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 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
798 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
801 // Do not attempt to promote non-128-bit vectors
802 if (!VT.is128BitVector()) {
805 setOperationAction(ISD::AND, SVT, Promote);
806 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
807 setOperationAction(ISD::OR, SVT, Promote);
808 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
809 setOperationAction(ISD::XOR, SVT, Promote);
810 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
811 setOperationAction(ISD::LOAD, SVT, Promote);
812 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
813 setOperationAction(ISD::SELECT, SVT, Promote);
814 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
817 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
819 // Custom lower v2i64 and v2f64 selects.
820 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
821 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
822 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
823 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
825 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
826 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
827 if (!DisableMMX && Subtarget->hasMMX()) {
828 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
829 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
833 if (Subtarget->hasSSE41()) {
834 // FIXME: Do we need to handle scalar-to-vector here?
835 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
837 // i8 and i16 vectors are custom , because the source register and source
838 // source memory operand types are not the same width. f32 vectors are
839 // custom since the immediate controlling the insert encodes additional
841 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
842 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
843 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
844 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
846 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
847 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
848 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
849 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
851 if (Subtarget->is64Bit()) {
852 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
853 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
857 if (Subtarget->hasSSE42()) {
858 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
861 if (!UseSoftFloat && Subtarget->hasAVX()) {
862 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
863 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
864 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
865 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
867 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
868 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
869 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
870 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
871 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
872 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
873 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
874 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
875 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
876 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
877 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
878 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
879 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
880 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
881 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
883 // Operations to consider commented out -v16i16 v32i8
884 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
885 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
886 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
887 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
888 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
889 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
890 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
891 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
892 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
893 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
894 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
895 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
896 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
897 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
899 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
900 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
901 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
902 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
904 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
905 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
906 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
907 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
908 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
910 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
911 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
912 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
913 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
914 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
915 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
918 // Not sure we want to do this since there are no 256-bit integer
921 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
922 // This includes 256-bit vectors
923 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
924 EVT VT = (MVT::SimpleValueType)i;
926 // Do not attempt to custom lower non-power-of-2 vectors
927 if (!isPowerOf2_32(VT.getVectorNumElements()))
930 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
931 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
932 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
935 if (Subtarget->is64Bit()) {
936 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
937 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
942 // Not sure we want to do this since there are no 256-bit integer
945 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
946 // Including 256-bit vectors
947 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
948 EVT VT = (MVT::SimpleValueType)i;
950 if (!VT.is256BitVector()) {
953 setOperationAction(ISD::AND, VT, Promote);
954 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
955 setOperationAction(ISD::OR, VT, Promote);
956 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
957 setOperationAction(ISD::XOR, VT, Promote);
958 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
959 setOperationAction(ISD::LOAD, VT, Promote);
960 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
961 setOperationAction(ISD::SELECT, VT, Promote);
962 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
965 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
969 // We want to custom lower some of our intrinsics.
970 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
972 // Add/Sub/Mul with overflow operations are custom lowered.
973 setOperationAction(ISD::SADDO, MVT::i32, Custom);
974 setOperationAction(ISD::SADDO, MVT::i64, Custom);
975 setOperationAction(ISD::UADDO, MVT::i32, Custom);
976 setOperationAction(ISD::UADDO, MVT::i64, Custom);
977 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
978 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
979 setOperationAction(ISD::USUBO, MVT::i32, Custom);
980 setOperationAction(ISD::USUBO, MVT::i64, Custom);
981 setOperationAction(ISD::SMULO, MVT::i32, Custom);
982 setOperationAction(ISD::SMULO, MVT::i64, Custom);
984 if (!Subtarget->is64Bit()) {
985 // These libcalls are not available in 32-bit.
986 setLibcallName(RTLIB::SHL_I128, 0);
987 setLibcallName(RTLIB::SRL_I128, 0);
988 setLibcallName(RTLIB::SRA_I128, 0);
991 // We have target-specific dag combine patterns for the following nodes:
992 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
993 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
994 setTargetDAGCombine(ISD::BUILD_VECTOR);
995 setTargetDAGCombine(ISD::SELECT);
996 setTargetDAGCombine(ISD::SHL);
997 setTargetDAGCombine(ISD::SRA);
998 setTargetDAGCombine(ISD::SRL);
999 setTargetDAGCombine(ISD::OR);
1000 setTargetDAGCombine(ISD::STORE);
1001 setTargetDAGCombine(ISD::MEMBARRIER);
1002 setTargetDAGCombine(ISD::ZERO_EXTEND);
1003 if (Subtarget->is64Bit())
1004 setTargetDAGCombine(ISD::MUL);
1006 computeRegisterProperties();
1008 // FIXME: These should be based on subtarget info. Plus, the values should
1009 // be smaller when we are in optimizing for size mode.
1010 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1011 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1012 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1013 setPrefLoopAlignment(16);
1014 benefitFromCodePlacementOpt = true;
1018 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1023 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1024 /// the desired ByVal argument alignment.
1025 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1028 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1029 if (VTy->getBitWidth() == 128)
1031 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1032 unsigned EltAlign = 0;
1033 getMaxByValAlign(ATy->getElementType(), EltAlign);
1034 if (EltAlign > MaxAlign)
1035 MaxAlign = EltAlign;
1036 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1037 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1038 unsigned EltAlign = 0;
1039 getMaxByValAlign(STy->getElementType(i), EltAlign);
1040 if (EltAlign > MaxAlign)
1041 MaxAlign = EltAlign;
1049 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1050 /// function arguments in the caller parameter area. For X86, aggregates
1051 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1052 /// are at 4-byte boundaries.
1053 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1054 if (Subtarget->is64Bit()) {
1055 // Max of 8 and alignment of type.
1056 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1063 if (Subtarget->hasSSE1())
1064 getMaxByValAlign(Ty, Align);
1068 /// getOptimalMemOpType - Returns the target specific optimal type for load
1069 /// and store operations as a result of memset, memcpy, and memmove
1070 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1073 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1074 bool isSrcConst, bool isSrcStr,
1075 SelectionDAG &DAG) const {
1076 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1077 // linux. This is because the stack realignment code can't handle certain
1078 // cases like PR2962. This should be removed when PR2962 is fixed.
1079 const Function *F = DAG.getMachineFunction().getFunction();
1080 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1081 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1082 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1084 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1087 if (Subtarget->is64Bit() && Size >= 8)
1092 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1093 /// current function. The returned value is a member of the
1094 /// MachineJumpTableInfo::JTEntryKind enum.
1095 unsigned X86TargetLowering::getJumpTableEncoding() const {
1096 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1098 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1099 Subtarget->isPICStyleGOT())
1100 return MachineJumpTableInfo::EK_Custom32;
1102 // Otherwise, use the normal jump table encoding heuristics.
1103 return TargetLowering::getJumpTableEncoding();
1106 /// getPICBaseSymbol - Return the X86-32 PIC base.
1108 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1109 MCContext &Ctx) const {
1110 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1111 return Ctx.GetOrCreateTemporarySymbol(Twine(MAI.getPrivateGlobalPrefix())+
1112 Twine(MF->getFunctionNumber())+"$pb");
1117 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1118 const MachineBasicBlock *MBB,
1119 unsigned uid,MCContext &Ctx) const{
1120 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1121 Subtarget->isPICStyleGOT());
1122 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1124 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1125 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1128 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1130 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1131 SelectionDAG &DAG) const {
1132 if (!Subtarget->is64Bit())
1133 // This doesn't have DebugLoc associated with it, but is not really the
1134 // same as a Register.
1135 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1140 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1141 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1143 const MCExpr *X86TargetLowering::
1144 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1145 MCContext &Ctx) const {
1146 // X86-64 uses RIP relative addressing based on the jump table label.
1147 if (Subtarget->isPICStyleRIPRel())
1148 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1150 // Otherwise, the reference is relative to the PIC base.
1151 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1154 /// getFunctionAlignment - Return the Log2 alignment of this function.
1155 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1156 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1159 //===----------------------------------------------------------------------===//
1160 // Return Value Calling Convention Implementation
1161 //===----------------------------------------------------------------------===//
1163 #include "X86GenCallingConv.inc"
1166 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1167 const SmallVectorImpl<EVT> &OutTys,
1168 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1169 SelectionDAG &DAG) {
1170 SmallVector<CCValAssign, 16> RVLocs;
1171 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1172 RVLocs, *DAG.getContext());
1173 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1177 X86TargetLowering::LowerReturn(SDValue Chain,
1178 CallingConv::ID CallConv, bool isVarArg,
1179 const SmallVectorImpl<ISD::OutputArg> &Outs,
1180 DebugLoc dl, SelectionDAG &DAG) {
1182 SmallVector<CCValAssign, 16> RVLocs;
1183 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1184 RVLocs, *DAG.getContext());
1185 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1187 // Add the regs to the liveout set for the function.
1188 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1189 for (unsigned i = 0; i != RVLocs.size(); ++i)
1190 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1191 MRI.addLiveOut(RVLocs[i].getLocReg());
1195 SmallVector<SDValue, 6> RetOps;
1196 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1197 // Operand #1 = Bytes To Pop
1198 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1200 // Copy the result values into the output registers.
1201 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1202 CCValAssign &VA = RVLocs[i];
1203 assert(VA.isRegLoc() && "Can only return in registers!");
1204 SDValue ValToCopy = Outs[i].Val;
1206 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1207 // the RET instruction and handled by the FP Stackifier.
1208 if (VA.getLocReg() == X86::ST0 ||
1209 VA.getLocReg() == X86::ST1) {
1210 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1211 // change the value to the FP stack register class.
1212 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1213 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1214 RetOps.push_back(ValToCopy);
1215 // Don't emit a copytoreg.
1219 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1220 // which is returned in RAX / RDX.
1221 if (Subtarget->is64Bit()) {
1222 EVT ValVT = ValToCopy.getValueType();
1223 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1224 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1225 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1226 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1230 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1231 Flag = Chain.getValue(1);
1234 // The x86-64 ABI for returning structs by value requires that we copy
1235 // the sret argument into %rax for the return. We saved the argument into
1236 // a virtual register in the entry block, so now we copy the value out
1238 if (Subtarget->is64Bit() &&
1239 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1240 MachineFunction &MF = DAG.getMachineFunction();
1241 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1242 unsigned Reg = FuncInfo->getSRetReturnReg();
1244 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1245 FuncInfo->setSRetReturnReg(Reg);
1247 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1249 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1250 Flag = Chain.getValue(1);
1252 // RAX now acts like a return value.
1253 MRI.addLiveOut(X86::RAX);
1256 RetOps[0] = Chain; // Update chain.
1258 // Add the flag if we have it.
1260 RetOps.push_back(Flag);
1262 return DAG.getNode(X86ISD::RET_FLAG, dl,
1263 MVT::Other, &RetOps[0], RetOps.size());
1266 /// LowerCallResult - Lower the result values of a call into the
1267 /// appropriate copies out of appropriate physical registers.
1270 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1271 CallingConv::ID CallConv, bool isVarArg,
1272 const SmallVectorImpl<ISD::InputArg> &Ins,
1273 DebugLoc dl, SelectionDAG &DAG,
1274 SmallVectorImpl<SDValue> &InVals) {
1276 // Assign locations to each value returned by this call.
1277 SmallVector<CCValAssign, 16> RVLocs;
1278 bool Is64Bit = Subtarget->is64Bit();
1279 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1280 RVLocs, *DAG.getContext());
1281 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1283 // Copy all of the result registers out of their specified physreg.
1284 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1285 CCValAssign &VA = RVLocs[i];
1286 EVT CopyVT = VA.getValVT();
1288 // If this is x86-64, and we disabled SSE, we can't return FP values
1289 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1290 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1291 llvm_report_error("SSE register return with SSE disabled");
1294 // If this is a call to a function that returns an fp value on the floating
1295 // point stack, but where we prefer to use the value in xmm registers, copy
1296 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1297 if ((VA.getLocReg() == X86::ST0 ||
1298 VA.getLocReg() == X86::ST1) &&
1299 isScalarFPTypeInSSEReg(VA.getValVT())) {
1304 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1305 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1306 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1307 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1308 MVT::v2i64, InFlag).getValue(1);
1309 Val = Chain.getValue(0);
1310 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1311 Val, DAG.getConstant(0, MVT::i64));
1313 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1314 MVT::i64, InFlag).getValue(1);
1315 Val = Chain.getValue(0);
1317 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1319 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1320 CopyVT, InFlag).getValue(1);
1321 Val = Chain.getValue(0);
1323 InFlag = Chain.getValue(2);
1325 if (CopyVT != VA.getValVT()) {
1326 // Round the F80 the right size, which also moves to the appropriate xmm
1328 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1329 // This truncation won't change the value.
1330 DAG.getIntPtrConstant(1));
1333 InVals.push_back(Val);
1340 //===----------------------------------------------------------------------===//
1341 // C & StdCall & Fast Calling Convention implementation
1342 //===----------------------------------------------------------------------===//
1343 // StdCall calling convention seems to be standard for many Windows' API
1344 // routines and around. It differs from C calling convention just a little:
1345 // callee should clean up the stack, not caller. Symbols should be also
1346 // decorated in some fancy way :) It doesn't support any vector arguments.
1347 // For info on fast calling convention see Fast Calling Convention (tail call)
1348 // implementation LowerX86_32FastCCCallTo.
1350 /// CallIsStructReturn - Determines whether a call uses struct return
1352 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1356 return Outs[0].Flags.isSRet();
1359 /// ArgsAreStructReturn - Determines whether a function uses struct
1360 /// return semantics.
1362 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1366 return Ins[0].Flags.isSRet();
1369 /// IsCalleePop - Determines whether the callee is required to pop its
1370 /// own arguments. Callee pop is necessary to support tail calls.
1371 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1375 switch (CallingConv) {
1378 case CallingConv::X86_StdCall:
1379 return !Subtarget->is64Bit();
1380 case CallingConv::X86_FastCall:
1381 return !Subtarget->is64Bit();
1382 case CallingConv::Fast:
1383 return GuaranteedTailCallOpt;
1384 case CallingConv::GHC:
1385 return GuaranteedTailCallOpt;
1389 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1390 /// given CallingConvention value.
1391 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1392 if (Subtarget->is64Bit()) {
1393 if (CC == CallingConv::GHC)
1394 return CC_X86_64_GHC;
1395 else if (Subtarget->isTargetWin64())
1396 return CC_X86_Win64_C;
1401 if (CC == CallingConv::X86_FastCall)
1402 return CC_X86_32_FastCall;
1403 else if (CC == CallingConv::Fast)
1404 return CC_X86_32_FastCC;
1405 else if (CC == CallingConv::GHC)
1406 return CC_X86_32_GHC;
1411 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1412 /// by "Src" to address "Dst" with size and alignment information specified by
1413 /// the specific parameter attribute. The copy will be passed as a byval
1414 /// function parameter.
1416 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1417 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1419 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1420 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1421 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1424 /// IsTailCallConvention - Return true if the calling convention is one that
1425 /// supports tail call optimization.
1426 static bool IsTailCallConvention(CallingConv::ID CC) {
1427 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1430 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1431 /// a tailcall target by changing its ABI.
1432 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1433 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1437 X86TargetLowering::LowerMemArgument(SDValue Chain,
1438 CallingConv::ID CallConv,
1439 const SmallVectorImpl<ISD::InputArg> &Ins,
1440 DebugLoc dl, SelectionDAG &DAG,
1441 const CCValAssign &VA,
1442 MachineFrameInfo *MFI,
1444 // Create the nodes corresponding to a load from this parameter slot.
1445 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1446 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1447 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1450 // If value is passed by pointer we have address passed instead of the value
1452 if (VA.getLocInfo() == CCValAssign::Indirect)
1453 ValVT = VA.getLocVT();
1455 ValVT = VA.getValVT();
1457 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1458 // changed with more analysis.
1459 // In case of tail call optimization mark all arguments mutable. Since they
1460 // could be overwritten by lowering of arguments in case of a tail call.
1461 if (Flags.isByVal()) {
1462 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1463 VA.getLocMemOffset(), isImmutable, false);
1464 return DAG.getFrameIndex(FI, getPointerTy());
1466 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1467 VA.getLocMemOffset(), isImmutable, false);
1468 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1469 return DAG.getLoad(ValVT, dl, Chain, FIN,
1470 PseudoSourceValue::getFixedStack(FI), 0,
1476 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1477 CallingConv::ID CallConv,
1479 const SmallVectorImpl<ISD::InputArg> &Ins,
1482 SmallVectorImpl<SDValue> &InVals) {
1483 MachineFunction &MF = DAG.getMachineFunction();
1484 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1486 const Function* Fn = MF.getFunction();
1487 if (Fn->hasExternalLinkage() &&
1488 Subtarget->isTargetCygMing() &&
1489 Fn->getName() == "main")
1490 FuncInfo->setForceFramePointer(true);
1492 MachineFrameInfo *MFI = MF.getFrameInfo();
1493 bool Is64Bit = Subtarget->is64Bit();
1494 bool IsWin64 = Subtarget->isTargetWin64();
1496 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1497 "Var args not supported with calling convention fastcc or ghc");
1499 // Assign locations to all of the incoming arguments.
1500 SmallVector<CCValAssign, 16> ArgLocs;
1501 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1502 ArgLocs, *DAG.getContext());
1503 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1505 unsigned LastVal = ~0U;
1507 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1508 CCValAssign &VA = ArgLocs[i];
1509 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1511 assert(VA.getValNo() != LastVal &&
1512 "Don't support value assigned to multiple locs yet");
1513 LastVal = VA.getValNo();
1515 if (VA.isRegLoc()) {
1516 EVT RegVT = VA.getLocVT();
1517 TargetRegisterClass *RC = NULL;
1518 if (RegVT == MVT::i32)
1519 RC = X86::GR32RegisterClass;
1520 else if (Is64Bit && RegVT == MVT::i64)
1521 RC = X86::GR64RegisterClass;
1522 else if (RegVT == MVT::f32)
1523 RC = X86::FR32RegisterClass;
1524 else if (RegVT == MVT::f64)
1525 RC = X86::FR64RegisterClass;
1526 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1527 RC = X86::VR128RegisterClass;
1528 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1529 RC = X86::VR64RegisterClass;
1531 llvm_unreachable("Unknown argument type!");
1533 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1534 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1536 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1537 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1539 if (VA.getLocInfo() == CCValAssign::SExt)
1540 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1541 DAG.getValueType(VA.getValVT()));
1542 else if (VA.getLocInfo() == CCValAssign::ZExt)
1543 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1544 DAG.getValueType(VA.getValVT()));
1545 else if (VA.getLocInfo() == CCValAssign::BCvt)
1546 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1548 if (VA.isExtInLoc()) {
1549 // Handle MMX values passed in XMM regs.
1550 if (RegVT.isVector()) {
1551 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1552 ArgValue, DAG.getConstant(0, MVT::i64));
1553 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1555 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1558 assert(VA.isMemLoc());
1559 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1562 // If value is passed via pointer - do a load.
1563 if (VA.getLocInfo() == CCValAssign::Indirect)
1564 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1567 InVals.push_back(ArgValue);
1570 // The x86-64 ABI for returning structs by value requires that we copy
1571 // the sret argument into %rax for the return. Save the argument into
1572 // a virtual register so that we can access it from the return points.
1573 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1574 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1575 unsigned Reg = FuncInfo->getSRetReturnReg();
1577 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1578 FuncInfo->setSRetReturnReg(Reg);
1580 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1581 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1584 unsigned StackSize = CCInfo.getNextStackOffset();
1585 // Align stack specially for tail calls.
1586 if (FuncIsMadeTailCallSafe(CallConv))
1587 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1589 // If the function takes variable number of arguments, make a frame index for
1590 // the start of the first vararg value... for expansion of llvm.va_start.
1592 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1593 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1596 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1598 // FIXME: We should really autogenerate these arrays
1599 static const unsigned GPR64ArgRegsWin64[] = {
1600 X86::RCX, X86::RDX, X86::R8, X86::R9
1602 static const unsigned XMMArgRegsWin64[] = {
1603 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1605 static const unsigned GPR64ArgRegs64Bit[] = {
1606 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1608 static const unsigned XMMArgRegs64Bit[] = {
1609 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1610 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1612 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1615 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1616 GPR64ArgRegs = GPR64ArgRegsWin64;
1617 XMMArgRegs = XMMArgRegsWin64;
1619 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1620 GPR64ArgRegs = GPR64ArgRegs64Bit;
1621 XMMArgRegs = XMMArgRegs64Bit;
1623 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1625 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1628 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1629 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1630 "SSE register cannot be used when SSE is disabled!");
1631 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1632 "SSE register cannot be used when SSE is disabled!");
1633 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1634 // Kernel mode asks for SSE to be disabled, so don't push them
1636 TotalNumXMMRegs = 0;
1638 // For X86-64, if there are vararg parameters that are passed via
1639 // registers, then we must store them to their spots on the stack so they
1640 // may be loaded by deferencing the result of va_next.
1641 VarArgsGPOffset = NumIntRegs * 8;
1642 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1643 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1644 TotalNumXMMRegs * 16, 16,
1647 // Store the integer parameter registers.
1648 SmallVector<SDValue, 8> MemOps;
1649 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1650 unsigned Offset = VarArgsGPOffset;
1651 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1652 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1653 DAG.getIntPtrConstant(Offset));
1654 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1655 X86::GR64RegisterClass);
1656 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1658 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1659 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1660 Offset, false, false, 0);
1661 MemOps.push_back(Store);
1665 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1666 // Now store the XMM (fp + vector) parameter registers.
1667 SmallVector<SDValue, 11> SaveXMMOps;
1668 SaveXMMOps.push_back(Chain);
1670 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1671 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1672 SaveXMMOps.push_back(ALVal);
1674 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1675 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1677 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1678 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1679 X86::VR128RegisterClass);
1680 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1681 SaveXMMOps.push_back(Val);
1683 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1685 &SaveXMMOps[0], SaveXMMOps.size()));
1688 if (!MemOps.empty())
1689 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1690 &MemOps[0], MemOps.size());
1694 // Some CCs need callee pop.
1695 if (IsCalleePop(isVarArg, CallConv)) {
1696 BytesToPopOnReturn = StackSize; // Callee pops everything.
1698 BytesToPopOnReturn = 0; // Callee pops nothing.
1699 // If this is an sret function, the return should pop the hidden pointer.
1700 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1701 BytesToPopOnReturn = 4;
1705 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1706 if (CallConv == CallingConv::X86_FastCall)
1707 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1710 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1716 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1717 SDValue StackPtr, SDValue Arg,
1718 DebugLoc dl, SelectionDAG &DAG,
1719 const CCValAssign &VA,
1720 ISD::ArgFlagsTy Flags) {
1721 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1722 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1723 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1724 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1725 if (Flags.isByVal()) {
1726 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1728 return DAG.getStore(Chain, dl, Arg, PtrOff,
1729 PseudoSourceValue::getStack(), LocMemOffset,
1733 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1734 /// optimization is performed and it is required.
1736 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1737 SDValue &OutRetAddr, SDValue Chain,
1738 bool IsTailCall, bool Is64Bit,
1739 int FPDiff, DebugLoc dl) {
1740 // Adjust the Return address stack slot.
1741 EVT VT = getPointerTy();
1742 OutRetAddr = getReturnAddressFrameIndex(DAG);
1744 // Load the "old" Return address.
1745 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1746 return SDValue(OutRetAddr.getNode(), 1);
1749 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1750 /// optimization is performed and it is required (FPDiff!=0).
1752 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1753 SDValue Chain, SDValue RetAddrFrIdx,
1754 bool Is64Bit, int FPDiff, DebugLoc dl) {
1755 // Store the return address to the appropriate stack slot.
1756 if (!FPDiff) return Chain;
1757 // Calculate the new stack slot for the return address.
1758 int SlotSize = Is64Bit ? 8 : 4;
1759 int NewReturnAddrFI =
1760 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1761 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1762 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1763 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1764 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1770 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1771 CallingConv::ID CallConv, bool isVarArg,
1773 const SmallVectorImpl<ISD::OutputArg> &Outs,
1774 const SmallVectorImpl<ISD::InputArg> &Ins,
1775 DebugLoc dl, SelectionDAG &DAG,
1776 SmallVectorImpl<SDValue> &InVals) {
1777 MachineFunction &MF = DAG.getMachineFunction();
1778 bool Is64Bit = Subtarget->is64Bit();
1779 bool IsStructRet = CallIsStructReturn(Outs);
1780 bool IsSibcall = false;
1783 // Check if it's really possible to do a tail call.
1784 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1785 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1788 // Sibcalls are automatically detected tailcalls which do not require
1790 if (!GuaranteedTailCallOpt && isTailCall)
1797 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1798 "Var args not supported with calling convention fastcc or ghc");
1800 // Analyze operands of the call, assigning locations to each operand.
1801 SmallVector<CCValAssign, 16> ArgLocs;
1802 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1803 ArgLocs, *DAG.getContext());
1804 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1806 // Get a count of how many bytes are to be pushed on the stack.
1807 unsigned NumBytes = CCInfo.getNextStackOffset();
1809 // This is a sibcall. The memory operands are available in caller's
1810 // own caller's stack.
1812 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1813 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1816 if (isTailCall && !IsSibcall) {
1817 // Lower arguments at fp - stackoffset + fpdiff.
1818 unsigned NumBytesCallerPushed =
1819 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1820 FPDiff = NumBytesCallerPushed - NumBytes;
1822 // Set the delta of movement of the returnaddr stackslot.
1823 // But only set if delta is greater than previous delta.
1824 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1825 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1829 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1831 SDValue RetAddrFrIdx;
1832 // Load return adress for tail calls.
1833 if (isTailCall && FPDiff)
1834 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1835 Is64Bit, FPDiff, dl);
1837 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1838 SmallVector<SDValue, 8> MemOpChains;
1841 // Walk the register/memloc assignments, inserting copies/loads. In the case
1842 // of tail call optimization arguments are handle later.
1843 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1844 CCValAssign &VA = ArgLocs[i];
1845 EVT RegVT = VA.getLocVT();
1846 SDValue Arg = Outs[i].Val;
1847 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1848 bool isByVal = Flags.isByVal();
1850 // Promote the value if needed.
1851 switch (VA.getLocInfo()) {
1852 default: llvm_unreachable("Unknown loc info!");
1853 case CCValAssign::Full: break;
1854 case CCValAssign::SExt:
1855 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1857 case CCValAssign::ZExt:
1858 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1860 case CCValAssign::AExt:
1861 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1862 // Special case: passing MMX values in XMM registers.
1863 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1864 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1865 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1867 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1869 case CCValAssign::BCvt:
1870 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1872 case CCValAssign::Indirect: {
1873 // Store the argument.
1874 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1875 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1876 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1877 PseudoSourceValue::getFixedStack(FI), 0,
1884 if (VA.isRegLoc()) {
1885 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1886 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1887 assert(VA.isMemLoc());
1888 if (StackPtr.getNode() == 0)
1889 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1890 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1891 dl, DAG, VA, Flags));
1895 if (!MemOpChains.empty())
1896 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1897 &MemOpChains[0], MemOpChains.size());
1899 // Build a sequence of copy-to-reg nodes chained together with token chain
1900 // and flag operands which copy the outgoing args into registers.
1902 // Tail call byval lowering might overwrite argument registers so in case of
1903 // tail call optimization the copies to registers are lowered later.
1905 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1906 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1907 RegsToPass[i].second, InFlag);
1908 InFlag = Chain.getValue(1);
1911 if (Subtarget->isPICStyleGOT()) {
1912 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1915 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1916 DAG.getNode(X86ISD::GlobalBaseReg,
1917 DebugLoc::getUnknownLoc(),
1920 InFlag = Chain.getValue(1);
1922 // If we are tail calling and generating PIC/GOT style code load the
1923 // address of the callee into ECX. The value in ecx is used as target of
1924 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1925 // for tail calls on PIC/GOT architectures. Normally we would just put the
1926 // address of GOT into ebx and then call target@PLT. But for tail calls
1927 // ebx would be restored (since ebx is callee saved) before jumping to the
1930 // Note: The actual moving to ECX is done further down.
1931 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1932 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1933 !G->getGlobal()->hasProtectedVisibility())
1934 Callee = LowerGlobalAddress(Callee, DAG);
1935 else if (isa<ExternalSymbolSDNode>(Callee))
1936 Callee = LowerExternalSymbol(Callee, DAG);
1940 if (Is64Bit && isVarArg) {
1941 // From AMD64 ABI document:
1942 // For calls that may call functions that use varargs or stdargs
1943 // (prototype-less calls or calls to functions containing ellipsis (...) in
1944 // the declaration) %al is used as hidden argument to specify the number
1945 // of SSE registers used. The contents of %al do not need to match exactly
1946 // the number of registers, but must be an ubound on the number of SSE
1947 // registers used and is in the range 0 - 8 inclusive.
1949 // FIXME: Verify this on Win64
1950 // Count the number of XMM registers allocated.
1951 static const unsigned XMMArgRegs[] = {
1952 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1953 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1955 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1956 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1957 && "SSE registers cannot be used when SSE is disabled");
1959 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1960 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1961 InFlag = Chain.getValue(1);
1965 // For tail calls lower the arguments to the 'real' stack slot.
1967 // Force all the incoming stack arguments to be loaded from the stack
1968 // before any new outgoing arguments are stored to the stack, because the
1969 // outgoing stack slots may alias the incoming argument stack slots, and
1970 // the alias isn't otherwise explicit. This is slightly more conservative
1971 // than necessary, because it means that each store effectively depends
1972 // on every argument instead of just those arguments it would clobber.
1973 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1975 SmallVector<SDValue, 8> MemOpChains2;
1978 // Do not flag preceeding copytoreg stuff together with the following stuff.
1980 if (GuaranteedTailCallOpt) {
1981 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1982 CCValAssign &VA = ArgLocs[i];
1985 assert(VA.isMemLoc());
1986 SDValue Arg = Outs[i].Val;
1987 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1988 // Create frame index.
1989 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1990 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1991 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1992 FIN = DAG.getFrameIndex(FI, getPointerTy());
1994 if (Flags.isByVal()) {
1995 // Copy relative to framepointer.
1996 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1997 if (StackPtr.getNode() == 0)
1998 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2000 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2002 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2006 // Store relative to framepointer.
2007 MemOpChains2.push_back(
2008 DAG.getStore(ArgChain, dl, Arg, FIN,
2009 PseudoSourceValue::getFixedStack(FI), 0,
2015 if (!MemOpChains2.empty())
2016 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2017 &MemOpChains2[0], MemOpChains2.size());
2019 // Copy arguments to their registers.
2020 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2021 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2022 RegsToPass[i].second, InFlag);
2023 InFlag = Chain.getValue(1);
2027 // Store the return address to the appropriate stack slot.
2028 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2032 bool WasGlobalOrExternal = false;
2033 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2034 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2035 // In the 64-bit large code model, we have to make all calls
2036 // through a register, since the call instruction's 32-bit
2037 // pc-relative offset may not be large enough to hold the whole
2039 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2040 WasGlobalOrExternal = true;
2041 // If the callee is a GlobalAddress node (quite common, every direct call
2042 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2045 // We should use extra load for direct calls to dllimported functions in
2047 GlobalValue *GV = G->getGlobal();
2048 if (!GV->hasDLLImportLinkage()) {
2049 unsigned char OpFlags = 0;
2051 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2052 // external symbols most go through the PLT in PIC mode. If the symbol
2053 // has hidden or protected visibility, or if it is static or local, then
2054 // we don't need to use the PLT - we can directly call it.
2055 if (Subtarget->isTargetELF() &&
2056 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2057 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2058 OpFlags = X86II::MO_PLT;
2059 } else if (Subtarget->isPICStyleStubAny() &&
2060 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2061 Subtarget->getDarwinVers() < 9) {
2062 // PC-relative references to external symbols should go through $stub,
2063 // unless we're building with the leopard linker or later, which
2064 // automatically synthesizes these stubs.
2065 OpFlags = X86II::MO_DARWIN_STUB;
2068 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2069 G->getOffset(), OpFlags);
2071 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2072 WasGlobalOrExternal = true;
2073 unsigned char OpFlags = 0;
2075 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2076 // symbols should go through the PLT.
2077 if (Subtarget->isTargetELF() &&
2078 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2079 OpFlags = X86II::MO_PLT;
2080 } else if (Subtarget->isPICStyleStubAny() &&
2081 Subtarget->getDarwinVers() < 9) {
2082 // PC-relative references to external symbols should go through $stub,
2083 // unless we're building with the leopard linker or later, which
2084 // automatically synthesizes these stubs.
2085 OpFlags = X86II::MO_DARWIN_STUB;
2088 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2092 // Returns a chain & a flag for retval copy to use.
2093 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2094 SmallVector<SDValue, 8> Ops;
2096 if (!IsSibcall && isTailCall) {
2097 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2098 DAG.getIntPtrConstant(0, true), InFlag);
2099 InFlag = Chain.getValue(1);
2102 Ops.push_back(Chain);
2103 Ops.push_back(Callee);
2106 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2108 // Add argument registers to the end of the list so that they are known live
2110 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2111 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2112 RegsToPass[i].second.getValueType()));
2114 // Add an implicit use GOT pointer in EBX.
2115 if (!isTailCall && Subtarget->isPICStyleGOT())
2116 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2118 // Add an implicit use of AL for x86 vararg functions.
2119 if (Is64Bit && isVarArg)
2120 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2122 if (InFlag.getNode())
2123 Ops.push_back(InFlag);
2126 // If this is the first return lowered for this function, add the regs
2127 // to the liveout set for the function.
2128 if (MF.getRegInfo().liveout_empty()) {
2129 SmallVector<CCValAssign, 16> RVLocs;
2130 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2132 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2133 for (unsigned i = 0; i != RVLocs.size(); ++i)
2134 if (RVLocs[i].isRegLoc())
2135 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2137 return DAG.getNode(X86ISD::TC_RETURN, dl,
2138 NodeTys, &Ops[0], Ops.size());
2141 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2142 InFlag = Chain.getValue(1);
2144 // Create the CALLSEQ_END node.
2145 unsigned NumBytesForCalleeToPush;
2146 if (IsCalleePop(isVarArg, CallConv))
2147 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2148 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2149 // If this is a call to a struct-return function, the callee
2150 // pops the hidden struct pointer, so we have to push it back.
2151 // This is common for Darwin/X86, Linux & Mingw32 targets.
2152 NumBytesForCalleeToPush = 4;
2154 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2156 // Returns a flag for retval copy to use.
2158 Chain = DAG.getCALLSEQ_END(Chain,
2159 DAG.getIntPtrConstant(NumBytes, true),
2160 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2163 InFlag = Chain.getValue(1);
2166 // Handle result values, copying them out of physregs into vregs that we
2168 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2169 Ins, dl, DAG, InVals);
2173 //===----------------------------------------------------------------------===//
2174 // Fast Calling Convention (tail call) implementation
2175 //===----------------------------------------------------------------------===//
2177 // Like std call, callee cleans arguments, convention except that ECX is
2178 // reserved for storing the tail called function address. Only 2 registers are
2179 // free for argument passing (inreg). Tail call optimization is performed
2181 // * tailcallopt is enabled
2182 // * caller/callee are fastcc
2183 // On X86_64 architecture with GOT-style position independent code only local
2184 // (within module) calls are supported at the moment.
2185 // To keep the stack aligned according to platform abi the function
2186 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2187 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2188 // If a tail called function callee has more arguments than the caller the
2189 // caller needs to make sure that there is room to move the RETADDR to. This is
2190 // achieved by reserving an area the size of the argument delta right after the
2191 // original REtADDR, but before the saved framepointer or the spilled registers
2192 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2204 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2205 /// for a 16 byte align requirement.
2206 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2207 SelectionDAG& DAG) {
2208 MachineFunction &MF = DAG.getMachineFunction();
2209 const TargetMachine &TM = MF.getTarget();
2210 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2211 unsigned StackAlignment = TFI.getStackAlignment();
2212 uint64_t AlignMask = StackAlignment - 1;
2213 int64_t Offset = StackSize;
2214 uint64_t SlotSize = TD->getPointerSize();
2215 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2216 // Number smaller than 12 so just add the difference.
2217 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2219 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2220 Offset = ((~AlignMask) & Offset) + StackAlignment +
2221 (StackAlignment-SlotSize);
2226 /// MatchingStackOffset - Return true if the given stack call argument is
2227 /// already available in the same position (relatively) of the caller's
2228 /// incoming argument stack.
2230 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2231 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2232 const X86InstrInfo *TII) {
2233 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2235 if (Arg.getOpcode() == ISD::CopyFromReg) {
2236 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2237 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2239 MachineInstr *Def = MRI->getVRegDef(VR);
2242 if (!Flags.isByVal()) {
2243 if (!TII->isLoadFromStackSlot(Def, FI))
2246 unsigned Opcode = Def->getOpcode();
2247 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2248 Def->getOperand(1).isFI()) {
2249 FI = Def->getOperand(1).getIndex();
2250 Bytes = Flags.getByValSize();
2254 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2255 if (Flags.isByVal())
2256 // ByVal argument is passed in as a pointer but it's now being
2257 // dereferenced. e.g.
2258 // define @foo(%struct.X* %A) {
2259 // tail call @bar(%struct.X* byval %A)
2262 SDValue Ptr = Ld->getBasePtr();
2263 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2266 FI = FINode->getIndex();
2270 assert(FI != INT_MAX);
2271 if (!MFI->isFixedObjectIndex(FI))
2273 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2276 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2277 /// for tail call optimization. Targets which want to do tail call
2278 /// optimization should implement this function.
2280 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2281 CallingConv::ID CalleeCC,
2283 bool isCalleeStructRet,
2284 bool isCallerStructRet,
2285 const SmallVectorImpl<ISD::OutputArg> &Outs,
2286 const SmallVectorImpl<ISD::InputArg> &Ins,
2287 SelectionDAG& DAG) const {
2288 if (!IsTailCallConvention(CalleeCC) &&
2289 CalleeCC != CallingConv::C)
2292 // If -tailcallopt is specified, make fastcc functions tail-callable.
2293 const Function *CallerF = DAG.getMachineFunction().getFunction();
2294 if (GuaranteedTailCallOpt) {
2295 if (IsTailCallConvention(CalleeCC) &&
2296 CallerF->getCallingConv() == CalleeCC)
2301 // Look for obvious safe cases to perform tail call optimization that does not
2302 // requite ABI changes. This is what gcc calls sibcall.
2304 // Do not sibcall optimize vararg calls for now.
2308 // Also avoid sibcall optimization if either caller or callee uses struct
2309 // return semantics.
2310 if (isCalleeStructRet || isCallerStructRet)
2313 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2314 // Therefore if it's not used by the call it is not safe to optimize this into
2316 bool Unused = false;
2317 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2324 SmallVector<CCValAssign, 16> RVLocs;
2325 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2326 RVLocs, *DAG.getContext());
2327 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2328 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2329 CCValAssign &VA = RVLocs[i];
2330 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2335 // If the callee takes no arguments then go on to check the results of the
2337 if (!Outs.empty()) {
2338 // Check if stack adjustment is needed. For now, do not do this if any
2339 // argument is passed on the stack.
2340 SmallVector<CCValAssign, 16> ArgLocs;
2341 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2342 ArgLocs, *DAG.getContext());
2343 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2344 if (CCInfo.getNextStackOffset()) {
2345 MachineFunction &MF = DAG.getMachineFunction();
2346 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2348 if (Subtarget->isTargetWin64())
2349 // Win64 ABI has additional complications.
2352 // Check if the arguments are already laid out in the right way as
2353 // the caller's fixed stack objects.
2354 MachineFrameInfo *MFI = MF.getFrameInfo();
2355 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2356 const X86InstrInfo *TII =
2357 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2358 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2359 CCValAssign &VA = ArgLocs[i];
2360 EVT RegVT = VA.getLocVT();
2361 SDValue Arg = Outs[i].Val;
2362 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2363 if (VA.getLocInfo() == CCValAssign::Indirect)
2365 if (!VA.isRegLoc()) {
2366 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2378 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2380 DenseMap<const Value *, unsigned> &vm,
2381 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2382 DenseMap<const AllocaInst *, int> &am
2384 , SmallSet<Instruction*, 8> &cil
2387 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2395 //===----------------------------------------------------------------------===//
2396 // Other Lowering Hooks
2397 //===----------------------------------------------------------------------===//
2400 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2401 MachineFunction &MF = DAG.getMachineFunction();
2402 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2403 int ReturnAddrIndex = FuncInfo->getRAIndex();
2405 if (ReturnAddrIndex == 0) {
2406 // Set up a frame object for the return address.
2407 uint64_t SlotSize = TD->getPointerSize();
2408 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2410 FuncInfo->setRAIndex(ReturnAddrIndex);
2413 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2417 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2418 bool hasSymbolicDisplacement) {
2419 // Offset should fit into 32 bit immediate field.
2420 if (!isInt32(Offset))
2423 // If we don't have a symbolic displacement - we don't have any extra
2425 if (!hasSymbolicDisplacement)
2428 // FIXME: Some tweaks might be needed for medium code model.
2429 if (M != CodeModel::Small && M != CodeModel::Kernel)
2432 // For small code model we assume that latest object is 16MB before end of 31
2433 // bits boundary. We may also accept pretty large negative constants knowing
2434 // that all objects are in the positive half of address space.
2435 if (M == CodeModel::Small && Offset < 16*1024*1024)
2438 // For kernel code model we know that all object resist in the negative half
2439 // of 32bits address space. We may not accept negative offsets, since they may
2440 // be just off and we may accept pretty large positive ones.
2441 if (M == CodeModel::Kernel && Offset > 0)
2447 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2448 /// specific condition code, returning the condition code and the LHS/RHS of the
2449 /// comparison to make.
2450 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2451 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2453 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2454 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2455 // X > -1 -> X == 0, jump !sign.
2456 RHS = DAG.getConstant(0, RHS.getValueType());
2457 return X86::COND_NS;
2458 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2459 // X < 0 -> X == 0, jump on sign.
2461 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2463 RHS = DAG.getConstant(0, RHS.getValueType());
2464 return X86::COND_LE;
2468 switch (SetCCOpcode) {
2469 default: llvm_unreachable("Invalid integer condition!");
2470 case ISD::SETEQ: return X86::COND_E;
2471 case ISD::SETGT: return X86::COND_G;
2472 case ISD::SETGE: return X86::COND_GE;
2473 case ISD::SETLT: return X86::COND_L;
2474 case ISD::SETLE: return X86::COND_LE;
2475 case ISD::SETNE: return X86::COND_NE;
2476 case ISD::SETULT: return X86::COND_B;
2477 case ISD::SETUGT: return X86::COND_A;
2478 case ISD::SETULE: return X86::COND_BE;
2479 case ISD::SETUGE: return X86::COND_AE;
2483 // First determine if it is required or is profitable to flip the operands.
2485 // If LHS is a foldable load, but RHS is not, flip the condition.
2486 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2487 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2488 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2489 std::swap(LHS, RHS);
2492 switch (SetCCOpcode) {
2498 std::swap(LHS, RHS);
2502 // On a floating point condition, the flags are set as follows:
2504 // 0 | 0 | 0 | X > Y
2505 // 0 | 0 | 1 | X < Y
2506 // 1 | 0 | 0 | X == Y
2507 // 1 | 1 | 1 | unordered
2508 switch (SetCCOpcode) {
2509 default: llvm_unreachable("Condcode should be pre-legalized away");
2511 case ISD::SETEQ: return X86::COND_E;
2512 case ISD::SETOLT: // flipped
2514 case ISD::SETGT: return X86::COND_A;
2515 case ISD::SETOLE: // flipped
2517 case ISD::SETGE: return X86::COND_AE;
2518 case ISD::SETUGT: // flipped
2520 case ISD::SETLT: return X86::COND_B;
2521 case ISD::SETUGE: // flipped
2523 case ISD::SETLE: return X86::COND_BE;
2525 case ISD::SETNE: return X86::COND_NE;
2526 case ISD::SETUO: return X86::COND_P;
2527 case ISD::SETO: return X86::COND_NP;
2529 case ISD::SETUNE: return X86::COND_INVALID;
2533 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2534 /// code. Current x86 isa includes the following FP cmov instructions:
2535 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2536 static bool hasFPCMov(unsigned X86CC) {
2552 /// isFPImmLegal - Returns true if the target can instruction select the
2553 /// specified FP immediate natively. If false, the legalizer will
2554 /// materialize the FP immediate as a load from a constant pool.
2555 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2556 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2557 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2563 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2564 /// the specified range (L, H].
2565 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2566 return (Val < 0) || (Val >= Low && Val < Hi);
2569 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2570 /// specified value.
2571 static bool isUndefOrEqual(int Val, int CmpVal) {
2572 if (Val < 0 || Val == CmpVal)
2577 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2578 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2579 /// the second operand.
2580 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2581 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2582 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2583 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2584 return (Mask[0] < 2 && Mask[1] < 2);
2588 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2589 SmallVector<int, 8> M;
2591 return ::isPSHUFDMask(M, N->getValueType(0));
2594 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2595 /// is suitable for input to PSHUFHW.
2596 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2597 if (VT != MVT::v8i16)
2600 // Lower quadword copied in order or undef.
2601 for (int i = 0; i != 4; ++i)
2602 if (Mask[i] >= 0 && Mask[i] != i)
2605 // Upper quadword shuffled.
2606 for (int i = 4; i != 8; ++i)
2607 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2613 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2614 SmallVector<int, 8> M;
2616 return ::isPSHUFHWMask(M, N->getValueType(0));
2619 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2620 /// is suitable for input to PSHUFLW.
2621 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2622 if (VT != MVT::v8i16)
2625 // Upper quadword copied in order.
2626 for (int i = 4; i != 8; ++i)
2627 if (Mask[i] >= 0 && Mask[i] != i)
2630 // Lower quadword shuffled.
2631 for (int i = 0; i != 4; ++i)
2638 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2639 SmallVector<int, 8> M;
2641 return ::isPSHUFLWMask(M, N->getValueType(0));
2644 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2645 /// is suitable for input to PALIGNR.
2646 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2648 int i, e = VT.getVectorNumElements();
2650 // Do not handle v2i64 / v2f64 shuffles with palignr.
2651 if (e < 4 || !hasSSSE3)
2654 for (i = 0; i != e; ++i)
2658 // All undef, not a palignr.
2662 // Determine if it's ok to perform a palignr with only the LHS, since we
2663 // don't have access to the actual shuffle elements to see if RHS is undef.
2664 bool Unary = Mask[i] < (int)e;
2665 bool NeedsUnary = false;
2667 int s = Mask[i] - i;
2669 // Check the rest of the elements to see if they are consecutive.
2670 for (++i; i != e; ++i) {
2675 Unary = Unary && (m < (int)e);
2676 NeedsUnary = NeedsUnary || (m < s);
2678 if (NeedsUnary && !Unary)
2680 if (Unary && m != ((s+i) & (e-1)))
2682 if (!Unary && m != (s+i))
2688 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2689 SmallVector<int, 8> M;
2691 return ::isPALIGNRMask(M, N->getValueType(0), true);
2694 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2695 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2696 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2697 int NumElems = VT.getVectorNumElements();
2698 if (NumElems != 2 && NumElems != 4)
2701 int Half = NumElems / 2;
2702 for (int i = 0; i < Half; ++i)
2703 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2705 for (int i = Half; i < NumElems; ++i)
2706 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2712 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2713 SmallVector<int, 8> M;
2715 return ::isSHUFPMask(M, N->getValueType(0));
2718 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2719 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2720 /// half elements to come from vector 1 (which would equal the dest.) and
2721 /// the upper half to come from vector 2.
2722 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2723 int NumElems = VT.getVectorNumElements();
2725 if (NumElems != 2 && NumElems != 4)
2728 int Half = NumElems / 2;
2729 for (int i = 0; i < Half; ++i)
2730 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2732 for (int i = Half; i < NumElems; ++i)
2733 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2738 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2739 SmallVector<int, 8> M;
2741 return isCommutedSHUFPMask(M, N->getValueType(0));
2744 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2745 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2746 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2747 if (N->getValueType(0).getVectorNumElements() != 4)
2750 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2751 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2752 isUndefOrEqual(N->getMaskElt(1), 7) &&
2753 isUndefOrEqual(N->getMaskElt(2), 2) &&
2754 isUndefOrEqual(N->getMaskElt(3), 3);
2757 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2758 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2760 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2761 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2766 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2767 isUndefOrEqual(N->getMaskElt(1), 3) &&
2768 isUndefOrEqual(N->getMaskElt(2), 2) &&
2769 isUndefOrEqual(N->getMaskElt(3), 3);
2772 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2773 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2774 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2775 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2777 if (NumElems != 2 && NumElems != 4)
2780 for (unsigned i = 0; i < NumElems/2; ++i)
2781 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2784 for (unsigned i = NumElems/2; i < NumElems; ++i)
2785 if (!isUndefOrEqual(N->getMaskElt(i), i))
2791 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2792 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2793 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2794 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2796 if (NumElems != 2 && NumElems != 4)
2799 for (unsigned i = 0; i < NumElems/2; ++i)
2800 if (!isUndefOrEqual(N->getMaskElt(i), i))
2803 for (unsigned i = 0; i < NumElems/2; ++i)
2804 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2810 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2811 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2812 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2813 bool V2IsSplat = false) {
2814 int NumElts = VT.getVectorNumElements();
2815 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2818 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2820 int BitI1 = Mask[i+1];
2821 if (!isUndefOrEqual(BitI, j))
2824 if (!isUndefOrEqual(BitI1, NumElts))
2827 if (!isUndefOrEqual(BitI1, j + NumElts))
2834 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2835 SmallVector<int, 8> M;
2837 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2840 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2841 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2842 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2843 bool V2IsSplat = false) {
2844 int NumElts = VT.getVectorNumElements();
2845 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2848 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2850 int BitI1 = Mask[i+1];
2851 if (!isUndefOrEqual(BitI, j + NumElts/2))
2854 if (isUndefOrEqual(BitI1, NumElts))
2857 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2864 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2865 SmallVector<int, 8> M;
2867 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2870 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2871 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2873 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2874 int NumElems = VT.getVectorNumElements();
2875 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2878 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2880 int BitI1 = Mask[i+1];
2881 if (!isUndefOrEqual(BitI, j))
2883 if (!isUndefOrEqual(BitI1, j))
2889 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2890 SmallVector<int, 8> M;
2892 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2895 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2896 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2898 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2899 int NumElems = VT.getVectorNumElements();
2900 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2903 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2905 int BitI1 = Mask[i+1];
2906 if (!isUndefOrEqual(BitI, j))
2908 if (!isUndefOrEqual(BitI1, j))
2914 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2915 SmallVector<int, 8> M;
2917 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2920 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2921 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2922 /// MOVSD, and MOVD, i.e. setting the lowest element.
2923 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2924 if (VT.getVectorElementType().getSizeInBits() < 32)
2927 int NumElts = VT.getVectorNumElements();
2929 if (!isUndefOrEqual(Mask[0], NumElts))
2932 for (int i = 1; i < NumElts; ++i)
2933 if (!isUndefOrEqual(Mask[i], i))
2939 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2940 SmallVector<int, 8> M;
2942 return ::isMOVLMask(M, N->getValueType(0));
2945 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2946 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2947 /// element of vector 2 and the other elements to come from vector 1 in order.
2948 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2949 bool V2IsSplat = false, bool V2IsUndef = false) {
2950 int NumOps = VT.getVectorNumElements();
2951 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2954 if (!isUndefOrEqual(Mask[0], 0))
2957 for (int i = 1; i < NumOps; ++i)
2958 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2959 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2960 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2966 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2967 bool V2IsUndef = false) {
2968 SmallVector<int, 8> M;
2970 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2973 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2974 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2975 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2976 if (N->getValueType(0).getVectorNumElements() != 4)
2979 // Expect 1, 1, 3, 3
2980 for (unsigned i = 0; i < 2; ++i) {
2981 int Elt = N->getMaskElt(i);
2982 if (Elt >= 0 && Elt != 1)
2987 for (unsigned i = 2; i < 4; ++i) {
2988 int Elt = N->getMaskElt(i);
2989 if (Elt >= 0 && Elt != 3)
2994 // Don't use movshdup if it can be done with a shufps.
2995 // FIXME: verify that matching u, u, 3, 3 is what we want.
2999 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3000 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3001 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3002 if (N->getValueType(0).getVectorNumElements() != 4)
3005 // Expect 0, 0, 2, 2
3006 for (unsigned i = 0; i < 2; ++i)
3007 if (N->getMaskElt(i) > 0)
3011 for (unsigned i = 2; i < 4; ++i) {
3012 int Elt = N->getMaskElt(i);
3013 if (Elt >= 0 && Elt != 2)
3018 // Don't use movsldup if it can be done with a shufps.
3022 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3023 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3024 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3025 int e = N->getValueType(0).getVectorNumElements() / 2;
3027 for (int i = 0; i < e; ++i)
3028 if (!isUndefOrEqual(N->getMaskElt(i), i))
3030 for (int i = 0; i < e; ++i)
3031 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3036 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3037 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3038 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3039 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3040 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3042 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3044 for (int i = 0; i < NumOperands; ++i) {
3045 int Val = SVOp->getMaskElt(NumOperands-i-1);
3046 if (Val < 0) Val = 0;
3047 if (Val >= NumOperands) Val -= NumOperands;
3049 if (i != NumOperands - 1)
3055 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3056 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3057 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3058 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3060 // 8 nodes, but we only care about the last 4.
3061 for (unsigned i = 7; i >= 4; --i) {
3062 int Val = SVOp->getMaskElt(i);
3071 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3072 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3073 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3074 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3076 // 8 nodes, but we only care about the first 4.
3077 for (int i = 3; i >= 0; --i) {
3078 int Val = SVOp->getMaskElt(i);
3087 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3088 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3089 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3090 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3091 EVT VVT = N->getValueType(0);
3092 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3096 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3097 Val = SVOp->getMaskElt(i);
3101 return (Val - i) * EltSize;
3104 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3106 bool X86::isZeroNode(SDValue Elt) {
3107 return ((isa<ConstantSDNode>(Elt) &&
3108 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3109 (isa<ConstantFPSDNode>(Elt) &&
3110 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3113 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3114 /// their permute mask.
3115 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3116 SelectionDAG &DAG) {
3117 EVT VT = SVOp->getValueType(0);
3118 unsigned NumElems = VT.getVectorNumElements();
3119 SmallVector<int, 8> MaskVec;
3121 for (unsigned i = 0; i != NumElems; ++i) {
3122 int idx = SVOp->getMaskElt(i);
3124 MaskVec.push_back(idx);
3125 else if (idx < (int)NumElems)
3126 MaskVec.push_back(idx + NumElems);
3128 MaskVec.push_back(idx - NumElems);
3130 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3131 SVOp->getOperand(0), &MaskVec[0]);
3134 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3135 /// the two vector operands have swapped position.
3136 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3137 unsigned NumElems = VT.getVectorNumElements();
3138 for (unsigned i = 0; i != NumElems; ++i) {
3142 else if (idx < (int)NumElems)
3143 Mask[i] = idx + NumElems;
3145 Mask[i] = idx - NumElems;
3149 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3150 /// match movhlps. The lower half elements should come from upper half of
3151 /// V1 (and in order), and the upper half elements should come from the upper
3152 /// half of V2 (and in order).
3153 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3154 if (Op->getValueType(0).getVectorNumElements() != 4)
3156 for (unsigned i = 0, e = 2; i != e; ++i)
3157 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3159 for (unsigned i = 2; i != 4; ++i)
3160 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3165 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3166 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3168 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3169 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3171 N = N->getOperand(0).getNode();
3172 if (!ISD::isNON_EXTLoad(N))
3175 *LD = cast<LoadSDNode>(N);
3179 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3180 /// match movlp{s|d}. The lower half elements should come from lower half of
3181 /// V1 (and in order), and the upper half elements should come from the upper
3182 /// half of V2 (and in order). And since V1 will become the source of the
3183 /// MOVLP, it must be either a vector load or a scalar load to vector.
3184 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3185 ShuffleVectorSDNode *Op) {
3186 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3188 // Is V2 is a vector load, don't do this transformation. We will try to use
3189 // load folding shufps op.
3190 if (ISD::isNON_EXTLoad(V2))
3193 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3195 if (NumElems != 2 && NumElems != 4)
3197 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3198 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3200 for (unsigned i = NumElems/2; i != NumElems; ++i)
3201 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3206 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3208 static bool isSplatVector(SDNode *N) {
3209 if (N->getOpcode() != ISD::BUILD_VECTOR)
3212 SDValue SplatValue = N->getOperand(0);
3213 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3214 if (N->getOperand(i) != SplatValue)
3219 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3220 /// to an zero vector.
3221 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3222 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3223 SDValue V1 = N->getOperand(0);
3224 SDValue V2 = N->getOperand(1);
3225 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3226 for (unsigned i = 0; i != NumElems; ++i) {
3227 int Idx = N->getMaskElt(i);
3228 if (Idx >= (int)NumElems) {
3229 unsigned Opc = V2.getOpcode();
3230 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3232 if (Opc != ISD::BUILD_VECTOR ||
3233 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3235 } else if (Idx >= 0) {
3236 unsigned Opc = V1.getOpcode();
3237 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3239 if (Opc != ISD::BUILD_VECTOR ||
3240 !X86::isZeroNode(V1.getOperand(Idx)))
3247 /// getZeroVector - Returns a vector of specified type with all zero elements.
3249 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3251 assert(VT.isVector() && "Expected a vector type");
3253 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3254 // type. This ensures they get CSE'd.
3256 if (VT.getSizeInBits() == 64) { // MMX
3257 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3258 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3259 } else if (HasSSE2) { // SSE2
3260 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3261 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3263 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3264 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3266 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3269 /// getOnesVector - Returns a vector of specified type with all bits set.
3271 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3272 assert(VT.isVector() && "Expected a vector type");
3274 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3275 // type. This ensures they get CSE'd.
3276 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3278 if (VT.getSizeInBits() == 64) // MMX
3279 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3281 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3282 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3286 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3287 /// that point to V2 points to its first element.
3288 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3289 EVT VT = SVOp->getValueType(0);
3290 unsigned NumElems = VT.getVectorNumElements();
3292 bool Changed = false;
3293 SmallVector<int, 8> MaskVec;
3294 SVOp->getMask(MaskVec);
3296 for (unsigned i = 0; i != NumElems; ++i) {
3297 if (MaskVec[i] > (int)NumElems) {
3298 MaskVec[i] = NumElems;
3303 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3304 SVOp->getOperand(1), &MaskVec[0]);
3305 return SDValue(SVOp, 0);
3308 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3309 /// operation of specified width.
3310 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3312 unsigned NumElems = VT.getVectorNumElements();
3313 SmallVector<int, 8> Mask;
3314 Mask.push_back(NumElems);
3315 for (unsigned i = 1; i != NumElems; ++i)
3317 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3320 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3321 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3323 unsigned NumElems = VT.getVectorNumElements();
3324 SmallVector<int, 8> Mask;
3325 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3327 Mask.push_back(i + NumElems);
3329 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3332 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3333 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3335 unsigned NumElems = VT.getVectorNumElements();
3336 unsigned Half = NumElems/2;
3337 SmallVector<int, 8> Mask;
3338 for (unsigned i = 0; i != Half; ++i) {
3339 Mask.push_back(i + Half);
3340 Mask.push_back(i + NumElems + Half);
3342 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3345 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3346 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3348 if (SV->getValueType(0).getVectorNumElements() <= 4)
3349 return SDValue(SV, 0);
3351 EVT PVT = MVT::v4f32;
3352 EVT VT = SV->getValueType(0);
3353 DebugLoc dl = SV->getDebugLoc();
3354 SDValue V1 = SV->getOperand(0);
3355 int NumElems = VT.getVectorNumElements();
3356 int EltNo = SV->getSplatIndex();
3358 // unpack elements to the correct location
3359 while (NumElems > 4) {
3360 if (EltNo < NumElems/2) {
3361 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3363 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3364 EltNo -= NumElems/2;
3369 // Perform the splat.
3370 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3371 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3372 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3373 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3376 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3377 /// vector of zero or undef vector. This produces a shuffle where the low
3378 /// element of V2 is swizzled into the zero/undef vector, landing at element
3379 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3380 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3381 bool isZero, bool HasSSE2,
3382 SelectionDAG &DAG) {
3383 EVT VT = V2.getValueType();
3385 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3386 unsigned NumElems = VT.getVectorNumElements();
3387 SmallVector<int, 16> MaskVec;
3388 for (unsigned i = 0; i != NumElems; ++i)
3389 // If this is the insertion idx, put the low elt of V2 here.
3390 MaskVec.push_back(i == Idx ? NumElems : i);
3391 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3394 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3395 /// a shuffle that is zero.
3397 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3398 bool Low, SelectionDAG &DAG) {
3399 unsigned NumZeros = 0;
3400 for (int i = 0; i < NumElems; ++i) {
3401 unsigned Index = Low ? i : NumElems-i-1;
3402 int Idx = SVOp->getMaskElt(Index);
3407 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3408 if (Elt.getNode() && X86::isZeroNode(Elt))
3416 /// isVectorShift - Returns true if the shuffle can be implemented as a
3417 /// logical left or right shift of a vector.
3418 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3419 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3420 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3421 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3424 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3427 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3431 bool SeenV1 = false;
3432 bool SeenV2 = false;
3433 for (int i = NumZeros; i < NumElems; ++i) {
3434 int Val = isLeft ? (i - NumZeros) : i;
3435 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3447 if (SeenV1 && SeenV2)
3450 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3456 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3458 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3459 unsigned NumNonZero, unsigned NumZero,
3460 SelectionDAG &DAG, TargetLowering &TLI) {
3464 DebugLoc dl = Op.getDebugLoc();
3467 for (unsigned i = 0; i < 16; ++i) {
3468 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3469 if (ThisIsNonZero && First) {
3471 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3473 V = DAG.getUNDEF(MVT::v8i16);
3478 SDValue ThisElt(0, 0), LastElt(0, 0);
3479 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3480 if (LastIsNonZero) {
3481 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3482 MVT::i16, Op.getOperand(i-1));
3484 if (ThisIsNonZero) {
3485 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3486 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3487 ThisElt, DAG.getConstant(8, MVT::i8));
3489 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3493 if (ThisElt.getNode())
3494 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3495 DAG.getIntPtrConstant(i/2));
3499 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3502 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3504 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3505 unsigned NumNonZero, unsigned NumZero,
3506 SelectionDAG &DAG, TargetLowering &TLI) {
3510 DebugLoc dl = Op.getDebugLoc();
3513 for (unsigned i = 0; i < 8; ++i) {
3514 bool isNonZero = (NonZeros & (1 << i)) != 0;
3518 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3520 V = DAG.getUNDEF(MVT::v8i16);
3523 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3524 MVT::v8i16, V, Op.getOperand(i),
3525 DAG.getIntPtrConstant(i));
3532 /// getVShift - Return a vector logical shift node.
3534 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3535 unsigned NumBits, SelectionDAG &DAG,
3536 const TargetLowering &TLI, DebugLoc dl) {
3537 bool isMMX = VT.getSizeInBits() == 64;
3538 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3539 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3540 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3541 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3542 DAG.getNode(Opc, dl, ShVT, SrcOp,
3543 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3547 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3548 SelectionDAG &DAG) {
3550 // Check if the scalar load can be widened into a vector load. And if
3551 // the address is "base + cst" see if the cst can be "absorbed" into
3552 // the shuffle mask.
3553 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3554 SDValue Ptr = LD->getBasePtr();
3555 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3557 EVT PVT = LD->getValueType(0);
3558 if (PVT != MVT::i32 && PVT != MVT::f32)
3563 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3564 FI = FINode->getIndex();
3566 } else if (Ptr.getOpcode() == ISD::ADD &&
3567 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3568 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3569 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3570 Offset = Ptr.getConstantOperandVal(1);
3571 Ptr = Ptr.getOperand(0);
3576 SDValue Chain = LD->getChain();
3577 // Make sure the stack object alignment is at least 16.
3578 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3579 if (DAG.InferPtrAlignment(Ptr) < 16) {
3580 if (MFI->isFixedObjectIndex(FI)) {
3581 // Can't change the alignment. FIXME: It's possible to compute
3582 // the exact stack offset and reference FI + adjust offset instead.
3583 // If someone *really* cares about this. That's the way to implement it.
3586 MFI->setObjectAlignment(FI, 16);
3590 // (Offset % 16) must be multiple of 4. Then address is then
3591 // Ptr + (Offset & ~15).
3594 if ((Offset % 16) & 3)
3596 int64_t StartOffset = Offset & ~15;
3598 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3599 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3601 int EltNo = (Offset - StartOffset) >> 2;
3602 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3603 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3604 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3606 // Canonicalize it to a v4i32 shuffle.
3607 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3608 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3609 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3610 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3616 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3617 DebugLoc &dl, SelectionDAG &DAG) {
3618 EVT EltVT = VT.getVectorElementType();
3619 unsigned NumElems = Elts.size();
3621 // FIXME: check for zeroes
3622 LoadSDNode *LDBase = NULL;
3623 unsigned LastLoadedElt = -1U;
3624 for (unsigned i = 0; i < NumElems; ++i) {
3625 SDValue Elt = Elts[i];
3627 if (!Elt.getNode() ||
3628 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3631 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3633 LDBase = cast<LoadSDNode>(Elt.getNode());
3637 if (Elt.getOpcode() == ISD::UNDEF)
3640 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3641 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3646 if (LastLoadedElt == NumElems - 1) {
3647 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3648 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3649 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3650 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3651 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3652 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3653 LDBase->isVolatile(), LDBase->isNonTemporal(),
3654 LDBase->getAlignment());
3655 } else if (NumElems == 4 && LastLoadedElt == 1) {
3656 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3657 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3658 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3659 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3665 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3666 DebugLoc dl = Op.getDebugLoc();
3667 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3668 if (ISD::isBuildVectorAllZeros(Op.getNode())
3669 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3670 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3671 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3672 // eliminated on x86-32 hosts.
3673 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3676 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3677 return getOnesVector(Op.getValueType(), DAG, dl);
3678 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3681 EVT VT = Op.getValueType();
3682 EVT ExtVT = VT.getVectorElementType();
3683 unsigned EVTBits = ExtVT.getSizeInBits();
3685 unsigned NumElems = Op.getNumOperands();
3686 unsigned NumZero = 0;
3687 unsigned NumNonZero = 0;
3688 unsigned NonZeros = 0;
3689 bool IsAllConstants = true;
3690 SmallSet<SDValue, 8> Values;
3691 for (unsigned i = 0; i < NumElems; ++i) {
3692 SDValue Elt = Op.getOperand(i);
3693 if (Elt.getOpcode() == ISD::UNDEF)
3696 if (Elt.getOpcode() != ISD::Constant &&
3697 Elt.getOpcode() != ISD::ConstantFP)
3698 IsAllConstants = false;
3699 if (X86::isZeroNode(Elt))
3702 NonZeros |= (1 << i);
3707 if (NumNonZero == 0) {
3708 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3709 return DAG.getUNDEF(VT);
3712 // Special case for single non-zero, non-undef, element.
3713 if (NumNonZero == 1) {
3714 unsigned Idx = CountTrailingZeros_32(NonZeros);
3715 SDValue Item = Op.getOperand(Idx);
3717 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3718 // the value are obviously zero, truncate the value to i32 and do the
3719 // insertion that way. Only do this if the value is non-constant or if the
3720 // value is a constant being inserted into element 0. It is cheaper to do
3721 // a constant pool load than it is to do a movd + shuffle.
3722 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3723 (!IsAllConstants || Idx == 0)) {
3724 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3725 // Handle MMX and SSE both.
3726 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3727 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3729 // Truncate the value (which may itself be a constant) to i32, and
3730 // convert it to a vector with movd (S2V+shuffle to zero extend).
3731 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3732 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3733 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3734 Subtarget->hasSSE2(), DAG);
3736 // Now we have our 32-bit value zero extended in the low element of
3737 // a vector. If Idx != 0, swizzle it into place.
3739 SmallVector<int, 4> Mask;
3740 Mask.push_back(Idx);
3741 for (unsigned i = 1; i != VecElts; ++i)
3743 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3744 DAG.getUNDEF(Item.getValueType()),
3747 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3751 // If we have a constant or non-constant insertion into the low element of
3752 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3753 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3754 // depending on what the source datatype is.
3757 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3758 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3759 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3760 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3761 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3762 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3764 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3765 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3766 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3767 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3768 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3769 Subtarget->hasSSE2(), DAG);
3770 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3774 // Is it a vector logical left shift?
3775 if (NumElems == 2 && Idx == 1 &&
3776 X86::isZeroNode(Op.getOperand(0)) &&
3777 !X86::isZeroNode(Op.getOperand(1))) {
3778 unsigned NumBits = VT.getSizeInBits();
3779 return getVShift(true, VT,
3780 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3781 VT, Op.getOperand(1)),
3782 NumBits/2, DAG, *this, dl);
3785 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3788 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3789 // is a non-constant being inserted into an element other than the low one,
3790 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3791 // movd/movss) to move this into the low element, then shuffle it into
3793 if (EVTBits == 32) {
3794 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3796 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3797 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3798 Subtarget->hasSSE2(), DAG);
3799 SmallVector<int, 8> MaskVec;
3800 for (unsigned i = 0; i < NumElems; i++)
3801 MaskVec.push_back(i == Idx ? 0 : 1);
3802 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3806 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3807 if (Values.size() == 1) {
3808 if (EVTBits == 32) {
3809 // Instead of a shuffle like this:
3810 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3811 // Check if it's possible to issue this instead.
3812 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3813 unsigned Idx = CountTrailingZeros_32(NonZeros);
3814 SDValue Item = Op.getOperand(Idx);
3815 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3816 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3821 // A vector full of immediates; various special cases are already
3822 // handled, so this is best done with a single constant-pool load.
3826 // Let legalizer expand 2-wide build_vectors.
3827 if (EVTBits == 64) {
3828 if (NumNonZero == 1) {
3829 // One half is zero or undef.
3830 unsigned Idx = CountTrailingZeros_32(NonZeros);
3831 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3832 Op.getOperand(Idx));
3833 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3834 Subtarget->hasSSE2(), DAG);
3839 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3840 if (EVTBits == 8 && NumElems == 16) {
3841 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3843 if (V.getNode()) return V;
3846 if (EVTBits == 16 && NumElems == 8) {
3847 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3849 if (V.getNode()) return V;
3852 // If element VT is == 32 bits, turn it into a number of shuffles.
3853 SmallVector<SDValue, 8> V;
3855 if (NumElems == 4 && NumZero > 0) {
3856 for (unsigned i = 0; i < 4; ++i) {
3857 bool isZero = !(NonZeros & (1 << i));
3859 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3861 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3864 for (unsigned i = 0; i < 2; ++i) {
3865 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3868 V[i] = V[i*2]; // Must be a zero vector.
3871 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3874 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3877 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3882 SmallVector<int, 8> MaskVec;
3883 bool Reverse = (NonZeros & 0x3) == 2;
3884 for (unsigned i = 0; i < 2; ++i)
3885 MaskVec.push_back(Reverse ? 1-i : i);
3886 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3887 for (unsigned i = 0; i < 2; ++i)
3888 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3889 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3892 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
3893 // Check for a build vector of consecutive loads.
3894 for (unsigned i = 0; i < NumElems; ++i)
3895 V[i] = Op.getOperand(i);
3897 // Check for elements which are consecutive loads.
3898 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
3902 // For SSE 4.1, use inserts into undef.
3903 if (getSubtarget()->hasSSE41()) {
3904 V[0] = DAG.getUNDEF(VT);
3905 for (unsigned i = 0; i < NumElems; ++i)
3906 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3907 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3908 Op.getOperand(i), DAG.getIntPtrConstant(i));
3912 // Otherwise, expand into a number of unpckl*
3914 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3915 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3916 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3917 for (unsigned i = 0; i < NumElems; ++i)
3918 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3920 while (NumElems != 0) {
3921 for (unsigned i = 0; i < NumElems; ++i)
3922 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3931 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3932 // We support concatenate two MMX registers and place them in a MMX
3933 // register. This is better than doing a stack convert.
3934 DebugLoc dl = Op.getDebugLoc();
3935 EVT ResVT = Op.getValueType();
3936 assert(Op.getNumOperands() == 2);
3937 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3938 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3940 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3941 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3942 InVec = Op.getOperand(1);
3943 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3944 unsigned NumElts = ResVT.getVectorNumElements();
3945 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3946 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3947 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3949 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3950 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3951 Mask[0] = 0; Mask[1] = 2;
3952 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3954 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3957 // v8i16 shuffles - Prefer shuffles in the following order:
3958 // 1. [all] pshuflw, pshufhw, optional move
3959 // 2. [ssse3] 1 x pshufb
3960 // 3. [ssse3] 2 x pshufb + 1 x por
3961 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3963 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3964 SelectionDAG &DAG, X86TargetLowering &TLI) {
3965 SDValue V1 = SVOp->getOperand(0);
3966 SDValue V2 = SVOp->getOperand(1);
3967 DebugLoc dl = SVOp->getDebugLoc();
3968 SmallVector<int, 8> MaskVals;
3970 // Determine if more than 1 of the words in each of the low and high quadwords
3971 // of the result come from the same quadword of one of the two inputs. Undef
3972 // mask values count as coming from any quadword, for better codegen.
3973 SmallVector<unsigned, 4> LoQuad(4);
3974 SmallVector<unsigned, 4> HiQuad(4);
3975 BitVector InputQuads(4);
3976 for (unsigned i = 0; i < 8; ++i) {
3977 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3978 int EltIdx = SVOp->getMaskElt(i);
3979 MaskVals.push_back(EltIdx);
3988 InputQuads.set(EltIdx / 4);
3991 int BestLoQuad = -1;
3992 unsigned MaxQuad = 1;
3993 for (unsigned i = 0; i < 4; ++i) {
3994 if (LoQuad[i] > MaxQuad) {
3996 MaxQuad = LoQuad[i];
4000 int BestHiQuad = -1;
4002 for (unsigned i = 0; i < 4; ++i) {
4003 if (HiQuad[i] > MaxQuad) {
4005 MaxQuad = HiQuad[i];
4009 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4010 // of the two input vectors, shuffle them into one input vector so only a
4011 // single pshufb instruction is necessary. If There are more than 2 input
4012 // quads, disable the next transformation since it does not help SSSE3.
4013 bool V1Used = InputQuads[0] || InputQuads[1];
4014 bool V2Used = InputQuads[2] || InputQuads[3];
4015 if (TLI.getSubtarget()->hasSSSE3()) {
4016 if (InputQuads.count() == 2 && V1Used && V2Used) {
4017 BestLoQuad = InputQuads.find_first();
4018 BestHiQuad = InputQuads.find_next(BestLoQuad);
4020 if (InputQuads.count() > 2) {
4026 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4027 // the shuffle mask. If a quad is scored as -1, that means that it contains
4028 // words from all 4 input quadwords.
4030 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4031 SmallVector<int, 8> MaskV;
4032 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4033 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4034 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4035 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4036 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4037 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4039 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4040 // source words for the shuffle, to aid later transformations.
4041 bool AllWordsInNewV = true;
4042 bool InOrder[2] = { true, true };
4043 for (unsigned i = 0; i != 8; ++i) {
4044 int idx = MaskVals[i];
4046 InOrder[i/4] = false;
4047 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4049 AllWordsInNewV = false;
4053 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4054 if (AllWordsInNewV) {
4055 for (int i = 0; i != 8; ++i) {
4056 int idx = MaskVals[i];
4059 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4060 if ((idx != i) && idx < 4)
4062 if ((idx != i) && idx > 3)
4071 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4072 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4073 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4074 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4075 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4079 // If we have SSSE3, and all words of the result are from 1 input vector,
4080 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4081 // is present, fall back to case 4.
4082 if (TLI.getSubtarget()->hasSSSE3()) {
4083 SmallVector<SDValue,16> pshufbMask;
4085 // If we have elements from both input vectors, set the high bit of the
4086 // shuffle mask element to zero out elements that come from V2 in the V1
4087 // mask, and elements that come from V1 in the V2 mask, so that the two
4088 // results can be OR'd together.
4089 bool TwoInputs = V1Used && V2Used;
4090 for (unsigned i = 0; i != 8; ++i) {
4091 int EltIdx = MaskVals[i] * 2;
4092 if (TwoInputs && (EltIdx >= 16)) {
4093 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4094 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4097 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4098 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4100 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4101 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4102 DAG.getNode(ISD::BUILD_VECTOR, dl,
4103 MVT::v16i8, &pshufbMask[0], 16));
4105 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4107 // Calculate the shuffle mask for the second input, shuffle it, and
4108 // OR it with the first shuffled input.
4110 for (unsigned i = 0; i != 8; ++i) {
4111 int EltIdx = MaskVals[i] * 2;
4113 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4114 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4117 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4118 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4120 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4121 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4122 DAG.getNode(ISD::BUILD_VECTOR, dl,
4123 MVT::v16i8, &pshufbMask[0], 16));
4124 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4125 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4128 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4129 // and update MaskVals with new element order.
4130 BitVector InOrder(8);
4131 if (BestLoQuad >= 0) {
4132 SmallVector<int, 8> MaskV;
4133 for (int i = 0; i != 4; ++i) {
4134 int idx = MaskVals[i];
4136 MaskV.push_back(-1);
4138 } else if ((idx / 4) == BestLoQuad) {
4139 MaskV.push_back(idx & 3);
4142 MaskV.push_back(-1);
4145 for (unsigned i = 4; i != 8; ++i)
4147 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4151 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4152 // and update MaskVals with the new element order.
4153 if (BestHiQuad >= 0) {
4154 SmallVector<int, 8> MaskV;
4155 for (unsigned i = 0; i != 4; ++i)
4157 for (unsigned i = 4; i != 8; ++i) {
4158 int idx = MaskVals[i];
4160 MaskV.push_back(-1);
4162 } else if ((idx / 4) == BestHiQuad) {
4163 MaskV.push_back((idx & 3) + 4);
4166 MaskV.push_back(-1);
4169 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4173 // In case BestHi & BestLo were both -1, which means each quadword has a word
4174 // from each of the four input quadwords, calculate the InOrder bitvector now
4175 // before falling through to the insert/extract cleanup.
4176 if (BestLoQuad == -1 && BestHiQuad == -1) {
4178 for (int i = 0; i != 8; ++i)
4179 if (MaskVals[i] < 0 || MaskVals[i] == i)
4183 // The other elements are put in the right place using pextrw and pinsrw.
4184 for (unsigned i = 0; i != 8; ++i) {
4187 int EltIdx = MaskVals[i];
4190 SDValue ExtOp = (EltIdx < 8)
4191 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4192 DAG.getIntPtrConstant(EltIdx))
4193 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4194 DAG.getIntPtrConstant(EltIdx - 8));
4195 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4196 DAG.getIntPtrConstant(i));
4201 // v16i8 shuffles - Prefer shuffles in the following order:
4202 // 1. [ssse3] 1 x pshufb
4203 // 2. [ssse3] 2 x pshufb + 1 x por
4204 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4206 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4207 SelectionDAG &DAG, X86TargetLowering &TLI) {
4208 SDValue V1 = SVOp->getOperand(0);
4209 SDValue V2 = SVOp->getOperand(1);
4210 DebugLoc dl = SVOp->getDebugLoc();
4211 SmallVector<int, 16> MaskVals;
4212 SVOp->getMask(MaskVals);
4214 // If we have SSSE3, case 1 is generated when all result bytes come from
4215 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4216 // present, fall back to case 3.
4217 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4220 for (unsigned i = 0; i < 16; ++i) {
4221 int EltIdx = MaskVals[i];
4230 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4231 if (TLI.getSubtarget()->hasSSSE3()) {
4232 SmallVector<SDValue,16> pshufbMask;
4234 // If all result elements are from one input vector, then only translate
4235 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4237 // Otherwise, we have elements from both input vectors, and must zero out
4238 // elements that come from V2 in the first mask, and V1 in the second mask
4239 // so that we can OR them together.
4240 bool TwoInputs = !(V1Only || V2Only);
4241 for (unsigned i = 0; i != 16; ++i) {
4242 int EltIdx = MaskVals[i];
4243 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4244 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4247 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4249 // If all the elements are from V2, assign it to V1 and return after
4250 // building the first pshufb.
4253 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4254 DAG.getNode(ISD::BUILD_VECTOR, dl,
4255 MVT::v16i8, &pshufbMask[0], 16));
4259 // Calculate the shuffle mask for the second input, shuffle it, and
4260 // OR it with the first shuffled input.
4262 for (unsigned i = 0; i != 16; ++i) {
4263 int EltIdx = MaskVals[i];
4265 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4268 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4270 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4271 DAG.getNode(ISD::BUILD_VECTOR, dl,
4272 MVT::v16i8, &pshufbMask[0], 16));
4273 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4276 // No SSSE3 - Calculate in place words and then fix all out of place words
4277 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4278 // the 16 different words that comprise the two doublequadword input vectors.
4279 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4280 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4281 SDValue NewV = V2Only ? V2 : V1;
4282 for (int i = 0; i != 8; ++i) {
4283 int Elt0 = MaskVals[i*2];
4284 int Elt1 = MaskVals[i*2+1];
4286 // This word of the result is all undef, skip it.
4287 if (Elt0 < 0 && Elt1 < 0)
4290 // This word of the result is already in the correct place, skip it.
4291 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4293 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4296 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4297 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4300 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4301 // using a single extract together, load it and store it.
4302 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4303 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4304 DAG.getIntPtrConstant(Elt1 / 2));
4305 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4306 DAG.getIntPtrConstant(i));
4310 // If Elt1 is defined, extract it from the appropriate source. If the
4311 // source byte is not also odd, shift the extracted word left 8 bits
4312 // otherwise clear the bottom 8 bits if we need to do an or.
4314 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4315 DAG.getIntPtrConstant(Elt1 / 2));
4316 if ((Elt1 & 1) == 0)
4317 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4318 DAG.getConstant(8, TLI.getShiftAmountTy()));
4320 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4321 DAG.getConstant(0xFF00, MVT::i16));
4323 // If Elt0 is defined, extract it from the appropriate source. If the
4324 // source byte is not also even, shift the extracted word right 8 bits. If
4325 // Elt1 was also defined, OR the extracted values together before
4326 // inserting them in the result.
4328 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4329 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4330 if ((Elt0 & 1) != 0)
4331 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4332 DAG.getConstant(8, TLI.getShiftAmountTy()));
4334 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4335 DAG.getConstant(0x00FF, MVT::i16));
4336 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4339 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4340 DAG.getIntPtrConstant(i));
4342 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4345 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4346 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4347 /// done when every pair / quad of shuffle mask elements point to elements in
4348 /// the right sequence. e.g.
4349 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4351 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4353 TargetLowering &TLI, DebugLoc dl) {
4354 EVT VT = SVOp->getValueType(0);
4355 SDValue V1 = SVOp->getOperand(0);
4356 SDValue V2 = SVOp->getOperand(1);
4357 unsigned NumElems = VT.getVectorNumElements();
4358 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4359 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4360 EVT MaskEltVT = MaskVT.getVectorElementType();
4362 switch (VT.getSimpleVT().SimpleTy) {
4363 default: assert(false && "Unexpected!");
4364 case MVT::v4f32: NewVT = MVT::v2f64; break;
4365 case MVT::v4i32: NewVT = MVT::v2i64; break;
4366 case MVT::v8i16: NewVT = MVT::v4i32; break;
4367 case MVT::v16i8: NewVT = MVT::v4i32; break;
4370 if (NewWidth == 2) {
4376 int Scale = NumElems / NewWidth;
4377 SmallVector<int, 8> MaskVec;
4378 for (unsigned i = 0; i < NumElems; i += Scale) {
4380 for (int j = 0; j < Scale; ++j) {
4381 int EltIdx = SVOp->getMaskElt(i+j);
4385 StartIdx = EltIdx - (EltIdx % Scale);
4386 if (EltIdx != StartIdx + j)
4390 MaskVec.push_back(-1);
4392 MaskVec.push_back(StartIdx / Scale);
4395 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4396 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4397 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4400 /// getVZextMovL - Return a zero-extending vector move low node.
4402 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4403 SDValue SrcOp, SelectionDAG &DAG,
4404 const X86Subtarget *Subtarget, DebugLoc dl) {
4405 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4406 LoadSDNode *LD = NULL;
4407 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4408 LD = dyn_cast<LoadSDNode>(SrcOp);
4410 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4412 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4413 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4414 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4415 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4416 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4418 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4419 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4420 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4421 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4429 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4430 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4431 DAG.getNode(ISD::BIT_CONVERT, dl,
4435 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4438 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4439 SDValue V1 = SVOp->getOperand(0);
4440 SDValue V2 = SVOp->getOperand(1);
4441 DebugLoc dl = SVOp->getDebugLoc();
4442 EVT VT = SVOp->getValueType(0);
4444 SmallVector<std::pair<int, int>, 8> Locs;
4446 SmallVector<int, 8> Mask1(4U, -1);
4447 SmallVector<int, 8> PermMask;
4448 SVOp->getMask(PermMask);
4452 for (unsigned i = 0; i != 4; ++i) {
4453 int Idx = PermMask[i];
4455 Locs[i] = std::make_pair(-1, -1);
4457 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4459 Locs[i] = std::make_pair(0, NumLo);
4463 Locs[i] = std::make_pair(1, NumHi);
4465 Mask1[2+NumHi] = Idx;
4471 if (NumLo <= 2 && NumHi <= 2) {
4472 // If no more than two elements come from either vector. This can be
4473 // implemented with two shuffles. First shuffle gather the elements.
4474 // The second shuffle, which takes the first shuffle as both of its
4475 // vector operands, put the elements into the right order.
4476 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4478 SmallVector<int, 8> Mask2(4U, -1);
4480 for (unsigned i = 0; i != 4; ++i) {
4481 if (Locs[i].first == -1)
4484 unsigned Idx = (i < 2) ? 0 : 4;
4485 Idx += Locs[i].first * 2 + Locs[i].second;
4490 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4491 } else if (NumLo == 3 || NumHi == 3) {
4492 // Otherwise, we must have three elements from one vector, call it X, and
4493 // one element from the other, call it Y. First, use a shufps to build an
4494 // intermediate vector with the one element from Y and the element from X
4495 // that will be in the same half in the final destination (the indexes don't
4496 // matter). Then, use a shufps to build the final vector, taking the half
4497 // containing the element from Y from the intermediate, and the other half
4500 // Normalize it so the 3 elements come from V1.
4501 CommuteVectorShuffleMask(PermMask, VT);
4505 // Find the element from V2.
4507 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4508 int Val = PermMask[HiIndex];
4515 Mask1[0] = PermMask[HiIndex];
4517 Mask1[2] = PermMask[HiIndex^1];
4519 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4522 Mask1[0] = PermMask[0];
4523 Mask1[1] = PermMask[1];
4524 Mask1[2] = HiIndex & 1 ? 6 : 4;
4525 Mask1[3] = HiIndex & 1 ? 4 : 6;
4526 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4528 Mask1[0] = HiIndex & 1 ? 2 : 0;
4529 Mask1[1] = HiIndex & 1 ? 0 : 2;
4530 Mask1[2] = PermMask[2];
4531 Mask1[3] = PermMask[3];
4536 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4540 // Break it into (shuffle shuffle_hi, shuffle_lo).
4542 SmallVector<int,8> LoMask(4U, -1);
4543 SmallVector<int,8> HiMask(4U, -1);
4545 SmallVector<int,8> *MaskPtr = &LoMask;
4546 unsigned MaskIdx = 0;
4549 for (unsigned i = 0; i != 4; ++i) {
4556 int Idx = PermMask[i];
4558 Locs[i] = std::make_pair(-1, -1);
4559 } else if (Idx < 4) {
4560 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4561 (*MaskPtr)[LoIdx] = Idx;
4564 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4565 (*MaskPtr)[HiIdx] = Idx;
4570 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4571 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4572 SmallVector<int, 8> MaskOps;
4573 for (unsigned i = 0; i != 4; ++i) {
4574 if (Locs[i].first == -1) {
4575 MaskOps.push_back(-1);
4577 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4578 MaskOps.push_back(Idx);
4581 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4585 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4586 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4587 SDValue V1 = Op.getOperand(0);
4588 SDValue V2 = Op.getOperand(1);
4589 EVT VT = Op.getValueType();
4590 DebugLoc dl = Op.getDebugLoc();
4591 unsigned NumElems = VT.getVectorNumElements();
4592 bool isMMX = VT.getSizeInBits() == 64;
4593 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4594 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4595 bool V1IsSplat = false;
4596 bool V2IsSplat = false;
4598 if (isZeroShuffle(SVOp))
4599 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4601 // Promote splats to v4f32.
4602 if (SVOp->isSplat()) {
4603 if (isMMX || NumElems < 4)
4605 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4608 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4610 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4611 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4612 if (NewOp.getNode())
4613 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4614 LowerVECTOR_SHUFFLE(NewOp, DAG));
4615 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4616 // FIXME: Figure out a cleaner way to do this.
4617 // Try to make use of movq to zero out the top part.
4618 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4619 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4620 if (NewOp.getNode()) {
4621 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4622 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4623 DAG, Subtarget, dl);
4625 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4626 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4627 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4628 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4629 DAG, Subtarget, dl);
4633 if (X86::isPSHUFDMask(SVOp))
4636 // Check if this can be converted into a logical shift.
4637 bool isLeft = false;
4640 bool isShift = getSubtarget()->hasSSE2() &&
4641 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4642 if (isShift && ShVal.hasOneUse()) {
4643 // If the shifted value has multiple uses, it may be cheaper to use
4644 // v_set0 + movlhps or movhlps, etc.
4645 EVT EltVT = VT.getVectorElementType();
4646 ShAmt *= EltVT.getSizeInBits();
4647 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4650 if (X86::isMOVLMask(SVOp)) {
4653 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4654 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4659 // FIXME: fold these into legal mask.
4660 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4661 X86::isMOVSLDUPMask(SVOp) ||
4662 X86::isMOVHLPSMask(SVOp) ||
4663 X86::isMOVLHPSMask(SVOp) ||
4664 X86::isMOVLPMask(SVOp)))
4667 if (ShouldXformToMOVHLPS(SVOp) ||
4668 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4669 return CommuteVectorShuffle(SVOp, DAG);
4672 // No better options. Use a vshl / vsrl.
4673 EVT EltVT = VT.getVectorElementType();
4674 ShAmt *= EltVT.getSizeInBits();
4675 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4678 bool Commuted = false;
4679 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4680 // 1,1,1,1 -> v8i16 though.
4681 V1IsSplat = isSplatVector(V1.getNode());
4682 V2IsSplat = isSplatVector(V2.getNode());
4684 // Canonicalize the splat or undef, if present, to be on the RHS.
4685 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4686 Op = CommuteVectorShuffle(SVOp, DAG);
4687 SVOp = cast<ShuffleVectorSDNode>(Op);
4688 V1 = SVOp->getOperand(0);
4689 V2 = SVOp->getOperand(1);
4690 std::swap(V1IsSplat, V2IsSplat);
4691 std::swap(V1IsUndef, V2IsUndef);
4695 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4696 // Shuffling low element of v1 into undef, just return v1.
4699 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4700 // the instruction selector will not match, so get a canonical MOVL with
4701 // swapped operands to undo the commute.
4702 return getMOVL(DAG, dl, VT, V2, V1);
4705 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4706 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4707 X86::isUNPCKLMask(SVOp) ||
4708 X86::isUNPCKHMask(SVOp))
4712 // Normalize mask so all entries that point to V2 points to its first
4713 // element then try to match unpck{h|l} again. If match, return a
4714 // new vector_shuffle with the corrected mask.
4715 SDValue NewMask = NormalizeMask(SVOp, DAG);
4716 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4717 if (NSVOp != SVOp) {
4718 if (X86::isUNPCKLMask(NSVOp, true)) {
4720 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4727 // Commute is back and try unpck* again.
4728 // FIXME: this seems wrong.
4729 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4730 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4731 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4732 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4733 X86::isUNPCKLMask(NewSVOp) ||
4734 X86::isUNPCKHMask(NewSVOp))
4738 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4740 // Normalize the node to match x86 shuffle ops if needed
4741 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4742 return CommuteVectorShuffle(SVOp, DAG);
4744 // Check for legal shuffle and return?
4745 SmallVector<int, 16> PermMask;
4746 SVOp->getMask(PermMask);
4747 if (isShuffleMaskLegal(PermMask, VT))
4750 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4751 if (VT == MVT::v8i16) {
4752 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4753 if (NewOp.getNode())
4757 if (VT == MVT::v16i8) {
4758 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4759 if (NewOp.getNode())
4763 // Handle all 4 wide cases with a number of shuffles except for MMX.
4764 if (NumElems == 4 && !isMMX)
4765 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4771 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4772 SelectionDAG &DAG) {
4773 EVT VT = Op.getValueType();
4774 DebugLoc dl = Op.getDebugLoc();
4775 if (VT.getSizeInBits() == 8) {
4776 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4777 Op.getOperand(0), Op.getOperand(1));
4778 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4779 DAG.getValueType(VT));
4780 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4781 } else if (VT.getSizeInBits() == 16) {
4782 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4783 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4785 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4786 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4787 DAG.getNode(ISD::BIT_CONVERT, dl,
4791 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4792 Op.getOperand(0), Op.getOperand(1));
4793 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4794 DAG.getValueType(VT));
4795 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4796 } else if (VT == MVT::f32) {
4797 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4798 // the result back to FR32 register. It's only worth matching if the
4799 // result has a single use which is a store or a bitcast to i32. And in
4800 // the case of a store, it's not worth it if the index is a constant 0,
4801 // because a MOVSSmr can be used instead, which is smaller and faster.
4802 if (!Op.hasOneUse())
4804 SDNode *User = *Op.getNode()->use_begin();
4805 if ((User->getOpcode() != ISD::STORE ||
4806 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4807 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4808 (User->getOpcode() != ISD::BIT_CONVERT ||
4809 User->getValueType(0) != MVT::i32))
4811 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4812 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4815 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4816 } else if (VT == MVT::i32) {
4817 // ExtractPS works with constant index.
4818 if (isa<ConstantSDNode>(Op.getOperand(1)))
4826 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4827 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4830 if (Subtarget->hasSSE41()) {
4831 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4836 EVT VT = Op.getValueType();
4837 DebugLoc dl = Op.getDebugLoc();
4838 // TODO: handle v16i8.
4839 if (VT.getSizeInBits() == 16) {
4840 SDValue Vec = Op.getOperand(0);
4841 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4843 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4844 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4845 DAG.getNode(ISD::BIT_CONVERT, dl,
4848 // Transform it so it match pextrw which produces a 32-bit result.
4849 EVT EltVT = MVT::i32;
4850 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4851 Op.getOperand(0), Op.getOperand(1));
4852 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4853 DAG.getValueType(VT));
4854 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4855 } else if (VT.getSizeInBits() == 32) {
4856 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4860 // SHUFPS the element to the lowest double word, then movss.
4861 int Mask[4] = { Idx, -1, -1, -1 };
4862 EVT VVT = Op.getOperand(0).getValueType();
4863 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4864 DAG.getUNDEF(VVT), Mask);
4865 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4866 DAG.getIntPtrConstant(0));
4867 } else if (VT.getSizeInBits() == 64) {
4868 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4869 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4870 // to match extract_elt for f64.
4871 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4875 // UNPCKHPD the element to the lowest double word, then movsd.
4876 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4877 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4878 int Mask[2] = { 1, -1 };
4879 EVT VVT = Op.getOperand(0).getValueType();
4880 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4881 DAG.getUNDEF(VVT), Mask);
4882 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4883 DAG.getIntPtrConstant(0));
4890 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4891 EVT VT = Op.getValueType();
4892 EVT EltVT = VT.getVectorElementType();
4893 DebugLoc dl = Op.getDebugLoc();
4895 SDValue N0 = Op.getOperand(0);
4896 SDValue N1 = Op.getOperand(1);
4897 SDValue N2 = Op.getOperand(2);
4899 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4900 isa<ConstantSDNode>(N2)) {
4902 if (VT == MVT::v8i16)
4903 Opc = X86ISD::PINSRW;
4904 else if (VT == MVT::v4i16)
4905 Opc = X86ISD::MMX_PINSRW;
4906 else if (VT == MVT::v16i8)
4907 Opc = X86ISD::PINSRB;
4909 Opc = X86ISD::PINSRB;
4911 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4913 if (N1.getValueType() != MVT::i32)
4914 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4915 if (N2.getValueType() != MVT::i32)
4916 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4917 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4918 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4919 // Bits [7:6] of the constant are the source select. This will always be
4920 // zero here. The DAG Combiner may combine an extract_elt index into these
4921 // bits. For example (insert (extract, 3), 2) could be matched by putting
4922 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4923 // Bits [5:4] of the constant are the destination select. This is the
4924 // value of the incoming immediate.
4925 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4926 // combine either bitwise AND or insert of float 0.0 to set these bits.
4927 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4928 // Create this as a scalar to vector..
4929 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4930 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4931 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4932 // PINSR* works with constant index.
4939 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4940 EVT VT = Op.getValueType();
4941 EVT EltVT = VT.getVectorElementType();
4943 if (Subtarget->hasSSE41())
4944 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4946 if (EltVT == MVT::i8)
4949 DebugLoc dl = Op.getDebugLoc();
4950 SDValue N0 = Op.getOperand(0);
4951 SDValue N1 = Op.getOperand(1);
4952 SDValue N2 = Op.getOperand(2);
4954 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4955 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4956 // as its second argument.
4957 if (N1.getValueType() != MVT::i32)
4958 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4959 if (N2.getValueType() != MVT::i32)
4960 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4961 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
4962 dl, VT, N0, N1, N2);
4968 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4969 DebugLoc dl = Op.getDebugLoc();
4970 if (Op.getValueType() == MVT::v2f32)
4971 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4972 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4973 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4974 Op.getOperand(0))));
4976 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4977 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4979 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4980 EVT VT = MVT::v2i32;
4981 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4988 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4989 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4992 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4993 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4994 // one of the above mentioned nodes. It has to be wrapped because otherwise
4995 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4996 // be used to form addressing mode. These wrapped nodes will be selected
4999 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
5000 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5002 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5004 unsigned char OpFlag = 0;
5005 unsigned WrapperKind = X86ISD::Wrapper;
5006 CodeModel::Model M = getTargetMachine().getCodeModel();
5008 if (Subtarget->isPICStyleRIPRel() &&
5009 (M == CodeModel::Small || M == CodeModel::Kernel))
5010 WrapperKind = X86ISD::WrapperRIP;
5011 else if (Subtarget->isPICStyleGOT())
5012 OpFlag = X86II::MO_GOTOFF;
5013 else if (Subtarget->isPICStyleStubPIC())
5014 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5016 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5018 CP->getOffset(), OpFlag);
5019 DebugLoc DL = CP->getDebugLoc();
5020 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5021 // With PIC, the address is actually $g + Offset.
5023 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5024 DAG.getNode(X86ISD::GlobalBaseReg,
5025 DebugLoc::getUnknownLoc(), getPointerTy()),
5032 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
5033 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5035 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5037 unsigned char OpFlag = 0;
5038 unsigned WrapperKind = X86ISD::Wrapper;
5039 CodeModel::Model M = getTargetMachine().getCodeModel();
5041 if (Subtarget->isPICStyleRIPRel() &&
5042 (M == CodeModel::Small || M == CodeModel::Kernel))
5043 WrapperKind = X86ISD::WrapperRIP;
5044 else if (Subtarget->isPICStyleGOT())
5045 OpFlag = X86II::MO_GOTOFF;
5046 else if (Subtarget->isPICStyleStubPIC())
5047 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5049 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5051 DebugLoc DL = JT->getDebugLoc();
5052 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5054 // With PIC, the address is actually $g + Offset.
5056 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5057 DAG.getNode(X86ISD::GlobalBaseReg,
5058 DebugLoc::getUnknownLoc(), getPointerTy()),
5066 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
5067 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5069 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5071 unsigned char OpFlag = 0;
5072 unsigned WrapperKind = X86ISD::Wrapper;
5073 CodeModel::Model M = getTargetMachine().getCodeModel();
5075 if (Subtarget->isPICStyleRIPRel() &&
5076 (M == CodeModel::Small || M == CodeModel::Kernel))
5077 WrapperKind = X86ISD::WrapperRIP;
5078 else if (Subtarget->isPICStyleGOT())
5079 OpFlag = X86II::MO_GOTOFF;
5080 else if (Subtarget->isPICStyleStubPIC())
5081 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5083 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5085 DebugLoc DL = Op.getDebugLoc();
5086 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5089 // With PIC, the address is actually $g + Offset.
5090 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5091 !Subtarget->is64Bit()) {
5092 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5093 DAG.getNode(X86ISD::GlobalBaseReg,
5094 DebugLoc::getUnknownLoc(),
5103 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5104 // Create the TargetBlockAddressAddress node.
5105 unsigned char OpFlags =
5106 Subtarget->ClassifyBlockAddressReference();
5107 CodeModel::Model M = getTargetMachine().getCodeModel();
5108 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5109 DebugLoc dl = Op.getDebugLoc();
5110 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5111 /*isTarget=*/true, OpFlags);
5113 if (Subtarget->isPICStyleRIPRel() &&
5114 (M == CodeModel::Small || M == CodeModel::Kernel))
5115 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5117 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5119 // With PIC, the address is actually $g + Offset.
5120 if (isGlobalRelativeToPICBase(OpFlags)) {
5121 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5122 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5130 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5132 SelectionDAG &DAG) const {
5133 // Create the TargetGlobalAddress node, folding in the constant
5134 // offset if it is legal.
5135 unsigned char OpFlags =
5136 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5137 CodeModel::Model M = getTargetMachine().getCodeModel();
5139 if (OpFlags == X86II::MO_NO_FLAG &&
5140 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5141 // A direct static reference to a global.
5142 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5145 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5148 if (Subtarget->isPICStyleRIPRel() &&
5149 (M == CodeModel::Small || M == CodeModel::Kernel))
5150 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5152 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5154 // With PIC, the address is actually $g + Offset.
5155 if (isGlobalRelativeToPICBase(OpFlags)) {
5156 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5157 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5161 // For globals that require a load from a stub to get the address, emit the
5163 if (isGlobalStubReference(OpFlags))
5164 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5165 PseudoSourceValue::getGOT(), 0, false, false, 0);
5167 // If there was a non-zero offset that we didn't fold, create an explicit
5170 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5171 DAG.getConstant(Offset, getPointerTy()));
5177 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5178 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5179 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5180 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5184 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5185 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5186 unsigned char OperandFlags) {
5187 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5188 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5189 DebugLoc dl = GA->getDebugLoc();
5190 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5191 GA->getValueType(0),
5195 SDValue Ops[] = { Chain, TGA, *InFlag };
5196 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5198 SDValue Ops[] = { Chain, TGA };
5199 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5202 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5203 MFI->setHasCalls(true);
5205 SDValue Flag = Chain.getValue(1);
5206 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5209 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5211 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5214 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5215 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5216 DAG.getNode(X86ISD::GlobalBaseReg,
5217 DebugLoc::getUnknownLoc(),
5219 InFlag = Chain.getValue(1);
5221 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5224 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5226 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5228 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5229 X86::RAX, X86II::MO_TLSGD);
5232 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5233 // "local exec" model.
5234 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5235 const EVT PtrVT, TLSModel::Model model,
5237 DebugLoc dl = GA->getDebugLoc();
5238 // Get the Thread Pointer
5239 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5240 DebugLoc::getUnknownLoc(), PtrVT,
5241 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5244 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5245 NULL, 0, false, false, 0);
5247 unsigned char OperandFlags = 0;
5248 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5250 unsigned WrapperKind = X86ISD::Wrapper;
5251 if (model == TLSModel::LocalExec) {
5252 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5253 } else if (is64Bit) {
5254 assert(model == TLSModel::InitialExec);
5255 OperandFlags = X86II::MO_GOTTPOFF;
5256 WrapperKind = X86ISD::WrapperRIP;
5258 assert(model == TLSModel::InitialExec);
5259 OperandFlags = X86II::MO_INDNTPOFF;
5262 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5264 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5265 GA->getOffset(), OperandFlags);
5266 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5268 if (model == TLSModel::InitialExec)
5269 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5270 PseudoSourceValue::getGOT(), 0, false, false, 0);
5272 // The address of the thread local variable is the add of the thread
5273 // pointer with the offset of the variable.
5274 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5278 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5279 // TODO: implement the "local dynamic" model
5280 // TODO: implement the "initial exec"model for pic executables
5281 assert(Subtarget->isTargetELF() &&
5282 "TLS not implemented for non-ELF targets");
5283 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5284 const GlobalValue *GV = GA->getGlobal();
5286 // If GV is an alias then use the aliasee for determining
5287 // thread-localness.
5288 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5289 GV = GA->resolveAliasedGlobal(false);
5291 TLSModel::Model model = getTLSModel(GV,
5292 getTargetMachine().getRelocationModel());
5295 case TLSModel::GeneralDynamic:
5296 case TLSModel::LocalDynamic: // not implemented
5297 if (Subtarget->is64Bit())
5298 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5299 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5301 case TLSModel::InitialExec:
5302 case TLSModel::LocalExec:
5303 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5304 Subtarget->is64Bit());
5307 llvm_unreachable("Unreachable");
5312 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5313 /// take a 2 x i32 value to shift plus a shift amount.
5314 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5315 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5316 EVT VT = Op.getValueType();
5317 unsigned VTBits = VT.getSizeInBits();
5318 DebugLoc dl = Op.getDebugLoc();
5319 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5320 SDValue ShOpLo = Op.getOperand(0);
5321 SDValue ShOpHi = Op.getOperand(1);
5322 SDValue ShAmt = Op.getOperand(2);
5323 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5324 DAG.getConstant(VTBits - 1, MVT::i8))
5325 : DAG.getConstant(0, VT);
5328 if (Op.getOpcode() == ISD::SHL_PARTS) {
5329 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5330 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5332 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5333 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5336 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5337 DAG.getConstant(VTBits, MVT::i8));
5338 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5339 AndNode, DAG.getConstant(0, MVT::i8));
5342 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5343 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5344 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5346 if (Op.getOpcode() == ISD::SHL_PARTS) {
5347 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5348 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5350 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5351 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5354 SDValue Ops[2] = { Lo, Hi };
5355 return DAG.getMergeValues(Ops, 2, dl);
5358 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5359 EVT SrcVT = Op.getOperand(0).getValueType();
5361 if (SrcVT.isVector()) {
5362 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5368 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5369 "Unknown SINT_TO_FP to lower!");
5371 // These are really Legal; return the operand so the caller accepts it as
5373 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5375 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5376 Subtarget->is64Bit()) {
5380 DebugLoc dl = Op.getDebugLoc();
5381 unsigned Size = SrcVT.getSizeInBits()/8;
5382 MachineFunction &MF = DAG.getMachineFunction();
5383 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5384 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5385 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5387 PseudoSourceValue::getFixedStack(SSFI), 0,
5389 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5392 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5394 SelectionDAG &DAG) {
5396 DebugLoc dl = Op.getDebugLoc();
5398 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5400 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5402 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5403 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5404 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5405 Tys, Ops, array_lengthof(Ops));
5408 Chain = Result.getValue(1);
5409 SDValue InFlag = Result.getValue(2);
5411 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5412 // shouldn't be necessary except that RFP cannot be live across
5413 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5414 MachineFunction &MF = DAG.getMachineFunction();
5415 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5416 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5417 Tys = DAG.getVTList(MVT::Other);
5419 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5421 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5422 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5423 PseudoSourceValue::getFixedStack(SSFI), 0,
5430 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5431 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5432 // This algorithm is not obvious. Here it is in C code, more or less:
5434 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5435 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5436 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5438 // Copy ints to xmm registers.
5439 __m128i xh = _mm_cvtsi32_si128( hi );
5440 __m128i xl = _mm_cvtsi32_si128( lo );
5442 // Combine into low half of a single xmm register.
5443 __m128i x = _mm_unpacklo_epi32( xh, xl );
5447 // Merge in appropriate exponents to give the integer bits the right
5449 x = _mm_unpacklo_epi32( x, exp );
5451 // Subtract away the biases to deal with the IEEE-754 double precision
5453 d = _mm_sub_pd( (__m128d) x, bias );
5455 // All conversions up to here are exact. The correctly rounded result is
5456 // calculated using the current rounding mode using the following
5458 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5459 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5460 // store doesn't really need to be here (except
5461 // maybe to zero the other double)
5466 DebugLoc dl = Op.getDebugLoc();
5467 LLVMContext *Context = DAG.getContext();
5469 // Build some magic constants.
5470 std::vector<Constant*> CV0;
5471 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5472 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5473 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5474 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5475 Constant *C0 = ConstantVector::get(CV0);
5476 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5478 std::vector<Constant*> CV1;
5480 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5482 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5483 Constant *C1 = ConstantVector::get(CV1);
5484 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5486 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5487 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5489 DAG.getIntPtrConstant(1)));
5490 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5491 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5493 DAG.getIntPtrConstant(0)));
5494 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5495 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5496 PseudoSourceValue::getConstantPool(), 0,
5498 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5499 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5500 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5501 PseudoSourceValue::getConstantPool(), 0,
5503 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5505 // Add the halves; easiest way is to swap them into another reg first.
5506 int ShufMask[2] = { 1, -1 };
5507 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5508 DAG.getUNDEF(MVT::v2f64), ShufMask);
5509 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5510 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5511 DAG.getIntPtrConstant(0));
5514 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5515 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5516 DebugLoc dl = Op.getDebugLoc();
5517 // FP constant to bias correct the final result.
5518 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5521 // Load the 32-bit value into an XMM register.
5522 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5523 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5525 DAG.getIntPtrConstant(0)));
5527 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5528 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5529 DAG.getIntPtrConstant(0));
5531 // Or the load with the bias.
5532 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5533 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5534 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5536 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5537 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5538 MVT::v2f64, Bias)));
5539 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5540 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5541 DAG.getIntPtrConstant(0));
5543 // Subtract the bias.
5544 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5546 // Handle final rounding.
5547 EVT DestVT = Op.getValueType();
5549 if (DestVT.bitsLT(MVT::f64)) {
5550 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5551 DAG.getIntPtrConstant(0));
5552 } else if (DestVT.bitsGT(MVT::f64)) {
5553 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5556 // Handle final rounding.
5560 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5561 SDValue N0 = Op.getOperand(0);
5562 DebugLoc dl = Op.getDebugLoc();
5564 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5565 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5566 // the optimization here.
5567 if (DAG.SignBitIsZero(N0))
5568 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5570 EVT SrcVT = N0.getValueType();
5571 if (SrcVT == MVT::i64) {
5572 // We only handle SSE2 f64 target here; caller can expand the rest.
5573 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5576 return LowerUINT_TO_FP_i64(Op, DAG);
5577 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5578 return LowerUINT_TO_FP_i32(Op, DAG);
5581 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5583 // Make a 64-bit buffer, and use it to build an FILD.
5584 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5585 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5586 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5587 getPointerTy(), StackSlot, WordOff);
5588 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5589 StackSlot, NULL, 0, false, false, 0);
5590 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5591 OffsetSlot, NULL, 0, false, false, 0);
5592 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5595 std::pair<SDValue,SDValue> X86TargetLowering::
5596 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5597 DebugLoc dl = Op.getDebugLoc();
5599 EVT DstTy = Op.getValueType();
5602 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5606 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5607 DstTy.getSimpleVT() >= MVT::i16 &&
5608 "Unknown FP_TO_SINT to lower!");
5610 // These are really Legal.
5611 if (DstTy == MVT::i32 &&
5612 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5613 return std::make_pair(SDValue(), SDValue());
5614 if (Subtarget->is64Bit() &&
5615 DstTy == MVT::i64 &&
5616 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5617 return std::make_pair(SDValue(), SDValue());
5619 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5621 MachineFunction &MF = DAG.getMachineFunction();
5622 unsigned MemSize = DstTy.getSizeInBits()/8;
5623 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5624 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5627 switch (DstTy.getSimpleVT().SimpleTy) {
5628 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5629 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5630 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5631 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5634 SDValue Chain = DAG.getEntryNode();
5635 SDValue Value = Op.getOperand(0);
5636 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5637 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5638 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5639 PseudoSourceValue::getFixedStack(SSFI), 0,
5641 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5643 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5645 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5646 Chain = Value.getValue(1);
5647 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5648 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5651 // Build the FP_TO_INT*_IN_MEM
5652 SDValue Ops[] = { Chain, Value, StackSlot };
5653 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5655 return std::make_pair(FIST, StackSlot);
5658 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5659 if (Op.getValueType().isVector()) {
5660 if (Op.getValueType() == MVT::v2i32 &&
5661 Op.getOperand(0).getValueType() == MVT::v2f64) {
5667 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5668 SDValue FIST = Vals.first, StackSlot = Vals.second;
5669 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5670 if (FIST.getNode() == 0) return Op;
5673 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5674 FIST, StackSlot, NULL, 0, false, false, 0);
5677 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5678 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5679 SDValue FIST = Vals.first, StackSlot = Vals.second;
5680 assert(FIST.getNode() && "Unexpected failure");
5683 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5684 FIST, StackSlot, NULL, 0, false, false, 0);
5687 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5688 LLVMContext *Context = DAG.getContext();
5689 DebugLoc dl = Op.getDebugLoc();
5690 EVT VT = Op.getValueType();
5693 EltVT = VT.getVectorElementType();
5694 std::vector<Constant*> CV;
5695 if (EltVT == MVT::f64) {
5696 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5700 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5706 Constant *C = ConstantVector::get(CV);
5707 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5708 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5709 PseudoSourceValue::getConstantPool(), 0,
5711 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5714 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5715 LLVMContext *Context = DAG.getContext();
5716 DebugLoc dl = Op.getDebugLoc();
5717 EVT VT = Op.getValueType();
5720 EltVT = VT.getVectorElementType();
5721 std::vector<Constant*> CV;
5722 if (EltVT == MVT::f64) {
5723 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5727 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5733 Constant *C = ConstantVector::get(CV);
5734 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5735 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5736 PseudoSourceValue::getConstantPool(), 0,
5738 if (VT.isVector()) {
5739 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5740 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5741 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5743 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5745 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5749 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5750 LLVMContext *Context = DAG.getContext();
5751 SDValue Op0 = Op.getOperand(0);
5752 SDValue Op1 = Op.getOperand(1);
5753 DebugLoc dl = Op.getDebugLoc();
5754 EVT VT = Op.getValueType();
5755 EVT SrcVT = Op1.getValueType();
5757 // If second operand is smaller, extend it first.
5758 if (SrcVT.bitsLT(VT)) {
5759 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5762 // And if it is bigger, shrink it first.
5763 if (SrcVT.bitsGT(VT)) {
5764 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5768 // At this point the operands and the result should have the same
5769 // type, and that won't be f80 since that is not custom lowered.
5771 // First get the sign bit of second operand.
5772 std::vector<Constant*> CV;
5773 if (SrcVT == MVT::f64) {
5774 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5775 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5777 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5778 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5779 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5780 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5782 Constant *C = ConstantVector::get(CV);
5783 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5784 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5785 PseudoSourceValue::getConstantPool(), 0,
5787 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5789 // Shift sign bit right or left if the two operands have different types.
5790 if (SrcVT.bitsGT(VT)) {
5791 // Op0 is MVT::f32, Op1 is MVT::f64.
5792 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5793 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5794 DAG.getConstant(32, MVT::i32));
5795 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5796 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5797 DAG.getIntPtrConstant(0));
5800 // Clear first operand sign bit.
5802 if (VT == MVT::f64) {
5803 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5804 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5806 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5807 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5808 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5809 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5811 C = ConstantVector::get(CV);
5812 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5813 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5814 PseudoSourceValue::getConstantPool(), 0,
5816 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5818 // Or the value with the sign bit.
5819 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5822 /// Emit nodes that will be selected as "test Op0,Op0", or something
5824 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5825 SelectionDAG &DAG) {
5826 DebugLoc dl = Op.getDebugLoc();
5828 // CF and OF aren't always set the way we want. Determine which
5829 // of these we need.
5830 bool NeedCF = false;
5831 bool NeedOF = false;
5833 case X86::COND_A: case X86::COND_AE:
5834 case X86::COND_B: case X86::COND_BE:
5837 case X86::COND_G: case X86::COND_GE:
5838 case X86::COND_L: case X86::COND_LE:
5839 case X86::COND_O: case X86::COND_NO:
5845 // See if we can use the EFLAGS value from the operand instead of
5846 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5847 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5848 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5849 unsigned Opcode = 0;
5850 unsigned NumOperands = 0;
5851 switch (Op.getNode()->getOpcode()) {
5853 // Due to an isel shortcoming, be conservative if this add is likely to
5854 // be selected as part of a load-modify-store instruction. When the root
5855 // node in a match is a store, isel doesn't know how to remap non-chain
5856 // non-flag uses of other nodes in the match, such as the ADD in this
5857 // case. This leads to the ADD being left around and reselected, with
5858 // the result being two adds in the output.
5859 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5860 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5861 if (UI->getOpcode() == ISD::STORE)
5863 if (ConstantSDNode *C =
5864 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5865 // An add of one will be selected as an INC.
5866 if (C->getAPIntValue() == 1) {
5867 Opcode = X86ISD::INC;
5871 // An add of negative one (subtract of one) will be selected as a DEC.
5872 if (C->getAPIntValue().isAllOnesValue()) {
5873 Opcode = X86ISD::DEC;
5878 // Otherwise use a regular EFLAGS-setting add.
5879 Opcode = X86ISD::ADD;
5883 // If the primary and result isn't used, don't bother using X86ISD::AND,
5884 // because a TEST instruction will be better.
5885 bool NonFlagUse = false;
5886 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5887 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5889 unsigned UOpNo = UI.getOperandNo();
5890 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5891 // Look pass truncate.
5892 UOpNo = User->use_begin().getOperandNo();
5893 User = *User->use_begin();
5895 if (User->getOpcode() != ISD::BRCOND &&
5896 User->getOpcode() != ISD::SETCC &&
5897 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5909 // Due to the ISEL shortcoming noted above, be conservative if this op is
5910 // likely to be selected as part of a load-modify-store instruction.
5911 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5912 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5913 if (UI->getOpcode() == ISD::STORE)
5915 // Otherwise use a regular EFLAGS-setting instruction.
5916 switch (Op.getNode()->getOpcode()) {
5917 case ISD::SUB: Opcode = X86ISD::SUB; break;
5918 case ISD::OR: Opcode = X86ISD::OR; break;
5919 case ISD::XOR: Opcode = X86ISD::XOR; break;
5920 case ISD::AND: Opcode = X86ISD::AND; break;
5921 default: llvm_unreachable("unexpected operator!");
5932 return SDValue(Op.getNode(), 1);
5938 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5939 SmallVector<SDValue, 4> Ops;
5940 for (unsigned i = 0; i != NumOperands; ++i)
5941 Ops.push_back(Op.getOperand(i));
5942 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5943 DAG.ReplaceAllUsesWith(Op, New);
5944 return SDValue(New.getNode(), 1);
5948 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5949 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5950 DAG.getConstant(0, Op.getValueType()));
5953 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5955 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5956 SelectionDAG &DAG) {
5957 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5958 if (C->getAPIntValue() == 0)
5959 return EmitTest(Op0, X86CC, DAG);
5961 DebugLoc dl = Op0.getDebugLoc();
5962 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5965 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5966 /// if it's possible.
5967 static SDValue LowerToBT(SDValue And, ISD::CondCode CC,
5968 DebugLoc dl, SelectionDAG &DAG) {
5969 SDValue Op0 = And.getOperand(0);
5970 SDValue Op1 = And.getOperand(1);
5971 if (Op0.getOpcode() == ISD::TRUNCATE)
5972 Op0 = Op0.getOperand(0);
5973 if (Op1.getOpcode() == ISD::TRUNCATE)
5974 Op1 = Op1.getOperand(0);
5977 if (Op1.getOpcode() == ISD::SHL) {
5978 if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
5979 if (And10C->getZExtValue() == 1) {
5981 RHS = Op1.getOperand(1);
5983 } else if (Op0.getOpcode() == ISD::SHL) {
5984 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
5985 if (And00C->getZExtValue() == 1) {
5987 RHS = Op0.getOperand(1);
5989 } else if (Op1.getOpcode() == ISD::Constant) {
5990 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
5991 SDValue AndLHS = Op0;
5992 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5993 LHS = AndLHS.getOperand(0);
5994 RHS = AndLHS.getOperand(1);
5998 if (LHS.getNode()) {
5999 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
6000 // instruction. Since the shift amount is in-range-or-undefined, we know
6001 // that doing a bittest on the i16 value is ok. We extend to i32 because
6002 // the encoding for the i16 version is larger than the i32 version.
6003 if (LHS.getValueType() == MVT::i8)
6004 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6006 // If the operand types disagree, extend the shift amount to match. Since
6007 // BT ignores high bits (like shifts) we can use anyextend.
6008 if (LHS.getValueType() != RHS.getValueType())
6009 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6011 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6012 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6013 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6014 DAG.getConstant(Cond, MVT::i8), BT);
6020 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
6021 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6022 SDValue Op0 = Op.getOperand(0);
6023 SDValue Op1 = Op.getOperand(1);
6024 DebugLoc dl = Op.getDebugLoc();
6025 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6027 // Optimize to BT if possible.
6028 // Lower (X & (1 << N)) == 0 to BT(X, N).
6029 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6030 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6031 if (Op0.getOpcode() == ISD::AND &&
6033 Op1.getOpcode() == ISD::Constant &&
6034 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
6035 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6036 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6037 if (NewSetCC.getNode())
6041 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6042 if (Op0.getOpcode() == X86ISD::SETCC &&
6043 Op1.getOpcode() == ISD::Constant &&
6044 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6045 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6046 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6047 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6048 bool Invert = (CC == ISD::SETNE) ^
6049 cast<ConstantSDNode>(Op1)->isNullValue();
6051 CCode = X86::GetOppositeBranchCondition(CCode);
6052 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6053 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6056 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6057 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6058 if (X86CC == X86::COND_INVALID)
6061 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6063 // Use sbb x, x to materialize carry bit into a GPR.
6064 if (X86CC == X86::COND_B)
6065 return DAG.getNode(ISD::AND, dl, MVT::i8,
6066 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6067 DAG.getConstant(X86CC, MVT::i8), Cond),
6068 DAG.getConstant(1, MVT::i8));
6070 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6071 DAG.getConstant(X86CC, MVT::i8), Cond);
6074 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
6076 SDValue Op0 = Op.getOperand(0);
6077 SDValue Op1 = Op.getOperand(1);
6078 SDValue CC = Op.getOperand(2);
6079 EVT VT = Op.getValueType();
6080 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6081 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6082 DebugLoc dl = Op.getDebugLoc();
6086 EVT VT0 = Op0.getValueType();
6087 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6088 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6091 switch (SetCCOpcode) {
6094 case ISD::SETEQ: SSECC = 0; break;
6096 case ISD::SETGT: Swap = true; // Fallthrough
6098 case ISD::SETOLT: SSECC = 1; break;
6100 case ISD::SETGE: Swap = true; // Fallthrough
6102 case ISD::SETOLE: SSECC = 2; break;
6103 case ISD::SETUO: SSECC = 3; break;
6105 case ISD::SETNE: SSECC = 4; break;
6106 case ISD::SETULE: Swap = true;
6107 case ISD::SETUGE: SSECC = 5; break;
6108 case ISD::SETULT: Swap = true;
6109 case ISD::SETUGT: SSECC = 6; break;
6110 case ISD::SETO: SSECC = 7; break;
6113 std::swap(Op0, Op1);
6115 // In the two special cases we can't handle, emit two comparisons.
6117 if (SetCCOpcode == ISD::SETUEQ) {
6119 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6120 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6121 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6123 else if (SetCCOpcode == ISD::SETONE) {
6125 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6126 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6127 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6129 llvm_unreachable("Illegal FP comparison");
6131 // Handle all other FP comparisons here.
6132 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6135 // We are handling one of the integer comparisons here. Since SSE only has
6136 // GT and EQ comparisons for integer, swapping operands and multiple
6137 // operations may be required for some comparisons.
6138 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6139 bool Swap = false, Invert = false, FlipSigns = false;
6141 switch (VT.getSimpleVT().SimpleTy) {
6144 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6146 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6148 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6149 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6152 switch (SetCCOpcode) {
6154 case ISD::SETNE: Invert = true;
6155 case ISD::SETEQ: Opc = EQOpc; break;
6156 case ISD::SETLT: Swap = true;
6157 case ISD::SETGT: Opc = GTOpc; break;
6158 case ISD::SETGE: Swap = true;
6159 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6160 case ISD::SETULT: Swap = true;
6161 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6162 case ISD::SETUGE: Swap = true;
6163 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6166 std::swap(Op0, Op1);
6168 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6169 // bits of the inputs before performing those operations.
6171 EVT EltVT = VT.getVectorElementType();
6172 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6174 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6175 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6177 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6178 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6181 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6183 // If the logical-not of the result is required, perform that now.
6185 Result = DAG.getNOT(dl, Result, VT);
6190 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6191 static bool isX86LogicalCmp(SDValue Op) {
6192 unsigned Opc = Op.getNode()->getOpcode();
6193 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6195 if (Op.getResNo() == 1 &&
6196 (Opc == X86ISD::ADD ||
6197 Opc == X86ISD::SUB ||
6198 Opc == X86ISD::SMUL ||
6199 Opc == X86ISD::UMUL ||
6200 Opc == X86ISD::INC ||
6201 Opc == X86ISD::DEC ||
6202 Opc == X86ISD::OR ||
6203 Opc == X86ISD::XOR ||
6204 Opc == X86ISD::AND))
6210 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6211 bool addTest = true;
6212 SDValue Cond = Op.getOperand(0);
6213 DebugLoc dl = Op.getDebugLoc();
6216 if (Cond.getOpcode() == ISD::SETCC) {
6217 SDValue NewCond = LowerSETCC(Cond, DAG);
6218 if (NewCond.getNode())
6222 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6223 SDValue Op1 = Op.getOperand(1);
6224 SDValue Op2 = Op.getOperand(2);
6225 if (Cond.getOpcode() == X86ISD::SETCC &&
6226 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6227 SDValue Cmp = Cond.getOperand(1);
6228 if (Cmp.getOpcode() == X86ISD::CMP) {
6229 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6230 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6231 ConstantSDNode *RHSC =
6232 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6233 if (N1C && N1C->isAllOnesValue() &&
6234 N2C && N2C->isNullValue() &&
6235 RHSC && RHSC->isNullValue()) {
6236 SDValue CmpOp0 = Cmp.getOperand(0);
6237 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6238 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6239 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6240 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6245 // Look pass (and (setcc_carry (cmp ...)), 1).
6246 if (Cond.getOpcode() == ISD::AND &&
6247 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6248 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6249 if (C && C->getAPIntValue() == 1)
6250 Cond = Cond.getOperand(0);
6253 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6254 // setting operand in place of the X86ISD::SETCC.
6255 if (Cond.getOpcode() == X86ISD::SETCC ||
6256 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6257 CC = Cond.getOperand(0);
6259 SDValue Cmp = Cond.getOperand(1);
6260 unsigned Opc = Cmp.getOpcode();
6261 EVT VT = Op.getValueType();
6263 bool IllegalFPCMov = false;
6264 if (VT.isFloatingPoint() && !VT.isVector() &&
6265 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6266 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6268 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6269 Opc == X86ISD::BT) { // FIXME
6276 // Look pass the truncate.
6277 if (Cond.getOpcode() == ISD::TRUNCATE)
6278 Cond = Cond.getOperand(0);
6280 // We know the result of AND is compared against zero. Try to match
6282 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6283 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6284 if (NewSetCC.getNode()) {
6285 CC = NewSetCC.getOperand(0);
6286 Cond = NewSetCC.getOperand(1);
6293 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6294 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6297 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6298 // condition is true.
6299 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6300 SDValue Ops[] = { Op2, Op1, CC, Cond };
6301 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6304 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6305 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6306 // from the AND / OR.
6307 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6308 Opc = Op.getOpcode();
6309 if (Opc != ISD::OR && Opc != ISD::AND)
6311 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6312 Op.getOperand(0).hasOneUse() &&
6313 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6314 Op.getOperand(1).hasOneUse());
6317 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6318 // 1 and that the SETCC node has a single use.
6319 static bool isXor1OfSetCC(SDValue Op) {
6320 if (Op.getOpcode() != ISD::XOR)
6322 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6323 if (N1C && N1C->getAPIntValue() == 1) {
6324 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6325 Op.getOperand(0).hasOneUse();
6330 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6331 bool addTest = true;
6332 SDValue Chain = Op.getOperand(0);
6333 SDValue Cond = Op.getOperand(1);
6334 SDValue Dest = Op.getOperand(2);
6335 DebugLoc dl = Op.getDebugLoc();
6338 if (Cond.getOpcode() == ISD::SETCC) {
6339 SDValue NewCond = LowerSETCC(Cond, DAG);
6340 if (NewCond.getNode())
6344 // FIXME: LowerXALUO doesn't handle these!!
6345 else if (Cond.getOpcode() == X86ISD::ADD ||
6346 Cond.getOpcode() == X86ISD::SUB ||
6347 Cond.getOpcode() == X86ISD::SMUL ||
6348 Cond.getOpcode() == X86ISD::UMUL)
6349 Cond = LowerXALUO(Cond, DAG);
6352 // Look pass (and (setcc_carry (cmp ...)), 1).
6353 if (Cond.getOpcode() == ISD::AND &&
6354 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6355 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6356 if (C && C->getAPIntValue() == 1)
6357 Cond = Cond.getOperand(0);
6360 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6361 // setting operand in place of the X86ISD::SETCC.
6362 if (Cond.getOpcode() == X86ISD::SETCC ||
6363 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6364 CC = Cond.getOperand(0);
6366 SDValue Cmp = Cond.getOperand(1);
6367 unsigned Opc = Cmp.getOpcode();
6368 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6369 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6373 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6377 // These can only come from an arithmetic instruction with overflow,
6378 // e.g. SADDO, UADDO.
6379 Cond = Cond.getNode()->getOperand(1);
6386 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6387 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6388 if (CondOpc == ISD::OR) {
6389 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6390 // two branches instead of an explicit OR instruction with a
6392 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6393 isX86LogicalCmp(Cmp)) {
6394 CC = Cond.getOperand(0).getOperand(0);
6395 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6396 Chain, Dest, CC, Cmp);
6397 CC = Cond.getOperand(1).getOperand(0);
6401 } else { // ISD::AND
6402 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6403 // two branches instead of an explicit AND instruction with a
6404 // separate test. However, we only do this if this block doesn't
6405 // have a fall-through edge, because this requires an explicit
6406 // jmp when the condition is false.
6407 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6408 isX86LogicalCmp(Cmp) &&
6409 Op.getNode()->hasOneUse()) {
6410 X86::CondCode CCode =
6411 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6412 CCode = X86::GetOppositeBranchCondition(CCode);
6413 CC = DAG.getConstant(CCode, MVT::i8);
6414 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6415 // Look for an unconditional branch following this conditional branch.
6416 // We need this because we need to reverse the successors in order
6417 // to implement FCMP_OEQ.
6418 if (User.getOpcode() == ISD::BR) {
6419 SDValue FalseBB = User.getOperand(1);
6421 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6422 assert(NewBR == User);
6425 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6426 Chain, Dest, CC, Cmp);
6427 X86::CondCode CCode =
6428 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6429 CCode = X86::GetOppositeBranchCondition(CCode);
6430 CC = DAG.getConstant(CCode, MVT::i8);
6436 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6437 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6438 // It should be transformed during dag combiner except when the condition
6439 // is set by a arithmetics with overflow node.
6440 X86::CondCode CCode =
6441 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6442 CCode = X86::GetOppositeBranchCondition(CCode);
6443 CC = DAG.getConstant(CCode, MVT::i8);
6444 Cond = Cond.getOperand(0).getOperand(1);
6450 // Look pass the truncate.
6451 if (Cond.getOpcode() == ISD::TRUNCATE)
6452 Cond = Cond.getOperand(0);
6454 // We know the result of AND is compared against zero. Try to match
6456 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6457 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6458 if (NewSetCC.getNode()) {
6459 CC = NewSetCC.getOperand(0);
6460 Cond = NewSetCC.getOperand(1);
6467 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6468 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6470 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6471 Chain, Dest, CC, Cond);
6475 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6476 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6477 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6478 // that the guard pages used by the OS virtual memory manager are allocated in
6479 // correct sequence.
6481 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6482 SelectionDAG &DAG) {
6483 assert(Subtarget->isTargetCygMing() &&
6484 "This should be used only on Cygwin/Mingw targets");
6485 DebugLoc dl = Op.getDebugLoc();
6488 SDValue Chain = Op.getOperand(0);
6489 SDValue Size = Op.getOperand(1);
6490 // FIXME: Ensure alignment here
6494 EVT IntPtr = getPointerTy();
6495 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6497 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6498 Flag = Chain.getValue(1);
6500 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6502 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6503 Flag = Chain.getValue(1);
6505 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6507 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6508 return DAG.getMergeValues(Ops1, 2, dl);
6512 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6514 SDValue Dst, SDValue Src,
6515 SDValue Size, unsigned Align,
6517 uint64_t DstSVOff) {
6518 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6520 // If not DWORD aligned or size is more than the threshold, call the library.
6521 // The libc version is likely to be faster for these cases. It can use the
6522 // address value and run time information about the CPU.
6523 if ((Align & 3) != 0 ||
6525 ConstantSize->getZExtValue() >
6526 getSubtarget()->getMaxInlineSizeThreshold()) {
6527 SDValue InFlag(0, 0);
6529 // Check to see if there is a specialized entry-point for memory zeroing.
6530 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6532 if (const char *bzeroEntry = V &&
6533 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6534 EVT IntPtr = getPointerTy();
6535 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6536 TargetLowering::ArgListTy Args;
6537 TargetLowering::ArgListEntry Entry;
6539 Entry.Ty = IntPtrTy;
6540 Args.push_back(Entry);
6542 Args.push_back(Entry);
6543 std::pair<SDValue,SDValue> CallResult =
6544 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6545 false, false, false, false,
6546 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6547 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
6548 return CallResult.second;
6551 // Otherwise have the target-independent code call memset.
6555 uint64_t SizeVal = ConstantSize->getZExtValue();
6556 SDValue InFlag(0, 0);
6559 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6560 unsigned BytesLeft = 0;
6561 bool TwoRepStos = false;
6564 uint64_t Val = ValC->getZExtValue() & 255;
6566 // If the value is a constant, then we can potentially use larger sets.
6567 switch (Align & 3) {
6568 case 2: // WORD aligned
6571 Val = (Val << 8) | Val;
6573 case 0: // DWORD aligned
6576 Val = (Val << 8) | Val;
6577 Val = (Val << 16) | Val;
6578 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6581 Val = (Val << 32) | Val;
6584 default: // Byte aligned
6587 Count = DAG.getIntPtrConstant(SizeVal);
6591 if (AVT.bitsGT(MVT::i8)) {
6592 unsigned UBytes = AVT.getSizeInBits() / 8;
6593 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6594 BytesLeft = SizeVal % UBytes;
6597 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6599 InFlag = Chain.getValue(1);
6602 Count = DAG.getIntPtrConstant(SizeVal);
6603 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6604 InFlag = Chain.getValue(1);
6607 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6610 InFlag = Chain.getValue(1);
6611 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6614 InFlag = Chain.getValue(1);
6616 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6617 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6618 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6621 InFlag = Chain.getValue(1);
6623 EVT CVT = Count.getValueType();
6624 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6625 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6626 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6629 InFlag = Chain.getValue(1);
6630 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6631 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6632 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6633 } else if (BytesLeft) {
6634 // Handle the last 1 - 7 bytes.
6635 unsigned Offset = SizeVal - BytesLeft;
6636 EVT AddrVT = Dst.getValueType();
6637 EVT SizeVT = Size.getValueType();
6639 Chain = DAG.getMemset(Chain, dl,
6640 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6641 DAG.getConstant(Offset, AddrVT)),
6643 DAG.getConstant(BytesLeft, SizeVT),
6644 Align, DstSV, DstSVOff + Offset);
6647 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6652 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6653 SDValue Chain, SDValue Dst, SDValue Src,
6654 SDValue Size, unsigned Align,
6656 const Value *DstSV, uint64_t DstSVOff,
6657 const Value *SrcSV, uint64_t SrcSVOff) {
6658 // This requires the copy size to be a constant, preferrably
6659 // within a subtarget-specific limit.
6660 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6663 uint64_t SizeVal = ConstantSize->getZExtValue();
6664 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6667 /// If not DWORD aligned, call the library.
6668 if ((Align & 3) != 0)
6673 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6676 unsigned UBytes = AVT.getSizeInBits() / 8;
6677 unsigned CountVal = SizeVal / UBytes;
6678 SDValue Count = DAG.getIntPtrConstant(CountVal);
6679 unsigned BytesLeft = SizeVal % UBytes;
6681 SDValue InFlag(0, 0);
6682 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6685 InFlag = Chain.getValue(1);
6686 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6689 InFlag = Chain.getValue(1);
6690 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6693 InFlag = Chain.getValue(1);
6695 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6696 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6697 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6698 array_lengthof(Ops));
6700 SmallVector<SDValue, 4> Results;
6701 Results.push_back(RepMovs);
6703 // Handle the last 1 - 7 bytes.
6704 unsigned Offset = SizeVal - BytesLeft;
6705 EVT DstVT = Dst.getValueType();
6706 EVT SrcVT = Src.getValueType();
6707 EVT SizeVT = Size.getValueType();
6708 Results.push_back(DAG.getMemcpy(Chain, dl,
6709 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6710 DAG.getConstant(Offset, DstVT)),
6711 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6712 DAG.getConstant(Offset, SrcVT)),
6713 DAG.getConstant(BytesLeft, SizeVT),
6714 Align, AlwaysInline,
6715 DstSV, DstSVOff + Offset,
6716 SrcSV, SrcSVOff + Offset));
6719 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6720 &Results[0], Results.size());
6723 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6724 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6725 DebugLoc dl = Op.getDebugLoc();
6727 if (!Subtarget->is64Bit()) {
6728 // vastart just stores the address of the VarArgsFrameIndex slot into the
6729 // memory location argument.
6730 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6731 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6736 // gp_offset (0 - 6 * 8)
6737 // fp_offset (48 - 48 + 8 * 16)
6738 // overflow_arg_area (point to parameters coming in memory).
6740 SmallVector<SDValue, 8> MemOps;
6741 SDValue FIN = Op.getOperand(1);
6743 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6744 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6745 FIN, SV, 0, false, false, 0);
6746 MemOps.push_back(Store);
6749 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6750 FIN, DAG.getIntPtrConstant(4));
6751 Store = DAG.getStore(Op.getOperand(0), dl,
6752 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6753 FIN, SV, 0, false, false, 0);
6754 MemOps.push_back(Store);
6756 // Store ptr to overflow_arg_area
6757 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6758 FIN, DAG.getIntPtrConstant(4));
6759 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6760 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6762 MemOps.push_back(Store);
6764 // Store ptr to reg_save_area.
6765 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6766 FIN, DAG.getIntPtrConstant(8));
6767 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6768 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6770 MemOps.push_back(Store);
6771 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6772 &MemOps[0], MemOps.size());
6775 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6776 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6777 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6778 SDValue Chain = Op.getOperand(0);
6779 SDValue SrcPtr = Op.getOperand(1);
6780 SDValue SrcSV = Op.getOperand(2);
6782 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6786 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6787 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6788 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6789 SDValue Chain = Op.getOperand(0);
6790 SDValue DstPtr = Op.getOperand(1);
6791 SDValue SrcPtr = Op.getOperand(2);
6792 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6793 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6794 DebugLoc dl = Op.getDebugLoc();
6796 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6797 DAG.getIntPtrConstant(24), 8, false,
6798 DstSV, 0, SrcSV, 0);
6802 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6803 DebugLoc dl = Op.getDebugLoc();
6804 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6806 default: return SDValue(); // Don't custom lower most intrinsics.
6807 // Comparison intrinsics.
6808 case Intrinsic::x86_sse_comieq_ss:
6809 case Intrinsic::x86_sse_comilt_ss:
6810 case Intrinsic::x86_sse_comile_ss:
6811 case Intrinsic::x86_sse_comigt_ss:
6812 case Intrinsic::x86_sse_comige_ss:
6813 case Intrinsic::x86_sse_comineq_ss:
6814 case Intrinsic::x86_sse_ucomieq_ss:
6815 case Intrinsic::x86_sse_ucomilt_ss:
6816 case Intrinsic::x86_sse_ucomile_ss:
6817 case Intrinsic::x86_sse_ucomigt_ss:
6818 case Intrinsic::x86_sse_ucomige_ss:
6819 case Intrinsic::x86_sse_ucomineq_ss:
6820 case Intrinsic::x86_sse2_comieq_sd:
6821 case Intrinsic::x86_sse2_comilt_sd:
6822 case Intrinsic::x86_sse2_comile_sd:
6823 case Intrinsic::x86_sse2_comigt_sd:
6824 case Intrinsic::x86_sse2_comige_sd:
6825 case Intrinsic::x86_sse2_comineq_sd:
6826 case Intrinsic::x86_sse2_ucomieq_sd:
6827 case Intrinsic::x86_sse2_ucomilt_sd:
6828 case Intrinsic::x86_sse2_ucomile_sd:
6829 case Intrinsic::x86_sse2_ucomigt_sd:
6830 case Intrinsic::x86_sse2_ucomige_sd:
6831 case Intrinsic::x86_sse2_ucomineq_sd: {
6833 ISD::CondCode CC = ISD::SETCC_INVALID;
6836 case Intrinsic::x86_sse_comieq_ss:
6837 case Intrinsic::x86_sse2_comieq_sd:
6841 case Intrinsic::x86_sse_comilt_ss:
6842 case Intrinsic::x86_sse2_comilt_sd:
6846 case Intrinsic::x86_sse_comile_ss:
6847 case Intrinsic::x86_sse2_comile_sd:
6851 case Intrinsic::x86_sse_comigt_ss:
6852 case Intrinsic::x86_sse2_comigt_sd:
6856 case Intrinsic::x86_sse_comige_ss:
6857 case Intrinsic::x86_sse2_comige_sd:
6861 case Intrinsic::x86_sse_comineq_ss:
6862 case Intrinsic::x86_sse2_comineq_sd:
6866 case Intrinsic::x86_sse_ucomieq_ss:
6867 case Intrinsic::x86_sse2_ucomieq_sd:
6868 Opc = X86ISD::UCOMI;
6871 case Intrinsic::x86_sse_ucomilt_ss:
6872 case Intrinsic::x86_sse2_ucomilt_sd:
6873 Opc = X86ISD::UCOMI;
6876 case Intrinsic::x86_sse_ucomile_ss:
6877 case Intrinsic::x86_sse2_ucomile_sd:
6878 Opc = X86ISD::UCOMI;
6881 case Intrinsic::x86_sse_ucomigt_ss:
6882 case Intrinsic::x86_sse2_ucomigt_sd:
6883 Opc = X86ISD::UCOMI;
6886 case Intrinsic::x86_sse_ucomige_ss:
6887 case Intrinsic::x86_sse2_ucomige_sd:
6888 Opc = X86ISD::UCOMI;
6891 case Intrinsic::x86_sse_ucomineq_ss:
6892 case Intrinsic::x86_sse2_ucomineq_sd:
6893 Opc = X86ISD::UCOMI;
6898 SDValue LHS = Op.getOperand(1);
6899 SDValue RHS = Op.getOperand(2);
6900 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6901 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6902 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6903 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6904 DAG.getConstant(X86CC, MVT::i8), Cond);
6905 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6907 // ptest intrinsics. The intrinsic these come from are designed to return
6908 // an integer value, not just an instruction so lower it to the ptest
6909 // pattern and a setcc for the result.
6910 case Intrinsic::x86_sse41_ptestz:
6911 case Intrinsic::x86_sse41_ptestc:
6912 case Intrinsic::x86_sse41_ptestnzc:{
6915 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6916 case Intrinsic::x86_sse41_ptestz:
6918 X86CC = X86::COND_E;
6920 case Intrinsic::x86_sse41_ptestc:
6922 X86CC = X86::COND_B;
6924 case Intrinsic::x86_sse41_ptestnzc:
6926 X86CC = X86::COND_A;
6930 SDValue LHS = Op.getOperand(1);
6931 SDValue RHS = Op.getOperand(2);
6932 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6933 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6934 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6935 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6938 // Fix vector shift instructions where the last operand is a non-immediate
6940 case Intrinsic::x86_sse2_pslli_w:
6941 case Intrinsic::x86_sse2_pslli_d:
6942 case Intrinsic::x86_sse2_pslli_q:
6943 case Intrinsic::x86_sse2_psrli_w:
6944 case Intrinsic::x86_sse2_psrli_d:
6945 case Intrinsic::x86_sse2_psrli_q:
6946 case Intrinsic::x86_sse2_psrai_w:
6947 case Intrinsic::x86_sse2_psrai_d:
6948 case Intrinsic::x86_mmx_pslli_w:
6949 case Intrinsic::x86_mmx_pslli_d:
6950 case Intrinsic::x86_mmx_pslli_q:
6951 case Intrinsic::x86_mmx_psrli_w:
6952 case Intrinsic::x86_mmx_psrli_d:
6953 case Intrinsic::x86_mmx_psrli_q:
6954 case Intrinsic::x86_mmx_psrai_w:
6955 case Intrinsic::x86_mmx_psrai_d: {
6956 SDValue ShAmt = Op.getOperand(2);
6957 if (isa<ConstantSDNode>(ShAmt))
6960 unsigned NewIntNo = 0;
6961 EVT ShAmtVT = MVT::v4i32;
6963 case Intrinsic::x86_sse2_pslli_w:
6964 NewIntNo = Intrinsic::x86_sse2_psll_w;
6966 case Intrinsic::x86_sse2_pslli_d:
6967 NewIntNo = Intrinsic::x86_sse2_psll_d;
6969 case Intrinsic::x86_sse2_pslli_q:
6970 NewIntNo = Intrinsic::x86_sse2_psll_q;
6972 case Intrinsic::x86_sse2_psrli_w:
6973 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6975 case Intrinsic::x86_sse2_psrli_d:
6976 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6978 case Intrinsic::x86_sse2_psrli_q:
6979 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6981 case Intrinsic::x86_sse2_psrai_w:
6982 NewIntNo = Intrinsic::x86_sse2_psra_w;
6984 case Intrinsic::x86_sse2_psrai_d:
6985 NewIntNo = Intrinsic::x86_sse2_psra_d;
6988 ShAmtVT = MVT::v2i32;
6990 case Intrinsic::x86_mmx_pslli_w:
6991 NewIntNo = Intrinsic::x86_mmx_psll_w;
6993 case Intrinsic::x86_mmx_pslli_d:
6994 NewIntNo = Intrinsic::x86_mmx_psll_d;
6996 case Intrinsic::x86_mmx_pslli_q:
6997 NewIntNo = Intrinsic::x86_mmx_psll_q;
6999 case Intrinsic::x86_mmx_psrli_w:
7000 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7002 case Intrinsic::x86_mmx_psrli_d:
7003 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7005 case Intrinsic::x86_mmx_psrli_q:
7006 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7008 case Intrinsic::x86_mmx_psrai_w:
7009 NewIntNo = Intrinsic::x86_mmx_psra_w;
7011 case Intrinsic::x86_mmx_psrai_d:
7012 NewIntNo = Intrinsic::x86_mmx_psra_d;
7014 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7020 // The vector shift intrinsics with scalars uses 32b shift amounts but
7021 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7025 ShOps[1] = DAG.getConstant(0, MVT::i32);
7026 if (ShAmtVT == MVT::v4i32) {
7027 ShOps[2] = DAG.getUNDEF(MVT::i32);
7028 ShOps[3] = DAG.getUNDEF(MVT::i32);
7029 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7031 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7034 EVT VT = Op.getValueType();
7035 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7036 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7037 DAG.getConstant(NewIntNo, MVT::i32),
7038 Op.getOperand(1), ShAmt);
7043 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
7044 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7045 DebugLoc dl = Op.getDebugLoc();
7048 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7050 DAG.getConstant(TD->getPointerSize(),
7051 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7052 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7053 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7055 NULL, 0, false, false, 0);
7058 // Just load the return address.
7059 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7060 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7061 RetAddrFI, NULL, 0, false, false, 0);
7064 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
7065 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7066 MFI->setFrameAddressIsTaken(true);
7067 EVT VT = Op.getValueType();
7068 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7069 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7070 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7071 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7073 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7078 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7079 SelectionDAG &DAG) {
7080 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7083 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
7085 MachineFunction &MF = DAG.getMachineFunction();
7086 SDValue Chain = Op.getOperand(0);
7087 SDValue Offset = Op.getOperand(1);
7088 SDValue Handler = Op.getOperand(2);
7089 DebugLoc dl = Op.getDebugLoc();
7091 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7093 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7095 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7096 DAG.getIntPtrConstant(-TD->getPointerSize()));
7097 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7098 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7099 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7100 MF.getRegInfo().addLiveOut(StoreAddrReg);
7102 return DAG.getNode(X86ISD::EH_RETURN, dl,
7104 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7107 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7108 SelectionDAG &DAG) {
7109 SDValue Root = Op.getOperand(0);
7110 SDValue Trmp = Op.getOperand(1); // trampoline
7111 SDValue FPtr = Op.getOperand(2); // nested function
7112 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7113 DebugLoc dl = Op.getDebugLoc();
7115 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7117 if (Subtarget->is64Bit()) {
7118 SDValue OutChains[6];
7120 // Large code-model.
7121 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7122 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7124 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7125 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7127 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7129 // Load the pointer to the nested function into R11.
7130 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7131 SDValue Addr = Trmp;
7132 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7133 Addr, TrmpAddr, 0, false, false, 0);
7135 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7136 DAG.getConstant(2, MVT::i64));
7137 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7140 // Load the 'nest' parameter value into R10.
7141 // R10 is specified in X86CallingConv.td
7142 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7143 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7144 DAG.getConstant(10, MVT::i64));
7145 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7146 Addr, TrmpAddr, 10, false, false, 0);
7148 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7149 DAG.getConstant(12, MVT::i64));
7150 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7153 // Jump to the nested function.
7154 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7155 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7156 DAG.getConstant(20, MVT::i64));
7157 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7158 Addr, TrmpAddr, 20, false, false, 0);
7160 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7161 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7162 DAG.getConstant(22, MVT::i64));
7163 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7164 TrmpAddr, 22, false, false, 0);
7167 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7168 return DAG.getMergeValues(Ops, 2, dl);
7170 const Function *Func =
7171 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7172 CallingConv::ID CC = Func->getCallingConv();
7177 llvm_unreachable("Unsupported calling convention");
7178 case CallingConv::C:
7179 case CallingConv::X86_StdCall: {
7180 // Pass 'nest' parameter in ECX.
7181 // Must be kept in sync with X86CallingConv.td
7184 // Check that ECX wasn't needed by an 'inreg' parameter.
7185 const FunctionType *FTy = Func->getFunctionType();
7186 const AttrListPtr &Attrs = Func->getAttributes();
7188 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7189 unsigned InRegCount = 0;
7192 for (FunctionType::param_iterator I = FTy->param_begin(),
7193 E = FTy->param_end(); I != E; ++I, ++Idx)
7194 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7195 // FIXME: should only count parameters that are lowered to integers.
7196 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7198 if (InRegCount > 2) {
7199 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7204 case CallingConv::X86_FastCall:
7205 case CallingConv::Fast:
7206 // Pass 'nest' parameter in EAX.
7207 // Must be kept in sync with X86CallingConv.td
7212 SDValue OutChains[4];
7215 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7216 DAG.getConstant(10, MVT::i32));
7217 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7219 // This is storing the opcode for MOV32ri.
7220 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7221 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7222 OutChains[0] = DAG.getStore(Root, dl,
7223 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7224 Trmp, TrmpAddr, 0, false, false, 0);
7226 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7227 DAG.getConstant(1, MVT::i32));
7228 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7231 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7232 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7233 DAG.getConstant(5, MVT::i32));
7234 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7235 TrmpAddr, 5, false, false, 1);
7237 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7238 DAG.getConstant(6, MVT::i32));
7239 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7243 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7244 return DAG.getMergeValues(Ops, 2, dl);
7248 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7250 The rounding mode is in bits 11:10 of FPSR, and has the following
7257 FLT_ROUNDS, on the other hand, expects the following:
7264 To perform the conversion, we do:
7265 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7268 MachineFunction &MF = DAG.getMachineFunction();
7269 const TargetMachine &TM = MF.getTarget();
7270 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7271 unsigned StackAlignment = TFI.getStackAlignment();
7272 EVT VT = Op.getValueType();
7273 DebugLoc dl = Op.getDebugLoc();
7275 // Save FP Control Word to stack slot
7276 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7277 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7279 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7280 DAG.getEntryNode(), StackSlot);
7282 // Load FP Control Word from stack slot
7283 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7286 // Transform as necessary
7288 DAG.getNode(ISD::SRL, dl, MVT::i16,
7289 DAG.getNode(ISD::AND, dl, MVT::i16,
7290 CWD, DAG.getConstant(0x800, MVT::i16)),
7291 DAG.getConstant(11, MVT::i8));
7293 DAG.getNode(ISD::SRL, dl, MVT::i16,
7294 DAG.getNode(ISD::AND, dl, MVT::i16,
7295 CWD, DAG.getConstant(0x400, MVT::i16)),
7296 DAG.getConstant(9, MVT::i8));
7299 DAG.getNode(ISD::AND, dl, MVT::i16,
7300 DAG.getNode(ISD::ADD, dl, MVT::i16,
7301 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7302 DAG.getConstant(1, MVT::i16)),
7303 DAG.getConstant(3, MVT::i16));
7306 return DAG.getNode((VT.getSizeInBits() < 16 ?
7307 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7310 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7311 EVT VT = Op.getValueType();
7313 unsigned NumBits = VT.getSizeInBits();
7314 DebugLoc dl = Op.getDebugLoc();
7316 Op = Op.getOperand(0);
7317 if (VT == MVT::i8) {
7318 // Zero extend to i32 since there is not an i8 bsr.
7320 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7323 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7324 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7325 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7327 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7330 DAG.getConstant(NumBits+NumBits-1, OpVT),
7331 DAG.getConstant(X86::COND_E, MVT::i8),
7334 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7336 // Finally xor with NumBits-1.
7337 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7340 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7344 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7345 EVT VT = Op.getValueType();
7347 unsigned NumBits = VT.getSizeInBits();
7348 DebugLoc dl = Op.getDebugLoc();
7350 Op = Op.getOperand(0);
7351 if (VT == MVT::i8) {
7353 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7356 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7357 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7358 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7360 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7363 DAG.getConstant(NumBits, OpVT),
7364 DAG.getConstant(X86::COND_E, MVT::i8),
7367 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7370 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7374 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7375 EVT VT = Op.getValueType();
7376 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7377 DebugLoc dl = Op.getDebugLoc();
7379 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7380 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7381 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7382 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7383 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7385 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7386 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7387 // return AloBlo + AloBhi + AhiBlo;
7389 SDValue A = Op.getOperand(0);
7390 SDValue B = Op.getOperand(1);
7392 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7393 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7394 A, DAG.getConstant(32, MVT::i32));
7395 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7396 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7397 B, DAG.getConstant(32, MVT::i32));
7398 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7399 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7401 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7402 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7404 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7405 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7407 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7408 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7409 AloBhi, DAG.getConstant(32, MVT::i32));
7410 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7411 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7412 AhiBlo, DAG.getConstant(32, MVT::i32));
7413 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7414 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7419 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7420 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7421 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7422 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7423 // has only one use.
7424 SDNode *N = Op.getNode();
7425 SDValue LHS = N->getOperand(0);
7426 SDValue RHS = N->getOperand(1);
7427 unsigned BaseOp = 0;
7429 DebugLoc dl = Op.getDebugLoc();
7431 switch (Op.getOpcode()) {
7432 default: llvm_unreachable("Unknown ovf instruction!");
7434 // A subtract of one will be selected as a INC. Note that INC doesn't
7435 // set CF, so we can't do this for UADDO.
7436 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7437 if (C->getAPIntValue() == 1) {
7438 BaseOp = X86ISD::INC;
7442 BaseOp = X86ISD::ADD;
7446 BaseOp = X86ISD::ADD;
7450 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7451 // set CF, so we can't do this for USUBO.
7452 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7453 if (C->getAPIntValue() == 1) {
7454 BaseOp = X86ISD::DEC;
7458 BaseOp = X86ISD::SUB;
7462 BaseOp = X86ISD::SUB;
7466 BaseOp = X86ISD::SMUL;
7470 BaseOp = X86ISD::UMUL;
7475 // Also sets EFLAGS.
7476 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7477 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7480 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7481 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7483 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7487 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7488 EVT T = Op.getValueType();
7489 DebugLoc dl = Op.getDebugLoc();
7492 switch(T.getSimpleVT().SimpleTy) {
7494 assert(false && "Invalid value type!");
7495 case MVT::i8: Reg = X86::AL; size = 1; break;
7496 case MVT::i16: Reg = X86::AX; size = 2; break;
7497 case MVT::i32: Reg = X86::EAX; size = 4; break;
7499 assert(Subtarget->is64Bit() && "Node not type legal!");
7500 Reg = X86::RAX; size = 8;
7503 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7504 Op.getOperand(2), SDValue());
7505 SDValue Ops[] = { cpIn.getValue(0),
7508 DAG.getTargetConstant(size, MVT::i8),
7510 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7511 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7513 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7517 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7518 SelectionDAG &DAG) {
7519 assert(Subtarget->is64Bit() && "Result not type legalized?");
7520 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7521 SDValue TheChain = Op.getOperand(0);
7522 DebugLoc dl = Op.getDebugLoc();
7523 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7524 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7525 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7527 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7528 DAG.getConstant(32, MVT::i8));
7530 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7533 return DAG.getMergeValues(Ops, 2, dl);
7536 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7537 SDNode *Node = Op.getNode();
7538 DebugLoc dl = Node->getDebugLoc();
7539 EVT T = Node->getValueType(0);
7540 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7541 DAG.getConstant(0, T), Node->getOperand(2));
7542 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7543 cast<AtomicSDNode>(Node)->getMemoryVT(),
7544 Node->getOperand(0),
7545 Node->getOperand(1), negOp,
7546 cast<AtomicSDNode>(Node)->getSrcValue(),
7547 cast<AtomicSDNode>(Node)->getAlignment());
7550 /// LowerOperation - Provide custom lowering hooks for some operations.
7552 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7553 switch (Op.getOpcode()) {
7554 default: llvm_unreachable("Should not custom lower this!");
7555 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7556 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7557 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7558 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7559 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7560 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7561 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7562 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7563 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7564 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7565 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7566 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7567 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7568 case ISD::SHL_PARTS:
7569 case ISD::SRA_PARTS:
7570 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7571 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7572 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7573 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7574 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7575 case ISD::FABS: return LowerFABS(Op, DAG);
7576 case ISD::FNEG: return LowerFNEG(Op, DAG);
7577 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7578 case ISD::SETCC: return LowerSETCC(Op, DAG);
7579 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7580 case ISD::SELECT: return LowerSELECT(Op, DAG);
7581 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7582 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7583 case ISD::VASTART: return LowerVASTART(Op, DAG);
7584 case ISD::VAARG: return LowerVAARG(Op, DAG);
7585 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7586 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7587 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7588 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7589 case ISD::FRAME_TO_ARGS_OFFSET:
7590 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7591 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7592 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7593 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7594 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7595 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7596 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7597 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7603 case ISD::UMULO: return LowerXALUO(Op, DAG);
7604 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7608 void X86TargetLowering::
7609 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7610 SelectionDAG &DAG, unsigned NewOp) {
7611 EVT T = Node->getValueType(0);
7612 DebugLoc dl = Node->getDebugLoc();
7613 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7615 SDValue Chain = Node->getOperand(0);
7616 SDValue In1 = Node->getOperand(1);
7617 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7618 Node->getOperand(2), DAG.getIntPtrConstant(0));
7619 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7620 Node->getOperand(2), DAG.getIntPtrConstant(1));
7621 SDValue Ops[] = { Chain, In1, In2L, In2H };
7622 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7624 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7625 cast<MemSDNode>(Node)->getMemOperand());
7626 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7627 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7628 Results.push_back(Result.getValue(2));
7631 /// ReplaceNodeResults - Replace a node with an illegal result type
7632 /// with a new node built out of custom code.
7633 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7634 SmallVectorImpl<SDValue>&Results,
7635 SelectionDAG &DAG) {
7636 DebugLoc dl = N->getDebugLoc();
7637 switch (N->getOpcode()) {
7639 assert(false && "Do not know how to custom type legalize this operation!");
7641 case ISD::FP_TO_SINT: {
7642 std::pair<SDValue,SDValue> Vals =
7643 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7644 SDValue FIST = Vals.first, StackSlot = Vals.second;
7645 if (FIST.getNode() != 0) {
7646 EVT VT = N->getValueType(0);
7647 // Return a load from the stack slot.
7648 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7653 case ISD::READCYCLECOUNTER: {
7654 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7655 SDValue TheChain = N->getOperand(0);
7656 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7657 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7659 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7661 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7662 SDValue Ops[] = { eax, edx };
7663 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7664 Results.push_back(edx.getValue(1));
7667 case ISD::ATOMIC_CMP_SWAP: {
7668 EVT T = N->getValueType(0);
7669 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7670 SDValue cpInL, cpInH;
7671 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7672 DAG.getConstant(0, MVT::i32));
7673 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7674 DAG.getConstant(1, MVT::i32));
7675 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7676 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7678 SDValue swapInL, swapInH;
7679 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7680 DAG.getConstant(0, MVT::i32));
7681 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7682 DAG.getConstant(1, MVT::i32));
7683 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7685 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7686 swapInL.getValue(1));
7687 SDValue Ops[] = { swapInH.getValue(0),
7689 swapInH.getValue(1) };
7690 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7691 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7692 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7693 MVT::i32, Result.getValue(1));
7694 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7695 MVT::i32, cpOutL.getValue(2));
7696 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7697 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7698 Results.push_back(cpOutH.getValue(1));
7701 case ISD::ATOMIC_LOAD_ADD:
7702 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7704 case ISD::ATOMIC_LOAD_AND:
7705 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7707 case ISD::ATOMIC_LOAD_NAND:
7708 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7710 case ISD::ATOMIC_LOAD_OR:
7711 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7713 case ISD::ATOMIC_LOAD_SUB:
7714 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7716 case ISD::ATOMIC_LOAD_XOR:
7717 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7719 case ISD::ATOMIC_SWAP:
7720 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7725 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7727 default: return NULL;
7728 case X86ISD::BSF: return "X86ISD::BSF";
7729 case X86ISD::BSR: return "X86ISD::BSR";
7730 case X86ISD::SHLD: return "X86ISD::SHLD";
7731 case X86ISD::SHRD: return "X86ISD::SHRD";
7732 case X86ISD::FAND: return "X86ISD::FAND";
7733 case X86ISD::FOR: return "X86ISD::FOR";
7734 case X86ISD::FXOR: return "X86ISD::FXOR";
7735 case X86ISD::FSRL: return "X86ISD::FSRL";
7736 case X86ISD::FILD: return "X86ISD::FILD";
7737 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7738 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7739 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7740 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7741 case X86ISD::FLD: return "X86ISD::FLD";
7742 case X86ISD::FST: return "X86ISD::FST";
7743 case X86ISD::CALL: return "X86ISD::CALL";
7744 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7745 case X86ISD::BT: return "X86ISD::BT";
7746 case X86ISD::CMP: return "X86ISD::CMP";
7747 case X86ISD::COMI: return "X86ISD::COMI";
7748 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7749 case X86ISD::SETCC: return "X86ISD::SETCC";
7750 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7751 case X86ISD::CMOV: return "X86ISD::CMOV";
7752 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7753 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7754 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7755 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7756 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7757 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7758 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7759 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7760 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7761 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7762 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7763 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7764 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7765 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7766 case X86ISD::FMAX: return "X86ISD::FMAX";
7767 case X86ISD::FMIN: return "X86ISD::FMIN";
7768 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7769 case X86ISD::FRCP: return "X86ISD::FRCP";
7770 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7771 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7772 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7773 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7774 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7775 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7776 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7777 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7778 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7779 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7780 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7781 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7782 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7783 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7784 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7785 case X86ISD::VSHL: return "X86ISD::VSHL";
7786 case X86ISD::VSRL: return "X86ISD::VSRL";
7787 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7788 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7789 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7790 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7791 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7792 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7793 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7794 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7795 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7796 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7797 case X86ISD::ADD: return "X86ISD::ADD";
7798 case X86ISD::SUB: return "X86ISD::SUB";
7799 case X86ISD::SMUL: return "X86ISD::SMUL";
7800 case X86ISD::UMUL: return "X86ISD::UMUL";
7801 case X86ISD::INC: return "X86ISD::INC";
7802 case X86ISD::DEC: return "X86ISD::DEC";
7803 case X86ISD::OR: return "X86ISD::OR";
7804 case X86ISD::XOR: return "X86ISD::XOR";
7805 case X86ISD::AND: return "X86ISD::AND";
7806 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7807 case X86ISD::PTEST: return "X86ISD::PTEST";
7808 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7809 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7813 // isLegalAddressingMode - Return true if the addressing mode represented
7814 // by AM is legal for this target, for a load/store of the specified type.
7815 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7816 const Type *Ty) const {
7817 // X86 supports extremely general addressing modes.
7818 CodeModel::Model M = getTargetMachine().getCodeModel();
7820 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7821 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7826 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7828 // If a reference to this global requires an extra load, we can't fold it.
7829 if (isGlobalStubReference(GVFlags))
7832 // If BaseGV requires a register for the PIC base, we cannot also have a
7833 // BaseReg specified.
7834 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7837 // If lower 4G is not available, then we must use rip-relative addressing.
7838 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7848 // These scales always work.
7853 // These scales are formed with basereg+scalereg. Only accept if there is
7858 default: // Other stuff never works.
7866 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7867 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7869 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7870 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7871 if (NumBits1 <= NumBits2)
7876 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7877 if (!VT1.isInteger() || !VT2.isInteger())
7879 unsigned NumBits1 = VT1.getSizeInBits();
7880 unsigned NumBits2 = VT2.getSizeInBits();
7881 if (NumBits1 <= NumBits2)
7886 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7887 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7888 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7891 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7892 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7893 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7896 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7897 // i16 instructions are longer (0x66 prefix) and potentially slower.
7898 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7901 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7902 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7903 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7904 /// are assumed to be legal.
7906 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7908 // Only do shuffles on 128-bit vector types for now.
7909 if (VT.getSizeInBits() == 64)
7912 // FIXME: pshufb, blends, shifts.
7913 return (VT.getVectorNumElements() == 2 ||
7914 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7915 isMOVLMask(M, VT) ||
7916 isSHUFPMask(M, VT) ||
7917 isPSHUFDMask(M, VT) ||
7918 isPSHUFHWMask(M, VT) ||
7919 isPSHUFLWMask(M, VT) ||
7920 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7921 isUNPCKLMask(M, VT) ||
7922 isUNPCKHMask(M, VT) ||
7923 isUNPCKL_v_undef_Mask(M, VT) ||
7924 isUNPCKH_v_undef_Mask(M, VT));
7928 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7930 unsigned NumElts = VT.getVectorNumElements();
7931 // FIXME: This collection of masks seems suspect.
7934 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7935 return (isMOVLMask(Mask, VT) ||
7936 isCommutedMOVLMask(Mask, VT, true) ||
7937 isSHUFPMask(Mask, VT) ||
7938 isCommutedSHUFPMask(Mask, VT));
7943 //===----------------------------------------------------------------------===//
7944 // X86 Scheduler Hooks
7945 //===----------------------------------------------------------------------===//
7947 // private utility function
7949 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7950 MachineBasicBlock *MBB,
7958 TargetRegisterClass *RC,
7959 bool invSrc) const {
7960 // For the atomic bitwise operator, we generate
7963 // ld t1 = [bitinstr.addr]
7964 // op t2 = t1, [bitinstr.val]
7966 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7968 // fallthrough -->nextMBB
7969 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7970 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7971 MachineFunction::iterator MBBIter = MBB;
7974 /// First build the CFG
7975 MachineFunction *F = MBB->getParent();
7976 MachineBasicBlock *thisMBB = MBB;
7977 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7978 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7979 F->insert(MBBIter, newMBB);
7980 F->insert(MBBIter, nextMBB);
7982 // Move all successors to thisMBB to nextMBB
7983 nextMBB->transferSuccessors(thisMBB);
7985 // Update thisMBB to fall through to newMBB
7986 thisMBB->addSuccessor(newMBB);
7988 // newMBB jumps to itself and fall through to nextMBB
7989 newMBB->addSuccessor(nextMBB);
7990 newMBB->addSuccessor(newMBB);
7992 // Insert instructions into newMBB based on incoming instruction
7993 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7994 "unexpected number of operands");
7995 DebugLoc dl = bInstr->getDebugLoc();
7996 MachineOperand& destOper = bInstr->getOperand(0);
7997 MachineOperand* argOpers[2 + X86AddrNumOperands];
7998 int numArgs = bInstr->getNumOperands() - 1;
7999 for (int i=0; i < numArgs; ++i)
8000 argOpers[i] = &bInstr->getOperand(i+1);
8002 // x86 address has 4 operands: base, index, scale, and displacement
8003 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8004 int valArgIndx = lastAddrIndx + 1;
8006 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8007 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8008 for (int i=0; i <= lastAddrIndx; ++i)
8009 (*MIB).addOperand(*argOpers[i]);
8011 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8013 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8018 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8019 assert((argOpers[valArgIndx]->isReg() ||
8020 argOpers[valArgIndx]->isImm()) &&
8022 if (argOpers[valArgIndx]->isReg())
8023 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8025 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8027 (*MIB).addOperand(*argOpers[valArgIndx]);
8029 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8032 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8033 for (int i=0; i <= lastAddrIndx; ++i)
8034 (*MIB).addOperand(*argOpers[i]);
8036 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8037 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8038 bInstr->memoperands_end());
8040 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8044 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8046 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8050 // private utility function: 64 bit atomics on 32 bit host.
8052 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8053 MachineBasicBlock *MBB,
8058 bool invSrc) const {
8059 // For the atomic bitwise operator, we generate
8060 // thisMBB (instructions are in pairs, except cmpxchg8b)
8061 // ld t1,t2 = [bitinstr.addr]
8063 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8064 // op t5, t6 <- out1, out2, [bitinstr.val]
8065 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8066 // mov ECX, EBX <- t5, t6
8067 // mov EAX, EDX <- t1, t2
8068 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8069 // mov t3, t4 <- EAX, EDX
8071 // result in out1, out2
8072 // fallthrough -->nextMBB
8074 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8075 const unsigned LoadOpc = X86::MOV32rm;
8076 const unsigned copyOpc = X86::MOV32rr;
8077 const unsigned NotOpc = X86::NOT32r;
8078 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8079 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8080 MachineFunction::iterator MBBIter = MBB;
8083 /// First build the CFG
8084 MachineFunction *F = MBB->getParent();
8085 MachineBasicBlock *thisMBB = MBB;
8086 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8087 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8088 F->insert(MBBIter, newMBB);
8089 F->insert(MBBIter, nextMBB);
8091 // Move all successors to thisMBB to nextMBB
8092 nextMBB->transferSuccessors(thisMBB);
8094 // Update thisMBB to fall through to newMBB
8095 thisMBB->addSuccessor(newMBB);
8097 // newMBB jumps to itself and fall through to nextMBB
8098 newMBB->addSuccessor(nextMBB);
8099 newMBB->addSuccessor(newMBB);
8101 DebugLoc dl = bInstr->getDebugLoc();
8102 // Insert instructions into newMBB based on incoming instruction
8103 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8104 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8105 "unexpected number of operands");
8106 MachineOperand& dest1Oper = bInstr->getOperand(0);
8107 MachineOperand& dest2Oper = bInstr->getOperand(1);
8108 MachineOperand* argOpers[2 + X86AddrNumOperands];
8109 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8110 argOpers[i] = &bInstr->getOperand(i+2);
8112 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8113 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8115 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8116 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8117 for (int i=0; i <= lastAddrIndx; ++i)
8118 (*MIB).addOperand(*argOpers[i]);
8119 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8120 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8121 // add 4 to displacement.
8122 for (int i=0; i <= lastAddrIndx-2; ++i)
8123 (*MIB).addOperand(*argOpers[i]);
8124 MachineOperand newOp3 = *(argOpers[3]);
8126 newOp3.setImm(newOp3.getImm()+4);
8128 newOp3.setOffset(newOp3.getOffset()+4);
8129 (*MIB).addOperand(newOp3);
8130 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8132 // t3/4 are defined later, at the bottom of the loop
8133 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8134 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8135 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8136 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8137 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8138 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8140 // The subsequent operations should be using the destination registers of
8141 //the PHI instructions.
8143 t1 = F->getRegInfo().createVirtualRegister(RC);
8144 t2 = F->getRegInfo().createVirtualRegister(RC);
8145 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8146 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8148 t1 = dest1Oper.getReg();
8149 t2 = dest2Oper.getReg();
8152 int valArgIndx = lastAddrIndx + 1;
8153 assert((argOpers[valArgIndx]->isReg() ||
8154 argOpers[valArgIndx]->isImm()) &&
8156 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8157 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8158 if (argOpers[valArgIndx]->isReg())
8159 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8161 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8162 if (regOpcL != X86::MOV32rr)
8164 (*MIB).addOperand(*argOpers[valArgIndx]);
8165 assert(argOpers[valArgIndx + 1]->isReg() ==
8166 argOpers[valArgIndx]->isReg());
8167 assert(argOpers[valArgIndx + 1]->isImm() ==
8168 argOpers[valArgIndx]->isImm());
8169 if (argOpers[valArgIndx + 1]->isReg())
8170 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8172 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8173 if (regOpcH != X86::MOV32rr)
8175 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8177 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8179 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8182 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8184 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8187 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8188 for (int i=0; i <= lastAddrIndx; ++i)
8189 (*MIB).addOperand(*argOpers[i]);
8191 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8192 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8193 bInstr->memoperands_end());
8195 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8196 MIB.addReg(X86::EAX);
8197 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8198 MIB.addReg(X86::EDX);
8201 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8203 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8207 // private utility function
8209 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8210 MachineBasicBlock *MBB,
8211 unsigned cmovOpc) const {
8212 // For the atomic min/max operator, we generate
8215 // ld t1 = [min/max.addr]
8216 // mov t2 = [min/max.val]
8218 // cmov[cond] t2 = t1
8220 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8222 // fallthrough -->nextMBB
8224 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8225 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8226 MachineFunction::iterator MBBIter = MBB;
8229 /// First build the CFG
8230 MachineFunction *F = MBB->getParent();
8231 MachineBasicBlock *thisMBB = MBB;
8232 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8233 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8234 F->insert(MBBIter, newMBB);
8235 F->insert(MBBIter, nextMBB);
8237 // Move all successors of thisMBB to nextMBB
8238 nextMBB->transferSuccessors(thisMBB);
8240 // Update thisMBB to fall through to newMBB
8241 thisMBB->addSuccessor(newMBB);
8243 // newMBB jumps to newMBB and fall through to nextMBB
8244 newMBB->addSuccessor(nextMBB);
8245 newMBB->addSuccessor(newMBB);
8247 DebugLoc dl = mInstr->getDebugLoc();
8248 // Insert instructions into newMBB based on incoming instruction
8249 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8250 "unexpected number of operands");
8251 MachineOperand& destOper = mInstr->getOperand(0);
8252 MachineOperand* argOpers[2 + X86AddrNumOperands];
8253 int numArgs = mInstr->getNumOperands() - 1;
8254 for (int i=0; i < numArgs; ++i)
8255 argOpers[i] = &mInstr->getOperand(i+1);
8257 // x86 address has 4 operands: base, index, scale, and displacement
8258 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8259 int valArgIndx = lastAddrIndx + 1;
8261 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8262 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8263 for (int i=0; i <= lastAddrIndx; ++i)
8264 (*MIB).addOperand(*argOpers[i]);
8266 // We only support register and immediate values
8267 assert((argOpers[valArgIndx]->isReg() ||
8268 argOpers[valArgIndx]->isImm()) &&
8271 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8272 if (argOpers[valArgIndx]->isReg())
8273 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8275 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8276 (*MIB).addOperand(*argOpers[valArgIndx]);
8278 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8281 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8286 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8287 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8291 // Cmp and exchange if none has modified the memory location
8292 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8293 for (int i=0; i <= lastAddrIndx; ++i)
8294 (*MIB).addOperand(*argOpers[i]);
8296 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8297 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8298 mInstr->memoperands_end());
8300 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8301 MIB.addReg(X86::EAX);
8304 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8306 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8310 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8311 // all of this code can be replaced with that in the .td file.
8313 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8314 unsigned numArgs, bool memArg) const {
8316 MachineFunction *F = BB->getParent();
8317 DebugLoc dl = MI->getDebugLoc();
8318 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8322 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8324 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8326 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8328 for (unsigned i = 0; i < numArgs; ++i) {
8329 MachineOperand &Op = MI->getOperand(i+1);
8331 if (!(Op.isReg() && Op.isImplicit()))
8335 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8338 F->DeleteMachineInstr(MI);
8344 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8346 MachineBasicBlock *MBB) const {
8347 // Emit code to save XMM registers to the stack. The ABI says that the
8348 // number of registers to save is given in %al, so it's theoretically
8349 // possible to do an indirect jump trick to avoid saving all of them,
8350 // however this code takes a simpler approach and just executes all
8351 // of the stores if %al is non-zero. It's less code, and it's probably
8352 // easier on the hardware branch predictor, and stores aren't all that
8353 // expensive anyway.
8355 // Create the new basic blocks. One block contains all the XMM stores,
8356 // and one block is the final destination regardless of whether any
8357 // stores were performed.
8358 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8359 MachineFunction *F = MBB->getParent();
8360 MachineFunction::iterator MBBIter = MBB;
8362 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8363 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8364 F->insert(MBBIter, XMMSaveMBB);
8365 F->insert(MBBIter, EndMBB);
8368 // Move any original successors of MBB to the end block.
8369 EndMBB->transferSuccessors(MBB);
8370 // The original block will now fall through to the XMM save block.
8371 MBB->addSuccessor(XMMSaveMBB);
8372 // The XMMSaveMBB will fall through to the end block.
8373 XMMSaveMBB->addSuccessor(EndMBB);
8375 // Now add the instructions.
8376 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8377 DebugLoc DL = MI->getDebugLoc();
8379 unsigned CountReg = MI->getOperand(0).getReg();
8380 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8381 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8383 if (!Subtarget->isTargetWin64()) {
8384 // If %al is 0, branch around the XMM save block.
8385 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8386 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8387 MBB->addSuccessor(EndMBB);
8390 // In the XMM save block, save all the XMM argument registers.
8391 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8392 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8393 MachineMemOperand *MMO =
8394 F->getMachineMemOperand(
8395 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8396 MachineMemOperand::MOStore, Offset,
8397 /*Size=*/16, /*Align=*/16);
8398 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8399 .addFrameIndex(RegSaveFrameIndex)
8400 .addImm(/*Scale=*/1)
8401 .addReg(/*IndexReg=*/0)
8402 .addImm(/*Disp=*/Offset)
8403 .addReg(/*Segment=*/0)
8404 .addReg(MI->getOperand(i).getReg())
8405 .addMemOperand(MMO);
8408 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8414 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8415 MachineBasicBlock *BB,
8416 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8417 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8418 DebugLoc DL = MI->getDebugLoc();
8420 // To "insert" a SELECT_CC instruction, we actually have to insert the
8421 // diamond control-flow pattern. The incoming instruction knows the
8422 // destination vreg to set, the condition code register to branch on, the
8423 // true/false values to select between, and a branch opcode to use.
8424 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8425 MachineFunction::iterator It = BB;
8431 // cmpTY ccX, r1, r2
8433 // fallthrough --> copy0MBB
8434 MachineBasicBlock *thisMBB = BB;
8435 MachineFunction *F = BB->getParent();
8436 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8437 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8439 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8440 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8441 F->insert(It, copy0MBB);
8442 F->insert(It, sinkMBB);
8443 // Update machine-CFG edges by first adding all successors of the current
8444 // block to the new block which will contain the Phi node for the select.
8445 // Also inform sdisel of the edge changes.
8446 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8447 E = BB->succ_end(); I != E; ++I) {
8448 EM->insert(std::make_pair(*I, sinkMBB));
8449 sinkMBB->addSuccessor(*I);
8451 // Next, remove all successors of the current block, and add the true
8452 // and fallthrough blocks as its successors.
8453 while (!BB->succ_empty())
8454 BB->removeSuccessor(BB->succ_begin());
8455 // Add the true and fallthrough blocks as its successors.
8456 BB->addSuccessor(copy0MBB);
8457 BB->addSuccessor(sinkMBB);
8460 // %FalseValue = ...
8461 // # fallthrough to sinkMBB
8464 // Update machine-CFG edges
8465 BB->addSuccessor(sinkMBB);
8468 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8471 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8472 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8473 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8475 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8480 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8481 MachineBasicBlock *BB,
8482 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8483 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8484 DebugLoc DL = MI->getDebugLoc();
8485 MachineFunction *F = BB->getParent();
8487 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8488 // non-trivial part is impdef of ESP.
8489 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8492 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8493 .addExternalSymbol("_alloca")
8494 .addReg(X86::EAX, RegState::Implicit)
8495 .addReg(X86::ESP, RegState::Implicit)
8496 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8497 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8499 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8504 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8505 MachineBasicBlock *BB,
8506 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8507 switch (MI->getOpcode()) {
8508 default: assert(false && "Unexpected instr type to insert");
8509 case X86::MINGW_ALLOCA:
8510 return EmitLoweredMingwAlloca(MI, BB, EM);
8512 case X86::CMOV_V1I64:
8513 case X86::CMOV_FR32:
8514 case X86::CMOV_FR64:
8515 case X86::CMOV_V4F32:
8516 case X86::CMOV_V2F64:
8517 case X86::CMOV_V2I64:
8518 case X86::CMOV_GR16:
8519 case X86::CMOV_GR32:
8520 case X86::CMOV_RFP32:
8521 case X86::CMOV_RFP64:
8522 case X86::CMOV_RFP80:
8523 return EmitLoweredSelect(MI, BB, EM);
8525 case X86::FP32_TO_INT16_IN_MEM:
8526 case X86::FP32_TO_INT32_IN_MEM:
8527 case X86::FP32_TO_INT64_IN_MEM:
8528 case X86::FP64_TO_INT16_IN_MEM:
8529 case X86::FP64_TO_INT32_IN_MEM:
8530 case X86::FP64_TO_INT64_IN_MEM:
8531 case X86::FP80_TO_INT16_IN_MEM:
8532 case X86::FP80_TO_INT32_IN_MEM:
8533 case X86::FP80_TO_INT64_IN_MEM: {
8534 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8535 DebugLoc DL = MI->getDebugLoc();
8537 // Change the floating point control register to use "round towards zero"
8538 // mode when truncating to an integer value.
8539 MachineFunction *F = BB->getParent();
8540 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8541 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8543 // Load the old value of the high byte of the control word...
8545 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8546 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8549 // Set the high part to be round to zero...
8550 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8553 // Reload the modified control word now...
8554 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8556 // Restore the memory image of control word to original value
8557 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8560 // Get the X86 opcode to use.
8562 switch (MI->getOpcode()) {
8563 default: llvm_unreachable("illegal opcode!");
8564 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8565 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8566 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8567 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8568 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8569 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8570 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8571 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8572 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8576 MachineOperand &Op = MI->getOperand(0);
8578 AM.BaseType = X86AddressMode::RegBase;
8579 AM.Base.Reg = Op.getReg();
8581 AM.BaseType = X86AddressMode::FrameIndexBase;
8582 AM.Base.FrameIndex = Op.getIndex();
8584 Op = MI->getOperand(1);
8586 AM.Scale = Op.getImm();
8587 Op = MI->getOperand(2);
8589 AM.IndexReg = Op.getImm();
8590 Op = MI->getOperand(3);
8591 if (Op.isGlobal()) {
8592 AM.GV = Op.getGlobal();
8594 AM.Disp = Op.getImm();
8596 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8597 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8599 // Reload the original control word now.
8600 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8602 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8605 // DBG_VALUE. Only the frame index case is done here.
8606 case X86::DBG_VALUE: {
8607 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8608 DebugLoc DL = MI->getDebugLoc();
8610 MachineFunction *F = BB->getParent();
8611 AM.BaseType = X86AddressMode::FrameIndexBase;
8612 AM.Base.FrameIndex = MI->getOperand(0).getImm();
8613 addFullAddress(BuildMI(BB, DL, TII->get(X86::DBG_VALUE)), AM).
8614 addImm(MI->getOperand(1).getImm()).
8615 addMetadata(MI->getOperand(2).getMetadata());
8616 F->DeleteMachineInstr(MI); // Remove pseudo.
8620 // String/text processing lowering.
8621 case X86::PCMPISTRM128REG:
8622 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8623 case X86::PCMPISTRM128MEM:
8624 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8625 case X86::PCMPESTRM128REG:
8626 return EmitPCMP(MI, BB, 5, false /* in mem */);
8627 case X86::PCMPESTRM128MEM:
8628 return EmitPCMP(MI, BB, 5, true /* in mem */);
8631 case X86::ATOMAND32:
8632 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8633 X86::AND32ri, X86::MOV32rm,
8634 X86::LCMPXCHG32, X86::MOV32rr,
8635 X86::NOT32r, X86::EAX,
8636 X86::GR32RegisterClass);
8638 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8639 X86::OR32ri, X86::MOV32rm,
8640 X86::LCMPXCHG32, X86::MOV32rr,
8641 X86::NOT32r, X86::EAX,
8642 X86::GR32RegisterClass);
8643 case X86::ATOMXOR32:
8644 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8645 X86::XOR32ri, X86::MOV32rm,
8646 X86::LCMPXCHG32, X86::MOV32rr,
8647 X86::NOT32r, X86::EAX,
8648 X86::GR32RegisterClass);
8649 case X86::ATOMNAND32:
8650 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8651 X86::AND32ri, X86::MOV32rm,
8652 X86::LCMPXCHG32, X86::MOV32rr,
8653 X86::NOT32r, X86::EAX,
8654 X86::GR32RegisterClass, true);
8655 case X86::ATOMMIN32:
8656 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8657 case X86::ATOMMAX32:
8658 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8659 case X86::ATOMUMIN32:
8660 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8661 case X86::ATOMUMAX32:
8662 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8664 case X86::ATOMAND16:
8665 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8666 X86::AND16ri, X86::MOV16rm,
8667 X86::LCMPXCHG16, X86::MOV16rr,
8668 X86::NOT16r, X86::AX,
8669 X86::GR16RegisterClass);
8671 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8672 X86::OR16ri, X86::MOV16rm,
8673 X86::LCMPXCHG16, X86::MOV16rr,
8674 X86::NOT16r, X86::AX,
8675 X86::GR16RegisterClass);
8676 case X86::ATOMXOR16:
8677 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8678 X86::XOR16ri, X86::MOV16rm,
8679 X86::LCMPXCHG16, X86::MOV16rr,
8680 X86::NOT16r, X86::AX,
8681 X86::GR16RegisterClass);
8682 case X86::ATOMNAND16:
8683 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8684 X86::AND16ri, X86::MOV16rm,
8685 X86::LCMPXCHG16, X86::MOV16rr,
8686 X86::NOT16r, X86::AX,
8687 X86::GR16RegisterClass, true);
8688 case X86::ATOMMIN16:
8689 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8690 case X86::ATOMMAX16:
8691 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8692 case X86::ATOMUMIN16:
8693 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8694 case X86::ATOMUMAX16:
8695 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8698 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8699 X86::AND8ri, X86::MOV8rm,
8700 X86::LCMPXCHG8, X86::MOV8rr,
8701 X86::NOT8r, X86::AL,
8702 X86::GR8RegisterClass);
8704 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8705 X86::OR8ri, X86::MOV8rm,
8706 X86::LCMPXCHG8, X86::MOV8rr,
8707 X86::NOT8r, X86::AL,
8708 X86::GR8RegisterClass);
8710 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8711 X86::XOR8ri, X86::MOV8rm,
8712 X86::LCMPXCHG8, X86::MOV8rr,
8713 X86::NOT8r, X86::AL,
8714 X86::GR8RegisterClass);
8715 case X86::ATOMNAND8:
8716 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8717 X86::AND8ri, X86::MOV8rm,
8718 X86::LCMPXCHG8, X86::MOV8rr,
8719 X86::NOT8r, X86::AL,
8720 X86::GR8RegisterClass, true);
8721 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8722 // This group is for 64-bit host.
8723 case X86::ATOMAND64:
8724 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8725 X86::AND64ri32, X86::MOV64rm,
8726 X86::LCMPXCHG64, X86::MOV64rr,
8727 X86::NOT64r, X86::RAX,
8728 X86::GR64RegisterClass);
8730 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8731 X86::OR64ri32, X86::MOV64rm,
8732 X86::LCMPXCHG64, X86::MOV64rr,
8733 X86::NOT64r, X86::RAX,
8734 X86::GR64RegisterClass);
8735 case X86::ATOMXOR64:
8736 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8737 X86::XOR64ri32, X86::MOV64rm,
8738 X86::LCMPXCHG64, X86::MOV64rr,
8739 X86::NOT64r, X86::RAX,
8740 X86::GR64RegisterClass);
8741 case X86::ATOMNAND64:
8742 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8743 X86::AND64ri32, X86::MOV64rm,
8744 X86::LCMPXCHG64, X86::MOV64rr,
8745 X86::NOT64r, X86::RAX,
8746 X86::GR64RegisterClass, true);
8747 case X86::ATOMMIN64:
8748 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8749 case X86::ATOMMAX64:
8750 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8751 case X86::ATOMUMIN64:
8752 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8753 case X86::ATOMUMAX64:
8754 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8756 // This group does 64-bit operations on a 32-bit host.
8757 case X86::ATOMAND6432:
8758 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8759 X86::AND32rr, X86::AND32rr,
8760 X86::AND32ri, X86::AND32ri,
8762 case X86::ATOMOR6432:
8763 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8764 X86::OR32rr, X86::OR32rr,
8765 X86::OR32ri, X86::OR32ri,
8767 case X86::ATOMXOR6432:
8768 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8769 X86::XOR32rr, X86::XOR32rr,
8770 X86::XOR32ri, X86::XOR32ri,
8772 case X86::ATOMNAND6432:
8773 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8774 X86::AND32rr, X86::AND32rr,
8775 X86::AND32ri, X86::AND32ri,
8777 case X86::ATOMADD6432:
8778 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8779 X86::ADD32rr, X86::ADC32rr,
8780 X86::ADD32ri, X86::ADC32ri,
8782 case X86::ATOMSUB6432:
8783 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8784 X86::SUB32rr, X86::SBB32rr,
8785 X86::SUB32ri, X86::SBB32ri,
8787 case X86::ATOMSWAP6432:
8788 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8789 X86::MOV32rr, X86::MOV32rr,
8790 X86::MOV32ri, X86::MOV32ri,
8792 case X86::VASTART_SAVE_XMM_REGS:
8793 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8797 //===----------------------------------------------------------------------===//
8798 // X86 Optimization Hooks
8799 //===----------------------------------------------------------------------===//
8801 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8805 const SelectionDAG &DAG,
8806 unsigned Depth) const {
8807 unsigned Opc = Op.getOpcode();
8808 assert((Opc >= ISD::BUILTIN_OP_END ||
8809 Opc == ISD::INTRINSIC_WO_CHAIN ||
8810 Opc == ISD::INTRINSIC_W_CHAIN ||
8811 Opc == ISD::INTRINSIC_VOID) &&
8812 "Should use MaskedValueIsZero if you don't know whether Op"
8813 " is a target node!");
8815 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8827 // These nodes' second result is a boolean.
8828 if (Op.getResNo() == 0)
8832 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8833 Mask.getBitWidth() - 1);
8838 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8839 /// node is a GlobalAddress + offset.
8840 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8841 GlobalValue* &GA, int64_t &Offset) const{
8842 if (N->getOpcode() == X86ISD::Wrapper) {
8843 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8844 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8845 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8849 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8852 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8853 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8854 /// if the load addresses are consecutive, non-overlapping, and in the right
8856 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8857 const TargetLowering &TLI) {
8858 DebugLoc dl = N->getDebugLoc();
8859 EVT VT = N->getValueType(0);
8860 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8862 if (VT.getSizeInBits() != 128)
8865 SmallVector<SDValue, 16> Elts;
8866 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8867 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8869 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8872 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8873 /// and convert it from being a bunch of shuffles and extracts to a simple
8874 /// store and scalar loads to extract the elements.
8875 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8876 const TargetLowering &TLI) {
8877 SDValue InputVector = N->getOperand(0);
8879 // Only operate on vectors of 4 elements, where the alternative shuffling
8880 // gets to be more expensive.
8881 if (InputVector.getValueType() != MVT::v4i32)
8884 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8885 // single use which is a sign-extend or zero-extend, and all elements are
8887 SmallVector<SDNode *, 4> Uses;
8888 unsigned ExtractedElements = 0;
8889 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8890 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8891 if (UI.getUse().getResNo() != InputVector.getResNo())
8894 SDNode *Extract = *UI;
8895 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8898 if (Extract->getValueType(0) != MVT::i32)
8900 if (!Extract->hasOneUse())
8902 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
8903 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
8905 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
8908 // Record which element was extracted.
8909 ExtractedElements |=
8910 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
8912 Uses.push_back(Extract);
8915 // If not all the elements were used, this may not be worthwhile.
8916 if (ExtractedElements != 15)
8919 // Ok, we've now decided to do the transformation.
8920 DebugLoc dl = InputVector.getDebugLoc();
8922 // Store the value to a temporary stack slot.
8923 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
8924 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
8927 // Replace each use (extract) with a load of the appropriate element.
8928 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
8929 UE = Uses.end(); UI != UE; ++UI) {
8930 SDNode *Extract = *UI;
8932 // Compute the element's address.
8933 SDValue Idx = Extract->getOperand(1);
8935 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
8936 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
8937 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
8939 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
8942 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
8943 NULL, 0, false, false, 0);
8945 // Replace the exact with the load.
8946 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
8949 // The replacement was made in place; don't return anything.
8953 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8954 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8955 const X86Subtarget *Subtarget) {
8956 DebugLoc DL = N->getDebugLoc();
8957 SDValue Cond = N->getOperand(0);
8958 // Get the LHS/RHS of the select.
8959 SDValue LHS = N->getOperand(1);
8960 SDValue RHS = N->getOperand(2);
8962 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8963 // instructions match the semantics of the common C idiom x<y?x:y but not
8964 // x<=y?x:y, because of how they handle negative zero (which can be
8965 // ignored in unsafe-math mode).
8966 if (Subtarget->hasSSE2() &&
8967 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8968 Cond.getOpcode() == ISD::SETCC) {
8969 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8971 unsigned Opcode = 0;
8972 // Check for x CC y ? x : y.
8973 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
8974 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
8978 // Converting this to a min would handle NaNs incorrectly, and swapping
8979 // the operands would cause it to handle comparisons between positive
8980 // and negative zero incorrectly.
8981 if (!FiniteOnlyFPMath() &&
8982 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
8983 if (!UnsafeFPMath &&
8984 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8986 std::swap(LHS, RHS);
8988 Opcode = X86ISD::FMIN;
8991 // Converting this to a min would handle comparisons between positive
8992 // and negative zero incorrectly.
8993 if (!UnsafeFPMath &&
8994 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
8996 Opcode = X86ISD::FMIN;
8999 // Converting this to a min would handle both negative zeros and NaNs
9000 // incorrectly, but we can swap the operands to fix both.
9001 std::swap(LHS, RHS);
9005 Opcode = X86ISD::FMIN;
9009 // Converting this to a max would handle comparisons between positive
9010 // and negative zero incorrectly.
9011 if (!UnsafeFPMath &&
9012 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9014 Opcode = X86ISD::FMAX;
9017 // Converting this to a max would handle NaNs incorrectly, and swapping
9018 // the operands would cause it to handle comparisons between positive
9019 // and negative zero incorrectly.
9020 if (!FiniteOnlyFPMath() &&
9021 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9022 if (!UnsafeFPMath &&
9023 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9025 std::swap(LHS, RHS);
9027 Opcode = X86ISD::FMAX;
9030 // Converting this to a max would handle both negative zeros and NaNs
9031 // incorrectly, but we can swap the operands to fix both.
9032 std::swap(LHS, RHS);
9036 Opcode = X86ISD::FMAX;
9039 // Check for x CC y ? y : x -- a min/max with reversed arms.
9040 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9041 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9045 // Converting this to a min would handle comparisons between positive
9046 // and negative zero incorrectly, and swapping the operands would
9047 // cause it to handle NaNs incorrectly.
9048 if (!UnsafeFPMath &&
9049 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9050 if (!FiniteOnlyFPMath() &&
9051 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9053 std::swap(LHS, RHS);
9055 Opcode = X86ISD::FMIN;
9058 // Converting this to a min would handle NaNs incorrectly.
9059 if (!UnsafeFPMath &&
9060 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9062 Opcode = X86ISD::FMIN;
9065 // Converting this to a min would handle both negative zeros and NaNs
9066 // incorrectly, but we can swap the operands to fix both.
9067 std::swap(LHS, RHS);
9071 Opcode = X86ISD::FMIN;
9075 // Converting this to a max would handle NaNs incorrectly.
9076 if (!FiniteOnlyFPMath() &&
9077 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9079 Opcode = X86ISD::FMAX;
9082 // Converting this to a max would handle comparisons between positive
9083 // and negative zero incorrectly, and swapping the operands would
9084 // cause it to handle NaNs incorrectly.
9085 if (!UnsafeFPMath &&
9086 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9087 if (!FiniteOnlyFPMath() &&
9088 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9090 std::swap(LHS, RHS);
9092 Opcode = X86ISD::FMAX;
9095 // Converting this to a max would handle both negative zeros and NaNs
9096 // incorrectly, but we can swap the operands to fix both.
9097 std::swap(LHS, RHS);
9101 Opcode = X86ISD::FMAX;
9107 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9110 // If this is a select between two integer constants, try to do some
9112 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9113 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9114 // Don't do this for crazy integer types.
9115 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9116 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9117 // so that TrueC (the true value) is larger than FalseC.
9118 bool NeedsCondInvert = false;
9120 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9121 // Efficiently invertible.
9122 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9123 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9124 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9125 NeedsCondInvert = true;
9126 std::swap(TrueC, FalseC);
9129 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9130 if (FalseC->getAPIntValue() == 0 &&
9131 TrueC->getAPIntValue().isPowerOf2()) {
9132 if (NeedsCondInvert) // Invert the condition if needed.
9133 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9134 DAG.getConstant(1, Cond.getValueType()));
9136 // Zero extend the condition if needed.
9137 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9139 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9140 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9141 DAG.getConstant(ShAmt, MVT::i8));
9144 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9145 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9146 if (NeedsCondInvert) // Invert the condition if needed.
9147 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9148 DAG.getConstant(1, Cond.getValueType()));
9150 // Zero extend the condition if needed.
9151 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9152 FalseC->getValueType(0), Cond);
9153 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9154 SDValue(FalseC, 0));
9157 // Optimize cases that will turn into an LEA instruction. This requires
9158 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9159 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9160 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9161 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9163 bool isFastMultiplier = false;
9165 switch ((unsigned char)Diff) {
9167 case 1: // result = add base, cond
9168 case 2: // result = lea base( , cond*2)
9169 case 3: // result = lea base(cond, cond*2)
9170 case 4: // result = lea base( , cond*4)
9171 case 5: // result = lea base(cond, cond*4)
9172 case 8: // result = lea base( , cond*8)
9173 case 9: // result = lea base(cond, cond*8)
9174 isFastMultiplier = true;
9179 if (isFastMultiplier) {
9180 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9181 if (NeedsCondInvert) // Invert the condition if needed.
9182 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9183 DAG.getConstant(1, Cond.getValueType()));
9185 // Zero extend the condition if needed.
9186 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9188 // Scale the condition by the difference.
9190 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9191 DAG.getConstant(Diff, Cond.getValueType()));
9193 // Add the base if non-zero.
9194 if (FalseC->getAPIntValue() != 0)
9195 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9196 SDValue(FalseC, 0));
9206 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9207 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9208 TargetLowering::DAGCombinerInfo &DCI) {
9209 DebugLoc DL = N->getDebugLoc();
9211 // If the flag operand isn't dead, don't touch this CMOV.
9212 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9215 // If this is a select between two integer constants, try to do some
9216 // optimizations. Note that the operands are ordered the opposite of SELECT
9218 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9219 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9220 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9221 // larger than FalseC (the false value).
9222 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9224 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9225 CC = X86::GetOppositeBranchCondition(CC);
9226 std::swap(TrueC, FalseC);
9229 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9230 // This is efficient for any integer data type (including i8/i16) and
9232 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9233 SDValue Cond = N->getOperand(3);
9234 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9235 DAG.getConstant(CC, MVT::i8), Cond);
9237 // Zero extend the condition if needed.
9238 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9240 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9241 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9242 DAG.getConstant(ShAmt, MVT::i8));
9243 if (N->getNumValues() == 2) // Dead flag value?
9244 return DCI.CombineTo(N, Cond, SDValue());
9248 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9249 // for any integer data type, including i8/i16.
9250 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9251 SDValue Cond = N->getOperand(3);
9252 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9253 DAG.getConstant(CC, MVT::i8), Cond);
9255 // Zero extend the condition if needed.
9256 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9257 FalseC->getValueType(0), Cond);
9258 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9259 SDValue(FalseC, 0));
9261 if (N->getNumValues() == 2) // Dead flag value?
9262 return DCI.CombineTo(N, Cond, SDValue());
9266 // Optimize cases that will turn into an LEA instruction. This requires
9267 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9268 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9269 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9270 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9272 bool isFastMultiplier = false;
9274 switch ((unsigned char)Diff) {
9276 case 1: // result = add base, cond
9277 case 2: // result = lea base( , cond*2)
9278 case 3: // result = lea base(cond, cond*2)
9279 case 4: // result = lea base( , cond*4)
9280 case 5: // result = lea base(cond, cond*4)
9281 case 8: // result = lea base( , cond*8)
9282 case 9: // result = lea base(cond, cond*8)
9283 isFastMultiplier = true;
9288 if (isFastMultiplier) {
9289 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9290 SDValue Cond = N->getOperand(3);
9291 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9292 DAG.getConstant(CC, MVT::i8), Cond);
9293 // Zero extend the condition if needed.
9294 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9296 // Scale the condition by the difference.
9298 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9299 DAG.getConstant(Diff, Cond.getValueType()));
9301 // Add the base if non-zero.
9302 if (FalseC->getAPIntValue() != 0)
9303 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9304 SDValue(FalseC, 0));
9305 if (N->getNumValues() == 2) // Dead flag value?
9306 return DCI.CombineTo(N, Cond, SDValue());
9316 /// PerformMulCombine - Optimize a single multiply with constant into two
9317 /// in order to implement it with two cheaper instructions, e.g.
9318 /// LEA + SHL, LEA + LEA.
9319 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9320 TargetLowering::DAGCombinerInfo &DCI) {
9321 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9324 EVT VT = N->getValueType(0);
9328 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9331 uint64_t MulAmt = C->getZExtValue();
9332 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9335 uint64_t MulAmt1 = 0;
9336 uint64_t MulAmt2 = 0;
9337 if ((MulAmt % 9) == 0) {
9339 MulAmt2 = MulAmt / 9;
9340 } else if ((MulAmt % 5) == 0) {
9342 MulAmt2 = MulAmt / 5;
9343 } else if ((MulAmt % 3) == 0) {
9345 MulAmt2 = MulAmt / 3;
9348 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9349 DebugLoc DL = N->getDebugLoc();
9351 if (isPowerOf2_64(MulAmt2) &&
9352 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9353 // If second multiplifer is pow2, issue it first. We want the multiply by
9354 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9356 std::swap(MulAmt1, MulAmt2);
9359 if (isPowerOf2_64(MulAmt1))
9360 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9361 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9363 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9364 DAG.getConstant(MulAmt1, VT));
9366 if (isPowerOf2_64(MulAmt2))
9367 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9368 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9370 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9371 DAG.getConstant(MulAmt2, VT));
9373 // Do not add new nodes to DAG combiner worklist.
9374 DCI.CombineTo(N, NewMul, false);
9379 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9380 SDValue N0 = N->getOperand(0);
9381 SDValue N1 = N->getOperand(1);
9382 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9383 EVT VT = N0.getValueType();
9385 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9386 // since the result of setcc_c is all zero's or all ones.
9387 if (N1C && N0.getOpcode() == ISD::AND &&
9388 N0.getOperand(1).getOpcode() == ISD::Constant) {
9389 SDValue N00 = N0.getOperand(0);
9390 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9391 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9392 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9393 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9394 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9395 APInt ShAmt = N1C->getAPIntValue();
9396 Mask = Mask.shl(ShAmt);
9398 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9399 N00, DAG.getConstant(Mask, VT));
9406 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9408 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9409 const X86Subtarget *Subtarget) {
9410 EVT VT = N->getValueType(0);
9411 if (!VT.isVector() && VT.isInteger() &&
9412 N->getOpcode() == ISD::SHL)
9413 return PerformSHLCombine(N, DAG);
9415 // On X86 with SSE2 support, we can transform this to a vector shift if
9416 // all elements are shifted by the same amount. We can't do this in legalize
9417 // because the a constant vector is typically transformed to a constant pool
9418 // so we have no knowledge of the shift amount.
9419 if (!Subtarget->hasSSE2())
9422 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9425 SDValue ShAmtOp = N->getOperand(1);
9426 EVT EltVT = VT.getVectorElementType();
9427 DebugLoc DL = N->getDebugLoc();
9428 SDValue BaseShAmt = SDValue();
9429 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9430 unsigned NumElts = VT.getVectorNumElements();
9432 for (; i != NumElts; ++i) {
9433 SDValue Arg = ShAmtOp.getOperand(i);
9434 if (Arg.getOpcode() == ISD::UNDEF) continue;
9438 for (; i != NumElts; ++i) {
9439 SDValue Arg = ShAmtOp.getOperand(i);
9440 if (Arg.getOpcode() == ISD::UNDEF) continue;
9441 if (Arg != BaseShAmt) {
9445 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9446 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9447 SDValue InVec = ShAmtOp.getOperand(0);
9448 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9449 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9451 for (; i != NumElts; ++i) {
9452 SDValue Arg = InVec.getOperand(i);
9453 if (Arg.getOpcode() == ISD::UNDEF) continue;
9457 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9458 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9459 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9460 if (C->getZExtValue() == SplatIdx)
9461 BaseShAmt = InVec.getOperand(1);
9464 if (BaseShAmt.getNode() == 0)
9465 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9466 DAG.getIntPtrConstant(0));
9470 // The shift amount is an i32.
9471 if (EltVT.bitsGT(MVT::i32))
9472 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9473 else if (EltVT.bitsLT(MVT::i32))
9474 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9476 // The shift amount is identical so we can do a vector shift.
9477 SDValue ValOp = N->getOperand(0);
9478 switch (N->getOpcode()) {
9480 llvm_unreachable("Unknown shift opcode!");
9483 if (VT == MVT::v2i64)
9484 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9485 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9487 if (VT == MVT::v4i32)
9488 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9489 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9491 if (VT == MVT::v8i16)
9492 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9493 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9497 if (VT == MVT::v4i32)
9498 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9499 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9501 if (VT == MVT::v8i16)
9502 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9503 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9507 if (VT == MVT::v2i64)
9508 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9509 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9511 if (VT == MVT::v4i32)
9512 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9513 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9515 if (VT == MVT::v8i16)
9516 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9517 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9524 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9525 const X86Subtarget *Subtarget) {
9526 EVT VT = N->getValueType(0);
9527 if (VT != MVT::i64 || !Subtarget->is64Bit())
9530 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9531 SDValue N0 = N->getOperand(0);
9532 SDValue N1 = N->getOperand(1);
9533 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9535 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9538 SDValue ShAmt0 = N0.getOperand(1);
9539 if (ShAmt0.getValueType() != MVT::i8)
9541 SDValue ShAmt1 = N1.getOperand(1);
9542 if (ShAmt1.getValueType() != MVT::i8)
9544 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9545 ShAmt0 = ShAmt0.getOperand(0);
9546 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9547 ShAmt1 = ShAmt1.getOperand(0);
9549 DebugLoc DL = N->getDebugLoc();
9550 unsigned Opc = X86ISD::SHLD;
9551 SDValue Op0 = N0.getOperand(0);
9552 SDValue Op1 = N1.getOperand(0);
9553 if (ShAmt0.getOpcode() == ISD::SUB) {
9555 std::swap(Op0, Op1);
9556 std::swap(ShAmt0, ShAmt1);
9559 if (ShAmt1.getOpcode() == ISD::SUB) {
9560 SDValue Sum = ShAmt1.getOperand(0);
9561 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9562 if (SumC->getSExtValue() == 64 &&
9563 ShAmt1.getOperand(1) == ShAmt0)
9564 return DAG.getNode(Opc, DL, VT,
9566 DAG.getNode(ISD::TRUNCATE, DL,
9569 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9570 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9572 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9573 return DAG.getNode(Opc, DL, VT,
9574 N0.getOperand(0), N1.getOperand(0),
9575 DAG.getNode(ISD::TRUNCATE, DL,
9582 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9583 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9584 const X86Subtarget *Subtarget) {
9585 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9586 // the FP state in cases where an emms may be missing.
9587 // A preferable solution to the general problem is to figure out the right
9588 // places to insert EMMS. This qualifies as a quick hack.
9590 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9591 StoreSDNode *St = cast<StoreSDNode>(N);
9592 EVT VT = St->getValue().getValueType();
9593 if (VT.getSizeInBits() != 64)
9596 const Function *F = DAG.getMachineFunction().getFunction();
9597 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9598 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9599 && Subtarget->hasSSE2();
9600 if ((VT.isVector() ||
9601 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9602 isa<LoadSDNode>(St->getValue()) &&
9603 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9604 St->getChain().hasOneUse() && !St->isVolatile()) {
9605 SDNode* LdVal = St->getValue().getNode();
9607 int TokenFactorIndex = -1;
9608 SmallVector<SDValue, 8> Ops;
9609 SDNode* ChainVal = St->getChain().getNode();
9610 // Must be a store of a load. We currently handle two cases: the load
9611 // is a direct child, and it's under an intervening TokenFactor. It is
9612 // possible to dig deeper under nested TokenFactors.
9613 if (ChainVal == LdVal)
9614 Ld = cast<LoadSDNode>(St->getChain());
9615 else if (St->getValue().hasOneUse() &&
9616 ChainVal->getOpcode() == ISD::TokenFactor) {
9617 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9618 if (ChainVal->getOperand(i).getNode() == LdVal) {
9619 TokenFactorIndex = i;
9620 Ld = cast<LoadSDNode>(St->getValue());
9622 Ops.push_back(ChainVal->getOperand(i));
9626 if (!Ld || !ISD::isNormalLoad(Ld))
9629 // If this is not the MMX case, i.e. we are just turning i64 load/store
9630 // into f64 load/store, avoid the transformation if there are multiple
9631 // uses of the loaded value.
9632 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9635 DebugLoc LdDL = Ld->getDebugLoc();
9636 DebugLoc StDL = N->getDebugLoc();
9637 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9638 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9640 if (Subtarget->is64Bit() || F64IsLegal) {
9641 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9642 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9643 Ld->getBasePtr(), Ld->getSrcValue(),
9644 Ld->getSrcValueOffset(), Ld->isVolatile(),
9645 Ld->isNonTemporal(), Ld->getAlignment());
9646 SDValue NewChain = NewLd.getValue(1);
9647 if (TokenFactorIndex != -1) {
9648 Ops.push_back(NewChain);
9649 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9652 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9653 St->getSrcValue(), St->getSrcValueOffset(),
9654 St->isVolatile(), St->isNonTemporal(),
9655 St->getAlignment());
9658 // Otherwise, lower to two pairs of 32-bit loads / stores.
9659 SDValue LoAddr = Ld->getBasePtr();
9660 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9661 DAG.getConstant(4, MVT::i32));
9663 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9664 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9665 Ld->isVolatile(), Ld->isNonTemporal(),
9666 Ld->getAlignment());
9667 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9668 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9669 Ld->isVolatile(), Ld->isNonTemporal(),
9670 MinAlign(Ld->getAlignment(), 4));
9672 SDValue NewChain = LoLd.getValue(1);
9673 if (TokenFactorIndex != -1) {
9674 Ops.push_back(LoLd);
9675 Ops.push_back(HiLd);
9676 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9680 LoAddr = St->getBasePtr();
9681 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9682 DAG.getConstant(4, MVT::i32));
9684 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9685 St->getSrcValue(), St->getSrcValueOffset(),
9686 St->isVolatile(), St->isNonTemporal(),
9687 St->getAlignment());
9688 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9690 St->getSrcValueOffset() + 4,
9692 St->isNonTemporal(),
9693 MinAlign(St->getAlignment(), 4));
9694 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9699 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9700 /// X86ISD::FXOR nodes.
9701 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9702 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9703 // F[X]OR(0.0, x) -> x
9704 // F[X]OR(x, 0.0) -> x
9705 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9706 if (C->getValueAPF().isPosZero())
9707 return N->getOperand(1);
9708 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9709 if (C->getValueAPF().isPosZero())
9710 return N->getOperand(0);
9714 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9715 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9716 // FAND(0.0, x) -> 0.0
9717 // FAND(x, 0.0) -> 0.0
9718 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9719 if (C->getValueAPF().isPosZero())
9720 return N->getOperand(0);
9721 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9722 if (C->getValueAPF().isPosZero())
9723 return N->getOperand(1);
9727 static SDValue PerformBTCombine(SDNode *N,
9729 TargetLowering::DAGCombinerInfo &DCI) {
9730 // BT ignores high bits in the bit index operand.
9731 SDValue Op1 = N->getOperand(1);
9732 if (Op1.hasOneUse()) {
9733 unsigned BitWidth = Op1.getValueSizeInBits();
9734 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9735 APInt KnownZero, KnownOne;
9736 TargetLowering::TargetLoweringOpt TLO(DAG);
9737 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9738 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9739 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9740 DCI.CommitTargetLoweringOpt(TLO);
9745 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9746 SDValue Op = N->getOperand(0);
9747 if (Op.getOpcode() == ISD::BIT_CONVERT)
9748 Op = Op.getOperand(0);
9749 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9750 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9751 VT.getVectorElementType().getSizeInBits() ==
9752 OpVT.getVectorElementType().getSizeInBits()) {
9753 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9758 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9759 // Locked instructions, in turn, have implicit fence semantics (all memory
9760 // operations are flushed before issuing the locked instruction, and the
9761 // are not buffered), so we can fold away the common pattern of
9762 // fence-atomic-fence.
9763 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9764 SDValue atomic = N->getOperand(0);
9765 switch (atomic.getOpcode()) {
9766 case ISD::ATOMIC_CMP_SWAP:
9767 case ISD::ATOMIC_SWAP:
9768 case ISD::ATOMIC_LOAD_ADD:
9769 case ISD::ATOMIC_LOAD_SUB:
9770 case ISD::ATOMIC_LOAD_AND:
9771 case ISD::ATOMIC_LOAD_OR:
9772 case ISD::ATOMIC_LOAD_XOR:
9773 case ISD::ATOMIC_LOAD_NAND:
9774 case ISD::ATOMIC_LOAD_MIN:
9775 case ISD::ATOMIC_LOAD_MAX:
9776 case ISD::ATOMIC_LOAD_UMIN:
9777 case ISD::ATOMIC_LOAD_UMAX:
9783 SDValue fence = atomic.getOperand(0);
9784 if (fence.getOpcode() != ISD::MEMBARRIER)
9787 switch (atomic.getOpcode()) {
9788 case ISD::ATOMIC_CMP_SWAP:
9789 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9790 atomic.getOperand(1), atomic.getOperand(2),
9791 atomic.getOperand(3));
9792 case ISD::ATOMIC_SWAP:
9793 case ISD::ATOMIC_LOAD_ADD:
9794 case ISD::ATOMIC_LOAD_SUB:
9795 case ISD::ATOMIC_LOAD_AND:
9796 case ISD::ATOMIC_LOAD_OR:
9797 case ISD::ATOMIC_LOAD_XOR:
9798 case ISD::ATOMIC_LOAD_NAND:
9799 case ISD::ATOMIC_LOAD_MIN:
9800 case ISD::ATOMIC_LOAD_MAX:
9801 case ISD::ATOMIC_LOAD_UMIN:
9802 case ISD::ATOMIC_LOAD_UMAX:
9803 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9804 atomic.getOperand(1), atomic.getOperand(2));
9810 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9811 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9812 // (and (i32 x86isd::setcc_carry), 1)
9813 // This eliminates the zext. This transformation is necessary because
9814 // ISD::SETCC is always legalized to i8.
9815 DebugLoc dl = N->getDebugLoc();
9816 SDValue N0 = N->getOperand(0);
9817 EVT VT = N->getValueType(0);
9818 if (N0.getOpcode() == ISD::AND &&
9820 N0.getOperand(0).hasOneUse()) {
9821 SDValue N00 = N0.getOperand(0);
9822 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9824 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9825 if (!C || C->getZExtValue() != 1)
9827 return DAG.getNode(ISD::AND, dl, VT,
9828 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9829 N00.getOperand(0), N00.getOperand(1)),
9830 DAG.getConstant(1, VT));
9836 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9837 DAGCombinerInfo &DCI) const {
9838 SelectionDAG &DAG = DCI.DAG;
9839 switch (N->getOpcode()) {
9841 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9842 case ISD::EXTRACT_VECTOR_ELT:
9843 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9844 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9845 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9846 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9849 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9850 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9851 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9853 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9854 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9855 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9856 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9857 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9858 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9864 //===----------------------------------------------------------------------===//
9865 // X86 Inline Assembly Support
9866 //===----------------------------------------------------------------------===//
9868 static bool LowerToBSwap(CallInst *CI) {
9869 // FIXME: this should verify that we are targetting a 486 or better. If not,
9870 // we will turn this bswap into something that will be lowered to logical ops
9871 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9872 // so don't worry about this.
9874 // Verify this is a simple bswap.
9875 if (CI->getNumOperands() != 2 ||
9876 CI->getType() != CI->getOperand(1)->getType() ||
9877 !CI->getType()->isIntegerTy())
9880 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9881 if (!Ty || Ty->getBitWidth() % 16 != 0)
9884 // Okay, we can do this xform, do so now.
9885 const Type *Tys[] = { Ty };
9886 Module *M = CI->getParent()->getParent()->getParent();
9887 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9889 Value *Op = CI->getOperand(1);
9890 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9892 CI->replaceAllUsesWith(Op);
9893 CI->eraseFromParent();
9897 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9898 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9899 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9901 std::string AsmStr = IA->getAsmString();
9903 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9904 SmallVector<StringRef, 4> AsmPieces;
9905 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9907 switch (AsmPieces.size()) {
9908 default: return false;
9910 AsmStr = AsmPieces[0];
9912 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9915 if (AsmPieces.size() == 2 &&
9916 (AsmPieces[0] == "bswap" ||
9917 AsmPieces[0] == "bswapq" ||
9918 AsmPieces[0] == "bswapl") &&
9919 (AsmPieces[1] == "$0" ||
9920 AsmPieces[1] == "${0:q}")) {
9921 // No need to check constraints, nothing other than the equivalent of
9922 // "=r,0" would be valid here.
9923 return LowerToBSwap(CI);
9925 // rorw $$8, ${0:w} --> llvm.bswap.i16
9926 if (CI->getType()->isIntegerTy(16) &&
9927 AsmPieces.size() == 3 &&
9928 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
9929 AsmPieces[1] == "$$8," &&
9930 AsmPieces[2] == "${0:w}" &&
9931 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
9933 const std::string &Constraints = IA->getConstraintString();
9934 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
9935 std::sort(AsmPieces.begin(), AsmPieces.end());
9936 if (AsmPieces.size() == 4 &&
9937 AsmPieces[0] == "~{cc}" &&
9938 AsmPieces[1] == "~{dirflag}" &&
9939 AsmPieces[2] == "~{flags}" &&
9940 AsmPieces[3] == "~{fpsr}") {
9941 return LowerToBSwap(CI);
9946 if (CI->getType()->isIntegerTy(64) &&
9947 Constraints.size() >= 2 &&
9948 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9949 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9950 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9951 SmallVector<StringRef, 4> Words;
9952 SplitString(AsmPieces[0], Words, " \t");
9953 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9955 SplitString(AsmPieces[1], Words, " \t");
9956 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9958 SplitString(AsmPieces[2], Words, " \t,");
9959 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9960 Words[2] == "%edx") {
9961 return LowerToBSwap(CI);
9973 /// getConstraintType - Given a constraint letter, return the type of
9974 /// constraint it is for this target.
9975 X86TargetLowering::ConstraintType
9976 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9977 if (Constraint.size() == 1) {
9978 switch (Constraint[0]) {
9990 return C_RegisterClass;
9998 return TargetLowering::getConstraintType(Constraint);
10001 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10002 /// with another that has more specific requirements based on the type of the
10003 /// corresponding operand.
10004 const char *X86TargetLowering::
10005 LowerXConstraint(EVT ConstraintVT) const {
10006 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10007 // 'f' like normal targets.
10008 if (ConstraintVT.isFloatingPoint()) {
10009 if (Subtarget->hasSSE2())
10011 if (Subtarget->hasSSE1())
10015 return TargetLowering::LowerXConstraint(ConstraintVT);
10018 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10019 /// vector. If it is invalid, don't add anything to Ops.
10020 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10023 std::vector<SDValue>&Ops,
10024 SelectionDAG &DAG) const {
10025 SDValue Result(0, 0);
10027 switch (Constraint) {
10030 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10031 if (C->getZExtValue() <= 31) {
10032 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10038 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10039 if (C->getZExtValue() <= 63) {
10040 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10046 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10047 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10048 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10054 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10055 if (C->getZExtValue() <= 255) {
10056 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10062 // 32-bit signed value
10063 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10064 const ConstantInt *CI = C->getConstantIntValue();
10065 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10066 C->getSExtValue())) {
10067 // Widen to 64 bits here to get it sign extended.
10068 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10071 // FIXME gcc accepts some relocatable values here too, but only in certain
10072 // memory models; it's complicated.
10077 // 32-bit unsigned value
10078 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10079 const ConstantInt *CI = C->getConstantIntValue();
10080 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10081 C->getZExtValue())) {
10082 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10086 // FIXME gcc accepts some relocatable values here too, but only in certain
10087 // memory models; it's complicated.
10091 // Literal immediates are always ok.
10092 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10093 // Widen to 64 bits here to get it sign extended.
10094 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10098 // If we are in non-pic codegen mode, we allow the address of a global (with
10099 // an optional displacement) to be used with 'i'.
10100 GlobalAddressSDNode *GA = 0;
10101 int64_t Offset = 0;
10103 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10105 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10106 Offset += GA->getOffset();
10108 } else if (Op.getOpcode() == ISD::ADD) {
10109 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10110 Offset += C->getZExtValue();
10111 Op = Op.getOperand(0);
10114 } else if (Op.getOpcode() == ISD::SUB) {
10115 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10116 Offset += -C->getZExtValue();
10117 Op = Op.getOperand(0);
10122 // Otherwise, this isn't something we can handle, reject it.
10126 GlobalValue *GV = GA->getGlobal();
10127 // If we require an extra load to get this address, as in PIC mode, we
10128 // can't accept it.
10129 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10130 getTargetMachine())))
10134 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
10136 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10142 if (Result.getNode()) {
10143 Ops.push_back(Result);
10146 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
10150 std::vector<unsigned> X86TargetLowering::
10151 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10153 if (Constraint.size() == 1) {
10154 // FIXME: not handling fp-stack yet!
10155 switch (Constraint[0]) { // GCC X86 Constraint Letters
10156 default: break; // Unknown constraint letter
10157 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10158 if (Subtarget->is64Bit()) {
10159 if (VT == MVT::i32)
10160 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10161 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10162 X86::R10D,X86::R11D,X86::R12D,
10163 X86::R13D,X86::R14D,X86::R15D,
10164 X86::EBP, X86::ESP, 0);
10165 else if (VT == MVT::i16)
10166 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10167 X86::SI, X86::DI, X86::R8W,X86::R9W,
10168 X86::R10W,X86::R11W,X86::R12W,
10169 X86::R13W,X86::R14W,X86::R15W,
10170 X86::BP, X86::SP, 0);
10171 else if (VT == MVT::i8)
10172 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10173 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10174 X86::R10B,X86::R11B,X86::R12B,
10175 X86::R13B,X86::R14B,X86::R15B,
10176 X86::BPL, X86::SPL, 0);
10178 else if (VT == MVT::i64)
10179 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10180 X86::RSI, X86::RDI, X86::R8, X86::R9,
10181 X86::R10, X86::R11, X86::R12,
10182 X86::R13, X86::R14, X86::R15,
10183 X86::RBP, X86::RSP, 0);
10187 // 32-bit fallthrough
10188 case 'Q': // Q_REGS
10189 if (VT == MVT::i32)
10190 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10191 else if (VT == MVT::i16)
10192 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10193 else if (VT == MVT::i8)
10194 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10195 else if (VT == MVT::i64)
10196 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10201 return std::vector<unsigned>();
10204 std::pair<unsigned, const TargetRegisterClass*>
10205 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10207 // First, see if this is a constraint that directly corresponds to an LLVM
10209 if (Constraint.size() == 1) {
10210 // GCC Constraint Letters
10211 switch (Constraint[0]) {
10213 case 'r': // GENERAL_REGS
10214 case 'l': // INDEX_REGS
10216 return std::make_pair(0U, X86::GR8RegisterClass);
10217 if (VT == MVT::i16)
10218 return std::make_pair(0U, X86::GR16RegisterClass);
10219 if (VT == MVT::i32 || !Subtarget->is64Bit())
10220 return std::make_pair(0U, X86::GR32RegisterClass);
10221 return std::make_pair(0U, X86::GR64RegisterClass);
10222 case 'R': // LEGACY_REGS
10224 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10225 if (VT == MVT::i16)
10226 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10227 if (VT == MVT::i32 || !Subtarget->is64Bit())
10228 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10229 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10230 case 'f': // FP Stack registers.
10231 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10232 // value to the correct fpstack register class.
10233 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10234 return std::make_pair(0U, X86::RFP32RegisterClass);
10235 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10236 return std::make_pair(0U, X86::RFP64RegisterClass);
10237 return std::make_pair(0U, X86::RFP80RegisterClass);
10238 case 'y': // MMX_REGS if MMX allowed.
10239 if (!Subtarget->hasMMX()) break;
10240 return std::make_pair(0U, X86::VR64RegisterClass);
10241 case 'Y': // SSE_REGS if SSE2 allowed
10242 if (!Subtarget->hasSSE2()) break;
10244 case 'x': // SSE_REGS if SSE1 allowed
10245 if (!Subtarget->hasSSE1()) break;
10247 switch (VT.getSimpleVT().SimpleTy) {
10249 // Scalar SSE types.
10252 return std::make_pair(0U, X86::FR32RegisterClass);
10255 return std::make_pair(0U, X86::FR64RegisterClass);
10263 return std::make_pair(0U, X86::VR128RegisterClass);
10269 // Use the default implementation in TargetLowering to convert the register
10270 // constraint into a member of a register class.
10271 std::pair<unsigned, const TargetRegisterClass*> Res;
10272 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10274 // Not found as a standard register?
10275 if (Res.second == 0) {
10276 // Map st(0) -> st(7) -> ST0
10277 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10278 tolower(Constraint[1]) == 's' &&
10279 tolower(Constraint[2]) == 't' &&
10280 Constraint[3] == '(' &&
10281 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10282 Constraint[5] == ')' &&
10283 Constraint[6] == '}') {
10285 Res.first = X86::ST0+Constraint[4]-'0';
10286 Res.second = X86::RFP80RegisterClass;
10290 // GCC allows "st(0)" to be called just plain "st".
10291 if (StringRef("{st}").equals_lower(Constraint)) {
10292 Res.first = X86::ST0;
10293 Res.second = X86::RFP80RegisterClass;
10298 if (StringRef("{flags}").equals_lower(Constraint)) {
10299 Res.first = X86::EFLAGS;
10300 Res.second = X86::CCRRegisterClass;
10304 // 'A' means EAX + EDX.
10305 if (Constraint == "A") {
10306 Res.first = X86::EAX;
10307 Res.second = X86::GR32_ADRegisterClass;
10313 // Otherwise, check to see if this is a register class of the wrong value
10314 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10315 // turn into {ax},{dx}.
10316 if (Res.second->hasType(VT))
10317 return Res; // Correct type already, nothing to do.
10319 // All of the single-register GCC register classes map their values onto
10320 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10321 // really want an 8-bit or 32-bit register, map to the appropriate register
10322 // class and return the appropriate register.
10323 if (Res.second == X86::GR16RegisterClass) {
10324 if (VT == MVT::i8) {
10325 unsigned DestReg = 0;
10326 switch (Res.first) {
10328 case X86::AX: DestReg = X86::AL; break;
10329 case X86::DX: DestReg = X86::DL; break;
10330 case X86::CX: DestReg = X86::CL; break;
10331 case X86::BX: DestReg = X86::BL; break;
10334 Res.first = DestReg;
10335 Res.second = X86::GR8RegisterClass;
10337 } else if (VT == MVT::i32) {
10338 unsigned DestReg = 0;
10339 switch (Res.first) {
10341 case X86::AX: DestReg = X86::EAX; break;
10342 case X86::DX: DestReg = X86::EDX; break;
10343 case X86::CX: DestReg = X86::ECX; break;
10344 case X86::BX: DestReg = X86::EBX; break;
10345 case X86::SI: DestReg = X86::ESI; break;
10346 case X86::DI: DestReg = X86::EDI; break;
10347 case X86::BP: DestReg = X86::EBP; break;
10348 case X86::SP: DestReg = X86::ESP; break;
10351 Res.first = DestReg;
10352 Res.second = X86::GR32RegisterClass;
10354 } else if (VT == MVT::i64) {
10355 unsigned DestReg = 0;
10356 switch (Res.first) {
10358 case X86::AX: DestReg = X86::RAX; break;
10359 case X86::DX: DestReg = X86::RDX; break;
10360 case X86::CX: DestReg = X86::RCX; break;
10361 case X86::BX: DestReg = X86::RBX; break;
10362 case X86::SI: DestReg = X86::RSI; break;
10363 case X86::DI: DestReg = X86::RDI; break;
10364 case X86::BP: DestReg = X86::RBP; break;
10365 case X86::SP: DestReg = X86::RSP; break;
10368 Res.first = DestReg;
10369 Res.second = X86::GR64RegisterClass;
10372 } else if (Res.second == X86::FR32RegisterClass ||
10373 Res.second == X86::FR64RegisterClass ||
10374 Res.second == X86::VR128RegisterClass) {
10375 // Handle references to XMM physical registers that got mapped into the
10376 // wrong class. This can happen with constraints like {xmm0} where the
10377 // target independent register mapper will just pick the first match it can
10378 // find, ignoring the required type.
10379 if (VT == MVT::f32)
10380 Res.second = X86::FR32RegisterClass;
10381 else if (VT == MVT::f64)
10382 Res.second = X86::FR64RegisterClass;
10383 else if (X86::VR128RegisterClass->hasType(VT))
10384 Res.second = X86::VR128RegisterClass;