRetReturnReg();
1604 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1605 FuncInfo->setSRetReturnReg(Reg);
1607 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1608 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1611 unsigned StackSize = CCInfo.getNextStackOffset();
1612 // Align stack specially for tail calls.
1613 if (FuncIsMadeTailCallSafe(CallConv))
1614 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1616 // If the function takes variable number of arguments, make a frame index for
1617 // the start of the first vararg value... for expansion of llvm.va_start.
1619 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1620 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1623 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1625 // FIXME: We should really autogenerate these arrays
1626 static const unsigned GPR64ArgRegsWin64[] = {
1627 X86::RCX, X86::RDX, X86::R8, X86::R9
1629 static const unsigned XMMArgRegsWin64[] = {
1630 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1632 static const unsigned GPR64ArgRegs64Bit[] = {
1633 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1635 static const unsigned XMMArgRegs64Bit[] = {
1636 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1637 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1639 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1642 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1643 GPR64ArgRegs = GPR64ArgRegsWin64;
1644 XMMArgRegs = XMMArgRegsWin64;
1646 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1647 GPR64ArgRegs = GPR64ArgRegs64Bit;
1648 XMMArgRegs = XMMArgRegs64Bit;
1650 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1652 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1655 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1656 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1657 "SSE register cannot be used when SSE is disabled!");
1658 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1659 "SSE register cannot be used when SSE is disabled!");
1660 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1661 // Kernel mode asks for SSE to be disabled, so don't push them
1663 TotalNumXMMRegs = 0;
1665 // For X86-64, if there are vararg parameters that are passed via
1666 // registers, then we must store them to their spots on the stack so they
1667 // may be loaded by deferencing the result of va_next.
1668 VarArgsGPOffset = NumIntRegs * 8;
1669 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1670 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1671 TotalNumXMMRegs * 16, 16,
1674 // Store the integer parameter registers.
1675 SmallVector<SDValue, 8> MemOps;
1676 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1677 unsigned Offset = VarArgsGPOffset;
1678 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1679 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1680 DAG.getIntPtrConstant(Offset));
1681 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1682 X86::GR64RegisterClass);
1683 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1685 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1686 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1687 Offset, false, false, 0);
1688 MemOps.push_back(Store);
1692 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1693 // Now store the XMM (fp + vector) parameter registers.
1694 SmallVector<SDValue, 11> SaveXMMOps;
1695 SaveXMMOps.push_back(Chain);
1697 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1698 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1699 SaveXMMOps.push_back(ALVal);
1701 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1702 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1704 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1705 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1706 X86::VR128RegisterClass);
1707 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1708 SaveXMMOps.push_back(Val);
1710 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1712 &SaveXMMOps[0], SaveXMMOps.size()));
1715 if (!MemOps.empty())
1716 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1717 &MemOps[0], MemOps.size());
1721 // Some CCs need callee pop.
1722 if (IsCalleePop(isVarArg, CallConv)) {
1723 BytesToPopOnReturn = StackSize; // Callee pops everything.
1725 BytesToPopOnReturn = 0; // Callee pops nothing.
1726 // If this is an sret function, the return should pop the hidden pointer.
1727 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1728 BytesToPopOnReturn = 4;
1732 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1733 if (CallConv == CallingConv::X86_FastCall)
1734 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1737 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1743 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1744 SDValue StackPtr, SDValue Arg,
1745 DebugLoc dl, SelectionDAG &DAG,
1746 const CCValAssign &VA,
1747 ISD::ArgFlagsTy Flags) {
1748 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1749 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1750 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1751 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1752 if (Flags.isByVal()) {
1753 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1755 return DAG.getStore(Chain, dl, Arg, PtrOff,
1756 PseudoSourceValue::getStack(), LocMemOffset,
1760 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1761 /// optimization is performed and it is required.
1763 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1764 SDValue &OutRetAddr, SDValue Chain,
1765 bool IsTailCall, bool Is64Bit,
1766 int FPDiff, DebugLoc dl) {
1767 // Adjust the Return address stack slot.
1768 EVT VT = getPointerTy();
1769 OutRetAddr = getReturnAddressFrameIndex(DAG);
1771 // Load the "old" Return address.
1772 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1773 return SDValue(OutRetAddr.getNode(), 1);
1776 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1777 /// optimization is performed and it is required (FPDiff!=0).
1779 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1780 SDValue Chain, SDValue RetAddrFrIdx,
1781 bool Is64Bit, int FPDiff, DebugLoc dl) {
1782 // Store the return address to the appropriate stack slot.
1783 if (!FPDiff) return Chain;
1784 // Calculate the new stack slot for the return address.
1785 int SlotSize = Is64Bit ? 8 : 4;
1786 int NewReturnAddrFI =
1787 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1788 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1789 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1790 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1791 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1797 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1798 CallingConv::ID CallConv, bool isVarArg,
1800 const SmallVectorImpl<ISD::OutputArg> &Outs,
1801 const SmallVectorImpl<ISD::InputArg> &Ins,
1802 DebugLoc dl, SelectionDAG &DAG,
1803 SmallVectorImpl<SDValue> &InVals) {
1804 MachineFunction &MF = DAG.getMachineFunction();
1805 bool Is64Bit = Subtarget->is64Bit();
1806 bool IsStructRet = CallIsStructReturn(Outs);
1807 bool IsSibcall = false;
1810 // Check if it's really possible to do a tail call.
1811 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1812 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1815 // Sibcalls are automatically detected tailcalls which do not require
1817 if (!GuaranteedTailCallOpt && isTailCall)
1824 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1825 "Var args not supported with calling convention fastcc or ghc");
1827 // Analyze operands of the call, assigning locations to each operand.
1828 SmallVector<CCValAssign, 16> ArgLocs;
1829 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1830 ArgLocs, *DAG.getContext());
1831 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1833 // Get a count of how many bytes are to be pushed on the stack.
1834 unsigned NumBytes = CCInfo.getNextStackOffset();
1836 // This is a sibcall. The memory operands are available in caller's
1837 // own caller's stack.
1839 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1840 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1843 if (isTailCall && !IsSibcall) {
1844 // Lower arguments at fp - stackoffset + fpdiff.
1845 unsigned NumBytesCallerPushed =
1846 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1847 FPDiff = NumBytesCallerPushed - NumBytes;
1849 // Set the delta of movement of the returnaddr stackslot.
1850 // But only set if delta is greater than previous delta.
1851 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1852 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1856 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1858 SDValue RetAddrFrIdx;
1859 // Load return adress for tail calls.
1860 if (isTailCall && FPDiff)
1861 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1862 Is64Bit, FPDiff, dl);
1864 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1865 SmallVector<SDValue, 8> MemOpChains;
1868 // Walk the register/memloc assignments, inserting copies/loads. In the case
1869 // of tail call optimization arguments are handle later.
1870 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1871 CCValAssign &VA = ArgLocs[i];
1872 EVT RegVT = VA.getLocVT();
1873 SDValue Arg = Outs[i].Val;
1874 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1875 bool isByVal = Flags.isByVal();
1877 // Promote the value if needed.
1878 switch (VA.getLocInfo()) {
1879 default: llvm_unreachable("Unknown loc info!");
1880 case CCValAssign::Full: break;
1881 case CCValAssign::SExt:
1882 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1884 case CCValAssign::ZExt:
1885 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1887 case CCValAssign::AExt:
1888 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1889 // Special case: passing MMX values in XMM registers.
1890 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1891 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1892 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1894 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1896 case CCValAssign::BCvt:
1897 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1899 case CCValAssign::Indirect: {
1900 // Store the argument.
1901 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1902 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1903 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1904 PseudoSourceValue::getFixedStack(FI), 0,
1911 if (VA.isRegLoc()) {
1912 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1913 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1914 assert(VA.isMemLoc());
1915 if (StackPtr.getNode() == 0)
1916 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1917 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1918 dl, DAG, VA, Flags));
1922 if (!MemOpChains.empty())
1923 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1924 &MemOpChains[0], MemOpChains.size());
1926 // Build a sequence of copy-to-reg nodes chained together with token chain
1927 // and flag operands which copy the outgoing args into registers.
1929 // Tail call byval lowering might overwrite argument registers so in case of
1930 // tail call optimization the copies to registers are lowered later.
1932 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1933 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1934 RegsToPass[i].second, InFlag);
1935 InFlag = Chain.getValue(1);
1938 if (Subtarget->isPICStyleGOT()) {
1939 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1942 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1943 DAG.getNode(X86ISD::GlobalBaseReg,
1944 DebugLoc(), getPointerTy()),
1946 InFlag = Chain.getValue(1);
1948 // If we are tail calling and generating PIC/GOT style code load the
1949 // address of the callee into ECX. The value in ecx is used as target of
1950 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1951 // for tail calls on PIC/GOT architectures. Normally we would just put the
1952 // address of GOT into ebx and then call target@PLT. But for tail calls
1953 // ebx would be restored (since ebx is callee saved) before jumping to the
1956 // Note: The actual moving to ECX is done further down.
1957 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1958 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1959 !G->getGlobal()->hasProtectedVisibility())
1960 Callee = LowerGlobalAddress(Callee, DAG);
1961 else if (isa<ExternalSymbolSDNode>(Callee))
1962 Callee = LowerExternalSymbol(Callee, DAG);
1966 if (Is64Bit && isVarArg) {
1967 // From AMD64 ABI document:
1968 // For calls that may call functions that use varargs or stdargs
1969 // (prototype-less calls or calls to functions containing ellipsis (...) in
1970 // the declaration) %al is used as hidden argument to specify the number
1971 // of SSE registers used. The contents of %al do not need to match exactly
1972 // the number of registers, but must be an ubound on the number of SSE
1973 // registers used and is in the range 0 - 8 inclusive.
1975 // FIXME: Verify this on Win64
1976 // Count the number of XMM registers allocated.
1977 static const unsigned XMMArgRegs[] = {
1978 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1979 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1981 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1982 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1983 && "SSE registers cannot be used when SSE is disabled");
1985 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1986 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1987 InFlag = Chain.getValue(1);
1991 // For tail calls lower the arguments to the 'real' stack slot.
1993 // Force all the incoming stack arguments to be loaded from the stack
1994 // before any new outgoing arguments are stored to the stack, because the
1995 // outgoing stack slots may alias the incoming argument stack slots, and
1996 // the alias isn't otherwise explicit. This is slightly more conservative
1997 // than necessary, because it means that each store effectively depends
1998 // on every argument instead of just those arguments it would clobber.
1999 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2001 SmallVector<SDValue, 8> MemOpChains2;
2004 // Do not flag preceeding copytoreg stuff together with the following stuff.
2006 if (GuaranteedTailCallOpt) {
2007 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2008 CCValAssign &VA = ArgLocs[i];
2011 assert(VA.isMemLoc());
2012 SDValue Arg = Outs[i].Val;
2013 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2014 // Create frame index.
2015 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2016 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2017 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
2018 FIN = DAG.getFrameIndex(FI, getPointerTy());
2020 if (Flags.isByVal()) {
2021 // Copy relative to framepointer.
2022 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2023 if (StackPtr.getNode() == 0)
2024 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2026 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2028 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2032 // Store relative to framepointer.
2033 MemOpChains2.push_back(
2034 DAG.getStore(ArgChain, dl, Arg, FIN,
2035 PseudoSourceValue::getFixedStack(FI), 0,
2041 if (!MemOpChains2.empty())
2042 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2043 &MemOpChains2[0], MemOpChains2.size());
2045 // Copy arguments to their registers.
2046 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2047 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2048 RegsToPass[i].second, InFlag);
2049 InFlag = Chain.getValue(1);
2053 // Store the return address to the appropriate stack slot.
2054 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2058 bool WasGlobalOrExternal = false;
2059 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2060 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2061 // In the 64-bit large code model, we have to make all calls
2062 // through a register, since the call instruction's 32-bit
2063 // pc-relative offset may not be large enough to hold the whole
2065 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2066 WasGlobalOrExternal = true;
2067 // If the callee is a GlobalAddress node (quite common, every direct call
2068 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2071 // We should use extra load for direct calls to dllimported functions in
2073 const GlobalValue *GV = G->getGlobal();
2074 if (!GV->hasDLLImportLinkage()) {
2075 unsigned char OpFlags = 0;
2077 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2078 // external symbols most go through the PLT in PIC mode. If the symbol
2079 // has hidden or protected visibility, or if it is static or local, then
2080 // we don't need to use the PLT - we can directly call it.
2081 if (Subtarget->isTargetELF() &&
2082 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2083 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2084 OpFlags = X86II::MO_PLT;
2085 } else if (Subtarget->isPICStyleStubAny() &&
2086 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2087 Subtarget->getDarwinVers() < 9) {
2088 // PC-relative references to external symbols should go through $stub,
2089 // unless we're building with the leopard linker or later, which
2090 // automatically synthesizes these stubs.
2091 OpFlags = X86II::MO_DARWIN_STUB;
2094 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2095 G->getOffset(), OpFlags);
2097 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2098 WasGlobalOrExternal = true;
2099 unsigned char OpFlags = 0;
2101 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2102 // symbols should go through the PLT.
2103 if (Subtarget->isTargetELF() &&
2104 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2105 OpFlags = X86II::MO_PLT;
2106 } else if (Subtarget->isPICStyleStubAny() &&
2107 Subtarget->getDarwinVers() < 9) {
2108 // PC-relative references to external symbols should go through $stub,
2109 // unless we're building with the leopard linker or later, which
2110 // automatically synthesizes these stubs.
2111 OpFlags = X86II::MO_DARWIN_STUB;
2114 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2118 // Returns a chain & a flag for retval copy to use.
2119 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2120 SmallVector<SDValue, 8> Ops;
2122 if (!IsSibcall && isTailCall) {
2123 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2124 DAG.getIntPtrConstant(0, true), InFlag);
2125 InFlag = Chain.getValue(1);
2128 Ops.push_back(Chain);
2129 Ops.push_back(Callee);
2132 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2134 // Add argument registers to the end of the list so that they are known live
2136 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2137 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2138 RegsToPass[i].second.getValueType()));
2140 // Add an implicit use GOT pointer in EBX.
2141 if (!isTailCall && Subtarget->isPICStyleGOT())
2142 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2144 // Add an implicit use of AL for x86 vararg functions.
2145 if (Is64Bit && isVarArg)
2146 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2148 if (InFlag.getNode())
2149 Ops.push_back(InFlag);
2152 // If this is the first return lowered for this function, add the regs
2153 // to the liveout set for the function.
2154 if (MF.getRegInfo().liveout_empty()) {
2155 SmallVector<CCValAssign, 16> RVLocs;
2156 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2158 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2159 for (unsigned i = 0; i != RVLocs.size(); ++i)
2160 if (RVLocs[i].isRegLoc())
2161 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2163 return DAG.getNode(X86ISD::TC_RETURN, dl,
2164 NodeTys, &Ops[0], Ops.size());
2167 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2168 InFlag = Chain.getValue(1);
2170 // Create the CALLSEQ_END node.
2171 unsigned NumBytesForCalleeToPush;
2172 if (IsCalleePop(isVarArg, CallConv))
2173 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2174 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2175 // If this is a call to a struct-return function, the callee
2176 // pops the hidden struct pointer, so we have to push it back.
2177 // This is common for Darwin/X86, Linux & Mingw32 targets.
2178 NumBytesForCalleeToPush = 4;
2180 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2182 // Returns a flag for retval copy to use.
2184 Chain = DAG.getCALLSEQ_END(Chain,
2185 DAG.getIntPtrConstant(NumBytes, true),
2186 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2189 InFlag = Chain.getValue(1);
2192 // Handle result values, copying them out of physregs into vregs that we
2194 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2195 Ins, dl, DAG, InVals);
2199 //===----------------------------------------------------------------------===//
2200 // Fast Calling Convention (tail call) implementation
2201 //===----------------------------------------------------------------------===//
2203 // Like std call, callee cleans arguments, convention except that ECX is
2204 // reserved for storing the tail called function address. Only 2 registers are
2205 // free for argument passing (inreg). Tail call optimization is performed
2207 // * tailcallopt is enabled
2208 // * caller/callee are fastcc
2209 // On X86_64 architecture with GOT-style position independent code only local
2210 // (within module) calls are supported at the moment.
2211 // To keep the stack aligned according to platform abi the function
2212 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2213 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2214 // If a tail called function callee has more arguments than the caller the
2215 // caller needs to make sure that there is room to move the RETADDR to. This is
2216 // achieved by reserving an area the size of the argument delta right after the
2217 // original REtADDR, but before the saved framepointer or the spilled registers
2218 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2230 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2231 /// for a 16 byte align requirement.
2232 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2233 SelectionDAG& DAG) {
2234 MachineFunction &MF = DAG.getMachineFunction();
2235 const TargetMachine &TM = MF.getTarget();
2236 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2237 unsigned StackAlignment = TFI.getStackAlignment();
2238 uint64_t AlignMask = StackAlignment - 1;
2239 int64_t Offset = StackSize;
2240 uint64_t SlotSize = TD->getPointerSize();
2241 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2242 // Number smaller than 12 so just add the difference.
2243 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2245 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2246 Offset = ((~AlignMask) & Offset) + StackAlignment +
2247 (StackAlignment-SlotSize);
2252 /// MatchingStackOffset - Return true if the given stack call argument is
2253 /// already available in the same position (relatively) of the caller's
2254 /// incoming argument stack.
2256 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2257 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2258 const X86InstrInfo *TII) {
2259 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2261 if (Arg.getOpcode() == ISD::CopyFromReg) {
2262 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2263 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2265 MachineInstr *Def = MRI->getVRegDef(VR);
2268 if (!Flags.isByVal()) {
2269 if (!TII->isLoadFromStackSlot(Def, FI))
2272 unsigned Opcode = Def->getOpcode();
2273 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2274 Def->getOperand(1).isFI()) {
2275 FI = Def->getOperand(1).getIndex();
2276 Bytes = Flags.getByValSize();
2280 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2281 if (Flags.isByVal())
2282 // ByVal argument is passed in as a pointer but it's now being
2283 // dereferenced. e.g.
2284 // define @foo(%struct.X* %A) {
2285 // tail call @bar(%struct.X* byval %A)
2288 SDValue Ptr = Ld->getBasePtr();
2289 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2292 FI = FINode->getIndex();
2296 assert(FI != INT_MAX);
2297 if (!MFI->isFixedObjectIndex(FI))
2299 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2302 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2303 /// for tail call optimization. Targets which want to do tail call
2304 /// optimization should implement this function.
2306 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2307 CallingConv::ID CalleeCC,
2309 bool isCalleeStructRet,
2310 bool isCallerStructRet,
2311 const SmallVectorImpl<ISD::OutputArg> &Outs,
2312 const SmallVectorImpl<ISD::InputArg> &Ins,
2313 SelectionDAG& DAG) const {
2314 if (!IsTailCallConvention(CalleeCC) &&
2315 CalleeCC != CallingConv::C)
2318 // If -tailcallopt is specified, make fastcc functions tail-callable.
2319 const MachineFunction &MF = DAG.getMachineFunction();
2320 const Function *CallerF = DAG.getMachineFunction().getFunction();
2321 if (GuaranteedTailCallOpt) {
2322 if (IsTailCallConvention(CalleeCC) &&
2323 CallerF->getCallingConv() == CalleeCC)
2328 // Look for obvious safe cases to perform tail call optimization that does not
2329 // requite ABI changes. This is what gcc calls sibcall.
2331 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2332 // emit a special epilogue.
2333 if (RegInfo->needsStackRealignment(MF))
2336 // Do not sibcall optimize vararg calls unless the call site is not passing any
2338 if (isVarArg && !Outs.empty())
2341 // Also avoid sibcall optimization if either caller or callee uses struct
2342 // return semantics.
2343 if (isCalleeStructRet || isCallerStructRet)
2346 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2347 // Therefore if it's not used by the call it is not safe to optimize this into
2349 bool Unused = false;
2350 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2357 SmallVector<CCValAssign, 16> RVLocs;
2358 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2359 RVLocs, *DAG.getContext());
2360 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2361 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2362 CCValAssign &VA = RVLocs[i];
2363 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2368 // If the callee takes no arguments then go on to check the results of the
2370 if (!Outs.empty()) {
2371 // Check if stack adjustment is needed. For now, do not do this if any
2372 // argument is passed on the stack.
2373 SmallVector<CCValAssign, 16> ArgLocs;
2374 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2375 ArgLocs, *DAG.getContext());
2376 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2377 if (CCInfo.getNextStackOffset()) {
2378 MachineFunction &MF = DAG.getMachineFunction();
2379 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2381 if (Subtarget->isTargetWin64())
2382 // Win64 ABI has additional complications.
2385 // Check if the arguments are already laid out in the right way as
2386 // the caller's fixed stack objects.
2387 MachineFrameInfo *MFI = MF.getFrameInfo();
2388 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2389 const X86InstrInfo *TII =
2390 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2391 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2392 CCValAssign &VA = ArgLocs[i];
2393 EVT RegVT = VA.getLocVT();
2394 SDValue Arg = Outs[i].Val;
2395 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2396 if (VA.getLocInfo() == CCValAssign::Indirect)
2398 if (!VA.isRegLoc()) {
2399 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2411 X86TargetLowering::createFastISel(MachineFunction &mf,
2412 DenseMap<const Value *, unsigned> &vm,
2413 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2414 DenseMap<const AllocaInst *, int> &am
2416 , SmallSet<const Instruction *, 8> &cil
2419 return X86::createFastISel(mf, vm, bm, am
2427 //===----------------------------------------------------------------------===//
2428 // Other Lowering Hooks
2429 //===----------------------------------------------------------------------===//
2432 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2433 MachineFunction &MF = DAG.getMachineFunction();
2434 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2435 int ReturnAddrIndex = FuncInfo->getRAIndex();
2437 if (ReturnAddrIndex == 0) {
2438 // Set up a frame object for the return address.
2439 uint64_t SlotSize = TD->getPointerSize();
2440 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2442 FuncInfo->setRAIndex(ReturnAddrIndex);
2445 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2449 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2450 bool hasSymbolicDisplacement) {
2451 // Offset should fit into 32 bit immediate field.
2452 if (!isInt<32>(Offset))
2455 // If we don't have a symbolic displacement - we don't have any extra
2457 if (!hasSymbolicDisplacement)
2460 // FIXME: Some tweaks might be needed for medium code model.
2461 if (M != CodeModel::Small && M != CodeModel::Kernel)
2464 // For small code model we assume that latest object is 16MB before end of 31
2465 // bits boundary. We may also accept pretty large negative constants knowing
2466 // that all objects are in the positive half of address space.
2467 if (M == CodeModel::Small && Offset < 16*1024*1024)
2470 // For kernel code model we know that all object resist in the negative half
2471 // of 32bits address space. We may not accept negative offsets, since they may
2472 // be just off and we may accept pretty large positive ones.
2473 if (M == CodeModel::Kernel && Offset > 0)
2479 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2480 /// specific condition code, returning the condition code and the LHS/RHS of the
2481 /// comparison to make.
2482 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2483 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2485 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2486 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2487 // X > -1 -> X == 0, jump !sign.
2488 RHS = DAG.getConstant(0, RHS.getValueType());
2489 return X86::COND_NS;
2490 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2491 // X < 0 -> X == 0, jump on sign.
2493 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2495 RHS = DAG.getConstant(0, RHS.getValueType());
2496 return X86::COND_LE;
2500 switch (SetCCOpcode) {
2501 default: llvm_unreachable("Invalid integer condition!");
2502 case ISD::SETEQ: return X86::COND_E;
2503 case ISD::SETGT: return X86::COND_G;
2504 case ISD::SETGE: return X86::COND_GE;
2505 case ISD::SETLT: return X86::COND_L;
2506 case ISD::SETLE: return X86::COND_LE;
2507 case ISD::SETNE: return X86::COND_NE;
2508 case ISD::SETULT: return X86::COND_B;
2509 case ISD::SETUGT: return X86::COND_A;
2510 case ISD::SETULE: return X86::COND_BE;
2511 case ISD::SETUGE: return X86::COND_AE;
2515 // First determine if it is required or is profitable to flip the operands.
2517 // If LHS is a foldable load, but RHS is not, flip the condition.
2518 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2519 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2520 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2521 std::swap(LHS, RHS);
2524 switch (SetCCOpcode) {
2530 std::swap(LHS, RHS);
2534 // On a floating point condition, the flags are set as follows:
2536 // 0 | 0 | 0 | X > Y
2537 // 0 | 0 | 1 | X < Y
2538 // 1 | 0 | 0 | X == Y
2539 // 1 | 1 | 1 | unordered
2540 switch (SetCCOpcode) {
2541 default: llvm_unreachable("Condcode should be pre-legalized away");
2543 case ISD::SETEQ: return X86::COND_E;
2544 case ISD::SETOLT: // flipped
2546 case ISD::SETGT: return X86::COND_A;
2547 case ISD::SETOLE: // flipped
2549 case ISD::SETGE: return X86::COND_AE;
2550 case ISD::SETUGT: // flipped
2552 case ISD::SETLT: return X86::COND_B;
2553 case ISD::SETUGE: // flipped
2555 case ISD::SETLE: return X86::COND_BE;
2557 case ISD::SETNE: return X86::COND_NE;
2558 case ISD::SETUO: return X86::COND_P;
2559 case ISD::SETO: return X86::COND_NP;
2561 case ISD::SETUNE: return X86::COND_INVALID;
2565 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2566 /// code. Current x86 isa includes the following FP cmov instructions:
2567 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2568 static bool hasFPCMov(unsigned X86CC) {
2584 /// isFPImmLegal - Returns true if the target can instruction select the
2585 /// specified FP immediate natively. If false, the legalizer will
2586 /// materialize the FP immediate as a load from a constant pool.
2587 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2588 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2589 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2595 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2596 /// the specified range (L, H].
2597 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2598 return (Val < 0) || (Val >= Low && Val < Hi);
2601 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2602 /// specified value.
2603 static bool isUndefOrEqual(int Val, int CmpVal) {
2604 if (Val < 0 || Val == CmpVal)
2609 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2610 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2611 /// the second operand.
2612 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2613 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2614 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2615 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2616 return (Mask[0] < 2 && Mask[1] < 2);
2620 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2621 SmallVector<int, 8> M;
2623 return ::isPSHUFDMask(M, N->getValueType(0));
2626 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2627 /// is suitable for input to PSHUFHW.
2628 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2629 if (VT != MVT::v8i16)
2632 // Lower quadword copied in order or undef.
2633 for (int i = 0; i != 4; ++i)
2634 if (Mask[i] >= 0 && Mask[i] != i)
2637 // Upper quadword shuffled.
2638 for (int i = 4; i != 8; ++i)
2639 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2645 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2646 SmallVector<int, 8> M;
2648 return ::isPSHUFHWMask(M, N->getValueType(0));
2651 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2652 /// is suitable for input to PSHUFLW.
2653 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2654 if (VT != MVT::v8i16)
2657 // Upper quadword copied in order.
2658 for (int i = 4; i != 8; ++i)
2659 if (Mask[i] >= 0 && Mask[i] != i)
2662 // Lower quadword shuffled.
2663 for (int i = 0; i != 4; ++i)
2670 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2671 SmallVector<int, 8> M;
2673 return ::isPSHUFLWMask(M, N->getValueType(0));
2676 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2677 /// is suitable for input to PALIGNR.
2678 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2680 int i, e = VT.getVectorNumElements();
2682 // Do not handle v2i64 / v2f64 shuffles with palignr.
2683 if (e < 4 || !hasSSSE3)
2686 for (i = 0; i != e; ++i)
2690 // All undef, not a palignr.
2694 // Determine if it's ok to perform a palignr with only the LHS, since we
2695 // don't have access to the actual shuffle elements to see if RHS is undef.
2696 bool Unary = Mask[i] < (int)e;
2697 bool NeedsUnary = false;
2699 int s = Mask[i] - i;
2701 // Check the rest of the elements to see if they are consecutive.
2702 for (++i; i != e; ++i) {
2707 Unary = Unary && (m < (int)e);
2708 NeedsUnary = NeedsUnary || (m < s);
2710 if (NeedsUnary && !Unary)
2712 if (Unary && m != ((s+i) & (e-1)))
2714 if (!Unary && m != (s+i))
2720 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2721 SmallVector<int, 8> M;
2723 return ::isPALIGNRMask(M, N->getValueType(0), true);
2726 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2727 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2728 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2729 int NumElems = VT.getVectorNumElements();
2730 if (NumElems != 2 && NumElems != 4)
2733 int Half = NumElems / 2;
2734 for (int i = 0; i < Half; ++i)
2735 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2737 for (int i = Half; i < NumElems; ++i)
2738 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2744 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2745 SmallVector<int, 8> M;
2747 return ::isSHUFPMask(M, N->getValueType(0));
2750 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2751 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2752 /// half elements to come from vector 1 (which would equal the dest.) and
2753 /// the upper half to come from vector 2.
2754 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2755 int NumElems = VT.getVectorNumElements();
2757 if (NumElems != 2 && NumElems != 4)
2760 int Half = NumElems / 2;
2761 for (int i = 0; i < Half; ++i)
2762 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2764 for (int i = Half; i < NumElems; ++i)
2765 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2770 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2771 SmallVector<int, 8> M;
2773 return isCommutedSHUFPMask(M, N->getValueType(0));
2776 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2777 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2778 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2779 if (N->getValueType(0).getVectorNumElements() != 4)
2782 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2783 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2784 isUndefOrEqual(N->getMaskElt(1), 7) &&
2785 isUndefOrEqual(N->getMaskElt(2), 2) &&
2786 isUndefOrEqual(N->getMaskElt(3), 3);
2789 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2790 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2792 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2793 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2798 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2799 isUndefOrEqual(N->getMaskElt(1), 3) &&
2800 isUndefOrEqual(N->getMaskElt(2), 2) &&
2801 isUndefOrEqual(N->getMaskElt(3), 3);
2804 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2805 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2806 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2807 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2809 if (NumElems != 2 && NumElems != 4)
2812 for (unsigned i = 0; i < NumElems/2; ++i)
2813 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2816 for (unsigned i = NumElems/2; i < NumElems; ++i)
2817 if (!isUndefOrEqual(N->getMaskElt(i), i))
2823 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2824 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2825 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2826 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2828 if (NumElems != 2 && NumElems != 4)
2831 for (unsigned i = 0; i < NumElems/2; ++i)
2832 if (!isUndefOrEqual(N->getMaskElt(i), i))
2835 for (unsigned i = 0; i < NumElems/2; ++i)
2836 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2842 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2843 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2844 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2845 bool V2IsSplat = false) {
2846 int NumElts = VT.getVectorNumElements();
2847 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2850 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2852 int BitI1 = Mask[i+1];
2853 if (!isUndefOrEqual(BitI, j))
2856 if (!isUndefOrEqual(BitI1, NumElts))
2859 if (!isUndefOrEqual(BitI1, j + NumElts))
2866 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2867 SmallVector<int, 8> M;
2869 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2872 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2873 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2874 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2875 bool V2IsSplat = false) {
2876 int NumElts = VT.getVectorNumElements();
2877 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2880 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2882 int BitI1 = Mask[i+1];
2883 if (!isUndefOrEqual(BitI, j + NumElts/2))
2886 if (isUndefOrEqual(BitI1, NumElts))
2889 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2896 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2897 SmallVector<int, 8> M;
2899 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2902 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2903 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2905 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2906 int NumElems = VT.getVectorNumElements();
2907 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2910 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2912 int BitI1 = Mask[i+1];
2913 if (!isUndefOrEqual(BitI, j))
2915 if (!isUndefOrEqual(BitI1, j))
2921 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2922 SmallVector<int, 8> M;
2924 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2927 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2928 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2930 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2931 int NumElems = VT.getVectorNumElements();
2932 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2935 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2937 int BitI1 = Mask[i+1];
2938 if (!isUndefOrEqual(BitI, j))
2940 if (!isUndefOrEqual(BitI1, j))
2946 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2947 SmallVector<int, 8> M;
2949 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2952 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2953 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2954 /// MOVSD, and MOVD, i.e. setting the lowest element.
2955 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2956 if (VT.getVectorElementType().getSizeInBits() < 32)
2959 int NumElts = VT.getVectorNumElements();
2961 if (!isUndefOrEqual(Mask[0], NumElts))
2964 for (int i = 1; i < NumElts; ++i)
2965 if (!isUndefOrEqual(Mask[i], i))
2971 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2972 SmallVector<int, 8> M;
2974 return ::isMOVLMask(M, N->getValueType(0));
2977 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2978 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2979 /// element of vector 2 and the other elements to come from vector 1 in order.
2980 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2981 bool V2IsSplat = false, bool V2IsUndef = false) {
2982 int NumOps = VT.getVectorNumElements();
2983 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2986 if (!isUndefOrEqual(Mask[0], 0))
2989 for (int i = 1; i < NumOps; ++i)
2990 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2991 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2992 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2998 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2999 bool V2IsUndef = false) {
3000 SmallVector<int, 8> M;
3002 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3005 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3006 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3007 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3008 if (N->getValueType(0).getVectorNumElements() != 4)
3011 // Expect 1, 1, 3, 3
3012 for (unsigned i = 0; i < 2; ++i) {
3013 int Elt = N->getMaskElt(i);
3014 if (Elt >= 0 && Elt != 1)
3019 for (unsigned i = 2; i < 4; ++i) {
3020 int Elt = N->getMaskElt(i);
3021 if (Elt >= 0 && Elt != 3)
3026 // Don't use movshdup if it can be done with a shufps.
3027 // FIXME: verify that matching u, u, 3, 3 is what we want.
3031 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3032 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3033 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3034 if (N->getValueType(0).getVectorNumElements() != 4)
3037 // Expect 0, 0, 2, 2
3038 for (unsigned i = 0; i < 2; ++i)
3039 if (N->getMaskElt(i) > 0)
3043 for (unsigned i = 2; i < 4; ++i) {
3044 int Elt = N->getMaskElt(i);
3045 if (Elt >= 0 && Elt != 2)
3050 // Don't use movsldup if it can be done with a shufps.
3054 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3055 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3056 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3057 int e = N->getValueType(0).getVectorNumElements() / 2;
3059 for (int i = 0; i < e; ++i)
3060 if (!isUndefOrEqual(N->getMaskElt(i), i))
3062 for (int i = 0; i < e; ++i)
3063 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3068 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3069 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3070 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3071 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3072 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3074 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3076 for (int i = 0; i < NumOperands; ++i) {
3077 int Val = SVOp->getMaskElt(NumOperands-i-1);
3078 if (Val < 0) Val = 0;
3079 if (Val >= NumOperands) Val -= NumOperands;
3081 if (i != NumOperands - 1)
3087 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3088 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3089 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3090 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3092 // 8 nodes, but we only care about the last 4.
3093 for (unsigned i = 7; i >= 4; --i) {
3094 int Val = SVOp->getMaskElt(i);
3103 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3104 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3105 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3106 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3108 // 8 nodes, but we only care about the first 4.
3109 for (int i = 3; i >= 0; --i) {
3110 int Val = SVOp->getMaskElt(i);
3119 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3120 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3121 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3122 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3123 EVT VVT = N->getValueType(0);
3124 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3128 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3129 Val = SVOp->getMaskElt(i);
3133 return (Val - i) * EltSize;
3136 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3138 bool X86::isZeroNode(SDValue Elt) {
3139 return ((isa<ConstantSDNode>(Elt) &&
3140 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3141 (isa<ConstantFPSDNode>(Elt) &&
3142 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3145 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3146 /// their permute mask.
3147 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3148 SelectionDAG &DAG) {
3149 EVT VT = SVOp->getValueType(0);
3150 unsigned NumElems = VT.getVectorNumElements();
3151 SmallVector<int, 8> MaskVec;
3153 for (unsigned i = 0; i != NumElems; ++i) {
3154 int idx = SVOp->getMaskElt(i);
3156 MaskVec.push_back(idx);
3157 else if (idx < (int)NumElems)
3158 MaskVec.push_back(idx + NumElems);
3160 MaskVec.push_back(idx - NumElems);
3162 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3163 SVOp->getOperand(0), &MaskVec[0]);
3166 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3167 /// the two vector operands have swapped position.
3168 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3169 unsigned NumElems = VT.getVectorNumElements();
3170 for (unsigned i = 0; i != NumElems; ++i) {
3174 else if (idx < (int)NumElems)
3175 Mask[i] = idx + NumElems;
3177 Mask[i] = idx - NumElems;
3181 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3182 /// match movhlps. The lower half elements should come from upper half of
3183 /// V1 (and in order), and the upper half elements should come from the upper
3184 /// half of V2 (and in order).
3185 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3186 if (Op->getValueType(0).getVectorNumElements() != 4)
3188 for (unsigned i = 0, e = 2; i != e; ++i)
3189 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3191 for (unsigned i = 2; i != 4; ++i)
3192 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3197 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3198 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3200 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3201 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3203 N = N->getOperand(0).getNode();
3204 if (!ISD::isNON_EXTLoad(N))
3207 *LD = cast<LoadSDNode>(N);
3211 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3212 /// match movlp{s|d}. The lower half elements should come from lower half of
3213 /// V1 (and in order), and the upper half elements should come from the upper
3214 /// half of V2 (and in order). And since V1 will become the source of the
3215 /// MOVLP, it must be either a vector load or a scalar load to vector.
3216 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3217 ShuffleVectorSDNode *Op) {
3218 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3220 // Is V2 is a vector load, don't do this transformation. We will try to use
3221 // load folding shufps op.
3222 if (ISD::isNON_EXTLoad(V2))
3225 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3227 if (NumElems != 2 && NumElems != 4)
3229 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3230 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3232 for (unsigned i = NumElems/2; i != NumElems; ++i)
3233 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3238 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3240 static bool isSplatVector(SDNode *N) {
3241 if (N->getOpcode() != ISD::BUILD_VECTOR)
3244 SDValue SplatValue = N->getOperand(0);
3245 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3246 if (N->getOperand(i) != SplatValue)
3251 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3252 /// to an zero vector.
3253 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3254 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3255 SDValue V1 = N->getOperand(0);
3256 SDValue V2 = N->getOperand(1);
3257 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3258 for (unsigned i = 0; i != NumElems; ++i) {
3259 int Idx = N->getMaskElt(i);
3260 if (Idx >= (int)NumElems) {
3261 unsigned Opc = V2.getOpcode();
3262 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3264 if (Opc != ISD::BUILD_VECTOR ||
3265 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3267 } else if (Idx >= 0) {
3268 unsigned Opc = V1.getOpcode();
3269 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3271 if (Opc != ISD::BUILD_VECTOR ||
3272 !X86::isZeroNode(V1.getOperand(Idx)))
3279 /// getZeroVector - Returns a vector of specified type with all zero elements.
3281 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3283 assert(VT.isVector() && "Expected a vector type");
3285 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3286 // type. This ensures they get CSE'd.
3288 if (VT.getSizeInBits() == 64) { // MMX
3289 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3290 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3291 } else if (HasSSE2) { // SSE2
3292 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3293 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3295 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3296 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3298 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3301 /// getOnesVector - Returns a vector of specified type with all bits set.
3303 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3304 assert(VT.isVector() && "Expected a vector type");
3306 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3307 // type. This ensures they get CSE'd.
3308 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3310 if (VT.getSizeInBits() == 64) // MMX
3311 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3313 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3314 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3318 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3319 /// that point to V2 points to its first element.
3320 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3321 EVT VT = SVOp->getValueType(0);
3322 unsigned NumElems = VT.getVectorNumElements();
3324 bool Changed = false;
3325 SmallVector<int, 8> MaskVec;
3326 SVOp->getMask(MaskVec);
3328 for (unsigned i = 0; i != NumElems; ++i) {
3329 if (MaskVec[i] > (int)NumElems) {
3330 MaskVec[i] = NumElems;
3335 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3336 SVOp->getOperand(1), &MaskVec[0]);
3337 return SDValue(SVOp, 0);
3340 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3341 /// operation of specified width.
3342 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3344 unsigned NumElems = VT.getVectorNumElements();
3345 SmallVector<int, 8> Mask;
3346 Mask.push_back(NumElems);
3347 for (unsigned i = 1; i != NumElems; ++i)
3349 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3352 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3353 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3355 unsigned NumElems = VT.getVectorNumElements();
3356 SmallVector<int, 8> Mask;
3357 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3359 Mask.push_back(i + NumElems);
3361 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3364 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3365 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3367 unsigned NumElems = VT.getVectorNumElements();
3368 unsigned Half = NumElems/2;
3369 SmallVector<int, 8> Mask;
3370 for (unsigned i = 0; i != Half; ++i) {
3371 Mask.push_back(i + Half);
3372 Mask.push_back(i + NumElems + Half);
3374 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3377 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3378 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3380 if (SV->getValueType(0).getVectorNumElements() <= 4)
3381 return SDValue(SV, 0);
3383 EVT PVT = MVT::v4f32;
3384 EVT VT = SV->getValueType(0);
3385 DebugLoc dl = SV->getDebugLoc();
3386 SDValue V1 = SV->getOperand(0);
3387 int NumElems = VT.getVectorNumElements();
3388 int EltNo = SV->getSplatIndex();
3390 // unpack elements to the correct location
3391 while (NumElems > 4) {
3392 if (EltNo < NumElems/2) {
3393 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3395 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3396 EltNo -= NumElems/2;
3401 // Perform the splat.
3402 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3403 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3404 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3405 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3408 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3409 /// vector of zero or undef vector. This produces a shuffle where the low
3410 /// element of V2 is swizzled into the zero/undef vector, landing at element
3411 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3412 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3413 bool isZero, bool HasSSE2,
3414 SelectionDAG &DAG) {
3415 EVT VT = V2.getValueType();
3417 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3418 unsigned NumElems = VT.getVectorNumElements();
3419 SmallVector<int, 16> MaskVec;
3420 for (unsigned i = 0; i != NumElems; ++i)
3421 // If this is the insertion idx, put the low elt of V2 here.
3422 MaskVec.push_back(i == Idx ? NumElems : i);
3423 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3426 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3427 /// a shuffle that is zero.
3429 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3430 bool Low, SelectionDAG &DAG) {
3431 unsigned NumZeros = 0;
3432 for (int i = 0; i < NumElems; ++i) {
3433 unsigned Index = Low ? i : NumElems-i-1;
3434 int Idx = SVOp->getMaskElt(Index);
3439 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3440 if (Elt.getNode() && X86::isZeroNode(Elt))
3448 /// isVectorShift - Returns true if the shuffle can be implemented as a
3449 /// logical left or right shift of a vector.
3450 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3451 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3452 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3453 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3456 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3459 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3463 bool SeenV1 = false;
3464 bool SeenV2 = false;
3465 for (unsigned i = NumZeros; i < NumElems; ++i) {
3466 unsigned Val = isLeft ? (i - NumZeros) : i;
3467 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3470 unsigned Idx = (unsigned) Idx_;
3480 if (SeenV1 && SeenV2)
3483 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3489 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3491 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3492 unsigned NumNonZero, unsigned NumZero,
3493 SelectionDAG &DAG, TargetLowering &TLI) {
3497 DebugLoc dl = Op.getDebugLoc();
3500 for (unsigned i = 0; i < 16; ++i) {
3501 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3502 if (ThisIsNonZero && First) {
3504 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3506 V = DAG.getUNDEF(MVT::v8i16);
3511 SDValue ThisElt(0, 0), LastElt(0, 0);
3512 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3513 if (LastIsNonZero) {
3514 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3515 MVT::i16, Op.getOperand(i-1));
3517 if (ThisIsNonZero) {
3518 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3519 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3520 ThisElt, DAG.getConstant(8, MVT::i8));
3522 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3526 if (ThisElt.getNode())
3527 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3528 DAG.getIntPtrConstant(i/2));
3532 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3535 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3537 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3538 unsigned NumNonZero, unsigned NumZero,
3539 SelectionDAG &DAG, TargetLowering &TLI) {
3543 DebugLoc dl = Op.getDebugLoc();
3546 for (unsigned i = 0; i < 8; ++i) {
3547 bool isNonZero = (NonZeros & (1 << i)) != 0;
3551 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3553 V = DAG.getUNDEF(MVT::v8i16);
3556 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3557 MVT::v8i16, V, Op.getOperand(i),
3558 DAG.getIntPtrConstant(i));
3565 /// getVShift - Return a vector logical shift node.
3567 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3568 unsigned NumBits, SelectionDAG &DAG,
3569 const TargetLowering &TLI, DebugLoc dl) {
3570 bool isMMX = VT.getSizeInBits() == 64;
3571 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3572 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3573 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3574 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3575 DAG.getNode(Opc, dl, ShVT, SrcOp,
3576 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3580 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3581 SelectionDAG &DAG) {
3583 // Check if the scalar load can be widened into a vector load. And if
3584 // the address is "base + cst" see if the cst can be "absorbed" into
3585 // the shuffle mask.
3586 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3587 SDValue Ptr = LD->getBasePtr();
3588 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3590 EVT PVT = LD->getValueType(0);
3591 if (PVT != MVT::i32 && PVT != MVT::f32)
3596 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3597 FI = FINode->getIndex();
3599 } else if (Ptr.getOpcode() == ISD::ADD &&
3600 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3601 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3602 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3603 Offset = Ptr.getConstantOperandVal(1);
3604 Ptr = Ptr.getOperand(0);
3609 SDValue Chain = LD->getChain();
3610 // Make sure the stack object alignment is at least 16.
3611 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3612 if (DAG.InferPtrAlignment(Ptr) < 16) {
3613 if (MFI->isFixedObjectIndex(FI)) {
3614 // Can't change the alignment. FIXME: It's possible to compute
3615 // the exact stack offset and reference FI + adjust offset instead.
3616 // If someone *really* cares about this. That's the way to implement it.
3619 MFI->setObjectAlignment(FI, 16);
3623 // (Offset % 16) must be multiple of 4. Then address is then
3624 // Ptr + (Offset & ~15).
3627 if ((Offset % 16) & 3)
3629 int64_t StartOffset = Offset & ~15;
3631 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3632 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3634 int EltNo = (Offset - StartOffset) >> 2;
3635 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3636 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3637 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3639 // Canonicalize it to a v4i32 shuffle.
3640 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3641 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3642 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3643 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3649 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3650 /// vector of type 'VT', see if the elements can be replaced by a single large
3651 /// load which has the same value as a build_vector whose operands are 'elts'.
3653 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3655 /// FIXME: we'd also like to handle the case where the last elements are zero
3656 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3657 /// There's even a handy isZeroNode for that purpose.
3658 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3659 DebugLoc &dl, SelectionDAG &DAG) {
3660 EVT EltVT = VT.getVectorElementType();
3661 unsigned NumElems = Elts.size();
3663 LoadSDNode *LDBase = NULL;
3664 unsigned LastLoadedElt = -1U;
3666 // For each element in the initializer, see if we've found a load or an undef.
3667 // If we don't find an initial load element, or later load elements are
3668 // non-consecutive, bail out.
3669 for (unsigned i = 0; i < NumElems; ++i) {
3670 SDValue Elt = Elts[i];
3672 if (!Elt.getNode() ||
3673 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3676 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3678 LDBase = cast<LoadSDNode>(Elt.getNode());
3682 if (Elt.getOpcode() == ISD::UNDEF)
3685 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3686 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3691 // If we have found an entire vector of loads and undefs, then return a large
3692 // load of the entire vector width starting at the base pointer. If we found
3693 // consecutive loads for the low half, generate a vzext_load node.
3694 if (LastLoadedElt == NumElems - 1) {
3695 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3696 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3697 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3698 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3699 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3700 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3701 LDBase->isVolatile(), LDBase->isNonTemporal(),
3702 LDBase->getAlignment());
3703 } else if (NumElems == 4 && LastLoadedElt == 1) {
3704 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3705 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3706 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3707 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3713 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3714 DebugLoc dl = Op.getDebugLoc();
3715 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3716 if (ISD::isBuildVectorAllZeros(Op.getNode())
3717 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3718 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3719 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3720 // eliminated on x86-32 hosts.
3721 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3724 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3725 return getOnesVector(Op.getValueType(), DAG, dl);
3726 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3729 EVT VT = Op.getValueType();
3730 EVT ExtVT = VT.getVectorElementType();
3731 unsigned EVTBits = ExtVT.getSizeInBits();
3733 unsigned NumElems = Op.getNumOperands();
3734 unsigned NumZero = 0;
3735 unsigned NumNonZero = 0;
3736 unsigned NonZeros = 0;
3737 bool IsAllConstants = true;
3738 SmallSet<SDValue, 8> Values;
3739 for (unsigned i = 0; i < NumElems; ++i) {
3740 SDValue Elt = Op.getOperand(i);
3741 if (Elt.getOpcode() == ISD::UNDEF)
3744 if (Elt.getOpcode() != ISD::Constant &&
3745 Elt.getOpcode() != ISD::ConstantFP)
3746 IsAllConstants = false;
3747 if (X86::isZeroNode(Elt))
3750 NonZeros |= (1 << i);
3755 if (NumNonZero == 0) {
3756 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3757 return DAG.getUNDEF(VT);
3760 // Special case for single non-zero, non-undef, element.
3761 if (NumNonZero == 1) {
3762 unsigned Idx = CountTrailingZeros_32(NonZeros);
3763 SDValue Item = Op.getOperand(Idx);
3765 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3766 // the value are obviously zero, truncate the value to i32 and do the
3767 // insertion that way. Only do this if the value is non-constant or if the
3768 // value is a constant being inserted into element 0. It is cheaper to do
3769 // a constant pool load than it is to do a movd + shuffle.
3770 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3771 (!IsAllConstants || Idx == 0)) {
3772 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3773 // Handle MMX and SSE both.
3774 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3775 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3777 // Truncate the value (which may itself be a constant) to i32, and
3778 // convert it to a vector with movd (S2V+shuffle to zero extend).
3779 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3780 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3781 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3782 Subtarget->hasSSE2(), DAG);
3784 // Now we have our 32-bit value zero extended in the low element of
3785 // a vector. If Idx != 0, swizzle it into place.
3787 SmallVector<int, 4> Mask;
3788 Mask.push_back(Idx);
3789 for (unsigned i = 1; i != VecElts; ++i)
3791 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3792 DAG.getUNDEF(Item.getValueType()),
3795 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3799 // If we have a constant or non-constant insertion into the low element of
3800 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3801 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3802 // depending on what the source datatype is.
3805 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3806 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3807 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3808 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3809 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3810 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3812 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3813 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3814 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3815 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3816 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3817 Subtarget->hasSSE2(), DAG);
3818 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3822 // Is it a vector logical left shift?
3823 if (NumElems == 2 && Idx == 1 &&
3824 X86::isZeroNode(Op.getOperand(0)) &&
3825 !X86::isZeroNode(Op.getOperand(1))) {
3826 unsigned NumBits = VT.getSizeInBits();
3827 return getVShift(true, VT,
3828 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3829 VT, Op.getOperand(1)),
3830 NumBits/2, DAG, *this, dl);
3833 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3836 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3837 // is a non-constant being inserted into an element other than the low one,
3838 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3839 // movd/movss) to move this into the low element, then shuffle it into
3841 if (EVTBits == 32) {
3842 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3844 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3845 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3846 Subtarget->hasSSE2(), DAG);
3847 SmallVector<int, 8> MaskVec;
3848 for (unsigned i = 0; i < NumElems; i++)
3849 MaskVec.push_back(i == Idx ? 0 : 1);
3850 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3854 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3855 if (Values.size() == 1) {
3856 if (EVTBits == 32) {
3857 // Instead of a shuffle like this:
3858 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3859 // Check if it's possible to issue this instead.
3860 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3861 unsigned Idx = CountTrailingZeros_32(NonZeros);
3862 SDValue Item = Op.getOperand(Idx);
3863 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3864 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3869 // A vector full of immediates; various special cases are already
3870 // handled, so this is best done with a single constant-pool load.
3874 // Let legalizer expand 2-wide build_vectors.
3875 if (EVTBits == 64) {
3876 if (NumNonZero == 1) {
3877 // One half is zero or undef.
3878 unsigned Idx = CountTrailingZeros_32(NonZeros);
3879 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3880 Op.getOperand(Idx));
3881 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3882 Subtarget->hasSSE2(), DAG);
3887 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3888 if (EVTBits == 8 && NumElems == 16) {
3889 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3891 if (V.getNode()) return V;
3894 if (EVTBits == 16 && NumElems == 8) {
3895 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3897 if (V.getNode()) return V;
3900 // If element VT is == 32 bits, turn it into a number of shuffles.
3901 SmallVector<SDValue, 8> V;
3903 if (NumElems == 4 && NumZero > 0) {
3904 for (unsigned i = 0; i < 4; ++i) {
3905 bool isZero = !(NonZeros & (1 << i));
3907 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3909 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3912 for (unsigned i = 0; i < 2; ++i) {
3913 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3916 V[i] = V[i*2]; // Must be a zero vector.
3919 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3922 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3925 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3930 SmallVector<int, 8> MaskVec;
3931 bool Reverse = (NonZeros & 0x3) == 2;
3932 for (unsigned i = 0; i < 2; ++i)
3933 MaskVec.push_back(Reverse ? 1-i : i);
3934 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3935 for (unsigned i = 0; i < 2; ++i)
3936 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3937 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3940 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
3941 // Check for a build vector of consecutive loads.
3942 for (unsigned i = 0; i < NumElems; ++i)
3943 V[i] = Op.getOperand(i);
3945 // Check for elements which are consecutive loads.
3946 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
3950 // For SSE 4.1, use inserts into undef.
3951 if (getSubtarget()->hasSSE41()) {
3952 V[0] = DAG.getUNDEF(VT);
3953 for (unsigned i = 0; i < NumElems; ++i)
3954 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3955 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3956 Op.getOperand(i), DAG.getIntPtrConstant(i));
3960 // Otherwise, expand into a number of unpckl*
3962 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3963 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3964 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3965 for (unsigned i = 0; i < NumElems; ++i)
3966 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3968 while (NumElems != 0) {
3969 for (unsigned i = 0; i < NumElems; ++i)
3970 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3979 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3980 // We support concatenate two MMX registers and place them in a MMX
3981 // register. This is better than doing a stack convert.
3982 DebugLoc dl = Op.getDebugLoc();
3983 EVT ResVT = Op.getValueType();
3984 assert(Op.getNumOperands() == 2);
3985 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3986 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3988 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3989 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3990 InVec = Op.getOperand(1);
3991 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3992 unsigned NumElts = ResVT.getVectorNumElements();
3993 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3994 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3995 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3997 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3998 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3999 Mask[0] = 0; Mask[1] = 2;
4000 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4002 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4005 // v8i16 shuffles - Prefer shuffles in the following order:
4006 // 1. [all] pshuflw, pshufhw, optional move
4007 // 2. [ssse3] 1 x pshufb
4008 // 3. [ssse3] 2 x pshufb + 1 x por
4009 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4011 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4012 SelectionDAG &DAG, X86TargetLowering &TLI) {
4013 SDValue V1 = SVOp->getOperand(0);
4014 SDValue V2 = SVOp->getOperand(1);
4015 DebugLoc dl = SVOp->getDebugLoc();
4016 SmallVector<int, 8> MaskVals;
4018 // Determine if more than 1 of the words in each of the low and high quadwords
4019 // of the result come from the same quadword of one of the two inputs. Undef
4020 // mask values count as coming from any quadword, for better codegen.
4021 SmallVector<unsigned, 4> LoQuad(4);
4022 SmallVector<unsigned, 4> HiQuad(4);
4023 BitVector InputQuads(4);
4024 for (unsigned i = 0; i < 8; ++i) {
4025 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4026 int EltIdx = SVOp->getMaskElt(i);
4027 MaskVals.push_back(EltIdx);
4036 InputQuads.set(EltIdx / 4);
4039 int BestLoQuad = -1;
4040 unsigned MaxQuad = 1;
4041 for (unsigned i = 0; i < 4; ++i) {
4042 if (LoQuad[i] > MaxQuad) {
4044 MaxQuad = LoQuad[i];
4048 int BestHiQuad = -1;
4050 for (unsigned i = 0; i < 4; ++i) {
4051 if (HiQuad[i] > MaxQuad) {
4053 MaxQuad = HiQuad[i];
4057 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4058 // of the two input vectors, shuffle them into one input vector so only a
4059 // single pshufb instruction is necessary. If There are more than 2 input
4060 // quads, disable the next transformation since it does not help SSSE3.
4061 bool V1Used = InputQuads[0] || InputQuads[1];
4062 bool V2Used = InputQuads[2] || InputQuads[3];
4063 if (TLI.getSubtarget()->hasSSSE3()) {
4064 if (InputQuads.count() == 2 && V1Used && V2Used) {
4065 BestLoQuad = InputQuads.find_first();
4066 BestHiQuad = InputQuads.find_next(BestLoQuad);
4068 if (InputQuads.count() > 2) {
4074 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4075 // the shuffle mask. If a quad is scored as -1, that means that it contains
4076 // words from all 4 input quadwords.
4078 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4079 SmallVector<int, 8> MaskV;
4080 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4081 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4082 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4083 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4084 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4085 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4087 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4088 // source words for the shuffle, to aid later transformations.
4089 bool AllWordsInNewV = true;
4090 bool InOrder[2] = { true, true };
4091 for (unsigned i = 0; i != 8; ++i) {
4092 int idx = MaskVals[i];
4094 InOrder[i/4] = false;
4095 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4097 AllWordsInNewV = false;
4101 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4102 if (AllWordsInNewV) {
4103 for (int i = 0; i != 8; ++i) {
4104 int idx = MaskVals[i];
4107 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4108 if ((idx != i) && idx < 4)
4110 if ((idx != i) && idx > 3)
4119 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4120 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4121 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4122 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4123 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4127 // If we have SSSE3, and all words of the result are from 1 input vector,
4128 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4129 // is present, fall back to case 4.
4130 if (TLI.getSubtarget()->hasSSSE3()) {
4131 SmallVector<SDValue,16> pshufbMask;
4133 // If we have elements from both input vectors, set the high bit of the
4134 // shuffle mask element to zero out elements that come from V2 in the V1
4135 // mask, and elements that come from V1 in the V2 mask, so that the two
4136 // results can be OR'd together.
4137 bool TwoInputs = V1Used && V2Used;
4138 for (unsigned i = 0; i != 8; ++i) {
4139 int EltIdx = MaskVals[i] * 2;
4140 if (TwoInputs && (EltIdx >= 16)) {
4141 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4142 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4145 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4146 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4148 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4149 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4150 DAG.getNode(ISD::BUILD_VECTOR, dl,
4151 MVT::v16i8, &pshufbMask[0], 16));
4153 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4155 // Calculate the shuffle mask for the second input, shuffle it, and
4156 // OR it with the first shuffled input.
4158 for (unsigned i = 0; i != 8; ++i) {
4159 int EltIdx = MaskVals[i] * 2;
4161 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4162 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4165 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4166 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4168 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4169 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4170 DAG.getNode(ISD::BUILD_VECTOR, dl,
4171 MVT::v16i8, &pshufbMask[0], 16));
4172 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4173 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4176 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4177 // and update MaskVals with new element order.
4178 BitVector InOrder(8);
4179 if (BestLoQuad >= 0) {
4180 SmallVector<int, 8> MaskV;
4181 for (int i = 0; i != 4; ++i) {
4182 int idx = MaskVals[i];
4184 MaskV.push_back(-1);
4186 } else if ((idx / 4) == BestLoQuad) {
4187 MaskV.push_back(idx & 3);
4190 MaskV.push_back(-1);
4193 for (unsigned i = 4; i != 8; ++i)
4195 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4199 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4200 // and update MaskVals with the new element order.
4201 if (BestHiQuad >= 0) {
4202 SmallVector<int, 8> MaskV;
4203 for (unsigned i = 0; i != 4; ++i)
4205 for (unsigned i = 4; i != 8; ++i) {
4206 int idx = MaskVals[i];
4208 MaskV.push_back(-1);
4210 } else if ((idx / 4) == BestHiQuad) {
4211 MaskV.push_back((idx & 3) + 4);
4214 MaskV.push_back(-1);
4217 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4221 // In case BestHi & BestLo were both -1, which means each quadword has a word
4222 // from each of the four input quadwords, calculate the InOrder bitvector now
4223 // before falling through to the insert/extract cleanup.
4224 if (BestLoQuad == -1 && BestHiQuad == -1) {
4226 for (int i = 0; i != 8; ++i)
4227 if (MaskVals[i] < 0 || MaskVals[i] == i)
4231 // The other elements are put in the right place using pextrw and pinsrw.
4232 for (unsigned i = 0; i != 8; ++i) {
4235 int EltIdx = MaskVals[i];
4238 SDValue ExtOp = (EltIdx < 8)
4239 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4240 DAG.getIntPtrConstant(EltIdx))
4241 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4242 DAG.getIntPtrConstant(EltIdx - 8));
4243 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4244 DAG.getIntPtrConstant(i));
4249 // v16i8 shuffles - Prefer shuffles in the following order:
4250 // 1. [ssse3] 1 x pshufb
4251 // 2. [ssse3] 2 x pshufb + 1 x por
4252 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4254 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4255 SelectionDAG &DAG, X86TargetLowering &TLI) {
4256 SDValue V1 = SVOp->getOperand(0);
4257 SDValue V2 = SVOp->getOperand(1);
4258 DebugLoc dl = SVOp->getDebugLoc();
4259 SmallVector<int, 16> MaskVals;
4260 SVOp->getMask(MaskVals);
4262 // If we have SSSE3, case 1 is generated when all result bytes come from
4263 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4264 // present, fall back to case 3.
4265 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4268 for (unsigned i = 0; i < 16; ++i) {
4269 int EltIdx = MaskVals[i];
4278 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4279 if (TLI.getSubtarget()->hasSSSE3()) {
4280 SmallVector<SDValue,16> pshufbMask;
4282 // If all result elements are from one input vector, then only translate
4283 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4285 // Otherwise, we have elements from both input vectors, and must zero out
4286 // elements that come from V2 in the first mask, and V1 in the second mask
4287 // so that we can OR them together.
4288 bool TwoInputs = !(V1Only || V2Only);
4289 for (unsigned i = 0; i != 16; ++i) {
4290 int EltIdx = MaskVals[i];
4291 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4292 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4295 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4297 // If all the elements are from V2, assign it to V1 and return after
4298 // building the first pshufb.
4301 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4302 DAG.getNode(ISD::BUILD_VECTOR, dl,
4303 MVT::v16i8, &pshufbMask[0], 16));
4307 // Calculate the shuffle mask for the second input, shuffle it, and
4308 // OR it with the first shuffled input.
4310 for (unsigned i = 0; i != 16; ++i) {
4311 int EltIdx = MaskVals[i];
4313 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4316 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4318 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4319 DAG.getNode(ISD::BUILD_VECTOR, dl,
4320 MVT::v16i8, &pshufbMask[0], 16));
4321 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4324 // No SSSE3 - Calculate in place words and then fix all out of place words
4325 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4326 // the 16 different words that comprise the two doublequadword input vectors.
4327 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4328 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4329 SDValue NewV = V2Only ? V2 : V1;
4330 for (int i = 0; i != 8; ++i) {
4331 int Elt0 = MaskVals[i*2];
4332 int Elt1 = MaskVals[i*2+1];
4334 // This word of the result is all undef, skip it.
4335 if (Elt0 < 0 && Elt1 < 0)
4338 // This word of the result is already in the correct place, skip it.
4339 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4341 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4344 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4345 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4348 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4349 // using a single extract together, load it and store it.
4350 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4351 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4352 DAG.getIntPtrConstant(Elt1 / 2));
4353 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4354 DAG.getIntPtrConstant(i));
4358 // If Elt1 is defined, extract it from the appropriate source. If the
4359 // source byte is not also odd, shift the extracted word left 8 bits
4360 // otherwise clear the bottom 8 bits if we need to do an or.
4362 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4363 DAG.getIntPtrConstant(Elt1 / 2));
4364 if ((Elt1 & 1) == 0)
4365 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4366 DAG.getConstant(8, TLI.getShiftAmountTy()));
4368 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4369 DAG.getConstant(0xFF00, MVT::i16));
4371 // If Elt0 is defined, extract it from the appropriate source. If the
4372 // source byte is not also even, shift the extracted word right 8 bits. If
4373 // Elt1 was also defined, OR the extracted values together before
4374 // inserting them in the result.
4376 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4377 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4378 if ((Elt0 & 1) != 0)
4379 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4380 DAG.getConstant(8, TLI.getShiftAmountTy()));
4382 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4383 DAG.getConstant(0x00FF, MVT::i16));
4384 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4387 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4388 DAG.getIntPtrConstant(i));
4390 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4393 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4394 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4395 /// done when every pair / quad of shuffle mask elements point to elements in
4396 /// the right sequence. e.g.
4397 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4399 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4401 TargetLowering &TLI, DebugLoc dl) {
4402 EVT VT = SVOp->getValueType(0);
4403 SDValue V1 = SVOp->getOperand(0);
4404 SDValue V2 = SVOp->getOperand(1);
4405 unsigned NumElems = VT.getVectorNumElements();
4406 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4407 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4408 EVT MaskEltVT = MaskVT.getVectorElementType();
4410 switch (VT.getSimpleVT().SimpleTy) {
4411 default: assert(false && "Unexpected!");
4412 case MVT::v4f32: NewVT = MVT::v2f64; break;
4413 case MVT::v4i32: NewVT = MVT::v2i64; break;
4414 case MVT::v8i16: NewVT = MVT::v4i32; break;
4415 case MVT::v16i8: NewVT = MVT::v4i32; break;
4418 if (NewWidth == 2) {
4424 int Scale = NumElems / NewWidth;
4425 SmallVector<int, 8> MaskVec;
4426 for (unsigned i = 0; i < NumElems; i += Scale) {
4428 for (int j = 0; j < Scale; ++j) {
4429 int EltIdx = SVOp->getMaskElt(i+j);
4433 StartIdx = EltIdx - (EltIdx % Scale);
4434 if (EltIdx != StartIdx + j)
4438 MaskVec.push_back(-1);
4440 MaskVec.push_back(StartIdx / Scale);
4443 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4444 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4445 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4448 /// getVZextMovL - Return a zero-extending vector move low node.
4450 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4451 SDValue SrcOp, SelectionDAG &DAG,
4452 const X86Subtarget *Subtarget, DebugLoc dl) {
4453 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4454 LoadSDNode *LD = NULL;
4455 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4456 LD = dyn_cast<LoadSDNode>(SrcOp);
4458 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4460 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4461 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4462 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4463 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4464 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4466 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4467 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4468 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4469 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4477 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4478 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4479 DAG.getNode(ISD::BIT_CONVERT, dl,
4483 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4486 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4487 SDValue V1 = SVOp->getOperand(0);
4488 SDValue V2 = SVOp->getOperand(1);
4489 DebugLoc dl = SVOp->getDebugLoc();
4490 EVT VT = SVOp->getValueType(0);
4492 SmallVector<std::pair<int, int>, 8> Locs;
4494 SmallVector<int, 8> Mask1(4U, -1);
4495 SmallVector<int, 8> PermMask;
4496 SVOp->getMask(PermMask);
4500 for (unsigned i = 0; i != 4; ++i) {
4501 int Idx = PermMask[i];
4503 Locs[i] = std::make_pair(-1, -1);
4505 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4507 Locs[i] = std::make_pair(0, NumLo);
4511 Locs[i] = std::make_pair(1, NumHi);
4513 Mask1[2+NumHi] = Idx;
4519 if (NumLo <= 2 && NumHi <= 2) {
4520 // If no more than two elements come from either vector. This can be
4521 // implemented with two shuffles. First shuffle gather the elements.
4522 // The second shuffle, which takes the first shuffle as both of its
4523 // vector operands, put the elements into the right order.
4524 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4526 SmallVector<int, 8> Mask2(4U, -1);
4528 for (unsigned i = 0; i != 4; ++i) {
4529 if (Locs[i].first == -1)
4532 unsigned Idx = (i < 2) ? 0 : 4;
4533 Idx += Locs[i].first * 2 + Locs[i].second;
4538 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4539 } else if (NumLo == 3 || NumHi == 3) {
4540 // Otherwise, we must have three elements from one vector, call it X, and
4541 // one element from the other, call it Y. First, use a shufps to build an
4542 // intermediate vector with the one element from Y and the element from X
4543 // that will be in the same half in the final destination (the indexes don't
4544 // matter). Then, use a shufps to build the final vector, taking the half
4545 // containing the element from Y from the intermediate, and the other half
4548 // Normalize it so the 3 elements come from V1.
4549 CommuteVectorShuffleMask(PermMask, VT);
4553 // Find the element from V2.
4555 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4556 int Val = PermMask[HiIndex];
4563 Mask1[0] = PermMask[HiIndex];
4565 Mask1[2] = PermMask[HiIndex^1];
4567 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4570 Mask1[0] = PermMask[0];
4571 Mask1[1] = PermMask[1];
4572 Mask1[2] = HiIndex & 1 ? 6 : 4;
4573 Mask1[3] = HiIndex & 1 ? 4 : 6;
4574 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4576 Mask1[0] = HiIndex & 1 ? 2 : 0;
4577 Mask1[1] = HiIndex & 1 ? 0 : 2;
4578 Mask1[2] = PermMask[2];
4579 Mask1[3] = PermMask[3];
4584 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4588 // Break it into (shuffle shuffle_hi, shuffle_lo).
4590 SmallVector<int,8> LoMask(4U, -1);
4591 SmallVector<int,8> HiMask(4U, -1);
4593 SmallVector<int,8> *MaskPtr = &LoMask;
4594 unsigned MaskIdx = 0;
4597 for (unsigned i = 0; i != 4; ++i) {
4604 int Idx = PermMask[i];
4606 Locs[i] = std::make_pair(-1, -1);
4607 } else if (Idx < 4) {
4608 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4609 (*MaskPtr)[LoIdx] = Idx;
4612 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4613 (*MaskPtr)[HiIdx] = Idx;
4618 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4619 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4620 SmallVector<int, 8> MaskOps;
4621 for (unsigned i = 0; i != 4; ++i) {
4622 if (Locs[i].first == -1) {
4623 MaskOps.push_back(-1);
4625 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4626 MaskOps.push_back(Idx);
4629 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4633 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4634 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4635 SDValue V1 = Op.getOperand(0);
4636 SDValue V2 = Op.getOperand(1);
4637 EVT VT = Op.getValueType();
4638 DebugLoc dl = Op.getDebugLoc();
4639 unsigned NumElems = VT.getVectorNumElements();
4640 bool isMMX = VT.getSizeInBits() == 64;
4641 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4642 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4643 bool V1IsSplat = false;
4644 bool V2IsSplat = false;
4646 if (isZeroShuffle(SVOp))
4647 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4649 // Promote splats to v4f32.
4650 if (SVOp->isSplat()) {
4651 if (isMMX || NumElems < 4)
4653 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4656 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4658 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4659 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4660 if (NewOp.getNode())
4661 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4662 LowerVECTOR_SHUFFLE(NewOp, DAG));
4663 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4664 // FIXME: Figure out a cleaner way to do this.
4665 // Try to make use of movq to zero out the top part.
4666 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4667 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4668 if (NewOp.getNode()) {
4669 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4670 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4671 DAG, Subtarget, dl);
4673 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4674 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4675 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4676 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4677 DAG, Subtarget, dl);
4681 if (X86::isPSHUFDMask(SVOp))
4684 // Check if this can be converted into a logical shift.
4685 bool isLeft = false;
4688 bool isShift = getSubtarget()->hasSSE2() &&
4689 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4690 if (isShift && ShVal.hasOneUse()) {
4691 // If the shifted value has multiple uses, it may be cheaper to use
4692 // v_set0 + movlhps or movhlps, etc.
4693 EVT EltVT = VT.getVectorElementType();
4694 ShAmt *= EltVT.getSizeInBits();
4695 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4698 if (X86::isMOVLMask(SVOp)) {
4701 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4702 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4707 // FIXME: fold these into legal mask.
4708 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4709 X86::isMOVSLDUPMask(SVOp) ||
4710 X86::isMOVHLPSMask(SVOp) ||
4711 X86::isMOVLHPSMask(SVOp) ||
4712 X86::isMOVLPMask(SVOp)))
4715 if (ShouldXformToMOVHLPS(SVOp) ||
4716 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4717 return CommuteVectorShuffle(SVOp, DAG);
4720 // No better options. Use a vshl / vsrl.
4721 EVT EltVT = VT.getVectorElementType();
4722 ShAmt *= EltVT.getSizeInBits();
4723 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4726 bool Commuted = false;
4727 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4728 // 1,1,1,1 -> v8i16 though.
4729 V1IsSplat = isSplatVector(V1.getNode());
4730 V2IsSplat = isSplatVector(V2.getNode());
4732 // Canonicalize the splat or undef, if present, to be on the RHS.
4733 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4734 Op = CommuteVectorShuffle(SVOp, DAG);
4735 SVOp = cast<ShuffleVectorSDNode>(Op);
4736 V1 = SVOp->getOperand(0);
4737 V2 = SVOp->getOperand(1);
4738 std::swap(V1IsSplat, V2IsSplat);
4739 std::swap(V1IsUndef, V2IsUndef);
4743 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4744 // Shuffling low element of v1 into undef, just return v1.
4747 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4748 // the instruction selector will not match, so get a canonical MOVL with
4749 // swapped operands to undo the commute.
4750 return getMOVL(DAG, dl, VT, V2, V1);
4753 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4754 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4755 X86::isUNPCKLMask(SVOp) ||
4756 X86::isUNPCKHMask(SVOp))
4760 // Normalize mask so all entries that point to V2 points to its first
4761 // element then try to match unpck{h|l} again. If match, return a
4762 // new vector_shuffle with the corrected mask.
4763 SDValue NewMask = NormalizeMask(SVOp, DAG);
4764 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4765 if (NSVOp != SVOp) {
4766 if (X86::isUNPCKLMask(NSVOp, true)) {
4768 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4775 // Commute is back and try unpck* again.
4776 // FIXME: this seems wrong.
4777 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4778 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4779 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4780 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4781 X86::isUNPCKLMask(NewSVOp) ||
4782 X86::isUNPCKHMask(NewSVOp))
4786 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4788 // Normalize the node to match x86 shuffle ops if needed
4789 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4790 return CommuteVectorShuffle(SVOp, DAG);
4792 // Check for legal shuffle and return?
4793 SmallVector<int, 16> PermMask;
4794 SVOp->getMask(PermMask);
4795 if (isShuffleMaskLegal(PermMask, VT))
4798 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4799 if (VT == MVT::v8i16) {
4800 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4801 if (NewOp.getNode())
4805 if (VT == MVT::v16i8) {
4806 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4807 if (NewOp.getNode())
4811 // Handle all 4 wide cases with a number of shuffles except for MMX.
4812 if (NumElems == 4 && !isMMX)
4813 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4819 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4820 SelectionDAG &DAG) {
4821 EVT VT = Op.getValueType();
4822 DebugLoc dl = Op.getDebugLoc();
4823 if (VT.getSizeInBits() == 8) {
4824 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4825 Op.getOperand(0), Op.getOperand(1));
4826 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4827 DAG.getValueType(VT));
4828 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4829 } else if (VT.getSizeInBits() == 16) {
4830 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4831 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4833 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4834 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4835 DAG.getNode(ISD::BIT_CONVERT, dl,
4839 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4840 Op.getOperand(0), Op.getOperand(1));
4841 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4842 DAG.getValueType(VT));
4843 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4844 } else if (VT == MVT::f32) {
4845 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4846 // the result back to FR32 register. It's only worth matching if the
4847 // result has a single use which is a store or a bitcast to i32. And in
4848 // the case of a store, it's not worth it if the index is a constant 0,
4849 // because a MOVSSmr can be used instead, which is smaller and faster.
4850 if (!Op.hasOneUse())
4852 SDNode *User = *Op.getNode()->use_begin();
4853 if ((User->getOpcode() != ISD::STORE ||
4854 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4855 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4856 (User->getOpcode() != ISD::BIT_CONVERT ||
4857 User->getValueType(0) != MVT::i32))
4859 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4860 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4863 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4864 } else if (VT == MVT::i32) {
4865 // ExtractPS works with constant index.
4866 if (isa<ConstantSDNode>(Op.getOperand(1)))
4874 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4875 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4878 if (Subtarget->hasSSE41()) {
4879 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4884 EVT VT = Op.getValueType();
4885 DebugLoc dl = Op.getDebugLoc();
4886 // TODO: handle v16i8.
4887 if (VT.getSizeInBits() == 16) {
4888 SDValue Vec = Op.getOperand(0);
4889 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4891 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4892 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4893 DAG.getNode(ISD::BIT_CONVERT, dl,
4896 // Transform it so it match pextrw which produces a 32-bit result.
4897 EVT EltVT = MVT::i32;
4898 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4899 Op.getOperand(0), Op.getOperand(1));
4900 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4901 DAG.getValueType(VT));
4902 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4903 } else if (VT.getSizeInBits() == 32) {
4904 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4908 // SHUFPS the element to the lowest double word, then movss.
4909 int Mask[4] = { Idx, -1, -1, -1 };
4910 EVT VVT = Op.getOperand(0).getValueType();
4911 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4912 DAG.getUNDEF(VVT), Mask);
4913 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4914 DAG.getIntPtrConstant(0));
4915 } else if (VT.getSizeInBits() == 64) {
4916 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4917 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4918 // to match extract_elt for f64.
4919 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4923 // UNPCKHPD the element to the lowest double word, then movsd.
4924 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4925 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4926 int Mask[2] = { 1, -1 };
4927 EVT VVT = Op.getOperand(0).getValueType();
4928 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4929 DAG.getUNDEF(VVT), Mask);
4930 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4931 DAG.getIntPtrConstant(0));
4938 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4939 EVT VT = Op.getValueType();
4940 EVT EltVT = VT.getVectorElementType();
4941 DebugLoc dl = Op.getDebugLoc();
4943 SDValue N0 = Op.getOperand(0);
4944 SDValue N1 = Op.getOperand(1);
4945 SDValue N2 = Op.getOperand(2);
4947 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4948 isa<ConstantSDNode>(N2)) {
4950 if (VT == MVT::v8i16)
4951 Opc = X86ISD::PINSRW;
4952 else if (VT == MVT::v4i16)
4953 Opc = X86ISD::MMX_PINSRW;
4954 else if (VT == MVT::v16i8)
4955 Opc = X86ISD::PINSRB;
4957 Opc = X86ISD::PINSRB;
4959 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4961 if (N1.getValueType() != MVT::i32)
4962 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4963 if (N2.getValueType() != MVT::i32)
4964 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4965 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4966 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4967 // Bits [7:6] of the constant are the source select. This will always be
4968 // zero here. The DAG Combiner may combine an extract_elt index into these
4969 // bits. For example (insert (extract, 3), 2) could be matched by putting
4970 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4971 // Bits [5:4] of the constant are the destination select. This is the
4972 // value of the incoming immediate.
4973 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4974 // combine either bitwise AND or insert of float 0.0 to set these bits.
4975 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4976 // Create this as a scalar to vector..
4977 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4978 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4979 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4980 // PINSR* works with constant index.
4987 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4988 EVT VT = Op.getValueType();
4989 EVT EltVT = VT.getVectorElementType();
4991 if (Subtarget->hasSSE41())
4992 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4994 if (EltVT == MVT::i8)
4997 DebugLoc dl = Op.getDebugLoc();
4998 SDValue N0 = Op.getOperand(0);
4999 SDValue N1 = Op.getOperand(1);
5000 SDValue N2 = Op.getOperand(2);
5002 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5003 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5004 // as its second argument.
5005 if (N1.getValueType() != MVT::i32)
5006 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5007 if (N2.getValueType() != MVT::i32)
5008 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5009 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5010 dl, VT, N0, N1, N2);
5016 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
5017 DebugLoc dl = Op.getDebugLoc();
5018 if (Op.getValueType() == MVT::v2f32)
5019 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
5020 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
5021 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
5022 Op.getOperand(0))));
5024 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
5025 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5027 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5028 EVT VT = MVT::v2i32;
5029 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5036 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5037 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5040 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5041 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5042 // one of the above mentioned nodes. It has to be wrapped because otherwise
5043 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5044 // be used to form addressing mode. These wrapped nodes will be selected
5047 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
5048 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5050 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5052 unsigned char OpFlag = 0;
5053 unsigned WrapperKind = X86ISD::Wrapper;
5054 CodeModel::Model M = getTargetMachine().getCodeModel();
5056 if (Subtarget->isPICStyleRIPRel() &&
5057 (M == CodeModel::Small || M == CodeModel::Kernel))
5058 WrapperKind = X86ISD::WrapperRIP;
5059 else if (Subtarget->isPICStyleGOT())
5060 OpFlag = X86II::MO_GOTOFF;
5061 else if (Subtarget->isPICStyleStubPIC())
5062 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5064 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5066 CP->getOffset(), OpFlag);
5067 DebugLoc DL = CP->getDebugLoc();
5068 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5069 // With PIC, the address is actually $g + Offset.
5071 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5072 DAG.getNode(X86ISD::GlobalBaseReg,
5073 DebugLoc(), getPointerTy()),
5080 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
5081 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5083 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5085 unsigned char OpFlag = 0;
5086 unsigned WrapperKind = X86ISD::Wrapper;
5087 CodeModel::Model M = getTargetMachine().getCodeModel();
5089 if (Subtarget->isPICStyleRIPRel() &&
5090 (M == CodeModel::Small || M == CodeModel::Kernel))
5091 WrapperKind = X86ISD::WrapperRIP;
5092 else if (Subtarget->isPICStyleGOT())
5093 OpFlag = X86II::MO_GOTOFF;
5094 else if (Subtarget->isPICStyleStubPIC())
5095 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5097 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5099 DebugLoc DL = JT->getDebugLoc();
5100 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5102 // With PIC, the address is actually $g + Offset.
5104 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5105 DAG.getNode(X86ISD::GlobalBaseReg,
5106 DebugLoc(), getPointerTy()),
5114 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
5115 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5117 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5119 unsigned char OpFlag = 0;
5120 unsigned WrapperKind = X86ISD::Wrapper;
5121 CodeModel::Model M = getTargetMachine().getCodeModel();
5123 if (Subtarget->isPICStyleRIPRel() &&
5124 (M == CodeModel::Small || M == CodeModel::Kernel))
5125 WrapperKind = X86ISD::WrapperRIP;
5126 else if (Subtarget->isPICStyleGOT())
5127 OpFlag = X86II::MO_GOTOFF;
5128 else if (Subtarget->isPICStyleStubPIC())
5129 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5131 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5133 DebugLoc DL = Op.getDebugLoc();
5134 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5137 // With PIC, the address is actually $g + Offset.
5138 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5139 !Subtarget->is64Bit()) {
5140 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5141 DAG.getNode(X86ISD::GlobalBaseReg,
5142 DebugLoc(), getPointerTy()),
5150 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5151 // Create the TargetBlockAddressAddress node.
5152 unsigned char OpFlags =
5153 Subtarget->ClassifyBlockAddressReference();
5154 CodeModel::Model M = getTargetMachine().getCodeModel();
5155 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5156 DebugLoc dl = Op.getDebugLoc();
5157 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5158 /*isTarget=*/true, OpFlags);
5160 if (Subtarget->isPICStyleRIPRel() &&
5161 (M == CodeModel::Small || M == CodeModel::Kernel))
5162 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5164 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5166 // With PIC, the address is actually $g + Offset.
5167 if (isGlobalRelativeToPICBase(OpFlags)) {
5168 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5169 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5177 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5179 SelectionDAG &DAG) const {
5180 // Create the TargetGlobalAddress node, folding in the constant
5181 // offset if it is legal.
5182 unsigned char OpFlags =
5183 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5184 CodeModel::Model M = getTargetMachine().getCodeModel();
5186 if (OpFlags == X86II::MO_NO_FLAG &&
5187 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5188 // A direct static reference to a global.
5189 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5192 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5195 if (Subtarget->isPICStyleRIPRel() &&
5196 (M == CodeModel::Small || M == CodeModel::Kernel))
5197 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5199 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5201 // With PIC, the address is actually $g + Offset.
5202 if (isGlobalRelativeToPICBase(OpFlags)) {
5203 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5204 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5208 // For globals that require a load from a stub to get the address, emit the
5210 if (isGlobalStubReference(OpFlags))
5211 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5212 PseudoSourceValue::getGOT(), 0, false, false, 0);
5214 // If there was a non-zero offset that we didn't fold, create an explicit
5217 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5218 DAG.getConstant(Offset, getPointerTy()));
5224 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5225 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5226 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5227 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5231 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5232 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5233 unsigned char OperandFlags) {
5234 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5235 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5236 DebugLoc dl = GA->getDebugLoc();
5237 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5238 GA->getValueType(0),
5242 SDValue Ops[] = { Chain, TGA, *InFlag };
5243 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5245 SDValue Ops[] = { Chain, TGA };
5246 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5249 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5250 MFI->setHasCalls(true);
5252 SDValue Flag = Chain.getValue(1);
5253 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5256 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5258 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5261 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5262 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5263 DAG.getNode(X86ISD::GlobalBaseReg,
5264 DebugLoc(), PtrVT), InFlag);
5265 InFlag = Chain.getValue(1);
5267 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5270 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5272 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5274 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5275 X86::RAX, X86II::MO_TLSGD);
5278 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5279 // "local exec" model.
5280 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5281 const EVT PtrVT, TLSModel::Model model,
5283 DebugLoc dl = GA->getDebugLoc();
5284 // Get the Thread Pointer
5285 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5287 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5290 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5291 NULL, 0, false, false, 0);
5293 unsigned char OperandFlags = 0;
5294 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5296 unsigned WrapperKind = X86ISD::Wrapper;
5297 if (model == TLSModel::LocalExec) {
5298 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5299 } else if (is64Bit) {
5300 assert(model == TLSModel::InitialExec);
5301 OperandFlags = X86II::MO_GOTTPOFF;
5302 WrapperKind = X86ISD::WrapperRIP;
5304 assert(model == TLSModel::InitialExec);
5305 OperandFlags = X86II::MO_INDNTPOFF;
5308 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5310 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5311 GA->getOffset(), OperandFlags);
5312 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5314 if (model == TLSModel::InitialExec)
5315 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5316 PseudoSourceValue::getGOT(), 0, false, false, 0);
5318 // The address of the thread local variable is the add of the thread
5319 // pointer with the offset of the variable.
5320 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5324 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5325 // TODO: implement the "local dynamic" model
5326 // TODO: implement the "initial exec"model for pic executables
5327 assert(Subtarget->isTargetELF() &&
5328 "TLS not implemented for non-ELF targets");
5329 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5330 const GlobalValue *GV = GA->getGlobal();
5332 // If GV is an alias then use the aliasee for determining
5333 // thread-localness.
5334 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5335 GV = GA->resolveAliasedGlobal(false);
5337 TLSModel::Model model = getTLSModel(GV,
5338 getTargetMachine().getRelocationModel());
5341 case TLSModel::GeneralDynamic:
5342 case TLSModel::LocalDynamic: // not implemented
5343 if (Subtarget->is64Bit())
5344 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5345 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5347 case TLSModel::InitialExec:
5348 case TLSModel::LocalExec:
5349 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5350 Subtarget->is64Bit());
5353 llvm_unreachable("Unreachable");
5358 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5359 /// take a 2 x i32 value to shift plus a shift amount.
5360 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5361 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5362 EVT VT = Op.getValueType();
5363 unsigned VTBits = VT.getSizeInBits();
5364 DebugLoc dl = Op.getDebugLoc();
5365 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5366 SDValue ShOpLo = Op.getOperand(0);
5367 SDValue ShOpHi = Op.getOperand(1);
5368 SDValue ShAmt = Op.getOperand(2);
5369 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5370 DAG.getConstant(VTBits - 1, MVT::i8))
5371 : DAG.getConstant(0, VT);
5374 if (Op.getOpcode() == ISD::SHL_PARTS) {
5375 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5376 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5378 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5379 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5382 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5383 DAG.getConstant(VTBits, MVT::i8));
5384 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5385 AndNode, DAG.getConstant(0, MVT::i8));
5388 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5389 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5390 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5392 if (Op.getOpcode() == ISD::SHL_PARTS) {
5393 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5394 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5396 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5397 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5400 SDValue Ops[2] = { Lo, Hi };
5401 return DAG.getMergeValues(Ops, 2, dl);
5404 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5405 EVT SrcVT = Op.getOperand(0).getValueType();
5407 if (SrcVT.isVector()) {
5408 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5414 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5415 "Unknown SINT_TO_FP to lower!");
5417 // These are really Legal; return the operand so the caller accepts it as
5419 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5421 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5422 Subtarget->is64Bit()) {
5426 DebugLoc dl = Op.getDebugLoc();
5427 unsigned Size = SrcVT.getSizeInBits()/8;
5428 MachineFunction &MF = DAG.getMachineFunction();
5429 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5430 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5431 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5433 PseudoSourceValue::getFixedStack(SSFI), 0,
5435 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5438 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5440 SelectionDAG &DAG) {
5442 DebugLoc dl = Op.getDebugLoc();
5444 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5446 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5448 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5449 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5450 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5451 Tys, Ops, array_lengthof(Ops));
5454 Chain = Result.getValue(1);
5455 SDValue InFlag = Result.getValue(2);
5457 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5458 // shouldn't be necessary except that RFP cannot be live across
5459 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5460 MachineFunction &MF = DAG.getMachineFunction();
5461 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5462 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5463 Tys = DAG.getVTList(MVT::Other);
5465 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5467 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5468 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5469 PseudoSourceValue::getFixedStack(SSFI), 0,
5476 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5477 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5478 // This algorithm is not obvious. Here it is in C code, more or less:
5480 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5481 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5482 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5484 // Copy ints to xmm registers.
5485 __m128i xh = _mm_cvtsi32_si128( hi );
5486 __m128i xl = _mm_cvtsi32_si128( lo );
5488 // Combine into low half of a single xmm register.
5489 __m128i x = _mm_unpacklo_epi32( xh, xl );
5493 // Merge in appropriate exponents to give the integer bits the right
5495 x = _mm_unpacklo_epi32( x, exp );
5497 // Subtract away the biases to deal with the IEEE-754 double precision
5499 d = _mm_sub_pd( (__m128d) x, bias );
5501 // All conversions up to here are exact. The correctly rounded result is
5502 // calculated using the current rounding mode using the following
5504 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5505 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5506 // store doesn't really need to be here (except
5507 // maybe to zero the other double)
5512 DebugLoc dl = Op.getDebugLoc();
5513 LLVMContext *Context = DAG.getContext();
5515 // Build some magic constants.
5516 std::vector<Constant*> CV0;
5517 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5518 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5519 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5520 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5521 Constant *C0 = ConstantVector::get(CV0);
5522 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5524 std::vector<Constant*> CV1;
5526 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5528 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5529 Constant *C1 = ConstantVector::get(CV1);
5530 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5532 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5533 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5535 DAG.getIntPtrConstant(1)));
5536 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5537 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5539 DAG.getIntPtrConstant(0)));
5540 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5541 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5542 PseudoSourceValue::getConstantPool(), 0,
5544 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5545 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5546 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5547 PseudoSourceValue::getConstantPool(), 0,
5549 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5551 // Add the halves; easiest way is to swap them into another reg first.
5552 int ShufMask[2] = { 1, -1 };
5553 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5554 DAG.getUNDEF(MVT::v2f64), ShufMask);
5555 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5556 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5557 DAG.getIntPtrConstant(0));
5560 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5561 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5562 DebugLoc dl = Op.getDebugLoc();
5563 // FP constant to bias correct the final result.
5564 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5567 // Load the 32-bit value into an XMM register.
5568 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5569 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5571 DAG.getIntPtrConstant(0)));
5573 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5574 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5575 DAG.getIntPtrConstant(0));
5577 // Or the load with the bias.
5578 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5579 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5580 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5582 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5583 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5584 MVT::v2f64, Bias)));
5585 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5586 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5587 DAG.getIntPtrConstant(0));
5589 // Subtract the bias.
5590 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5592 // Handle final rounding.
5593 EVT DestVT = Op.getValueType();
5595 if (DestVT.bitsLT(MVT::f64)) {
5596 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5597 DAG.getIntPtrConstant(0));
5598 } else if (DestVT.bitsGT(MVT::f64)) {
5599 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5602 // Handle final rounding.
5606 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5607 SDValue N0 = Op.getOperand(0);
5608 DebugLoc dl = Op.getDebugLoc();
5610 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5611 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5612 // the optimization here.
5613 if (DAG.SignBitIsZero(N0))
5614 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5616 EVT SrcVT = N0.getValueType();
5617 if (SrcVT == MVT::i64) {
5618 // We only handle SSE2 f64 target here; caller can expand the rest.
5619 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5622 return LowerUINT_TO_FP_i64(Op, DAG);
5623 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5624 return LowerUINT_TO_FP_i32(Op, DAG);
5627 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5629 // Make a 64-bit buffer, and use it to build an FILD.
5630 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5631 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5632 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5633 getPointerTy(), StackSlot, WordOff);
5634 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5635 StackSlot, NULL, 0, false, false, 0);
5636 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5637 OffsetSlot, NULL, 0, false, false, 0);
5638 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5641 std::pair<SDValue,SDValue> X86TargetLowering::
5642 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5643 DebugLoc dl = Op.getDebugLoc();
5645 EVT DstTy = Op.getValueType();
5648 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5652 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5653 DstTy.getSimpleVT() >= MVT::i16 &&
5654 "Unknown FP_TO_SINT to lower!");
5656 // These are really Legal.
5657 if (DstTy == MVT::i32 &&
5658 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5659 return std::make_pair(SDValue(), SDValue());
5660 if (Subtarget->is64Bit() &&
5661 DstTy == MVT::i64 &&
5662 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5663 return std::make_pair(SDValue(), SDValue());
5665 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5667 MachineFunction &MF = DAG.getMachineFunction();
5668 unsigned MemSize = DstTy.getSizeInBits()/8;
5669 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5670 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5673 switch (DstTy.getSimpleVT().SimpleTy) {
5674 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5675 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5676 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5677 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5680 SDValue Chain = DAG.getEntryNode();
5681 SDValue Value = Op.getOperand(0);
5682 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5683 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5684 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5685 PseudoSourceValue::getFixedStack(SSFI), 0,
5687 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5689 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5691 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5692 Chain = Value.getValue(1);
5693 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5694 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5697 // Build the FP_TO_INT*_IN_MEM
5698 SDValue Ops[] = { Chain, Value, StackSlot };
5699 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5701 return std::make_pair(FIST, StackSlot);
5704 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5705 if (Op.getValueType().isVector()) {
5706 if (Op.getValueType() == MVT::v2i32 &&
5707 Op.getOperand(0).getValueType() == MVT::v2f64) {
5713 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5714 SDValue FIST = Vals.first, StackSlot = Vals.second;
5715 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5716 if (FIST.getNode() == 0) return Op;
5719 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5720 FIST, StackSlot, NULL, 0, false, false, 0);
5723 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5724 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5725 SDValue FIST = Vals.first, StackSlot = Vals.second;
5726 assert(FIST.getNode() && "Unexpected failure");
5729 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5730 FIST, StackSlot, NULL, 0, false, false, 0);
5733 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5734 LLVMContext *Context = DAG.getContext();
5735 DebugLoc dl = Op.getDebugLoc();
5736 EVT VT = Op.getValueType();
5739 EltVT = VT.getVectorElementType();
5740 std::vector<Constant*> CV;
5741 if (EltVT == MVT::f64) {
5742 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5746 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5752 Constant *C = ConstantVector::get(CV);
5753 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5754 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5755 PseudoSourceValue::getConstantPool(), 0,
5757 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5760 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5761 LLVMContext *Context = DAG.getContext();
5762 DebugLoc dl = Op.getDebugLoc();
5763 EVT VT = Op.getValueType();
5766 EltVT = VT.getVectorElementType();
5767 std::vector<Constant*> CV;
5768 if (EltVT == MVT::f64) {
5769 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5773 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5779 Constant *C = ConstantVector::get(CV);
5780 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5781 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5782 PseudoSourceValue::getConstantPool(), 0,
5784 if (VT.isVector()) {
5785 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5786 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5787 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5789 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5791 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5795 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5796 LLVMContext *Context = DAG.getContext();
5797 SDValue Op0 = Op.getOperand(0);
5798 SDValue Op1 = Op.getOperand(1);
5799 DebugLoc dl = Op.getDebugLoc();
5800 EVT VT = Op.getValueType();
5801 EVT SrcVT = Op1.getValueType();
5803 // If second operand is smaller, extend it first.
5804 if (SrcVT.bitsLT(VT)) {
5805 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5808 // And if it is bigger, shrink it first.
5809 if (SrcVT.bitsGT(VT)) {
5810 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5814 // At this point the operands and the result should have the same
5815 // type, and that won't be f80 since that is not custom lowered.
5817 // First get the sign bit of second operand.
5818 std::vector<Constant*> CV;
5819 if (SrcVT == MVT::f64) {
5820 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5821 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5823 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5824 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5825 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5826 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5828 Constant *C = ConstantVector::get(CV);
5829 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5830 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5831 PseudoSourceValue::getConstantPool(), 0,
5833 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5835 // Shift sign bit right or left if the two operands have different types.
5836 if (SrcVT.bitsGT(VT)) {
5837 // Op0 is MVT::f32, Op1 is MVT::f64.
5838 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5839 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5840 DAG.getConstant(32, MVT::i32));
5841 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5842 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5843 DAG.getIntPtrConstant(0));
5846 // Clear first operand sign bit.
5848 if (VT == MVT::f64) {
5849 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5850 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5852 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5853 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5854 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5855 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5857 C = ConstantVector::get(CV);
5858 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5859 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5860 PseudoSourceValue::getConstantPool(), 0,
5862 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5864 // Or the value with the sign bit.
5865 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5868 /// Emit nodes that will be selected as "test Op0,Op0", or something
5870 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5871 SelectionDAG &DAG) {
5872 DebugLoc dl = Op.getDebugLoc();
5874 // CF and OF aren't always set the way we want. Determine which
5875 // of these we need.
5876 bool NeedCF = false;
5877 bool NeedOF = false;
5879 case X86::COND_A: case X86::COND_AE:
5880 case X86::COND_B: case X86::COND_BE:
5883 case X86::COND_G: case X86::COND_GE:
5884 case X86::COND_L: case X86::COND_LE:
5885 case X86::COND_O: case X86::COND_NO:
5891 // See if we can use the EFLAGS value from the operand instead of
5892 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5893 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5894 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5895 unsigned Opcode = 0;
5896 unsigned NumOperands = 0;
5897 switch (Op.getNode()->getOpcode()) {
5899 // Due to an isel shortcoming, be conservative if this add is likely to
5900 // be selected as part of a load-modify-store instruction. When the root
5901 // node in a match is a store, isel doesn't know how to remap non-chain
5902 // non-flag uses of other nodes in the match, such as the ADD in this
5903 // case. This leads to the ADD being left around and reselected, with
5904 // the result being two adds in the output.
5905 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5906 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5907 if (UI->getOpcode() == ISD::STORE)
5909 if (ConstantSDNode *C =
5910 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5911 // An add of one will be selected as an INC.
5912 if (C->getAPIntValue() == 1) {
5913 Opcode = X86ISD::INC;
5917 // An add of negative one (subtract of one) will be selected as a DEC.
5918 if (C->getAPIntValue().isAllOnesValue()) {
5919 Opcode = X86ISD::DEC;
5924 // Otherwise use a regular EFLAGS-setting add.
5925 Opcode = X86ISD::ADD;
5929 // If the primary and result isn't used, don't bother using X86ISD::AND,
5930 // because a TEST instruction will be better.
5931 bool NonFlagUse = false;
5932 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5933 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5935 unsigned UOpNo = UI.getOperandNo();
5936 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5937 // Look pass truncate.
5938 UOpNo = User->use_begin().getOperandNo();
5939 User = *User->use_begin();
5941 if (User->getOpcode() != ISD::BRCOND &&
5942 User->getOpcode() != ISD::SETCC &&
5943 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5955 // Due to the ISEL shortcoming noted above, be conservative if this op is
5956 // likely to be selected as part of a load-modify-store instruction.
5957 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5958 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5959 if (UI->getOpcode() == ISD::STORE)
5961 // Otherwise use a regular EFLAGS-setting instruction.
5962 switch (Op.getNode()->getOpcode()) {
5963 case ISD::SUB: Opcode = X86ISD::SUB; break;
5964 case ISD::OR: Opcode = X86ISD::OR; break;
5965 case ISD::XOR: Opcode = X86ISD::XOR; break;
5966 case ISD::AND: Opcode = X86ISD::AND; break;
5967 default: llvm_unreachable("unexpected operator!");
5978 return SDValue(Op.getNode(), 1);
5984 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5985 SmallVector<SDValue, 4> Ops;
5986 for (unsigned i = 0; i != NumOperands; ++i)
5987 Ops.push_back(Op.getOperand(i));
5988 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5989 DAG.ReplaceAllUsesWith(Op, New);
5990 return SDValue(New.getNode(), 1);
5994 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5995 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5996 DAG.getConstant(0, Op.getValueType()));
5999 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6001 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6002 SelectionDAG &DAG) {
6003 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6004 if (C->getAPIntValue() == 0)
6005 return EmitTest(Op0, X86CC, DAG);
6007 DebugLoc dl = Op0.getDebugLoc();
6008 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6011 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6012 /// if it's possible.
6013 static SDValue LowerToBT(SDValue And, ISD::CondCode CC,
6014 DebugLoc dl, SelectionDAG &DAG) {
6015 SDValue Op0 = And.getOperand(0);
6016 SDValue Op1 = And.getOperand(1);
6017 if (Op0.getOpcode() == ISD::TRUNCATE)
6018 Op0 = Op0.getOperand(0);
6019 if (Op1.getOpcode() == ISD::TRUNCATE)
6020 Op1 = Op1.getOperand(0);
6023 if (Op1.getOpcode() == ISD::SHL) {
6024 if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
6025 if (And10C->getZExtValue() == 1) {
6027 RHS = Op1.getOperand(1);
6029 } else if (Op0.getOpcode() == ISD::SHL) {
6030 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6031 if (And00C->getZExtValue() == 1) {
6033 RHS = Op0.getOperand(1);
6035 } else if (Op1.getOpcode() == ISD::Constant) {
6036 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6037 SDValue AndLHS = Op0;
6038 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6039 LHS = AndLHS.getOperand(0);
6040 RHS = AndLHS.getOperand(1);
6044 if (LHS.getNode()) {
6045 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
6046 // instruction. Since the shift amount is in-range-or-undefined, we know
6047 // that doing a bittest on the i16 value is ok. We extend to i32 because
6048 // the encoding for the i16 version is larger than the i32 version.
6049 if (LHS.getValueType() == MVT::i8)
6050 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6052 // If the operand types disagree, extend the shift amount to match. Since
6053 // BT ignores high bits (like shifts) we can use anyextend.
6054 if (LHS.getValueType() != RHS.getValueType())
6055 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6057 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6058 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6059 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6060 DAG.getConstant(Cond, MVT::i8), BT);
6066 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
6067 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6068 SDValue Op0 = Op.getOperand(0);
6069 SDValue Op1 = Op.getOperand(1);
6070 DebugLoc dl = Op.getDebugLoc();
6071 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6073 // Optimize to BT if possible.
6074 // Lower (X & (1 << N)) == 0 to BT(X, N).
6075 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6076 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6077 if (Op0.getOpcode() == ISD::AND &&
6079 Op1.getOpcode() == ISD::Constant &&
6080 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
6081 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6082 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6083 if (NewSetCC.getNode())
6087 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6088 if (Op0.getOpcode() == X86ISD::SETCC &&
6089 Op1.getOpcode() == ISD::Constant &&
6090 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6091 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6092 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6093 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6094 bool Invert = (CC == ISD::SETNE) ^
6095 cast<ConstantSDNode>(Op1)->isNullValue();
6097 CCode = X86::GetOppositeBranchCondition(CCode);
6098 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6099 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6102 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6103 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6104 if (X86CC == X86::COND_INVALID)
6107 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6109 // Use sbb x, x to materialize carry bit into a GPR.
6110 if (X86CC == X86::COND_B)
6111 return DAG.getNode(ISD::AND, dl, MVT::i8,
6112 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6113 DAG.getConstant(X86CC, MVT::i8), Cond),
6114 DAG.getConstant(1, MVT::i8));
6116 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6117 DAG.getConstant(X86CC, MVT::i8), Cond);
6120 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
6122 SDValue Op0 = Op.getOperand(0);
6123 SDValue Op1 = Op.getOperand(1);
6124 SDValue CC = Op.getOperand(2);
6125 EVT VT = Op.getValueType();
6126 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6127 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6128 DebugLoc dl = Op.getDebugLoc();
6132 EVT VT0 = Op0.getValueType();
6133 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6134 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6137 switch (SetCCOpcode) {
6140 case ISD::SETEQ: SSECC = 0; break;
6142 case ISD::SETGT: Swap = true; // Fallthrough
6144 case ISD::SETOLT: SSECC = 1; break;
6146 case ISD::SETGE: Swap = true; // Fallthrough
6148 case ISD::SETOLE: SSECC = 2; break;
6149 case ISD::SETUO: SSECC = 3; break;
6151 case ISD::SETNE: SSECC = 4; break;
6152 case ISD::SETULE: Swap = true;
6153 case ISD::SETUGE: SSECC = 5; break;
6154 case ISD::SETULT: Swap = true;
6155 case ISD::SETUGT: SSECC = 6; break;
6156 case ISD::SETO: SSECC = 7; break;
6159 std::swap(Op0, Op1);
6161 // In the two special cases we can't handle, emit two comparisons.
6163 if (SetCCOpcode == ISD::SETUEQ) {
6165 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6166 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6167 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6169 else if (SetCCOpcode == ISD::SETONE) {
6171 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6172 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6173 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6175 llvm_unreachable("Illegal FP comparison");
6177 // Handle all other FP comparisons here.
6178 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6181 // We are handling one of the integer comparisons here. Since SSE only has
6182 // GT and EQ comparisons for integer, swapping operands and multiple
6183 // operations may be required for some comparisons.
6184 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6185 bool Swap = false, Invert = false, FlipSigns = false;
6187 switch (VT.getSimpleVT().SimpleTy) {
6190 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6192 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6194 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6195 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6198 switch (SetCCOpcode) {
6200 case ISD::SETNE: Invert = true;
6201 case ISD::SETEQ: Opc = EQOpc; break;
6202 case ISD::SETLT: Swap = true;
6203 case ISD::SETGT: Opc = GTOpc; break;
6204 case ISD::SETGE: Swap = true;
6205 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6206 case ISD::SETULT: Swap = true;
6207 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6208 case ISD::SETUGE: Swap = true;
6209 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6212 std::swap(Op0, Op1);
6214 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6215 // bits of the inputs before performing those operations.
6217 EVT EltVT = VT.getVectorElementType();
6218 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6220 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6221 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6223 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6224 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6227 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6229 // If the logical-not of the result is required, perform that now.
6231 Result = DAG.getNOT(dl, Result, VT);
6236 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6237 static bool isX86LogicalCmp(SDValue Op) {
6238 unsigned Opc = Op.getNode()->getOpcode();
6239 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6241 if (Op.getResNo() == 1 &&
6242 (Opc == X86ISD::ADD ||
6243 Opc == X86ISD::SUB ||
6244 Opc == X86ISD::SMUL ||
6245 Opc == X86ISD::UMUL ||
6246 Opc == X86ISD::INC ||
6247 Opc == X86ISD::DEC ||
6248 Opc == X86ISD::OR ||
6249 Opc == X86ISD::XOR ||
6250 Opc == X86ISD::AND))
6256 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6257 bool addTest = true;
6258 SDValue Cond = Op.getOperand(0);
6259 DebugLoc dl = Op.getDebugLoc();
6262 if (Cond.getOpcode() == ISD::SETCC) {
6263 SDValue NewCond = LowerSETCC(Cond, DAG);
6264 if (NewCond.getNode())
6268 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6269 SDValue Op1 = Op.getOperand(1);
6270 SDValue Op2 = Op.getOperand(2);
6271 if (Cond.getOpcode() == X86ISD::SETCC &&
6272 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6273 SDValue Cmp = Cond.getOperand(1);
6274 if (Cmp.getOpcode() == X86ISD::CMP) {
6275 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6276 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6277 ConstantSDNode *RHSC =
6278 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6279 if (N1C && N1C->isAllOnesValue() &&
6280 N2C && N2C->isNullValue() &&
6281 RHSC && RHSC->isNullValue()) {
6282 SDValue CmpOp0 = Cmp.getOperand(0);
6283 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6284 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6285 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6286 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6291 // Look pass (and (setcc_carry (cmp ...)), 1).
6292 if (Cond.getOpcode() == ISD::AND &&
6293 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6294 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6295 if (C && C->getAPIntValue() == 1)
6296 Cond = Cond.getOperand(0);
6299 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6300 // setting operand in place of the X86ISD::SETCC.
6301 if (Cond.getOpcode() == X86ISD::SETCC ||
6302 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6303 CC = Cond.getOperand(0);
6305 SDValue Cmp = Cond.getOperand(1);
6306 unsigned Opc = Cmp.getOpcode();
6307 EVT VT = Op.getValueType();
6309 bool IllegalFPCMov = false;
6310 if (VT.isFloatingPoint() && !VT.isVector() &&
6311 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6312 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6314 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6315 Opc == X86ISD::BT) { // FIXME
6322 // Look pass the truncate.
6323 if (Cond.getOpcode() == ISD::TRUNCATE)
6324 Cond = Cond.getOperand(0);
6326 // We know the result of AND is compared against zero. Try to match
6328 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6329 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6330 if (NewSetCC.getNode()) {
6331 CC = NewSetCC.getOperand(0);
6332 Cond = NewSetCC.getOperand(1);
6339 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6340 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6343 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6344 // condition is true.
6345 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6346 SDValue Ops[] = { Op2, Op1, CC, Cond };
6347 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6350 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6351 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6352 // from the AND / OR.
6353 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6354 Opc = Op.getOpcode();
6355 if (Opc != ISD::OR && Opc != ISD::AND)
6357 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6358 Op.getOperand(0).hasOneUse() &&
6359 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6360 Op.getOperand(1).hasOneUse());
6363 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6364 // 1 and that the SETCC node has a single use.
6365 static bool isXor1OfSetCC(SDValue Op) {
6366 if (Op.getOpcode() != ISD::XOR)
6368 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6369 if (N1C && N1C->getAPIntValue() == 1) {
6370 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6371 Op.getOperand(0).hasOneUse();
6376 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6377 bool addTest = true;
6378 SDValue Chain = Op.getOperand(0);
6379 SDValue Cond = Op.getOperand(1);
6380 SDValue Dest = Op.getOperand(2);
6381 DebugLoc dl = Op.getDebugLoc();
6384 if (Cond.getOpcode() == ISD::SETCC) {
6385 SDValue NewCond = LowerSETCC(Cond, DAG);
6386 if (NewCond.getNode())
6390 // FIXME: LowerXALUO doesn't handle these!!
6391 else if (Cond.getOpcode() == X86ISD::ADD ||
6392 Cond.getOpcode() == X86ISD::SUB ||
6393 Cond.getOpcode() == X86ISD::SMUL ||
6394 Cond.getOpcode() == X86ISD::UMUL)
6395 Cond = LowerXALUO(Cond, DAG);
6398 // Look pass (and (setcc_carry (cmp ...)), 1).
6399 if (Cond.getOpcode() == ISD::AND &&
6400 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6401 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6402 if (C && C->getAPIntValue() == 1)
6403 Cond = Cond.getOperand(0);
6406 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6407 // setting operand in place of the X86ISD::SETCC.
6408 if (Cond.getOpcode() == X86ISD::SETCC ||
6409 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6410 CC = Cond.getOperand(0);
6412 SDValue Cmp = Cond.getOperand(1);
6413 unsigned Opc = Cmp.getOpcode();
6414 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6415 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6419 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6423 // These can only come from an arithmetic instruction with overflow,
6424 // e.g. SADDO, UADDO.
6425 Cond = Cond.getNode()->getOperand(1);
6432 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6433 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6434 if (CondOpc == ISD::OR) {
6435 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6436 // two branches instead of an explicit OR instruction with a
6438 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6439 isX86LogicalCmp(Cmp)) {
6440 CC = Cond.getOperand(0).getOperand(0);
6441 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6442 Chain, Dest, CC, Cmp);
6443 CC = Cond.getOperand(1).getOperand(0);
6447 } else { // ISD::AND
6448 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6449 // two branches instead of an explicit AND instruction with a
6450 // separate test. However, we only do this if this block doesn't
6451 // have a fall-through edge, because this requires an explicit
6452 // jmp when the condition is false.
6453 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6454 isX86LogicalCmp(Cmp) &&
6455 Op.getNode()->hasOneUse()) {
6456 X86::CondCode CCode =
6457 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6458 CCode = X86::GetOppositeBranchCondition(CCode);
6459 CC = DAG.getConstant(CCode, MVT::i8);
6460 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6461 // Look for an unconditional branch following this conditional branch.
6462 // We need this because we need to reverse the successors in order
6463 // to implement FCMP_OEQ.
6464 if (User.getOpcode() == ISD::BR) {
6465 SDValue FalseBB = User.getOperand(1);
6467 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6468 assert(NewBR == User);
6471 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6472 Chain, Dest, CC, Cmp);
6473 X86::CondCode CCode =
6474 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6475 CCode = X86::GetOppositeBranchCondition(CCode);
6476 CC = DAG.getConstant(CCode, MVT::i8);
6482 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6483 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6484 // It should be transformed during dag combiner except when the condition
6485 // is set by a arithmetics with overflow node.
6486 X86::CondCode CCode =
6487 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6488 CCode = X86::GetOppositeBranchCondition(CCode);
6489 CC = DAG.getConstant(CCode, MVT::i8);
6490 Cond = Cond.getOperand(0).getOperand(1);
6496 // Look pass the truncate.
6497 if (Cond.getOpcode() == ISD::TRUNCATE)
6498 Cond = Cond.getOperand(0);
6500 // We know the result of AND is compared against zero. Try to match
6502 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6503 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6504 if (NewSetCC.getNode()) {
6505 CC = NewSetCC.getOperand(0);
6506 Cond = NewSetCC.getOperand(1);
6513 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6514 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6516 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6517 Chain, Dest, CC, Cond);
6521 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6522 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6523 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6524 // that the guard pages used by the OS virtual memory manager are allocated in
6525 // correct sequence.
6527 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6528 SelectionDAG &DAG) {
6529 assert(Subtarget->isTargetCygMing() &&
6530 "This should be used only on Cygwin/Mingw targets");
6531 DebugLoc dl = Op.getDebugLoc();
6534 SDValue Chain = Op.getOperand(0);
6535 SDValue Size = Op.getOperand(1);
6536 // FIXME: Ensure alignment here
6540 EVT IntPtr = getPointerTy();
6541 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6543 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6544 Flag = Chain.getValue(1);
6546 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6548 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6549 Flag = Chain.getValue(1);
6551 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6553 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6554 return DAG.getMergeValues(Ops1, 2, dl);
6558 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6560 SDValue Dst, SDValue Src,
6561 SDValue Size, unsigned Align,
6564 uint64_t DstSVOff) {
6565 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6567 // If not DWORD aligned or size is more than the threshold, call the library.
6568 // The libc version is likely to be faster for these cases. It can use the
6569 // address value and run time information about the CPU.
6570 if ((Align & 3) != 0 ||
6572 ConstantSize->getZExtValue() >
6573 getSubtarget()->getMaxInlineSizeThreshold()) {
6574 SDValue InFlag(0, 0);
6576 // Check to see if there is a specialized entry-point for memory zeroing.
6577 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6579 if (const char *bzeroEntry = V &&
6580 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6581 EVT IntPtr = getPointerTy();
6582 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6583 TargetLowering::ArgListTy Args;
6584 TargetLowering::ArgListEntry Entry;
6586 Entry.Ty = IntPtrTy;
6587 Args.push_back(Entry);
6589 Args.push_back(Entry);
6590 std::pair<SDValue,SDValue> CallResult =
6591 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6592 false, false, false, false,
6593 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6594 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
6595 return CallResult.second;
6598 // Otherwise have the target-independent code call memset.
6602 uint64_t SizeVal = ConstantSize->getZExtValue();
6603 SDValue InFlag(0, 0);
6606 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6607 unsigned BytesLeft = 0;
6608 bool TwoRepStos = false;
6611 uint64_t Val = ValC->getZExtValue() & 255;
6613 // If the value is a constant, then we can potentially use larger sets.
6614 switch (Align & 3) {
6615 case 2: // WORD aligned
6618 Val = (Val << 8) | Val;
6620 case 0: // DWORD aligned
6623 Val = (Val << 8) | Val;
6624 Val = (Val << 16) | Val;
6625 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6628 Val = (Val << 32) | Val;
6631 default: // Byte aligned
6634 Count = DAG.getIntPtrConstant(SizeVal);
6638 if (AVT.bitsGT(MVT::i8)) {
6639 unsigned UBytes = AVT.getSizeInBits() / 8;
6640 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6641 BytesLeft = SizeVal % UBytes;
6644 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6646 InFlag = Chain.getValue(1);
6649 Count = DAG.getIntPtrConstant(SizeVal);
6650 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6651 InFlag = Chain.getValue(1);
6654 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6657 InFlag = Chain.getValue(1);
6658 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6661 InFlag = Chain.getValue(1);
6663 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6664 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6665 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6668 InFlag = Chain.getValue(1);
6670 EVT CVT = Count.getValueType();
6671 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6672 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6673 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6676 InFlag = Chain.getValue(1);
6677 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6678 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6679 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6680 } else if (BytesLeft) {
6681 // Handle the last 1 - 7 bytes.
6682 unsigned Offset = SizeVal - BytesLeft;
6683 EVT AddrVT = Dst.getValueType();
6684 EVT SizeVT = Size.getValueType();
6686 Chain = DAG.getMemset(Chain, dl,
6687 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6688 DAG.getConstant(Offset, AddrVT)),
6690 DAG.getConstant(BytesLeft, SizeVT),
6691 Align, isVolatile, DstSV, DstSVOff + Offset);
6694 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6699 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6700 SDValue Chain, SDValue Dst, SDValue Src,
6701 SDValue Size, unsigned Align,
6702 bool isVolatile, bool AlwaysInline,
6703 const Value *DstSV, uint64_t DstSVOff,
6704 const Value *SrcSV, uint64_t SrcSVOff) {
6705 // This requires the copy size to be a constant, preferrably
6706 // within a subtarget-specific limit.
6707 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6710 uint64_t SizeVal = ConstantSize->getZExtValue();
6711 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6714 /// If not DWORD aligned, call the library.
6715 if ((Align & 3) != 0)
6720 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6723 unsigned UBytes = AVT.getSizeInBits() / 8;
6724 unsigned CountVal = SizeVal / UBytes;
6725 SDValue Count = DAG.getIntPtrConstant(CountVal);
6726 unsigned BytesLeft = SizeVal % UBytes;
6728 SDValue InFlag(0, 0);
6729 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6732 InFlag = Chain.getValue(1);
6733 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6736 InFlag = Chain.getValue(1);
6737 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6740 InFlag = Chain.getValue(1);
6742 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6743 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6744 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6745 array_lengthof(Ops));
6747 SmallVector<SDValue, 4> Results;
6748 Results.push_back(RepMovs);
6750 // Handle the last 1 - 7 bytes.
6751 unsigned Offset = SizeVal - BytesLeft;
6752 EVT DstVT = Dst.getValueType();
6753 EVT SrcVT = Src.getValueType();
6754 EVT SizeVT = Size.getValueType();
6755 Results.push_back(DAG.getMemcpy(Chain, dl,
6756 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6757 DAG.getConstant(Offset, DstVT)),
6758 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6759 DAG.getConstant(Offset, SrcVT)),
6760 DAG.getConstant(BytesLeft, SizeVT),
6761 Align, isVolatile, AlwaysInline,
6762 DstSV, DstSVOff + Offset,
6763 SrcSV, SrcSVOff + Offset));
6766 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6767 &Results[0], Results.size());
6770 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6771 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6772 DebugLoc dl = Op.getDebugLoc();
6774 if (!Subtarget->is64Bit()) {
6775 // vastart just stores the address of the VarArgsFrameIndex slot into the
6776 // memory location argument.
6777 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6778 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6783 // gp_offset (0 - 6 * 8)
6784 // fp_offset (48 - 48 + 8 * 16)
6785 // overflow_arg_area (point to parameters coming in memory).
6787 SmallVector<SDValue, 8> MemOps;
6788 SDValue FIN = Op.getOperand(1);
6790 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6791 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6792 FIN, SV, 0, false, false, 0);
6793 MemOps.push_back(Store);
6796 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6797 FIN, DAG.getIntPtrConstant(4));
6798 Store = DAG.getStore(Op.getOperand(0), dl,
6799 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6800 FIN, SV, 0, false, false, 0);
6801 MemOps.push_back(Store);
6803 // Store ptr to overflow_arg_area
6804 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6805 FIN, DAG.getIntPtrConstant(4));
6806 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6807 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6809 MemOps.push_back(Store);
6811 // Store ptr to reg_save_area.
6812 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6813 FIN, DAG.getIntPtrConstant(8));
6814 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6815 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6817 MemOps.push_back(Store);
6818 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6819 &MemOps[0], MemOps.size());
6822 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6823 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6824 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6825 SDValue Chain = Op.getOperand(0);
6826 SDValue SrcPtr = Op.getOperand(1);
6827 SDValue SrcSV = Op.getOperand(2);
6829 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6833 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6834 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6835 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6836 SDValue Chain = Op.getOperand(0);
6837 SDValue DstPtr = Op.getOperand(1);
6838 SDValue SrcPtr = Op.getOperand(2);
6839 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6840 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6841 DebugLoc dl = Op.getDebugLoc();
6843 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6844 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6845 false, DstSV, 0, SrcSV, 0);
6849 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6850 DebugLoc dl = Op.getDebugLoc();
6851 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6853 default: return SDValue(); // Don't custom lower most intrinsics.
6854 // Comparison intrinsics.
6855 case Intrinsic::x86_sse_comieq_ss:
6856 case Intrinsic::x86_sse_comilt_ss:
6857 case Intrinsic::x86_sse_comile_ss:
6858 case Intrinsic::x86_sse_comigt_ss:
6859 case Intrinsic::x86_sse_comige_ss:
6860 case Intrinsic::x86_sse_comineq_ss:
6861 case Intrinsic::x86_sse_ucomieq_ss:
6862 case Intrinsic::x86_sse_ucomilt_ss:
6863 case Intrinsic::x86_sse_ucomile_ss:
6864 case Intrinsic::x86_sse_ucomigt_ss:
6865 case Intrinsic::x86_sse_ucomige_ss:
6866 case Intrinsic::x86_sse_ucomineq_ss:
6867 case Intrinsic::x86_sse2_comieq_sd:
6868 case Intrinsic::x86_sse2_comilt_sd:
6869 case Intrinsic::x86_sse2_comile_sd:
6870 case Intrinsic::x86_sse2_comigt_sd:
6871 case Intrinsic::x86_sse2_comige_sd:
6872 case Intrinsic::x86_sse2_comineq_sd:
6873 case Intrinsic::x86_sse2_ucomieq_sd:
6874 case Intrinsic::x86_sse2_ucomilt_sd:
6875 case Intrinsic::x86_sse2_ucomile_sd:
6876 case Intrinsic::x86_sse2_ucomigt_sd:
6877 case Intrinsic::x86_sse2_ucomige_sd:
6878 case Intrinsic::x86_sse2_ucomineq_sd: {
6880 ISD::CondCode CC = ISD::SETCC_INVALID;
6883 case Intrinsic::x86_sse_comieq_ss:
6884 case Intrinsic::x86_sse2_comieq_sd:
6888 case Intrinsic::x86_sse_comilt_ss:
6889 case Intrinsic::x86_sse2_comilt_sd:
6893 case Intrinsic::x86_sse_comile_ss:
6894 case Intrinsic::x86_sse2_comile_sd:
6898 case Intrinsic::x86_sse_comigt_ss:
6899 case Intrinsic::x86_sse2_comigt_sd:
6903 case Intrinsic::x86_sse_comige_ss:
6904 case Intrinsic::x86_sse2_comige_sd:
6908 case Intrinsic::x86_sse_comineq_ss:
6909 case Intrinsic::x86_sse2_comineq_sd:
6913 case Intrinsic::x86_sse_ucomieq_ss:
6914 case Intrinsic::x86_sse2_ucomieq_sd:
6915 Opc = X86ISD::UCOMI;
6918 case Intrinsic::x86_sse_ucomilt_ss:
6919 case Intrinsic::x86_sse2_ucomilt_sd:
6920 Opc = X86ISD::UCOMI;
6923 case Intrinsic::x86_sse_ucomile_ss:
6924 case Intrinsic::x86_sse2_ucomile_sd:
6925 Opc = X86ISD::UCOMI;
6928 case Intrinsic::x86_sse_ucomigt_ss:
6929 case Intrinsic::x86_sse2_ucomigt_sd:
6930 Opc = X86ISD::UCOMI;
6933 case Intrinsic::x86_sse_ucomige_ss:
6934 case Intrinsic::x86_sse2_ucomige_sd:
6935 Opc = X86ISD::UCOMI;
6938 case Intrinsic::x86_sse_ucomineq_ss:
6939 case Intrinsic::x86_sse2_ucomineq_sd:
6940 Opc = X86ISD::UCOMI;
6945 SDValue LHS = Op.getOperand(1);
6946 SDValue RHS = Op.getOperand(2);
6947 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6948 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6949 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6950 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6951 DAG.getConstant(X86CC, MVT::i8), Cond);
6952 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6954 // ptest intrinsics. The intrinsic these come from are designed to return
6955 // an integer value, not just an instruction so lower it to the ptest
6956 // pattern and a setcc for the result.
6957 case Intrinsic::x86_sse41_ptestz:
6958 case Intrinsic::x86_sse41_ptestc:
6959 case Intrinsic::x86_sse41_ptestnzc:{
6962 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6963 case Intrinsic::x86_sse41_ptestz:
6965 X86CC = X86::COND_E;
6967 case Intrinsic::x86_sse41_ptestc:
6969 X86CC = X86::COND_B;
6971 case Intrinsic::x86_sse41_ptestnzc:
6973 X86CC = X86::COND_A;
6977 SDValue LHS = Op.getOperand(1);
6978 SDValue RHS = Op.getOperand(2);
6979 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6980 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6981 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6982 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6985 // Fix vector shift instructions where the last operand is a non-immediate
6987 case Intrinsic::x86_sse2_pslli_w:
6988 case Intrinsic::x86_sse2_pslli_d:
6989 case Intrinsic::x86_sse2_pslli_q:
6990 case Intrinsic::x86_sse2_psrli_w:
6991 case Intrinsic::x86_sse2_psrli_d:
6992 case Intrinsic::x86_sse2_psrli_q:
6993 case Intrinsic::x86_sse2_psrai_w:
6994 case Intrinsic::x86_sse2_psrai_d:
6995 case Intrinsic::x86_mmx_pslli_w:
6996 case Intrinsic::x86_mmx_pslli_d:
6997 case Intrinsic::x86_mmx_pslli_q:
6998 case Intrinsic::x86_mmx_psrli_w:
6999 case Intrinsic::x86_mmx_psrli_d:
7000 case Intrinsic::x86_mmx_psrli_q:
7001 case Intrinsic::x86_mmx_psrai_w:
7002 case Intrinsic::x86_mmx_psrai_d: {
7003 SDValue ShAmt = Op.getOperand(2);
7004 if (isa<ConstantSDNode>(ShAmt))
7007 unsigned NewIntNo = 0;
7008 EVT ShAmtVT = MVT::v4i32;
7010 case Intrinsic::x86_sse2_pslli_w:
7011 NewIntNo = Intrinsic::x86_sse2_psll_w;
7013 case Intrinsic::x86_sse2_pslli_d:
7014 NewIntNo = Intrinsic::x86_sse2_psll_d;
7016 case Intrinsic::x86_sse2_pslli_q:
7017 NewIntNo = Intrinsic::x86_sse2_psll_q;
7019 case Intrinsic::x86_sse2_psrli_w:
7020 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7022 case Intrinsic::x86_sse2_psrli_d:
7023 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7025 case Intrinsic::x86_sse2_psrli_q:
7026 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7028 case Intrinsic::x86_sse2_psrai_w:
7029 NewIntNo = Intrinsic::x86_sse2_psra_w;
7031 case Intrinsic::x86_sse2_psrai_d:
7032 NewIntNo = Intrinsic::x86_sse2_psra_d;
7035 ShAmtVT = MVT::v2i32;
7037 case Intrinsic::x86_mmx_pslli_w:
7038 NewIntNo = Intrinsic::x86_mmx_psll_w;
7040 case Intrinsic::x86_mmx_pslli_d:
7041 NewIntNo = Intrinsic::x86_mmx_psll_d;
7043 case Intrinsic::x86_mmx_pslli_q:
7044 NewIntNo = Intrinsic::x86_mmx_psll_q;
7046 case Intrinsic::x86_mmx_psrli_w:
7047 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7049 case Intrinsic::x86_mmx_psrli_d:
7050 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7052 case Intrinsic::x86_mmx_psrli_q:
7053 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7055 case Intrinsic::x86_mmx_psrai_w:
7056 NewIntNo = Intrinsic::x86_mmx_psra_w;
7058 case Intrinsic::x86_mmx_psrai_d:
7059 NewIntNo = Intrinsic::x86_mmx_psra_d;
7061 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7067 // The vector shift intrinsics with scalars uses 32b shift amounts but
7068 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7072 ShOps[1] = DAG.getConstant(0, MVT::i32);
7073 if (ShAmtVT == MVT::v4i32) {
7074 ShOps[2] = DAG.getUNDEF(MVT::i32);
7075 ShOps[3] = DAG.getUNDEF(MVT::i32);
7076 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7078 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7081 EVT VT = Op.getValueType();
7082 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7083 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7084 DAG.getConstant(NewIntNo, MVT::i32),
7085 Op.getOperand(1), ShAmt);
7090 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
7091 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7092 DebugLoc dl = Op.getDebugLoc();
7095 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7097 DAG.getConstant(TD->getPointerSize(),
7098 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7099 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7100 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7102 NULL, 0, false, false, 0);
7105 // Just load the return address.
7106 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7107 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7108 RetAddrFI, NULL, 0, false, false, 0);
7111 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
7112 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7113 MFI->setFrameAddressIsTaken(true);
7114 EVT VT = Op.getValueType();
7115 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7116 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7117 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7118 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7120 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7125 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7126 SelectionDAG &DAG) {
7127 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7130 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
7132 MachineFunction &MF = DAG.getMachineFunction();
7133 SDValue Chain = Op.getOperand(0);
7134 SDValue Offset = Op.getOperand(1);
7135 SDValue Handler = Op.getOperand(2);
7136 DebugLoc dl = Op.getDebugLoc();
7138 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7140 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7142 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7143 DAG.getIntPtrConstant(-TD->getPointerSize()));
7144 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7145 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7146 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7147 MF.getRegInfo().addLiveOut(StoreAddrReg);
7149 return DAG.getNode(X86ISD::EH_RETURN, dl,
7151 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7154 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7155 SelectionDAG &DAG) {
7156 SDValue Root = Op.getOperand(0);
7157 SDValue Trmp = Op.getOperand(1); // trampoline
7158 SDValue FPtr = Op.getOperand(2); // nested function
7159 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7160 DebugLoc dl = Op.getDebugLoc();
7162 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7164 if (Subtarget->is64Bit()) {
7165 SDValue OutChains[6];
7167 // Large code-model.
7168 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7169 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7171 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7172 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7174 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7176 // Load the pointer to the nested function into R11.
7177 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7178 SDValue Addr = Trmp;
7179 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7180 Addr, TrmpAddr, 0, false, false, 0);
7182 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7183 DAG.getConstant(2, MVT::i64));
7184 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7187 // Load the 'nest' parameter value into R10.
7188 // R10 is specified in X86CallingConv.td
7189 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7190 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7191 DAG.getConstant(10, MVT::i64));
7192 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7193 Addr, TrmpAddr, 10, false, false, 0);
7195 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7196 DAG.getConstant(12, MVT::i64));
7197 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7200 // Jump to the nested function.
7201 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7202 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7203 DAG.getConstant(20, MVT::i64));
7204 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7205 Addr, TrmpAddr, 20, false, false, 0);
7207 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7208 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7209 DAG.getConstant(22, MVT::i64));
7210 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7211 TrmpAddr, 22, false, false, 0);
7214 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7215 return DAG.getMergeValues(Ops, 2, dl);
7217 const Function *Func =
7218 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7219 CallingConv::ID CC = Func->getCallingConv();
7224 llvm_unreachable("Unsupported calling convention");
7225 case CallingConv::C:
7226 case CallingConv::X86_StdCall: {
7227 // Pass 'nest' parameter in ECX.
7228 // Must be kept in sync with X86CallingConv.td
7231 // Check that ECX wasn't needed by an 'inreg' parameter.
7232 const FunctionType *FTy = Func->getFunctionType();
7233 const AttrListPtr &Attrs = Func->getAttributes();
7235 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7236 unsigned InRegCount = 0;
7239 for (FunctionType::param_iterator I = FTy->param_begin(),
7240 E = FTy->param_end(); I != E; ++I, ++Idx)
7241 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7242 // FIXME: should only count parameters that are lowered to integers.
7243 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7245 if (InRegCount > 2) {
7246 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7251 case CallingConv::X86_FastCall:
7252 case CallingConv::Fast:
7253 // Pass 'nest' parameter in EAX.
7254 // Must be kept in sync with X86CallingConv.td
7259 SDValue OutChains[4];
7262 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7263 DAG.getConstant(10, MVT::i32));
7264 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7266 // This is storing the opcode for MOV32ri.
7267 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7268 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7269 OutChains[0] = DAG.getStore(Root, dl,
7270 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7271 Trmp, TrmpAddr, 0, false, false, 0);
7273 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7274 DAG.getConstant(1, MVT::i32));
7275 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7278 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7279 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7280 DAG.getConstant(5, MVT::i32));
7281 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7282 TrmpAddr, 5, false, false, 1);
7284 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7285 DAG.getConstant(6, MVT::i32));
7286 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7290 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7291 return DAG.getMergeValues(Ops, 2, dl);
7295 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7297 The rounding mode is in bits 11:10 of FPSR, and has the following
7304 FLT_ROUNDS, on the other hand, expects the following:
7311 To perform the conversion, we do:
7312 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7315 MachineFunction &MF = DAG.getMachineFunction();
7316 const TargetMachine &TM = MF.getTarget();
7317 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7318 unsigned StackAlignment = TFI.getStackAlignment();
7319 EVT VT = Op.getValueType();
7320 DebugLoc dl = Op.getDebugLoc();
7322 // Save FP Control Word to stack slot
7323 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7324 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7326 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7327 DAG.getEntryNode(), StackSlot);
7329 // Load FP Control Word from stack slot
7330 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7333 // Transform as necessary
7335 DAG.getNode(ISD::SRL, dl, MVT::i16,
7336 DAG.getNode(ISD::AND, dl, MVT::i16,
7337 CWD, DAG.getConstant(0x800, MVT::i16)),
7338 DAG.getConstant(11, MVT::i8));
7340 DAG.getNode(ISD::SRL, dl, MVT::i16,
7341 DAG.getNode(ISD::AND, dl, MVT::i16,
7342 CWD, DAG.getConstant(0x400, MVT::i16)),
7343 DAG.getConstant(9, MVT::i8));
7346 DAG.getNode(ISD::AND, dl, MVT::i16,
7347 DAG.getNode(ISD::ADD, dl, MVT::i16,
7348 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7349 DAG.getConstant(1, MVT::i16)),
7350 DAG.getConstant(3, MVT::i16));
7353 return DAG.getNode((VT.getSizeInBits() < 16 ?
7354 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7357 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7358 EVT VT = Op.getValueType();
7360 unsigned NumBits = VT.getSizeInBits();
7361 DebugLoc dl = Op.getDebugLoc();
7363 Op = Op.getOperand(0);
7364 if (VT == MVT::i8) {
7365 // Zero extend to i32 since there is not an i8 bsr.
7367 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7370 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7371 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7372 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7374 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7377 DAG.getConstant(NumBits+NumBits-1, OpVT),
7378 DAG.getConstant(X86::COND_E, MVT::i8),
7381 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7383 // Finally xor with NumBits-1.
7384 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7387 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7391 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7392 EVT VT = Op.getValueType();
7394 unsigned NumBits = VT.getSizeInBits();
7395 DebugLoc dl = Op.getDebugLoc();
7397 Op = Op.getOperand(0);
7398 if (VT == MVT::i8) {
7400 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7403 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7404 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7405 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7407 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7410 DAG.getConstant(NumBits, OpVT),
7411 DAG.getConstant(X86::COND_E, MVT::i8),
7414 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7417 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7421 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7422 EVT VT = Op.getValueType();
7423 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7424 DebugLoc dl = Op.getDebugLoc();
7426 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7427 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7428 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7429 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7430 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7432 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7433 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7434 // return AloBlo + AloBhi + AhiBlo;
7436 SDValue A = Op.getOperand(0);
7437 SDValue B = Op.getOperand(1);
7439 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7440 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7441 A, DAG.getConstant(32, MVT::i32));
7442 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7443 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7444 B, DAG.getConstant(32, MVT::i32));
7445 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7446 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7448 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7449 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7451 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7452 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7454 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7455 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7456 AloBhi, DAG.getConstant(32, MVT::i32));
7457 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7458 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7459 AhiBlo, DAG.getConstant(32, MVT::i32));
7460 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7461 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7466 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7467 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7468 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7469 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7470 // has only one use.
7471 SDNode *N = Op.getNode();
7472 SDValue LHS = N->getOperand(0);
7473 SDValue RHS = N->getOperand(1);
7474 unsigned BaseOp = 0;
7476 DebugLoc dl = Op.getDebugLoc();
7478 switch (Op.getOpcode()) {
7479 default: llvm_unreachable("Unknown ovf instruction!");
7481 // A subtract of one will be selected as a INC. Note that INC doesn't
7482 // set CF, so we can't do this for UADDO.
7483 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7484 if (C->getAPIntValue() == 1) {
7485 BaseOp = X86ISD::INC;
7489 BaseOp = X86ISD::ADD;
7493 BaseOp = X86ISD::ADD;
7497 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7498 // set CF, so we can't do this for USUBO.
7499 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7500 if (C->getAPIntValue() == 1) {
7501 BaseOp = X86ISD::DEC;
7505 BaseOp = X86ISD::SUB;
7509 BaseOp = X86ISD::SUB;
7513 BaseOp = X86ISD::SMUL;
7517 BaseOp = X86ISD::UMUL;
7522 // Also sets EFLAGS.
7523 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7524 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7527 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7528 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7530 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7534 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7535 EVT T = Op.getValueType();
7536 DebugLoc dl = Op.getDebugLoc();
7539 switch(T.getSimpleVT().SimpleTy) {
7541 assert(false && "Invalid value type!");
7542 case MVT::i8: Reg = X86::AL; size = 1; break;
7543 case MVT::i16: Reg = X86::AX; size = 2; break;
7544 case MVT::i32: Reg = X86::EAX; size = 4; break;
7546 assert(Subtarget->is64Bit() && "Node not type legal!");
7547 Reg = X86::RAX; size = 8;
7550 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7551 Op.getOperand(2), SDValue());
7552 SDValue Ops[] = { cpIn.getValue(0),
7555 DAG.getTargetConstant(size, MVT::i8),
7557 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7558 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7560 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7564 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7565 SelectionDAG &DAG) {
7566 assert(Subtarget->is64Bit() && "Result not type legalized?");
7567 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7568 SDValue TheChain = Op.getOperand(0);
7569 DebugLoc dl = Op.getDebugLoc();
7570 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7571 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7572 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7574 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7575 DAG.getConstant(32, MVT::i8));
7577 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7580 return DAG.getMergeValues(Ops, 2, dl);
7583 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7584 SDNode *Node = Op.getNode();
7585 DebugLoc dl = Node->getDebugLoc();
7586 EVT T = Node->getValueType(0);
7587 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7588 DAG.getConstant(0, T), Node->getOperand(2));
7589 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7590 cast<AtomicSDNode>(Node)->getMemoryVT(),
7591 Node->getOperand(0),
7592 Node->getOperand(1), negOp,
7593 cast<AtomicSDNode>(Node)->getSrcValue(),
7594 cast<AtomicSDNode>(Node)->getAlignment());
7597 /// LowerOperation - Provide custom lowering hooks for some operations.
7599 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7600 switch (Op.getOpcode()) {
7601 default: llvm_unreachable("Should not custom lower this!");
7602 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7603 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7604 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7605 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7606 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7607 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7608 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7609 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7610 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7611 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7612 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7613 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7614 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7615 case ISD::SHL_PARTS:
7616 case ISD::SRA_PARTS:
7617 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7618 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7619 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7620 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7621 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7622 case ISD::FABS: return LowerFABS(Op, DAG);
7623 case ISD::FNEG: return LowerFNEG(Op, DAG);
7624 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7625 case ISD::SETCC: return LowerSETCC(Op, DAG);
7626 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7627 case ISD::SELECT: return LowerSELECT(Op, DAG);
7628 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7629 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7630 case ISD::VASTART: return LowerVASTART(Op, DAG);
7631 case ISD::VAARG: return LowerVAARG(Op, DAG);
7632 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7633 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7634 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7635 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7636 case ISD::FRAME_TO_ARGS_OFFSET:
7637 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7638 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7639 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7640 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7641 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7642 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7643 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7644 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7650 case ISD::UMULO: return LowerXALUO(Op, DAG);
7651 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7655 void X86TargetLowering::
7656 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7657 SelectionDAG &DAG, unsigned NewOp) {
7658 EVT T = Node->getValueType(0);
7659 DebugLoc dl = Node->getDebugLoc();
7660 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7662 SDValue Chain = Node->getOperand(0);
7663 SDValue In1 = Node->getOperand(1);
7664 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7665 Node->getOperand(2), DAG.getIntPtrConstant(0));
7666 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7667 Node->getOperand(2), DAG.getIntPtrConstant(1));
7668 SDValue Ops[] = { Chain, In1, In2L, In2H };
7669 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7671 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7672 cast<MemSDNode>(Node)->getMemOperand());
7673 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7674 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7675 Results.push_back(Result.getValue(2));
7678 /// ReplaceNodeResults - Replace a node with an illegal result type
7679 /// with a new node built out of custom code.
7680 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7681 SmallVectorImpl<SDValue>&Results,
7682 SelectionDAG &DAG) {
7683 DebugLoc dl = N->getDebugLoc();
7684 switch (N->getOpcode()) {
7686 assert(false && "Do not know how to custom type legalize this operation!");
7688 case ISD::FP_TO_SINT: {
7689 std::pair<SDValue,SDValue> Vals =
7690 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7691 SDValue FIST = Vals.first, StackSlot = Vals.second;
7692 if (FIST.getNode() != 0) {
7693 EVT VT = N->getValueType(0);
7694 // Return a load from the stack slot.
7695 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7700 case ISD::READCYCLECOUNTER: {
7701 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7702 SDValue TheChain = N->getOperand(0);
7703 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7704 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7706 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7708 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7709 SDValue Ops[] = { eax, edx };
7710 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7711 Results.push_back(edx.getValue(1));
7714 case ISD::ATOMIC_CMP_SWAP: {
7715 EVT T = N->getValueType(0);
7716 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7717 SDValue cpInL, cpInH;
7718 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7719 DAG.getConstant(0, MVT::i32));
7720 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7721 DAG.getConstant(1, MVT::i32));
7722 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7723 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7725 SDValue swapInL, swapInH;
7726 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7727 DAG.getConstant(0, MVT::i32));
7728 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7729 DAG.getConstant(1, MVT::i32));
7730 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7732 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7733 swapInL.getValue(1));
7734 SDValue Ops[] = { swapInH.getValue(0),
7736 swapInH.getValue(1) };
7737 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7738 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7739 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7740 MVT::i32, Result.getValue(1));
7741 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7742 MVT::i32, cpOutL.getValue(2));
7743 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7744 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7745 Results.push_back(cpOutH.getValue(1));
7748 case ISD::ATOMIC_LOAD_ADD:
7749 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7751 case ISD::ATOMIC_LOAD_AND:
7752 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7754 case ISD::ATOMIC_LOAD_NAND:
7755 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7757 case ISD::ATOMIC_LOAD_OR:
7758 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7760 case ISD::ATOMIC_LOAD_SUB:
7761 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7763 case ISD::ATOMIC_LOAD_XOR:
7764 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7766 case ISD::ATOMIC_SWAP:
7767 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7772 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7774 default: return NULL;
7775 case X86ISD::BSF: return "X86ISD::BSF";
7776 case X86ISD::BSR: return "X86ISD::BSR";
7777 case X86ISD::SHLD: return "X86ISD::SHLD";
7778 case X86ISD::SHRD: return "X86ISD::SHRD";
7779 case X86ISD::FAND: return "X86ISD::FAND";
7780 case X86ISD::FOR: return "X86ISD::FOR";
7781 case X86ISD::FXOR: return "X86ISD::FXOR";
7782 case X86ISD::FSRL: return "X86ISD::FSRL";
7783 case X86ISD::FILD: return "X86ISD::FILD";
7784 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7785 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7786 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7787 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7788 case X86ISD::FLD: return "X86ISD::FLD";
7789 case X86ISD::FST: return "X86ISD::FST";
7790 case X86ISD::CALL: return "X86ISD::CALL";
7791 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7792 case X86ISD::BT: return "X86ISD::BT";
7793 case X86ISD::CMP: return "X86ISD::CMP";
7794 case X86ISD::COMI: return "X86ISD::COMI";
7795 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7796 case X86ISD::SETCC: return "X86ISD::SETCC";
7797 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7798 case X86ISD::CMOV: return "X86ISD::CMOV";
7799 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7800 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7801 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7802 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7803 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7804 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7805 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7806 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7807 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7808 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7809 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7810 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7811 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7812 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7813 case X86ISD::FMAX: return "X86ISD::FMAX";
7814 case X86ISD::FMIN: return "X86ISD::FMIN";
7815 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7816 case X86ISD::FRCP: return "X86ISD::FRCP";
7817 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7818 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7819 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7820 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7821 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7822 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7823 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7824 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7825 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7826 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7827 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7828 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7829 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7830 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7831 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7832 case X86ISD::VSHL: return "X86ISD::VSHL";
7833 case X86ISD::VSRL: return "X86ISD::VSRL";
7834 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7835 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7836 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7837 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7838 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7839 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7840 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7841 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7842 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7843 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7844 case X86ISD::ADD: return "X86ISD::ADD";
7845 case X86ISD::SUB: return "X86ISD::SUB";
7846 case X86ISD::SMUL: return "X86ISD::SMUL";
7847 case X86ISD::UMUL: return "X86ISD::UMUL";
7848 case X86ISD::INC: return "X86ISD::INC";
7849 case X86ISD::DEC: return "X86ISD::DEC";
7850 case X86ISD::OR: return "X86ISD::OR";
7851 case X86ISD::XOR: return "X86ISD::XOR";
7852 case X86ISD::AND: return "X86ISD::AND";
7853 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7854 case X86ISD::PTEST: return "X86ISD::PTEST";
7855 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7856 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7860 // isLegalAddressingMode - Return true if the addressing mode represented
7861 // by AM is legal for this target, for a load/store of the specified type.
7862 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7863 const Type *Ty) const {
7864 // X86 supports extremely general addressing modes.
7865 CodeModel::Model M = getTargetMachine().getCodeModel();
7867 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7868 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7873 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7875 // If a reference to this global requires an extra load, we can't fold it.
7876 if (isGlobalStubReference(GVFlags))
7879 // If BaseGV requires a register for the PIC base, we cannot also have a
7880 // BaseReg specified.
7881 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7884 // If lower 4G is not available, then we must use rip-relative addressing.
7885 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7895 // These scales always work.
7900 // These scales are formed with basereg+scalereg. Only accept if there is
7905 default: // Other stuff never works.
7913 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7914 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7916 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7917 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7918 if (NumBits1 <= NumBits2)
7923 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7924 if (!VT1.isInteger() || !VT2.isInteger())
7926 unsigned NumBits1 = VT1.getSizeInBits();
7927 unsigned NumBits2 = VT2.getSizeInBits();
7928 if (NumBits1 <= NumBits2)
7933 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7934 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7935 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7938 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7939 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7940 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7943 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7944 // i16 instructions are longer (0x66 prefix) and potentially slower.
7945 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7948 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7949 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7950 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7951 /// are assumed to be legal.
7953 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7955 // Very little shuffling can be done for 64-bit vectors right now.
7956 if (VT.getSizeInBits() == 64)
7957 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
7959 // FIXME: pshufb, blends, shifts.
7960 return (VT.getVectorNumElements() == 2 ||
7961 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7962 isMOVLMask(M, VT) ||
7963 isSHUFPMask(M, VT) ||
7964 isPSHUFDMask(M, VT) ||
7965 isPSHUFHWMask(M, VT) ||
7966 isPSHUFLWMask(M, VT) ||
7967 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7968 isUNPCKLMask(M, VT) ||
7969 isUNPCKHMask(M, VT) ||
7970 isUNPCKL_v_undef_Mask(M, VT) ||
7971 isUNPCKH_v_undef_Mask(M, VT));
7975 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7977 unsigned NumElts = VT.getVectorNumElements();
7978 // FIXME: This collection of masks seems suspect.
7981 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7982 return (isMOVLMask(Mask, VT) ||
7983 isCommutedMOVLMask(Mask, VT, true) ||
7984 isSHUFPMask(Mask, VT) ||
7985 isCommutedSHUFPMask(Mask, VT));
7990 //===----------------------------------------------------------------------===//
7991 // X86 Scheduler Hooks
7992 //===----------------------------------------------------------------------===//
7994 // private utility function
7996 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7997 MachineBasicBlock *MBB,
8005 TargetRegisterClass *RC,
8006 bool invSrc) const {
8007 // For the atomic bitwise operator, we generate
8010 // ld t1 = [bitinstr.addr]
8011 // op t2 = t1, [bitinstr.val]
8013 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8015 // fallthrough -->nextMBB
8016 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8017 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8018 MachineFunction::iterator MBBIter = MBB;
8021 /// First build the CFG
8022 MachineFunction *F = MBB->getParent();
8023 MachineBasicBlock *thisMBB = MBB;
8024 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8025 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8026 F->insert(MBBIter, newMBB);
8027 F->insert(MBBIter, nextMBB);
8029 // Move all successors to thisMBB to nextMBB
8030 nextMBB->transferSuccessors(thisMBB);
8032 // Update thisMBB to fall through to newMBB
8033 thisMBB->addSuccessor(newMBB);
8035 // newMBB jumps to itself and fall through to nextMBB
8036 newMBB->addSuccessor(nextMBB);
8037 newMBB->addSuccessor(newMBB);
8039 // Insert instructions into newMBB based on incoming instruction
8040 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8041 "unexpected number of operands");
8042 DebugLoc dl = bInstr->getDebugLoc();
8043 MachineOperand& destOper = bInstr->getOperand(0);
8044 MachineOperand* argOpers[2 + X86AddrNumOperands];
8045 int numArgs = bInstr->getNumOperands() - 1;
8046 for (int i=0; i < numArgs; ++i)
8047 argOpers[i] = &bInstr->getOperand(i+1);
8049 // x86 address has 4 operands: base, index, scale, and displacement
8050 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8051 int valArgIndx = lastAddrIndx + 1;
8053 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8054 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8055 for (int i=0; i <= lastAddrIndx; ++i)
8056 (*MIB).addOperand(*argOpers[i]);
8058 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8060 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8065 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8066 assert((argOpers[valArgIndx]->isReg() ||
8067 argOpers[valArgIndx]->isImm()) &&
8069 if (argOpers[valArgIndx]->isReg())
8070 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8072 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8074 (*MIB).addOperand(*argOpers[valArgIndx]);
8076 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8079 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8080 for (int i=0; i <= lastAddrIndx; ++i)
8081 (*MIB).addOperand(*argOpers[i]);
8083 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8084 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8085 bInstr->memoperands_end());
8087 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8091 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8093 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8097 // private utility function: 64 bit atomics on 32 bit host.
8099 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8100 MachineBasicBlock *MBB,
8105 bool invSrc) const {
8106 // For the atomic bitwise operator, we generate
8107 // thisMBB (instructions are in pairs, except cmpxchg8b)
8108 // ld t1,t2 = [bitinstr.addr]
8110 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8111 // op t5, t6 <- out1, out2, [bitinstr.val]
8112 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8113 // mov ECX, EBX <- t5, t6
8114 // mov EAX, EDX <- t1, t2
8115 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8116 // mov t3, t4 <- EAX, EDX
8118 // result in out1, out2
8119 // fallthrough -->nextMBB
8121 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8122 const unsigned LoadOpc = X86::MOV32rm;
8123 const unsigned copyOpc = X86::MOV32rr;
8124 const unsigned NotOpc = X86::NOT32r;
8125 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8126 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8127 MachineFunction::iterator MBBIter = MBB;
8130 /// First build the CFG
8131 MachineFunction *F = MBB->getParent();
8132 MachineBasicBlock *thisMBB = MBB;
8133 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8134 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8135 F->insert(MBBIter, newMBB);
8136 F->insert(MBBIter, nextMBB);
8138 // Move all successors to thisMBB to nextMBB
8139 nextMBB->transferSuccessors(thisMBB);
8141 // Update thisMBB to fall through to newMBB
8142 thisMBB->addSuccessor(newMBB);
8144 // newMBB jumps to itself and fall through to nextMBB
8145 newMBB->addSuccessor(nextMBB);
8146 newMBB->addSuccessor(newMBB);
8148 DebugLoc dl = bInstr->getDebugLoc();
8149 // Insert instructions into newMBB based on incoming instruction
8150 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8151 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8152 "unexpected number of operands");
8153 MachineOperand& dest1Oper = bInstr->getOperand(0);
8154 MachineOperand& dest2Oper = bInstr->getOperand(1);
8155 MachineOperand* argOpers[2 + X86AddrNumOperands];
8156 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8157 argOpers[i] = &bInstr->getOperand(i+2);
8159 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8160 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8162 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8163 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8164 for (int i=0; i <= lastAddrIndx; ++i)
8165 (*MIB).addOperand(*argOpers[i]);
8166 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8167 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8168 // add 4 to displacement.
8169 for (int i=0; i <= lastAddrIndx-2; ++i)
8170 (*MIB).addOperand(*argOpers[i]);
8171 MachineOperand newOp3 = *(argOpers[3]);
8173 newOp3.setImm(newOp3.getImm()+4);
8175 newOp3.setOffset(newOp3.getOffset()+4);
8176 (*MIB).addOperand(newOp3);
8177 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8179 // t3/4 are defined later, at the bottom of the loop
8180 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8181 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8182 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8183 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8184 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8185 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8187 // The subsequent operations should be using the destination registers of
8188 //the PHI instructions.
8190 t1 = F->getRegInfo().createVirtualRegister(RC);
8191 t2 = F->getRegInfo().createVirtualRegister(RC);
8192 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8193 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8195 t1 = dest1Oper.getReg();
8196 t2 = dest2Oper.getReg();
8199 int valArgIndx = lastAddrIndx + 1;
8200 assert((argOpers[valArgIndx]->isReg() ||
8201 argOpers[valArgIndx]->isImm()) &&
8203 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8204 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8205 if (argOpers[valArgIndx]->isReg())
8206 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8208 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8209 if (regOpcL != X86::MOV32rr)
8211 (*MIB).addOperand(*argOpers[valArgIndx]);
8212 assert(argOpers[valArgIndx + 1]->isReg() ==
8213 argOpers[valArgIndx]->isReg());
8214 assert(argOpers[valArgIndx + 1]->isImm() ==
8215 argOpers[valArgIndx]->isImm());
8216 if (argOpers[valArgIndx + 1]->isReg())
8217 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8219 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8220 if (regOpcH != X86::MOV32rr)
8222 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8224 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8226 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8229 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8231 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8234 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8235 for (int i=0; i <= lastAddrIndx; ++i)
8236 (*MIB).addOperand(*argOpers[i]);
8238 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8239 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8240 bInstr->memoperands_end());
8242 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8243 MIB.addReg(X86::EAX);
8244 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8245 MIB.addReg(X86::EDX);
8248 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8250 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8254 // private utility function
8256 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8257 MachineBasicBlock *MBB,
8258 unsigned cmovOpc) const {
8259 // For the atomic min/max operator, we generate
8262 // ld t1 = [min/max.addr]
8263 // mov t2 = [min/max.val]
8265 // cmov[cond] t2 = t1
8267 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8269 // fallthrough -->nextMBB
8271 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8272 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8273 MachineFunction::iterator MBBIter = MBB;
8276 /// First build the CFG
8277 MachineFunction *F = MBB->getParent();
8278 MachineBasicBlock *thisMBB = MBB;
8279 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8280 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8281 F->insert(MBBIter, newMBB);
8282 F->insert(MBBIter, nextMBB);
8284 // Move all successors of thisMBB to nextMBB
8285 nextMBB->transferSuccessors(thisMBB);
8287 // Update thisMBB to fall through to newMBB
8288 thisMBB->addSuccessor(newMBB);
8290 // newMBB jumps to newMBB and fall through to nextMBB
8291 newMBB->addSuccessor(nextMBB);
8292 newMBB->addSuccessor(newMBB);
8294 DebugLoc dl = mInstr->getDebugLoc();
8295 // Insert instructions into newMBB based on incoming instruction
8296 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8297 "unexpected number of operands");
8298 MachineOperand& destOper = mInstr->getOperand(0);
8299 MachineOperand* argOpers[2 + X86AddrNumOperands];
8300 int numArgs = mInstr->getNumOperands() - 1;
8301 for (int i=0; i < numArgs; ++i)
8302 argOpers[i] = &mInstr->getOperand(i+1);
8304 // x86 address has 4 operands: base, index, scale, and displacement
8305 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8306 int valArgIndx = lastAddrIndx + 1;
8308 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8309 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8310 for (int i=0; i <= lastAddrIndx; ++i)
8311 (*MIB).addOperand(*argOpers[i]);
8313 // We only support register and immediate values
8314 assert((argOpers[valArgIndx]->isReg() ||
8315 argOpers[valArgIndx]->isImm()) &&
8318 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8319 if (argOpers[valArgIndx]->isReg())
8320 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8322 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8323 (*MIB).addOperand(*argOpers[valArgIndx]);
8325 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8328 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8333 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8334 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8338 // Cmp and exchange if none has modified the memory location
8339 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8340 for (int i=0; i <= lastAddrIndx; ++i)
8341 (*MIB).addOperand(*argOpers[i]);
8343 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8344 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8345 mInstr->memoperands_end());
8347 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8348 MIB.addReg(X86::EAX);
8351 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8353 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8357 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8358 // all of this code can be replaced with that in the .td file.
8360 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8361 unsigned numArgs, bool memArg) const {
8363 MachineFunction *F = BB->getParent();
8364 DebugLoc dl = MI->getDebugLoc();
8365 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8369 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8371 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8373 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8375 for (unsigned i = 0; i < numArgs; ++i) {
8376 MachineOperand &Op = MI->getOperand(i+1);
8378 if (!(Op.isReg() && Op.isImplicit()))
8382 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8385 F->DeleteMachineInstr(MI);
8391 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8393 MachineBasicBlock *MBB) const {
8394 // Emit code to save XMM registers to the stack. The ABI says that the
8395 // number of registers to save is given in %al, so it's theoretically
8396 // possible to do an indirect jump trick to avoid saving all of them,
8397 // however this code takes a simpler approach and just executes all
8398 // of the stores if %al is non-zero. It's less code, and it's probably
8399 // easier on the hardware branch predictor, and stores aren't all that
8400 // expensive anyway.
8402 // Create the new basic blocks. One block contains all the XMM stores,
8403 // and one block is the final destination regardless of whether any
8404 // stores were performed.
8405 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8406 MachineFunction *F = MBB->getParent();
8407 MachineFunction::iterator MBBIter = MBB;
8409 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8410 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8411 F->insert(MBBIter, XMMSaveMBB);
8412 F->insert(MBBIter, EndMBB);
8415 // Move any original successors of MBB to the end block.
8416 EndMBB->transferSuccessors(MBB);
8417 // The original block will now fall through to the XMM save block.
8418 MBB->addSuccessor(XMMSaveMBB);
8419 // The XMMSaveMBB will fall through to the end block.
8420 XMMSaveMBB->addSuccessor(EndMBB);
8422 // Now add the instructions.
8423 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8424 DebugLoc DL = MI->getDebugLoc();
8426 unsigned CountReg = MI->getOperand(0).getReg();
8427 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8428 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8430 if (!Subtarget->isTargetWin64()) {
8431 // If %al is 0, branch around the XMM save block.
8432 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8433 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8434 MBB->addSuccessor(EndMBB);
8437 // In the XMM save block, save all the XMM argument registers.
8438 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8439 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8440 MachineMemOperand *MMO =
8441 F->getMachineMemOperand(
8442 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8443 MachineMemOperand::MOStore, Offset,
8444 /*Size=*/16, /*Align=*/16);
8445 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8446 .addFrameIndex(RegSaveFrameIndex)
8447 .addImm(/*Scale=*/1)
8448 .addReg(/*IndexReg=*/0)
8449 .addImm(/*Disp=*/Offset)
8450 .addReg(/*Segment=*/0)
8451 .addReg(MI->getOperand(i).getReg())
8452 .addMemOperand(MMO);
8455 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8461 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8462 MachineBasicBlock *BB,
8463 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8464 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8465 DebugLoc DL = MI->getDebugLoc();
8467 // To "insert" a SELECT_CC instruction, we actually have to insert the
8468 // diamond control-flow pattern. The incoming instruction knows the
8469 // destination vreg to set, the condition code register to branch on, the
8470 // true/false values to select between, and a branch opcode to use.
8471 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8472 MachineFunction::iterator It = BB;
8478 // cmpTY ccX, r1, r2
8480 // fallthrough --> copy0MBB
8481 MachineBasicBlock *thisMBB = BB;
8482 MachineFunction *F = BB->getParent();
8483 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8484 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8486 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8487 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8488 F->insert(It, copy0MBB);
8489 F->insert(It, sinkMBB);
8490 // Update machine-CFG edges by first adding all successors of the current
8491 // block to the new block which will contain the Phi node for the select.
8492 // Also inform sdisel of the edge changes.
8493 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8494 E = BB->succ_end(); I != E; ++I) {
8495 EM->insert(std::make_pair(*I, sinkMBB));
8496 sinkMBB->addSuccessor(*I);
8498 // Next, remove all successors of the current block, and add the true
8499 // and fallthrough blocks as its successors.
8500 while (!BB->succ_empty())
8501 BB->removeSuccessor(BB->succ_begin());
8502 // Add the true and fallthrough blocks as its successors.
8503 BB->addSuccessor(copy0MBB);
8504 BB->addSuccessor(sinkMBB);
8507 // %FalseValue = ...
8508 // # fallthrough to sinkMBB
8511 // Update machine-CFG edges
8512 BB->addSuccessor(sinkMBB);
8515 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8518 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8519 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8520 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8522 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8527 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8528 MachineBasicBlock *BB,
8529 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8530 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8531 DebugLoc DL = MI->getDebugLoc();
8532 MachineFunction *F = BB->getParent();
8534 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8535 // non-trivial part is impdef of ESP.
8536 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8539 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8540 .addExternalSymbol("_alloca")
8541 .addReg(X86::EAX, RegState::Implicit)
8542 .addReg(X86::ESP, RegState::Implicit)
8543 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8544 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8546 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8551 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8552 MachineBasicBlock *BB,
8553 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8554 switch (MI->getOpcode()) {
8555 default: assert(false && "Unexpected instr type to insert");
8556 case X86::MINGW_ALLOCA:
8557 return EmitLoweredMingwAlloca(MI, BB, EM);
8559 case X86::CMOV_V1I64:
8560 case X86::CMOV_FR32:
8561 case X86::CMOV_FR64:
8562 case X86::CMOV_V4F32:
8563 case X86::CMOV_V2F64:
8564 case X86::CMOV_V2I64:
8565 case X86::CMOV_GR16:
8566 case X86::CMOV_GR32:
8567 case X86::CMOV_RFP32:
8568 case X86::CMOV_RFP64:
8569 case X86::CMOV_RFP80:
8570 return EmitLoweredSelect(MI, BB, EM);
8572 case X86::FP32_TO_INT16_IN_MEM:
8573 case X86::FP32_TO_INT32_IN_MEM:
8574 case X86::FP32_TO_INT64_IN_MEM:
8575 case X86::FP64_TO_INT16_IN_MEM:
8576 case X86::FP64_TO_INT32_IN_MEM:
8577 case X86::FP64_TO_INT64_IN_MEM:
8578 case X86::FP80_TO_INT16_IN_MEM:
8579 case X86::FP80_TO_INT32_IN_MEM:
8580 case X86::FP80_TO_INT64_IN_MEM: {
8581 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8582 DebugLoc DL = MI->getDebugLoc();
8584 // Change the floating point control register to use "round towards zero"
8585 // mode when truncating to an integer value.
8586 MachineFunction *F = BB->getParent();
8587 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8588 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8590 // Load the old value of the high byte of the control word...
8592 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8593 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8596 // Set the high part to be round to zero...
8597 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8600 // Reload the modified control word now...
8601 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8603 // Restore the memory image of control word to original value
8604 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8607 // Get the X86 opcode to use.
8609 switch (MI->getOpcode()) {
8610 default: llvm_unreachable("illegal opcode!");
8611 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8612 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8613 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8614 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8615 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8616 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8617 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8618 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8619 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8623 MachineOperand &Op = MI->getOperand(0);
8625 AM.BaseType = X86AddressMode::RegBase;
8626 AM.Base.Reg = Op.getReg();
8628 AM.BaseType = X86AddressMode::FrameIndexBase;
8629 AM.Base.FrameIndex = Op.getIndex();
8631 Op = MI->getOperand(1);
8633 AM.Scale = Op.getImm();
8634 Op = MI->getOperand(2);
8636 AM.IndexReg = Op.getImm();
8637 Op = MI->getOperand(3);
8638 if (Op.isGlobal()) {
8639 AM.GV = Op.getGlobal();
8641 AM.Disp = Op.getImm();
8643 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8644 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8646 // Reload the original control word now.
8647 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8649 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8652 // DBG_VALUE. Only the frame index case is done here.
8653 case X86::DBG_VALUE: {
8654 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8655 DebugLoc DL = MI->getDebugLoc();
8657 MachineFunction *F = BB->getParent();
8658 AM.BaseType = X86AddressMode::FrameIndexBase;
8659 AM.Base.FrameIndex = MI->getOperand(0).getImm();
8660 addFullAddress(BuildMI(BB, DL, TII->get(X86::DBG_VALUE)), AM).
8661 addImm(MI->getOperand(1).getImm()).
8662 addMetadata(MI->getOperand(2).getMetadata());
8663 F->DeleteMachineInstr(MI); // Remove pseudo.
8667 // String/text processing lowering.
8668 case X86::PCMPISTRM128REG:
8669 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8670 case X86::PCMPISTRM128MEM:
8671 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8672 case X86::PCMPESTRM128REG:
8673 return EmitPCMP(MI, BB, 5, false /* in mem */);
8674 case X86::PCMPESTRM128MEM:
8675 return EmitPCMP(MI, BB, 5, true /* in mem */);
8678 case X86::ATOMAND32:
8679 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8680 X86::AND32ri, X86::MOV32rm,
8681 X86::LCMPXCHG32, X86::MOV32rr,
8682 X86::NOT32r, X86::EAX,
8683 X86::GR32RegisterClass);
8685 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8686 X86::OR32ri, X86::MOV32rm,
8687 X86::LCMPXCHG32, X86::MOV32rr,
8688 X86::NOT32r, X86::EAX,
8689 X86::GR32RegisterClass);
8690 case X86::ATOMXOR32:
8691 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8692 X86::XOR32ri, X86::MOV32rm,
8693 X86::LCMPXCHG32, X86::MOV32rr,
8694 X86::NOT32r, X86::EAX,
8695 X86::GR32RegisterClass);
8696 case X86::ATOMNAND32:
8697 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8698 X86::AND32ri, X86::MOV32rm,
8699 X86::LCMPXCHG32, X86::MOV32rr,
8700 X86::NOT32r, X86::EAX,
8701 X86::GR32RegisterClass, true);
8702 case X86::ATOMMIN32:
8703 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8704 case X86::ATOMMAX32:
8705 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8706 case X86::ATOMUMIN32:
8707 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8708 case X86::ATOMUMAX32:
8709 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8711 case X86::ATOMAND16:
8712 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8713 X86::AND16ri, X86::MOV16rm,
8714 X86::LCMPXCHG16, X86::MOV16rr,
8715 X86::NOT16r, X86::AX,
8716 X86::GR16RegisterClass);
8718 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8719 X86::OR16ri, X86::MOV16rm,
8720 X86::LCMPXCHG16, X86::MOV16rr,
8721 X86::NOT16r, X86::AX,
8722 X86::GR16RegisterClass);
8723 case X86::ATOMXOR16:
8724 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8725 X86::XOR16ri, X86::MOV16rm,
8726 X86::LCMPXCHG16, X86::MOV16rr,
8727 X86::NOT16r, X86::AX,
8728 X86::GR16RegisterClass);
8729 case X86::ATOMNAND16:
8730 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8731 X86::AND16ri, X86::MOV16rm,
8732 X86::LCMPXCHG16, X86::MOV16rr,
8733 X86::NOT16r, X86::AX,
8734 X86::GR16RegisterClass, true);
8735 case X86::ATOMMIN16:
8736 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8737 case X86::ATOMMAX16:
8738 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8739 case X86::ATOMUMIN16:
8740 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8741 case X86::ATOMUMAX16:
8742 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8745 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8746 X86::AND8ri, X86::MOV8rm,
8747 X86::LCMPXCHG8, X86::MOV8rr,
8748 X86::NOT8r, X86::AL,
8749 X86::GR8RegisterClass);
8751 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8752 X86::OR8ri, X86::MOV8rm,
8753 X86::LCMPXCHG8, X86::MOV8rr,
8754 X86::NOT8r, X86::AL,
8755 X86::GR8RegisterClass);
8757 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8758 X86::XOR8ri, X86::MOV8rm,
8759 X86::LCMPXCHG8, X86::MOV8rr,
8760 X86::NOT8r, X86::AL,
8761 X86::GR8RegisterClass);
8762 case X86::ATOMNAND8:
8763 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8764 X86::AND8ri, X86::MOV8rm,
8765 X86::LCMPXCHG8, X86::MOV8rr,
8766 X86::NOT8r, X86::AL,
8767 X86::GR8RegisterClass, true);
8768 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8769 // This group is for 64-bit host.
8770 case X86::ATOMAND64:
8771 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8772 X86::AND64ri32, X86::MOV64rm,
8773 X86::LCMPXCHG64, X86::MOV64rr,
8774 X86::NOT64r, X86::RAX,
8775 X86::GR64RegisterClass);
8777 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8778 X86::OR64ri32, X86::MOV64rm,
8779 X86::LCMPXCHG64, X86::MOV64rr,
8780 X86::NOT64r, X86::RAX,
8781 X86::GR64RegisterClass);
8782 case X86::ATOMXOR64:
8783 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8784 X86::XOR64ri32, X86::MOV64rm,
8785 X86::LCMPXCHG64, X86::MOV64rr,
8786 X86::NOT64r, X86::RAX,
8787 X86::GR64RegisterClass);
8788 case X86::ATOMNAND64:
8789 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8790 X86::AND64ri32, X86::MOV64rm,
8791 X86::LCMPXCHG64, X86::MOV64rr,
8792 X86::NOT64r, X86::RAX,
8793 X86::GR64RegisterClass, true);
8794 case X86::ATOMMIN64:
8795 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8796 case X86::ATOMMAX64:
8797 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8798 case X86::ATOMUMIN64:
8799 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8800 case X86::ATOMUMAX64:
8801 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8803 // This group does 64-bit operations on a 32-bit host.
8804 case X86::ATOMAND6432:
8805 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8806 X86::AND32rr, X86::AND32rr,
8807 X86::AND32ri, X86::AND32ri,
8809 case X86::ATOMOR6432:
8810 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8811 X86::OR32rr, X86::OR32rr,
8812 X86::OR32ri, X86::OR32ri,
8814 case X86::ATOMXOR6432:
8815 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8816 X86::XOR32rr, X86::XOR32rr,
8817 X86::XOR32ri, X86::XOR32ri,
8819 case X86::ATOMNAND6432:
8820 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8821 X86::AND32rr, X86::AND32rr,
8822 X86::AND32ri, X86::AND32ri,
8824 case X86::ATOMADD6432:
8825 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8826 X86::ADD32rr, X86::ADC32rr,
8827 X86::ADD32ri, X86::ADC32ri,
8829 case X86::ATOMSUB6432:
8830 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8831 X86::SUB32rr, X86::SBB32rr,
8832 X86::SUB32ri, X86::SBB32ri,
8834 case X86::ATOMSWAP6432:
8835 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8836 X86::MOV32rr, X86::MOV32rr,
8837 X86::MOV32ri, X86::MOV32ri,
8839 case X86::VASTART_SAVE_XMM_REGS:
8840 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8844 //===----------------------------------------------------------------------===//
8845 // X86 Optimization Hooks
8846 //===----------------------------------------------------------------------===//
8848 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8852 const SelectionDAG &DAG,
8853 unsigned Depth) const {
8854 unsigned Opc = Op.getOpcode();
8855 assert((Opc >= ISD::BUILTIN_OP_END ||
8856 Opc == ISD::INTRINSIC_WO_CHAIN ||
8857 Opc == ISD::INTRINSIC_W_CHAIN ||
8858 Opc == ISD::INTRINSIC_VOID) &&
8859 "Should use MaskedValueIsZero if you don't know whether Op"
8860 " is a target node!");
8862 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8874 // These nodes' second result is a boolean.
8875 if (Op.getResNo() == 0)
8879 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8880 Mask.getBitWidth() - 1);
8885 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8886 /// node is a GlobalAddress + offset.
8887 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8888 const GlobalValue* &GA,
8889 int64_t &Offset) const {
8890 if (N->getOpcode() == X86ISD::Wrapper) {
8891 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8892 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8893 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8897 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8900 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8901 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8902 /// if the load addresses are consecutive, non-overlapping, and in the right
8904 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8905 const TargetLowering &TLI) {
8906 DebugLoc dl = N->getDebugLoc();
8907 EVT VT = N->getValueType(0);
8908 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8910 if (VT.getSizeInBits() != 128)
8913 SmallVector<SDValue, 16> Elts;
8914 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8915 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8917 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8920 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8921 /// and convert it from being a bunch of shuffles and extracts to a simple
8922 /// store and scalar loads to extract the elements.
8923 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8924 const TargetLowering &TLI) {
8925 SDValue InputVector = N->getOperand(0);
8927 // Only operate on vectors of 4 elements, where the alternative shuffling
8928 // gets to be more expensive.
8929 if (InputVector.getValueType() != MVT::v4i32)
8932 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8933 // single use which is a sign-extend or zero-extend, and all elements are
8935 SmallVector<SDNode *, 4> Uses;
8936 unsigned ExtractedElements = 0;
8937 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8938 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8939 if (UI.getUse().getResNo() != InputVector.getResNo())
8942 SDNode *Extract = *UI;
8943 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8946 if (Extract->getValueType(0) != MVT::i32)
8948 if (!Extract->hasOneUse())
8950 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
8951 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
8953 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
8956 // Record which element was extracted.
8957 ExtractedElements |=
8958 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
8960 Uses.push_back(Extract);
8963 // If not all the elements were used, this may not be worthwhile.
8964 if (ExtractedElements != 15)
8967 // Ok, we've now decided to do the transformation.
8968 DebugLoc dl = InputVector.getDebugLoc();
8970 // Store the value to a temporary stack slot.
8971 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
8972 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
8975 // Replace each use (extract) with a load of the appropriate element.
8976 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
8977 UE = Uses.end(); UI != UE; ++UI) {
8978 SDNode *Extract = *UI;
8980 // Compute the element's address.
8981 SDValue Idx = Extract->getOperand(1);
8983 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
8984 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
8985 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
8987 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
8990 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
8991 NULL, 0, false, false, 0);
8993 // Replace the exact with the load.
8994 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
8997 // The replacement was made in place; don't return anything.
9001 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9002 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9003 const X86Subtarget *Subtarget) {
9004 DebugLoc DL = N->getDebugLoc();
9005 SDValue Cond = N->getOperand(0);
9006 // Get the LHS/RHS of the select.
9007 SDValue LHS = N->getOperand(1);
9008 SDValue RHS = N->getOperand(2);
9010 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9011 // instructions match the semantics of the common C idiom x<y?x:y but not
9012 // x<=y?x:y, because of how they handle negative zero (which can be
9013 // ignored in unsafe-math mode).
9014 if (Subtarget->hasSSE2() &&
9015 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9016 Cond.getOpcode() == ISD::SETCC) {
9017 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9019 unsigned Opcode = 0;
9020 // Check for x CC y ? x : y.
9021 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9022 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9026 // Converting this to a min would handle NaNs incorrectly, and swapping
9027 // the operands would cause it to handle comparisons between positive
9028 // and negative zero incorrectly.
9029 if (!FiniteOnlyFPMath() &&
9030 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9031 if (!UnsafeFPMath &&
9032 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9034 std::swap(LHS, RHS);
9036 Opcode = X86ISD::FMIN;
9039 // Converting this to a min would handle comparisons between positive
9040 // and negative zero incorrectly.
9041 if (!UnsafeFPMath &&
9042 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9044 Opcode = X86ISD::FMIN;
9047 // Converting this to a min would handle both negative zeros and NaNs
9048 // incorrectly, but we can swap the operands to fix both.
9049 std::swap(LHS, RHS);
9053 Opcode = X86ISD::FMIN;
9057 // Converting this to a max would handle comparisons between positive
9058 // and negative zero incorrectly.
9059 if (!UnsafeFPMath &&
9060 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9062 Opcode = X86ISD::FMAX;
9065 // Converting this to a max would handle NaNs incorrectly, and swapping
9066 // the operands would cause it to handle comparisons between positive
9067 // and negative zero incorrectly.
9068 if (!FiniteOnlyFPMath() &&
9069 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9070 if (!UnsafeFPMath &&
9071 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9073 std::swap(LHS, RHS);
9075 Opcode = X86ISD::FMAX;
9078 // Converting this to a max would handle both negative zeros and NaNs
9079 // incorrectly, but we can swap the operands to fix both.
9080 std::swap(LHS, RHS);
9084 Opcode = X86ISD::FMAX;
9087 // Check for x CC y ? y : x -- a min/max with reversed arms.
9088 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9089 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9093 // Converting this to a min would handle comparisons between positive
9094 // and negative zero incorrectly, and swapping the operands would
9095 // cause it to handle NaNs incorrectly.
9096 if (!UnsafeFPMath &&
9097 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9098 if (!FiniteOnlyFPMath() &&
9099 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9101 std::swap(LHS, RHS);
9103 Opcode = X86ISD::FMIN;
9106 // Converting this to a min would handle NaNs incorrectly.
9107 if (!UnsafeFPMath &&
9108 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9110 Opcode = X86ISD::FMIN;
9113 // Converting this to a min would handle both negative zeros and NaNs
9114 // incorrectly, but we can swap the operands to fix both.
9115 std::swap(LHS, RHS);
9119 Opcode = X86ISD::FMIN;
9123 // Converting this to a max would handle NaNs incorrectly.
9124 if (!FiniteOnlyFPMath() &&
9125 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9127 Opcode = X86ISD::FMAX;
9130 // Converting this to a max would handle comparisons between positive
9131 // and negative zero incorrectly, and swapping the operands would
9132 // cause it to handle NaNs incorrectly.
9133 if (!UnsafeFPMath &&
9134 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9135 if (!FiniteOnlyFPMath() &&
9136 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9138 std::swap(LHS, RHS);
9140 Opcode = X86ISD::FMAX;
9143 // Converting this to a max would handle both negative zeros and NaNs
9144 // incorrectly, but we can swap the operands to fix both.
9145 std::swap(LHS, RHS);
9149 Opcode = X86ISD::FMAX;
9155 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9158 // If this is a select between two integer constants, try to do some
9160 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9161 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9162 // Don't do this for crazy integer types.
9163 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9164 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9165 // so that TrueC (the true value) is larger than FalseC.
9166 bool NeedsCondInvert = false;
9168 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9169 // Efficiently invertible.
9170 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9171 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9172 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9173 NeedsCondInvert = true;
9174 std::swap(TrueC, FalseC);
9177 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9178 if (FalseC->getAPIntValue() == 0 &&
9179 TrueC->getAPIntValue().isPowerOf2()) {
9180 if (NeedsCondInvert) // Invert the condition if needed.
9181 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9182 DAG.getConstant(1, Cond.getValueType()));
9184 // Zero extend the condition if needed.
9185 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9187 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9188 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9189 DAG.getConstant(ShAmt, MVT::i8));
9192 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9193 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9194 if (NeedsCondInvert) // Invert the condition if needed.
9195 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9196 DAG.getConstant(1, Cond.getValueType()));
9198 // Zero extend the condition if needed.
9199 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9200 FalseC->getValueType(0), Cond);
9201 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9202 SDValue(FalseC, 0));
9205 // Optimize cases that will turn into an LEA instruction. This requires
9206 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9207 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9208 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9209 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9211 bool isFastMultiplier = false;
9213 switch ((unsigned char)Diff) {
9215 case 1: // result = add base, cond
9216 case 2: // result = lea base( , cond*2)
9217 case 3: // result = lea base(cond, cond*2)
9218 case 4: // result = lea base( , cond*4)
9219 case 5: // result = lea base(cond, cond*4)
9220 case 8: // result = lea base( , cond*8)
9221 case 9: // result = lea base(cond, cond*8)
9222 isFastMultiplier = true;
9227 if (isFastMultiplier) {
9228 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9229 if (NeedsCondInvert) // Invert the condition if needed.
9230 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9231 DAG.getConstant(1, Cond.getValueType()));
9233 // Zero extend the condition if needed.
9234 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9236 // Scale the condition by the difference.
9238 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9239 DAG.getConstant(Diff, Cond.getValueType()));
9241 // Add the base if non-zero.
9242 if (FalseC->getAPIntValue() != 0)
9243 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9244 SDValue(FalseC, 0));
9254 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9255 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9256 TargetLowering::DAGCombinerInfo &DCI) {
9257 DebugLoc DL = N->getDebugLoc();
9259 // If the flag operand isn't dead, don't touch this CMOV.
9260 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9263 // If this is a select between two integer constants, try to do some
9264 // optimizations. Note that the operands are ordered the opposite of SELECT
9266 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9267 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9268 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9269 // larger than FalseC (the false value).
9270 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9272 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9273 CC = X86::GetOppositeBranchCondition(CC);
9274 std::swap(TrueC, FalseC);
9277 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9278 // This is efficient for any integer data type (including i8/i16) and
9280 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9281 SDValue Cond = N->getOperand(3);
9282 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9283 DAG.getConstant(CC, MVT::i8), Cond);
9285 // Zero extend the condition if needed.
9286 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9288 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9289 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9290 DAG.getConstant(ShAmt, MVT::i8));
9291 if (N->getNumValues() == 2) // Dead flag value?
9292 return DCI.CombineTo(N, Cond, SDValue());
9296 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9297 // for any integer data type, including i8/i16.
9298 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9299 SDValue Cond = N->getOperand(3);
9300 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9301 DAG.getConstant(CC, MVT::i8), Cond);
9303 // Zero extend the condition if needed.
9304 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9305 FalseC->getValueType(0), Cond);
9306 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9307 SDValue(FalseC, 0));
9309 if (N->getNumValues() == 2) // Dead flag value?
9310 return DCI.CombineTo(N, Cond, SDValue());
9314 // Optimize cases that will turn into an LEA instruction. This requires
9315 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9316 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9317 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9318 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9320 bool isFastMultiplier = false;
9322 switch ((unsigned char)Diff) {
9324 case 1: // result = add base, cond
9325 case 2: // result = lea base( , cond*2)
9326 case 3: // result = lea base(cond, cond*2)
9327 case 4: // result = lea base( , cond*4)
9328 case 5: // result = lea base(cond, cond*4)
9329 case 8: // result = lea base( , cond*8)
9330 case 9: // result = lea base(cond, cond*8)
9331 isFastMultiplier = true;
9336 if (isFastMultiplier) {
9337 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9338 SDValue Cond = N->getOperand(3);
9339 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9340 DAG.getConstant(CC, MVT::i8), Cond);
9341 // Zero extend the condition if needed.
9342 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9344 // Scale the condition by the difference.
9346 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9347 DAG.getConstant(Diff, Cond.getValueType()));
9349 // Add the base if non-zero.
9350 if (FalseC->getAPIntValue() != 0)
9351 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9352 SDValue(FalseC, 0));
9353 if (N->getNumValues() == 2) // Dead flag value?
9354 return DCI.CombineTo(N, Cond, SDValue());
9364 /// PerformMulCombine - Optimize a single multiply with constant into two
9365 /// in order to implement it with two cheaper instructions, e.g.
9366 /// LEA + SHL, LEA + LEA.
9367 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9368 TargetLowering::DAGCombinerInfo &DCI) {
9369 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9372 EVT VT = N->getValueType(0);
9376 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9379 uint64_t MulAmt = C->getZExtValue();
9380 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9383 uint64_t MulAmt1 = 0;
9384 uint64_t MulAmt2 = 0;
9385 if ((MulAmt % 9) == 0) {
9387 MulAmt2 = MulAmt / 9;
9388 } else if ((MulAmt % 5) == 0) {
9390 MulAmt2 = MulAmt / 5;
9391 } else if ((MulAmt % 3) == 0) {
9393 MulAmt2 = MulAmt / 3;
9396 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9397 DebugLoc DL = N->getDebugLoc();
9399 if (isPowerOf2_64(MulAmt2) &&
9400 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9401 // If second multiplifer is pow2, issue it first. We want the multiply by
9402 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9404 std::swap(MulAmt1, MulAmt2);
9407 if (isPowerOf2_64(MulAmt1))
9408 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9409 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9411 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9412 DAG.getConstant(MulAmt1, VT));
9414 if (isPowerOf2_64(MulAmt2))
9415 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9416 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9418 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9419 DAG.getConstant(MulAmt2, VT));
9421 // Do not add new nodes to DAG combiner worklist.
9422 DCI.CombineTo(N, NewMul, false);
9427 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9428 SDValue N0 = N->getOperand(0);
9429 SDValue N1 = N->getOperand(1);
9430 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9431 EVT VT = N0.getValueType();
9433 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9434 // since the result of setcc_c is all zero's or all ones.
9435 if (N1C && N0.getOpcode() == ISD::AND &&
9436 N0.getOperand(1).getOpcode() == ISD::Constant) {
9437 SDValue N00 = N0.getOperand(0);
9438 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9439 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9440 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9441 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9442 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9443 APInt ShAmt = N1C->getAPIntValue();
9444 Mask = Mask.shl(ShAmt);
9446 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9447 N00, DAG.getConstant(Mask, VT));
9454 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9456 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9457 const X86Subtarget *Subtarget) {
9458 EVT VT = N->getValueType(0);
9459 if (!VT.isVector() && VT.isInteger() &&
9460 N->getOpcode() == ISD::SHL)
9461 return PerformSHLCombine(N, DAG);
9463 // On X86 with SSE2 support, we can transform this to a vector shift if
9464 // all elements are shifted by the same amount. We can't do this in legalize
9465 // because the a constant vector is typically transformed to a constant pool
9466 // so we have no knowledge of the shift amount.
9467 if (!Subtarget->hasSSE2())
9470 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9473 SDValue ShAmtOp = N->getOperand(1);
9474 EVT EltVT = VT.getVectorElementType();
9475 DebugLoc DL = N->getDebugLoc();
9476 SDValue BaseShAmt = SDValue();
9477 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9478 unsigned NumElts = VT.getVectorNumElements();
9480 for (; i != NumElts; ++i) {
9481 SDValue Arg = ShAmtOp.getOperand(i);
9482 if (Arg.getOpcode() == ISD::UNDEF) continue;
9486 for (; i != NumElts; ++i) {
9487 SDValue Arg = ShAmtOp.getOperand(i);
9488 if (Arg.getOpcode() == ISD::UNDEF) continue;
9489 if (Arg != BaseShAmt) {
9493 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9494 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9495 SDValue InVec = ShAmtOp.getOperand(0);
9496 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9497 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9499 for (; i != NumElts; ++i) {
9500 SDValue Arg = InVec.getOperand(i);
9501 if (Arg.getOpcode() == ISD::UNDEF) continue;
9505 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9506 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9507 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9508 if (C->getZExtValue() == SplatIdx)
9509 BaseShAmt = InVec.getOperand(1);
9512 if (BaseShAmt.getNode() == 0)
9513 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9514 DAG.getIntPtrConstant(0));
9518 // The shift amount is an i32.
9519 if (EltVT.bitsGT(MVT::i32))
9520 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9521 else if (EltVT.bitsLT(MVT::i32))
9522 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9524 // The shift amount is identical so we can do a vector shift.
9525 SDValue ValOp = N->getOperand(0);
9526 switch (N->getOpcode()) {
9528 llvm_unreachable("Unknown shift opcode!");
9531 if (VT == MVT::v2i64)
9532 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9533 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9535 if (VT == MVT::v4i32)
9536 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9537 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9539 if (VT == MVT::v8i16)
9540 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9541 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9545 if (VT == MVT::v4i32)
9546 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9547 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9549 if (VT == MVT::v8i16)
9550 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9551 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9555 if (VT == MVT::v2i64)
9556 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9557 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9559 if (VT == MVT::v4i32)
9560 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9561 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9563 if (VT == MVT::v8i16)
9564 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9565 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9572 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9573 const X86Subtarget *Subtarget) {
9574 EVT VT = N->getValueType(0);
9575 if (VT != MVT::i64 || !Subtarget->is64Bit())
9578 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9579 SDValue N0 = N->getOperand(0);
9580 SDValue N1 = N->getOperand(1);
9581 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9583 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9586 SDValue ShAmt0 = N0.getOperand(1);
9587 if (ShAmt0.getValueType() != MVT::i8)
9589 SDValue ShAmt1 = N1.getOperand(1);
9590 if (ShAmt1.getValueType() != MVT::i8)
9592 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9593 ShAmt0 = ShAmt0.getOperand(0);
9594 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9595 ShAmt1 = ShAmt1.getOperand(0);
9597 DebugLoc DL = N->getDebugLoc();
9598 unsigned Opc = X86ISD::SHLD;
9599 SDValue Op0 = N0.getOperand(0);
9600 SDValue Op1 = N1.getOperand(0);
9601 if (ShAmt0.getOpcode() == ISD::SUB) {
9603 std::swap(Op0, Op1);
9604 std::swap(ShAmt0, ShAmt1);
9607 if (ShAmt1.getOpcode() == ISD::SUB) {
9608 SDValue Sum = ShAmt1.getOperand(0);
9609 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9610 if (SumC->getSExtValue() == 64 &&
9611 ShAmt1.getOperand(1) == ShAmt0)
9612 return DAG.getNode(Opc, DL, VT,
9614 DAG.getNode(ISD::TRUNCATE, DL,
9617 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9618 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9620 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9621 return DAG.getNode(Opc, DL, VT,
9622 N0.getOperand(0), N1.getOperand(0),
9623 DAG.getNode(ISD::TRUNCATE, DL,
9630 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9631 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9632 const X86Subtarget *Subtarget) {
9633 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9634 // the FP state in cases where an emms may be missing.
9635 // A preferable solution to the general problem is to figure out the right
9636 // places to insert EMMS. This qualifies as a quick hack.
9638 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9639 StoreSDNode *St = cast<StoreSDNode>(N);
9640 EVT VT = St->getValue().getValueType();
9641 if (VT.getSizeInBits() != 64)
9644 const Function *F = DAG.getMachineFunction().getFunction();
9645 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9646 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9647 && Subtarget->hasSSE2();
9648 if ((VT.isVector() ||
9649 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9650 isa<LoadSDNode>(St->getValue()) &&
9651 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9652 St->getChain().hasOneUse() && !St->isVolatile()) {
9653 SDNode* LdVal = St->getValue().getNode();
9655 int TokenFactorIndex = -1;
9656 SmallVector<SDValue, 8> Ops;
9657 SDNode* ChainVal = St->getChain().getNode();
9658 // Must be a store of a load. We currently handle two cases: the load
9659 // is a direct child, and it's under an intervening TokenFactor. It is
9660 // possible to dig deeper under nested TokenFactors.
9661 if (ChainVal == LdVal)
9662 Ld = cast<LoadSDNode>(St->getChain());
9663 else if (St->getValue().hasOneUse() &&
9664 ChainVal->getOpcode() == ISD::TokenFactor) {
9665 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9666 if (ChainVal->getOperand(i).getNode() == LdVal) {
9667 TokenFactorIndex = i;
9668 Ld = cast<LoadSDNode>(St->getValue());
9670 Ops.push_back(ChainVal->getOperand(i));
9674 if (!Ld || !ISD::isNormalLoad(Ld))
9677 // If this is not the MMX case, i.e. we are just turning i64 load/store
9678 // into f64 load/store, avoid the transformation if there are multiple
9679 // uses of the loaded value.
9680 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9683 DebugLoc LdDL = Ld->getDebugLoc();
9684 DebugLoc StDL = N->getDebugLoc();
9685 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9686 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9688 if (Subtarget->is64Bit() || F64IsLegal) {
9689 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9690 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9691 Ld->getBasePtr(), Ld->getSrcValue(),
9692 Ld->getSrcValueOffset(), Ld->isVolatile(),
9693 Ld->isNonTemporal(), Ld->getAlignment());
9694 SDValue NewChain = NewLd.getValue(1);
9695 if (TokenFactorIndex != -1) {
9696 Ops.push_back(NewChain);
9697 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9700 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9701 St->getSrcValue(), St->getSrcValueOffset(),
9702 St->isVolatile(), St->isNonTemporal(),
9703 St->getAlignment());
9706 // Otherwise, lower to two pairs of 32-bit loads / stores.
9707 SDValue LoAddr = Ld->getBasePtr();
9708 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9709 DAG.getConstant(4, MVT::i32));
9711 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9712 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9713 Ld->isVolatile(), Ld->isNonTemporal(),
9714 Ld->getAlignment());
9715 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9716 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9717 Ld->isVolatile(), Ld->isNonTemporal(),
9718 MinAlign(Ld->getAlignment(), 4));
9720 SDValue NewChain = LoLd.getValue(1);
9721 if (TokenFactorIndex != -1) {
9722 Ops.push_back(LoLd);
9723 Ops.push_back(HiLd);
9724 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9728 LoAddr = St->getBasePtr();
9729 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9730 DAG.getConstant(4, MVT::i32));
9732 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9733 St->getSrcValue(), St->getSrcValueOffset(),
9734 St->isVolatile(), St->isNonTemporal(),
9735 St->getAlignment());
9736 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9738 St->getSrcValueOffset() + 4,
9740 St->isNonTemporal(),
9741 MinAlign(St->getAlignment(), 4));
9742 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9747 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9748 /// X86ISD::FXOR nodes.
9749 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9750 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9751 // F[X]OR(0.0, x) -> x
9752 // F[X]OR(x, 0.0) -> x
9753 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9754 if (C->getValueAPF().isPosZero())
9755 return N->getOperand(1);
9756 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9757 if (C->getValueAPF().isPosZero())
9758 return N->getOperand(0);
9762 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9763 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9764 // FAND(0.0, x) -> 0.0
9765 // FAND(x, 0.0) -> 0.0
9766 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9767 if (C->getValueAPF().isPosZero())
9768 return N->getOperand(0);
9769 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9770 if (C->getValueAPF().isPosZero())
9771 return N->getOperand(1);
9775 static SDValue PerformBTCombine(SDNode *N,
9777 TargetLowering::DAGCombinerInfo &DCI) {
9778 // BT ignores high bits in the bit index operand.
9779 SDValue Op1 = N->getOperand(1);
9780 if (Op1.hasOneUse()) {
9781 unsigned BitWidth = Op1.getValueSizeInBits();
9782 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9783 APInt KnownZero, KnownOne;
9784 TargetLowering::TargetLoweringOpt TLO(DAG);
9785 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9786 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9787 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9788 DCI.CommitTargetLoweringOpt(TLO);
9793 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9794 SDValue Op = N->getOperand(0);
9795 if (Op.getOpcode() == ISD::BIT_CONVERT)
9796 Op = Op.getOperand(0);
9797 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9798 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9799 VT.getVectorElementType().getSizeInBits() ==
9800 OpVT.getVectorElementType().getSizeInBits()) {
9801 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9806 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9807 // Locked instructions, in turn, have implicit fence semantics (all memory
9808 // operations are flushed before issuing the locked instruction, and the
9809 // are not buffered), so we can fold away the common pattern of
9810 // fence-atomic-fence.
9811 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9812 SDValue atomic = N->getOperand(0);
9813 switch (atomic.getOpcode()) {
9814 case ISD::ATOMIC_CMP_SWAP:
9815 case ISD::ATOMIC_SWAP:
9816 case ISD::ATOMIC_LOAD_ADD:
9817 case ISD::ATOMIC_LOAD_SUB:
9818 case ISD::ATOMIC_LOAD_AND:
9819 case ISD::ATOMIC_LOAD_OR:
9820 case ISD::ATOMIC_LOAD_XOR:
9821 case ISD::ATOMIC_LOAD_NAND:
9822 case ISD::ATOMIC_LOAD_MIN:
9823 case ISD::ATOMIC_LOAD_MAX:
9824 case ISD::ATOMIC_LOAD_UMIN:
9825 case ISD::ATOMIC_LOAD_UMAX:
9831 SDValue fence = atomic.getOperand(0);
9832 if (fence.getOpcode() != ISD::MEMBARRIER)
9835 switch (atomic.getOpcode()) {
9836 case ISD::ATOMIC_CMP_SWAP:
9837 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9838 atomic.getOperand(1), atomic.getOperand(2),
9839 atomic.getOperand(3));
9840 case ISD::ATOMIC_SWAP:
9841 case ISD::ATOMIC_LOAD_ADD:
9842 case ISD::ATOMIC_LOAD_SUB:
9843 case ISD::ATOMIC_LOAD_AND:
9844 case ISD::ATOMIC_LOAD_OR:
9845 case ISD::ATOMIC_LOAD_XOR:
9846 case ISD::ATOMIC_LOAD_NAND:
9847 case ISD::ATOMIC_LOAD_MIN:
9848 case ISD::ATOMIC_LOAD_MAX:
9849 case ISD::ATOMIC_LOAD_UMIN:
9850 case ISD::ATOMIC_LOAD_UMAX:
9851 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9852 atomic.getOperand(1), atomic.getOperand(2));
9858 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9859 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9860 // (and (i32 x86isd::setcc_carry), 1)
9861 // This eliminates the zext. This transformation is necessary because
9862 // ISD::SETCC is always legalized to i8.
9863 DebugLoc dl = N->getDebugLoc();
9864 SDValue N0 = N->getOperand(0);
9865 EVT VT = N->getValueType(0);
9866 if (N0.getOpcode() == ISD::AND &&
9868 N0.getOperand(0).hasOneUse()) {
9869 SDValue N00 = N0.getOperand(0);
9870 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9872 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9873 if (!C || C->getZExtValue() != 1)
9875 return DAG.getNode(ISD::AND, dl, VT,
9876 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9877 N00.getOperand(0), N00.getOperand(1)),
9878 DAG.getConstant(1, VT));
9884 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9885 DAGCombinerInfo &DCI) const {
9886 SelectionDAG &DAG = DCI.DAG;
9887 switch (N->getOpcode()) {
9889 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9890 case ISD::EXTRACT_VECTOR_ELT:
9891 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9892 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9893 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9894 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9897 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9898 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9899 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9901 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9902 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9903 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9904 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9905 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9906 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9912 /// PerformDAGCombinePromotion - This method query the target whether it is
9913 /// beneficial for dag combiner to promote the specified node. If true, it
9914 /// should return the desired promotion type by reference.
9915 bool X86TargetLowering::PerformDAGCombinePromotion(SDValue Op, EVT &PVT) const {
9919 EVT VT = Op.getValueType();
9923 bool Commute = true;
9924 switch (Op.getOpcode()) {
9925 default: return false;
9934 SDValue N0 = Op.getOperand(0);
9935 SDValue N1 = Op.getOperand(1);
9936 if (!Commute && isa<LoadSDNode>(N1))
9938 // Avoid disabling potential load folding opportunities.
9939 if ((isa<LoadSDNode>(N0) && N0.hasOneUse()) && !isa<ConstantSDNode>(N1))
9941 if ((isa<LoadSDNode>(N1) && N1.hasOneUse()) && !isa<ConstantSDNode>(N0))
9950 //===----------------------------------------------------------------------===//
9951 // X86 Inline Assembly Support
9952 //===----------------------------------------------------------------------===//
9954 static bool LowerToBSwap(CallInst *CI) {
9955 // FIXME: this should verify that we are targetting a 486 or better. If not,
9956 // we will turn this bswap into something that will be lowered to logical ops
9957 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9958 // so don't worry about this.
9960 // Verify this is a simple bswap.
9961 if (CI->getNumOperands() != 2 ||
9962 CI->getType() != CI->getOperand(1)->getType() ||
9963 !CI->getType()->isIntegerTy())
9966 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9967 if (!Ty || Ty->getBitWidth() % 16 != 0)
9970 // Okay, we can do this xform, do so now.
9971 const Type *Tys[] = { Ty };
9972 Module *M = CI->getParent()->getParent()->getParent();
9973 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9975 Value *Op = CI->getOperand(1);
9976 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9978 CI->replaceAllUsesWith(Op);
9979 CI->eraseFromParent();
9983 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9984 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9985 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9987 std::string AsmStr = IA->getAsmString();
9989 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9990 SmallVector<StringRef, 4> AsmPieces;
9991 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9993 switch (AsmPieces.size()) {
9994 default: return false;
9996 AsmStr = AsmPieces[0];
9998 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10001 if (AsmPieces.size() == 2 &&
10002 (AsmPieces[0] == "bswap" ||
10003 AsmPieces[0] == "bswapq" ||
10004 AsmPieces[0] == "bswapl") &&
10005 (AsmPieces[1] == "$0" ||
10006 AsmPieces[1] == "${0:q}")) {
10007 // No need to check constraints, nothing other than the equivalent of
10008 // "=r,0" would be valid here.
10009 return LowerToBSwap(CI);
10011 // rorw $$8, ${0:w} --> llvm.bswap.i16
10012 if (CI->getType()->isIntegerTy(16) &&
10013 AsmPieces.size() == 3 &&
10014 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10015 AsmPieces[1] == "$$8," &&
10016 AsmPieces[2] == "${0:w}" &&
10017 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10019 const std::string &Constraints = IA->getConstraintString();
10020 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10021 std::sort(AsmPieces.begin(), AsmPieces.end());
10022 if (AsmPieces.size() == 4 &&
10023 AsmPieces[0] == "~{cc}" &&
10024 AsmPieces[1] == "~{dirflag}" &&
10025 AsmPieces[2] == "~{flags}" &&
10026 AsmPieces[3] == "~{fpsr}") {
10027 return LowerToBSwap(CI);
10032 if (CI->getType()->isIntegerTy(64) &&
10033 Constraints.size() >= 2 &&
10034 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10035 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10036 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10037 SmallVector<StringRef, 4> Words;
10038 SplitString(AsmPieces[0], Words, " \t");
10039 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10041 SplitString(AsmPieces[1], Words, " \t");
10042 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10044 SplitString(AsmPieces[2], Words, " \t,");
10045 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10046 Words[2] == "%edx") {
10047 return LowerToBSwap(CI);
10059 /// getConstraintType - Given a constraint letter, return the type of
10060 /// constraint it is for this target.
10061 X86TargetLowering::ConstraintType
10062 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10063 if (Constraint.size() == 1) {
10064 switch (Constraint[0]) {
10076 return C_RegisterClass;
10084 return TargetLowering::getConstraintType(Constraint);
10087 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10088 /// with another that has more specific requirements based on the type of the
10089 /// corresponding operand.
10090 const char *X86TargetLowering::
10091 LowerXConstraint(EVT ConstraintVT) const {
10092 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10093 // 'f' like normal targets.
10094 if (ConstraintVT.isFloatingPoint()) {
10095 if (Subtarget->hasSSE2())
10097 if (Subtarget->hasSSE1())
10101 return TargetLowering::LowerXConstraint(ConstraintVT);
10104 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10105 /// vector. If it is invalid, don't add anything to Ops.
10106 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10109 std::vector<SDValue>&Ops,
10110 SelectionDAG &DAG) const {
10111 SDValue Result(0, 0);
10113 switch (Constraint) {
10116 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10117 if (C->getZExtValue() <= 31) {
10118 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10124 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10125 if (C->getZExtValue() <= 63) {
10126 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10132 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10133 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10134 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10140 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10141 if (C->getZExtValue() <= 255) {
10142 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10148 // 32-bit signed value
10149 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10150 const ConstantInt *CI = C->getConstantIntValue();
10151 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10152 C->getSExtValue())) {
10153 // Widen to 64 bits here to get it sign extended.
10154 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10157 // FIXME gcc accepts some relocatable values here too, but only in certain
10158 // memory models; it's complicated.
10163 // 32-bit unsigned value
10164 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10165 const ConstantInt *CI = C->getConstantIntValue();
10166 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10167 C->getZExtValue())) {
10168 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10172 // FIXME gcc accepts some relocatable values here too, but only in certain
10173 // memory models; it's complicated.
10177 // Literal immediates are always ok.
10178 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10179 // Widen to 64 bits here to get it sign extended.
10180 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10184 // If we are in non-pic codegen mode, we allow the address of a global (with
10185 // an optional displacement) to be used with 'i'.
10186 GlobalAddressSDNode *GA = 0;
10187 int64_t Offset = 0;
10189 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10191 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10192 Offset += GA->getOffset();
10194 } else if (Op.getOpcode() == ISD::ADD) {
10195 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10196 Offset += C->getZExtValue();
10197 Op = Op.getOperand(0);
10200 } else if (Op.getOpcode() == ISD::SUB) {
10201 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10202 Offset += -C->getZExtValue();
10203 Op = Op.getOperand(0);
10208 // Otherwise, this isn't something we can handle, reject it.
10212 const GlobalValue *GV = GA->getGlobal();
10213 // If we require an extra load to get this address, as in PIC mode, we
10214 // can't accept it.
10215 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10216 getTargetMachine())))
10220 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
10222 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10228 if (Result.getNode()) {
10229 Ops.push_back(Result);
10232 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
10236 std::vector<unsigned> X86TargetLowering::
10237 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10239 if (Constraint.size() == 1) {
10240 // FIXME: not handling fp-stack yet!
10241 switch (Constraint[0]) { // GCC X86 Constraint Letters
10242 default: break; // Unknown constraint letter
10243 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10244 if (Subtarget->is64Bit()) {
10245 if (VT == MVT::i32)
10246 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10247 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10248 X86::R10D,X86::R11D,X86::R12D,
10249 X86::R13D,X86::R14D,X86::R15D,
10250 X86::EBP, X86::ESP, 0);
10251 else if (VT == MVT::i16)
10252 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10253 X86::SI, X86::DI, X86::R8W,X86::R9W,
10254 X86::R10W,X86::R11W,X86::R12W,
10255 X86::R13W,X86::R14W,X86::R15W,
10256 X86::BP, X86::SP, 0);
10257 else if (VT == MVT::i8)
10258 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10259 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10260 X86::R10B,X86::R11B,X86::R12B,
10261 X86::R13B,X86::R14B,X86::R15B,
10262 X86::BPL, X86::SPL, 0);
10264 else if (VT == MVT::i64)
10265 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10266 X86::RSI, X86::RDI, X86::R8, X86::R9,
10267 X86::R10, X86::R11, X86::R12,
10268 X86::R13, X86::R14, X86::R15,
10269 X86::RBP, X86::RSP, 0);
10273 // 32-bit fallthrough
10274 case 'Q': // Q_REGS
10275 if (VT == MVT::i32)
10276 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10277 else if (VT == MVT::i16)
10278 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10279 else if (VT == MVT::i8)
10280 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10281 else if (VT == MVT::i64)
10282 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10287 return std::vector<unsigned>();
10290 std::pair<unsigned, const TargetRegisterClass*>
10291 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10293 // First, see if this is a constraint that directly corresponds to an LLVM
10295 if (Constraint.size() == 1) {
10296 // GCC Constraint Letters
10297 switch (Constraint[0]) {
10299 case 'r': // GENERAL_REGS
10300 case 'l': // INDEX_REGS
10302 return std::make_pair(0U, X86::GR8RegisterClass);
10303 if (VT == MVT::i16)
10304 return std::make_pair(0U, X86::GR16RegisterClass);
10305 if (VT == MVT::i32 || !Subtarget->is64Bit())
10306 return std::make_pair(0U, X86::GR32RegisterClass);
10307 return std::make_pair(0U, X86::GR64RegisterClass);
10308 case 'R': // LEGACY_REGS
10310 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10311 if (VT == MVT::i16)
10312 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10313 if (VT == MVT::i32 || !Subtarget->is64Bit())
10314 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10315 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10316 case 'f': // FP Stack registers.
10317 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10318 // value to the correct fpstack register class.
10319 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10320 return std::make_pair(0U, X86::RFP32RegisterClass);
10321 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10322 return std::make_pair(0U, X86::RFP64RegisterClass);
10323 return std::make_pair(0U, X86::RFP80RegisterClass);
10324 case 'y': // MMX_REGS if MMX allowed.
10325 if (!Subtarget->hasMMX()) break;
10326 return std::make_pair(0U, X86::VR64RegisterClass);
10327 case 'Y': // SSE_REGS if SSE2 allowed
10328 if (!Subtarget->hasSSE2()) break;
10330 case 'x': // SSE_REGS if SSE1 allowed
10331 if (!Subtarget->hasSSE1()) break;
10333 switch (VT.getSimpleVT().SimpleTy) {
10335 // Scalar SSE types.
10338 return std::make_pair(0U, X86::FR32RegisterClass);
10341 return std::make_pair(0U, X86::FR64RegisterClass);
10349 return std::make_pair(0U, X86::VR128RegisterClass);
10355 // Use the default implementation in TargetLowering to convert the register
10356 // constraint into a member of a register class.
10357 std::pair<unsigned, const TargetRegisterClass*> Res;
10358 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10360 // Not found as a standard register?
10361 if (Res.second == 0) {
10362 // Map st(0) -> st(7) -> ST0
10363 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10364 tolower(Constraint[1]) == 's' &&
10365 tolower(Constraint[2]) == 't' &&
10366 Constraint[3] == '(' &&
10367 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10368 Constraint[5] == ')' &&
10369 Constraint[6] == '}') {
10371 Res.first = X86::ST0+Constraint[4]-'0';
10372 Res.second = X86::RFP80RegisterClass;
10376 // GCC allows "st(0)" to be called just plain "st".
10377 if (StringRef("{st}").equals_lower(Constraint)) {
10378 Res.first = X86::ST0;
10379 Res.second = X86::RFP80RegisterClass;
10384 if (StringRef("{flags}").equals_lower(Constraint)) {
10385 Res.first = X86::EFLAGS;
10386 Res.second = X86::CCRRegisterClass;
10390 // 'A' means EAX + EDX.
10391 if (Constraint == "A") {
10392 Res.first = X86::EAX;
10393 Res.second = X86::GR32_ADRegisterClass;
10399 // Otherwise, check to see if this is a register class of the wrong value
10400 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10401 // turn into {ax},{dx}.
10402 if (Res.second->hasType(VT))
10403 return Res; // Correct type already, nothing to do.
10405 // All of the single-register GCC register classes map their values onto
10406 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10407 // really want an 8-bit or 32-bit register, map to the appropriate register
10408 // class and return the appropriate register.
10409 if (Res.second == X86::GR16RegisterClass) {
10410 if (VT == MVT::i8) {
10411 unsigned DestReg = 0;
10412 switch (Res.first) {
10414 case X86::AX: DestReg = X86::AL; break;
10415 case X86::DX: DestReg = X86::DL; break;
10416 case X86::CX: DestReg = X86::CL; break;
10417 case X86::BX: DestReg = X86::BL; break;
10420 Res.first = DestReg;
10421 Res.second = X86::GR8RegisterClass;
10423 } else if (VT == MVT::i32) {
10424 unsigned DestReg = 0;
10425 switch (Res.first) {
10427 case X86::AX: DestReg = X86::EAX; break;
10428 case X86::DX: DestReg = X86::EDX; break;
10429 case X86::CX: DestReg = X86::ECX; break;
10430 case X86::BX: DestReg = X86::EBX; break;
10431 case X86::SI: DestReg = X86::ESI; break;
10432 case X86::DI: DestReg = X86::EDI; break;
10433 case X86::BP: DestReg = X86::EBP; break;
10434 case X86::SP: DestReg = X86::ESP; break;
10437 Res.first = DestReg;
10438 Res.second = X86::GR32RegisterClass;
10440 } else if (VT == MVT::i64) {
10441 unsigned DestReg = 0;
10442 switch (Res.first) {
10444 case X86::AX: DestReg = X86::RAX; break;
10445 case X86::DX: DestReg = X86::RDX; break;
10446 case X86::CX: DestReg = X86::RCX; break;
10447 case X86::BX: DestReg = X86::RBX; break;
10448 case X86::SI: DestReg = X86::RSI; break;
10449 case X86::DI: DestReg = X86::RDI; break;
10450 case X86::BP: DestReg = X86::RBP; break;
10451 case X86::SP: DestReg = X86::RSP; break;
10454 Res.first = DestReg;
10455 Res.second = X86::GR64RegisterClass;
10458 } else if (Res.second == X86::FR32RegisterClass ||
10459 Res.second == X86::FR64RegisterClass ||
10460 Res.second == X86::VR128RegisterClass) {
10461 // Handle references to XMM physical registers that got mapped into the
10462 // wrong class. This can happen with constraints like {xmm0} where the
10463 // target independent register mapper will just pick the first match it can
10464 // find, ignoring the required type.
10465 if (VT == MVT::f32)
10466 Res.second = X86::FR32RegisterClass;
10467 else if (VT == MVT::f64)
10468 Res.second = X86::FR64RegisterClass;
10469 else if (X86::VR128RegisterClass->hasType(VT))
10470 Res.second = X86::VR128RegisterClass;