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);
627 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
628 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
629 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
630 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
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 // FIXME: This produces lots of inefficiencies in isel since
798 // we then need notice that most of our operands have been implicitly
799 // converted to v2i64.
800 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
801 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
804 // Do not attempt to promote non-128-bit vectors
805 if (!VT.is128BitVector()) {
809 setOperationAction(ISD::AND, SVT, Promote);
810 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
811 setOperationAction(ISD::OR, SVT, Promote);
812 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
813 setOperationAction(ISD::XOR, SVT, Promote);
814 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
815 setOperationAction(ISD::LOAD, SVT, Promote);
816 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
817 setOperationAction(ISD::SELECT, SVT, Promote);
818 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
821 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
823 // Custom lower v2i64 and v2f64 selects.
824 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
825 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
826 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
827 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
829 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
830 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
831 if (!DisableMMX && Subtarget->hasMMX()) {
832 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
833 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
837 if (Subtarget->hasSSE41()) {
838 // FIXME: Do we need to handle scalar-to-vector here?
839 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
841 // i8 and i16 vectors are custom , because the source register and source
842 // source memory operand types are not the same width. f32 vectors are
843 // custom since the immediate controlling the insert encodes additional
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
848 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
851 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
852 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
853 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
855 if (Subtarget->is64Bit()) {
856 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
857 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
861 if (Subtarget->hasSSE42()) {
862 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
865 if (!UseSoftFloat && Subtarget->hasAVX()) {
866 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
867 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
868 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
869 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
871 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
872 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
873 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
874 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
875 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
876 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
877 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
878 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
879 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
880 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
881 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
882 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
883 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
884 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
885 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
887 // Operations to consider commented out -v16i16 v32i8
888 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
889 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
890 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
891 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
892 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
893 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
894 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
895 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
896 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
897 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
898 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
899 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
900 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
901 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
903 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
904 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
905 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
906 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
908 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
909 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
910 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
912 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
914 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
915 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
916 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
917 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
918 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
919 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
922 // Not sure we want to do this since there are no 256-bit integer
925 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
926 // This includes 256-bit vectors
927 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
928 EVT VT = (MVT::SimpleValueType)i;
930 // Do not attempt to custom lower non-power-of-2 vectors
931 if (!isPowerOf2_32(VT.getVectorNumElements()))
934 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
935 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
936 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
939 if (Subtarget->is64Bit()) {
940 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
941 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
946 // Not sure we want to do this since there are no 256-bit integer
949 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
950 // Including 256-bit vectors
951 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
952 EVT VT = (MVT::SimpleValueType)i;
954 if (!VT.is256BitVector()) {
957 setOperationAction(ISD::AND, VT, Promote);
958 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
959 setOperationAction(ISD::OR, VT, Promote);
960 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
961 setOperationAction(ISD::XOR, VT, Promote);
962 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
963 setOperationAction(ISD::LOAD, VT, Promote);
964 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
965 setOperationAction(ISD::SELECT, VT, Promote);
966 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
969 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
973 // We want to custom lower some of our intrinsics.
974 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
976 // Add/Sub/Mul with overflow operations are custom lowered.
977 setOperationAction(ISD::SADDO, MVT::i32, Custom);
978 setOperationAction(ISD::SADDO, MVT::i64, Custom);
979 setOperationAction(ISD::UADDO, MVT::i32, Custom);
980 setOperationAction(ISD::UADDO, MVT::i64, Custom);
981 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
982 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
983 setOperationAction(ISD::USUBO, MVT::i32, Custom);
984 setOperationAction(ISD::USUBO, MVT::i64, Custom);
985 setOperationAction(ISD::SMULO, MVT::i32, Custom);
986 setOperationAction(ISD::SMULO, MVT::i64, Custom);
988 if (!Subtarget->is64Bit()) {
989 // These libcalls are not available in 32-bit.
990 setLibcallName(RTLIB::SHL_I128, 0);
991 setLibcallName(RTLIB::SRL_I128, 0);
992 setLibcallName(RTLIB::SRA_I128, 0);
995 // We have target-specific dag combine patterns for the following nodes:
996 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
997 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
998 setTargetDAGCombine(ISD::BUILD_VECTOR);
999 setTargetDAGCombine(ISD::SELECT);
1000 setTargetDAGCombine(ISD::SHL);
1001 setTargetDAGCombine(ISD::SRA);
1002 setTargetDAGCombine(ISD::SRL);
1003 setTargetDAGCombine(ISD::OR);
1004 setTargetDAGCombine(ISD::STORE);
1005 setTargetDAGCombine(ISD::MEMBARRIER);
1006 setTargetDAGCombine(ISD::ZERO_EXTEND);
1007 if (Subtarget->is64Bit())
1008 setTargetDAGCombine(ISD::MUL);
1010 computeRegisterProperties();
1012 // FIXME: These should be based on subtarget info. Plus, the values should
1013 // be smaller when we are in optimizing for size mode.
1014 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1015 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1016 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1017 setPrefLoopAlignment(16);
1018 benefitFromCodePlacementOpt = true;
1022 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1027 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1028 /// the desired ByVal argument alignment.
1029 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1032 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1033 if (VTy->getBitWidth() == 128)
1035 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1036 unsigned EltAlign = 0;
1037 getMaxByValAlign(ATy->getElementType(), EltAlign);
1038 if (EltAlign > MaxAlign)
1039 MaxAlign = EltAlign;
1040 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1041 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1042 unsigned EltAlign = 0;
1043 getMaxByValAlign(STy->getElementType(i), EltAlign);
1044 if (EltAlign > MaxAlign)
1045 MaxAlign = EltAlign;
1053 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1054 /// function arguments in the caller parameter area. For X86, aggregates
1055 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1056 /// are at 4-byte boundaries.
1057 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1058 if (Subtarget->is64Bit()) {
1059 // Max of 8 and alignment of type.
1060 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1067 if (Subtarget->hasSSE1())
1068 getMaxByValAlign(Ty, Align);
1072 /// getOptimalMemOpType - Returns the target specific optimal type for load
1073 /// and store operations as a result of memset, memcpy, and memmove
1074 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1077 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1078 bool isSrcConst, bool isSrcStr,
1079 SelectionDAG &DAG) const {
1080 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1081 // linux. This is because the stack realignment code can't handle certain
1082 // cases like PR2962. This should be removed when PR2962 is fixed.
1083 const Function *F = DAG.getMachineFunction().getFunction();
1084 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1085 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1086 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1088 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1091 if (Subtarget->is64Bit() && Size >= 8)
1096 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1097 /// current function. The returned value is a member of the
1098 /// MachineJumpTableInfo::JTEntryKind enum.
1099 unsigned X86TargetLowering::getJumpTableEncoding() const {
1100 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1102 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1103 Subtarget->isPICStyleGOT())
1104 return MachineJumpTableInfo::EK_Custom32;
1106 // Otherwise, use the normal jump table encoding heuristics.
1107 return TargetLowering::getJumpTableEncoding();
1110 /// getPICBaseSymbol - Return the X86-32 PIC base.
1112 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1113 MCContext &Ctx) const {
1114 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1115 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1116 Twine(MF->getFunctionNumber())+"$pb");
1121 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1122 const MachineBasicBlock *MBB,
1123 unsigned uid,MCContext &Ctx) const{
1124 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1125 Subtarget->isPICStyleGOT());
1126 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1128 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1129 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1132 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1134 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1135 SelectionDAG &DAG) const {
1136 if (!Subtarget->is64Bit())
1137 // This doesn't have DebugLoc associated with it, but is not really the
1138 // same as a Register.
1139 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1144 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1145 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1147 const MCExpr *X86TargetLowering::
1148 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1149 MCContext &Ctx) const {
1150 // X86-64 uses RIP relative addressing based on the jump table label.
1151 if (Subtarget->isPICStyleRIPRel())
1152 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1154 // Otherwise, the reference is relative to the PIC base.
1155 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1158 /// getFunctionAlignment - Return the Log2 alignment of this function.
1159 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1160 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1163 //===----------------------------------------------------------------------===//
1164 // Return Value Calling Convention Implementation
1165 //===----------------------------------------------------------------------===//
1167 #include "X86GenCallingConv.inc"
1170 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1171 const SmallVectorImpl<EVT> &OutTys,
1172 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1173 SelectionDAG &DAG) {
1174 SmallVector<CCValAssign, 16> RVLocs;
1175 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1176 RVLocs, *DAG.getContext());
1177 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1181 X86TargetLowering::LowerReturn(SDValue Chain,
1182 CallingConv::ID CallConv, bool isVarArg,
1183 const SmallVectorImpl<ISD::OutputArg> &Outs,
1184 DebugLoc dl, SelectionDAG &DAG) {
1186 SmallVector<CCValAssign, 16> RVLocs;
1187 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1188 RVLocs, *DAG.getContext());
1189 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1191 // Add the regs to the liveout set for the function.
1192 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1193 for (unsigned i = 0; i != RVLocs.size(); ++i)
1194 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1195 MRI.addLiveOut(RVLocs[i].getLocReg());
1199 SmallVector<SDValue, 6> RetOps;
1200 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1201 // Operand #1 = Bytes To Pop
1202 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1204 // Copy the result values into the output registers.
1205 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1206 CCValAssign &VA = RVLocs[i];
1207 assert(VA.isRegLoc() && "Can only return in registers!");
1208 SDValue ValToCopy = Outs[i].Val;
1210 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1211 // the RET instruction and handled by the FP Stackifier.
1212 if (VA.getLocReg() == X86::ST0 ||
1213 VA.getLocReg() == X86::ST1) {
1214 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1215 // change the value to the FP stack register class.
1216 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1217 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1218 RetOps.push_back(ValToCopy);
1219 // Don't emit a copytoreg.
1223 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1224 // which is returned in RAX / RDX.
1225 if (Subtarget->is64Bit()) {
1226 EVT ValVT = ValToCopy.getValueType();
1227 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1228 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1229 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1230 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1234 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1235 Flag = Chain.getValue(1);
1238 // The x86-64 ABI for returning structs by value requires that we copy
1239 // the sret argument into %rax for the return. We saved the argument into
1240 // a virtual register in the entry block, so now we copy the value out
1242 if (Subtarget->is64Bit() &&
1243 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1244 MachineFunction &MF = DAG.getMachineFunction();
1245 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1246 unsigned Reg = FuncInfo->getSRetReturnReg();
1248 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1249 FuncInfo->setSRetReturnReg(Reg);
1251 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1253 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1254 Flag = Chain.getValue(1);
1256 // RAX now acts like a return value.
1257 MRI.addLiveOut(X86::RAX);
1260 RetOps[0] = Chain; // Update chain.
1262 // Add the flag if we have it.
1264 RetOps.push_back(Flag);
1266 return DAG.getNode(X86ISD::RET_FLAG, dl,
1267 MVT::Other, &RetOps[0], RetOps.size());
1270 /// LowerCallResult - Lower the result values of a call into the
1271 /// appropriate copies out of appropriate physical registers.
1274 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1275 CallingConv::ID CallConv, bool isVarArg,
1276 const SmallVectorImpl<ISD::InputArg> &Ins,
1277 DebugLoc dl, SelectionDAG &DAG,
1278 SmallVectorImpl<SDValue> &InVals) {
1280 // Assign locations to each value returned by this call.
1281 SmallVector<CCValAssign, 16> RVLocs;
1282 bool Is64Bit = Subtarget->is64Bit();
1283 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1284 RVLocs, *DAG.getContext());
1285 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1287 // Copy all of the result registers out of their specified physreg.
1288 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1289 CCValAssign &VA = RVLocs[i];
1290 EVT CopyVT = VA.getValVT();
1292 // If this is x86-64, and we disabled SSE, we can't return FP values
1293 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1294 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1295 llvm_report_error("SSE register return with SSE disabled");
1298 // If this is a call to a function that returns an fp value on the floating
1299 // point stack, but where we prefer to use the value in xmm registers, copy
1300 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1301 if ((VA.getLocReg() == X86::ST0 ||
1302 VA.getLocReg() == X86::ST1) &&
1303 isScalarFPTypeInSSEReg(VA.getValVT())) {
1308 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1309 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1310 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1311 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1312 MVT::v2i64, InFlag).getValue(1);
1313 Val = Chain.getValue(0);
1314 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1315 Val, DAG.getConstant(0, MVT::i64));
1317 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1318 MVT::i64, InFlag).getValue(1);
1319 Val = Chain.getValue(0);
1321 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1323 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1324 CopyVT, InFlag).getValue(1);
1325 Val = Chain.getValue(0);
1327 InFlag = Chain.getValue(2);
1329 if (CopyVT != VA.getValVT()) {
1330 // Round the F80 the right size, which also moves to the appropriate xmm
1332 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1333 // This truncation won't change the value.
1334 DAG.getIntPtrConstant(1));
1337 InVals.push_back(Val);
1344 //===----------------------------------------------------------------------===//
1345 // C & StdCall & Fast Calling Convention implementation
1346 //===----------------------------------------------------------------------===//
1347 // StdCall calling convention seems to be standard for many Windows' API
1348 // routines and around. It differs from C calling convention just a little:
1349 // callee should clean up the stack, not caller. Symbols should be also
1350 // decorated in some fancy way :) It doesn't support any vector arguments.
1351 // For info on fast calling convention see Fast Calling Convention (tail call)
1352 // implementation LowerX86_32FastCCCallTo.
1354 /// CallIsStructReturn - Determines whether a call uses struct return
1356 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1360 return Outs[0].Flags.isSRet();
1363 /// ArgsAreStructReturn - Determines whether a function uses struct
1364 /// return semantics.
1366 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1370 return Ins[0].Flags.isSRet();
1373 /// IsCalleePop - Determines whether the callee is required to pop its
1374 /// own arguments. Callee pop is necessary to support tail calls.
1375 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1379 switch (CallingConv) {
1382 case CallingConv::X86_StdCall:
1383 return !Subtarget->is64Bit();
1384 case CallingConv::X86_FastCall:
1385 return !Subtarget->is64Bit();
1386 case CallingConv::Fast:
1387 return GuaranteedTailCallOpt;
1388 case CallingConv::GHC:
1389 return GuaranteedTailCallOpt;
1393 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1394 /// given CallingConvention value.
1395 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1396 if (Subtarget->is64Bit()) {
1397 if (CC == CallingConv::GHC)
1398 return CC_X86_64_GHC;
1399 else if (Subtarget->isTargetWin64())
1400 return CC_X86_Win64_C;
1405 if (CC == CallingConv::X86_FastCall)
1406 return CC_X86_32_FastCall;
1407 else if (CC == CallingConv::Fast)
1408 return CC_X86_32_FastCC;
1409 else if (CC == CallingConv::GHC)
1410 return CC_X86_32_GHC;
1415 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1416 /// by "Src" to address "Dst" with size and alignment information specified by
1417 /// the specific parameter attribute. The copy will be passed as a byval
1418 /// function parameter.
1420 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1421 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1423 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1424 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1425 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1428 /// IsTailCallConvention - Return true if the calling convention is one that
1429 /// supports tail call optimization.
1430 static bool IsTailCallConvention(CallingConv::ID CC) {
1431 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1434 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1435 /// a tailcall target by changing its ABI.
1436 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1437 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1441 X86TargetLowering::LowerMemArgument(SDValue Chain,
1442 CallingConv::ID CallConv,
1443 const SmallVectorImpl<ISD::InputArg> &Ins,
1444 DebugLoc dl, SelectionDAG &DAG,
1445 const CCValAssign &VA,
1446 MachineFrameInfo *MFI,
1448 // Create the nodes corresponding to a load from this parameter slot.
1449 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1450 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1451 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1454 // If value is passed by pointer we have address passed instead of the value
1456 if (VA.getLocInfo() == CCValAssign::Indirect)
1457 ValVT = VA.getLocVT();
1459 ValVT = VA.getValVT();
1461 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1462 // changed with more analysis.
1463 // In case of tail call optimization mark all arguments mutable. Since they
1464 // could be overwritten by lowering of arguments in case of a tail call.
1465 if (Flags.isByVal()) {
1466 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1467 VA.getLocMemOffset(), isImmutable, false);
1468 return DAG.getFrameIndex(FI, getPointerTy());
1470 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1471 VA.getLocMemOffset(), isImmutable, false);
1472 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1473 return DAG.getLoad(ValVT, dl, Chain, FIN,
1474 PseudoSourceValue::getFixedStack(FI), 0,
1480 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1481 CallingConv::ID CallConv,
1483 const SmallVectorImpl<ISD::InputArg> &Ins,
1486 SmallVectorImpl<SDValue> &InVals) {
1487 MachineFunction &MF = DAG.getMachineFunction();
1488 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1490 const Function* Fn = MF.getFunction();
1491 if (Fn->hasExternalLinkage() &&
1492 Subtarget->isTargetCygMing() &&
1493 Fn->getName() == "main")
1494 FuncInfo->setForceFramePointer(true);
1496 MachineFrameInfo *MFI = MF.getFrameInfo();
1497 bool Is64Bit = Subtarget->is64Bit();
1498 bool IsWin64 = Subtarget->isTargetWin64();
1500 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1501 "Var args not supported with calling convention fastcc or ghc");
1503 // Assign locations to all of the incoming arguments.
1504 SmallVector<CCValAssign, 16> ArgLocs;
1505 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1506 ArgLocs, *DAG.getContext());
1507 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1509 unsigned LastVal = ~0U;
1511 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1512 CCValAssign &VA = ArgLocs[i];
1513 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1515 assert(VA.getValNo() != LastVal &&
1516 "Don't support value assigned to multiple locs yet");
1517 LastVal = VA.getValNo();
1519 if (VA.isRegLoc()) {
1520 EVT RegVT = VA.getLocVT();
1521 TargetRegisterClass *RC = NULL;
1522 if (RegVT == MVT::i32)
1523 RC = X86::GR32RegisterClass;
1524 else if (Is64Bit && RegVT == MVT::i64)
1525 RC = X86::GR64RegisterClass;
1526 else if (RegVT == MVT::f32)
1527 RC = X86::FR32RegisterClass;
1528 else if (RegVT == MVT::f64)
1529 RC = X86::FR64RegisterClass;
1530 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1531 RC = X86::VR128RegisterClass;
1532 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1533 RC = X86::VR64RegisterClass;
1535 llvm_unreachable("Unknown argument type!");
1537 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1538 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1540 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1541 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1543 if (VA.getLocInfo() == CCValAssign::SExt)
1544 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1545 DAG.getValueType(VA.getValVT()));
1546 else if (VA.getLocInfo() == CCValAssign::ZExt)
1547 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1548 DAG.getValueType(VA.getValVT()));
1549 else if (VA.getLocInfo() == CCValAssign::BCvt)
1550 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1552 if (VA.isExtInLoc()) {
1553 // Handle MMX values passed in XMM regs.
1554 if (RegVT.isVector()) {
1555 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1556 ArgValue, DAG.getConstant(0, MVT::i64));
1557 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1559 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1562 assert(VA.isMemLoc());
1563 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1566 // If value is passed via pointer - do a load.
1567 if (VA.getLocInfo() == CCValAssign::Indirect)
1568 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1571 InVals.push_back(ArgValue);
1574 // The x86-64 ABI for returning structs by value requires that we copy
1575 // the sret argument into %rax for the return. Save the argument into
1576 // a virtual register so that we can access it from the return points.
1577 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1578 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1579 unsigned Reg = FuncInfo->getSRetReturnReg();
1581 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1582 FuncInfo->setSRetReturnReg(Reg);
1584 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1585 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1588 unsigned StackSize = CCInfo.getNextStackOffset();
1589 // Align stack specially for tail calls.
1590 if (FuncIsMadeTailCallSafe(CallConv))
1591 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1593 // If the function takes variable number of arguments, make a frame index for
1594 // the start of the first vararg value... for expansion of llvm.va_start.
1596 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1597 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1600 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1602 // FIXME: We should really autogenerate these arrays
1603 static const unsigned GPR64ArgRegsWin64[] = {
1604 X86::RCX, X86::RDX, X86::R8, X86::R9
1606 static const unsigned XMMArgRegsWin64[] = {
1607 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1609 static const unsigned GPR64ArgRegs64Bit[] = {
1610 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1612 static const unsigned XMMArgRegs64Bit[] = {
1613 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1614 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1616 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1619 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1620 GPR64ArgRegs = GPR64ArgRegsWin64;
1621 XMMArgRegs = XMMArgRegsWin64;
1623 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1624 GPR64ArgRegs = GPR64ArgRegs64Bit;
1625 XMMArgRegs = XMMArgRegs64Bit;
1627 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1629 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1632 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1633 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1634 "SSE register cannot be used when SSE is disabled!");
1635 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1636 "SSE register cannot be used when SSE is disabled!");
1637 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1638 // Kernel mode asks for SSE to be disabled, so don't push them
1640 TotalNumXMMRegs = 0;
1642 // For X86-64, if there are vararg parameters that are passed via
1643 // registers, then we must store them to their spots on the stack so they
1644 // may be loaded by deferencing the result of va_next.
1645 VarArgsGPOffset = NumIntRegs * 8;
1646 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1647 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1648 TotalNumXMMRegs * 16, 16,
1651 // Store the integer parameter registers.
1652 SmallVector<SDValue, 8> MemOps;
1653 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1654 unsigned Offset = VarArgsGPOffset;
1655 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1656 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1657 DAG.getIntPtrConstant(Offset));
1658 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1659 X86::GR64RegisterClass);
1660 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1662 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1663 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1664 Offset, false, false, 0);
1665 MemOps.push_back(Store);
1669 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1670 // Now store the XMM (fp + vector) parameter registers.
1671 SmallVector<SDValue, 11> SaveXMMOps;
1672 SaveXMMOps.push_back(Chain);
1674 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1675 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1676 SaveXMMOps.push_back(ALVal);
1678 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1679 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1681 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1682 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1683 X86::VR128RegisterClass);
1684 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1685 SaveXMMOps.push_back(Val);
1687 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1689 &SaveXMMOps[0], SaveXMMOps.size()));
1692 if (!MemOps.empty())
1693 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1694 &MemOps[0], MemOps.size());
1698 // Some CCs need callee pop.
1699 if (IsCalleePop(isVarArg, CallConv)) {
1700 BytesToPopOnReturn = StackSize; // Callee pops everything.
1702 BytesToPopOnReturn = 0; // Callee pops nothing.
1703 // If this is an sret function, the return should pop the hidden pointer.
1704 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1705 BytesToPopOnReturn = 4;
1709 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1710 if (CallConv == CallingConv::X86_FastCall)
1711 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1714 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1720 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1721 SDValue StackPtr, SDValue Arg,
1722 DebugLoc dl, SelectionDAG &DAG,
1723 const CCValAssign &VA,
1724 ISD::ArgFlagsTy Flags) {
1725 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1726 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1727 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1728 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1729 if (Flags.isByVal()) {
1730 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1732 return DAG.getStore(Chain, dl, Arg, PtrOff,
1733 PseudoSourceValue::getStack(), LocMemOffset,
1737 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1738 /// optimization is performed and it is required.
1740 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1741 SDValue &OutRetAddr, SDValue Chain,
1742 bool IsTailCall, bool Is64Bit,
1743 int FPDiff, DebugLoc dl) {
1744 // Adjust the Return address stack slot.
1745 EVT VT = getPointerTy();
1746 OutRetAddr = getReturnAddressFrameIndex(DAG);
1748 // Load the "old" Return address.
1749 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1750 return SDValue(OutRetAddr.getNode(), 1);
1753 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1754 /// optimization is performed and it is required (FPDiff!=0).
1756 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1757 SDValue Chain, SDValue RetAddrFrIdx,
1758 bool Is64Bit, int FPDiff, DebugLoc dl) {
1759 // Store the return address to the appropriate stack slot.
1760 if (!FPDiff) return Chain;
1761 // Calculate the new stack slot for the return address.
1762 int SlotSize = Is64Bit ? 8 : 4;
1763 int NewReturnAddrFI =
1764 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1765 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1766 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1767 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1768 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1774 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1775 CallingConv::ID CallConv, bool isVarArg,
1777 const SmallVectorImpl<ISD::OutputArg> &Outs,
1778 const SmallVectorImpl<ISD::InputArg> &Ins,
1779 DebugLoc dl, SelectionDAG &DAG,
1780 SmallVectorImpl<SDValue> &InVals) {
1781 MachineFunction &MF = DAG.getMachineFunction();
1782 bool Is64Bit = Subtarget->is64Bit();
1783 bool IsStructRet = CallIsStructReturn(Outs);
1784 bool IsSibcall = false;
1787 // Check if it's really possible to do a tail call.
1788 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1789 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1792 // Sibcalls are automatically detected tailcalls which do not require
1794 if (!GuaranteedTailCallOpt && isTailCall)
1801 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1802 "Var args not supported with calling convention fastcc or ghc");
1804 // Analyze operands of the call, assigning locations to each operand.
1805 SmallVector<CCValAssign, 16> ArgLocs;
1806 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1807 ArgLocs, *DAG.getContext());
1808 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1810 // Get a count of how many bytes are to be pushed on the stack.
1811 unsigned NumBytes = CCInfo.getNextStackOffset();
1813 // This is a sibcall. The memory operands are available in caller's
1814 // own caller's stack.
1816 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1817 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1820 if (isTailCall && !IsSibcall) {
1821 // Lower arguments at fp - stackoffset + fpdiff.
1822 unsigned NumBytesCallerPushed =
1823 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1824 FPDiff = NumBytesCallerPushed - NumBytes;
1826 // Set the delta of movement of the returnaddr stackslot.
1827 // But only set if delta is greater than previous delta.
1828 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1829 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1833 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1835 SDValue RetAddrFrIdx;
1836 // Load return adress for tail calls.
1837 if (isTailCall && FPDiff)
1838 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1839 Is64Bit, FPDiff, dl);
1841 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1842 SmallVector<SDValue, 8> MemOpChains;
1845 // Walk the register/memloc assignments, inserting copies/loads. In the case
1846 // of tail call optimization arguments are handle later.
1847 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1848 CCValAssign &VA = ArgLocs[i];
1849 EVT RegVT = VA.getLocVT();
1850 SDValue Arg = Outs[i].Val;
1851 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1852 bool isByVal = Flags.isByVal();
1854 // Promote the value if needed.
1855 switch (VA.getLocInfo()) {
1856 default: llvm_unreachable("Unknown loc info!");
1857 case CCValAssign::Full: break;
1858 case CCValAssign::SExt:
1859 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1861 case CCValAssign::ZExt:
1862 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1864 case CCValAssign::AExt:
1865 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1866 // Special case: passing MMX values in XMM registers.
1867 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1868 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1869 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1871 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1873 case CCValAssign::BCvt:
1874 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1876 case CCValAssign::Indirect: {
1877 // Store the argument.
1878 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1879 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1880 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1881 PseudoSourceValue::getFixedStack(FI), 0,
1888 if (VA.isRegLoc()) {
1889 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1890 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1891 assert(VA.isMemLoc());
1892 if (StackPtr.getNode() == 0)
1893 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1894 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1895 dl, DAG, VA, Flags));
1899 if (!MemOpChains.empty())
1900 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1901 &MemOpChains[0], MemOpChains.size());
1903 // Build a sequence of copy-to-reg nodes chained together with token chain
1904 // and flag operands which copy the outgoing args into registers.
1906 // Tail call byval lowering might overwrite argument registers so in case of
1907 // tail call optimization the copies to registers are lowered later.
1909 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1910 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1911 RegsToPass[i].second, InFlag);
1912 InFlag = Chain.getValue(1);
1915 if (Subtarget->isPICStyleGOT()) {
1916 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1919 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1920 DAG.getNode(X86ISD::GlobalBaseReg,
1921 DebugLoc::getUnknownLoc(),
1924 InFlag = Chain.getValue(1);
1926 // If we are tail calling and generating PIC/GOT style code load the
1927 // address of the callee into ECX. The value in ecx is used as target of
1928 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1929 // for tail calls on PIC/GOT architectures. Normally we would just put the
1930 // address of GOT into ebx and then call target@PLT. But for tail calls
1931 // ebx would be restored (since ebx is callee saved) before jumping to the
1934 // Note: The actual moving to ECX is done further down.
1935 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1936 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1937 !G->getGlobal()->hasProtectedVisibility())
1938 Callee = LowerGlobalAddress(Callee, DAG);
1939 else if (isa<ExternalSymbolSDNode>(Callee))
1940 Callee = LowerExternalSymbol(Callee, DAG);
1944 if (Is64Bit && isVarArg) {
1945 // From AMD64 ABI document:
1946 // For calls that may call functions that use varargs or stdargs
1947 // (prototype-less calls or calls to functions containing ellipsis (...) in
1948 // the declaration) %al is used as hidden argument to specify the number
1949 // of SSE registers used. The contents of %al do not need to match exactly
1950 // the number of registers, but must be an ubound on the number of SSE
1951 // registers used and is in the range 0 - 8 inclusive.
1953 // FIXME: Verify this on Win64
1954 // Count the number of XMM registers allocated.
1955 static const unsigned XMMArgRegs[] = {
1956 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1957 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1959 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1960 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1961 && "SSE registers cannot be used when SSE is disabled");
1963 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1964 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1965 InFlag = Chain.getValue(1);
1969 // For tail calls lower the arguments to the 'real' stack slot.
1971 // Force all the incoming stack arguments to be loaded from the stack
1972 // before any new outgoing arguments are stored to the stack, because the
1973 // outgoing stack slots may alias the incoming argument stack slots, and
1974 // the alias isn't otherwise explicit. This is slightly more conservative
1975 // than necessary, because it means that each store effectively depends
1976 // on every argument instead of just those arguments it would clobber.
1977 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1979 SmallVector<SDValue, 8> MemOpChains2;
1982 // Do not flag preceeding copytoreg stuff together with the following stuff.
1984 if (GuaranteedTailCallOpt) {
1985 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1986 CCValAssign &VA = ArgLocs[i];
1989 assert(VA.isMemLoc());
1990 SDValue Arg = Outs[i].Val;
1991 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1992 // Create frame index.
1993 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1994 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1995 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1996 FIN = DAG.getFrameIndex(FI, getPointerTy());
1998 if (Flags.isByVal()) {
1999 // Copy relative to framepointer.
2000 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2001 if (StackPtr.getNode() == 0)
2002 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2004 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2006 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2010 // Store relative to framepointer.
2011 MemOpChains2.push_back(
2012 DAG.getStore(ArgChain, dl, Arg, FIN,
2013 PseudoSourceValue::getFixedStack(FI), 0,
2019 if (!MemOpChains2.empty())
2020 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2021 &MemOpChains2[0], MemOpChains2.size());
2023 // Copy arguments to their registers.
2024 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2025 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2026 RegsToPass[i].second, InFlag);
2027 InFlag = Chain.getValue(1);
2031 // Store the return address to the appropriate stack slot.
2032 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2036 bool WasGlobalOrExternal = false;
2037 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2038 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2039 // In the 64-bit large code model, we have to make all calls
2040 // through a register, since the call instruction's 32-bit
2041 // pc-relative offset may not be large enough to hold the whole
2043 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2044 WasGlobalOrExternal = true;
2045 // If the callee is a GlobalAddress node (quite common, every direct call
2046 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2049 // We should use extra load for direct calls to dllimported functions in
2051 GlobalValue *GV = G->getGlobal();
2052 if (!GV->hasDLLImportLinkage()) {
2053 unsigned char OpFlags = 0;
2055 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2056 // external symbols most go through the PLT in PIC mode. If the symbol
2057 // has hidden or protected visibility, or if it is static or local, then
2058 // we don't need to use the PLT - we can directly call it.
2059 if (Subtarget->isTargetELF() &&
2060 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2061 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2062 OpFlags = X86II::MO_PLT;
2063 } else if (Subtarget->isPICStyleStubAny() &&
2064 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2065 Subtarget->getDarwinVers() < 9) {
2066 // PC-relative references to external symbols should go through $stub,
2067 // unless we're building with the leopard linker or later, which
2068 // automatically synthesizes these stubs.
2069 OpFlags = X86II::MO_DARWIN_STUB;
2072 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2073 G->getOffset(), OpFlags);
2075 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2076 WasGlobalOrExternal = true;
2077 unsigned char OpFlags = 0;
2079 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2080 // symbols should go through the PLT.
2081 if (Subtarget->isTargetELF() &&
2082 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2083 OpFlags = X86II::MO_PLT;
2084 } else if (Subtarget->isPICStyleStubAny() &&
2085 Subtarget->getDarwinVers() < 9) {
2086 // PC-relative references to external symbols should go through $stub,
2087 // unless we're building with the leopard linker or later, which
2088 // automatically synthesizes these stubs.
2089 OpFlags = X86II::MO_DARWIN_STUB;
2092 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2096 // Returns a chain & a flag for retval copy to use.
2097 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2098 SmallVector<SDValue, 8> Ops;
2100 if (!IsSibcall && isTailCall) {
2101 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2102 DAG.getIntPtrConstant(0, true), InFlag);
2103 InFlag = Chain.getValue(1);
2106 Ops.push_back(Chain);
2107 Ops.push_back(Callee);
2110 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2112 // Add argument registers to the end of the list so that they are known live
2114 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2115 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2116 RegsToPass[i].second.getValueType()));
2118 // Add an implicit use GOT pointer in EBX.
2119 if (!isTailCall && Subtarget->isPICStyleGOT())
2120 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2122 // Add an implicit use of AL for x86 vararg functions.
2123 if (Is64Bit && isVarArg)
2124 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2126 if (InFlag.getNode())
2127 Ops.push_back(InFlag);
2130 // If this is the first return lowered for this function, add the regs
2131 // to the liveout set for the function.
2132 if (MF.getRegInfo().liveout_empty()) {
2133 SmallVector<CCValAssign, 16> RVLocs;
2134 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2136 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2137 for (unsigned i = 0; i != RVLocs.size(); ++i)
2138 if (RVLocs[i].isRegLoc())
2139 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2141 return DAG.getNode(X86ISD::TC_RETURN, dl,
2142 NodeTys, &Ops[0], Ops.size());
2145 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2146 InFlag = Chain.getValue(1);
2148 // Create the CALLSEQ_END node.
2149 unsigned NumBytesForCalleeToPush;
2150 if (IsCalleePop(isVarArg, CallConv))
2151 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2152 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2153 // If this is a call to a struct-return function, the callee
2154 // pops the hidden struct pointer, so we have to push it back.
2155 // This is common for Darwin/X86, Linux & Mingw32 targets.
2156 NumBytesForCalleeToPush = 4;
2158 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2160 // Returns a flag for retval copy to use.
2162 Chain = DAG.getCALLSEQ_END(Chain,
2163 DAG.getIntPtrConstant(NumBytes, true),
2164 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2167 InFlag = Chain.getValue(1);
2170 // Handle result values, copying them out of physregs into vregs that we
2172 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2173 Ins, dl, DAG, InVals);
2177 //===----------------------------------------------------------------------===//
2178 // Fast Calling Convention (tail call) implementation
2179 //===----------------------------------------------------------------------===//
2181 // Like std call, callee cleans arguments, convention except that ECX is
2182 // reserved for storing the tail called function address. Only 2 registers are
2183 // free for argument passing (inreg). Tail call optimization is performed
2185 // * tailcallopt is enabled
2186 // * caller/callee are fastcc
2187 // On X86_64 architecture with GOT-style position independent code only local
2188 // (within module) calls are supported at the moment.
2189 // To keep the stack aligned according to platform abi the function
2190 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2191 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2192 // If a tail called function callee has more arguments than the caller the
2193 // caller needs to make sure that there is room to move the RETADDR to. This is
2194 // achieved by reserving an area the size of the argument delta right after the
2195 // original REtADDR, but before the saved framepointer or the spilled registers
2196 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2208 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2209 /// for a 16 byte align requirement.
2210 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2211 SelectionDAG& DAG) {
2212 MachineFunction &MF = DAG.getMachineFunction();
2213 const TargetMachine &TM = MF.getTarget();
2214 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2215 unsigned StackAlignment = TFI.getStackAlignment();
2216 uint64_t AlignMask = StackAlignment - 1;
2217 int64_t Offset = StackSize;
2218 uint64_t SlotSize = TD->getPointerSize();
2219 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2220 // Number smaller than 12 so just add the difference.
2221 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2223 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2224 Offset = ((~AlignMask) & Offset) + StackAlignment +
2225 (StackAlignment-SlotSize);
2230 /// MatchingStackOffset - Return true if the given stack call argument is
2231 /// already available in the same position (relatively) of the caller's
2232 /// incoming argument stack.
2234 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2235 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2236 const X86InstrInfo *TII) {
2237 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2239 if (Arg.getOpcode() == ISD::CopyFromReg) {
2240 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2241 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2243 MachineInstr *Def = MRI->getVRegDef(VR);
2246 if (!Flags.isByVal()) {
2247 if (!TII->isLoadFromStackSlot(Def, FI))
2250 unsigned Opcode = Def->getOpcode();
2251 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2252 Def->getOperand(1).isFI()) {
2253 FI = Def->getOperand(1).getIndex();
2254 Bytes = Flags.getByValSize();
2258 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2259 if (Flags.isByVal())
2260 // ByVal argument is passed in as a pointer but it's now being
2261 // dereferenced. e.g.
2262 // define @foo(%struct.X* %A) {
2263 // tail call @bar(%struct.X* byval %A)
2266 SDValue Ptr = Ld->getBasePtr();
2267 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2270 FI = FINode->getIndex();
2274 assert(FI != INT_MAX);
2275 if (!MFI->isFixedObjectIndex(FI))
2277 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2280 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2281 /// for tail call optimization. Targets which want to do tail call
2282 /// optimization should implement this function.
2284 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2285 CallingConv::ID CalleeCC,
2287 bool isCalleeStructRet,
2288 bool isCallerStructRet,
2289 const SmallVectorImpl<ISD::OutputArg> &Outs,
2290 const SmallVectorImpl<ISD::InputArg> &Ins,
2291 SelectionDAG& DAG) const {
2292 if (!IsTailCallConvention(CalleeCC) &&
2293 CalleeCC != CallingConv::C)
2296 // If -tailcallopt is specified, make fastcc functions tail-callable.
2297 const MachineFunction &MF = DAG.getMachineFunction();
2298 const Function *CallerF = DAG.getMachineFunction().getFunction();
2299 if (GuaranteedTailCallOpt) {
2300 if (IsTailCallConvention(CalleeCC) &&
2301 CallerF->getCallingConv() == CalleeCC)
2306 // Look for obvious safe cases to perform tail call optimization that does not
2307 // requite ABI changes. This is what gcc calls sibcall.
2309 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2310 // emit a special epilogue.
2311 if (RegInfo->needsStackRealignment(MF))
2314 // Do not sibcall optimize vararg calls unless the call site is not passing any
2316 if (isVarArg && !Outs.empty())
2319 // Also avoid sibcall optimization if either caller or callee uses struct
2320 // return semantics.
2321 if (isCalleeStructRet || isCallerStructRet)
2324 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2325 // Therefore if it's not used by the call it is not safe to optimize this into
2327 bool Unused = false;
2328 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2335 SmallVector<CCValAssign, 16> RVLocs;
2336 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2337 RVLocs, *DAG.getContext());
2338 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2339 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2340 CCValAssign &VA = RVLocs[i];
2341 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2346 // If the callee takes no arguments then go on to check the results of the
2348 if (!Outs.empty()) {
2349 // Check if stack adjustment is needed. For now, do not do this if any
2350 // argument is passed on the stack.
2351 SmallVector<CCValAssign, 16> ArgLocs;
2352 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2353 ArgLocs, *DAG.getContext());
2354 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2355 if (CCInfo.getNextStackOffset()) {
2356 MachineFunction &MF = DAG.getMachineFunction();
2357 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2359 if (Subtarget->isTargetWin64())
2360 // Win64 ABI has additional complications.
2363 // Check if the arguments are already laid out in the right way as
2364 // the caller's fixed stack objects.
2365 MachineFrameInfo *MFI = MF.getFrameInfo();
2366 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2367 const X86InstrInfo *TII =
2368 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2369 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2370 CCValAssign &VA = ArgLocs[i];
2371 EVT RegVT = VA.getLocVT();
2372 SDValue Arg = Outs[i].Val;
2373 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2374 if (VA.getLocInfo() == CCValAssign::Indirect)
2376 if (!VA.isRegLoc()) {
2377 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2389 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2391 DenseMap<const Value *, unsigned> &vm,
2392 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2393 DenseMap<const AllocaInst *, int> &am
2395 , SmallSet<Instruction*, 8> &cil
2398 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2406 //===----------------------------------------------------------------------===//
2407 // Other Lowering Hooks
2408 //===----------------------------------------------------------------------===//
2411 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2412 MachineFunction &MF = DAG.getMachineFunction();
2413 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2414 int ReturnAddrIndex = FuncInfo->getRAIndex();
2416 if (ReturnAddrIndex == 0) {
2417 // Set up a frame object for the return address.
2418 uint64_t SlotSize = TD->getPointerSize();
2419 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2421 FuncInfo->setRAIndex(ReturnAddrIndex);
2424 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2428 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2429 bool hasSymbolicDisplacement) {
2430 // Offset should fit into 32 bit immediate field.
2431 if (!isInt<32>(Offset))
2434 // If we don't have a symbolic displacement - we don't have any extra
2436 if (!hasSymbolicDisplacement)
2439 // FIXME: Some tweaks might be needed for medium code model.
2440 if (M != CodeModel::Small && M != CodeModel::Kernel)
2443 // For small code model we assume that latest object is 16MB before end of 31
2444 // bits boundary. We may also accept pretty large negative constants knowing
2445 // that all objects are in the positive half of address space.
2446 if (M == CodeModel::Small && Offset < 16*1024*1024)
2449 // For kernel code model we know that all object resist in the negative half
2450 // of 32bits address space. We may not accept negative offsets, since they may
2451 // be just off and we may accept pretty large positive ones.
2452 if (M == CodeModel::Kernel && Offset > 0)
2458 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2459 /// specific condition code, returning the condition code and the LHS/RHS of the
2460 /// comparison to make.
2461 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2462 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2464 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2465 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2466 // X > -1 -> X == 0, jump !sign.
2467 RHS = DAG.getConstant(0, RHS.getValueType());
2468 return X86::COND_NS;
2469 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2470 // X < 0 -> X == 0, jump on sign.
2472 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2474 RHS = DAG.getConstant(0, RHS.getValueType());
2475 return X86::COND_LE;
2479 switch (SetCCOpcode) {
2480 default: llvm_unreachable("Invalid integer condition!");
2481 case ISD::SETEQ: return X86::COND_E;
2482 case ISD::SETGT: return X86::COND_G;
2483 case ISD::SETGE: return X86::COND_GE;
2484 case ISD::SETLT: return X86::COND_L;
2485 case ISD::SETLE: return X86::COND_LE;
2486 case ISD::SETNE: return X86::COND_NE;
2487 case ISD::SETULT: return X86::COND_B;
2488 case ISD::SETUGT: return X86::COND_A;
2489 case ISD::SETULE: return X86::COND_BE;
2490 case ISD::SETUGE: return X86::COND_AE;
2494 // First determine if it is required or is profitable to flip the operands.
2496 // If LHS is a foldable load, but RHS is not, flip the condition.
2497 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2498 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2499 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2500 std::swap(LHS, RHS);
2503 switch (SetCCOpcode) {
2509 std::swap(LHS, RHS);
2513 // On a floating point condition, the flags are set as follows:
2515 // 0 | 0 | 0 | X > Y
2516 // 0 | 0 | 1 | X < Y
2517 // 1 | 0 | 0 | X == Y
2518 // 1 | 1 | 1 | unordered
2519 switch (SetCCOpcode) {
2520 default: llvm_unreachable("Condcode should be pre-legalized away");
2522 case ISD::SETEQ: return X86::COND_E;
2523 case ISD::SETOLT: // flipped
2525 case ISD::SETGT: return X86::COND_A;
2526 case ISD::SETOLE: // flipped
2528 case ISD::SETGE: return X86::COND_AE;
2529 case ISD::SETUGT: // flipped
2531 case ISD::SETLT: return X86::COND_B;
2532 case ISD::SETUGE: // flipped
2534 case ISD::SETLE: return X86::COND_BE;
2536 case ISD::SETNE: return X86::COND_NE;
2537 case ISD::SETUO: return X86::COND_P;
2538 case ISD::SETO: return X86::COND_NP;
2540 case ISD::SETUNE: return X86::COND_INVALID;
2544 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2545 /// code. Current x86 isa includes the following FP cmov instructions:
2546 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2547 static bool hasFPCMov(unsigned X86CC) {
2563 /// isFPImmLegal - Returns true if the target can instruction select the
2564 /// specified FP immediate natively. If false, the legalizer will
2565 /// materialize the FP immediate as a load from a constant pool.
2566 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2567 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2568 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2574 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2575 /// the specified range (L, H].
2576 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2577 return (Val < 0) || (Val >= Low && Val < Hi);
2580 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2581 /// specified value.
2582 static bool isUndefOrEqual(int Val, int CmpVal) {
2583 if (Val < 0 || Val == CmpVal)
2588 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2589 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2590 /// the second operand.
2591 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2592 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2593 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2594 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2595 return (Mask[0] < 2 && Mask[1] < 2);
2599 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2600 SmallVector<int, 8> M;
2602 return ::isPSHUFDMask(M, N->getValueType(0));
2605 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2606 /// is suitable for input to PSHUFHW.
2607 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2608 if (VT != MVT::v8i16)
2611 // Lower quadword copied in order or undef.
2612 for (int i = 0; i != 4; ++i)
2613 if (Mask[i] >= 0 && Mask[i] != i)
2616 // Upper quadword shuffled.
2617 for (int i = 4; i != 8; ++i)
2618 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2624 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2625 SmallVector<int, 8> M;
2627 return ::isPSHUFHWMask(M, N->getValueType(0));
2630 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2631 /// is suitable for input to PSHUFLW.
2632 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2633 if (VT != MVT::v8i16)
2636 // Upper quadword copied in order.
2637 for (int i = 4; i != 8; ++i)
2638 if (Mask[i] >= 0 && Mask[i] != i)
2641 // Lower quadword shuffled.
2642 for (int i = 0; i != 4; ++i)
2649 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2650 SmallVector<int, 8> M;
2652 return ::isPSHUFLWMask(M, N->getValueType(0));
2655 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2656 /// is suitable for input to PALIGNR.
2657 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2659 int i, e = VT.getVectorNumElements();
2661 // Do not handle v2i64 / v2f64 shuffles with palignr.
2662 if (e < 4 || !hasSSSE3)
2665 for (i = 0; i != e; ++i)
2669 // All undef, not a palignr.
2673 // Determine if it's ok to perform a palignr with only the LHS, since we
2674 // don't have access to the actual shuffle elements to see if RHS is undef.
2675 bool Unary = Mask[i] < (int)e;
2676 bool NeedsUnary = false;
2678 int s = Mask[i] - i;
2680 // Check the rest of the elements to see if they are consecutive.
2681 for (++i; i != e; ++i) {
2686 Unary = Unary && (m < (int)e);
2687 NeedsUnary = NeedsUnary || (m < s);
2689 if (NeedsUnary && !Unary)
2691 if (Unary && m != ((s+i) & (e-1)))
2693 if (!Unary && m != (s+i))
2699 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2700 SmallVector<int, 8> M;
2702 return ::isPALIGNRMask(M, N->getValueType(0), true);
2705 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2706 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2707 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2708 int NumElems = VT.getVectorNumElements();
2709 if (NumElems != 2 && NumElems != 4)
2712 int Half = NumElems / 2;
2713 for (int i = 0; i < Half; ++i)
2714 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2716 for (int i = Half; i < NumElems; ++i)
2717 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2723 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2724 SmallVector<int, 8> M;
2726 return ::isSHUFPMask(M, N->getValueType(0));
2729 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2730 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2731 /// half elements to come from vector 1 (which would equal the dest.) and
2732 /// the upper half to come from vector 2.
2733 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2734 int NumElems = VT.getVectorNumElements();
2736 if (NumElems != 2 && NumElems != 4)
2739 int Half = NumElems / 2;
2740 for (int i = 0; i < Half; ++i)
2741 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2743 for (int i = Half; i < NumElems; ++i)
2744 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2749 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2750 SmallVector<int, 8> M;
2752 return isCommutedSHUFPMask(M, N->getValueType(0));
2755 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2756 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2757 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2758 if (N->getValueType(0).getVectorNumElements() != 4)
2761 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2762 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2763 isUndefOrEqual(N->getMaskElt(1), 7) &&
2764 isUndefOrEqual(N->getMaskElt(2), 2) &&
2765 isUndefOrEqual(N->getMaskElt(3), 3);
2768 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2769 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2771 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2772 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2777 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2778 isUndefOrEqual(N->getMaskElt(1), 3) &&
2779 isUndefOrEqual(N->getMaskElt(2), 2) &&
2780 isUndefOrEqual(N->getMaskElt(3), 3);
2783 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2784 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2785 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2786 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2788 if (NumElems != 2 && NumElems != 4)
2791 for (unsigned i = 0; i < NumElems/2; ++i)
2792 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2795 for (unsigned i = NumElems/2; i < NumElems; ++i)
2796 if (!isUndefOrEqual(N->getMaskElt(i), i))
2802 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2803 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2804 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2805 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2807 if (NumElems != 2 && NumElems != 4)
2810 for (unsigned i = 0; i < NumElems/2; ++i)
2811 if (!isUndefOrEqual(N->getMaskElt(i), i))
2814 for (unsigned i = 0; i < NumElems/2; ++i)
2815 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2821 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2822 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2823 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2824 bool V2IsSplat = false) {
2825 int NumElts = VT.getVectorNumElements();
2826 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2829 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2831 int BitI1 = Mask[i+1];
2832 if (!isUndefOrEqual(BitI, j))
2835 if (!isUndefOrEqual(BitI1, NumElts))
2838 if (!isUndefOrEqual(BitI1, j + NumElts))
2845 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2846 SmallVector<int, 8> M;
2848 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2851 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2852 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2853 static bool isUNPCKHMask(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 + NumElts/2))
2865 if (isUndefOrEqual(BitI1, NumElts))
2868 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2875 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2876 SmallVector<int, 8> M;
2878 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2881 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2882 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2884 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2885 int NumElems = VT.getVectorNumElements();
2886 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2889 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2891 int BitI1 = Mask[i+1];
2892 if (!isUndefOrEqual(BitI, j))
2894 if (!isUndefOrEqual(BitI1, j))
2900 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2901 SmallVector<int, 8> M;
2903 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2906 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2907 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2909 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2910 int NumElems = VT.getVectorNumElements();
2911 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2914 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2916 int BitI1 = Mask[i+1];
2917 if (!isUndefOrEqual(BitI, j))
2919 if (!isUndefOrEqual(BitI1, j))
2925 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2926 SmallVector<int, 8> M;
2928 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2931 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2932 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2933 /// MOVSD, and MOVD, i.e. setting the lowest element.
2934 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2935 if (VT.getVectorElementType().getSizeInBits() < 32)
2938 int NumElts = VT.getVectorNumElements();
2940 if (!isUndefOrEqual(Mask[0], NumElts))
2943 for (int i = 1; i < NumElts; ++i)
2944 if (!isUndefOrEqual(Mask[i], i))
2950 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2951 SmallVector<int, 8> M;
2953 return ::isMOVLMask(M, N->getValueType(0));
2956 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2957 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2958 /// element of vector 2 and the other elements to come from vector 1 in order.
2959 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2960 bool V2IsSplat = false, bool V2IsUndef = false) {
2961 int NumOps = VT.getVectorNumElements();
2962 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2965 if (!isUndefOrEqual(Mask[0], 0))
2968 for (int i = 1; i < NumOps; ++i)
2969 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2970 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2971 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2977 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2978 bool V2IsUndef = false) {
2979 SmallVector<int, 8> M;
2981 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2984 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2985 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2986 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2987 if (N->getValueType(0).getVectorNumElements() != 4)
2990 // Expect 1, 1, 3, 3
2991 for (unsigned i = 0; i < 2; ++i) {
2992 int Elt = N->getMaskElt(i);
2993 if (Elt >= 0 && Elt != 1)
2998 for (unsigned i = 2; i < 4; ++i) {
2999 int Elt = N->getMaskElt(i);
3000 if (Elt >= 0 && Elt != 3)
3005 // Don't use movshdup if it can be done with a shufps.
3006 // FIXME: verify that matching u, u, 3, 3 is what we want.
3010 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3011 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3012 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3013 if (N->getValueType(0).getVectorNumElements() != 4)
3016 // Expect 0, 0, 2, 2
3017 for (unsigned i = 0; i < 2; ++i)
3018 if (N->getMaskElt(i) > 0)
3022 for (unsigned i = 2; i < 4; ++i) {
3023 int Elt = N->getMaskElt(i);
3024 if (Elt >= 0 && Elt != 2)
3029 // Don't use movsldup if it can be done with a shufps.
3033 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3034 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3035 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3036 int e = N->getValueType(0).getVectorNumElements() / 2;
3038 for (int i = 0; i < e; ++i)
3039 if (!isUndefOrEqual(N->getMaskElt(i), i))
3041 for (int i = 0; i < e; ++i)
3042 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3047 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3048 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3049 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3050 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3051 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3053 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3055 for (int i = 0; i < NumOperands; ++i) {
3056 int Val = SVOp->getMaskElt(NumOperands-i-1);
3057 if (Val < 0) Val = 0;
3058 if (Val >= NumOperands) Val -= NumOperands;
3060 if (i != NumOperands - 1)
3066 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3067 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3068 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3069 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3071 // 8 nodes, but we only care about the last 4.
3072 for (unsigned i = 7; i >= 4; --i) {
3073 int Val = SVOp->getMaskElt(i);
3082 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3083 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3084 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3085 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3087 // 8 nodes, but we only care about the first 4.
3088 for (int i = 3; i >= 0; --i) {
3089 int Val = SVOp->getMaskElt(i);
3098 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3099 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3100 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3101 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3102 EVT VVT = N->getValueType(0);
3103 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3107 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3108 Val = SVOp->getMaskElt(i);
3112 return (Val - i) * EltSize;
3115 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3117 bool X86::isZeroNode(SDValue Elt) {
3118 return ((isa<ConstantSDNode>(Elt) &&
3119 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3120 (isa<ConstantFPSDNode>(Elt) &&
3121 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3124 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3125 /// their permute mask.
3126 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3127 SelectionDAG &DAG) {
3128 EVT VT = SVOp->getValueType(0);
3129 unsigned NumElems = VT.getVectorNumElements();
3130 SmallVector<int, 8> MaskVec;
3132 for (unsigned i = 0; i != NumElems; ++i) {
3133 int idx = SVOp->getMaskElt(i);
3135 MaskVec.push_back(idx);
3136 else if (idx < (int)NumElems)
3137 MaskVec.push_back(idx + NumElems);
3139 MaskVec.push_back(idx - NumElems);
3141 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3142 SVOp->getOperand(0), &MaskVec[0]);
3145 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3146 /// the two vector operands have swapped position.
3147 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3148 unsigned NumElems = VT.getVectorNumElements();
3149 for (unsigned i = 0; i != NumElems; ++i) {
3153 else if (idx < (int)NumElems)
3154 Mask[i] = idx + NumElems;
3156 Mask[i] = idx - NumElems;
3160 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3161 /// match movhlps. The lower half elements should come from upper half of
3162 /// V1 (and in order), and the upper half elements should come from the upper
3163 /// half of V2 (and in order).
3164 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3165 if (Op->getValueType(0).getVectorNumElements() != 4)
3167 for (unsigned i = 0, e = 2; i != e; ++i)
3168 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3170 for (unsigned i = 2; i != 4; ++i)
3171 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3176 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3177 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3179 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3180 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3182 N = N->getOperand(0).getNode();
3183 if (!ISD::isNON_EXTLoad(N))
3186 *LD = cast<LoadSDNode>(N);
3190 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3191 /// match movlp{s|d}. The lower half elements should come from lower half of
3192 /// V1 (and in order), and the upper half elements should come from the upper
3193 /// half of V2 (and in order). And since V1 will become the source of the
3194 /// MOVLP, it must be either a vector load or a scalar load to vector.
3195 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3196 ShuffleVectorSDNode *Op) {
3197 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3199 // Is V2 is a vector load, don't do this transformation. We will try to use
3200 // load folding shufps op.
3201 if (ISD::isNON_EXTLoad(V2))
3204 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3206 if (NumElems != 2 && NumElems != 4)
3208 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3209 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3211 for (unsigned i = NumElems/2; i != NumElems; ++i)
3212 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3217 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3219 static bool isSplatVector(SDNode *N) {
3220 if (N->getOpcode() != ISD::BUILD_VECTOR)
3223 SDValue SplatValue = N->getOperand(0);
3224 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3225 if (N->getOperand(i) != SplatValue)
3230 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3231 /// to an zero vector.
3232 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3233 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3234 SDValue V1 = N->getOperand(0);
3235 SDValue V2 = N->getOperand(1);
3236 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3237 for (unsigned i = 0; i != NumElems; ++i) {
3238 int Idx = N->getMaskElt(i);
3239 if (Idx >= (int)NumElems) {
3240 unsigned Opc = V2.getOpcode();
3241 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3243 if (Opc != ISD::BUILD_VECTOR ||
3244 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3246 } else if (Idx >= 0) {
3247 unsigned Opc = V1.getOpcode();
3248 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3250 if (Opc != ISD::BUILD_VECTOR ||
3251 !X86::isZeroNode(V1.getOperand(Idx)))
3258 /// getZeroVector - Returns a vector of specified type with all zero elements.
3260 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3262 assert(VT.isVector() && "Expected a vector type");
3264 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3265 // type. This ensures they get CSE'd.
3267 if (VT.getSizeInBits() == 64) { // MMX
3268 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3269 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3270 } else if (HasSSE2) { // SSE2
3271 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3272 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3274 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3275 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3277 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3280 /// getOnesVector - Returns a vector of specified type with all bits set.
3282 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3283 assert(VT.isVector() && "Expected a vector type");
3285 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3286 // type. This ensures they get CSE'd.
3287 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3289 if (VT.getSizeInBits() == 64) // MMX
3290 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3292 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3293 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3297 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3298 /// that point to V2 points to its first element.
3299 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3300 EVT VT = SVOp->getValueType(0);
3301 unsigned NumElems = VT.getVectorNumElements();
3303 bool Changed = false;
3304 SmallVector<int, 8> MaskVec;
3305 SVOp->getMask(MaskVec);
3307 for (unsigned i = 0; i != NumElems; ++i) {
3308 if (MaskVec[i] > (int)NumElems) {
3309 MaskVec[i] = NumElems;
3314 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3315 SVOp->getOperand(1), &MaskVec[0]);
3316 return SDValue(SVOp, 0);
3319 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3320 /// operation of specified width.
3321 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3323 unsigned NumElems = VT.getVectorNumElements();
3324 SmallVector<int, 8> Mask;
3325 Mask.push_back(NumElems);
3326 for (unsigned i = 1; i != NumElems; ++i)
3328 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3331 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3332 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3334 unsigned NumElems = VT.getVectorNumElements();
3335 SmallVector<int, 8> Mask;
3336 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3338 Mask.push_back(i + NumElems);
3340 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3343 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3344 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3346 unsigned NumElems = VT.getVectorNumElements();
3347 unsigned Half = NumElems/2;
3348 SmallVector<int, 8> Mask;
3349 for (unsigned i = 0; i != Half; ++i) {
3350 Mask.push_back(i + Half);
3351 Mask.push_back(i + NumElems + Half);
3353 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3356 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3357 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3359 if (SV->getValueType(0).getVectorNumElements() <= 4)
3360 return SDValue(SV, 0);
3362 EVT PVT = MVT::v4f32;
3363 EVT VT = SV->getValueType(0);
3364 DebugLoc dl = SV->getDebugLoc();
3365 SDValue V1 = SV->getOperand(0);
3366 int NumElems = VT.getVectorNumElements();
3367 int EltNo = SV->getSplatIndex();
3369 // unpack elements to the correct location
3370 while (NumElems > 4) {
3371 if (EltNo < NumElems/2) {
3372 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3374 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3375 EltNo -= NumElems/2;
3380 // Perform the splat.
3381 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3382 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3383 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3384 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3387 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3388 /// vector of zero or undef vector. This produces a shuffle where the low
3389 /// element of V2 is swizzled into the zero/undef vector, landing at element
3390 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3391 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3392 bool isZero, bool HasSSE2,
3393 SelectionDAG &DAG) {
3394 EVT VT = V2.getValueType();
3396 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3397 unsigned NumElems = VT.getVectorNumElements();
3398 SmallVector<int, 16> MaskVec;
3399 for (unsigned i = 0; i != NumElems; ++i)
3400 // If this is the insertion idx, put the low elt of V2 here.
3401 MaskVec.push_back(i == Idx ? NumElems : i);
3402 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3405 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3406 /// a shuffle that is zero.
3408 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3409 bool Low, SelectionDAG &DAG) {
3410 unsigned NumZeros = 0;
3411 for (int i = 0; i < NumElems; ++i) {
3412 unsigned Index = Low ? i : NumElems-i-1;
3413 int Idx = SVOp->getMaskElt(Index);
3418 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3419 if (Elt.getNode() && X86::isZeroNode(Elt))
3427 /// isVectorShift - Returns true if the shuffle can be implemented as a
3428 /// logical left or right shift of a vector.
3429 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3430 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3431 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3432 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3435 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3438 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3442 bool SeenV1 = false;
3443 bool SeenV2 = false;
3444 for (int i = NumZeros; i < NumElems; ++i) {
3445 int Val = isLeft ? (i - NumZeros) : i;
3446 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3458 if (SeenV1 && SeenV2)
3461 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3467 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3469 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3470 unsigned NumNonZero, unsigned NumZero,
3471 SelectionDAG &DAG, TargetLowering &TLI) {
3475 DebugLoc dl = Op.getDebugLoc();
3478 for (unsigned i = 0; i < 16; ++i) {
3479 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3480 if (ThisIsNonZero && First) {
3482 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3484 V = DAG.getUNDEF(MVT::v8i16);
3489 SDValue ThisElt(0, 0), LastElt(0, 0);
3490 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3491 if (LastIsNonZero) {
3492 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3493 MVT::i16, Op.getOperand(i-1));
3495 if (ThisIsNonZero) {
3496 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3497 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3498 ThisElt, DAG.getConstant(8, MVT::i8));
3500 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3504 if (ThisElt.getNode())
3505 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3506 DAG.getIntPtrConstant(i/2));
3510 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3513 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3515 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3516 unsigned NumNonZero, unsigned NumZero,
3517 SelectionDAG &DAG, TargetLowering &TLI) {
3521 DebugLoc dl = Op.getDebugLoc();
3524 for (unsigned i = 0; i < 8; ++i) {
3525 bool isNonZero = (NonZeros & (1 << i)) != 0;
3529 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3531 V = DAG.getUNDEF(MVT::v8i16);
3534 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3535 MVT::v8i16, V, Op.getOperand(i),
3536 DAG.getIntPtrConstant(i));
3543 /// getVShift - Return a vector logical shift node.
3545 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3546 unsigned NumBits, SelectionDAG &DAG,
3547 const TargetLowering &TLI, DebugLoc dl) {
3548 bool isMMX = VT.getSizeInBits() == 64;
3549 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3550 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3551 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3552 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3553 DAG.getNode(Opc, dl, ShVT, SrcOp,
3554 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3558 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3559 SelectionDAG &DAG) {
3561 // Check if the scalar load can be widened into a vector load. And if
3562 // the address is "base + cst" see if the cst can be "absorbed" into
3563 // the shuffle mask.
3564 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3565 SDValue Ptr = LD->getBasePtr();
3566 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3568 EVT PVT = LD->getValueType(0);
3569 if (PVT != MVT::i32 && PVT != MVT::f32)
3574 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3575 FI = FINode->getIndex();
3577 } else if (Ptr.getOpcode() == ISD::ADD &&
3578 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3579 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3580 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3581 Offset = Ptr.getConstantOperandVal(1);
3582 Ptr = Ptr.getOperand(0);
3587 SDValue Chain = LD->getChain();
3588 // Make sure the stack object alignment is at least 16.
3589 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3590 if (DAG.InferPtrAlignment(Ptr) < 16) {
3591 if (MFI->isFixedObjectIndex(FI)) {
3592 // Can't change the alignment. FIXME: It's possible to compute
3593 // the exact stack offset and reference FI + adjust offset instead.
3594 // If someone *really* cares about this. That's the way to implement it.
3597 MFI->setObjectAlignment(FI, 16);
3601 // (Offset % 16) must be multiple of 4. Then address is then
3602 // Ptr + (Offset & ~15).
3605 if ((Offset % 16) & 3)
3607 int64_t StartOffset = Offset & ~15;
3609 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3610 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3612 int EltNo = (Offset - StartOffset) >> 2;
3613 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3614 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3615 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3617 // Canonicalize it to a v4i32 shuffle.
3618 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3619 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3620 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3621 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3627 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3628 /// vector of type 'VT', see if the elements can be replaced by a single large
3629 /// load which has the same value as a build_vector whose operands are 'elts'.
3631 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3633 /// FIXME: we'd also like to handle the case where the last elements are zero
3634 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3635 /// There's even a handy isZeroNode for that purpose.
3636 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3637 DebugLoc &dl, SelectionDAG &DAG) {
3638 EVT EltVT = VT.getVectorElementType();
3639 unsigned NumElems = Elts.size();
3641 LoadSDNode *LDBase = NULL;
3642 unsigned LastLoadedElt = -1U;
3644 // For each element in the initializer, see if we've found a load or an undef.
3645 // If we don't find an initial load element, or later load elements are
3646 // non-consecutive, bail out.
3647 for (unsigned i = 0; i < NumElems; ++i) {
3648 SDValue Elt = Elts[i];
3650 if (!Elt.getNode() ||
3651 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3654 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3656 LDBase = cast<LoadSDNode>(Elt.getNode());
3660 if (Elt.getOpcode() == ISD::UNDEF)
3663 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3664 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3669 // If we have found an entire vector of loads and undefs, then return a large
3670 // load of the entire vector width starting at the base pointer. If we found
3671 // consecutive loads for the low half, generate a vzext_load node.
3672 if (LastLoadedElt == NumElems - 1) {
3673 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3674 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3675 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3676 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3677 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3678 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3679 LDBase->isVolatile(), LDBase->isNonTemporal(),
3680 LDBase->getAlignment());
3681 } else if (NumElems == 4 && LastLoadedElt == 1) {
3682 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3683 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3684 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3685 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3691 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3692 DebugLoc dl = Op.getDebugLoc();
3693 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3694 if (ISD::isBuildVectorAllZeros(Op.getNode())
3695 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3696 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3697 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3698 // eliminated on x86-32 hosts.
3699 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3702 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3703 return getOnesVector(Op.getValueType(), DAG, dl);
3704 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3707 EVT VT = Op.getValueType();
3708 EVT ExtVT = VT.getVectorElementType();
3709 unsigned EVTBits = ExtVT.getSizeInBits();
3711 unsigned NumElems = Op.getNumOperands();
3712 unsigned NumZero = 0;
3713 unsigned NumNonZero = 0;
3714 unsigned NonZeros = 0;
3715 bool IsAllConstants = true;
3716 SmallSet<SDValue, 8> Values;
3717 for (unsigned i = 0; i < NumElems; ++i) {
3718 SDValue Elt = Op.getOperand(i);
3719 if (Elt.getOpcode() == ISD::UNDEF)
3722 if (Elt.getOpcode() != ISD::Constant &&
3723 Elt.getOpcode() != ISD::ConstantFP)
3724 IsAllConstants = false;
3725 if (X86::isZeroNode(Elt))
3728 NonZeros |= (1 << i);
3733 if (NumNonZero == 0) {
3734 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3735 return DAG.getUNDEF(VT);
3738 // Special case for single non-zero, non-undef, element.
3739 if (NumNonZero == 1) {
3740 unsigned Idx = CountTrailingZeros_32(NonZeros);
3741 SDValue Item = Op.getOperand(Idx);
3743 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3744 // the value are obviously zero, truncate the value to i32 and do the
3745 // insertion that way. Only do this if the value is non-constant or if the
3746 // value is a constant being inserted into element 0. It is cheaper to do
3747 // a constant pool load than it is to do a movd + shuffle.
3748 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3749 (!IsAllConstants || Idx == 0)) {
3750 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3751 // Handle MMX and SSE both.
3752 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3753 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3755 // Truncate the value (which may itself be a constant) to i32, and
3756 // convert it to a vector with movd (S2V+shuffle to zero extend).
3757 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3758 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3759 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3760 Subtarget->hasSSE2(), DAG);
3762 // Now we have our 32-bit value zero extended in the low element of
3763 // a vector. If Idx != 0, swizzle it into place.
3765 SmallVector<int, 4> Mask;
3766 Mask.push_back(Idx);
3767 for (unsigned i = 1; i != VecElts; ++i)
3769 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3770 DAG.getUNDEF(Item.getValueType()),
3773 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3777 // If we have a constant or non-constant insertion into the low element of
3778 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3779 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3780 // depending on what the source datatype is.
3783 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3784 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3785 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3786 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3787 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3788 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3790 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3791 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3792 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3793 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3794 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3795 Subtarget->hasSSE2(), DAG);
3796 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3800 // Is it a vector logical left shift?
3801 if (NumElems == 2 && Idx == 1 &&
3802 X86::isZeroNode(Op.getOperand(0)) &&
3803 !X86::isZeroNode(Op.getOperand(1))) {
3804 unsigned NumBits = VT.getSizeInBits();
3805 return getVShift(true, VT,
3806 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3807 VT, Op.getOperand(1)),
3808 NumBits/2, DAG, *this, dl);
3811 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3814 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3815 // is a non-constant being inserted into an element other than the low one,
3816 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3817 // movd/movss) to move this into the low element, then shuffle it into
3819 if (EVTBits == 32) {
3820 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3822 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3823 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3824 Subtarget->hasSSE2(), DAG);
3825 SmallVector<int, 8> MaskVec;
3826 for (unsigned i = 0; i < NumElems; i++)
3827 MaskVec.push_back(i == Idx ? 0 : 1);
3828 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3832 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3833 if (Values.size() == 1) {
3834 if (EVTBits == 32) {
3835 // Instead of a shuffle like this:
3836 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3837 // Check if it's possible to issue this instead.
3838 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3839 unsigned Idx = CountTrailingZeros_32(NonZeros);
3840 SDValue Item = Op.getOperand(Idx);
3841 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3842 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3847 // A vector full of immediates; various special cases are already
3848 // handled, so this is best done with a single constant-pool load.
3852 // Let legalizer expand 2-wide build_vectors.
3853 if (EVTBits == 64) {
3854 if (NumNonZero == 1) {
3855 // One half is zero or undef.
3856 unsigned Idx = CountTrailingZeros_32(NonZeros);
3857 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3858 Op.getOperand(Idx));
3859 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3860 Subtarget->hasSSE2(), DAG);
3865 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3866 if (EVTBits == 8 && NumElems == 16) {
3867 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3869 if (V.getNode()) return V;
3872 if (EVTBits == 16 && NumElems == 8) {
3873 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3875 if (V.getNode()) return V;
3878 // If element VT is == 32 bits, turn it into a number of shuffles.
3879 SmallVector<SDValue, 8> V;
3881 if (NumElems == 4 && NumZero > 0) {
3882 for (unsigned i = 0; i < 4; ++i) {
3883 bool isZero = !(NonZeros & (1 << i));
3885 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3887 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3890 for (unsigned i = 0; i < 2; ++i) {
3891 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3894 V[i] = V[i*2]; // Must be a zero vector.
3897 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3900 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3903 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3908 SmallVector<int, 8> MaskVec;
3909 bool Reverse = (NonZeros & 0x3) == 2;
3910 for (unsigned i = 0; i < 2; ++i)
3911 MaskVec.push_back(Reverse ? 1-i : i);
3912 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3913 for (unsigned i = 0; i < 2; ++i)
3914 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3915 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3918 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
3919 // Check for a build vector of consecutive loads.
3920 for (unsigned i = 0; i < NumElems; ++i)
3921 V[i] = Op.getOperand(i);
3923 // Check for elements which are consecutive loads.
3924 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
3928 // For SSE 4.1, use inserts into undef.
3929 if (getSubtarget()->hasSSE41()) {
3930 V[0] = DAG.getUNDEF(VT);
3931 for (unsigned i = 0; i < NumElems; ++i)
3932 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3933 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3934 Op.getOperand(i), DAG.getIntPtrConstant(i));
3938 // Otherwise, expand into a number of unpckl*
3940 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3941 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3942 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3943 for (unsigned i = 0; i < NumElems; ++i)
3944 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3946 while (NumElems != 0) {
3947 for (unsigned i = 0; i < NumElems; ++i)
3948 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3957 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3958 // We support concatenate two MMX registers and place them in a MMX
3959 // register. This is better than doing a stack convert.
3960 DebugLoc dl = Op.getDebugLoc();
3961 EVT ResVT = Op.getValueType();
3962 assert(Op.getNumOperands() == 2);
3963 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3964 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3966 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3967 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3968 InVec = Op.getOperand(1);
3969 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3970 unsigned NumElts = ResVT.getVectorNumElements();
3971 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3972 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3973 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3975 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3976 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3977 Mask[0] = 0; Mask[1] = 2;
3978 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3980 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3983 // v8i16 shuffles - Prefer shuffles in the following order:
3984 // 1. [all] pshuflw, pshufhw, optional move
3985 // 2. [ssse3] 1 x pshufb
3986 // 3. [ssse3] 2 x pshufb + 1 x por
3987 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3989 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3990 SelectionDAG &DAG, X86TargetLowering &TLI) {
3991 SDValue V1 = SVOp->getOperand(0);
3992 SDValue V2 = SVOp->getOperand(1);
3993 DebugLoc dl = SVOp->getDebugLoc();
3994 SmallVector<int, 8> MaskVals;
3996 // Determine if more than 1 of the words in each of the low and high quadwords
3997 // of the result come from the same quadword of one of the two inputs. Undef
3998 // mask values count as coming from any quadword, for better codegen.
3999 SmallVector<unsigned, 4> LoQuad(4);
4000 SmallVector<unsigned, 4> HiQuad(4);
4001 BitVector InputQuads(4);
4002 for (unsigned i = 0; i < 8; ++i) {
4003 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4004 int EltIdx = SVOp->getMaskElt(i);
4005 MaskVals.push_back(EltIdx);
4014 InputQuads.set(EltIdx / 4);
4017 int BestLoQuad = -1;
4018 unsigned MaxQuad = 1;
4019 for (unsigned i = 0; i < 4; ++i) {
4020 if (LoQuad[i] > MaxQuad) {
4022 MaxQuad = LoQuad[i];
4026 int BestHiQuad = -1;
4028 for (unsigned i = 0; i < 4; ++i) {
4029 if (HiQuad[i] > MaxQuad) {
4031 MaxQuad = HiQuad[i];
4035 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4036 // of the two input vectors, shuffle them into one input vector so only a
4037 // single pshufb instruction is necessary. If There are more than 2 input
4038 // quads, disable the next transformation since it does not help SSSE3.
4039 bool V1Used = InputQuads[0] || InputQuads[1];
4040 bool V2Used = InputQuads[2] || InputQuads[3];
4041 if (TLI.getSubtarget()->hasSSSE3()) {
4042 if (InputQuads.count() == 2 && V1Used && V2Used) {
4043 BestLoQuad = InputQuads.find_first();
4044 BestHiQuad = InputQuads.find_next(BestLoQuad);
4046 if (InputQuads.count() > 2) {
4052 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4053 // the shuffle mask. If a quad is scored as -1, that means that it contains
4054 // words from all 4 input quadwords.
4056 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4057 SmallVector<int, 8> MaskV;
4058 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4059 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4060 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4061 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4062 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4063 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4065 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4066 // source words for the shuffle, to aid later transformations.
4067 bool AllWordsInNewV = true;
4068 bool InOrder[2] = { true, true };
4069 for (unsigned i = 0; i != 8; ++i) {
4070 int idx = MaskVals[i];
4072 InOrder[i/4] = false;
4073 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4075 AllWordsInNewV = false;
4079 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4080 if (AllWordsInNewV) {
4081 for (int i = 0; i != 8; ++i) {
4082 int idx = MaskVals[i];
4085 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4086 if ((idx != i) && idx < 4)
4088 if ((idx != i) && idx > 3)
4097 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4098 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4099 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4100 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4101 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4105 // If we have SSSE3, and all words of the result are from 1 input vector,
4106 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4107 // is present, fall back to case 4.
4108 if (TLI.getSubtarget()->hasSSSE3()) {
4109 SmallVector<SDValue,16> pshufbMask;
4111 // If we have elements from both input vectors, set the high bit of the
4112 // shuffle mask element to zero out elements that come from V2 in the V1
4113 // mask, and elements that come from V1 in the V2 mask, so that the two
4114 // results can be OR'd together.
4115 bool TwoInputs = V1Used && V2Used;
4116 for (unsigned i = 0; i != 8; ++i) {
4117 int EltIdx = MaskVals[i] * 2;
4118 if (TwoInputs && (EltIdx >= 16)) {
4119 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4120 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4123 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4124 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4126 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4127 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4128 DAG.getNode(ISD::BUILD_VECTOR, dl,
4129 MVT::v16i8, &pshufbMask[0], 16));
4131 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4133 // Calculate the shuffle mask for the second input, shuffle it, and
4134 // OR it with the first shuffled input.
4136 for (unsigned i = 0; i != 8; ++i) {
4137 int EltIdx = MaskVals[i] * 2;
4139 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4140 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4143 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4144 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4146 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4147 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4148 DAG.getNode(ISD::BUILD_VECTOR, dl,
4149 MVT::v16i8, &pshufbMask[0], 16));
4150 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4151 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4154 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4155 // and update MaskVals with new element order.
4156 BitVector InOrder(8);
4157 if (BestLoQuad >= 0) {
4158 SmallVector<int, 8> MaskV;
4159 for (int i = 0; i != 4; ++i) {
4160 int idx = MaskVals[i];
4162 MaskV.push_back(-1);
4164 } else if ((idx / 4) == BestLoQuad) {
4165 MaskV.push_back(idx & 3);
4168 MaskV.push_back(-1);
4171 for (unsigned i = 4; i != 8; ++i)
4173 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4177 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4178 // and update MaskVals with the new element order.
4179 if (BestHiQuad >= 0) {
4180 SmallVector<int, 8> MaskV;
4181 for (unsigned i = 0; i != 4; ++i)
4183 for (unsigned i = 4; i != 8; ++i) {
4184 int idx = MaskVals[i];
4186 MaskV.push_back(-1);
4188 } else if ((idx / 4) == BestHiQuad) {
4189 MaskV.push_back((idx & 3) + 4);
4192 MaskV.push_back(-1);
4195 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4199 // In case BestHi & BestLo were both -1, which means each quadword has a word
4200 // from each of the four input quadwords, calculate the InOrder bitvector now
4201 // before falling through to the insert/extract cleanup.
4202 if (BestLoQuad == -1 && BestHiQuad == -1) {
4204 for (int i = 0; i != 8; ++i)
4205 if (MaskVals[i] < 0 || MaskVals[i] == i)
4209 // The other elements are put in the right place using pextrw and pinsrw.
4210 for (unsigned i = 0; i != 8; ++i) {
4213 int EltIdx = MaskVals[i];
4216 SDValue ExtOp = (EltIdx < 8)
4217 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4218 DAG.getIntPtrConstant(EltIdx))
4219 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4220 DAG.getIntPtrConstant(EltIdx - 8));
4221 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4222 DAG.getIntPtrConstant(i));
4227 // v16i8 shuffles - Prefer shuffles in the following order:
4228 // 1. [ssse3] 1 x pshufb
4229 // 2. [ssse3] 2 x pshufb + 1 x por
4230 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4232 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4233 SelectionDAG &DAG, X86TargetLowering &TLI) {
4234 SDValue V1 = SVOp->getOperand(0);
4235 SDValue V2 = SVOp->getOperand(1);
4236 DebugLoc dl = SVOp->getDebugLoc();
4237 SmallVector<int, 16> MaskVals;
4238 SVOp->getMask(MaskVals);
4240 // If we have SSSE3, case 1 is generated when all result bytes come from
4241 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4242 // present, fall back to case 3.
4243 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4246 for (unsigned i = 0; i < 16; ++i) {
4247 int EltIdx = MaskVals[i];
4256 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4257 if (TLI.getSubtarget()->hasSSSE3()) {
4258 SmallVector<SDValue,16> pshufbMask;
4260 // If all result elements are from one input vector, then only translate
4261 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4263 // Otherwise, we have elements from both input vectors, and must zero out
4264 // elements that come from V2 in the first mask, and V1 in the second mask
4265 // so that we can OR them together.
4266 bool TwoInputs = !(V1Only || V2Only);
4267 for (unsigned i = 0; i != 16; ++i) {
4268 int EltIdx = MaskVals[i];
4269 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4270 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4273 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4275 // If all the elements are from V2, assign it to V1 and return after
4276 // building the first pshufb.
4279 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4280 DAG.getNode(ISD::BUILD_VECTOR, dl,
4281 MVT::v16i8, &pshufbMask[0], 16));
4285 // Calculate the shuffle mask for the second input, shuffle it, and
4286 // OR it with the first shuffled input.
4288 for (unsigned i = 0; i != 16; ++i) {
4289 int EltIdx = MaskVals[i];
4291 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4294 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4296 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4297 DAG.getNode(ISD::BUILD_VECTOR, dl,
4298 MVT::v16i8, &pshufbMask[0], 16));
4299 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4302 // No SSSE3 - Calculate in place words and then fix all out of place words
4303 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4304 // the 16 different words that comprise the two doublequadword input vectors.
4305 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4306 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4307 SDValue NewV = V2Only ? V2 : V1;
4308 for (int i = 0; i != 8; ++i) {
4309 int Elt0 = MaskVals[i*2];
4310 int Elt1 = MaskVals[i*2+1];
4312 // This word of the result is all undef, skip it.
4313 if (Elt0 < 0 && Elt1 < 0)
4316 // This word of the result is already in the correct place, skip it.
4317 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4319 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4322 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4323 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4326 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4327 // using a single extract together, load it and store it.
4328 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4329 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4330 DAG.getIntPtrConstant(Elt1 / 2));
4331 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4332 DAG.getIntPtrConstant(i));
4336 // If Elt1 is defined, extract it from the appropriate source. If the
4337 // source byte is not also odd, shift the extracted word left 8 bits
4338 // otherwise clear the bottom 8 bits if we need to do an or.
4340 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4341 DAG.getIntPtrConstant(Elt1 / 2));
4342 if ((Elt1 & 1) == 0)
4343 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4344 DAG.getConstant(8, TLI.getShiftAmountTy()));
4346 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4347 DAG.getConstant(0xFF00, MVT::i16));
4349 // If Elt0 is defined, extract it from the appropriate source. If the
4350 // source byte is not also even, shift the extracted word right 8 bits. If
4351 // Elt1 was also defined, OR the extracted values together before
4352 // inserting them in the result.
4354 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4355 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4356 if ((Elt0 & 1) != 0)
4357 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4358 DAG.getConstant(8, TLI.getShiftAmountTy()));
4360 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4361 DAG.getConstant(0x00FF, MVT::i16));
4362 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4365 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4366 DAG.getIntPtrConstant(i));
4368 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4371 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4372 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4373 /// done when every pair / quad of shuffle mask elements point to elements in
4374 /// the right sequence. e.g.
4375 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4377 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4379 TargetLowering &TLI, DebugLoc dl) {
4380 EVT VT = SVOp->getValueType(0);
4381 SDValue V1 = SVOp->getOperand(0);
4382 SDValue V2 = SVOp->getOperand(1);
4383 unsigned NumElems = VT.getVectorNumElements();
4384 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4385 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4386 EVT MaskEltVT = MaskVT.getVectorElementType();
4388 switch (VT.getSimpleVT().SimpleTy) {
4389 default: assert(false && "Unexpected!");
4390 case MVT::v4f32: NewVT = MVT::v2f64; break;
4391 case MVT::v4i32: NewVT = MVT::v2i64; break;
4392 case MVT::v8i16: NewVT = MVT::v4i32; break;
4393 case MVT::v16i8: NewVT = MVT::v4i32; break;
4396 if (NewWidth == 2) {
4402 int Scale = NumElems / NewWidth;
4403 SmallVector<int, 8> MaskVec;
4404 for (unsigned i = 0; i < NumElems; i += Scale) {
4406 for (int j = 0; j < Scale; ++j) {
4407 int EltIdx = SVOp->getMaskElt(i+j);
4411 StartIdx = EltIdx - (EltIdx % Scale);
4412 if (EltIdx != StartIdx + j)
4416 MaskVec.push_back(-1);
4418 MaskVec.push_back(StartIdx / Scale);
4421 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4422 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4423 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4426 /// getVZextMovL - Return a zero-extending vector move low node.
4428 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4429 SDValue SrcOp, SelectionDAG &DAG,
4430 const X86Subtarget *Subtarget, DebugLoc dl) {
4431 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4432 LoadSDNode *LD = NULL;
4433 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4434 LD = dyn_cast<LoadSDNode>(SrcOp);
4436 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4438 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4439 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4440 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4441 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4442 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4444 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4445 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4446 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4447 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4455 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4456 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4457 DAG.getNode(ISD::BIT_CONVERT, dl,
4461 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4464 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4465 SDValue V1 = SVOp->getOperand(0);
4466 SDValue V2 = SVOp->getOperand(1);
4467 DebugLoc dl = SVOp->getDebugLoc();
4468 EVT VT = SVOp->getValueType(0);
4470 SmallVector<std::pair<int, int>, 8> Locs;
4472 SmallVector<int, 8> Mask1(4U, -1);
4473 SmallVector<int, 8> PermMask;
4474 SVOp->getMask(PermMask);
4478 for (unsigned i = 0; i != 4; ++i) {
4479 int Idx = PermMask[i];
4481 Locs[i] = std::make_pair(-1, -1);
4483 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4485 Locs[i] = std::make_pair(0, NumLo);
4489 Locs[i] = std::make_pair(1, NumHi);
4491 Mask1[2+NumHi] = Idx;
4497 if (NumLo <= 2 && NumHi <= 2) {
4498 // If no more than two elements come from either vector. This can be
4499 // implemented with two shuffles. First shuffle gather the elements.
4500 // The second shuffle, which takes the first shuffle as both of its
4501 // vector operands, put the elements into the right order.
4502 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4504 SmallVector<int, 8> Mask2(4U, -1);
4506 for (unsigned i = 0; i != 4; ++i) {
4507 if (Locs[i].first == -1)
4510 unsigned Idx = (i < 2) ? 0 : 4;
4511 Idx += Locs[i].first * 2 + Locs[i].second;
4516 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4517 } else if (NumLo == 3 || NumHi == 3) {
4518 // Otherwise, we must have three elements from one vector, call it X, and
4519 // one element from the other, call it Y. First, use a shufps to build an
4520 // intermediate vector with the one element from Y and the element from X
4521 // that will be in the same half in the final destination (the indexes don't
4522 // matter). Then, use a shufps to build the final vector, taking the half
4523 // containing the element from Y from the intermediate, and the other half
4526 // Normalize it so the 3 elements come from V1.
4527 CommuteVectorShuffleMask(PermMask, VT);
4531 // Find the element from V2.
4533 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4534 int Val = PermMask[HiIndex];
4541 Mask1[0] = PermMask[HiIndex];
4543 Mask1[2] = PermMask[HiIndex^1];
4545 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4548 Mask1[0] = PermMask[0];
4549 Mask1[1] = PermMask[1];
4550 Mask1[2] = HiIndex & 1 ? 6 : 4;
4551 Mask1[3] = HiIndex & 1 ? 4 : 6;
4552 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4554 Mask1[0] = HiIndex & 1 ? 2 : 0;
4555 Mask1[1] = HiIndex & 1 ? 0 : 2;
4556 Mask1[2] = PermMask[2];
4557 Mask1[3] = PermMask[3];
4562 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4566 // Break it into (shuffle shuffle_hi, shuffle_lo).
4568 SmallVector<int,8> LoMask(4U, -1);
4569 SmallVector<int,8> HiMask(4U, -1);
4571 SmallVector<int,8> *MaskPtr = &LoMask;
4572 unsigned MaskIdx = 0;
4575 for (unsigned i = 0; i != 4; ++i) {
4582 int Idx = PermMask[i];
4584 Locs[i] = std::make_pair(-1, -1);
4585 } else if (Idx < 4) {
4586 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4587 (*MaskPtr)[LoIdx] = Idx;
4590 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4591 (*MaskPtr)[HiIdx] = Idx;
4596 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4597 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4598 SmallVector<int, 8> MaskOps;
4599 for (unsigned i = 0; i != 4; ++i) {
4600 if (Locs[i].first == -1) {
4601 MaskOps.push_back(-1);
4603 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4604 MaskOps.push_back(Idx);
4607 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4611 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4612 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4613 SDValue V1 = Op.getOperand(0);
4614 SDValue V2 = Op.getOperand(1);
4615 EVT VT = Op.getValueType();
4616 DebugLoc dl = Op.getDebugLoc();
4617 unsigned NumElems = VT.getVectorNumElements();
4618 bool isMMX = VT.getSizeInBits() == 64;
4619 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4620 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4621 bool V1IsSplat = false;
4622 bool V2IsSplat = false;
4624 if (isZeroShuffle(SVOp))
4625 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4627 // Promote splats to v4f32.
4628 if (SVOp->isSplat()) {
4629 if (isMMX || NumElems < 4)
4631 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4634 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4636 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4637 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4638 if (NewOp.getNode())
4639 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4640 LowerVECTOR_SHUFFLE(NewOp, DAG));
4641 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4642 // FIXME: Figure out a cleaner way to do this.
4643 // Try to make use of movq to zero out the top part.
4644 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4645 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4646 if (NewOp.getNode()) {
4647 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4648 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4649 DAG, Subtarget, dl);
4651 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4652 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4653 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4654 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4655 DAG, Subtarget, dl);
4659 if (X86::isPSHUFDMask(SVOp))
4662 // Check if this can be converted into a logical shift.
4663 bool isLeft = false;
4666 bool isShift = getSubtarget()->hasSSE2() &&
4667 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4668 if (isShift && ShVal.hasOneUse()) {
4669 // If the shifted value has multiple uses, it may be cheaper to use
4670 // v_set0 + movlhps or movhlps, etc.
4671 EVT EltVT = VT.getVectorElementType();
4672 ShAmt *= EltVT.getSizeInBits();
4673 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4676 if (X86::isMOVLMask(SVOp)) {
4679 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4680 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4685 // FIXME: fold these into legal mask.
4686 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4687 X86::isMOVSLDUPMask(SVOp) ||
4688 X86::isMOVHLPSMask(SVOp) ||
4689 X86::isMOVLHPSMask(SVOp) ||
4690 X86::isMOVLPMask(SVOp)))
4693 if (ShouldXformToMOVHLPS(SVOp) ||
4694 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4695 return CommuteVectorShuffle(SVOp, DAG);
4698 // No better options. Use a vshl / vsrl.
4699 EVT EltVT = VT.getVectorElementType();
4700 ShAmt *= EltVT.getSizeInBits();
4701 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4704 bool Commuted = false;
4705 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4706 // 1,1,1,1 -> v8i16 though.
4707 V1IsSplat = isSplatVector(V1.getNode());
4708 V2IsSplat = isSplatVector(V2.getNode());
4710 // Canonicalize the splat or undef, if present, to be on the RHS.
4711 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4712 Op = CommuteVectorShuffle(SVOp, DAG);
4713 SVOp = cast<ShuffleVectorSDNode>(Op);
4714 V1 = SVOp->getOperand(0);
4715 V2 = SVOp->getOperand(1);
4716 std::swap(V1IsSplat, V2IsSplat);
4717 std::swap(V1IsUndef, V2IsUndef);
4721 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4722 // Shuffling low element of v1 into undef, just return v1.
4725 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4726 // the instruction selector will not match, so get a canonical MOVL with
4727 // swapped operands to undo the commute.
4728 return getMOVL(DAG, dl, VT, V2, V1);
4731 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4732 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4733 X86::isUNPCKLMask(SVOp) ||
4734 X86::isUNPCKHMask(SVOp))
4738 // Normalize mask so all entries that point to V2 points to its first
4739 // element then try to match unpck{h|l} again. If match, return a
4740 // new vector_shuffle with the corrected mask.
4741 SDValue NewMask = NormalizeMask(SVOp, DAG);
4742 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4743 if (NSVOp != SVOp) {
4744 if (X86::isUNPCKLMask(NSVOp, true)) {
4746 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4753 // Commute is back and try unpck* again.
4754 // FIXME: this seems wrong.
4755 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4756 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4757 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4758 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4759 X86::isUNPCKLMask(NewSVOp) ||
4760 X86::isUNPCKHMask(NewSVOp))
4764 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4766 // Normalize the node to match x86 shuffle ops if needed
4767 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4768 return CommuteVectorShuffle(SVOp, DAG);
4770 // Check for legal shuffle and return?
4771 SmallVector<int, 16> PermMask;
4772 SVOp->getMask(PermMask);
4773 if (isShuffleMaskLegal(PermMask, VT))
4776 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4777 if (VT == MVT::v8i16) {
4778 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4779 if (NewOp.getNode())
4783 if (VT == MVT::v16i8) {
4784 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4785 if (NewOp.getNode())
4789 // Handle all 4 wide cases with a number of shuffles except for MMX.
4790 if (NumElems == 4 && !isMMX)
4791 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4797 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4798 SelectionDAG &DAG) {
4799 EVT VT = Op.getValueType();
4800 DebugLoc dl = Op.getDebugLoc();
4801 if (VT.getSizeInBits() == 8) {
4802 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4803 Op.getOperand(0), Op.getOperand(1));
4804 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4805 DAG.getValueType(VT));
4806 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4807 } else if (VT.getSizeInBits() == 16) {
4808 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4809 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4811 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4812 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4813 DAG.getNode(ISD::BIT_CONVERT, dl,
4817 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4818 Op.getOperand(0), Op.getOperand(1));
4819 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4820 DAG.getValueType(VT));
4821 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4822 } else if (VT == MVT::f32) {
4823 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4824 // the result back to FR32 register. It's only worth matching if the
4825 // result has a single use which is a store or a bitcast to i32. And in
4826 // the case of a store, it's not worth it if the index is a constant 0,
4827 // because a MOVSSmr can be used instead, which is smaller and faster.
4828 if (!Op.hasOneUse())
4830 SDNode *User = *Op.getNode()->use_begin();
4831 if ((User->getOpcode() != ISD::STORE ||
4832 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4833 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4834 (User->getOpcode() != ISD::BIT_CONVERT ||
4835 User->getValueType(0) != MVT::i32))
4837 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4838 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4841 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4842 } else if (VT == MVT::i32) {
4843 // ExtractPS works with constant index.
4844 if (isa<ConstantSDNode>(Op.getOperand(1)))
4852 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4853 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4856 if (Subtarget->hasSSE41()) {
4857 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4862 EVT VT = Op.getValueType();
4863 DebugLoc dl = Op.getDebugLoc();
4864 // TODO: handle v16i8.
4865 if (VT.getSizeInBits() == 16) {
4866 SDValue Vec = Op.getOperand(0);
4867 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4869 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4870 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4871 DAG.getNode(ISD::BIT_CONVERT, dl,
4874 // Transform it so it match pextrw which produces a 32-bit result.
4875 EVT EltVT = MVT::i32;
4876 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4877 Op.getOperand(0), Op.getOperand(1));
4878 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4879 DAG.getValueType(VT));
4880 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4881 } else if (VT.getSizeInBits() == 32) {
4882 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4886 // SHUFPS the element to the lowest double word, then movss.
4887 int Mask[4] = { Idx, -1, -1, -1 };
4888 EVT VVT = Op.getOperand(0).getValueType();
4889 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4890 DAG.getUNDEF(VVT), Mask);
4891 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4892 DAG.getIntPtrConstant(0));
4893 } else if (VT.getSizeInBits() == 64) {
4894 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4895 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4896 // to match extract_elt for f64.
4897 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4901 // UNPCKHPD the element to the lowest double word, then movsd.
4902 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4903 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4904 int Mask[2] = { 1, -1 };
4905 EVT VVT = Op.getOperand(0).getValueType();
4906 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4907 DAG.getUNDEF(VVT), Mask);
4908 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4909 DAG.getIntPtrConstant(0));
4916 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4917 EVT VT = Op.getValueType();
4918 EVT EltVT = VT.getVectorElementType();
4919 DebugLoc dl = Op.getDebugLoc();
4921 SDValue N0 = Op.getOperand(0);
4922 SDValue N1 = Op.getOperand(1);
4923 SDValue N2 = Op.getOperand(2);
4925 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4926 isa<ConstantSDNode>(N2)) {
4928 if (VT == MVT::v8i16)
4929 Opc = X86ISD::PINSRW;
4930 else if (VT == MVT::v4i16)
4931 Opc = X86ISD::MMX_PINSRW;
4932 else if (VT == MVT::v16i8)
4933 Opc = X86ISD::PINSRB;
4935 Opc = X86ISD::PINSRB;
4937 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4939 if (N1.getValueType() != MVT::i32)
4940 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4941 if (N2.getValueType() != MVT::i32)
4942 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4943 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4944 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4945 // Bits [7:6] of the constant are the source select. This will always be
4946 // zero here. The DAG Combiner may combine an extract_elt index into these
4947 // bits. For example (insert (extract, 3), 2) could be matched by putting
4948 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4949 // Bits [5:4] of the constant are the destination select. This is the
4950 // value of the incoming immediate.
4951 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4952 // combine either bitwise AND or insert of float 0.0 to set these bits.
4953 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4954 // Create this as a scalar to vector..
4955 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4956 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4957 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4958 // PINSR* works with constant index.
4965 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4966 EVT VT = Op.getValueType();
4967 EVT EltVT = VT.getVectorElementType();
4969 if (Subtarget->hasSSE41())
4970 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4972 if (EltVT == MVT::i8)
4975 DebugLoc dl = Op.getDebugLoc();
4976 SDValue N0 = Op.getOperand(0);
4977 SDValue N1 = Op.getOperand(1);
4978 SDValue N2 = Op.getOperand(2);
4980 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4981 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4982 // as its second argument.
4983 if (N1.getValueType() != MVT::i32)
4984 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4985 if (N2.getValueType() != MVT::i32)
4986 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4987 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
4988 dl, VT, N0, N1, N2);
4994 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4995 DebugLoc dl = Op.getDebugLoc();
4996 if (Op.getValueType() == MVT::v2f32)
4997 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4998 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4999 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
5000 Op.getOperand(0))));
5002 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
5003 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5005 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5006 EVT VT = MVT::v2i32;
5007 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5014 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5015 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5018 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5019 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5020 // one of the above mentioned nodes. It has to be wrapped because otherwise
5021 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5022 // be used to form addressing mode. These wrapped nodes will be selected
5025 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
5026 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5028 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5030 unsigned char OpFlag = 0;
5031 unsigned WrapperKind = X86ISD::Wrapper;
5032 CodeModel::Model M = getTargetMachine().getCodeModel();
5034 if (Subtarget->isPICStyleRIPRel() &&
5035 (M == CodeModel::Small || M == CodeModel::Kernel))
5036 WrapperKind = X86ISD::WrapperRIP;
5037 else if (Subtarget->isPICStyleGOT())
5038 OpFlag = X86II::MO_GOTOFF;
5039 else if (Subtarget->isPICStyleStubPIC())
5040 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5042 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5044 CP->getOffset(), OpFlag);
5045 DebugLoc DL = CP->getDebugLoc();
5046 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5047 // With PIC, the address is actually $g + Offset.
5049 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5050 DAG.getNode(X86ISD::GlobalBaseReg,
5051 DebugLoc::getUnknownLoc(), getPointerTy()),
5058 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
5059 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5061 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5063 unsigned char OpFlag = 0;
5064 unsigned WrapperKind = X86ISD::Wrapper;
5065 CodeModel::Model M = getTargetMachine().getCodeModel();
5067 if (Subtarget->isPICStyleRIPRel() &&
5068 (M == CodeModel::Small || M == CodeModel::Kernel))
5069 WrapperKind = X86ISD::WrapperRIP;
5070 else if (Subtarget->isPICStyleGOT())
5071 OpFlag = X86II::MO_GOTOFF;
5072 else if (Subtarget->isPICStyleStubPIC())
5073 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5075 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5077 DebugLoc DL = JT->getDebugLoc();
5078 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5080 // With PIC, the address is actually $g + Offset.
5082 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5083 DAG.getNode(X86ISD::GlobalBaseReg,
5084 DebugLoc::getUnknownLoc(), getPointerTy()),
5092 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
5093 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5095 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5097 unsigned char OpFlag = 0;
5098 unsigned WrapperKind = X86ISD::Wrapper;
5099 CodeModel::Model M = getTargetMachine().getCodeModel();
5101 if (Subtarget->isPICStyleRIPRel() &&
5102 (M == CodeModel::Small || M == CodeModel::Kernel))
5103 WrapperKind = X86ISD::WrapperRIP;
5104 else if (Subtarget->isPICStyleGOT())
5105 OpFlag = X86II::MO_GOTOFF;
5106 else if (Subtarget->isPICStyleStubPIC())
5107 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5109 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5111 DebugLoc DL = Op.getDebugLoc();
5112 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5115 // With PIC, the address is actually $g + Offset.
5116 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5117 !Subtarget->is64Bit()) {
5118 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5119 DAG.getNode(X86ISD::GlobalBaseReg,
5120 DebugLoc::getUnknownLoc(),
5129 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5130 // Create the TargetBlockAddressAddress node.
5131 unsigned char OpFlags =
5132 Subtarget->ClassifyBlockAddressReference();
5133 CodeModel::Model M = getTargetMachine().getCodeModel();
5134 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5135 DebugLoc dl = Op.getDebugLoc();
5136 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5137 /*isTarget=*/true, OpFlags);
5139 if (Subtarget->isPICStyleRIPRel() &&
5140 (M == CodeModel::Small || M == CodeModel::Kernel))
5141 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5143 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5145 // With PIC, the address is actually $g + Offset.
5146 if (isGlobalRelativeToPICBase(OpFlags)) {
5147 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5148 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5156 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5158 SelectionDAG &DAG) const {
5159 // Create the TargetGlobalAddress node, folding in the constant
5160 // offset if it is legal.
5161 unsigned char OpFlags =
5162 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5163 CodeModel::Model M = getTargetMachine().getCodeModel();
5165 if (OpFlags == X86II::MO_NO_FLAG &&
5166 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5167 // A direct static reference to a global.
5168 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5171 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5174 if (Subtarget->isPICStyleRIPRel() &&
5175 (M == CodeModel::Small || M == CodeModel::Kernel))
5176 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5178 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5180 // With PIC, the address is actually $g + Offset.
5181 if (isGlobalRelativeToPICBase(OpFlags)) {
5182 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5183 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5187 // For globals that require a load from a stub to get the address, emit the
5189 if (isGlobalStubReference(OpFlags))
5190 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5191 PseudoSourceValue::getGOT(), 0, false, false, 0);
5193 // If there was a non-zero offset that we didn't fold, create an explicit
5196 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5197 DAG.getConstant(Offset, getPointerTy()));
5203 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5204 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5205 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5206 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5210 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5211 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5212 unsigned char OperandFlags) {
5213 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5214 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5215 DebugLoc dl = GA->getDebugLoc();
5216 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5217 GA->getValueType(0),
5221 SDValue Ops[] = { Chain, TGA, *InFlag };
5222 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5224 SDValue Ops[] = { Chain, TGA };
5225 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5228 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5229 MFI->setHasCalls(true);
5231 SDValue Flag = Chain.getValue(1);
5232 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5235 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5237 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5240 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5241 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5242 DAG.getNode(X86ISD::GlobalBaseReg,
5243 DebugLoc::getUnknownLoc(),
5245 InFlag = Chain.getValue(1);
5247 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5250 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5252 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5254 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5255 X86::RAX, X86II::MO_TLSGD);
5258 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5259 // "local exec" model.
5260 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5261 const EVT PtrVT, TLSModel::Model model,
5263 DebugLoc dl = GA->getDebugLoc();
5264 // Get the Thread Pointer
5265 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5266 DebugLoc::getUnknownLoc(), PtrVT,
5267 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5270 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5271 NULL, 0, false, false, 0);
5273 unsigned char OperandFlags = 0;
5274 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5276 unsigned WrapperKind = X86ISD::Wrapper;
5277 if (model == TLSModel::LocalExec) {
5278 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5279 } else if (is64Bit) {
5280 assert(model == TLSModel::InitialExec);
5281 OperandFlags = X86II::MO_GOTTPOFF;
5282 WrapperKind = X86ISD::WrapperRIP;
5284 assert(model == TLSModel::InitialExec);
5285 OperandFlags = X86II::MO_INDNTPOFF;
5288 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5290 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5291 GA->getOffset(), OperandFlags);
5292 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5294 if (model == TLSModel::InitialExec)
5295 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5296 PseudoSourceValue::getGOT(), 0, false, false, 0);
5298 // The address of the thread local variable is the add of the thread
5299 // pointer with the offset of the variable.
5300 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5304 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5305 // TODO: implement the "local dynamic" model
5306 // TODO: implement the "initial exec"model for pic executables
5307 assert(Subtarget->isTargetELF() &&
5308 "TLS not implemented for non-ELF targets");
5309 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5310 const GlobalValue *GV = GA->getGlobal();
5312 // If GV is an alias then use the aliasee for determining
5313 // thread-localness.
5314 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5315 GV = GA->resolveAliasedGlobal(false);
5317 TLSModel::Model model = getTLSModel(GV,
5318 getTargetMachine().getRelocationModel());
5321 case TLSModel::GeneralDynamic:
5322 case TLSModel::LocalDynamic: // not implemented
5323 if (Subtarget->is64Bit())
5324 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5325 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5327 case TLSModel::InitialExec:
5328 case TLSModel::LocalExec:
5329 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5330 Subtarget->is64Bit());
5333 llvm_unreachable("Unreachable");
5338 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5339 /// take a 2 x i32 value to shift plus a shift amount.
5340 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5341 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5342 EVT VT = Op.getValueType();
5343 unsigned VTBits = VT.getSizeInBits();
5344 DebugLoc dl = Op.getDebugLoc();
5345 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5346 SDValue ShOpLo = Op.getOperand(0);
5347 SDValue ShOpHi = Op.getOperand(1);
5348 SDValue ShAmt = Op.getOperand(2);
5349 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5350 DAG.getConstant(VTBits - 1, MVT::i8))
5351 : DAG.getConstant(0, VT);
5354 if (Op.getOpcode() == ISD::SHL_PARTS) {
5355 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5356 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5358 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5359 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5362 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5363 DAG.getConstant(VTBits, MVT::i8));
5364 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5365 AndNode, DAG.getConstant(0, MVT::i8));
5368 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5369 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5370 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5372 if (Op.getOpcode() == ISD::SHL_PARTS) {
5373 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5374 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5376 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5377 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5380 SDValue Ops[2] = { Lo, Hi };
5381 return DAG.getMergeValues(Ops, 2, dl);
5384 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5385 EVT SrcVT = Op.getOperand(0).getValueType();
5387 if (SrcVT.isVector()) {
5388 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5394 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5395 "Unknown SINT_TO_FP to lower!");
5397 // These are really Legal; return the operand so the caller accepts it as
5399 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5401 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5402 Subtarget->is64Bit()) {
5406 DebugLoc dl = Op.getDebugLoc();
5407 unsigned Size = SrcVT.getSizeInBits()/8;
5408 MachineFunction &MF = DAG.getMachineFunction();
5409 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5410 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5411 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5413 PseudoSourceValue::getFixedStack(SSFI), 0,
5415 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5418 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5420 SelectionDAG &DAG) {
5422 DebugLoc dl = Op.getDebugLoc();
5424 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5426 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5428 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5429 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5430 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5431 Tys, Ops, array_lengthof(Ops));
5434 Chain = Result.getValue(1);
5435 SDValue InFlag = Result.getValue(2);
5437 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5438 // shouldn't be necessary except that RFP cannot be live across
5439 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5440 MachineFunction &MF = DAG.getMachineFunction();
5441 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5442 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5443 Tys = DAG.getVTList(MVT::Other);
5445 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5447 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5448 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5449 PseudoSourceValue::getFixedStack(SSFI), 0,
5456 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5457 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5458 // This algorithm is not obvious. Here it is in C code, more or less:
5460 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5461 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5462 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5464 // Copy ints to xmm registers.
5465 __m128i xh = _mm_cvtsi32_si128( hi );
5466 __m128i xl = _mm_cvtsi32_si128( lo );
5468 // Combine into low half of a single xmm register.
5469 __m128i x = _mm_unpacklo_epi32( xh, xl );
5473 // Merge in appropriate exponents to give the integer bits the right
5475 x = _mm_unpacklo_epi32( x, exp );
5477 // Subtract away the biases to deal with the IEEE-754 double precision
5479 d = _mm_sub_pd( (__m128d) x, bias );
5481 // All conversions up to here are exact. The correctly rounded result is
5482 // calculated using the current rounding mode using the following
5484 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5485 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5486 // store doesn't really need to be here (except
5487 // maybe to zero the other double)
5492 DebugLoc dl = Op.getDebugLoc();
5493 LLVMContext *Context = DAG.getContext();
5495 // Build some magic constants.
5496 std::vector<Constant*> CV0;
5497 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5498 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5499 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5500 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5501 Constant *C0 = ConstantVector::get(CV0);
5502 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5504 std::vector<Constant*> CV1;
5506 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5508 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5509 Constant *C1 = ConstantVector::get(CV1);
5510 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5512 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5513 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5515 DAG.getIntPtrConstant(1)));
5516 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5517 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5519 DAG.getIntPtrConstant(0)));
5520 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5521 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5522 PseudoSourceValue::getConstantPool(), 0,
5524 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5525 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5526 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5527 PseudoSourceValue::getConstantPool(), 0,
5529 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5531 // Add the halves; easiest way is to swap them into another reg first.
5532 int ShufMask[2] = { 1, -1 };
5533 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5534 DAG.getUNDEF(MVT::v2f64), ShufMask);
5535 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5536 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5537 DAG.getIntPtrConstant(0));
5540 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5541 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5542 DebugLoc dl = Op.getDebugLoc();
5543 // FP constant to bias correct the final result.
5544 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5547 // Load the 32-bit value into an XMM register.
5548 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5549 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5551 DAG.getIntPtrConstant(0)));
5553 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5554 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5555 DAG.getIntPtrConstant(0));
5557 // Or the load with the bias.
5558 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5559 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5560 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5562 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5563 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5564 MVT::v2f64, Bias)));
5565 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5566 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5567 DAG.getIntPtrConstant(0));
5569 // Subtract the bias.
5570 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5572 // Handle final rounding.
5573 EVT DestVT = Op.getValueType();
5575 if (DestVT.bitsLT(MVT::f64)) {
5576 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5577 DAG.getIntPtrConstant(0));
5578 } else if (DestVT.bitsGT(MVT::f64)) {
5579 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5582 // Handle final rounding.
5586 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5587 SDValue N0 = Op.getOperand(0);
5588 DebugLoc dl = Op.getDebugLoc();
5590 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5591 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5592 // the optimization here.
5593 if (DAG.SignBitIsZero(N0))
5594 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5596 EVT SrcVT = N0.getValueType();
5597 if (SrcVT == MVT::i64) {
5598 // We only handle SSE2 f64 target here; caller can expand the rest.
5599 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5602 return LowerUINT_TO_FP_i64(Op, DAG);
5603 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5604 return LowerUINT_TO_FP_i32(Op, DAG);
5607 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5609 // Make a 64-bit buffer, and use it to build an FILD.
5610 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5611 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5612 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5613 getPointerTy(), StackSlot, WordOff);
5614 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5615 StackSlot, NULL, 0, false, false, 0);
5616 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5617 OffsetSlot, NULL, 0, false, false, 0);
5618 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5621 std::pair<SDValue,SDValue> X86TargetLowering::
5622 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5623 DebugLoc dl = Op.getDebugLoc();
5625 EVT DstTy = Op.getValueType();
5628 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5632 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5633 DstTy.getSimpleVT() >= MVT::i16 &&
5634 "Unknown FP_TO_SINT to lower!");
5636 // These are really Legal.
5637 if (DstTy == MVT::i32 &&
5638 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5639 return std::make_pair(SDValue(), SDValue());
5640 if (Subtarget->is64Bit() &&
5641 DstTy == MVT::i64 &&
5642 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5643 return std::make_pair(SDValue(), SDValue());
5645 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5647 MachineFunction &MF = DAG.getMachineFunction();
5648 unsigned MemSize = DstTy.getSizeInBits()/8;
5649 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5650 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5653 switch (DstTy.getSimpleVT().SimpleTy) {
5654 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5655 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5656 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5657 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5660 SDValue Chain = DAG.getEntryNode();
5661 SDValue Value = Op.getOperand(0);
5662 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5663 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5664 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5665 PseudoSourceValue::getFixedStack(SSFI), 0,
5667 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5669 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5671 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5672 Chain = Value.getValue(1);
5673 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5674 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5677 // Build the FP_TO_INT*_IN_MEM
5678 SDValue Ops[] = { Chain, Value, StackSlot };
5679 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5681 return std::make_pair(FIST, StackSlot);
5684 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5685 if (Op.getValueType().isVector()) {
5686 if (Op.getValueType() == MVT::v2i32 &&
5687 Op.getOperand(0).getValueType() == MVT::v2f64) {
5693 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5694 SDValue FIST = Vals.first, StackSlot = Vals.second;
5695 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5696 if (FIST.getNode() == 0) return Op;
5699 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5700 FIST, StackSlot, NULL, 0, false, false, 0);
5703 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5704 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5705 SDValue FIST = Vals.first, StackSlot = Vals.second;
5706 assert(FIST.getNode() && "Unexpected failure");
5709 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5710 FIST, StackSlot, NULL, 0, false, false, 0);
5713 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5714 LLVMContext *Context = DAG.getContext();
5715 DebugLoc dl = Op.getDebugLoc();
5716 EVT VT = Op.getValueType();
5719 EltVT = VT.getVectorElementType();
5720 std::vector<Constant*> CV;
5721 if (EltVT == MVT::f64) {
5722 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5726 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5732 Constant *C = ConstantVector::get(CV);
5733 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5734 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5735 PseudoSourceValue::getConstantPool(), 0,
5737 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5740 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5741 LLVMContext *Context = DAG.getContext();
5742 DebugLoc dl = Op.getDebugLoc();
5743 EVT VT = Op.getValueType();
5746 EltVT = VT.getVectorElementType();
5747 std::vector<Constant*> CV;
5748 if (EltVT == MVT::f64) {
5749 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5753 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5759 Constant *C = ConstantVector::get(CV);
5760 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5761 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5762 PseudoSourceValue::getConstantPool(), 0,
5764 if (VT.isVector()) {
5765 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5766 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5767 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5769 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5771 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5775 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5776 LLVMContext *Context = DAG.getContext();
5777 SDValue Op0 = Op.getOperand(0);
5778 SDValue Op1 = Op.getOperand(1);
5779 DebugLoc dl = Op.getDebugLoc();
5780 EVT VT = Op.getValueType();
5781 EVT SrcVT = Op1.getValueType();
5783 // If second operand is smaller, extend it first.
5784 if (SrcVT.bitsLT(VT)) {
5785 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5788 // And if it is bigger, shrink it first.
5789 if (SrcVT.bitsGT(VT)) {
5790 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5794 // At this point the operands and the result should have the same
5795 // type, and that won't be f80 since that is not custom lowered.
5797 // First get the sign bit of second operand.
5798 std::vector<Constant*> CV;
5799 if (SrcVT == MVT::f64) {
5800 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5801 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5803 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5804 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5805 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5806 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5808 Constant *C = ConstantVector::get(CV);
5809 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5810 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5811 PseudoSourceValue::getConstantPool(), 0,
5813 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5815 // Shift sign bit right or left if the two operands have different types.
5816 if (SrcVT.bitsGT(VT)) {
5817 // Op0 is MVT::f32, Op1 is MVT::f64.
5818 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5819 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5820 DAG.getConstant(32, MVT::i32));
5821 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5822 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5823 DAG.getIntPtrConstant(0));
5826 // Clear first operand sign bit.
5828 if (VT == MVT::f64) {
5829 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5830 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5832 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5833 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5834 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5835 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5837 C = ConstantVector::get(CV);
5838 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5839 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5840 PseudoSourceValue::getConstantPool(), 0,
5842 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5844 // Or the value with the sign bit.
5845 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5848 /// Emit nodes that will be selected as "test Op0,Op0", or something
5850 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5851 SelectionDAG &DAG) {
5852 DebugLoc dl = Op.getDebugLoc();
5854 // CF and OF aren't always set the way we want. Determine which
5855 // of these we need.
5856 bool NeedCF = false;
5857 bool NeedOF = false;
5859 case X86::COND_A: case X86::COND_AE:
5860 case X86::COND_B: case X86::COND_BE:
5863 case X86::COND_G: case X86::COND_GE:
5864 case X86::COND_L: case X86::COND_LE:
5865 case X86::COND_O: case X86::COND_NO:
5871 // See if we can use the EFLAGS value from the operand instead of
5872 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5873 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5874 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5875 unsigned Opcode = 0;
5876 unsigned NumOperands = 0;
5877 switch (Op.getNode()->getOpcode()) {
5879 // Due to an isel shortcoming, be conservative if this add is likely to
5880 // be selected as part of a load-modify-store instruction. When the root
5881 // node in a match is a store, isel doesn't know how to remap non-chain
5882 // non-flag uses of other nodes in the match, such as the ADD in this
5883 // case. This leads to the ADD being left around and reselected, with
5884 // the result being two adds in the output.
5885 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5886 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5887 if (UI->getOpcode() == ISD::STORE)
5889 if (ConstantSDNode *C =
5890 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5891 // An add of one will be selected as an INC.
5892 if (C->getAPIntValue() == 1) {
5893 Opcode = X86ISD::INC;
5897 // An add of negative one (subtract of one) will be selected as a DEC.
5898 if (C->getAPIntValue().isAllOnesValue()) {
5899 Opcode = X86ISD::DEC;
5904 // Otherwise use a regular EFLAGS-setting add.
5905 Opcode = X86ISD::ADD;
5909 // If the primary and result isn't used, don't bother using X86ISD::AND,
5910 // because a TEST instruction will be better.
5911 bool NonFlagUse = false;
5912 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5913 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5915 unsigned UOpNo = UI.getOperandNo();
5916 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5917 // Look pass truncate.
5918 UOpNo = User->use_begin().getOperandNo();
5919 User = *User->use_begin();
5921 if (User->getOpcode() != ISD::BRCOND &&
5922 User->getOpcode() != ISD::SETCC &&
5923 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5935 // Due to the ISEL shortcoming noted above, be conservative if this op is
5936 // likely to be selected as part of a load-modify-store instruction.
5937 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5938 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5939 if (UI->getOpcode() == ISD::STORE)
5941 // Otherwise use a regular EFLAGS-setting instruction.
5942 switch (Op.getNode()->getOpcode()) {
5943 case ISD::SUB: Opcode = X86ISD::SUB; break;
5944 case ISD::OR: Opcode = X86ISD::OR; break;
5945 case ISD::XOR: Opcode = X86ISD::XOR; break;
5946 case ISD::AND: Opcode = X86ISD::AND; break;
5947 default: llvm_unreachable("unexpected operator!");
5958 return SDValue(Op.getNode(), 1);
5964 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5965 SmallVector<SDValue, 4> Ops;
5966 for (unsigned i = 0; i != NumOperands; ++i)
5967 Ops.push_back(Op.getOperand(i));
5968 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5969 DAG.ReplaceAllUsesWith(Op, New);
5970 return SDValue(New.getNode(), 1);
5974 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5975 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5976 DAG.getConstant(0, Op.getValueType()));
5979 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5981 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5982 SelectionDAG &DAG) {
5983 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5984 if (C->getAPIntValue() == 0)
5985 return EmitTest(Op0, X86CC, DAG);
5987 DebugLoc dl = Op0.getDebugLoc();
5988 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5991 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5992 /// if it's possible.
5993 static SDValue LowerToBT(SDValue And, ISD::CondCode CC,
5994 DebugLoc dl, SelectionDAG &DAG) {
5995 SDValue Op0 = And.getOperand(0);
5996 SDValue Op1 = And.getOperand(1);
5997 if (Op0.getOpcode() == ISD::TRUNCATE)
5998 Op0 = Op0.getOperand(0);
5999 if (Op1.getOpcode() == ISD::TRUNCATE)
6000 Op1 = Op1.getOperand(0);
6003 if (Op1.getOpcode() == ISD::SHL) {
6004 if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
6005 if (And10C->getZExtValue() == 1) {
6007 RHS = Op1.getOperand(1);
6009 } else if (Op0.getOpcode() == ISD::SHL) {
6010 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6011 if (And00C->getZExtValue() == 1) {
6013 RHS = Op0.getOperand(1);
6015 } else if (Op1.getOpcode() == ISD::Constant) {
6016 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6017 SDValue AndLHS = Op0;
6018 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6019 LHS = AndLHS.getOperand(0);
6020 RHS = AndLHS.getOperand(1);
6024 if (LHS.getNode()) {
6025 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
6026 // instruction. Since the shift amount is in-range-or-undefined, we know
6027 // that doing a bittest on the i16 value is ok. We extend to i32 because
6028 // the encoding for the i16 version is larger than the i32 version.
6029 if (LHS.getValueType() == MVT::i8)
6030 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6032 // If the operand types disagree, extend the shift amount to match. Since
6033 // BT ignores high bits (like shifts) we can use anyextend.
6034 if (LHS.getValueType() != RHS.getValueType())
6035 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6037 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6038 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6039 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6040 DAG.getConstant(Cond, MVT::i8), BT);
6046 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
6047 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6048 SDValue Op0 = Op.getOperand(0);
6049 SDValue Op1 = Op.getOperand(1);
6050 DebugLoc dl = Op.getDebugLoc();
6051 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6053 // Optimize to BT if possible.
6054 // Lower (X & (1 << N)) == 0 to BT(X, N).
6055 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6056 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6057 if (Op0.getOpcode() == ISD::AND &&
6059 Op1.getOpcode() == ISD::Constant &&
6060 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
6061 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6062 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6063 if (NewSetCC.getNode())
6067 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6068 if (Op0.getOpcode() == X86ISD::SETCC &&
6069 Op1.getOpcode() == ISD::Constant &&
6070 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6071 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6072 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6073 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6074 bool Invert = (CC == ISD::SETNE) ^
6075 cast<ConstantSDNode>(Op1)->isNullValue();
6077 CCode = X86::GetOppositeBranchCondition(CCode);
6078 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6079 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6082 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6083 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6084 if (X86CC == X86::COND_INVALID)
6087 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6089 // Use sbb x, x to materialize carry bit into a GPR.
6090 if (X86CC == X86::COND_B)
6091 return DAG.getNode(ISD::AND, dl, MVT::i8,
6092 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6093 DAG.getConstant(X86CC, MVT::i8), Cond),
6094 DAG.getConstant(1, MVT::i8));
6096 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6097 DAG.getConstant(X86CC, MVT::i8), Cond);
6100 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
6102 SDValue Op0 = Op.getOperand(0);
6103 SDValue Op1 = Op.getOperand(1);
6104 SDValue CC = Op.getOperand(2);
6105 EVT VT = Op.getValueType();
6106 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6107 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6108 DebugLoc dl = Op.getDebugLoc();
6112 EVT VT0 = Op0.getValueType();
6113 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6114 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6117 switch (SetCCOpcode) {
6120 case ISD::SETEQ: SSECC = 0; break;
6122 case ISD::SETGT: Swap = true; // Fallthrough
6124 case ISD::SETOLT: SSECC = 1; break;
6126 case ISD::SETGE: Swap = true; // Fallthrough
6128 case ISD::SETOLE: SSECC = 2; break;
6129 case ISD::SETUO: SSECC = 3; break;
6131 case ISD::SETNE: SSECC = 4; break;
6132 case ISD::SETULE: Swap = true;
6133 case ISD::SETUGE: SSECC = 5; break;
6134 case ISD::SETULT: Swap = true;
6135 case ISD::SETUGT: SSECC = 6; break;
6136 case ISD::SETO: SSECC = 7; break;
6139 std::swap(Op0, Op1);
6141 // In the two special cases we can't handle, emit two comparisons.
6143 if (SetCCOpcode == ISD::SETUEQ) {
6145 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6146 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6147 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6149 else if (SetCCOpcode == ISD::SETONE) {
6151 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6152 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6153 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6155 llvm_unreachable("Illegal FP comparison");
6157 // Handle all other FP comparisons here.
6158 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6161 // We are handling one of the integer comparisons here. Since SSE only has
6162 // GT and EQ comparisons for integer, swapping operands and multiple
6163 // operations may be required for some comparisons.
6164 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6165 bool Swap = false, Invert = false, FlipSigns = false;
6167 switch (VT.getSimpleVT().SimpleTy) {
6170 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6172 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6174 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6175 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6178 switch (SetCCOpcode) {
6180 case ISD::SETNE: Invert = true;
6181 case ISD::SETEQ: Opc = EQOpc; break;
6182 case ISD::SETLT: Swap = true;
6183 case ISD::SETGT: Opc = GTOpc; break;
6184 case ISD::SETGE: Swap = true;
6185 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6186 case ISD::SETULT: Swap = true;
6187 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6188 case ISD::SETUGE: Swap = true;
6189 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6192 std::swap(Op0, Op1);
6194 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6195 // bits of the inputs before performing those operations.
6197 EVT EltVT = VT.getVectorElementType();
6198 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6200 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6201 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6203 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6204 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6207 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6209 // If the logical-not of the result is required, perform that now.
6211 Result = DAG.getNOT(dl, Result, VT);
6216 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6217 static bool isX86LogicalCmp(SDValue Op) {
6218 unsigned Opc = Op.getNode()->getOpcode();
6219 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6221 if (Op.getResNo() == 1 &&
6222 (Opc == X86ISD::ADD ||
6223 Opc == X86ISD::SUB ||
6224 Opc == X86ISD::SMUL ||
6225 Opc == X86ISD::UMUL ||
6226 Opc == X86ISD::INC ||
6227 Opc == X86ISD::DEC ||
6228 Opc == X86ISD::OR ||
6229 Opc == X86ISD::XOR ||
6230 Opc == X86ISD::AND))
6236 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6237 bool addTest = true;
6238 SDValue Cond = Op.getOperand(0);
6239 DebugLoc dl = Op.getDebugLoc();
6242 if (Cond.getOpcode() == ISD::SETCC) {
6243 SDValue NewCond = LowerSETCC(Cond, DAG);
6244 if (NewCond.getNode())
6248 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6249 SDValue Op1 = Op.getOperand(1);
6250 SDValue Op2 = Op.getOperand(2);
6251 if (Cond.getOpcode() == X86ISD::SETCC &&
6252 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6253 SDValue Cmp = Cond.getOperand(1);
6254 if (Cmp.getOpcode() == X86ISD::CMP) {
6255 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6256 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6257 ConstantSDNode *RHSC =
6258 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6259 if (N1C && N1C->isAllOnesValue() &&
6260 N2C && N2C->isNullValue() &&
6261 RHSC && RHSC->isNullValue()) {
6262 SDValue CmpOp0 = Cmp.getOperand(0);
6263 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6264 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6265 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6266 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6271 // Look pass (and (setcc_carry (cmp ...)), 1).
6272 if (Cond.getOpcode() == ISD::AND &&
6273 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6274 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6275 if (C && C->getAPIntValue() == 1)
6276 Cond = Cond.getOperand(0);
6279 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6280 // setting operand in place of the X86ISD::SETCC.
6281 if (Cond.getOpcode() == X86ISD::SETCC ||
6282 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6283 CC = Cond.getOperand(0);
6285 SDValue Cmp = Cond.getOperand(1);
6286 unsigned Opc = Cmp.getOpcode();
6287 EVT VT = Op.getValueType();
6289 bool IllegalFPCMov = false;
6290 if (VT.isFloatingPoint() && !VT.isVector() &&
6291 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6292 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6294 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6295 Opc == X86ISD::BT) { // FIXME
6302 // Look pass the truncate.
6303 if (Cond.getOpcode() == ISD::TRUNCATE)
6304 Cond = Cond.getOperand(0);
6306 // We know the result of AND is compared against zero. Try to match
6308 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6309 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6310 if (NewSetCC.getNode()) {
6311 CC = NewSetCC.getOperand(0);
6312 Cond = NewSetCC.getOperand(1);
6319 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6320 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6323 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6324 // condition is true.
6325 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6326 SDValue Ops[] = { Op2, Op1, CC, Cond };
6327 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6330 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6331 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6332 // from the AND / OR.
6333 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6334 Opc = Op.getOpcode();
6335 if (Opc != ISD::OR && Opc != ISD::AND)
6337 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6338 Op.getOperand(0).hasOneUse() &&
6339 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6340 Op.getOperand(1).hasOneUse());
6343 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6344 // 1 and that the SETCC node has a single use.
6345 static bool isXor1OfSetCC(SDValue Op) {
6346 if (Op.getOpcode() != ISD::XOR)
6348 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6349 if (N1C && N1C->getAPIntValue() == 1) {
6350 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6351 Op.getOperand(0).hasOneUse();
6356 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6357 bool addTest = true;
6358 SDValue Chain = Op.getOperand(0);
6359 SDValue Cond = Op.getOperand(1);
6360 SDValue Dest = Op.getOperand(2);
6361 DebugLoc dl = Op.getDebugLoc();
6364 if (Cond.getOpcode() == ISD::SETCC) {
6365 SDValue NewCond = LowerSETCC(Cond, DAG);
6366 if (NewCond.getNode())
6370 // FIXME: LowerXALUO doesn't handle these!!
6371 else if (Cond.getOpcode() == X86ISD::ADD ||
6372 Cond.getOpcode() == X86ISD::SUB ||
6373 Cond.getOpcode() == X86ISD::SMUL ||
6374 Cond.getOpcode() == X86ISD::UMUL)
6375 Cond = LowerXALUO(Cond, DAG);
6378 // Look pass (and (setcc_carry (cmp ...)), 1).
6379 if (Cond.getOpcode() == ISD::AND &&
6380 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6381 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6382 if (C && C->getAPIntValue() == 1)
6383 Cond = Cond.getOperand(0);
6386 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6387 // setting operand in place of the X86ISD::SETCC.
6388 if (Cond.getOpcode() == X86ISD::SETCC ||
6389 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6390 CC = Cond.getOperand(0);
6392 SDValue Cmp = Cond.getOperand(1);
6393 unsigned Opc = Cmp.getOpcode();
6394 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6395 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6399 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6403 // These can only come from an arithmetic instruction with overflow,
6404 // e.g. SADDO, UADDO.
6405 Cond = Cond.getNode()->getOperand(1);
6412 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6413 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6414 if (CondOpc == ISD::OR) {
6415 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6416 // two branches instead of an explicit OR instruction with a
6418 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6419 isX86LogicalCmp(Cmp)) {
6420 CC = Cond.getOperand(0).getOperand(0);
6421 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6422 Chain, Dest, CC, Cmp);
6423 CC = Cond.getOperand(1).getOperand(0);
6427 } else { // ISD::AND
6428 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6429 // two branches instead of an explicit AND instruction with a
6430 // separate test. However, we only do this if this block doesn't
6431 // have a fall-through edge, because this requires an explicit
6432 // jmp when the condition is false.
6433 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6434 isX86LogicalCmp(Cmp) &&
6435 Op.getNode()->hasOneUse()) {
6436 X86::CondCode CCode =
6437 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6438 CCode = X86::GetOppositeBranchCondition(CCode);
6439 CC = DAG.getConstant(CCode, MVT::i8);
6440 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6441 // Look for an unconditional branch following this conditional branch.
6442 // We need this because we need to reverse the successors in order
6443 // to implement FCMP_OEQ.
6444 if (User.getOpcode() == ISD::BR) {
6445 SDValue FalseBB = User.getOperand(1);
6447 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6448 assert(NewBR == User);
6451 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6452 Chain, Dest, CC, Cmp);
6453 X86::CondCode CCode =
6454 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6455 CCode = X86::GetOppositeBranchCondition(CCode);
6456 CC = DAG.getConstant(CCode, MVT::i8);
6462 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6463 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6464 // It should be transformed during dag combiner except when the condition
6465 // is set by a arithmetics with overflow node.
6466 X86::CondCode CCode =
6467 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6468 CCode = X86::GetOppositeBranchCondition(CCode);
6469 CC = DAG.getConstant(CCode, MVT::i8);
6470 Cond = Cond.getOperand(0).getOperand(1);
6476 // Look pass the truncate.
6477 if (Cond.getOpcode() == ISD::TRUNCATE)
6478 Cond = Cond.getOperand(0);
6480 // We know the result of AND is compared against zero. Try to match
6482 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6483 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6484 if (NewSetCC.getNode()) {
6485 CC = NewSetCC.getOperand(0);
6486 Cond = NewSetCC.getOperand(1);
6493 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6494 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6496 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6497 Chain, Dest, CC, Cond);
6501 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6502 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6503 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6504 // that the guard pages used by the OS virtual memory manager are allocated in
6505 // correct sequence.
6507 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6508 SelectionDAG &DAG) {
6509 assert(Subtarget->isTargetCygMing() &&
6510 "This should be used only on Cygwin/Mingw targets");
6511 DebugLoc dl = Op.getDebugLoc();
6514 SDValue Chain = Op.getOperand(0);
6515 SDValue Size = Op.getOperand(1);
6516 // FIXME: Ensure alignment here
6520 EVT IntPtr = getPointerTy();
6521 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6523 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6524 Flag = Chain.getValue(1);
6526 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6528 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6529 Flag = Chain.getValue(1);
6531 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6533 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6534 return DAG.getMergeValues(Ops1, 2, dl);
6538 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6540 SDValue Dst, SDValue Src,
6541 SDValue Size, unsigned Align,
6543 uint64_t DstSVOff) {
6544 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6546 // If not DWORD aligned or size is more than the threshold, call the library.
6547 // The libc version is likely to be faster for these cases. It can use the
6548 // address value and run time information about the CPU.
6549 if ((Align & 3) != 0 ||
6551 ConstantSize->getZExtValue() >
6552 getSubtarget()->getMaxInlineSizeThreshold()) {
6553 SDValue InFlag(0, 0);
6555 // Check to see if there is a specialized entry-point for memory zeroing.
6556 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6558 if (const char *bzeroEntry = V &&
6559 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6560 EVT IntPtr = getPointerTy();
6561 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6562 TargetLowering::ArgListTy Args;
6563 TargetLowering::ArgListEntry Entry;
6565 Entry.Ty = IntPtrTy;
6566 Args.push_back(Entry);
6568 Args.push_back(Entry);
6569 std::pair<SDValue,SDValue> CallResult =
6570 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6571 false, false, false, false,
6572 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6573 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
6574 return CallResult.second;
6577 // Otherwise have the target-independent code call memset.
6581 uint64_t SizeVal = ConstantSize->getZExtValue();
6582 SDValue InFlag(0, 0);
6585 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6586 unsigned BytesLeft = 0;
6587 bool TwoRepStos = false;
6590 uint64_t Val = ValC->getZExtValue() & 255;
6592 // If the value is a constant, then we can potentially use larger sets.
6593 switch (Align & 3) {
6594 case 2: // WORD aligned
6597 Val = (Val << 8) | Val;
6599 case 0: // DWORD aligned
6602 Val = (Val << 8) | Val;
6603 Val = (Val << 16) | Val;
6604 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6607 Val = (Val << 32) | Val;
6610 default: // Byte aligned
6613 Count = DAG.getIntPtrConstant(SizeVal);
6617 if (AVT.bitsGT(MVT::i8)) {
6618 unsigned UBytes = AVT.getSizeInBits() / 8;
6619 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6620 BytesLeft = SizeVal % UBytes;
6623 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6625 InFlag = Chain.getValue(1);
6628 Count = DAG.getIntPtrConstant(SizeVal);
6629 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6630 InFlag = Chain.getValue(1);
6633 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6636 InFlag = Chain.getValue(1);
6637 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6640 InFlag = Chain.getValue(1);
6642 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6643 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6644 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6647 InFlag = Chain.getValue(1);
6649 EVT CVT = Count.getValueType();
6650 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6651 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6652 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6655 InFlag = Chain.getValue(1);
6656 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6657 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6658 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6659 } else if (BytesLeft) {
6660 // Handle the last 1 - 7 bytes.
6661 unsigned Offset = SizeVal - BytesLeft;
6662 EVT AddrVT = Dst.getValueType();
6663 EVT SizeVT = Size.getValueType();
6665 Chain = DAG.getMemset(Chain, dl,
6666 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6667 DAG.getConstant(Offset, AddrVT)),
6669 DAG.getConstant(BytesLeft, SizeVT),
6670 Align, DstSV, DstSVOff + Offset);
6673 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6678 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6679 SDValue Chain, SDValue Dst, SDValue Src,
6680 SDValue Size, unsigned Align,
6682 const Value *DstSV, uint64_t DstSVOff,
6683 const Value *SrcSV, uint64_t SrcSVOff) {
6684 // This requires the copy size to be a constant, preferrably
6685 // within a subtarget-specific limit.
6686 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6689 uint64_t SizeVal = ConstantSize->getZExtValue();
6690 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6693 /// If not DWORD aligned, call the library.
6694 if ((Align & 3) != 0)
6699 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6702 unsigned UBytes = AVT.getSizeInBits() / 8;
6703 unsigned CountVal = SizeVal / UBytes;
6704 SDValue Count = DAG.getIntPtrConstant(CountVal);
6705 unsigned BytesLeft = SizeVal % UBytes;
6707 SDValue InFlag(0, 0);
6708 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6711 InFlag = Chain.getValue(1);
6712 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6715 InFlag = Chain.getValue(1);
6716 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6719 InFlag = Chain.getValue(1);
6721 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6722 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6723 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6724 array_lengthof(Ops));
6726 SmallVector<SDValue, 4> Results;
6727 Results.push_back(RepMovs);
6729 // Handle the last 1 - 7 bytes.
6730 unsigned Offset = SizeVal - BytesLeft;
6731 EVT DstVT = Dst.getValueType();
6732 EVT SrcVT = Src.getValueType();
6733 EVT SizeVT = Size.getValueType();
6734 Results.push_back(DAG.getMemcpy(Chain, dl,
6735 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6736 DAG.getConstant(Offset, DstVT)),
6737 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6738 DAG.getConstant(Offset, SrcVT)),
6739 DAG.getConstant(BytesLeft, SizeVT),
6740 Align, AlwaysInline,
6741 DstSV, DstSVOff + Offset,
6742 SrcSV, SrcSVOff + Offset));
6745 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6746 &Results[0], Results.size());
6749 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6750 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6751 DebugLoc dl = Op.getDebugLoc();
6753 if (!Subtarget->is64Bit()) {
6754 // vastart just stores the address of the VarArgsFrameIndex slot into the
6755 // memory location argument.
6756 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6757 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6762 // gp_offset (0 - 6 * 8)
6763 // fp_offset (48 - 48 + 8 * 16)
6764 // overflow_arg_area (point to parameters coming in memory).
6766 SmallVector<SDValue, 8> MemOps;
6767 SDValue FIN = Op.getOperand(1);
6769 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6770 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6771 FIN, SV, 0, false, false, 0);
6772 MemOps.push_back(Store);
6775 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6776 FIN, DAG.getIntPtrConstant(4));
6777 Store = DAG.getStore(Op.getOperand(0), dl,
6778 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6779 FIN, SV, 0, false, false, 0);
6780 MemOps.push_back(Store);
6782 // Store ptr to overflow_arg_area
6783 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6784 FIN, DAG.getIntPtrConstant(4));
6785 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6786 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6788 MemOps.push_back(Store);
6790 // Store ptr to reg_save_area.
6791 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6792 FIN, DAG.getIntPtrConstant(8));
6793 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6794 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6796 MemOps.push_back(Store);
6797 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6798 &MemOps[0], MemOps.size());
6801 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6802 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6803 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6804 SDValue Chain = Op.getOperand(0);
6805 SDValue SrcPtr = Op.getOperand(1);
6806 SDValue SrcSV = Op.getOperand(2);
6808 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6812 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6813 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6814 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6815 SDValue Chain = Op.getOperand(0);
6816 SDValue DstPtr = Op.getOperand(1);
6817 SDValue SrcPtr = Op.getOperand(2);
6818 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6819 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6820 DebugLoc dl = Op.getDebugLoc();
6822 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6823 DAG.getIntPtrConstant(24), 8, false,
6824 DstSV, 0, SrcSV, 0);
6828 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6829 DebugLoc dl = Op.getDebugLoc();
6830 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6832 default: return SDValue(); // Don't custom lower most intrinsics.
6833 // Comparison intrinsics.
6834 case Intrinsic::x86_sse_comieq_ss:
6835 case Intrinsic::x86_sse_comilt_ss:
6836 case Intrinsic::x86_sse_comile_ss:
6837 case Intrinsic::x86_sse_comigt_ss:
6838 case Intrinsic::x86_sse_comige_ss:
6839 case Intrinsic::x86_sse_comineq_ss:
6840 case Intrinsic::x86_sse_ucomieq_ss:
6841 case Intrinsic::x86_sse_ucomilt_ss:
6842 case Intrinsic::x86_sse_ucomile_ss:
6843 case Intrinsic::x86_sse_ucomigt_ss:
6844 case Intrinsic::x86_sse_ucomige_ss:
6845 case Intrinsic::x86_sse_ucomineq_ss:
6846 case Intrinsic::x86_sse2_comieq_sd:
6847 case Intrinsic::x86_sse2_comilt_sd:
6848 case Intrinsic::x86_sse2_comile_sd:
6849 case Intrinsic::x86_sse2_comigt_sd:
6850 case Intrinsic::x86_sse2_comige_sd:
6851 case Intrinsic::x86_sse2_comineq_sd:
6852 case Intrinsic::x86_sse2_ucomieq_sd:
6853 case Intrinsic::x86_sse2_ucomilt_sd:
6854 case Intrinsic::x86_sse2_ucomile_sd:
6855 case Intrinsic::x86_sse2_ucomigt_sd:
6856 case Intrinsic::x86_sse2_ucomige_sd:
6857 case Intrinsic::x86_sse2_ucomineq_sd: {
6859 ISD::CondCode CC = ISD::SETCC_INVALID;
6862 case Intrinsic::x86_sse_comieq_ss:
6863 case Intrinsic::x86_sse2_comieq_sd:
6867 case Intrinsic::x86_sse_comilt_ss:
6868 case Intrinsic::x86_sse2_comilt_sd:
6872 case Intrinsic::x86_sse_comile_ss:
6873 case Intrinsic::x86_sse2_comile_sd:
6877 case Intrinsic::x86_sse_comigt_ss:
6878 case Intrinsic::x86_sse2_comigt_sd:
6882 case Intrinsic::x86_sse_comige_ss:
6883 case Intrinsic::x86_sse2_comige_sd:
6887 case Intrinsic::x86_sse_comineq_ss:
6888 case Intrinsic::x86_sse2_comineq_sd:
6892 case Intrinsic::x86_sse_ucomieq_ss:
6893 case Intrinsic::x86_sse2_ucomieq_sd:
6894 Opc = X86ISD::UCOMI;
6897 case Intrinsic::x86_sse_ucomilt_ss:
6898 case Intrinsic::x86_sse2_ucomilt_sd:
6899 Opc = X86ISD::UCOMI;
6902 case Intrinsic::x86_sse_ucomile_ss:
6903 case Intrinsic::x86_sse2_ucomile_sd:
6904 Opc = X86ISD::UCOMI;
6907 case Intrinsic::x86_sse_ucomigt_ss:
6908 case Intrinsic::x86_sse2_ucomigt_sd:
6909 Opc = X86ISD::UCOMI;
6912 case Intrinsic::x86_sse_ucomige_ss:
6913 case Intrinsic::x86_sse2_ucomige_sd:
6914 Opc = X86ISD::UCOMI;
6917 case Intrinsic::x86_sse_ucomineq_ss:
6918 case Intrinsic::x86_sse2_ucomineq_sd:
6919 Opc = X86ISD::UCOMI;
6924 SDValue LHS = Op.getOperand(1);
6925 SDValue RHS = Op.getOperand(2);
6926 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6927 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6928 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6929 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6930 DAG.getConstant(X86CC, MVT::i8), Cond);
6931 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6933 // ptest intrinsics. The intrinsic these come from are designed to return
6934 // an integer value, not just an instruction so lower it to the ptest
6935 // pattern and a setcc for the result.
6936 case Intrinsic::x86_sse41_ptestz:
6937 case Intrinsic::x86_sse41_ptestc:
6938 case Intrinsic::x86_sse41_ptestnzc:{
6941 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6942 case Intrinsic::x86_sse41_ptestz:
6944 X86CC = X86::COND_E;
6946 case Intrinsic::x86_sse41_ptestc:
6948 X86CC = X86::COND_B;
6950 case Intrinsic::x86_sse41_ptestnzc:
6952 X86CC = X86::COND_A;
6956 SDValue LHS = Op.getOperand(1);
6957 SDValue RHS = Op.getOperand(2);
6958 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6959 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6960 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6961 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6964 // Fix vector shift instructions where the last operand is a non-immediate
6966 case Intrinsic::x86_sse2_pslli_w:
6967 case Intrinsic::x86_sse2_pslli_d:
6968 case Intrinsic::x86_sse2_pslli_q:
6969 case Intrinsic::x86_sse2_psrli_w:
6970 case Intrinsic::x86_sse2_psrli_d:
6971 case Intrinsic::x86_sse2_psrli_q:
6972 case Intrinsic::x86_sse2_psrai_w:
6973 case Intrinsic::x86_sse2_psrai_d:
6974 case Intrinsic::x86_mmx_pslli_w:
6975 case Intrinsic::x86_mmx_pslli_d:
6976 case Intrinsic::x86_mmx_pslli_q:
6977 case Intrinsic::x86_mmx_psrli_w:
6978 case Intrinsic::x86_mmx_psrli_d:
6979 case Intrinsic::x86_mmx_psrli_q:
6980 case Intrinsic::x86_mmx_psrai_w:
6981 case Intrinsic::x86_mmx_psrai_d: {
6982 SDValue ShAmt = Op.getOperand(2);
6983 if (isa<ConstantSDNode>(ShAmt))
6986 unsigned NewIntNo = 0;
6987 EVT ShAmtVT = MVT::v4i32;
6989 case Intrinsic::x86_sse2_pslli_w:
6990 NewIntNo = Intrinsic::x86_sse2_psll_w;
6992 case Intrinsic::x86_sse2_pslli_d:
6993 NewIntNo = Intrinsic::x86_sse2_psll_d;
6995 case Intrinsic::x86_sse2_pslli_q:
6996 NewIntNo = Intrinsic::x86_sse2_psll_q;
6998 case Intrinsic::x86_sse2_psrli_w:
6999 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7001 case Intrinsic::x86_sse2_psrli_d:
7002 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7004 case Intrinsic::x86_sse2_psrli_q:
7005 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7007 case Intrinsic::x86_sse2_psrai_w:
7008 NewIntNo = Intrinsic::x86_sse2_psra_w;
7010 case Intrinsic::x86_sse2_psrai_d:
7011 NewIntNo = Intrinsic::x86_sse2_psra_d;
7014 ShAmtVT = MVT::v2i32;
7016 case Intrinsic::x86_mmx_pslli_w:
7017 NewIntNo = Intrinsic::x86_mmx_psll_w;
7019 case Intrinsic::x86_mmx_pslli_d:
7020 NewIntNo = Intrinsic::x86_mmx_psll_d;
7022 case Intrinsic::x86_mmx_pslli_q:
7023 NewIntNo = Intrinsic::x86_mmx_psll_q;
7025 case Intrinsic::x86_mmx_psrli_w:
7026 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7028 case Intrinsic::x86_mmx_psrli_d:
7029 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7031 case Intrinsic::x86_mmx_psrli_q:
7032 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7034 case Intrinsic::x86_mmx_psrai_w:
7035 NewIntNo = Intrinsic::x86_mmx_psra_w;
7037 case Intrinsic::x86_mmx_psrai_d:
7038 NewIntNo = Intrinsic::x86_mmx_psra_d;
7040 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7046 // The vector shift intrinsics with scalars uses 32b shift amounts but
7047 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7051 ShOps[1] = DAG.getConstant(0, MVT::i32);
7052 if (ShAmtVT == MVT::v4i32) {
7053 ShOps[2] = DAG.getUNDEF(MVT::i32);
7054 ShOps[3] = DAG.getUNDEF(MVT::i32);
7055 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7057 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7060 EVT VT = Op.getValueType();
7061 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7062 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7063 DAG.getConstant(NewIntNo, MVT::i32),
7064 Op.getOperand(1), ShAmt);
7069 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
7070 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7071 DebugLoc dl = Op.getDebugLoc();
7074 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7076 DAG.getConstant(TD->getPointerSize(),
7077 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7078 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7079 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7081 NULL, 0, false, false, 0);
7084 // Just load the return address.
7085 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7086 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7087 RetAddrFI, NULL, 0, false, false, 0);
7090 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
7091 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7092 MFI->setFrameAddressIsTaken(true);
7093 EVT VT = Op.getValueType();
7094 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7095 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7096 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7097 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7099 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7104 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7105 SelectionDAG &DAG) {
7106 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7109 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
7111 MachineFunction &MF = DAG.getMachineFunction();
7112 SDValue Chain = Op.getOperand(0);
7113 SDValue Offset = Op.getOperand(1);
7114 SDValue Handler = Op.getOperand(2);
7115 DebugLoc dl = Op.getDebugLoc();
7117 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7119 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7121 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7122 DAG.getIntPtrConstant(-TD->getPointerSize()));
7123 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7124 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7125 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7126 MF.getRegInfo().addLiveOut(StoreAddrReg);
7128 return DAG.getNode(X86ISD::EH_RETURN, dl,
7130 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7133 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7134 SelectionDAG &DAG) {
7135 SDValue Root = Op.getOperand(0);
7136 SDValue Trmp = Op.getOperand(1); // trampoline
7137 SDValue FPtr = Op.getOperand(2); // nested function
7138 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7139 DebugLoc dl = Op.getDebugLoc();
7141 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7143 if (Subtarget->is64Bit()) {
7144 SDValue OutChains[6];
7146 // Large code-model.
7147 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7148 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7150 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7151 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7153 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7155 // Load the pointer to the nested function into R11.
7156 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7157 SDValue Addr = Trmp;
7158 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7159 Addr, TrmpAddr, 0, false, false, 0);
7161 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7162 DAG.getConstant(2, MVT::i64));
7163 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7166 // Load the 'nest' parameter value into R10.
7167 // R10 is specified in X86CallingConv.td
7168 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7169 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7170 DAG.getConstant(10, MVT::i64));
7171 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7172 Addr, TrmpAddr, 10, false, false, 0);
7174 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7175 DAG.getConstant(12, MVT::i64));
7176 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7179 // Jump to the nested function.
7180 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7181 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7182 DAG.getConstant(20, MVT::i64));
7183 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7184 Addr, TrmpAddr, 20, false, false, 0);
7186 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7187 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7188 DAG.getConstant(22, MVT::i64));
7189 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7190 TrmpAddr, 22, false, false, 0);
7193 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7194 return DAG.getMergeValues(Ops, 2, dl);
7196 const Function *Func =
7197 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7198 CallingConv::ID CC = Func->getCallingConv();
7203 llvm_unreachable("Unsupported calling convention");
7204 case CallingConv::C:
7205 case CallingConv::X86_StdCall: {
7206 // Pass 'nest' parameter in ECX.
7207 // Must be kept in sync with X86CallingConv.td
7210 // Check that ECX wasn't needed by an 'inreg' parameter.
7211 const FunctionType *FTy = Func->getFunctionType();
7212 const AttrListPtr &Attrs = Func->getAttributes();
7214 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7215 unsigned InRegCount = 0;
7218 for (FunctionType::param_iterator I = FTy->param_begin(),
7219 E = FTy->param_end(); I != E; ++I, ++Idx)
7220 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7221 // FIXME: should only count parameters that are lowered to integers.
7222 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7224 if (InRegCount > 2) {
7225 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7230 case CallingConv::X86_FastCall:
7231 case CallingConv::Fast:
7232 // Pass 'nest' parameter in EAX.
7233 // Must be kept in sync with X86CallingConv.td
7238 SDValue OutChains[4];
7241 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7242 DAG.getConstant(10, MVT::i32));
7243 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7245 // This is storing the opcode for MOV32ri.
7246 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7247 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7248 OutChains[0] = DAG.getStore(Root, dl,
7249 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7250 Trmp, TrmpAddr, 0, false, false, 0);
7252 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7253 DAG.getConstant(1, MVT::i32));
7254 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7257 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7258 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7259 DAG.getConstant(5, MVT::i32));
7260 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7261 TrmpAddr, 5, false, false, 1);
7263 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7264 DAG.getConstant(6, MVT::i32));
7265 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7269 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7270 return DAG.getMergeValues(Ops, 2, dl);
7274 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7276 The rounding mode is in bits 11:10 of FPSR, and has the following
7283 FLT_ROUNDS, on the other hand, expects the following:
7290 To perform the conversion, we do:
7291 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7294 MachineFunction &MF = DAG.getMachineFunction();
7295 const TargetMachine &TM = MF.getTarget();
7296 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7297 unsigned StackAlignment = TFI.getStackAlignment();
7298 EVT VT = Op.getValueType();
7299 DebugLoc dl = Op.getDebugLoc();
7301 // Save FP Control Word to stack slot
7302 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7303 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7305 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7306 DAG.getEntryNode(), StackSlot);
7308 // Load FP Control Word from stack slot
7309 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7312 // Transform as necessary
7314 DAG.getNode(ISD::SRL, dl, MVT::i16,
7315 DAG.getNode(ISD::AND, dl, MVT::i16,
7316 CWD, DAG.getConstant(0x800, MVT::i16)),
7317 DAG.getConstant(11, MVT::i8));
7319 DAG.getNode(ISD::SRL, dl, MVT::i16,
7320 DAG.getNode(ISD::AND, dl, MVT::i16,
7321 CWD, DAG.getConstant(0x400, MVT::i16)),
7322 DAG.getConstant(9, MVT::i8));
7325 DAG.getNode(ISD::AND, dl, MVT::i16,
7326 DAG.getNode(ISD::ADD, dl, MVT::i16,
7327 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7328 DAG.getConstant(1, MVT::i16)),
7329 DAG.getConstant(3, MVT::i16));
7332 return DAG.getNode((VT.getSizeInBits() < 16 ?
7333 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7336 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7337 EVT VT = Op.getValueType();
7339 unsigned NumBits = VT.getSizeInBits();
7340 DebugLoc dl = Op.getDebugLoc();
7342 Op = Op.getOperand(0);
7343 if (VT == MVT::i8) {
7344 // Zero extend to i32 since there is not an i8 bsr.
7346 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7349 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7350 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7351 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7353 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7356 DAG.getConstant(NumBits+NumBits-1, OpVT),
7357 DAG.getConstant(X86::COND_E, MVT::i8),
7360 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7362 // Finally xor with NumBits-1.
7363 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7366 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7370 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7371 EVT VT = Op.getValueType();
7373 unsigned NumBits = VT.getSizeInBits();
7374 DebugLoc dl = Op.getDebugLoc();
7376 Op = Op.getOperand(0);
7377 if (VT == MVT::i8) {
7379 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7382 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7383 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7384 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7386 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7389 DAG.getConstant(NumBits, OpVT),
7390 DAG.getConstant(X86::COND_E, MVT::i8),
7393 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7396 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7400 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7401 EVT VT = Op.getValueType();
7402 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7403 DebugLoc dl = Op.getDebugLoc();
7405 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7406 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7407 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7408 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7409 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7411 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7412 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7413 // return AloBlo + AloBhi + AhiBlo;
7415 SDValue A = Op.getOperand(0);
7416 SDValue B = Op.getOperand(1);
7418 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7419 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7420 A, DAG.getConstant(32, MVT::i32));
7421 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7422 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7423 B, DAG.getConstant(32, MVT::i32));
7424 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7425 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7427 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7428 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7430 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7431 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7433 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7434 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7435 AloBhi, DAG.getConstant(32, MVT::i32));
7436 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7437 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7438 AhiBlo, DAG.getConstant(32, MVT::i32));
7439 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7440 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7445 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7446 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7447 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7448 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7449 // has only one use.
7450 SDNode *N = Op.getNode();
7451 SDValue LHS = N->getOperand(0);
7452 SDValue RHS = N->getOperand(1);
7453 unsigned BaseOp = 0;
7455 DebugLoc dl = Op.getDebugLoc();
7457 switch (Op.getOpcode()) {
7458 default: llvm_unreachable("Unknown ovf instruction!");
7460 // A subtract of one will be selected as a INC. Note that INC doesn't
7461 // set CF, so we can't do this for UADDO.
7462 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7463 if (C->getAPIntValue() == 1) {
7464 BaseOp = X86ISD::INC;
7468 BaseOp = X86ISD::ADD;
7472 BaseOp = X86ISD::ADD;
7476 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7477 // set CF, so we can't do this for USUBO.
7478 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7479 if (C->getAPIntValue() == 1) {
7480 BaseOp = X86ISD::DEC;
7484 BaseOp = X86ISD::SUB;
7488 BaseOp = X86ISD::SUB;
7492 BaseOp = X86ISD::SMUL;
7496 BaseOp = X86ISD::UMUL;
7501 // Also sets EFLAGS.
7502 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7503 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7506 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7507 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7509 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7513 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7514 EVT T = Op.getValueType();
7515 DebugLoc dl = Op.getDebugLoc();
7518 switch(T.getSimpleVT().SimpleTy) {
7520 assert(false && "Invalid value type!");
7521 case MVT::i8: Reg = X86::AL; size = 1; break;
7522 case MVT::i16: Reg = X86::AX; size = 2; break;
7523 case MVT::i32: Reg = X86::EAX; size = 4; break;
7525 assert(Subtarget->is64Bit() && "Node not type legal!");
7526 Reg = X86::RAX; size = 8;
7529 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7530 Op.getOperand(2), SDValue());
7531 SDValue Ops[] = { cpIn.getValue(0),
7534 DAG.getTargetConstant(size, MVT::i8),
7536 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7537 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7539 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7543 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7544 SelectionDAG &DAG) {
7545 assert(Subtarget->is64Bit() && "Result not type legalized?");
7546 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7547 SDValue TheChain = Op.getOperand(0);
7548 DebugLoc dl = Op.getDebugLoc();
7549 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7550 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7551 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7553 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7554 DAG.getConstant(32, MVT::i8));
7556 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7559 return DAG.getMergeValues(Ops, 2, dl);
7562 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7563 SDNode *Node = Op.getNode();
7564 DebugLoc dl = Node->getDebugLoc();
7565 EVT T = Node->getValueType(0);
7566 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7567 DAG.getConstant(0, T), Node->getOperand(2));
7568 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7569 cast<AtomicSDNode>(Node)->getMemoryVT(),
7570 Node->getOperand(0),
7571 Node->getOperand(1), negOp,
7572 cast<AtomicSDNode>(Node)->getSrcValue(),
7573 cast<AtomicSDNode>(Node)->getAlignment());
7576 /// LowerOperation - Provide custom lowering hooks for some operations.
7578 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7579 switch (Op.getOpcode()) {
7580 default: llvm_unreachable("Should not custom lower this!");
7581 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7582 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7583 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7584 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7585 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7586 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7587 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7588 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7589 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7590 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7591 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7592 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7593 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7594 case ISD::SHL_PARTS:
7595 case ISD::SRA_PARTS:
7596 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7597 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7598 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7599 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7600 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7601 case ISD::FABS: return LowerFABS(Op, DAG);
7602 case ISD::FNEG: return LowerFNEG(Op, DAG);
7603 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7604 case ISD::SETCC: return LowerSETCC(Op, DAG);
7605 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7606 case ISD::SELECT: return LowerSELECT(Op, DAG);
7607 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7608 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7609 case ISD::VASTART: return LowerVASTART(Op, DAG);
7610 case ISD::VAARG: return LowerVAARG(Op, DAG);
7611 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7612 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7613 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7614 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7615 case ISD::FRAME_TO_ARGS_OFFSET:
7616 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7617 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7618 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7619 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7620 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7621 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7622 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7623 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7629 case ISD::UMULO: return LowerXALUO(Op, DAG);
7630 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7634 void X86TargetLowering::
7635 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7636 SelectionDAG &DAG, unsigned NewOp) {
7637 EVT T = Node->getValueType(0);
7638 DebugLoc dl = Node->getDebugLoc();
7639 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7641 SDValue Chain = Node->getOperand(0);
7642 SDValue In1 = Node->getOperand(1);
7643 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7644 Node->getOperand(2), DAG.getIntPtrConstant(0));
7645 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7646 Node->getOperand(2), DAG.getIntPtrConstant(1));
7647 SDValue Ops[] = { Chain, In1, In2L, In2H };
7648 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7650 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7651 cast<MemSDNode>(Node)->getMemOperand());
7652 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7653 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7654 Results.push_back(Result.getValue(2));
7657 /// ReplaceNodeResults - Replace a node with an illegal result type
7658 /// with a new node built out of custom code.
7659 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7660 SmallVectorImpl<SDValue>&Results,
7661 SelectionDAG &DAG) {
7662 DebugLoc dl = N->getDebugLoc();
7663 switch (N->getOpcode()) {
7665 assert(false && "Do not know how to custom type legalize this operation!");
7667 case ISD::FP_TO_SINT: {
7668 std::pair<SDValue,SDValue> Vals =
7669 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7670 SDValue FIST = Vals.first, StackSlot = Vals.second;
7671 if (FIST.getNode() != 0) {
7672 EVT VT = N->getValueType(0);
7673 // Return a load from the stack slot.
7674 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7679 case ISD::READCYCLECOUNTER: {
7680 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7681 SDValue TheChain = N->getOperand(0);
7682 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7683 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7685 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7687 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7688 SDValue Ops[] = { eax, edx };
7689 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7690 Results.push_back(edx.getValue(1));
7693 case ISD::ATOMIC_CMP_SWAP: {
7694 EVT T = N->getValueType(0);
7695 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7696 SDValue cpInL, cpInH;
7697 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7698 DAG.getConstant(0, MVT::i32));
7699 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7700 DAG.getConstant(1, MVT::i32));
7701 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7702 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7704 SDValue swapInL, swapInH;
7705 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7706 DAG.getConstant(0, MVT::i32));
7707 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7708 DAG.getConstant(1, MVT::i32));
7709 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7711 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7712 swapInL.getValue(1));
7713 SDValue Ops[] = { swapInH.getValue(0),
7715 swapInH.getValue(1) };
7716 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7717 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7718 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7719 MVT::i32, Result.getValue(1));
7720 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7721 MVT::i32, cpOutL.getValue(2));
7722 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7723 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7724 Results.push_back(cpOutH.getValue(1));
7727 case ISD::ATOMIC_LOAD_ADD:
7728 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7730 case ISD::ATOMIC_LOAD_AND:
7731 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7733 case ISD::ATOMIC_LOAD_NAND:
7734 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7736 case ISD::ATOMIC_LOAD_OR:
7737 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7739 case ISD::ATOMIC_LOAD_SUB:
7740 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7742 case ISD::ATOMIC_LOAD_XOR:
7743 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7745 case ISD::ATOMIC_SWAP:
7746 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7751 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7753 default: return NULL;
7754 case X86ISD::BSF: return "X86ISD::BSF";
7755 case X86ISD::BSR: return "X86ISD::BSR";
7756 case X86ISD::SHLD: return "X86ISD::SHLD";
7757 case X86ISD::SHRD: return "X86ISD::SHRD";
7758 case X86ISD::FAND: return "X86ISD::FAND";
7759 case X86ISD::FOR: return "X86ISD::FOR";
7760 case X86ISD::FXOR: return "X86ISD::FXOR";
7761 case X86ISD::FSRL: return "X86ISD::FSRL";
7762 case X86ISD::FILD: return "X86ISD::FILD";
7763 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7764 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7765 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7766 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7767 case X86ISD::FLD: return "X86ISD::FLD";
7768 case X86ISD::FST: return "X86ISD::FST";
7769 case X86ISD::CALL: return "X86ISD::CALL";
7770 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7771 case X86ISD::BT: return "X86ISD::BT";
7772 case X86ISD::CMP: return "X86ISD::CMP";
7773 case X86ISD::COMI: return "X86ISD::COMI";
7774 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7775 case X86ISD::SETCC: return "X86ISD::SETCC";
7776 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7777 case X86ISD::CMOV: return "X86ISD::CMOV";
7778 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7779 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7780 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7781 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7782 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7783 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7784 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7785 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7786 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7787 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7788 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7789 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7790 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7791 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7792 case X86ISD::FMAX: return "X86ISD::FMAX";
7793 case X86ISD::FMIN: return "X86ISD::FMIN";
7794 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7795 case X86ISD::FRCP: return "X86ISD::FRCP";
7796 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7797 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7798 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7799 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7800 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7801 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7802 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7803 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7804 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7805 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7806 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7807 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7808 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7809 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7810 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7811 case X86ISD::VSHL: return "X86ISD::VSHL";
7812 case X86ISD::VSRL: return "X86ISD::VSRL";
7813 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7814 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7815 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7816 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7817 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7818 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7819 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7820 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7821 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7822 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7823 case X86ISD::ADD: return "X86ISD::ADD";
7824 case X86ISD::SUB: return "X86ISD::SUB";
7825 case X86ISD::SMUL: return "X86ISD::SMUL";
7826 case X86ISD::UMUL: return "X86ISD::UMUL";
7827 case X86ISD::INC: return "X86ISD::INC";
7828 case X86ISD::DEC: return "X86ISD::DEC";
7829 case X86ISD::OR: return "X86ISD::OR";
7830 case X86ISD::XOR: return "X86ISD::XOR";
7831 case X86ISD::AND: return "X86ISD::AND";
7832 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7833 case X86ISD::PTEST: return "X86ISD::PTEST";
7834 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7835 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7839 // isLegalAddressingMode - Return true if the addressing mode represented
7840 // by AM is legal for this target, for a load/store of the specified type.
7841 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7842 const Type *Ty) const {
7843 // X86 supports extremely general addressing modes.
7844 CodeModel::Model M = getTargetMachine().getCodeModel();
7846 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7847 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7852 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7854 // If a reference to this global requires an extra load, we can't fold it.
7855 if (isGlobalStubReference(GVFlags))
7858 // If BaseGV requires a register for the PIC base, we cannot also have a
7859 // BaseReg specified.
7860 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7863 // If lower 4G is not available, then we must use rip-relative addressing.
7864 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7874 // These scales always work.
7879 // These scales are formed with basereg+scalereg. Only accept if there is
7884 default: // Other stuff never works.
7892 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7893 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7895 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7896 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7897 if (NumBits1 <= NumBits2)
7902 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7903 if (!VT1.isInteger() || !VT2.isInteger())
7905 unsigned NumBits1 = VT1.getSizeInBits();
7906 unsigned NumBits2 = VT2.getSizeInBits();
7907 if (NumBits1 <= NumBits2)
7912 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7913 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7914 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7917 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7918 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7919 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7922 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7923 // i16 instructions are longer (0x66 prefix) and potentially slower.
7924 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7927 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7928 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7929 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7930 /// are assumed to be legal.
7932 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7934 // Only do shuffles on 128-bit vector types for now.
7935 if (VT.getSizeInBits() == 64)
7938 // FIXME: pshufb, blends, shifts.
7939 return (VT.getVectorNumElements() == 2 ||
7940 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7941 isMOVLMask(M, VT) ||
7942 isSHUFPMask(M, VT) ||
7943 isPSHUFDMask(M, VT) ||
7944 isPSHUFHWMask(M, VT) ||
7945 isPSHUFLWMask(M, VT) ||
7946 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7947 isUNPCKLMask(M, VT) ||
7948 isUNPCKHMask(M, VT) ||
7949 isUNPCKL_v_undef_Mask(M, VT) ||
7950 isUNPCKH_v_undef_Mask(M, VT));
7954 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7956 unsigned NumElts = VT.getVectorNumElements();
7957 // FIXME: This collection of masks seems suspect.
7960 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7961 return (isMOVLMask(Mask, VT) ||
7962 isCommutedMOVLMask(Mask, VT, true) ||
7963 isSHUFPMask(Mask, VT) ||
7964 isCommutedSHUFPMask(Mask, VT));
7969 //===----------------------------------------------------------------------===//
7970 // X86 Scheduler Hooks
7971 //===----------------------------------------------------------------------===//
7973 // private utility function
7975 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7976 MachineBasicBlock *MBB,
7984 TargetRegisterClass *RC,
7985 bool invSrc) const {
7986 // For the atomic bitwise operator, we generate
7989 // ld t1 = [bitinstr.addr]
7990 // op t2 = t1, [bitinstr.val]
7992 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7994 // fallthrough -->nextMBB
7995 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7996 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7997 MachineFunction::iterator MBBIter = MBB;
8000 /// First build the CFG
8001 MachineFunction *F = MBB->getParent();
8002 MachineBasicBlock *thisMBB = MBB;
8003 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8004 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8005 F->insert(MBBIter, newMBB);
8006 F->insert(MBBIter, nextMBB);
8008 // Move all successors to thisMBB to nextMBB
8009 nextMBB->transferSuccessors(thisMBB);
8011 // Update thisMBB to fall through to newMBB
8012 thisMBB->addSuccessor(newMBB);
8014 // newMBB jumps to itself and fall through to nextMBB
8015 newMBB->addSuccessor(nextMBB);
8016 newMBB->addSuccessor(newMBB);
8018 // Insert instructions into newMBB based on incoming instruction
8019 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8020 "unexpected number of operands");
8021 DebugLoc dl = bInstr->getDebugLoc();
8022 MachineOperand& destOper = bInstr->getOperand(0);
8023 MachineOperand* argOpers[2 + X86AddrNumOperands];
8024 int numArgs = bInstr->getNumOperands() - 1;
8025 for (int i=0; i < numArgs; ++i)
8026 argOpers[i] = &bInstr->getOperand(i+1);
8028 // x86 address has 4 operands: base, index, scale, and displacement
8029 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8030 int valArgIndx = lastAddrIndx + 1;
8032 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8033 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8034 for (int i=0; i <= lastAddrIndx; ++i)
8035 (*MIB).addOperand(*argOpers[i]);
8037 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8039 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8044 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8045 assert((argOpers[valArgIndx]->isReg() ||
8046 argOpers[valArgIndx]->isImm()) &&
8048 if (argOpers[valArgIndx]->isReg())
8049 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8051 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8053 (*MIB).addOperand(*argOpers[valArgIndx]);
8055 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8058 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8059 for (int i=0; i <= lastAddrIndx; ++i)
8060 (*MIB).addOperand(*argOpers[i]);
8062 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8063 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8064 bInstr->memoperands_end());
8066 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8070 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8072 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8076 // private utility function: 64 bit atomics on 32 bit host.
8078 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8079 MachineBasicBlock *MBB,
8084 bool invSrc) const {
8085 // For the atomic bitwise operator, we generate
8086 // thisMBB (instructions are in pairs, except cmpxchg8b)
8087 // ld t1,t2 = [bitinstr.addr]
8089 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8090 // op t5, t6 <- out1, out2, [bitinstr.val]
8091 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8092 // mov ECX, EBX <- t5, t6
8093 // mov EAX, EDX <- t1, t2
8094 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8095 // mov t3, t4 <- EAX, EDX
8097 // result in out1, out2
8098 // fallthrough -->nextMBB
8100 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8101 const unsigned LoadOpc = X86::MOV32rm;
8102 const unsigned copyOpc = X86::MOV32rr;
8103 const unsigned NotOpc = X86::NOT32r;
8104 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8105 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8106 MachineFunction::iterator MBBIter = MBB;
8109 /// First build the CFG
8110 MachineFunction *F = MBB->getParent();
8111 MachineBasicBlock *thisMBB = MBB;
8112 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8113 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8114 F->insert(MBBIter, newMBB);
8115 F->insert(MBBIter, nextMBB);
8117 // Move all successors to thisMBB to nextMBB
8118 nextMBB->transferSuccessors(thisMBB);
8120 // Update thisMBB to fall through to newMBB
8121 thisMBB->addSuccessor(newMBB);
8123 // newMBB jumps to itself and fall through to nextMBB
8124 newMBB->addSuccessor(nextMBB);
8125 newMBB->addSuccessor(newMBB);
8127 DebugLoc dl = bInstr->getDebugLoc();
8128 // Insert instructions into newMBB based on incoming instruction
8129 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8130 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8131 "unexpected number of operands");
8132 MachineOperand& dest1Oper = bInstr->getOperand(0);
8133 MachineOperand& dest2Oper = bInstr->getOperand(1);
8134 MachineOperand* argOpers[2 + X86AddrNumOperands];
8135 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8136 argOpers[i] = &bInstr->getOperand(i+2);
8138 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8139 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8141 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8142 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8143 for (int i=0; i <= lastAddrIndx; ++i)
8144 (*MIB).addOperand(*argOpers[i]);
8145 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8146 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8147 // add 4 to displacement.
8148 for (int i=0; i <= lastAddrIndx-2; ++i)
8149 (*MIB).addOperand(*argOpers[i]);
8150 MachineOperand newOp3 = *(argOpers[3]);
8152 newOp3.setImm(newOp3.getImm()+4);
8154 newOp3.setOffset(newOp3.getOffset()+4);
8155 (*MIB).addOperand(newOp3);
8156 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8158 // t3/4 are defined later, at the bottom of the loop
8159 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8160 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8161 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8162 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8163 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8164 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8166 // The subsequent operations should be using the destination registers of
8167 //the PHI instructions.
8169 t1 = F->getRegInfo().createVirtualRegister(RC);
8170 t2 = F->getRegInfo().createVirtualRegister(RC);
8171 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8172 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8174 t1 = dest1Oper.getReg();
8175 t2 = dest2Oper.getReg();
8178 int valArgIndx = lastAddrIndx + 1;
8179 assert((argOpers[valArgIndx]->isReg() ||
8180 argOpers[valArgIndx]->isImm()) &&
8182 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8183 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8184 if (argOpers[valArgIndx]->isReg())
8185 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8187 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8188 if (regOpcL != X86::MOV32rr)
8190 (*MIB).addOperand(*argOpers[valArgIndx]);
8191 assert(argOpers[valArgIndx + 1]->isReg() ==
8192 argOpers[valArgIndx]->isReg());
8193 assert(argOpers[valArgIndx + 1]->isImm() ==
8194 argOpers[valArgIndx]->isImm());
8195 if (argOpers[valArgIndx + 1]->isReg())
8196 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8198 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8199 if (regOpcH != X86::MOV32rr)
8201 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8203 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8205 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8208 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8210 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8213 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8214 for (int i=0; i <= lastAddrIndx; ++i)
8215 (*MIB).addOperand(*argOpers[i]);
8217 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8218 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8219 bInstr->memoperands_end());
8221 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8222 MIB.addReg(X86::EAX);
8223 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8224 MIB.addReg(X86::EDX);
8227 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8229 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8233 // private utility function
8235 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8236 MachineBasicBlock *MBB,
8237 unsigned cmovOpc) const {
8238 // For the atomic min/max operator, we generate
8241 // ld t1 = [min/max.addr]
8242 // mov t2 = [min/max.val]
8244 // cmov[cond] t2 = t1
8246 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8248 // fallthrough -->nextMBB
8250 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8251 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8252 MachineFunction::iterator MBBIter = MBB;
8255 /// First build the CFG
8256 MachineFunction *F = MBB->getParent();
8257 MachineBasicBlock *thisMBB = MBB;
8258 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8259 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8260 F->insert(MBBIter, newMBB);
8261 F->insert(MBBIter, nextMBB);
8263 // Move all successors of thisMBB to nextMBB
8264 nextMBB->transferSuccessors(thisMBB);
8266 // Update thisMBB to fall through to newMBB
8267 thisMBB->addSuccessor(newMBB);
8269 // newMBB jumps to newMBB and fall through to nextMBB
8270 newMBB->addSuccessor(nextMBB);
8271 newMBB->addSuccessor(newMBB);
8273 DebugLoc dl = mInstr->getDebugLoc();
8274 // Insert instructions into newMBB based on incoming instruction
8275 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8276 "unexpected number of operands");
8277 MachineOperand& destOper = mInstr->getOperand(0);
8278 MachineOperand* argOpers[2 + X86AddrNumOperands];
8279 int numArgs = mInstr->getNumOperands() - 1;
8280 for (int i=0; i < numArgs; ++i)
8281 argOpers[i] = &mInstr->getOperand(i+1);
8283 // x86 address has 4 operands: base, index, scale, and displacement
8284 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8285 int valArgIndx = lastAddrIndx + 1;
8287 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8288 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8289 for (int i=0; i <= lastAddrIndx; ++i)
8290 (*MIB).addOperand(*argOpers[i]);
8292 // We only support register and immediate values
8293 assert((argOpers[valArgIndx]->isReg() ||
8294 argOpers[valArgIndx]->isImm()) &&
8297 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8298 if (argOpers[valArgIndx]->isReg())
8299 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8301 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8302 (*MIB).addOperand(*argOpers[valArgIndx]);
8304 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8307 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8312 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8313 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8317 // Cmp and exchange if none has modified the memory location
8318 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8319 for (int i=0; i <= lastAddrIndx; ++i)
8320 (*MIB).addOperand(*argOpers[i]);
8322 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8323 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8324 mInstr->memoperands_end());
8326 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8327 MIB.addReg(X86::EAX);
8330 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8332 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8336 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8337 // all of this code can be replaced with that in the .td file.
8339 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8340 unsigned numArgs, bool memArg) const {
8342 MachineFunction *F = BB->getParent();
8343 DebugLoc dl = MI->getDebugLoc();
8344 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8348 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8350 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8352 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8354 for (unsigned i = 0; i < numArgs; ++i) {
8355 MachineOperand &Op = MI->getOperand(i+1);
8357 if (!(Op.isReg() && Op.isImplicit()))
8361 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8364 F->DeleteMachineInstr(MI);
8370 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8372 MachineBasicBlock *MBB) const {
8373 // Emit code to save XMM registers to the stack. The ABI says that the
8374 // number of registers to save is given in %al, so it's theoretically
8375 // possible to do an indirect jump trick to avoid saving all of them,
8376 // however this code takes a simpler approach and just executes all
8377 // of the stores if %al is non-zero. It's less code, and it's probably
8378 // easier on the hardware branch predictor, and stores aren't all that
8379 // expensive anyway.
8381 // Create the new basic blocks. One block contains all the XMM stores,
8382 // and one block is the final destination regardless of whether any
8383 // stores were performed.
8384 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8385 MachineFunction *F = MBB->getParent();
8386 MachineFunction::iterator MBBIter = MBB;
8388 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8389 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8390 F->insert(MBBIter, XMMSaveMBB);
8391 F->insert(MBBIter, EndMBB);
8394 // Move any original successors of MBB to the end block.
8395 EndMBB->transferSuccessors(MBB);
8396 // The original block will now fall through to the XMM save block.
8397 MBB->addSuccessor(XMMSaveMBB);
8398 // The XMMSaveMBB will fall through to the end block.
8399 XMMSaveMBB->addSuccessor(EndMBB);
8401 // Now add the instructions.
8402 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8403 DebugLoc DL = MI->getDebugLoc();
8405 unsigned CountReg = MI->getOperand(0).getReg();
8406 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8407 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8409 if (!Subtarget->isTargetWin64()) {
8410 // If %al is 0, branch around the XMM save block.
8411 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8412 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8413 MBB->addSuccessor(EndMBB);
8416 // In the XMM save block, save all the XMM argument registers.
8417 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8418 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8419 MachineMemOperand *MMO =
8420 F->getMachineMemOperand(
8421 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8422 MachineMemOperand::MOStore, Offset,
8423 /*Size=*/16, /*Align=*/16);
8424 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8425 .addFrameIndex(RegSaveFrameIndex)
8426 .addImm(/*Scale=*/1)
8427 .addReg(/*IndexReg=*/0)
8428 .addImm(/*Disp=*/Offset)
8429 .addReg(/*Segment=*/0)
8430 .addReg(MI->getOperand(i).getReg())
8431 .addMemOperand(MMO);
8434 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8440 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8441 MachineBasicBlock *BB,
8442 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8443 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8444 DebugLoc DL = MI->getDebugLoc();
8446 // To "insert" a SELECT_CC instruction, we actually have to insert the
8447 // diamond control-flow pattern. The incoming instruction knows the
8448 // destination vreg to set, the condition code register to branch on, the
8449 // true/false values to select between, and a branch opcode to use.
8450 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8451 MachineFunction::iterator It = BB;
8457 // cmpTY ccX, r1, r2
8459 // fallthrough --> copy0MBB
8460 MachineBasicBlock *thisMBB = BB;
8461 MachineFunction *F = BB->getParent();
8462 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8463 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8465 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8466 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8467 F->insert(It, copy0MBB);
8468 F->insert(It, sinkMBB);
8469 // Update machine-CFG edges by first adding all successors of the current
8470 // block to the new block which will contain the Phi node for the select.
8471 // Also inform sdisel of the edge changes.
8472 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8473 E = BB->succ_end(); I != E; ++I) {
8474 EM->insert(std::make_pair(*I, sinkMBB));
8475 sinkMBB->addSuccessor(*I);
8477 // Next, remove all successors of the current block, and add the true
8478 // and fallthrough blocks as its successors.
8479 while (!BB->succ_empty())
8480 BB->removeSuccessor(BB->succ_begin());
8481 // Add the true and fallthrough blocks as its successors.
8482 BB->addSuccessor(copy0MBB);
8483 BB->addSuccessor(sinkMBB);
8486 // %FalseValue = ...
8487 // # fallthrough to sinkMBB
8490 // Update machine-CFG edges
8491 BB->addSuccessor(sinkMBB);
8494 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8497 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8498 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8499 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8501 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8506 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8507 MachineBasicBlock *BB,
8508 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8509 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8510 DebugLoc DL = MI->getDebugLoc();
8511 MachineFunction *F = BB->getParent();
8513 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8514 // non-trivial part is impdef of ESP.
8515 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8518 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8519 .addExternalSymbol("_alloca")
8520 .addReg(X86::EAX, RegState::Implicit)
8521 .addReg(X86::ESP, RegState::Implicit)
8522 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8523 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8525 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8530 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8531 MachineBasicBlock *BB,
8532 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8533 switch (MI->getOpcode()) {
8534 default: assert(false && "Unexpected instr type to insert");
8535 case X86::MINGW_ALLOCA:
8536 return EmitLoweredMingwAlloca(MI, BB, EM);
8538 case X86::CMOV_V1I64:
8539 case X86::CMOV_FR32:
8540 case X86::CMOV_FR64:
8541 case X86::CMOV_V4F32:
8542 case X86::CMOV_V2F64:
8543 case X86::CMOV_V2I64:
8544 case X86::CMOV_GR16:
8545 case X86::CMOV_GR32:
8546 case X86::CMOV_RFP32:
8547 case X86::CMOV_RFP64:
8548 case X86::CMOV_RFP80:
8549 return EmitLoweredSelect(MI, BB, EM);
8551 case X86::FP32_TO_INT16_IN_MEM:
8552 case X86::FP32_TO_INT32_IN_MEM:
8553 case X86::FP32_TO_INT64_IN_MEM:
8554 case X86::FP64_TO_INT16_IN_MEM:
8555 case X86::FP64_TO_INT32_IN_MEM:
8556 case X86::FP64_TO_INT64_IN_MEM:
8557 case X86::FP80_TO_INT16_IN_MEM:
8558 case X86::FP80_TO_INT32_IN_MEM:
8559 case X86::FP80_TO_INT64_IN_MEM: {
8560 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8561 DebugLoc DL = MI->getDebugLoc();
8563 // Change the floating point control register to use "round towards zero"
8564 // mode when truncating to an integer value.
8565 MachineFunction *F = BB->getParent();
8566 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8567 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8569 // Load the old value of the high byte of the control word...
8571 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8572 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8575 // Set the high part to be round to zero...
8576 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8579 // Reload the modified control word now...
8580 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8582 // Restore the memory image of control word to original value
8583 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8586 // Get the X86 opcode to use.
8588 switch (MI->getOpcode()) {
8589 default: llvm_unreachable("illegal opcode!");
8590 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8591 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8592 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8593 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8594 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8595 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8596 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8597 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8598 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8602 MachineOperand &Op = MI->getOperand(0);
8604 AM.BaseType = X86AddressMode::RegBase;
8605 AM.Base.Reg = Op.getReg();
8607 AM.BaseType = X86AddressMode::FrameIndexBase;
8608 AM.Base.FrameIndex = Op.getIndex();
8610 Op = MI->getOperand(1);
8612 AM.Scale = Op.getImm();
8613 Op = MI->getOperand(2);
8615 AM.IndexReg = Op.getImm();
8616 Op = MI->getOperand(3);
8617 if (Op.isGlobal()) {
8618 AM.GV = Op.getGlobal();
8620 AM.Disp = Op.getImm();
8622 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8623 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8625 // Reload the original control word now.
8626 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8628 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8631 // DBG_VALUE. Only the frame index case is done here.
8632 case X86::DBG_VALUE: {
8633 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8634 DebugLoc DL = MI->getDebugLoc();
8636 MachineFunction *F = BB->getParent();
8637 AM.BaseType = X86AddressMode::FrameIndexBase;
8638 AM.Base.FrameIndex = MI->getOperand(0).getImm();
8639 addFullAddress(BuildMI(BB, DL, TII->get(X86::DBG_VALUE)), AM).
8640 addImm(MI->getOperand(1).getImm()).
8641 addMetadata(MI->getOperand(2).getMetadata());
8642 F->DeleteMachineInstr(MI); // Remove pseudo.
8646 // String/text processing lowering.
8647 case X86::PCMPISTRM128REG:
8648 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8649 case X86::PCMPISTRM128MEM:
8650 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8651 case X86::PCMPESTRM128REG:
8652 return EmitPCMP(MI, BB, 5, false /* in mem */);
8653 case X86::PCMPESTRM128MEM:
8654 return EmitPCMP(MI, BB, 5, true /* in mem */);
8657 case X86::ATOMAND32:
8658 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8659 X86::AND32ri, X86::MOV32rm,
8660 X86::LCMPXCHG32, X86::MOV32rr,
8661 X86::NOT32r, X86::EAX,
8662 X86::GR32RegisterClass);
8664 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8665 X86::OR32ri, X86::MOV32rm,
8666 X86::LCMPXCHG32, X86::MOV32rr,
8667 X86::NOT32r, X86::EAX,
8668 X86::GR32RegisterClass);
8669 case X86::ATOMXOR32:
8670 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8671 X86::XOR32ri, X86::MOV32rm,
8672 X86::LCMPXCHG32, X86::MOV32rr,
8673 X86::NOT32r, X86::EAX,
8674 X86::GR32RegisterClass);
8675 case X86::ATOMNAND32:
8676 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8677 X86::AND32ri, X86::MOV32rm,
8678 X86::LCMPXCHG32, X86::MOV32rr,
8679 X86::NOT32r, X86::EAX,
8680 X86::GR32RegisterClass, true);
8681 case X86::ATOMMIN32:
8682 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8683 case X86::ATOMMAX32:
8684 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8685 case X86::ATOMUMIN32:
8686 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8687 case X86::ATOMUMAX32:
8688 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8690 case X86::ATOMAND16:
8691 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8692 X86::AND16ri, X86::MOV16rm,
8693 X86::LCMPXCHG16, X86::MOV16rr,
8694 X86::NOT16r, X86::AX,
8695 X86::GR16RegisterClass);
8697 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8698 X86::OR16ri, X86::MOV16rm,
8699 X86::LCMPXCHG16, X86::MOV16rr,
8700 X86::NOT16r, X86::AX,
8701 X86::GR16RegisterClass);
8702 case X86::ATOMXOR16:
8703 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8704 X86::XOR16ri, X86::MOV16rm,
8705 X86::LCMPXCHG16, X86::MOV16rr,
8706 X86::NOT16r, X86::AX,
8707 X86::GR16RegisterClass);
8708 case X86::ATOMNAND16:
8709 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8710 X86::AND16ri, X86::MOV16rm,
8711 X86::LCMPXCHG16, X86::MOV16rr,
8712 X86::NOT16r, X86::AX,
8713 X86::GR16RegisterClass, true);
8714 case X86::ATOMMIN16:
8715 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8716 case X86::ATOMMAX16:
8717 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8718 case X86::ATOMUMIN16:
8719 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8720 case X86::ATOMUMAX16:
8721 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8724 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8725 X86::AND8ri, X86::MOV8rm,
8726 X86::LCMPXCHG8, X86::MOV8rr,
8727 X86::NOT8r, X86::AL,
8728 X86::GR8RegisterClass);
8730 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8731 X86::OR8ri, X86::MOV8rm,
8732 X86::LCMPXCHG8, X86::MOV8rr,
8733 X86::NOT8r, X86::AL,
8734 X86::GR8RegisterClass);
8736 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8737 X86::XOR8ri, X86::MOV8rm,
8738 X86::LCMPXCHG8, X86::MOV8rr,
8739 X86::NOT8r, X86::AL,
8740 X86::GR8RegisterClass);
8741 case X86::ATOMNAND8:
8742 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8743 X86::AND8ri, X86::MOV8rm,
8744 X86::LCMPXCHG8, X86::MOV8rr,
8745 X86::NOT8r, X86::AL,
8746 X86::GR8RegisterClass, true);
8747 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8748 // This group is for 64-bit host.
8749 case X86::ATOMAND64:
8750 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8751 X86::AND64ri32, X86::MOV64rm,
8752 X86::LCMPXCHG64, X86::MOV64rr,
8753 X86::NOT64r, X86::RAX,
8754 X86::GR64RegisterClass);
8756 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8757 X86::OR64ri32, X86::MOV64rm,
8758 X86::LCMPXCHG64, X86::MOV64rr,
8759 X86::NOT64r, X86::RAX,
8760 X86::GR64RegisterClass);
8761 case X86::ATOMXOR64:
8762 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8763 X86::XOR64ri32, X86::MOV64rm,
8764 X86::LCMPXCHG64, X86::MOV64rr,
8765 X86::NOT64r, X86::RAX,
8766 X86::GR64RegisterClass);
8767 case X86::ATOMNAND64:
8768 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8769 X86::AND64ri32, X86::MOV64rm,
8770 X86::LCMPXCHG64, X86::MOV64rr,
8771 X86::NOT64r, X86::RAX,
8772 X86::GR64RegisterClass, true);
8773 case X86::ATOMMIN64:
8774 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8775 case X86::ATOMMAX64:
8776 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8777 case X86::ATOMUMIN64:
8778 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8779 case X86::ATOMUMAX64:
8780 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8782 // This group does 64-bit operations on a 32-bit host.
8783 case X86::ATOMAND6432:
8784 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8785 X86::AND32rr, X86::AND32rr,
8786 X86::AND32ri, X86::AND32ri,
8788 case X86::ATOMOR6432:
8789 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8790 X86::OR32rr, X86::OR32rr,
8791 X86::OR32ri, X86::OR32ri,
8793 case X86::ATOMXOR6432:
8794 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8795 X86::XOR32rr, X86::XOR32rr,
8796 X86::XOR32ri, X86::XOR32ri,
8798 case X86::ATOMNAND6432:
8799 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8800 X86::AND32rr, X86::AND32rr,
8801 X86::AND32ri, X86::AND32ri,
8803 case X86::ATOMADD6432:
8804 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8805 X86::ADD32rr, X86::ADC32rr,
8806 X86::ADD32ri, X86::ADC32ri,
8808 case X86::ATOMSUB6432:
8809 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8810 X86::SUB32rr, X86::SBB32rr,
8811 X86::SUB32ri, X86::SBB32ri,
8813 case X86::ATOMSWAP6432:
8814 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8815 X86::MOV32rr, X86::MOV32rr,
8816 X86::MOV32ri, X86::MOV32ri,
8818 case X86::VASTART_SAVE_XMM_REGS:
8819 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8823 //===----------------------------------------------------------------------===//
8824 // X86 Optimization Hooks
8825 //===----------------------------------------------------------------------===//
8827 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8831 const SelectionDAG &DAG,
8832 unsigned Depth) const {
8833 unsigned Opc = Op.getOpcode();
8834 assert((Opc >= ISD::BUILTIN_OP_END ||
8835 Opc == ISD::INTRINSIC_WO_CHAIN ||
8836 Opc == ISD::INTRINSIC_W_CHAIN ||
8837 Opc == ISD::INTRINSIC_VOID) &&
8838 "Should use MaskedValueIsZero if you don't know whether Op"
8839 " is a target node!");
8841 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8853 // These nodes' second result is a boolean.
8854 if (Op.getResNo() == 0)
8858 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8859 Mask.getBitWidth() - 1);
8864 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8865 /// node is a GlobalAddress + offset.
8866 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8867 GlobalValue* &GA, int64_t &Offset) const{
8868 if (N->getOpcode() == X86ISD::Wrapper) {
8869 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8870 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8871 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8875 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8878 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8879 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8880 /// if the load addresses are consecutive, non-overlapping, and in the right
8882 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8883 const TargetLowering &TLI) {
8884 DebugLoc dl = N->getDebugLoc();
8885 EVT VT = N->getValueType(0);
8886 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8888 if (VT.getSizeInBits() != 128)
8891 SmallVector<SDValue, 16> Elts;
8892 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8893 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8895 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8898 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8899 /// and convert it from being a bunch of shuffles and extracts to a simple
8900 /// store and scalar loads to extract the elements.
8901 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8902 const TargetLowering &TLI) {
8903 SDValue InputVector = N->getOperand(0);
8905 // Only operate on vectors of 4 elements, where the alternative shuffling
8906 // gets to be more expensive.
8907 if (InputVector.getValueType() != MVT::v4i32)
8910 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8911 // single use which is a sign-extend or zero-extend, and all elements are
8913 SmallVector<SDNode *, 4> Uses;
8914 unsigned ExtractedElements = 0;
8915 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8916 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8917 if (UI.getUse().getResNo() != InputVector.getResNo())
8920 SDNode *Extract = *UI;
8921 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8924 if (Extract->getValueType(0) != MVT::i32)
8926 if (!Extract->hasOneUse())
8928 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
8929 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
8931 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
8934 // Record which element was extracted.
8935 ExtractedElements |=
8936 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
8938 Uses.push_back(Extract);
8941 // If not all the elements were used, this may not be worthwhile.
8942 if (ExtractedElements != 15)
8945 // Ok, we've now decided to do the transformation.
8946 DebugLoc dl = InputVector.getDebugLoc();
8948 // Store the value to a temporary stack slot.
8949 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
8950 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
8953 // Replace each use (extract) with a load of the appropriate element.
8954 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
8955 UE = Uses.end(); UI != UE; ++UI) {
8956 SDNode *Extract = *UI;
8958 // Compute the element's address.
8959 SDValue Idx = Extract->getOperand(1);
8961 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
8962 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
8963 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
8965 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
8968 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
8969 NULL, 0, false, false, 0);
8971 // Replace the exact with the load.
8972 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
8975 // The replacement was made in place; don't return anything.
8979 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8980 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8981 const X86Subtarget *Subtarget) {
8982 DebugLoc DL = N->getDebugLoc();
8983 SDValue Cond = N->getOperand(0);
8984 // Get the LHS/RHS of the select.
8985 SDValue LHS = N->getOperand(1);
8986 SDValue RHS = N->getOperand(2);
8988 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8989 // instructions match the semantics of the common C idiom x<y?x:y but not
8990 // x<=y?x:y, because of how they handle negative zero (which can be
8991 // ignored in unsafe-math mode).
8992 if (Subtarget->hasSSE2() &&
8993 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8994 Cond.getOpcode() == ISD::SETCC) {
8995 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8997 unsigned Opcode = 0;
8998 // Check for x CC y ? x : y.
8999 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9000 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9004 // Converting this to a min would handle NaNs incorrectly, and swapping
9005 // the operands would cause it to handle comparisons between positive
9006 // and negative zero incorrectly.
9007 if (!FiniteOnlyFPMath() &&
9008 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9009 if (!UnsafeFPMath &&
9010 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9012 std::swap(LHS, RHS);
9014 Opcode = X86ISD::FMIN;
9017 // Converting this to a min would handle comparisons between positive
9018 // and negative zero incorrectly.
9019 if (!UnsafeFPMath &&
9020 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9022 Opcode = X86ISD::FMIN;
9025 // Converting this to a min would handle both negative zeros and NaNs
9026 // incorrectly, but we can swap the operands to fix both.
9027 std::swap(LHS, RHS);
9031 Opcode = X86ISD::FMIN;
9035 // Converting this to a max would handle comparisons between positive
9036 // and negative zero incorrectly.
9037 if (!UnsafeFPMath &&
9038 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9040 Opcode = X86ISD::FMAX;
9043 // Converting this to a max would handle NaNs incorrectly, and swapping
9044 // the operands would cause it to handle comparisons between positive
9045 // and negative zero incorrectly.
9046 if (!FiniteOnlyFPMath() &&
9047 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9048 if (!UnsafeFPMath &&
9049 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9051 std::swap(LHS, RHS);
9053 Opcode = X86ISD::FMAX;
9056 // Converting this to a max would handle both negative zeros and NaNs
9057 // incorrectly, but we can swap the operands to fix both.
9058 std::swap(LHS, RHS);
9062 Opcode = X86ISD::FMAX;
9065 // Check for x CC y ? y : x -- a min/max with reversed arms.
9066 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9067 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9071 // Converting this to a min would handle comparisons between positive
9072 // and negative zero incorrectly, and swapping the operands would
9073 // cause it to handle NaNs incorrectly.
9074 if (!UnsafeFPMath &&
9075 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9076 if (!FiniteOnlyFPMath() &&
9077 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9079 std::swap(LHS, RHS);
9081 Opcode = X86ISD::FMIN;
9084 // Converting this to a min would handle NaNs incorrectly.
9085 if (!UnsafeFPMath &&
9086 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9088 Opcode = X86ISD::FMIN;
9091 // Converting this to a min would handle both negative zeros and NaNs
9092 // incorrectly, but we can swap the operands to fix both.
9093 std::swap(LHS, RHS);
9097 Opcode = X86ISD::FMIN;
9101 // Converting this to a max would handle NaNs incorrectly.
9102 if (!FiniteOnlyFPMath() &&
9103 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9105 Opcode = X86ISD::FMAX;
9108 // Converting this to a max would handle comparisons between positive
9109 // and negative zero incorrectly, and swapping the operands would
9110 // cause it to handle NaNs incorrectly.
9111 if (!UnsafeFPMath &&
9112 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9113 if (!FiniteOnlyFPMath() &&
9114 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9116 std::swap(LHS, RHS);
9118 Opcode = X86ISD::FMAX;
9121 // Converting this to a max would handle both negative zeros and NaNs
9122 // incorrectly, but we can swap the operands to fix both.
9123 std::swap(LHS, RHS);
9127 Opcode = X86ISD::FMAX;
9133 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9136 // If this is a select between two integer constants, try to do some
9138 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9139 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9140 // Don't do this for crazy integer types.
9141 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9142 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9143 // so that TrueC (the true value) is larger than FalseC.
9144 bool NeedsCondInvert = false;
9146 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9147 // Efficiently invertible.
9148 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9149 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9150 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9151 NeedsCondInvert = true;
9152 std::swap(TrueC, FalseC);
9155 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9156 if (FalseC->getAPIntValue() == 0 &&
9157 TrueC->getAPIntValue().isPowerOf2()) {
9158 if (NeedsCondInvert) // Invert the condition if needed.
9159 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9160 DAG.getConstant(1, Cond.getValueType()));
9162 // Zero extend the condition if needed.
9163 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9165 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9166 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9167 DAG.getConstant(ShAmt, MVT::i8));
9170 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9171 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9172 if (NeedsCondInvert) // Invert the condition if needed.
9173 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9174 DAG.getConstant(1, Cond.getValueType()));
9176 // Zero extend the condition if needed.
9177 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9178 FalseC->getValueType(0), Cond);
9179 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9180 SDValue(FalseC, 0));
9183 // Optimize cases that will turn into an LEA instruction. This requires
9184 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9185 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9186 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9187 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9189 bool isFastMultiplier = false;
9191 switch ((unsigned char)Diff) {
9193 case 1: // result = add base, cond
9194 case 2: // result = lea base( , cond*2)
9195 case 3: // result = lea base(cond, cond*2)
9196 case 4: // result = lea base( , cond*4)
9197 case 5: // result = lea base(cond, cond*4)
9198 case 8: // result = lea base( , cond*8)
9199 case 9: // result = lea base(cond, cond*8)
9200 isFastMultiplier = true;
9205 if (isFastMultiplier) {
9206 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9207 if (NeedsCondInvert) // Invert the condition if needed.
9208 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9209 DAG.getConstant(1, Cond.getValueType()));
9211 // Zero extend the condition if needed.
9212 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9214 // Scale the condition by the difference.
9216 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9217 DAG.getConstant(Diff, Cond.getValueType()));
9219 // Add the base if non-zero.
9220 if (FalseC->getAPIntValue() != 0)
9221 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9222 SDValue(FalseC, 0));
9232 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9233 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9234 TargetLowering::DAGCombinerInfo &DCI) {
9235 DebugLoc DL = N->getDebugLoc();
9237 // If the flag operand isn't dead, don't touch this CMOV.
9238 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9241 // If this is a select between two integer constants, try to do some
9242 // optimizations. Note that the operands are ordered the opposite of SELECT
9244 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9245 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9246 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9247 // larger than FalseC (the false value).
9248 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9250 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9251 CC = X86::GetOppositeBranchCondition(CC);
9252 std::swap(TrueC, FalseC);
9255 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9256 // This is efficient for any integer data type (including i8/i16) and
9258 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9259 SDValue Cond = N->getOperand(3);
9260 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9261 DAG.getConstant(CC, MVT::i8), Cond);
9263 // Zero extend the condition if needed.
9264 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9266 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9267 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9268 DAG.getConstant(ShAmt, MVT::i8));
9269 if (N->getNumValues() == 2) // Dead flag value?
9270 return DCI.CombineTo(N, Cond, SDValue());
9274 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9275 // for any integer data type, including i8/i16.
9276 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9277 SDValue Cond = N->getOperand(3);
9278 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9279 DAG.getConstant(CC, MVT::i8), Cond);
9281 // Zero extend the condition if needed.
9282 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9283 FalseC->getValueType(0), Cond);
9284 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9285 SDValue(FalseC, 0));
9287 if (N->getNumValues() == 2) // Dead flag value?
9288 return DCI.CombineTo(N, Cond, SDValue());
9292 // Optimize cases that will turn into an LEA instruction. This requires
9293 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9294 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9295 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9296 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9298 bool isFastMultiplier = false;
9300 switch ((unsigned char)Diff) {
9302 case 1: // result = add base, cond
9303 case 2: // result = lea base( , cond*2)
9304 case 3: // result = lea base(cond, cond*2)
9305 case 4: // result = lea base( , cond*4)
9306 case 5: // result = lea base(cond, cond*4)
9307 case 8: // result = lea base( , cond*8)
9308 case 9: // result = lea base(cond, cond*8)
9309 isFastMultiplier = true;
9314 if (isFastMultiplier) {
9315 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9316 SDValue Cond = N->getOperand(3);
9317 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9318 DAG.getConstant(CC, MVT::i8), Cond);
9319 // Zero extend the condition if needed.
9320 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9322 // Scale the condition by the difference.
9324 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9325 DAG.getConstant(Diff, Cond.getValueType()));
9327 // Add the base if non-zero.
9328 if (FalseC->getAPIntValue() != 0)
9329 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9330 SDValue(FalseC, 0));
9331 if (N->getNumValues() == 2) // Dead flag value?
9332 return DCI.CombineTo(N, Cond, SDValue());
9342 /// PerformMulCombine - Optimize a single multiply with constant into two
9343 /// in order to implement it with two cheaper instructions, e.g.
9344 /// LEA + SHL, LEA + LEA.
9345 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9346 TargetLowering::DAGCombinerInfo &DCI) {
9347 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9350 EVT VT = N->getValueType(0);
9354 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9357 uint64_t MulAmt = C->getZExtValue();
9358 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9361 uint64_t MulAmt1 = 0;
9362 uint64_t MulAmt2 = 0;
9363 if ((MulAmt % 9) == 0) {
9365 MulAmt2 = MulAmt / 9;
9366 } else if ((MulAmt % 5) == 0) {
9368 MulAmt2 = MulAmt / 5;
9369 } else if ((MulAmt % 3) == 0) {
9371 MulAmt2 = MulAmt / 3;
9374 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9375 DebugLoc DL = N->getDebugLoc();
9377 if (isPowerOf2_64(MulAmt2) &&
9378 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9379 // If second multiplifer is pow2, issue it first. We want the multiply by
9380 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9382 std::swap(MulAmt1, MulAmt2);
9385 if (isPowerOf2_64(MulAmt1))
9386 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9387 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9389 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9390 DAG.getConstant(MulAmt1, VT));
9392 if (isPowerOf2_64(MulAmt2))
9393 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9394 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9396 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9397 DAG.getConstant(MulAmt2, VT));
9399 // Do not add new nodes to DAG combiner worklist.
9400 DCI.CombineTo(N, NewMul, false);
9405 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9406 SDValue N0 = N->getOperand(0);
9407 SDValue N1 = N->getOperand(1);
9408 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9409 EVT VT = N0.getValueType();
9411 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9412 // since the result of setcc_c is all zero's or all ones.
9413 if (N1C && N0.getOpcode() == ISD::AND &&
9414 N0.getOperand(1).getOpcode() == ISD::Constant) {
9415 SDValue N00 = N0.getOperand(0);
9416 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9417 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9418 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9419 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9420 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9421 APInt ShAmt = N1C->getAPIntValue();
9422 Mask = Mask.shl(ShAmt);
9424 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9425 N00, DAG.getConstant(Mask, VT));
9432 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9434 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9435 const X86Subtarget *Subtarget) {
9436 EVT VT = N->getValueType(0);
9437 if (!VT.isVector() && VT.isInteger() &&
9438 N->getOpcode() == ISD::SHL)
9439 return PerformSHLCombine(N, DAG);
9441 // On X86 with SSE2 support, we can transform this to a vector shift if
9442 // all elements are shifted by the same amount. We can't do this in legalize
9443 // because the a constant vector is typically transformed to a constant pool
9444 // so we have no knowledge of the shift amount.
9445 if (!Subtarget->hasSSE2())
9448 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9451 SDValue ShAmtOp = N->getOperand(1);
9452 EVT EltVT = VT.getVectorElementType();
9453 DebugLoc DL = N->getDebugLoc();
9454 SDValue BaseShAmt = SDValue();
9455 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9456 unsigned NumElts = VT.getVectorNumElements();
9458 for (; i != NumElts; ++i) {
9459 SDValue Arg = ShAmtOp.getOperand(i);
9460 if (Arg.getOpcode() == ISD::UNDEF) continue;
9464 for (; i != NumElts; ++i) {
9465 SDValue Arg = ShAmtOp.getOperand(i);
9466 if (Arg.getOpcode() == ISD::UNDEF) continue;
9467 if (Arg != BaseShAmt) {
9471 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9472 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9473 SDValue InVec = ShAmtOp.getOperand(0);
9474 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9475 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9477 for (; i != NumElts; ++i) {
9478 SDValue Arg = InVec.getOperand(i);
9479 if (Arg.getOpcode() == ISD::UNDEF) continue;
9483 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9484 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9485 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9486 if (C->getZExtValue() == SplatIdx)
9487 BaseShAmt = InVec.getOperand(1);
9490 if (BaseShAmt.getNode() == 0)
9491 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9492 DAG.getIntPtrConstant(0));
9496 // The shift amount is an i32.
9497 if (EltVT.bitsGT(MVT::i32))
9498 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9499 else if (EltVT.bitsLT(MVT::i32))
9500 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9502 // The shift amount is identical so we can do a vector shift.
9503 SDValue ValOp = N->getOperand(0);
9504 switch (N->getOpcode()) {
9506 llvm_unreachable("Unknown shift opcode!");
9509 if (VT == MVT::v2i64)
9510 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9511 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9513 if (VT == MVT::v4i32)
9514 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9515 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9517 if (VT == MVT::v8i16)
9518 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9519 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9523 if (VT == MVT::v4i32)
9524 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9525 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9527 if (VT == MVT::v8i16)
9528 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9529 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9533 if (VT == MVT::v2i64)
9534 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9535 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9537 if (VT == MVT::v4i32)
9538 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9539 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9541 if (VT == MVT::v8i16)
9542 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9543 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9550 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9551 const X86Subtarget *Subtarget) {
9552 EVT VT = N->getValueType(0);
9553 if (VT != MVT::i64 || !Subtarget->is64Bit())
9556 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9557 SDValue N0 = N->getOperand(0);
9558 SDValue N1 = N->getOperand(1);
9559 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9561 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9564 SDValue ShAmt0 = N0.getOperand(1);
9565 if (ShAmt0.getValueType() != MVT::i8)
9567 SDValue ShAmt1 = N1.getOperand(1);
9568 if (ShAmt1.getValueType() != MVT::i8)
9570 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9571 ShAmt0 = ShAmt0.getOperand(0);
9572 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9573 ShAmt1 = ShAmt1.getOperand(0);
9575 DebugLoc DL = N->getDebugLoc();
9576 unsigned Opc = X86ISD::SHLD;
9577 SDValue Op0 = N0.getOperand(0);
9578 SDValue Op1 = N1.getOperand(0);
9579 if (ShAmt0.getOpcode() == ISD::SUB) {
9581 std::swap(Op0, Op1);
9582 std::swap(ShAmt0, ShAmt1);
9585 if (ShAmt1.getOpcode() == ISD::SUB) {
9586 SDValue Sum = ShAmt1.getOperand(0);
9587 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9588 if (SumC->getSExtValue() == 64 &&
9589 ShAmt1.getOperand(1) == ShAmt0)
9590 return DAG.getNode(Opc, DL, VT,
9592 DAG.getNode(ISD::TRUNCATE, DL,
9595 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9596 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9598 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9599 return DAG.getNode(Opc, DL, VT,
9600 N0.getOperand(0), N1.getOperand(0),
9601 DAG.getNode(ISD::TRUNCATE, DL,
9608 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9609 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9610 const X86Subtarget *Subtarget) {
9611 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9612 // the FP state in cases where an emms may be missing.
9613 // A preferable solution to the general problem is to figure out the right
9614 // places to insert EMMS. This qualifies as a quick hack.
9616 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9617 StoreSDNode *St = cast<StoreSDNode>(N);
9618 EVT VT = St->getValue().getValueType();
9619 if (VT.getSizeInBits() != 64)
9622 const Function *F = DAG.getMachineFunction().getFunction();
9623 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9624 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9625 && Subtarget->hasSSE2();
9626 if ((VT.isVector() ||
9627 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9628 isa<LoadSDNode>(St->getValue()) &&
9629 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9630 St->getChain().hasOneUse() && !St->isVolatile()) {
9631 SDNode* LdVal = St->getValue().getNode();
9633 int TokenFactorIndex = -1;
9634 SmallVector<SDValue, 8> Ops;
9635 SDNode* ChainVal = St->getChain().getNode();
9636 // Must be a store of a load. We currently handle two cases: the load
9637 // is a direct child, and it's under an intervening TokenFactor. It is
9638 // possible to dig deeper under nested TokenFactors.
9639 if (ChainVal == LdVal)
9640 Ld = cast<LoadSDNode>(St->getChain());
9641 else if (St->getValue().hasOneUse() &&
9642 ChainVal->getOpcode() == ISD::TokenFactor) {
9643 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9644 if (ChainVal->getOperand(i).getNode() == LdVal) {
9645 TokenFactorIndex = i;
9646 Ld = cast<LoadSDNode>(St->getValue());
9648 Ops.push_back(ChainVal->getOperand(i));
9652 if (!Ld || !ISD::isNormalLoad(Ld))
9655 // If this is not the MMX case, i.e. we are just turning i64 load/store
9656 // into f64 load/store, avoid the transformation if there are multiple
9657 // uses of the loaded value.
9658 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9661 DebugLoc LdDL = Ld->getDebugLoc();
9662 DebugLoc StDL = N->getDebugLoc();
9663 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9664 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9666 if (Subtarget->is64Bit() || F64IsLegal) {
9667 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9668 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9669 Ld->getBasePtr(), Ld->getSrcValue(),
9670 Ld->getSrcValueOffset(), Ld->isVolatile(),
9671 Ld->isNonTemporal(), Ld->getAlignment());
9672 SDValue NewChain = NewLd.getValue(1);
9673 if (TokenFactorIndex != -1) {
9674 Ops.push_back(NewChain);
9675 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9678 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9679 St->getSrcValue(), St->getSrcValueOffset(),
9680 St->isVolatile(), St->isNonTemporal(),
9681 St->getAlignment());
9684 // Otherwise, lower to two pairs of 32-bit loads / stores.
9685 SDValue LoAddr = Ld->getBasePtr();
9686 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9687 DAG.getConstant(4, MVT::i32));
9689 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9690 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9691 Ld->isVolatile(), Ld->isNonTemporal(),
9692 Ld->getAlignment());
9693 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9694 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9695 Ld->isVolatile(), Ld->isNonTemporal(),
9696 MinAlign(Ld->getAlignment(), 4));
9698 SDValue NewChain = LoLd.getValue(1);
9699 if (TokenFactorIndex != -1) {
9700 Ops.push_back(LoLd);
9701 Ops.push_back(HiLd);
9702 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9706 LoAddr = St->getBasePtr();
9707 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9708 DAG.getConstant(4, MVT::i32));
9710 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9711 St->getSrcValue(), St->getSrcValueOffset(),
9712 St->isVolatile(), St->isNonTemporal(),
9713 St->getAlignment());
9714 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9716 St->getSrcValueOffset() + 4,
9718 St->isNonTemporal(),
9719 MinAlign(St->getAlignment(), 4));
9720 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9725 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9726 /// X86ISD::FXOR nodes.
9727 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9728 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9729 // F[X]OR(0.0, x) -> x
9730 // F[X]OR(x, 0.0) -> x
9731 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9732 if (C->getValueAPF().isPosZero())
9733 return N->getOperand(1);
9734 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9735 if (C->getValueAPF().isPosZero())
9736 return N->getOperand(0);
9740 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9741 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9742 // FAND(0.0, x) -> 0.0
9743 // FAND(x, 0.0) -> 0.0
9744 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9745 if (C->getValueAPF().isPosZero())
9746 return N->getOperand(0);
9747 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9748 if (C->getValueAPF().isPosZero())
9749 return N->getOperand(1);
9753 static SDValue PerformBTCombine(SDNode *N,
9755 TargetLowering::DAGCombinerInfo &DCI) {
9756 // BT ignores high bits in the bit index operand.
9757 SDValue Op1 = N->getOperand(1);
9758 if (Op1.hasOneUse()) {
9759 unsigned BitWidth = Op1.getValueSizeInBits();
9760 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9761 APInt KnownZero, KnownOne;
9762 TargetLowering::TargetLoweringOpt TLO(DAG);
9763 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9764 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9765 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9766 DCI.CommitTargetLoweringOpt(TLO);
9771 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9772 SDValue Op = N->getOperand(0);
9773 if (Op.getOpcode() == ISD::BIT_CONVERT)
9774 Op = Op.getOperand(0);
9775 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9776 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9777 VT.getVectorElementType().getSizeInBits() ==
9778 OpVT.getVectorElementType().getSizeInBits()) {
9779 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9784 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9785 // Locked instructions, in turn, have implicit fence semantics (all memory
9786 // operations are flushed before issuing the locked instruction, and the
9787 // are not buffered), so we can fold away the common pattern of
9788 // fence-atomic-fence.
9789 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9790 SDValue atomic = N->getOperand(0);
9791 switch (atomic.getOpcode()) {
9792 case ISD::ATOMIC_CMP_SWAP:
9793 case ISD::ATOMIC_SWAP:
9794 case ISD::ATOMIC_LOAD_ADD:
9795 case ISD::ATOMIC_LOAD_SUB:
9796 case ISD::ATOMIC_LOAD_AND:
9797 case ISD::ATOMIC_LOAD_OR:
9798 case ISD::ATOMIC_LOAD_XOR:
9799 case ISD::ATOMIC_LOAD_NAND:
9800 case ISD::ATOMIC_LOAD_MIN:
9801 case ISD::ATOMIC_LOAD_MAX:
9802 case ISD::ATOMIC_LOAD_UMIN:
9803 case ISD::ATOMIC_LOAD_UMAX:
9809 SDValue fence = atomic.getOperand(0);
9810 if (fence.getOpcode() != ISD::MEMBARRIER)
9813 switch (atomic.getOpcode()) {
9814 case ISD::ATOMIC_CMP_SWAP:
9815 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9816 atomic.getOperand(1), atomic.getOperand(2),
9817 atomic.getOperand(3));
9818 case ISD::ATOMIC_SWAP:
9819 case ISD::ATOMIC_LOAD_ADD:
9820 case ISD::ATOMIC_LOAD_SUB:
9821 case ISD::ATOMIC_LOAD_AND:
9822 case ISD::ATOMIC_LOAD_OR:
9823 case ISD::ATOMIC_LOAD_XOR:
9824 case ISD::ATOMIC_LOAD_NAND:
9825 case ISD::ATOMIC_LOAD_MIN:
9826 case ISD::ATOMIC_LOAD_MAX:
9827 case ISD::ATOMIC_LOAD_UMIN:
9828 case ISD::ATOMIC_LOAD_UMAX:
9829 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9830 atomic.getOperand(1), atomic.getOperand(2));
9836 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9837 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9838 // (and (i32 x86isd::setcc_carry), 1)
9839 // This eliminates the zext. This transformation is necessary because
9840 // ISD::SETCC is always legalized to i8.
9841 DebugLoc dl = N->getDebugLoc();
9842 SDValue N0 = N->getOperand(0);
9843 EVT VT = N->getValueType(0);
9844 if (N0.getOpcode() == ISD::AND &&
9846 N0.getOperand(0).hasOneUse()) {
9847 SDValue N00 = N0.getOperand(0);
9848 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9850 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9851 if (!C || C->getZExtValue() != 1)
9853 return DAG.getNode(ISD::AND, dl, VT,
9854 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9855 N00.getOperand(0), N00.getOperand(1)),
9856 DAG.getConstant(1, VT));
9862 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9863 DAGCombinerInfo &DCI) const {
9864 SelectionDAG &DAG = DCI.DAG;
9865 switch (N->getOpcode()) {
9867 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9868 case ISD::EXTRACT_VECTOR_ELT:
9869 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9870 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9871 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9872 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9875 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9876 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9877 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9879 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9880 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9881 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9882 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9883 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9884 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9890 //===----------------------------------------------------------------------===//
9891 // X86 Inline Assembly Support
9892 //===----------------------------------------------------------------------===//
9894 static bool LowerToBSwap(CallInst *CI) {
9895 // FIXME: this should verify that we are targetting a 486 or better. If not,
9896 // we will turn this bswap into something that will be lowered to logical ops
9897 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9898 // so don't worry about this.
9900 // Verify this is a simple bswap.
9901 if (CI->getNumOperands() != 2 ||
9902 CI->getType() != CI->getOperand(1)->getType() ||
9903 !CI->getType()->isIntegerTy())
9906 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9907 if (!Ty || Ty->getBitWidth() % 16 != 0)
9910 // Okay, we can do this xform, do so now.
9911 const Type *Tys[] = { Ty };
9912 Module *M = CI->getParent()->getParent()->getParent();
9913 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9915 Value *Op = CI->getOperand(1);
9916 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9918 CI->replaceAllUsesWith(Op);
9919 CI->eraseFromParent();
9923 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9924 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9925 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9927 std::string AsmStr = IA->getAsmString();
9929 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9930 SmallVector<StringRef, 4> AsmPieces;
9931 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9933 switch (AsmPieces.size()) {
9934 default: return false;
9936 AsmStr = AsmPieces[0];
9938 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9941 if (AsmPieces.size() == 2 &&
9942 (AsmPieces[0] == "bswap" ||
9943 AsmPieces[0] == "bswapq" ||
9944 AsmPieces[0] == "bswapl") &&
9945 (AsmPieces[1] == "$0" ||
9946 AsmPieces[1] == "${0:q}")) {
9947 // No need to check constraints, nothing other than the equivalent of
9948 // "=r,0" would be valid here.
9949 return LowerToBSwap(CI);
9951 // rorw $$8, ${0:w} --> llvm.bswap.i16
9952 if (CI->getType()->isIntegerTy(16) &&
9953 AsmPieces.size() == 3 &&
9954 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
9955 AsmPieces[1] == "$$8," &&
9956 AsmPieces[2] == "${0:w}" &&
9957 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
9959 const std::string &Constraints = IA->getConstraintString();
9960 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
9961 std::sort(AsmPieces.begin(), AsmPieces.end());
9962 if (AsmPieces.size() == 4 &&
9963 AsmPieces[0] == "~{cc}" &&
9964 AsmPieces[1] == "~{dirflag}" &&
9965 AsmPieces[2] == "~{flags}" &&
9966 AsmPieces[3] == "~{fpsr}") {
9967 return LowerToBSwap(CI);
9972 if (CI->getType()->isIntegerTy(64) &&
9973 Constraints.size() >= 2 &&
9974 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9975 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9976 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9977 SmallVector<StringRef, 4> Words;
9978 SplitString(AsmPieces[0], Words, " \t");
9979 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9981 SplitString(AsmPieces[1], Words, " \t");
9982 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9984 SplitString(AsmPieces[2], Words, " \t,");
9985 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9986 Words[2] == "%edx") {
9987 return LowerToBSwap(CI);
9999 /// getConstraintType - Given a constraint letter, return the type of
10000 /// constraint it is for this target.
10001 X86TargetLowering::ConstraintType
10002 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10003 if (Constraint.size() == 1) {
10004 switch (Constraint[0]) {
10016 return C_RegisterClass;
10024 return TargetLowering::getConstraintType(Constraint);
10027 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10028 /// with another that has more specific requirements based on the type of the
10029 /// corresponding operand.
10030 const char *X86TargetLowering::
10031 LowerXConstraint(EVT ConstraintVT) const {
10032 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10033 // 'f' like normal targets.
10034 if (ConstraintVT.isFloatingPoint()) {
10035 if (Subtarget->hasSSE2())
10037 if (Subtarget->hasSSE1())
10041 return TargetLowering::LowerXConstraint(ConstraintVT);
10044 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10045 /// vector. If it is invalid, don't add anything to Ops.
10046 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10049 std::vector<SDValue>&Ops,
10050 SelectionDAG &DAG) const {
10051 SDValue Result(0, 0);
10053 switch (Constraint) {
10056 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10057 if (C->getZExtValue() <= 31) {
10058 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10064 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10065 if (C->getZExtValue() <= 63) {
10066 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10072 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10073 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10074 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10080 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10081 if (C->getZExtValue() <= 255) {
10082 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10088 // 32-bit signed value
10089 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10090 const ConstantInt *CI = C->getConstantIntValue();
10091 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10092 C->getSExtValue())) {
10093 // Widen to 64 bits here to get it sign extended.
10094 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10097 // FIXME gcc accepts some relocatable values here too, but only in certain
10098 // memory models; it's complicated.
10103 // 32-bit unsigned value
10104 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10105 const ConstantInt *CI = C->getConstantIntValue();
10106 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10107 C->getZExtValue())) {
10108 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10112 // FIXME gcc accepts some relocatable values here too, but only in certain
10113 // memory models; it's complicated.
10117 // Literal immediates are always ok.
10118 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10119 // Widen to 64 bits here to get it sign extended.
10120 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10124 // If we are in non-pic codegen mode, we allow the address of a global (with
10125 // an optional displacement) to be used with 'i'.
10126 GlobalAddressSDNode *GA = 0;
10127 int64_t Offset = 0;
10129 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10131 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10132 Offset += GA->getOffset();
10134 } else if (Op.getOpcode() == ISD::ADD) {
10135 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10136 Offset += C->getZExtValue();
10137 Op = Op.getOperand(0);
10140 } else if (Op.getOpcode() == ISD::SUB) {
10141 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10142 Offset += -C->getZExtValue();
10143 Op = Op.getOperand(0);
10148 // Otherwise, this isn't something we can handle, reject it.
10152 GlobalValue *GV = GA->getGlobal();
10153 // If we require an extra load to get this address, as in PIC mode, we
10154 // can't accept it.
10155 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10156 getTargetMachine())))
10160 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
10162 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10168 if (Result.getNode()) {
10169 Ops.push_back(Result);
10172 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
10176 std::vector<unsigned> X86TargetLowering::
10177 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10179 if (Constraint.size() == 1) {
10180 // FIXME: not handling fp-stack yet!
10181 switch (Constraint[0]) { // GCC X86 Constraint Letters
10182 default: break; // Unknown constraint letter
10183 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10184 if (Subtarget->is64Bit()) {
10185 if (VT == MVT::i32)
10186 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10187 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10188 X86::R10D,X86::R11D,X86::R12D,
10189 X86::R13D,X86::R14D,X86::R15D,
10190 X86::EBP, X86::ESP, 0);
10191 else if (VT == MVT::i16)
10192 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10193 X86::SI, X86::DI, X86::R8W,X86::R9W,
10194 X86::R10W,X86::R11W,X86::R12W,
10195 X86::R13W,X86::R14W,X86::R15W,
10196 X86::BP, X86::SP, 0);
10197 else if (VT == MVT::i8)
10198 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10199 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10200 X86::R10B,X86::R11B,X86::R12B,
10201 X86::R13B,X86::R14B,X86::R15B,
10202 X86::BPL, X86::SPL, 0);
10204 else if (VT == MVT::i64)
10205 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10206 X86::RSI, X86::RDI, X86::R8, X86::R9,
10207 X86::R10, X86::R11, X86::R12,
10208 X86::R13, X86::R14, X86::R15,
10209 X86::RBP, X86::RSP, 0);
10213 // 32-bit fallthrough
10214 case 'Q': // Q_REGS
10215 if (VT == MVT::i32)
10216 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10217 else if (VT == MVT::i16)
10218 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10219 else if (VT == MVT::i8)
10220 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10221 else if (VT == MVT::i64)
10222 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10227 return std::vector<unsigned>();
10230 std::pair<unsigned, const TargetRegisterClass*>
10231 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10233 // First, see if this is a constraint that directly corresponds to an LLVM
10235 if (Constraint.size() == 1) {
10236 // GCC Constraint Letters
10237 switch (Constraint[0]) {
10239 case 'r': // GENERAL_REGS
10240 case 'l': // INDEX_REGS
10242 return std::make_pair(0U, X86::GR8RegisterClass);
10243 if (VT == MVT::i16)
10244 return std::make_pair(0U, X86::GR16RegisterClass);
10245 if (VT == MVT::i32 || !Subtarget->is64Bit())
10246 return std::make_pair(0U, X86::GR32RegisterClass);
10247 return std::make_pair(0U, X86::GR64RegisterClass);
10248 case 'R': // LEGACY_REGS
10250 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10251 if (VT == MVT::i16)
10252 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10253 if (VT == MVT::i32 || !Subtarget->is64Bit())
10254 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10255 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10256 case 'f': // FP Stack registers.
10257 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10258 // value to the correct fpstack register class.
10259 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10260 return std::make_pair(0U, X86::RFP32RegisterClass);
10261 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10262 return std::make_pair(0U, X86::RFP64RegisterClass);
10263 return std::make_pair(0U, X86::RFP80RegisterClass);
10264 case 'y': // MMX_REGS if MMX allowed.
10265 if (!Subtarget->hasMMX()) break;
10266 return std::make_pair(0U, X86::VR64RegisterClass);
10267 case 'Y': // SSE_REGS if SSE2 allowed
10268 if (!Subtarget->hasSSE2()) break;
10270 case 'x': // SSE_REGS if SSE1 allowed
10271 if (!Subtarget->hasSSE1()) break;
10273 switch (VT.getSimpleVT().SimpleTy) {
10275 // Scalar SSE types.
10278 return std::make_pair(0U, X86::FR32RegisterClass);
10281 return std::make_pair(0U, X86::FR64RegisterClass);
10289 return std::make_pair(0U, X86::VR128RegisterClass);
10295 // Use the default implementation in TargetLowering to convert the register
10296 // constraint into a member of a register class.
10297 std::pair<unsigned, const TargetRegisterClass*> Res;
10298 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10300 // Not found as a standard register?
10301 if (Res.second == 0) {
10302 // Map st(0) -> st(7) -> ST0
10303 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10304 tolower(Constraint[1]) == 's' &&
10305 tolower(Constraint[2]) == 't' &&
10306 Constraint[3] == '(' &&
10307 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10308 Constraint[5] == ')' &&
10309 Constraint[6] == '}') {
10311 Res.first = X86::ST0+Constraint[4]-'0';
10312 Res.second = X86::RFP80RegisterClass;
10316 // GCC allows "st(0)" to be called just plain "st".
10317 if (StringRef("{st}").equals_lower(Constraint)) {
10318 Res.first = X86::ST0;
10319 Res.second = X86::RFP80RegisterClass;
10324 if (StringRef("{flags}").equals_lower(Constraint)) {
10325 Res.first = X86::EFLAGS;
10326 Res.second = X86::CCRRegisterClass;
10330 // 'A' means EAX + EDX.
10331 if (Constraint == "A") {
10332 Res.first = X86::EAX;
10333 Res.second = X86::GR32_ADRegisterClass;
10339 // Otherwise, check to see if this is a register class of the wrong value
10340 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10341 // turn into {ax},{dx}.
10342 if (Res.second->hasType(VT))
10343 return Res; // Correct type already, nothing to do.
10345 // All of the single-register GCC register classes map their values onto
10346 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10347 // really want an 8-bit or 32-bit register, map to the appropriate register
10348 // class and return the appropriate register.
10349 if (Res.second == X86::GR16RegisterClass) {
10350 if (VT == MVT::i8) {
10351 unsigned DestReg = 0;
10352 switch (Res.first) {
10354 case X86::AX: DestReg = X86::AL; break;
10355 case X86::DX: DestReg = X86::DL; break;
10356 case X86::CX: DestReg = X86::CL; break;
10357 case X86::BX: DestReg = X86::BL; break;
10360 Res.first = DestReg;
10361 Res.second = X86::GR8RegisterClass;
10363 } else if (VT == MVT::i32) {
10364 unsigned DestReg = 0;
10365 switch (Res.first) {
10367 case X86::AX: DestReg = X86::EAX; break;
10368 case X86::DX: DestReg = X86::EDX; break;
10369 case X86::CX: DestReg = X86::ECX; break;
10370 case X86::BX: DestReg = X86::EBX; break;
10371 case X86::SI: DestReg = X86::ESI; break;
10372 case X86::DI: DestReg = X86::EDI; break;
10373 case X86::BP: DestReg = X86::EBP; break;
10374 case X86::SP: DestReg = X86::ESP; break;
10377 Res.first = DestReg;
10378 Res.second = X86::GR32RegisterClass;
10380 } else if (VT == MVT::i64) {
10381 unsigned DestReg = 0;
10382 switch (Res.first) {
10384 case X86::AX: DestReg = X86::RAX; break;
10385 case X86::DX: DestReg = X86::RDX; break;
10386 case X86::CX: DestReg = X86::RCX; break;
10387 case X86::BX: DestReg = X86::RBX; break;
10388 case X86::SI: DestReg = X86::RSI; break;
10389 case X86::DI: DestReg = X86::RDI; break;
10390 case X86::BP: DestReg = X86::RBP; break;
10391 case X86::SP: DestReg = X86::RSP; break;
10394 Res.first = DestReg;
10395 Res.second = X86::GR64RegisterClass;
10398 } else if (Res.second == X86::FR32RegisterClass ||
10399 Res.second == X86::FR64RegisterClass ||
10400 Res.second == X86::VR128RegisterClass) {
10401 // Handle references to XMM physical registers that got mapped into the
10402 // wrong class. This can happen with constraints like {xmm0} where the
10403 // target independent register mapper will just pick the first match it can
10404 // find, ignoring the required type.
10405 if (VT == MVT::f32)
10406 Res.second = X86::FR32RegisterClass;
10407 else if (VT == MVT::f64)
10408 Res.second = X86::FR64RegisterClass;
10409 else if (X86::VR128RegisterClass->hasType(VT))
10410 Res.second = X86::VR128RegisterClass;