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 // Forward declarations.
69 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
72 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
73 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
74 default: llvm_unreachable("unknown subtarget type");
75 case X86Subtarget::isDarwin:
76 if (TM.getSubtarget<X86Subtarget>().is64Bit())
77 return new X8664_MachoTargetObjectFile();
78 return new TargetLoweringObjectFileMachO();
79 case X86Subtarget::isELF:
80 if (TM.getSubtarget<X86Subtarget>().is64Bit())
81 return new X8664_ELFTargetObjectFile(TM);
82 return new X8632_ELFTargetObjectFile(TM);
83 case X86Subtarget::isMingw:
84 case X86Subtarget::isCygwin:
85 case X86Subtarget::isWindows:
86 return new TargetLoweringObjectFileCOFF();
90 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
91 : TargetLowering(TM, createTLOF(TM)) {
92 Subtarget = &TM.getSubtarget<X86Subtarget>();
93 X86ScalarSSEf64 = Subtarget->hasSSE2();
94 X86ScalarSSEf32 = Subtarget->hasSSE1();
95 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
97 RegInfo = TM.getRegisterInfo();
100 // Set up the TargetLowering object.
102 // X86 is weird, it always uses i8 for shift amounts and setcc results.
103 setShiftAmountType(MVT::i8);
104 setBooleanContents(ZeroOrOneBooleanContent);
105 setSchedulingPreference(SchedulingForRegPressure);
106 setStackPointerRegisterToSaveRestore(X86StackPtr);
108 if (Subtarget->isTargetDarwin()) {
109 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
110 setUseUnderscoreSetJmp(false);
111 setUseUnderscoreLongJmp(false);
112 } else if (Subtarget->isTargetMingw()) {
113 // MS runtime is weird: it exports _setjmp, but longjmp!
114 setUseUnderscoreSetJmp(true);
115 setUseUnderscoreLongJmp(false);
117 setUseUnderscoreSetJmp(true);
118 setUseUnderscoreLongJmp(true);
121 // Set up the register classes.
122 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
124 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
125 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
126 if (Subtarget->is64Bit())
127 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
129 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
131 // We don't accept any truncstore of integer registers.
132 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
134 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
135 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
137 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
138 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
139 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
141 // SETOEQ and SETUNE require checking two conditions.
142 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
143 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
144 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
145 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
146 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
147 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
149 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
151 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
152 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
153 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
155 if (Subtarget->is64Bit()) {
156 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
157 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
158 } else if (!UseSoftFloat) {
159 if (X86ScalarSSEf64) {
160 // We have an impenetrably clever algorithm for ui64->double only.
161 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
163 // We have an algorithm for SSE2, and we turn this into a 64-bit
164 // FILD for other targets.
165 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
168 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
170 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
171 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
174 // SSE has no i16 to fp conversion, only i32
175 if (X86ScalarSSEf32) {
176 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
177 // f32 and f64 cases are Legal, f80 case is not
178 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
180 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
181 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
184 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
185 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
188 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
189 // are Legal, f80 is custom lowered.
190 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
191 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
193 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
195 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
196 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
198 if (X86ScalarSSEf32) {
199 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
200 // f32 and f64 cases are Legal, f80 case is not
201 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
203 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
204 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
207 // Handle FP_TO_UINT by promoting the destination to a larger signed
209 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
210 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
211 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
213 if (Subtarget->is64Bit()) {
214 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
215 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
216 } else if (!UseSoftFloat) {
217 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
218 // Expand FP_TO_UINT into a select.
219 // FIXME: We would like to use a Custom expander here eventually to do
220 // the optimal thing for SSE vs. the default expansion in the legalizer.
221 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
223 // With SSE3 we can use fisttpll to convert to a signed i64; without
224 // SSE, we're stuck with a fistpll.
225 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
228 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
229 if (!X86ScalarSSEf64) {
230 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
231 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
234 // Scalar integer divide and remainder are lowered to use operations that
235 // produce two results, to match the available instructions. This exposes
236 // the two-result form to trivial CSE, which is able to combine x/y and x%y
237 // into a single instruction.
239 // Scalar integer multiply-high is also lowered to use two-result
240 // operations, to match the available instructions. However, plain multiply
241 // (low) operations are left as Legal, as there are single-result
242 // instructions for this in x86. Using the two-result multiply instructions
243 // when both high and low results are needed must be arranged by dagcombine.
244 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
245 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
246 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
247 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
248 setOperationAction(ISD::SREM , MVT::i8 , Expand);
249 setOperationAction(ISD::UREM , MVT::i8 , Expand);
250 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
251 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
252 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
253 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
254 setOperationAction(ISD::SREM , MVT::i16 , Expand);
255 setOperationAction(ISD::UREM , MVT::i16 , Expand);
256 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
257 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
258 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
259 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
260 setOperationAction(ISD::SREM , MVT::i32 , Expand);
261 setOperationAction(ISD::UREM , MVT::i32 , Expand);
262 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
263 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
264 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
265 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
266 setOperationAction(ISD::SREM , MVT::i64 , Expand);
267 setOperationAction(ISD::UREM , MVT::i64 , Expand);
269 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
270 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
271 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
272 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
273 if (Subtarget->is64Bit())
274 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
275 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
276 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
277 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
278 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
279 setOperationAction(ISD::FREM , MVT::f32 , Expand);
280 setOperationAction(ISD::FREM , MVT::f64 , Expand);
281 setOperationAction(ISD::FREM , MVT::f80 , Expand);
282 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
287 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
290 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
292 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
293 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
295 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
296 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
297 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
298 if (Subtarget->is64Bit()) {
299 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
300 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
301 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
304 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
305 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
307 // These should be promoted to a larger select which is supported.
308 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
309 // X86 wants to expand cmov itself.
310 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
312 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
314 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
315 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
316 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
317 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
318 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
319 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
321 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
323 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
324 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
325 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
326 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
327 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
328 if (Subtarget->is64Bit()) {
329 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
330 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
332 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
335 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
336 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
337 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
338 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
339 if (Subtarget->is64Bit())
340 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
341 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
342 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
343 if (Subtarget->is64Bit()) {
344 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
345 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
346 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
347 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
348 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
350 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
351 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
352 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
353 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
354 if (Subtarget->is64Bit()) {
355 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
356 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
357 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
360 if (Subtarget->hasSSE1())
361 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
363 if (!Subtarget->hasSSE2())
364 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
366 // Expand certain atomics
367 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
368 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
369 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
370 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
374 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
375 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
377 if (!Subtarget->is64Bit()) {
378 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
379 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
380 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
381 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
382 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
383 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
384 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
387 // FIXME - use subtarget debug flags
388 if (!Subtarget->isTargetDarwin() &&
389 !Subtarget->isTargetELF() &&
390 !Subtarget->isTargetCygMing()) {
391 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
394 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
395 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
396 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
397 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
398 if (Subtarget->is64Bit()) {
399 setExceptionPointerRegister(X86::RAX);
400 setExceptionSelectorRegister(X86::RDX);
402 setExceptionPointerRegister(X86::EAX);
403 setExceptionSelectorRegister(X86::EDX);
405 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
406 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
408 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
410 setOperationAction(ISD::TRAP, MVT::Other, Legal);
412 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
413 setOperationAction(ISD::VASTART , MVT::Other, Custom);
414 setOperationAction(ISD::VAEND , MVT::Other, Expand);
415 if (Subtarget->is64Bit()) {
416 setOperationAction(ISD::VAARG , MVT::Other, Custom);
417 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
419 setOperationAction(ISD::VAARG , MVT::Other, Expand);
420 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
423 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
424 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
425 if (Subtarget->is64Bit())
426 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
427 if (Subtarget->isTargetCygMing())
428 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
430 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
432 if (!UseSoftFloat && X86ScalarSSEf64) {
433 // f32 and f64 use SSE.
434 // Set up the FP register classes.
435 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
436 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
438 // Use ANDPD to simulate FABS.
439 setOperationAction(ISD::FABS , MVT::f64, Custom);
440 setOperationAction(ISD::FABS , MVT::f32, Custom);
442 // Use XORP to simulate FNEG.
443 setOperationAction(ISD::FNEG , MVT::f64, Custom);
444 setOperationAction(ISD::FNEG , MVT::f32, Custom);
446 // Use ANDPD and ORPD to simulate FCOPYSIGN.
447 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
448 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
450 // We don't support sin/cos/fmod
451 setOperationAction(ISD::FSIN , MVT::f64, Expand);
452 setOperationAction(ISD::FCOS , MVT::f64, Expand);
453 setOperationAction(ISD::FSIN , MVT::f32, Expand);
454 setOperationAction(ISD::FCOS , MVT::f32, Expand);
456 // Expand FP immediates into loads from the stack, except for the special
458 addLegalFPImmediate(APFloat(+0.0)); // xorpd
459 addLegalFPImmediate(APFloat(+0.0f)); // xorps
460 } else if (!UseSoftFloat && X86ScalarSSEf32) {
461 // Use SSE for f32, x87 for f64.
462 // Set up the FP register classes.
463 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
464 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
466 // Use ANDPS to simulate FABS.
467 setOperationAction(ISD::FABS , MVT::f32, Custom);
469 // Use XORP to simulate FNEG.
470 setOperationAction(ISD::FNEG , MVT::f32, Custom);
472 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
474 // Use ANDPS and ORPS to simulate FCOPYSIGN.
475 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
476 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
478 // We don't support sin/cos/fmod
479 setOperationAction(ISD::FSIN , MVT::f32, Expand);
480 setOperationAction(ISD::FCOS , MVT::f32, Expand);
482 // Special cases we handle for FP constants.
483 addLegalFPImmediate(APFloat(+0.0f)); // xorps
484 addLegalFPImmediate(APFloat(+0.0)); // FLD0
485 addLegalFPImmediate(APFloat(+1.0)); // FLD1
486 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
487 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
490 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
491 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
493 } else if (!UseSoftFloat) {
494 // f32 and f64 in x87.
495 // Set up the FP register classes.
496 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
497 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
499 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
500 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
501 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
502 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
505 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
506 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
508 addLegalFPImmediate(APFloat(+0.0)); // FLD0
509 addLegalFPImmediate(APFloat(+1.0)); // FLD1
510 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
511 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
512 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
513 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
514 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
515 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
518 // Long double always uses X87.
520 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
521 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
522 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
525 APFloat TmpFlt(+0.0);
526 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
528 addLegalFPImmediate(TmpFlt); // FLD0
530 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
531 APFloat TmpFlt2(+1.0);
532 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
534 addLegalFPImmediate(TmpFlt2); // FLD1
535 TmpFlt2.changeSign();
536 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
540 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
541 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
545 // Always use a library call for pow.
546 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
547 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
548 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
550 setOperationAction(ISD::FLOG, MVT::f80, Expand);
551 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
552 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
553 setOperationAction(ISD::FEXP, MVT::f80, Expand);
554 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
556 // First set operation action for all vector types to either promote
557 // (for widening) or expand (for scalarization). Then we will selectively
558 // turn on ones that can be effectively codegen'd.
559 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
560 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
561 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
576 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
577 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
607 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
608 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
609 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
610 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
611 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
612 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
613 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
614 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
615 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
616 setTruncStoreAction((MVT::SimpleValueType)VT,
617 (MVT::SimpleValueType)InnerVT, Expand);
618 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
619 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
620 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
623 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
624 // with -msoft-float, disable use of MMX as well.
625 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
626 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
627 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
628 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
629 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass, false);
630 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
632 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
633 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
634 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
635 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
637 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
638 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
639 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
640 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
642 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
643 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
645 setOperationAction(ISD::AND, MVT::v8i8, Promote);
646 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
647 setOperationAction(ISD::AND, MVT::v4i16, Promote);
648 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
649 setOperationAction(ISD::AND, MVT::v2i32, Promote);
650 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
651 setOperationAction(ISD::AND, MVT::v1i64, Legal);
653 setOperationAction(ISD::OR, MVT::v8i8, Promote);
654 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
655 setOperationAction(ISD::OR, MVT::v4i16, Promote);
656 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
657 setOperationAction(ISD::OR, MVT::v2i32, Promote);
658 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
659 setOperationAction(ISD::OR, MVT::v1i64, Legal);
661 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
662 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
663 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
664 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
665 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
666 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
667 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
669 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
670 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
671 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
672 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
673 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
674 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
675 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
676 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
677 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
679 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
680 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
681 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
682 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
683 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
685 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
686 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
687 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
688 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
690 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
691 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
692 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
693 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
695 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
697 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
698 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
699 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
700 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
701 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
702 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
703 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
706 if (!UseSoftFloat && Subtarget->hasSSE1()) {
707 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
709 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
710 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
711 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
712 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
713 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
714 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
715 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
716 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
717 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
718 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
719 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
720 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
723 if (!UseSoftFloat && Subtarget->hasSSE2()) {
724 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
726 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
727 // registers cannot be used even for integer operations.
728 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
729 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
730 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
731 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
733 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
734 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
735 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
736 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
737 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
738 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
739 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
740 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
741 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
742 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
743 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
744 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
745 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
746 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
747 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
748 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
750 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
751 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
752 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
753 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
755 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
756 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
757 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
758 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
759 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
761 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
762 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
763 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
764 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
765 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
767 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
768 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
769 EVT VT = (MVT::SimpleValueType)i;
770 // Do not attempt to custom lower non-power-of-2 vectors
771 if (!isPowerOf2_32(VT.getVectorNumElements()))
773 // Do not attempt to custom lower non-128-bit vectors
774 if (!VT.is128BitVector())
776 setOperationAction(ISD::BUILD_VECTOR,
777 VT.getSimpleVT().SimpleTy, Custom);
778 setOperationAction(ISD::VECTOR_SHUFFLE,
779 VT.getSimpleVT().SimpleTy, Custom);
780 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
781 VT.getSimpleVT().SimpleTy, Custom);
784 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
785 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
786 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
787 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
788 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
789 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
791 if (Subtarget->is64Bit()) {
792 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
793 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
796 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
797 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
798 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
801 // Do not attempt to promote non-128-bit vectors
802 if (!VT.is128BitVector()) {
806 setOperationAction(ISD::AND, SVT, Promote);
807 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
808 setOperationAction(ISD::OR, SVT, Promote);
809 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
810 setOperationAction(ISD::XOR, SVT, Promote);
811 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
812 setOperationAction(ISD::LOAD, SVT, Promote);
813 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
814 setOperationAction(ISD::SELECT, SVT, Promote);
815 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
818 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
820 // Custom lower v2i64 and v2f64 selects.
821 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
822 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
823 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
824 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
826 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
827 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
828 if (!DisableMMX && Subtarget->hasMMX()) {
829 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
830 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
834 if (Subtarget->hasSSE41()) {
835 // FIXME: Do we need to handle scalar-to-vector here?
836 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
838 // i8 and i16 vectors are custom , because the source register and source
839 // source memory operand types are not the same width. f32 vectors are
840 // custom since the immediate controlling the insert encodes additional
842 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
843 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
844 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
847 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
848 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
849 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
852 if (Subtarget->is64Bit()) {
853 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
854 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
858 if (Subtarget->hasSSE42()) {
859 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
862 if (!UseSoftFloat && Subtarget->hasAVX()) {
863 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
864 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
865 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
866 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
868 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
869 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
870 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
871 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
872 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
873 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
874 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
875 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
876 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
877 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
878 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
879 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
880 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
881 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
882 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
884 // Operations to consider commented out -v16i16 v32i8
885 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
886 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
887 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
888 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
889 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
890 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
891 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
892 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
893 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
894 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
895 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
896 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
897 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
898 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
900 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
901 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
902 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
903 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
905 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
906 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
907 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
908 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
909 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
911 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
912 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
913 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
914 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
915 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
916 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
919 // Not sure we want to do this since there are no 256-bit integer
922 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
923 // This includes 256-bit vectors
924 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
925 EVT VT = (MVT::SimpleValueType)i;
927 // Do not attempt to custom lower non-power-of-2 vectors
928 if (!isPowerOf2_32(VT.getVectorNumElements()))
931 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
932 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
933 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
936 if (Subtarget->is64Bit()) {
937 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
938 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
943 // Not sure we want to do this since there are no 256-bit integer
946 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
947 // Including 256-bit vectors
948 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
949 EVT VT = (MVT::SimpleValueType)i;
951 if (!VT.is256BitVector()) {
954 setOperationAction(ISD::AND, VT, Promote);
955 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
956 setOperationAction(ISD::OR, VT, Promote);
957 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
958 setOperationAction(ISD::XOR, VT, Promote);
959 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
960 setOperationAction(ISD::LOAD, VT, Promote);
961 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
962 setOperationAction(ISD::SELECT, VT, Promote);
963 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
966 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
970 // We want to custom lower some of our intrinsics.
971 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
973 // Add/Sub/Mul with overflow operations are custom lowered.
974 setOperationAction(ISD::SADDO, MVT::i32, Custom);
975 setOperationAction(ISD::SADDO, MVT::i64, Custom);
976 setOperationAction(ISD::UADDO, MVT::i32, Custom);
977 setOperationAction(ISD::UADDO, MVT::i64, Custom);
978 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
979 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
980 setOperationAction(ISD::USUBO, MVT::i32, Custom);
981 setOperationAction(ISD::USUBO, MVT::i64, Custom);
982 setOperationAction(ISD::SMULO, MVT::i32, Custom);
983 setOperationAction(ISD::SMULO, MVT::i64, Custom);
985 if (!Subtarget->is64Bit()) {
986 // These libcalls are not available in 32-bit.
987 setLibcallName(RTLIB::SHL_I128, 0);
988 setLibcallName(RTLIB::SRL_I128, 0);
989 setLibcallName(RTLIB::SRA_I128, 0);
992 // We have target-specific dag combine patterns for the following nodes:
993 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
994 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
995 setTargetDAGCombine(ISD::BUILD_VECTOR);
996 setTargetDAGCombine(ISD::SELECT);
997 setTargetDAGCombine(ISD::SHL);
998 setTargetDAGCombine(ISD::SRA);
999 setTargetDAGCombine(ISD::SRL);
1000 setTargetDAGCombine(ISD::OR);
1001 setTargetDAGCombine(ISD::STORE);
1002 setTargetDAGCombine(ISD::MEMBARRIER);
1003 setTargetDAGCombine(ISD::ZERO_EXTEND);
1004 if (Subtarget->is64Bit())
1005 setTargetDAGCombine(ISD::MUL);
1007 computeRegisterProperties();
1009 // FIXME: These should be based on subtarget info. Plus, the values should
1010 // be smaller when we are in optimizing for size mode.
1011 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1012 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1013 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1014 setPrefLoopAlignment(16);
1015 benefitFromCodePlacementOpt = true;
1019 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1024 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1025 /// the desired ByVal argument alignment.
1026 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1029 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1030 if (VTy->getBitWidth() == 128)
1032 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1033 unsigned EltAlign = 0;
1034 getMaxByValAlign(ATy->getElementType(), EltAlign);
1035 if (EltAlign > MaxAlign)
1036 MaxAlign = EltAlign;
1037 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1038 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1039 unsigned EltAlign = 0;
1040 getMaxByValAlign(STy->getElementType(i), EltAlign);
1041 if (EltAlign > MaxAlign)
1042 MaxAlign = EltAlign;
1050 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1051 /// function arguments in the caller parameter area. For X86, aggregates
1052 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1053 /// are at 4-byte boundaries.
1054 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1055 if (Subtarget->is64Bit()) {
1056 // Max of 8 and alignment of type.
1057 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1064 if (Subtarget->hasSSE1())
1065 getMaxByValAlign(Ty, Align);
1069 /// getOptimalMemOpType - Returns the target specific optimal type for load
1070 /// and store operations as a result of memset, memcpy, and memmove
1071 /// lowering. If DstAlign is zero that means it's safe to destination
1072 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1073 /// means there isn't a need to check it against alignment requirement,
1074 /// probably because the source does not need to be loaded. If
1075 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1076 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1077 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1078 /// constant so it does not need to be loaded.
1079 /// It returns EVT::Other if the type should be determined using generic
1080 /// target-independent logic.
1082 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1083 unsigned DstAlign, unsigned SrcAlign,
1084 bool NonScalarIntSafe,
1086 MachineFunction &MF) const {
1087 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1088 // linux. This is because the stack realignment code can't handle certain
1089 // cases like PR2962. This should be removed when PR2962 is fixed.
1090 const Function *F = MF.getFunction();
1091 if (NonScalarIntSafe &&
1092 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1094 (Subtarget->isUnalignedMemAccessFast() ||
1095 ((DstAlign == 0 || DstAlign >= 16) &&
1096 (SrcAlign == 0 || SrcAlign >= 16))) &&
1097 Subtarget->getStackAlignment() >= 16) {
1098 if (Subtarget->hasSSE2())
1100 if (Subtarget->hasSSE1())
1102 } else if (!MemcpyStrSrc && Size >= 8 &&
1103 !Subtarget->is64Bit() &&
1104 Subtarget->getStackAlignment() >= 8 &&
1105 Subtarget->hasSSE2()) {
1106 // Do not use f64 to lower memcpy if source is string constant. It's
1107 // better to use i32 to avoid the loads.
1111 if (Subtarget->is64Bit() && Size >= 8)
1116 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1117 /// current function. The returned value is a member of the
1118 /// MachineJumpTableInfo::JTEntryKind enum.
1119 unsigned X86TargetLowering::getJumpTableEncoding() const {
1120 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1122 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1123 Subtarget->isPICStyleGOT())
1124 return MachineJumpTableInfo::EK_Custom32;
1126 // Otherwise, use the normal jump table encoding heuristics.
1127 return TargetLowering::getJumpTableEncoding();
1130 /// getPICBaseSymbol - Return the X86-32 PIC base.
1132 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1133 MCContext &Ctx) const {
1134 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1135 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1136 Twine(MF->getFunctionNumber())+"$pb");
1141 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1142 const MachineBasicBlock *MBB,
1143 unsigned uid,MCContext &Ctx) const{
1144 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1145 Subtarget->isPICStyleGOT());
1146 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1148 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1149 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1152 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1154 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1155 SelectionDAG &DAG) const {
1156 if (!Subtarget->is64Bit())
1157 // This doesn't have DebugLoc associated with it, but is not really the
1158 // same as a Register.
1159 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1163 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1164 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1166 const MCExpr *X86TargetLowering::
1167 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1168 MCContext &Ctx) const {
1169 // X86-64 uses RIP relative addressing based on the jump table label.
1170 if (Subtarget->isPICStyleRIPRel())
1171 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1173 // Otherwise, the reference is relative to the PIC base.
1174 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1177 /// getFunctionAlignment - Return the Log2 alignment of this function.
1178 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1179 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1182 //===----------------------------------------------------------------------===//
1183 // Return Value Calling Convention Implementation
1184 //===----------------------------------------------------------------------===//
1186 #include "X86GenCallingConv.inc"
1189 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1190 const SmallVectorImpl<EVT> &OutTys,
1191 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1192 SelectionDAG &DAG) const {
1193 SmallVector<CCValAssign, 16> RVLocs;
1194 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1195 RVLocs, *DAG.getContext());
1196 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1200 X86TargetLowering::LowerReturn(SDValue Chain,
1201 CallingConv::ID CallConv, bool isVarArg,
1202 const SmallVectorImpl<ISD::OutputArg> &Outs,
1203 DebugLoc dl, SelectionDAG &DAG) const {
1204 MachineFunction &MF = DAG.getMachineFunction();
1205 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1207 SmallVector<CCValAssign, 16> RVLocs;
1208 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1209 RVLocs, *DAG.getContext());
1210 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1212 // Add the regs to the liveout set for the function.
1213 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1214 for (unsigned i = 0; i != RVLocs.size(); ++i)
1215 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1216 MRI.addLiveOut(RVLocs[i].getLocReg());
1220 SmallVector<SDValue, 6> RetOps;
1221 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1222 // Operand #1 = Bytes To Pop
1223 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
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) const {
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,
1398 CallingConv::ID CallingConv) const {
1402 switch (CallingConv) {
1405 case CallingConv::X86_StdCall:
1406 return !Subtarget->is64Bit();
1407 case CallingConv::X86_FastCall:
1408 return !Subtarget->is64Bit();
1409 case CallingConv::Fast:
1410 return GuaranteedTailCallOpt;
1411 case CallingConv::GHC:
1412 return GuaranteedTailCallOpt;
1416 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1417 /// given CallingConvention value.
1418 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1419 if (Subtarget->is64Bit()) {
1420 if (CC == CallingConv::GHC)
1421 return CC_X86_64_GHC;
1422 else if (Subtarget->isTargetWin64())
1423 return CC_X86_Win64_C;
1428 if (CC == CallingConv::X86_FastCall)
1429 return CC_X86_32_FastCall;
1430 else if (CC == CallingConv::Fast)
1431 return CC_X86_32_FastCC;
1432 else if (CC == CallingConv::GHC)
1433 return CC_X86_32_GHC;
1438 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1439 /// by "Src" to address "Dst" with size and alignment information specified by
1440 /// the specific parameter attribute. The copy will be passed as a byval
1441 /// function parameter.
1443 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1444 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1446 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1447 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1448 /*isVolatile*/false, /*AlwaysInline=*/true,
1452 /// IsTailCallConvention - Return true if the calling convention is one that
1453 /// supports tail call optimization.
1454 static bool IsTailCallConvention(CallingConv::ID CC) {
1455 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1458 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1459 /// a tailcall target by changing its ABI.
1460 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1461 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1465 X86TargetLowering::LowerMemArgument(SDValue Chain,
1466 CallingConv::ID CallConv,
1467 const SmallVectorImpl<ISD::InputArg> &Ins,
1468 DebugLoc dl, SelectionDAG &DAG,
1469 const CCValAssign &VA,
1470 MachineFrameInfo *MFI,
1472 // Create the nodes corresponding to a load from this parameter slot.
1473 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1474 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1475 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1478 // If value is passed by pointer we have address passed instead of the value
1480 if (VA.getLocInfo() == CCValAssign::Indirect)
1481 ValVT = VA.getLocVT();
1483 ValVT = VA.getValVT();
1485 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1486 // changed with more analysis.
1487 // In case of tail call optimization mark all arguments mutable. Since they
1488 // could be overwritten by lowering of arguments in case of a tail call.
1489 if (Flags.isByVal()) {
1490 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1491 VA.getLocMemOffset(), isImmutable, false);
1492 return DAG.getFrameIndex(FI, getPointerTy());
1494 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1495 VA.getLocMemOffset(), isImmutable, false);
1496 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1497 return DAG.getLoad(ValVT, dl, Chain, FIN,
1498 PseudoSourceValue::getFixedStack(FI), 0,
1504 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1505 CallingConv::ID CallConv,
1507 const SmallVectorImpl<ISD::InputArg> &Ins,
1510 SmallVectorImpl<SDValue> &InVals)
1512 MachineFunction &MF = DAG.getMachineFunction();
1513 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1515 const Function* Fn = MF.getFunction();
1516 if (Fn->hasExternalLinkage() &&
1517 Subtarget->isTargetCygMing() &&
1518 Fn->getName() == "main")
1519 FuncInfo->setForceFramePointer(true);
1521 MachineFrameInfo *MFI = MF.getFrameInfo();
1522 bool Is64Bit = Subtarget->is64Bit();
1523 bool IsWin64 = Subtarget->isTargetWin64();
1525 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1526 "Var args not supported with calling convention fastcc or ghc");
1528 // Assign locations to all of the incoming arguments.
1529 SmallVector<CCValAssign, 16> ArgLocs;
1530 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1531 ArgLocs, *DAG.getContext());
1532 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1534 unsigned LastVal = ~0U;
1536 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1537 CCValAssign &VA = ArgLocs[i];
1538 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1540 assert(VA.getValNo() != LastVal &&
1541 "Don't support value assigned to multiple locs yet");
1542 LastVal = VA.getValNo();
1544 if (VA.isRegLoc()) {
1545 EVT RegVT = VA.getLocVT();
1546 TargetRegisterClass *RC = NULL;
1547 if (RegVT == MVT::i32)
1548 RC = X86::GR32RegisterClass;
1549 else if (Is64Bit && RegVT == MVT::i64)
1550 RC = X86::GR64RegisterClass;
1551 else if (RegVT == MVT::f32)
1552 RC = X86::FR32RegisterClass;
1553 else if (RegVT == MVT::f64)
1554 RC = X86::FR64RegisterClass;
1555 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1556 RC = X86::VR128RegisterClass;
1557 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1558 RC = X86::VR64RegisterClass;
1560 llvm_unreachable("Unknown argument type!");
1562 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1563 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1565 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1566 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1568 if (VA.getLocInfo() == CCValAssign::SExt)
1569 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1570 DAG.getValueType(VA.getValVT()));
1571 else if (VA.getLocInfo() == CCValAssign::ZExt)
1572 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1573 DAG.getValueType(VA.getValVT()));
1574 else if (VA.getLocInfo() == CCValAssign::BCvt)
1575 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1577 if (VA.isExtInLoc()) {
1578 // Handle MMX values passed in XMM regs.
1579 if (RegVT.isVector()) {
1580 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1581 ArgValue, DAG.getConstant(0, MVT::i64));
1582 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1584 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1587 assert(VA.isMemLoc());
1588 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1591 // If value is passed via pointer - do a load.
1592 if (VA.getLocInfo() == CCValAssign::Indirect)
1593 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1596 InVals.push_back(ArgValue);
1599 // The x86-64 ABI for returning structs by value requires that we copy
1600 // the sret argument into %rax for the return. Save the argument into
1601 // a virtual register so that we can access it from the return points.
1602 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1603 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1604 unsigned Reg = FuncInfo->getSRetReturnReg();
1606 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1607 FuncInfo->setSRetReturnReg(Reg);
1609 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1610 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1613 unsigned StackSize = CCInfo.getNextStackOffset();
1614 // Align stack specially for tail calls.
1615 if (FuncIsMadeTailCallSafe(CallConv))
1616 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1618 // If the function takes variable number of arguments, make a frame index for
1619 // the start of the first vararg value... for expansion of llvm.va_start.
1621 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1622 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,
1626 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1628 // FIXME: We should really autogenerate these arrays
1629 static const unsigned GPR64ArgRegsWin64[] = {
1630 X86::RCX, X86::RDX, X86::R8, X86::R9
1632 static const unsigned XMMArgRegsWin64[] = {
1633 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1635 static const unsigned GPR64ArgRegs64Bit[] = {
1636 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1638 static const unsigned XMMArgRegs64Bit[] = {
1639 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1640 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1642 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1645 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1646 GPR64ArgRegs = GPR64ArgRegsWin64;
1647 XMMArgRegs = XMMArgRegsWin64;
1649 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1650 GPR64ArgRegs = GPR64ArgRegs64Bit;
1651 XMMArgRegs = XMMArgRegs64Bit;
1653 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1655 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1658 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1659 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1660 "SSE register cannot be used when SSE is disabled!");
1661 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1662 "SSE register cannot be used when SSE is disabled!");
1663 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1664 // Kernel mode asks for SSE to be disabled, so don't push them
1666 TotalNumXMMRegs = 0;
1668 // For X86-64, if there are vararg parameters that are passed via
1669 // registers, then we must store them to their spots on the stack so they
1670 // may be loaded by deferencing the result of va_next.
1671 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1672 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1673 FuncInfo->setRegSaveFrameIndex(
1674 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1677 // Store the integer parameter registers.
1678 SmallVector<SDValue, 8> MemOps;
1679 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1681 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1682 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1683 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1684 DAG.getIntPtrConstant(Offset));
1685 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1686 X86::GR64RegisterClass);
1687 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1689 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1690 PseudoSourceValue::getFixedStack(
1691 FuncInfo->getRegSaveFrameIndex()),
1692 Offset, false, false, 0);
1693 MemOps.push_back(Store);
1697 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1698 // Now store the XMM (fp + vector) parameter registers.
1699 SmallVector<SDValue, 11> SaveXMMOps;
1700 SaveXMMOps.push_back(Chain);
1702 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1703 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1704 SaveXMMOps.push_back(ALVal);
1706 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1707 FuncInfo->getRegSaveFrameIndex()));
1708 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1709 FuncInfo->getVarArgsFPOffset()));
1711 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1712 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1713 X86::VR128RegisterClass);
1714 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1715 SaveXMMOps.push_back(Val);
1717 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1719 &SaveXMMOps[0], SaveXMMOps.size()));
1722 if (!MemOps.empty())
1723 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1724 &MemOps[0], MemOps.size());
1728 // Some CCs need callee pop.
1729 if (IsCalleePop(isVarArg, CallConv)) {
1730 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1732 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1733 // If this is an sret function, the return should pop the hidden pointer.
1734 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1735 FuncInfo->setBytesToPopOnReturn(4);
1739 // RegSaveFrameIndex is X86-64 only.
1740 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1741 if (CallConv == CallingConv::X86_FastCall)
1742 // fastcc functions can't have varargs.
1743 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1750 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1751 SDValue StackPtr, SDValue Arg,
1752 DebugLoc dl, SelectionDAG &DAG,
1753 const CCValAssign &VA,
1754 ISD::ArgFlagsTy Flags) const {
1755 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1756 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1757 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1758 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1759 if (Flags.isByVal()) {
1760 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1762 return DAG.getStore(Chain, dl, Arg, PtrOff,
1763 PseudoSourceValue::getStack(), LocMemOffset,
1767 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1768 /// optimization is performed and it is required.
1770 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1771 SDValue &OutRetAddr, SDValue Chain,
1772 bool IsTailCall, bool Is64Bit,
1773 int FPDiff, DebugLoc dl) const {
1774 // Adjust the Return address stack slot.
1775 EVT VT = getPointerTy();
1776 OutRetAddr = getReturnAddressFrameIndex(DAG);
1778 // Load the "old" Return address.
1779 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1780 return SDValue(OutRetAddr.getNode(), 1);
1783 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1784 /// optimization is performed and it is required (FPDiff!=0).
1786 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1787 SDValue Chain, SDValue RetAddrFrIdx,
1788 bool Is64Bit, int FPDiff, DebugLoc dl) {
1789 // Store the return address to the appropriate stack slot.
1790 if (!FPDiff) return Chain;
1791 // Calculate the new stack slot for the return address.
1792 int SlotSize = Is64Bit ? 8 : 4;
1793 int NewReturnAddrFI =
1794 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1795 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1796 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1797 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1798 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1804 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1805 CallingConv::ID CallConv, bool isVarArg,
1807 const SmallVectorImpl<ISD::OutputArg> &Outs,
1808 const SmallVectorImpl<ISD::InputArg> &Ins,
1809 DebugLoc dl, SelectionDAG &DAG,
1810 SmallVectorImpl<SDValue> &InVals) const {
1811 MachineFunction &MF = DAG.getMachineFunction();
1812 bool Is64Bit = Subtarget->is64Bit();
1813 bool IsStructRet = CallIsStructReturn(Outs);
1814 bool IsSibcall = false;
1817 // Check if it's really possible to do a tail call.
1818 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1819 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1822 // Sibcalls are automatically detected tailcalls which do not require
1824 if (!GuaranteedTailCallOpt && isTailCall)
1831 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1832 "Var args not supported with calling convention fastcc or ghc");
1834 // Analyze operands of the call, assigning locations to each operand.
1835 SmallVector<CCValAssign, 16> ArgLocs;
1836 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1837 ArgLocs, *DAG.getContext());
1838 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1840 // Get a count of how many bytes are to be pushed on the stack.
1841 unsigned NumBytes = CCInfo.getNextStackOffset();
1843 // This is a sibcall. The memory operands are available in caller's
1844 // own caller's stack.
1846 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1847 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1850 if (isTailCall && !IsSibcall) {
1851 // Lower arguments at fp - stackoffset + fpdiff.
1852 unsigned NumBytesCallerPushed =
1853 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1854 FPDiff = NumBytesCallerPushed - NumBytes;
1856 // Set the delta of movement of the returnaddr stackslot.
1857 // But only set if delta is greater than previous delta.
1858 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1859 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1863 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1865 SDValue RetAddrFrIdx;
1866 // Load return adress for tail calls.
1867 if (isTailCall && FPDiff)
1868 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1869 Is64Bit, FPDiff, dl);
1871 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1872 SmallVector<SDValue, 8> MemOpChains;
1875 // Walk the register/memloc assignments, inserting copies/loads. In the case
1876 // of tail call optimization arguments are handle later.
1877 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1878 CCValAssign &VA = ArgLocs[i];
1879 EVT RegVT = VA.getLocVT();
1880 SDValue Arg = Outs[i].Val;
1881 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1882 bool isByVal = Flags.isByVal();
1884 // Promote the value if needed.
1885 switch (VA.getLocInfo()) {
1886 default: llvm_unreachable("Unknown loc info!");
1887 case CCValAssign::Full: break;
1888 case CCValAssign::SExt:
1889 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1891 case CCValAssign::ZExt:
1892 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1894 case CCValAssign::AExt:
1895 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1896 // Special case: passing MMX values in XMM registers.
1897 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1898 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1899 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1901 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1903 case CCValAssign::BCvt:
1904 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1906 case CCValAssign::Indirect: {
1907 // Store the argument.
1908 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1909 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1910 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1911 PseudoSourceValue::getFixedStack(FI), 0,
1918 if (VA.isRegLoc()) {
1919 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1920 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1921 assert(VA.isMemLoc());
1922 if (StackPtr.getNode() == 0)
1923 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1924 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1925 dl, DAG, VA, Flags));
1929 if (!MemOpChains.empty())
1930 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1931 &MemOpChains[0], MemOpChains.size());
1933 // Build a sequence of copy-to-reg nodes chained together with token chain
1934 // and flag operands which copy the outgoing args into registers.
1936 // Tail call byval lowering might overwrite argument registers so in case of
1937 // tail call optimization the copies to registers are lowered later.
1939 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1940 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1941 RegsToPass[i].second, InFlag);
1942 InFlag = Chain.getValue(1);
1945 if (Subtarget->isPICStyleGOT()) {
1946 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1949 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1950 DAG.getNode(X86ISD::GlobalBaseReg,
1951 DebugLoc(), getPointerTy()),
1953 InFlag = Chain.getValue(1);
1955 // If we are tail calling and generating PIC/GOT style code load the
1956 // address of the callee into ECX. The value in ecx is used as target of
1957 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1958 // for tail calls on PIC/GOT architectures. Normally we would just put the
1959 // address of GOT into ebx and then call target@PLT. But for tail calls
1960 // ebx would be restored (since ebx is callee saved) before jumping to the
1963 // Note: The actual moving to ECX is done further down.
1964 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1965 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1966 !G->getGlobal()->hasProtectedVisibility())
1967 Callee = LowerGlobalAddress(Callee, DAG);
1968 else if (isa<ExternalSymbolSDNode>(Callee))
1969 Callee = LowerExternalSymbol(Callee, DAG);
1973 if (Is64Bit && isVarArg) {
1974 // From AMD64 ABI document:
1975 // For calls that may call functions that use varargs or stdargs
1976 // (prototype-less calls or calls to functions containing ellipsis (...) in
1977 // the declaration) %al is used as hidden argument to specify the number
1978 // of SSE registers used. The contents of %al do not need to match exactly
1979 // the number of registers, but must be an ubound on the number of SSE
1980 // registers used and is in the range 0 - 8 inclusive.
1982 // FIXME: Verify this on Win64
1983 // Count the number of XMM registers allocated.
1984 static const unsigned XMMArgRegs[] = {
1985 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1986 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1988 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1989 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1990 && "SSE registers cannot be used when SSE is disabled");
1992 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1993 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1994 InFlag = Chain.getValue(1);
1998 // For tail calls lower the arguments to the 'real' stack slot.
2000 // Force all the incoming stack arguments to be loaded from the stack
2001 // before any new outgoing arguments are stored to the stack, because the
2002 // outgoing stack slots may alias the incoming argument stack slots, and
2003 // the alias isn't otherwise explicit. This is slightly more conservative
2004 // than necessary, because it means that each store effectively depends
2005 // on every argument instead of just those arguments it would clobber.
2006 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2008 SmallVector<SDValue, 8> MemOpChains2;
2011 // Do not flag preceeding copytoreg stuff together with the following stuff.
2013 if (GuaranteedTailCallOpt) {
2014 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2015 CCValAssign &VA = ArgLocs[i];
2018 assert(VA.isMemLoc());
2019 SDValue Arg = Outs[i].Val;
2020 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2021 // Create frame index.
2022 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2023 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2024 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
2025 FIN = DAG.getFrameIndex(FI, getPointerTy());
2027 if (Flags.isByVal()) {
2028 // Copy relative to framepointer.
2029 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2030 if (StackPtr.getNode() == 0)
2031 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2033 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2035 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2039 // Store relative to framepointer.
2040 MemOpChains2.push_back(
2041 DAG.getStore(ArgChain, dl, Arg, FIN,
2042 PseudoSourceValue::getFixedStack(FI), 0,
2048 if (!MemOpChains2.empty())
2049 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2050 &MemOpChains2[0], MemOpChains2.size());
2052 // Copy arguments to their registers.
2053 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2054 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2055 RegsToPass[i].second, InFlag);
2056 InFlag = Chain.getValue(1);
2060 // Store the return address to the appropriate stack slot.
2061 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2065 bool WasGlobalOrExternal = false;
2066 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2067 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2068 // In the 64-bit large code model, we have to make all calls
2069 // through a register, since the call instruction's 32-bit
2070 // pc-relative offset may not be large enough to hold the whole
2072 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2073 WasGlobalOrExternal = true;
2074 // If the callee is a GlobalAddress node (quite common, every direct call
2075 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2078 // We should use extra load for direct calls to dllimported functions in
2080 const GlobalValue *GV = G->getGlobal();
2081 if (!GV->hasDLLImportLinkage()) {
2082 unsigned char OpFlags = 0;
2084 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2085 // external symbols most go through the PLT in PIC mode. If the symbol
2086 // has hidden or protected visibility, or if it is static or local, then
2087 // we don't need to use the PLT - we can directly call it.
2088 if (Subtarget->isTargetELF() &&
2089 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2090 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2091 OpFlags = X86II::MO_PLT;
2092 } else if (Subtarget->isPICStyleStubAny() &&
2093 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2094 Subtarget->getDarwinVers() < 9) {
2095 // PC-relative references to external symbols should go through $stub,
2096 // unless we're building with the leopard linker or later, which
2097 // automatically synthesizes these stubs.
2098 OpFlags = X86II::MO_DARWIN_STUB;
2101 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2102 G->getOffset(), OpFlags);
2104 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2105 WasGlobalOrExternal = true;
2106 unsigned char OpFlags = 0;
2108 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2109 // symbols should go through the PLT.
2110 if (Subtarget->isTargetELF() &&
2111 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2112 OpFlags = X86II::MO_PLT;
2113 } else if (Subtarget->isPICStyleStubAny() &&
2114 Subtarget->getDarwinVers() < 9) {
2115 // PC-relative references to external symbols should go through $stub,
2116 // unless we're building with the leopard linker or later, which
2117 // automatically synthesizes these stubs.
2118 OpFlags = X86II::MO_DARWIN_STUB;
2121 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2125 // Returns a chain & a flag for retval copy to use.
2126 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2127 SmallVector<SDValue, 8> Ops;
2129 if (!IsSibcall && isTailCall) {
2130 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2131 DAG.getIntPtrConstant(0, true), InFlag);
2132 InFlag = Chain.getValue(1);
2135 Ops.push_back(Chain);
2136 Ops.push_back(Callee);
2139 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2141 // Add argument registers to the end of the list so that they are known live
2143 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2144 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2145 RegsToPass[i].second.getValueType()));
2147 // Add an implicit use GOT pointer in EBX.
2148 if (!isTailCall && Subtarget->isPICStyleGOT())
2149 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2151 // Add an implicit use of AL for x86 vararg functions.
2152 if (Is64Bit && isVarArg)
2153 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2155 if (InFlag.getNode())
2156 Ops.push_back(InFlag);
2159 // If this is the first return lowered for this function, add the regs
2160 // to the liveout set for the function.
2161 if (MF.getRegInfo().liveout_empty()) {
2162 SmallVector<CCValAssign, 16> RVLocs;
2163 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2165 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2166 for (unsigned i = 0; i != RVLocs.size(); ++i)
2167 if (RVLocs[i].isRegLoc())
2168 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2170 return DAG.getNode(X86ISD::TC_RETURN, dl,
2171 NodeTys, &Ops[0], Ops.size());
2174 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2175 InFlag = Chain.getValue(1);
2177 // Create the CALLSEQ_END node.
2178 unsigned NumBytesForCalleeToPush;
2179 if (IsCalleePop(isVarArg, CallConv))
2180 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2181 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2182 // If this is a call to a struct-return function, the callee
2183 // pops the hidden struct pointer, so we have to push it back.
2184 // This is common for Darwin/X86, Linux & Mingw32 targets.
2185 NumBytesForCalleeToPush = 4;
2187 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2189 // Returns a flag for retval copy to use.
2191 Chain = DAG.getCALLSEQ_END(Chain,
2192 DAG.getIntPtrConstant(NumBytes, true),
2193 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2196 InFlag = Chain.getValue(1);
2199 // Handle result values, copying them out of physregs into vregs that we
2201 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2202 Ins, dl, DAG, InVals);
2206 //===----------------------------------------------------------------------===//
2207 // Fast Calling Convention (tail call) implementation
2208 //===----------------------------------------------------------------------===//
2210 // Like std call, callee cleans arguments, convention except that ECX is
2211 // reserved for storing the tail called function address. Only 2 registers are
2212 // free for argument passing (inreg). Tail call optimization is performed
2214 // * tailcallopt is enabled
2215 // * caller/callee are fastcc
2216 // On X86_64 architecture with GOT-style position independent code only local
2217 // (within module) calls are supported at the moment.
2218 // To keep the stack aligned according to platform abi the function
2219 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2220 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2221 // If a tail called function callee has more arguments than the caller the
2222 // caller needs to make sure that there is room to move the RETADDR to. This is
2223 // achieved by reserving an area the size of the argument delta right after the
2224 // original REtADDR, but before the saved framepointer or the spilled registers
2225 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2237 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2238 /// for a 16 byte align requirement.
2240 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2241 SelectionDAG& DAG) const {
2242 MachineFunction &MF = DAG.getMachineFunction();
2243 const TargetMachine &TM = MF.getTarget();
2244 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2245 unsigned StackAlignment = TFI.getStackAlignment();
2246 uint64_t AlignMask = StackAlignment - 1;
2247 int64_t Offset = StackSize;
2248 uint64_t SlotSize = TD->getPointerSize();
2249 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2250 // Number smaller than 12 so just add the difference.
2251 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2253 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2254 Offset = ((~AlignMask) & Offset) + StackAlignment +
2255 (StackAlignment-SlotSize);
2260 /// MatchingStackOffset - Return true if the given stack call argument is
2261 /// already available in the same position (relatively) of the caller's
2262 /// incoming argument stack.
2264 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2265 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2266 const X86InstrInfo *TII) {
2267 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2269 if (Arg.getOpcode() == ISD::CopyFromReg) {
2270 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2271 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2273 MachineInstr *Def = MRI->getVRegDef(VR);
2276 if (!Flags.isByVal()) {
2277 if (!TII->isLoadFromStackSlot(Def, FI))
2280 unsigned Opcode = Def->getOpcode();
2281 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2282 Def->getOperand(1).isFI()) {
2283 FI = Def->getOperand(1).getIndex();
2284 Bytes = Flags.getByValSize();
2288 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2289 if (Flags.isByVal())
2290 // ByVal argument is passed in as a pointer but it's now being
2291 // dereferenced. e.g.
2292 // define @foo(%struct.X* %A) {
2293 // tail call @bar(%struct.X* byval %A)
2296 SDValue Ptr = Ld->getBasePtr();
2297 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2300 FI = FINode->getIndex();
2304 assert(FI != INT_MAX);
2305 if (!MFI->isFixedObjectIndex(FI))
2307 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2310 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2311 /// for tail call optimization. Targets which want to do tail call
2312 /// optimization should implement this function.
2314 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2315 CallingConv::ID CalleeCC,
2317 bool isCalleeStructRet,
2318 bool isCallerStructRet,
2319 const SmallVectorImpl<ISD::OutputArg> &Outs,
2320 const SmallVectorImpl<ISD::InputArg> &Ins,
2321 SelectionDAG& DAG) const {
2322 if (!IsTailCallConvention(CalleeCC) &&
2323 CalleeCC != CallingConv::C)
2326 // If -tailcallopt is specified, make fastcc functions tail-callable.
2327 const MachineFunction &MF = DAG.getMachineFunction();
2328 const Function *CallerF = DAG.getMachineFunction().getFunction();
2329 if (GuaranteedTailCallOpt) {
2330 if (IsTailCallConvention(CalleeCC) &&
2331 CallerF->getCallingConv() == CalleeCC)
2336 // Look for obvious safe cases to perform tail call optimization that does not
2337 // requite ABI changes. This is what gcc calls sibcall.
2339 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2340 // emit a special epilogue.
2341 if (RegInfo->needsStackRealignment(MF))
2344 // Do not sibcall optimize vararg calls unless the call site is not passing any
2346 if (isVarArg && !Outs.empty())
2349 // Also avoid sibcall optimization if either caller or callee uses struct
2350 // return semantics.
2351 if (isCalleeStructRet || isCallerStructRet)
2354 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2355 // Therefore if it's not used by the call it is not safe to optimize this into
2357 bool Unused = false;
2358 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2365 SmallVector<CCValAssign, 16> RVLocs;
2366 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2367 RVLocs, *DAG.getContext());
2368 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2369 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2370 CCValAssign &VA = RVLocs[i];
2371 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2376 // If the callee takes no arguments then go on to check the results of the
2378 if (!Outs.empty()) {
2379 // Check if stack adjustment is needed. For now, do not do this if any
2380 // argument is passed on the stack.
2381 SmallVector<CCValAssign, 16> ArgLocs;
2382 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2383 ArgLocs, *DAG.getContext());
2384 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2385 if (CCInfo.getNextStackOffset()) {
2386 MachineFunction &MF = DAG.getMachineFunction();
2387 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2389 if (Subtarget->isTargetWin64())
2390 // Win64 ABI has additional complications.
2393 // Check if the arguments are already laid out in the right way as
2394 // the caller's fixed stack objects.
2395 MachineFrameInfo *MFI = MF.getFrameInfo();
2396 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2397 const X86InstrInfo *TII =
2398 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2399 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2400 CCValAssign &VA = ArgLocs[i];
2401 EVT RegVT = VA.getLocVT();
2402 SDValue Arg = Outs[i].Val;
2403 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2404 if (VA.getLocInfo() == CCValAssign::Indirect)
2406 if (!VA.isRegLoc()) {
2407 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2419 X86TargetLowering::createFastISel(MachineFunction &mf,
2420 DenseMap<const Value *, unsigned> &vm,
2421 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2422 DenseMap<const AllocaInst *, int> &am,
2423 std::vector<std::pair<MachineInstr*, unsigned> > &pn
2425 , SmallSet<const Instruction *, 8> &cil
2428 return X86::createFastISel(mf, vm, bm, am, pn
2436 //===----------------------------------------------------------------------===//
2437 // Other Lowering Hooks
2438 //===----------------------------------------------------------------------===//
2441 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2442 MachineFunction &MF = DAG.getMachineFunction();
2443 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2444 int ReturnAddrIndex = FuncInfo->getRAIndex();
2446 if (ReturnAddrIndex == 0) {
2447 // Set up a frame object for the return address.
2448 uint64_t SlotSize = TD->getPointerSize();
2449 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2451 FuncInfo->setRAIndex(ReturnAddrIndex);
2454 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2458 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2459 bool hasSymbolicDisplacement) {
2460 // Offset should fit into 32 bit immediate field.
2461 if (!isInt<32>(Offset))
2464 // If we don't have a symbolic displacement - we don't have any extra
2466 if (!hasSymbolicDisplacement)
2469 // FIXME: Some tweaks might be needed for medium code model.
2470 if (M != CodeModel::Small && M != CodeModel::Kernel)
2473 // For small code model we assume that latest object is 16MB before end of 31
2474 // bits boundary. We may also accept pretty large negative constants knowing
2475 // that all objects are in the positive half of address space.
2476 if (M == CodeModel::Small && Offset < 16*1024*1024)
2479 // For kernel code model we know that all object resist in the negative half
2480 // of 32bits address space. We may not accept negative offsets, since they may
2481 // be just off and we may accept pretty large positive ones.
2482 if (M == CodeModel::Kernel && Offset > 0)
2488 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2489 /// specific condition code, returning the condition code and the LHS/RHS of the
2490 /// comparison to make.
2491 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2492 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2494 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2495 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2496 // X > -1 -> X == 0, jump !sign.
2497 RHS = DAG.getConstant(0, RHS.getValueType());
2498 return X86::COND_NS;
2499 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2500 // X < 0 -> X == 0, jump on sign.
2502 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2504 RHS = DAG.getConstant(0, RHS.getValueType());
2505 return X86::COND_LE;
2509 switch (SetCCOpcode) {
2510 default: llvm_unreachable("Invalid integer condition!");
2511 case ISD::SETEQ: return X86::COND_E;
2512 case ISD::SETGT: return X86::COND_G;
2513 case ISD::SETGE: return X86::COND_GE;
2514 case ISD::SETLT: return X86::COND_L;
2515 case ISD::SETLE: return X86::COND_LE;
2516 case ISD::SETNE: return X86::COND_NE;
2517 case ISD::SETULT: return X86::COND_B;
2518 case ISD::SETUGT: return X86::COND_A;
2519 case ISD::SETULE: return X86::COND_BE;
2520 case ISD::SETUGE: return X86::COND_AE;
2524 // First determine if it is required or is profitable to flip the operands.
2526 // If LHS is a foldable load, but RHS is not, flip the condition.
2527 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2528 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2529 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2530 std::swap(LHS, RHS);
2533 switch (SetCCOpcode) {
2539 std::swap(LHS, RHS);
2543 // On a floating point condition, the flags are set as follows:
2545 // 0 | 0 | 0 | X > Y
2546 // 0 | 0 | 1 | X < Y
2547 // 1 | 0 | 0 | X == Y
2548 // 1 | 1 | 1 | unordered
2549 switch (SetCCOpcode) {
2550 default: llvm_unreachable("Condcode should be pre-legalized away");
2552 case ISD::SETEQ: return X86::COND_E;
2553 case ISD::SETOLT: // flipped
2555 case ISD::SETGT: return X86::COND_A;
2556 case ISD::SETOLE: // flipped
2558 case ISD::SETGE: return X86::COND_AE;
2559 case ISD::SETUGT: // flipped
2561 case ISD::SETLT: return X86::COND_B;
2562 case ISD::SETUGE: // flipped
2564 case ISD::SETLE: return X86::COND_BE;
2566 case ISD::SETNE: return X86::COND_NE;
2567 case ISD::SETUO: return X86::COND_P;
2568 case ISD::SETO: return X86::COND_NP;
2570 case ISD::SETUNE: return X86::COND_INVALID;
2574 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2575 /// code. Current x86 isa includes the following FP cmov instructions:
2576 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2577 static bool hasFPCMov(unsigned X86CC) {
2593 /// isFPImmLegal - Returns true if the target can instruction select the
2594 /// specified FP immediate natively. If false, the legalizer will
2595 /// materialize the FP immediate as a load from a constant pool.
2596 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2597 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2598 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2604 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2605 /// the specified range (L, H].
2606 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2607 return (Val < 0) || (Val >= Low && Val < Hi);
2610 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2611 /// specified value.
2612 static bool isUndefOrEqual(int Val, int CmpVal) {
2613 if (Val < 0 || Val == CmpVal)
2618 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2619 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2620 /// the second operand.
2621 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2622 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2623 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2624 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2625 return (Mask[0] < 2 && Mask[1] < 2);
2629 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2630 SmallVector<int, 8> M;
2632 return ::isPSHUFDMask(M, N->getValueType(0));
2635 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2636 /// is suitable for input to PSHUFHW.
2637 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2638 if (VT != MVT::v8i16)
2641 // Lower quadword copied in order or undef.
2642 for (int i = 0; i != 4; ++i)
2643 if (Mask[i] >= 0 && Mask[i] != i)
2646 // Upper quadword shuffled.
2647 for (int i = 4; i != 8; ++i)
2648 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2654 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2655 SmallVector<int, 8> M;
2657 return ::isPSHUFHWMask(M, N->getValueType(0));
2660 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2661 /// is suitable for input to PSHUFLW.
2662 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2663 if (VT != MVT::v8i16)
2666 // Upper quadword copied in order.
2667 for (int i = 4; i != 8; ++i)
2668 if (Mask[i] >= 0 && Mask[i] != i)
2671 // Lower quadword shuffled.
2672 for (int i = 0; i != 4; ++i)
2679 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2680 SmallVector<int, 8> M;
2682 return ::isPSHUFLWMask(M, N->getValueType(0));
2685 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2686 /// is suitable for input to PALIGNR.
2687 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2689 int i, e = VT.getVectorNumElements();
2691 // Do not handle v2i64 / v2f64 shuffles with palignr.
2692 if (e < 4 || !hasSSSE3)
2695 for (i = 0; i != e; ++i)
2699 // All undef, not a palignr.
2703 // Determine if it's ok to perform a palignr with only the LHS, since we
2704 // don't have access to the actual shuffle elements to see if RHS is undef.
2705 bool Unary = Mask[i] < (int)e;
2706 bool NeedsUnary = false;
2708 int s = Mask[i] - i;
2710 // Check the rest of the elements to see if they are consecutive.
2711 for (++i; i != e; ++i) {
2716 Unary = Unary && (m < (int)e);
2717 NeedsUnary = NeedsUnary || (m < s);
2719 if (NeedsUnary && !Unary)
2721 if (Unary && m != ((s+i) & (e-1)))
2723 if (!Unary && m != (s+i))
2729 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2730 SmallVector<int, 8> M;
2732 return ::isPALIGNRMask(M, N->getValueType(0), true);
2735 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2736 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2737 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2738 int NumElems = VT.getVectorNumElements();
2739 if (NumElems != 2 && NumElems != 4)
2742 int Half = NumElems / 2;
2743 for (int i = 0; i < Half; ++i)
2744 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2746 for (int i = Half; i < NumElems; ++i)
2747 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2753 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2754 SmallVector<int, 8> M;
2756 return ::isSHUFPMask(M, N->getValueType(0));
2759 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2760 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2761 /// half elements to come from vector 1 (which would equal the dest.) and
2762 /// the upper half to come from vector 2.
2763 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2764 int NumElems = VT.getVectorNumElements();
2766 if (NumElems != 2 && NumElems != 4)
2769 int Half = NumElems / 2;
2770 for (int i = 0; i < Half; ++i)
2771 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2773 for (int i = Half; i < NumElems; ++i)
2774 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2779 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2780 SmallVector<int, 8> M;
2782 return isCommutedSHUFPMask(M, N->getValueType(0));
2785 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2786 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2787 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2788 if (N->getValueType(0).getVectorNumElements() != 4)
2791 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2792 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2793 isUndefOrEqual(N->getMaskElt(1), 7) &&
2794 isUndefOrEqual(N->getMaskElt(2), 2) &&
2795 isUndefOrEqual(N->getMaskElt(3), 3);
2798 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2799 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2801 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2802 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2807 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2808 isUndefOrEqual(N->getMaskElt(1), 3) &&
2809 isUndefOrEqual(N->getMaskElt(2), 2) &&
2810 isUndefOrEqual(N->getMaskElt(3), 3);
2813 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2814 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2815 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2816 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2818 if (NumElems != 2 && NumElems != 4)
2821 for (unsigned i = 0; i < NumElems/2; ++i)
2822 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2825 for (unsigned i = NumElems/2; i < NumElems; ++i)
2826 if (!isUndefOrEqual(N->getMaskElt(i), i))
2832 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2833 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2834 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2835 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2837 if (NumElems != 2 && NumElems != 4)
2840 for (unsigned i = 0; i < NumElems/2; ++i)
2841 if (!isUndefOrEqual(N->getMaskElt(i), i))
2844 for (unsigned i = 0; i < NumElems/2; ++i)
2845 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2851 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2852 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2853 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2854 bool V2IsSplat = false) {
2855 int NumElts = VT.getVectorNumElements();
2856 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2859 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2861 int BitI1 = Mask[i+1];
2862 if (!isUndefOrEqual(BitI, j))
2865 if (!isUndefOrEqual(BitI1, NumElts))
2868 if (!isUndefOrEqual(BitI1, j + NumElts))
2875 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2876 SmallVector<int, 8> M;
2878 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2881 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2882 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2883 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2884 bool V2IsSplat = false) {
2885 int NumElts = VT.getVectorNumElements();
2886 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2889 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2891 int BitI1 = Mask[i+1];
2892 if (!isUndefOrEqual(BitI, j + NumElts/2))
2895 if (isUndefOrEqual(BitI1, NumElts))
2898 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2905 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2906 SmallVector<int, 8> M;
2908 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2911 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2912 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2914 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2915 int NumElems = VT.getVectorNumElements();
2916 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2919 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2921 int BitI1 = Mask[i+1];
2922 if (!isUndefOrEqual(BitI, j))
2924 if (!isUndefOrEqual(BitI1, j))
2930 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2931 SmallVector<int, 8> M;
2933 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2936 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2937 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2939 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2940 int NumElems = VT.getVectorNumElements();
2941 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2944 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2946 int BitI1 = Mask[i+1];
2947 if (!isUndefOrEqual(BitI, j))
2949 if (!isUndefOrEqual(BitI1, j))
2955 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2956 SmallVector<int, 8> M;
2958 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2961 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2962 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2963 /// MOVSD, and MOVD, i.e. setting the lowest element.
2964 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2965 if (VT.getVectorElementType().getSizeInBits() < 32)
2968 int NumElts = VT.getVectorNumElements();
2970 if (!isUndefOrEqual(Mask[0], NumElts))
2973 for (int i = 1; i < NumElts; ++i)
2974 if (!isUndefOrEqual(Mask[i], i))
2980 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2981 SmallVector<int, 8> M;
2983 return ::isMOVLMask(M, N->getValueType(0));
2986 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2987 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2988 /// element of vector 2 and the other elements to come from vector 1 in order.
2989 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2990 bool V2IsSplat = false, bool V2IsUndef = false) {
2991 int NumOps = VT.getVectorNumElements();
2992 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2995 if (!isUndefOrEqual(Mask[0], 0))
2998 for (int i = 1; i < NumOps; ++i)
2999 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3000 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3001 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3007 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3008 bool V2IsUndef = false) {
3009 SmallVector<int, 8> M;
3011 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3014 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3015 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3016 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3017 if (N->getValueType(0).getVectorNumElements() != 4)
3020 // Expect 1, 1, 3, 3
3021 for (unsigned i = 0; i < 2; ++i) {
3022 int Elt = N->getMaskElt(i);
3023 if (Elt >= 0 && Elt != 1)
3028 for (unsigned i = 2; i < 4; ++i) {
3029 int Elt = N->getMaskElt(i);
3030 if (Elt >= 0 && Elt != 3)
3035 // Don't use movshdup if it can be done with a shufps.
3036 // FIXME: verify that matching u, u, 3, 3 is what we want.
3040 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3041 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3042 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3043 if (N->getValueType(0).getVectorNumElements() != 4)
3046 // Expect 0, 0, 2, 2
3047 for (unsigned i = 0; i < 2; ++i)
3048 if (N->getMaskElt(i) > 0)
3052 for (unsigned i = 2; i < 4; ++i) {
3053 int Elt = N->getMaskElt(i);
3054 if (Elt >= 0 && Elt != 2)
3059 // Don't use movsldup if it can be done with a shufps.
3063 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3064 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3065 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3066 int e = N->getValueType(0).getVectorNumElements() / 2;
3068 for (int i = 0; i < e; ++i)
3069 if (!isUndefOrEqual(N->getMaskElt(i), i))
3071 for (int i = 0; i < e; ++i)
3072 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3077 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3078 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3079 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3080 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3081 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3083 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3085 for (int i = 0; i < NumOperands; ++i) {
3086 int Val = SVOp->getMaskElt(NumOperands-i-1);
3087 if (Val < 0) Val = 0;
3088 if (Val >= NumOperands) Val -= NumOperands;
3090 if (i != NumOperands - 1)
3096 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3097 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3098 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3099 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3101 // 8 nodes, but we only care about the last 4.
3102 for (unsigned i = 7; i >= 4; --i) {
3103 int Val = SVOp->getMaskElt(i);
3112 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3113 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3114 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3115 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3117 // 8 nodes, but we only care about the first 4.
3118 for (int i = 3; i >= 0; --i) {
3119 int Val = SVOp->getMaskElt(i);
3128 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3129 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3130 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3131 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3132 EVT VVT = N->getValueType(0);
3133 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3137 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3138 Val = SVOp->getMaskElt(i);
3142 return (Val - i) * EltSize;
3145 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3147 bool X86::isZeroNode(SDValue Elt) {
3148 return ((isa<ConstantSDNode>(Elt) &&
3149 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3150 (isa<ConstantFPSDNode>(Elt) &&
3151 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3154 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3155 /// their permute mask.
3156 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3157 SelectionDAG &DAG) {
3158 EVT VT = SVOp->getValueType(0);
3159 unsigned NumElems = VT.getVectorNumElements();
3160 SmallVector<int, 8> MaskVec;
3162 for (unsigned i = 0; i != NumElems; ++i) {
3163 int idx = SVOp->getMaskElt(i);
3165 MaskVec.push_back(idx);
3166 else if (idx < (int)NumElems)
3167 MaskVec.push_back(idx + NumElems);
3169 MaskVec.push_back(idx - NumElems);
3171 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3172 SVOp->getOperand(0), &MaskVec[0]);
3175 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3176 /// the two vector operands have swapped position.
3177 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3178 unsigned NumElems = VT.getVectorNumElements();
3179 for (unsigned i = 0; i != NumElems; ++i) {
3183 else if (idx < (int)NumElems)
3184 Mask[i] = idx + NumElems;
3186 Mask[i] = idx - NumElems;
3190 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3191 /// match movhlps. The lower half elements should come from upper half of
3192 /// V1 (and in order), and the upper half elements should come from the upper
3193 /// half of V2 (and in order).
3194 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3195 if (Op->getValueType(0).getVectorNumElements() != 4)
3197 for (unsigned i = 0, e = 2; i != e; ++i)
3198 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3200 for (unsigned i = 2; i != 4; ++i)
3201 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3206 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3207 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3209 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3210 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3212 N = N->getOperand(0).getNode();
3213 if (!ISD::isNON_EXTLoad(N))
3216 *LD = cast<LoadSDNode>(N);
3220 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3221 /// match movlp{s|d}. The lower half elements should come from lower half of
3222 /// V1 (and in order), and the upper half elements should come from the upper
3223 /// half of V2 (and in order). And since V1 will become the source of the
3224 /// MOVLP, it must be either a vector load or a scalar load to vector.
3225 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3226 ShuffleVectorSDNode *Op) {
3227 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3229 // Is V2 is a vector load, don't do this transformation. We will try to use
3230 // load folding shufps op.
3231 if (ISD::isNON_EXTLoad(V2))
3234 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3236 if (NumElems != 2 && NumElems != 4)
3238 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3239 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3241 for (unsigned i = NumElems/2; i != NumElems; ++i)
3242 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3247 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3249 static bool isSplatVector(SDNode *N) {
3250 if (N->getOpcode() != ISD::BUILD_VECTOR)
3253 SDValue SplatValue = N->getOperand(0);
3254 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3255 if (N->getOperand(i) != SplatValue)
3260 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3261 /// to an zero vector.
3262 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3263 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3264 SDValue V1 = N->getOperand(0);
3265 SDValue V2 = N->getOperand(1);
3266 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3267 for (unsigned i = 0; i != NumElems; ++i) {
3268 int Idx = N->getMaskElt(i);
3269 if (Idx >= (int)NumElems) {
3270 unsigned Opc = V2.getOpcode();
3271 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3273 if (Opc != ISD::BUILD_VECTOR ||
3274 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3276 } else if (Idx >= 0) {
3277 unsigned Opc = V1.getOpcode();
3278 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3280 if (Opc != ISD::BUILD_VECTOR ||
3281 !X86::isZeroNode(V1.getOperand(Idx)))
3288 /// getZeroVector - Returns a vector of specified type with all zero elements.
3290 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3292 assert(VT.isVector() && "Expected a vector type");
3294 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3295 // type. This ensures they get CSE'd.
3297 if (VT.getSizeInBits() == 64) { // MMX
3298 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3299 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3300 } else if (HasSSE2) { // SSE2
3301 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3302 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3304 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3305 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3307 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3310 /// getOnesVector - Returns a vector of specified type with all bits set.
3312 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3313 assert(VT.isVector() && "Expected a vector type");
3315 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3316 // type. This ensures they get CSE'd.
3317 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3319 if (VT.getSizeInBits() == 64) // MMX
3320 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3322 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3323 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3327 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3328 /// that point to V2 points to its first element.
3329 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3330 EVT VT = SVOp->getValueType(0);
3331 unsigned NumElems = VT.getVectorNumElements();
3333 bool Changed = false;
3334 SmallVector<int, 8> MaskVec;
3335 SVOp->getMask(MaskVec);
3337 for (unsigned i = 0; i != NumElems; ++i) {
3338 if (MaskVec[i] > (int)NumElems) {
3339 MaskVec[i] = NumElems;
3344 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3345 SVOp->getOperand(1), &MaskVec[0]);
3346 return SDValue(SVOp, 0);
3349 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3350 /// operation of specified width.
3351 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3353 unsigned NumElems = VT.getVectorNumElements();
3354 SmallVector<int, 8> Mask;
3355 Mask.push_back(NumElems);
3356 for (unsigned i = 1; i != NumElems; ++i)
3358 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3361 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3362 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3364 unsigned NumElems = VT.getVectorNumElements();
3365 SmallVector<int, 8> Mask;
3366 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3368 Mask.push_back(i + NumElems);
3370 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3373 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3374 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3376 unsigned NumElems = VT.getVectorNumElements();
3377 unsigned Half = NumElems/2;
3378 SmallVector<int, 8> Mask;
3379 for (unsigned i = 0; i != Half; ++i) {
3380 Mask.push_back(i + Half);
3381 Mask.push_back(i + NumElems + Half);
3383 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3386 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3387 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3389 if (SV->getValueType(0).getVectorNumElements() <= 4)
3390 return SDValue(SV, 0);
3392 EVT PVT = MVT::v4f32;
3393 EVT VT = SV->getValueType(0);
3394 DebugLoc dl = SV->getDebugLoc();
3395 SDValue V1 = SV->getOperand(0);
3396 int NumElems = VT.getVectorNumElements();
3397 int EltNo = SV->getSplatIndex();
3399 // unpack elements to the correct location
3400 while (NumElems > 4) {
3401 if (EltNo < NumElems/2) {
3402 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3404 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3405 EltNo -= NumElems/2;
3410 // Perform the splat.
3411 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3412 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3413 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3414 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3417 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3418 /// vector of zero or undef vector. This produces a shuffle where the low
3419 /// element of V2 is swizzled into the zero/undef vector, landing at element
3420 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3421 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3422 bool isZero, bool HasSSE2,
3423 SelectionDAG &DAG) {
3424 EVT VT = V2.getValueType();
3426 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3427 unsigned NumElems = VT.getVectorNumElements();
3428 SmallVector<int, 16> MaskVec;
3429 for (unsigned i = 0; i != NumElems; ++i)
3430 // If this is the insertion idx, put the low elt of V2 here.
3431 MaskVec.push_back(i == Idx ? NumElems : i);
3432 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3435 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3436 /// a shuffle that is zero.
3438 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3439 bool Low, SelectionDAG &DAG) {
3440 unsigned NumZeros = 0;
3441 for (int i = 0; i < NumElems; ++i) {
3442 unsigned Index = Low ? i : NumElems-i-1;
3443 int Idx = SVOp->getMaskElt(Index);
3448 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3449 if (Elt.getNode() && X86::isZeroNode(Elt))
3457 /// isVectorShift - Returns true if the shuffle can be implemented as a
3458 /// logical left or right shift of a vector.
3459 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3460 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3461 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3462 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3465 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3468 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3472 bool SeenV1 = false;
3473 bool SeenV2 = false;
3474 for (unsigned i = NumZeros; i < NumElems; ++i) {
3475 unsigned Val = isLeft ? (i - NumZeros) : i;
3476 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3479 unsigned Idx = (unsigned) Idx_;
3489 if (SeenV1 && SeenV2)
3492 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3498 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3500 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3501 unsigned NumNonZero, unsigned NumZero,
3503 const TargetLowering &TLI) {
3507 DebugLoc dl = Op.getDebugLoc();
3510 for (unsigned i = 0; i < 16; ++i) {
3511 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3512 if (ThisIsNonZero && First) {
3514 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3516 V = DAG.getUNDEF(MVT::v8i16);
3521 SDValue ThisElt(0, 0), LastElt(0, 0);
3522 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3523 if (LastIsNonZero) {
3524 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3525 MVT::i16, Op.getOperand(i-1));
3527 if (ThisIsNonZero) {
3528 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3529 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3530 ThisElt, DAG.getConstant(8, MVT::i8));
3532 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3536 if (ThisElt.getNode())
3537 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3538 DAG.getIntPtrConstant(i/2));
3542 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3545 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3547 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3548 unsigned NumNonZero, unsigned NumZero,
3550 const TargetLowering &TLI) {
3554 DebugLoc dl = Op.getDebugLoc();
3557 for (unsigned i = 0; i < 8; ++i) {
3558 bool isNonZero = (NonZeros & (1 << i)) != 0;
3562 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3564 V = DAG.getUNDEF(MVT::v8i16);
3567 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3568 MVT::v8i16, V, Op.getOperand(i),
3569 DAG.getIntPtrConstant(i));
3576 /// getVShift - Return a vector logical shift node.
3578 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3579 unsigned NumBits, SelectionDAG &DAG,
3580 const TargetLowering &TLI, DebugLoc dl) {
3581 bool isMMX = VT.getSizeInBits() == 64;
3582 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3583 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3584 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3585 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3586 DAG.getNode(Opc, dl, ShVT, SrcOp,
3587 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3591 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3592 SelectionDAG &DAG) const {
3594 // Check if the scalar load can be widened into a vector load. And if
3595 // the address is "base + cst" see if the cst can be "absorbed" into
3596 // the shuffle mask.
3597 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3598 SDValue Ptr = LD->getBasePtr();
3599 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3601 EVT PVT = LD->getValueType(0);
3602 if (PVT != MVT::i32 && PVT != MVT::f32)
3607 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3608 FI = FINode->getIndex();
3610 } else if (Ptr.getOpcode() == ISD::ADD &&
3611 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3612 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3613 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3614 Offset = Ptr.getConstantOperandVal(1);
3615 Ptr = Ptr.getOperand(0);
3620 SDValue Chain = LD->getChain();
3621 // Make sure the stack object alignment is at least 16.
3622 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3623 if (DAG.InferPtrAlignment(Ptr) < 16) {
3624 if (MFI->isFixedObjectIndex(FI)) {
3625 // Can't change the alignment. FIXME: It's possible to compute
3626 // the exact stack offset and reference FI + adjust offset instead.
3627 // If someone *really* cares about this. That's the way to implement it.
3630 MFI->setObjectAlignment(FI, 16);
3634 // (Offset % 16) must be multiple of 4. Then address is then
3635 // Ptr + (Offset & ~15).
3638 if ((Offset % 16) & 3)
3640 int64_t StartOffset = Offset & ~15;
3642 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3643 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3645 int EltNo = (Offset - StartOffset) >> 2;
3646 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3647 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3648 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3650 // Canonicalize it to a v4i32 shuffle.
3651 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3652 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3653 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3654 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3660 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3661 /// vector of type 'VT', see if the elements can be replaced by a single large
3662 /// load which has the same value as a build_vector whose operands are 'elts'.
3664 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3666 /// FIXME: we'd also like to handle the case where the last elements are zero
3667 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3668 /// There's even a handy isZeroNode for that purpose.
3669 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3670 DebugLoc &dl, SelectionDAG &DAG) {
3671 EVT EltVT = VT.getVectorElementType();
3672 unsigned NumElems = Elts.size();
3674 LoadSDNode *LDBase = NULL;
3675 unsigned LastLoadedElt = -1U;
3677 // For each element in the initializer, see if we've found a load or an undef.
3678 // If we don't find an initial load element, or later load elements are
3679 // non-consecutive, bail out.
3680 for (unsigned i = 0; i < NumElems; ++i) {
3681 SDValue Elt = Elts[i];
3683 if (!Elt.getNode() ||
3684 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3687 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3689 LDBase = cast<LoadSDNode>(Elt.getNode());
3693 if (Elt.getOpcode() == ISD::UNDEF)
3696 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3697 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3702 // If we have found an entire vector of loads and undefs, then return a large
3703 // load of the entire vector width starting at the base pointer. If we found
3704 // consecutive loads for the low half, generate a vzext_load node.
3705 if (LastLoadedElt == NumElems - 1) {
3706 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3707 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3708 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3709 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3710 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3711 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3712 LDBase->isVolatile(), LDBase->isNonTemporal(),
3713 LDBase->getAlignment());
3714 } else if (NumElems == 4 && LastLoadedElt == 1) {
3715 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3716 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3717 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3718 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3724 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3725 DebugLoc dl = Op.getDebugLoc();
3726 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3727 if (ISD::isBuildVectorAllZeros(Op.getNode())
3728 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3729 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3730 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3731 // eliminated on x86-32 hosts.
3732 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3735 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3736 return getOnesVector(Op.getValueType(), DAG, dl);
3737 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3740 EVT VT = Op.getValueType();
3741 EVT ExtVT = VT.getVectorElementType();
3742 unsigned EVTBits = ExtVT.getSizeInBits();
3744 unsigned NumElems = Op.getNumOperands();
3745 unsigned NumZero = 0;
3746 unsigned NumNonZero = 0;
3747 unsigned NonZeros = 0;
3748 bool IsAllConstants = true;
3749 SmallSet<SDValue, 8> Values;
3750 for (unsigned i = 0; i < NumElems; ++i) {
3751 SDValue Elt = Op.getOperand(i);
3752 if (Elt.getOpcode() == ISD::UNDEF)
3755 if (Elt.getOpcode() != ISD::Constant &&
3756 Elt.getOpcode() != ISD::ConstantFP)
3757 IsAllConstants = false;
3758 if (X86::isZeroNode(Elt))
3761 NonZeros |= (1 << i);
3766 if (NumNonZero == 0) {
3767 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3768 return DAG.getUNDEF(VT);
3771 // Special case for single non-zero, non-undef, element.
3772 if (NumNonZero == 1) {
3773 unsigned Idx = CountTrailingZeros_32(NonZeros);
3774 SDValue Item = Op.getOperand(Idx);
3776 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3777 // the value are obviously zero, truncate the value to i32 and do the
3778 // insertion that way. Only do this if the value is non-constant or if the
3779 // value is a constant being inserted into element 0. It is cheaper to do
3780 // a constant pool load than it is to do a movd + shuffle.
3781 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3782 (!IsAllConstants || Idx == 0)) {
3783 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3784 // Handle MMX and SSE both.
3785 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3786 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3788 // Truncate the value (which may itself be a constant) to i32, and
3789 // convert it to a vector with movd (S2V+shuffle to zero extend).
3790 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3791 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3792 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3793 Subtarget->hasSSE2(), DAG);
3795 // Now we have our 32-bit value zero extended in the low element of
3796 // a vector. If Idx != 0, swizzle it into place.
3798 SmallVector<int, 4> Mask;
3799 Mask.push_back(Idx);
3800 for (unsigned i = 1; i != VecElts; ++i)
3802 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3803 DAG.getUNDEF(Item.getValueType()),
3806 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3810 // If we have a constant or non-constant insertion into the low element of
3811 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3812 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3813 // depending on what the source datatype is.
3816 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3817 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3818 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3819 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3820 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3821 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3823 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3824 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3825 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3826 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3827 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3828 Subtarget->hasSSE2(), DAG);
3829 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3833 // Is it a vector logical left shift?
3834 if (NumElems == 2 && Idx == 1 &&
3835 X86::isZeroNode(Op.getOperand(0)) &&
3836 !X86::isZeroNode(Op.getOperand(1))) {
3837 unsigned NumBits = VT.getSizeInBits();
3838 return getVShift(true, VT,
3839 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3840 VT, Op.getOperand(1)),
3841 NumBits/2, DAG, *this, dl);
3844 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3847 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3848 // is a non-constant being inserted into an element other than the low one,
3849 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3850 // movd/movss) to move this into the low element, then shuffle it into
3852 if (EVTBits == 32) {
3853 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3855 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3856 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3857 Subtarget->hasSSE2(), DAG);
3858 SmallVector<int, 8> MaskVec;
3859 for (unsigned i = 0; i < NumElems; i++)
3860 MaskVec.push_back(i == Idx ? 0 : 1);
3861 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3865 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3866 if (Values.size() == 1) {
3867 if (EVTBits == 32) {
3868 // Instead of a shuffle like this:
3869 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3870 // Check if it's possible to issue this instead.
3871 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3872 unsigned Idx = CountTrailingZeros_32(NonZeros);
3873 SDValue Item = Op.getOperand(Idx);
3874 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3875 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3880 // A vector full of immediates; various special cases are already
3881 // handled, so this is best done with a single constant-pool load.
3885 // Let legalizer expand 2-wide build_vectors.
3886 if (EVTBits == 64) {
3887 if (NumNonZero == 1) {
3888 // One half is zero or undef.
3889 unsigned Idx = CountTrailingZeros_32(NonZeros);
3890 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3891 Op.getOperand(Idx));
3892 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3893 Subtarget->hasSSE2(), DAG);
3898 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3899 if (EVTBits == 8 && NumElems == 16) {
3900 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3902 if (V.getNode()) return V;
3905 if (EVTBits == 16 && NumElems == 8) {
3906 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3908 if (V.getNode()) return V;
3911 // If element VT is == 32 bits, turn it into a number of shuffles.
3912 SmallVector<SDValue, 8> V;
3914 if (NumElems == 4 && NumZero > 0) {
3915 for (unsigned i = 0; i < 4; ++i) {
3916 bool isZero = !(NonZeros & (1 << i));
3918 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3920 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3923 for (unsigned i = 0; i < 2; ++i) {
3924 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3927 V[i] = V[i*2]; // Must be a zero vector.
3930 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3933 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3936 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3941 SmallVector<int, 8> MaskVec;
3942 bool Reverse = (NonZeros & 0x3) == 2;
3943 for (unsigned i = 0; i < 2; ++i)
3944 MaskVec.push_back(Reverse ? 1-i : i);
3945 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3946 for (unsigned i = 0; i < 2; ++i)
3947 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3948 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3951 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
3952 // Check for a build vector of consecutive loads.
3953 for (unsigned i = 0; i < NumElems; ++i)
3954 V[i] = Op.getOperand(i);
3956 // Check for elements which are consecutive loads.
3957 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
3961 // For SSE 4.1, use inserts into undef.
3962 if (getSubtarget()->hasSSE41()) {
3963 V[0] = DAG.getUNDEF(VT);
3964 for (unsigned i = 0; i < NumElems; ++i)
3965 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3966 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3967 Op.getOperand(i), DAG.getIntPtrConstant(i));
3971 // Otherwise, expand into a number of unpckl*
3973 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3974 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3975 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3976 for (unsigned i = 0; i < NumElems; ++i)
3977 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3979 while (NumElems != 0) {
3980 for (unsigned i = 0; i < NumElems; ++i)
3981 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3990 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
3991 // We support concatenate two MMX registers and place them in a MMX
3992 // register. This is better than doing a stack convert.
3993 DebugLoc dl = Op.getDebugLoc();
3994 EVT ResVT = Op.getValueType();
3995 assert(Op.getNumOperands() == 2);
3996 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3997 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3999 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4000 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4001 InVec = Op.getOperand(1);
4002 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4003 unsigned NumElts = ResVT.getVectorNumElements();
4004 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4005 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4006 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4008 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4009 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4010 Mask[0] = 0; Mask[1] = 2;
4011 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4013 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4016 // v8i16 shuffles - Prefer shuffles in the following order:
4017 // 1. [all] pshuflw, pshufhw, optional move
4018 // 2. [ssse3] 1 x pshufb
4019 // 3. [ssse3] 2 x pshufb + 1 x por
4020 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4022 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4024 const X86TargetLowering &TLI) {
4025 SDValue V1 = SVOp->getOperand(0);
4026 SDValue V2 = SVOp->getOperand(1);
4027 DebugLoc dl = SVOp->getDebugLoc();
4028 SmallVector<int, 8> MaskVals;
4030 // Determine if more than 1 of the words in each of the low and high quadwords
4031 // of the result come from the same quadword of one of the two inputs. Undef
4032 // mask values count as coming from any quadword, for better codegen.
4033 SmallVector<unsigned, 4> LoQuad(4);
4034 SmallVector<unsigned, 4> HiQuad(4);
4035 BitVector InputQuads(4);
4036 for (unsigned i = 0; i < 8; ++i) {
4037 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4038 int EltIdx = SVOp->getMaskElt(i);
4039 MaskVals.push_back(EltIdx);
4048 InputQuads.set(EltIdx / 4);
4051 int BestLoQuad = -1;
4052 unsigned MaxQuad = 1;
4053 for (unsigned i = 0; i < 4; ++i) {
4054 if (LoQuad[i] > MaxQuad) {
4056 MaxQuad = LoQuad[i];
4060 int BestHiQuad = -1;
4062 for (unsigned i = 0; i < 4; ++i) {
4063 if (HiQuad[i] > MaxQuad) {
4065 MaxQuad = HiQuad[i];
4069 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4070 // of the two input vectors, shuffle them into one input vector so only a
4071 // single pshufb instruction is necessary. If There are more than 2 input
4072 // quads, disable the next transformation since it does not help SSSE3.
4073 bool V1Used = InputQuads[0] || InputQuads[1];
4074 bool V2Used = InputQuads[2] || InputQuads[3];
4075 if (TLI.getSubtarget()->hasSSSE3()) {
4076 if (InputQuads.count() == 2 && V1Used && V2Used) {
4077 BestLoQuad = InputQuads.find_first();
4078 BestHiQuad = InputQuads.find_next(BestLoQuad);
4080 if (InputQuads.count() > 2) {
4086 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4087 // the shuffle mask. If a quad is scored as -1, that means that it contains
4088 // words from all 4 input quadwords.
4090 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4091 SmallVector<int, 8> MaskV;
4092 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4093 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4094 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4095 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4096 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4097 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4099 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4100 // source words for the shuffle, to aid later transformations.
4101 bool AllWordsInNewV = true;
4102 bool InOrder[2] = { true, true };
4103 for (unsigned i = 0; i != 8; ++i) {
4104 int idx = MaskVals[i];
4106 InOrder[i/4] = false;
4107 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4109 AllWordsInNewV = false;
4113 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4114 if (AllWordsInNewV) {
4115 for (int i = 0; i != 8; ++i) {
4116 int idx = MaskVals[i];
4119 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4120 if ((idx != i) && idx < 4)
4122 if ((idx != i) && idx > 3)
4131 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4132 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4133 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4134 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4135 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4139 // If we have SSSE3, and all words of the result are from 1 input vector,
4140 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4141 // is present, fall back to case 4.
4142 if (TLI.getSubtarget()->hasSSSE3()) {
4143 SmallVector<SDValue,16> pshufbMask;
4145 // If we have elements from both input vectors, set the high bit of the
4146 // shuffle mask element to zero out elements that come from V2 in the V1
4147 // mask, and elements that come from V1 in the V2 mask, so that the two
4148 // results can be OR'd together.
4149 bool TwoInputs = V1Used && V2Used;
4150 for (unsigned i = 0; i != 8; ++i) {
4151 int EltIdx = MaskVals[i] * 2;
4152 if (TwoInputs && (EltIdx >= 16)) {
4153 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4154 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4157 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4158 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4160 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4161 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4162 DAG.getNode(ISD::BUILD_VECTOR, dl,
4163 MVT::v16i8, &pshufbMask[0], 16));
4165 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4167 // Calculate the shuffle mask for the second input, shuffle it, and
4168 // OR it with the first shuffled input.
4170 for (unsigned i = 0; i != 8; ++i) {
4171 int EltIdx = MaskVals[i] * 2;
4173 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4174 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4177 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4178 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4180 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4181 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4182 DAG.getNode(ISD::BUILD_VECTOR, dl,
4183 MVT::v16i8, &pshufbMask[0], 16));
4184 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4185 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4188 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4189 // and update MaskVals with new element order.
4190 BitVector InOrder(8);
4191 if (BestLoQuad >= 0) {
4192 SmallVector<int, 8> MaskV;
4193 for (int i = 0; i != 4; ++i) {
4194 int idx = MaskVals[i];
4196 MaskV.push_back(-1);
4198 } else if ((idx / 4) == BestLoQuad) {
4199 MaskV.push_back(idx & 3);
4202 MaskV.push_back(-1);
4205 for (unsigned i = 4; i != 8; ++i)
4207 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4211 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4212 // and update MaskVals with the new element order.
4213 if (BestHiQuad >= 0) {
4214 SmallVector<int, 8> MaskV;
4215 for (unsigned i = 0; i != 4; ++i)
4217 for (unsigned i = 4; i != 8; ++i) {
4218 int idx = MaskVals[i];
4220 MaskV.push_back(-1);
4222 } else if ((idx / 4) == BestHiQuad) {
4223 MaskV.push_back((idx & 3) + 4);
4226 MaskV.push_back(-1);
4229 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4233 // In case BestHi & BestLo were both -1, which means each quadword has a word
4234 // from each of the four input quadwords, calculate the InOrder bitvector now
4235 // before falling through to the insert/extract cleanup.
4236 if (BestLoQuad == -1 && BestHiQuad == -1) {
4238 for (int i = 0; i != 8; ++i)
4239 if (MaskVals[i] < 0 || MaskVals[i] == i)
4243 // The other elements are put in the right place using pextrw and pinsrw.
4244 for (unsigned i = 0; i != 8; ++i) {
4247 int EltIdx = MaskVals[i];
4250 SDValue ExtOp = (EltIdx < 8)
4251 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4252 DAG.getIntPtrConstant(EltIdx))
4253 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4254 DAG.getIntPtrConstant(EltIdx - 8));
4255 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4256 DAG.getIntPtrConstant(i));
4261 // v16i8 shuffles - Prefer shuffles in the following order:
4262 // 1. [ssse3] 1 x pshufb
4263 // 2. [ssse3] 2 x pshufb + 1 x por
4264 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4266 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4268 const X86TargetLowering &TLI) {
4269 SDValue V1 = SVOp->getOperand(0);
4270 SDValue V2 = SVOp->getOperand(1);
4271 DebugLoc dl = SVOp->getDebugLoc();
4272 SmallVector<int, 16> MaskVals;
4273 SVOp->getMask(MaskVals);
4275 // If we have SSSE3, case 1 is generated when all result bytes come from
4276 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4277 // present, fall back to case 3.
4278 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4281 for (unsigned i = 0; i < 16; ++i) {
4282 int EltIdx = MaskVals[i];
4291 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4292 if (TLI.getSubtarget()->hasSSSE3()) {
4293 SmallVector<SDValue,16> pshufbMask;
4295 // If all result elements are from one input vector, then only translate
4296 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4298 // Otherwise, we have elements from both input vectors, and must zero out
4299 // elements that come from V2 in the first mask, and V1 in the second mask
4300 // so that we can OR them together.
4301 bool TwoInputs = !(V1Only || V2Only);
4302 for (unsigned i = 0; i != 16; ++i) {
4303 int EltIdx = MaskVals[i];
4304 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4305 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4308 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4310 // If all the elements are from V2, assign it to V1 and return after
4311 // building the first pshufb.
4314 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4315 DAG.getNode(ISD::BUILD_VECTOR, dl,
4316 MVT::v16i8, &pshufbMask[0], 16));
4320 // Calculate the shuffle mask for the second input, shuffle it, and
4321 // OR it with the first shuffled input.
4323 for (unsigned i = 0; i != 16; ++i) {
4324 int EltIdx = MaskVals[i];
4326 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4329 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4331 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4332 DAG.getNode(ISD::BUILD_VECTOR, dl,
4333 MVT::v16i8, &pshufbMask[0], 16));
4334 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4337 // No SSSE3 - Calculate in place words and then fix all out of place words
4338 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4339 // the 16 different words that comprise the two doublequadword input vectors.
4340 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4341 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4342 SDValue NewV = V2Only ? V2 : V1;
4343 for (int i = 0; i != 8; ++i) {
4344 int Elt0 = MaskVals[i*2];
4345 int Elt1 = MaskVals[i*2+1];
4347 // This word of the result is all undef, skip it.
4348 if (Elt0 < 0 && Elt1 < 0)
4351 // This word of the result is already in the correct place, skip it.
4352 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4354 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4357 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4358 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4361 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4362 // using a single extract together, load it and store it.
4363 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4364 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4365 DAG.getIntPtrConstant(Elt1 / 2));
4366 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4367 DAG.getIntPtrConstant(i));
4371 // If Elt1 is defined, extract it from the appropriate source. If the
4372 // source byte is not also odd, shift the extracted word left 8 bits
4373 // otherwise clear the bottom 8 bits if we need to do an or.
4375 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4376 DAG.getIntPtrConstant(Elt1 / 2));
4377 if ((Elt1 & 1) == 0)
4378 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4379 DAG.getConstant(8, TLI.getShiftAmountTy()));
4381 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4382 DAG.getConstant(0xFF00, MVT::i16));
4384 // If Elt0 is defined, extract it from the appropriate source. If the
4385 // source byte is not also even, shift the extracted word right 8 bits. If
4386 // Elt1 was also defined, OR the extracted values together before
4387 // inserting them in the result.
4389 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4390 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4391 if ((Elt0 & 1) != 0)
4392 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4393 DAG.getConstant(8, TLI.getShiftAmountTy()));
4395 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4396 DAG.getConstant(0x00FF, MVT::i16));
4397 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4400 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4401 DAG.getIntPtrConstant(i));
4403 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4406 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4407 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4408 /// done when every pair / quad of shuffle mask elements point to elements in
4409 /// the right sequence. e.g.
4410 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4412 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4414 const TargetLowering &TLI, DebugLoc dl) {
4415 EVT VT = SVOp->getValueType(0);
4416 SDValue V1 = SVOp->getOperand(0);
4417 SDValue V2 = SVOp->getOperand(1);
4418 unsigned NumElems = VT.getVectorNumElements();
4419 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4420 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4421 EVT MaskEltVT = MaskVT.getVectorElementType();
4423 switch (VT.getSimpleVT().SimpleTy) {
4424 default: assert(false && "Unexpected!");
4425 case MVT::v4f32: NewVT = MVT::v2f64; break;
4426 case MVT::v4i32: NewVT = MVT::v2i64; break;
4427 case MVT::v8i16: NewVT = MVT::v4i32; break;
4428 case MVT::v16i8: NewVT = MVT::v4i32; break;
4431 if (NewWidth == 2) {
4437 int Scale = NumElems / NewWidth;
4438 SmallVector<int, 8> MaskVec;
4439 for (unsigned i = 0; i < NumElems; i += Scale) {
4441 for (int j = 0; j < Scale; ++j) {
4442 int EltIdx = SVOp->getMaskElt(i+j);
4446 StartIdx = EltIdx - (EltIdx % Scale);
4447 if (EltIdx != StartIdx + j)
4451 MaskVec.push_back(-1);
4453 MaskVec.push_back(StartIdx / Scale);
4456 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4457 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4458 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4461 /// getVZextMovL - Return a zero-extending vector move low node.
4463 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4464 SDValue SrcOp, SelectionDAG &DAG,
4465 const X86Subtarget *Subtarget, DebugLoc dl) {
4466 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4467 LoadSDNode *LD = NULL;
4468 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4469 LD = dyn_cast<LoadSDNode>(SrcOp);
4471 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4473 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4474 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4475 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4476 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4477 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4479 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4480 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4481 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4482 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4490 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4491 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4492 DAG.getNode(ISD::BIT_CONVERT, dl,
4496 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4499 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4500 SDValue V1 = SVOp->getOperand(0);
4501 SDValue V2 = SVOp->getOperand(1);
4502 DebugLoc dl = SVOp->getDebugLoc();
4503 EVT VT = SVOp->getValueType(0);
4505 SmallVector<std::pair<int, int>, 8> Locs;
4507 SmallVector<int, 8> Mask1(4U, -1);
4508 SmallVector<int, 8> PermMask;
4509 SVOp->getMask(PermMask);
4513 for (unsigned i = 0; i != 4; ++i) {
4514 int Idx = PermMask[i];
4516 Locs[i] = std::make_pair(-1, -1);
4518 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4520 Locs[i] = std::make_pair(0, NumLo);
4524 Locs[i] = std::make_pair(1, NumHi);
4526 Mask1[2+NumHi] = Idx;
4532 if (NumLo <= 2 && NumHi <= 2) {
4533 // If no more than two elements come from either vector. This can be
4534 // implemented with two shuffles. First shuffle gather the elements.
4535 // The second shuffle, which takes the first shuffle as both of its
4536 // vector operands, put the elements into the right order.
4537 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4539 SmallVector<int, 8> Mask2(4U, -1);
4541 for (unsigned i = 0; i != 4; ++i) {
4542 if (Locs[i].first == -1)
4545 unsigned Idx = (i < 2) ? 0 : 4;
4546 Idx += Locs[i].first * 2 + Locs[i].second;
4551 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4552 } else if (NumLo == 3 || NumHi == 3) {
4553 // Otherwise, we must have three elements from one vector, call it X, and
4554 // one element from the other, call it Y. First, use a shufps to build an
4555 // intermediate vector with the one element from Y and the element from X
4556 // that will be in the same half in the final destination (the indexes don't
4557 // matter). Then, use a shufps to build the final vector, taking the half
4558 // containing the element from Y from the intermediate, and the other half
4561 // Normalize it so the 3 elements come from V1.
4562 CommuteVectorShuffleMask(PermMask, VT);
4566 // Find the element from V2.
4568 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4569 int Val = PermMask[HiIndex];
4576 Mask1[0] = PermMask[HiIndex];
4578 Mask1[2] = PermMask[HiIndex^1];
4580 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4583 Mask1[0] = PermMask[0];
4584 Mask1[1] = PermMask[1];
4585 Mask1[2] = HiIndex & 1 ? 6 : 4;
4586 Mask1[3] = HiIndex & 1 ? 4 : 6;
4587 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4589 Mask1[0] = HiIndex & 1 ? 2 : 0;
4590 Mask1[1] = HiIndex & 1 ? 0 : 2;
4591 Mask1[2] = PermMask[2];
4592 Mask1[3] = PermMask[3];
4597 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4601 // Break it into (shuffle shuffle_hi, shuffle_lo).
4603 SmallVector<int,8> LoMask(4U, -1);
4604 SmallVector<int,8> HiMask(4U, -1);
4606 SmallVector<int,8> *MaskPtr = &LoMask;
4607 unsigned MaskIdx = 0;
4610 for (unsigned i = 0; i != 4; ++i) {
4617 int Idx = PermMask[i];
4619 Locs[i] = std::make_pair(-1, -1);
4620 } else if (Idx < 4) {
4621 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4622 (*MaskPtr)[LoIdx] = Idx;
4625 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4626 (*MaskPtr)[HiIdx] = Idx;
4631 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4632 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4633 SmallVector<int, 8> MaskOps;
4634 for (unsigned i = 0; i != 4; ++i) {
4635 if (Locs[i].first == -1) {
4636 MaskOps.push_back(-1);
4638 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4639 MaskOps.push_back(Idx);
4642 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4646 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4647 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4648 SDValue V1 = Op.getOperand(0);
4649 SDValue V2 = Op.getOperand(1);
4650 EVT VT = Op.getValueType();
4651 DebugLoc dl = Op.getDebugLoc();
4652 unsigned NumElems = VT.getVectorNumElements();
4653 bool isMMX = VT.getSizeInBits() == 64;
4654 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4655 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4656 bool V1IsSplat = false;
4657 bool V2IsSplat = false;
4659 if (isZeroShuffle(SVOp))
4660 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4662 // Promote splats to v4f32.
4663 if (SVOp->isSplat()) {
4664 if (isMMX || NumElems < 4)
4666 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4669 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4671 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4672 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4673 if (NewOp.getNode())
4674 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4675 LowerVECTOR_SHUFFLE(NewOp, DAG));
4676 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4677 // FIXME: Figure out a cleaner way to do this.
4678 // Try to make use of movq to zero out the top part.
4679 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4680 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4681 if (NewOp.getNode()) {
4682 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4683 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4684 DAG, Subtarget, dl);
4686 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4687 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4688 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4689 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4690 DAG, Subtarget, dl);
4694 if (X86::isPSHUFDMask(SVOp))
4697 // Check if this can be converted into a logical shift.
4698 bool isLeft = false;
4701 bool isShift = getSubtarget()->hasSSE2() &&
4702 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4703 if (isShift && ShVal.hasOneUse()) {
4704 // If the shifted value has multiple uses, it may be cheaper to use
4705 // v_set0 + movlhps or movhlps, etc.
4706 EVT EltVT = VT.getVectorElementType();
4707 ShAmt *= EltVT.getSizeInBits();
4708 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4711 if (X86::isMOVLMask(SVOp)) {
4714 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4715 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4720 // FIXME: fold these into legal mask.
4721 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4722 X86::isMOVSLDUPMask(SVOp) ||
4723 X86::isMOVHLPSMask(SVOp) ||
4724 X86::isMOVLHPSMask(SVOp) ||
4725 X86::isMOVLPMask(SVOp)))
4728 if (ShouldXformToMOVHLPS(SVOp) ||
4729 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4730 return CommuteVectorShuffle(SVOp, DAG);
4733 // No better options. Use a vshl / vsrl.
4734 EVT EltVT = VT.getVectorElementType();
4735 ShAmt *= EltVT.getSizeInBits();
4736 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4739 bool Commuted = false;
4740 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4741 // 1,1,1,1 -> v8i16 though.
4742 V1IsSplat = isSplatVector(V1.getNode());
4743 V2IsSplat = isSplatVector(V2.getNode());
4745 // Canonicalize the splat or undef, if present, to be on the RHS.
4746 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4747 Op = CommuteVectorShuffle(SVOp, DAG);
4748 SVOp = cast<ShuffleVectorSDNode>(Op);
4749 V1 = SVOp->getOperand(0);
4750 V2 = SVOp->getOperand(1);
4751 std::swap(V1IsSplat, V2IsSplat);
4752 std::swap(V1IsUndef, V2IsUndef);
4756 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4757 // Shuffling low element of v1 into undef, just return v1.
4760 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4761 // the instruction selector will not match, so get a canonical MOVL with
4762 // swapped operands to undo the commute.
4763 return getMOVL(DAG, dl, VT, V2, V1);
4766 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4767 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4768 X86::isUNPCKLMask(SVOp) ||
4769 X86::isUNPCKHMask(SVOp))
4773 // Normalize mask so all entries that point to V2 points to its first
4774 // element then try to match unpck{h|l} again. If match, return a
4775 // new vector_shuffle with the corrected mask.
4776 SDValue NewMask = NormalizeMask(SVOp, DAG);
4777 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4778 if (NSVOp != SVOp) {
4779 if (X86::isUNPCKLMask(NSVOp, true)) {
4781 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4788 // Commute is back and try unpck* again.
4789 // FIXME: this seems wrong.
4790 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4791 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4792 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4793 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4794 X86::isUNPCKLMask(NewSVOp) ||
4795 X86::isUNPCKHMask(NewSVOp))
4799 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4801 // Normalize the node to match x86 shuffle ops if needed
4802 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4803 return CommuteVectorShuffle(SVOp, DAG);
4805 // Check for legal shuffle and return?
4806 SmallVector<int, 16> PermMask;
4807 SVOp->getMask(PermMask);
4808 if (isShuffleMaskLegal(PermMask, VT))
4811 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4812 if (VT == MVT::v8i16) {
4813 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4814 if (NewOp.getNode())
4818 if (VT == MVT::v16i8) {
4819 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4820 if (NewOp.getNode())
4824 // Handle all 4 wide cases with a number of shuffles except for MMX.
4825 if (NumElems == 4 && !isMMX)
4826 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4832 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4833 SelectionDAG &DAG) const {
4834 EVT VT = Op.getValueType();
4835 DebugLoc dl = Op.getDebugLoc();
4836 if (VT.getSizeInBits() == 8) {
4837 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4838 Op.getOperand(0), Op.getOperand(1));
4839 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4840 DAG.getValueType(VT));
4841 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4842 } else if (VT.getSizeInBits() == 16) {
4843 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4844 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4846 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4847 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4848 DAG.getNode(ISD::BIT_CONVERT, dl,
4852 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4853 Op.getOperand(0), Op.getOperand(1));
4854 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4855 DAG.getValueType(VT));
4856 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4857 } else if (VT == MVT::f32) {
4858 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4859 // the result back to FR32 register. It's only worth matching if the
4860 // result has a single use which is a store or a bitcast to i32. And in
4861 // the case of a store, it's not worth it if the index is a constant 0,
4862 // because a MOVSSmr can be used instead, which is smaller and faster.
4863 if (!Op.hasOneUse())
4865 SDNode *User = *Op.getNode()->use_begin();
4866 if ((User->getOpcode() != ISD::STORE ||
4867 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4868 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4869 (User->getOpcode() != ISD::BIT_CONVERT ||
4870 User->getValueType(0) != MVT::i32))
4872 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4873 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4876 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4877 } else if (VT == MVT::i32) {
4878 // ExtractPS works with constant index.
4879 if (isa<ConstantSDNode>(Op.getOperand(1)))
4887 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4888 SelectionDAG &DAG) const {
4889 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4892 if (Subtarget->hasSSE41()) {
4893 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4898 EVT VT = Op.getValueType();
4899 DebugLoc dl = Op.getDebugLoc();
4900 // TODO: handle v16i8.
4901 if (VT.getSizeInBits() == 16) {
4902 SDValue Vec = Op.getOperand(0);
4903 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4905 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4906 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4907 DAG.getNode(ISD::BIT_CONVERT, dl,
4910 // Transform it so it match pextrw which produces a 32-bit result.
4911 EVT EltVT = MVT::i32;
4912 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4913 Op.getOperand(0), Op.getOperand(1));
4914 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4915 DAG.getValueType(VT));
4916 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4917 } else if (VT.getSizeInBits() == 32) {
4918 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4922 // SHUFPS the element to the lowest double word, then movss.
4923 int Mask[4] = { Idx, -1, -1, -1 };
4924 EVT VVT = Op.getOperand(0).getValueType();
4925 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4926 DAG.getUNDEF(VVT), Mask);
4927 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4928 DAG.getIntPtrConstant(0));
4929 } else if (VT.getSizeInBits() == 64) {
4930 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4931 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4932 // to match extract_elt for f64.
4933 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4937 // UNPCKHPD the element to the lowest double word, then movsd.
4938 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4939 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4940 int Mask[2] = { 1, -1 };
4941 EVT VVT = Op.getOperand(0).getValueType();
4942 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4943 DAG.getUNDEF(VVT), Mask);
4944 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4945 DAG.getIntPtrConstant(0));
4952 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
4953 SelectionDAG &DAG) const {
4954 EVT VT = Op.getValueType();
4955 EVT EltVT = VT.getVectorElementType();
4956 DebugLoc dl = Op.getDebugLoc();
4958 SDValue N0 = Op.getOperand(0);
4959 SDValue N1 = Op.getOperand(1);
4960 SDValue N2 = Op.getOperand(2);
4962 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4963 isa<ConstantSDNode>(N2)) {
4965 if (VT == MVT::v8i16)
4966 Opc = X86ISD::PINSRW;
4967 else if (VT == MVT::v4i16)
4968 Opc = X86ISD::MMX_PINSRW;
4969 else if (VT == MVT::v16i8)
4970 Opc = X86ISD::PINSRB;
4972 Opc = X86ISD::PINSRB;
4974 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4976 if (N1.getValueType() != MVT::i32)
4977 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4978 if (N2.getValueType() != MVT::i32)
4979 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4980 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4981 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4982 // Bits [7:6] of the constant are the source select. This will always be
4983 // zero here. The DAG Combiner may combine an extract_elt index into these
4984 // bits. For example (insert (extract, 3), 2) could be matched by putting
4985 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4986 // Bits [5:4] of the constant are the destination select. This is the
4987 // value of the incoming immediate.
4988 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4989 // combine either bitwise AND or insert of float 0.0 to set these bits.
4990 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4991 // Create this as a scalar to vector..
4992 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4993 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4994 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4995 // PINSR* works with constant index.
5002 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5003 EVT VT = Op.getValueType();
5004 EVT EltVT = VT.getVectorElementType();
5006 if (Subtarget->hasSSE41())
5007 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5009 if (EltVT == MVT::i8)
5012 DebugLoc dl = Op.getDebugLoc();
5013 SDValue N0 = Op.getOperand(0);
5014 SDValue N1 = Op.getOperand(1);
5015 SDValue N2 = Op.getOperand(2);
5017 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5018 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5019 // as its second argument.
5020 if (N1.getValueType() != MVT::i32)
5021 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5022 if (N2.getValueType() != MVT::i32)
5023 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5024 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5025 dl, VT, N0, N1, N2);
5031 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5032 DebugLoc dl = Op.getDebugLoc();
5033 if (Op.getValueType() == MVT::v2f32)
5034 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
5035 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
5036 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
5037 Op.getOperand(0))));
5039 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
5040 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5042 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5043 EVT VT = MVT::v2i32;
5044 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5051 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5052 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5055 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5056 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5057 // one of the above mentioned nodes. It has to be wrapped because otherwise
5058 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5059 // be used to form addressing mode. These wrapped nodes will be selected
5062 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5063 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5065 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5067 unsigned char OpFlag = 0;
5068 unsigned WrapperKind = X86ISD::Wrapper;
5069 CodeModel::Model M = getTargetMachine().getCodeModel();
5071 if (Subtarget->isPICStyleRIPRel() &&
5072 (M == CodeModel::Small || M == CodeModel::Kernel))
5073 WrapperKind = X86ISD::WrapperRIP;
5074 else if (Subtarget->isPICStyleGOT())
5075 OpFlag = X86II::MO_GOTOFF;
5076 else if (Subtarget->isPICStyleStubPIC())
5077 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5079 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5081 CP->getOffset(), OpFlag);
5082 DebugLoc DL = CP->getDebugLoc();
5083 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5084 // With PIC, the address is actually $g + Offset.
5086 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5087 DAG.getNode(X86ISD::GlobalBaseReg,
5088 DebugLoc(), getPointerTy()),
5095 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5096 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5098 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5100 unsigned char OpFlag = 0;
5101 unsigned WrapperKind = X86ISD::Wrapper;
5102 CodeModel::Model M = getTargetMachine().getCodeModel();
5104 if (Subtarget->isPICStyleRIPRel() &&
5105 (M == CodeModel::Small || M == CodeModel::Kernel))
5106 WrapperKind = X86ISD::WrapperRIP;
5107 else if (Subtarget->isPICStyleGOT())
5108 OpFlag = X86II::MO_GOTOFF;
5109 else if (Subtarget->isPICStyleStubPIC())
5110 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5112 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5114 DebugLoc DL = JT->getDebugLoc();
5115 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5117 // With PIC, the address is actually $g + Offset.
5119 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5120 DAG.getNode(X86ISD::GlobalBaseReg,
5121 DebugLoc(), getPointerTy()),
5129 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5130 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5132 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5134 unsigned char OpFlag = 0;
5135 unsigned WrapperKind = X86ISD::Wrapper;
5136 CodeModel::Model M = getTargetMachine().getCodeModel();
5138 if (Subtarget->isPICStyleRIPRel() &&
5139 (M == CodeModel::Small || M == CodeModel::Kernel))
5140 WrapperKind = X86ISD::WrapperRIP;
5141 else if (Subtarget->isPICStyleGOT())
5142 OpFlag = X86II::MO_GOTOFF;
5143 else if (Subtarget->isPICStyleStubPIC())
5144 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5146 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5148 DebugLoc DL = Op.getDebugLoc();
5149 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5152 // With PIC, the address is actually $g + Offset.
5153 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5154 !Subtarget->is64Bit()) {
5155 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5156 DAG.getNode(X86ISD::GlobalBaseReg,
5157 DebugLoc(), getPointerTy()),
5165 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5166 // Create the TargetBlockAddressAddress node.
5167 unsigned char OpFlags =
5168 Subtarget->ClassifyBlockAddressReference();
5169 CodeModel::Model M = getTargetMachine().getCodeModel();
5170 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5171 DebugLoc dl = Op.getDebugLoc();
5172 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5173 /*isTarget=*/true, OpFlags);
5175 if (Subtarget->isPICStyleRIPRel() &&
5176 (M == CodeModel::Small || M == CodeModel::Kernel))
5177 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5179 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5181 // With PIC, the address is actually $g + Offset.
5182 if (isGlobalRelativeToPICBase(OpFlags)) {
5183 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5184 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5192 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5194 SelectionDAG &DAG) const {
5195 // Create the TargetGlobalAddress node, folding in the constant
5196 // offset if it is legal.
5197 unsigned char OpFlags =
5198 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5199 CodeModel::Model M = getTargetMachine().getCodeModel();
5201 if (OpFlags == X86II::MO_NO_FLAG &&
5202 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5203 // A direct static reference to a global.
5204 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5207 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5210 if (Subtarget->isPICStyleRIPRel() &&
5211 (M == CodeModel::Small || M == CodeModel::Kernel))
5212 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5214 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5216 // With PIC, the address is actually $g + Offset.
5217 if (isGlobalRelativeToPICBase(OpFlags)) {
5218 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5219 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5223 // For globals that require a load from a stub to get the address, emit the
5225 if (isGlobalStubReference(OpFlags))
5226 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5227 PseudoSourceValue::getGOT(), 0, false, false, 0);
5229 // If there was a non-zero offset that we didn't fold, create an explicit
5232 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5233 DAG.getConstant(Offset, getPointerTy()));
5239 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5240 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5241 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5242 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5246 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5247 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5248 unsigned char OperandFlags) {
5249 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5250 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5251 DebugLoc dl = GA->getDebugLoc();
5252 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5253 GA->getValueType(0),
5257 SDValue Ops[] = { Chain, TGA, *InFlag };
5258 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5260 SDValue Ops[] = { Chain, TGA };
5261 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5264 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5265 MFI->setHasCalls(true);
5267 SDValue Flag = Chain.getValue(1);
5268 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5271 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5273 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5276 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5277 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5278 DAG.getNode(X86ISD::GlobalBaseReg,
5279 DebugLoc(), PtrVT), InFlag);
5280 InFlag = Chain.getValue(1);
5282 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5285 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5287 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5289 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5290 X86::RAX, X86II::MO_TLSGD);
5293 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5294 // "local exec" model.
5295 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5296 const EVT PtrVT, TLSModel::Model model,
5298 DebugLoc dl = GA->getDebugLoc();
5299 // Get the Thread Pointer
5300 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5302 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5305 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5306 NULL, 0, false, false, 0);
5308 unsigned char OperandFlags = 0;
5309 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5311 unsigned WrapperKind = X86ISD::Wrapper;
5312 if (model == TLSModel::LocalExec) {
5313 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5314 } else if (is64Bit) {
5315 assert(model == TLSModel::InitialExec);
5316 OperandFlags = X86II::MO_GOTTPOFF;
5317 WrapperKind = X86ISD::WrapperRIP;
5319 assert(model == TLSModel::InitialExec);
5320 OperandFlags = X86II::MO_INDNTPOFF;
5323 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5325 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5326 GA->getOffset(), OperandFlags);
5327 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5329 if (model == TLSModel::InitialExec)
5330 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5331 PseudoSourceValue::getGOT(), 0, false, false, 0);
5333 // The address of the thread local variable is the add of the thread
5334 // pointer with the offset of the variable.
5335 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5339 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5340 // TODO: implement the "local dynamic" model
5341 // TODO: implement the "initial exec"model for pic executables
5342 assert(Subtarget->isTargetELF() &&
5343 "TLS not implemented for non-ELF targets");
5344 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5345 const GlobalValue *GV = GA->getGlobal();
5347 // If GV is an alias then use the aliasee for determining
5348 // thread-localness.
5349 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5350 GV = GA->resolveAliasedGlobal(false);
5352 TLSModel::Model model = getTLSModel(GV,
5353 getTargetMachine().getRelocationModel());
5356 case TLSModel::GeneralDynamic:
5357 case TLSModel::LocalDynamic: // not implemented
5358 if (Subtarget->is64Bit())
5359 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5360 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5362 case TLSModel::InitialExec:
5363 case TLSModel::LocalExec:
5364 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5365 Subtarget->is64Bit());
5368 llvm_unreachable("Unreachable");
5373 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5374 /// take a 2 x i32 value to shift plus a shift amount.
5375 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5376 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5377 EVT VT = Op.getValueType();
5378 unsigned VTBits = VT.getSizeInBits();
5379 DebugLoc dl = Op.getDebugLoc();
5380 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5381 SDValue ShOpLo = Op.getOperand(0);
5382 SDValue ShOpHi = Op.getOperand(1);
5383 SDValue ShAmt = Op.getOperand(2);
5384 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5385 DAG.getConstant(VTBits - 1, MVT::i8))
5386 : DAG.getConstant(0, VT);
5389 if (Op.getOpcode() == ISD::SHL_PARTS) {
5390 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5391 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5393 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5394 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5397 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5398 DAG.getConstant(VTBits, MVT::i8));
5399 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5400 AndNode, DAG.getConstant(0, MVT::i8));
5403 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5404 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5405 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5407 if (Op.getOpcode() == ISD::SHL_PARTS) {
5408 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5409 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5411 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5412 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5415 SDValue Ops[2] = { Lo, Hi };
5416 return DAG.getMergeValues(Ops, 2, dl);
5419 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5420 SelectionDAG &DAG) const {
5421 EVT SrcVT = Op.getOperand(0).getValueType();
5423 if (SrcVT.isVector()) {
5424 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5430 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5431 "Unknown SINT_TO_FP to lower!");
5433 // These are really Legal; return the operand so the caller accepts it as
5435 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5437 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5438 Subtarget->is64Bit()) {
5442 DebugLoc dl = Op.getDebugLoc();
5443 unsigned Size = SrcVT.getSizeInBits()/8;
5444 MachineFunction &MF = DAG.getMachineFunction();
5445 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5446 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5447 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5449 PseudoSourceValue::getFixedStack(SSFI), 0,
5451 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5454 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5456 SelectionDAG &DAG) const {
5458 DebugLoc dl = Op.getDebugLoc();
5460 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5462 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5464 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5465 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5466 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5467 Tys, Ops, array_lengthof(Ops));
5470 Chain = Result.getValue(1);
5471 SDValue InFlag = Result.getValue(2);
5473 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5474 // shouldn't be necessary except that RFP cannot be live across
5475 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5476 MachineFunction &MF = DAG.getMachineFunction();
5477 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5478 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5479 Tys = DAG.getVTList(MVT::Other);
5481 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5483 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5484 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5485 PseudoSourceValue::getFixedStack(SSFI), 0,
5492 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5493 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5494 SelectionDAG &DAG) const {
5495 // This algorithm is not obvious. Here it is in C code, more or less:
5497 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5498 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5499 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5501 // Copy ints to xmm registers.
5502 __m128i xh = _mm_cvtsi32_si128( hi );
5503 __m128i xl = _mm_cvtsi32_si128( lo );
5505 // Combine into low half of a single xmm register.
5506 __m128i x = _mm_unpacklo_epi32( xh, xl );
5510 // Merge in appropriate exponents to give the integer bits the right
5512 x = _mm_unpacklo_epi32( x, exp );
5514 // Subtract away the biases to deal with the IEEE-754 double precision
5516 d = _mm_sub_pd( (__m128d) x, bias );
5518 // All conversions up to here are exact. The correctly rounded result is
5519 // calculated using the current rounding mode using the following
5521 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5522 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5523 // store doesn't really need to be here (except
5524 // maybe to zero the other double)
5529 DebugLoc dl = Op.getDebugLoc();
5530 LLVMContext *Context = DAG.getContext();
5532 // Build some magic constants.
5533 std::vector<Constant*> CV0;
5534 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5535 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5536 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5537 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5538 Constant *C0 = ConstantVector::get(CV0);
5539 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5541 std::vector<Constant*> CV1;
5543 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5545 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5546 Constant *C1 = ConstantVector::get(CV1);
5547 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5549 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5550 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5552 DAG.getIntPtrConstant(1)));
5553 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5554 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5556 DAG.getIntPtrConstant(0)));
5557 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5558 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5559 PseudoSourceValue::getConstantPool(), 0,
5561 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5562 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5563 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5564 PseudoSourceValue::getConstantPool(), 0,
5566 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5568 // Add the halves; easiest way is to swap them into another reg first.
5569 int ShufMask[2] = { 1, -1 };
5570 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5571 DAG.getUNDEF(MVT::v2f64), ShufMask);
5572 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5573 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5574 DAG.getIntPtrConstant(0));
5577 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5578 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5579 SelectionDAG &DAG) const {
5580 DebugLoc dl = Op.getDebugLoc();
5581 // FP constant to bias correct the final result.
5582 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5585 // Load the 32-bit value into an XMM register.
5586 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5587 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5589 DAG.getIntPtrConstant(0)));
5591 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5592 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5593 DAG.getIntPtrConstant(0));
5595 // Or the load with the bias.
5596 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5597 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5598 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5600 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5601 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5602 MVT::v2f64, Bias)));
5603 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5604 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5605 DAG.getIntPtrConstant(0));
5607 // Subtract the bias.
5608 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5610 // Handle final rounding.
5611 EVT DestVT = Op.getValueType();
5613 if (DestVT.bitsLT(MVT::f64)) {
5614 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5615 DAG.getIntPtrConstant(0));
5616 } else if (DestVT.bitsGT(MVT::f64)) {
5617 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5620 // Handle final rounding.
5624 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5625 SelectionDAG &DAG) const {
5626 SDValue N0 = Op.getOperand(0);
5627 DebugLoc dl = Op.getDebugLoc();
5629 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5630 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5631 // the optimization here.
5632 if (DAG.SignBitIsZero(N0))
5633 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5635 EVT SrcVT = N0.getValueType();
5636 if (SrcVT == MVT::i64) {
5637 // We only handle SSE2 f64 target here; caller can expand the rest.
5638 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5641 return LowerUINT_TO_FP_i64(Op, DAG);
5642 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5643 return LowerUINT_TO_FP_i32(Op, DAG);
5646 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5648 // Make a 64-bit buffer, and use it to build an FILD.
5649 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5650 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5651 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5652 getPointerTy(), StackSlot, WordOff);
5653 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5654 StackSlot, NULL, 0, false, false, 0);
5655 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5656 OffsetSlot, NULL, 0, false, false, 0);
5657 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5660 std::pair<SDValue,SDValue> X86TargetLowering::
5661 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5662 DebugLoc dl = Op.getDebugLoc();
5664 EVT DstTy = Op.getValueType();
5667 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5671 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5672 DstTy.getSimpleVT() >= MVT::i16 &&
5673 "Unknown FP_TO_SINT to lower!");
5675 // These are really Legal.
5676 if (DstTy == MVT::i32 &&
5677 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5678 return std::make_pair(SDValue(), SDValue());
5679 if (Subtarget->is64Bit() &&
5680 DstTy == MVT::i64 &&
5681 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5682 return std::make_pair(SDValue(), SDValue());
5684 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5686 MachineFunction &MF = DAG.getMachineFunction();
5687 unsigned MemSize = DstTy.getSizeInBits()/8;
5688 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5689 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5692 switch (DstTy.getSimpleVT().SimpleTy) {
5693 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5694 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5695 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5696 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5699 SDValue Chain = DAG.getEntryNode();
5700 SDValue Value = Op.getOperand(0);
5701 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5702 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5703 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5704 PseudoSourceValue::getFixedStack(SSFI), 0,
5706 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5708 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5710 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5711 Chain = Value.getValue(1);
5712 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5713 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5716 // Build the FP_TO_INT*_IN_MEM
5717 SDValue Ops[] = { Chain, Value, StackSlot };
5718 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5720 return std::make_pair(FIST, StackSlot);
5723 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5724 SelectionDAG &DAG) const {
5725 if (Op.getValueType().isVector()) {
5726 if (Op.getValueType() == MVT::v2i32 &&
5727 Op.getOperand(0).getValueType() == MVT::v2f64) {
5733 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5734 SDValue FIST = Vals.first, StackSlot = Vals.second;
5735 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5736 if (FIST.getNode() == 0) return Op;
5739 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5740 FIST, StackSlot, NULL, 0, false, false, 0);
5743 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5744 SelectionDAG &DAG) const {
5745 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5746 SDValue FIST = Vals.first, StackSlot = Vals.second;
5747 assert(FIST.getNode() && "Unexpected failure");
5750 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5751 FIST, StackSlot, NULL, 0, false, false, 0);
5754 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5755 SelectionDAG &DAG) const {
5756 LLVMContext *Context = DAG.getContext();
5757 DebugLoc dl = Op.getDebugLoc();
5758 EVT VT = Op.getValueType();
5761 EltVT = VT.getVectorElementType();
5762 std::vector<Constant*> CV;
5763 if (EltVT == MVT::f64) {
5764 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5768 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5774 Constant *C = ConstantVector::get(CV);
5775 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5776 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5777 PseudoSourceValue::getConstantPool(), 0,
5779 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5782 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5783 LLVMContext *Context = DAG.getContext();
5784 DebugLoc dl = Op.getDebugLoc();
5785 EVT VT = Op.getValueType();
5788 EltVT = VT.getVectorElementType();
5789 std::vector<Constant*> CV;
5790 if (EltVT == MVT::f64) {
5791 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5795 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5801 Constant *C = ConstantVector::get(CV);
5802 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5803 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5804 PseudoSourceValue::getConstantPool(), 0,
5806 if (VT.isVector()) {
5807 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5808 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5809 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5811 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5813 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5817 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5818 LLVMContext *Context = DAG.getContext();
5819 SDValue Op0 = Op.getOperand(0);
5820 SDValue Op1 = Op.getOperand(1);
5821 DebugLoc dl = Op.getDebugLoc();
5822 EVT VT = Op.getValueType();
5823 EVT SrcVT = Op1.getValueType();
5825 // If second operand is smaller, extend it first.
5826 if (SrcVT.bitsLT(VT)) {
5827 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5830 // And if it is bigger, shrink it first.
5831 if (SrcVT.bitsGT(VT)) {
5832 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5836 // At this point the operands and the result should have the same
5837 // type, and that won't be f80 since that is not custom lowered.
5839 // First get the sign bit of second operand.
5840 std::vector<Constant*> CV;
5841 if (SrcVT == MVT::f64) {
5842 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5843 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5845 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5846 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5847 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5848 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5850 Constant *C = ConstantVector::get(CV);
5851 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5852 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5853 PseudoSourceValue::getConstantPool(), 0,
5855 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5857 // Shift sign bit right or left if the two operands have different types.
5858 if (SrcVT.bitsGT(VT)) {
5859 // Op0 is MVT::f32, Op1 is MVT::f64.
5860 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5861 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5862 DAG.getConstant(32, MVT::i32));
5863 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5864 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5865 DAG.getIntPtrConstant(0));
5868 // Clear first operand sign bit.
5870 if (VT == MVT::f64) {
5871 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5872 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5874 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5875 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5876 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5877 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5879 C = ConstantVector::get(CV);
5880 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5881 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5882 PseudoSourceValue::getConstantPool(), 0,
5884 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5886 // Or the value with the sign bit.
5887 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5890 /// Emit nodes that will be selected as "test Op0,Op0", or something
5892 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5893 SelectionDAG &DAG) const {
5894 DebugLoc dl = Op.getDebugLoc();
5896 // CF and OF aren't always set the way we want. Determine which
5897 // of these we need.
5898 bool NeedCF = false;
5899 bool NeedOF = false;
5901 case X86::COND_A: case X86::COND_AE:
5902 case X86::COND_B: case X86::COND_BE:
5905 case X86::COND_G: case X86::COND_GE:
5906 case X86::COND_L: case X86::COND_LE:
5907 case X86::COND_O: case X86::COND_NO:
5913 // See if we can use the EFLAGS value from the operand instead of
5914 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5915 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5916 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5917 unsigned Opcode = 0;
5918 unsigned NumOperands = 0;
5919 switch (Op.getNode()->getOpcode()) {
5921 // Due to an isel shortcoming, be conservative if this add is likely to
5922 // be selected as part of a load-modify-store instruction. When the root
5923 // node in a match is a store, isel doesn't know how to remap non-chain
5924 // non-flag uses of other nodes in the match, such as the ADD in this
5925 // case. This leads to the ADD being left around and reselected, with
5926 // the result being two adds in the output.
5927 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5928 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5929 if (UI->getOpcode() == ISD::STORE)
5931 if (ConstantSDNode *C =
5932 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5933 // An add of one will be selected as an INC.
5934 if (C->getAPIntValue() == 1) {
5935 Opcode = X86ISD::INC;
5939 // An add of negative one (subtract of one) will be selected as a DEC.
5940 if (C->getAPIntValue().isAllOnesValue()) {
5941 Opcode = X86ISD::DEC;
5946 // Otherwise use a regular EFLAGS-setting add.
5947 Opcode = X86ISD::ADD;
5951 // If the primary and result isn't used, don't bother using X86ISD::AND,
5952 // because a TEST instruction will be better.
5953 bool NonFlagUse = false;
5954 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5955 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5957 unsigned UOpNo = UI.getOperandNo();
5958 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5959 // Look pass truncate.
5960 UOpNo = User->use_begin().getOperandNo();
5961 User = *User->use_begin();
5963 if (User->getOpcode() != ISD::BRCOND &&
5964 User->getOpcode() != ISD::SETCC &&
5965 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5977 // Due to the ISEL shortcoming noted above, be conservative if this op is
5978 // likely to be selected as part of a load-modify-store instruction.
5979 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5980 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5981 if (UI->getOpcode() == ISD::STORE)
5983 // Otherwise use a regular EFLAGS-setting instruction.
5984 switch (Op.getNode()->getOpcode()) {
5985 case ISD::SUB: Opcode = X86ISD::SUB; break;
5986 case ISD::OR: Opcode = X86ISD::OR; break;
5987 case ISD::XOR: Opcode = X86ISD::XOR; break;
5988 case ISD::AND: Opcode = X86ISD::AND; break;
5989 default: llvm_unreachable("unexpected operator!");
6000 return SDValue(Op.getNode(), 1);
6006 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6007 SmallVector<SDValue, 4> Ops;
6008 for (unsigned i = 0; i != NumOperands; ++i)
6009 Ops.push_back(Op.getOperand(i));
6010 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6011 DAG.ReplaceAllUsesWith(Op, New);
6012 return SDValue(New.getNode(), 1);
6016 // Otherwise just emit a CMP with 0, which is the TEST pattern.
6017 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6018 DAG.getConstant(0, Op.getValueType()));
6021 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6023 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6024 SelectionDAG &DAG) const {
6025 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6026 if (C->getAPIntValue() == 0)
6027 return EmitTest(Op0, X86CC, DAG);
6029 DebugLoc dl = Op0.getDebugLoc();
6030 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6033 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6034 /// if it's possible.
6035 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6036 DebugLoc dl, SelectionDAG &DAG) const {
6037 SDValue Op0 = And.getOperand(0);
6038 SDValue Op1 = And.getOperand(1);
6039 if (Op0.getOpcode() == ISD::TRUNCATE)
6040 Op0 = Op0.getOperand(0);
6041 if (Op1.getOpcode() == ISD::TRUNCATE)
6042 Op1 = Op1.getOperand(0);
6045 if (Op1.getOpcode() == ISD::SHL) {
6046 if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
6047 if (And10C->getZExtValue() == 1) {
6049 RHS = Op1.getOperand(1);
6051 } else if (Op0.getOpcode() == ISD::SHL) {
6052 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6053 if (And00C->getZExtValue() == 1) {
6055 RHS = Op0.getOperand(1);
6057 } else if (Op1.getOpcode() == ISD::Constant) {
6058 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6059 SDValue AndLHS = Op0;
6060 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6061 LHS = AndLHS.getOperand(0);
6062 RHS = AndLHS.getOperand(1);
6066 if (LHS.getNode()) {
6067 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6068 // instruction. Since the shift amount is in-range-or-undefined, we know
6069 // that doing a bittest on the i32 value is ok. We extend to i32 because
6070 // the encoding for the i16 version is larger than the i32 version.
6071 // Also promote i16 to i32 for performance / code size reason.
6072 if (LHS.getValueType() == MVT::i8 ||
6073 (Subtarget->shouldPromote16Bit() && LHS.getValueType() == MVT::i16))
6074 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6076 // If the operand types disagree, extend the shift amount to match. Since
6077 // BT ignores high bits (like shifts) we can use anyextend.
6078 if (LHS.getValueType() != RHS.getValueType())
6079 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6081 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6082 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6083 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6084 DAG.getConstant(Cond, MVT::i8), BT);
6090 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6091 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6092 SDValue Op0 = Op.getOperand(0);
6093 SDValue Op1 = Op.getOperand(1);
6094 DebugLoc dl = Op.getDebugLoc();
6095 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6097 // Optimize to BT if possible.
6098 // Lower (X & (1 << N)) == 0 to BT(X, N).
6099 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6100 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6101 if (Op0.getOpcode() == ISD::AND &&
6103 Op1.getOpcode() == ISD::Constant &&
6104 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
6105 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6106 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6107 if (NewSetCC.getNode())
6111 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6112 if (Op0.getOpcode() == X86ISD::SETCC &&
6113 Op1.getOpcode() == ISD::Constant &&
6114 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6115 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6116 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6117 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6118 bool Invert = (CC == ISD::SETNE) ^
6119 cast<ConstantSDNode>(Op1)->isNullValue();
6121 CCode = X86::GetOppositeBranchCondition(CCode);
6122 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6123 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6126 bool isFP = Op1.getValueType().isFloatingPoint();
6127 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6128 if (X86CC == X86::COND_INVALID)
6131 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6133 // Use sbb x, x to materialize carry bit into a GPR.
6134 if (X86CC == X86::COND_B)
6135 return DAG.getNode(ISD::AND, dl, MVT::i8,
6136 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6137 DAG.getConstant(X86CC, MVT::i8), Cond),
6138 DAG.getConstant(1, MVT::i8));
6140 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6141 DAG.getConstant(X86CC, MVT::i8), Cond);
6144 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6146 SDValue Op0 = Op.getOperand(0);
6147 SDValue Op1 = Op.getOperand(1);
6148 SDValue CC = Op.getOperand(2);
6149 EVT VT = Op.getValueType();
6150 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6151 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6152 DebugLoc dl = Op.getDebugLoc();
6156 EVT VT0 = Op0.getValueType();
6157 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6158 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6161 switch (SetCCOpcode) {
6164 case ISD::SETEQ: SSECC = 0; break;
6166 case ISD::SETGT: Swap = true; // Fallthrough
6168 case ISD::SETOLT: SSECC = 1; break;
6170 case ISD::SETGE: Swap = true; // Fallthrough
6172 case ISD::SETOLE: SSECC = 2; break;
6173 case ISD::SETUO: SSECC = 3; break;
6175 case ISD::SETNE: SSECC = 4; break;
6176 case ISD::SETULE: Swap = true;
6177 case ISD::SETUGE: SSECC = 5; break;
6178 case ISD::SETULT: Swap = true;
6179 case ISD::SETUGT: SSECC = 6; break;
6180 case ISD::SETO: SSECC = 7; break;
6183 std::swap(Op0, Op1);
6185 // In the two special cases we can't handle, emit two comparisons.
6187 if (SetCCOpcode == ISD::SETUEQ) {
6189 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6190 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6191 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6193 else if (SetCCOpcode == ISD::SETONE) {
6195 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6196 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6197 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6199 llvm_unreachable("Illegal FP comparison");
6201 // Handle all other FP comparisons here.
6202 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6205 // We are handling one of the integer comparisons here. Since SSE only has
6206 // GT and EQ comparisons for integer, swapping operands and multiple
6207 // operations may be required for some comparisons.
6208 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6209 bool Swap = false, Invert = false, FlipSigns = false;
6211 switch (VT.getSimpleVT().SimpleTy) {
6214 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6216 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6218 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6219 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6222 switch (SetCCOpcode) {
6224 case ISD::SETNE: Invert = true;
6225 case ISD::SETEQ: Opc = EQOpc; break;
6226 case ISD::SETLT: Swap = true;
6227 case ISD::SETGT: Opc = GTOpc; break;
6228 case ISD::SETGE: Swap = true;
6229 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6230 case ISD::SETULT: Swap = true;
6231 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6232 case ISD::SETUGE: Swap = true;
6233 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6236 std::swap(Op0, Op1);
6238 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6239 // bits of the inputs before performing those operations.
6241 EVT EltVT = VT.getVectorElementType();
6242 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6244 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6245 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6247 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6248 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6251 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6253 // If the logical-not of the result is required, perform that now.
6255 Result = DAG.getNOT(dl, Result, VT);
6260 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6261 static bool isX86LogicalCmp(SDValue Op) {
6262 unsigned Opc = Op.getNode()->getOpcode();
6263 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6265 if (Op.getResNo() == 1 &&
6266 (Opc == X86ISD::ADD ||
6267 Opc == X86ISD::SUB ||
6268 Opc == X86ISD::SMUL ||
6269 Opc == X86ISD::UMUL ||
6270 Opc == X86ISD::INC ||
6271 Opc == X86ISD::DEC ||
6272 Opc == X86ISD::OR ||
6273 Opc == X86ISD::XOR ||
6274 Opc == X86ISD::AND))
6280 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6281 bool addTest = true;
6282 SDValue Cond = Op.getOperand(0);
6283 DebugLoc dl = Op.getDebugLoc();
6286 if (Cond.getOpcode() == ISD::SETCC) {
6287 SDValue NewCond = LowerSETCC(Cond, DAG);
6288 if (NewCond.getNode())
6292 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6293 SDValue Op1 = Op.getOperand(1);
6294 SDValue Op2 = Op.getOperand(2);
6295 if (Cond.getOpcode() == X86ISD::SETCC &&
6296 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6297 SDValue Cmp = Cond.getOperand(1);
6298 if (Cmp.getOpcode() == X86ISD::CMP) {
6299 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6300 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6301 ConstantSDNode *RHSC =
6302 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6303 if (N1C && N1C->isAllOnesValue() &&
6304 N2C && N2C->isNullValue() &&
6305 RHSC && RHSC->isNullValue()) {
6306 SDValue CmpOp0 = Cmp.getOperand(0);
6307 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6308 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6309 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6310 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6315 // Look pass (and (setcc_carry (cmp ...)), 1).
6316 if (Cond.getOpcode() == ISD::AND &&
6317 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6318 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6319 if (C && C->getAPIntValue() == 1)
6320 Cond = Cond.getOperand(0);
6323 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6324 // setting operand in place of the X86ISD::SETCC.
6325 if (Cond.getOpcode() == X86ISD::SETCC ||
6326 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6327 CC = Cond.getOperand(0);
6329 SDValue Cmp = Cond.getOperand(1);
6330 unsigned Opc = Cmp.getOpcode();
6331 EVT VT = Op.getValueType();
6333 bool IllegalFPCMov = false;
6334 if (VT.isFloatingPoint() && !VT.isVector() &&
6335 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6336 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6338 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6339 Opc == X86ISD::BT) { // FIXME
6346 // Look pass the truncate.
6347 if (Cond.getOpcode() == ISD::TRUNCATE)
6348 Cond = Cond.getOperand(0);
6350 // We know the result of AND is compared against zero. Try to match
6352 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6353 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6354 if (NewSetCC.getNode()) {
6355 CC = NewSetCC.getOperand(0);
6356 Cond = NewSetCC.getOperand(1);
6363 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6364 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6367 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6368 // condition is true.
6369 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6370 SDValue Ops[] = { Op2, Op1, CC, Cond };
6371 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6374 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6375 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6376 // from the AND / OR.
6377 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6378 Opc = Op.getOpcode();
6379 if (Opc != ISD::OR && Opc != ISD::AND)
6381 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6382 Op.getOperand(0).hasOneUse() &&
6383 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6384 Op.getOperand(1).hasOneUse());
6387 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6388 // 1 and that the SETCC node has a single use.
6389 static bool isXor1OfSetCC(SDValue Op) {
6390 if (Op.getOpcode() != ISD::XOR)
6392 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6393 if (N1C && N1C->getAPIntValue() == 1) {
6394 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6395 Op.getOperand(0).hasOneUse();
6400 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6401 bool addTest = true;
6402 SDValue Chain = Op.getOperand(0);
6403 SDValue Cond = Op.getOperand(1);
6404 SDValue Dest = Op.getOperand(2);
6405 DebugLoc dl = Op.getDebugLoc();
6408 if (Cond.getOpcode() == ISD::SETCC) {
6409 SDValue NewCond = LowerSETCC(Cond, DAG);
6410 if (NewCond.getNode())
6414 // FIXME: LowerXALUO doesn't handle these!!
6415 else if (Cond.getOpcode() == X86ISD::ADD ||
6416 Cond.getOpcode() == X86ISD::SUB ||
6417 Cond.getOpcode() == X86ISD::SMUL ||
6418 Cond.getOpcode() == X86ISD::UMUL)
6419 Cond = LowerXALUO(Cond, DAG);
6422 // Look pass (and (setcc_carry (cmp ...)), 1).
6423 if (Cond.getOpcode() == ISD::AND &&
6424 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6425 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6426 if (C && C->getAPIntValue() == 1)
6427 Cond = Cond.getOperand(0);
6430 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6431 // setting operand in place of the X86ISD::SETCC.
6432 if (Cond.getOpcode() == X86ISD::SETCC ||
6433 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6434 CC = Cond.getOperand(0);
6436 SDValue Cmp = Cond.getOperand(1);
6437 unsigned Opc = Cmp.getOpcode();
6438 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6439 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6443 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6447 // These can only come from an arithmetic instruction with overflow,
6448 // e.g. SADDO, UADDO.
6449 Cond = Cond.getNode()->getOperand(1);
6456 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6457 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6458 if (CondOpc == ISD::OR) {
6459 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6460 // two branches instead of an explicit OR instruction with a
6462 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6463 isX86LogicalCmp(Cmp)) {
6464 CC = Cond.getOperand(0).getOperand(0);
6465 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6466 Chain, Dest, CC, Cmp);
6467 CC = Cond.getOperand(1).getOperand(0);
6471 } else { // ISD::AND
6472 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6473 // two branches instead of an explicit AND instruction with a
6474 // separate test. However, we only do this if this block doesn't
6475 // have a fall-through edge, because this requires an explicit
6476 // jmp when the condition is false.
6477 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6478 isX86LogicalCmp(Cmp) &&
6479 Op.getNode()->hasOneUse()) {
6480 X86::CondCode CCode =
6481 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6482 CCode = X86::GetOppositeBranchCondition(CCode);
6483 CC = DAG.getConstant(CCode, MVT::i8);
6484 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6485 // Look for an unconditional branch following this conditional branch.
6486 // We need this because we need to reverse the successors in order
6487 // to implement FCMP_OEQ.
6488 if (User.getOpcode() == ISD::BR) {
6489 SDValue FalseBB = User.getOperand(1);
6491 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6492 assert(NewBR == User);
6495 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6496 Chain, Dest, CC, Cmp);
6497 X86::CondCode CCode =
6498 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6499 CCode = X86::GetOppositeBranchCondition(CCode);
6500 CC = DAG.getConstant(CCode, MVT::i8);
6506 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6507 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6508 // It should be transformed during dag combiner except when the condition
6509 // is set by a arithmetics with overflow node.
6510 X86::CondCode CCode =
6511 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6512 CCode = X86::GetOppositeBranchCondition(CCode);
6513 CC = DAG.getConstant(CCode, MVT::i8);
6514 Cond = Cond.getOperand(0).getOperand(1);
6520 // Look pass the truncate.
6521 if (Cond.getOpcode() == ISD::TRUNCATE)
6522 Cond = Cond.getOperand(0);
6524 // We know the result of AND is compared against zero. Try to match
6526 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6527 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6528 if (NewSetCC.getNode()) {
6529 CC = NewSetCC.getOperand(0);
6530 Cond = NewSetCC.getOperand(1);
6537 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6538 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6540 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6541 Chain, Dest, CC, Cond);
6545 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6546 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6547 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6548 // that the guard pages used by the OS virtual memory manager are allocated in
6549 // correct sequence.
6551 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6552 SelectionDAG &DAG) const {
6553 assert(Subtarget->isTargetCygMing() &&
6554 "This should be used only on Cygwin/Mingw targets");
6555 DebugLoc dl = Op.getDebugLoc();
6558 SDValue Chain = Op.getOperand(0);
6559 SDValue Size = Op.getOperand(1);
6560 // FIXME: Ensure alignment here
6564 EVT IntPtr = getPointerTy();
6565 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6567 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6568 Flag = Chain.getValue(1);
6570 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6572 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6573 Flag = Chain.getValue(1);
6575 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6577 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6578 return DAG.getMergeValues(Ops1, 2, dl);
6582 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6584 SDValue Dst, SDValue Src,
6585 SDValue Size, unsigned Align,
6588 uint64_t DstSVOff) const {
6589 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6591 // If not DWORD aligned or size is more than the threshold, call the library.
6592 // The libc version is likely to be faster for these cases. It can use the
6593 // address value and run time information about the CPU.
6594 if ((Align & 3) != 0 ||
6596 ConstantSize->getZExtValue() >
6597 getSubtarget()->getMaxInlineSizeThreshold()) {
6598 SDValue InFlag(0, 0);
6600 // Check to see if there is a specialized entry-point for memory zeroing.
6601 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6603 if (const char *bzeroEntry = V &&
6604 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6605 EVT IntPtr = getPointerTy();
6606 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6607 TargetLowering::ArgListTy Args;
6608 TargetLowering::ArgListEntry Entry;
6610 Entry.Ty = IntPtrTy;
6611 Args.push_back(Entry);
6613 Args.push_back(Entry);
6614 std::pair<SDValue,SDValue> CallResult =
6615 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6616 false, false, false, false,
6617 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6618 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
6619 return CallResult.second;
6622 // Otherwise have the target-independent code call memset.
6626 uint64_t SizeVal = ConstantSize->getZExtValue();
6627 SDValue InFlag(0, 0);
6630 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6631 unsigned BytesLeft = 0;
6632 bool TwoRepStos = false;
6635 uint64_t Val = ValC->getZExtValue() & 255;
6637 // If the value is a constant, then we can potentially use larger sets.
6638 switch (Align & 3) {
6639 case 2: // WORD aligned
6642 Val = (Val << 8) | Val;
6644 case 0: // DWORD aligned
6647 Val = (Val << 8) | Val;
6648 Val = (Val << 16) | Val;
6649 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6652 Val = (Val << 32) | Val;
6655 default: // Byte aligned
6658 Count = DAG.getIntPtrConstant(SizeVal);
6662 if (AVT.bitsGT(MVT::i8)) {
6663 unsigned UBytes = AVT.getSizeInBits() / 8;
6664 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6665 BytesLeft = SizeVal % UBytes;
6668 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6670 InFlag = Chain.getValue(1);
6673 Count = DAG.getIntPtrConstant(SizeVal);
6674 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6675 InFlag = Chain.getValue(1);
6678 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6681 InFlag = Chain.getValue(1);
6682 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6685 InFlag = Chain.getValue(1);
6687 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6688 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6689 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6692 InFlag = Chain.getValue(1);
6694 EVT CVT = Count.getValueType();
6695 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6696 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6697 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6700 InFlag = Chain.getValue(1);
6701 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6702 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6703 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6704 } else if (BytesLeft) {
6705 // Handle the last 1 - 7 bytes.
6706 unsigned Offset = SizeVal - BytesLeft;
6707 EVT AddrVT = Dst.getValueType();
6708 EVT SizeVT = Size.getValueType();
6710 Chain = DAG.getMemset(Chain, dl,
6711 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6712 DAG.getConstant(Offset, AddrVT)),
6714 DAG.getConstant(BytesLeft, SizeVT),
6715 Align, isVolatile, DstSV, DstSVOff + Offset);
6718 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6723 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6724 SDValue Chain, SDValue Dst, SDValue Src,
6725 SDValue Size, unsigned Align,
6726 bool isVolatile, bool AlwaysInline,
6730 uint64_t SrcSVOff) const {
6731 // This requires the copy size to be a constant, preferrably
6732 // within a subtarget-specific limit.
6733 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6736 uint64_t SizeVal = ConstantSize->getZExtValue();
6737 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6740 /// If not DWORD aligned, call the library.
6741 if ((Align & 3) != 0)
6746 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6749 unsigned UBytes = AVT.getSizeInBits() / 8;
6750 unsigned CountVal = SizeVal / UBytes;
6751 SDValue Count = DAG.getIntPtrConstant(CountVal);
6752 unsigned BytesLeft = SizeVal % UBytes;
6754 SDValue InFlag(0, 0);
6755 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6758 InFlag = Chain.getValue(1);
6759 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6762 InFlag = Chain.getValue(1);
6763 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6766 InFlag = Chain.getValue(1);
6768 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6769 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6770 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6771 array_lengthof(Ops));
6773 SmallVector<SDValue, 4> Results;
6774 Results.push_back(RepMovs);
6776 // Handle the last 1 - 7 bytes.
6777 unsigned Offset = SizeVal - BytesLeft;
6778 EVT DstVT = Dst.getValueType();
6779 EVT SrcVT = Src.getValueType();
6780 EVT SizeVT = Size.getValueType();
6781 Results.push_back(DAG.getMemcpy(Chain, dl,
6782 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6783 DAG.getConstant(Offset, DstVT)),
6784 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6785 DAG.getConstant(Offset, SrcVT)),
6786 DAG.getConstant(BytesLeft, SizeVT),
6787 Align, isVolatile, AlwaysInline,
6788 DstSV, DstSVOff + Offset,
6789 SrcSV, SrcSVOff + Offset));
6792 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6793 &Results[0], Results.size());
6796 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6797 MachineFunction &MF = DAG.getMachineFunction();
6798 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6800 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6801 DebugLoc dl = Op.getDebugLoc();
6803 if (!Subtarget->is64Bit()) {
6804 // vastart just stores the address of the VarArgsFrameIndex slot into the
6805 // memory location argument.
6806 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6808 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6813 // gp_offset (0 - 6 * 8)
6814 // fp_offset (48 - 48 + 8 * 16)
6815 // overflow_arg_area (point to parameters coming in memory).
6817 SmallVector<SDValue, 8> MemOps;
6818 SDValue FIN = Op.getOperand(1);
6820 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6821 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6823 FIN, SV, 0, false, false, 0);
6824 MemOps.push_back(Store);
6827 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6828 FIN, DAG.getIntPtrConstant(4));
6829 Store = DAG.getStore(Op.getOperand(0), dl,
6830 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6832 FIN, SV, 0, false, false, 0);
6833 MemOps.push_back(Store);
6835 // Store ptr to overflow_arg_area
6836 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6837 FIN, DAG.getIntPtrConstant(4));
6838 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6840 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6842 MemOps.push_back(Store);
6844 // Store ptr to reg_save_area.
6845 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6846 FIN, DAG.getIntPtrConstant(8));
6847 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6849 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6851 MemOps.push_back(Store);
6852 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6853 &MemOps[0], MemOps.size());
6856 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6857 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6858 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6859 SDValue Chain = Op.getOperand(0);
6860 SDValue SrcPtr = Op.getOperand(1);
6861 SDValue SrcSV = Op.getOperand(2);
6863 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6867 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6868 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6869 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6870 SDValue Chain = Op.getOperand(0);
6871 SDValue DstPtr = Op.getOperand(1);
6872 SDValue SrcPtr = Op.getOperand(2);
6873 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6874 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6875 DebugLoc dl = Op.getDebugLoc();
6877 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6878 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6879 false, DstSV, 0, SrcSV, 0);
6883 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6884 DebugLoc dl = Op.getDebugLoc();
6885 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6887 default: return SDValue(); // Don't custom lower most intrinsics.
6888 // Comparison intrinsics.
6889 case Intrinsic::x86_sse_comieq_ss:
6890 case Intrinsic::x86_sse_comilt_ss:
6891 case Intrinsic::x86_sse_comile_ss:
6892 case Intrinsic::x86_sse_comigt_ss:
6893 case Intrinsic::x86_sse_comige_ss:
6894 case Intrinsic::x86_sse_comineq_ss:
6895 case Intrinsic::x86_sse_ucomieq_ss:
6896 case Intrinsic::x86_sse_ucomilt_ss:
6897 case Intrinsic::x86_sse_ucomile_ss:
6898 case Intrinsic::x86_sse_ucomigt_ss:
6899 case Intrinsic::x86_sse_ucomige_ss:
6900 case Intrinsic::x86_sse_ucomineq_ss:
6901 case Intrinsic::x86_sse2_comieq_sd:
6902 case Intrinsic::x86_sse2_comilt_sd:
6903 case Intrinsic::x86_sse2_comile_sd:
6904 case Intrinsic::x86_sse2_comigt_sd:
6905 case Intrinsic::x86_sse2_comige_sd:
6906 case Intrinsic::x86_sse2_comineq_sd:
6907 case Intrinsic::x86_sse2_ucomieq_sd:
6908 case Intrinsic::x86_sse2_ucomilt_sd:
6909 case Intrinsic::x86_sse2_ucomile_sd:
6910 case Intrinsic::x86_sse2_ucomigt_sd:
6911 case Intrinsic::x86_sse2_ucomige_sd:
6912 case Intrinsic::x86_sse2_ucomineq_sd: {
6914 ISD::CondCode CC = ISD::SETCC_INVALID;
6917 case Intrinsic::x86_sse_comieq_ss:
6918 case Intrinsic::x86_sse2_comieq_sd:
6922 case Intrinsic::x86_sse_comilt_ss:
6923 case Intrinsic::x86_sse2_comilt_sd:
6927 case Intrinsic::x86_sse_comile_ss:
6928 case Intrinsic::x86_sse2_comile_sd:
6932 case Intrinsic::x86_sse_comigt_ss:
6933 case Intrinsic::x86_sse2_comigt_sd:
6937 case Intrinsic::x86_sse_comige_ss:
6938 case Intrinsic::x86_sse2_comige_sd:
6942 case Intrinsic::x86_sse_comineq_ss:
6943 case Intrinsic::x86_sse2_comineq_sd:
6947 case Intrinsic::x86_sse_ucomieq_ss:
6948 case Intrinsic::x86_sse2_ucomieq_sd:
6949 Opc = X86ISD::UCOMI;
6952 case Intrinsic::x86_sse_ucomilt_ss:
6953 case Intrinsic::x86_sse2_ucomilt_sd:
6954 Opc = X86ISD::UCOMI;
6957 case Intrinsic::x86_sse_ucomile_ss:
6958 case Intrinsic::x86_sse2_ucomile_sd:
6959 Opc = X86ISD::UCOMI;
6962 case Intrinsic::x86_sse_ucomigt_ss:
6963 case Intrinsic::x86_sse2_ucomigt_sd:
6964 Opc = X86ISD::UCOMI;
6967 case Intrinsic::x86_sse_ucomige_ss:
6968 case Intrinsic::x86_sse2_ucomige_sd:
6969 Opc = X86ISD::UCOMI;
6972 case Intrinsic::x86_sse_ucomineq_ss:
6973 case Intrinsic::x86_sse2_ucomineq_sd:
6974 Opc = X86ISD::UCOMI;
6979 SDValue LHS = Op.getOperand(1);
6980 SDValue RHS = Op.getOperand(2);
6981 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6982 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6983 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6984 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6985 DAG.getConstant(X86CC, MVT::i8), Cond);
6986 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6988 // ptest intrinsics. The intrinsic these come from are designed to return
6989 // an integer value, not just an instruction so lower it to the ptest
6990 // pattern and a setcc for the result.
6991 case Intrinsic::x86_sse41_ptestz:
6992 case Intrinsic::x86_sse41_ptestc:
6993 case Intrinsic::x86_sse41_ptestnzc:{
6996 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6997 case Intrinsic::x86_sse41_ptestz:
6999 X86CC = X86::COND_E;
7001 case Intrinsic::x86_sse41_ptestc:
7003 X86CC = X86::COND_B;
7005 case Intrinsic::x86_sse41_ptestnzc:
7007 X86CC = X86::COND_A;
7011 SDValue LHS = Op.getOperand(1);
7012 SDValue RHS = Op.getOperand(2);
7013 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
7014 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7015 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7016 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7019 // Fix vector shift instructions where the last operand is a non-immediate
7021 case Intrinsic::x86_sse2_pslli_w:
7022 case Intrinsic::x86_sse2_pslli_d:
7023 case Intrinsic::x86_sse2_pslli_q:
7024 case Intrinsic::x86_sse2_psrli_w:
7025 case Intrinsic::x86_sse2_psrli_d:
7026 case Intrinsic::x86_sse2_psrli_q:
7027 case Intrinsic::x86_sse2_psrai_w:
7028 case Intrinsic::x86_sse2_psrai_d:
7029 case Intrinsic::x86_mmx_pslli_w:
7030 case Intrinsic::x86_mmx_pslli_d:
7031 case Intrinsic::x86_mmx_pslli_q:
7032 case Intrinsic::x86_mmx_psrli_w:
7033 case Intrinsic::x86_mmx_psrli_d:
7034 case Intrinsic::x86_mmx_psrli_q:
7035 case Intrinsic::x86_mmx_psrai_w:
7036 case Intrinsic::x86_mmx_psrai_d: {
7037 SDValue ShAmt = Op.getOperand(2);
7038 if (isa<ConstantSDNode>(ShAmt))
7041 unsigned NewIntNo = 0;
7042 EVT ShAmtVT = MVT::v4i32;
7044 case Intrinsic::x86_sse2_pslli_w:
7045 NewIntNo = Intrinsic::x86_sse2_psll_w;
7047 case Intrinsic::x86_sse2_pslli_d:
7048 NewIntNo = Intrinsic::x86_sse2_psll_d;
7050 case Intrinsic::x86_sse2_pslli_q:
7051 NewIntNo = Intrinsic::x86_sse2_psll_q;
7053 case Intrinsic::x86_sse2_psrli_w:
7054 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7056 case Intrinsic::x86_sse2_psrli_d:
7057 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7059 case Intrinsic::x86_sse2_psrli_q:
7060 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7062 case Intrinsic::x86_sse2_psrai_w:
7063 NewIntNo = Intrinsic::x86_sse2_psra_w;
7065 case Intrinsic::x86_sse2_psrai_d:
7066 NewIntNo = Intrinsic::x86_sse2_psra_d;
7069 ShAmtVT = MVT::v2i32;
7071 case Intrinsic::x86_mmx_pslli_w:
7072 NewIntNo = Intrinsic::x86_mmx_psll_w;
7074 case Intrinsic::x86_mmx_pslli_d:
7075 NewIntNo = Intrinsic::x86_mmx_psll_d;
7077 case Intrinsic::x86_mmx_pslli_q:
7078 NewIntNo = Intrinsic::x86_mmx_psll_q;
7080 case Intrinsic::x86_mmx_psrli_w:
7081 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7083 case Intrinsic::x86_mmx_psrli_d:
7084 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7086 case Intrinsic::x86_mmx_psrli_q:
7087 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7089 case Intrinsic::x86_mmx_psrai_w:
7090 NewIntNo = Intrinsic::x86_mmx_psra_w;
7092 case Intrinsic::x86_mmx_psrai_d:
7093 NewIntNo = Intrinsic::x86_mmx_psra_d;
7095 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7101 // The vector shift intrinsics with scalars uses 32b shift amounts but
7102 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7106 ShOps[1] = DAG.getConstant(0, MVT::i32);
7107 if (ShAmtVT == MVT::v4i32) {
7108 ShOps[2] = DAG.getUNDEF(MVT::i32);
7109 ShOps[3] = DAG.getUNDEF(MVT::i32);
7110 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7112 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7115 EVT VT = Op.getValueType();
7116 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7117 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7118 DAG.getConstant(NewIntNo, MVT::i32),
7119 Op.getOperand(1), ShAmt);
7124 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7125 SelectionDAG &DAG) const {
7126 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7127 DebugLoc dl = Op.getDebugLoc();
7130 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7132 DAG.getConstant(TD->getPointerSize(),
7133 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7134 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7135 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7137 NULL, 0, false, false, 0);
7140 // Just load the return address.
7141 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7142 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7143 RetAddrFI, NULL, 0, false, false, 0);
7146 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7147 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7148 MFI->setFrameAddressIsTaken(true);
7149 EVT VT = Op.getValueType();
7150 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7151 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7152 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7153 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7155 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7160 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7161 SelectionDAG &DAG) const {
7162 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7165 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7166 MachineFunction &MF = DAG.getMachineFunction();
7167 SDValue Chain = Op.getOperand(0);
7168 SDValue Offset = Op.getOperand(1);
7169 SDValue Handler = Op.getOperand(2);
7170 DebugLoc dl = Op.getDebugLoc();
7172 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7174 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7176 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7177 DAG.getIntPtrConstant(-TD->getPointerSize()));
7178 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7179 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7180 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7181 MF.getRegInfo().addLiveOut(StoreAddrReg);
7183 return DAG.getNode(X86ISD::EH_RETURN, dl,
7185 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7188 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7189 SelectionDAG &DAG) const {
7190 SDValue Root = Op.getOperand(0);
7191 SDValue Trmp = Op.getOperand(1); // trampoline
7192 SDValue FPtr = Op.getOperand(2); // nested function
7193 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7194 DebugLoc dl = Op.getDebugLoc();
7196 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7198 if (Subtarget->is64Bit()) {
7199 SDValue OutChains[6];
7201 // Large code-model.
7202 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7203 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7205 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7206 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7208 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7210 // Load the pointer to the nested function into R11.
7211 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7212 SDValue Addr = Trmp;
7213 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7214 Addr, TrmpAddr, 0, false, false, 0);
7216 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7217 DAG.getConstant(2, MVT::i64));
7218 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7221 // Load the 'nest' parameter value into R10.
7222 // R10 is specified in X86CallingConv.td
7223 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7224 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7225 DAG.getConstant(10, MVT::i64));
7226 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7227 Addr, TrmpAddr, 10, false, false, 0);
7229 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7230 DAG.getConstant(12, MVT::i64));
7231 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7234 // Jump to the nested function.
7235 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7236 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7237 DAG.getConstant(20, MVT::i64));
7238 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7239 Addr, TrmpAddr, 20, false, false, 0);
7241 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7242 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7243 DAG.getConstant(22, MVT::i64));
7244 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7245 TrmpAddr, 22, false, false, 0);
7248 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7249 return DAG.getMergeValues(Ops, 2, dl);
7251 const Function *Func =
7252 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7253 CallingConv::ID CC = Func->getCallingConv();
7258 llvm_unreachable("Unsupported calling convention");
7259 case CallingConv::C:
7260 case CallingConv::X86_StdCall: {
7261 // Pass 'nest' parameter in ECX.
7262 // Must be kept in sync with X86CallingConv.td
7265 // Check that ECX wasn't needed by an 'inreg' parameter.
7266 const FunctionType *FTy = Func->getFunctionType();
7267 const AttrListPtr &Attrs = Func->getAttributes();
7269 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7270 unsigned InRegCount = 0;
7273 for (FunctionType::param_iterator I = FTy->param_begin(),
7274 E = FTy->param_end(); I != E; ++I, ++Idx)
7275 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7276 // FIXME: should only count parameters that are lowered to integers.
7277 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7279 if (InRegCount > 2) {
7280 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7285 case CallingConv::X86_FastCall:
7286 case CallingConv::Fast:
7287 // Pass 'nest' parameter in EAX.
7288 // Must be kept in sync with X86CallingConv.td
7293 SDValue OutChains[4];
7296 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7297 DAG.getConstant(10, MVT::i32));
7298 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7300 // This is storing the opcode for MOV32ri.
7301 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7302 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7303 OutChains[0] = DAG.getStore(Root, dl,
7304 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7305 Trmp, TrmpAddr, 0, false, false, 0);
7307 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7308 DAG.getConstant(1, MVT::i32));
7309 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7312 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7313 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7314 DAG.getConstant(5, MVT::i32));
7315 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7316 TrmpAddr, 5, false, false, 1);
7318 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7319 DAG.getConstant(6, MVT::i32));
7320 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7324 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7325 return DAG.getMergeValues(Ops, 2, dl);
7329 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7330 SelectionDAG &DAG) const {
7332 The rounding mode is in bits 11:10 of FPSR, and has the following
7339 FLT_ROUNDS, on the other hand, expects the following:
7346 To perform the conversion, we do:
7347 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7350 MachineFunction &MF = DAG.getMachineFunction();
7351 const TargetMachine &TM = MF.getTarget();
7352 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7353 unsigned StackAlignment = TFI.getStackAlignment();
7354 EVT VT = Op.getValueType();
7355 DebugLoc dl = Op.getDebugLoc();
7357 // Save FP Control Word to stack slot
7358 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7359 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7361 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7362 DAG.getEntryNode(), StackSlot);
7364 // Load FP Control Word from stack slot
7365 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7368 // Transform as necessary
7370 DAG.getNode(ISD::SRL, dl, MVT::i16,
7371 DAG.getNode(ISD::AND, dl, MVT::i16,
7372 CWD, DAG.getConstant(0x800, MVT::i16)),
7373 DAG.getConstant(11, MVT::i8));
7375 DAG.getNode(ISD::SRL, dl, MVT::i16,
7376 DAG.getNode(ISD::AND, dl, MVT::i16,
7377 CWD, DAG.getConstant(0x400, MVT::i16)),
7378 DAG.getConstant(9, MVT::i8));
7381 DAG.getNode(ISD::AND, dl, MVT::i16,
7382 DAG.getNode(ISD::ADD, dl, MVT::i16,
7383 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7384 DAG.getConstant(1, MVT::i16)),
7385 DAG.getConstant(3, MVT::i16));
7388 return DAG.getNode((VT.getSizeInBits() < 16 ?
7389 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7392 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7393 EVT VT = Op.getValueType();
7395 unsigned NumBits = VT.getSizeInBits();
7396 DebugLoc dl = Op.getDebugLoc();
7398 Op = Op.getOperand(0);
7399 if (VT == MVT::i8) {
7400 // Zero extend to i32 since there is not an i8 bsr.
7402 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7405 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7406 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7407 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7409 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7412 DAG.getConstant(NumBits+NumBits-1, OpVT),
7413 DAG.getConstant(X86::COND_E, MVT::i8),
7416 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7418 // Finally xor with NumBits-1.
7419 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7422 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7426 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7427 EVT VT = Op.getValueType();
7429 unsigned NumBits = VT.getSizeInBits();
7430 DebugLoc dl = Op.getDebugLoc();
7432 Op = Op.getOperand(0);
7433 if (VT == MVT::i8) {
7435 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7438 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7439 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7440 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7442 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7445 DAG.getConstant(NumBits, OpVT),
7446 DAG.getConstant(X86::COND_E, MVT::i8),
7449 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7452 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7456 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7457 EVT VT = Op.getValueType();
7458 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7459 DebugLoc dl = Op.getDebugLoc();
7461 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7462 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7463 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7464 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7465 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7467 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7468 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7469 // return AloBlo + AloBhi + AhiBlo;
7471 SDValue A = Op.getOperand(0);
7472 SDValue B = Op.getOperand(1);
7474 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7475 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7476 A, DAG.getConstant(32, MVT::i32));
7477 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7478 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7479 B, DAG.getConstant(32, MVT::i32));
7480 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7481 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7483 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7484 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7486 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7487 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7489 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7490 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7491 AloBhi, DAG.getConstant(32, MVT::i32));
7492 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7493 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7494 AhiBlo, DAG.getConstant(32, MVT::i32));
7495 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7496 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7501 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7502 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7503 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7504 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7505 // has only one use.
7506 SDNode *N = Op.getNode();
7507 SDValue LHS = N->getOperand(0);
7508 SDValue RHS = N->getOperand(1);
7509 unsigned BaseOp = 0;
7511 DebugLoc dl = Op.getDebugLoc();
7513 switch (Op.getOpcode()) {
7514 default: llvm_unreachable("Unknown ovf instruction!");
7516 // A subtract of one will be selected as a INC. Note that INC doesn't
7517 // set CF, so we can't do this for UADDO.
7518 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7519 if (C->getAPIntValue() == 1) {
7520 BaseOp = X86ISD::INC;
7524 BaseOp = X86ISD::ADD;
7528 BaseOp = X86ISD::ADD;
7532 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7533 // set CF, so we can't do this for USUBO.
7534 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7535 if (C->getAPIntValue() == 1) {
7536 BaseOp = X86ISD::DEC;
7540 BaseOp = X86ISD::SUB;
7544 BaseOp = X86ISD::SUB;
7548 BaseOp = X86ISD::SMUL;
7552 BaseOp = X86ISD::UMUL;
7557 // Also sets EFLAGS.
7558 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7559 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7562 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7563 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7565 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7569 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7570 EVT T = Op.getValueType();
7571 DebugLoc dl = Op.getDebugLoc();
7574 switch(T.getSimpleVT().SimpleTy) {
7576 assert(false && "Invalid value type!");
7577 case MVT::i8: Reg = X86::AL; size = 1; break;
7578 case MVT::i16: Reg = X86::AX; size = 2; break;
7579 case MVT::i32: Reg = X86::EAX; size = 4; break;
7581 assert(Subtarget->is64Bit() && "Node not type legal!");
7582 Reg = X86::RAX; size = 8;
7585 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7586 Op.getOperand(2), SDValue());
7587 SDValue Ops[] = { cpIn.getValue(0),
7590 DAG.getTargetConstant(size, MVT::i8),
7592 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7593 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7595 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7599 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7600 SelectionDAG &DAG) const {
7601 assert(Subtarget->is64Bit() && "Result not type legalized?");
7602 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7603 SDValue TheChain = Op.getOperand(0);
7604 DebugLoc dl = Op.getDebugLoc();
7605 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7606 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7607 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7609 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7610 DAG.getConstant(32, MVT::i8));
7612 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7615 return DAG.getMergeValues(Ops, 2, dl);
7618 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7619 SDNode *Node = Op.getNode();
7620 DebugLoc dl = Node->getDebugLoc();
7621 EVT T = Node->getValueType(0);
7622 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7623 DAG.getConstant(0, T), Node->getOperand(2));
7624 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7625 cast<AtomicSDNode>(Node)->getMemoryVT(),
7626 Node->getOperand(0),
7627 Node->getOperand(1), negOp,
7628 cast<AtomicSDNode>(Node)->getSrcValue(),
7629 cast<AtomicSDNode>(Node)->getAlignment());
7632 /// LowerOperation - Provide custom lowering hooks for some operations.
7634 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7635 switch (Op.getOpcode()) {
7636 default: llvm_unreachable("Should not custom lower this!");
7637 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7638 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7639 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7640 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7641 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7642 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7643 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7644 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7645 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7646 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7647 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7648 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7649 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7650 case ISD::SHL_PARTS:
7651 case ISD::SRA_PARTS:
7652 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7653 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7654 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7655 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7656 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7657 case ISD::FABS: return LowerFABS(Op, DAG);
7658 case ISD::FNEG: return LowerFNEG(Op, DAG);
7659 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7660 case ISD::SETCC: return LowerSETCC(Op, DAG);
7661 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7662 case ISD::SELECT: return LowerSELECT(Op, DAG);
7663 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7664 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7665 case ISD::VASTART: return LowerVASTART(Op, DAG);
7666 case ISD::VAARG: return LowerVAARG(Op, DAG);
7667 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7668 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7669 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7670 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7671 case ISD::FRAME_TO_ARGS_OFFSET:
7672 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7673 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7674 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7675 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7676 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7677 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7678 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7679 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7685 case ISD::UMULO: return LowerXALUO(Op, DAG);
7686 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7690 void X86TargetLowering::
7691 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7692 SelectionDAG &DAG, unsigned NewOp) const {
7693 EVT T = Node->getValueType(0);
7694 DebugLoc dl = Node->getDebugLoc();
7695 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7697 SDValue Chain = Node->getOperand(0);
7698 SDValue In1 = Node->getOperand(1);
7699 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7700 Node->getOperand(2), DAG.getIntPtrConstant(0));
7701 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7702 Node->getOperand(2), DAG.getIntPtrConstant(1));
7703 SDValue Ops[] = { Chain, In1, In2L, In2H };
7704 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7706 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7707 cast<MemSDNode>(Node)->getMemOperand());
7708 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7709 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7710 Results.push_back(Result.getValue(2));
7713 /// ReplaceNodeResults - Replace a node with an illegal result type
7714 /// with a new node built out of custom code.
7715 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7716 SmallVectorImpl<SDValue>&Results,
7717 SelectionDAG &DAG) const {
7718 DebugLoc dl = N->getDebugLoc();
7719 switch (N->getOpcode()) {
7721 assert(false && "Do not know how to custom type legalize this operation!");
7723 case ISD::FP_TO_SINT: {
7724 std::pair<SDValue,SDValue> Vals =
7725 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7726 SDValue FIST = Vals.first, StackSlot = Vals.second;
7727 if (FIST.getNode() != 0) {
7728 EVT VT = N->getValueType(0);
7729 // Return a load from the stack slot.
7730 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7735 case ISD::READCYCLECOUNTER: {
7736 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7737 SDValue TheChain = N->getOperand(0);
7738 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7739 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7741 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7743 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7744 SDValue Ops[] = { eax, edx };
7745 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7746 Results.push_back(edx.getValue(1));
7749 case ISD::ATOMIC_CMP_SWAP: {
7750 EVT T = N->getValueType(0);
7751 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7752 SDValue cpInL, cpInH;
7753 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7754 DAG.getConstant(0, MVT::i32));
7755 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7756 DAG.getConstant(1, MVT::i32));
7757 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7758 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7760 SDValue swapInL, swapInH;
7761 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7762 DAG.getConstant(0, MVT::i32));
7763 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7764 DAG.getConstant(1, MVT::i32));
7765 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7767 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7768 swapInL.getValue(1));
7769 SDValue Ops[] = { swapInH.getValue(0),
7771 swapInH.getValue(1) };
7772 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7773 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7774 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7775 MVT::i32, Result.getValue(1));
7776 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7777 MVT::i32, cpOutL.getValue(2));
7778 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7779 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7780 Results.push_back(cpOutH.getValue(1));
7783 case ISD::ATOMIC_LOAD_ADD:
7784 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7786 case ISD::ATOMIC_LOAD_AND:
7787 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7789 case ISD::ATOMIC_LOAD_NAND:
7790 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7792 case ISD::ATOMIC_LOAD_OR:
7793 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7795 case ISD::ATOMIC_LOAD_SUB:
7796 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7798 case ISD::ATOMIC_LOAD_XOR:
7799 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7801 case ISD::ATOMIC_SWAP:
7802 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7807 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7809 default: return NULL;
7810 case X86ISD::BSF: return "X86ISD::BSF";
7811 case X86ISD::BSR: return "X86ISD::BSR";
7812 case X86ISD::SHLD: return "X86ISD::SHLD";
7813 case X86ISD::SHRD: return "X86ISD::SHRD";
7814 case X86ISD::FAND: return "X86ISD::FAND";
7815 case X86ISD::FOR: return "X86ISD::FOR";
7816 case X86ISD::FXOR: return "X86ISD::FXOR";
7817 case X86ISD::FSRL: return "X86ISD::FSRL";
7818 case X86ISD::FILD: return "X86ISD::FILD";
7819 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7820 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7821 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7822 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7823 case X86ISD::FLD: return "X86ISD::FLD";
7824 case X86ISD::FST: return "X86ISD::FST";
7825 case X86ISD::CALL: return "X86ISD::CALL";
7826 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7827 case X86ISD::BT: return "X86ISD::BT";
7828 case X86ISD::CMP: return "X86ISD::CMP";
7829 case X86ISD::COMI: return "X86ISD::COMI";
7830 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7831 case X86ISD::SETCC: return "X86ISD::SETCC";
7832 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7833 case X86ISD::CMOV: return "X86ISD::CMOV";
7834 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7835 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7836 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7837 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7838 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7839 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7840 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7841 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7842 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7843 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7844 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7845 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7846 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7847 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7848 case X86ISD::FMAX: return "X86ISD::FMAX";
7849 case X86ISD::FMIN: return "X86ISD::FMIN";
7850 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7851 case X86ISD::FRCP: return "X86ISD::FRCP";
7852 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7853 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7854 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7855 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7856 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7857 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7858 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7859 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7860 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7861 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7862 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7863 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7864 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7865 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7866 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7867 case X86ISD::VSHL: return "X86ISD::VSHL";
7868 case X86ISD::VSRL: return "X86ISD::VSRL";
7869 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7870 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7871 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7872 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7873 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7874 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7875 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7876 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7877 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7878 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7879 case X86ISD::ADD: return "X86ISD::ADD";
7880 case X86ISD::SUB: return "X86ISD::SUB";
7881 case X86ISD::SMUL: return "X86ISD::SMUL";
7882 case X86ISD::UMUL: return "X86ISD::UMUL";
7883 case X86ISD::INC: return "X86ISD::INC";
7884 case X86ISD::DEC: return "X86ISD::DEC";
7885 case X86ISD::OR: return "X86ISD::OR";
7886 case X86ISD::XOR: return "X86ISD::XOR";
7887 case X86ISD::AND: return "X86ISD::AND";
7888 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7889 case X86ISD::PTEST: return "X86ISD::PTEST";
7890 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7891 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7895 // isLegalAddressingMode - Return true if the addressing mode represented
7896 // by AM is legal for this target, for a load/store of the specified type.
7897 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7898 const Type *Ty) const {
7899 // X86 supports extremely general addressing modes.
7900 CodeModel::Model M = getTargetMachine().getCodeModel();
7902 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7903 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7908 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7910 // If a reference to this global requires an extra load, we can't fold it.
7911 if (isGlobalStubReference(GVFlags))
7914 // If BaseGV requires a register for the PIC base, we cannot also have a
7915 // BaseReg specified.
7916 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7919 // If lower 4G is not available, then we must use rip-relative addressing.
7920 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7930 // These scales always work.
7935 // These scales are formed with basereg+scalereg. Only accept if there is
7940 default: // Other stuff never works.
7948 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7949 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7951 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7952 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7953 if (NumBits1 <= NumBits2)
7958 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7959 if (!VT1.isInteger() || !VT2.isInteger())
7961 unsigned NumBits1 = VT1.getSizeInBits();
7962 unsigned NumBits2 = VT2.getSizeInBits();
7963 if (NumBits1 <= NumBits2)
7968 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7969 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7970 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7973 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7974 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7975 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7978 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7979 // i16 instructions are longer (0x66 prefix) and potentially slower.
7980 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7983 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7984 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7985 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7986 /// are assumed to be legal.
7988 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7990 // Very little shuffling can be done for 64-bit vectors right now.
7991 if (VT.getSizeInBits() == 64)
7992 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
7994 // FIXME: pshufb, blends, shifts.
7995 return (VT.getVectorNumElements() == 2 ||
7996 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7997 isMOVLMask(M, VT) ||
7998 isSHUFPMask(M, VT) ||
7999 isPSHUFDMask(M, VT) ||
8000 isPSHUFHWMask(M, VT) ||
8001 isPSHUFLWMask(M, VT) ||
8002 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8003 isUNPCKLMask(M, VT) ||
8004 isUNPCKHMask(M, VT) ||
8005 isUNPCKL_v_undef_Mask(M, VT) ||
8006 isUNPCKH_v_undef_Mask(M, VT));
8010 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8012 unsigned NumElts = VT.getVectorNumElements();
8013 // FIXME: This collection of masks seems suspect.
8016 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8017 return (isMOVLMask(Mask, VT) ||
8018 isCommutedMOVLMask(Mask, VT, true) ||
8019 isSHUFPMask(Mask, VT) ||
8020 isCommutedSHUFPMask(Mask, VT));
8025 //===----------------------------------------------------------------------===//
8026 // X86 Scheduler Hooks
8027 //===----------------------------------------------------------------------===//
8029 // private utility function
8031 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8032 MachineBasicBlock *MBB,
8040 TargetRegisterClass *RC,
8041 bool invSrc) const {
8042 // For the atomic bitwise operator, we generate
8045 // ld t1 = [bitinstr.addr]
8046 // op t2 = t1, [bitinstr.val]
8048 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8050 // fallthrough -->nextMBB
8051 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8052 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8053 MachineFunction::iterator MBBIter = MBB;
8056 /// First build the CFG
8057 MachineFunction *F = MBB->getParent();
8058 MachineBasicBlock *thisMBB = MBB;
8059 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8060 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8061 F->insert(MBBIter, newMBB);
8062 F->insert(MBBIter, nextMBB);
8064 // Move all successors to thisMBB to nextMBB
8065 nextMBB->transferSuccessors(thisMBB);
8067 // Update thisMBB to fall through to newMBB
8068 thisMBB->addSuccessor(newMBB);
8070 // newMBB jumps to itself and fall through to nextMBB
8071 newMBB->addSuccessor(nextMBB);
8072 newMBB->addSuccessor(newMBB);
8074 // Insert instructions into newMBB based on incoming instruction
8075 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8076 "unexpected number of operands");
8077 DebugLoc dl = bInstr->getDebugLoc();
8078 MachineOperand& destOper = bInstr->getOperand(0);
8079 MachineOperand* argOpers[2 + X86AddrNumOperands];
8080 int numArgs = bInstr->getNumOperands() - 1;
8081 for (int i=0; i < numArgs; ++i)
8082 argOpers[i] = &bInstr->getOperand(i+1);
8084 // x86 address has 4 operands: base, index, scale, and displacement
8085 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8086 int valArgIndx = lastAddrIndx + 1;
8088 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8089 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8090 for (int i=0; i <= lastAddrIndx; ++i)
8091 (*MIB).addOperand(*argOpers[i]);
8093 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8095 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8100 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8101 assert((argOpers[valArgIndx]->isReg() ||
8102 argOpers[valArgIndx]->isImm()) &&
8104 if (argOpers[valArgIndx]->isReg())
8105 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8107 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8109 (*MIB).addOperand(*argOpers[valArgIndx]);
8111 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8114 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8115 for (int i=0; i <= lastAddrIndx; ++i)
8116 (*MIB).addOperand(*argOpers[i]);
8118 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8119 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8120 bInstr->memoperands_end());
8122 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8126 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8128 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8132 // private utility function: 64 bit atomics on 32 bit host.
8134 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8135 MachineBasicBlock *MBB,
8140 bool invSrc) const {
8141 // For the atomic bitwise operator, we generate
8142 // thisMBB (instructions are in pairs, except cmpxchg8b)
8143 // ld t1,t2 = [bitinstr.addr]
8145 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8146 // op t5, t6 <- out1, out2, [bitinstr.val]
8147 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8148 // mov ECX, EBX <- t5, t6
8149 // mov EAX, EDX <- t1, t2
8150 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8151 // mov t3, t4 <- EAX, EDX
8153 // result in out1, out2
8154 // fallthrough -->nextMBB
8156 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8157 const unsigned LoadOpc = X86::MOV32rm;
8158 const unsigned copyOpc = X86::MOV32rr;
8159 const unsigned NotOpc = X86::NOT32r;
8160 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8161 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8162 MachineFunction::iterator MBBIter = MBB;
8165 /// First build the CFG
8166 MachineFunction *F = MBB->getParent();
8167 MachineBasicBlock *thisMBB = MBB;
8168 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8169 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8170 F->insert(MBBIter, newMBB);
8171 F->insert(MBBIter, nextMBB);
8173 // Move all successors to thisMBB to nextMBB
8174 nextMBB->transferSuccessors(thisMBB);
8176 // Update thisMBB to fall through to newMBB
8177 thisMBB->addSuccessor(newMBB);
8179 // newMBB jumps to itself and fall through to nextMBB
8180 newMBB->addSuccessor(nextMBB);
8181 newMBB->addSuccessor(newMBB);
8183 DebugLoc dl = bInstr->getDebugLoc();
8184 // Insert instructions into newMBB based on incoming instruction
8185 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8186 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8187 "unexpected number of operands");
8188 MachineOperand& dest1Oper = bInstr->getOperand(0);
8189 MachineOperand& dest2Oper = bInstr->getOperand(1);
8190 MachineOperand* argOpers[2 + X86AddrNumOperands];
8191 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8192 argOpers[i] = &bInstr->getOperand(i+2);
8194 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8195 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8197 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8198 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8199 for (int i=0; i <= lastAddrIndx; ++i)
8200 (*MIB).addOperand(*argOpers[i]);
8201 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8202 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8203 // add 4 to displacement.
8204 for (int i=0; i <= lastAddrIndx-2; ++i)
8205 (*MIB).addOperand(*argOpers[i]);
8206 MachineOperand newOp3 = *(argOpers[3]);
8208 newOp3.setImm(newOp3.getImm()+4);
8210 newOp3.setOffset(newOp3.getOffset()+4);
8211 (*MIB).addOperand(newOp3);
8212 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8214 // t3/4 are defined later, at the bottom of the loop
8215 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8216 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8217 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8218 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8219 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8220 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8222 // The subsequent operations should be using the destination registers of
8223 //the PHI instructions.
8225 t1 = F->getRegInfo().createVirtualRegister(RC);
8226 t2 = F->getRegInfo().createVirtualRegister(RC);
8227 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8228 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8230 t1 = dest1Oper.getReg();
8231 t2 = dest2Oper.getReg();
8234 int valArgIndx = lastAddrIndx + 1;
8235 assert((argOpers[valArgIndx]->isReg() ||
8236 argOpers[valArgIndx]->isImm()) &&
8238 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8239 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8240 if (argOpers[valArgIndx]->isReg())
8241 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8243 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8244 if (regOpcL != X86::MOV32rr)
8246 (*MIB).addOperand(*argOpers[valArgIndx]);
8247 assert(argOpers[valArgIndx + 1]->isReg() ==
8248 argOpers[valArgIndx]->isReg());
8249 assert(argOpers[valArgIndx + 1]->isImm() ==
8250 argOpers[valArgIndx]->isImm());
8251 if (argOpers[valArgIndx + 1]->isReg())
8252 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8254 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8255 if (regOpcH != X86::MOV32rr)
8257 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8259 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8261 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8264 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8266 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8269 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8270 for (int i=0; i <= lastAddrIndx; ++i)
8271 (*MIB).addOperand(*argOpers[i]);
8273 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8274 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8275 bInstr->memoperands_end());
8277 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8278 MIB.addReg(X86::EAX);
8279 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8280 MIB.addReg(X86::EDX);
8283 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8285 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8289 // private utility function
8291 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8292 MachineBasicBlock *MBB,
8293 unsigned cmovOpc) const {
8294 // For the atomic min/max operator, we generate
8297 // ld t1 = [min/max.addr]
8298 // mov t2 = [min/max.val]
8300 // cmov[cond] t2 = t1
8302 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8304 // fallthrough -->nextMBB
8306 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8307 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8308 MachineFunction::iterator MBBIter = MBB;
8311 /// First build the CFG
8312 MachineFunction *F = MBB->getParent();
8313 MachineBasicBlock *thisMBB = MBB;
8314 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8315 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8316 F->insert(MBBIter, newMBB);
8317 F->insert(MBBIter, nextMBB);
8319 // Move all successors of thisMBB to nextMBB
8320 nextMBB->transferSuccessors(thisMBB);
8322 // Update thisMBB to fall through to newMBB
8323 thisMBB->addSuccessor(newMBB);
8325 // newMBB jumps to newMBB and fall through to nextMBB
8326 newMBB->addSuccessor(nextMBB);
8327 newMBB->addSuccessor(newMBB);
8329 DebugLoc dl = mInstr->getDebugLoc();
8330 // Insert instructions into newMBB based on incoming instruction
8331 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8332 "unexpected number of operands");
8333 MachineOperand& destOper = mInstr->getOperand(0);
8334 MachineOperand* argOpers[2 + X86AddrNumOperands];
8335 int numArgs = mInstr->getNumOperands() - 1;
8336 for (int i=0; i < numArgs; ++i)
8337 argOpers[i] = &mInstr->getOperand(i+1);
8339 // x86 address has 4 operands: base, index, scale, and displacement
8340 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8341 int valArgIndx = lastAddrIndx + 1;
8343 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8344 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8345 for (int i=0; i <= lastAddrIndx; ++i)
8346 (*MIB).addOperand(*argOpers[i]);
8348 // We only support register and immediate values
8349 assert((argOpers[valArgIndx]->isReg() ||
8350 argOpers[valArgIndx]->isImm()) &&
8353 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8354 if (argOpers[valArgIndx]->isReg())
8355 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8357 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8358 (*MIB).addOperand(*argOpers[valArgIndx]);
8360 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8363 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8368 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8369 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8373 // Cmp and exchange if none has modified the memory location
8374 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8375 for (int i=0; i <= lastAddrIndx; ++i)
8376 (*MIB).addOperand(*argOpers[i]);
8378 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8379 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8380 mInstr->memoperands_end());
8382 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8383 MIB.addReg(X86::EAX);
8386 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8388 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8392 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8393 // all of this code can be replaced with that in the .td file.
8395 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8396 unsigned numArgs, bool memArg) const {
8398 MachineFunction *F = BB->getParent();
8399 DebugLoc dl = MI->getDebugLoc();
8400 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8404 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8406 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8408 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8410 for (unsigned i = 0; i < numArgs; ++i) {
8411 MachineOperand &Op = MI->getOperand(i+1);
8413 if (!(Op.isReg() && Op.isImplicit()))
8417 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8420 F->DeleteMachineInstr(MI);
8426 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8428 MachineBasicBlock *MBB) const {
8429 // Emit code to save XMM registers to the stack. The ABI says that the
8430 // number of registers to save is given in %al, so it's theoretically
8431 // possible to do an indirect jump trick to avoid saving all of them,
8432 // however this code takes a simpler approach and just executes all
8433 // of the stores if %al is non-zero. It's less code, and it's probably
8434 // easier on the hardware branch predictor, and stores aren't all that
8435 // expensive anyway.
8437 // Create the new basic blocks. One block contains all the XMM stores,
8438 // and one block is the final destination regardless of whether any
8439 // stores were performed.
8440 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8441 MachineFunction *F = MBB->getParent();
8442 MachineFunction::iterator MBBIter = MBB;
8444 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8445 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8446 F->insert(MBBIter, XMMSaveMBB);
8447 F->insert(MBBIter, EndMBB);
8450 // Move any original successors of MBB to the end block.
8451 EndMBB->transferSuccessors(MBB);
8452 // The original block will now fall through to the XMM save block.
8453 MBB->addSuccessor(XMMSaveMBB);
8454 // The XMMSaveMBB will fall through to the end block.
8455 XMMSaveMBB->addSuccessor(EndMBB);
8457 // Now add the instructions.
8458 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8459 DebugLoc DL = MI->getDebugLoc();
8461 unsigned CountReg = MI->getOperand(0).getReg();
8462 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8463 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8465 if (!Subtarget->isTargetWin64()) {
8466 // If %al is 0, branch around the XMM save block.
8467 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8468 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8469 MBB->addSuccessor(EndMBB);
8472 // In the XMM save block, save all the XMM argument registers.
8473 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8474 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8475 MachineMemOperand *MMO =
8476 F->getMachineMemOperand(
8477 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8478 MachineMemOperand::MOStore, Offset,
8479 /*Size=*/16, /*Align=*/16);
8480 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8481 .addFrameIndex(RegSaveFrameIndex)
8482 .addImm(/*Scale=*/1)
8483 .addReg(/*IndexReg=*/0)
8484 .addImm(/*Disp=*/Offset)
8485 .addReg(/*Segment=*/0)
8486 .addReg(MI->getOperand(i).getReg())
8487 .addMemOperand(MMO);
8490 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8496 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8497 MachineBasicBlock *BB,
8498 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8499 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8500 DebugLoc DL = MI->getDebugLoc();
8502 // To "insert" a SELECT_CC instruction, we actually have to insert the
8503 // diamond control-flow pattern. The incoming instruction knows the
8504 // destination vreg to set, the condition code register to branch on, the
8505 // true/false values to select between, and a branch opcode to use.
8506 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8507 MachineFunction::iterator It = BB;
8513 // cmpTY ccX, r1, r2
8515 // fallthrough --> copy0MBB
8516 MachineBasicBlock *thisMBB = BB;
8517 MachineFunction *F = BB->getParent();
8518 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8519 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8521 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8522 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8523 F->insert(It, copy0MBB);
8524 F->insert(It, sinkMBB);
8525 // Update machine-CFG edges by first adding all successors of the current
8526 // block to the new block which will contain the Phi node for the select.
8527 // Also inform sdisel of the edge changes.
8528 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8529 E = BB->succ_end(); I != E; ++I) {
8530 EM->insert(std::make_pair(*I, sinkMBB));
8531 sinkMBB->addSuccessor(*I);
8533 // Next, remove all successors of the current block, and add the true
8534 // and fallthrough blocks as its successors.
8535 while (!BB->succ_empty())
8536 BB->removeSuccessor(BB->succ_begin());
8537 // Add the true and fallthrough blocks as its successors.
8538 BB->addSuccessor(copy0MBB);
8539 BB->addSuccessor(sinkMBB);
8542 // %FalseValue = ...
8543 // # fallthrough to sinkMBB
8546 // Update machine-CFG edges
8547 BB->addSuccessor(sinkMBB);
8550 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8553 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8554 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8555 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8557 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8562 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8563 MachineBasicBlock *BB,
8564 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8565 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8566 DebugLoc DL = MI->getDebugLoc();
8567 MachineFunction *F = BB->getParent();
8569 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8570 // non-trivial part is impdef of ESP.
8571 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8574 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8575 .addExternalSymbol("_alloca")
8576 .addReg(X86::EAX, RegState::Implicit)
8577 .addReg(X86::ESP, RegState::Implicit)
8578 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8579 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8581 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8586 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8587 MachineBasicBlock *BB,
8588 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8589 switch (MI->getOpcode()) {
8590 default: assert(false && "Unexpected instr type to insert");
8591 case X86::MINGW_ALLOCA:
8592 return EmitLoweredMingwAlloca(MI, BB, EM);
8594 case X86::CMOV_V1I64:
8595 case X86::CMOV_FR32:
8596 case X86::CMOV_FR64:
8597 case X86::CMOV_V4F32:
8598 case X86::CMOV_V2F64:
8599 case X86::CMOV_V2I64:
8600 case X86::CMOV_GR16:
8601 case X86::CMOV_GR32:
8602 case X86::CMOV_RFP32:
8603 case X86::CMOV_RFP64:
8604 case X86::CMOV_RFP80:
8605 return EmitLoweredSelect(MI, BB, EM);
8607 case X86::FP32_TO_INT16_IN_MEM:
8608 case X86::FP32_TO_INT32_IN_MEM:
8609 case X86::FP32_TO_INT64_IN_MEM:
8610 case X86::FP64_TO_INT16_IN_MEM:
8611 case X86::FP64_TO_INT32_IN_MEM:
8612 case X86::FP64_TO_INT64_IN_MEM:
8613 case X86::FP80_TO_INT16_IN_MEM:
8614 case X86::FP80_TO_INT32_IN_MEM:
8615 case X86::FP80_TO_INT64_IN_MEM: {
8616 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8617 DebugLoc DL = MI->getDebugLoc();
8619 // Change the floating point control register to use "round towards zero"
8620 // mode when truncating to an integer value.
8621 MachineFunction *F = BB->getParent();
8622 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8623 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8625 // Load the old value of the high byte of the control word...
8627 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8628 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8631 // Set the high part to be round to zero...
8632 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8635 // Reload the modified control word now...
8636 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8638 // Restore the memory image of control word to original value
8639 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8642 // Get the X86 opcode to use.
8644 switch (MI->getOpcode()) {
8645 default: llvm_unreachable("illegal opcode!");
8646 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8647 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8648 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8649 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8650 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8651 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8652 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8653 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8654 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8658 MachineOperand &Op = MI->getOperand(0);
8660 AM.BaseType = X86AddressMode::RegBase;
8661 AM.Base.Reg = Op.getReg();
8663 AM.BaseType = X86AddressMode::FrameIndexBase;
8664 AM.Base.FrameIndex = Op.getIndex();
8666 Op = MI->getOperand(1);
8668 AM.Scale = Op.getImm();
8669 Op = MI->getOperand(2);
8671 AM.IndexReg = Op.getImm();
8672 Op = MI->getOperand(3);
8673 if (Op.isGlobal()) {
8674 AM.GV = Op.getGlobal();
8676 AM.Disp = Op.getImm();
8678 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8679 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8681 // Reload the original control word now.
8682 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8684 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8687 // String/text processing lowering.
8688 case X86::PCMPISTRM128REG:
8689 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8690 case X86::PCMPISTRM128MEM:
8691 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8692 case X86::PCMPESTRM128REG:
8693 return EmitPCMP(MI, BB, 5, false /* in mem */);
8694 case X86::PCMPESTRM128MEM:
8695 return EmitPCMP(MI, BB, 5, true /* in mem */);
8698 case X86::ATOMAND32:
8699 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8700 X86::AND32ri, X86::MOV32rm,
8701 X86::LCMPXCHG32, X86::MOV32rr,
8702 X86::NOT32r, X86::EAX,
8703 X86::GR32RegisterClass);
8705 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8706 X86::OR32ri, X86::MOV32rm,
8707 X86::LCMPXCHG32, X86::MOV32rr,
8708 X86::NOT32r, X86::EAX,
8709 X86::GR32RegisterClass);
8710 case X86::ATOMXOR32:
8711 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8712 X86::XOR32ri, X86::MOV32rm,
8713 X86::LCMPXCHG32, X86::MOV32rr,
8714 X86::NOT32r, X86::EAX,
8715 X86::GR32RegisterClass);
8716 case X86::ATOMNAND32:
8717 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8718 X86::AND32ri, X86::MOV32rm,
8719 X86::LCMPXCHG32, X86::MOV32rr,
8720 X86::NOT32r, X86::EAX,
8721 X86::GR32RegisterClass, true);
8722 case X86::ATOMMIN32:
8723 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8724 case X86::ATOMMAX32:
8725 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8726 case X86::ATOMUMIN32:
8727 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8728 case X86::ATOMUMAX32:
8729 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8731 case X86::ATOMAND16:
8732 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8733 X86::AND16ri, X86::MOV16rm,
8734 X86::LCMPXCHG16, X86::MOV16rr,
8735 X86::NOT16r, X86::AX,
8736 X86::GR16RegisterClass);
8738 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8739 X86::OR16ri, X86::MOV16rm,
8740 X86::LCMPXCHG16, X86::MOV16rr,
8741 X86::NOT16r, X86::AX,
8742 X86::GR16RegisterClass);
8743 case X86::ATOMXOR16:
8744 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8745 X86::XOR16ri, X86::MOV16rm,
8746 X86::LCMPXCHG16, X86::MOV16rr,
8747 X86::NOT16r, X86::AX,
8748 X86::GR16RegisterClass);
8749 case X86::ATOMNAND16:
8750 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8751 X86::AND16ri, X86::MOV16rm,
8752 X86::LCMPXCHG16, X86::MOV16rr,
8753 X86::NOT16r, X86::AX,
8754 X86::GR16RegisterClass, true);
8755 case X86::ATOMMIN16:
8756 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8757 case X86::ATOMMAX16:
8758 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8759 case X86::ATOMUMIN16:
8760 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8761 case X86::ATOMUMAX16:
8762 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8765 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8766 X86::AND8ri, X86::MOV8rm,
8767 X86::LCMPXCHG8, X86::MOV8rr,
8768 X86::NOT8r, X86::AL,
8769 X86::GR8RegisterClass);
8771 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8772 X86::OR8ri, X86::MOV8rm,
8773 X86::LCMPXCHG8, X86::MOV8rr,
8774 X86::NOT8r, X86::AL,
8775 X86::GR8RegisterClass);
8777 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8778 X86::XOR8ri, X86::MOV8rm,
8779 X86::LCMPXCHG8, X86::MOV8rr,
8780 X86::NOT8r, X86::AL,
8781 X86::GR8RegisterClass);
8782 case X86::ATOMNAND8:
8783 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8784 X86::AND8ri, X86::MOV8rm,
8785 X86::LCMPXCHG8, X86::MOV8rr,
8786 X86::NOT8r, X86::AL,
8787 X86::GR8RegisterClass, true);
8788 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8789 // This group is for 64-bit host.
8790 case X86::ATOMAND64:
8791 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8792 X86::AND64ri32, X86::MOV64rm,
8793 X86::LCMPXCHG64, X86::MOV64rr,
8794 X86::NOT64r, X86::RAX,
8795 X86::GR64RegisterClass);
8797 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8798 X86::OR64ri32, X86::MOV64rm,
8799 X86::LCMPXCHG64, X86::MOV64rr,
8800 X86::NOT64r, X86::RAX,
8801 X86::GR64RegisterClass);
8802 case X86::ATOMXOR64:
8803 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8804 X86::XOR64ri32, X86::MOV64rm,
8805 X86::LCMPXCHG64, X86::MOV64rr,
8806 X86::NOT64r, X86::RAX,
8807 X86::GR64RegisterClass);
8808 case X86::ATOMNAND64:
8809 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8810 X86::AND64ri32, X86::MOV64rm,
8811 X86::LCMPXCHG64, X86::MOV64rr,
8812 X86::NOT64r, X86::RAX,
8813 X86::GR64RegisterClass, true);
8814 case X86::ATOMMIN64:
8815 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8816 case X86::ATOMMAX64:
8817 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8818 case X86::ATOMUMIN64:
8819 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8820 case X86::ATOMUMAX64:
8821 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8823 // This group does 64-bit operations on a 32-bit host.
8824 case X86::ATOMAND6432:
8825 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8826 X86::AND32rr, X86::AND32rr,
8827 X86::AND32ri, X86::AND32ri,
8829 case X86::ATOMOR6432:
8830 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8831 X86::OR32rr, X86::OR32rr,
8832 X86::OR32ri, X86::OR32ri,
8834 case X86::ATOMXOR6432:
8835 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8836 X86::XOR32rr, X86::XOR32rr,
8837 X86::XOR32ri, X86::XOR32ri,
8839 case X86::ATOMNAND6432:
8840 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8841 X86::AND32rr, X86::AND32rr,
8842 X86::AND32ri, X86::AND32ri,
8844 case X86::ATOMADD6432:
8845 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8846 X86::ADD32rr, X86::ADC32rr,
8847 X86::ADD32ri, X86::ADC32ri,
8849 case X86::ATOMSUB6432:
8850 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8851 X86::SUB32rr, X86::SBB32rr,
8852 X86::SUB32ri, X86::SBB32ri,
8854 case X86::ATOMSWAP6432:
8855 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8856 X86::MOV32rr, X86::MOV32rr,
8857 X86::MOV32ri, X86::MOV32ri,
8859 case X86::VASTART_SAVE_XMM_REGS:
8860 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8864 //===----------------------------------------------------------------------===//
8865 // X86 Optimization Hooks
8866 //===----------------------------------------------------------------------===//
8868 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8872 const SelectionDAG &DAG,
8873 unsigned Depth) const {
8874 unsigned Opc = Op.getOpcode();
8875 assert((Opc >= ISD::BUILTIN_OP_END ||
8876 Opc == ISD::INTRINSIC_WO_CHAIN ||
8877 Opc == ISD::INTRINSIC_W_CHAIN ||
8878 Opc == ISD::INTRINSIC_VOID) &&
8879 "Should use MaskedValueIsZero if you don't know whether Op"
8880 " is a target node!");
8882 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8894 // These nodes' second result is a boolean.
8895 if (Op.getResNo() == 0)
8899 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8900 Mask.getBitWidth() - 1);
8905 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8906 /// node is a GlobalAddress + offset.
8907 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8908 const GlobalValue* &GA,
8909 int64_t &Offset) const {
8910 if (N->getOpcode() == X86ISD::Wrapper) {
8911 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8912 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8913 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8917 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8920 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8921 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8922 /// if the load addresses are consecutive, non-overlapping, and in the right
8924 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8925 const TargetLowering &TLI) {
8926 DebugLoc dl = N->getDebugLoc();
8927 EVT VT = N->getValueType(0);
8928 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8930 if (VT.getSizeInBits() != 128)
8933 SmallVector<SDValue, 16> Elts;
8934 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8935 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8937 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8940 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8941 /// and convert it from being a bunch of shuffles and extracts to a simple
8942 /// store and scalar loads to extract the elements.
8943 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8944 const TargetLowering &TLI) {
8945 SDValue InputVector = N->getOperand(0);
8947 // Only operate on vectors of 4 elements, where the alternative shuffling
8948 // gets to be more expensive.
8949 if (InputVector.getValueType() != MVT::v4i32)
8952 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8953 // single use which is a sign-extend or zero-extend, and all elements are
8955 SmallVector<SDNode *, 4> Uses;
8956 unsigned ExtractedElements = 0;
8957 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8958 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8959 if (UI.getUse().getResNo() != InputVector.getResNo())
8962 SDNode *Extract = *UI;
8963 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8966 if (Extract->getValueType(0) != MVT::i32)
8968 if (!Extract->hasOneUse())
8970 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
8971 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
8973 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
8976 // Record which element was extracted.
8977 ExtractedElements |=
8978 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
8980 Uses.push_back(Extract);
8983 // If not all the elements were used, this may not be worthwhile.
8984 if (ExtractedElements != 15)
8987 // Ok, we've now decided to do the transformation.
8988 DebugLoc dl = InputVector.getDebugLoc();
8990 // Store the value to a temporary stack slot.
8991 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
8992 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
8995 // Replace each use (extract) with a load of the appropriate element.
8996 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
8997 UE = Uses.end(); UI != UE; ++UI) {
8998 SDNode *Extract = *UI;
9000 // Compute the element's address.
9001 SDValue Idx = Extract->getOperand(1);
9003 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9004 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9005 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9007 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
9010 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
9011 NULL, 0, false, false, 0);
9013 // Replace the exact with the load.
9014 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9017 // The replacement was made in place; don't return anything.
9021 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9022 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9023 const X86Subtarget *Subtarget) {
9024 DebugLoc DL = N->getDebugLoc();
9025 SDValue Cond = N->getOperand(0);
9026 // Get the LHS/RHS of the select.
9027 SDValue LHS = N->getOperand(1);
9028 SDValue RHS = N->getOperand(2);
9030 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9031 // instructions match the semantics of the common C idiom x<y?x:y but not
9032 // x<=y?x:y, because of how they handle negative zero (which can be
9033 // ignored in unsafe-math mode).
9034 if (Subtarget->hasSSE2() &&
9035 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9036 Cond.getOpcode() == ISD::SETCC) {
9037 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9039 unsigned Opcode = 0;
9040 // Check for x CC y ? x : y.
9041 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9042 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9046 // Converting this to a min would handle NaNs incorrectly, and swapping
9047 // the operands would cause it to handle comparisons between positive
9048 // and negative zero incorrectly.
9049 if (!FiniteOnlyFPMath() &&
9050 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9051 if (!UnsafeFPMath &&
9052 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9054 std::swap(LHS, RHS);
9056 Opcode = X86ISD::FMIN;
9059 // Converting this to a min would handle comparisons between positive
9060 // and negative zero incorrectly.
9061 if (!UnsafeFPMath &&
9062 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9064 Opcode = X86ISD::FMIN;
9067 // Converting this to a min would handle both negative zeros and NaNs
9068 // incorrectly, but we can swap the operands to fix both.
9069 std::swap(LHS, RHS);
9073 Opcode = X86ISD::FMIN;
9077 // Converting this to a max would handle comparisons between positive
9078 // and negative zero incorrectly.
9079 if (!UnsafeFPMath &&
9080 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9082 Opcode = X86ISD::FMAX;
9085 // Converting this to a max would handle NaNs incorrectly, and swapping
9086 // the operands would cause it to handle comparisons between positive
9087 // and negative zero incorrectly.
9088 if (!FiniteOnlyFPMath() &&
9089 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9090 if (!UnsafeFPMath &&
9091 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9093 std::swap(LHS, RHS);
9095 Opcode = X86ISD::FMAX;
9098 // Converting this to a max would handle both negative zeros and NaNs
9099 // incorrectly, but we can swap the operands to fix both.
9100 std::swap(LHS, RHS);
9104 Opcode = X86ISD::FMAX;
9107 // Check for x CC y ? y : x -- a min/max with reversed arms.
9108 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9109 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9113 // Converting this to a min would handle comparisons between positive
9114 // and negative zero incorrectly, and swapping the operands would
9115 // cause it to handle NaNs incorrectly.
9116 if (!UnsafeFPMath &&
9117 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9118 if (!FiniteOnlyFPMath() &&
9119 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9121 std::swap(LHS, RHS);
9123 Opcode = X86ISD::FMIN;
9126 // Converting this to a min would handle NaNs incorrectly.
9127 if (!UnsafeFPMath &&
9128 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9130 Opcode = X86ISD::FMIN;
9133 // Converting this to a min would handle both negative zeros and NaNs
9134 // incorrectly, but we can swap the operands to fix both.
9135 std::swap(LHS, RHS);
9139 Opcode = X86ISD::FMIN;
9143 // Converting this to a max would handle NaNs incorrectly.
9144 if (!FiniteOnlyFPMath() &&
9145 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9147 Opcode = X86ISD::FMAX;
9150 // Converting this to a max would handle comparisons between positive
9151 // and negative zero incorrectly, and swapping the operands would
9152 // cause it to handle NaNs incorrectly.
9153 if (!UnsafeFPMath &&
9154 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9155 if (!FiniteOnlyFPMath() &&
9156 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9158 std::swap(LHS, RHS);
9160 Opcode = X86ISD::FMAX;
9163 // Converting this to a max would handle both negative zeros and NaNs
9164 // incorrectly, but we can swap the operands to fix both.
9165 std::swap(LHS, RHS);
9169 Opcode = X86ISD::FMAX;
9175 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9178 // If this is a select between two integer constants, try to do some
9180 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9181 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9182 // Don't do this for crazy integer types.
9183 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9184 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9185 // so that TrueC (the true value) is larger than FalseC.
9186 bool NeedsCondInvert = false;
9188 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9189 // Efficiently invertible.
9190 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9191 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9192 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9193 NeedsCondInvert = true;
9194 std::swap(TrueC, FalseC);
9197 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9198 if (FalseC->getAPIntValue() == 0 &&
9199 TrueC->getAPIntValue().isPowerOf2()) {
9200 if (NeedsCondInvert) // Invert the condition if needed.
9201 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9202 DAG.getConstant(1, Cond.getValueType()));
9204 // Zero extend the condition if needed.
9205 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9207 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9208 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9209 DAG.getConstant(ShAmt, MVT::i8));
9212 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9213 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9214 if (NeedsCondInvert) // Invert the condition if needed.
9215 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9216 DAG.getConstant(1, Cond.getValueType()));
9218 // Zero extend the condition if needed.
9219 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9220 FalseC->getValueType(0), Cond);
9221 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9222 SDValue(FalseC, 0));
9225 // Optimize cases that will turn into an LEA instruction. This requires
9226 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9227 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9228 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9229 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9231 bool isFastMultiplier = false;
9233 switch ((unsigned char)Diff) {
9235 case 1: // result = add base, cond
9236 case 2: // result = lea base( , cond*2)
9237 case 3: // result = lea base(cond, cond*2)
9238 case 4: // result = lea base( , cond*4)
9239 case 5: // result = lea base(cond, cond*4)
9240 case 8: // result = lea base( , cond*8)
9241 case 9: // result = lea base(cond, cond*8)
9242 isFastMultiplier = true;
9247 if (isFastMultiplier) {
9248 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9249 if (NeedsCondInvert) // Invert the condition if needed.
9250 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9251 DAG.getConstant(1, Cond.getValueType()));
9253 // Zero extend the condition if needed.
9254 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9256 // Scale the condition by the difference.
9258 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9259 DAG.getConstant(Diff, Cond.getValueType()));
9261 // Add the base if non-zero.
9262 if (FalseC->getAPIntValue() != 0)
9263 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9264 SDValue(FalseC, 0));
9274 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9275 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9276 TargetLowering::DAGCombinerInfo &DCI) {
9277 DebugLoc DL = N->getDebugLoc();
9279 // If the flag operand isn't dead, don't touch this CMOV.
9280 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9283 // If this is a select between two integer constants, try to do some
9284 // optimizations. Note that the operands are ordered the opposite of SELECT
9286 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9287 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9288 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9289 // larger than FalseC (the false value).
9290 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9292 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9293 CC = X86::GetOppositeBranchCondition(CC);
9294 std::swap(TrueC, FalseC);
9297 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9298 // This is efficient for any integer data type (including i8/i16) and
9300 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9301 SDValue Cond = N->getOperand(3);
9302 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9303 DAG.getConstant(CC, MVT::i8), Cond);
9305 // Zero extend the condition if needed.
9306 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9308 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9309 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9310 DAG.getConstant(ShAmt, MVT::i8));
9311 if (N->getNumValues() == 2) // Dead flag value?
9312 return DCI.CombineTo(N, Cond, SDValue());
9316 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9317 // for any integer data type, including i8/i16.
9318 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9319 SDValue Cond = N->getOperand(3);
9320 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9321 DAG.getConstant(CC, MVT::i8), Cond);
9323 // Zero extend the condition if needed.
9324 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9325 FalseC->getValueType(0), Cond);
9326 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9327 SDValue(FalseC, 0));
9329 if (N->getNumValues() == 2) // Dead flag value?
9330 return DCI.CombineTo(N, Cond, SDValue());
9334 // Optimize cases that will turn into an LEA instruction. This requires
9335 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9336 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9337 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9338 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9340 bool isFastMultiplier = false;
9342 switch ((unsigned char)Diff) {
9344 case 1: // result = add base, cond
9345 case 2: // result = lea base( , cond*2)
9346 case 3: // result = lea base(cond, cond*2)
9347 case 4: // result = lea base( , cond*4)
9348 case 5: // result = lea base(cond, cond*4)
9349 case 8: // result = lea base( , cond*8)
9350 case 9: // result = lea base(cond, cond*8)
9351 isFastMultiplier = true;
9356 if (isFastMultiplier) {
9357 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9358 SDValue Cond = N->getOperand(3);
9359 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9360 DAG.getConstant(CC, MVT::i8), Cond);
9361 // Zero extend the condition if needed.
9362 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9364 // Scale the condition by the difference.
9366 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9367 DAG.getConstant(Diff, Cond.getValueType()));
9369 // Add the base if non-zero.
9370 if (FalseC->getAPIntValue() != 0)
9371 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9372 SDValue(FalseC, 0));
9373 if (N->getNumValues() == 2) // Dead flag value?
9374 return DCI.CombineTo(N, Cond, SDValue());
9384 /// PerformMulCombine - Optimize a single multiply with constant into two
9385 /// in order to implement it with two cheaper instructions, e.g.
9386 /// LEA + SHL, LEA + LEA.
9387 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9388 TargetLowering::DAGCombinerInfo &DCI) {
9389 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9392 EVT VT = N->getValueType(0);
9396 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9399 uint64_t MulAmt = C->getZExtValue();
9400 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9403 uint64_t MulAmt1 = 0;
9404 uint64_t MulAmt2 = 0;
9405 if ((MulAmt % 9) == 0) {
9407 MulAmt2 = MulAmt / 9;
9408 } else if ((MulAmt % 5) == 0) {
9410 MulAmt2 = MulAmt / 5;
9411 } else if ((MulAmt % 3) == 0) {
9413 MulAmt2 = MulAmt / 3;
9416 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9417 DebugLoc DL = N->getDebugLoc();
9419 if (isPowerOf2_64(MulAmt2) &&
9420 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9421 // If second multiplifer is pow2, issue it first. We want the multiply by
9422 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9424 std::swap(MulAmt1, MulAmt2);
9427 if (isPowerOf2_64(MulAmt1))
9428 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9429 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9431 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9432 DAG.getConstant(MulAmt1, VT));
9434 if (isPowerOf2_64(MulAmt2))
9435 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9436 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9438 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9439 DAG.getConstant(MulAmt2, VT));
9441 // Do not add new nodes to DAG combiner worklist.
9442 DCI.CombineTo(N, NewMul, false);
9447 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9448 SDValue N0 = N->getOperand(0);
9449 SDValue N1 = N->getOperand(1);
9450 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9451 EVT VT = N0.getValueType();
9453 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9454 // since the result of setcc_c is all zero's or all ones.
9455 if (N1C && N0.getOpcode() == ISD::AND &&
9456 N0.getOperand(1).getOpcode() == ISD::Constant) {
9457 SDValue N00 = N0.getOperand(0);
9458 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9459 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9460 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9461 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9462 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9463 APInt ShAmt = N1C->getAPIntValue();
9464 Mask = Mask.shl(ShAmt);
9466 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9467 N00, DAG.getConstant(Mask, VT));
9474 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9476 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9477 const X86Subtarget *Subtarget) {
9478 EVT VT = N->getValueType(0);
9479 if (!VT.isVector() && VT.isInteger() &&
9480 N->getOpcode() == ISD::SHL)
9481 return PerformSHLCombine(N, DAG);
9483 // On X86 with SSE2 support, we can transform this to a vector shift if
9484 // all elements are shifted by the same amount. We can't do this in legalize
9485 // because the a constant vector is typically transformed to a constant pool
9486 // so we have no knowledge of the shift amount.
9487 if (!Subtarget->hasSSE2())
9490 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9493 SDValue ShAmtOp = N->getOperand(1);
9494 EVT EltVT = VT.getVectorElementType();
9495 DebugLoc DL = N->getDebugLoc();
9496 SDValue BaseShAmt = SDValue();
9497 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9498 unsigned NumElts = VT.getVectorNumElements();
9500 for (; i != NumElts; ++i) {
9501 SDValue Arg = ShAmtOp.getOperand(i);
9502 if (Arg.getOpcode() == ISD::UNDEF) continue;
9506 for (; i != NumElts; ++i) {
9507 SDValue Arg = ShAmtOp.getOperand(i);
9508 if (Arg.getOpcode() == ISD::UNDEF) continue;
9509 if (Arg != BaseShAmt) {
9513 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9514 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9515 SDValue InVec = ShAmtOp.getOperand(0);
9516 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9517 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9519 for (; i != NumElts; ++i) {
9520 SDValue Arg = InVec.getOperand(i);
9521 if (Arg.getOpcode() == ISD::UNDEF) continue;
9525 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9526 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9527 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9528 if (C->getZExtValue() == SplatIdx)
9529 BaseShAmt = InVec.getOperand(1);
9532 if (BaseShAmt.getNode() == 0)
9533 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9534 DAG.getIntPtrConstant(0));
9538 // The shift amount is an i32.
9539 if (EltVT.bitsGT(MVT::i32))
9540 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9541 else if (EltVT.bitsLT(MVT::i32))
9542 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9544 // The shift amount is identical so we can do a vector shift.
9545 SDValue ValOp = N->getOperand(0);
9546 switch (N->getOpcode()) {
9548 llvm_unreachable("Unknown shift opcode!");
9551 if (VT == MVT::v2i64)
9552 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9553 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9555 if (VT == MVT::v4i32)
9556 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9557 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9559 if (VT == MVT::v8i16)
9560 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9561 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9565 if (VT == MVT::v4i32)
9566 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9567 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9569 if (VT == MVT::v8i16)
9570 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9571 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9575 if (VT == MVT::v2i64)
9576 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9577 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9579 if (VT == MVT::v4i32)
9580 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9581 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9583 if (VT == MVT::v8i16)
9584 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9585 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9592 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9593 const X86Subtarget *Subtarget) {
9594 EVT VT = N->getValueType(0);
9595 if (VT != MVT::i64 || !Subtarget->is64Bit())
9598 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9599 SDValue N0 = N->getOperand(0);
9600 SDValue N1 = N->getOperand(1);
9601 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9603 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9606 SDValue ShAmt0 = N0.getOperand(1);
9607 if (ShAmt0.getValueType() != MVT::i8)
9609 SDValue ShAmt1 = N1.getOperand(1);
9610 if (ShAmt1.getValueType() != MVT::i8)
9612 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9613 ShAmt0 = ShAmt0.getOperand(0);
9614 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9615 ShAmt1 = ShAmt1.getOperand(0);
9617 DebugLoc DL = N->getDebugLoc();
9618 unsigned Opc = X86ISD::SHLD;
9619 SDValue Op0 = N0.getOperand(0);
9620 SDValue Op1 = N1.getOperand(0);
9621 if (ShAmt0.getOpcode() == ISD::SUB) {
9623 std::swap(Op0, Op1);
9624 std::swap(ShAmt0, ShAmt1);
9627 if (ShAmt1.getOpcode() == ISD::SUB) {
9628 SDValue Sum = ShAmt1.getOperand(0);
9629 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9630 if (SumC->getSExtValue() == 64 &&
9631 ShAmt1.getOperand(1) == ShAmt0)
9632 return DAG.getNode(Opc, DL, VT,
9634 DAG.getNode(ISD::TRUNCATE, DL,
9637 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9638 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9640 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9641 return DAG.getNode(Opc, DL, VT,
9642 N0.getOperand(0), N1.getOperand(0),
9643 DAG.getNode(ISD::TRUNCATE, DL,
9650 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9651 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9652 const X86Subtarget *Subtarget) {
9653 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9654 // the FP state in cases where an emms may be missing.
9655 // A preferable solution to the general problem is to figure out the right
9656 // places to insert EMMS. This qualifies as a quick hack.
9658 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9659 StoreSDNode *St = cast<StoreSDNode>(N);
9660 EVT VT = St->getValue().getValueType();
9661 if (VT.getSizeInBits() != 64)
9664 const Function *F = DAG.getMachineFunction().getFunction();
9665 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9666 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9667 && Subtarget->hasSSE2();
9668 if ((VT.isVector() ||
9669 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9670 isa<LoadSDNode>(St->getValue()) &&
9671 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9672 St->getChain().hasOneUse() && !St->isVolatile()) {
9673 SDNode* LdVal = St->getValue().getNode();
9675 int TokenFactorIndex = -1;
9676 SmallVector<SDValue, 8> Ops;
9677 SDNode* ChainVal = St->getChain().getNode();
9678 // Must be a store of a load. We currently handle two cases: the load
9679 // is a direct child, and it's under an intervening TokenFactor. It is
9680 // possible to dig deeper under nested TokenFactors.
9681 if (ChainVal == LdVal)
9682 Ld = cast<LoadSDNode>(St->getChain());
9683 else if (St->getValue().hasOneUse() &&
9684 ChainVal->getOpcode() == ISD::TokenFactor) {
9685 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9686 if (ChainVal->getOperand(i).getNode() == LdVal) {
9687 TokenFactorIndex = i;
9688 Ld = cast<LoadSDNode>(St->getValue());
9690 Ops.push_back(ChainVal->getOperand(i));
9694 if (!Ld || !ISD::isNormalLoad(Ld))
9697 // If this is not the MMX case, i.e. we are just turning i64 load/store
9698 // into f64 load/store, avoid the transformation if there are multiple
9699 // uses of the loaded value.
9700 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9703 DebugLoc LdDL = Ld->getDebugLoc();
9704 DebugLoc StDL = N->getDebugLoc();
9705 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9706 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9708 if (Subtarget->is64Bit() || F64IsLegal) {
9709 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9710 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9711 Ld->getBasePtr(), Ld->getSrcValue(),
9712 Ld->getSrcValueOffset(), Ld->isVolatile(),
9713 Ld->isNonTemporal(), Ld->getAlignment());
9714 SDValue NewChain = NewLd.getValue(1);
9715 if (TokenFactorIndex != -1) {
9716 Ops.push_back(NewChain);
9717 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9720 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9721 St->getSrcValue(), St->getSrcValueOffset(),
9722 St->isVolatile(), St->isNonTemporal(),
9723 St->getAlignment());
9726 // Otherwise, lower to two pairs of 32-bit loads / stores.
9727 SDValue LoAddr = Ld->getBasePtr();
9728 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9729 DAG.getConstant(4, MVT::i32));
9731 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9732 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9733 Ld->isVolatile(), Ld->isNonTemporal(),
9734 Ld->getAlignment());
9735 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9736 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9737 Ld->isVolatile(), Ld->isNonTemporal(),
9738 MinAlign(Ld->getAlignment(), 4));
9740 SDValue NewChain = LoLd.getValue(1);
9741 if (TokenFactorIndex != -1) {
9742 Ops.push_back(LoLd);
9743 Ops.push_back(HiLd);
9744 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9748 LoAddr = St->getBasePtr();
9749 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9750 DAG.getConstant(4, MVT::i32));
9752 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9753 St->getSrcValue(), St->getSrcValueOffset(),
9754 St->isVolatile(), St->isNonTemporal(),
9755 St->getAlignment());
9756 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9758 St->getSrcValueOffset() + 4,
9760 St->isNonTemporal(),
9761 MinAlign(St->getAlignment(), 4));
9762 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9767 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9768 /// X86ISD::FXOR nodes.
9769 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9770 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9771 // F[X]OR(0.0, x) -> x
9772 // F[X]OR(x, 0.0) -> x
9773 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9774 if (C->getValueAPF().isPosZero())
9775 return N->getOperand(1);
9776 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9777 if (C->getValueAPF().isPosZero())
9778 return N->getOperand(0);
9782 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9783 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9784 // FAND(0.0, x) -> 0.0
9785 // FAND(x, 0.0) -> 0.0
9786 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9787 if (C->getValueAPF().isPosZero())
9788 return N->getOperand(0);
9789 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9790 if (C->getValueAPF().isPosZero())
9791 return N->getOperand(1);
9795 static SDValue PerformBTCombine(SDNode *N,
9797 TargetLowering::DAGCombinerInfo &DCI) {
9798 // BT ignores high bits in the bit index operand.
9799 SDValue Op1 = N->getOperand(1);
9800 if (Op1.hasOneUse()) {
9801 unsigned BitWidth = Op1.getValueSizeInBits();
9802 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9803 APInt KnownZero, KnownOne;
9804 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9805 !DCI.isBeforeLegalizeOps());
9806 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9807 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9808 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9809 DCI.CommitTargetLoweringOpt(TLO);
9814 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9815 SDValue Op = N->getOperand(0);
9816 if (Op.getOpcode() == ISD::BIT_CONVERT)
9817 Op = Op.getOperand(0);
9818 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9819 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9820 VT.getVectorElementType().getSizeInBits() ==
9821 OpVT.getVectorElementType().getSizeInBits()) {
9822 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9827 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9828 // Locked instructions, in turn, have implicit fence semantics (all memory
9829 // operations are flushed before issuing the locked instruction, and the
9830 // are not buffered), so we can fold away the common pattern of
9831 // fence-atomic-fence.
9832 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9833 SDValue atomic = N->getOperand(0);
9834 switch (atomic.getOpcode()) {
9835 case ISD::ATOMIC_CMP_SWAP:
9836 case ISD::ATOMIC_SWAP:
9837 case ISD::ATOMIC_LOAD_ADD:
9838 case ISD::ATOMIC_LOAD_SUB:
9839 case ISD::ATOMIC_LOAD_AND:
9840 case ISD::ATOMIC_LOAD_OR:
9841 case ISD::ATOMIC_LOAD_XOR:
9842 case ISD::ATOMIC_LOAD_NAND:
9843 case ISD::ATOMIC_LOAD_MIN:
9844 case ISD::ATOMIC_LOAD_MAX:
9845 case ISD::ATOMIC_LOAD_UMIN:
9846 case ISD::ATOMIC_LOAD_UMAX:
9852 SDValue fence = atomic.getOperand(0);
9853 if (fence.getOpcode() != ISD::MEMBARRIER)
9856 switch (atomic.getOpcode()) {
9857 case ISD::ATOMIC_CMP_SWAP:
9858 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9859 atomic.getOperand(1), atomic.getOperand(2),
9860 atomic.getOperand(3));
9861 case ISD::ATOMIC_SWAP:
9862 case ISD::ATOMIC_LOAD_ADD:
9863 case ISD::ATOMIC_LOAD_SUB:
9864 case ISD::ATOMIC_LOAD_AND:
9865 case ISD::ATOMIC_LOAD_OR:
9866 case ISD::ATOMIC_LOAD_XOR:
9867 case ISD::ATOMIC_LOAD_NAND:
9868 case ISD::ATOMIC_LOAD_MIN:
9869 case ISD::ATOMIC_LOAD_MAX:
9870 case ISD::ATOMIC_LOAD_UMIN:
9871 case ISD::ATOMIC_LOAD_UMAX:
9872 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9873 atomic.getOperand(1), atomic.getOperand(2));
9879 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9880 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9881 // (and (i32 x86isd::setcc_carry), 1)
9882 // This eliminates the zext. This transformation is necessary because
9883 // ISD::SETCC is always legalized to i8.
9884 DebugLoc dl = N->getDebugLoc();
9885 SDValue N0 = N->getOperand(0);
9886 EVT VT = N->getValueType(0);
9887 if (N0.getOpcode() == ISD::AND &&
9889 N0.getOperand(0).hasOneUse()) {
9890 SDValue N00 = N0.getOperand(0);
9891 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9893 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9894 if (!C || C->getZExtValue() != 1)
9896 return DAG.getNode(ISD::AND, dl, VT,
9897 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9898 N00.getOperand(0), N00.getOperand(1)),
9899 DAG.getConstant(1, VT));
9905 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9906 DAGCombinerInfo &DCI) const {
9907 SelectionDAG &DAG = DCI.DAG;
9908 switch (N->getOpcode()) {
9910 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9911 case ISD::EXTRACT_VECTOR_ELT:
9912 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9913 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9914 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9915 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9918 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9919 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9920 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9922 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9923 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9924 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9925 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9926 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9927 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9933 /// isTypeDesirableForOp - Return true if the target has native support for
9934 /// the specified value type and it is 'desirable' to use the type for the
9935 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9936 /// instruction encodings are longer and some i16 instructions are slow.
9937 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9938 if (!isTypeLegal(VT))
9940 if (!Subtarget->shouldPromote16Bit() || VT != MVT::i16)
9947 case ISD::SIGN_EXTEND:
9948 case ISD::ZERO_EXTEND:
9949 case ISD::ANY_EXTEND:
9963 static bool MayFoldLoad(SDValue Op) {
9964 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
9967 static bool MayFoldIntoStore(SDValue Op) {
9968 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
9971 /// IsDesirableToPromoteOp - This method query the target whether it is
9972 /// beneficial for dag combiner to promote the specified node. If true, it
9973 /// should return the desired promotion type by reference.
9974 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
9975 if (!Subtarget->shouldPromote16Bit())
9978 EVT VT = Op.getValueType();
9982 bool Promote = false;
9983 bool Commute = false;
9984 switch (Op.getOpcode()) {
9987 LoadSDNode *LD = cast<LoadSDNode>(Op);
9988 // If the non-extending load has a single use and it's not live out, then it
9990 if (LD->getExtensionType() == ISD::NON_EXTLOAD &&
9992 Op.getNode()->use_begin()->getOpcode() != ISD::CopyToReg)
9997 case ISD::SIGN_EXTEND:
9998 case ISD::ZERO_EXTEND:
9999 case ISD::ANY_EXTEND:
10005 SDValue N0 = Op.getOperand(0);
10006 // Look out for (store (shl (load), x)).
10007 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10020 SDValue N0 = Op.getOperand(0);
10021 SDValue N1 = Op.getOperand(1);
10022 if (!Commute && MayFoldLoad(N1))
10024 // Avoid disabling potential load folding opportunities.
10025 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10027 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10037 //===----------------------------------------------------------------------===//
10038 // X86 Inline Assembly Support
10039 //===----------------------------------------------------------------------===//
10041 static bool LowerToBSwap(CallInst *CI) {
10042 // FIXME: this should verify that we are targetting a 486 or better. If not,
10043 // we will turn this bswap into something that will be lowered to logical ops
10044 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10045 // so don't worry about this.
10047 // Verify this is a simple bswap.
10048 if (CI->getNumOperands() != 2 ||
10049 CI->getType() != CI->getOperand(1)->getType() ||
10050 !CI->getType()->isIntegerTy())
10053 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10054 if (!Ty || Ty->getBitWidth() % 16 != 0)
10057 // Okay, we can do this xform, do so now.
10058 const Type *Tys[] = { Ty };
10059 Module *M = CI->getParent()->getParent()->getParent();
10060 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10062 Value *Op = CI->getOperand(1);
10063 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10065 CI->replaceAllUsesWith(Op);
10066 CI->eraseFromParent();
10070 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10071 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10072 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10074 std::string AsmStr = IA->getAsmString();
10076 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10077 SmallVector<StringRef, 4> AsmPieces;
10078 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10080 switch (AsmPieces.size()) {
10081 default: return false;
10083 AsmStr = AsmPieces[0];
10085 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10088 if (AsmPieces.size() == 2 &&
10089 (AsmPieces[0] == "bswap" ||
10090 AsmPieces[0] == "bswapq" ||
10091 AsmPieces[0] == "bswapl") &&
10092 (AsmPieces[1] == "$0" ||
10093 AsmPieces[1] == "${0:q}")) {
10094 // No need to check constraints, nothing other than the equivalent of
10095 // "=r,0" would be valid here.
10096 return LowerToBSwap(CI);
10098 // rorw $$8, ${0:w} --> llvm.bswap.i16
10099 if (CI->getType()->isIntegerTy(16) &&
10100 AsmPieces.size() == 3 &&
10101 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10102 AsmPieces[1] == "$$8," &&
10103 AsmPieces[2] == "${0:w}" &&
10104 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10106 const std::string &Constraints = IA->getConstraintString();
10107 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10108 std::sort(AsmPieces.begin(), AsmPieces.end());
10109 if (AsmPieces.size() == 4 &&
10110 AsmPieces[0] == "~{cc}" &&
10111 AsmPieces[1] == "~{dirflag}" &&
10112 AsmPieces[2] == "~{flags}" &&
10113 AsmPieces[3] == "~{fpsr}") {
10114 return LowerToBSwap(CI);
10119 if (CI->getType()->isIntegerTy(64) &&
10120 Constraints.size() >= 2 &&
10121 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10122 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10123 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10124 SmallVector<StringRef, 4> Words;
10125 SplitString(AsmPieces[0], Words, " \t");
10126 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10128 SplitString(AsmPieces[1], Words, " \t");
10129 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10131 SplitString(AsmPieces[2], Words, " \t,");
10132 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10133 Words[2] == "%edx") {
10134 return LowerToBSwap(CI);
10146 /// getConstraintType - Given a constraint letter, return the type of
10147 /// constraint it is for this target.
10148 X86TargetLowering::ConstraintType
10149 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10150 if (Constraint.size() == 1) {
10151 switch (Constraint[0]) {
10163 return C_RegisterClass;
10171 return TargetLowering::getConstraintType(Constraint);
10174 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10175 /// with another that has more specific requirements based on the type of the
10176 /// corresponding operand.
10177 const char *X86TargetLowering::
10178 LowerXConstraint(EVT ConstraintVT) const {
10179 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10180 // 'f' like normal targets.
10181 if (ConstraintVT.isFloatingPoint()) {
10182 if (Subtarget->hasSSE2())
10184 if (Subtarget->hasSSE1())
10188 return TargetLowering::LowerXConstraint(ConstraintVT);
10191 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10192 /// vector. If it is invalid, don't add anything to Ops.
10193 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10196 std::vector<SDValue>&Ops,
10197 SelectionDAG &DAG) const {
10198 SDValue Result(0, 0);
10200 switch (Constraint) {
10203 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10204 if (C->getZExtValue() <= 31) {
10205 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10211 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10212 if (C->getZExtValue() <= 63) {
10213 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10219 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10220 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10221 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10227 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10228 if (C->getZExtValue() <= 255) {
10229 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10235 // 32-bit signed value
10236 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10237 const ConstantInt *CI = C->getConstantIntValue();
10238 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10239 C->getSExtValue())) {
10240 // Widen to 64 bits here to get it sign extended.
10241 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10244 // FIXME gcc accepts some relocatable values here too, but only in certain
10245 // memory models; it's complicated.
10250 // 32-bit unsigned value
10251 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10252 const ConstantInt *CI = C->getConstantIntValue();
10253 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10254 C->getZExtValue())) {
10255 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10259 // FIXME gcc accepts some relocatable values here too, but only in certain
10260 // memory models; it's complicated.
10264 // Literal immediates are always ok.
10265 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10266 // Widen to 64 bits here to get it sign extended.
10267 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10271 // If we are in non-pic codegen mode, we allow the address of a global (with
10272 // an optional displacement) to be used with 'i'.
10273 GlobalAddressSDNode *GA = 0;
10274 int64_t Offset = 0;
10276 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10278 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10279 Offset += GA->getOffset();
10281 } else if (Op.getOpcode() == ISD::ADD) {
10282 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10283 Offset += C->getZExtValue();
10284 Op = Op.getOperand(0);
10287 } else if (Op.getOpcode() == ISD::SUB) {
10288 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10289 Offset += -C->getZExtValue();
10290 Op = Op.getOperand(0);
10295 // Otherwise, this isn't something we can handle, reject it.
10299 const GlobalValue *GV = GA->getGlobal();
10300 // If we require an extra load to get this address, as in PIC mode, we
10301 // can't accept it.
10302 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10303 getTargetMachine())))
10307 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
10309 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10315 if (Result.getNode()) {
10316 Ops.push_back(Result);
10319 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
10323 std::vector<unsigned> X86TargetLowering::
10324 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10326 if (Constraint.size() == 1) {
10327 // FIXME: not handling fp-stack yet!
10328 switch (Constraint[0]) { // GCC X86 Constraint Letters
10329 default: break; // Unknown constraint letter
10330 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10331 if (Subtarget->is64Bit()) {
10332 if (VT == MVT::i32)
10333 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10334 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10335 X86::R10D,X86::R11D,X86::R12D,
10336 X86::R13D,X86::R14D,X86::R15D,
10337 X86::EBP, X86::ESP, 0);
10338 else if (VT == MVT::i16)
10339 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10340 X86::SI, X86::DI, X86::R8W,X86::R9W,
10341 X86::R10W,X86::R11W,X86::R12W,
10342 X86::R13W,X86::R14W,X86::R15W,
10343 X86::BP, X86::SP, 0);
10344 else if (VT == MVT::i8)
10345 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10346 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10347 X86::R10B,X86::R11B,X86::R12B,
10348 X86::R13B,X86::R14B,X86::R15B,
10349 X86::BPL, X86::SPL, 0);
10351 else if (VT == MVT::i64)
10352 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10353 X86::RSI, X86::RDI, X86::R8, X86::R9,
10354 X86::R10, X86::R11, X86::R12,
10355 X86::R13, X86::R14, X86::R15,
10356 X86::RBP, X86::RSP, 0);
10360 // 32-bit fallthrough
10361 case 'Q': // Q_REGS
10362 if (VT == MVT::i32)
10363 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10364 else if (VT == MVT::i16)
10365 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10366 else if (VT == MVT::i8)
10367 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10368 else if (VT == MVT::i64)
10369 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10374 return std::vector<unsigned>();
10377 std::pair<unsigned, const TargetRegisterClass*>
10378 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10380 // First, see if this is a constraint that directly corresponds to an LLVM
10382 if (Constraint.size() == 1) {
10383 // GCC Constraint Letters
10384 switch (Constraint[0]) {
10386 case 'r': // GENERAL_REGS
10387 case 'l': // INDEX_REGS
10389 return std::make_pair(0U, X86::GR8RegisterClass);
10390 if (VT == MVT::i16)
10391 return std::make_pair(0U, X86::GR16RegisterClass);
10392 if (VT == MVT::i32 || !Subtarget->is64Bit())
10393 return std::make_pair(0U, X86::GR32RegisterClass);
10394 return std::make_pair(0U, X86::GR64RegisterClass);
10395 case 'R': // LEGACY_REGS
10397 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10398 if (VT == MVT::i16)
10399 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10400 if (VT == MVT::i32 || !Subtarget->is64Bit())
10401 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10402 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10403 case 'f': // FP Stack registers.
10404 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10405 // value to the correct fpstack register class.
10406 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10407 return std::make_pair(0U, X86::RFP32RegisterClass);
10408 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10409 return std::make_pair(0U, X86::RFP64RegisterClass);
10410 return std::make_pair(0U, X86::RFP80RegisterClass);
10411 case 'y': // MMX_REGS if MMX allowed.
10412 if (!Subtarget->hasMMX()) break;
10413 return std::make_pair(0U, X86::VR64RegisterClass);
10414 case 'Y': // SSE_REGS if SSE2 allowed
10415 if (!Subtarget->hasSSE2()) break;
10417 case 'x': // SSE_REGS if SSE1 allowed
10418 if (!Subtarget->hasSSE1()) break;
10420 switch (VT.getSimpleVT().SimpleTy) {
10422 // Scalar SSE types.
10425 return std::make_pair(0U, X86::FR32RegisterClass);
10428 return std::make_pair(0U, X86::FR64RegisterClass);
10436 return std::make_pair(0U, X86::VR128RegisterClass);
10442 // Use the default implementation in TargetLowering to convert the register
10443 // constraint into a member of a register class.
10444 std::pair<unsigned, const TargetRegisterClass*> Res;
10445 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10447 // Not found as a standard register?
10448 if (Res.second == 0) {
10449 // Map st(0) -> st(7) -> ST0
10450 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10451 tolower(Constraint[1]) == 's' &&
10452 tolower(Constraint[2]) == 't' &&
10453 Constraint[3] == '(' &&
10454 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10455 Constraint[5] == ')' &&
10456 Constraint[6] == '}') {
10458 Res.first = X86::ST0+Constraint[4]-'0';
10459 Res.second = X86::RFP80RegisterClass;
10463 // GCC allows "st(0)" to be called just plain "st".
10464 if (StringRef("{st}").equals_lower(Constraint)) {
10465 Res.first = X86::ST0;
10466 Res.second = X86::RFP80RegisterClass;
10471 if (StringRef("{flags}").equals_lower(Constraint)) {
10472 Res.first = X86::EFLAGS;
10473 Res.second = X86::CCRRegisterClass;
10477 // 'A' means EAX + EDX.
10478 if (Constraint == "A") {
10479 Res.first = X86::EAX;
10480 Res.second = X86::GR32_ADRegisterClass;
10486 // Otherwise, check to see if this is a register class of the wrong value
10487 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10488 // turn into {ax},{dx}.
10489 if (Res.second->hasType(VT))
10490 return Res; // Correct type already, nothing to do.
10492 // All of the single-register GCC register classes map their values onto
10493 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10494 // really want an 8-bit or 32-bit register, map to the appropriate register
10495 // class and return the appropriate register.
10496 if (Res.second == X86::GR16RegisterClass) {
10497 if (VT == MVT::i8) {
10498 unsigned DestReg = 0;
10499 switch (Res.first) {
10501 case X86::AX: DestReg = X86::AL; break;
10502 case X86::DX: DestReg = X86::DL; break;
10503 case X86::CX: DestReg = X86::CL; break;
10504 case X86::BX: DestReg = X86::BL; break;
10507 Res.first = DestReg;
10508 Res.second = X86::GR8RegisterClass;
10510 } else if (VT == MVT::i32) {
10511 unsigned DestReg = 0;
10512 switch (Res.first) {
10514 case X86::AX: DestReg = X86::EAX; break;
10515 case X86::DX: DestReg = X86::EDX; break;
10516 case X86::CX: DestReg = X86::ECX; break;
10517 case X86::BX: DestReg = X86::EBX; break;
10518 case X86::SI: DestReg = X86::ESI; break;
10519 case X86::DI: DestReg = X86::EDI; break;
10520 case X86::BP: DestReg = X86::EBP; break;
10521 case X86::SP: DestReg = X86::ESP; break;
10524 Res.first = DestReg;
10525 Res.second = X86::GR32RegisterClass;
10527 } else if (VT == MVT::i64) {
10528 unsigned DestReg = 0;
10529 switch (Res.first) {
10531 case X86::AX: DestReg = X86::RAX; break;
10532 case X86::DX: DestReg = X86::RDX; break;
10533 case X86::CX: DestReg = X86::RCX; break;
10534 case X86::BX: DestReg = X86::RBX; break;
10535 case X86::SI: DestReg = X86::RSI; break;
10536 case X86::DI: DestReg = X86::RDI; break;
10537 case X86::BP: DestReg = X86::RBP; break;
10538 case X86::SP: DestReg = X86::RSP; break;
10541 Res.first = DestReg;
10542 Res.second = X86::GR64RegisterClass;
10545 } else if (Res.second == X86::FR32RegisterClass ||
10546 Res.second == X86::FR64RegisterClass ||
10547 Res.second == X86::VR128RegisterClass) {
10548 // Handle references to XMM physical registers that got mapped into the
10549 // wrong class. This can happen with constraints like {xmm0} where the
10550 // target independent register mapper will just pick the first match it can
10551 // find, ignoring the required type.
10552 if (VT == MVT::f32)
10553 Res.second = X86::FR32RegisterClass;
10554 else if (VT == MVT::f64)
10555 Res.second = X86::FR64RegisterClass;
10556 else if (X86::VR128RegisterClass->hasType(VT))
10557 Res.second = X86::VR128RegisterClass;