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 "X86MCTargetExpr.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineJumpTableInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/PseudoSourceValue.h"
38 #include "llvm/MC/MCAsmInfo.h"
39 #include "llvm/MC/MCContext.h"
40 #include "llvm/MC/MCSectionMachO.h"
41 #include "llvm/MC/MCSymbol.h"
42 #include "llvm/ADT/BitVector.h"
43 #include "llvm/ADT/SmallSet.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/ADT/StringExtras.h"
46 #include "llvm/ADT/VectorExtras.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/Dwarf.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
54 using namespace dwarf;
56 STATISTIC(NumTailCalls, "Number of tail calls");
59 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
61 // Disable16Bit - 16-bit operations typically have a larger encoding than
62 // corresponding 32-bit instructions, and 16-bit code is slow on some
63 // processors. This is an experimental flag to disable 16-bit operations
64 // (which forces them to be Legalized to 32-bit operations).
66 Disable16Bit("disable-16bit", cl::Hidden,
67 cl::desc("Disable use of 16-bit instructions"));
69 // Forward declarations.
70 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
73 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
74 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
75 default: llvm_unreachable("unknown subtarget type");
76 case X86Subtarget::isDarwin:
77 if (TM.getSubtarget<X86Subtarget>().is64Bit())
78 return new X8664_MachoTargetObjectFile();
80 return new X86_MachoTargetObjectFile();
82 if (TM.getSubtarget<X86Subtarget>().is64Bit())
83 return new X8664_MachoTargetObjectFile();
84 return new TargetLoweringObjectFileMachO();
85 case X86Subtarget::isELF:
86 if (TM.getSubtarget<X86Subtarget>().is64Bit())
87 return new X8664_ELFTargetObjectFile(TM);
88 return new X8632_ELFTargetObjectFile(TM);
89 case X86Subtarget::isMingw:
90 case X86Subtarget::isCygwin:
91 case X86Subtarget::isWindows:
92 return new TargetLoweringObjectFileCOFF();
96 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
97 : TargetLowering(TM, createTLOF(TM)) {
98 Subtarget = &TM.getSubtarget<X86Subtarget>();
99 X86ScalarSSEf64 = Subtarget->hasSSE2();
100 X86ScalarSSEf32 = Subtarget->hasSSE1();
101 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
103 RegInfo = TM.getRegisterInfo();
104 TD = getTargetData();
106 // Set up the TargetLowering object.
108 // X86 is weird, it always uses i8 for shift amounts and setcc results.
109 setShiftAmountType(MVT::i8);
110 setBooleanContents(ZeroOrOneBooleanContent);
111 setSchedulingPreference(SchedulingForRegPressure);
112 setStackPointerRegisterToSaveRestore(X86StackPtr);
114 if (Subtarget->isTargetDarwin()) {
115 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
116 setUseUnderscoreSetJmp(false);
117 setUseUnderscoreLongJmp(false);
118 } else if (Subtarget->isTargetMingw()) {
119 // MS runtime is weird: it exports _setjmp, but longjmp!
120 setUseUnderscoreSetJmp(true);
121 setUseUnderscoreLongJmp(false);
123 setUseUnderscoreSetJmp(true);
124 setUseUnderscoreLongJmp(true);
127 // Set up the register classes.
128 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
130 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
131 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
132 if (Subtarget->is64Bit())
133 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
135 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
137 // We don't accept any truncstore of integer registers.
138 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
140 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
141 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
143 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
144 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
145 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
147 // SETOEQ and SETUNE require checking two conditions.
148 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
149 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
150 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
151 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
152 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
153 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
155 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
157 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
158 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
159 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
161 if (Subtarget->is64Bit()) {
162 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
163 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
164 } else if (!UseSoftFloat) {
165 if (X86ScalarSSEf64) {
166 // We have an impenetrably clever algorithm for ui64->double only.
167 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
169 // We have an algorithm for SSE2, and we turn this into a 64-bit
170 // FILD for other targets.
171 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
174 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
176 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
180 // SSE has no i16 to fp conversion, only i32
181 if (X86ScalarSSEf32) {
182 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
183 // f32 and f64 cases are Legal, f80 case is not
184 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
186 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
187 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
190 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
191 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
194 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
195 // are Legal, f80 is custom lowered.
196 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
197 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
199 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
201 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
202 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
204 if (X86ScalarSSEf32) {
205 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
206 // f32 and f64 cases are Legal, f80 case is not
207 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
209 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
210 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
213 // Handle FP_TO_UINT by promoting the destination to a larger signed
215 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
216 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
217 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
219 if (Subtarget->is64Bit()) {
220 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
221 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
222 } else if (!UseSoftFloat) {
223 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
224 // Expand FP_TO_UINT into a select.
225 // FIXME: We would like to use a Custom expander here eventually to do
226 // the optimal thing for SSE vs. the default expansion in the legalizer.
227 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
229 // With SSE3 we can use fisttpll to convert to a signed i64; without
230 // SSE, we're stuck with a fistpll.
231 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
234 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
235 if (!X86ScalarSSEf64) {
236 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
237 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
240 // Scalar integer divide and remainder are lowered to use operations that
241 // produce two results, to match the available instructions. This exposes
242 // the two-result form to trivial CSE, which is able to combine x/y and x%y
243 // into a single instruction.
245 // Scalar integer multiply-high is also lowered to use two-result
246 // operations, to match the available instructions. However, plain multiply
247 // (low) operations are left as Legal, as there are single-result
248 // instructions for this in x86. Using the two-result multiply instructions
249 // when both high and low results are needed must be arranged by dagcombine.
250 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
251 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
252 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
253 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
254 setOperationAction(ISD::SREM , MVT::i8 , Expand);
255 setOperationAction(ISD::UREM , MVT::i8 , Expand);
256 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
257 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
258 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
259 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
260 setOperationAction(ISD::SREM , MVT::i16 , Expand);
261 setOperationAction(ISD::UREM , MVT::i16 , Expand);
262 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
263 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
264 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
265 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
266 setOperationAction(ISD::SREM , MVT::i32 , Expand);
267 setOperationAction(ISD::UREM , MVT::i32 , Expand);
268 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
269 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
270 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
271 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
272 setOperationAction(ISD::SREM , MVT::i64 , Expand);
273 setOperationAction(ISD::UREM , MVT::i64 , Expand);
275 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
276 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
277 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
278 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
279 if (Subtarget->is64Bit())
280 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
281 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
282 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
283 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
284 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
285 setOperationAction(ISD::FREM , MVT::f32 , Expand);
286 setOperationAction(ISD::FREM , MVT::f64 , Expand);
287 setOperationAction(ISD::FREM , MVT::f80 , Expand);
288 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
290 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
291 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
292 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
293 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
295 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
296 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
298 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
299 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
301 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
302 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
303 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
304 if (Subtarget->is64Bit()) {
305 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
306 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
307 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
310 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
311 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
313 // These should be promoted to a larger select which is supported.
314 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
315 // X86 wants to expand cmov itself.
316 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
318 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
320 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
321 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
322 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
323 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
324 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
325 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
327 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
329 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
330 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
331 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
332 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
333 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
334 if (Subtarget->is64Bit()) {
335 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
336 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
338 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
341 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
342 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
343 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
344 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
345 if (Subtarget->is64Bit())
346 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
347 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
348 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
349 if (Subtarget->is64Bit()) {
350 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
351 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
352 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
353 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
354 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
356 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
357 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
358 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
359 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
360 if (Subtarget->is64Bit()) {
361 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
362 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
363 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
366 if (Subtarget->hasSSE1())
367 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
369 if (!Subtarget->hasSSE2())
370 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
372 // Expand certain atomics
373 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
374 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
375 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
376 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
378 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
379 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
380 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
381 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
383 if (!Subtarget->is64Bit()) {
384 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
385 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
386 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
387 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
388 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
389 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
390 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
393 // FIXME - use subtarget debug flags
394 if (!Subtarget->isTargetDarwin() &&
395 !Subtarget->isTargetELF() &&
396 !Subtarget->isTargetCygMing()) {
397 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
400 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
401 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
402 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
403 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
404 if (Subtarget->is64Bit()) {
405 setExceptionPointerRegister(X86::RAX);
406 setExceptionSelectorRegister(X86::RDX);
408 setExceptionPointerRegister(X86::EAX);
409 setExceptionSelectorRegister(X86::EDX);
411 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
412 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
414 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
416 setOperationAction(ISD::TRAP, MVT::Other, Legal);
418 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
419 setOperationAction(ISD::VASTART , MVT::Other, Custom);
420 setOperationAction(ISD::VAEND , MVT::Other, Expand);
421 if (Subtarget->is64Bit()) {
422 setOperationAction(ISD::VAARG , MVT::Other, Custom);
423 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
425 setOperationAction(ISD::VAARG , MVT::Other, Expand);
426 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
429 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
430 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
431 if (Subtarget->is64Bit())
432 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
433 if (Subtarget->isTargetCygMing())
434 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
436 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
438 if (!UseSoftFloat && X86ScalarSSEf64) {
439 // f32 and f64 use SSE.
440 // Set up the FP register classes.
441 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
442 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
444 // Use ANDPD to simulate FABS.
445 setOperationAction(ISD::FABS , MVT::f64, Custom);
446 setOperationAction(ISD::FABS , MVT::f32, Custom);
448 // Use XORP to simulate FNEG.
449 setOperationAction(ISD::FNEG , MVT::f64, Custom);
450 setOperationAction(ISD::FNEG , MVT::f32, Custom);
452 // Use ANDPD and ORPD to simulate FCOPYSIGN.
453 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
454 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
456 // We don't support sin/cos/fmod
457 setOperationAction(ISD::FSIN , MVT::f64, Expand);
458 setOperationAction(ISD::FCOS , MVT::f64, Expand);
459 setOperationAction(ISD::FSIN , MVT::f32, Expand);
460 setOperationAction(ISD::FCOS , MVT::f32, Expand);
462 // Expand FP immediates into loads from the stack, except for the special
464 addLegalFPImmediate(APFloat(+0.0)); // xorpd
465 addLegalFPImmediate(APFloat(+0.0f)); // xorps
466 } else if (!UseSoftFloat && X86ScalarSSEf32) {
467 // Use SSE for f32, x87 for f64.
468 // Set up the FP register classes.
469 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
470 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
472 // Use ANDPS to simulate FABS.
473 setOperationAction(ISD::FABS , MVT::f32, Custom);
475 // Use XORP to simulate FNEG.
476 setOperationAction(ISD::FNEG , MVT::f32, Custom);
478 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
480 // Use ANDPS and ORPS to simulate FCOPYSIGN.
481 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
482 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
484 // We don't support sin/cos/fmod
485 setOperationAction(ISD::FSIN , MVT::f32, Expand);
486 setOperationAction(ISD::FCOS , MVT::f32, Expand);
488 // Special cases we handle for FP constants.
489 addLegalFPImmediate(APFloat(+0.0f)); // xorps
490 addLegalFPImmediate(APFloat(+0.0)); // FLD0
491 addLegalFPImmediate(APFloat(+1.0)); // FLD1
492 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
493 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
496 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
497 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
499 } else if (!UseSoftFloat) {
500 // f32 and f64 in x87.
501 // Set up the FP register classes.
502 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
503 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
505 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
506 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
507 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
508 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
511 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
512 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
514 addLegalFPImmediate(APFloat(+0.0)); // FLD0
515 addLegalFPImmediate(APFloat(+1.0)); // FLD1
516 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
517 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
518 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
519 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
520 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
521 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
524 // Long double always uses X87.
526 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
527 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
528 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
531 APFloat TmpFlt(+0.0);
532 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
534 addLegalFPImmediate(TmpFlt); // FLD0
536 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
537 APFloat TmpFlt2(+1.0);
538 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
540 addLegalFPImmediate(TmpFlt2); // FLD1
541 TmpFlt2.changeSign();
542 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
546 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
547 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
551 // Always use a library call for pow.
552 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
553 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
554 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
556 setOperationAction(ISD::FLOG, MVT::f80, Expand);
557 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
558 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
559 setOperationAction(ISD::FEXP, MVT::f80, Expand);
560 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
562 // First set operation action for all vector types to either promote
563 // (for widening) or expand (for scalarization). Then we will selectively
564 // turn on ones that can be effectively codegen'd.
565 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
566 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
567 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
582 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
583 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
607 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
608 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
609 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
610 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
611 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
612 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
613 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
614 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
615 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
616 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
617 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
618 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
619 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
620 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
621 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
622 setTruncStoreAction((MVT::SimpleValueType)VT,
623 (MVT::SimpleValueType)InnerVT, Expand);
624 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
625 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
626 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
629 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
630 // with -msoft-float, disable use of MMX as well.
631 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
632 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
633 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
634 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
635 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
636 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
638 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
639 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
640 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
641 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
643 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
644 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
645 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
646 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
648 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
649 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
651 setOperationAction(ISD::AND, MVT::v8i8, Promote);
652 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
653 setOperationAction(ISD::AND, MVT::v4i16, Promote);
654 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
655 setOperationAction(ISD::AND, MVT::v2i32, Promote);
656 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
657 setOperationAction(ISD::AND, MVT::v1i64, Legal);
659 setOperationAction(ISD::OR, MVT::v8i8, Promote);
660 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
661 setOperationAction(ISD::OR, MVT::v4i16, Promote);
662 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
663 setOperationAction(ISD::OR, MVT::v2i32, Promote);
664 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
665 setOperationAction(ISD::OR, MVT::v1i64, Legal);
667 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
668 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
669 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
670 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
671 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
672 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
673 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
675 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
676 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
677 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
678 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
679 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
680 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
681 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
682 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
683 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
685 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
686 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
687 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
688 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
689 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
691 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
692 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
693 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
694 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
696 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
697 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
698 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
699 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
701 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
703 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
704 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
705 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
706 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
707 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
708 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
709 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
712 if (!UseSoftFloat && Subtarget->hasSSE1()) {
713 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
715 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
716 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
717 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
718 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
719 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
720 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
721 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
722 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
723 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
724 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
725 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
726 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
729 if (!UseSoftFloat && Subtarget->hasSSE2()) {
730 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
732 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
733 // registers cannot be used even for integer operations.
734 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
735 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
736 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
737 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
739 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
740 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
741 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
742 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
743 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
744 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
745 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
746 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
747 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
748 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
749 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
750 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
751 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
752 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
753 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
754 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
756 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
757 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
758 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
759 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
761 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
762 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
763 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
764 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
765 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
767 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
768 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
769 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
770 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
771 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
773 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
774 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
775 EVT VT = (MVT::SimpleValueType)i;
776 // Do not attempt to custom lower non-power-of-2 vectors
777 if (!isPowerOf2_32(VT.getVectorNumElements()))
779 // Do not attempt to custom lower non-128-bit vectors
780 if (!VT.is128BitVector())
782 setOperationAction(ISD::BUILD_VECTOR,
783 VT.getSimpleVT().SimpleTy, Custom);
784 setOperationAction(ISD::VECTOR_SHUFFLE,
785 VT.getSimpleVT().SimpleTy, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
787 VT.getSimpleVT().SimpleTy, Custom);
790 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
791 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
792 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
793 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
794 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
795 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
797 if (Subtarget->is64Bit()) {
798 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
799 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
802 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
803 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
804 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
807 // Do not attempt to promote non-128-bit vectors
808 if (!VT.is128BitVector()) {
811 setOperationAction(ISD::AND, SVT, Promote);
812 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
813 setOperationAction(ISD::OR, SVT, Promote);
814 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
815 setOperationAction(ISD::XOR, SVT, Promote);
816 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
817 setOperationAction(ISD::LOAD, SVT, Promote);
818 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
819 setOperationAction(ISD::SELECT, SVT, Promote);
820 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
823 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
825 // Custom lower v2i64 and v2f64 selects.
826 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
827 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
828 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
829 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
831 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
832 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
833 if (!DisableMMX && Subtarget->hasMMX()) {
834 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
835 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
839 if (Subtarget->hasSSE41()) {
840 // FIXME: Do we need to handle scalar-to-vector here?
841 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
843 // i8 and i16 vectors are custom , because the source register and source
844 // source memory operand types are not the same width. f32 vectors are
845 // custom since the immediate controlling the insert encodes additional
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
848 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
850 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
852 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
853 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
854 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
855 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
857 if (Subtarget->is64Bit()) {
858 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
859 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
863 if (Subtarget->hasSSE42()) {
864 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
867 if (!UseSoftFloat && Subtarget->hasAVX()) {
868 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
869 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
870 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
871 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
873 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
874 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
875 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
876 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
877 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
878 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
879 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
880 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
881 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
882 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
883 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
884 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
885 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
886 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
887 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
889 // Operations to consider commented out -v16i16 v32i8
890 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
891 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
892 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
893 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
894 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
895 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
896 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
897 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
898 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
899 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
900 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
901 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
902 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
903 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
905 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
906 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
907 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
908 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
910 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
911 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
912 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
913 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
914 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
916 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
917 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
918 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
919 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
920 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
921 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
924 // Not sure we want to do this since there are no 256-bit integer
927 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
928 // This includes 256-bit vectors
929 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
930 EVT VT = (MVT::SimpleValueType)i;
932 // Do not attempt to custom lower non-power-of-2 vectors
933 if (!isPowerOf2_32(VT.getVectorNumElements()))
936 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
937 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
938 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
941 if (Subtarget->is64Bit()) {
942 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
943 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
948 // Not sure we want to do this since there are no 256-bit integer
951 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
952 // Including 256-bit vectors
953 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
954 EVT VT = (MVT::SimpleValueType)i;
956 if (!VT.is256BitVector()) {
959 setOperationAction(ISD::AND, VT, Promote);
960 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
961 setOperationAction(ISD::OR, VT, Promote);
962 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
963 setOperationAction(ISD::XOR, VT, Promote);
964 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
965 setOperationAction(ISD::LOAD, VT, Promote);
966 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
967 setOperationAction(ISD::SELECT, VT, Promote);
968 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
971 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
975 // We want to custom lower some of our intrinsics.
976 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
978 // Add/Sub/Mul with overflow operations are custom lowered.
979 setOperationAction(ISD::SADDO, MVT::i32, Custom);
980 setOperationAction(ISD::SADDO, MVT::i64, Custom);
981 setOperationAction(ISD::UADDO, MVT::i32, Custom);
982 setOperationAction(ISD::UADDO, MVT::i64, Custom);
983 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
984 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
985 setOperationAction(ISD::USUBO, MVT::i32, Custom);
986 setOperationAction(ISD::USUBO, MVT::i64, Custom);
987 setOperationAction(ISD::SMULO, MVT::i32, Custom);
988 setOperationAction(ISD::SMULO, MVT::i64, Custom);
990 if (!Subtarget->is64Bit()) {
991 // These libcalls are not available in 32-bit.
992 setLibcallName(RTLIB::SHL_I128, 0);
993 setLibcallName(RTLIB::SRL_I128, 0);
994 setLibcallName(RTLIB::SRA_I128, 0);
997 // We have target-specific dag combine patterns for the following nodes:
998 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
999 setTargetDAGCombine(ISD::BUILD_VECTOR);
1000 setTargetDAGCombine(ISD::SELECT);
1001 setTargetDAGCombine(ISD::SHL);
1002 setTargetDAGCombine(ISD::SRA);
1003 setTargetDAGCombine(ISD::SRL);
1004 setTargetDAGCombine(ISD::OR);
1005 setTargetDAGCombine(ISD::STORE);
1006 setTargetDAGCombine(ISD::MEMBARRIER);
1007 setTargetDAGCombine(ISD::ZERO_EXTEND);
1008 if (Subtarget->is64Bit())
1009 setTargetDAGCombine(ISD::MUL);
1011 computeRegisterProperties();
1013 // FIXME: These should be based on subtarget info. Plus, the values should
1014 // be smaller when we are in optimizing for size mode.
1015 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1016 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1017 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1018 setPrefLoopAlignment(16);
1019 benefitFromCodePlacementOpt = true;
1023 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1028 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1029 /// the desired ByVal argument alignment.
1030 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1033 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1034 if (VTy->getBitWidth() == 128)
1036 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1037 unsigned EltAlign = 0;
1038 getMaxByValAlign(ATy->getElementType(), EltAlign);
1039 if (EltAlign > MaxAlign)
1040 MaxAlign = EltAlign;
1041 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1042 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1043 unsigned EltAlign = 0;
1044 getMaxByValAlign(STy->getElementType(i), EltAlign);
1045 if (EltAlign > MaxAlign)
1046 MaxAlign = EltAlign;
1054 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1055 /// function arguments in the caller parameter area. For X86, aggregates
1056 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1057 /// are at 4-byte boundaries.
1058 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1059 if (Subtarget->is64Bit()) {
1060 // Max of 8 and alignment of type.
1061 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1068 if (Subtarget->hasSSE1())
1069 getMaxByValAlign(Ty, Align);
1073 /// getOptimalMemOpType - Returns the target specific optimal type for load
1074 /// and store operations as a result of memset, memcpy, and memmove
1075 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1078 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1079 bool isSrcConst, bool isSrcStr,
1080 SelectionDAG &DAG) const {
1081 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1082 // linux. This is because the stack realignment code can't handle certain
1083 // cases like PR2962. This should be removed when PR2962 is fixed.
1084 const Function *F = DAG.getMachineFunction().getFunction();
1085 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1086 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1087 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1089 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1092 if (Subtarget->is64Bit() && Size >= 8)
1097 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1098 /// current function. The returned value is a member of the
1099 /// MachineJumpTableInfo::JTEntryKind enum.
1100 unsigned X86TargetLowering::getJumpTableEncoding() const {
1101 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1103 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1104 Subtarget->isPICStyleGOT())
1105 return MachineJumpTableInfo::EK_Custom32;
1107 // Otherwise, use the normal jump table encoding heuristics.
1108 return TargetLowering::getJumpTableEncoding();
1111 /// getPICBaseSymbol - Return the X86-32 PIC base.
1113 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1114 MCContext &Ctx) const {
1115 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1116 return Ctx.GetOrCreateTemporarySymbol(Twine(MAI.getPrivateGlobalPrefix())+
1117 Twine(MF->getFunctionNumber())+"$pb");
1122 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1123 const MachineBasicBlock *MBB,
1124 unsigned uid,MCContext &Ctx) const{
1125 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1126 Subtarget->isPICStyleGOT());
1127 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1129 return X86MCTargetExpr::Create(MBB->getSymbol(),
1130 X86MCTargetExpr::GOTOFF, Ctx);
1133 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1135 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1136 SelectionDAG &DAG) const {
1137 if (!Subtarget->is64Bit())
1138 // This doesn't have DebugLoc associated with it, but is not really the
1139 // same as a Register.
1140 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1145 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1146 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1148 const MCExpr *X86TargetLowering::
1149 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1150 MCContext &Ctx) const {
1151 // X86-64 uses RIP relative addressing based on the jump table label.
1152 if (Subtarget->isPICStyleRIPRel())
1153 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1155 // Otherwise, the reference is relative to the PIC base.
1156 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1159 /// getFunctionAlignment - Return the Log2 alignment of this function.
1160 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1161 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1164 //===----------------------------------------------------------------------===//
1165 // Return Value Calling Convention Implementation
1166 //===----------------------------------------------------------------------===//
1168 #include "X86GenCallingConv.inc"
1171 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1172 const SmallVectorImpl<EVT> &OutTys,
1173 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1174 SelectionDAG &DAG) {
1175 SmallVector<CCValAssign, 16> RVLocs;
1176 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1177 RVLocs, *DAG.getContext());
1178 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1182 X86TargetLowering::LowerReturn(SDValue Chain,
1183 CallingConv::ID CallConv, bool isVarArg,
1184 const SmallVectorImpl<ISD::OutputArg> &Outs,
1185 DebugLoc dl, SelectionDAG &DAG) {
1187 SmallVector<CCValAssign, 16> RVLocs;
1188 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1189 RVLocs, *DAG.getContext());
1190 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1192 // Add the regs to the liveout set for the function.
1193 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1194 for (unsigned i = 0; i != RVLocs.size(); ++i)
1195 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1196 MRI.addLiveOut(RVLocs[i].getLocReg());
1200 SmallVector<SDValue, 6> RetOps;
1201 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1202 // Operand #1 = Bytes To Pop
1203 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1205 // Copy the result values into the output registers.
1206 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1207 CCValAssign &VA = RVLocs[i];
1208 assert(VA.isRegLoc() && "Can only return in registers!");
1209 SDValue ValToCopy = Outs[i].Val;
1211 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1212 // the RET instruction and handled by the FP Stackifier.
1213 if (VA.getLocReg() == X86::ST0 ||
1214 VA.getLocReg() == X86::ST1) {
1215 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1216 // change the value to the FP stack register class.
1217 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1218 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1219 RetOps.push_back(ValToCopy);
1220 // Don't emit a copytoreg.
1224 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1225 // which is returned in RAX / RDX.
1226 if (Subtarget->is64Bit()) {
1227 EVT ValVT = ValToCopy.getValueType();
1228 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1229 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1230 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1231 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1235 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1236 Flag = Chain.getValue(1);
1239 // The x86-64 ABI for returning structs by value requires that we copy
1240 // the sret argument into %rax for the return. We saved the argument into
1241 // a virtual register in the entry block, so now we copy the value out
1243 if (Subtarget->is64Bit() &&
1244 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1245 MachineFunction &MF = DAG.getMachineFunction();
1246 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1247 unsigned Reg = FuncInfo->getSRetReturnReg();
1249 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1250 FuncInfo->setSRetReturnReg(Reg);
1252 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1254 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1255 Flag = Chain.getValue(1);
1257 // RAX now acts like a return value.
1258 MRI.addLiveOut(X86::RAX);
1261 RetOps[0] = Chain; // Update chain.
1263 // Add the flag if we have it.
1265 RetOps.push_back(Flag);
1267 return DAG.getNode(X86ISD::RET_FLAG, dl,
1268 MVT::Other, &RetOps[0], RetOps.size());
1271 /// LowerCallResult - Lower the result values of a call into the
1272 /// appropriate copies out of appropriate physical registers.
1275 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1276 CallingConv::ID CallConv, bool isVarArg,
1277 const SmallVectorImpl<ISD::InputArg> &Ins,
1278 DebugLoc dl, SelectionDAG &DAG,
1279 SmallVectorImpl<SDValue> &InVals) {
1281 // Assign locations to each value returned by this call.
1282 SmallVector<CCValAssign, 16> RVLocs;
1283 bool Is64Bit = Subtarget->is64Bit();
1284 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1285 RVLocs, *DAG.getContext());
1286 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1288 // Copy all of the result registers out of their specified physreg.
1289 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1290 CCValAssign &VA = RVLocs[i];
1291 EVT CopyVT = VA.getValVT();
1293 // If this is x86-64, and we disabled SSE, we can't return FP values
1294 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1295 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1296 llvm_report_error("SSE register return with SSE disabled");
1299 // If this is a call to a function that returns an fp value on the floating
1300 // point stack, but where we prefer to use the value in xmm registers, copy
1301 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1302 if ((VA.getLocReg() == X86::ST0 ||
1303 VA.getLocReg() == X86::ST1) &&
1304 isScalarFPTypeInSSEReg(VA.getValVT())) {
1309 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1310 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1311 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1312 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1313 MVT::v2i64, InFlag).getValue(1);
1314 Val = Chain.getValue(0);
1315 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1316 Val, DAG.getConstant(0, MVT::i64));
1318 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1319 MVT::i64, InFlag).getValue(1);
1320 Val = Chain.getValue(0);
1322 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1324 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1325 CopyVT, InFlag).getValue(1);
1326 Val = Chain.getValue(0);
1328 InFlag = Chain.getValue(2);
1330 if (CopyVT != VA.getValVT()) {
1331 // Round the F80 the right size, which also moves to the appropriate xmm
1333 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1334 // This truncation won't change the value.
1335 DAG.getIntPtrConstant(1));
1338 InVals.push_back(Val);
1345 //===----------------------------------------------------------------------===//
1346 // C & StdCall & Fast Calling Convention implementation
1347 //===----------------------------------------------------------------------===//
1348 // StdCall calling convention seems to be standard for many Windows' API
1349 // routines and around. It differs from C calling convention just a little:
1350 // callee should clean up the stack, not caller. Symbols should be also
1351 // decorated in some fancy way :) It doesn't support any vector arguments.
1352 // For info on fast calling convention see Fast Calling Convention (tail call)
1353 // implementation LowerX86_32FastCCCallTo.
1355 /// CallIsStructReturn - Determines whether a call uses struct return
1357 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1361 return Outs[0].Flags.isSRet();
1364 /// ArgsAreStructReturn - Determines whether a function uses struct
1365 /// return semantics.
1367 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1371 return Ins[0].Flags.isSRet();
1374 /// IsCalleePop - Determines whether the callee is required to pop its
1375 /// own arguments. Callee pop is necessary to support tail calls.
1376 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1380 switch (CallingConv) {
1383 case CallingConv::X86_StdCall:
1384 return !Subtarget->is64Bit();
1385 case CallingConv::X86_FastCall:
1386 return !Subtarget->is64Bit();
1387 case CallingConv::Fast:
1388 return GuaranteedTailCallOpt;
1389 case CallingConv::GHC:
1390 return GuaranteedTailCallOpt;
1394 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1395 /// given CallingConvention value.
1396 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1397 if (Subtarget->is64Bit()) {
1398 if (CC == CallingConv::GHC)
1399 return CC_X86_64_GHC;
1400 else if (Subtarget->isTargetWin64())
1401 return CC_X86_Win64_C;
1406 if (CC == CallingConv::X86_FastCall)
1407 return CC_X86_32_FastCall;
1408 else if (CC == CallingConv::Fast)
1409 return CC_X86_32_FastCC;
1410 else if (CC == CallingConv::GHC)
1411 return CC_X86_32_GHC;
1416 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1417 /// by "Src" to address "Dst" with size and alignment information specified by
1418 /// the specific parameter attribute. The copy will be passed as a byval
1419 /// function parameter.
1421 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1422 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1424 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1425 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1426 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1429 /// IsTailCallConvention - Return true if the calling convention is one that
1430 /// supports tail call optimization.
1431 static bool IsTailCallConvention(CallingConv::ID CC) {
1432 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1435 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1436 /// a tailcall target by changing its ABI.
1437 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1438 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1442 X86TargetLowering::LowerMemArgument(SDValue Chain,
1443 CallingConv::ID CallConv,
1444 const SmallVectorImpl<ISD::InputArg> &Ins,
1445 DebugLoc dl, SelectionDAG &DAG,
1446 const CCValAssign &VA,
1447 MachineFrameInfo *MFI,
1449 // Create the nodes corresponding to a load from this parameter slot.
1450 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1451 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1452 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1455 // If value is passed by pointer we have address passed instead of the value
1457 if (VA.getLocInfo() == CCValAssign::Indirect)
1458 ValVT = VA.getLocVT();
1460 ValVT = VA.getValVT();
1462 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1463 // changed with more analysis.
1464 // In case of tail call optimization mark all arguments mutable. Since they
1465 // could be overwritten by lowering of arguments in case of a tail call.
1466 if (Flags.isByVal()) {
1467 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1468 VA.getLocMemOffset(), isImmutable, false);
1469 return DAG.getFrameIndex(FI, getPointerTy());
1471 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1472 VA.getLocMemOffset(), isImmutable, false);
1473 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1474 return DAG.getLoad(ValVT, dl, Chain, FIN,
1475 PseudoSourceValue::getFixedStack(FI), 0,
1481 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1482 CallingConv::ID CallConv,
1484 const SmallVectorImpl<ISD::InputArg> &Ins,
1487 SmallVectorImpl<SDValue> &InVals) {
1488 MachineFunction &MF = DAG.getMachineFunction();
1489 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1491 const Function* Fn = MF.getFunction();
1492 if (Fn->hasExternalLinkage() &&
1493 Subtarget->isTargetCygMing() &&
1494 Fn->getName() == "main")
1495 FuncInfo->setForceFramePointer(true);
1497 MachineFrameInfo *MFI = MF.getFrameInfo();
1498 bool Is64Bit = Subtarget->is64Bit();
1499 bool IsWin64 = Subtarget->isTargetWin64();
1501 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1502 "Var args not supported with calling convention fastcc or ghc");
1504 // Assign locations to all of the incoming arguments.
1505 SmallVector<CCValAssign, 16> ArgLocs;
1506 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1507 ArgLocs, *DAG.getContext());
1508 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1510 unsigned LastVal = ~0U;
1512 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1513 CCValAssign &VA = ArgLocs[i];
1514 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1516 assert(VA.getValNo() != LastVal &&
1517 "Don't support value assigned to multiple locs yet");
1518 LastVal = VA.getValNo();
1520 if (VA.isRegLoc()) {
1521 EVT RegVT = VA.getLocVT();
1522 TargetRegisterClass *RC = NULL;
1523 if (RegVT == MVT::i32)
1524 RC = X86::GR32RegisterClass;
1525 else if (Is64Bit && RegVT == MVT::i64)
1526 RC = X86::GR64RegisterClass;
1527 else if (RegVT == MVT::f32)
1528 RC = X86::FR32RegisterClass;
1529 else if (RegVT == MVT::f64)
1530 RC = X86::FR64RegisterClass;
1531 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1532 RC = X86::VR128RegisterClass;
1533 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1534 RC = X86::VR64RegisterClass;
1536 llvm_unreachable("Unknown argument type!");
1538 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1539 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1541 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1542 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1544 if (VA.getLocInfo() == CCValAssign::SExt)
1545 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1546 DAG.getValueType(VA.getValVT()));
1547 else if (VA.getLocInfo() == CCValAssign::ZExt)
1548 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1549 DAG.getValueType(VA.getValVT()));
1550 else if (VA.getLocInfo() == CCValAssign::BCvt)
1551 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1553 if (VA.isExtInLoc()) {
1554 // Handle MMX values passed in XMM regs.
1555 if (RegVT.isVector()) {
1556 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1557 ArgValue, DAG.getConstant(0, MVT::i64));
1558 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1560 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1563 assert(VA.isMemLoc());
1564 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1567 // If value is passed via pointer - do a load.
1568 if (VA.getLocInfo() == CCValAssign::Indirect)
1569 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1572 InVals.push_back(ArgValue);
1575 // The x86-64 ABI for returning structs by value requires that we copy
1576 // the sret argument into %rax for the return. Save the argument into
1577 // a virtual register so that we can access it from the return points.
1578 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1579 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1580 unsigned Reg = FuncInfo->getSRetReturnReg();
1582 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1583 FuncInfo->setSRetReturnReg(Reg);
1585 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1586 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1589 unsigned StackSize = CCInfo.getNextStackOffset();
1590 // Align stack specially for tail calls.
1591 if (FuncIsMadeTailCallSafe(CallConv))
1592 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1594 // If the function takes variable number of arguments, make a frame index for
1595 // the start of the first vararg value... for expansion of llvm.va_start.
1597 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1598 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1601 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1603 // FIXME: We should really autogenerate these arrays
1604 static const unsigned GPR64ArgRegsWin64[] = {
1605 X86::RCX, X86::RDX, X86::R8, X86::R9
1607 static const unsigned XMMArgRegsWin64[] = {
1608 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1610 static const unsigned GPR64ArgRegs64Bit[] = {
1611 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1613 static const unsigned XMMArgRegs64Bit[] = {
1614 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1615 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1617 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1620 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1621 GPR64ArgRegs = GPR64ArgRegsWin64;
1622 XMMArgRegs = XMMArgRegsWin64;
1624 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1625 GPR64ArgRegs = GPR64ArgRegs64Bit;
1626 XMMArgRegs = XMMArgRegs64Bit;
1628 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1630 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1633 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1634 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1635 "SSE register cannot be used when SSE is disabled!");
1636 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1637 "SSE register cannot be used when SSE is disabled!");
1638 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1639 // Kernel mode asks for SSE to be disabled, so don't push them
1641 TotalNumXMMRegs = 0;
1643 // For X86-64, if there are vararg parameters that are passed via
1644 // registers, then we must store them to their spots on the stack so they
1645 // may be loaded by deferencing the result of va_next.
1646 VarArgsGPOffset = NumIntRegs * 8;
1647 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1648 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1649 TotalNumXMMRegs * 16, 16,
1652 // Store the integer parameter registers.
1653 SmallVector<SDValue, 8> MemOps;
1654 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1655 unsigned Offset = VarArgsGPOffset;
1656 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1657 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1658 DAG.getIntPtrConstant(Offset));
1659 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1660 X86::GR64RegisterClass);
1661 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1663 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1664 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1665 Offset, false, false, 0);
1666 MemOps.push_back(Store);
1670 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1671 // Now store the XMM (fp + vector) parameter registers.
1672 SmallVector<SDValue, 11> SaveXMMOps;
1673 SaveXMMOps.push_back(Chain);
1675 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1676 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1677 SaveXMMOps.push_back(ALVal);
1679 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1680 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1682 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1683 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1684 X86::VR128RegisterClass);
1685 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1686 SaveXMMOps.push_back(Val);
1688 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1690 &SaveXMMOps[0], SaveXMMOps.size()));
1693 if (!MemOps.empty())
1694 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1695 &MemOps[0], MemOps.size());
1699 // Some CCs need callee pop.
1700 if (IsCalleePop(isVarArg, CallConv)) {
1701 BytesToPopOnReturn = StackSize; // Callee pops everything.
1703 BytesToPopOnReturn = 0; // Callee pops nothing.
1704 // If this is an sret function, the return should pop the hidden pointer.
1705 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1706 BytesToPopOnReturn = 4;
1710 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1711 if (CallConv == CallingConv::X86_FastCall)
1712 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1715 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1721 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1722 SDValue StackPtr, SDValue Arg,
1723 DebugLoc dl, SelectionDAG &DAG,
1724 const CCValAssign &VA,
1725 ISD::ArgFlagsTy Flags) {
1726 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1727 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1728 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1729 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1730 if (Flags.isByVal()) {
1731 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1733 return DAG.getStore(Chain, dl, Arg, PtrOff,
1734 PseudoSourceValue::getStack(), LocMemOffset,
1738 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1739 /// optimization is performed and it is required.
1741 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1742 SDValue &OutRetAddr, SDValue Chain,
1743 bool IsTailCall, bool Is64Bit,
1744 int FPDiff, DebugLoc dl) {
1745 // Adjust the Return address stack slot.
1746 EVT VT = getPointerTy();
1747 OutRetAddr = getReturnAddressFrameIndex(DAG);
1749 // Load the "old" Return address.
1750 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1751 return SDValue(OutRetAddr.getNode(), 1);
1754 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1755 /// optimization is performed and it is required (FPDiff!=0).
1757 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1758 SDValue Chain, SDValue RetAddrFrIdx,
1759 bool Is64Bit, int FPDiff, DebugLoc dl) {
1760 // Store the return address to the appropriate stack slot.
1761 if (!FPDiff) return Chain;
1762 // Calculate the new stack slot for the return address.
1763 int SlotSize = Is64Bit ? 8 : 4;
1764 int NewReturnAddrFI =
1765 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1766 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1767 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1768 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1769 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1775 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1776 CallingConv::ID CallConv, bool isVarArg,
1778 const SmallVectorImpl<ISD::OutputArg> &Outs,
1779 const SmallVectorImpl<ISD::InputArg> &Ins,
1780 DebugLoc dl, SelectionDAG &DAG,
1781 SmallVectorImpl<SDValue> &InVals) {
1782 MachineFunction &MF = DAG.getMachineFunction();
1783 bool Is64Bit = Subtarget->is64Bit();
1784 bool IsStructRet = CallIsStructReturn(Outs);
1785 bool IsSibcall = false;
1788 // Check if it's really possible to do a tail call.
1789 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1790 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1793 // Sibcalls are automatically detected tailcalls which do not require
1795 if (!GuaranteedTailCallOpt && isTailCall)
1802 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1803 "Var args not supported with calling convention fastcc or ghc");
1805 // Analyze operands of the call, assigning locations to each operand.
1806 SmallVector<CCValAssign, 16> ArgLocs;
1807 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1808 ArgLocs, *DAG.getContext());
1809 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1811 // Get a count of how many bytes are to be pushed on the stack.
1812 unsigned NumBytes = CCInfo.getNextStackOffset();
1814 // This is a sibcall. The memory operands are available in caller's
1815 // own caller's stack.
1817 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1818 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1821 if (isTailCall && !IsSibcall) {
1822 // Lower arguments at fp - stackoffset + fpdiff.
1823 unsigned NumBytesCallerPushed =
1824 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1825 FPDiff = NumBytesCallerPushed - NumBytes;
1827 // Set the delta of movement of the returnaddr stackslot.
1828 // But only set if delta is greater than previous delta.
1829 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1830 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1834 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1836 SDValue RetAddrFrIdx;
1837 // Load return adress for tail calls.
1838 if (isTailCall && FPDiff)
1839 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1840 Is64Bit, FPDiff, dl);
1842 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1843 SmallVector<SDValue, 8> MemOpChains;
1846 // Walk the register/memloc assignments, inserting copies/loads. In the case
1847 // of tail call optimization arguments are handle later.
1848 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1849 CCValAssign &VA = ArgLocs[i];
1850 EVT RegVT = VA.getLocVT();
1851 SDValue Arg = Outs[i].Val;
1852 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1853 bool isByVal = Flags.isByVal();
1855 // Promote the value if needed.
1856 switch (VA.getLocInfo()) {
1857 default: llvm_unreachable("Unknown loc info!");
1858 case CCValAssign::Full: break;
1859 case CCValAssign::SExt:
1860 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1862 case CCValAssign::ZExt:
1863 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1865 case CCValAssign::AExt:
1866 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1867 // Special case: passing MMX values in XMM registers.
1868 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1869 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1870 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1872 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1874 case CCValAssign::BCvt:
1875 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1877 case CCValAssign::Indirect: {
1878 // Store the argument.
1879 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1880 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1881 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1882 PseudoSourceValue::getFixedStack(FI), 0,
1889 if (VA.isRegLoc()) {
1890 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1891 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1892 assert(VA.isMemLoc());
1893 if (StackPtr.getNode() == 0)
1894 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1895 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1896 dl, DAG, VA, Flags));
1900 if (!MemOpChains.empty())
1901 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1902 &MemOpChains[0], MemOpChains.size());
1904 // Build a sequence of copy-to-reg nodes chained together with token chain
1905 // and flag operands which copy the outgoing args into registers.
1907 // Tail call byval lowering might overwrite argument registers so in case of
1908 // tail call optimization the copies to registers are lowered later.
1910 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1911 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1912 RegsToPass[i].second, InFlag);
1913 InFlag = Chain.getValue(1);
1916 if (Subtarget->isPICStyleGOT()) {
1917 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1920 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1921 DAG.getNode(X86ISD::GlobalBaseReg,
1922 DebugLoc::getUnknownLoc(),
1925 InFlag = Chain.getValue(1);
1927 // If we are tail calling and generating PIC/GOT style code load the
1928 // address of the callee into ECX. The value in ecx is used as target of
1929 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1930 // for tail calls on PIC/GOT architectures. Normally we would just put the
1931 // address of GOT into ebx and then call target@PLT. But for tail calls
1932 // ebx would be restored (since ebx is callee saved) before jumping to the
1935 // Note: The actual moving to ECX is done further down.
1936 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1937 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1938 !G->getGlobal()->hasProtectedVisibility())
1939 Callee = LowerGlobalAddress(Callee, DAG);
1940 else if (isa<ExternalSymbolSDNode>(Callee))
1941 Callee = LowerExternalSymbol(Callee, DAG);
1945 if (Is64Bit && isVarArg) {
1946 // From AMD64 ABI document:
1947 // For calls that may call functions that use varargs or stdargs
1948 // (prototype-less calls or calls to functions containing ellipsis (...) in
1949 // the declaration) %al is used as hidden argument to specify the number
1950 // of SSE registers used. The contents of %al do not need to match exactly
1951 // the number of registers, but must be an ubound on the number of SSE
1952 // registers used and is in the range 0 - 8 inclusive.
1954 // FIXME: Verify this on Win64
1955 // Count the number of XMM registers allocated.
1956 static const unsigned XMMArgRegs[] = {
1957 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1958 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1960 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1961 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1962 && "SSE registers cannot be used when SSE is disabled");
1964 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1965 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1966 InFlag = Chain.getValue(1);
1970 // For tail calls lower the arguments to the 'real' stack slot.
1972 // Force all the incoming stack arguments to be loaded from the stack
1973 // before any new outgoing arguments are stored to the stack, because the
1974 // outgoing stack slots may alias the incoming argument stack slots, and
1975 // the alias isn't otherwise explicit. This is slightly more conservative
1976 // than necessary, because it means that each store effectively depends
1977 // on every argument instead of just those arguments it would clobber.
1978 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1980 SmallVector<SDValue, 8> MemOpChains2;
1983 // Do not flag preceeding copytoreg stuff together with the following stuff.
1985 if (GuaranteedTailCallOpt) {
1986 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1987 CCValAssign &VA = ArgLocs[i];
1990 assert(VA.isMemLoc());
1991 SDValue Arg = Outs[i].Val;
1992 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1993 // Create frame index.
1994 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1995 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1996 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1997 FIN = DAG.getFrameIndex(FI, getPointerTy());
1999 if (Flags.isByVal()) {
2000 // Copy relative to framepointer.
2001 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2002 if (StackPtr.getNode() == 0)
2003 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2005 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2007 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2011 // Store relative to framepointer.
2012 MemOpChains2.push_back(
2013 DAG.getStore(ArgChain, dl, Arg, FIN,
2014 PseudoSourceValue::getFixedStack(FI), 0,
2020 if (!MemOpChains2.empty())
2021 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2022 &MemOpChains2[0], MemOpChains2.size());
2024 // Copy arguments to their registers.
2025 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2026 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2027 RegsToPass[i].second, InFlag);
2028 InFlag = Chain.getValue(1);
2032 // Store the return address to the appropriate stack slot.
2033 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2037 bool WasGlobalOrExternal = false;
2038 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2039 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2040 // In the 64-bit large code model, we have to make all calls
2041 // through a register, since the call instruction's 32-bit
2042 // pc-relative offset may not be large enough to hold the whole
2044 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2045 WasGlobalOrExternal = true;
2046 // If the callee is a GlobalAddress node (quite common, every direct call
2047 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2050 // We should use extra load for direct calls to dllimported functions in
2052 GlobalValue *GV = G->getGlobal();
2053 if (!GV->hasDLLImportLinkage()) {
2054 unsigned char OpFlags = 0;
2056 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2057 // external symbols most go through the PLT in PIC mode. If the symbol
2058 // has hidden or protected visibility, or if it is static or local, then
2059 // we don't need to use the PLT - we can directly call it.
2060 if (Subtarget->isTargetELF() &&
2061 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2062 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2063 OpFlags = X86II::MO_PLT;
2064 } else if (Subtarget->isPICStyleStubAny() &&
2065 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2066 Subtarget->getDarwinVers() < 9) {
2067 // PC-relative references to external symbols should go through $stub,
2068 // unless we're building with the leopard linker or later, which
2069 // automatically synthesizes these stubs.
2070 OpFlags = X86II::MO_DARWIN_STUB;
2073 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2074 G->getOffset(), OpFlags);
2076 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2077 WasGlobalOrExternal = true;
2078 unsigned char OpFlags = 0;
2080 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2081 // symbols should go through the PLT.
2082 if (Subtarget->isTargetELF() &&
2083 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2084 OpFlags = X86II::MO_PLT;
2085 } else if (Subtarget->isPICStyleStubAny() &&
2086 Subtarget->getDarwinVers() < 9) {
2087 // PC-relative references to external symbols should go through $stub,
2088 // unless we're building with the leopard linker or later, which
2089 // automatically synthesizes these stubs.
2090 OpFlags = X86II::MO_DARWIN_STUB;
2093 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2097 // Returns a chain & a flag for retval copy to use.
2098 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2099 SmallVector<SDValue, 8> Ops;
2101 if (!IsSibcall && isTailCall) {
2102 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2103 DAG.getIntPtrConstant(0, true), InFlag);
2104 InFlag = Chain.getValue(1);
2107 Ops.push_back(Chain);
2108 Ops.push_back(Callee);
2111 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2113 // Add argument registers to the end of the list so that they are known live
2115 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2116 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2117 RegsToPass[i].second.getValueType()));
2119 // Add an implicit use GOT pointer in EBX.
2120 if (!isTailCall && Subtarget->isPICStyleGOT())
2121 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2123 // Add an implicit use of AL for x86 vararg functions.
2124 if (Is64Bit && isVarArg)
2125 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2127 if (InFlag.getNode())
2128 Ops.push_back(InFlag);
2131 // If this is the first return lowered for this function, add the regs
2132 // to the liveout set for the function.
2133 if (MF.getRegInfo().liveout_empty()) {
2134 SmallVector<CCValAssign, 16> RVLocs;
2135 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2137 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2138 for (unsigned i = 0; i != RVLocs.size(); ++i)
2139 if (RVLocs[i].isRegLoc())
2140 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2142 return DAG.getNode(X86ISD::TC_RETURN, dl,
2143 NodeTys, &Ops[0], Ops.size());
2146 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2147 InFlag = Chain.getValue(1);
2149 // Create the CALLSEQ_END node.
2150 unsigned NumBytesForCalleeToPush;
2151 if (IsCalleePop(isVarArg, CallConv))
2152 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2153 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2154 // If this is a call to a struct-return function, the callee
2155 // pops the hidden struct pointer, so we have to push it back.
2156 // This is common for Darwin/X86, Linux & Mingw32 targets.
2157 NumBytesForCalleeToPush = 4;
2159 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2161 // Returns a flag for retval copy to use.
2163 Chain = DAG.getCALLSEQ_END(Chain,
2164 DAG.getIntPtrConstant(NumBytes, true),
2165 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2168 InFlag = Chain.getValue(1);
2171 // Handle result values, copying them out of physregs into vregs that we
2173 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2174 Ins, dl, DAG, InVals);
2178 //===----------------------------------------------------------------------===//
2179 // Fast Calling Convention (tail call) implementation
2180 //===----------------------------------------------------------------------===//
2182 // Like std call, callee cleans arguments, convention except that ECX is
2183 // reserved for storing the tail called function address. Only 2 registers are
2184 // free for argument passing (inreg). Tail call optimization is performed
2186 // * tailcallopt is enabled
2187 // * caller/callee are fastcc
2188 // On X86_64 architecture with GOT-style position independent code only local
2189 // (within module) calls are supported at the moment.
2190 // To keep the stack aligned according to platform abi the function
2191 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2192 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2193 // If a tail called function callee has more arguments than the caller the
2194 // caller needs to make sure that there is room to move the RETADDR to. This is
2195 // achieved by reserving an area the size of the argument delta right after the
2196 // original REtADDR, but before the saved framepointer or the spilled registers
2197 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2209 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2210 /// for a 16 byte align requirement.
2211 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2212 SelectionDAG& DAG) {
2213 MachineFunction &MF = DAG.getMachineFunction();
2214 const TargetMachine &TM = MF.getTarget();
2215 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2216 unsigned StackAlignment = TFI.getStackAlignment();
2217 uint64_t AlignMask = StackAlignment - 1;
2218 int64_t Offset = StackSize;
2219 uint64_t SlotSize = TD->getPointerSize();
2220 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2221 // Number smaller than 12 so just add the difference.
2222 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2224 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2225 Offset = ((~AlignMask) & Offset) + StackAlignment +
2226 (StackAlignment-SlotSize);
2231 /// MatchingStackOffset - Return true if the given stack call argument is
2232 /// already available in the same position (relatively) of the caller's
2233 /// incoming argument stack.
2235 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2236 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2237 const X86InstrInfo *TII) {
2238 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2240 if (Arg.getOpcode() == ISD::CopyFromReg) {
2241 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2242 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2244 MachineInstr *Def = MRI->getVRegDef(VR);
2247 if (!Flags.isByVal()) {
2248 if (!TII->isLoadFromStackSlot(Def, FI))
2251 unsigned Opcode = Def->getOpcode();
2252 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2253 Def->getOperand(1).isFI()) {
2254 FI = Def->getOperand(1).getIndex();
2255 Bytes = Flags.getByValSize();
2259 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2260 if (Flags.isByVal())
2261 // ByVal argument is passed in as a pointer but it's now being
2262 // dereferenced. e.g.
2263 // define @foo(%struct.X* %A) {
2264 // tail call @bar(%struct.X* byval %A)
2267 SDValue Ptr = Ld->getBasePtr();
2268 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2271 FI = FINode->getIndex();
2275 assert(FI != INT_MAX);
2276 if (!MFI->isFixedObjectIndex(FI))
2278 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2281 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2282 /// for tail call optimization. Targets which want to do tail call
2283 /// optimization should implement this function.
2285 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2286 CallingConv::ID CalleeCC,
2288 bool isCalleeStructRet,
2289 bool isCallerStructRet,
2290 const SmallVectorImpl<ISD::OutputArg> &Outs,
2291 const SmallVectorImpl<ISD::InputArg> &Ins,
2292 SelectionDAG& DAG) const {
2293 if (!IsTailCallConvention(CalleeCC) &&
2294 CalleeCC != CallingConv::C)
2297 // If -tailcallopt is specified, make fastcc functions tail-callable.
2298 const Function *CallerF = DAG.getMachineFunction().getFunction();
2299 if (GuaranteedTailCallOpt) {
2300 if (IsTailCallConvention(CalleeCC) &&
2301 CallerF->getCallingConv() == CalleeCC)
2306 // Look for obvious safe cases to perform tail call optimization that does not
2307 // requite ABI changes. This is what gcc calls sibcall.
2309 // Do not sibcall optimize vararg calls for now.
2313 // Also avoid sibcall optimization if either caller or callee uses struct
2314 // return semantics.
2315 if (isCalleeStructRet || isCallerStructRet)
2318 // If the callee takes no arguments then go on to check the results of the
2320 if (!Outs.empty()) {
2321 // Check if stack adjustment is needed. For now, do not do this if any
2322 // argument is passed on the stack.
2323 SmallVector<CCValAssign, 16> ArgLocs;
2324 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2325 ArgLocs, *DAG.getContext());
2326 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2327 if (CCInfo.getNextStackOffset()) {
2328 MachineFunction &MF = DAG.getMachineFunction();
2329 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2331 if (Subtarget->isTargetWin64())
2332 // Win64 ABI has additional complications.
2335 // Check if the arguments are already laid out in the right way as
2336 // the caller's fixed stack objects.
2337 MachineFrameInfo *MFI = MF.getFrameInfo();
2338 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2339 const X86InstrInfo *TII =
2340 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2341 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2342 CCValAssign &VA = ArgLocs[i];
2343 EVT RegVT = VA.getLocVT();
2344 SDValue Arg = Outs[i].Val;
2345 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2346 if (VA.getLocInfo() == CCValAssign::Indirect)
2348 if (!VA.isRegLoc()) {
2349 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2361 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2363 DenseMap<const Value *, unsigned> &vm,
2364 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2365 DenseMap<const AllocaInst *, int> &am
2367 , SmallSet<Instruction*, 8> &cil
2370 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2378 //===----------------------------------------------------------------------===//
2379 // Other Lowering Hooks
2380 //===----------------------------------------------------------------------===//
2383 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2384 MachineFunction &MF = DAG.getMachineFunction();
2385 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2386 int ReturnAddrIndex = FuncInfo->getRAIndex();
2388 if (ReturnAddrIndex == 0) {
2389 // Set up a frame object for the return address.
2390 uint64_t SlotSize = TD->getPointerSize();
2391 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2393 FuncInfo->setRAIndex(ReturnAddrIndex);
2396 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2400 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2401 bool hasSymbolicDisplacement) {
2402 // Offset should fit into 32 bit immediate field.
2403 if (!isInt32(Offset))
2406 // If we don't have a symbolic displacement - we don't have any extra
2408 if (!hasSymbolicDisplacement)
2411 // FIXME: Some tweaks might be needed for medium code model.
2412 if (M != CodeModel::Small && M != CodeModel::Kernel)
2415 // For small code model we assume that latest object is 16MB before end of 31
2416 // bits boundary. We may also accept pretty large negative constants knowing
2417 // that all objects are in the positive half of address space.
2418 if (M == CodeModel::Small && Offset < 16*1024*1024)
2421 // For kernel code model we know that all object resist in the negative half
2422 // of 32bits address space. We may not accept negative offsets, since they may
2423 // be just off and we may accept pretty large positive ones.
2424 if (M == CodeModel::Kernel && Offset > 0)
2430 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2431 /// specific condition code, returning the condition code and the LHS/RHS of the
2432 /// comparison to make.
2433 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2434 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2436 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2437 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2438 // X > -1 -> X == 0, jump !sign.
2439 RHS = DAG.getConstant(0, RHS.getValueType());
2440 return X86::COND_NS;
2441 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2442 // X < 0 -> X == 0, jump on sign.
2444 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2446 RHS = DAG.getConstant(0, RHS.getValueType());
2447 return X86::COND_LE;
2451 switch (SetCCOpcode) {
2452 default: llvm_unreachable("Invalid integer condition!");
2453 case ISD::SETEQ: return X86::COND_E;
2454 case ISD::SETGT: return X86::COND_G;
2455 case ISD::SETGE: return X86::COND_GE;
2456 case ISD::SETLT: return X86::COND_L;
2457 case ISD::SETLE: return X86::COND_LE;
2458 case ISD::SETNE: return X86::COND_NE;
2459 case ISD::SETULT: return X86::COND_B;
2460 case ISD::SETUGT: return X86::COND_A;
2461 case ISD::SETULE: return X86::COND_BE;
2462 case ISD::SETUGE: return X86::COND_AE;
2466 // First determine if it is required or is profitable to flip the operands.
2468 // If LHS is a foldable load, but RHS is not, flip the condition.
2469 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2470 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2471 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2472 std::swap(LHS, RHS);
2475 switch (SetCCOpcode) {
2481 std::swap(LHS, RHS);
2485 // On a floating point condition, the flags are set as follows:
2487 // 0 | 0 | 0 | X > Y
2488 // 0 | 0 | 1 | X < Y
2489 // 1 | 0 | 0 | X == Y
2490 // 1 | 1 | 1 | unordered
2491 switch (SetCCOpcode) {
2492 default: llvm_unreachable("Condcode should be pre-legalized away");
2494 case ISD::SETEQ: return X86::COND_E;
2495 case ISD::SETOLT: // flipped
2497 case ISD::SETGT: return X86::COND_A;
2498 case ISD::SETOLE: // flipped
2500 case ISD::SETGE: return X86::COND_AE;
2501 case ISD::SETUGT: // flipped
2503 case ISD::SETLT: return X86::COND_B;
2504 case ISD::SETUGE: // flipped
2506 case ISD::SETLE: return X86::COND_BE;
2508 case ISD::SETNE: return X86::COND_NE;
2509 case ISD::SETUO: return X86::COND_P;
2510 case ISD::SETO: return X86::COND_NP;
2512 case ISD::SETUNE: return X86::COND_INVALID;
2516 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2517 /// code. Current x86 isa includes the following FP cmov instructions:
2518 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2519 static bool hasFPCMov(unsigned X86CC) {
2535 /// isFPImmLegal - Returns true if the target can instruction select the
2536 /// specified FP immediate natively. If false, the legalizer will
2537 /// materialize the FP immediate as a load from a constant pool.
2538 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2539 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2540 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2546 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2547 /// the specified range (L, H].
2548 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2549 return (Val < 0) || (Val >= Low && Val < Hi);
2552 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2553 /// specified value.
2554 static bool isUndefOrEqual(int Val, int CmpVal) {
2555 if (Val < 0 || Val == CmpVal)
2560 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2561 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2562 /// the second operand.
2563 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2564 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2565 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2566 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2567 return (Mask[0] < 2 && Mask[1] < 2);
2571 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2572 SmallVector<int, 8> M;
2574 return ::isPSHUFDMask(M, N->getValueType(0));
2577 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2578 /// is suitable for input to PSHUFHW.
2579 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2580 if (VT != MVT::v8i16)
2583 // Lower quadword copied in order or undef.
2584 for (int i = 0; i != 4; ++i)
2585 if (Mask[i] >= 0 && Mask[i] != i)
2588 // Upper quadword shuffled.
2589 for (int i = 4; i != 8; ++i)
2590 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2596 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2597 SmallVector<int, 8> M;
2599 return ::isPSHUFHWMask(M, N->getValueType(0));
2602 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2603 /// is suitable for input to PSHUFLW.
2604 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2605 if (VT != MVT::v8i16)
2608 // Upper quadword copied in order.
2609 for (int i = 4; i != 8; ++i)
2610 if (Mask[i] >= 0 && Mask[i] != i)
2613 // Lower quadword shuffled.
2614 for (int i = 0; i != 4; ++i)
2621 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2622 SmallVector<int, 8> M;
2624 return ::isPSHUFLWMask(M, N->getValueType(0));
2627 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2628 /// is suitable for input to PALIGNR.
2629 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2631 int i, e = VT.getVectorNumElements();
2633 // Do not handle v2i64 / v2f64 shuffles with palignr.
2634 if (e < 4 || !hasSSSE3)
2637 for (i = 0; i != e; ++i)
2641 // All undef, not a palignr.
2645 // Determine if it's ok to perform a palignr with only the LHS, since we
2646 // don't have access to the actual shuffle elements to see if RHS is undef.
2647 bool Unary = Mask[i] < (int)e;
2648 bool NeedsUnary = false;
2650 int s = Mask[i] - i;
2652 // Check the rest of the elements to see if they are consecutive.
2653 for (++i; i != e; ++i) {
2658 Unary = Unary && (m < (int)e);
2659 NeedsUnary = NeedsUnary || (m < s);
2661 if (NeedsUnary && !Unary)
2663 if (Unary && m != ((s+i) & (e-1)))
2665 if (!Unary && m != (s+i))
2671 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2672 SmallVector<int, 8> M;
2674 return ::isPALIGNRMask(M, N->getValueType(0), true);
2677 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2678 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2679 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2680 int NumElems = VT.getVectorNumElements();
2681 if (NumElems != 2 && NumElems != 4)
2684 int Half = NumElems / 2;
2685 for (int i = 0; i < Half; ++i)
2686 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2688 for (int i = Half; i < NumElems; ++i)
2689 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2695 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2696 SmallVector<int, 8> M;
2698 return ::isSHUFPMask(M, N->getValueType(0));
2701 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2702 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2703 /// half elements to come from vector 1 (which would equal the dest.) and
2704 /// the upper half to come from vector 2.
2705 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2706 int NumElems = VT.getVectorNumElements();
2708 if (NumElems != 2 && NumElems != 4)
2711 int Half = NumElems / 2;
2712 for (int i = 0; i < Half; ++i)
2713 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2715 for (int i = Half; i < NumElems; ++i)
2716 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2721 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2722 SmallVector<int, 8> M;
2724 return isCommutedSHUFPMask(M, N->getValueType(0));
2727 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2728 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2729 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2730 if (N->getValueType(0).getVectorNumElements() != 4)
2733 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2734 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2735 isUndefOrEqual(N->getMaskElt(1), 7) &&
2736 isUndefOrEqual(N->getMaskElt(2), 2) &&
2737 isUndefOrEqual(N->getMaskElt(3), 3);
2740 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2741 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2743 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2744 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2749 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2750 isUndefOrEqual(N->getMaskElt(1), 3) &&
2751 isUndefOrEqual(N->getMaskElt(2), 2) &&
2752 isUndefOrEqual(N->getMaskElt(3), 3);
2755 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2756 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2757 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2758 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2760 if (NumElems != 2 && NumElems != 4)
2763 for (unsigned i = 0; i < NumElems/2; ++i)
2764 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2767 for (unsigned i = NumElems/2; i < NumElems; ++i)
2768 if (!isUndefOrEqual(N->getMaskElt(i), i))
2774 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2775 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2776 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2777 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2779 if (NumElems != 2 && NumElems != 4)
2782 for (unsigned i = 0; i < NumElems/2; ++i)
2783 if (!isUndefOrEqual(N->getMaskElt(i), i))
2786 for (unsigned i = 0; i < NumElems/2; ++i)
2787 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2793 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2794 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2795 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2796 bool V2IsSplat = false) {
2797 int NumElts = VT.getVectorNumElements();
2798 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2801 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2803 int BitI1 = Mask[i+1];
2804 if (!isUndefOrEqual(BitI, j))
2807 if (!isUndefOrEqual(BitI1, NumElts))
2810 if (!isUndefOrEqual(BitI1, j + NumElts))
2817 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2818 SmallVector<int, 8> M;
2820 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2823 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2824 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2825 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2826 bool V2IsSplat = false) {
2827 int NumElts = VT.getVectorNumElements();
2828 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2831 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2833 int BitI1 = Mask[i+1];
2834 if (!isUndefOrEqual(BitI, j + NumElts/2))
2837 if (isUndefOrEqual(BitI1, NumElts))
2840 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2847 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2848 SmallVector<int, 8> M;
2850 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2853 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2854 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2856 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2857 int NumElems = VT.getVectorNumElements();
2858 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2861 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2863 int BitI1 = Mask[i+1];
2864 if (!isUndefOrEqual(BitI, j))
2866 if (!isUndefOrEqual(BitI1, j))
2872 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2873 SmallVector<int, 8> M;
2875 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2878 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2879 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2881 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2882 int NumElems = VT.getVectorNumElements();
2883 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2886 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2888 int BitI1 = Mask[i+1];
2889 if (!isUndefOrEqual(BitI, j))
2891 if (!isUndefOrEqual(BitI1, j))
2897 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2898 SmallVector<int, 8> M;
2900 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2903 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2904 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2905 /// MOVSD, and MOVD, i.e. setting the lowest element.
2906 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2907 if (VT.getVectorElementType().getSizeInBits() < 32)
2910 int NumElts = VT.getVectorNumElements();
2912 if (!isUndefOrEqual(Mask[0], NumElts))
2915 for (int i = 1; i < NumElts; ++i)
2916 if (!isUndefOrEqual(Mask[i], i))
2922 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2923 SmallVector<int, 8> M;
2925 return ::isMOVLMask(M, N->getValueType(0));
2928 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2929 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2930 /// element of vector 2 and the other elements to come from vector 1 in order.
2931 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2932 bool V2IsSplat = false, bool V2IsUndef = false) {
2933 int NumOps = VT.getVectorNumElements();
2934 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2937 if (!isUndefOrEqual(Mask[0], 0))
2940 for (int i = 1; i < NumOps; ++i)
2941 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2942 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2943 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2949 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2950 bool V2IsUndef = false) {
2951 SmallVector<int, 8> M;
2953 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2956 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2957 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2958 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2959 if (N->getValueType(0).getVectorNumElements() != 4)
2962 // Expect 1, 1, 3, 3
2963 for (unsigned i = 0; i < 2; ++i) {
2964 int Elt = N->getMaskElt(i);
2965 if (Elt >= 0 && Elt != 1)
2970 for (unsigned i = 2; i < 4; ++i) {
2971 int Elt = N->getMaskElt(i);
2972 if (Elt >= 0 && Elt != 3)
2977 // Don't use movshdup if it can be done with a shufps.
2978 // FIXME: verify that matching u, u, 3, 3 is what we want.
2982 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2983 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2984 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2985 if (N->getValueType(0).getVectorNumElements() != 4)
2988 // Expect 0, 0, 2, 2
2989 for (unsigned i = 0; i < 2; ++i)
2990 if (N->getMaskElt(i) > 0)
2994 for (unsigned i = 2; i < 4; ++i) {
2995 int Elt = N->getMaskElt(i);
2996 if (Elt >= 0 && Elt != 2)
3001 // Don't use movsldup if it can be done with a shufps.
3005 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3006 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3007 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3008 int e = N->getValueType(0).getVectorNumElements() / 2;
3010 for (int i = 0; i < e; ++i)
3011 if (!isUndefOrEqual(N->getMaskElt(i), i))
3013 for (int i = 0; i < e; ++i)
3014 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3019 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3020 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3021 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3022 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3023 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3025 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3027 for (int i = 0; i < NumOperands; ++i) {
3028 int Val = SVOp->getMaskElt(NumOperands-i-1);
3029 if (Val < 0) Val = 0;
3030 if (Val >= NumOperands) Val -= NumOperands;
3032 if (i != NumOperands - 1)
3038 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3039 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3040 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3041 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3043 // 8 nodes, but we only care about the last 4.
3044 for (unsigned i = 7; i >= 4; --i) {
3045 int Val = SVOp->getMaskElt(i);
3054 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3055 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3056 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3057 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3059 // 8 nodes, but we only care about the first 4.
3060 for (int i = 3; i >= 0; --i) {
3061 int Val = SVOp->getMaskElt(i);
3070 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3071 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3072 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3073 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3074 EVT VVT = N->getValueType(0);
3075 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3079 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3080 Val = SVOp->getMaskElt(i);
3084 return (Val - i) * EltSize;
3087 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3089 bool X86::isZeroNode(SDValue Elt) {
3090 return ((isa<ConstantSDNode>(Elt) &&
3091 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3092 (isa<ConstantFPSDNode>(Elt) &&
3093 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3096 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3097 /// their permute mask.
3098 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3099 SelectionDAG &DAG) {
3100 EVT VT = SVOp->getValueType(0);
3101 unsigned NumElems = VT.getVectorNumElements();
3102 SmallVector<int, 8> MaskVec;
3104 for (unsigned i = 0; i != NumElems; ++i) {
3105 int idx = SVOp->getMaskElt(i);
3107 MaskVec.push_back(idx);
3108 else if (idx < (int)NumElems)
3109 MaskVec.push_back(idx + NumElems);
3111 MaskVec.push_back(idx - NumElems);
3113 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3114 SVOp->getOperand(0), &MaskVec[0]);
3117 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3118 /// the two vector operands have swapped position.
3119 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3120 unsigned NumElems = VT.getVectorNumElements();
3121 for (unsigned i = 0; i != NumElems; ++i) {
3125 else if (idx < (int)NumElems)
3126 Mask[i] = idx + NumElems;
3128 Mask[i] = idx - NumElems;
3132 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3133 /// match movhlps. The lower half elements should come from upper half of
3134 /// V1 (and in order), and the upper half elements should come from the upper
3135 /// half of V2 (and in order).
3136 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3137 if (Op->getValueType(0).getVectorNumElements() != 4)
3139 for (unsigned i = 0, e = 2; i != e; ++i)
3140 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3142 for (unsigned i = 2; i != 4; ++i)
3143 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3148 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3149 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3151 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3152 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3154 N = N->getOperand(0).getNode();
3155 if (!ISD::isNON_EXTLoad(N))
3158 *LD = cast<LoadSDNode>(N);
3162 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3163 /// match movlp{s|d}. The lower half elements should come from lower half of
3164 /// V1 (and in order), and the upper half elements should come from the upper
3165 /// half of V2 (and in order). And since V1 will become the source of the
3166 /// MOVLP, it must be either a vector load or a scalar load to vector.
3167 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3168 ShuffleVectorSDNode *Op) {
3169 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3171 // Is V2 is a vector load, don't do this transformation. We will try to use
3172 // load folding shufps op.
3173 if (ISD::isNON_EXTLoad(V2))
3176 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3178 if (NumElems != 2 && NumElems != 4)
3180 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3181 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3183 for (unsigned i = NumElems/2; i != NumElems; ++i)
3184 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3189 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3191 static bool isSplatVector(SDNode *N) {
3192 if (N->getOpcode() != ISD::BUILD_VECTOR)
3195 SDValue SplatValue = N->getOperand(0);
3196 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3197 if (N->getOperand(i) != SplatValue)
3202 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3203 /// to an zero vector.
3204 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3205 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3206 SDValue V1 = N->getOperand(0);
3207 SDValue V2 = N->getOperand(1);
3208 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3209 for (unsigned i = 0; i != NumElems; ++i) {
3210 int Idx = N->getMaskElt(i);
3211 if (Idx >= (int)NumElems) {
3212 unsigned Opc = V2.getOpcode();
3213 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3215 if (Opc != ISD::BUILD_VECTOR ||
3216 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3218 } else if (Idx >= 0) {
3219 unsigned Opc = V1.getOpcode();
3220 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3222 if (Opc != ISD::BUILD_VECTOR ||
3223 !X86::isZeroNode(V1.getOperand(Idx)))
3230 /// getZeroVector - Returns a vector of specified type with all zero elements.
3232 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3234 assert(VT.isVector() && "Expected a vector type");
3236 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3237 // type. This ensures they get CSE'd.
3239 if (VT.getSizeInBits() == 64) { // MMX
3240 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3241 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3242 } else if (HasSSE2) { // SSE2
3243 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3244 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3246 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3247 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3249 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3252 /// getOnesVector - Returns a vector of specified type with all bits set.
3254 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3255 assert(VT.isVector() && "Expected a vector type");
3257 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3258 // type. This ensures they get CSE'd.
3259 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3261 if (VT.getSizeInBits() == 64) // MMX
3262 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3264 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3265 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3269 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3270 /// that point to V2 points to its first element.
3271 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3272 EVT VT = SVOp->getValueType(0);
3273 unsigned NumElems = VT.getVectorNumElements();
3275 bool Changed = false;
3276 SmallVector<int, 8> MaskVec;
3277 SVOp->getMask(MaskVec);
3279 for (unsigned i = 0; i != NumElems; ++i) {
3280 if (MaskVec[i] > (int)NumElems) {
3281 MaskVec[i] = NumElems;
3286 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3287 SVOp->getOperand(1), &MaskVec[0]);
3288 return SDValue(SVOp, 0);
3291 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3292 /// operation of specified width.
3293 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3295 unsigned NumElems = VT.getVectorNumElements();
3296 SmallVector<int, 8> Mask;
3297 Mask.push_back(NumElems);
3298 for (unsigned i = 1; i != NumElems; ++i)
3300 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3303 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3304 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3306 unsigned NumElems = VT.getVectorNumElements();
3307 SmallVector<int, 8> Mask;
3308 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3310 Mask.push_back(i + NumElems);
3312 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3315 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3316 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3318 unsigned NumElems = VT.getVectorNumElements();
3319 unsigned Half = NumElems/2;
3320 SmallVector<int, 8> Mask;
3321 for (unsigned i = 0; i != Half; ++i) {
3322 Mask.push_back(i + Half);
3323 Mask.push_back(i + NumElems + Half);
3325 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3328 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3329 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3331 if (SV->getValueType(0).getVectorNumElements() <= 4)
3332 return SDValue(SV, 0);
3334 EVT PVT = MVT::v4f32;
3335 EVT VT = SV->getValueType(0);
3336 DebugLoc dl = SV->getDebugLoc();
3337 SDValue V1 = SV->getOperand(0);
3338 int NumElems = VT.getVectorNumElements();
3339 int EltNo = SV->getSplatIndex();
3341 // unpack elements to the correct location
3342 while (NumElems > 4) {
3343 if (EltNo < NumElems/2) {
3344 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3346 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3347 EltNo -= NumElems/2;
3352 // Perform the splat.
3353 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3354 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3355 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3356 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3359 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3360 /// vector of zero or undef vector. This produces a shuffle where the low
3361 /// element of V2 is swizzled into the zero/undef vector, landing at element
3362 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3363 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3364 bool isZero, bool HasSSE2,
3365 SelectionDAG &DAG) {
3366 EVT VT = V2.getValueType();
3368 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3369 unsigned NumElems = VT.getVectorNumElements();
3370 SmallVector<int, 16> MaskVec;
3371 for (unsigned i = 0; i != NumElems; ++i)
3372 // If this is the insertion idx, put the low elt of V2 here.
3373 MaskVec.push_back(i == Idx ? NumElems : i);
3374 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3377 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3378 /// a shuffle that is zero.
3380 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3381 bool Low, SelectionDAG &DAG) {
3382 unsigned NumZeros = 0;
3383 for (int i = 0; i < NumElems; ++i) {
3384 unsigned Index = Low ? i : NumElems-i-1;
3385 int Idx = SVOp->getMaskElt(Index);
3390 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3391 if (Elt.getNode() && X86::isZeroNode(Elt))
3399 /// isVectorShift - Returns true if the shuffle can be implemented as a
3400 /// logical left or right shift of a vector.
3401 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3402 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3403 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3404 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3407 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3410 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3414 bool SeenV1 = false;
3415 bool SeenV2 = false;
3416 for (int i = NumZeros; i < NumElems; ++i) {
3417 int Val = isLeft ? (i - NumZeros) : i;
3418 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3430 if (SeenV1 && SeenV2)
3433 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3439 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3441 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3442 unsigned NumNonZero, unsigned NumZero,
3443 SelectionDAG &DAG, TargetLowering &TLI) {
3447 DebugLoc dl = Op.getDebugLoc();
3450 for (unsigned i = 0; i < 16; ++i) {
3451 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3452 if (ThisIsNonZero && First) {
3454 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3456 V = DAG.getUNDEF(MVT::v8i16);
3461 SDValue ThisElt(0, 0), LastElt(0, 0);
3462 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3463 if (LastIsNonZero) {
3464 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3465 MVT::i16, Op.getOperand(i-1));
3467 if (ThisIsNonZero) {
3468 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3469 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3470 ThisElt, DAG.getConstant(8, MVT::i8));
3472 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3476 if (ThisElt.getNode())
3477 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3478 DAG.getIntPtrConstant(i/2));
3482 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3485 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3487 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3488 unsigned NumNonZero, unsigned NumZero,
3489 SelectionDAG &DAG, TargetLowering &TLI) {
3493 DebugLoc dl = Op.getDebugLoc();
3496 for (unsigned i = 0; i < 8; ++i) {
3497 bool isNonZero = (NonZeros & (1 << i)) != 0;
3501 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3503 V = DAG.getUNDEF(MVT::v8i16);
3506 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3507 MVT::v8i16, V, Op.getOperand(i),
3508 DAG.getIntPtrConstant(i));
3515 /// getVShift - Return a vector logical shift node.
3517 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3518 unsigned NumBits, SelectionDAG &DAG,
3519 const TargetLowering &TLI, DebugLoc dl) {
3520 bool isMMX = VT.getSizeInBits() == 64;
3521 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3522 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3523 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3524 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3525 DAG.getNode(Opc, dl, ShVT, SrcOp,
3526 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3530 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3531 SelectionDAG &DAG) {
3533 // Check if the scalar load can be widened into a vector load. And if
3534 // the address is "base + cst" see if the cst can be "absorbed" into
3535 // the shuffle mask.
3536 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3537 SDValue Ptr = LD->getBasePtr();
3538 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3540 EVT PVT = LD->getValueType(0);
3541 if (PVT != MVT::i32 && PVT != MVT::f32)
3546 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3547 FI = FINode->getIndex();
3549 } else if (Ptr.getOpcode() == ISD::ADD &&
3550 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3551 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3552 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3553 Offset = Ptr.getConstantOperandVal(1);
3554 Ptr = Ptr.getOperand(0);
3559 SDValue Chain = LD->getChain();
3560 // Make sure the stack object alignment is at least 16.
3561 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3562 if (DAG.InferPtrAlignment(Ptr) < 16) {
3563 if (MFI->isFixedObjectIndex(FI)) {
3564 // Can't change the alignment. FIXME: It's possible to compute
3565 // the exact stack offset and reference FI + adjust offset instead.
3566 // If someone *really* cares about this. That's the way to implement it.
3569 MFI->setObjectAlignment(FI, 16);
3573 // (Offset % 16) must be multiple of 4. Then address is then
3574 // Ptr + (Offset & ~15).
3577 if ((Offset % 16) & 3)
3579 int64_t StartOffset = Offset & ~15;
3581 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3582 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3584 int EltNo = (Offset - StartOffset) >> 2;
3585 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3586 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3587 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3589 // Canonicalize it to a v4i32 shuffle.
3590 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3591 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3592 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3593 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3600 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3601 DebugLoc dl = Op.getDebugLoc();
3602 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3603 if (ISD::isBuildVectorAllZeros(Op.getNode())
3604 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3605 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3606 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3607 // eliminated on x86-32 hosts.
3608 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3611 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3612 return getOnesVector(Op.getValueType(), DAG, dl);
3613 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3616 EVT VT = Op.getValueType();
3617 EVT ExtVT = VT.getVectorElementType();
3618 unsigned EVTBits = ExtVT.getSizeInBits();
3620 unsigned NumElems = Op.getNumOperands();
3621 unsigned NumZero = 0;
3622 unsigned NumNonZero = 0;
3623 unsigned NonZeros = 0;
3624 bool IsAllConstants = true;
3625 SmallSet<SDValue, 8> Values;
3626 for (unsigned i = 0; i < NumElems; ++i) {
3627 SDValue Elt = Op.getOperand(i);
3628 if (Elt.getOpcode() == ISD::UNDEF)
3631 if (Elt.getOpcode() != ISD::Constant &&
3632 Elt.getOpcode() != ISD::ConstantFP)
3633 IsAllConstants = false;
3634 if (X86::isZeroNode(Elt))
3637 NonZeros |= (1 << i);
3642 if (NumNonZero == 0) {
3643 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3644 return DAG.getUNDEF(VT);
3647 // Special case for single non-zero, non-undef, element.
3648 if (NumNonZero == 1) {
3649 unsigned Idx = CountTrailingZeros_32(NonZeros);
3650 SDValue Item = Op.getOperand(Idx);
3652 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3653 // the value are obviously zero, truncate the value to i32 and do the
3654 // insertion that way. Only do this if the value is non-constant or if the
3655 // value is a constant being inserted into element 0. It is cheaper to do
3656 // a constant pool load than it is to do a movd + shuffle.
3657 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3658 (!IsAllConstants || Idx == 0)) {
3659 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3660 // Handle MMX and SSE both.
3661 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3662 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3664 // Truncate the value (which may itself be a constant) to i32, and
3665 // convert it to a vector with movd (S2V+shuffle to zero extend).
3666 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3667 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3668 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3669 Subtarget->hasSSE2(), DAG);
3671 // Now we have our 32-bit value zero extended in the low element of
3672 // a vector. If Idx != 0, swizzle it into place.
3674 SmallVector<int, 4> Mask;
3675 Mask.push_back(Idx);
3676 for (unsigned i = 1; i != VecElts; ++i)
3678 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3679 DAG.getUNDEF(Item.getValueType()),
3682 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3686 // If we have a constant or non-constant insertion into the low element of
3687 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3688 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3689 // depending on what the source datatype is.
3692 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3693 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3694 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3695 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3696 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3697 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3699 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3700 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3701 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3702 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3703 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3704 Subtarget->hasSSE2(), DAG);
3705 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3709 // Is it a vector logical left shift?
3710 if (NumElems == 2 && Idx == 1 &&
3711 X86::isZeroNode(Op.getOperand(0)) &&
3712 !X86::isZeroNode(Op.getOperand(1))) {
3713 unsigned NumBits = VT.getSizeInBits();
3714 return getVShift(true, VT,
3715 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3716 VT, Op.getOperand(1)),
3717 NumBits/2, DAG, *this, dl);
3720 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3723 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3724 // is a non-constant being inserted into an element other than the low one,
3725 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3726 // movd/movss) to move this into the low element, then shuffle it into
3728 if (EVTBits == 32) {
3729 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3731 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3732 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3733 Subtarget->hasSSE2(), DAG);
3734 SmallVector<int, 8> MaskVec;
3735 for (unsigned i = 0; i < NumElems; i++)
3736 MaskVec.push_back(i == Idx ? 0 : 1);
3737 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3741 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3742 if (Values.size() == 1) {
3743 if (EVTBits == 32) {
3744 // Instead of a shuffle like this:
3745 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3746 // Check if it's possible to issue this instead.
3747 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3748 unsigned Idx = CountTrailingZeros_32(NonZeros);
3749 SDValue Item = Op.getOperand(Idx);
3750 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3751 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3756 // A vector full of immediates; various special cases are already
3757 // handled, so this is best done with a single constant-pool load.
3761 // Let legalizer expand 2-wide build_vectors.
3762 if (EVTBits == 64) {
3763 if (NumNonZero == 1) {
3764 // One half is zero or undef.
3765 unsigned Idx = CountTrailingZeros_32(NonZeros);
3766 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3767 Op.getOperand(Idx));
3768 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3769 Subtarget->hasSSE2(), DAG);
3774 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3775 if (EVTBits == 8 && NumElems == 16) {
3776 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3778 if (V.getNode()) return V;
3781 if (EVTBits == 16 && NumElems == 8) {
3782 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3784 if (V.getNode()) return V;
3787 // If element VT is == 32 bits, turn it into a number of shuffles.
3788 SmallVector<SDValue, 8> V;
3790 if (NumElems == 4 && NumZero > 0) {
3791 for (unsigned i = 0; i < 4; ++i) {
3792 bool isZero = !(NonZeros & (1 << i));
3794 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3796 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3799 for (unsigned i = 0; i < 2; ++i) {
3800 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3803 V[i] = V[i*2]; // Must be a zero vector.
3806 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3809 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3812 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3817 SmallVector<int, 8> MaskVec;
3818 bool Reverse = (NonZeros & 0x3) == 2;
3819 for (unsigned i = 0; i < 2; ++i)
3820 MaskVec.push_back(Reverse ? 1-i : i);
3821 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3822 for (unsigned i = 0; i < 2; ++i)
3823 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3824 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3827 if (Values.size() > 2) {
3828 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3829 // values to be inserted is equal to the number of elements, in which case
3830 // use the unpack code below in the hopes of matching the consecutive elts
3831 // load merge pattern for shuffles.
3832 // FIXME: We could probably just check that here directly.
3833 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3834 getSubtarget()->hasSSE41()) {
3835 V[0] = DAG.getUNDEF(VT);
3836 for (unsigned i = 0; i < NumElems; ++i)
3837 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3838 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3839 Op.getOperand(i), DAG.getIntPtrConstant(i));
3842 // Expand into a number of unpckl*.
3844 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3845 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3846 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3847 for (unsigned i = 0; i < NumElems; ++i)
3848 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3850 while (NumElems != 0) {
3851 for (unsigned i = 0; i < NumElems; ++i)
3852 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3862 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3863 // We support concatenate two MMX registers and place them in a MMX
3864 // register. This is better than doing a stack convert.
3865 DebugLoc dl = Op.getDebugLoc();
3866 EVT ResVT = Op.getValueType();
3867 assert(Op.getNumOperands() == 2);
3868 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3869 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3871 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3872 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3873 InVec = Op.getOperand(1);
3874 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3875 unsigned NumElts = ResVT.getVectorNumElements();
3876 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3877 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3878 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3880 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3881 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3882 Mask[0] = 0; Mask[1] = 2;
3883 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3885 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3888 // v8i16 shuffles - Prefer shuffles in the following order:
3889 // 1. [all] pshuflw, pshufhw, optional move
3890 // 2. [ssse3] 1 x pshufb
3891 // 3. [ssse3] 2 x pshufb + 1 x por
3892 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3894 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3895 SelectionDAG &DAG, X86TargetLowering &TLI) {
3896 SDValue V1 = SVOp->getOperand(0);
3897 SDValue V2 = SVOp->getOperand(1);
3898 DebugLoc dl = SVOp->getDebugLoc();
3899 SmallVector<int, 8> MaskVals;
3901 // Determine if more than 1 of the words in each of the low and high quadwords
3902 // of the result come from the same quadword of one of the two inputs. Undef
3903 // mask values count as coming from any quadword, for better codegen.
3904 SmallVector<unsigned, 4> LoQuad(4);
3905 SmallVector<unsigned, 4> HiQuad(4);
3906 BitVector InputQuads(4);
3907 for (unsigned i = 0; i < 8; ++i) {
3908 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3909 int EltIdx = SVOp->getMaskElt(i);
3910 MaskVals.push_back(EltIdx);
3919 InputQuads.set(EltIdx / 4);
3922 int BestLoQuad = -1;
3923 unsigned MaxQuad = 1;
3924 for (unsigned i = 0; i < 4; ++i) {
3925 if (LoQuad[i] > MaxQuad) {
3927 MaxQuad = LoQuad[i];
3931 int BestHiQuad = -1;
3933 for (unsigned i = 0; i < 4; ++i) {
3934 if (HiQuad[i] > MaxQuad) {
3936 MaxQuad = HiQuad[i];
3940 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3941 // of the two input vectors, shuffle them into one input vector so only a
3942 // single pshufb instruction is necessary. If There are more than 2 input
3943 // quads, disable the next transformation since it does not help SSSE3.
3944 bool V1Used = InputQuads[0] || InputQuads[1];
3945 bool V2Used = InputQuads[2] || InputQuads[3];
3946 if (TLI.getSubtarget()->hasSSSE3()) {
3947 if (InputQuads.count() == 2 && V1Used && V2Used) {
3948 BestLoQuad = InputQuads.find_first();
3949 BestHiQuad = InputQuads.find_next(BestLoQuad);
3951 if (InputQuads.count() > 2) {
3957 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3958 // the shuffle mask. If a quad is scored as -1, that means that it contains
3959 // words from all 4 input quadwords.
3961 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3962 SmallVector<int, 8> MaskV;
3963 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3964 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3965 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3966 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3967 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3968 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3970 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3971 // source words for the shuffle, to aid later transformations.
3972 bool AllWordsInNewV = true;
3973 bool InOrder[2] = { true, true };
3974 for (unsigned i = 0; i != 8; ++i) {
3975 int idx = MaskVals[i];
3977 InOrder[i/4] = false;
3978 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3980 AllWordsInNewV = false;
3984 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3985 if (AllWordsInNewV) {
3986 for (int i = 0; i != 8; ++i) {
3987 int idx = MaskVals[i];
3990 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3991 if ((idx != i) && idx < 4)
3993 if ((idx != i) && idx > 3)
4002 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4003 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4004 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4005 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4006 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4010 // If we have SSSE3, and all words of the result are from 1 input vector,
4011 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4012 // is present, fall back to case 4.
4013 if (TLI.getSubtarget()->hasSSSE3()) {
4014 SmallVector<SDValue,16> pshufbMask;
4016 // If we have elements from both input vectors, set the high bit of the
4017 // shuffle mask element to zero out elements that come from V2 in the V1
4018 // mask, and elements that come from V1 in the V2 mask, so that the two
4019 // results can be OR'd together.
4020 bool TwoInputs = V1Used && V2Used;
4021 for (unsigned i = 0; i != 8; ++i) {
4022 int EltIdx = MaskVals[i] * 2;
4023 if (TwoInputs && (EltIdx >= 16)) {
4024 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4025 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4028 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4029 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4031 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4032 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4033 DAG.getNode(ISD::BUILD_VECTOR, dl,
4034 MVT::v16i8, &pshufbMask[0], 16));
4036 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4038 // Calculate the shuffle mask for the second input, shuffle it, and
4039 // OR it with the first shuffled input.
4041 for (unsigned i = 0; i != 8; ++i) {
4042 int EltIdx = MaskVals[i] * 2;
4044 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4045 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4048 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4049 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4051 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4052 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4053 DAG.getNode(ISD::BUILD_VECTOR, dl,
4054 MVT::v16i8, &pshufbMask[0], 16));
4055 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4056 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4059 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4060 // and update MaskVals with new element order.
4061 BitVector InOrder(8);
4062 if (BestLoQuad >= 0) {
4063 SmallVector<int, 8> MaskV;
4064 for (int i = 0; i != 4; ++i) {
4065 int idx = MaskVals[i];
4067 MaskV.push_back(-1);
4069 } else if ((idx / 4) == BestLoQuad) {
4070 MaskV.push_back(idx & 3);
4073 MaskV.push_back(-1);
4076 for (unsigned i = 4; i != 8; ++i)
4078 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4082 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4083 // and update MaskVals with the new element order.
4084 if (BestHiQuad >= 0) {
4085 SmallVector<int, 8> MaskV;
4086 for (unsigned i = 0; i != 4; ++i)
4088 for (unsigned i = 4; i != 8; ++i) {
4089 int idx = MaskVals[i];
4091 MaskV.push_back(-1);
4093 } else if ((idx / 4) == BestHiQuad) {
4094 MaskV.push_back((idx & 3) + 4);
4097 MaskV.push_back(-1);
4100 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4104 // In case BestHi & BestLo were both -1, which means each quadword has a word
4105 // from each of the four input quadwords, calculate the InOrder bitvector now
4106 // before falling through to the insert/extract cleanup.
4107 if (BestLoQuad == -1 && BestHiQuad == -1) {
4109 for (int i = 0; i != 8; ++i)
4110 if (MaskVals[i] < 0 || MaskVals[i] == i)
4114 // The other elements are put in the right place using pextrw and pinsrw.
4115 for (unsigned i = 0; i != 8; ++i) {
4118 int EltIdx = MaskVals[i];
4121 SDValue ExtOp = (EltIdx < 8)
4122 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4123 DAG.getIntPtrConstant(EltIdx))
4124 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4125 DAG.getIntPtrConstant(EltIdx - 8));
4126 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4127 DAG.getIntPtrConstant(i));
4132 // v16i8 shuffles - Prefer shuffles in the following order:
4133 // 1. [ssse3] 1 x pshufb
4134 // 2. [ssse3] 2 x pshufb + 1 x por
4135 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4137 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4138 SelectionDAG &DAG, X86TargetLowering &TLI) {
4139 SDValue V1 = SVOp->getOperand(0);
4140 SDValue V2 = SVOp->getOperand(1);
4141 DebugLoc dl = SVOp->getDebugLoc();
4142 SmallVector<int, 16> MaskVals;
4143 SVOp->getMask(MaskVals);
4145 // If we have SSSE3, case 1 is generated when all result bytes come from
4146 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4147 // present, fall back to case 3.
4148 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4151 for (unsigned i = 0; i < 16; ++i) {
4152 int EltIdx = MaskVals[i];
4161 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4162 if (TLI.getSubtarget()->hasSSSE3()) {
4163 SmallVector<SDValue,16> pshufbMask;
4165 // If all result elements are from one input vector, then only translate
4166 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4168 // Otherwise, we have elements from both input vectors, and must zero out
4169 // elements that come from V2 in the first mask, and V1 in the second mask
4170 // so that we can OR them together.
4171 bool TwoInputs = !(V1Only || V2Only);
4172 for (unsigned i = 0; i != 16; ++i) {
4173 int EltIdx = MaskVals[i];
4174 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4175 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4178 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4180 // If all the elements are from V2, assign it to V1 and return after
4181 // building the first pshufb.
4184 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4185 DAG.getNode(ISD::BUILD_VECTOR, dl,
4186 MVT::v16i8, &pshufbMask[0], 16));
4190 // Calculate the shuffle mask for the second input, shuffle it, and
4191 // OR it with the first shuffled input.
4193 for (unsigned i = 0; i != 16; ++i) {
4194 int EltIdx = MaskVals[i];
4196 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4199 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4201 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4202 DAG.getNode(ISD::BUILD_VECTOR, dl,
4203 MVT::v16i8, &pshufbMask[0], 16));
4204 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4207 // No SSSE3 - Calculate in place words and then fix all out of place words
4208 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4209 // the 16 different words that comprise the two doublequadword input vectors.
4210 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4211 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4212 SDValue NewV = V2Only ? V2 : V1;
4213 for (int i = 0; i != 8; ++i) {
4214 int Elt0 = MaskVals[i*2];
4215 int Elt1 = MaskVals[i*2+1];
4217 // This word of the result is all undef, skip it.
4218 if (Elt0 < 0 && Elt1 < 0)
4221 // This word of the result is already in the correct place, skip it.
4222 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4224 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4227 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4228 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4231 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4232 // using a single extract together, load it and store it.
4233 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4234 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4235 DAG.getIntPtrConstant(Elt1 / 2));
4236 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4237 DAG.getIntPtrConstant(i));
4241 // If Elt1 is defined, extract it from the appropriate source. If the
4242 // source byte is not also odd, shift the extracted word left 8 bits
4243 // otherwise clear the bottom 8 bits if we need to do an or.
4245 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4246 DAG.getIntPtrConstant(Elt1 / 2));
4247 if ((Elt1 & 1) == 0)
4248 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4249 DAG.getConstant(8, TLI.getShiftAmountTy()));
4251 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4252 DAG.getConstant(0xFF00, MVT::i16));
4254 // If Elt0 is defined, extract it from the appropriate source. If the
4255 // source byte is not also even, shift the extracted word right 8 bits. If
4256 // Elt1 was also defined, OR the extracted values together before
4257 // inserting them in the result.
4259 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4260 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4261 if ((Elt0 & 1) != 0)
4262 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4263 DAG.getConstant(8, TLI.getShiftAmountTy()));
4265 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4266 DAG.getConstant(0x00FF, MVT::i16));
4267 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4270 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4271 DAG.getIntPtrConstant(i));
4273 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4276 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4277 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4278 /// done when every pair / quad of shuffle mask elements point to elements in
4279 /// the right sequence. e.g.
4280 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4282 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4284 TargetLowering &TLI, DebugLoc dl) {
4285 EVT VT = SVOp->getValueType(0);
4286 SDValue V1 = SVOp->getOperand(0);
4287 SDValue V2 = SVOp->getOperand(1);
4288 unsigned NumElems = VT.getVectorNumElements();
4289 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4290 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4291 EVT MaskEltVT = MaskVT.getVectorElementType();
4293 switch (VT.getSimpleVT().SimpleTy) {
4294 default: assert(false && "Unexpected!");
4295 case MVT::v4f32: NewVT = MVT::v2f64; break;
4296 case MVT::v4i32: NewVT = MVT::v2i64; break;
4297 case MVT::v8i16: NewVT = MVT::v4i32; break;
4298 case MVT::v16i8: NewVT = MVT::v4i32; break;
4301 if (NewWidth == 2) {
4307 int Scale = NumElems / NewWidth;
4308 SmallVector<int, 8> MaskVec;
4309 for (unsigned i = 0; i < NumElems; i += Scale) {
4311 for (int j = 0; j < Scale; ++j) {
4312 int EltIdx = SVOp->getMaskElt(i+j);
4316 StartIdx = EltIdx - (EltIdx % Scale);
4317 if (EltIdx != StartIdx + j)
4321 MaskVec.push_back(-1);
4323 MaskVec.push_back(StartIdx / Scale);
4326 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4327 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4328 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4331 /// getVZextMovL - Return a zero-extending vector move low node.
4333 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4334 SDValue SrcOp, SelectionDAG &DAG,
4335 const X86Subtarget *Subtarget, DebugLoc dl) {
4336 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4337 LoadSDNode *LD = NULL;
4338 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4339 LD = dyn_cast<LoadSDNode>(SrcOp);
4341 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4343 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4344 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4345 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4346 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4347 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4349 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4350 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4351 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4352 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4360 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4361 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4362 DAG.getNode(ISD::BIT_CONVERT, dl,
4366 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4369 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4370 SDValue V1 = SVOp->getOperand(0);
4371 SDValue V2 = SVOp->getOperand(1);
4372 DebugLoc dl = SVOp->getDebugLoc();
4373 EVT VT = SVOp->getValueType(0);
4375 SmallVector<std::pair<int, int>, 8> Locs;
4377 SmallVector<int, 8> Mask1(4U, -1);
4378 SmallVector<int, 8> PermMask;
4379 SVOp->getMask(PermMask);
4383 for (unsigned i = 0; i != 4; ++i) {
4384 int Idx = PermMask[i];
4386 Locs[i] = std::make_pair(-1, -1);
4388 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4390 Locs[i] = std::make_pair(0, NumLo);
4394 Locs[i] = std::make_pair(1, NumHi);
4396 Mask1[2+NumHi] = Idx;
4402 if (NumLo <= 2 && NumHi <= 2) {
4403 // If no more than two elements come from either vector. This can be
4404 // implemented with two shuffles. First shuffle gather the elements.
4405 // The second shuffle, which takes the first shuffle as both of its
4406 // vector operands, put the elements into the right order.
4407 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4409 SmallVector<int, 8> Mask2(4U, -1);
4411 for (unsigned i = 0; i != 4; ++i) {
4412 if (Locs[i].first == -1)
4415 unsigned Idx = (i < 2) ? 0 : 4;
4416 Idx += Locs[i].first * 2 + Locs[i].second;
4421 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4422 } else if (NumLo == 3 || NumHi == 3) {
4423 // Otherwise, we must have three elements from one vector, call it X, and
4424 // one element from the other, call it Y. First, use a shufps to build an
4425 // intermediate vector with the one element from Y and the element from X
4426 // that will be in the same half in the final destination (the indexes don't
4427 // matter). Then, use a shufps to build the final vector, taking the half
4428 // containing the element from Y from the intermediate, and the other half
4431 // Normalize it so the 3 elements come from V1.
4432 CommuteVectorShuffleMask(PermMask, VT);
4436 // Find the element from V2.
4438 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4439 int Val = PermMask[HiIndex];
4446 Mask1[0] = PermMask[HiIndex];
4448 Mask1[2] = PermMask[HiIndex^1];
4450 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4453 Mask1[0] = PermMask[0];
4454 Mask1[1] = PermMask[1];
4455 Mask1[2] = HiIndex & 1 ? 6 : 4;
4456 Mask1[3] = HiIndex & 1 ? 4 : 6;
4457 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4459 Mask1[0] = HiIndex & 1 ? 2 : 0;
4460 Mask1[1] = HiIndex & 1 ? 0 : 2;
4461 Mask1[2] = PermMask[2];
4462 Mask1[3] = PermMask[3];
4467 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4471 // Break it into (shuffle shuffle_hi, shuffle_lo).
4473 SmallVector<int,8> LoMask(4U, -1);
4474 SmallVector<int,8> HiMask(4U, -1);
4476 SmallVector<int,8> *MaskPtr = &LoMask;
4477 unsigned MaskIdx = 0;
4480 for (unsigned i = 0; i != 4; ++i) {
4487 int Idx = PermMask[i];
4489 Locs[i] = std::make_pair(-1, -1);
4490 } else if (Idx < 4) {
4491 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4492 (*MaskPtr)[LoIdx] = Idx;
4495 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4496 (*MaskPtr)[HiIdx] = Idx;
4501 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4502 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4503 SmallVector<int, 8> MaskOps;
4504 for (unsigned i = 0; i != 4; ++i) {
4505 if (Locs[i].first == -1) {
4506 MaskOps.push_back(-1);
4508 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4509 MaskOps.push_back(Idx);
4512 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4516 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4517 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4518 SDValue V1 = Op.getOperand(0);
4519 SDValue V2 = Op.getOperand(1);
4520 EVT VT = Op.getValueType();
4521 DebugLoc dl = Op.getDebugLoc();
4522 unsigned NumElems = VT.getVectorNumElements();
4523 bool isMMX = VT.getSizeInBits() == 64;
4524 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4525 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4526 bool V1IsSplat = false;
4527 bool V2IsSplat = false;
4529 if (isZeroShuffle(SVOp))
4530 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4532 // Promote splats to v4f32.
4533 if (SVOp->isSplat()) {
4534 if (isMMX || NumElems < 4)
4536 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4539 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4541 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4542 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4543 if (NewOp.getNode())
4544 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4545 LowerVECTOR_SHUFFLE(NewOp, DAG));
4546 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4547 // FIXME: Figure out a cleaner way to do this.
4548 // Try to make use of movq to zero out the top part.
4549 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4550 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4551 if (NewOp.getNode()) {
4552 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4553 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4554 DAG, Subtarget, dl);
4556 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4557 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4558 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4559 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4560 DAG, Subtarget, dl);
4564 if (X86::isPSHUFDMask(SVOp))
4567 // Check if this can be converted into a logical shift.
4568 bool isLeft = false;
4571 bool isShift = getSubtarget()->hasSSE2() &&
4572 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4573 if (isShift && ShVal.hasOneUse()) {
4574 // If the shifted value has multiple uses, it may be cheaper to use
4575 // v_set0 + movlhps or movhlps, etc.
4576 EVT EltVT = VT.getVectorElementType();
4577 ShAmt *= EltVT.getSizeInBits();
4578 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4581 if (X86::isMOVLMask(SVOp)) {
4584 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4585 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4590 // FIXME: fold these into legal mask.
4591 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4592 X86::isMOVSLDUPMask(SVOp) ||
4593 X86::isMOVHLPSMask(SVOp) ||
4594 X86::isMOVLHPSMask(SVOp) ||
4595 X86::isMOVLPMask(SVOp)))
4598 if (ShouldXformToMOVHLPS(SVOp) ||
4599 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4600 return CommuteVectorShuffle(SVOp, DAG);
4603 // No better options. Use a vshl / vsrl.
4604 EVT EltVT = VT.getVectorElementType();
4605 ShAmt *= EltVT.getSizeInBits();
4606 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4609 bool Commuted = false;
4610 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4611 // 1,1,1,1 -> v8i16 though.
4612 V1IsSplat = isSplatVector(V1.getNode());
4613 V2IsSplat = isSplatVector(V2.getNode());
4615 // Canonicalize the splat or undef, if present, to be on the RHS.
4616 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4617 Op = CommuteVectorShuffle(SVOp, DAG);
4618 SVOp = cast<ShuffleVectorSDNode>(Op);
4619 V1 = SVOp->getOperand(0);
4620 V2 = SVOp->getOperand(1);
4621 std::swap(V1IsSplat, V2IsSplat);
4622 std::swap(V1IsUndef, V2IsUndef);
4626 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4627 // Shuffling low element of v1 into undef, just return v1.
4630 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4631 // the instruction selector will not match, so get a canonical MOVL with
4632 // swapped operands to undo the commute.
4633 return getMOVL(DAG, dl, VT, V2, V1);
4636 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4637 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4638 X86::isUNPCKLMask(SVOp) ||
4639 X86::isUNPCKHMask(SVOp))
4643 // Normalize mask so all entries that point to V2 points to its first
4644 // element then try to match unpck{h|l} again. If match, return a
4645 // new vector_shuffle with the corrected mask.
4646 SDValue NewMask = NormalizeMask(SVOp, DAG);
4647 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4648 if (NSVOp != SVOp) {
4649 if (X86::isUNPCKLMask(NSVOp, true)) {
4651 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4658 // Commute is back and try unpck* again.
4659 // FIXME: this seems wrong.
4660 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4661 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4662 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4663 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4664 X86::isUNPCKLMask(NewSVOp) ||
4665 X86::isUNPCKHMask(NewSVOp))
4669 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4671 // Normalize the node to match x86 shuffle ops if needed
4672 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4673 return CommuteVectorShuffle(SVOp, DAG);
4675 // Check for legal shuffle and return?
4676 SmallVector<int, 16> PermMask;
4677 SVOp->getMask(PermMask);
4678 if (isShuffleMaskLegal(PermMask, VT))
4681 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4682 if (VT == MVT::v8i16) {
4683 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4684 if (NewOp.getNode())
4688 if (VT == MVT::v16i8) {
4689 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4690 if (NewOp.getNode())
4694 // Handle all 4 wide cases with a number of shuffles except for MMX.
4695 if (NumElems == 4 && !isMMX)
4696 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4702 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4703 SelectionDAG &DAG) {
4704 EVT VT = Op.getValueType();
4705 DebugLoc dl = Op.getDebugLoc();
4706 if (VT.getSizeInBits() == 8) {
4707 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4708 Op.getOperand(0), Op.getOperand(1));
4709 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4710 DAG.getValueType(VT));
4711 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4712 } else if (VT.getSizeInBits() == 16) {
4713 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4714 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4716 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4717 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4718 DAG.getNode(ISD::BIT_CONVERT, dl,
4722 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4723 Op.getOperand(0), Op.getOperand(1));
4724 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4725 DAG.getValueType(VT));
4726 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4727 } else if (VT == MVT::f32) {
4728 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4729 // the result back to FR32 register. It's only worth matching if the
4730 // result has a single use which is a store or a bitcast to i32. And in
4731 // the case of a store, it's not worth it if the index is a constant 0,
4732 // because a MOVSSmr can be used instead, which is smaller and faster.
4733 if (!Op.hasOneUse())
4735 SDNode *User = *Op.getNode()->use_begin();
4736 if ((User->getOpcode() != ISD::STORE ||
4737 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4738 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4739 (User->getOpcode() != ISD::BIT_CONVERT ||
4740 User->getValueType(0) != MVT::i32))
4742 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4743 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4746 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4747 } else if (VT == MVT::i32) {
4748 // ExtractPS works with constant index.
4749 if (isa<ConstantSDNode>(Op.getOperand(1)))
4757 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4758 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4761 if (Subtarget->hasSSE41()) {
4762 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4767 EVT VT = Op.getValueType();
4768 DebugLoc dl = Op.getDebugLoc();
4769 // TODO: handle v16i8.
4770 if (VT.getSizeInBits() == 16) {
4771 SDValue Vec = Op.getOperand(0);
4772 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4774 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4775 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4776 DAG.getNode(ISD::BIT_CONVERT, dl,
4779 // Transform it so it match pextrw which produces a 32-bit result.
4780 EVT EltVT = MVT::i32;
4781 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4782 Op.getOperand(0), Op.getOperand(1));
4783 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4784 DAG.getValueType(VT));
4785 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4786 } else if (VT.getSizeInBits() == 32) {
4787 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4791 // SHUFPS the element to the lowest double word, then movss.
4792 int Mask[4] = { Idx, -1, -1, -1 };
4793 EVT VVT = Op.getOperand(0).getValueType();
4794 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4795 DAG.getUNDEF(VVT), Mask);
4796 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4797 DAG.getIntPtrConstant(0));
4798 } else if (VT.getSizeInBits() == 64) {
4799 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4800 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4801 // to match extract_elt for f64.
4802 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4806 // UNPCKHPD the element to the lowest double word, then movsd.
4807 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4808 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4809 int Mask[2] = { 1, -1 };
4810 EVT VVT = Op.getOperand(0).getValueType();
4811 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4812 DAG.getUNDEF(VVT), Mask);
4813 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4814 DAG.getIntPtrConstant(0));
4821 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4822 EVT VT = Op.getValueType();
4823 EVT EltVT = VT.getVectorElementType();
4824 DebugLoc dl = Op.getDebugLoc();
4826 SDValue N0 = Op.getOperand(0);
4827 SDValue N1 = Op.getOperand(1);
4828 SDValue N2 = Op.getOperand(2);
4830 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4831 isa<ConstantSDNode>(N2)) {
4833 if (VT == MVT::v8i16)
4834 Opc = X86ISD::PINSRW;
4835 else if (VT == MVT::v4i16)
4836 Opc = X86ISD::MMX_PINSRW;
4837 else if (VT == MVT::v16i8)
4838 Opc = X86ISD::PINSRB;
4840 Opc = X86ISD::PINSRB;
4842 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4844 if (N1.getValueType() != MVT::i32)
4845 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4846 if (N2.getValueType() != MVT::i32)
4847 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4848 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4849 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4850 // Bits [7:6] of the constant are the source select. This will always be
4851 // zero here. The DAG Combiner may combine an extract_elt index into these
4852 // bits. For example (insert (extract, 3), 2) could be matched by putting
4853 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4854 // Bits [5:4] of the constant are the destination select. This is the
4855 // value of the incoming immediate.
4856 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4857 // combine either bitwise AND or insert of float 0.0 to set these bits.
4858 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4859 // Create this as a scalar to vector..
4860 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4861 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4862 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4863 // PINSR* works with constant index.
4870 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4871 EVT VT = Op.getValueType();
4872 EVT EltVT = VT.getVectorElementType();
4874 if (Subtarget->hasSSE41())
4875 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4877 if (EltVT == MVT::i8)
4880 DebugLoc dl = Op.getDebugLoc();
4881 SDValue N0 = Op.getOperand(0);
4882 SDValue N1 = Op.getOperand(1);
4883 SDValue N2 = Op.getOperand(2);
4885 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4886 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4887 // as its second argument.
4888 if (N1.getValueType() != MVT::i32)
4889 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4890 if (N2.getValueType() != MVT::i32)
4891 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4892 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
4893 dl, VT, N0, N1, N2);
4899 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4900 DebugLoc dl = Op.getDebugLoc();
4901 if (Op.getValueType() == MVT::v2f32)
4902 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4903 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4904 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4905 Op.getOperand(0))));
4907 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4908 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4910 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4911 EVT VT = MVT::v2i32;
4912 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4919 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4920 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4923 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4924 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4925 // one of the above mentioned nodes. It has to be wrapped because otherwise
4926 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4927 // be used to form addressing mode. These wrapped nodes will be selected
4930 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4931 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4933 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4935 unsigned char OpFlag = 0;
4936 unsigned WrapperKind = X86ISD::Wrapper;
4937 CodeModel::Model M = getTargetMachine().getCodeModel();
4939 if (Subtarget->isPICStyleRIPRel() &&
4940 (M == CodeModel::Small || M == CodeModel::Kernel))
4941 WrapperKind = X86ISD::WrapperRIP;
4942 else if (Subtarget->isPICStyleGOT())
4943 OpFlag = X86II::MO_GOTOFF;
4944 else if (Subtarget->isPICStyleStubPIC())
4945 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4947 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4949 CP->getOffset(), OpFlag);
4950 DebugLoc DL = CP->getDebugLoc();
4951 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4952 // With PIC, the address is actually $g + Offset.
4954 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4955 DAG.getNode(X86ISD::GlobalBaseReg,
4956 DebugLoc::getUnknownLoc(), getPointerTy()),
4963 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4964 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4966 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4968 unsigned char OpFlag = 0;
4969 unsigned WrapperKind = X86ISD::Wrapper;
4970 CodeModel::Model M = getTargetMachine().getCodeModel();
4972 if (Subtarget->isPICStyleRIPRel() &&
4973 (M == CodeModel::Small || M == CodeModel::Kernel))
4974 WrapperKind = X86ISD::WrapperRIP;
4975 else if (Subtarget->isPICStyleGOT())
4976 OpFlag = X86II::MO_GOTOFF;
4977 else if (Subtarget->isPICStyleStubPIC())
4978 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4980 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4982 DebugLoc DL = JT->getDebugLoc();
4983 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4985 // With PIC, the address is actually $g + Offset.
4987 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4988 DAG.getNode(X86ISD::GlobalBaseReg,
4989 DebugLoc::getUnknownLoc(), getPointerTy()),
4997 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4998 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5000 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5002 unsigned char OpFlag = 0;
5003 unsigned WrapperKind = X86ISD::Wrapper;
5004 CodeModel::Model M = getTargetMachine().getCodeModel();
5006 if (Subtarget->isPICStyleRIPRel() &&
5007 (M == CodeModel::Small || M == CodeModel::Kernel))
5008 WrapperKind = X86ISD::WrapperRIP;
5009 else if (Subtarget->isPICStyleGOT())
5010 OpFlag = X86II::MO_GOTOFF;
5011 else if (Subtarget->isPICStyleStubPIC())
5012 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5014 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5016 DebugLoc DL = Op.getDebugLoc();
5017 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5020 // With PIC, the address is actually $g + Offset.
5021 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5022 !Subtarget->is64Bit()) {
5023 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5024 DAG.getNode(X86ISD::GlobalBaseReg,
5025 DebugLoc::getUnknownLoc(),
5034 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5035 // Create the TargetBlockAddressAddress node.
5036 unsigned char OpFlags =
5037 Subtarget->ClassifyBlockAddressReference();
5038 CodeModel::Model M = getTargetMachine().getCodeModel();
5039 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5040 DebugLoc dl = Op.getDebugLoc();
5041 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5042 /*isTarget=*/true, OpFlags);
5044 if (Subtarget->isPICStyleRIPRel() &&
5045 (M == CodeModel::Small || M == CodeModel::Kernel))
5046 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5048 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5050 // With PIC, the address is actually $g + Offset.
5051 if (isGlobalRelativeToPICBase(OpFlags)) {
5052 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5053 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5061 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5063 SelectionDAG &DAG) const {
5064 // Create the TargetGlobalAddress node, folding in the constant
5065 // offset if it is legal.
5066 unsigned char OpFlags =
5067 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5068 CodeModel::Model M = getTargetMachine().getCodeModel();
5070 if (OpFlags == X86II::MO_NO_FLAG &&
5071 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5072 // A direct static reference to a global.
5073 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5076 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5079 if (Subtarget->isPICStyleRIPRel() &&
5080 (M == CodeModel::Small || M == CodeModel::Kernel))
5081 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5083 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5085 // With PIC, the address is actually $g + Offset.
5086 if (isGlobalRelativeToPICBase(OpFlags)) {
5087 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5088 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5092 // For globals that require a load from a stub to get the address, emit the
5094 if (isGlobalStubReference(OpFlags))
5095 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5096 PseudoSourceValue::getGOT(), 0, false, false, 0);
5098 // If there was a non-zero offset that we didn't fold, create an explicit
5101 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5102 DAG.getConstant(Offset, getPointerTy()));
5108 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5109 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5110 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5111 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5115 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5116 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5117 unsigned char OperandFlags) {
5118 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5119 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5120 DebugLoc dl = GA->getDebugLoc();
5121 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5122 GA->getValueType(0),
5126 SDValue Ops[] = { Chain, TGA, *InFlag };
5127 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5129 SDValue Ops[] = { Chain, TGA };
5130 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5133 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5134 MFI->setHasCalls(true);
5136 SDValue Flag = Chain.getValue(1);
5137 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5140 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5142 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5145 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5146 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5147 DAG.getNode(X86ISD::GlobalBaseReg,
5148 DebugLoc::getUnknownLoc(),
5150 InFlag = Chain.getValue(1);
5152 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5155 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5157 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5159 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5160 X86::RAX, X86II::MO_TLSGD);
5163 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5164 // "local exec" model.
5165 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5166 const EVT PtrVT, TLSModel::Model model,
5168 DebugLoc dl = GA->getDebugLoc();
5169 // Get the Thread Pointer
5170 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5171 DebugLoc::getUnknownLoc(), PtrVT,
5172 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5175 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5176 NULL, 0, false, false, 0);
5178 unsigned char OperandFlags = 0;
5179 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5181 unsigned WrapperKind = X86ISD::Wrapper;
5182 if (model == TLSModel::LocalExec) {
5183 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5184 } else if (is64Bit) {
5185 assert(model == TLSModel::InitialExec);
5186 OperandFlags = X86II::MO_GOTTPOFF;
5187 WrapperKind = X86ISD::WrapperRIP;
5189 assert(model == TLSModel::InitialExec);
5190 OperandFlags = X86II::MO_INDNTPOFF;
5193 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5195 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5196 GA->getOffset(), OperandFlags);
5197 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5199 if (model == TLSModel::InitialExec)
5200 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5201 PseudoSourceValue::getGOT(), 0, false, false, 0);
5203 // The address of the thread local variable is the add of the thread
5204 // pointer with the offset of the variable.
5205 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5209 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5210 // TODO: implement the "local dynamic" model
5211 // TODO: implement the "initial exec"model for pic executables
5212 assert(Subtarget->isTargetELF() &&
5213 "TLS not implemented for non-ELF targets");
5214 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5215 const GlobalValue *GV = GA->getGlobal();
5217 // If GV is an alias then use the aliasee for determining
5218 // thread-localness.
5219 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5220 GV = GA->resolveAliasedGlobal(false);
5222 TLSModel::Model model = getTLSModel(GV,
5223 getTargetMachine().getRelocationModel());
5226 case TLSModel::GeneralDynamic:
5227 case TLSModel::LocalDynamic: // not implemented
5228 if (Subtarget->is64Bit())
5229 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5230 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5232 case TLSModel::InitialExec:
5233 case TLSModel::LocalExec:
5234 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5235 Subtarget->is64Bit());
5238 llvm_unreachable("Unreachable");
5243 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5244 /// take a 2 x i32 value to shift plus a shift amount.
5245 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5246 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5247 EVT VT = Op.getValueType();
5248 unsigned VTBits = VT.getSizeInBits();
5249 DebugLoc dl = Op.getDebugLoc();
5250 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5251 SDValue ShOpLo = Op.getOperand(0);
5252 SDValue ShOpHi = Op.getOperand(1);
5253 SDValue ShAmt = Op.getOperand(2);
5254 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5255 DAG.getConstant(VTBits - 1, MVT::i8))
5256 : DAG.getConstant(0, VT);
5259 if (Op.getOpcode() == ISD::SHL_PARTS) {
5260 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5261 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5263 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5264 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5267 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5268 DAG.getConstant(VTBits, MVT::i8));
5269 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5270 AndNode, DAG.getConstant(0, MVT::i8));
5273 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5274 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5275 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5277 if (Op.getOpcode() == ISD::SHL_PARTS) {
5278 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5279 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5281 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5282 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5285 SDValue Ops[2] = { Lo, Hi };
5286 return DAG.getMergeValues(Ops, 2, dl);
5289 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5290 EVT SrcVT = Op.getOperand(0).getValueType();
5292 if (SrcVT.isVector()) {
5293 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5299 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5300 "Unknown SINT_TO_FP to lower!");
5302 // These are really Legal; return the operand so the caller accepts it as
5304 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5306 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5307 Subtarget->is64Bit()) {
5311 DebugLoc dl = Op.getDebugLoc();
5312 unsigned Size = SrcVT.getSizeInBits()/8;
5313 MachineFunction &MF = DAG.getMachineFunction();
5314 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5315 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5316 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5318 PseudoSourceValue::getFixedStack(SSFI), 0,
5320 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5323 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5325 SelectionDAG &DAG) {
5327 DebugLoc dl = Op.getDebugLoc();
5329 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5331 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5333 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5334 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5335 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5336 Tys, Ops, array_lengthof(Ops));
5339 Chain = Result.getValue(1);
5340 SDValue InFlag = Result.getValue(2);
5342 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5343 // shouldn't be necessary except that RFP cannot be live across
5344 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5345 MachineFunction &MF = DAG.getMachineFunction();
5346 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5347 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5348 Tys = DAG.getVTList(MVT::Other);
5350 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5352 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5353 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5354 PseudoSourceValue::getFixedStack(SSFI), 0,
5361 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5362 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5363 // This algorithm is not obvious. Here it is in C code, more or less:
5365 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5366 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5367 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5369 // Copy ints to xmm registers.
5370 __m128i xh = _mm_cvtsi32_si128( hi );
5371 __m128i xl = _mm_cvtsi32_si128( lo );
5373 // Combine into low half of a single xmm register.
5374 __m128i x = _mm_unpacklo_epi32( xh, xl );
5378 // Merge in appropriate exponents to give the integer bits the right
5380 x = _mm_unpacklo_epi32( x, exp );
5382 // Subtract away the biases to deal with the IEEE-754 double precision
5384 d = _mm_sub_pd( (__m128d) x, bias );
5386 // All conversions up to here are exact. The correctly rounded result is
5387 // calculated using the current rounding mode using the following
5389 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5390 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5391 // store doesn't really need to be here (except
5392 // maybe to zero the other double)
5397 DebugLoc dl = Op.getDebugLoc();
5398 LLVMContext *Context = DAG.getContext();
5400 // Build some magic constants.
5401 std::vector<Constant*> CV0;
5402 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5403 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5404 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5405 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5406 Constant *C0 = ConstantVector::get(CV0);
5407 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5409 std::vector<Constant*> CV1;
5411 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5413 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5414 Constant *C1 = ConstantVector::get(CV1);
5415 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5417 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5418 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5420 DAG.getIntPtrConstant(1)));
5421 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5422 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5424 DAG.getIntPtrConstant(0)));
5425 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5426 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5427 PseudoSourceValue::getConstantPool(), 0,
5429 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5430 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5431 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5432 PseudoSourceValue::getConstantPool(), 0,
5434 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5436 // Add the halves; easiest way is to swap them into another reg first.
5437 int ShufMask[2] = { 1, -1 };
5438 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5439 DAG.getUNDEF(MVT::v2f64), ShufMask);
5440 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5441 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5442 DAG.getIntPtrConstant(0));
5445 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5446 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5447 DebugLoc dl = Op.getDebugLoc();
5448 // FP constant to bias correct the final result.
5449 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5452 // Load the 32-bit value into an XMM register.
5453 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5454 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5456 DAG.getIntPtrConstant(0)));
5458 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5459 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5460 DAG.getIntPtrConstant(0));
5462 // Or the load with the bias.
5463 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5464 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5465 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5467 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5468 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5469 MVT::v2f64, Bias)));
5470 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5471 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5472 DAG.getIntPtrConstant(0));
5474 // Subtract the bias.
5475 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5477 // Handle final rounding.
5478 EVT DestVT = Op.getValueType();
5480 if (DestVT.bitsLT(MVT::f64)) {
5481 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5482 DAG.getIntPtrConstant(0));
5483 } else if (DestVT.bitsGT(MVT::f64)) {
5484 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5487 // Handle final rounding.
5491 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5492 SDValue N0 = Op.getOperand(0);
5493 DebugLoc dl = Op.getDebugLoc();
5495 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5496 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5497 // the optimization here.
5498 if (DAG.SignBitIsZero(N0))
5499 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5501 EVT SrcVT = N0.getValueType();
5502 if (SrcVT == MVT::i64) {
5503 // We only handle SSE2 f64 target here; caller can expand the rest.
5504 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5507 return LowerUINT_TO_FP_i64(Op, DAG);
5508 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5509 return LowerUINT_TO_FP_i32(Op, DAG);
5512 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5514 // Make a 64-bit buffer, and use it to build an FILD.
5515 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5516 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5517 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5518 getPointerTy(), StackSlot, WordOff);
5519 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5520 StackSlot, NULL, 0, false, false, 0);
5521 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5522 OffsetSlot, NULL, 0, false, false, 0);
5523 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5526 std::pair<SDValue,SDValue> X86TargetLowering::
5527 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5528 DebugLoc dl = Op.getDebugLoc();
5530 EVT DstTy = Op.getValueType();
5533 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5537 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5538 DstTy.getSimpleVT() >= MVT::i16 &&
5539 "Unknown FP_TO_SINT to lower!");
5541 // These are really Legal.
5542 if (DstTy == MVT::i32 &&
5543 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5544 return std::make_pair(SDValue(), SDValue());
5545 if (Subtarget->is64Bit() &&
5546 DstTy == MVT::i64 &&
5547 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5548 return std::make_pair(SDValue(), SDValue());
5550 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5552 MachineFunction &MF = DAG.getMachineFunction();
5553 unsigned MemSize = DstTy.getSizeInBits()/8;
5554 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5555 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5558 switch (DstTy.getSimpleVT().SimpleTy) {
5559 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5560 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5561 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5562 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5565 SDValue Chain = DAG.getEntryNode();
5566 SDValue Value = Op.getOperand(0);
5567 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5568 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5569 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5570 PseudoSourceValue::getFixedStack(SSFI), 0,
5572 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5574 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5576 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5577 Chain = Value.getValue(1);
5578 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5579 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5582 // Build the FP_TO_INT*_IN_MEM
5583 SDValue Ops[] = { Chain, Value, StackSlot };
5584 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5586 return std::make_pair(FIST, StackSlot);
5589 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5590 if (Op.getValueType().isVector()) {
5591 if (Op.getValueType() == MVT::v2i32 &&
5592 Op.getOperand(0).getValueType() == MVT::v2f64) {
5598 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5599 SDValue FIST = Vals.first, StackSlot = Vals.second;
5600 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5601 if (FIST.getNode() == 0) return Op;
5604 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5605 FIST, StackSlot, NULL, 0, false, false, 0);
5608 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5609 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5610 SDValue FIST = Vals.first, StackSlot = Vals.second;
5611 assert(FIST.getNode() && "Unexpected failure");
5614 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5615 FIST, StackSlot, NULL, 0, false, false, 0);
5618 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5619 LLVMContext *Context = DAG.getContext();
5620 DebugLoc dl = Op.getDebugLoc();
5621 EVT VT = Op.getValueType();
5624 EltVT = VT.getVectorElementType();
5625 std::vector<Constant*> CV;
5626 if (EltVT == MVT::f64) {
5627 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5631 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5637 Constant *C = ConstantVector::get(CV);
5638 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5639 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5640 PseudoSourceValue::getConstantPool(), 0,
5642 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5645 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5646 LLVMContext *Context = DAG.getContext();
5647 DebugLoc dl = Op.getDebugLoc();
5648 EVT VT = Op.getValueType();
5651 EltVT = VT.getVectorElementType();
5652 std::vector<Constant*> CV;
5653 if (EltVT == MVT::f64) {
5654 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5658 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5664 Constant *C = ConstantVector::get(CV);
5665 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5666 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5667 PseudoSourceValue::getConstantPool(), 0,
5669 if (VT.isVector()) {
5670 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5671 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5672 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5674 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5676 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5680 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5681 LLVMContext *Context = DAG.getContext();
5682 SDValue Op0 = Op.getOperand(0);
5683 SDValue Op1 = Op.getOperand(1);
5684 DebugLoc dl = Op.getDebugLoc();
5685 EVT VT = Op.getValueType();
5686 EVT SrcVT = Op1.getValueType();
5688 // If second operand is smaller, extend it first.
5689 if (SrcVT.bitsLT(VT)) {
5690 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5693 // And if it is bigger, shrink it first.
5694 if (SrcVT.bitsGT(VT)) {
5695 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5699 // At this point the operands and the result should have the same
5700 // type, and that won't be f80 since that is not custom lowered.
5702 // First get the sign bit of second operand.
5703 std::vector<Constant*> CV;
5704 if (SrcVT == MVT::f64) {
5705 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5706 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5708 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5709 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5710 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5711 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5713 Constant *C = ConstantVector::get(CV);
5714 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5715 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5716 PseudoSourceValue::getConstantPool(), 0,
5718 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5720 // Shift sign bit right or left if the two operands have different types.
5721 if (SrcVT.bitsGT(VT)) {
5722 // Op0 is MVT::f32, Op1 is MVT::f64.
5723 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5724 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5725 DAG.getConstant(32, MVT::i32));
5726 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5727 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5728 DAG.getIntPtrConstant(0));
5731 // Clear first operand sign bit.
5733 if (VT == MVT::f64) {
5734 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5735 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5737 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5738 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5739 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5740 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5742 C = ConstantVector::get(CV);
5743 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5744 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5745 PseudoSourceValue::getConstantPool(), 0,
5747 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5749 // Or the value with the sign bit.
5750 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5753 /// Emit nodes that will be selected as "test Op0,Op0", or something
5755 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5756 SelectionDAG &DAG) {
5757 DebugLoc dl = Op.getDebugLoc();
5759 // CF and OF aren't always set the way we want. Determine which
5760 // of these we need.
5761 bool NeedCF = false;
5762 bool NeedOF = false;
5764 case X86::COND_A: case X86::COND_AE:
5765 case X86::COND_B: case X86::COND_BE:
5768 case X86::COND_G: case X86::COND_GE:
5769 case X86::COND_L: case X86::COND_LE:
5770 case X86::COND_O: case X86::COND_NO:
5776 // See if we can use the EFLAGS value from the operand instead of
5777 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5778 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5779 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5780 unsigned Opcode = 0;
5781 unsigned NumOperands = 0;
5782 switch (Op.getNode()->getOpcode()) {
5784 // Due to an isel shortcoming, be conservative if this add is likely to
5785 // be selected as part of a load-modify-store instruction. When the root
5786 // node in a match is a store, isel doesn't know how to remap non-chain
5787 // non-flag uses of other nodes in the match, such as the ADD in this
5788 // case. This leads to the ADD being left around and reselected, with
5789 // the result being two adds in the output.
5790 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5791 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5792 if (UI->getOpcode() == ISD::STORE)
5794 if (ConstantSDNode *C =
5795 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5796 // An add of one will be selected as an INC.
5797 if (C->getAPIntValue() == 1) {
5798 Opcode = X86ISD::INC;
5802 // An add of negative one (subtract of one) will be selected as a DEC.
5803 if (C->getAPIntValue().isAllOnesValue()) {
5804 Opcode = X86ISD::DEC;
5809 // Otherwise use a regular EFLAGS-setting add.
5810 Opcode = X86ISD::ADD;
5814 // If the primary and result isn't used, don't bother using X86ISD::AND,
5815 // because a TEST instruction will be better.
5816 bool NonFlagUse = false;
5817 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5818 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5820 unsigned UOpNo = UI.getOperandNo();
5821 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5822 // Look pass truncate.
5823 UOpNo = User->use_begin().getOperandNo();
5824 User = *User->use_begin();
5826 if (User->getOpcode() != ISD::BRCOND &&
5827 User->getOpcode() != ISD::SETCC &&
5828 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5840 // Due to the ISEL shortcoming noted above, be conservative if this op is
5841 // likely to be selected as part of a load-modify-store instruction.
5842 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5843 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5844 if (UI->getOpcode() == ISD::STORE)
5846 // Otherwise use a regular EFLAGS-setting instruction.
5847 switch (Op.getNode()->getOpcode()) {
5848 case ISD::SUB: Opcode = X86ISD::SUB; break;
5849 case ISD::OR: Opcode = X86ISD::OR; break;
5850 case ISD::XOR: Opcode = X86ISD::XOR; break;
5851 case ISD::AND: Opcode = X86ISD::AND; break;
5852 default: llvm_unreachable("unexpected operator!");
5863 return SDValue(Op.getNode(), 1);
5869 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5870 SmallVector<SDValue, 4> Ops;
5871 for (unsigned i = 0; i != NumOperands; ++i)
5872 Ops.push_back(Op.getOperand(i));
5873 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5874 DAG.ReplaceAllUsesWith(Op, New);
5875 return SDValue(New.getNode(), 1);
5879 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5880 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5881 DAG.getConstant(0, Op.getValueType()));
5884 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5886 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5887 SelectionDAG &DAG) {
5888 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5889 if (C->getAPIntValue() == 0)
5890 return EmitTest(Op0, X86CC, DAG);
5892 DebugLoc dl = Op0.getDebugLoc();
5893 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5896 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5897 /// if it's possible.
5898 static SDValue LowerToBT(SDValue And, ISD::CondCode CC,
5899 DebugLoc dl, SelectionDAG &DAG) {
5900 SDValue Op0 = And.getOperand(0);
5901 SDValue Op1 = And.getOperand(1);
5902 if (Op0.getOpcode() == ISD::TRUNCATE)
5903 Op0 = Op0.getOperand(0);
5904 if (Op1.getOpcode() == ISD::TRUNCATE)
5905 Op1 = Op1.getOperand(0);
5908 if (Op1.getOpcode() == ISD::SHL) {
5909 if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
5910 if (And10C->getZExtValue() == 1) {
5912 RHS = Op1.getOperand(1);
5914 } else if (Op0.getOpcode() == ISD::SHL) {
5915 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
5916 if (And00C->getZExtValue() == 1) {
5918 RHS = Op0.getOperand(1);
5920 } else if (Op1.getOpcode() == ISD::Constant) {
5921 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
5922 SDValue AndLHS = Op0;
5923 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5924 LHS = AndLHS.getOperand(0);
5925 RHS = AndLHS.getOperand(1);
5929 if (LHS.getNode()) {
5930 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5931 // instruction. Since the shift amount is in-range-or-undefined, we know
5932 // that doing a bittest on the i16 value is ok. We extend to i32 because
5933 // the encoding for the i16 version is larger than the i32 version.
5934 if (LHS.getValueType() == MVT::i8)
5935 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5937 // If the operand types disagree, extend the shift amount to match. Since
5938 // BT ignores high bits (like shifts) we can use anyextend.
5939 if (LHS.getValueType() != RHS.getValueType())
5940 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5942 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5943 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5944 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5945 DAG.getConstant(Cond, MVT::i8), BT);
5951 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5952 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5953 SDValue Op0 = Op.getOperand(0);
5954 SDValue Op1 = Op.getOperand(1);
5955 DebugLoc dl = Op.getDebugLoc();
5956 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5958 // Optimize to BT if possible.
5959 // Lower (X & (1 << N)) == 0 to BT(X, N).
5960 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5961 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5962 if (Op0.getOpcode() == ISD::AND &&
5964 Op1.getOpcode() == ISD::Constant &&
5965 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5966 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5967 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5968 if (NewSetCC.getNode())
5972 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
5973 if (Op0.getOpcode() == X86ISD::SETCC &&
5974 Op1.getOpcode() == ISD::Constant &&
5975 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
5976 cast<ConstantSDNode>(Op1)->isNullValue()) &&
5977 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5978 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
5979 bool Invert = (CC == ISD::SETNE) ^
5980 cast<ConstantSDNode>(Op1)->isNullValue();
5982 CCode = X86::GetOppositeBranchCondition(CCode);
5983 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5984 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
5987 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5988 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5989 if (X86CC == X86::COND_INVALID)
5992 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5994 // Use sbb x, x to materialize carry bit into a GPR.
5995 if (X86CC == X86::COND_B)
5996 return DAG.getNode(ISD::AND, dl, MVT::i8,
5997 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5998 DAG.getConstant(X86CC, MVT::i8), Cond),
5999 DAG.getConstant(1, MVT::i8));
6001 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6002 DAG.getConstant(X86CC, MVT::i8), Cond);
6005 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
6007 SDValue Op0 = Op.getOperand(0);
6008 SDValue Op1 = Op.getOperand(1);
6009 SDValue CC = Op.getOperand(2);
6010 EVT VT = Op.getValueType();
6011 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6012 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6013 DebugLoc dl = Op.getDebugLoc();
6017 EVT VT0 = Op0.getValueType();
6018 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6019 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6022 switch (SetCCOpcode) {
6025 case ISD::SETEQ: SSECC = 0; break;
6027 case ISD::SETGT: Swap = true; // Fallthrough
6029 case ISD::SETOLT: SSECC = 1; break;
6031 case ISD::SETGE: Swap = true; // Fallthrough
6033 case ISD::SETOLE: SSECC = 2; break;
6034 case ISD::SETUO: SSECC = 3; break;
6036 case ISD::SETNE: SSECC = 4; break;
6037 case ISD::SETULE: Swap = true;
6038 case ISD::SETUGE: SSECC = 5; break;
6039 case ISD::SETULT: Swap = true;
6040 case ISD::SETUGT: SSECC = 6; break;
6041 case ISD::SETO: SSECC = 7; break;
6044 std::swap(Op0, Op1);
6046 // In the two special cases we can't handle, emit two comparisons.
6048 if (SetCCOpcode == ISD::SETUEQ) {
6050 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6051 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6052 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6054 else if (SetCCOpcode == ISD::SETONE) {
6056 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6057 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6058 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6060 llvm_unreachable("Illegal FP comparison");
6062 // Handle all other FP comparisons here.
6063 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6066 // We are handling one of the integer comparisons here. Since SSE only has
6067 // GT and EQ comparisons for integer, swapping operands and multiple
6068 // operations may be required for some comparisons.
6069 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6070 bool Swap = false, Invert = false, FlipSigns = false;
6072 switch (VT.getSimpleVT().SimpleTy) {
6075 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6077 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6079 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6080 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6083 switch (SetCCOpcode) {
6085 case ISD::SETNE: Invert = true;
6086 case ISD::SETEQ: Opc = EQOpc; break;
6087 case ISD::SETLT: Swap = true;
6088 case ISD::SETGT: Opc = GTOpc; break;
6089 case ISD::SETGE: Swap = true;
6090 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6091 case ISD::SETULT: Swap = true;
6092 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6093 case ISD::SETUGE: Swap = true;
6094 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6097 std::swap(Op0, Op1);
6099 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6100 // bits of the inputs before performing those operations.
6102 EVT EltVT = VT.getVectorElementType();
6103 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6105 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6106 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6108 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6109 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6112 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6114 // If the logical-not of the result is required, perform that now.
6116 Result = DAG.getNOT(dl, Result, VT);
6121 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6122 static bool isX86LogicalCmp(SDValue Op) {
6123 unsigned Opc = Op.getNode()->getOpcode();
6124 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6126 if (Op.getResNo() == 1 &&
6127 (Opc == X86ISD::ADD ||
6128 Opc == X86ISD::SUB ||
6129 Opc == X86ISD::SMUL ||
6130 Opc == X86ISD::UMUL ||
6131 Opc == X86ISD::INC ||
6132 Opc == X86ISD::DEC ||
6133 Opc == X86ISD::OR ||
6134 Opc == X86ISD::XOR ||
6135 Opc == X86ISD::AND))
6141 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6142 bool addTest = true;
6143 SDValue Cond = Op.getOperand(0);
6144 DebugLoc dl = Op.getDebugLoc();
6147 if (Cond.getOpcode() == ISD::SETCC) {
6148 SDValue NewCond = LowerSETCC(Cond, DAG);
6149 if (NewCond.getNode())
6153 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6154 SDValue Op1 = Op.getOperand(1);
6155 SDValue Op2 = Op.getOperand(2);
6156 if (Cond.getOpcode() == X86ISD::SETCC &&
6157 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6158 SDValue Cmp = Cond.getOperand(1);
6159 if (Cmp.getOpcode() == X86ISD::CMP) {
6160 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6161 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6162 ConstantSDNode *RHSC =
6163 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6164 if (N1C && N1C->isAllOnesValue() &&
6165 N2C && N2C->isNullValue() &&
6166 RHSC && RHSC->isNullValue()) {
6167 SDValue CmpOp0 = Cmp.getOperand(0);
6168 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6169 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6170 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6171 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6176 // Look pass (and (setcc_carry (cmp ...)), 1).
6177 if (Cond.getOpcode() == ISD::AND &&
6178 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6179 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6180 if (C && C->getAPIntValue() == 1)
6181 Cond = Cond.getOperand(0);
6184 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6185 // setting operand in place of the X86ISD::SETCC.
6186 if (Cond.getOpcode() == X86ISD::SETCC ||
6187 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6188 CC = Cond.getOperand(0);
6190 SDValue Cmp = Cond.getOperand(1);
6191 unsigned Opc = Cmp.getOpcode();
6192 EVT VT = Op.getValueType();
6194 bool IllegalFPCMov = false;
6195 if (VT.isFloatingPoint() && !VT.isVector() &&
6196 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6197 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6199 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6200 Opc == X86ISD::BT) { // FIXME
6207 // Look pass the truncate.
6208 if (Cond.getOpcode() == ISD::TRUNCATE)
6209 Cond = Cond.getOperand(0);
6211 // We know the result of AND is compared against zero. Try to match
6213 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6214 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6215 if (NewSetCC.getNode()) {
6216 CC = NewSetCC.getOperand(0);
6217 Cond = NewSetCC.getOperand(1);
6224 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6225 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6228 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6229 // condition is true.
6230 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6231 SDValue Ops[] = { Op2, Op1, CC, Cond };
6232 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6235 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6236 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6237 // from the AND / OR.
6238 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6239 Opc = Op.getOpcode();
6240 if (Opc != ISD::OR && Opc != ISD::AND)
6242 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6243 Op.getOperand(0).hasOneUse() &&
6244 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6245 Op.getOperand(1).hasOneUse());
6248 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6249 // 1 and that the SETCC node has a single use.
6250 static bool isXor1OfSetCC(SDValue Op) {
6251 if (Op.getOpcode() != ISD::XOR)
6253 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6254 if (N1C && N1C->getAPIntValue() == 1) {
6255 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6256 Op.getOperand(0).hasOneUse();
6261 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6262 bool addTest = true;
6263 SDValue Chain = Op.getOperand(0);
6264 SDValue Cond = Op.getOperand(1);
6265 SDValue Dest = Op.getOperand(2);
6266 DebugLoc dl = Op.getDebugLoc();
6269 if (Cond.getOpcode() == ISD::SETCC) {
6270 SDValue NewCond = LowerSETCC(Cond, DAG);
6271 if (NewCond.getNode())
6275 // FIXME: LowerXALUO doesn't handle these!!
6276 else if (Cond.getOpcode() == X86ISD::ADD ||
6277 Cond.getOpcode() == X86ISD::SUB ||
6278 Cond.getOpcode() == X86ISD::SMUL ||
6279 Cond.getOpcode() == X86ISD::UMUL)
6280 Cond = LowerXALUO(Cond, DAG);
6283 // Look pass (and (setcc_carry (cmp ...)), 1).
6284 if (Cond.getOpcode() == ISD::AND &&
6285 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6286 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6287 if (C && C->getAPIntValue() == 1)
6288 Cond = Cond.getOperand(0);
6291 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6292 // setting operand in place of the X86ISD::SETCC.
6293 if (Cond.getOpcode() == X86ISD::SETCC ||
6294 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6295 CC = Cond.getOperand(0);
6297 SDValue Cmp = Cond.getOperand(1);
6298 unsigned Opc = Cmp.getOpcode();
6299 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6300 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6304 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6308 // These can only come from an arithmetic instruction with overflow,
6309 // e.g. SADDO, UADDO.
6310 Cond = Cond.getNode()->getOperand(1);
6317 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6318 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6319 if (CondOpc == ISD::OR) {
6320 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6321 // two branches instead of an explicit OR instruction with a
6323 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6324 isX86LogicalCmp(Cmp)) {
6325 CC = Cond.getOperand(0).getOperand(0);
6326 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6327 Chain, Dest, CC, Cmp);
6328 CC = Cond.getOperand(1).getOperand(0);
6332 } else { // ISD::AND
6333 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6334 // two branches instead of an explicit AND instruction with a
6335 // separate test. However, we only do this if this block doesn't
6336 // have a fall-through edge, because this requires an explicit
6337 // jmp when the condition is false.
6338 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6339 isX86LogicalCmp(Cmp) &&
6340 Op.getNode()->hasOneUse()) {
6341 X86::CondCode CCode =
6342 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6343 CCode = X86::GetOppositeBranchCondition(CCode);
6344 CC = DAG.getConstant(CCode, MVT::i8);
6345 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6346 // Look for an unconditional branch following this conditional branch.
6347 // We need this because we need to reverse the successors in order
6348 // to implement FCMP_OEQ.
6349 if (User.getOpcode() == ISD::BR) {
6350 SDValue FalseBB = User.getOperand(1);
6352 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6353 assert(NewBR == User);
6356 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6357 Chain, Dest, CC, Cmp);
6358 X86::CondCode CCode =
6359 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6360 CCode = X86::GetOppositeBranchCondition(CCode);
6361 CC = DAG.getConstant(CCode, MVT::i8);
6367 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6368 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6369 // It should be transformed during dag combiner except when the condition
6370 // is set by a arithmetics with overflow node.
6371 X86::CondCode CCode =
6372 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6373 CCode = X86::GetOppositeBranchCondition(CCode);
6374 CC = DAG.getConstant(CCode, MVT::i8);
6375 Cond = Cond.getOperand(0).getOperand(1);
6381 // Look pass the truncate.
6382 if (Cond.getOpcode() == ISD::TRUNCATE)
6383 Cond = Cond.getOperand(0);
6385 // We know the result of AND is compared against zero. Try to match
6387 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6388 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6389 if (NewSetCC.getNode()) {
6390 CC = NewSetCC.getOperand(0);
6391 Cond = NewSetCC.getOperand(1);
6398 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6399 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6401 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6402 Chain, Dest, CC, Cond);
6406 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6407 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6408 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6409 // that the guard pages used by the OS virtual memory manager are allocated in
6410 // correct sequence.
6412 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6413 SelectionDAG &DAG) {
6414 assert(Subtarget->isTargetCygMing() &&
6415 "This should be used only on Cygwin/Mingw targets");
6416 DebugLoc dl = Op.getDebugLoc();
6419 SDValue Chain = Op.getOperand(0);
6420 SDValue Size = Op.getOperand(1);
6421 // FIXME: Ensure alignment here
6425 EVT IntPtr = getPointerTy();
6426 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6428 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6429 Flag = Chain.getValue(1);
6431 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6433 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6434 Flag = Chain.getValue(1);
6436 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6438 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6439 return DAG.getMergeValues(Ops1, 2, dl);
6443 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6445 SDValue Dst, SDValue Src,
6446 SDValue Size, unsigned Align,
6448 uint64_t DstSVOff) {
6449 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6451 // If not DWORD aligned or size is more than the threshold, call the library.
6452 // The libc version is likely to be faster for these cases. It can use the
6453 // address value and run time information about the CPU.
6454 if ((Align & 3) != 0 ||
6456 ConstantSize->getZExtValue() >
6457 getSubtarget()->getMaxInlineSizeThreshold()) {
6458 SDValue InFlag(0, 0);
6460 // Check to see if there is a specialized entry-point for memory zeroing.
6461 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6463 if (const char *bzeroEntry = V &&
6464 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6465 EVT IntPtr = getPointerTy();
6466 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6467 TargetLowering::ArgListTy Args;
6468 TargetLowering::ArgListEntry Entry;
6470 Entry.Ty = IntPtrTy;
6471 Args.push_back(Entry);
6473 Args.push_back(Entry);
6474 std::pair<SDValue,SDValue> CallResult =
6475 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6476 false, false, false, false,
6477 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6478 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
6479 return CallResult.second;
6482 // Otherwise have the target-independent code call memset.
6486 uint64_t SizeVal = ConstantSize->getZExtValue();
6487 SDValue InFlag(0, 0);
6490 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6491 unsigned BytesLeft = 0;
6492 bool TwoRepStos = false;
6495 uint64_t Val = ValC->getZExtValue() & 255;
6497 // If the value is a constant, then we can potentially use larger sets.
6498 switch (Align & 3) {
6499 case 2: // WORD aligned
6502 Val = (Val << 8) | Val;
6504 case 0: // DWORD aligned
6507 Val = (Val << 8) | Val;
6508 Val = (Val << 16) | Val;
6509 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6512 Val = (Val << 32) | Val;
6515 default: // Byte aligned
6518 Count = DAG.getIntPtrConstant(SizeVal);
6522 if (AVT.bitsGT(MVT::i8)) {
6523 unsigned UBytes = AVT.getSizeInBits() / 8;
6524 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6525 BytesLeft = SizeVal % UBytes;
6528 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6530 InFlag = Chain.getValue(1);
6533 Count = DAG.getIntPtrConstant(SizeVal);
6534 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6535 InFlag = Chain.getValue(1);
6538 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6541 InFlag = Chain.getValue(1);
6542 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6545 InFlag = Chain.getValue(1);
6547 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6548 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6549 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6552 InFlag = Chain.getValue(1);
6554 EVT CVT = Count.getValueType();
6555 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6556 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6557 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6560 InFlag = Chain.getValue(1);
6561 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6562 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6563 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6564 } else if (BytesLeft) {
6565 // Handle the last 1 - 7 bytes.
6566 unsigned Offset = SizeVal - BytesLeft;
6567 EVT AddrVT = Dst.getValueType();
6568 EVT SizeVT = Size.getValueType();
6570 Chain = DAG.getMemset(Chain, dl,
6571 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6572 DAG.getConstant(Offset, AddrVT)),
6574 DAG.getConstant(BytesLeft, SizeVT),
6575 Align, DstSV, DstSVOff + Offset);
6578 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6583 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6584 SDValue Chain, SDValue Dst, SDValue Src,
6585 SDValue Size, unsigned Align,
6587 const Value *DstSV, uint64_t DstSVOff,
6588 const Value *SrcSV, uint64_t SrcSVOff) {
6589 // This requires the copy size to be a constant, preferrably
6590 // within a subtarget-specific limit.
6591 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6594 uint64_t SizeVal = ConstantSize->getZExtValue();
6595 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6598 /// If not DWORD aligned, call the library.
6599 if ((Align & 3) != 0)
6604 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6607 unsigned UBytes = AVT.getSizeInBits() / 8;
6608 unsigned CountVal = SizeVal / UBytes;
6609 SDValue Count = DAG.getIntPtrConstant(CountVal);
6610 unsigned BytesLeft = SizeVal % UBytes;
6612 SDValue InFlag(0, 0);
6613 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6616 InFlag = Chain.getValue(1);
6617 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6620 InFlag = Chain.getValue(1);
6621 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6624 InFlag = Chain.getValue(1);
6626 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6627 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6628 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6629 array_lengthof(Ops));
6631 SmallVector<SDValue, 4> Results;
6632 Results.push_back(RepMovs);
6634 // Handle the last 1 - 7 bytes.
6635 unsigned Offset = SizeVal - BytesLeft;
6636 EVT DstVT = Dst.getValueType();
6637 EVT SrcVT = Src.getValueType();
6638 EVT SizeVT = Size.getValueType();
6639 Results.push_back(DAG.getMemcpy(Chain, dl,
6640 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6641 DAG.getConstant(Offset, DstVT)),
6642 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6643 DAG.getConstant(Offset, SrcVT)),
6644 DAG.getConstant(BytesLeft, SizeVT),
6645 Align, AlwaysInline,
6646 DstSV, DstSVOff + Offset,
6647 SrcSV, SrcSVOff + Offset));
6650 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6651 &Results[0], Results.size());
6654 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6655 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6656 DebugLoc dl = Op.getDebugLoc();
6658 if (!Subtarget->is64Bit()) {
6659 // vastart just stores the address of the VarArgsFrameIndex slot into the
6660 // memory location argument.
6661 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6662 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6667 // gp_offset (0 - 6 * 8)
6668 // fp_offset (48 - 48 + 8 * 16)
6669 // overflow_arg_area (point to parameters coming in memory).
6671 SmallVector<SDValue, 8> MemOps;
6672 SDValue FIN = Op.getOperand(1);
6674 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6675 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6676 FIN, SV, 0, false, false, 0);
6677 MemOps.push_back(Store);
6680 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6681 FIN, DAG.getIntPtrConstant(4));
6682 Store = DAG.getStore(Op.getOperand(0), dl,
6683 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6684 FIN, SV, 0, false, false, 0);
6685 MemOps.push_back(Store);
6687 // Store ptr to overflow_arg_area
6688 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6689 FIN, DAG.getIntPtrConstant(4));
6690 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6691 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6693 MemOps.push_back(Store);
6695 // Store ptr to reg_save_area.
6696 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6697 FIN, DAG.getIntPtrConstant(8));
6698 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6699 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6701 MemOps.push_back(Store);
6702 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6703 &MemOps[0], MemOps.size());
6706 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6707 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6708 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6709 SDValue Chain = Op.getOperand(0);
6710 SDValue SrcPtr = Op.getOperand(1);
6711 SDValue SrcSV = Op.getOperand(2);
6713 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6717 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6718 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6719 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6720 SDValue Chain = Op.getOperand(0);
6721 SDValue DstPtr = Op.getOperand(1);
6722 SDValue SrcPtr = Op.getOperand(2);
6723 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6724 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6725 DebugLoc dl = Op.getDebugLoc();
6727 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6728 DAG.getIntPtrConstant(24), 8, false,
6729 DstSV, 0, SrcSV, 0);
6733 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6734 DebugLoc dl = Op.getDebugLoc();
6735 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6737 default: return SDValue(); // Don't custom lower most intrinsics.
6738 // Comparison intrinsics.
6739 case Intrinsic::x86_sse_comieq_ss:
6740 case Intrinsic::x86_sse_comilt_ss:
6741 case Intrinsic::x86_sse_comile_ss:
6742 case Intrinsic::x86_sse_comigt_ss:
6743 case Intrinsic::x86_sse_comige_ss:
6744 case Intrinsic::x86_sse_comineq_ss:
6745 case Intrinsic::x86_sse_ucomieq_ss:
6746 case Intrinsic::x86_sse_ucomilt_ss:
6747 case Intrinsic::x86_sse_ucomile_ss:
6748 case Intrinsic::x86_sse_ucomigt_ss:
6749 case Intrinsic::x86_sse_ucomige_ss:
6750 case Intrinsic::x86_sse_ucomineq_ss:
6751 case Intrinsic::x86_sse2_comieq_sd:
6752 case Intrinsic::x86_sse2_comilt_sd:
6753 case Intrinsic::x86_sse2_comile_sd:
6754 case Intrinsic::x86_sse2_comigt_sd:
6755 case Intrinsic::x86_sse2_comige_sd:
6756 case Intrinsic::x86_sse2_comineq_sd:
6757 case Intrinsic::x86_sse2_ucomieq_sd:
6758 case Intrinsic::x86_sse2_ucomilt_sd:
6759 case Intrinsic::x86_sse2_ucomile_sd:
6760 case Intrinsic::x86_sse2_ucomigt_sd:
6761 case Intrinsic::x86_sse2_ucomige_sd:
6762 case Intrinsic::x86_sse2_ucomineq_sd: {
6764 ISD::CondCode CC = ISD::SETCC_INVALID;
6767 case Intrinsic::x86_sse_comieq_ss:
6768 case Intrinsic::x86_sse2_comieq_sd:
6772 case Intrinsic::x86_sse_comilt_ss:
6773 case Intrinsic::x86_sse2_comilt_sd:
6777 case Intrinsic::x86_sse_comile_ss:
6778 case Intrinsic::x86_sse2_comile_sd:
6782 case Intrinsic::x86_sse_comigt_ss:
6783 case Intrinsic::x86_sse2_comigt_sd:
6787 case Intrinsic::x86_sse_comige_ss:
6788 case Intrinsic::x86_sse2_comige_sd:
6792 case Intrinsic::x86_sse_comineq_ss:
6793 case Intrinsic::x86_sse2_comineq_sd:
6797 case Intrinsic::x86_sse_ucomieq_ss:
6798 case Intrinsic::x86_sse2_ucomieq_sd:
6799 Opc = X86ISD::UCOMI;
6802 case Intrinsic::x86_sse_ucomilt_ss:
6803 case Intrinsic::x86_sse2_ucomilt_sd:
6804 Opc = X86ISD::UCOMI;
6807 case Intrinsic::x86_sse_ucomile_ss:
6808 case Intrinsic::x86_sse2_ucomile_sd:
6809 Opc = X86ISD::UCOMI;
6812 case Intrinsic::x86_sse_ucomigt_ss:
6813 case Intrinsic::x86_sse2_ucomigt_sd:
6814 Opc = X86ISD::UCOMI;
6817 case Intrinsic::x86_sse_ucomige_ss:
6818 case Intrinsic::x86_sse2_ucomige_sd:
6819 Opc = X86ISD::UCOMI;
6822 case Intrinsic::x86_sse_ucomineq_ss:
6823 case Intrinsic::x86_sse2_ucomineq_sd:
6824 Opc = X86ISD::UCOMI;
6829 SDValue LHS = Op.getOperand(1);
6830 SDValue RHS = Op.getOperand(2);
6831 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6832 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6833 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6834 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6835 DAG.getConstant(X86CC, MVT::i8), Cond);
6836 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6838 // ptest intrinsics. The intrinsic these come from are designed to return
6839 // an integer value, not just an instruction so lower it to the ptest
6840 // pattern and a setcc for the result.
6841 case Intrinsic::x86_sse41_ptestz:
6842 case Intrinsic::x86_sse41_ptestc:
6843 case Intrinsic::x86_sse41_ptestnzc:{
6846 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6847 case Intrinsic::x86_sse41_ptestz:
6849 X86CC = X86::COND_E;
6851 case Intrinsic::x86_sse41_ptestc:
6853 X86CC = X86::COND_B;
6855 case Intrinsic::x86_sse41_ptestnzc:
6857 X86CC = X86::COND_A;
6861 SDValue LHS = Op.getOperand(1);
6862 SDValue RHS = Op.getOperand(2);
6863 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6864 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6865 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6866 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6869 // Fix vector shift instructions where the last operand is a non-immediate
6871 case Intrinsic::x86_sse2_pslli_w:
6872 case Intrinsic::x86_sse2_pslli_d:
6873 case Intrinsic::x86_sse2_pslli_q:
6874 case Intrinsic::x86_sse2_psrli_w:
6875 case Intrinsic::x86_sse2_psrli_d:
6876 case Intrinsic::x86_sse2_psrli_q:
6877 case Intrinsic::x86_sse2_psrai_w:
6878 case Intrinsic::x86_sse2_psrai_d:
6879 case Intrinsic::x86_mmx_pslli_w:
6880 case Intrinsic::x86_mmx_pslli_d:
6881 case Intrinsic::x86_mmx_pslli_q:
6882 case Intrinsic::x86_mmx_psrli_w:
6883 case Intrinsic::x86_mmx_psrli_d:
6884 case Intrinsic::x86_mmx_psrli_q:
6885 case Intrinsic::x86_mmx_psrai_w:
6886 case Intrinsic::x86_mmx_psrai_d: {
6887 SDValue ShAmt = Op.getOperand(2);
6888 if (isa<ConstantSDNode>(ShAmt))
6891 unsigned NewIntNo = 0;
6892 EVT ShAmtVT = MVT::v4i32;
6894 case Intrinsic::x86_sse2_pslli_w:
6895 NewIntNo = Intrinsic::x86_sse2_psll_w;
6897 case Intrinsic::x86_sse2_pslli_d:
6898 NewIntNo = Intrinsic::x86_sse2_psll_d;
6900 case Intrinsic::x86_sse2_pslli_q:
6901 NewIntNo = Intrinsic::x86_sse2_psll_q;
6903 case Intrinsic::x86_sse2_psrli_w:
6904 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6906 case Intrinsic::x86_sse2_psrli_d:
6907 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6909 case Intrinsic::x86_sse2_psrli_q:
6910 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6912 case Intrinsic::x86_sse2_psrai_w:
6913 NewIntNo = Intrinsic::x86_sse2_psra_w;
6915 case Intrinsic::x86_sse2_psrai_d:
6916 NewIntNo = Intrinsic::x86_sse2_psra_d;
6919 ShAmtVT = MVT::v2i32;
6921 case Intrinsic::x86_mmx_pslli_w:
6922 NewIntNo = Intrinsic::x86_mmx_psll_w;
6924 case Intrinsic::x86_mmx_pslli_d:
6925 NewIntNo = Intrinsic::x86_mmx_psll_d;
6927 case Intrinsic::x86_mmx_pslli_q:
6928 NewIntNo = Intrinsic::x86_mmx_psll_q;
6930 case Intrinsic::x86_mmx_psrli_w:
6931 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6933 case Intrinsic::x86_mmx_psrli_d:
6934 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6936 case Intrinsic::x86_mmx_psrli_q:
6937 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6939 case Intrinsic::x86_mmx_psrai_w:
6940 NewIntNo = Intrinsic::x86_mmx_psra_w;
6942 case Intrinsic::x86_mmx_psrai_d:
6943 NewIntNo = Intrinsic::x86_mmx_psra_d;
6945 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6951 // The vector shift intrinsics with scalars uses 32b shift amounts but
6952 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6956 ShOps[1] = DAG.getConstant(0, MVT::i32);
6957 if (ShAmtVT == MVT::v4i32) {
6958 ShOps[2] = DAG.getUNDEF(MVT::i32);
6959 ShOps[3] = DAG.getUNDEF(MVT::i32);
6960 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6962 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6965 EVT VT = Op.getValueType();
6966 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6967 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6968 DAG.getConstant(NewIntNo, MVT::i32),
6969 Op.getOperand(1), ShAmt);
6974 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6975 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6976 DebugLoc dl = Op.getDebugLoc();
6979 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6981 DAG.getConstant(TD->getPointerSize(),
6982 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6983 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6984 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6986 NULL, 0, false, false, 0);
6989 // Just load the return address.
6990 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6991 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6992 RetAddrFI, NULL, 0, false, false, 0);
6995 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6996 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6997 MFI->setFrameAddressIsTaken(true);
6998 EVT VT = Op.getValueType();
6999 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7000 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7001 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7002 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7004 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7009 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7010 SelectionDAG &DAG) {
7011 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7014 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
7016 MachineFunction &MF = DAG.getMachineFunction();
7017 SDValue Chain = Op.getOperand(0);
7018 SDValue Offset = Op.getOperand(1);
7019 SDValue Handler = Op.getOperand(2);
7020 DebugLoc dl = Op.getDebugLoc();
7022 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7024 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7026 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7027 DAG.getIntPtrConstant(-TD->getPointerSize()));
7028 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7029 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7030 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7031 MF.getRegInfo().addLiveOut(StoreAddrReg);
7033 return DAG.getNode(X86ISD::EH_RETURN, dl,
7035 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7038 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7039 SelectionDAG &DAG) {
7040 SDValue Root = Op.getOperand(0);
7041 SDValue Trmp = Op.getOperand(1); // trampoline
7042 SDValue FPtr = Op.getOperand(2); // nested function
7043 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7044 DebugLoc dl = Op.getDebugLoc();
7046 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7048 if (Subtarget->is64Bit()) {
7049 SDValue OutChains[6];
7051 // Large code-model.
7052 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7053 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7055 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7056 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7058 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7060 // Load the pointer to the nested function into R11.
7061 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7062 SDValue Addr = Trmp;
7063 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7064 Addr, TrmpAddr, 0, false, false, 0);
7066 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7067 DAG.getConstant(2, MVT::i64));
7068 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7071 // Load the 'nest' parameter value into R10.
7072 // R10 is specified in X86CallingConv.td
7073 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7074 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7075 DAG.getConstant(10, MVT::i64));
7076 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7077 Addr, TrmpAddr, 10, false, false, 0);
7079 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7080 DAG.getConstant(12, MVT::i64));
7081 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7084 // Jump to the nested function.
7085 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7086 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7087 DAG.getConstant(20, MVT::i64));
7088 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7089 Addr, TrmpAddr, 20, false, false, 0);
7091 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7092 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7093 DAG.getConstant(22, MVT::i64));
7094 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7095 TrmpAddr, 22, false, false, 0);
7098 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7099 return DAG.getMergeValues(Ops, 2, dl);
7101 const Function *Func =
7102 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7103 CallingConv::ID CC = Func->getCallingConv();
7108 llvm_unreachable("Unsupported calling convention");
7109 case CallingConv::C:
7110 case CallingConv::X86_StdCall: {
7111 // Pass 'nest' parameter in ECX.
7112 // Must be kept in sync with X86CallingConv.td
7115 // Check that ECX wasn't needed by an 'inreg' parameter.
7116 const FunctionType *FTy = Func->getFunctionType();
7117 const AttrListPtr &Attrs = Func->getAttributes();
7119 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7120 unsigned InRegCount = 0;
7123 for (FunctionType::param_iterator I = FTy->param_begin(),
7124 E = FTy->param_end(); I != E; ++I, ++Idx)
7125 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7126 // FIXME: should only count parameters that are lowered to integers.
7127 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7129 if (InRegCount > 2) {
7130 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7135 case CallingConv::X86_FastCall:
7136 case CallingConv::Fast:
7137 // Pass 'nest' parameter in EAX.
7138 // Must be kept in sync with X86CallingConv.td
7143 SDValue OutChains[4];
7146 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7147 DAG.getConstant(10, MVT::i32));
7148 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7150 // This is storing the opcode for MOV32ri.
7151 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7152 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7153 OutChains[0] = DAG.getStore(Root, dl,
7154 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7155 Trmp, TrmpAddr, 0, false, false, 0);
7157 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7158 DAG.getConstant(1, MVT::i32));
7159 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7162 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7163 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7164 DAG.getConstant(5, MVT::i32));
7165 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7166 TrmpAddr, 5, false, false, 1);
7168 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7169 DAG.getConstant(6, MVT::i32));
7170 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7174 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7175 return DAG.getMergeValues(Ops, 2, dl);
7179 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7181 The rounding mode is in bits 11:10 of FPSR, and has the following
7188 FLT_ROUNDS, on the other hand, expects the following:
7195 To perform the conversion, we do:
7196 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7199 MachineFunction &MF = DAG.getMachineFunction();
7200 const TargetMachine &TM = MF.getTarget();
7201 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7202 unsigned StackAlignment = TFI.getStackAlignment();
7203 EVT VT = Op.getValueType();
7204 DebugLoc dl = Op.getDebugLoc();
7206 // Save FP Control Word to stack slot
7207 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7208 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7210 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7211 DAG.getEntryNode(), StackSlot);
7213 // Load FP Control Word from stack slot
7214 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7217 // Transform as necessary
7219 DAG.getNode(ISD::SRL, dl, MVT::i16,
7220 DAG.getNode(ISD::AND, dl, MVT::i16,
7221 CWD, DAG.getConstant(0x800, MVT::i16)),
7222 DAG.getConstant(11, MVT::i8));
7224 DAG.getNode(ISD::SRL, dl, MVT::i16,
7225 DAG.getNode(ISD::AND, dl, MVT::i16,
7226 CWD, DAG.getConstant(0x400, MVT::i16)),
7227 DAG.getConstant(9, MVT::i8));
7230 DAG.getNode(ISD::AND, dl, MVT::i16,
7231 DAG.getNode(ISD::ADD, dl, MVT::i16,
7232 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7233 DAG.getConstant(1, MVT::i16)),
7234 DAG.getConstant(3, MVT::i16));
7237 return DAG.getNode((VT.getSizeInBits() < 16 ?
7238 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7241 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7242 EVT VT = Op.getValueType();
7244 unsigned NumBits = VT.getSizeInBits();
7245 DebugLoc dl = Op.getDebugLoc();
7247 Op = Op.getOperand(0);
7248 if (VT == MVT::i8) {
7249 // Zero extend to i32 since there is not an i8 bsr.
7251 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7254 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7255 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7256 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7258 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7261 DAG.getConstant(NumBits+NumBits-1, OpVT),
7262 DAG.getConstant(X86::COND_E, MVT::i8),
7265 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7267 // Finally xor with NumBits-1.
7268 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7271 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7275 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7276 EVT VT = Op.getValueType();
7278 unsigned NumBits = VT.getSizeInBits();
7279 DebugLoc dl = Op.getDebugLoc();
7281 Op = Op.getOperand(0);
7282 if (VT == MVT::i8) {
7284 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7287 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7288 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7289 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7291 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7294 DAG.getConstant(NumBits, OpVT),
7295 DAG.getConstant(X86::COND_E, MVT::i8),
7298 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7301 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7305 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7306 EVT VT = Op.getValueType();
7307 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7308 DebugLoc dl = Op.getDebugLoc();
7310 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7311 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7312 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7313 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7314 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7316 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7317 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7318 // return AloBlo + AloBhi + AhiBlo;
7320 SDValue A = Op.getOperand(0);
7321 SDValue B = Op.getOperand(1);
7323 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7324 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7325 A, DAG.getConstant(32, MVT::i32));
7326 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7327 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7328 B, DAG.getConstant(32, MVT::i32));
7329 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7330 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7332 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7333 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7335 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7336 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7338 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7339 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7340 AloBhi, DAG.getConstant(32, MVT::i32));
7341 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7342 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7343 AhiBlo, DAG.getConstant(32, MVT::i32));
7344 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7345 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7350 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7351 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7352 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7353 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7354 // has only one use.
7355 SDNode *N = Op.getNode();
7356 SDValue LHS = N->getOperand(0);
7357 SDValue RHS = N->getOperand(1);
7358 unsigned BaseOp = 0;
7360 DebugLoc dl = Op.getDebugLoc();
7362 switch (Op.getOpcode()) {
7363 default: llvm_unreachable("Unknown ovf instruction!");
7365 // A subtract of one will be selected as a INC. Note that INC doesn't
7366 // set CF, so we can't do this for UADDO.
7367 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7368 if (C->getAPIntValue() == 1) {
7369 BaseOp = X86ISD::INC;
7373 BaseOp = X86ISD::ADD;
7377 BaseOp = X86ISD::ADD;
7381 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7382 // set CF, so we can't do this for USUBO.
7383 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7384 if (C->getAPIntValue() == 1) {
7385 BaseOp = X86ISD::DEC;
7389 BaseOp = X86ISD::SUB;
7393 BaseOp = X86ISD::SUB;
7397 BaseOp = X86ISD::SMUL;
7401 BaseOp = X86ISD::UMUL;
7406 // Also sets EFLAGS.
7407 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7408 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7411 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7412 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7414 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7418 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7419 EVT T = Op.getValueType();
7420 DebugLoc dl = Op.getDebugLoc();
7423 switch(T.getSimpleVT().SimpleTy) {
7425 assert(false && "Invalid value type!");
7426 case MVT::i8: Reg = X86::AL; size = 1; break;
7427 case MVT::i16: Reg = X86::AX; size = 2; break;
7428 case MVT::i32: Reg = X86::EAX; size = 4; break;
7430 assert(Subtarget->is64Bit() && "Node not type legal!");
7431 Reg = X86::RAX; size = 8;
7434 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7435 Op.getOperand(2), SDValue());
7436 SDValue Ops[] = { cpIn.getValue(0),
7439 DAG.getTargetConstant(size, MVT::i8),
7441 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7442 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7444 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7448 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7449 SelectionDAG &DAG) {
7450 assert(Subtarget->is64Bit() && "Result not type legalized?");
7451 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7452 SDValue TheChain = Op.getOperand(0);
7453 DebugLoc dl = Op.getDebugLoc();
7454 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7455 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7456 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7458 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7459 DAG.getConstant(32, MVT::i8));
7461 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7464 return DAG.getMergeValues(Ops, 2, dl);
7467 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7468 SDNode *Node = Op.getNode();
7469 DebugLoc dl = Node->getDebugLoc();
7470 EVT T = Node->getValueType(0);
7471 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7472 DAG.getConstant(0, T), Node->getOperand(2));
7473 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7474 cast<AtomicSDNode>(Node)->getMemoryVT(),
7475 Node->getOperand(0),
7476 Node->getOperand(1), negOp,
7477 cast<AtomicSDNode>(Node)->getSrcValue(),
7478 cast<AtomicSDNode>(Node)->getAlignment());
7481 /// LowerOperation - Provide custom lowering hooks for some operations.
7483 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7484 switch (Op.getOpcode()) {
7485 default: llvm_unreachable("Should not custom lower this!");
7486 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7487 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7488 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7489 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7490 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7491 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7492 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7493 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7494 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7495 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7496 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7497 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7498 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7499 case ISD::SHL_PARTS:
7500 case ISD::SRA_PARTS:
7501 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7502 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7503 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7504 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7505 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7506 case ISD::FABS: return LowerFABS(Op, DAG);
7507 case ISD::FNEG: return LowerFNEG(Op, DAG);
7508 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7509 case ISD::SETCC: return LowerSETCC(Op, DAG);
7510 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7511 case ISD::SELECT: return LowerSELECT(Op, DAG);
7512 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7513 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7514 case ISD::VASTART: return LowerVASTART(Op, DAG);
7515 case ISD::VAARG: return LowerVAARG(Op, DAG);
7516 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7517 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7518 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7519 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7520 case ISD::FRAME_TO_ARGS_OFFSET:
7521 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7522 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7523 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7524 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7525 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7526 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7527 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7528 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7534 case ISD::UMULO: return LowerXALUO(Op, DAG);
7535 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7539 void X86TargetLowering::
7540 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7541 SelectionDAG &DAG, unsigned NewOp) {
7542 EVT T = Node->getValueType(0);
7543 DebugLoc dl = Node->getDebugLoc();
7544 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7546 SDValue Chain = Node->getOperand(0);
7547 SDValue In1 = Node->getOperand(1);
7548 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7549 Node->getOperand(2), DAG.getIntPtrConstant(0));
7550 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7551 Node->getOperand(2), DAG.getIntPtrConstant(1));
7552 SDValue Ops[] = { Chain, In1, In2L, In2H };
7553 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7555 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7556 cast<MemSDNode>(Node)->getMemOperand());
7557 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7558 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7559 Results.push_back(Result.getValue(2));
7562 /// ReplaceNodeResults - Replace a node with an illegal result type
7563 /// with a new node built out of custom code.
7564 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7565 SmallVectorImpl<SDValue>&Results,
7566 SelectionDAG &DAG) {
7567 DebugLoc dl = N->getDebugLoc();
7568 switch (N->getOpcode()) {
7570 assert(false && "Do not know how to custom type legalize this operation!");
7572 case ISD::FP_TO_SINT: {
7573 std::pair<SDValue,SDValue> Vals =
7574 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7575 SDValue FIST = Vals.first, StackSlot = Vals.second;
7576 if (FIST.getNode() != 0) {
7577 EVT VT = N->getValueType(0);
7578 // Return a load from the stack slot.
7579 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7584 case ISD::READCYCLECOUNTER: {
7585 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7586 SDValue TheChain = N->getOperand(0);
7587 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7588 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7590 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7592 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7593 SDValue Ops[] = { eax, edx };
7594 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7595 Results.push_back(edx.getValue(1));
7598 case ISD::ATOMIC_CMP_SWAP: {
7599 EVT T = N->getValueType(0);
7600 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7601 SDValue cpInL, cpInH;
7602 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7603 DAG.getConstant(0, MVT::i32));
7604 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7605 DAG.getConstant(1, MVT::i32));
7606 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7607 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7609 SDValue swapInL, swapInH;
7610 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7611 DAG.getConstant(0, MVT::i32));
7612 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7613 DAG.getConstant(1, MVT::i32));
7614 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7616 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7617 swapInL.getValue(1));
7618 SDValue Ops[] = { swapInH.getValue(0),
7620 swapInH.getValue(1) };
7621 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7622 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7623 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7624 MVT::i32, Result.getValue(1));
7625 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7626 MVT::i32, cpOutL.getValue(2));
7627 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7628 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7629 Results.push_back(cpOutH.getValue(1));
7632 case ISD::ATOMIC_LOAD_ADD:
7633 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7635 case ISD::ATOMIC_LOAD_AND:
7636 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7638 case ISD::ATOMIC_LOAD_NAND:
7639 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7641 case ISD::ATOMIC_LOAD_OR:
7642 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7644 case ISD::ATOMIC_LOAD_SUB:
7645 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7647 case ISD::ATOMIC_LOAD_XOR:
7648 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7650 case ISD::ATOMIC_SWAP:
7651 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7656 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7658 default: return NULL;
7659 case X86ISD::BSF: return "X86ISD::BSF";
7660 case X86ISD::BSR: return "X86ISD::BSR";
7661 case X86ISD::SHLD: return "X86ISD::SHLD";
7662 case X86ISD::SHRD: return "X86ISD::SHRD";
7663 case X86ISD::FAND: return "X86ISD::FAND";
7664 case X86ISD::FOR: return "X86ISD::FOR";
7665 case X86ISD::FXOR: return "X86ISD::FXOR";
7666 case X86ISD::FSRL: return "X86ISD::FSRL";
7667 case X86ISD::FILD: return "X86ISD::FILD";
7668 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7669 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7670 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7671 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7672 case X86ISD::FLD: return "X86ISD::FLD";
7673 case X86ISD::FST: return "X86ISD::FST";
7674 case X86ISD::CALL: return "X86ISD::CALL";
7675 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7676 case X86ISD::BT: return "X86ISD::BT";
7677 case X86ISD::CMP: return "X86ISD::CMP";
7678 case X86ISD::COMI: return "X86ISD::COMI";
7679 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7680 case X86ISD::SETCC: return "X86ISD::SETCC";
7681 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7682 case X86ISD::CMOV: return "X86ISD::CMOV";
7683 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7684 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7685 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7686 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7687 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7688 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7689 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7690 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7691 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7692 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7693 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7694 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7695 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7696 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7697 case X86ISD::FMAX: return "X86ISD::FMAX";
7698 case X86ISD::FMIN: return "X86ISD::FMIN";
7699 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7700 case X86ISD::FRCP: return "X86ISD::FRCP";
7701 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7702 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7703 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7704 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7705 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7706 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7707 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7708 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7709 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7710 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7711 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7712 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7713 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7714 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7715 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7716 case X86ISD::VSHL: return "X86ISD::VSHL";
7717 case X86ISD::VSRL: return "X86ISD::VSRL";
7718 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7719 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7720 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7721 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7722 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7723 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7724 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7725 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7726 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7727 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7728 case X86ISD::ADD: return "X86ISD::ADD";
7729 case X86ISD::SUB: return "X86ISD::SUB";
7730 case X86ISD::SMUL: return "X86ISD::SMUL";
7731 case X86ISD::UMUL: return "X86ISD::UMUL";
7732 case X86ISD::INC: return "X86ISD::INC";
7733 case X86ISD::DEC: return "X86ISD::DEC";
7734 case X86ISD::OR: return "X86ISD::OR";
7735 case X86ISD::XOR: return "X86ISD::XOR";
7736 case X86ISD::AND: return "X86ISD::AND";
7737 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7738 case X86ISD::PTEST: return "X86ISD::PTEST";
7739 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7740 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7744 // isLegalAddressingMode - Return true if the addressing mode represented
7745 // by AM is legal for this target, for a load/store of the specified type.
7746 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7747 const Type *Ty) const {
7748 // X86 supports extremely general addressing modes.
7749 CodeModel::Model M = getTargetMachine().getCodeModel();
7751 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7752 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7757 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7759 // If a reference to this global requires an extra load, we can't fold it.
7760 if (isGlobalStubReference(GVFlags))
7763 // If BaseGV requires a register for the PIC base, we cannot also have a
7764 // BaseReg specified.
7765 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7768 // If lower 4G is not available, then we must use rip-relative addressing.
7769 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7779 // These scales always work.
7784 // These scales are formed with basereg+scalereg. Only accept if there is
7789 default: // Other stuff never works.
7797 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7798 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7800 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7801 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7802 if (NumBits1 <= NumBits2)
7807 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7808 if (!VT1.isInteger() || !VT2.isInteger())
7810 unsigned NumBits1 = VT1.getSizeInBits();
7811 unsigned NumBits2 = VT2.getSizeInBits();
7812 if (NumBits1 <= NumBits2)
7817 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7818 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7819 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7822 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7823 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7824 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7827 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7828 // i16 instructions are longer (0x66 prefix) and potentially slower.
7829 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7832 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7833 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7834 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7835 /// are assumed to be legal.
7837 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7839 // Only do shuffles on 128-bit vector types for now.
7840 if (VT.getSizeInBits() == 64)
7843 // FIXME: pshufb, blends, shifts.
7844 return (VT.getVectorNumElements() == 2 ||
7845 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7846 isMOVLMask(M, VT) ||
7847 isSHUFPMask(M, VT) ||
7848 isPSHUFDMask(M, VT) ||
7849 isPSHUFHWMask(M, VT) ||
7850 isPSHUFLWMask(M, VT) ||
7851 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7852 isUNPCKLMask(M, VT) ||
7853 isUNPCKHMask(M, VT) ||
7854 isUNPCKL_v_undef_Mask(M, VT) ||
7855 isUNPCKH_v_undef_Mask(M, VT));
7859 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7861 unsigned NumElts = VT.getVectorNumElements();
7862 // FIXME: This collection of masks seems suspect.
7865 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7866 return (isMOVLMask(Mask, VT) ||
7867 isCommutedMOVLMask(Mask, VT, true) ||
7868 isSHUFPMask(Mask, VT) ||
7869 isCommutedSHUFPMask(Mask, VT));
7874 //===----------------------------------------------------------------------===//
7875 // X86 Scheduler Hooks
7876 //===----------------------------------------------------------------------===//
7878 // private utility function
7880 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7881 MachineBasicBlock *MBB,
7889 TargetRegisterClass *RC,
7890 bool invSrc) const {
7891 // For the atomic bitwise operator, we generate
7894 // ld t1 = [bitinstr.addr]
7895 // op t2 = t1, [bitinstr.val]
7897 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7899 // fallthrough -->nextMBB
7900 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7901 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7902 MachineFunction::iterator MBBIter = MBB;
7905 /// First build the CFG
7906 MachineFunction *F = MBB->getParent();
7907 MachineBasicBlock *thisMBB = MBB;
7908 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7909 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7910 F->insert(MBBIter, newMBB);
7911 F->insert(MBBIter, nextMBB);
7913 // Move all successors to thisMBB to nextMBB
7914 nextMBB->transferSuccessors(thisMBB);
7916 // Update thisMBB to fall through to newMBB
7917 thisMBB->addSuccessor(newMBB);
7919 // newMBB jumps to itself and fall through to nextMBB
7920 newMBB->addSuccessor(nextMBB);
7921 newMBB->addSuccessor(newMBB);
7923 // Insert instructions into newMBB based on incoming instruction
7924 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7925 "unexpected number of operands");
7926 DebugLoc dl = bInstr->getDebugLoc();
7927 MachineOperand& destOper = bInstr->getOperand(0);
7928 MachineOperand* argOpers[2 + X86AddrNumOperands];
7929 int numArgs = bInstr->getNumOperands() - 1;
7930 for (int i=0; i < numArgs; ++i)
7931 argOpers[i] = &bInstr->getOperand(i+1);
7933 // x86 address has 4 operands: base, index, scale, and displacement
7934 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7935 int valArgIndx = lastAddrIndx + 1;
7937 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7938 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7939 for (int i=0; i <= lastAddrIndx; ++i)
7940 (*MIB).addOperand(*argOpers[i]);
7942 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7944 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7949 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7950 assert((argOpers[valArgIndx]->isReg() ||
7951 argOpers[valArgIndx]->isImm()) &&
7953 if (argOpers[valArgIndx]->isReg())
7954 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7956 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7958 (*MIB).addOperand(*argOpers[valArgIndx]);
7960 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7963 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7964 for (int i=0; i <= lastAddrIndx; ++i)
7965 (*MIB).addOperand(*argOpers[i]);
7967 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7968 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7969 bInstr->memoperands_end());
7971 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7975 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
7977 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7981 // private utility function: 64 bit atomics on 32 bit host.
7983 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7984 MachineBasicBlock *MBB,
7989 bool invSrc) const {
7990 // For the atomic bitwise operator, we generate
7991 // thisMBB (instructions are in pairs, except cmpxchg8b)
7992 // ld t1,t2 = [bitinstr.addr]
7994 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7995 // op t5, t6 <- out1, out2, [bitinstr.val]
7996 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7997 // mov ECX, EBX <- t5, t6
7998 // mov EAX, EDX <- t1, t2
7999 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8000 // mov t3, t4 <- EAX, EDX
8002 // result in out1, out2
8003 // fallthrough -->nextMBB
8005 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8006 const unsigned LoadOpc = X86::MOV32rm;
8007 const unsigned copyOpc = X86::MOV32rr;
8008 const unsigned NotOpc = X86::NOT32r;
8009 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8010 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8011 MachineFunction::iterator MBBIter = MBB;
8014 /// First build the CFG
8015 MachineFunction *F = MBB->getParent();
8016 MachineBasicBlock *thisMBB = MBB;
8017 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8018 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8019 F->insert(MBBIter, newMBB);
8020 F->insert(MBBIter, nextMBB);
8022 // Move all successors to thisMBB to nextMBB
8023 nextMBB->transferSuccessors(thisMBB);
8025 // Update thisMBB to fall through to newMBB
8026 thisMBB->addSuccessor(newMBB);
8028 // newMBB jumps to itself and fall through to nextMBB
8029 newMBB->addSuccessor(nextMBB);
8030 newMBB->addSuccessor(newMBB);
8032 DebugLoc dl = bInstr->getDebugLoc();
8033 // Insert instructions into newMBB based on incoming instruction
8034 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8035 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8036 "unexpected number of operands");
8037 MachineOperand& dest1Oper = bInstr->getOperand(0);
8038 MachineOperand& dest2Oper = bInstr->getOperand(1);
8039 MachineOperand* argOpers[2 + X86AddrNumOperands];
8040 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8041 argOpers[i] = &bInstr->getOperand(i+2);
8043 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8044 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8046 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8047 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8048 for (int i=0; i <= lastAddrIndx; ++i)
8049 (*MIB).addOperand(*argOpers[i]);
8050 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8051 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8052 // add 4 to displacement.
8053 for (int i=0; i <= lastAddrIndx-2; ++i)
8054 (*MIB).addOperand(*argOpers[i]);
8055 MachineOperand newOp3 = *(argOpers[3]);
8057 newOp3.setImm(newOp3.getImm()+4);
8059 newOp3.setOffset(newOp3.getOffset()+4);
8060 (*MIB).addOperand(newOp3);
8061 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8063 // t3/4 are defined later, at the bottom of the loop
8064 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8065 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8066 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8067 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8068 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8069 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8071 // The subsequent operations should be using the destination registers of
8072 //the PHI instructions.
8074 t1 = F->getRegInfo().createVirtualRegister(RC);
8075 t2 = F->getRegInfo().createVirtualRegister(RC);
8076 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8077 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8079 t1 = dest1Oper.getReg();
8080 t2 = dest2Oper.getReg();
8083 int valArgIndx = lastAddrIndx + 1;
8084 assert((argOpers[valArgIndx]->isReg() ||
8085 argOpers[valArgIndx]->isImm()) &&
8087 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8088 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8089 if (argOpers[valArgIndx]->isReg())
8090 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8092 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8093 if (regOpcL != X86::MOV32rr)
8095 (*MIB).addOperand(*argOpers[valArgIndx]);
8096 assert(argOpers[valArgIndx + 1]->isReg() ==
8097 argOpers[valArgIndx]->isReg());
8098 assert(argOpers[valArgIndx + 1]->isImm() ==
8099 argOpers[valArgIndx]->isImm());
8100 if (argOpers[valArgIndx + 1]->isReg())
8101 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8103 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8104 if (regOpcH != X86::MOV32rr)
8106 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8108 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8110 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8113 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8115 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8118 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8119 for (int i=0; i <= lastAddrIndx; ++i)
8120 (*MIB).addOperand(*argOpers[i]);
8122 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8123 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8124 bInstr->memoperands_end());
8126 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8127 MIB.addReg(X86::EAX);
8128 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8129 MIB.addReg(X86::EDX);
8132 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8134 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8138 // private utility function
8140 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8141 MachineBasicBlock *MBB,
8142 unsigned cmovOpc) const {
8143 // For the atomic min/max operator, we generate
8146 // ld t1 = [min/max.addr]
8147 // mov t2 = [min/max.val]
8149 // cmov[cond] t2 = t1
8151 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8153 // fallthrough -->nextMBB
8155 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8156 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8157 MachineFunction::iterator MBBIter = MBB;
8160 /// First build the CFG
8161 MachineFunction *F = MBB->getParent();
8162 MachineBasicBlock *thisMBB = MBB;
8163 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8164 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8165 F->insert(MBBIter, newMBB);
8166 F->insert(MBBIter, nextMBB);
8168 // Move all successors of thisMBB to nextMBB
8169 nextMBB->transferSuccessors(thisMBB);
8171 // Update thisMBB to fall through to newMBB
8172 thisMBB->addSuccessor(newMBB);
8174 // newMBB jumps to newMBB and fall through to nextMBB
8175 newMBB->addSuccessor(nextMBB);
8176 newMBB->addSuccessor(newMBB);
8178 DebugLoc dl = mInstr->getDebugLoc();
8179 // Insert instructions into newMBB based on incoming instruction
8180 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8181 "unexpected number of operands");
8182 MachineOperand& destOper = mInstr->getOperand(0);
8183 MachineOperand* argOpers[2 + X86AddrNumOperands];
8184 int numArgs = mInstr->getNumOperands() - 1;
8185 for (int i=0; i < numArgs; ++i)
8186 argOpers[i] = &mInstr->getOperand(i+1);
8188 // x86 address has 4 operands: base, index, scale, and displacement
8189 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8190 int valArgIndx = lastAddrIndx + 1;
8192 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8193 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8194 for (int i=0; i <= lastAddrIndx; ++i)
8195 (*MIB).addOperand(*argOpers[i]);
8197 // We only support register and immediate values
8198 assert((argOpers[valArgIndx]->isReg() ||
8199 argOpers[valArgIndx]->isImm()) &&
8202 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8203 if (argOpers[valArgIndx]->isReg())
8204 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8206 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8207 (*MIB).addOperand(*argOpers[valArgIndx]);
8209 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8212 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8217 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8218 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8222 // Cmp and exchange if none has modified the memory location
8223 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8224 for (int i=0; i <= lastAddrIndx; ++i)
8225 (*MIB).addOperand(*argOpers[i]);
8227 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8228 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8229 mInstr->memoperands_end());
8231 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8232 MIB.addReg(X86::EAX);
8235 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8237 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8241 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8242 // all of this code can be replaced with that in the .td file.
8244 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8245 unsigned numArgs, bool memArg) const {
8247 MachineFunction *F = BB->getParent();
8248 DebugLoc dl = MI->getDebugLoc();
8249 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8253 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8255 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8257 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8259 for (unsigned i = 0; i < numArgs; ++i) {
8260 MachineOperand &Op = MI->getOperand(i+1);
8262 if (!(Op.isReg() && Op.isImplicit()))
8266 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8269 F->DeleteMachineInstr(MI);
8275 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8277 MachineBasicBlock *MBB) const {
8278 // Emit code to save XMM registers to the stack. The ABI says that the
8279 // number of registers to save is given in %al, so it's theoretically
8280 // possible to do an indirect jump trick to avoid saving all of them,
8281 // however this code takes a simpler approach and just executes all
8282 // of the stores if %al is non-zero. It's less code, and it's probably
8283 // easier on the hardware branch predictor, and stores aren't all that
8284 // expensive anyway.
8286 // Create the new basic blocks. One block contains all the XMM stores,
8287 // and one block is the final destination regardless of whether any
8288 // stores were performed.
8289 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8290 MachineFunction *F = MBB->getParent();
8291 MachineFunction::iterator MBBIter = MBB;
8293 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8294 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8295 F->insert(MBBIter, XMMSaveMBB);
8296 F->insert(MBBIter, EndMBB);
8299 // Move any original successors of MBB to the end block.
8300 EndMBB->transferSuccessors(MBB);
8301 // The original block will now fall through to the XMM save block.
8302 MBB->addSuccessor(XMMSaveMBB);
8303 // The XMMSaveMBB will fall through to the end block.
8304 XMMSaveMBB->addSuccessor(EndMBB);
8306 // Now add the instructions.
8307 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8308 DebugLoc DL = MI->getDebugLoc();
8310 unsigned CountReg = MI->getOperand(0).getReg();
8311 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8312 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8314 if (!Subtarget->isTargetWin64()) {
8315 // If %al is 0, branch around the XMM save block.
8316 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8317 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8318 MBB->addSuccessor(EndMBB);
8321 // In the XMM save block, save all the XMM argument registers.
8322 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8323 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8324 MachineMemOperand *MMO =
8325 F->getMachineMemOperand(
8326 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8327 MachineMemOperand::MOStore, Offset,
8328 /*Size=*/16, /*Align=*/16);
8329 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8330 .addFrameIndex(RegSaveFrameIndex)
8331 .addImm(/*Scale=*/1)
8332 .addReg(/*IndexReg=*/0)
8333 .addImm(/*Disp=*/Offset)
8334 .addReg(/*Segment=*/0)
8335 .addReg(MI->getOperand(i).getReg())
8336 .addMemOperand(MMO);
8339 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8345 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8346 MachineBasicBlock *BB,
8347 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8348 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8349 DebugLoc DL = MI->getDebugLoc();
8351 // To "insert" a SELECT_CC instruction, we actually have to insert the
8352 // diamond control-flow pattern. The incoming instruction knows the
8353 // destination vreg to set, the condition code register to branch on, the
8354 // true/false values to select between, and a branch opcode to use.
8355 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8356 MachineFunction::iterator It = BB;
8362 // cmpTY ccX, r1, r2
8364 // fallthrough --> copy0MBB
8365 MachineBasicBlock *thisMBB = BB;
8366 MachineFunction *F = BB->getParent();
8367 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8368 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8370 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8371 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8372 F->insert(It, copy0MBB);
8373 F->insert(It, sinkMBB);
8374 // Update machine-CFG edges by first adding all successors of the current
8375 // block to the new block which will contain the Phi node for the select.
8376 // Also inform sdisel of the edge changes.
8377 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8378 E = BB->succ_end(); I != E; ++I) {
8379 EM->insert(std::make_pair(*I, sinkMBB));
8380 sinkMBB->addSuccessor(*I);
8382 // Next, remove all successors of the current block, and add the true
8383 // and fallthrough blocks as its successors.
8384 while (!BB->succ_empty())
8385 BB->removeSuccessor(BB->succ_begin());
8386 // Add the true and fallthrough blocks as its successors.
8387 BB->addSuccessor(copy0MBB);
8388 BB->addSuccessor(sinkMBB);
8391 // %FalseValue = ...
8392 // # fallthrough to sinkMBB
8395 // Update machine-CFG edges
8396 BB->addSuccessor(sinkMBB);
8399 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8402 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8403 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8404 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8406 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8411 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8412 MachineBasicBlock *BB,
8413 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8414 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8415 DebugLoc DL = MI->getDebugLoc();
8416 MachineFunction *F = BB->getParent();
8418 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8419 // non-trivial part is impdef of ESP.
8420 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8423 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8424 .addExternalSymbol("_alloca")
8425 .addReg(X86::EAX, RegState::Implicit)
8426 .addReg(X86::ESP, RegState::Implicit)
8427 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8428 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8430 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8435 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8436 MachineBasicBlock *BB,
8437 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8438 switch (MI->getOpcode()) {
8439 default: assert(false && "Unexpected instr type to insert");
8440 case X86::MINGW_ALLOCA:
8441 return EmitLoweredMingwAlloca(MI, BB, EM);
8443 case X86::CMOV_V1I64:
8444 case X86::CMOV_FR32:
8445 case X86::CMOV_FR64:
8446 case X86::CMOV_V4F32:
8447 case X86::CMOV_V2F64:
8448 case X86::CMOV_V2I64:
8449 case X86::CMOV_GR16:
8450 case X86::CMOV_GR32:
8451 case X86::CMOV_RFP32:
8452 case X86::CMOV_RFP64:
8453 case X86::CMOV_RFP80:
8454 return EmitLoweredSelect(MI, BB, EM);
8456 case X86::FP32_TO_INT16_IN_MEM:
8457 case X86::FP32_TO_INT32_IN_MEM:
8458 case X86::FP32_TO_INT64_IN_MEM:
8459 case X86::FP64_TO_INT16_IN_MEM:
8460 case X86::FP64_TO_INT32_IN_MEM:
8461 case X86::FP64_TO_INT64_IN_MEM:
8462 case X86::FP80_TO_INT16_IN_MEM:
8463 case X86::FP80_TO_INT32_IN_MEM:
8464 case X86::FP80_TO_INT64_IN_MEM: {
8465 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8466 DebugLoc DL = MI->getDebugLoc();
8468 // Change the floating point control register to use "round towards zero"
8469 // mode when truncating to an integer value.
8470 MachineFunction *F = BB->getParent();
8471 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8472 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8474 // Load the old value of the high byte of the control word...
8476 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8477 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8480 // Set the high part to be round to zero...
8481 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8484 // Reload the modified control word now...
8485 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8487 // Restore the memory image of control word to original value
8488 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8491 // Get the X86 opcode to use.
8493 switch (MI->getOpcode()) {
8494 default: llvm_unreachable("illegal opcode!");
8495 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8496 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8497 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8498 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8499 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8500 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8501 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8502 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8503 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8507 MachineOperand &Op = MI->getOperand(0);
8509 AM.BaseType = X86AddressMode::RegBase;
8510 AM.Base.Reg = Op.getReg();
8512 AM.BaseType = X86AddressMode::FrameIndexBase;
8513 AM.Base.FrameIndex = Op.getIndex();
8515 Op = MI->getOperand(1);
8517 AM.Scale = Op.getImm();
8518 Op = MI->getOperand(2);
8520 AM.IndexReg = Op.getImm();
8521 Op = MI->getOperand(3);
8522 if (Op.isGlobal()) {
8523 AM.GV = Op.getGlobal();
8525 AM.Disp = Op.getImm();
8527 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8528 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8530 // Reload the original control word now.
8531 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8533 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8536 // DBG_VALUE. Only the frame index case is done here.
8537 case X86::DBG_VALUE: {
8538 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8539 DebugLoc DL = MI->getDebugLoc();
8541 MachineFunction *F = BB->getParent();
8542 AM.BaseType = X86AddressMode::FrameIndexBase;
8543 AM.Base.FrameIndex = MI->getOperand(0).getImm();
8544 addFullAddress(BuildMI(BB, DL, TII->get(X86::DBG_VALUE)), AM).
8545 addImm(MI->getOperand(1).getImm()).
8546 addMetadata(MI->getOperand(2).getMetadata());
8547 F->DeleteMachineInstr(MI); // Remove pseudo.
8551 // String/text processing lowering.
8552 case X86::PCMPISTRM128REG:
8553 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8554 case X86::PCMPISTRM128MEM:
8555 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8556 case X86::PCMPESTRM128REG:
8557 return EmitPCMP(MI, BB, 5, false /* in mem */);
8558 case X86::PCMPESTRM128MEM:
8559 return EmitPCMP(MI, BB, 5, true /* in mem */);
8562 case X86::ATOMAND32:
8563 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8564 X86::AND32ri, X86::MOV32rm,
8565 X86::LCMPXCHG32, X86::MOV32rr,
8566 X86::NOT32r, X86::EAX,
8567 X86::GR32RegisterClass);
8569 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8570 X86::OR32ri, X86::MOV32rm,
8571 X86::LCMPXCHG32, X86::MOV32rr,
8572 X86::NOT32r, X86::EAX,
8573 X86::GR32RegisterClass);
8574 case X86::ATOMXOR32:
8575 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8576 X86::XOR32ri, X86::MOV32rm,
8577 X86::LCMPXCHG32, X86::MOV32rr,
8578 X86::NOT32r, X86::EAX,
8579 X86::GR32RegisterClass);
8580 case X86::ATOMNAND32:
8581 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8582 X86::AND32ri, X86::MOV32rm,
8583 X86::LCMPXCHG32, X86::MOV32rr,
8584 X86::NOT32r, X86::EAX,
8585 X86::GR32RegisterClass, true);
8586 case X86::ATOMMIN32:
8587 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8588 case X86::ATOMMAX32:
8589 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8590 case X86::ATOMUMIN32:
8591 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8592 case X86::ATOMUMAX32:
8593 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8595 case X86::ATOMAND16:
8596 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8597 X86::AND16ri, X86::MOV16rm,
8598 X86::LCMPXCHG16, X86::MOV16rr,
8599 X86::NOT16r, X86::AX,
8600 X86::GR16RegisterClass);
8602 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8603 X86::OR16ri, X86::MOV16rm,
8604 X86::LCMPXCHG16, X86::MOV16rr,
8605 X86::NOT16r, X86::AX,
8606 X86::GR16RegisterClass);
8607 case X86::ATOMXOR16:
8608 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8609 X86::XOR16ri, X86::MOV16rm,
8610 X86::LCMPXCHG16, X86::MOV16rr,
8611 X86::NOT16r, X86::AX,
8612 X86::GR16RegisterClass);
8613 case X86::ATOMNAND16:
8614 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8615 X86::AND16ri, X86::MOV16rm,
8616 X86::LCMPXCHG16, X86::MOV16rr,
8617 X86::NOT16r, X86::AX,
8618 X86::GR16RegisterClass, true);
8619 case X86::ATOMMIN16:
8620 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8621 case X86::ATOMMAX16:
8622 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8623 case X86::ATOMUMIN16:
8624 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8625 case X86::ATOMUMAX16:
8626 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8629 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8630 X86::AND8ri, X86::MOV8rm,
8631 X86::LCMPXCHG8, X86::MOV8rr,
8632 X86::NOT8r, X86::AL,
8633 X86::GR8RegisterClass);
8635 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8636 X86::OR8ri, X86::MOV8rm,
8637 X86::LCMPXCHG8, X86::MOV8rr,
8638 X86::NOT8r, X86::AL,
8639 X86::GR8RegisterClass);
8641 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8642 X86::XOR8ri, X86::MOV8rm,
8643 X86::LCMPXCHG8, X86::MOV8rr,
8644 X86::NOT8r, X86::AL,
8645 X86::GR8RegisterClass);
8646 case X86::ATOMNAND8:
8647 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8648 X86::AND8ri, X86::MOV8rm,
8649 X86::LCMPXCHG8, X86::MOV8rr,
8650 X86::NOT8r, X86::AL,
8651 X86::GR8RegisterClass, true);
8652 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8653 // This group is for 64-bit host.
8654 case X86::ATOMAND64:
8655 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8656 X86::AND64ri32, X86::MOV64rm,
8657 X86::LCMPXCHG64, X86::MOV64rr,
8658 X86::NOT64r, X86::RAX,
8659 X86::GR64RegisterClass);
8661 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8662 X86::OR64ri32, X86::MOV64rm,
8663 X86::LCMPXCHG64, X86::MOV64rr,
8664 X86::NOT64r, X86::RAX,
8665 X86::GR64RegisterClass);
8666 case X86::ATOMXOR64:
8667 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8668 X86::XOR64ri32, X86::MOV64rm,
8669 X86::LCMPXCHG64, X86::MOV64rr,
8670 X86::NOT64r, X86::RAX,
8671 X86::GR64RegisterClass);
8672 case X86::ATOMNAND64:
8673 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8674 X86::AND64ri32, X86::MOV64rm,
8675 X86::LCMPXCHG64, X86::MOV64rr,
8676 X86::NOT64r, X86::RAX,
8677 X86::GR64RegisterClass, true);
8678 case X86::ATOMMIN64:
8679 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8680 case X86::ATOMMAX64:
8681 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8682 case X86::ATOMUMIN64:
8683 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8684 case X86::ATOMUMAX64:
8685 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8687 // This group does 64-bit operations on a 32-bit host.
8688 case X86::ATOMAND6432:
8689 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8690 X86::AND32rr, X86::AND32rr,
8691 X86::AND32ri, X86::AND32ri,
8693 case X86::ATOMOR6432:
8694 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8695 X86::OR32rr, X86::OR32rr,
8696 X86::OR32ri, X86::OR32ri,
8698 case X86::ATOMXOR6432:
8699 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8700 X86::XOR32rr, X86::XOR32rr,
8701 X86::XOR32ri, X86::XOR32ri,
8703 case X86::ATOMNAND6432:
8704 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8705 X86::AND32rr, X86::AND32rr,
8706 X86::AND32ri, X86::AND32ri,
8708 case X86::ATOMADD6432:
8709 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8710 X86::ADD32rr, X86::ADC32rr,
8711 X86::ADD32ri, X86::ADC32ri,
8713 case X86::ATOMSUB6432:
8714 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8715 X86::SUB32rr, X86::SBB32rr,
8716 X86::SUB32ri, X86::SBB32ri,
8718 case X86::ATOMSWAP6432:
8719 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8720 X86::MOV32rr, X86::MOV32rr,
8721 X86::MOV32ri, X86::MOV32ri,
8723 case X86::VASTART_SAVE_XMM_REGS:
8724 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8728 //===----------------------------------------------------------------------===//
8729 // X86 Optimization Hooks
8730 //===----------------------------------------------------------------------===//
8732 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8736 const SelectionDAG &DAG,
8737 unsigned Depth) const {
8738 unsigned Opc = Op.getOpcode();
8739 assert((Opc >= ISD::BUILTIN_OP_END ||
8740 Opc == ISD::INTRINSIC_WO_CHAIN ||
8741 Opc == ISD::INTRINSIC_W_CHAIN ||
8742 Opc == ISD::INTRINSIC_VOID) &&
8743 "Should use MaskedValueIsZero if you don't know whether Op"
8744 " is a target node!");
8746 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8758 // These nodes' second result is a boolean.
8759 if (Op.getResNo() == 0)
8763 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8764 Mask.getBitWidth() - 1);
8769 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8770 /// node is a GlobalAddress + offset.
8771 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8772 GlobalValue* &GA, int64_t &Offset) const{
8773 if (N->getOpcode() == X86ISD::Wrapper) {
8774 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8775 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8776 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8780 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8783 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8784 EVT EltVT, LoadSDNode *&LDBase,
8785 unsigned &LastLoadedElt,
8786 SelectionDAG &DAG, MachineFrameInfo *MFI,
8787 const TargetLowering &TLI) {
8789 LastLoadedElt = -1U;
8790 for (unsigned i = 0; i < NumElems; ++i) {
8791 if (N->getMaskElt(i) < 0) {
8797 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8798 if (!Elt.getNode() ||
8799 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8802 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8804 LDBase = cast<LoadSDNode>(Elt.getNode());
8808 if (Elt.getOpcode() == ISD::UNDEF)
8811 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8812 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8819 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8820 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8821 /// if the load addresses are consecutive, non-overlapping, and in the right
8822 /// order. In the case of v2i64, it will see if it can rewrite the
8823 /// shuffle to be an appropriate build vector so it can take advantage of
8824 // performBuildVectorCombine.
8825 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8826 const TargetLowering &TLI) {
8827 DebugLoc dl = N->getDebugLoc();
8828 EVT VT = N->getValueType(0);
8829 EVT EltVT = VT.getVectorElementType();
8830 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8831 unsigned NumElems = VT.getVectorNumElements();
8833 if (VT.getSizeInBits() != 128)
8836 // Try to combine a vector_shuffle into a 128-bit load.
8837 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8838 LoadSDNode *LD = NULL;
8839 unsigned LastLoadedElt;
8840 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8844 if (LastLoadedElt == NumElems - 1) {
8845 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8846 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8847 LD->getSrcValue(), LD->getSrcValueOffset(),
8848 LD->isVolatile(), LD->isNonTemporal(), 0);
8849 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8850 LD->getSrcValue(), LD->getSrcValueOffset(),
8851 LD->isVolatile(), LD->isNonTemporal(),
8852 LD->getAlignment());
8853 } else if (NumElems == 4 && LastLoadedElt == 1) {
8854 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8855 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8856 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8857 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8862 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8863 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8864 const X86Subtarget *Subtarget) {
8865 DebugLoc DL = N->getDebugLoc();
8866 SDValue Cond = N->getOperand(0);
8867 // Get the LHS/RHS of the select.
8868 SDValue LHS = N->getOperand(1);
8869 SDValue RHS = N->getOperand(2);
8871 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8872 // instructions match the semantics of the common C idiom x<y?x:y but not
8873 // x<=y?x:y, because of how they handle negative zero (which can be
8874 // ignored in unsafe-math mode).
8875 if (Subtarget->hasSSE2() &&
8876 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8877 Cond.getOpcode() == ISD::SETCC) {
8878 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8880 unsigned Opcode = 0;
8881 // Check for x CC y ? x : y.
8882 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
8883 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
8887 // Converting this to a min would handle NaNs incorrectly, and swapping
8888 // the operands would cause it to handle comparisons between positive
8889 // and negative zero incorrectly.
8890 if (!FiniteOnlyFPMath() &&
8891 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
8892 if (!UnsafeFPMath &&
8893 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8895 std::swap(LHS, RHS);
8897 Opcode = X86ISD::FMIN;
8900 // Converting this to a min would handle comparisons between positive
8901 // and negative zero incorrectly.
8902 if (!UnsafeFPMath &&
8903 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
8905 Opcode = X86ISD::FMIN;
8908 // Converting this to a min would handle both negative zeros and NaNs
8909 // incorrectly, but we can swap the operands to fix both.
8910 std::swap(LHS, RHS);
8914 Opcode = X86ISD::FMIN;
8918 // Converting this to a max would handle comparisons between positive
8919 // and negative zero incorrectly.
8920 if (!UnsafeFPMath &&
8921 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
8923 Opcode = X86ISD::FMAX;
8926 // Converting this to a max would handle NaNs incorrectly, and swapping
8927 // the operands would cause it to handle comparisons between positive
8928 // and negative zero incorrectly.
8929 if (!FiniteOnlyFPMath() &&
8930 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
8931 if (!UnsafeFPMath &&
8932 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8934 std::swap(LHS, RHS);
8936 Opcode = X86ISD::FMAX;
8939 // Converting this to a max would handle both negative zeros and NaNs
8940 // incorrectly, but we can swap the operands to fix both.
8941 std::swap(LHS, RHS);
8945 Opcode = X86ISD::FMAX;
8948 // Check for x CC y ? y : x -- a min/max with reversed arms.
8949 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
8950 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
8954 // Converting this to a min would handle comparisons between positive
8955 // and negative zero incorrectly, and swapping the operands would
8956 // cause it to handle NaNs incorrectly.
8957 if (!UnsafeFPMath &&
8958 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
8959 if (!FiniteOnlyFPMath() &&
8960 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8962 std::swap(LHS, RHS);
8964 Opcode = X86ISD::FMIN;
8967 // Converting this to a min would handle NaNs incorrectly.
8968 if (!UnsafeFPMath &&
8969 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8971 Opcode = X86ISD::FMIN;
8974 // Converting this to a min would handle both negative zeros and NaNs
8975 // incorrectly, but we can swap the operands to fix both.
8976 std::swap(LHS, RHS);
8980 Opcode = X86ISD::FMIN;
8984 // Converting this to a max would handle NaNs incorrectly.
8985 if (!FiniteOnlyFPMath() &&
8986 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8988 Opcode = X86ISD::FMAX;
8991 // Converting this to a max would handle comparisons between positive
8992 // and negative zero incorrectly, and swapping the operands would
8993 // cause it to handle NaNs incorrectly.
8994 if (!UnsafeFPMath &&
8995 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
8996 if (!FiniteOnlyFPMath() &&
8997 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8999 std::swap(LHS, RHS);
9001 Opcode = X86ISD::FMAX;
9004 // Converting this to a max would handle both negative zeros and NaNs
9005 // incorrectly, but we can swap the operands to fix both.
9006 std::swap(LHS, RHS);
9010 Opcode = X86ISD::FMAX;
9016 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9019 // If this is a select between two integer constants, try to do some
9021 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9022 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9023 // Don't do this for crazy integer types.
9024 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9025 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9026 // so that TrueC (the true value) is larger than FalseC.
9027 bool NeedsCondInvert = false;
9029 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9030 // Efficiently invertible.
9031 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9032 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9033 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9034 NeedsCondInvert = true;
9035 std::swap(TrueC, FalseC);
9038 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9039 if (FalseC->getAPIntValue() == 0 &&
9040 TrueC->getAPIntValue().isPowerOf2()) {
9041 if (NeedsCondInvert) // Invert the condition if needed.
9042 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9043 DAG.getConstant(1, Cond.getValueType()));
9045 // Zero extend the condition if needed.
9046 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9048 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9049 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9050 DAG.getConstant(ShAmt, MVT::i8));
9053 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9054 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9055 if (NeedsCondInvert) // Invert the condition if needed.
9056 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9057 DAG.getConstant(1, Cond.getValueType()));
9059 // Zero extend the condition if needed.
9060 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9061 FalseC->getValueType(0), Cond);
9062 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9063 SDValue(FalseC, 0));
9066 // Optimize cases that will turn into an LEA instruction. This requires
9067 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9068 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9069 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9070 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9072 bool isFastMultiplier = false;
9074 switch ((unsigned char)Diff) {
9076 case 1: // result = add base, cond
9077 case 2: // result = lea base( , cond*2)
9078 case 3: // result = lea base(cond, cond*2)
9079 case 4: // result = lea base( , cond*4)
9080 case 5: // result = lea base(cond, cond*4)
9081 case 8: // result = lea base( , cond*8)
9082 case 9: // result = lea base(cond, cond*8)
9083 isFastMultiplier = true;
9088 if (isFastMultiplier) {
9089 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9090 if (NeedsCondInvert) // Invert the condition if needed.
9091 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9092 DAG.getConstant(1, Cond.getValueType()));
9094 // Zero extend the condition if needed.
9095 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9097 // Scale the condition by the difference.
9099 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9100 DAG.getConstant(Diff, Cond.getValueType()));
9102 // Add the base if non-zero.
9103 if (FalseC->getAPIntValue() != 0)
9104 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9105 SDValue(FalseC, 0));
9115 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9116 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9117 TargetLowering::DAGCombinerInfo &DCI) {
9118 DebugLoc DL = N->getDebugLoc();
9120 // If the flag operand isn't dead, don't touch this CMOV.
9121 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9124 // If this is a select between two integer constants, try to do some
9125 // optimizations. Note that the operands are ordered the opposite of SELECT
9127 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9128 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9129 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9130 // larger than FalseC (the false value).
9131 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9133 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9134 CC = X86::GetOppositeBranchCondition(CC);
9135 std::swap(TrueC, FalseC);
9138 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9139 // This is efficient for any integer data type (including i8/i16) and
9141 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9142 SDValue Cond = N->getOperand(3);
9143 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9144 DAG.getConstant(CC, MVT::i8), Cond);
9146 // Zero extend the condition if needed.
9147 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9149 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9150 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9151 DAG.getConstant(ShAmt, MVT::i8));
9152 if (N->getNumValues() == 2) // Dead flag value?
9153 return DCI.CombineTo(N, Cond, SDValue());
9157 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9158 // for any integer data type, including i8/i16.
9159 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9160 SDValue Cond = N->getOperand(3);
9161 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9162 DAG.getConstant(CC, MVT::i8), Cond);
9164 // Zero extend the condition if needed.
9165 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9166 FalseC->getValueType(0), Cond);
9167 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9168 SDValue(FalseC, 0));
9170 if (N->getNumValues() == 2) // Dead flag value?
9171 return DCI.CombineTo(N, Cond, SDValue());
9175 // Optimize cases that will turn into an LEA instruction. This requires
9176 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9177 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9178 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9179 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9181 bool isFastMultiplier = false;
9183 switch ((unsigned char)Diff) {
9185 case 1: // result = add base, cond
9186 case 2: // result = lea base( , cond*2)
9187 case 3: // result = lea base(cond, cond*2)
9188 case 4: // result = lea base( , cond*4)
9189 case 5: // result = lea base(cond, cond*4)
9190 case 8: // result = lea base( , cond*8)
9191 case 9: // result = lea base(cond, cond*8)
9192 isFastMultiplier = true;
9197 if (isFastMultiplier) {
9198 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9199 SDValue Cond = N->getOperand(3);
9200 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9201 DAG.getConstant(CC, MVT::i8), Cond);
9202 // Zero extend the condition if needed.
9203 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9205 // Scale the condition by the difference.
9207 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9208 DAG.getConstant(Diff, Cond.getValueType()));
9210 // Add the base if non-zero.
9211 if (FalseC->getAPIntValue() != 0)
9212 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9213 SDValue(FalseC, 0));
9214 if (N->getNumValues() == 2) // Dead flag value?
9215 return DCI.CombineTo(N, Cond, SDValue());
9225 /// PerformMulCombine - Optimize a single multiply with constant into two
9226 /// in order to implement it with two cheaper instructions, e.g.
9227 /// LEA + SHL, LEA + LEA.
9228 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9229 TargetLowering::DAGCombinerInfo &DCI) {
9230 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9233 EVT VT = N->getValueType(0);
9237 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9240 uint64_t MulAmt = C->getZExtValue();
9241 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9244 uint64_t MulAmt1 = 0;
9245 uint64_t MulAmt2 = 0;
9246 if ((MulAmt % 9) == 0) {
9248 MulAmt2 = MulAmt / 9;
9249 } else if ((MulAmt % 5) == 0) {
9251 MulAmt2 = MulAmt / 5;
9252 } else if ((MulAmt % 3) == 0) {
9254 MulAmt2 = MulAmt / 3;
9257 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9258 DebugLoc DL = N->getDebugLoc();
9260 if (isPowerOf2_64(MulAmt2) &&
9261 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9262 // If second multiplifer is pow2, issue it first. We want the multiply by
9263 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9265 std::swap(MulAmt1, MulAmt2);
9268 if (isPowerOf2_64(MulAmt1))
9269 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9270 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9272 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9273 DAG.getConstant(MulAmt1, VT));
9275 if (isPowerOf2_64(MulAmt2))
9276 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9277 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9279 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9280 DAG.getConstant(MulAmt2, VT));
9282 // Do not add new nodes to DAG combiner worklist.
9283 DCI.CombineTo(N, NewMul, false);
9288 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9289 SDValue N0 = N->getOperand(0);
9290 SDValue N1 = N->getOperand(1);
9291 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9292 EVT VT = N0.getValueType();
9294 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9295 // since the result of setcc_c is all zero's or all ones.
9296 if (N1C && N0.getOpcode() == ISD::AND &&
9297 N0.getOperand(1).getOpcode() == ISD::Constant) {
9298 SDValue N00 = N0.getOperand(0);
9299 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9300 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9301 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9302 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9303 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9304 APInt ShAmt = N1C->getAPIntValue();
9305 Mask = Mask.shl(ShAmt);
9307 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9308 N00, DAG.getConstant(Mask, VT));
9315 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9317 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9318 const X86Subtarget *Subtarget) {
9319 EVT VT = N->getValueType(0);
9320 if (!VT.isVector() && VT.isInteger() &&
9321 N->getOpcode() == ISD::SHL)
9322 return PerformSHLCombine(N, DAG);
9324 // On X86 with SSE2 support, we can transform this to a vector shift if
9325 // all elements are shifted by the same amount. We can't do this in legalize
9326 // because the a constant vector is typically transformed to a constant pool
9327 // so we have no knowledge of the shift amount.
9328 if (!Subtarget->hasSSE2())
9331 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9334 SDValue ShAmtOp = N->getOperand(1);
9335 EVT EltVT = VT.getVectorElementType();
9336 DebugLoc DL = N->getDebugLoc();
9337 SDValue BaseShAmt = SDValue();
9338 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9339 unsigned NumElts = VT.getVectorNumElements();
9341 for (; i != NumElts; ++i) {
9342 SDValue Arg = ShAmtOp.getOperand(i);
9343 if (Arg.getOpcode() == ISD::UNDEF) continue;
9347 for (; i != NumElts; ++i) {
9348 SDValue Arg = ShAmtOp.getOperand(i);
9349 if (Arg.getOpcode() == ISD::UNDEF) continue;
9350 if (Arg != BaseShAmt) {
9354 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9355 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9356 SDValue InVec = ShAmtOp.getOperand(0);
9357 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9358 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9360 for (; i != NumElts; ++i) {
9361 SDValue Arg = InVec.getOperand(i);
9362 if (Arg.getOpcode() == ISD::UNDEF) continue;
9366 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9367 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9368 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9369 if (C->getZExtValue() == SplatIdx)
9370 BaseShAmt = InVec.getOperand(1);
9373 if (BaseShAmt.getNode() == 0)
9374 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9375 DAG.getIntPtrConstant(0));
9379 // The shift amount is an i32.
9380 if (EltVT.bitsGT(MVT::i32))
9381 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9382 else if (EltVT.bitsLT(MVT::i32))
9383 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9385 // The shift amount is identical so we can do a vector shift.
9386 SDValue ValOp = N->getOperand(0);
9387 switch (N->getOpcode()) {
9389 llvm_unreachable("Unknown shift opcode!");
9392 if (VT == MVT::v2i64)
9393 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9394 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9396 if (VT == MVT::v4i32)
9397 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9398 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9400 if (VT == MVT::v8i16)
9401 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9402 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9406 if (VT == MVT::v4i32)
9407 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9408 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9410 if (VT == MVT::v8i16)
9411 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9412 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9416 if (VT == MVT::v2i64)
9417 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9418 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9420 if (VT == MVT::v4i32)
9421 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9422 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9424 if (VT == MVT::v8i16)
9425 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9426 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9433 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9434 const X86Subtarget *Subtarget) {
9435 EVT VT = N->getValueType(0);
9436 if (VT != MVT::i64 || !Subtarget->is64Bit())
9439 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9440 SDValue N0 = N->getOperand(0);
9441 SDValue N1 = N->getOperand(1);
9442 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9444 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9447 SDValue ShAmt0 = N0.getOperand(1);
9448 if (ShAmt0.getValueType() != MVT::i8)
9450 SDValue ShAmt1 = N1.getOperand(1);
9451 if (ShAmt1.getValueType() != MVT::i8)
9453 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9454 ShAmt0 = ShAmt0.getOperand(0);
9455 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9456 ShAmt1 = ShAmt1.getOperand(0);
9458 DebugLoc DL = N->getDebugLoc();
9459 unsigned Opc = X86ISD::SHLD;
9460 SDValue Op0 = N0.getOperand(0);
9461 SDValue Op1 = N1.getOperand(0);
9462 if (ShAmt0.getOpcode() == ISD::SUB) {
9464 std::swap(Op0, Op1);
9465 std::swap(ShAmt0, ShAmt1);
9468 if (ShAmt1.getOpcode() == ISD::SUB) {
9469 SDValue Sum = ShAmt1.getOperand(0);
9470 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9471 if (SumC->getSExtValue() == 64 &&
9472 ShAmt1.getOperand(1) == ShAmt0)
9473 return DAG.getNode(Opc, DL, VT,
9475 DAG.getNode(ISD::TRUNCATE, DL,
9478 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9479 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9481 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9482 return DAG.getNode(Opc, DL, VT,
9483 N0.getOperand(0), N1.getOperand(0),
9484 DAG.getNode(ISD::TRUNCATE, DL,
9491 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9492 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9493 const X86Subtarget *Subtarget) {
9494 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9495 // the FP state in cases where an emms may be missing.
9496 // A preferable solution to the general problem is to figure out the right
9497 // places to insert EMMS. This qualifies as a quick hack.
9499 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9500 StoreSDNode *St = cast<StoreSDNode>(N);
9501 EVT VT = St->getValue().getValueType();
9502 if (VT.getSizeInBits() != 64)
9505 const Function *F = DAG.getMachineFunction().getFunction();
9506 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9507 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9508 && Subtarget->hasSSE2();
9509 if ((VT.isVector() ||
9510 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9511 isa<LoadSDNode>(St->getValue()) &&
9512 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9513 St->getChain().hasOneUse() && !St->isVolatile()) {
9514 SDNode* LdVal = St->getValue().getNode();
9516 int TokenFactorIndex = -1;
9517 SmallVector<SDValue, 8> Ops;
9518 SDNode* ChainVal = St->getChain().getNode();
9519 // Must be a store of a load. We currently handle two cases: the load
9520 // is a direct child, and it's under an intervening TokenFactor. It is
9521 // possible to dig deeper under nested TokenFactors.
9522 if (ChainVal == LdVal)
9523 Ld = cast<LoadSDNode>(St->getChain());
9524 else if (St->getValue().hasOneUse() &&
9525 ChainVal->getOpcode() == ISD::TokenFactor) {
9526 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9527 if (ChainVal->getOperand(i).getNode() == LdVal) {
9528 TokenFactorIndex = i;
9529 Ld = cast<LoadSDNode>(St->getValue());
9531 Ops.push_back(ChainVal->getOperand(i));
9535 if (!Ld || !ISD::isNormalLoad(Ld))
9538 // If this is not the MMX case, i.e. we are just turning i64 load/store
9539 // into f64 load/store, avoid the transformation if there are multiple
9540 // uses of the loaded value.
9541 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9544 DebugLoc LdDL = Ld->getDebugLoc();
9545 DebugLoc StDL = N->getDebugLoc();
9546 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9547 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9549 if (Subtarget->is64Bit() || F64IsLegal) {
9550 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9551 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9552 Ld->getBasePtr(), Ld->getSrcValue(),
9553 Ld->getSrcValueOffset(), Ld->isVolatile(),
9554 Ld->isNonTemporal(), Ld->getAlignment());
9555 SDValue NewChain = NewLd.getValue(1);
9556 if (TokenFactorIndex != -1) {
9557 Ops.push_back(NewChain);
9558 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9561 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9562 St->getSrcValue(), St->getSrcValueOffset(),
9563 St->isVolatile(), St->isNonTemporal(),
9564 St->getAlignment());
9567 // Otherwise, lower to two pairs of 32-bit loads / stores.
9568 SDValue LoAddr = Ld->getBasePtr();
9569 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9570 DAG.getConstant(4, MVT::i32));
9572 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9573 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9574 Ld->isVolatile(), Ld->isNonTemporal(),
9575 Ld->getAlignment());
9576 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9577 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9578 Ld->isVolatile(), Ld->isNonTemporal(),
9579 MinAlign(Ld->getAlignment(), 4));
9581 SDValue NewChain = LoLd.getValue(1);
9582 if (TokenFactorIndex != -1) {
9583 Ops.push_back(LoLd);
9584 Ops.push_back(HiLd);
9585 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9589 LoAddr = St->getBasePtr();
9590 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9591 DAG.getConstant(4, MVT::i32));
9593 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9594 St->getSrcValue(), St->getSrcValueOffset(),
9595 St->isVolatile(), St->isNonTemporal(),
9596 St->getAlignment());
9597 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9599 St->getSrcValueOffset() + 4,
9601 St->isNonTemporal(),
9602 MinAlign(St->getAlignment(), 4));
9603 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9608 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9609 /// X86ISD::FXOR nodes.
9610 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9611 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9612 // F[X]OR(0.0, x) -> x
9613 // F[X]OR(x, 0.0) -> x
9614 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9615 if (C->getValueAPF().isPosZero())
9616 return N->getOperand(1);
9617 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9618 if (C->getValueAPF().isPosZero())
9619 return N->getOperand(0);
9623 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9624 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9625 // FAND(0.0, x) -> 0.0
9626 // FAND(x, 0.0) -> 0.0
9627 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9628 if (C->getValueAPF().isPosZero())
9629 return N->getOperand(0);
9630 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9631 if (C->getValueAPF().isPosZero())
9632 return N->getOperand(1);
9636 static SDValue PerformBTCombine(SDNode *N,
9638 TargetLowering::DAGCombinerInfo &DCI) {
9639 // BT ignores high bits in the bit index operand.
9640 SDValue Op1 = N->getOperand(1);
9641 if (Op1.hasOneUse()) {
9642 unsigned BitWidth = Op1.getValueSizeInBits();
9643 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9644 APInt KnownZero, KnownOne;
9645 TargetLowering::TargetLoweringOpt TLO(DAG);
9646 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9647 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9648 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9649 DCI.CommitTargetLoweringOpt(TLO);
9654 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9655 SDValue Op = N->getOperand(0);
9656 if (Op.getOpcode() == ISD::BIT_CONVERT)
9657 Op = Op.getOperand(0);
9658 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9659 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9660 VT.getVectorElementType().getSizeInBits() ==
9661 OpVT.getVectorElementType().getSizeInBits()) {
9662 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9667 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9668 // Locked instructions, in turn, have implicit fence semantics (all memory
9669 // operations are flushed before issuing the locked instruction, and the
9670 // are not buffered), so we can fold away the common pattern of
9671 // fence-atomic-fence.
9672 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9673 SDValue atomic = N->getOperand(0);
9674 switch (atomic.getOpcode()) {
9675 case ISD::ATOMIC_CMP_SWAP:
9676 case ISD::ATOMIC_SWAP:
9677 case ISD::ATOMIC_LOAD_ADD:
9678 case ISD::ATOMIC_LOAD_SUB:
9679 case ISD::ATOMIC_LOAD_AND:
9680 case ISD::ATOMIC_LOAD_OR:
9681 case ISD::ATOMIC_LOAD_XOR:
9682 case ISD::ATOMIC_LOAD_NAND:
9683 case ISD::ATOMIC_LOAD_MIN:
9684 case ISD::ATOMIC_LOAD_MAX:
9685 case ISD::ATOMIC_LOAD_UMIN:
9686 case ISD::ATOMIC_LOAD_UMAX:
9692 SDValue fence = atomic.getOperand(0);
9693 if (fence.getOpcode() != ISD::MEMBARRIER)
9696 switch (atomic.getOpcode()) {
9697 case ISD::ATOMIC_CMP_SWAP:
9698 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9699 atomic.getOperand(1), atomic.getOperand(2),
9700 atomic.getOperand(3));
9701 case ISD::ATOMIC_SWAP:
9702 case ISD::ATOMIC_LOAD_ADD:
9703 case ISD::ATOMIC_LOAD_SUB:
9704 case ISD::ATOMIC_LOAD_AND:
9705 case ISD::ATOMIC_LOAD_OR:
9706 case ISD::ATOMIC_LOAD_XOR:
9707 case ISD::ATOMIC_LOAD_NAND:
9708 case ISD::ATOMIC_LOAD_MIN:
9709 case ISD::ATOMIC_LOAD_MAX:
9710 case ISD::ATOMIC_LOAD_UMIN:
9711 case ISD::ATOMIC_LOAD_UMAX:
9712 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9713 atomic.getOperand(1), atomic.getOperand(2));
9719 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9720 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9721 // (and (i32 x86isd::setcc_carry), 1)
9722 // This eliminates the zext. This transformation is necessary because
9723 // ISD::SETCC is always legalized to i8.
9724 DebugLoc dl = N->getDebugLoc();
9725 SDValue N0 = N->getOperand(0);
9726 EVT VT = N->getValueType(0);
9727 if (N0.getOpcode() == ISD::AND &&
9729 N0.getOperand(0).hasOneUse()) {
9730 SDValue N00 = N0.getOperand(0);
9731 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9733 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9734 if (!C || C->getZExtValue() != 1)
9736 return DAG.getNode(ISD::AND, dl, VT,
9737 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9738 N00.getOperand(0), N00.getOperand(1)),
9739 DAG.getConstant(1, VT));
9745 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9746 DAGCombinerInfo &DCI) const {
9747 SelectionDAG &DAG = DCI.DAG;
9748 switch (N->getOpcode()) {
9750 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9751 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9752 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9753 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9756 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9757 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9758 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9760 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9761 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9762 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9763 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9764 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9765 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9771 //===----------------------------------------------------------------------===//
9772 // X86 Inline Assembly Support
9773 //===----------------------------------------------------------------------===//
9775 static bool LowerToBSwap(CallInst *CI) {
9776 // FIXME: this should verify that we are targetting a 486 or better. If not,
9777 // we will turn this bswap into something that will be lowered to logical ops
9778 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9779 // so don't worry about this.
9781 // Verify this is a simple bswap.
9782 if (CI->getNumOperands() != 2 ||
9783 CI->getType() != CI->getOperand(1)->getType() ||
9784 !CI->getType()->isIntegerTy())
9787 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9788 if (!Ty || Ty->getBitWidth() % 16 != 0)
9791 // Okay, we can do this xform, do so now.
9792 const Type *Tys[] = { Ty };
9793 Module *M = CI->getParent()->getParent()->getParent();
9794 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9796 Value *Op = CI->getOperand(1);
9797 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9799 CI->replaceAllUsesWith(Op);
9800 CI->eraseFromParent();
9804 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9805 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9806 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9808 std::string AsmStr = IA->getAsmString();
9810 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9811 SmallVector<StringRef, 4> AsmPieces;
9812 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9814 switch (AsmPieces.size()) {
9815 default: return false;
9817 AsmStr = AsmPieces[0];
9819 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9822 if (AsmPieces.size() == 2 &&
9823 (AsmPieces[0] == "bswap" ||
9824 AsmPieces[0] == "bswapq" ||
9825 AsmPieces[0] == "bswapl") &&
9826 (AsmPieces[1] == "$0" ||
9827 AsmPieces[1] == "${0:q}")) {
9828 // No need to check constraints, nothing other than the equivalent of
9829 // "=r,0" would be valid here.
9830 return LowerToBSwap(CI);
9832 // rorw $$8, ${0:w} --> llvm.bswap.i16
9833 if (CI->getType()->isIntegerTy(16) &&
9834 AsmPieces.size() == 3 &&
9835 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
9836 AsmPieces[1] == "$$8," &&
9837 AsmPieces[2] == "${0:w}" &&
9838 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
9840 const std::string &Constraints = IA->getConstraintString();
9841 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
9842 std::sort(AsmPieces.begin(), AsmPieces.end());
9843 if (AsmPieces.size() == 4 &&
9844 AsmPieces[0] == "~{cc}" &&
9845 AsmPieces[1] == "~{dirflag}" &&
9846 AsmPieces[2] == "~{flags}" &&
9847 AsmPieces[3] == "~{fpsr}") {
9848 return LowerToBSwap(CI);
9853 if (CI->getType()->isIntegerTy(64) &&
9854 Constraints.size() >= 2 &&
9855 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9856 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9857 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9858 SmallVector<StringRef, 4> Words;
9859 SplitString(AsmPieces[0], Words, " \t");
9860 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9862 SplitString(AsmPieces[1], Words, " \t");
9863 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9865 SplitString(AsmPieces[2], Words, " \t,");
9866 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9867 Words[2] == "%edx") {
9868 return LowerToBSwap(CI);
9880 /// getConstraintType - Given a constraint letter, return the type of
9881 /// constraint it is for this target.
9882 X86TargetLowering::ConstraintType
9883 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9884 if (Constraint.size() == 1) {
9885 switch (Constraint[0]) {
9897 return C_RegisterClass;
9905 return TargetLowering::getConstraintType(Constraint);
9908 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9909 /// with another that has more specific requirements based on the type of the
9910 /// corresponding operand.
9911 const char *X86TargetLowering::
9912 LowerXConstraint(EVT ConstraintVT) const {
9913 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9914 // 'f' like normal targets.
9915 if (ConstraintVT.isFloatingPoint()) {
9916 if (Subtarget->hasSSE2())
9918 if (Subtarget->hasSSE1())
9922 return TargetLowering::LowerXConstraint(ConstraintVT);
9925 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9926 /// vector. If it is invalid, don't add anything to Ops.
9927 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9930 std::vector<SDValue>&Ops,
9931 SelectionDAG &DAG) const {
9932 SDValue Result(0, 0);
9934 switch (Constraint) {
9937 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9938 if (C->getZExtValue() <= 31) {
9939 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9945 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9946 if (C->getZExtValue() <= 63) {
9947 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9953 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9954 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9955 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9961 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9962 if (C->getZExtValue() <= 255) {
9963 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9969 // 32-bit signed value
9970 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9971 const ConstantInt *CI = C->getConstantIntValue();
9972 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9973 C->getSExtValue())) {
9974 // Widen to 64 bits here to get it sign extended.
9975 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9978 // FIXME gcc accepts some relocatable values here too, but only in certain
9979 // memory models; it's complicated.
9984 // 32-bit unsigned value
9985 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9986 const ConstantInt *CI = C->getConstantIntValue();
9987 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9988 C->getZExtValue())) {
9989 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9993 // FIXME gcc accepts some relocatable values here too, but only in certain
9994 // memory models; it's complicated.
9998 // Literal immediates are always ok.
9999 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10000 // Widen to 64 bits here to get it sign extended.
10001 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10005 // If we are in non-pic codegen mode, we allow the address of a global (with
10006 // an optional displacement) to be used with 'i'.
10007 GlobalAddressSDNode *GA = 0;
10008 int64_t Offset = 0;
10010 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10012 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10013 Offset += GA->getOffset();
10015 } else if (Op.getOpcode() == ISD::ADD) {
10016 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10017 Offset += C->getZExtValue();
10018 Op = Op.getOperand(0);
10021 } else if (Op.getOpcode() == ISD::SUB) {
10022 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10023 Offset += -C->getZExtValue();
10024 Op = Op.getOperand(0);
10029 // Otherwise, this isn't something we can handle, reject it.
10033 GlobalValue *GV = GA->getGlobal();
10034 // If we require an extra load to get this address, as in PIC mode, we
10035 // can't accept it.
10036 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10037 getTargetMachine())))
10041 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
10043 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10049 if (Result.getNode()) {
10050 Ops.push_back(Result);
10053 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
10057 std::vector<unsigned> X86TargetLowering::
10058 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10060 if (Constraint.size() == 1) {
10061 // FIXME: not handling fp-stack yet!
10062 switch (Constraint[0]) { // GCC X86 Constraint Letters
10063 default: break; // Unknown constraint letter
10064 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10065 if (Subtarget->is64Bit()) {
10066 if (VT == MVT::i32)
10067 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10068 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10069 X86::R10D,X86::R11D,X86::R12D,
10070 X86::R13D,X86::R14D,X86::R15D,
10071 X86::EBP, X86::ESP, 0);
10072 else if (VT == MVT::i16)
10073 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10074 X86::SI, X86::DI, X86::R8W,X86::R9W,
10075 X86::R10W,X86::R11W,X86::R12W,
10076 X86::R13W,X86::R14W,X86::R15W,
10077 X86::BP, X86::SP, 0);
10078 else if (VT == MVT::i8)
10079 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10080 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10081 X86::R10B,X86::R11B,X86::R12B,
10082 X86::R13B,X86::R14B,X86::R15B,
10083 X86::BPL, X86::SPL, 0);
10085 else if (VT == MVT::i64)
10086 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10087 X86::RSI, X86::RDI, X86::R8, X86::R9,
10088 X86::R10, X86::R11, X86::R12,
10089 X86::R13, X86::R14, X86::R15,
10090 X86::RBP, X86::RSP, 0);
10094 // 32-bit fallthrough
10095 case 'Q': // Q_REGS
10096 if (VT == MVT::i32)
10097 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10098 else if (VT == MVT::i16)
10099 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10100 else if (VT == MVT::i8)
10101 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10102 else if (VT == MVT::i64)
10103 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10108 return std::vector<unsigned>();
10111 std::pair<unsigned, const TargetRegisterClass*>
10112 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10114 // First, see if this is a constraint that directly corresponds to an LLVM
10116 if (Constraint.size() == 1) {
10117 // GCC Constraint Letters
10118 switch (Constraint[0]) {
10120 case 'r': // GENERAL_REGS
10121 case 'l': // INDEX_REGS
10123 return std::make_pair(0U, X86::GR8RegisterClass);
10124 if (VT == MVT::i16)
10125 return std::make_pair(0U, X86::GR16RegisterClass);
10126 if (VT == MVT::i32 || !Subtarget->is64Bit())
10127 return std::make_pair(0U, X86::GR32RegisterClass);
10128 return std::make_pair(0U, X86::GR64RegisterClass);
10129 case 'R': // LEGACY_REGS
10131 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10132 if (VT == MVT::i16)
10133 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10134 if (VT == MVT::i32 || !Subtarget->is64Bit())
10135 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10136 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10137 case 'f': // FP Stack registers.
10138 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10139 // value to the correct fpstack register class.
10140 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10141 return std::make_pair(0U, X86::RFP32RegisterClass);
10142 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10143 return std::make_pair(0U, X86::RFP64RegisterClass);
10144 return std::make_pair(0U, X86::RFP80RegisterClass);
10145 case 'y': // MMX_REGS if MMX allowed.
10146 if (!Subtarget->hasMMX()) break;
10147 return std::make_pair(0U, X86::VR64RegisterClass);
10148 case 'Y': // SSE_REGS if SSE2 allowed
10149 if (!Subtarget->hasSSE2()) break;
10151 case 'x': // SSE_REGS if SSE1 allowed
10152 if (!Subtarget->hasSSE1()) break;
10154 switch (VT.getSimpleVT().SimpleTy) {
10156 // Scalar SSE types.
10159 return std::make_pair(0U, X86::FR32RegisterClass);
10162 return std::make_pair(0U, X86::FR64RegisterClass);
10170 return std::make_pair(0U, X86::VR128RegisterClass);
10176 // Use the default implementation in TargetLowering to convert the register
10177 // constraint into a member of a register class.
10178 std::pair<unsigned, const TargetRegisterClass*> Res;
10179 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10181 // Not found as a standard register?
10182 if (Res.second == 0) {
10183 // Map st(0) -> st(7) -> ST0
10184 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10185 tolower(Constraint[1]) == 's' &&
10186 tolower(Constraint[2]) == 't' &&
10187 Constraint[3] == '(' &&
10188 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10189 Constraint[5] == ')' &&
10190 Constraint[6] == '}') {
10192 Res.first = X86::ST0+Constraint[4]-'0';
10193 Res.second = X86::RFP80RegisterClass;
10197 // GCC allows "st(0)" to be called just plain "st".
10198 if (StringRef("{st}").equals_lower(Constraint)) {
10199 Res.first = X86::ST0;
10200 Res.second = X86::RFP80RegisterClass;
10205 if (StringRef("{flags}").equals_lower(Constraint)) {
10206 Res.first = X86::EFLAGS;
10207 Res.second = X86::CCRRegisterClass;
10211 // 'A' means EAX + EDX.
10212 if (Constraint == "A") {
10213 Res.first = X86::EAX;
10214 Res.second = X86::GR32_ADRegisterClass;
10220 // Otherwise, check to see if this is a register class of the wrong value
10221 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10222 // turn into {ax},{dx}.
10223 if (Res.second->hasType(VT))
10224 return Res; // Correct type already, nothing to do.
10226 // All of the single-register GCC register classes map their values onto
10227 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10228 // really want an 8-bit or 32-bit register, map to the appropriate register
10229 // class and return the appropriate register.
10230 if (Res.second == X86::GR16RegisterClass) {
10231 if (VT == MVT::i8) {
10232 unsigned DestReg = 0;
10233 switch (Res.first) {
10235 case X86::AX: DestReg = X86::AL; break;
10236 case X86::DX: DestReg = X86::DL; break;
10237 case X86::CX: DestReg = X86::CL; break;
10238 case X86::BX: DestReg = X86::BL; break;
10241 Res.first = DestReg;
10242 Res.second = X86::GR8RegisterClass;
10244 } else if (VT == MVT::i32) {
10245 unsigned DestReg = 0;
10246 switch (Res.first) {
10248 case X86::AX: DestReg = X86::EAX; break;
10249 case X86::DX: DestReg = X86::EDX; break;
10250 case X86::CX: DestReg = X86::ECX; break;
10251 case X86::BX: DestReg = X86::EBX; break;
10252 case X86::SI: DestReg = X86::ESI; break;
10253 case X86::DI: DestReg = X86::EDI; break;
10254 case X86::BP: DestReg = X86::EBP; break;
10255 case X86::SP: DestReg = X86::ESP; break;
10258 Res.first = DestReg;
10259 Res.second = X86::GR32RegisterClass;
10261 } else if (VT == MVT::i64) {
10262 unsigned DestReg = 0;
10263 switch (Res.first) {
10265 case X86::AX: DestReg = X86::RAX; break;
10266 case X86::DX: DestReg = X86::RDX; break;
10267 case X86::CX: DestReg = X86::RCX; break;
10268 case X86::BX: DestReg = X86::RBX; break;
10269 case X86::SI: DestReg = X86::RSI; break;
10270 case X86::DI: DestReg = X86::RDI; break;
10271 case X86::BP: DestReg = X86::RBP; break;
10272 case X86::SP: DestReg = X86::RSP; break;
10275 Res.first = DestReg;
10276 Res.second = X86::GR64RegisterClass;
10279 } else if (Res.second == X86::FR32RegisterClass ||
10280 Res.second == X86::FR64RegisterClass ||
10281 Res.second == X86::VR128RegisterClass) {
10282 // Handle references to XMM physical registers that got mapped into the
10283 // wrong class. This can happen with constraints like {xmm0} where the
10284 // target independent register mapper will just pick the first match it can
10285 // find, ignoring the required type.
10286 if (VT == MVT::f32)
10287 Res.second = X86::FR32RegisterClass;
10288 else if (VT == MVT::f64)
10289 Res.second = X86::FR64RegisterClass;
10290 else if (X86::VR128RegisterClass->hasType(VT))
10291 Res.second = X86::VR128RegisterClass;