1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalAlias.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/LLVMContext.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/Dwarf.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/raw_ostream.h"
53 using namespace dwarf;
55 STATISTIC(NumTailCalls, "Number of tail calls");
58 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
60 // Forward declarations.
61 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
64 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
65 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
66 default: llvm_unreachable("unknown subtarget type");
67 case X86Subtarget::isDarwin:
68 if (TM.getSubtarget<X86Subtarget>().is64Bit())
69 return new X8664_MachoTargetObjectFile();
70 return new TargetLoweringObjectFileMachO();
71 case X86Subtarget::isELF:
72 if (TM.getSubtarget<X86Subtarget>().is64Bit())
73 return new X8664_ELFTargetObjectFile(TM);
74 return new X8632_ELFTargetObjectFile(TM);
75 case X86Subtarget::isMingw:
76 case X86Subtarget::isCygwin:
77 case X86Subtarget::isWindows:
78 return new TargetLoweringObjectFileCOFF();
82 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
83 : TargetLowering(TM, createTLOF(TM)) {
84 Subtarget = &TM.getSubtarget<X86Subtarget>();
85 X86ScalarSSEf64 = Subtarget->hasSSE2();
86 X86ScalarSSEf32 = Subtarget->hasSSE1();
87 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
89 RegInfo = TM.getRegisterInfo();
92 // Set up the TargetLowering object.
94 // X86 is weird, it always uses i8 for shift amounts and setcc results.
95 setShiftAmountType(MVT::i8);
96 setBooleanContents(ZeroOrOneBooleanContent);
97 setSchedulingPreference(Sched::RegPressure);
98 setStackPointerRegisterToSaveRestore(X86StackPtr);
100 if (Subtarget->isTargetDarwin()) {
101 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
102 setUseUnderscoreSetJmp(false);
103 setUseUnderscoreLongJmp(false);
104 } else if (Subtarget->isTargetMingw()) {
105 // MS runtime is weird: it exports _setjmp, but longjmp!
106 setUseUnderscoreSetJmp(true);
107 setUseUnderscoreLongJmp(false);
109 setUseUnderscoreSetJmp(true);
110 setUseUnderscoreLongJmp(true);
113 // Set up the register classes.
114 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
115 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
116 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
117 if (Subtarget->is64Bit())
118 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
120 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
122 // We don't accept any truncstore of integer registers.
123 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
124 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
125 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
127 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
128 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
130 // SETOEQ and SETUNE require checking two conditions.
131 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
132 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
133 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
135 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
136 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
138 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
140 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
141 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
142 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
144 if (Subtarget->is64Bit()) {
145 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
146 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
147 } else if (!UseSoftFloat) {
148 // We have an algorithm for SSE2->double, and we turn this into a
149 // 64-bit FILD followed by conditional FADD for other targets.
150 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
151 // We have an algorithm for SSE2, and we turn this into a 64-bit
152 // FILD for other targets.
153 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
156 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
158 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
159 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
162 // SSE has no i16 to fp conversion, only i32
163 if (X86ScalarSSEf32) {
164 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
165 // f32 and f64 cases are Legal, f80 case is not
166 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
168 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
169 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
172 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
173 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
176 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
177 // are Legal, f80 is custom lowered.
178 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
179 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
181 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
183 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
184 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
186 if (X86ScalarSSEf32) {
187 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
188 // f32 and f64 cases are Legal, f80 case is not
189 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
191 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
192 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
195 // Handle FP_TO_UINT by promoting the destination to a larger signed
197 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
198 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
199 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
201 if (Subtarget->is64Bit()) {
202 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
203 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
204 } else if (!UseSoftFloat) {
205 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
206 // Expand FP_TO_UINT into a select.
207 // FIXME: We would like to use a Custom expander here eventually to do
208 // the optimal thing for SSE vs. the default expansion in the legalizer.
209 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
211 // With SSE3 we can use fisttpll to convert to a signed i64; without
212 // SSE, we're stuck with a fistpll.
213 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
216 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
217 if (!X86ScalarSSEf64) {
218 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
219 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
220 if (Subtarget->is64Bit()) {
221 setOperationAction(ISD::BIT_CONVERT , MVT::f64 , Expand);
222 // Without SSE, i64->f64 goes through memory; i64->MMX is Legal.
223 if (Subtarget->hasMMX() && !DisableMMX)
224 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Custom);
226 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Expand);
230 // Scalar integer divide and remainder are lowered to use operations that
231 // produce two results, to match the available instructions. This exposes
232 // the two-result form to trivial CSE, which is able to combine x/y and x%y
233 // into a single instruction.
235 // Scalar integer multiply-high is also lowered to use two-result
236 // operations, to match the available instructions. However, plain multiply
237 // (low) operations are left as Legal, as there are single-result
238 // instructions for this in x86. Using the two-result multiply instructions
239 // when both high and low results are needed must be arranged by dagcombine.
240 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
241 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
242 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
243 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
244 setOperationAction(ISD::SREM , MVT::i8 , Expand);
245 setOperationAction(ISD::UREM , MVT::i8 , Expand);
246 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
247 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
248 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
249 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
250 setOperationAction(ISD::SREM , MVT::i16 , Expand);
251 setOperationAction(ISD::UREM , MVT::i16 , Expand);
252 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
253 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
254 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
255 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
256 setOperationAction(ISD::SREM , MVT::i32 , Expand);
257 setOperationAction(ISD::UREM , MVT::i32 , Expand);
258 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
259 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
260 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
261 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
262 setOperationAction(ISD::SREM , MVT::i64 , Expand);
263 setOperationAction(ISD::UREM , MVT::i64 , Expand);
265 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
266 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
267 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
268 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
269 if (Subtarget->is64Bit())
270 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
272 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
273 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
274 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
275 setOperationAction(ISD::FREM , MVT::f32 , Expand);
276 setOperationAction(ISD::FREM , MVT::f64 , Expand);
277 setOperationAction(ISD::FREM , MVT::f80 , Expand);
278 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
280 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
281 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
282 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
283 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
284 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
285 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
286 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
287 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
288 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
289 if (Subtarget->is64Bit()) {
290 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
291 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
292 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
295 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
296 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
298 // These should be promoted to a larger select which is supported.
299 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
300 // X86 wants to expand cmov itself.
301 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
302 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
303 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
304 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
305 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
306 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
307 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
308 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
309 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
310 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
311 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
312 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
313 if (Subtarget->is64Bit()) {
314 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
315 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
317 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
320 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
321 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
322 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
323 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
324 if (Subtarget->is64Bit())
325 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
326 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
327 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
328 if (Subtarget->is64Bit()) {
329 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
330 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
331 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
332 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
333 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
335 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
336 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
337 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
338 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
339 if (Subtarget->is64Bit()) {
340 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
341 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
342 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
345 if (Subtarget->hasSSE1())
346 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
348 if (!Subtarget->hasSSE2())
349 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
351 // Expand certain atomics
352 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
353 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
354 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
355 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
357 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
358 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
359 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
360 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
362 if (!Subtarget->is64Bit()) {
363 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
365 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
366 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
367 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
372 // FIXME - use subtarget debug flags
373 if (!Subtarget->isTargetDarwin() &&
374 !Subtarget->isTargetELF() &&
375 !Subtarget->isTargetCygMing()) {
376 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
379 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
380 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
381 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
382 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
383 if (Subtarget->is64Bit()) {
384 setExceptionPointerRegister(X86::RAX);
385 setExceptionSelectorRegister(X86::RDX);
387 setExceptionPointerRegister(X86::EAX);
388 setExceptionSelectorRegister(X86::EDX);
390 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
391 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
393 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
395 setOperationAction(ISD::TRAP, MVT::Other, Legal);
397 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
398 setOperationAction(ISD::VASTART , MVT::Other, Custom);
399 setOperationAction(ISD::VAEND , MVT::Other, Expand);
400 if (Subtarget->is64Bit()) {
401 setOperationAction(ISD::VAARG , MVT::Other, Custom);
402 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
404 setOperationAction(ISD::VAARG , MVT::Other, Expand);
405 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
408 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
409 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
410 if (Subtarget->is64Bit())
411 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
412 if (Subtarget->isTargetCygMing())
413 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
415 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
417 if (!UseSoftFloat && X86ScalarSSEf64) {
418 // f32 and f64 use SSE.
419 // Set up the FP register classes.
420 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
421 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
423 // Use ANDPD to simulate FABS.
424 setOperationAction(ISD::FABS , MVT::f64, Custom);
425 setOperationAction(ISD::FABS , MVT::f32, Custom);
427 // Use XORP to simulate FNEG.
428 setOperationAction(ISD::FNEG , MVT::f64, Custom);
429 setOperationAction(ISD::FNEG , MVT::f32, Custom);
431 // Use ANDPD and ORPD to simulate FCOPYSIGN.
432 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
433 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
435 // We don't support sin/cos/fmod
436 setOperationAction(ISD::FSIN , MVT::f64, Expand);
437 setOperationAction(ISD::FCOS , MVT::f64, Expand);
438 setOperationAction(ISD::FSIN , MVT::f32, Expand);
439 setOperationAction(ISD::FCOS , MVT::f32, Expand);
441 // Expand FP immediates into loads from the stack, except for the special
443 addLegalFPImmediate(APFloat(+0.0)); // xorpd
444 addLegalFPImmediate(APFloat(+0.0f)); // xorps
445 } else if (!UseSoftFloat && X86ScalarSSEf32) {
446 // Use SSE for f32, x87 for f64.
447 // Set up the FP register classes.
448 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
449 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
451 // Use ANDPS to simulate FABS.
452 setOperationAction(ISD::FABS , MVT::f32, Custom);
454 // Use XORP to simulate FNEG.
455 setOperationAction(ISD::FNEG , MVT::f32, Custom);
457 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
459 // Use ANDPS and ORPS to simulate FCOPYSIGN.
460 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
461 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
463 // We don't support sin/cos/fmod
464 setOperationAction(ISD::FSIN , MVT::f32, Expand);
465 setOperationAction(ISD::FCOS , MVT::f32, Expand);
467 // Special cases we handle for FP constants.
468 addLegalFPImmediate(APFloat(+0.0f)); // xorps
469 addLegalFPImmediate(APFloat(+0.0)); // FLD0
470 addLegalFPImmediate(APFloat(+1.0)); // FLD1
471 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
472 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
475 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
476 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
478 } else if (!UseSoftFloat) {
479 // f32 and f64 in x87.
480 // Set up the FP register classes.
481 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
482 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
484 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
485 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
486 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
487 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
490 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
491 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
493 addLegalFPImmediate(APFloat(+0.0)); // FLD0
494 addLegalFPImmediate(APFloat(+1.0)); // FLD1
495 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
496 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
497 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
498 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
499 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
500 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
503 // Long double always uses X87.
505 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
506 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
507 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
510 APFloat TmpFlt(+0.0);
511 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
513 addLegalFPImmediate(TmpFlt); // FLD0
515 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
516 APFloat TmpFlt2(+1.0);
517 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
519 addLegalFPImmediate(TmpFlt2); // FLD1
520 TmpFlt2.changeSign();
521 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
525 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
526 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
530 // Always use a library call for pow.
531 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
532 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
533 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
535 setOperationAction(ISD::FLOG, MVT::f80, Expand);
536 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
537 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
538 setOperationAction(ISD::FEXP, MVT::f80, Expand);
539 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
541 // First set operation action for all vector types to either promote
542 // (for widening) or expand (for scalarization). Then we will selectively
543 // turn on ones that can be effectively codegen'd.
544 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
545 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
546 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
561 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
562 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
595 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
599 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
600 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
601 setTruncStoreAction((MVT::SimpleValueType)VT,
602 (MVT::SimpleValueType)InnerVT, Expand);
603 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
604 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
605 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
608 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
609 // with -msoft-float, disable use of MMX as well.
610 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
611 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
612 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
613 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
614 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass, false);
615 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
617 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
618 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
619 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
620 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
622 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
623 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
624 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
625 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
627 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
628 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
630 setOperationAction(ISD::AND, MVT::v8i8, Promote);
631 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
632 setOperationAction(ISD::AND, MVT::v4i16, Promote);
633 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
634 setOperationAction(ISD::AND, MVT::v2i32, Promote);
635 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
636 setOperationAction(ISD::AND, MVT::v1i64, Legal);
638 setOperationAction(ISD::OR, MVT::v8i8, Promote);
639 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
640 setOperationAction(ISD::OR, MVT::v4i16, Promote);
641 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
642 setOperationAction(ISD::OR, MVT::v2i32, Promote);
643 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
644 setOperationAction(ISD::OR, MVT::v1i64, Legal);
646 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
647 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
648 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
649 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
650 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
651 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
652 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
654 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
655 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
656 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
657 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
658 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
659 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
660 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
661 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
662 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
664 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
665 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
666 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
667 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
670 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
671 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
672 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
675 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
676 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
677 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
678 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
680 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
682 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
683 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
684 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
685 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
686 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
690 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
691 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
692 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
693 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
694 setOperationAction(ISD::BIT_CONVERT, MVT::v2f32, Custom);
695 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
699 if (!UseSoftFloat && Subtarget->hasSSE1()) {
700 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
702 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
703 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
704 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
705 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
706 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
707 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
708 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
709 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
710 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
711 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
712 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
713 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
716 if (!UseSoftFloat && Subtarget->hasSSE2()) {
717 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
719 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
720 // registers cannot be used even for integer operations.
721 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
722 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
723 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
724 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
726 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
727 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
728 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
729 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
730 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
731 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
732 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
733 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
734 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
735 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
736 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
737 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
738 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
739 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
740 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
741 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
743 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
744 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
745 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
746 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
748 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
749 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
750 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
751 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
752 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
754 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
755 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
756 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
757 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
760 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
761 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
762 EVT VT = (MVT::SimpleValueType)i;
763 // Do not attempt to custom lower non-power-of-2 vectors
764 if (!isPowerOf2_32(VT.getVectorNumElements()))
766 // Do not attempt to custom lower non-128-bit vectors
767 if (!VT.is128BitVector())
769 setOperationAction(ISD::BUILD_VECTOR,
770 VT.getSimpleVT().SimpleTy, Custom);
771 setOperationAction(ISD::VECTOR_SHUFFLE,
772 VT.getSimpleVT().SimpleTy, Custom);
773 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
774 VT.getSimpleVT().SimpleTy, Custom);
777 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
778 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
779 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
780 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
781 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
782 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
784 if (Subtarget->is64Bit()) {
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
789 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
790 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
791 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
794 // Do not attempt to promote non-128-bit vectors
795 if (!VT.is128BitVector()) {
799 setOperationAction(ISD::AND, SVT, Promote);
800 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
801 setOperationAction(ISD::OR, SVT, Promote);
802 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
803 setOperationAction(ISD::XOR, SVT, Promote);
804 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
805 setOperationAction(ISD::LOAD, SVT, Promote);
806 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
807 setOperationAction(ISD::SELECT, SVT, Promote);
808 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
811 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
813 // Custom lower v2i64 and v2f64 selects.
814 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
815 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
816 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
817 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
819 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
820 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
821 if (!DisableMMX && Subtarget->hasMMX()) {
822 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
823 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
827 if (Subtarget->hasSSE41()) {
828 // FIXME: Do we need to handle scalar-to-vector here?
829 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
831 // i8 and i16 vectors are custom , because the source register and source
832 // source memory operand types are not the same width. f32 vectors are
833 // custom since the immediate controlling the insert encodes additional
835 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
836 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
837 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
838 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
840 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
841 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
842 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
843 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
845 if (Subtarget->is64Bit()) {
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
847 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
851 if (Subtarget->hasSSE42()) {
852 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
855 if (!UseSoftFloat && Subtarget->hasAVX()) {
856 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
857 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
858 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
859 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
861 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
862 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
863 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
864 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
865 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
866 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
867 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
868 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
869 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
870 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
871 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
872 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
873 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
874 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
875 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
877 // Operations to consider commented out -v16i16 v32i8
878 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
879 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
880 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
881 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
882 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
883 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
884 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
885 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
886 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
887 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
888 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
889 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
890 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
891 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
893 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
894 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
895 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
896 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
898 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
899 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
900 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
901 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
902 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
904 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
905 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
906 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
907 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
908 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
909 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
912 // Not sure we want to do this since there are no 256-bit integer
915 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
916 // This includes 256-bit vectors
917 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
918 EVT VT = (MVT::SimpleValueType)i;
920 // Do not attempt to custom lower non-power-of-2 vectors
921 if (!isPowerOf2_32(VT.getVectorNumElements()))
924 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
925 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
926 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
929 if (Subtarget->is64Bit()) {
930 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
931 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
936 // Not sure we want to do this since there are no 256-bit integer
939 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
940 // Including 256-bit vectors
941 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
942 EVT VT = (MVT::SimpleValueType)i;
944 if (!VT.is256BitVector()) {
947 setOperationAction(ISD::AND, VT, Promote);
948 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
949 setOperationAction(ISD::OR, VT, Promote);
950 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
951 setOperationAction(ISD::XOR, VT, Promote);
952 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
953 setOperationAction(ISD::LOAD, VT, Promote);
954 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
955 setOperationAction(ISD::SELECT, VT, Promote);
956 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
959 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
963 // We want to custom lower some of our intrinsics.
964 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
966 // Add/Sub/Mul with overflow operations are custom lowered.
967 setOperationAction(ISD::SADDO, MVT::i32, Custom);
968 setOperationAction(ISD::SADDO, MVT::i64, Custom);
969 setOperationAction(ISD::UADDO, MVT::i32, Custom);
970 setOperationAction(ISD::UADDO, MVT::i64, Custom);
971 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
972 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
973 setOperationAction(ISD::USUBO, MVT::i32, Custom);
974 setOperationAction(ISD::USUBO, MVT::i64, Custom);
975 setOperationAction(ISD::SMULO, MVT::i32, Custom);
976 setOperationAction(ISD::SMULO, MVT::i64, Custom);
978 if (!Subtarget->is64Bit()) {
979 // These libcalls are not available in 32-bit.
980 setLibcallName(RTLIB::SHL_I128, 0);
981 setLibcallName(RTLIB::SRL_I128, 0);
982 setLibcallName(RTLIB::SRA_I128, 0);
985 // We have target-specific dag combine patterns for the following nodes:
986 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
987 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
988 setTargetDAGCombine(ISD::BUILD_VECTOR);
989 setTargetDAGCombine(ISD::SELECT);
990 setTargetDAGCombine(ISD::SHL);
991 setTargetDAGCombine(ISD::SRA);
992 setTargetDAGCombine(ISD::SRL);
993 setTargetDAGCombine(ISD::OR);
994 setTargetDAGCombine(ISD::STORE);
995 setTargetDAGCombine(ISD::MEMBARRIER);
996 setTargetDAGCombine(ISD::ZERO_EXTEND);
997 if (Subtarget->is64Bit())
998 setTargetDAGCombine(ISD::MUL);
1000 computeRegisterProperties();
1002 // FIXME: These should be based on subtarget info. Plus, the values should
1003 // be smaller when we are in optimizing for size mode.
1004 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1005 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1006 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1007 setPrefLoopAlignment(16);
1008 benefitFromCodePlacementOpt = true;
1012 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1017 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1018 /// the desired ByVal argument alignment.
1019 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1022 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1023 if (VTy->getBitWidth() == 128)
1025 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1026 unsigned EltAlign = 0;
1027 getMaxByValAlign(ATy->getElementType(), EltAlign);
1028 if (EltAlign > MaxAlign)
1029 MaxAlign = EltAlign;
1030 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1031 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1032 unsigned EltAlign = 0;
1033 getMaxByValAlign(STy->getElementType(i), EltAlign);
1034 if (EltAlign > MaxAlign)
1035 MaxAlign = EltAlign;
1043 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1044 /// function arguments in the caller parameter area. For X86, aggregates
1045 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1046 /// are at 4-byte boundaries.
1047 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1048 if (Subtarget->is64Bit()) {
1049 // Max of 8 and alignment of type.
1050 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1057 if (Subtarget->hasSSE1())
1058 getMaxByValAlign(Ty, Align);
1062 /// getOptimalMemOpType - Returns the target specific optimal type for load
1063 /// and store operations as a result of memset, memcpy, and memmove
1064 /// lowering. If DstAlign is zero that means it's safe to destination
1065 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1066 /// means there isn't a need to check it against alignment requirement,
1067 /// probably because the source does not need to be loaded. If
1068 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1069 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1070 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1071 /// constant so it does not need to be loaded.
1072 /// It returns EVT::Other if the type should be determined using generic
1073 /// target-independent logic.
1075 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1076 unsigned DstAlign, unsigned SrcAlign,
1077 bool NonScalarIntSafe,
1079 MachineFunction &MF) const {
1080 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1081 // linux. This is because the stack realignment code can't handle certain
1082 // cases like PR2962. This should be removed when PR2962 is fixed.
1083 const Function *F = MF.getFunction();
1084 if (NonScalarIntSafe &&
1085 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1087 (Subtarget->isUnalignedMemAccessFast() ||
1088 ((DstAlign == 0 || DstAlign >= 16) &&
1089 (SrcAlign == 0 || SrcAlign >= 16))) &&
1090 Subtarget->getStackAlignment() >= 16) {
1091 if (Subtarget->hasSSE2())
1093 if (Subtarget->hasSSE1())
1095 } else if (!MemcpyStrSrc && Size >= 8 &&
1096 !Subtarget->is64Bit() &&
1097 Subtarget->getStackAlignment() >= 8 &&
1098 Subtarget->hasSSE2()) {
1099 // Do not use f64 to lower memcpy if source is string constant. It's
1100 // better to use i32 to avoid the loads.
1104 if (Subtarget->is64Bit() && Size >= 8)
1109 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1110 /// current function. The returned value is a member of the
1111 /// MachineJumpTableInfo::JTEntryKind enum.
1112 unsigned X86TargetLowering::getJumpTableEncoding() const {
1113 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1115 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1116 Subtarget->isPICStyleGOT())
1117 return MachineJumpTableInfo::EK_Custom32;
1119 // Otherwise, use the normal jump table encoding heuristics.
1120 return TargetLowering::getJumpTableEncoding();
1123 /// getPICBaseSymbol - Return the X86-32 PIC base.
1125 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1126 MCContext &Ctx) const {
1127 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1128 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1129 Twine(MF->getFunctionNumber())+"$pb");
1134 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1135 const MachineBasicBlock *MBB,
1136 unsigned uid,MCContext &Ctx) const{
1137 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1138 Subtarget->isPICStyleGOT());
1139 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1141 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1142 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1145 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1147 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1148 SelectionDAG &DAG) const {
1149 if (!Subtarget->is64Bit())
1150 // This doesn't have DebugLoc associated with it, but is not really the
1151 // same as a Register.
1152 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1156 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1157 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1159 const MCExpr *X86TargetLowering::
1160 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1161 MCContext &Ctx) const {
1162 // X86-64 uses RIP relative addressing based on the jump table label.
1163 if (Subtarget->isPICStyleRIPRel())
1164 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1166 // Otherwise, the reference is relative to the PIC base.
1167 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1170 /// getFunctionAlignment - Return the Log2 alignment of this function.
1171 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1172 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1175 //===----------------------------------------------------------------------===//
1176 // Return Value Calling Convention Implementation
1177 //===----------------------------------------------------------------------===//
1179 #include "X86GenCallingConv.inc"
1182 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1183 const SmallVectorImpl<EVT> &OutTys,
1184 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1185 SelectionDAG &DAG) const {
1186 SmallVector<CCValAssign, 16> RVLocs;
1187 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1188 RVLocs, *DAG.getContext());
1189 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1193 X86TargetLowering::LowerReturn(SDValue Chain,
1194 CallingConv::ID CallConv, bool isVarArg,
1195 const SmallVectorImpl<ISD::OutputArg> &Outs,
1196 DebugLoc dl, SelectionDAG &DAG) const {
1197 MachineFunction &MF = DAG.getMachineFunction();
1198 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1200 SmallVector<CCValAssign, 16> RVLocs;
1201 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1202 RVLocs, *DAG.getContext());
1203 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1205 // Add the regs to the liveout set for the function.
1206 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1207 for (unsigned i = 0; i != RVLocs.size(); ++i)
1208 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1209 MRI.addLiveOut(RVLocs[i].getLocReg());
1213 SmallVector<SDValue, 6> RetOps;
1214 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1215 // Operand #1 = Bytes To Pop
1216 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1219 // Copy the result values into the output registers.
1220 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1221 CCValAssign &VA = RVLocs[i];
1222 assert(VA.isRegLoc() && "Can only return in registers!");
1223 SDValue ValToCopy = Outs[i].Val;
1225 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1226 // the RET instruction and handled by the FP Stackifier.
1227 if (VA.getLocReg() == X86::ST0 ||
1228 VA.getLocReg() == X86::ST1) {
1229 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1230 // change the value to the FP stack register class.
1231 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1232 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1233 RetOps.push_back(ValToCopy);
1234 // Don't emit a copytoreg.
1238 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1239 // which is returned in RAX / RDX.
1240 if (Subtarget->is64Bit()) {
1241 EVT ValVT = ValToCopy.getValueType();
1242 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1243 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1244 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1245 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1249 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1250 Flag = Chain.getValue(1);
1253 // The x86-64 ABI for returning structs by value requires that we copy
1254 // the sret argument into %rax for the return. We saved the argument into
1255 // a virtual register in the entry block, so now we copy the value out
1257 if (Subtarget->is64Bit() &&
1258 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1259 MachineFunction &MF = DAG.getMachineFunction();
1260 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1261 unsigned Reg = FuncInfo->getSRetReturnReg();
1263 "SRetReturnReg should have been set in LowerFormalArguments().");
1264 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1266 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1267 Flag = Chain.getValue(1);
1269 // RAX now acts like a return value.
1270 MRI.addLiveOut(X86::RAX);
1273 RetOps[0] = Chain; // Update chain.
1275 // Add the flag if we have it.
1277 RetOps.push_back(Flag);
1279 return DAG.getNode(X86ISD::RET_FLAG, dl,
1280 MVT::Other, &RetOps[0], RetOps.size());
1283 /// LowerCallResult - Lower the result values of a call into the
1284 /// appropriate copies out of appropriate physical registers.
1287 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1288 CallingConv::ID CallConv, bool isVarArg,
1289 const SmallVectorImpl<ISD::InputArg> &Ins,
1290 DebugLoc dl, SelectionDAG &DAG,
1291 SmallVectorImpl<SDValue> &InVals) const {
1293 // Assign locations to each value returned by this call.
1294 SmallVector<CCValAssign, 16> RVLocs;
1295 bool Is64Bit = Subtarget->is64Bit();
1296 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1297 RVLocs, *DAG.getContext());
1298 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1300 // Copy all of the result registers out of their specified physreg.
1301 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1302 CCValAssign &VA = RVLocs[i];
1303 EVT CopyVT = VA.getValVT();
1305 // If this is x86-64, and we disabled SSE, we can't return FP values
1306 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1307 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1308 report_fatal_error("SSE register return with SSE disabled");
1311 // If this is a call to a function that returns an fp value on the floating
1312 // point stack, but where we prefer to use the value in xmm registers, copy
1313 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1314 if ((VA.getLocReg() == X86::ST0 ||
1315 VA.getLocReg() == X86::ST1) &&
1316 isScalarFPTypeInSSEReg(VA.getValVT())) {
1321 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1322 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1323 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1324 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1325 MVT::v2i64, InFlag).getValue(1);
1326 Val = Chain.getValue(0);
1327 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1328 Val, DAG.getConstant(0, MVT::i64));
1330 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1331 MVT::i64, InFlag).getValue(1);
1332 Val = Chain.getValue(0);
1334 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1336 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1337 CopyVT, InFlag).getValue(1);
1338 Val = Chain.getValue(0);
1340 InFlag = Chain.getValue(2);
1342 if (CopyVT != VA.getValVT()) {
1343 // Round the F80 the right size, which also moves to the appropriate xmm
1345 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1346 // This truncation won't change the value.
1347 DAG.getIntPtrConstant(1));
1350 InVals.push_back(Val);
1357 //===----------------------------------------------------------------------===//
1358 // C & StdCall & Fast Calling Convention implementation
1359 //===----------------------------------------------------------------------===//
1360 // StdCall calling convention seems to be standard for many Windows' API
1361 // routines and around. It differs from C calling convention just a little:
1362 // callee should clean up the stack, not caller. Symbols should be also
1363 // decorated in some fancy way :) It doesn't support any vector arguments.
1364 // For info on fast calling convention see Fast Calling Convention (tail call)
1365 // implementation LowerX86_32FastCCCallTo.
1367 /// CallIsStructReturn - Determines whether a call uses struct return
1369 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1373 return Outs[0].Flags.isSRet();
1376 /// ArgsAreStructReturn - Determines whether a function uses struct
1377 /// return semantics.
1379 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1383 return Ins[0].Flags.isSRet();
1386 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1387 /// given CallingConvention value.
1388 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1389 if (Subtarget->is64Bit()) {
1390 if (CC == CallingConv::GHC)
1391 return CC_X86_64_GHC;
1392 else if (Subtarget->isTargetWin64())
1393 return CC_X86_Win64_C;
1398 if (CC == CallingConv::X86_FastCall)
1399 return CC_X86_32_FastCall;
1400 else if (CC == CallingConv::X86_ThisCall)
1401 return CC_X86_32_ThisCall;
1402 else if (CC == CallingConv::Fast)
1403 return CC_X86_32_FastCC;
1404 else if (CC == CallingConv::GHC)
1405 return CC_X86_32_GHC;
1410 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1411 /// by "Src" to address "Dst" with size and alignment information specified by
1412 /// the specific parameter attribute. The copy will be passed as a byval
1413 /// function parameter.
1415 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1416 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1418 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1419 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1420 /*isVolatile*/false, /*AlwaysInline=*/true,
1424 /// IsTailCallConvention - Return true if the calling convention is one that
1425 /// supports tail call optimization.
1426 static bool IsTailCallConvention(CallingConv::ID CC) {
1427 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1430 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1431 /// a tailcall target by changing its ABI.
1432 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1433 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1437 X86TargetLowering::LowerMemArgument(SDValue Chain,
1438 CallingConv::ID CallConv,
1439 const SmallVectorImpl<ISD::InputArg> &Ins,
1440 DebugLoc dl, SelectionDAG &DAG,
1441 const CCValAssign &VA,
1442 MachineFrameInfo *MFI,
1444 // Create the nodes corresponding to a load from this parameter slot.
1445 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1446 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1447 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1450 // If value is passed by pointer we have address passed instead of the value
1452 if (VA.getLocInfo() == CCValAssign::Indirect)
1453 ValVT = VA.getLocVT();
1455 ValVT = VA.getValVT();
1457 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1458 // changed with more analysis.
1459 // In case of tail call optimization mark all arguments mutable. Since they
1460 // could be overwritten by lowering of arguments in case of a tail call.
1461 if (Flags.isByVal()) {
1462 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1463 VA.getLocMemOffset(), isImmutable, false);
1464 return DAG.getFrameIndex(FI, getPointerTy());
1466 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1467 VA.getLocMemOffset(), isImmutable, false);
1468 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1469 return DAG.getLoad(ValVT, dl, Chain, FIN,
1470 PseudoSourceValue::getFixedStack(FI), 0,
1476 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1477 CallingConv::ID CallConv,
1479 const SmallVectorImpl<ISD::InputArg> &Ins,
1482 SmallVectorImpl<SDValue> &InVals)
1484 MachineFunction &MF = DAG.getMachineFunction();
1485 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1487 const Function* Fn = MF.getFunction();
1488 if (Fn->hasExternalLinkage() &&
1489 Subtarget->isTargetCygMing() &&
1490 Fn->getName() == "main")
1491 FuncInfo->setForceFramePointer(true);
1493 MachineFrameInfo *MFI = MF.getFrameInfo();
1494 bool Is64Bit = Subtarget->is64Bit();
1495 bool IsWin64 = Subtarget->isTargetWin64();
1497 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1498 "Var args not supported with calling convention fastcc or ghc");
1500 // Assign locations to all of the incoming arguments.
1501 SmallVector<CCValAssign, 16> ArgLocs;
1502 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1503 ArgLocs, *DAG.getContext());
1504 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1506 unsigned LastVal = ~0U;
1508 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1509 CCValAssign &VA = ArgLocs[i];
1510 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1512 assert(VA.getValNo() != LastVal &&
1513 "Don't support value assigned to multiple locs yet");
1514 LastVal = VA.getValNo();
1516 if (VA.isRegLoc()) {
1517 EVT RegVT = VA.getLocVT();
1518 TargetRegisterClass *RC = NULL;
1519 if (RegVT == MVT::i32)
1520 RC = X86::GR32RegisterClass;
1521 else if (Is64Bit && RegVT == MVT::i64)
1522 RC = X86::GR64RegisterClass;
1523 else if (RegVT == MVT::f32)
1524 RC = X86::FR32RegisterClass;
1525 else if (RegVT == MVT::f64)
1526 RC = X86::FR64RegisterClass;
1527 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1528 RC = X86::VR128RegisterClass;
1529 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1530 RC = X86::VR64RegisterClass;
1532 llvm_unreachable("Unknown argument type!");
1534 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1535 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1537 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1538 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1540 if (VA.getLocInfo() == CCValAssign::SExt)
1541 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1542 DAG.getValueType(VA.getValVT()));
1543 else if (VA.getLocInfo() == CCValAssign::ZExt)
1544 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1545 DAG.getValueType(VA.getValVT()));
1546 else if (VA.getLocInfo() == CCValAssign::BCvt)
1547 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1549 if (VA.isExtInLoc()) {
1550 // Handle MMX values passed in XMM regs.
1551 if (RegVT.isVector()) {
1552 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1553 ArgValue, DAG.getConstant(0, MVT::i64));
1554 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1556 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1559 assert(VA.isMemLoc());
1560 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1563 // If value is passed via pointer - do a load.
1564 if (VA.getLocInfo() == CCValAssign::Indirect)
1565 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1568 InVals.push_back(ArgValue);
1571 // The x86-64 ABI for returning structs by value requires that we copy
1572 // the sret argument into %rax for the return. Save the argument into
1573 // a virtual register so that we can access it from the return points.
1574 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1575 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1576 unsigned Reg = FuncInfo->getSRetReturnReg();
1578 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1579 FuncInfo->setSRetReturnReg(Reg);
1581 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1582 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1585 unsigned StackSize = CCInfo.getNextStackOffset();
1586 // Align stack specially for tail calls.
1587 if (FuncIsMadeTailCallSafe(CallConv))
1588 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1590 // If the function takes variable number of arguments, make a frame index for
1591 // the start of the first vararg value... for expansion of llvm.va_start.
1593 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1594 CallConv != CallingConv::X86_ThisCall)) {
1595 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,
1599 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1601 // FIXME: We should really autogenerate these arrays
1602 static const unsigned GPR64ArgRegsWin64[] = {
1603 X86::RCX, X86::RDX, X86::R8, X86::R9
1605 static const unsigned XMMArgRegsWin64[] = {
1606 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1608 static const unsigned GPR64ArgRegs64Bit[] = {
1609 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1611 static const unsigned XMMArgRegs64Bit[] = {
1612 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1613 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1615 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1618 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1619 GPR64ArgRegs = GPR64ArgRegsWin64;
1620 XMMArgRegs = XMMArgRegsWin64;
1622 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1623 GPR64ArgRegs = GPR64ArgRegs64Bit;
1624 XMMArgRegs = XMMArgRegs64Bit;
1626 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1628 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1631 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1632 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1633 "SSE register cannot be used when SSE is disabled!");
1634 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1635 "SSE register cannot be used when SSE is disabled!");
1636 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1637 // Kernel mode asks for SSE to be disabled, so don't push them
1639 TotalNumXMMRegs = 0;
1641 // For X86-64, if there are vararg parameters that are passed via
1642 // registers, then we must store them to their spots on the stack so they
1643 // may be loaded by deferencing the result of va_next.
1644 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1645 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1646 FuncInfo->setRegSaveFrameIndex(
1647 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1650 // Store the integer parameter registers.
1651 SmallVector<SDValue, 8> MemOps;
1652 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1654 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1655 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1656 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1657 DAG.getIntPtrConstant(Offset));
1658 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1659 X86::GR64RegisterClass);
1660 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1662 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1663 PseudoSourceValue::getFixedStack(
1664 FuncInfo->getRegSaveFrameIndex()),
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(
1680 FuncInfo->getRegSaveFrameIndex()));
1681 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1682 FuncInfo->getVarArgsFPOffset()));
1684 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1685 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1686 X86::VR128RegisterClass);
1687 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1688 SaveXMMOps.push_back(Val);
1690 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1692 &SaveXMMOps[0], SaveXMMOps.size()));
1695 if (!MemOps.empty())
1696 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1697 &MemOps[0], MemOps.size());
1701 // Some CCs need callee pop.
1702 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1703 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1705 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1706 // If this is an sret function, the return should pop the hidden pointer.
1707 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1708 FuncInfo->setBytesToPopOnReturn(4);
1712 // RegSaveFrameIndex is X86-64 only.
1713 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1714 if (CallConv == CallingConv::X86_FastCall ||
1715 CallConv == CallingConv::X86_ThisCall)
1716 // fastcc functions can't have varargs.
1717 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1724 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1725 SDValue StackPtr, SDValue Arg,
1726 DebugLoc dl, SelectionDAG &DAG,
1727 const CCValAssign &VA,
1728 ISD::ArgFlagsTy Flags) const {
1729 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1730 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1731 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1732 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1733 if (Flags.isByVal()) {
1734 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1736 return DAG.getStore(Chain, dl, Arg, PtrOff,
1737 PseudoSourceValue::getStack(), LocMemOffset,
1741 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1742 /// optimization is performed and it is required.
1744 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1745 SDValue &OutRetAddr, SDValue Chain,
1746 bool IsTailCall, bool Is64Bit,
1747 int FPDiff, DebugLoc dl) const {
1748 // Adjust the Return address stack slot.
1749 EVT VT = getPointerTy();
1750 OutRetAddr = getReturnAddressFrameIndex(DAG);
1752 // Load the "old" Return address.
1753 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1754 return SDValue(OutRetAddr.getNode(), 1);
1757 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1758 /// optimization is performed and it is required (FPDiff!=0).
1760 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1761 SDValue Chain, SDValue RetAddrFrIdx,
1762 bool Is64Bit, int FPDiff, DebugLoc dl) {
1763 // Store the return address to the appropriate stack slot.
1764 if (!FPDiff) return Chain;
1765 // Calculate the new stack slot for the return address.
1766 int SlotSize = Is64Bit ? 8 : 4;
1767 int NewReturnAddrFI =
1768 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1769 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1770 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1771 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1772 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1778 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1779 CallingConv::ID CallConv, bool isVarArg,
1781 const SmallVectorImpl<ISD::OutputArg> &Outs,
1782 const SmallVectorImpl<ISD::InputArg> &Ins,
1783 DebugLoc dl, SelectionDAG &DAG,
1784 SmallVectorImpl<SDValue> &InVals) const {
1785 MachineFunction &MF = DAG.getMachineFunction();
1786 bool Is64Bit = Subtarget->is64Bit();
1787 bool IsStructRet = CallIsStructReturn(Outs);
1788 bool IsSibcall = false;
1791 // Check if it's really possible to do a tail call.
1792 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1793 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1796 // Sibcalls are automatically detected tailcalls which do not require
1798 if (!GuaranteedTailCallOpt && isTailCall)
1805 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1806 "Var args not supported with calling convention fastcc or ghc");
1808 // Analyze operands of the call, assigning locations to each operand.
1809 SmallVector<CCValAssign, 16> ArgLocs;
1810 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1811 ArgLocs, *DAG.getContext());
1812 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1814 // Get a count of how many bytes are to be pushed on the stack.
1815 unsigned NumBytes = CCInfo.getNextStackOffset();
1817 // This is a sibcall. The memory operands are available in caller's
1818 // own caller's stack.
1820 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1821 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1824 if (isTailCall && !IsSibcall) {
1825 // Lower arguments at fp - stackoffset + fpdiff.
1826 unsigned NumBytesCallerPushed =
1827 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1828 FPDiff = NumBytesCallerPushed - NumBytes;
1830 // Set the delta of movement of the returnaddr stackslot.
1831 // But only set if delta is greater than previous delta.
1832 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1833 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1837 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1839 SDValue RetAddrFrIdx;
1840 // Load return adress for tail calls.
1841 if (isTailCall && FPDiff)
1842 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1843 Is64Bit, FPDiff, dl);
1845 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1846 SmallVector<SDValue, 8> MemOpChains;
1849 // Walk the register/memloc assignments, inserting copies/loads. In the case
1850 // of tail call optimization arguments are handle later.
1851 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1852 CCValAssign &VA = ArgLocs[i];
1853 EVT RegVT = VA.getLocVT();
1854 SDValue Arg = Outs[i].Val;
1855 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1856 bool isByVal = Flags.isByVal();
1858 // Promote the value if needed.
1859 switch (VA.getLocInfo()) {
1860 default: llvm_unreachable("Unknown loc info!");
1861 case CCValAssign::Full: break;
1862 case CCValAssign::SExt:
1863 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1865 case CCValAssign::ZExt:
1866 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1868 case CCValAssign::AExt:
1869 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1870 // Special case: passing MMX values in XMM registers.
1871 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1872 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1873 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1875 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1877 case CCValAssign::BCvt:
1878 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1880 case CCValAssign::Indirect: {
1881 // Store the argument.
1882 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1883 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1884 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1885 PseudoSourceValue::getFixedStack(FI), 0,
1892 if (VA.isRegLoc()) {
1893 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1894 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1895 assert(VA.isMemLoc());
1896 if (StackPtr.getNode() == 0)
1897 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1898 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1899 dl, DAG, VA, Flags));
1903 if (!MemOpChains.empty())
1904 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1905 &MemOpChains[0], MemOpChains.size());
1907 // Build a sequence of copy-to-reg nodes chained together with token chain
1908 // and flag operands which copy the outgoing args into registers.
1910 // Tail call byval lowering might overwrite argument registers so in case of
1911 // tail call optimization the copies to registers are lowered later.
1913 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1914 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1915 RegsToPass[i].second, InFlag);
1916 InFlag = Chain.getValue(1);
1919 if (Subtarget->isPICStyleGOT()) {
1920 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1923 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1924 DAG.getNode(X86ISD::GlobalBaseReg,
1925 DebugLoc(), getPointerTy()),
1927 InFlag = Chain.getValue(1);
1929 // If we are tail calling and generating PIC/GOT style code load the
1930 // address of the callee into ECX. The value in ecx is used as target of
1931 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1932 // for tail calls on PIC/GOT architectures. Normally we would just put the
1933 // address of GOT into ebx and then call target@PLT. But for tail calls
1934 // ebx would be restored (since ebx is callee saved) before jumping to the
1937 // Note: The actual moving to ECX is done further down.
1938 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1939 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1940 !G->getGlobal()->hasProtectedVisibility())
1941 Callee = LowerGlobalAddress(Callee, DAG);
1942 else if (isa<ExternalSymbolSDNode>(Callee))
1943 Callee = LowerExternalSymbol(Callee, DAG);
1947 if (Is64Bit && isVarArg) {
1948 // From AMD64 ABI document:
1949 // For calls that may call functions that use varargs or stdargs
1950 // (prototype-less calls or calls to functions containing ellipsis (...) in
1951 // the declaration) %al is used as hidden argument to specify the number
1952 // of SSE registers used. The contents of %al do not need to match exactly
1953 // the number of registers, but must be an ubound on the number of SSE
1954 // registers used and is in the range 0 - 8 inclusive.
1956 // FIXME: Verify this on Win64
1957 // Count the number of XMM registers allocated.
1958 static const unsigned XMMArgRegs[] = {
1959 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1960 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1962 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1963 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1964 && "SSE registers cannot be used when SSE is disabled");
1966 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1967 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1968 InFlag = Chain.getValue(1);
1972 // For tail calls lower the arguments to the 'real' stack slot.
1974 // Force all the incoming stack arguments to be loaded from the stack
1975 // before any new outgoing arguments are stored to the stack, because the
1976 // outgoing stack slots may alias the incoming argument stack slots, and
1977 // the alias isn't otherwise explicit. This is slightly more conservative
1978 // than necessary, because it means that each store effectively depends
1979 // on every argument instead of just those arguments it would clobber.
1980 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1982 SmallVector<SDValue, 8> MemOpChains2;
1985 // Do not flag preceeding copytoreg stuff together with the following stuff.
1987 if (GuaranteedTailCallOpt) {
1988 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1989 CCValAssign &VA = ArgLocs[i];
1992 assert(VA.isMemLoc());
1993 SDValue Arg = Outs[i].Val;
1994 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1995 // Create frame index.
1996 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1997 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1998 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1999 FIN = DAG.getFrameIndex(FI, getPointerTy());
2001 if (Flags.isByVal()) {
2002 // Copy relative to framepointer.
2003 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2004 if (StackPtr.getNode() == 0)
2005 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2007 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2009 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2013 // Store relative to framepointer.
2014 MemOpChains2.push_back(
2015 DAG.getStore(ArgChain, dl, Arg, FIN,
2016 PseudoSourceValue::getFixedStack(FI), 0,
2022 if (!MemOpChains2.empty())
2023 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2024 &MemOpChains2[0], MemOpChains2.size());
2026 // Copy arguments to their registers.
2027 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2028 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2029 RegsToPass[i].second, InFlag);
2030 InFlag = Chain.getValue(1);
2034 // Store the return address to the appropriate stack slot.
2035 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2039 bool WasGlobalOrExternal = false;
2040 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2041 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2042 // In the 64-bit large code model, we have to make all calls
2043 // through a register, since the call instruction's 32-bit
2044 // pc-relative offset may not be large enough to hold the whole
2046 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2047 WasGlobalOrExternal = true;
2048 // If the callee is a GlobalAddress node (quite common, every direct call
2049 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2052 // We should use extra load for direct calls to dllimported functions in
2054 const GlobalValue *GV = G->getGlobal();
2055 if (!GV->hasDLLImportLinkage()) {
2056 unsigned char OpFlags = 0;
2058 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2059 // external symbols most go through the PLT in PIC mode. If the symbol
2060 // has hidden or protected visibility, or if it is static or local, then
2061 // we don't need to use the PLT - we can directly call it.
2062 if (Subtarget->isTargetELF() &&
2063 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2064 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2065 OpFlags = X86II::MO_PLT;
2066 } else if (Subtarget->isPICStyleStubAny() &&
2067 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2068 Subtarget->getDarwinVers() < 9) {
2069 // PC-relative references to external symbols should go through $stub,
2070 // unless we're building with the leopard linker or later, which
2071 // automatically synthesizes these stubs.
2072 OpFlags = X86II::MO_DARWIN_STUB;
2075 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2076 G->getOffset(), OpFlags);
2078 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2079 WasGlobalOrExternal = true;
2080 unsigned char OpFlags = 0;
2082 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2083 // symbols should go through the PLT.
2084 if (Subtarget->isTargetELF() &&
2085 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2086 OpFlags = X86II::MO_PLT;
2087 } else if (Subtarget->isPICStyleStubAny() &&
2088 Subtarget->getDarwinVers() < 9) {
2089 // PC-relative references to external symbols should go through $stub,
2090 // unless we're building with the leopard linker or later, which
2091 // automatically synthesizes these stubs.
2092 OpFlags = X86II::MO_DARWIN_STUB;
2095 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2099 // Returns a chain & a flag for retval copy to use.
2100 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2101 SmallVector<SDValue, 8> Ops;
2103 if (!IsSibcall && isTailCall) {
2104 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2105 DAG.getIntPtrConstant(0, true), InFlag);
2106 InFlag = Chain.getValue(1);
2109 Ops.push_back(Chain);
2110 Ops.push_back(Callee);
2113 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2115 // Add argument registers to the end of the list so that they are known live
2117 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2118 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2119 RegsToPass[i].second.getValueType()));
2121 // Add an implicit use GOT pointer in EBX.
2122 if (!isTailCall && Subtarget->isPICStyleGOT())
2123 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2125 // Add an implicit use of AL for x86 vararg functions.
2126 if (Is64Bit && isVarArg)
2127 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2129 if (InFlag.getNode())
2130 Ops.push_back(InFlag);
2133 // If this is the first return lowered for this function, add the regs
2134 // to the liveout set for the function.
2135 if (MF.getRegInfo().liveout_empty()) {
2136 SmallVector<CCValAssign, 16> RVLocs;
2137 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2139 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2140 for (unsigned i = 0; i != RVLocs.size(); ++i)
2141 if (RVLocs[i].isRegLoc())
2142 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2144 return DAG.getNode(X86ISD::TC_RETURN, dl,
2145 NodeTys, &Ops[0], Ops.size());
2148 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2149 InFlag = Chain.getValue(1);
2151 // Create the CALLSEQ_END node.
2152 unsigned NumBytesForCalleeToPush;
2153 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2154 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2155 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2156 // If this is a call to a struct-return function, the callee
2157 // pops the hidden struct pointer, so we have to push it back.
2158 // This is common for Darwin/X86, Linux & Mingw32 targets.
2159 NumBytesForCalleeToPush = 4;
2161 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2163 // Returns a flag for retval copy to use.
2165 Chain = DAG.getCALLSEQ_END(Chain,
2166 DAG.getIntPtrConstant(NumBytes, true),
2167 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2170 InFlag = Chain.getValue(1);
2173 // Handle result values, copying them out of physregs into vregs that we
2175 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2176 Ins, dl, DAG, InVals);
2180 //===----------------------------------------------------------------------===//
2181 // Fast Calling Convention (tail call) implementation
2182 //===----------------------------------------------------------------------===//
2184 // Like std call, callee cleans arguments, convention except that ECX is
2185 // reserved for storing the tail called function address. Only 2 registers are
2186 // free for argument passing (inreg). Tail call optimization is performed
2188 // * tailcallopt is enabled
2189 // * caller/callee are fastcc
2190 // On X86_64 architecture with GOT-style position independent code only local
2191 // (within module) calls are supported at the moment.
2192 // To keep the stack aligned according to platform abi the function
2193 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2194 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2195 // If a tail called function callee has more arguments than the caller the
2196 // caller needs to make sure that there is room to move the RETADDR to. This is
2197 // achieved by reserving an area the size of the argument delta right after the
2198 // original REtADDR, but before the saved framepointer or the spilled registers
2199 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2211 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2212 /// for a 16 byte align requirement.
2214 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2215 SelectionDAG& DAG) const {
2216 MachineFunction &MF = DAG.getMachineFunction();
2217 const TargetMachine &TM = MF.getTarget();
2218 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2219 unsigned StackAlignment = TFI.getStackAlignment();
2220 uint64_t AlignMask = StackAlignment - 1;
2221 int64_t Offset = StackSize;
2222 uint64_t SlotSize = TD->getPointerSize();
2223 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2224 // Number smaller than 12 so just add the difference.
2225 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2227 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2228 Offset = ((~AlignMask) & Offset) + StackAlignment +
2229 (StackAlignment-SlotSize);
2234 /// MatchingStackOffset - Return true if the given stack call argument is
2235 /// already available in the same position (relatively) of the caller's
2236 /// incoming argument stack.
2238 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2239 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2240 const X86InstrInfo *TII) {
2241 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2243 if (Arg.getOpcode() == ISD::CopyFromReg) {
2244 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2245 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2247 MachineInstr *Def = MRI->getVRegDef(VR);
2250 if (!Flags.isByVal()) {
2251 if (!TII->isLoadFromStackSlot(Def, FI))
2254 unsigned Opcode = Def->getOpcode();
2255 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2256 Def->getOperand(1).isFI()) {
2257 FI = Def->getOperand(1).getIndex();
2258 Bytes = Flags.getByValSize();
2262 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2263 if (Flags.isByVal())
2264 // ByVal argument is passed in as a pointer but it's now being
2265 // dereferenced. e.g.
2266 // define @foo(%struct.X* %A) {
2267 // tail call @bar(%struct.X* byval %A)
2270 SDValue Ptr = Ld->getBasePtr();
2271 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2274 FI = FINode->getIndex();
2278 assert(FI != INT_MAX);
2279 if (!MFI->isFixedObjectIndex(FI))
2281 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2284 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2285 /// for tail call optimization. Targets which want to do tail call
2286 /// optimization should implement this function.
2288 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2289 CallingConv::ID CalleeCC,
2291 bool isCalleeStructRet,
2292 bool isCallerStructRet,
2293 const SmallVectorImpl<ISD::OutputArg> &Outs,
2294 const SmallVectorImpl<ISD::InputArg> &Ins,
2295 SelectionDAG& DAG) const {
2296 if (!IsTailCallConvention(CalleeCC) &&
2297 CalleeCC != CallingConv::C)
2300 // If -tailcallopt is specified, make fastcc functions tail-callable.
2301 const MachineFunction &MF = DAG.getMachineFunction();
2302 const Function *CallerF = DAG.getMachineFunction().getFunction();
2303 CallingConv::ID CallerCC = CallerF->getCallingConv();
2304 bool CCMatch = CallerCC == CalleeCC;
2306 if (GuaranteedTailCallOpt) {
2307 if (IsTailCallConvention(CalleeCC) && CCMatch)
2312 // Look for obvious safe cases to perform tail call optimization that does not
2313 // requite ABI changes. This is what gcc calls sibcall.
2315 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2316 // emit a special epilogue.
2317 if (RegInfo->needsStackRealignment(MF))
2320 // Do not sibcall optimize vararg calls unless the call site is not passing any
2322 if (isVarArg && !Outs.empty())
2325 // Also avoid sibcall optimization if either caller or callee uses struct
2326 // return semantics.
2327 if (isCalleeStructRet || isCallerStructRet)
2330 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2331 // Therefore if it's not used by the call it is not safe to optimize this into
2333 bool Unused = false;
2334 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2341 SmallVector<CCValAssign, 16> RVLocs;
2342 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2343 RVLocs, *DAG.getContext());
2344 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2345 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2346 CCValAssign &VA = RVLocs[i];
2347 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2352 // If the calling conventions do not match, then we'd better make sure the
2353 // results are returned in the same way as what the caller expects.
2355 SmallVector<CCValAssign, 16> RVLocs1;
2356 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2357 RVLocs1, *DAG.getContext());
2358 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2360 SmallVector<CCValAssign, 16> RVLocs2;
2361 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2362 RVLocs2, *DAG.getContext());
2363 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2365 if (RVLocs1.size() != RVLocs2.size())
2367 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2368 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2370 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2372 if (RVLocs1[i].isRegLoc()) {
2373 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2376 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2382 // If the callee takes no arguments then go on to check the results of the
2384 if (!Outs.empty()) {
2385 // Check if stack adjustment is needed. For now, do not do this if any
2386 // argument is passed on the stack.
2387 SmallVector<CCValAssign, 16> ArgLocs;
2388 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2389 ArgLocs, *DAG.getContext());
2390 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2391 if (CCInfo.getNextStackOffset()) {
2392 MachineFunction &MF = DAG.getMachineFunction();
2393 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2395 if (Subtarget->isTargetWin64())
2396 // Win64 ABI has additional complications.
2399 // Check if the arguments are already laid out in the right way as
2400 // the caller's fixed stack objects.
2401 MachineFrameInfo *MFI = MF.getFrameInfo();
2402 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2403 const X86InstrInfo *TII =
2404 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2405 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2406 CCValAssign &VA = ArgLocs[i];
2407 EVT RegVT = VA.getLocVT();
2408 SDValue Arg = Outs[i].Val;
2409 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2410 if (VA.getLocInfo() == CCValAssign::Indirect)
2412 if (!VA.isRegLoc()) {
2413 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2425 X86TargetLowering::createFastISel(MachineFunction &mf,
2426 DenseMap<const Value *, unsigned> &vm,
2427 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2428 DenseMap<const AllocaInst *, int> &am,
2429 std::vector<std::pair<MachineInstr*, unsigned> > &pn
2431 , SmallSet<const Instruction *, 8> &cil
2434 return X86::createFastISel(mf, vm, bm, am, pn
2442 //===----------------------------------------------------------------------===//
2443 // Other Lowering Hooks
2444 //===----------------------------------------------------------------------===//
2447 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2448 MachineFunction &MF = DAG.getMachineFunction();
2449 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2450 int ReturnAddrIndex = FuncInfo->getRAIndex();
2452 if (ReturnAddrIndex == 0) {
2453 // Set up a frame object for the return address.
2454 uint64_t SlotSize = TD->getPointerSize();
2455 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2457 FuncInfo->setRAIndex(ReturnAddrIndex);
2460 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2464 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2465 bool hasSymbolicDisplacement) {
2466 // Offset should fit into 32 bit immediate field.
2467 if (!isInt<32>(Offset))
2470 // If we don't have a symbolic displacement - we don't have any extra
2472 if (!hasSymbolicDisplacement)
2475 // FIXME: Some tweaks might be needed for medium code model.
2476 if (M != CodeModel::Small && M != CodeModel::Kernel)
2479 // For small code model we assume that latest object is 16MB before end of 31
2480 // bits boundary. We may also accept pretty large negative constants knowing
2481 // that all objects are in the positive half of address space.
2482 if (M == CodeModel::Small && Offset < 16*1024*1024)
2485 // For kernel code model we know that all object resist in the negative half
2486 // of 32bits address space. We may not accept negative offsets, since they may
2487 // be just off and we may accept pretty large positive ones.
2488 if (M == CodeModel::Kernel && Offset > 0)
2494 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2495 /// specific condition code, returning the condition code and the LHS/RHS of the
2496 /// comparison to make.
2497 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2498 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2500 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2501 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2502 // X > -1 -> X == 0, jump !sign.
2503 RHS = DAG.getConstant(0, RHS.getValueType());
2504 return X86::COND_NS;
2505 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2506 // X < 0 -> X == 0, jump on sign.
2508 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2510 RHS = DAG.getConstant(0, RHS.getValueType());
2511 return X86::COND_LE;
2515 switch (SetCCOpcode) {
2516 default: llvm_unreachable("Invalid integer condition!");
2517 case ISD::SETEQ: return X86::COND_E;
2518 case ISD::SETGT: return X86::COND_G;
2519 case ISD::SETGE: return X86::COND_GE;
2520 case ISD::SETLT: return X86::COND_L;
2521 case ISD::SETLE: return X86::COND_LE;
2522 case ISD::SETNE: return X86::COND_NE;
2523 case ISD::SETULT: return X86::COND_B;
2524 case ISD::SETUGT: return X86::COND_A;
2525 case ISD::SETULE: return X86::COND_BE;
2526 case ISD::SETUGE: return X86::COND_AE;
2530 // First determine if it is required or is profitable to flip the operands.
2532 // If LHS is a foldable load, but RHS is not, flip the condition.
2533 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2534 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2535 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2536 std::swap(LHS, RHS);
2539 switch (SetCCOpcode) {
2545 std::swap(LHS, RHS);
2549 // On a floating point condition, the flags are set as follows:
2551 // 0 | 0 | 0 | X > Y
2552 // 0 | 0 | 1 | X < Y
2553 // 1 | 0 | 0 | X == Y
2554 // 1 | 1 | 1 | unordered
2555 switch (SetCCOpcode) {
2556 default: llvm_unreachable("Condcode should be pre-legalized away");
2558 case ISD::SETEQ: return X86::COND_E;
2559 case ISD::SETOLT: // flipped
2561 case ISD::SETGT: return X86::COND_A;
2562 case ISD::SETOLE: // flipped
2564 case ISD::SETGE: return X86::COND_AE;
2565 case ISD::SETUGT: // flipped
2567 case ISD::SETLT: return X86::COND_B;
2568 case ISD::SETUGE: // flipped
2570 case ISD::SETLE: return X86::COND_BE;
2572 case ISD::SETNE: return X86::COND_NE;
2573 case ISD::SETUO: return X86::COND_P;
2574 case ISD::SETO: return X86::COND_NP;
2576 case ISD::SETUNE: return X86::COND_INVALID;
2580 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2581 /// code. Current x86 isa includes the following FP cmov instructions:
2582 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2583 static bool hasFPCMov(unsigned X86CC) {
2599 /// isFPImmLegal - Returns true if the target can instruction select the
2600 /// specified FP immediate natively. If false, the legalizer will
2601 /// materialize the FP immediate as a load from a constant pool.
2602 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2603 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2604 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2610 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2611 /// the specified range (L, H].
2612 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2613 return (Val < 0) || (Val >= Low && Val < Hi);
2616 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2617 /// specified value.
2618 static bool isUndefOrEqual(int Val, int CmpVal) {
2619 if (Val < 0 || Val == CmpVal)
2624 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2625 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2626 /// the second operand.
2627 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2628 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2629 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2630 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2631 return (Mask[0] < 2 && Mask[1] < 2);
2635 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2636 SmallVector<int, 8> M;
2638 return ::isPSHUFDMask(M, N->getValueType(0));
2641 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2642 /// is suitable for input to PSHUFHW.
2643 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2644 if (VT != MVT::v8i16)
2647 // Lower quadword copied in order or undef.
2648 for (int i = 0; i != 4; ++i)
2649 if (Mask[i] >= 0 && Mask[i] != i)
2652 // Upper quadword shuffled.
2653 for (int i = 4; i != 8; ++i)
2654 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2660 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2661 SmallVector<int, 8> M;
2663 return ::isPSHUFHWMask(M, N->getValueType(0));
2666 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2667 /// is suitable for input to PSHUFLW.
2668 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2669 if (VT != MVT::v8i16)
2672 // Upper quadword copied in order.
2673 for (int i = 4; i != 8; ++i)
2674 if (Mask[i] >= 0 && Mask[i] != i)
2677 // Lower quadword shuffled.
2678 for (int i = 0; i != 4; ++i)
2685 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2686 SmallVector<int, 8> M;
2688 return ::isPSHUFLWMask(M, N->getValueType(0));
2691 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2692 /// is suitable for input to PALIGNR.
2693 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2695 int i, e = VT.getVectorNumElements();
2697 // Do not handle v2i64 / v2f64 shuffles with palignr.
2698 if (e < 4 || !hasSSSE3)
2701 for (i = 0; i != e; ++i)
2705 // All undef, not a palignr.
2709 // Determine if it's ok to perform a palignr with only the LHS, since we
2710 // don't have access to the actual shuffle elements to see if RHS is undef.
2711 bool Unary = Mask[i] < (int)e;
2712 bool NeedsUnary = false;
2714 int s = Mask[i] - i;
2716 // Check the rest of the elements to see if they are consecutive.
2717 for (++i; i != e; ++i) {
2722 Unary = Unary && (m < (int)e);
2723 NeedsUnary = NeedsUnary || (m < s);
2725 if (NeedsUnary && !Unary)
2727 if (Unary && m != ((s+i) & (e-1)))
2729 if (!Unary && m != (s+i))
2735 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2736 SmallVector<int, 8> M;
2738 return ::isPALIGNRMask(M, N->getValueType(0), true);
2741 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2742 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2743 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2744 int NumElems = VT.getVectorNumElements();
2745 if (NumElems != 2 && NumElems != 4)
2748 int Half = NumElems / 2;
2749 for (int i = 0; i < Half; ++i)
2750 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2752 for (int i = Half; i < NumElems; ++i)
2753 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2759 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2760 SmallVector<int, 8> M;
2762 return ::isSHUFPMask(M, N->getValueType(0));
2765 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2766 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2767 /// half elements to come from vector 1 (which would equal the dest.) and
2768 /// the upper half to come from vector 2.
2769 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2770 int NumElems = VT.getVectorNumElements();
2772 if (NumElems != 2 && NumElems != 4)
2775 int Half = NumElems / 2;
2776 for (int i = 0; i < Half; ++i)
2777 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2779 for (int i = Half; i < NumElems; ++i)
2780 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2785 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2786 SmallVector<int, 8> M;
2788 return isCommutedSHUFPMask(M, N->getValueType(0));
2791 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2792 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2793 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2794 if (N->getValueType(0).getVectorNumElements() != 4)
2797 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2798 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2799 isUndefOrEqual(N->getMaskElt(1), 7) &&
2800 isUndefOrEqual(N->getMaskElt(2), 2) &&
2801 isUndefOrEqual(N->getMaskElt(3), 3);
2804 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2805 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2807 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2808 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2813 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2814 isUndefOrEqual(N->getMaskElt(1), 3) &&
2815 isUndefOrEqual(N->getMaskElt(2), 2) &&
2816 isUndefOrEqual(N->getMaskElt(3), 3);
2819 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2820 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2821 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2822 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2824 if (NumElems != 2 && NumElems != 4)
2827 for (unsigned i = 0; i < NumElems/2; ++i)
2828 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2831 for (unsigned i = NumElems/2; i < NumElems; ++i)
2832 if (!isUndefOrEqual(N->getMaskElt(i), i))
2838 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2839 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2840 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2841 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2843 if (NumElems != 2 && NumElems != 4)
2846 for (unsigned i = 0; i < NumElems/2; ++i)
2847 if (!isUndefOrEqual(N->getMaskElt(i), i))
2850 for (unsigned i = 0; i < NumElems/2; ++i)
2851 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2857 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2858 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2859 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2860 bool V2IsSplat = false) {
2861 int NumElts = VT.getVectorNumElements();
2862 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2865 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2867 int BitI1 = Mask[i+1];
2868 if (!isUndefOrEqual(BitI, j))
2871 if (!isUndefOrEqual(BitI1, NumElts))
2874 if (!isUndefOrEqual(BitI1, j + NumElts))
2881 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2882 SmallVector<int, 8> M;
2884 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2887 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2888 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2889 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2890 bool V2IsSplat = false) {
2891 int NumElts = VT.getVectorNumElements();
2892 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2895 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2897 int BitI1 = Mask[i+1];
2898 if (!isUndefOrEqual(BitI, j + NumElts/2))
2901 if (isUndefOrEqual(BitI1, NumElts))
2904 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2911 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2912 SmallVector<int, 8> M;
2914 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2917 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2918 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2920 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2921 int NumElems = VT.getVectorNumElements();
2922 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2925 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2927 int BitI1 = Mask[i+1];
2928 if (!isUndefOrEqual(BitI, j))
2930 if (!isUndefOrEqual(BitI1, j))
2936 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2937 SmallVector<int, 8> M;
2939 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2942 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2943 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2945 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2946 int NumElems = VT.getVectorNumElements();
2947 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2950 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2952 int BitI1 = Mask[i+1];
2953 if (!isUndefOrEqual(BitI, j))
2955 if (!isUndefOrEqual(BitI1, j))
2961 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2962 SmallVector<int, 8> M;
2964 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2967 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2968 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2969 /// MOVSD, and MOVD, i.e. setting the lowest element.
2970 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2971 if (VT.getVectorElementType().getSizeInBits() < 32)
2974 int NumElts = VT.getVectorNumElements();
2976 if (!isUndefOrEqual(Mask[0], NumElts))
2979 for (int i = 1; i < NumElts; ++i)
2980 if (!isUndefOrEqual(Mask[i], i))
2986 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2987 SmallVector<int, 8> M;
2989 return ::isMOVLMask(M, N->getValueType(0));
2992 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2993 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2994 /// element of vector 2 and the other elements to come from vector 1 in order.
2995 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2996 bool V2IsSplat = false, bool V2IsUndef = false) {
2997 int NumOps = VT.getVectorNumElements();
2998 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3001 if (!isUndefOrEqual(Mask[0], 0))
3004 for (int i = 1; i < NumOps; ++i)
3005 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3006 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3007 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3013 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3014 bool V2IsUndef = false) {
3015 SmallVector<int, 8> M;
3017 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3020 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3021 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3022 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3023 if (N->getValueType(0).getVectorNumElements() != 4)
3026 // Expect 1, 1, 3, 3
3027 for (unsigned i = 0; i < 2; ++i) {
3028 int Elt = N->getMaskElt(i);
3029 if (Elt >= 0 && Elt != 1)
3034 for (unsigned i = 2; i < 4; ++i) {
3035 int Elt = N->getMaskElt(i);
3036 if (Elt >= 0 && Elt != 3)
3041 // Don't use movshdup if it can be done with a shufps.
3042 // FIXME: verify that matching u, u, 3, 3 is what we want.
3046 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3047 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3048 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3049 if (N->getValueType(0).getVectorNumElements() != 4)
3052 // Expect 0, 0, 2, 2
3053 for (unsigned i = 0; i < 2; ++i)
3054 if (N->getMaskElt(i) > 0)
3058 for (unsigned i = 2; i < 4; ++i) {
3059 int Elt = N->getMaskElt(i);
3060 if (Elt >= 0 && Elt != 2)
3065 // Don't use movsldup if it can be done with a shufps.
3069 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3070 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3071 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3072 int e = N->getValueType(0).getVectorNumElements() / 2;
3074 for (int i = 0; i < e; ++i)
3075 if (!isUndefOrEqual(N->getMaskElt(i), i))
3077 for (int i = 0; i < e; ++i)
3078 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3083 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3084 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3085 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3086 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3087 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3089 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3091 for (int i = 0; i < NumOperands; ++i) {
3092 int Val = SVOp->getMaskElt(NumOperands-i-1);
3093 if (Val < 0) Val = 0;
3094 if (Val >= NumOperands) Val -= NumOperands;
3096 if (i != NumOperands - 1)
3102 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3103 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3104 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3105 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3107 // 8 nodes, but we only care about the last 4.
3108 for (unsigned i = 7; i >= 4; --i) {
3109 int Val = SVOp->getMaskElt(i);
3118 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3119 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3120 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3121 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3123 // 8 nodes, but we only care about the first 4.
3124 for (int i = 3; i >= 0; --i) {
3125 int Val = SVOp->getMaskElt(i);
3134 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3135 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3136 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3137 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3138 EVT VVT = N->getValueType(0);
3139 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3143 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3144 Val = SVOp->getMaskElt(i);
3148 return (Val - i) * EltSize;
3151 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3153 bool X86::isZeroNode(SDValue Elt) {
3154 return ((isa<ConstantSDNode>(Elt) &&
3155 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3156 (isa<ConstantFPSDNode>(Elt) &&
3157 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3160 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3161 /// their permute mask.
3162 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3163 SelectionDAG &DAG) {
3164 EVT VT = SVOp->getValueType(0);
3165 unsigned NumElems = VT.getVectorNumElements();
3166 SmallVector<int, 8> MaskVec;
3168 for (unsigned i = 0; i != NumElems; ++i) {
3169 int idx = SVOp->getMaskElt(i);
3171 MaskVec.push_back(idx);
3172 else if (idx < (int)NumElems)
3173 MaskVec.push_back(idx + NumElems);
3175 MaskVec.push_back(idx - NumElems);
3177 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3178 SVOp->getOperand(0), &MaskVec[0]);
3181 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3182 /// the two vector operands have swapped position.
3183 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3184 unsigned NumElems = VT.getVectorNumElements();
3185 for (unsigned i = 0; i != NumElems; ++i) {
3189 else if (idx < (int)NumElems)
3190 Mask[i] = idx + NumElems;
3192 Mask[i] = idx - NumElems;
3196 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3197 /// match movhlps. The lower half elements should come from upper half of
3198 /// V1 (and in order), and the upper half elements should come from the upper
3199 /// half of V2 (and in order).
3200 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3201 if (Op->getValueType(0).getVectorNumElements() != 4)
3203 for (unsigned i = 0, e = 2; i != e; ++i)
3204 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3206 for (unsigned i = 2; i != 4; ++i)
3207 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3212 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3213 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3215 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3216 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3218 N = N->getOperand(0).getNode();
3219 if (!ISD::isNON_EXTLoad(N))
3222 *LD = cast<LoadSDNode>(N);
3226 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3227 /// match movlp{s|d}. The lower half elements should come from lower half of
3228 /// V1 (and in order), and the upper half elements should come from the upper
3229 /// half of V2 (and in order). And since V1 will become the source of the
3230 /// MOVLP, it must be either a vector load or a scalar load to vector.
3231 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3232 ShuffleVectorSDNode *Op) {
3233 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3235 // Is V2 is a vector load, don't do this transformation. We will try to use
3236 // load folding shufps op.
3237 if (ISD::isNON_EXTLoad(V2))
3240 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3242 if (NumElems != 2 && NumElems != 4)
3244 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3245 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3247 for (unsigned i = NumElems/2; i != NumElems; ++i)
3248 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3253 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3255 static bool isSplatVector(SDNode *N) {
3256 if (N->getOpcode() != ISD::BUILD_VECTOR)
3259 SDValue SplatValue = N->getOperand(0);
3260 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3261 if (N->getOperand(i) != SplatValue)
3266 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3267 /// to an zero vector.
3268 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3269 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3270 SDValue V1 = N->getOperand(0);
3271 SDValue V2 = N->getOperand(1);
3272 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3273 for (unsigned i = 0; i != NumElems; ++i) {
3274 int Idx = N->getMaskElt(i);
3275 if (Idx >= (int)NumElems) {
3276 unsigned Opc = V2.getOpcode();
3277 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3279 if (Opc != ISD::BUILD_VECTOR ||
3280 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3282 } else if (Idx >= 0) {
3283 unsigned Opc = V1.getOpcode();
3284 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3286 if (Opc != ISD::BUILD_VECTOR ||
3287 !X86::isZeroNode(V1.getOperand(Idx)))
3294 /// getZeroVector - Returns a vector of specified type with all zero elements.
3296 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3298 assert(VT.isVector() && "Expected a vector type");
3300 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3301 // type. This ensures they get CSE'd.
3303 if (VT.getSizeInBits() == 64) { // MMX
3304 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3305 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3306 } else if (HasSSE2) { // SSE2
3307 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3308 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3310 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3311 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3313 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3316 /// getOnesVector - Returns a vector of specified type with all bits set.
3318 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3319 assert(VT.isVector() && "Expected a vector type");
3321 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3322 // type. This ensures they get CSE'd.
3323 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3325 if (VT.getSizeInBits() == 64) // MMX
3326 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3328 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3329 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3333 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3334 /// that point to V2 points to its first element.
3335 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3336 EVT VT = SVOp->getValueType(0);
3337 unsigned NumElems = VT.getVectorNumElements();
3339 bool Changed = false;
3340 SmallVector<int, 8> MaskVec;
3341 SVOp->getMask(MaskVec);
3343 for (unsigned i = 0; i != NumElems; ++i) {
3344 if (MaskVec[i] > (int)NumElems) {
3345 MaskVec[i] = NumElems;
3350 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3351 SVOp->getOperand(1), &MaskVec[0]);
3352 return SDValue(SVOp, 0);
3355 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3356 /// operation of specified width.
3357 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3359 unsigned NumElems = VT.getVectorNumElements();
3360 SmallVector<int, 8> Mask;
3361 Mask.push_back(NumElems);
3362 for (unsigned i = 1; i != NumElems; ++i)
3364 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3367 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3368 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3370 unsigned NumElems = VT.getVectorNumElements();
3371 SmallVector<int, 8> Mask;
3372 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3374 Mask.push_back(i + NumElems);
3376 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3379 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3380 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3382 unsigned NumElems = VT.getVectorNumElements();
3383 unsigned Half = NumElems/2;
3384 SmallVector<int, 8> Mask;
3385 for (unsigned i = 0; i != Half; ++i) {
3386 Mask.push_back(i + Half);
3387 Mask.push_back(i + NumElems + Half);
3389 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3392 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3393 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3395 if (SV->getValueType(0).getVectorNumElements() <= 4)
3396 return SDValue(SV, 0);
3398 EVT PVT = MVT::v4f32;
3399 EVT VT = SV->getValueType(0);
3400 DebugLoc dl = SV->getDebugLoc();
3401 SDValue V1 = SV->getOperand(0);
3402 int NumElems = VT.getVectorNumElements();
3403 int EltNo = SV->getSplatIndex();
3405 // unpack elements to the correct location
3406 while (NumElems > 4) {
3407 if (EltNo < NumElems/2) {
3408 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3410 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3411 EltNo -= NumElems/2;
3416 // Perform the splat.
3417 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3418 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3419 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3420 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3423 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3424 /// vector of zero or undef vector. This produces a shuffle where the low
3425 /// element of V2 is swizzled into the zero/undef vector, landing at element
3426 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3427 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3428 bool isZero, bool HasSSE2,
3429 SelectionDAG &DAG) {
3430 EVT VT = V2.getValueType();
3432 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3433 unsigned NumElems = VT.getVectorNumElements();
3434 SmallVector<int, 16> MaskVec;
3435 for (unsigned i = 0; i != NumElems; ++i)
3436 // If this is the insertion idx, put the low elt of V2 here.
3437 MaskVec.push_back(i == Idx ? NumElems : i);
3438 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3441 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3442 /// a shuffle that is zero.
3444 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3445 bool Low, SelectionDAG &DAG) {
3446 unsigned NumZeros = 0;
3447 for (int i = 0; i < NumElems; ++i) {
3448 unsigned Index = Low ? i : NumElems-i-1;
3449 int Idx = SVOp->getMaskElt(Index);
3454 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3455 if (Elt.getNode() && X86::isZeroNode(Elt))
3463 /// isVectorShift - Returns true if the shuffle can be implemented as a
3464 /// logical left or right shift of a vector.
3465 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3466 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3467 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3468 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3471 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3474 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3478 bool SeenV1 = false;
3479 bool SeenV2 = false;
3480 for (unsigned i = NumZeros; i < NumElems; ++i) {
3481 unsigned Val = isLeft ? (i - NumZeros) : i;
3482 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3485 unsigned Idx = (unsigned) Idx_;
3495 if (SeenV1 && SeenV2)
3498 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3504 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3506 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3507 unsigned NumNonZero, unsigned NumZero,
3509 const TargetLowering &TLI) {
3513 DebugLoc dl = Op.getDebugLoc();
3516 for (unsigned i = 0; i < 16; ++i) {
3517 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3518 if (ThisIsNonZero && First) {
3520 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3522 V = DAG.getUNDEF(MVT::v8i16);
3527 SDValue ThisElt(0, 0), LastElt(0, 0);
3528 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3529 if (LastIsNonZero) {
3530 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3531 MVT::i16, Op.getOperand(i-1));
3533 if (ThisIsNonZero) {
3534 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3535 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3536 ThisElt, DAG.getConstant(8, MVT::i8));
3538 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3542 if (ThisElt.getNode())
3543 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3544 DAG.getIntPtrConstant(i/2));
3548 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3551 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3553 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3554 unsigned NumNonZero, unsigned NumZero,
3556 const TargetLowering &TLI) {
3560 DebugLoc dl = Op.getDebugLoc();
3563 for (unsigned i = 0; i < 8; ++i) {
3564 bool isNonZero = (NonZeros & (1 << i)) != 0;
3568 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3570 V = DAG.getUNDEF(MVT::v8i16);
3573 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3574 MVT::v8i16, V, Op.getOperand(i),
3575 DAG.getIntPtrConstant(i));
3582 /// getVShift - Return a vector logical shift node.
3584 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3585 unsigned NumBits, SelectionDAG &DAG,
3586 const TargetLowering &TLI, DebugLoc dl) {
3587 bool isMMX = VT.getSizeInBits() == 64;
3588 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3589 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3590 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3591 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3592 DAG.getNode(Opc, dl, ShVT, SrcOp,
3593 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3597 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3598 SelectionDAG &DAG) const {
3600 // Check if the scalar load can be widened into a vector load. And if
3601 // the address is "base + cst" see if the cst can be "absorbed" into
3602 // the shuffle mask.
3603 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3604 SDValue Ptr = LD->getBasePtr();
3605 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3607 EVT PVT = LD->getValueType(0);
3608 if (PVT != MVT::i32 && PVT != MVT::f32)
3613 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3614 FI = FINode->getIndex();
3616 } else if (Ptr.getOpcode() == ISD::ADD &&
3617 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3618 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3619 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3620 Offset = Ptr.getConstantOperandVal(1);
3621 Ptr = Ptr.getOperand(0);
3626 SDValue Chain = LD->getChain();
3627 // Make sure the stack object alignment is at least 16.
3628 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3629 if (DAG.InferPtrAlignment(Ptr) < 16) {
3630 if (MFI->isFixedObjectIndex(FI)) {
3631 // Can't change the alignment. FIXME: It's possible to compute
3632 // the exact stack offset and reference FI + adjust offset instead.
3633 // If someone *really* cares about this. That's the way to implement it.
3636 MFI->setObjectAlignment(FI, 16);
3640 // (Offset % 16) must be multiple of 4. Then address is then
3641 // Ptr + (Offset & ~15).
3644 if ((Offset % 16) & 3)
3646 int64_t StartOffset = Offset & ~15;
3648 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3649 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3651 int EltNo = (Offset - StartOffset) >> 2;
3652 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3653 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3654 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3656 // Canonicalize it to a v4i32 shuffle.
3657 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3658 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3659 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3660 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3666 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3667 /// vector of type 'VT', see if the elements can be replaced by a single large
3668 /// load which has the same value as a build_vector whose operands are 'elts'.
3670 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3672 /// FIXME: we'd also like to handle the case where the last elements are zero
3673 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3674 /// There's even a handy isZeroNode for that purpose.
3675 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3676 DebugLoc &dl, SelectionDAG &DAG) {
3677 EVT EltVT = VT.getVectorElementType();
3678 unsigned NumElems = Elts.size();
3680 LoadSDNode *LDBase = NULL;
3681 unsigned LastLoadedElt = -1U;
3683 // For each element in the initializer, see if we've found a load or an undef.
3684 // If we don't find an initial load element, or later load elements are
3685 // non-consecutive, bail out.
3686 for (unsigned i = 0; i < NumElems; ++i) {
3687 SDValue Elt = Elts[i];
3689 if (!Elt.getNode() ||
3690 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3693 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3695 LDBase = cast<LoadSDNode>(Elt.getNode());
3699 if (Elt.getOpcode() == ISD::UNDEF)
3702 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3703 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3708 // If we have found an entire vector of loads and undefs, then return a large
3709 // load of the entire vector width starting at the base pointer. If we found
3710 // consecutive loads for the low half, generate a vzext_load node.
3711 if (LastLoadedElt == NumElems - 1) {
3712 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3713 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3714 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3715 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3716 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3717 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3718 LDBase->isVolatile(), LDBase->isNonTemporal(),
3719 LDBase->getAlignment());
3720 } else if (NumElems == 4 && LastLoadedElt == 1) {
3721 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3722 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3723 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3724 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3730 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3731 DebugLoc dl = Op.getDebugLoc();
3732 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3733 if (ISD::isBuildVectorAllZeros(Op.getNode())
3734 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3735 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3736 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3737 // eliminated on x86-32 hosts.
3738 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3741 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3742 return getOnesVector(Op.getValueType(), DAG, dl);
3743 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3746 EVT VT = Op.getValueType();
3747 EVT ExtVT = VT.getVectorElementType();
3748 unsigned EVTBits = ExtVT.getSizeInBits();
3750 unsigned NumElems = Op.getNumOperands();
3751 unsigned NumZero = 0;
3752 unsigned NumNonZero = 0;
3753 unsigned NonZeros = 0;
3754 bool IsAllConstants = true;
3755 SmallSet<SDValue, 8> Values;
3756 for (unsigned i = 0; i < NumElems; ++i) {
3757 SDValue Elt = Op.getOperand(i);
3758 if (Elt.getOpcode() == ISD::UNDEF)
3761 if (Elt.getOpcode() != ISD::Constant &&
3762 Elt.getOpcode() != ISD::ConstantFP)
3763 IsAllConstants = false;
3764 if (X86::isZeroNode(Elt))
3767 NonZeros |= (1 << i);
3772 if (NumNonZero == 0) {
3773 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3774 return DAG.getUNDEF(VT);
3777 // Special case for single non-zero, non-undef, element.
3778 if (NumNonZero == 1) {
3779 unsigned Idx = CountTrailingZeros_32(NonZeros);
3780 SDValue Item = Op.getOperand(Idx);
3782 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3783 // the value are obviously zero, truncate the value to i32 and do the
3784 // insertion that way. Only do this if the value is non-constant or if the
3785 // value is a constant being inserted into element 0. It is cheaper to do
3786 // a constant pool load than it is to do a movd + shuffle.
3787 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3788 (!IsAllConstants || Idx == 0)) {
3789 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3790 // Handle MMX and SSE both.
3791 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3792 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3794 // Truncate the value (which may itself be a constant) to i32, and
3795 // convert it to a vector with movd (S2V+shuffle to zero extend).
3796 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3797 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3798 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3799 Subtarget->hasSSE2(), DAG);
3801 // Now we have our 32-bit value zero extended in the low element of
3802 // a vector. If Idx != 0, swizzle it into place.
3804 SmallVector<int, 4> Mask;
3805 Mask.push_back(Idx);
3806 for (unsigned i = 1; i != VecElts; ++i)
3808 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3809 DAG.getUNDEF(Item.getValueType()),
3812 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3816 // If we have a constant or non-constant insertion into the low element of
3817 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3818 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3819 // depending on what the source datatype is.
3822 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3823 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3824 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3825 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3826 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3827 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3829 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3830 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3831 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3832 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3833 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3834 Subtarget->hasSSE2(), DAG);
3835 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3839 // Is it a vector logical left shift?
3840 if (NumElems == 2 && Idx == 1 &&
3841 X86::isZeroNode(Op.getOperand(0)) &&
3842 !X86::isZeroNode(Op.getOperand(1))) {
3843 unsigned NumBits = VT.getSizeInBits();
3844 return getVShift(true, VT,
3845 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3846 VT, Op.getOperand(1)),
3847 NumBits/2, DAG, *this, dl);
3850 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3853 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3854 // is a non-constant being inserted into an element other than the low one,
3855 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3856 // movd/movss) to move this into the low element, then shuffle it into
3858 if (EVTBits == 32) {
3859 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3861 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3862 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3863 Subtarget->hasSSE2(), DAG);
3864 SmallVector<int, 8> MaskVec;
3865 for (unsigned i = 0; i < NumElems; i++)
3866 MaskVec.push_back(i == Idx ? 0 : 1);
3867 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3871 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3872 if (Values.size() == 1) {
3873 if (EVTBits == 32) {
3874 // Instead of a shuffle like this:
3875 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3876 // Check if it's possible to issue this instead.
3877 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3878 unsigned Idx = CountTrailingZeros_32(NonZeros);
3879 SDValue Item = Op.getOperand(Idx);
3880 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3881 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3886 // A vector full of immediates; various special cases are already
3887 // handled, so this is best done with a single constant-pool load.
3891 // Let legalizer expand 2-wide build_vectors.
3892 if (EVTBits == 64) {
3893 if (NumNonZero == 1) {
3894 // One half is zero or undef.
3895 unsigned Idx = CountTrailingZeros_32(NonZeros);
3896 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3897 Op.getOperand(Idx));
3898 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3899 Subtarget->hasSSE2(), DAG);
3904 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3905 if (EVTBits == 8 && NumElems == 16) {
3906 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3908 if (V.getNode()) return V;
3911 if (EVTBits == 16 && NumElems == 8) {
3912 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3914 if (V.getNode()) return V;
3917 // If element VT is == 32 bits, turn it into a number of shuffles.
3918 SmallVector<SDValue, 8> V;
3920 if (NumElems == 4 && NumZero > 0) {
3921 for (unsigned i = 0; i < 4; ++i) {
3922 bool isZero = !(NonZeros & (1 << i));
3924 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3926 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3929 for (unsigned i = 0; i < 2; ++i) {
3930 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3933 V[i] = V[i*2]; // Must be a zero vector.
3936 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3939 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3942 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3947 SmallVector<int, 8> MaskVec;
3948 bool Reverse = (NonZeros & 0x3) == 2;
3949 for (unsigned i = 0; i < 2; ++i)
3950 MaskVec.push_back(Reverse ? 1-i : i);
3951 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3952 for (unsigned i = 0; i < 2; ++i)
3953 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3954 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3957 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
3958 // Check for a build vector of consecutive loads.
3959 for (unsigned i = 0; i < NumElems; ++i)
3960 V[i] = Op.getOperand(i);
3962 // Check for elements which are consecutive loads.
3963 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
3967 // For SSE 4.1, use inserts into undef.
3968 if (getSubtarget()->hasSSE41()) {
3969 V[0] = DAG.getUNDEF(VT);
3970 for (unsigned i = 0; i < NumElems; ++i)
3971 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3972 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3973 Op.getOperand(i), DAG.getIntPtrConstant(i));
3977 // Otherwise, expand into a number of unpckl*
3979 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3980 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3981 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3982 for (unsigned i = 0; i < NumElems; ++i)
3983 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3985 while (NumElems != 0) {
3986 for (unsigned i = 0; i < NumElems; ++i)
3987 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3996 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
3997 // We support concatenate two MMX registers and place them in a MMX
3998 // register. This is better than doing a stack convert.
3999 DebugLoc dl = Op.getDebugLoc();
4000 EVT ResVT = Op.getValueType();
4001 assert(Op.getNumOperands() == 2);
4002 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4003 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4005 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4006 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4007 InVec = Op.getOperand(1);
4008 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4009 unsigned NumElts = ResVT.getVectorNumElements();
4010 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4011 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4012 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4014 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4015 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4016 Mask[0] = 0; Mask[1] = 2;
4017 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4019 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4022 // v8i16 shuffles - Prefer shuffles in the following order:
4023 // 1. [all] pshuflw, pshufhw, optional move
4024 // 2. [ssse3] 1 x pshufb
4025 // 3. [ssse3] 2 x pshufb + 1 x por
4026 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4028 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4030 const X86TargetLowering &TLI) {
4031 SDValue V1 = SVOp->getOperand(0);
4032 SDValue V2 = SVOp->getOperand(1);
4033 DebugLoc dl = SVOp->getDebugLoc();
4034 SmallVector<int, 8> MaskVals;
4036 // Determine if more than 1 of the words in each of the low and high quadwords
4037 // of the result come from the same quadword of one of the two inputs. Undef
4038 // mask values count as coming from any quadword, for better codegen.
4039 SmallVector<unsigned, 4> LoQuad(4);
4040 SmallVector<unsigned, 4> HiQuad(4);
4041 BitVector InputQuads(4);
4042 for (unsigned i = 0; i < 8; ++i) {
4043 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4044 int EltIdx = SVOp->getMaskElt(i);
4045 MaskVals.push_back(EltIdx);
4054 InputQuads.set(EltIdx / 4);
4057 int BestLoQuad = -1;
4058 unsigned MaxQuad = 1;
4059 for (unsigned i = 0; i < 4; ++i) {
4060 if (LoQuad[i] > MaxQuad) {
4062 MaxQuad = LoQuad[i];
4066 int BestHiQuad = -1;
4068 for (unsigned i = 0; i < 4; ++i) {
4069 if (HiQuad[i] > MaxQuad) {
4071 MaxQuad = HiQuad[i];
4075 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4076 // of the two input vectors, shuffle them into one input vector so only a
4077 // single pshufb instruction is necessary. If There are more than 2 input
4078 // quads, disable the next transformation since it does not help SSSE3.
4079 bool V1Used = InputQuads[0] || InputQuads[1];
4080 bool V2Used = InputQuads[2] || InputQuads[3];
4081 if (TLI.getSubtarget()->hasSSSE3()) {
4082 if (InputQuads.count() == 2 && V1Used && V2Used) {
4083 BestLoQuad = InputQuads.find_first();
4084 BestHiQuad = InputQuads.find_next(BestLoQuad);
4086 if (InputQuads.count() > 2) {
4092 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4093 // the shuffle mask. If a quad is scored as -1, that means that it contains
4094 // words from all 4 input quadwords.
4096 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4097 SmallVector<int, 8> MaskV;
4098 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4099 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4100 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4101 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4102 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4103 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4105 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4106 // source words for the shuffle, to aid later transformations.
4107 bool AllWordsInNewV = true;
4108 bool InOrder[2] = { true, true };
4109 for (unsigned i = 0; i != 8; ++i) {
4110 int idx = MaskVals[i];
4112 InOrder[i/4] = false;
4113 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4115 AllWordsInNewV = false;
4119 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4120 if (AllWordsInNewV) {
4121 for (int i = 0; i != 8; ++i) {
4122 int idx = MaskVals[i];
4125 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4126 if ((idx != i) && idx < 4)
4128 if ((idx != i) && idx > 3)
4137 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4138 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4139 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4140 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4141 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4145 // If we have SSSE3, and all words of the result are from 1 input vector,
4146 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4147 // is present, fall back to case 4.
4148 if (TLI.getSubtarget()->hasSSSE3()) {
4149 SmallVector<SDValue,16> pshufbMask;
4151 // If we have elements from both input vectors, set the high bit of the
4152 // shuffle mask element to zero out elements that come from V2 in the V1
4153 // mask, and elements that come from V1 in the V2 mask, so that the two
4154 // results can be OR'd together.
4155 bool TwoInputs = V1Used && V2Used;
4156 for (unsigned i = 0; i != 8; ++i) {
4157 int EltIdx = MaskVals[i] * 2;
4158 if (TwoInputs && (EltIdx >= 16)) {
4159 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4160 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4163 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4164 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4166 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4167 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4168 DAG.getNode(ISD::BUILD_VECTOR, dl,
4169 MVT::v16i8, &pshufbMask[0], 16));
4171 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4173 // Calculate the shuffle mask for the second input, shuffle it, and
4174 // OR it with the first shuffled input.
4176 for (unsigned i = 0; i != 8; ++i) {
4177 int EltIdx = MaskVals[i] * 2;
4179 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4180 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4183 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4184 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4186 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4187 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4188 DAG.getNode(ISD::BUILD_VECTOR, dl,
4189 MVT::v16i8, &pshufbMask[0], 16));
4190 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4191 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4194 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4195 // and update MaskVals with new element order.
4196 BitVector InOrder(8);
4197 if (BestLoQuad >= 0) {
4198 SmallVector<int, 8> MaskV;
4199 for (int i = 0; i != 4; ++i) {
4200 int idx = MaskVals[i];
4202 MaskV.push_back(-1);
4204 } else if ((idx / 4) == BestLoQuad) {
4205 MaskV.push_back(idx & 3);
4208 MaskV.push_back(-1);
4211 for (unsigned i = 4; i != 8; ++i)
4213 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4217 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4218 // and update MaskVals with the new element order.
4219 if (BestHiQuad >= 0) {
4220 SmallVector<int, 8> MaskV;
4221 for (unsigned i = 0; i != 4; ++i)
4223 for (unsigned i = 4; i != 8; ++i) {
4224 int idx = MaskVals[i];
4226 MaskV.push_back(-1);
4228 } else if ((idx / 4) == BestHiQuad) {
4229 MaskV.push_back((idx & 3) + 4);
4232 MaskV.push_back(-1);
4235 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4239 // In case BestHi & BestLo were both -1, which means each quadword has a word
4240 // from each of the four input quadwords, calculate the InOrder bitvector now
4241 // before falling through to the insert/extract cleanup.
4242 if (BestLoQuad == -1 && BestHiQuad == -1) {
4244 for (int i = 0; i != 8; ++i)
4245 if (MaskVals[i] < 0 || MaskVals[i] == i)
4249 // The other elements are put in the right place using pextrw and pinsrw.
4250 for (unsigned i = 0; i != 8; ++i) {
4253 int EltIdx = MaskVals[i];
4256 SDValue ExtOp = (EltIdx < 8)
4257 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4258 DAG.getIntPtrConstant(EltIdx))
4259 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4260 DAG.getIntPtrConstant(EltIdx - 8));
4261 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4262 DAG.getIntPtrConstant(i));
4267 // v16i8 shuffles - Prefer shuffles in the following order:
4268 // 1. [ssse3] 1 x pshufb
4269 // 2. [ssse3] 2 x pshufb + 1 x por
4270 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4272 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4274 const X86TargetLowering &TLI) {
4275 SDValue V1 = SVOp->getOperand(0);
4276 SDValue V2 = SVOp->getOperand(1);
4277 DebugLoc dl = SVOp->getDebugLoc();
4278 SmallVector<int, 16> MaskVals;
4279 SVOp->getMask(MaskVals);
4281 // If we have SSSE3, case 1 is generated when all result bytes come from
4282 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4283 // present, fall back to case 3.
4284 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4287 for (unsigned i = 0; i < 16; ++i) {
4288 int EltIdx = MaskVals[i];
4297 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4298 if (TLI.getSubtarget()->hasSSSE3()) {
4299 SmallVector<SDValue,16> pshufbMask;
4301 // If all result elements are from one input vector, then only translate
4302 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4304 // Otherwise, we have elements from both input vectors, and must zero out
4305 // elements that come from V2 in the first mask, and V1 in the second mask
4306 // so that we can OR them together.
4307 bool TwoInputs = !(V1Only || V2Only);
4308 for (unsigned i = 0; i != 16; ++i) {
4309 int EltIdx = MaskVals[i];
4310 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4311 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4314 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4316 // If all the elements are from V2, assign it to V1 and return after
4317 // building the first pshufb.
4320 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4321 DAG.getNode(ISD::BUILD_VECTOR, dl,
4322 MVT::v16i8, &pshufbMask[0], 16));
4326 // Calculate the shuffle mask for the second input, shuffle it, and
4327 // OR it with the first shuffled input.
4329 for (unsigned i = 0; i != 16; ++i) {
4330 int EltIdx = MaskVals[i];
4332 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4335 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4337 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4338 DAG.getNode(ISD::BUILD_VECTOR, dl,
4339 MVT::v16i8, &pshufbMask[0], 16));
4340 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4343 // No SSSE3 - Calculate in place words and then fix all out of place words
4344 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4345 // the 16 different words that comprise the two doublequadword input vectors.
4346 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4347 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4348 SDValue NewV = V2Only ? V2 : V1;
4349 for (int i = 0; i != 8; ++i) {
4350 int Elt0 = MaskVals[i*2];
4351 int Elt1 = MaskVals[i*2+1];
4353 // This word of the result is all undef, skip it.
4354 if (Elt0 < 0 && Elt1 < 0)
4357 // This word of the result is already in the correct place, skip it.
4358 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4360 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4363 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4364 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4367 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4368 // using a single extract together, load it and store it.
4369 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4370 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4371 DAG.getIntPtrConstant(Elt1 / 2));
4372 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4373 DAG.getIntPtrConstant(i));
4377 // If Elt1 is defined, extract it from the appropriate source. If the
4378 // source byte is not also odd, shift the extracted word left 8 bits
4379 // otherwise clear the bottom 8 bits if we need to do an or.
4381 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4382 DAG.getIntPtrConstant(Elt1 / 2));
4383 if ((Elt1 & 1) == 0)
4384 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4385 DAG.getConstant(8, TLI.getShiftAmountTy()));
4387 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4388 DAG.getConstant(0xFF00, MVT::i16));
4390 // If Elt0 is defined, extract it from the appropriate source. If the
4391 // source byte is not also even, shift the extracted word right 8 bits. If
4392 // Elt1 was also defined, OR the extracted values together before
4393 // inserting them in the result.
4395 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4396 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4397 if ((Elt0 & 1) != 0)
4398 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4399 DAG.getConstant(8, TLI.getShiftAmountTy()));
4401 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4402 DAG.getConstant(0x00FF, MVT::i16));
4403 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4406 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4407 DAG.getIntPtrConstant(i));
4409 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4412 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4413 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4414 /// done when every pair / quad of shuffle mask elements point to elements in
4415 /// the right sequence. e.g.
4416 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4418 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4420 const TargetLowering &TLI, DebugLoc dl) {
4421 EVT VT = SVOp->getValueType(0);
4422 SDValue V1 = SVOp->getOperand(0);
4423 SDValue V2 = SVOp->getOperand(1);
4424 unsigned NumElems = VT.getVectorNumElements();
4425 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4426 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4427 EVT MaskEltVT = MaskVT.getVectorElementType();
4429 switch (VT.getSimpleVT().SimpleTy) {
4430 default: assert(false && "Unexpected!");
4431 case MVT::v4f32: NewVT = MVT::v2f64; break;
4432 case MVT::v4i32: NewVT = MVT::v2i64; break;
4433 case MVT::v8i16: NewVT = MVT::v4i32; break;
4434 case MVT::v16i8: NewVT = MVT::v4i32; break;
4437 if (NewWidth == 2) {
4443 int Scale = NumElems / NewWidth;
4444 SmallVector<int, 8> MaskVec;
4445 for (unsigned i = 0; i < NumElems; i += Scale) {
4447 for (int j = 0; j < Scale; ++j) {
4448 int EltIdx = SVOp->getMaskElt(i+j);
4452 StartIdx = EltIdx - (EltIdx % Scale);
4453 if (EltIdx != StartIdx + j)
4457 MaskVec.push_back(-1);
4459 MaskVec.push_back(StartIdx / Scale);
4462 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4463 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4464 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4467 /// getVZextMovL - Return a zero-extending vector move low node.
4469 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4470 SDValue SrcOp, SelectionDAG &DAG,
4471 const X86Subtarget *Subtarget, DebugLoc dl) {
4472 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4473 LoadSDNode *LD = NULL;
4474 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4475 LD = dyn_cast<LoadSDNode>(SrcOp);
4477 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4479 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4480 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4481 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4482 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4483 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4485 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4486 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4487 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4488 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4496 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4497 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4498 DAG.getNode(ISD::BIT_CONVERT, dl,
4502 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4505 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4506 SDValue V1 = SVOp->getOperand(0);
4507 SDValue V2 = SVOp->getOperand(1);
4508 DebugLoc dl = SVOp->getDebugLoc();
4509 EVT VT = SVOp->getValueType(0);
4511 SmallVector<std::pair<int, int>, 8> Locs;
4513 SmallVector<int, 8> Mask1(4U, -1);
4514 SmallVector<int, 8> PermMask;
4515 SVOp->getMask(PermMask);
4519 for (unsigned i = 0; i != 4; ++i) {
4520 int Idx = PermMask[i];
4522 Locs[i] = std::make_pair(-1, -1);
4524 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4526 Locs[i] = std::make_pair(0, NumLo);
4530 Locs[i] = std::make_pair(1, NumHi);
4532 Mask1[2+NumHi] = Idx;
4538 if (NumLo <= 2 && NumHi <= 2) {
4539 // If no more than two elements come from either vector. This can be
4540 // implemented with two shuffles. First shuffle gather the elements.
4541 // The second shuffle, which takes the first shuffle as both of its
4542 // vector operands, put the elements into the right order.
4543 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4545 SmallVector<int, 8> Mask2(4U, -1);
4547 for (unsigned i = 0; i != 4; ++i) {
4548 if (Locs[i].first == -1)
4551 unsigned Idx = (i < 2) ? 0 : 4;
4552 Idx += Locs[i].first * 2 + Locs[i].second;
4557 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4558 } else if (NumLo == 3 || NumHi == 3) {
4559 // Otherwise, we must have three elements from one vector, call it X, and
4560 // one element from the other, call it Y. First, use a shufps to build an
4561 // intermediate vector with the one element from Y and the element from X
4562 // that will be in the same half in the final destination (the indexes don't
4563 // matter). Then, use a shufps to build the final vector, taking the half
4564 // containing the element from Y from the intermediate, and the other half
4567 // Normalize it so the 3 elements come from V1.
4568 CommuteVectorShuffleMask(PermMask, VT);
4572 // Find the element from V2.
4574 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4575 int Val = PermMask[HiIndex];
4582 Mask1[0] = PermMask[HiIndex];
4584 Mask1[2] = PermMask[HiIndex^1];
4586 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4589 Mask1[0] = PermMask[0];
4590 Mask1[1] = PermMask[1];
4591 Mask1[2] = HiIndex & 1 ? 6 : 4;
4592 Mask1[3] = HiIndex & 1 ? 4 : 6;
4593 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4595 Mask1[0] = HiIndex & 1 ? 2 : 0;
4596 Mask1[1] = HiIndex & 1 ? 0 : 2;
4597 Mask1[2] = PermMask[2];
4598 Mask1[3] = PermMask[3];
4603 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4607 // Break it into (shuffle shuffle_hi, shuffle_lo).
4609 SmallVector<int,8> LoMask(4U, -1);
4610 SmallVector<int,8> HiMask(4U, -1);
4612 SmallVector<int,8> *MaskPtr = &LoMask;
4613 unsigned MaskIdx = 0;
4616 for (unsigned i = 0; i != 4; ++i) {
4623 int Idx = PermMask[i];
4625 Locs[i] = std::make_pair(-1, -1);
4626 } else if (Idx < 4) {
4627 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4628 (*MaskPtr)[LoIdx] = Idx;
4631 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4632 (*MaskPtr)[HiIdx] = Idx;
4637 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4638 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4639 SmallVector<int, 8> MaskOps;
4640 for (unsigned i = 0; i != 4; ++i) {
4641 if (Locs[i].first == -1) {
4642 MaskOps.push_back(-1);
4644 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4645 MaskOps.push_back(Idx);
4648 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4652 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4653 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4654 SDValue V1 = Op.getOperand(0);
4655 SDValue V2 = Op.getOperand(1);
4656 EVT VT = Op.getValueType();
4657 DebugLoc dl = Op.getDebugLoc();
4658 unsigned NumElems = VT.getVectorNumElements();
4659 bool isMMX = VT.getSizeInBits() == 64;
4660 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4661 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4662 bool V1IsSplat = false;
4663 bool V2IsSplat = false;
4665 if (isZeroShuffle(SVOp))
4666 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4668 // Promote splats to v4f32.
4669 if (SVOp->isSplat()) {
4670 if (isMMX || NumElems < 4)
4672 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4675 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4677 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4678 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4679 if (NewOp.getNode())
4680 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4681 LowerVECTOR_SHUFFLE(NewOp, DAG));
4682 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4683 // FIXME: Figure out a cleaner way to do this.
4684 // Try to make use of movq to zero out the top part.
4685 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4686 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4687 if (NewOp.getNode()) {
4688 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4689 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4690 DAG, Subtarget, dl);
4692 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4693 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4694 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4695 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4696 DAG, Subtarget, dl);
4700 if (X86::isPSHUFDMask(SVOp))
4703 // Check if this can be converted into a logical shift.
4704 bool isLeft = false;
4707 bool isShift = getSubtarget()->hasSSE2() &&
4708 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4709 if (isShift && ShVal.hasOneUse()) {
4710 // If the shifted value has multiple uses, it may be cheaper to use
4711 // v_set0 + movlhps or movhlps, etc.
4712 EVT EltVT = VT.getVectorElementType();
4713 ShAmt *= EltVT.getSizeInBits();
4714 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4717 if (X86::isMOVLMask(SVOp)) {
4720 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4721 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4726 // FIXME: fold these into legal mask.
4727 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4728 X86::isMOVSLDUPMask(SVOp) ||
4729 X86::isMOVHLPSMask(SVOp) ||
4730 X86::isMOVLHPSMask(SVOp) ||
4731 X86::isMOVLPMask(SVOp)))
4734 if (ShouldXformToMOVHLPS(SVOp) ||
4735 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4736 return CommuteVectorShuffle(SVOp, DAG);
4739 // No better options. Use a vshl / vsrl.
4740 EVT EltVT = VT.getVectorElementType();
4741 ShAmt *= EltVT.getSizeInBits();
4742 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4745 bool Commuted = false;
4746 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4747 // 1,1,1,1 -> v8i16 though.
4748 V1IsSplat = isSplatVector(V1.getNode());
4749 V2IsSplat = isSplatVector(V2.getNode());
4751 // Canonicalize the splat or undef, if present, to be on the RHS.
4752 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4753 Op = CommuteVectorShuffle(SVOp, DAG);
4754 SVOp = cast<ShuffleVectorSDNode>(Op);
4755 V1 = SVOp->getOperand(0);
4756 V2 = SVOp->getOperand(1);
4757 std::swap(V1IsSplat, V2IsSplat);
4758 std::swap(V1IsUndef, V2IsUndef);
4762 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4763 // Shuffling low element of v1 into undef, just return v1.
4766 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4767 // the instruction selector will not match, so get a canonical MOVL with
4768 // swapped operands to undo the commute.
4769 return getMOVL(DAG, dl, VT, V2, V1);
4772 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4773 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4774 X86::isUNPCKLMask(SVOp) ||
4775 X86::isUNPCKHMask(SVOp))
4779 // Normalize mask so all entries that point to V2 points to its first
4780 // element then try to match unpck{h|l} again. If match, return a
4781 // new vector_shuffle with the corrected mask.
4782 SDValue NewMask = NormalizeMask(SVOp, DAG);
4783 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4784 if (NSVOp != SVOp) {
4785 if (X86::isUNPCKLMask(NSVOp, true)) {
4787 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4794 // Commute is back and try unpck* again.
4795 // FIXME: this seems wrong.
4796 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4797 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4798 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4799 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4800 X86::isUNPCKLMask(NewSVOp) ||
4801 X86::isUNPCKHMask(NewSVOp))
4805 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4807 // Normalize the node to match x86 shuffle ops if needed
4808 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4809 return CommuteVectorShuffle(SVOp, DAG);
4811 // Check for legal shuffle and return?
4812 SmallVector<int, 16> PermMask;
4813 SVOp->getMask(PermMask);
4814 if (isShuffleMaskLegal(PermMask, VT))
4817 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4818 if (VT == MVT::v8i16) {
4819 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4820 if (NewOp.getNode())
4824 if (VT == MVT::v16i8) {
4825 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4826 if (NewOp.getNode())
4830 // Handle all 4 wide cases with a number of shuffles except for MMX.
4831 if (NumElems == 4 && !isMMX)
4832 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4838 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4839 SelectionDAG &DAG) const {
4840 EVT VT = Op.getValueType();
4841 DebugLoc dl = Op.getDebugLoc();
4842 if (VT.getSizeInBits() == 8) {
4843 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4844 Op.getOperand(0), Op.getOperand(1));
4845 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4846 DAG.getValueType(VT));
4847 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4848 } else if (VT.getSizeInBits() == 16) {
4849 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4850 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4852 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4853 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4854 DAG.getNode(ISD::BIT_CONVERT, dl,
4858 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4859 Op.getOperand(0), Op.getOperand(1));
4860 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4861 DAG.getValueType(VT));
4862 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4863 } else if (VT == MVT::f32) {
4864 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4865 // the result back to FR32 register. It's only worth matching if the
4866 // result has a single use which is a store or a bitcast to i32. And in
4867 // the case of a store, it's not worth it if the index is a constant 0,
4868 // because a MOVSSmr can be used instead, which is smaller and faster.
4869 if (!Op.hasOneUse())
4871 SDNode *User = *Op.getNode()->use_begin();
4872 if ((User->getOpcode() != ISD::STORE ||
4873 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4874 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4875 (User->getOpcode() != ISD::BIT_CONVERT ||
4876 User->getValueType(0) != MVT::i32))
4878 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4879 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4882 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4883 } else if (VT == MVT::i32) {
4884 // ExtractPS works with constant index.
4885 if (isa<ConstantSDNode>(Op.getOperand(1)))
4893 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4894 SelectionDAG &DAG) const {
4895 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4898 if (Subtarget->hasSSE41()) {
4899 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4904 EVT VT = Op.getValueType();
4905 DebugLoc dl = Op.getDebugLoc();
4906 // TODO: handle v16i8.
4907 if (VT.getSizeInBits() == 16) {
4908 SDValue Vec = Op.getOperand(0);
4909 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4911 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4912 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4913 DAG.getNode(ISD::BIT_CONVERT, dl,
4916 // Transform it so it match pextrw which produces a 32-bit result.
4917 EVT EltVT = MVT::i32;
4918 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4919 Op.getOperand(0), Op.getOperand(1));
4920 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4921 DAG.getValueType(VT));
4922 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4923 } else if (VT.getSizeInBits() == 32) {
4924 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4928 // SHUFPS the element to the lowest double word, then movss.
4929 int Mask[4] = { Idx, -1, -1, -1 };
4930 EVT VVT = Op.getOperand(0).getValueType();
4931 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4932 DAG.getUNDEF(VVT), Mask);
4933 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4934 DAG.getIntPtrConstant(0));
4935 } else if (VT.getSizeInBits() == 64) {
4936 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4937 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4938 // to match extract_elt for f64.
4939 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4943 // UNPCKHPD the element to the lowest double word, then movsd.
4944 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4945 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4946 int Mask[2] = { 1, -1 };
4947 EVT VVT = Op.getOperand(0).getValueType();
4948 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4949 DAG.getUNDEF(VVT), Mask);
4950 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4951 DAG.getIntPtrConstant(0));
4958 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
4959 SelectionDAG &DAG) const {
4960 EVT VT = Op.getValueType();
4961 EVT EltVT = VT.getVectorElementType();
4962 DebugLoc dl = Op.getDebugLoc();
4964 SDValue N0 = Op.getOperand(0);
4965 SDValue N1 = Op.getOperand(1);
4966 SDValue N2 = Op.getOperand(2);
4968 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4969 isa<ConstantSDNode>(N2)) {
4971 if (VT == MVT::v8i16)
4972 Opc = X86ISD::PINSRW;
4973 else if (VT == MVT::v4i16)
4974 Opc = X86ISD::MMX_PINSRW;
4975 else if (VT == MVT::v16i8)
4976 Opc = X86ISD::PINSRB;
4978 Opc = X86ISD::PINSRB;
4980 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4982 if (N1.getValueType() != MVT::i32)
4983 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4984 if (N2.getValueType() != MVT::i32)
4985 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4986 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4987 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4988 // Bits [7:6] of the constant are the source select. This will always be
4989 // zero here. The DAG Combiner may combine an extract_elt index into these
4990 // bits. For example (insert (extract, 3), 2) could be matched by putting
4991 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4992 // Bits [5:4] of the constant are the destination select. This is the
4993 // value of the incoming immediate.
4994 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4995 // combine either bitwise AND or insert of float 0.0 to set these bits.
4996 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4997 // Create this as a scalar to vector..
4998 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4999 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5000 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5001 // PINSR* works with constant index.
5008 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5009 EVT VT = Op.getValueType();
5010 EVT EltVT = VT.getVectorElementType();
5012 if (Subtarget->hasSSE41())
5013 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5015 if (EltVT == MVT::i8)
5018 DebugLoc dl = Op.getDebugLoc();
5019 SDValue N0 = Op.getOperand(0);
5020 SDValue N1 = Op.getOperand(1);
5021 SDValue N2 = Op.getOperand(2);
5023 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5024 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5025 // as its second argument.
5026 if (N1.getValueType() != MVT::i32)
5027 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5028 if (N2.getValueType() != MVT::i32)
5029 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5030 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5031 dl, VT, N0, N1, N2);
5037 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5038 DebugLoc dl = Op.getDebugLoc();
5039 if (Op.getValueType() == MVT::v2f32)
5040 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
5041 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
5042 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
5043 Op.getOperand(0))));
5045 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
5046 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5048 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5049 EVT VT = MVT::v2i32;
5050 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5057 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5058 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5061 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5062 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5063 // one of the above mentioned nodes. It has to be wrapped because otherwise
5064 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5065 // be used to form addressing mode. These wrapped nodes will be selected
5068 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5069 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5071 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5073 unsigned char OpFlag = 0;
5074 unsigned WrapperKind = X86ISD::Wrapper;
5075 CodeModel::Model M = getTargetMachine().getCodeModel();
5077 if (Subtarget->isPICStyleRIPRel() &&
5078 (M == CodeModel::Small || M == CodeModel::Kernel))
5079 WrapperKind = X86ISD::WrapperRIP;
5080 else if (Subtarget->isPICStyleGOT())
5081 OpFlag = X86II::MO_GOTOFF;
5082 else if (Subtarget->isPICStyleStubPIC())
5083 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5085 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5087 CP->getOffset(), OpFlag);
5088 DebugLoc DL = CP->getDebugLoc();
5089 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5090 // With PIC, the address is actually $g + Offset.
5092 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5093 DAG.getNode(X86ISD::GlobalBaseReg,
5094 DebugLoc(), getPointerTy()),
5101 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5102 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5104 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5106 unsigned char OpFlag = 0;
5107 unsigned WrapperKind = X86ISD::Wrapper;
5108 CodeModel::Model M = getTargetMachine().getCodeModel();
5110 if (Subtarget->isPICStyleRIPRel() &&
5111 (M == CodeModel::Small || M == CodeModel::Kernel))
5112 WrapperKind = X86ISD::WrapperRIP;
5113 else if (Subtarget->isPICStyleGOT())
5114 OpFlag = X86II::MO_GOTOFF;
5115 else if (Subtarget->isPICStyleStubPIC())
5116 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5118 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5120 DebugLoc DL = JT->getDebugLoc();
5121 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5123 // With PIC, the address is actually $g + Offset.
5125 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5126 DAG.getNode(X86ISD::GlobalBaseReg,
5127 DebugLoc(), getPointerTy()),
5135 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5136 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5138 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5140 unsigned char OpFlag = 0;
5141 unsigned WrapperKind = X86ISD::Wrapper;
5142 CodeModel::Model M = getTargetMachine().getCodeModel();
5144 if (Subtarget->isPICStyleRIPRel() &&
5145 (M == CodeModel::Small || M == CodeModel::Kernel))
5146 WrapperKind = X86ISD::WrapperRIP;
5147 else if (Subtarget->isPICStyleGOT())
5148 OpFlag = X86II::MO_GOTOFF;
5149 else if (Subtarget->isPICStyleStubPIC())
5150 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5152 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5154 DebugLoc DL = Op.getDebugLoc();
5155 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5158 // With PIC, the address is actually $g + Offset.
5159 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5160 !Subtarget->is64Bit()) {
5161 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5162 DAG.getNode(X86ISD::GlobalBaseReg,
5163 DebugLoc(), getPointerTy()),
5171 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5172 // Create the TargetBlockAddressAddress node.
5173 unsigned char OpFlags =
5174 Subtarget->ClassifyBlockAddressReference();
5175 CodeModel::Model M = getTargetMachine().getCodeModel();
5176 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5177 DebugLoc dl = Op.getDebugLoc();
5178 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5179 /*isTarget=*/true, OpFlags);
5181 if (Subtarget->isPICStyleRIPRel() &&
5182 (M == CodeModel::Small || M == CodeModel::Kernel))
5183 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5185 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5187 // With PIC, the address is actually $g + Offset.
5188 if (isGlobalRelativeToPICBase(OpFlags)) {
5189 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5190 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5198 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5200 SelectionDAG &DAG) const {
5201 // Create the TargetGlobalAddress node, folding in the constant
5202 // offset if it is legal.
5203 unsigned char OpFlags =
5204 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5205 CodeModel::Model M = getTargetMachine().getCodeModel();
5207 if (OpFlags == X86II::MO_NO_FLAG &&
5208 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5209 // A direct static reference to a global.
5210 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5213 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5216 if (Subtarget->isPICStyleRIPRel() &&
5217 (M == CodeModel::Small || M == CodeModel::Kernel))
5218 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5220 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5222 // With PIC, the address is actually $g + Offset.
5223 if (isGlobalRelativeToPICBase(OpFlags)) {
5224 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5225 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5229 // For globals that require a load from a stub to get the address, emit the
5231 if (isGlobalStubReference(OpFlags))
5232 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5233 PseudoSourceValue::getGOT(), 0, false, false, 0);
5235 // If there was a non-zero offset that we didn't fold, create an explicit
5238 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5239 DAG.getConstant(Offset, getPointerTy()));
5245 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5246 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5247 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5248 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5252 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5253 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5254 unsigned char OperandFlags) {
5255 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5256 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5257 DebugLoc dl = GA->getDebugLoc();
5258 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5259 GA->getValueType(0),
5263 SDValue Ops[] = { Chain, TGA, *InFlag };
5264 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5266 SDValue Ops[] = { Chain, TGA };
5267 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5270 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5271 MFI->setAdjustsStack(true);
5273 SDValue Flag = Chain.getValue(1);
5274 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5277 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5279 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5282 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5283 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5284 DAG.getNode(X86ISD::GlobalBaseReg,
5285 DebugLoc(), PtrVT), InFlag);
5286 InFlag = Chain.getValue(1);
5288 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5291 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5293 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5295 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5296 X86::RAX, X86II::MO_TLSGD);
5299 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5300 // "local exec" model.
5301 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5302 const EVT PtrVT, TLSModel::Model model,
5304 DebugLoc dl = GA->getDebugLoc();
5305 // Get the Thread Pointer
5306 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5308 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5311 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5312 NULL, 0, false, false, 0);
5314 unsigned char OperandFlags = 0;
5315 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5317 unsigned WrapperKind = X86ISD::Wrapper;
5318 if (model == TLSModel::LocalExec) {
5319 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5320 } else if (is64Bit) {
5321 assert(model == TLSModel::InitialExec);
5322 OperandFlags = X86II::MO_GOTTPOFF;
5323 WrapperKind = X86ISD::WrapperRIP;
5325 assert(model == TLSModel::InitialExec);
5326 OperandFlags = X86II::MO_INDNTPOFF;
5329 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5331 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5332 GA->getOffset(), OperandFlags);
5333 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5335 if (model == TLSModel::InitialExec)
5336 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5337 PseudoSourceValue::getGOT(), 0, false, false, 0);
5339 // The address of the thread local variable is the add of the thread
5340 // pointer with the offset of the variable.
5341 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5345 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5346 // TODO: implement the "local dynamic" model
5347 // TODO: implement the "initial exec"model for pic executables
5348 assert(Subtarget->isTargetELF() &&
5349 "TLS not implemented for non-ELF targets");
5350 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5351 const GlobalValue *GV = GA->getGlobal();
5353 // If GV is an alias then use the aliasee for determining
5354 // thread-localness.
5355 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5356 GV = GA->resolveAliasedGlobal(false);
5358 TLSModel::Model model = getTLSModel(GV,
5359 getTargetMachine().getRelocationModel());
5362 case TLSModel::GeneralDynamic:
5363 case TLSModel::LocalDynamic: // not implemented
5364 if (Subtarget->is64Bit())
5365 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5366 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5368 case TLSModel::InitialExec:
5369 case TLSModel::LocalExec:
5370 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5371 Subtarget->is64Bit());
5374 llvm_unreachable("Unreachable");
5379 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5380 /// take a 2 x i32 value to shift plus a shift amount.
5381 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5382 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5383 EVT VT = Op.getValueType();
5384 unsigned VTBits = VT.getSizeInBits();
5385 DebugLoc dl = Op.getDebugLoc();
5386 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5387 SDValue ShOpLo = Op.getOperand(0);
5388 SDValue ShOpHi = Op.getOperand(1);
5389 SDValue ShAmt = Op.getOperand(2);
5390 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5391 DAG.getConstant(VTBits - 1, MVT::i8))
5392 : DAG.getConstant(0, VT);
5395 if (Op.getOpcode() == ISD::SHL_PARTS) {
5396 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5397 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5399 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5400 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5403 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5404 DAG.getConstant(VTBits, MVT::i8));
5405 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5406 AndNode, DAG.getConstant(0, MVT::i8));
5409 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5410 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5411 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5413 if (Op.getOpcode() == ISD::SHL_PARTS) {
5414 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5415 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5417 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5418 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5421 SDValue Ops[2] = { Lo, Hi };
5422 return DAG.getMergeValues(Ops, 2, dl);
5425 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5426 SelectionDAG &DAG) const {
5427 EVT SrcVT = Op.getOperand(0).getValueType();
5429 if (SrcVT.isVector()) {
5430 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5436 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5437 "Unknown SINT_TO_FP to lower!");
5439 // These are really Legal; return the operand so the caller accepts it as
5441 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5443 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5444 Subtarget->is64Bit()) {
5448 DebugLoc dl = Op.getDebugLoc();
5449 unsigned Size = SrcVT.getSizeInBits()/8;
5450 MachineFunction &MF = DAG.getMachineFunction();
5451 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5452 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5453 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5455 PseudoSourceValue::getFixedStack(SSFI), 0,
5457 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5460 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5462 SelectionDAG &DAG) const {
5464 DebugLoc dl = Op.getDebugLoc();
5466 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5468 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5470 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5471 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5472 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5473 Tys, Ops, array_lengthof(Ops));
5476 Chain = Result.getValue(1);
5477 SDValue InFlag = Result.getValue(2);
5479 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5480 // shouldn't be necessary except that RFP cannot be live across
5481 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5482 MachineFunction &MF = DAG.getMachineFunction();
5483 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5484 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5485 Tys = DAG.getVTList(MVT::Other);
5487 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5489 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5490 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5491 PseudoSourceValue::getFixedStack(SSFI), 0,
5498 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5499 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5500 SelectionDAG &DAG) const {
5501 // This algorithm is not obvious. Here it is in C code, more or less:
5503 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5504 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5505 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5507 // Copy ints to xmm registers.
5508 __m128i xh = _mm_cvtsi32_si128( hi );
5509 __m128i xl = _mm_cvtsi32_si128( lo );
5511 // Combine into low half of a single xmm register.
5512 __m128i x = _mm_unpacklo_epi32( xh, xl );
5516 // Merge in appropriate exponents to give the integer bits the right
5518 x = _mm_unpacklo_epi32( x, exp );
5520 // Subtract away the biases to deal with the IEEE-754 double precision
5522 d = _mm_sub_pd( (__m128d) x, bias );
5524 // All conversions up to here are exact. The correctly rounded result is
5525 // calculated using the current rounding mode using the following
5527 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5528 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5529 // store doesn't really need to be here (except
5530 // maybe to zero the other double)
5535 DebugLoc dl = Op.getDebugLoc();
5536 LLVMContext *Context = DAG.getContext();
5538 // Build some magic constants.
5539 std::vector<Constant*> CV0;
5540 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5541 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5542 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5543 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5544 Constant *C0 = ConstantVector::get(CV0);
5545 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5547 std::vector<Constant*> CV1;
5549 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5551 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5552 Constant *C1 = ConstantVector::get(CV1);
5553 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5555 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5556 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5558 DAG.getIntPtrConstant(1)));
5559 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5560 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5562 DAG.getIntPtrConstant(0)));
5563 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5564 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5565 PseudoSourceValue::getConstantPool(), 0,
5567 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5568 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5569 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5570 PseudoSourceValue::getConstantPool(), 0,
5572 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5574 // Add the halves; easiest way is to swap them into another reg first.
5575 int ShufMask[2] = { 1, -1 };
5576 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5577 DAG.getUNDEF(MVT::v2f64), ShufMask);
5578 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5579 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5580 DAG.getIntPtrConstant(0));
5583 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5584 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5585 SelectionDAG &DAG) const {
5586 DebugLoc dl = Op.getDebugLoc();
5587 // FP constant to bias correct the final result.
5588 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5591 // Load the 32-bit value into an XMM register.
5592 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5593 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5595 DAG.getIntPtrConstant(0)));
5597 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5598 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5599 DAG.getIntPtrConstant(0));
5601 // Or the load with the bias.
5602 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5603 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5604 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5606 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5607 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5608 MVT::v2f64, Bias)));
5609 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5610 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5611 DAG.getIntPtrConstant(0));
5613 // Subtract the bias.
5614 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5616 // Handle final rounding.
5617 EVT DestVT = Op.getValueType();
5619 if (DestVT.bitsLT(MVT::f64)) {
5620 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5621 DAG.getIntPtrConstant(0));
5622 } else if (DestVT.bitsGT(MVT::f64)) {
5623 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5626 // Handle final rounding.
5630 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5631 SelectionDAG &DAG) const {
5632 SDValue N0 = Op.getOperand(0);
5633 DebugLoc dl = Op.getDebugLoc();
5635 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5636 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5637 // the optimization here.
5638 if (DAG.SignBitIsZero(N0))
5639 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5641 EVT SrcVT = N0.getValueType();
5642 EVT DstVT = Op.getValueType();
5643 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5644 return LowerUINT_TO_FP_i64(Op, DAG);
5645 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5646 return LowerUINT_TO_FP_i32(Op, DAG);
5648 // Make a 64-bit buffer, and use it to build an FILD.
5649 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5650 if (SrcVT == MVT::i32) {
5651 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5652 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5653 getPointerTy(), StackSlot, WordOff);
5654 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5655 StackSlot, NULL, 0, false, false, 0);
5656 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5657 OffsetSlot, NULL, 0, false, false, 0);
5658 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5662 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5663 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5664 StackSlot, NULL, 0, false, false, 0);
5665 // For i64 source, we need to add the appropriate power of 2 if the input
5666 // was negative. This is the same as the optimization in
5667 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5668 // we must be careful to do the computation in x87 extended precision, not
5669 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5670 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5671 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5672 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5674 APInt FF(32, 0x5F800000ULL);
5676 // Check whether the sign bit is set.
5677 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5678 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5681 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5682 SDValue FudgePtr = DAG.getConstantPool(
5683 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5686 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5687 SDValue Zero = DAG.getIntPtrConstant(0);
5688 SDValue Four = DAG.getIntPtrConstant(4);
5689 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5691 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5693 // Load the value out, extending it from f32 to f80.
5694 // FIXME: Avoid the extend by constructing the right constant pool?
5695 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
5696 FudgePtr, PseudoSourceValue::getConstantPool(),
5697 0, MVT::f32, false, false, 4);
5698 // Extend everything to 80 bits to force it to be done on x87.
5699 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5700 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5703 std::pair<SDValue,SDValue> X86TargetLowering::
5704 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5705 DebugLoc dl = Op.getDebugLoc();
5707 EVT DstTy = Op.getValueType();
5710 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5714 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5715 DstTy.getSimpleVT() >= MVT::i16 &&
5716 "Unknown FP_TO_SINT to lower!");
5718 // These are really Legal.
5719 if (DstTy == MVT::i32 &&
5720 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5721 return std::make_pair(SDValue(), SDValue());
5722 if (Subtarget->is64Bit() &&
5723 DstTy == MVT::i64 &&
5724 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5725 return std::make_pair(SDValue(), SDValue());
5727 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5729 MachineFunction &MF = DAG.getMachineFunction();
5730 unsigned MemSize = DstTy.getSizeInBits()/8;
5731 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5732 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5735 switch (DstTy.getSimpleVT().SimpleTy) {
5736 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5737 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5738 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5739 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5742 SDValue Chain = DAG.getEntryNode();
5743 SDValue Value = Op.getOperand(0);
5744 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5745 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5746 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5747 PseudoSourceValue::getFixedStack(SSFI), 0,
5749 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5751 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5753 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5754 Chain = Value.getValue(1);
5755 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5756 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5759 // Build the FP_TO_INT*_IN_MEM
5760 SDValue Ops[] = { Chain, Value, StackSlot };
5761 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5763 return std::make_pair(FIST, StackSlot);
5766 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5767 SelectionDAG &DAG) const {
5768 if (Op.getValueType().isVector()) {
5769 if (Op.getValueType() == MVT::v2i32 &&
5770 Op.getOperand(0).getValueType() == MVT::v2f64) {
5776 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5777 SDValue FIST = Vals.first, StackSlot = Vals.second;
5778 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5779 if (FIST.getNode() == 0) return Op;
5782 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5783 FIST, StackSlot, NULL, 0, false, false, 0);
5786 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5787 SelectionDAG &DAG) const {
5788 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5789 SDValue FIST = Vals.first, StackSlot = Vals.second;
5790 assert(FIST.getNode() && "Unexpected failure");
5793 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5794 FIST, StackSlot, NULL, 0, false, false, 0);
5797 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5798 SelectionDAG &DAG) const {
5799 LLVMContext *Context = DAG.getContext();
5800 DebugLoc dl = Op.getDebugLoc();
5801 EVT VT = Op.getValueType();
5804 EltVT = VT.getVectorElementType();
5805 std::vector<Constant*> CV;
5806 if (EltVT == MVT::f64) {
5807 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5811 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5817 Constant *C = ConstantVector::get(CV);
5818 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5819 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5820 PseudoSourceValue::getConstantPool(), 0,
5822 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5825 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5826 LLVMContext *Context = DAG.getContext();
5827 DebugLoc dl = Op.getDebugLoc();
5828 EVT VT = Op.getValueType();
5831 EltVT = VT.getVectorElementType();
5832 std::vector<Constant*> CV;
5833 if (EltVT == MVT::f64) {
5834 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5838 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5844 Constant *C = ConstantVector::get(CV);
5845 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5846 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5847 PseudoSourceValue::getConstantPool(), 0,
5849 if (VT.isVector()) {
5850 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5851 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5852 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5854 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5856 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5860 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5861 LLVMContext *Context = DAG.getContext();
5862 SDValue Op0 = Op.getOperand(0);
5863 SDValue Op1 = Op.getOperand(1);
5864 DebugLoc dl = Op.getDebugLoc();
5865 EVT VT = Op.getValueType();
5866 EVT SrcVT = Op1.getValueType();
5868 // If second operand is smaller, extend it first.
5869 if (SrcVT.bitsLT(VT)) {
5870 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5873 // And if it is bigger, shrink it first.
5874 if (SrcVT.bitsGT(VT)) {
5875 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5879 // At this point the operands and the result should have the same
5880 // type, and that won't be f80 since that is not custom lowered.
5882 // First get the sign bit of second operand.
5883 std::vector<Constant*> CV;
5884 if (SrcVT == MVT::f64) {
5885 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5886 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5888 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5889 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5890 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5891 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5893 Constant *C = ConstantVector::get(CV);
5894 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5895 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5896 PseudoSourceValue::getConstantPool(), 0,
5898 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5900 // Shift sign bit right or left if the two operands have different types.
5901 if (SrcVT.bitsGT(VT)) {
5902 // Op0 is MVT::f32, Op1 is MVT::f64.
5903 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5904 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5905 DAG.getConstant(32, MVT::i32));
5906 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5907 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5908 DAG.getIntPtrConstant(0));
5911 // Clear first operand sign bit.
5913 if (VT == MVT::f64) {
5914 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5915 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5917 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5918 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5919 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5920 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5922 C = ConstantVector::get(CV);
5923 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5924 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5925 PseudoSourceValue::getConstantPool(), 0,
5927 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5929 // Or the value with the sign bit.
5930 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5933 /// Emit nodes that will be selected as "test Op0,Op0", or something
5935 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5936 SelectionDAG &DAG) const {
5937 DebugLoc dl = Op.getDebugLoc();
5939 // CF and OF aren't always set the way we want. Determine which
5940 // of these we need.
5941 bool NeedCF = false;
5942 bool NeedOF = false;
5944 case X86::COND_A: case X86::COND_AE:
5945 case X86::COND_B: case X86::COND_BE:
5948 case X86::COND_G: case X86::COND_GE:
5949 case X86::COND_L: case X86::COND_LE:
5950 case X86::COND_O: case X86::COND_NO:
5956 // See if we can use the EFLAGS value from the operand instead of
5957 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5958 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5959 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5960 unsigned Opcode = 0;
5961 unsigned NumOperands = 0;
5962 switch (Op.getNode()->getOpcode()) {
5964 // Due to an isel shortcoming, be conservative if this add is
5965 // likely to be selected as part of a load-modify-store
5966 // instruction. When the root node in a match is a store, isel
5967 // doesn't know how to remap non-chain non-flag uses of other
5968 // nodes in the match, such as the ADD in this case. This leads
5969 // to the ADD being left around and reselected, with the result
5970 // being two adds in the output. Alas, even if none our users
5971 // are stores, that doesn't prove we're O.K. Ergo, if we have
5972 // any parents that aren't CopyToReg or SETCC, eschew INC/DEC.
5973 // A better fix seems to require climbing the DAG back to the
5974 // root, and it doesn't seem to be worth the effort.
5975 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5976 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5977 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
5979 if (ConstantSDNode *C =
5980 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5981 // An add of one will be selected as an INC.
5982 if (C->getAPIntValue() == 1) {
5983 Opcode = X86ISD::INC;
5987 // An add of negative one (subtract of one) will be selected as a DEC.
5988 if (C->getAPIntValue().isAllOnesValue()) {
5989 Opcode = X86ISD::DEC;
5994 // Otherwise use a regular EFLAGS-setting add.
5995 Opcode = X86ISD::ADD;
5999 // If the primary and result isn't used, don't bother using X86ISD::AND,
6000 // because a TEST instruction will be better.
6001 bool NonFlagUse = false;
6002 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6003 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6005 unsigned UOpNo = UI.getOperandNo();
6006 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6007 // Look pass truncate.
6008 UOpNo = User->use_begin().getOperandNo();
6009 User = *User->use_begin();
6011 if (User->getOpcode() != ISD::BRCOND &&
6012 User->getOpcode() != ISD::SETCC &&
6013 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6025 // Due to the ISEL shortcoming noted above, be conservative if this op is
6026 // likely to be selected as part of a load-modify-store instruction.
6027 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6028 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6029 if (UI->getOpcode() == ISD::STORE)
6031 // Otherwise use a regular EFLAGS-setting instruction.
6032 switch (Op.getNode()->getOpcode()) {
6033 case ISD::SUB: Opcode = X86ISD::SUB; break;
6034 case ISD::OR: Opcode = X86ISD::OR; break;
6035 case ISD::XOR: Opcode = X86ISD::XOR; break;
6036 case ISD::AND: Opcode = X86ISD::AND; break;
6037 default: llvm_unreachable("unexpected operator!");
6048 return SDValue(Op.getNode(), 1);
6054 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6055 SmallVector<SDValue, 4> Ops;
6056 for (unsigned i = 0; i != NumOperands; ++i)
6057 Ops.push_back(Op.getOperand(i));
6058 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6059 DAG.ReplaceAllUsesWith(Op, New);
6060 return SDValue(New.getNode(), 1);
6064 // Otherwise just emit a CMP with 0, which is the TEST pattern.
6065 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6066 DAG.getConstant(0, Op.getValueType()));
6069 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6071 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6072 SelectionDAG &DAG) const {
6073 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6074 if (C->getAPIntValue() == 0)
6075 return EmitTest(Op0, X86CC, DAG);
6077 DebugLoc dl = Op0.getDebugLoc();
6078 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6081 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6082 /// if it's possible.
6083 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6084 DebugLoc dl, SelectionDAG &DAG) const {
6085 SDValue Op0 = And.getOperand(0);
6086 SDValue Op1 = And.getOperand(1);
6087 if (Op0.getOpcode() == ISD::TRUNCATE)
6088 Op0 = Op0.getOperand(0);
6089 if (Op1.getOpcode() == ISD::TRUNCATE)
6090 Op1 = Op1.getOperand(0);
6093 if (Op1.getOpcode() == ISD::SHL) {
6094 if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
6095 if (And10C->getZExtValue() == 1) {
6097 RHS = Op1.getOperand(1);
6099 } else if (Op0.getOpcode() == ISD::SHL) {
6100 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6101 if (And00C->getZExtValue() == 1) {
6103 RHS = Op0.getOperand(1);
6105 } else if (Op1.getOpcode() == ISD::Constant) {
6106 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6107 SDValue AndLHS = Op0;
6108 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6109 LHS = AndLHS.getOperand(0);
6110 RHS = AndLHS.getOperand(1);
6114 if (LHS.getNode()) {
6115 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6116 // instruction. Since the shift amount is in-range-or-undefined, we know
6117 // that doing a bittest on the i32 value is ok. We extend to i32 because
6118 // the encoding for the i16 version is larger than the i32 version.
6119 // Also promote i16 to i32 for performance / code size reason.
6120 if (LHS.getValueType() == MVT::i8 ||
6121 LHS.getValueType() == MVT::i16)
6122 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6124 // If the operand types disagree, extend the shift amount to match. Since
6125 // BT ignores high bits (like shifts) we can use anyextend.
6126 if (LHS.getValueType() != RHS.getValueType())
6127 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6129 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6130 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6131 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6132 DAG.getConstant(Cond, MVT::i8), BT);
6138 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6139 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6140 SDValue Op0 = Op.getOperand(0);
6141 SDValue Op1 = Op.getOperand(1);
6142 DebugLoc dl = Op.getDebugLoc();
6143 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6145 // Optimize to BT if possible.
6146 // Lower (X & (1 << N)) == 0 to BT(X, N).
6147 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6148 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6149 if (Op0.getOpcode() == ISD::AND &&
6151 Op1.getOpcode() == ISD::Constant &&
6152 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
6153 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6154 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6155 if (NewSetCC.getNode())
6159 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6160 if (Op0.getOpcode() == X86ISD::SETCC &&
6161 Op1.getOpcode() == ISD::Constant &&
6162 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6163 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6164 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6165 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6166 bool Invert = (CC == ISD::SETNE) ^
6167 cast<ConstantSDNode>(Op1)->isNullValue();
6169 CCode = X86::GetOppositeBranchCondition(CCode);
6170 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6171 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6174 bool isFP = Op1.getValueType().isFloatingPoint();
6175 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6176 if (X86CC == X86::COND_INVALID)
6179 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6181 // Use sbb x, x to materialize carry bit into a GPR.
6182 if (X86CC == X86::COND_B)
6183 return DAG.getNode(ISD::AND, dl, MVT::i8,
6184 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6185 DAG.getConstant(X86CC, MVT::i8), Cond),
6186 DAG.getConstant(1, MVT::i8));
6188 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6189 DAG.getConstant(X86CC, MVT::i8), Cond);
6192 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6194 SDValue Op0 = Op.getOperand(0);
6195 SDValue Op1 = Op.getOperand(1);
6196 SDValue CC = Op.getOperand(2);
6197 EVT VT = Op.getValueType();
6198 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6199 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6200 DebugLoc dl = Op.getDebugLoc();
6204 EVT VT0 = Op0.getValueType();
6205 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6206 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6209 switch (SetCCOpcode) {
6212 case ISD::SETEQ: SSECC = 0; break;
6214 case ISD::SETGT: Swap = true; // Fallthrough
6216 case ISD::SETOLT: SSECC = 1; break;
6218 case ISD::SETGE: Swap = true; // Fallthrough
6220 case ISD::SETOLE: SSECC = 2; break;
6221 case ISD::SETUO: SSECC = 3; break;
6223 case ISD::SETNE: SSECC = 4; break;
6224 case ISD::SETULE: Swap = true;
6225 case ISD::SETUGE: SSECC = 5; break;
6226 case ISD::SETULT: Swap = true;
6227 case ISD::SETUGT: SSECC = 6; break;
6228 case ISD::SETO: SSECC = 7; break;
6231 std::swap(Op0, Op1);
6233 // In the two special cases we can't handle, emit two comparisons.
6235 if (SetCCOpcode == ISD::SETUEQ) {
6237 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6238 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6239 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6241 else if (SetCCOpcode == ISD::SETONE) {
6243 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6244 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6245 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6247 llvm_unreachable("Illegal FP comparison");
6249 // Handle all other FP comparisons here.
6250 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6253 // We are handling one of the integer comparisons here. Since SSE only has
6254 // GT and EQ comparisons for integer, swapping operands and multiple
6255 // operations may be required for some comparisons.
6256 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6257 bool Swap = false, Invert = false, FlipSigns = false;
6259 switch (VT.getSimpleVT().SimpleTy) {
6262 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6264 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6266 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6267 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6270 switch (SetCCOpcode) {
6272 case ISD::SETNE: Invert = true;
6273 case ISD::SETEQ: Opc = EQOpc; break;
6274 case ISD::SETLT: Swap = true;
6275 case ISD::SETGT: Opc = GTOpc; break;
6276 case ISD::SETGE: Swap = true;
6277 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6278 case ISD::SETULT: Swap = true;
6279 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6280 case ISD::SETUGE: Swap = true;
6281 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6284 std::swap(Op0, Op1);
6286 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6287 // bits of the inputs before performing those operations.
6289 EVT EltVT = VT.getVectorElementType();
6290 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6292 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6293 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6295 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6296 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6299 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6301 // If the logical-not of the result is required, perform that now.
6303 Result = DAG.getNOT(dl, Result, VT);
6308 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6309 static bool isX86LogicalCmp(SDValue Op) {
6310 unsigned Opc = Op.getNode()->getOpcode();
6311 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6313 if (Op.getResNo() == 1 &&
6314 (Opc == X86ISD::ADD ||
6315 Opc == X86ISD::SUB ||
6316 Opc == X86ISD::SMUL ||
6317 Opc == X86ISD::UMUL ||
6318 Opc == X86ISD::INC ||
6319 Opc == X86ISD::DEC ||
6320 Opc == X86ISD::OR ||
6321 Opc == X86ISD::XOR ||
6322 Opc == X86ISD::AND))
6328 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6329 bool addTest = true;
6330 SDValue Cond = Op.getOperand(0);
6331 DebugLoc dl = Op.getDebugLoc();
6334 if (Cond.getOpcode() == ISD::SETCC) {
6335 SDValue NewCond = LowerSETCC(Cond, DAG);
6336 if (NewCond.getNode())
6340 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6341 SDValue Op1 = Op.getOperand(1);
6342 SDValue Op2 = Op.getOperand(2);
6343 if (Cond.getOpcode() == X86ISD::SETCC &&
6344 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6345 SDValue Cmp = Cond.getOperand(1);
6346 if (Cmp.getOpcode() == X86ISD::CMP) {
6347 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6348 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6349 ConstantSDNode *RHSC =
6350 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6351 if (N1C && N1C->isAllOnesValue() &&
6352 N2C && N2C->isNullValue() &&
6353 RHSC && RHSC->isNullValue()) {
6354 SDValue CmpOp0 = Cmp.getOperand(0);
6355 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6356 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6357 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6358 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6363 // Look pass (and (setcc_carry (cmp ...)), 1).
6364 if (Cond.getOpcode() == ISD::AND &&
6365 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6366 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6367 if (C && C->getAPIntValue() == 1)
6368 Cond = Cond.getOperand(0);
6371 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6372 // setting operand in place of the X86ISD::SETCC.
6373 if (Cond.getOpcode() == X86ISD::SETCC ||
6374 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6375 CC = Cond.getOperand(0);
6377 SDValue Cmp = Cond.getOperand(1);
6378 unsigned Opc = Cmp.getOpcode();
6379 EVT VT = Op.getValueType();
6381 bool IllegalFPCMov = false;
6382 if (VT.isFloatingPoint() && !VT.isVector() &&
6383 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6384 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6386 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6387 Opc == X86ISD::BT) { // FIXME
6394 // Look pass the truncate.
6395 if (Cond.getOpcode() == ISD::TRUNCATE)
6396 Cond = Cond.getOperand(0);
6398 // We know the result of AND is compared against zero. Try to match
6400 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6401 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6402 if (NewSetCC.getNode()) {
6403 CC = NewSetCC.getOperand(0);
6404 Cond = NewSetCC.getOperand(1);
6411 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6412 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6415 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6416 // condition is true.
6417 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6418 SDValue Ops[] = { Op2, Op1, CC, Cond };
6419 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6422 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6423 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6424 // from the AND / OR.
6425 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6426 Opc = Op.getOpcode();
6427 if (Opc != ISD::OR && Opc != ISD::AND)
6429 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6430 Op.getOperand(0).hasOneUse() &&
6431 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6432 Op.getOperand(1).hasOneUse());
6435 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6436 // 1 and that the SETCC node has a single use.
6437 static bool isXor1OfSetCC(SDValue Op) {
6438 if (Op.getOpcode() != ISD::XOR)
6440 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6441 if (N1C && N1C->getAPIntValue() == 1) {
6442 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6443 Op.getOperand(0).hasOneUse();
6448 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6449 bool addTest = true;
6450 SDValue Chain = Op.getOperand(0);
6451 SDValue Cond = Op.getOperand(1);
6452 SDValue Dest = Op.getOperand(2);
6453 DebugLoc dl = Op.getDebugLoc();
6456 if (Cond.getOpcode() == ISD::SETCC) {
6457 SDValue NewCond = LowerSETCC(Cond, DAG);
6458 if (NewCond.getNode())
6462 // FIXME: LowerXALUO doesn't handle these!!
6463 else if (Cond.getOpcode() == X86ISD::ADD ||
6464 Cond.getOpcode() == X86ISD::SUB ||
6465 Cond.getOpcode() == X86ISD::SMUL ||
6466 Cond.getOpcode() == X86ISD::UMUL)
6467 Cond = LowerXALUO(Cond, DAG);
6470 // Look pass (and (setcc_carry (cmp ...)), 1).
6471 if (Cond.getOpcode() == ISD::AND &&
6472 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6473 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6474 if (C && C->getAPIntValue() == 1)
6475 Cond = Cond.getOperand(0);
6478 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6479 // setting operand in place of the X86ISD::SETCC.
6480 if (Cond.getOpcode() == X86ISD::SETCC ||
6481 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6482 CC = Cond.getOperand(0);
6484 SDValue Cmp = Cond.getOperand(1);
6485 unsigned Opc = Cmp.getOpcode();
6486 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6487 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6491 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6495 // These can only come from an arithmetic instruction with overflow,
6496 // e.g. SADDO, UADDO.
6497 Cond = Cond.getNode()->getOperand(1);
6504 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6505 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6506 if (CondOpc == ISD::OR) {
6507 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6508 // two branches instead of an explicit OR instruction with a
6510 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6511 isX86LogicalCmp(Cmp)) {
6512 CC = Cond.getOperand(0).getOperand(0);
6513 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6514 Chain, Dest, CC, Cmp);
6515 CC = Cond.getOperand(1).getOperand(0);
6519 } else { // ISD::AND
6520 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6521 // two branches instead of an explicit AND instruction with a
6522 // separate test. However, we only do this if this block doesn't
6523 // have a fall-through edge, because this requires an explicit
6524 // jmp when the condition is false.
6525 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6526 isX86LogicalCmp(Cmp) &&
6527 Op.getNode()->hasOneUse()) {
6528 X86::CondCode CCode =
6529 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6530 CCode = X86::GetOppositeBranchCondition(CCode);
6531 CC = DAG.getConstant(CCode, MVT::i8);
6532 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6533 // Look for an unconditional branch following this conditional branch.
6534 // We need this because we need to reverse the successors in order
6535 // to implement FCMP_OEQ.
6536 if (User.getOpcode() == ISD::BR) {
6537 SDValue FalseBB = User.getOperand(1);
6539 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6540 assert(NewBR == User);
6543 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6544 Chain, Dest, CC, Cmp);
6545 X86::CondCode CCode =
6546 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6547 CCode = X86::GetOppositeBranchCondition(CCode);
6548 CC = DAG.getConstant(CCode, MVT::i8);
6554 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6555 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6556 // It should be transformed during dag combiner except when the condition
6557 // is set by a arithmetics with overflow node.
6558 X86::CondCode CCode =
6559 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6560 CCode = X86::GetOppositeBranchCondition(CCode);
6561 CC = DAG.getConstant(CCode, MVT::i8);
6562 Cond = Cond.getOperand(0).getOperand(1);
6568 // Look pass the truncate.
6569 if (Cond.getOpcode() == ISD::TRUNCATE)
6570 Cond = Cond.getOperand(0);
6572 // We know the result of AND is compared against zero. Try to match
6574 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6575 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6576 if (NewSetCC.getNode()) {
6577 CC = NewSetCC.getOperand(0);
6578 Cond = NewSetCC.getOperand(1);
6585 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6586 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6588 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6589 Chain, Dest, CC, Cond);
6593 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6594 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6595 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6596 // that the guard pages used by the OS virtual memory manager are allocated in
6597 // correct sequence.
6599 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6600 SelectionDAG &DAG) const {
6601 assert(Subtarget->isTargetCygMing() &&
6602 "This should be used only on Cygwin/Mingw targets");
6603 DebugLoc dl = Op.getDebugLoc();
6606 SDValue Chain = Op.getOperand(0);
6607 SDValue Size = Op.getOperand(1);
6608 // FIXME: Ensure alignment here
6612 EVT IntPtr = getPointerTy();
6613 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6615 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6616 Flag = Chain.getValue(1);
6618 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6620 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6621 Flag = Chain.getValue(1);
6623 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6625 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6626 return DAG.getMergeValues(Ops1, 2, dl);
6629 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6630 MachineFunction &MF = DAG.getMachineFunction();
6631 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6633 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6634 DebugLoc dl = Op.getDebugLoc();
6636 if (!Subtarget->is64Bit()) {
6637 // vastart just stores the address of the VarArgsFrameIndex slot into the
6638 // memory location argument.
6639 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6641 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6646 // gp_offset (0 - 6 * 8)
6647 // fp_offset (48 - 48 + 8 * 16)
6648 // overflow_arg_area (point to parameters coming in memory).
6650 SmallVector<SDValue, 8> MemOps;
6651 SDValue FIN = Op.getOperand(1);
6653 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6654 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6656 FIN, SV, 0, false, false, 0);
6657 MemOps.push_back(Store);
6660 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6661 FIN, DAG.getIntPtrConstant(4));
6662 Store = DAG.getStore(Op.getOperand(0), dl,
6663 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6665 FIN, SV, 0, false, false, 0);
6666 MemOps.push_back(Store);
6668 // Store ptr to overflow_arg_area
6669 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6670 FIN, DAG.getIntPtrConstant(4));
6671 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6673 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6675 MemOps.push_back(Store);
6677 // Store ptr to reg_save_area.
6678 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6679 FIN, DAG.getIntPtrConstant(8));
6680 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6682 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6684 MemOps.push_back(Store);
6685 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6686 &MemOps[0], MemOps.size());
6689 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6690 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6691 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6692 SDValue Chain = Op.getOperand(0);
6693 SDValue SrcPtr = Op.getOperand(1);
6694 SDValue SrcSV = Op.getOperand(2);
6696 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6700 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6701 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6702 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6703 SDValue Chain = Op.getOperand(0);
6704 SDValue DstPtr = Op.getOperand(1);
6705 SDValue SrcPtr = Op.getOperand(2);
6706 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6707 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6708 DebugLoc dl = Op.getDebugLoc();
6710 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6711 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6712 false, DstSV, 0, SrcSV, 0);
6716 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6717 DebugLoc dl = Op.getDebugLoc();
6718 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6720 default: return SDValue(); // Don't custom lower most intrinsics.
6721 // Comparison intrinsics.
6722 case Intrinsic::x86_sse_comieq_ss:
6723 case Intrinsic::x86_sse_comilt_ss:
6724 case Intrinsic::x86_sse_comile_ss:
6725 case Intrinsic::x86_sse_comigt_ss:
6726 case Intrinsic::x86_sse_comige_ss:
6727 case Intrinsic::x86_sse_comineq_ss:
6728 case Intrinsic::x86_sse_ucomieq_ss:
6729 case Intrinsic::x86_sse_ucomilt_ss:
6730 case Intrinsic::x86_sse_ucomile_ss:
6731 case Intrinsic::x86_sse_ucomigt_ss:
6732 case Intrinsic::x86_sse_ucomige_ss:
6733 case Intrinsic::x86_sse_ucomineq_ss:
6734 case Intrinsic::x86_sse2_comieq_sd:
6735 case Intrinsic::x86_sse2_comilt_sd:
6736 case Intrinsic::x86_sse2_comile_sd:
6737 case Intrinsic::x86_sse2_comigt_sd:
6738 case Intrinsic::x86_sse2_comige_sd:
6739 case Intrinsic::x86_sse2_comineq_sd:
6740 case Intrinsic::x86_sse2_ucomieq_sd:
6741 case Intrinsic::x86_sse2_ucomilt_sd:
6742 case Intrinsic::x86_sse2_ucomile_sd:
6743 case Intrinsic::x86_sse2_ucomigt_sd:
6744 case Intrinsic::x86_sse2_ucomige_sd:
6745 case Intrinsic::x86_sse2_ucomineq_sd: {
6747 ISD::CondCode CC = ISD::SETCC_INVALID;
6750 case Intrinsic::x86_sse_comieq_ss:
6751 case Intrinsic::x86_sse2_comieq_sd:
6755 case Intrinsic::x86_sse_comilt_ss:
6756 case Intrinsic::x86_sse2_comilt_sd:
6760 case Intrinsic::x86_sse_comile_ss:
6761 case Intrinsic::x86_sse2_comile_sd:
6765 case Intrinsic::x86_sse_comigt_ss:
6766 case Intrinsic::x86_sse2_comigt_sd:
6770 case Intrinsic::x86_sse_comige_ss:
6771 case Intrinsic::x86_sse2_comige_sd:
6775 case Intrinsic::x86_sse_comineq_ss:
6776 case Intrinsic::x86_sse2_comineq_sd:
6780 case Intrinsic::x86_sse_ucomieq_ss:
6781 case Intrinsic::x86_sse2_ucomieq_sd:
6782 Opc = X86ISD::UCOMI;
6785 case Intrinsic::x86_sse_ucomilt_ss:
6786 case Intrinsic::x86_sse2_ucomilt_sd:
6787 Opc = X86ISD::UCOMI;
6790 case Intrinsic::x86_sse_ucomile_ss:
6791 case Intrinsic::x86_sse2_ucomile_sd:
6792 Opc = X86ISD::UCOMI;
6795 case Intrinsic::x86_sse_ucomigt_ss:
6796 case Intrinsic::x86_sse2_ucomigt_sd:
6797 Opc = X86ISD::UCOMI;
6800 case Intrinsic::x86_sse_ucomige_ss:
6801 case Intrinsic::x86_sse2_ucomige_sd:
6802 Opc = X86ISD::UCOMI;
6805 case Intrinsic::x86_sse_ucomineq_ss:
6806 case Intrinsic::x86_sse2_ucomineq_sd:
6807 Opc = X86ISD::UCOMI;
6812 SDValue LHS = Op.getOperand(1);
6813 SDValue RHS = Op.getOperand(2);
6814 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6815 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6816 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6817 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6818 DAG.getConstant(X86CC, MVT::i8), Cond);
6819 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6821 // ptest intrinsics. The intrinsic these come from are designed to return
6822 // an integer value, not just an instruction so lower it to the ptest
6823 // pattern and a setcc for the result.
6824 case Intrinsic::x86_sse41_ptestz:
6825 case Intrinsic::x86_sse41_ptestc:
6826 case Intrinsic::x86_sse41_ptestnzc:{
6829 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6830 case Intrinsic::x86_sse41_ptestz:
6832 X86CC = X86::COND_E;
6834 case Intrinsic::x86_sse41_ptestc:
6836 X86CC = X86::COND_B;
6838 case Intrinsic::x86_sse41_ptestnzc:
6840 X86CC = X86::COND_A;
6844 SDValue LHS = Op.getOperand(1);
6845 SDValue RHS = Op.getOperand(2);
6846 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6847 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6848 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6849 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6852 // Fix vector shift instructions where the last operand is a non-immediate
6854 case Intrinsic::x86_sse2_pslli_w:
6855 case Intrinsic::x86_sse2_pslli_d:
6856 case Intrinsic::x86_sse2_pslli_q:
6857 case Intrinsic::x86_sse2_psrli_w:
6858 case Intrinsic::x86_sse2_psrli_d:
6859 case Intrinsic::x86_sse2_psrli_q:
6860 case Intrinsic::x86_sse2_psrai_w:
6861 case Intrinsic::x86_sse2_psrai_d:
6862 case Intrinsic::x86_mmx_pslli_w:
6863 case Intrinsic::x86_mmx_pslli_d:
6864 case Intrinsic::x86_mmx_pslli_q:
6865 case Intrinsic::x86_mmx_psrli_w:
6866 case Intrinsic::x86_mmx_psrli_d:
6867 case Intrinsic::x86_mmx_psrli_q:
6868 case Intrinsic::x86_mmx_psrai_w:
6869 case Intrinsic::x86_mmx_psrai_d: {
6870 SDValue ShAmt = Op.getOperand(2);
6871 if (isa<ConstantSDNode>(ShAmt))
6874 unsigned NewIntNo = 0;
6875 EVT ShAmtVT = MVT::v4i32;
6877 case Intrinsic::x86_sse2_pslli_w:
6878 NewIntNo = Intrinsic::x86_sse2_psll_w;
6880 case Intrinsic::x86_sse2_pslli_d:
6881 NewIntNo = Intrinsic::x86_sse2_psll_d;
6883 case Intrinsic::x86_sse2_pslli_q:
6884 NewIntNo = Intrinsic::x86_sse2_psll_q;
6886 case Intrinsic::x86_sse2_psrli_w:
6887 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6889 case Intrinsic::x86_sse2_psrli_d:
6890 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6892 case Intrinsic::x86_sse2_psrli_q:
6893 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6895 case Intrinsic::x86_sse2_psrai_w:
6896 NewIntNo = Intrinsic::x86_sse2_psra_w;
6898 case Intrinsic::x86_sse2_psrai_d:
6899 NewIntNo = Intrinsic::x86_sse2_psra_d;
6902 ShAmtVT = MVT::v2i32;
6904 case Intrinsic::x86_mmx_pslli_w:
6905 NewIntNo = Intrinsic::x86_mmx_psll_w;
6907 case Intrinsic::x86_mmx_pslli_d:
6908 NewIntNo = Intrinsic::x86_mmx_psll_d;
6910 case Intrinsic::x86_mmx_pslli_q:
6911 NewIntNo = Intrinsic::x86_mmx_psll_q;
6913 case Intrinsic::x86_mmx_psrli_w:
6914 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6916 case Intrinsic::x86_mmx_psrli_d:
6917 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6919 case Intrinsic::x86_mmx_psrli_q:
6920 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6922 case Intrinsic::x86_mmx_psrai_w:
6923 NewIntNo = Intrinsic::x86_mmx_psra_w;
6925 case Intrinsic::x86_mmx_psrai_d:
6926 NewIntNo = Intrinsic::x86_mmx_psra_d;
6928 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6934 // The vector shift intrinsics with scalars uses 32b shift amounts but
6935 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6939 ShOps[1] = DAG.getConstant(0, MVT::i32);
6940 if (ShAmtVT == MVT::v4i32) {
6941 ShOps[2] = DAG.getUNDEF(MVT::i32);
6942 ShOps[3] = DAG.getUNDEF(MVT::i32);
6943 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6945 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6948 EVT VT = Op.getValueType();
6949 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6950 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6951 DAG.getConstant(NewIntNo, MVT::i32),
6952 Op.getOperand(1), ShAmt);
6957 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
6958 SelectionDAG &DAG) const {
6959 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6960 MFI->setReturnAddressIsTaken(true);
6962 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6963 DebugLoc dl = Op.getDebugLoc();
6966 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6968 DAG.getConstant(TD->getPointerSize(),
6969 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6970 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6971 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6973 NULL, 0, false, false, 0);
6976 // Just load the return address.
6977 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6978 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6979 RetAddrFI, NULL, 0, false, false, 0);
6982 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
6983 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6984 MFI->setFrameAddressIsTaken(true);
6986 EVT VT = Op.getValueType();
6987 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6988 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6989 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6990 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6992 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
6997 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6998 SelectionDAG &DAG) const {
6999 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7002 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7003 MachineFunction &MF = DAG.getMachineFunction();
7004 SDValue Chain = Op.getOperand(0);
7005 SDValue Offset = Op.getOperand(1);
7006 SDValue Handler = Op.getOperand(2);
7007 DebugLoc dl = Op.getDebugLoc();
7009 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7011 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7013 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7014 DAG.getIntPtrConstant(-TD->getPointerSize()));
7015 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7016 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7017 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7018 MF.getRegInfo().addLiveOut(StoreAddrReg);
7020 return DAG.getNode(X86ISD::EH_RETURN, dl,
7022 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7025 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7026 SelectionDAG &DAG) const {
7027 SDValue Root = Op.getOperand(0);
7028 SDValue Trmp = Op.getOperand(1); // trampoline
7029 SDValue FPtr = Op.getOperand(2); // nested function
7030 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7031 DebugLoc dl = Op.getDebugLoc();
7033 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7035 if (Subtarget->is64Bit()) {
7036 SDValue OutChains[6];
7038 // Large code-model.
7039 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7040 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7042 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7043 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7045 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7047 // Load the pointer to the nested function into R11.
7048 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7049 SDValue Addr = Trmp;
7050 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7051 Addr, TrmpAddr, 0, false, false, 0);
7053 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7054 DAG.getConstant(2, MVT::i64));
7055 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7058 // Load the 'nest' parameter value into R10.
7059 // R10 is specified in X86CallingConv.td
7060 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7061 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7062 DAG.getConstant(10, MVT::i64));
7063 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7064 Addr, TrmpAddr, 10, false, false, 0);
7066 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7067 DAG.getConstant(12, MVT::i64));
7068 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7071 // Jump to the nested function.
7072 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7073 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7074 DAG.getConstant(20, MVT::i64));
7075 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7076 Addr, TrmpAddr, 20, false, false, 0);
7078 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7079 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7080 DAG.getConstant(22, MVT::i64));
7081 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7082 TrmpAddr, 22, false, false, 0);
7085 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7086 return DAG.getMergeValues(Ops, 2, dl);
7088 const Function *Func =
7089 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7090 CallingConv::ID CC = Func->getCallingConv();
7095 llvm_unreachable("Unsupported calling convention");
7096 case CallingConv::C:
7097 case CallingConv::X86_StdCall: {
7098 // Pass 'nest' parameter in ECX.
7099 // Must be kept in sync with X86CallingConv.td
7102 // Check that ECX wasn't needed by an 'inreg' parameter.
7103 const FunctionType *FTy = Func->getFunctionType();
7104 const AttrListPtr &Attrs = Func->getAttributes();
7106 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7107 unsigned InRegCount = 0;
7110 for (FunctionType::param_iterator I = FTy->param_begin(),
7111 E = FTy->param_end(); I != E; ++I, ++Idx)
7112 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7113 // FIXME: should only count parameters that are lowered to integers.
7114 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7116 if (InRegCount > 2) {
7117 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7122 case CallingConv::X86_FastCall:
7123 case CallingConv::X86_ThisCall:
7124 case CallingConv::Fast:
7125 // Pass 'nest' parameter in EAX.
7126 // Must be kept in sync with X86CallingConv.td
7131 SDValue OutChains[4];
7134 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7135 DAG.getConstant(10, MVT::i32));
7136 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7138 // This is storing the opcode for MOV32ri.
7139 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7140 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7141 OutChains[0] = DAG.getStore(Root, dl,
7142 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7143 Trmp, TrmpAddr, 0, false, false, 0);
7145 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7146 DAG.getConstant(1, MVT::i32));
7147 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7150 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7151 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7152 DAG.getConstant(5, MVT::i32));
7153 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7154 TrmpAddr, 5, false, false, 1);
7156 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7157 DAG.getConstant(6, MVT::i32));
7158 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7162 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7163 return DAG.getMergeValues(Ops, 2, dl);
7167 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7168 SelectionDAG &DAG) const {
7170 The rounding mode is in bits 11:10 of FPSR, and has the following
7177 FLT_ROUNDS, on the other hand, expects the following:
7184 To perform the conversion, we do:
7185 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7188 MachineFunction &MF = DAG.getMachineFunction();
7189 const TargetMachine &TM = MF.getTarget();
7190 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7191 unsigned StackAlignment = TFI.getStackAlignment();
7192 EVT VT = Op.getValueType();
7193 DebugLoc dl = Op.getDebugLoc();
7195 // Save FP Control Word to stack slot
7196 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7197 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7199 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7200 DAG.getEntryNode(), StackSlot);
7202 // Load FP Control Word from stack slot
7203 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7206 // Transform as necessary
7208 DAG.getNode(ISD::SRL, dl, MVT::i16,
7209 DAG.getNode(ISD::AND, dl, MVT::i16,
7210 CWD, DAG.getConstant(0x800, MVT::i16)),
7211 DAG.getConstant(11, MVT::i8));
7213 DAG.getNode(ISD::SRL, dl, MVT::i16,
7214 DAG.getNode(ISD::AND, dl, MVT::i16,
7215 CWD, DAG.getConstant(0x400, MVT::i16)),
7216 DAG.getConstant(9, MVT::i8));
7219 DAG.getNode(ISD::AND, dl, MVT::i16,
7220 DAG.getNode(ISD::ADD, dl, MVT::i16,
7221 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7222 DAG.getConstant(1, MVT::i16)),
7223 DAG.getConstant(3, MVT::i16));
7226 return DAG.getNode((VT.getSizeInBits() < 16 ?
7227 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7230 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7231 EVT VT = Op.getValueType();
7233 unsigned NumBits = VT.getSizeInBits();
7234 DebugLoc dl = Op.getDebugLoc();
7236 Op = Op.getOperand(0);
7237 if (VT == MVT::i8) {
7238 // Zero extend to i32 since there is not an i8 bsr.
7240 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7243 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7244 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7245 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7247 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7250 DAG.getConstant(NumBits+NumBits-1, OpVT),
7251 DAG.getConstant(X86::COND_E, MVT::i8),
7254 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7256 // Finally xor with NumBits-1.
7257 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7260 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7264 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7265 EVT VT = Op.getValueType();
7267 unsigned NumBits = VT.getSizeInBits();
7268 DebugLoc dl = Op.getDebugLoc();
7270 Op = Op.getOperand(0);
7271 if (VT == MVT::i8) {
7273 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7276 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7277 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7278 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7280 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7283 DAG.getConstant(NumBits, OpVT),
7284 DAG.getConstant(X86::COND_E, MVT::i8),
7287 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7290 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7294 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7295 EVT VT = Op.getValueType();
7296 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7297 DebugLoc dl = Op.getDebugLoc();
7299 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7300 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7301 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7302 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7303 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7305 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7306 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7307 // return AloBlo + AloBhi + AhiBlo;
7309 SDValue A = Op.getOperand(0);
7310 SDValue B = Op.getOperand(1);
7312 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7313 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7314 A, DAG.getConstant(32, MVT::i32));
7315 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7316 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7317 B, DAG.getConstant(32, MVT::i32));
7318 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7319 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7321 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7322 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7324 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7325 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7327 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7328 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7329 AloBhi, DAG.getConstant(32, MVT::i32));
7330 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7331 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7332 AhiBlo, DAG.getConstant(32, MVT::i32));
7333 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7334 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7339 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7340 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7341 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7342 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7343 // has only one use.
7344 SDNode *N = Op.getNode();
7345 SDValue LHS = N->getOperand(0);
7346 SDValue RHS = N->getOperand(1);
7347 unsigned BaseOp = 0;
7349 DebugLoc dl = Op.getDebugLoc();
7351 switch (Op.getOpcode()) {
7352 default: llvm_unreachable("Unknown ovf instruction!");
7354 // A subtract of one will be selected as a INC. Note that INC doesn't
7355 // set CF, so we can't do this for UADDO.
7356 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7357 if (C->getAPIntValue() == 1) {
7358 BaseOp = X86ISD::INC;
7362 BaseOp = X86ISD::ADD;
7366 BaseOp = X86ISD::ADD;
7370 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7371 // set CF, so we can't do this for USUBO.
7372 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7373 if (C->getAPIntValue() == 1) {
7374 BaseOp = X86ISD::DEC;
7378 BaseOp = X86ISD::SUB;
7382 BaseOp = X86ISD::SUB;
7386 BaseOp = X86ISD::SMUL;
7390 BaseOp = X86ISD::UMUL;
7395 // Also sets EFLAGS.
7396 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7397 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7400 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7401 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7403 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7407 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7408 EVT T = Op.getValueType();
7409 DebugLoc dl = Op.getDebugLoc();
7412 switch(T.getSimpleVT().SimpleTy) {
7414 assert(false && "Invalid value type!");
7415 case MVT::i8: Reg = X86::AL; size = 1; break;
7416 case MVT::i16: Reg = X86::AX; size = 2; break;
7417 case MVT::i32: Reg = X86::EAX; size = 4; break;
7419 assert(Subtarget->is64Bit() && "Node not type legal!");
7420 Reg = X86::RAX; size = 8;
7423 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7424 Op.getOperand(2), SDValue());
7425 SDValue Ops[] = { cpIn.getValue(0),
7428 DAG.getTargetConstant(size, MVT::i8),
7430 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7431 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7433 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7437 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7438 SelectionDAG &DAG) const {
7439 assert(Subtarget->is64Bit() && "Result not type legalized?");
7440 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7441 SDValue TheChain = Op.getOperand(0);
7442 DebugLoc dl = Op.getDebugLoc();
7443 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7444 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7445 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7447 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7448 DAG.getConstant(32, MVT::i8));
7450 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7453 return DAG.getMergeValues(Ops, 2, dl);
7456 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7457 SelectionDAG &DAG) const {
7458 EVT SrcVT = Op.getOperand(0).getValueType();
7459 EVT DstVT = Op.getValueType();
7460 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7461 Subtarget->hasMMX() && !DisableMMX) &&
7462 "Unexpected custom BIT_CONVERT");
7463 assert((DstVT == MVT::i64 ||
7464 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7465 "Unexpected custom BIT_CONVERT");
7466 // i64 <=> MMX conversions are Legal.
7467 if (SrcVT==MVT::i64 && DstVT.isVector())
7469 if (DstVT==MVT::i64 && SrcVT.isVector())
7471 // MMX <=> MMX conversions are Legal.
7472 if (SrcVT.isVector() && DstVT.isVector())
7474 // All other conversions need to be expanded.
7477 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7478 SDNode *Node = Op.getNode();
7479 DebugLoc dl = Node->getDebugLoc();
7480 EVT T = Node->getValueType(0);
7481 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7482 DAG.getConstant(0, T), Node->getOperand(2));
7483 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7484 cast<AtomicSDNode>(Node)->getMemoryVT(),
7485 Node->getOperand(0),
7486 Node->getOperand(1), negOp,
7487 cast<AtomicSDNode>(Node)->getSrcValue(),
7488 cast<AtomicSDNode>(Node)->getAlignment());
7491 /// LowerOperation - Provide custom lowering hooks for some operations.
7493 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7494 switch (Op.getOpcode()) {
7495 default: llvm_unreachable("Should not custom lower this!");
7496 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7497 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7498 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7499 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7500 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7501 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7502 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7503 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7504 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7505 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7506 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7507 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7508 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7509 case ISD::SHL_PARTS:
7510 case ISD::SRA_PARTS:
7511 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7512 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7513 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7514 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7515 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7516 case ISD::FABS: return LowerFABS(Op, DAG);
7517 case ISD::FNEG: return LowerFNEG(Op, DAG);
7518 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7519 case ISD::SETCC: return LowerSETCC(Op, DAG);
7520 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7521 case ISD::SELECT: return LowerSELECT(Op, DAG);
7522 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7523 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7524 case ISD::VASTART: return LowerVASTART(Op, DAG);
7525 case ISD::VAARG: return LowerVAARG(Op, DAG);
7526 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7527 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7528 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7529 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7530 case ISD::FRAME_TO_ARGS_OFFSET:
7531 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7532 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7533 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7534 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7535 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7536 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7537 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7538 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7544 case ISD::UMULO: return LowerXALUO(Op, DAG);
7545 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7546 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7550 void X86TargetLowering::
7551 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7552 SelectionDAG &DAG, unsigned NewOp) const {
7553 EVT T = Node->getValueType(0);
7554 DebugLoc dl = Node->getDebugLoc();
7555 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7557 SDValue Chain = Node->getOperand(0);
7558 SDValue In1 = Node->getOperand(1);
7559 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7560 Node->getOperand(2), DAG.getIntPtrConstant(0));
7561 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7562 Node->getOperand(2), DAG.getIntPtrConstant(1));
7563 SDValue Ops[] = { Chain, In1, In2L, In2H };
7564 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7566 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7567 cast<MemSDNode>(Node)->getMemOperand());
7568 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7569 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7570 Results.push_back(Result.getValue(2));
7573 /// ReplaceNodeResults - Replace a node with an illegal result type
7574 /// with a new node built out of custom code.
7575 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7576 SmallVectorImpl<SDValue>&Results,
7577 SelectionDAG &DAG) const {
7578 DebugLoc dl = N->getDebugLoc();
7579 switch (N->getOpcode()) {
7581 assert(false && "Do not know how to custom type legalize this operation!");
7583 case ISD::FP_TO_SINT: {
7584 std::pair<SDValue,SDValue> Vals =
7585 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7586 SDValue FIST = Vals.first, StackSlot = Vals.second;
7587 if (FIST.getNode() != 0) {
7588 EVT VT = N->getValueType(0);
7589 // Return a load from the stack slot.
7590 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7595 case ISD::READCYCLECOUNTER: {
7596 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7597 SDValue TheChain = N->getOperand(0);
7598 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7599 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7601 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7603 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7604 SDValue Ops[] = { eax, edx };
7605 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7606 Results.push_back(edx.getValue(1));
7609 case ISD::ATOMIC_CMP_SWAP: {
7610 EVT T = N->getValueType(0);
7611 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7612 SDValue cpInL, cpInH;
7613 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7614 DAG.getConstant(0, MVT::i32));
7615 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7616 DAG.getConstant(1, MVT::i32));
7617 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7618 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7620 SDValue swapInL, swapInH;
7621 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7622 DAG.getConstant(0, MVT::i32));
7623 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7624 DAG.getConstant(1, MVT::i32));
7625 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7627 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7628 swapInL.getValue(1));
7629 SDValue Ops[] = { swapInH.getValue(0),
7631 swapInH.getValue(1) };
7632 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7633 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7634 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7635 MVT::i32, Result.getValue(1));
7636 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7637 MVT::i32, cpOutL.getValue(2));
7638 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7639 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7640 Results.push_back(cpOutH.getValue(1));
7643 case ISD::ATOMIC_LOAD_ADD:
7644 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7646 case ISD::ATOMIC_LOAD_AND:
7647 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7649 case ISD::ATOMIC_LOAD_NAND:
7650 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7652 case ISD::ATOMIC_LOAD_OR:
7653 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7655 case ISD::ATOMIC_LOAD_SUB:
7656 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7658 case ISD::ATOMIC_LOAD_XOR:
7659 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7661 case ISD::ATOMIC_SWAP:
7662 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7667 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7669 default: return NULL;
7670 case X86ISD::BSF: return "X86ISD::BSF";
7671 case X86ISD::BSR: return "X86ISD::BSR";
7672 case X86ISD::SHLD: return "X86ISD::SHLD";
7673 case X86ISD::SHRD: return "X86ISD::SHRD";
7674 case X86ISD::FAND: return "X86ISD::FAND";
7675 case X86ISD::FOR: return "X86ISD::FOR";
7676 case X86ISD::FXOR: return "X86ISD::FXOR";
7677 case X86ISD::FSRL: return "X86ISD::FSRL";
7678 case X86ISD::FILD: return "X86ISD::FILD";
7679 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7680 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7681 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7682 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7683 case X86ISD::FLD: return "X86ISD::FLD";
7684 case X86ISD::FST: return "X86ISD::FST";
7685 case X86ISD::CALL: return "X86ISD::CALL";
7686 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7687 case X86ISD::BT: return "X86ISD::BT";
7688 case X86ISD::CMP: return "X86ISD::CMP";
7689 case X86ISD::COMI: return "X86ISD::COMI";
7690 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7691 case X86ISD::SETCC: return "X86ISD::SETCC";
7692 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7693 case X86ISD::CMOV: return "X86ISD::CMOV";
7694 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7695 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7696 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7697 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7698 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7699 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7700 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7701 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7702 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7703 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7704 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7705 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7706 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7707 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7708 case X86ISD::FMAX: return "X86ISD::FMAX";
7709 case X86ISD::FMIN: return "X86ISD::FMIN";
7710 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7711 case X86ISD::FRCP: return "X86ISD::FRCP";
7712 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7713 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7714 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7715 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7716 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7717 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7718 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7719 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7720 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7721 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7722 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7723 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7724 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7725 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7726 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7727 case X86ISD::VSHL: return "X86ISD::VSHL";
7728 case X86ISD::VSRL: return "X86ISD::VSRL";
7729 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7730 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7731 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7732 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7733 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7734 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7735 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7736 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7737 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7738 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7739 case X86ISD::ADD: return "X86ISD::ADD";
7740 case X86ISD::SUB: return "X86ISD::SUB";
7741 case X86ISD::SMUL: return "X86ISD::SMUL";
7742 case X86ISD::UMUL: return "X86ISD::UMUL";
7743 case X86ISD::INC: return "X86ISD::INC";
7744 case X86ISD::DEC: return "X86ISD::DEC";
7745 case X86ISD::OR: return "X86ISD::OR";
7746 case X86ISD::XOR: return "X86ISD::XOR";
7747 case X86ISD::AND: return "X86ISD::AND";
7748 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7749 case X86ISD::PTEST: return "X86ISD::PTEST";
7750 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7751 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7755 // isLegalAddressingMode - Return true if the addressing mode represented
7756 // by AM is legal for this target, for a load/store of the specified type.
7757 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7758 const Type *Ty) const {
7759 // X86 supports extremely general addressing modes.
7760 CodeModel::Model M = getTargetMachine().getCodeModel();
7762 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7763 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7768 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7770 // If a reference to this global requires an extra load, we can't fold it.
7771 if (isGlobalStubReference(GVFlags))
7774 // If BaseGV requires a register for the PIC base, we cannot also have a
7775 // BaseReg specified.
7776 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7779 // If lower 4G is not available, then we must use rip-relative addressing.
7780 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7790 // These scales always work.
7795 // These scales are formed with basereg+scalereg. Only accept if there is
7800 default: // Other stuff never works.
7808 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7809 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7811 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7812 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7813 if (NumBits1 <= NumBits2)
7818 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7819 if (!VT1.isInteger() || !VT2.isInteger())
7821 unsigned NumBits1 = VT1.getSizeInBits();
7822 unsigned NumBits2 = VT2.getSizeInBits();
7823 if (NumBits1 <= NumBits2)
7828 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7829 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7830 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7833 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7834 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7835 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7838 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7839 // i16 instructions are longer (0x66 prefix) and potentially slower.
7840 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7843 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7844 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7845 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7846 /// are assumed to be legal.
7848 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7850 // Very little shuffling can be done for 64-bit vectors right now.
7851 if (VT.getSizeInBits() == 64)
7852 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
7854 // FIXME: pshufb, blends, shifts.
7855 return (VT.getVectorNumElements() == 2 ||
7856 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7857 isMOVLMask(M, VT) ||
7858 isSHUFPMask(M, VT) ||
7859 isPSHUFDMask(M, VT) ||
7860 isPSHUFHWMask(M, VT) ||
7861 isPSHUFLWMask(M, VT) ||
7862 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7863 isUNPCKLMask(M, VT) ||
7864 isUNPCKHMask(M, VT) ||
7865 isUNPCKL_v_undef_Mask(M, VT) ||
7866 isUNPCKH_v_undef_Mask(M, VT));
7870 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7872 unsigned NumElts = VT.getVectorNumElements();
7873 // FIXME: This collection of masks seems suspect.
7876 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7877 return (isMOVLMask(Mask, VT) ||
7878 isCommutedMOVLMask(Mask, VT, true) ||
7879 isSHUFPMask(Mask, VT) ||
7880 isCommutedSHUFPMask(Mask, VT));
7885 //===----------------------------------------------------------------------===//
7886 // X86 Scheduler Hooks
7887 //===----------------------------------------------------------------------===//
7889 // private utility function
7891 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7892 MachineBasicBlock *MBB,
7900 TargetRegisterClass *RC,
7901 bool invSrc) const {
7902 // For the atomic bitwise operator, we generate
7905 // ld t1 = [bitinstr.addr]
7906 // op t2 = t1, [bitinstr.val]
7908 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7910 // fallthrough -->nextMBB
7911 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7912 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7913 MachineFunction::iterator MBBIter = MBB;
7916 /// First build the CFG
7917 MachineFunction *F = MBB->getParent();
7918 MachineBasicBlock *thisMBB = MBB;
7919 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7920 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7921 F->insert(MBBIter, newMBB);
7922 F->insert(MBBIter, nextMBB);
7924 // Move all successors to thisMBB to nextMBB
7925 nextMBB->transferSuccessors(thisMBB);
7927 // Update thisMBB to fall through to newMBB
7928 thisMBB->addSuccessor(newMBB);
7930 // newMBB jumps to itself and fall through to nextMBB
7931 newMBB->addSuccessor(nextMBB);
7932 newMBB->addSuccessor(newMBB);
7934 // Insert instructions into newMBB based on incoming instruction
7935 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7936 "unexpected number of operands");
7937 DebugLoc dl = bInstr->getDebugLoc();
7938 MachineOperand& destOper = bInstr->getOperand(0);
7939 MachineOperand* argOpers[2 + X86AddrNumOperands];
7940 int numArgs = bInstr->getNumOperands() - 1;
7941 for (int i=0; i < numArgs; ++i)
7942 argOpers[i] = &bInstr->getOperand(i+1);
7944 // x86 address has 4 operands: base, index, scale, and displacement
7945 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7946 int valArgIndx = lastAddrIndx + 1;
7948 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7949 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7950 for (int i=0; i <= lastAddrIndx; ++i)
7951 (*MIB).addOperand(*argOpers[i]);
7953 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7955 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7960 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7961 assert((argOpers[valArgIndx]->isReg() ||
7962 argOpers[valArgIndx]->isImm()) &&
7964 if (argOpers[valArgIndx]->isReg())
7965 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7967 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7969 (*MIB).addOperand(*argOpers[valArgIndx]);
7971 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7974 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7975 for (int i=0; i <= lastAddrIndx; ++i)
7976 (*MIB).addOperand(*argOpers[i]);
7978 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7979 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7980 bInstr->memoperands_end());
7982 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7986 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
7988 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7992 // private utility function: 64 bit atomics on 32 bit host.
7994 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7995 MachineBasicBlock *MBB,
8000 bool invSrc) const {
8001 // For the atomic bitwise operator, we generate
8002 // thisMBB (instructions are in pairs, except cmpxchg8b)
8003 // ld t1,t2 = [bitinstr.addr]
8005 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8006 // op t5, t6 <- out1, out2, [bitinstr.val]
8007 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8008 // mov ECX, EBX <- t5, t6
8009 // mov EAX, EDX <- t1, t2
8010 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8011 // mov t3, t4 <- EAX, EDX
8013 // result in out1, out2
8014 // fallthrough -->nextMBB
8016 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8017 const unsigned LoadOpc = X86::MOV32rm;
8018 const unsigned copyOpc = X86::MOV32rr;
8019 const unsigned NotOpc = X86::NOT32r;
8020 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8021 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8022 MachineFunction::iterator MBBIter = MBB;
8025 /// First build the CFG
8026 MachineFunction *F = MBB->getParent();
8027 MachineBasicBlock *thisMBB = MBB;
8028 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8029 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8030 F->insert(MBBIter, newMBB);
8031 F->insert(MBBIter, nextMBB);
8033 // Move all successors to thisMBB to nextMBB
8034 nextMBB->transferSuccessors(thisMBB);
8036 // Update thisMBB to fall through to newMBB
8037 thisMBB->addSuccessor(newMBB);
8039 // newMBB jumps to itself and fall through to nextMBB
8040 newMBB->addSuccessor(nextMBB);
8041 newMBB->addSuccessor(newMBB);
8043 DebugLoc dl = bInstr->getDebugLoc();
8044 // Insert instructions into newMBB based on incoming instruction
8045 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8046 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8047 "unexpected number of operands");
8048 MachineOperand& dest1Oper = bInstr->getOperand(0);
8049 MachineOperand& dest2Oper = bInstr->getOperand(1);
8050 MachineOperand* argOpers[2 + X86AddrNumOperands];
8051 for (int i=0; i < 2 + X86AddrNumOperands; ++i) {
8052 argOpers[i] = &bInstr->getOperand(i+2);
8054 // We use some of the operands multiple times, so conservatively just
8055 // clear any kill flags that might be present.
8056 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8057 argOpers[i]->setIsKill(false);
8060 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8061 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8063 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8064 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8065 for (int i=0; i <= lastAddrIndx; ++i)
8066 (*MIB).addOperand(*argOpers[i]);
8067 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8068 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8069 // add 4 to displacement.
8070 for (int i=0; i <= lastAddrIndx-2; ++i)
8071 (*MIB).addOperand(*argOpers[i]);
8072 MachineOperand newOp3 = *(argOpers[3]);
8074 newOp3.setImm(newOp3.getImm()+4);
8076 newOp3.setOffset(newOp3.getOffset()+4);
8077 (*MIB).addOperand(newOp3);
8078 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8080 // t3/4 are defined later, at the bottom of the loop
8081 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8082 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8083 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8084 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8085 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8086 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8088 // The subsequent operations should be using the destination registers of
8089 //the PHI instructions.
8091 t1 = F->getRegInfo().createVirtualRegister(RC);
8092 t2 = F->getRegInfo().createVirtualRegister(RC);
8093 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8094 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8096 t1 = dest1Oper.getReg();
8097 t2 = dest2Oper.getReg();
8100 int valArgIndx = lastAddrIndx + 1;
8101 assert((argOpers[valArgIndx]->isReg() ||
8102 argOpers[valArgIndx]->isImm()) &&
8104 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8105 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8106 if (argOpers[valArgIndx]->isReg())
8107 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8109 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8110 if (regOpcL != X86::MOV32rr)
8112 (*MIB).addOperand(*argOpers[valArgIndx]);
8113 assert(argOpers[valArgIndx + 1]->isReg() ==
8114 argOpers[valArgIndx]->isReg());
8115 assert(argOpers[valArgIndx + 1]->isImm() ==
8116 argOpers[valArgIndx]->isImm());
8117 if (argOpers[valArgIndx + 1]->isReg())
8118 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8120 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8121 if (regOpcH != X86::MOV32rr)
8123 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8125 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8127 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8130 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8132 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8135 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8136 for (int i=0; i <= lastAddrIndx; ++i)
8137 (*MIB).addOperand(*argOpers[i]);
8139 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8140 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8141 bInstr->memoperands_end());
8143 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8144 MIB.addReg(X86::EAX);
8145 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8146 MIB.addReg(X86::EDX);
8149 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8151 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8155 // private utility function
8157 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8158 MachineBasicBlock *MBB,
8159 unsigned cmovOpc) const {
8160 // For the atomic min/max operator, we generate
8163 // ld t1 = [min/max.addr]
8164 // mov t2 = [min/max.val]
8166 // cmov[cond] t2 = t1
8168 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8170 // fallthrough -->nextMBB
8172 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8173 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8174 MachineFunction::iterator MBBIter = MBB;
8177 /// First build the CFG
8178 MachineFunction *F = MBB->getParent();
8179 MachineBasicBlock *thisMBB = MBB;
8180 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8181 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8182 F->insert(MBBIter, newMBB);
8183 F->insert(MBBIter, nextMBB);
8185 // Move all successors of thisMBB to nextMBB
8186 nextMBB->transferSuccessors(thisMBB);
8188 // Update thisMBB to fall through to newMBB
8189 thisMBB->addSuccessor(newMBB);
8191 // newMBB jumps to newMBB and fall through to nextMBB
8192 newMBB->addSuccessor(nextMBB);
8193 newMBB->addSuccessor(newMBB);
8195 DebugLoc dl = mInstr->getDebugLoc();
8196 // Insert instructions into newMBB based on incoming instruction
8197 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8198 "unexpected number of operands");
8199 MachineOperand& destOper = mInstr->getOperand(0);
8200 MachineOperand* argOpers[2 + X86AddrNumOperands];
8201 int numArgs = mInstr->getNumOperands() - 1;
8202 for (int i=0; i < numArgs; ++i)
8203 argOpers[i] = &mInstr->getOperand(i+1);
8205 // x86 address has 4 operands: base, index, scale, and displacement
8206 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8207 int valArgIndx = lastAddrIndx + 1;
8209 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8210 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8211 for (int i=0; i <= lastAddrIndx; ++i)
8212 (*MIB).addOperand(*argOpers[i]);
8214 // We only support register and immediate values
8215 assert((argOpers[valArgIndx]->isReg() ||
8216 argOpers[valArgIndx]->isImm()) &&
8219 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8220 if (argOpers[valArgIndx]->isReg())
8221 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8223 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8224 (*MIB).addOperand(*argOpers[valArgIndx]);
8226 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8229 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8234 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8235 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8239 // Cmp and exchange if none has modified the memory location
8240 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8241 for (int i=0; i <= lastAddrIndx; ++i)
8242 (*MIB).addOperand(*argOpers[i]);
8244 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8245 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8246 mInstr->memoperands_end());
8248 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8249 MIB.addReg(X86::EAX);
8252 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8254 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8258 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8259 // all of this code can be replaced with that in the .td file.
8261 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8262 unsigned numArgs, bool memArg) const {
8264 MachineFunction *F = BB->getParent();
8265 DebugLoc dl = MI->getDebugLoc();
8266 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8270 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8272 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8274 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8276 for (unsigned i = 0; i < numArgs; ++i) {
8277 MachineOperand &Op = MI->getOperand(i+1);
8279 if (!(Op.isReg() && Op.isImplicit()))
8283 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8286 F->DeleteMachineInstr(MI);
8292 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8294 MachineBasicBlock *MBB) const {
8295 // Emit code to save XMM registers to the stack. The ABI says that the
8296 // number of registers to save is given in %al, so it's theoretically
8297 // possible to do an indirect jump trick to avoid saving all of them,
8298 // however this code takes a simpler approach and just executes all
8299 // of the stores if %al is non-zero. It's less code, and it's probably
8300 // easier on the hardware branch predictor, and stores aren't all that
8301 // expensive anyway.
8303 // Create the new basic blocks. One block contains all the XMM stores,
8304 // and one block is the final destination regardless of whether any
8305 // stores were performed.
8306 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8307 MachineFunction *F = MBB->getParent();
8308 MachineFunction::iterator MBBIter = MBB;
8310 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8311 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8312 F->insert(MBBIter, XMMSaveMBB);
8313 F->insert(MBBIter, EndMBB);
8316 // Move any original successors of MBB to the end block.
8317 EndMBB->transferSuccessors(MBB);
8318 // The original block will now fall through to the XMM save block.
8319 MBB->addSuccessor(XMMSaveMBB);
8320 // The XMMSaveMBB will fall through to the end block.
8321 XMMSaveMBB->addSuccessor(EndMBB);
8323 // Now add the instructions.
8324 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8325 DebugLoc DL = MI->getDebugLoc();
8327 unsigned CountReg = MI->getOperand(0).getReg();
8328 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8329 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8331 if (!Subtarget->isTargetWin64()) {
8332 // If %al is 0, branch around the XMM save block.
8333 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8334 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8335 MBB->addSuccessor(EndMBB);
8338 // In the XMM save block, save all the XMM argument registers.
8339 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8340 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8341 MachineMemOperand *MMO =
8342 F->getMachineMemOperand(
8343 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8344 MachineMemOperand::MOStore, Offset,
8345 /*Size=*/16, /*Align=*/16);
8346 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8347 .addFrameIndex(RegSaveFrameIndex)
8348 .addImm(/*Scale=*/1)
8349 .addReg(/*IndexReg=*/0)
8350 .addImm(/*Disp=*/Offset)
8351 .addReg(/*Segment=*/0)
8352 .addReg(MI->getOperand(i).getReg())
8353 .addMemOperand(MMO);
8356 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8362 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8363 MachineBasicBlock *BB) const {
8364 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8365 DebugLoc DL = MI->getDebugLoc();
8367 // To "insert" a SELECT_CC instruction, we actually have to insert the
8368 // diamond control-flow pattern. The incoming instruction knows the
8369 // destination vreg to set, the condition code register to branch on, the
8370 // true/false values to select between, and a branch opcode to use.
8371 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8372 MachineFunction::iterator It = BB;
8378 // cmpTY ccX, r1, r2
8380 // fallthrough --> copy0MBB
8381 MachineBasicBlock *thisMBB = BB;
8382 MachineFunction *F = BB->getParent();
8383 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8384 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8386 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8387 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8388 F->insert(It, copy0MBB);
8389 F->insert(It, sinkMBB);
8390 // Update machine-CFG edges by first adding all successors of the current
8391 // block to the new block which will contain the Phi node for the select.
8392 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8393 E = BB->succ_end(); I != E; ++I)
8394 sinkMBB->addSuccessor(*I);
8395 // Next, remove all successors of the current block, and add the true
8396 // and fallthrough blocks as its successors.
8397 while (!BB->succ_empty())
8398 BB->removeSuccessor(BB->succ_begin());
8399 // Add the true and fallthrough blocks as its successors.
8400 BB->addSuccessor(copy0MBB);
8401 BB->addSuccessor(sinkMBB);
8404 // %FalseValue = ...
8405 // # fallthrough to sinkMBB
8406 copy0MBB->addSuccessor(sinkMBB);
8409 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8411 BuildMI(sinkMBB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8412 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8413 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8415 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8420 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8421 MachineBasicBlock *BB) const {
8422 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8423 DebugLoc DL = MI->getDebugLoc();
8424 MachineFunction *F = BB->getParent();
8426 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8427 // non-trivial part is impdef of ESP.
8428 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8431 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8432 .addExternalSymbol("_alloca")
8433 .addReg(X86::EAX, RegState::Implicit)
8434 .addReg(X86::ESP, RegState::Implicit)
8435 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8436 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8438 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8443 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8444 MachineBasicBlock *BB) const {
8445 switch (MI->getOpcode()) {
8446 default: assert(false && "Unexpected instr type to insert");
8447 case X86::MINGW_ALLOCA:
8448 return EmitLoweredMingwAlloca(MI, BB);
8450 case X86::CMOV_V1I64:
8451 case X86::CMOV_FR32:
8452 case X86::CMOV_FR64:
8453 case X86::CMOV_V4F32:
8454 case X86::CMOV_V2F64:
8455 case X86::CMOV_V2I64:
8456 case X86::CMOV_GR16:
8457 case X86::CMOV_GR32:
8458 case X86::CMOV_RFP32:
8459 case X86::CMOV_RFP64:
8460 case X86::CMOV_RFP80:
8461 return EmitLoweredSelect(MI, BB);
8463 case X86::FP32_TO_INT16_IN_MEM:
8464 case X86::FP32_TO_INT32_IN_MEM:
8465 case X86::FP32_TO_INT64_IN_MEM:
8466 case X86::FP64_TO_INT16_IN_MEM:
8467 case X86::FP64_TO_INT32_IN_MEM:
8468 case X86::FP64_TO_INT64_IN_MEM:
8469 case X86::FP80_TO_INT16_IN_MEM:
8470 case X86::FP80_TO_INT32_IN_MEM:
8471 case X86::FP80_TO_INT64_IN_MEM: {
8472 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8473 DebugLoc DL = MI->getDebugLoc();
8475 // Change the floating point control register to use "round towards zero"
8476 // mode when truncating to an integer value.
8477 MachineFunction *F = BB->getParent();
8478 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8479 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8481 // Load the old value of the high byte of the control word...
8483 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8484 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8487 // Set the high part to be round to zero...
8488 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8491 // Reload the modified control word now...
8492 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8494 // Restore the memory image of control word to original value
8495 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8498 // Get the X86 opcode to use.
8500 switch (MI->getOpcode()) {
8501 default: llvm_unreachable("illegal opcode!");
8502 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8503 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8504 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8505 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8506 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8507 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8508 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8509 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8510 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8514 MachineOperand &Op = MI->getOperand(0);
8516 AM.BaseType = X86AddressMode::RegBase;
8517 AM.Base.Reg = Op.getReg();
8519 AM.BaseType = X86AddressMode::FrameIndexBase;
8520 AM.Base.FrameIndex = Op.getIndex();
8522 Op = MI->getOperand(1);
8524 AM.Scale = Op.getImm();
8525 Op = MI->getOperand(2);
8527 AM.IndexReg = Op.getImm();
8528 Op = MI->getOperand(3);
8529 if (Op.isGlobal()) {
8530 AM.GV = Op.getGlobal();
8532 AM.Disp = Op.getImm();
8534 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8535 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8537 // Reload the original control word now.
8538 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8540 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8543 // String/text processing lowering.
8544 case X86::PCMPISTRM128REG:
8545 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8546 case X86::PCMPISTRM128MEM:
8547 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8548 case X86::PCMPESTRM128REG:
8549 return EmitPCMP(MI, BB, 5, false /* in mem */);
8550 case X86::PCMPESTRM128MEM:
8551 return EmitPCMP(MI, BB, 5, true /* in mem */);
8554 case X86::ATOMAND32:
8555 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8556 X86::AND32ri, X86::MOV32rm,
8557 X86::LCMPXCHG32, X86::MOV32rr,
8558 X86::NOT32r, X86::EAX,
8559 X86::GR32RegisterClass);
8561 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8562 X86::OR32ri, X86::MOV32rm,
8563 X86::LCMPXCHG32, X86::MOV32rr,
8564 X86::NOT32r, X86::EAX,
8565 X86::GR32RegisterClass);
8566 case X86::ATOMXOR32:
8567 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8568 X86::XOR32ri, X86::MOV32rm,
8569 X86::LCMPXCHG32, X86::MOV32rr,
8570 X86::NOT32r, X86::EAX,
8571 X86::GR32RegisterClass);
8572 case X86::ATOMNAND32:
8573 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8574 X86::AND32ri, X86::MOV32rm,
8575 X86::LCMPXCHG32, X86::MOV32rr,
8576 X86::NOT32r, X86::EAX,
8577 X86::GR32RegisterClass, true);
8578 case X86::ATOMMIN32:
8579 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8580 case X86::ATOMMAX32:
8581 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8582 case X86::ATOMUMIN32:
8583 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8584 case X86::ATOMUMAX32:
8585 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8587 case X86::ATOMAND16:
8588 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8589 X86::AND16ri, X86::MOV16rm,
8590 X86::LCMPXCHG16, X86::MOV16rr,
8591 X86::NOT16r, X86::AX,
8592 X86::GR16RegisterClass);
8594 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8595 X86::OR16ri, X86::MOV16rm,
8596 X86::LCMPXCHG16, X86::MOV16rr,
8597 X86::NOT16r, X86::AX,
8598 X86::GR16RegisterClass);
8599 case X86::ATOMXOR16:
8600 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8601 X86::XOR16ri, X86::MOV16rm,
8602 X86::LCMPXCHG16, X86::MOV16rr,
8603 X86::NOT16r, X86::AX,
8604 X86::GR16RegisterClass);
8605 case X86::ATOMNAND16:
8606 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8607 X86::AND16ri, X86::MOV16rm,
8608 X86::LCMPXCHG16, X86::MOV16rr,
8609 X86::NOT16r, X86::AX,
8610 X86::GR16RegisterClass, true);
8611 case X86::ATOMMIN16:
8612 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8613 case X86::ATOMMAX16:
8614 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8615 case X86::ATOMUMIN16:
8616 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8617 case X86::ATOMUMAX16:
8618 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8621 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8622 X86::AND8ri, X86::MOV8rm,
8623 X86::LCMPXCHG8, X86::MOV8rr,
8624 X86::NOT8r, X86::AL,
8625 X86::GR8RegisterClass);
8627 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8628 X86::OR8ri, X86::MOV8rm,
8629 X86::LCMPXCHG8, X86::MOV8rr,
8630 X86::NOT8r, X86::AL,
8631 X86::GR8RegisterClass);
8633 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8634 X86::XOR8ri, X86::MOV8rm,
8635 X86::LCMPXCHG8, X86::MOV8rr,
8636 X86::NOT8r, X86::AL,
8637 X86::GR8RegisterClass);
8638 case X86::ATOMNAND8:
8639 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8640 X86::AND8ri, X86::MOV8rm,
8641 X86::LCMPXCHG8, X86::MOV8rr,
8642 X86::NOT8r, X86::AL,
8643 X86::GR8RegisterClass, true);
8644 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8645 // This group is for 64-bit host.
8646 case X86::ATOMAND64:
8647 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8648 X86::AND64ri32, X86::MOV64rm,
8649 X86::LCMPXCHG64, X86::MOV64rr,
8650 X86::NOT64r, X86::RAX,
8651 X86::GR64RegisterClass);
8653 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8654 X86::OR64ri32, X86::MOV64rm,
8655 X86::LCMPXCHG64, X86::MOV64rr,
8656 X86::NOT64r, X86::RAX,
8657 X86::GR64RegisterClass);
8658 case X86::ATOMXOR64:
8659 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8660 X86::XOR64ri32, X86::MOV64rm,
8661 X86::LCMPXCHG64, X86::MOV64rr,
8662 X86::NOT64r, X86::RAX,
8663 X86::GR64RegisterClass);
8664 case X86::ATOMNAND64:
8665 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8666 X86::AND64ri32, X86::MOV64rm,
8667 X86::LCMPXCHG64, X86::MOV64rr,
8668 X86::NOT64r, X86::RAX,
8669 X86::GR64RegisterClass, true);
8670 case X86::ATOMMIN64:
8671 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8672 case X86::ATOMMAX64:
8673 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8674 case X86::ATOMUMIN64:
8675 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8676 case X86::ATOMUMAX64:
8677 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8679 // This group does 64-bit operations on a 32-bit host.
8680 case X86::ATOMAND6432:
8681 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8682 X86::AND32rr, X86::AND32rr,
8683 X86::AND32ri, X86::AND32ri,
8685 case X86::ATOMOR6432:
8686 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8687 X86::OR32rr, X86::OR32rr,
8688 X86::OR32ri, X86::OR32ri,
8690 case X86::ATOMXOR6432:
8691 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8692 X86::XOR32rr, X86::XOR32rr,
8693 X86::XOR32ri, X86::XOR32ri,
8695 case X86::ATOMNAND6432:
8696 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8697 X86::AND32rr, X86::AND32rr,
8698 X86::AND32ri, X86::AND32ri,
8700 case X86::ATOMADD6432:
8701 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8702 X86::ADD32rr, X86::ADC32rr,
8703 X86::ADD32ri, X86::ADC32ri,
8705 case X86::ATOMSUB6432:
8706 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8707 X86::SUB32rr, X86::SBB32rr,
8708 X86::SUB32ri, X86::SBB32ri,
8710 case X86::ATOMSWAP6432:
8711 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8712 X86::MOV32rr, X86::MOV32rr,
8713 X86::MOV32ri, X86::MOV32ri,
8715 case X86::VASTART_SAVE_XMM_REGS:
8716 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8720 //===----------------------------------------------------------------------===//
8721 // X86 Optimization Hooks
8722 //===----------------------------------------------------------------------===//
8724 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8728 const SelectionDAG &DAG,
8729 unsigned Depth) const {
8730 unsigned Opc = Op.getOpcode();
8731 assert((Opc >= ISD::BUILTIN_OP_END ||
8732 Opc == ISD::INTRINSIC_WO_CHAIN ||
8733 Opc == ISD::INTRINSIC_W_CHAIN ||
8734 Opc == ISD::INTRINSIC_VOID) &&
8735 "Should use MaskedValueIsZero if you don't know whether Op"
8736 " is a target node!");
8738 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8750 // These nodes' second result is a boolean.
8751 if (Op.getResNo() == 0)
8755 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8756 Mask.getBitWidth() - 1);
8761 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8762 /// node is a GlobalAddress + offset.
8763 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8764 const GlobalValue* &GA,
8765 int64_t &Offset) const {
8766 if (N->getOpcode() == X86ISD::Wrapper) {
8767 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8768 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8769 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8773 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8776 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8777 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8778 /// if the load addresses are consecutive, non-overlapping, and in the right
8780 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8781 const TargetLowering &TLI) {
8782 DebugLoc dl = N->getDebugLoc();
8783 EVT VT = N->getValueType(0);
8784 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8786 if (VT.getSizeInBits() != 128)
8789 SmallVector<SDValue, 16> Elts;
8790 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8791 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8793 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8796 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8797 /// and convert it from being a bunch of shuffles and extracts to a simple
8798 /// store and scalar loads to extract the elements.
8799 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8800 const TargetLowering &TLI) {
8801 SDValue InputVector = N->getOperand(0);
8803 // Only operate on vectors of 4 elements, where the alternative shuffling
8804 // gets to be more expensive.
8805 if (InputVector.getValueType() != MVT::v4i32)
8808 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8809 // single use which is a sign-extend or zero-extend, and all elements are
8811 SmallVector<SDNode *, 4> Uses;
8812 unsigned ExtractedElements = 0;
8813 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8814 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8815 if (UI.getUse().getResNo() != InputVector.getResNo())
8818 SDNode *Extract = *UI;
8819 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8822 if (Extract->getValueType(0) != MVT::i32)
8824 if (!Extract->hasOneUse())
8826 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
8827 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
8829 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
8832 // Record which element was extracted.
8833 ExtractedElements |=
8834 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
8836 Uses.push_back(Extract);
8839 // If not all the elements were used, this may not be worthwhile.
8840 if (ExtractedElements != 15)
8843 // Ok, we've now decided to do the transformation.
8844 DebugLoc dl = InputVector.getDebugLoc();
8846 // Store the value to a temporary stack slot.
8847 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
8848 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
8851 // Replace each use (extract) with a load of the appropriate element.
8852 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
8853 UE = Uses.end(); UI != UE; ++UI) {
8854 SDNode *Extract = *UI;
8856 // Compute the element's address.
8857 SDValue Idx = Extract->getOperand(1);
8859 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
8860 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
8861 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
8863 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
8866 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
8867 NULL, 0, false, false, 0);
8869 // Replace the exact with the load.
8870 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
8873 // The replacement was made in place; don't return anything.
8877 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8878 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8879 const X86Subtarget *Subtarget) {
8880 DebugLoc DL = N->getDebugLoc();
8881 SDValue Cond = N->getOperand(0);
8882 // Get the LHS/RHS of the select.
8883 SDValue LHS = N->getOperand(1);
8884 SDValue RHS = N->getOperand(2);
8886 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8887 // instructions match the semantics of the common C idiom x<y?x:y but not
8888 // x<=y?x:y, because of how they handle negative zero (which can be
8889 // ignored in unsafe-math mode).
8890 if (Subtarget->hasSSE2() &&
8891 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8892 Cond.getOpcode() == ISD::SETCC) {
8893 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8895 unsigned Opcode = 0;
8896 // Check for x CC y ? x : y.
8897 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
8898 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
8902 // Converting this to a min would handle NaNs incorrectly, and swapping
8903 // the operands would cause it to handle comparisons between positive
8904 // and negative zero incorrectly.
8905 if (!FiniteOnlyFPMath() &&
8906 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
8907 if (!UnsafeFPMath &&
8908 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8910 std::swap(LHS, RHS);
8912 Opcode = X86ISD::FMIN;
8915 // Converting this to a min would handle comparisons between positive
8916 // and negative zero incorrectly.
8917 if (!UnsafeFPMath &&
8918 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
8920 Opcode = X86ISD::FMIN;
8923 // Converting this to a min would handle both negative zeros and NaNs
8924 // incorrectly, but we can swap the operands to fix both.
8925 std::swap(LHS, RHS);
8929 Opcode = X86ISD::FMIN;
8933 // Converting this to a max would handle comparisons between positive
8934 // and negative zero incorrectly.
8935 if (!UnsafeFPMath &&
8936 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
8938 Opcode = X86ISD::FMAX;
8941 // Converting this to a max would handle NaNs incorrectly, and swapping
8942 // the operands would cause it to handle comparisons between positive
8943 // and negative zero incorrectly.
8944 if (!FiniteOnlyFPMath() &&
8945 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
8946 if (!UnsafeFPMath &&
8947 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8949 std::swap(LHS, RHS);
8951 Opcode = X86ISD::FMAX;
8954 // Converting this to a max would handle both negative zeros and NaNs
8955 // incorrectly, but we can swap the operands to fix both.
8956 std::swap(LHS, RHS);
8960 Opcode = X86ISD::FMAX;
8963 // Check for x CC y ? y : x -- a min/max with reversed arms.
8964 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
8965 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
8969 // Converting this to a min would handle comparisons between positive
8970 // and negative zero incorrectly, and swapping the operands would
8971 // cause it to handle NaNs incorrectly.
8972 if (!UnsafeFPMath &&
8973 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
8974 if (!FiniteOnlyFPMath() &&
8975 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8977 std::swap(LHS, RHS);
8979 Opcode = X86ISD::FMIN;
8982 // Converting this to a min would handle NaNs incorrectly.
8983 if (!UnsafeFPMath &&
8984 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8986 Opcode = X86ISD::FMIN;
8989 // Converting this to a min would handle both negative zeros and NaNs
8990 // incorrectly, but we can swap the operands to fix both.
8991 std::swap(LHS, RHS);
8995 Opcode = X86ISD::FMIN;
8999 // Converting this to a max would handle NaNs incorrectly.
9000 if (!FiniteOnlyFPMath() &&
9001 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9003 Opcode = X86ISD::FMAX;
9006 // Converting this to a max would handle comparisons between positive
9007 // and negative zero incorrectly, and swapping the operands would
9008 // cause it to handle NaNs incorrectly.
9009 if (!UnsafeFPMath &&
9010 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9011 if (!FiniteOnlyFPMath() &&
9012 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9014 std::swap(LHS, RHS);
9016 Opcode = X86ISD::FMAX;
9019 // Converting this to a max would handle both negative zeros and NaNs
9020 // incorrectly, but we can swap the operands to fix both.
9021 std::swap(LHS, RHS);
9025 Opcode = X86ISD::FMAX;
9031 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9034 // If this is a select between two integer constants, try to do some
9036 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9037 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9038 // Don't do this for crazy integer types.
9039 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9040 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9041 // so that TrueC (the true value) is larger than FalseC.
9042 bool NeedsCondInvert = false;
9044 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9045 // Efficiently invertible.
9046 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9047 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9048 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9049 NeedsCondInvert = true;
9050 std::swap(TrueC, FalseC);
9053 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9054 if (FalseC->getAPIntValue() == 0 &&
9055 TrueC->getAPIntValue().isPowerOf2()) {
9056 if (NeedsCondInvert) // Invert the condition if needed.
9057 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9058 DAG.getConstant(1, Cond.getValueType()));
9060 // Zero extend the condition if needed.
9061 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9063 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9064 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9065 DAG.getConstant(ShAmt, MVT::i8));
9068 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9069 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9070 if (NeedsCondInvert) // Invert the condition if needed.
9071 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9072 DAG.getConstant(1, Cond.getValueType()));
9074 // Zero extend the condition if needed.
9075 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9076 FalseC->getValueType(0), Cond);
9077 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9078 SDValue(FalseC, 0));
9081 // Optimize cases that will turn into an LEA instruction. This requires
9082 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9083 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9084 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9085 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9087 bool isFastMultiplier = false;
9089 switch ((unsigned char)Diff) {
9091 case 1: // result = add base, cond
9092 case 2: // result = lea base( , cond*2)
9093 case 3: // result = lea base(cond, cond*2)
9094 case 4: // result = lea base( , cond*4)
9095 case 5: // result = lea base(cond, cond*4)
9096 case 8: // result = lea base( , cond*8)
9097 case 9: // result = lea base(cond, cond*8)
9098 isFastMultiplier = true;
9103 if (isFastMultiplier) {
9104 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9105 if (NeedsCondInvert) // Invert the condition if needed.
9106 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9107 DAG.getConstant(1, Cond.getValueType()));
9109 // Zero extend the condition if needed.
9110 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9112 // Scale the condition by the difference.
9114 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9115 DAG.getConstant(Diff, Cond.getValueType()));
9117 // Add the base if non-zero.
9118 if (FalseC->getAPIntValue() != 0)
9119 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9120 SDValue(FalseC, 0));
9130 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9131 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9132 TargetLowering::DAGCombinerInfo &DCI) {
9133 DebugLoc DL = N->getDebugLoc();
9135 // If the flag operand isn't dead, don't touch this CMOV.
9136 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9139 // If this is a select between two integer constants, try to do some
9140 // optimizations. Note that the operands are ordered the opposite of SELECT
9142 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9143 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9144 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9145 // larger than FalseC (the false value).
9146 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9148 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9149 CC = X86::GetOppositeBranchCondition(CC);
9150 std::swap(TrueC, FalseC);
9153 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9154 // This is efficient for any integer data type (including i8/i16) and
9156 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9157 SDValue Cond = N->getOperand(3);
9158 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9159 DAG.getConstant(CC, MVT::i8), Cond);
9161 // Zero extend the condition if needed.
9162 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9164 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9165 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9166 DAG.getConstant(ShAmt, MVT::i8));
9167 if (N->getNumValues() == 2) // Dead flag value?
9168 return DCI.CombineTo(N, Cond, SDValue());
9172 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9173 // for any integer data type, including i8/i16.
9174 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9175 SDValue Cond = N->getOperand(3);
9176 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9177 DAG.getConstant(CC, MVT::i8), Cond);
9179 // Zero extend the condition if needed.
9180 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9181 FalseC->getValueType(0), Cond);
9182 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9183 SDValue(FalseC, 0));
9185 if (N->getNumValues() == 2) // Dead flag value?
9186 return DCI.CombineTo(N, Cond, SDValue());
9190 // Optimize cases that will turn into an LEA instruction. This requires
9191 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9192 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9193 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9194 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9196 bool isFastMultiplier = false;
9198 switch ((unsigned char)Diff) {
9200 case 1: // result = add base, cond
9201 case 2: // result = lea base( , cond*2)
9202 case 3: // result = lea base(cond, cond*2)
9203 case 4: // result = lea base( , cond*4)
9204 case 5: // result = lea base(cond, cond*4)
9205 case 8: // result = lea base( , cond*8)
9206 case 9: // result = lea base(cond, cond*8)
9207 isFastMultiplier = true;
9212 if (isFastMultiplier) {
9213 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9214 SDValue Cond = N->getOperand(3);
9215 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9216 DAG.getConstant(CC, MVT::i8), Cond);
9217 // Zero extend the condition if needed.
9218 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9220 // Scale the condition by the difference.
9222 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9223 DAG.getConstant(Diff, Cond.getValueType()));
9225 // Add the base if non-zero.
9226 if (FalseC->getAPIntValue() != 0)
9227 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9228 SDValue(FalseC, 0));
9229 if (N->getNumValues() == 2) // Dead flag value?
9230 return DCI.CombineTo(N, Cond, SDValue());
9240 /// PerformMulCombine - Optimize a single multiply with constant into two
9241 /// in order to implement it with two cheaper instructions, e.g.
9242 /// LEA + SHL, LEA + LEA.
9243 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9244 TargetLowering::DAGCombinerInfo &DCI) {
9245 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9248 EVT VT = N->getValueType(0);
9252 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9255 uint64_t MulAmt = C->getZExtValue();
9256 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9259 uint64_t MulAmt1 = 0;
9260 uint64_t MulAmt2 = 0;
9261 if ((MulAmt % 9) == 0) {
9263 MulAmt2 = MulAmt / 9;
9264 } else if ((MulAmt % 5) == 0) {
9266 MulAmt2 = MulAmt / 5;
9267 } else if ((MulAmt % 3) == 0) {
9269 MulAmt2 = MulAmt / 3;
9272 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9273 DebugLoc DL = N->getDebugLoc();
9275 if (isPowerOf2_64(MulAmt2) &&
9276 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9277 // If second multiplifer is pow2, issue it first. We want the multiply by
9278 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9280 std::swap(MulAmt1, MulAmt2);
9283 if (isPowerOf2_64(MulAmt1))
9284 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9285 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9287 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9288 DAG.getConstant(MulAmt1, VT));
9290 if (isPowerOf2_64(MulAmt2))
9291 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9292 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9294 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9295 DAG.getConstant(MulAmt2, VT));
9297 // Do not add new nodes to DAG combiner worklist.
9298 DCI.CombineTo(N, NewMul, false);
9303 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9304 SDValue N0 = N->getOperand(0);
9305 SDValue N1 = N->getOperand(1);
9306 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9307 EVT VT = N0.getValueType();
9309 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9310 // since the result of setcc_c is all zero's or all ones.
9311 if (N1C && N0.getOpcode() == ISD::AND &&
9312 N0.getOperand(1).getOpcode() == ISD::Constant) {
9313 SDValue N00 = N0.getOperand(0);
9314 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9315 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9316 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9317 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9318 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9319 APInt ShAmt = N1C->getAPIntValue();
9320 Mask = Mask.shl(ShAmt);
9322 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9323 N00, DAG.getConstant(Mask, VT));
9330 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9332 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9333 const X86Subtarget *Subtarget) {
9334 EVT VT = N->getValueType(0);
9335 if (!VT.isVector() && VT.isInteger() &&
9336 N->getOpcode() == ISD::SHL)
9337 return PerformSHLCombine(N, DAG);
9339 // On X86 with SSE2 support, we can transform this to a vector shift if
9340 // all elements are shifted by the same amount. We can't do this in legalize
9341 // because the a constant vector is typically transformed to a constant pool
9342 // so we have no knowledge of the shift amount.
9343 if (!Subtarget->hasSSE2())
9346 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9349 SDValue ShAmtOp = N->getOperand(1);
9350 EVT EltVT = VT.getVectorElementType();
9351 DebugLoc DL = N->getDebugLoc();
9352 SDValue BaseShAmt = SDValue();
9353 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9354 unsigned NumElts = VT.getVectorNumElements();
9356 for (; i != NumElts; ++i) {
9357 SDValue Arg = ShAmtOp.getOperand(i);
9358 if (Arg.getOpcode() == ISD::UNDEF) continue;
9362 for (; i != NumElts; ++i) {
9363 SDValue Arg = ShAmtOp.getOperand(i);
9364 if (Arg.getOpcode() == ISD::UNDEF) continue;
9365 if (Arg != BaseShAmt) {
9369 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9370 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9371 SDValue InVec = ShAmtOp.getOperand(0);
9372 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9373 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9375 for (; i != NumElts; ++i) {
9376 SDValue Arg = InVec.getOperand(i);
9377 if (Arg.getOpcode() == ISD::UNDEF) continue;
9381 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9382 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9383 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9384 if (C->getZExtValue() == SplatIdx)
9385 BaseShAmt = InVec.getOperand(1);
9388 if (BaseShAmt.getNode() == 0)
9389 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9390 DAG.getIntPtrConstant(0));
9394 // The shift amount is an i32.
9395 if (EltVT.bitsGT(MVT::i32))
9396 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9397 else if (EltVT.bitsLT(MVT::i32))
9398 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9400 // The shift amount is identical so we can do a vector shift.
9401 SDValue ValOp = N->getOperand(0);
9402 switch (N->getOpcode()) {
9404 llvm_unreachable("Unknown shift opcode!");
9407 if (VT == MVT::v2i64)
9408 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9409 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9411 if (VT == MVT::v4i32)
9412 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9413 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9415 if (VT == MVT::v8i16)
9416 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9417 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9421 if (VT == MVT::v4i32)
9422 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9423 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9425 if (VT == MVT::v8i16)
9426 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9427 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9431 if (VT == MVT::v2i64)
9432 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9433 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9435 if (VT == MVT::v4i32)
9436 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9437 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9439 if (VT == MVT::v8i16)
9440 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9441 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9448 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9449 TargetLowering::DAGCombinerInfo &DCI,
9450 const X86Subtarget *Subtarget) {
9451 if (DCI.isBeforeLegalizeOps())
9454 EVT VT = N->getValueType(0);
9455 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9458 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9459 SDValue N0 = N->getOperand(0);
9460 SDValue N1 = N->getOperand(1);
9461 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9463 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9465 if (!N0.hasOneUse() || !N1.hasOneUse())
9468 SDValue ShAmt0 = N0.getOperand(1);
9469 if (ShAmt0.getValueType() != MVT::i8)
9471 SDValue ShAmt1 = N1.getOperand(1);
9472 if (ShAmt1.getValueType() != MVT::i8)
9474 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9475 ShAmt0 = ShAmt0.getOperand(0);
9476 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9477 ShAmt1 = ShAmt1.getOperand(0);
9479 DebugLoc DL = N->getDebugLoc();
9480 unsigned Opc = X86ISD::SHLD;
9481 SDValue Op0 = N0.getOperand(0);
9482 SDValue Op1 = N1.getOperand(0);
9483 if (ShAmt0.getOpcode() == ISD::SUB) {
9485 std::swap(Op0, Op1);
9486 std::swap(ShAmt0, ShAmt1);
9489 unsigned Bits = VT.getSizeInBits();
9490 if (ShAmt1.getOpcode() == ISD::SUB) {
9491 SDValue Sum = ShAmt1.getOperand(0);
9492 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9493 if (SumC->getSExtValue() == Bits &&
9494 ShAmt1.getOperand(1) == ShAmt0)
9495 return DAG.getNode(Opc, DL, VT,
9497 DAG.getNode(ISD::TRUNCATE, DL,
9500 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9501 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9503 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9504 return DAG.getNode(Opc, DL, VT,
9505 N0.getOperand(0), N1.getOperand(0),
9506 DAG.getNode(ISD::TRUNCATE, DL,
9513 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9514 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9515 const X86Subtarget *Subtarget) {
9516 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9517 // the FP state in cases where an emms may be missing.
9518 // A preferable solution to the general problem is to figure out the right
9519 // places to insert EMMS. This qualifies as a quick hack.
9521 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9522 StoreSDNode *St = cast<StoreSDNode>(N);
9523 EVT VT = St->getValue().getValueType();
9524 if (VT.getSizeInBits() != 64)
9527 const Function *F = DAG.getMachineFunction().getFunction();
9528 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9529 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9530 && Subtarget->hasSSE2();
9531 if ((VT.isVector() ||
9532 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9533 isa<LoadSDNode>(St->getValue()) &&
9534 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9535 St->getChain().hasOneUse() && !St->isVolatile()) {
9536 SDNode* LdVal = St->getValue().getNode();
9538 int TokenFactorIndex = -1;
9539 SmallVector<SDValue, 8> Ops;
9540 SDNode* ChainVal = St->getChain().getNode();
9541 // Must be a store of a load. We currently handle two cases: the load
9542 // is a direct child, and it's under an intervening TokenFactor. It is
9543 // possible to dig deeper under nested TokenFactors.
9544 if (ChainVal == LdVal)
9545 Ld = cast<LoadSDNode>(St->getChain());
9546 else if (St->getValue().hasOneUse() &&
9547 ChainVal->getOpcode() == ISD::TokenFactor) {
9548 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9549 if (ChainVal->getOperand(i).getNode() == LdVal) {
9550 TokenFactorIndex = i;
9551 Ld = cast<LoadSDNode>(St->getValue());
9553 Ops.push_back(ChainVal->getOperand(i));
9557 if (!Ld || !ISD::isNormalLoad(Ld))
9560 // If this is not the MMX case, i.e. we are just turning i64 load/store
9561 // into f64 load/store, avoid the transformation if there are multiple
9562 // uses of the loaded value.
9563 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9566 DebugLoc LdDL = Ld->getDebugLoc();
9567 DebugLoc StDL = N->getDebugLoc();
9568 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9569 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9571 if (Subtarget->is64Bit() || F64IsLegal) {
9572 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9573 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9574 Ld->getBasePtr(), Ld->getSrcValue(),
9575 Ld->getSrcValueOffset(), Ld->isVolatile(),
9576 Ld->isNonTemporal(), Ld->getAlignment());
9577 SDValue NewChain = NewLd.getValue(1);
9578 if (TokenFactorIndex != -1) {
9579 Ops.push_back(NewChain);
9580 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9583 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9584 St->getSrcValue(), St->getSrcValueOffset(),
9585 St->isVolatile(), St->isNonTemporal(),
9586 St->getAlignment());
9589 // Otherwise, lower to two pairs of 32-bit loads / stores.
9590 SDValue LoAddr = Ld->getBasePtr();
9591 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9592 DAG.getConstant(4, MVT::i32));
9594 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9595 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9596 Ld->isVolatile(), Ld->isNonTemporal(),
9597 Ld->getAlignment());
9598 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9599 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9600 Ld->isVolatile(), Ld->isNonTemporal(),
9601 MinAlign(Ld->getAlignment(), 4));
9603 SDValue NewChain = LoLd.getValue(1);
9604 if (TokenFactorIndex != -1) {
9605 Ops.push_back(LoLd);
9606 Ops.push_back(HiLd);
9607 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9611 LoAddr = St->getBasePtr();
9612 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9613 DAG.getConstant(4, MVT::i32));
9615 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9616 St->getSrcValue(), St->getSrcValueOffset(),
9617 St->isVolatile(), St->isNonTemporal(),
9618 St->getAlignment());
9619 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9621 St->getSrcValueOffset() + 4,
9623 St->isNonTemporal(),
9624 MinAlign(St->getAlignment(), 4));
9625 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9630 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9631 /// X86ISD::FXOR nodes.
9632 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9633 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9634 // F[X]OR(0.0, x) -> x
9635 // F[X]OR(x, 0.0) -> x
9636 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9637 if (C->getValueAPF().isPosZero())
9638 return N->getOperand(1);
9639 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9640 if (C->getValueAPF().isPosZero())
9641 return N->getOperand(0);
9645 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9646 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9647 // FAND(0.0, x) -> 0.0
9648 // FAND(x, 0.0) -> 0.0
9649 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9650 if (C->getValueAPF().isPosZero())
9651 return N->getOperand(0);
9652 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9653 if (C->getValueAPF().isPosZero())
9654 return N->getOperand(1);
9658 static SDValue PerformBTCombine(SDNode *N,
9660 TargetLowering::DAGCombinerInfo &DCI) {
9661 // BT ignores high bits in the bit index operand.
9662 SDValue Op1 = N->getOperand(1);
9663 if (Op1.hasOneUse()) {
9664 unsigned BitWidth = Op1.getValueSizeInBits();
9665 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9666 APInt KnownZero, KnownOne;
9667 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9668 !DCI.isBeforeLegalizeOps());
9669 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9670 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9671 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9672 DCI.CommitTargetLoweringOpt(TLO);
9677 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9678 SDValue Op = N->getOperand(0);
9679 if (Op.getOpcode() == ISD::BIT_CONVERT)
9680 Op = Op.getOperand(0);
9681 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9682 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9683 VT.getVectorElementType().getSizeInBits() ==
9684 OpVT.getVectorElementType().getSizeInBits()) {
9685 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9690 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9691 // Locked instructions, in turn, have implicit fence semantics (all memory
9692 // operations are flushed before issuing the locked instruction, and the
9693 // are not buffered), so we can fold away the common pattern of
9694 // fence-atomic-fence.
9695 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9696 SDValue atomic = N->getOperand(0);
9697 switch (atomic.getOpcode()) {
9698 case ISD::ATOMIC_CMP_SWAP:
9699 case ISD::ATOMIC_SWAP:
9700 case ISD::ATOMIC_LOAD_ADD:
9701 case ISD::ATOMIC_LOAD_SUB:
9702 case ISD::ATOMIC_LOAD_AND:
9703 case ISD::ATOMIC_LOAD_OR:
9704 case ISD::ATOMIC_LOAD_XOR:
9705 case ISD::ATOMIC_LOAD_NAND:
9706 case ISD::ATOMIC_LOAD_MIN:
9707 case ISD::ATOMIC_LOAD_MAX:
9708 case ISD::ATOMIC_LOAD_UMIN:
9709 case ISD::ATOMIC_LOAD_UMAX:
9715 SDValue fence = atomic.getOperand(0);
9716 if (fence.getOpcode() != ISD::MEMBARRIER)
9719 switch (atomic.getOpcode()) {
9720 case ISD::ATOMIC_CMP_SWAP:
9721 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9722 atomic.getOperand(1), atomic.getOperand(2),
9723 atomic.getOperand(3));
9724 case ISD::ATOMIC_SWAP:
9725 case ISD::ATOMIC_LOAD_ADD:
9726 case ISD::ATOMIC_LOAD_SUB:
9727 case ISD::ATOMIC_LOAD_AND:
9728 case ISD::ATOMIC_LOAD_OR:
9729 case ISD::ATOMIC_LOAD_XOR:
9730 case ISD::ATOMIC_LOAD_NAND:
9731 case ISD::ATOMIC_LOAD_MIN:
9732 case ISD::ATOMIC_LOAD_MAX:
9733 case ISD::ATOMIC_LOAD_UMIN:
9734 case ISD::ATOMIC_LOAD_UMAX:
9735 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9736 atomic.getOperand(1), atomic.getOperand(2));
9742 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9743 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9744 // (and (i32 x86isd::setcc_carry), 1)
9745 // This eliminates the zext. This transformation is necessary because
9746 // ISD::SETCC is always legalized to i8.
9747 DebugLoc dl = N->getDebugLoc();
9748 SDValue N0 = N->getOperand(0);
9749 EVT VT = N->getValueType(0);
9750 if (N0.getOpcode() == ISD::AND &&
9752 N0.getOperand(0).hasOneUse()) {
9753 SDValue N00 = N0.getOperand(0);
9754 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9756 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9757 if (!C || C->getZExtValue() != 1)
9759 return DAG.getNode(ISD::AND, dl, VT,
9760 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9761 N00.getOperand(0), N00.getOperand(1)),
9762 DAG.getConstant(1, VT));
9768 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9769 DAGCombinerInfo &DCI) const {
9770 SelectionDAG &DAG = DCI.DAG;
9771 switch (N->getOpcode()) {
9773 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9774 case ISD::EXTRACT_VECTOR_ELT:
9775 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9776 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9777 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9778 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9781 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9782 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
9783 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9785 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9786 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9787 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9788 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9789 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9790 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9796 /// isTypeDesirableForOp - Return true if the target has native support for
9797 /// the specified value type and it is 'desirable' to use the type for the
9798 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9799 /// instruction encodings are longer and some i16 instructions are slow.
9800 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9801 if (!isTypeLegal(VT))
9810 case ISD::SIGN_EXTEND:
9811 case ISD::ZERO_EXTEND:
9812 case ISD::ANY_EXTEND:
9825 static bool MayFoldLoad(SDValue Op) {
9826 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
9829 static bool MayFoldIntoStore(SDValue Op) {
9830 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
9833 /// IsDesirableToPromoteOp - This method query the target whether it is
9834 /// beneficial for dag combiner to promote the specified node. If true, it
9835 /// should return the desired promotion type by reference.
9836 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
9837 EVT VT = Op.getValueType();
9841 bool Promote = false;
9842 bool Commute = false;
9843 switch (Op.getOpcode()) {
9846 LoadSDNode *LD = cast<LoadSDNode>(Op);
9847 // If the non-extending load has a single use and it's not live out, then it
9849 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
9851 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9852 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
9853 // The only case where we'd want to promote LOAD (rather then it being
9854 // promoted as an operand is when it's only use is liveout.
9855 if (UI->getOpcode() != ISD::CopyToReg)
9862 case ISD::SIGN_EXTEND:
9863 case ISD::ZERO_EXTEND:
9864 case ISD::ANY_EXTEND:
9869 SDValue N0 = Op.getOperand(0);
9870 // Look out for (store (shl (load), x)).
9871 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
9884 SDValue N0 = Op.getOperand(0);
9885 SDValue N1 = Op.getOperand(1);
9886 if (!Commute && MayFoldLoad(N1))
9888 // Avoid disabling potential load folding opportunities.
9889 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
9891 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
9901 //===----------------------------------------------------------------------===//
9902 // X86 Inline Assembly Support
9903 //===----------------------------------------------------------------------===//
9905 static bool LowerToBSwap(CallInst *CI) {
9906 // FIXME: this should verify that we are targetting a 486 or better. If not,
9907 // we will turn this bswap into something that will be lowered to logical ops
9908 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9909 // so don't worry about this.
9911 // Verify this is a simple bswap.
9912 if (CI->getNumOperands() != 2 ||
9913 CI->getType() != CI->getOperand(1)->getType() ||
9914 !CI->getType()->isIntegerTy())
9917 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9918 if (!Ty || Ty->getBitWidth() % 16 != 0)
9921 // Okay, we can do this xform, do so now.
9922 const Type *Tys[] = { Ty };
9923 Module *M = CI->getParent()->getParent()->getParent();
9924 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9926 Value *Op = CI->getOperand(1);
9927 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9929 CI->replaceAllUsesWith(Op);
9930 CI->eraseFromParent();
9934 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9935 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9936 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9938 std::string AsmStr = IA->getAsmString();
9940 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9941 SmallVector<StringRef, 4> AsmPieces;
9942 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9944 switch (AsmPieces.size()) {
9945 default: return false;
9947 AsmStr = AsmPieces[0];
9949 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9952 if (AsmPieces.size() == 2 &&
9953 (AsmPieces[0] == "bswap" ||
9954 AsmPieces[0] == "bswapq" ||
9955 AsmPieces[0] == "bswapl") &&
9956 (AsmPieces[1] == "$0" ||
9957 AsmPieces[1] == "${0:q}")) {
9958 // No need to check constraints, nothing other than the equivalent of
9959 // "=r,0" would be valid here.
9960 return LowerToBSwap(CI);
9962 // rorw $$8, ${0:w} --> llvm.bswap.i16
9963 if (CI->getType()->isIntegerTy(16) &&
9964 AsmPieces.size() == 3 &&
9965 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
9966 AsmPieces[1] == "$$8," &&
9967 AsmPieces[2] == "${0:w}" &&
9968 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
9970 const std::string &Constraints = IA->getConstraintString();
9971 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
9972 std::sort(AsmPieces.begin(), AsmPieces.end());
9973 if (AsmPieces.size() == 4 &&
9974 AsmPieces[0] == "~{cc}" &&
9975 AsmPieces[1] == "~{dirflag}" &&
9976 AsmPieces[2] == "~{flags}" &&
9977 AsmPieces[3] == "~{fpsr}") {
9978 return LowerToBSwap(CI);
9983 if (CI->getType()->isIntegerTy(64) &&
9984 Constraints.size() >= 2 &&
9985 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9986 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9987 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9988 SmallVector<StringRef, 4> Words;
9989 SplitString(AsmPieces[0], Words, " \t");
9990 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9992 SplitString(AsmPieces[1], Words, " \t");
9993 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9995 SplitString(AsmPieces[2], Words, " \t,");
9996 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9997 Words[2] == "%edx") {
9998 return LowerToBSwap(CI);
10010 /// getConstraintType - Given a constraint letter, return the type of
10011 /// constraint it is for this target.
10012 X86TargetLowering::ConstraintType
10013 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10014 if (Constraint.size() == 1) {
10015 switch (Constraint[0]) {
10027 return C_RegisterClass;
10035 return TargetLowering::getConstraintType(Constraint);
10038 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10039 /// with another that has more specific requirements based on the type of the
10040 /// corresponding operand.
10041 const char *X86TargetLowering::
10042 LowerXConstraint(EVT ConstraintVT) const {
10043 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10044 // 'f' like normal targets.
10045 if (ConstraintVT.isFloatingPoint()) {
10046 if (Subtarget->hasSSE2())
10048 if (Subtarget->hasSSE1())
10052 return TargetLowering::LowerXConstraint(ConstraintVT);
10055 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10056 /// vector. If it is invalid, don't add anything to Ops.
10057 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10060 std::vector<SDValue>&Ops,
10061 SelectionDAG &DAG) const {
10062 SDValue Result(0, 0);
10064 switch (Constraint) {
10067 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10068 if (C->getZExtValue() <= 31) {
10069 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10075 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10076 if (C->getZExtValue() <= 63) {
10077 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10083 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10084 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10085 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10091 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10092 if (C->getZExtValue() <= 255) {
10093 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10099 // 32-bit signed value
10100 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10101 const ConstantInt *CI = C->getConstantIntValue();
10102 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10103 C->getSExtValue())) {
10104 // Widen to 64 bits here to get it sign extended.
10105 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10108 // FIXME gcc accepts some relocatable values here too, but only in certain
10109 // memory models; it's complicated.
10114 // 32-bit unsigned value
10115 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10116 const ConstantInt *CI = C->getConstantIntValue();
10117 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10118 C->getZExtValue())) {
10119 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10123 // FIXME gcc accepts some relocatable values here too, but only in certain
10124 // memory models; it's complicated.
10128 // Literal immediates are always ok.
10129 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10130 // Widen to 64 bits here to get it sign extended.
10131 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10135 // If we are in non-pic codegen mode, we allow the address of a global (with
10136 // an optional displacement) to be used with 'i'.
10137 GlobalAddressSDNode *GA = 0;
10138 int64_t Offset = 0;
10140 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10142 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10143 Offset += GA->getOffset();
10145 } else if (Op.getOpcode() == ISD::ADD) {
10146 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10147 Offset += C->getZExtValue();
10148 Op = Op.getOperand(0);
10151 } else if (Op.getOpcode() == ISD::SUB) {
10152 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10153 Offset += -C->getZExtValue();
10154 Op = Op.getOperand(0);
10159 // Otherwise, this isn't something we can handle, reject it.
10163 const GlobalValue *GV = GA->getGlobal();
10164 // If we require an extra load to get this address, as in PIC mode, we
10165 // can't accept it.
10166 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10167 getTargetMachine())))
10171 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
10173 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10179 if (Result.getNode()) {
10180 Ops.push_back(Result);
10183 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
10187 std::vector<unsigned> X86TargetLowering::
10188 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10190 if (Constraint.size() == 1) {
10191 // FIXME: not handling fp-stack yet!
10192 switch (Constraint[0]) { // GCC X86 Constraint Letters
10193 default: break; // Unknown constraint letter
10194 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10195 if (Subtarget->is64Bit()) {
10196 if (VT == MVT::i32)
10197 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10198 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10199 X86::R10D,X86::R11D,X86::R12D,
10200 X86::R13D,X86::R14D,X86::R15D,
10201 X86::EBP, X86::ESP, 0);
10202 else if (VT == MVT::i16)
10203 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10204 X86::SI, X86::DI, X86::R8W,X86::R9W,
10205 X86::R10W,X86::R11W,X86::R12W,
10206 X86::R13W,X86::R14W,X86::R15W,
10207 X86::BP, X86::SP, 0);
10208 else if (VT == MVT::i8)
10209 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10210 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10211 X86::R10B,X86::R11B,X86::R12B,
10212 X86::R13B,X86::R14B,X86::R15B,
10213 X86::BPL, X86::SPL, 0);
10215 else if (VT == MVT::i64)
10216 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10217 X86::RSI, X86::RDI, X86::R8, X86::R9,
10218 X86::R10, X86::R11, X86::R12,
10219 X86::R13, X86::R14, X86::R15,
10220 X86::RBP, X86::RSP, 0);
10224 // 32-bit fallthrough
10225 case 'Q': // Q_REGS
10226 if (VT == MVT::i32)
10227 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10228 else if (VT == MVT::i16)
10229 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10230 else if (VT == MVT::i8)
10231 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10232 else if (VT == MVT::i64)
10233 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10238 return std::vector<unsigned>();
10241 std::pair<unsigned, const TargetRegisterClass*>
10242 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10244 // First, see if this is a constraint that directly corresponds to an LLVM
10246 if (Constraint.size() == 1) {
10247 // GCC Constraint Letters
10248 switch (Constraint[0]) {
10250 case 'r': // GENERAL_REGS
10251 case 'l': // INDEX_REGS
10253 return std::make_pair(0U, X86::GR8RegisterClass);
10254 if (VT == MVT::i16)
10255 return std::make_pair(0U, X86::GR16RegisterClass);
10256 if (VT == MVT::i32 || !Subtarget->is64Bit())
10257 return std::make_pair(0U, X86::GR32RegisterClass);
10258 return std::make_pair(0U, X86::GR64RegisterClass);
10259 case 'R': // LEGACY_REGS
10261 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10262 if (VT == MVT::i16)
10263 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10264 if (VT == MVT::i32 || !Subtarget->is64Bit())
10265 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10266 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10267 case 'f': // FP Stack registers.
10268 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10269 // value to the correct fpstack register class.
10270 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10271 return std::make_pair(0U, X86::RFP32RegisterClass);
10272 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10273 return std::make_pair(0U, X86::RFP64RegisterClass);
10274 return std::make_pair(0U, X86::RFP80RegisterClass);
10275 case 'y': // MMX_REGS if MMX allowed.
10276 if (!Subtarget->hasMMX()) break;
10277 return std::make_pair(0U, X86::VR64RegisterClass);
10278 case 'Y': // SSE_REGS if SSE2 allowed
10279 if (!Subtarget->hasSSE2()) break;
10281 case 'x': // SSE_REGS if SSE1 allowed
10282 if (!Subtarget->hasSSE1()) break;
10284 switch (VT.getSimpleVT().SimpleTy) {
10286 // Scalar SSE types.
10289 return std::make_pair(0U, X86::FR32RegisterClass);
10292 return std::make_pair(0U, X86::FR64RegisterClass);
10300 return std::make_pair(0U, X86::VR128RegisterClass);
10306 // Use the default implementation in TargetLowering to convert the register
10307 // constraint into a member of a register class.
10308 std::pair<unsigned, const TargetRegisterClass*> Res;
10309 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10311 // Not found as a standard register?
10312 if (Res.second == 0) {
10313 // Map st(0) -> st(7) -> ST0
10314 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10315 tolower(Constraint[1]) == 's' &&
10316 tolower(Constraint[2]) == 't' &&
10317 Constraint[3] == '(' &&
10318 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10319 Constraint[5] == ')' &&
10320 Constraint[6] == '}') {
10322 Res.first = X86::ST0+Constraint[4]-'0';
10323 Res.second = X86::RFP80RegisterClass;
10327 // GCC allows "st(0)" to be called just plain "st".
10328 if (StringRef("{st}").equals_lower(Constraint)) {
10329 Res.first = X86::ST0;
10330 Res.second = X86::RFP80RegisterClass;
10335 if (StringRef("{flags}").equals_lower(Constraint)) {
10336 Res.first = X86::EFLAGS;
10337 Res.second = X86::CCRRegisterClass;
10341 // 'A' means EAX + EDX.
10342 if (Constraint == "A") {
10343 Res.first = X86::EAX;
10344 Res.second = X86::GR32_ADRegisterClass;
10350 // Otherwise, check to see if this is a register class of the wrong value
10351 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10352 // turn into {ax},{dx}.
10353 if (Res.second->hasType(VT))
10354 return Res; // Correct type already, nothing to do.
10356 // All of the single-register GCC register classes map their values onto
10357 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10358 // really want an 8-bit or 32-bit register, map to the appropriate register
10359 // class and return the appropriate register.
10360 if (Res.second == X86::GR16RegisterClass) {
10361 if (VT == MVT::i8) {
10362 unsigned DestReg = 0;
10363 switch (Res.first) {
10365 case X86::AX: DestReg = X86::AL; break;
10366 case X86::DX: DestReg = X86::DL; break;
10367 case X86::CX: DestReg = X86::CL; break;
10368 case X86::BX: DestReg = X86::BL; break;
10371 Res.first = DestReg;
10372 Res.second = X86::GR8RegisterClass;
10374 } else if (VT == MVT::i32) {
10375 unsigned DestReg = 0;
10376 switch (Res.first) {
10378 case X86::AX: DestReg = X86::EAX; break;
10379 case X86::DX: DestReg = X86::EDX; break;
10380 case X86::CX: DestReg = X86::ECX; break;
10381 case X86::BX: DestReg = X86::EBX; break;
10382 case X86::SI: DestReg = X86::ESI; break;
10383 case X86::DI: DestReg = X86::EDI; break;
10384 case X86::BP: DestReg = X86::EBP; break;
10385 case X86::SP: DestReg = X86::ESP; break;
10388 Res.first = DestReg;
10389 Res.second = X86::GR32RegisterClass;
10391 } else if (VT == MVT::i64) {
10392 unsigned DestReg = 0;
10393 switch (Res.first) {
10395 case X86::AX: DestReg = X86::RAX; break;
10396 case X86::DX: DestReg = X86::RDX; break;
10397 case X86::CX: DestReg = X86::RCX; break;
10398 case X86::BX: DestReg = X86::RBX; break;
10399 case X86::SI: DestReg = X86::RSI; break;
10400 case X86::DI: DestReg = X86::RDI; break;
10401 case X86::BP: DestReg = X86::RBP; break;
10402 case X86::SP: DestReg = X86::RSP; break;
10405 Res.first = DestReg;
10406 Res.second = X86::GR64RegisterClass;
10409 } else if (Res.second == X86::FR32RegisterClass ||
10410 Res.second == X86::FR64RegisterClass ||
10411 Res.second == X86::VR128RegisterClass) {
10412 // Handle references to XMM physical registers that got mapped into the
10413 // wrong class. This can happen with constraints like {xmm0} where the
10414 // target independent register mapper will just pick the first match it can
10415 // find, ignoring the required type.
10416 if (VT == MVT::f32)
10417 Res.second = X86::FR32RegisterClass;
10418 else if (VT == MVT::f64)
10419 Res.second = X86::FR64RegisterClass;
10420 else if (X86::VR128RegisterClass->hasType(VT))
10421 Res.second = X86::VR128RegisterClass;