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);
350 // On X86 and X86-64, atomic operations are lowered to locked instructions.
351 // Locked instructions, in turn, have implicit fence semantics (all memory
352 // operations are flushed before issuing the locked instruction, and they
353 // are not buffered), so we can fold away the common pattern of
354 // fence-atomic-fence.
355 setShouldFoldAtomicFences(true);
357 // Expand certain atomics
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
360 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
361 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
365 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
366 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
368 if (!Subtarget->is64Bit()) {
369 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
374 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
375 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
378 // FIXME - use subtarget debug flags
379 if (!Subtarget->isTargetDarwin() &&
380 !Subtarget->isTargetELF() &&
381 !Subtarget->isTargetCygMing()) {
382 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
385 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
386 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
387 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
388 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
389 if (Subtarget->is64Bit()) {
390 setExceptionPointerRegister(X86::RAX);
391 setExceptionSelectorRegister(X86::RDX);
393 setExceptionPointerRegister(X86::EAX);
394 setExceptionSelectorRegister(X86::EDX);
396 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
397 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
399 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
401 setOperationAction(ISD::TRAP, MVT::Other, Legal);
403 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
404 setOperationAction(ISD::VASTART , MVT::Other, Custom);
405 setOperationAction(ISD::VAEND , MVT::Other, Expand);
406 if (Subtarget->is64Bit()) {
407 setOperationAction(ISD::VAARG , MVT::Other, Custom);
408 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
410 setOperationAction(ISD::VAARG , MVT::Other, Expand);
411 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
414 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
415 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
416 if (Subtarget->is64Bit())
417 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
418 if (Subtarget->isTargetCygMing())
419 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
421 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
423 if (!UseSoftFloat && X86ScalarSSEf64) {
424 // f32 and f64 use SSE.
425 // Set up the FP register classes.
426 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
427 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
429 // Use ANDPD to simulate FABS.
430 setOperationAction(ISD::FABS , MVT::f64, Custom);
431 setOperationAction(ISD::FABS , MVT::f32, Custom);
433 // Use XORP to simulate FNEG.
434 setOperationAction(ISD::FNEG , MVT::f64, Custom);
435 setOperationAction(ISD::FNEG , MVT::f32, Custom);
437 // Use ANDPD and ORPD to simulate FCOPYSIGN.
438 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
439 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
441 // We don't support sin/cos/fmod
442 setOperationAction(ISD::FSIN , MVT::f64, Expand);
443 setOperationAction(ISD::FCOS , MVT::f64, Expand);
444 setOperationAction(ISD::FSIN , MVT::f32, Expand);
445 setOperationAction(ISD::FCOS , MVT::f32, Expand);
447 // Expand FP immediates into loads from the stack, except for the special
449 addLegalFPImmediate(APFloat(+0.0)); // xorpd
450 addLegalFPImmediate(APFloat(+0.0f)); // xorps
451 } else if (!UseSoftFloat && X86ScalarSSEf32) {
452 // Use SSE for f32, x87 for f64.
453 // Set up the FP register classes.
454 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
455 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
457 // Use ANDPS to simulate FABS.
458 setOperationAction(ISD::FABS , MVT::f32, Custom);
460 // Use XORP to simulate FNEG.
461 setOperationAction(ISD::FNEG , MVT::f32, Custom);
463 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
465 // Use ANDPS and ORPS to simulate FCOPYSIGN.
466 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
467 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
469 // We don't support sin/cos/fmod
470 setOperationAction(ISD::FSIN , MVT::f32, Expand);
471 setOperationAction(ISD::FCOS , MVT::f32, Expand);
473 // Special cases we handle for FP constants.
474 addLegalFPImmediate(APFloat(+0.0f)); // xorps
475 addLegalFPImmediate(APFloat(+0.0)); // FLD0
476 addLegalFPImmediate(APFloat(+1.0)); // FLD1
477 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
478 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
481 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
482 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
484 } else if (!UseSoftFloat) {
485 // f32 and f64 in x87.
486 // Set up the FP register classes.
487 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
488 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
490 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
491 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
492 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
493 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
496 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
497 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
499 addLegalFPImmediate(APFloat(+0.0)); // FLD0
500 addLegalFPImmediate(APFloat(+1.0)); // FLD1
501 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
502 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
503 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
504 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
505 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
506 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
509 // Long double always uses X87.
511 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
512 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
513 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
516 APFloat TmpFlt(+0.0);
517 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
519 addLegalFPImmediate(TmpFlt); // FLD0
521 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
522 APFloat TmpFlt2(+1.0);
523 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
525 addLegalFPImmediate(TmpFlt2); // FLD1
526 TmpFlt2.changeSign();
527 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
531 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
532 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
536 // Always use a library call for pow.
537 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
538 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
539 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
541 setOperationAction(ISD::FLOG, MVT::f80, Expand);
542 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
543 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
544 setOperationAction(ISD::FEXP, MVT::f80, Expand);
545 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
547 // First set operation action for all vector types to either promote
548 // (for widening) or expand (for scalarization). Then we will selectively
549 // turn on ones that can be effectively codegen'd.
550 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
551 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
552 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
567 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
568 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
601 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
605 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
606 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
607 setTruncStoreAction((MVT::SimpleValueType)VT,
608 (MVT::SimpleValueType)InnerVT, Expand);
609 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
610 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
611 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
614 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
615 // with -msoft-float, disable use of MMX as well.
616 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
617 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
618 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
619 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
621 // FIXME: v2f32 isn't an MMX type. We currently claim that it is legal
622 // because of some ABI issue, but this isn't the right fix.
623 bool IsV2F32Legal = !Subtarget->is64Bit();
625 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass, false);
626 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
628 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
629 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
630 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
631 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
633 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
634 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
635 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
636 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
638 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
639 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
641 setOperationAction(ISD::AND, MVT::v8i8, Promote);
642 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
643 setOperationAction(ISD::AND, MVT::v4i16, Promote);
644 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
645 setOperationAction(ISD::AND, MVT::v2i32, Promote);
646 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
647 setOperationAction(ISD::AND, MVT::v1i64, Legal);
649 setOperationAction(ISD::OR, MVT::v8i8, Promote);
650 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
651 setOperationAction(ISD::OR, MVT::v4i16, Promote);
652 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
653 setOperationAction(ISD::OR, MVT::v2i32, Promote);
654 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
655 setOperationAction(ISD::OR, MVT::v1i64, Legal);
657 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
658 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
659 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
660 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
661 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
662 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
663 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
665 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
666 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
667 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
668 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
669 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
670 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
672 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
673 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
675 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
677 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
678 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
679 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
681 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
682 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
684 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
685 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
686 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
687 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
690 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
691 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
692 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
693 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
695 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
697 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
698 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
699 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
700 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
701 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
702 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
703 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
705 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
706 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
707 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
708 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
710 setOperationAction(ISD::BIT_CONVERT, MVT::v2f32, Custom);
711 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
715 if (!UseSoftFloat && Subtarget->hasSSE1()) {
716 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
718 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
719 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
720 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
721 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
722 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
723 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
724 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
725 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
726 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
727 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
728 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
729 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
732 if (!UseSoftFloat && Subtarget->hasSSE2()) {
733 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
735 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
736 // registers cannot be used even for integer operations.
737 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
738 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
739 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
740 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
742 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
743 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
744 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
745 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
746 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
747 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
748 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
749 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
750 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
751 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
752 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
753 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
754 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
755 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
756 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
757 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
759 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
760 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
761 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
762 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
764 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
765 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
766 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
767 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
768 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
770 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
771 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
772 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
773 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
774 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
776 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
777 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
778 EVT VT = (MVT::SimpleValueType)i;
779 // Do not attempt to custom lower non-power-of-2 vectors
780 if (!isPowerOf2_32(VT.getVectorNumElements()))
782 // Do not attempt to custom lower non-128-bit vectors
783 if (!VT.is128BitVector())
785 setOperationAction(ISD::BUILD_VECTOR,
786 VT.getSimpleVT().SimpleTy, Custom);
787 setOperationAction(ISD::VECTOR_SHUFFLE,
788 VT.getSimpleVT().SimpleTy, Custom);
789 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
790 VT.getSimpleVT().SimpleTy, Custom);
793 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
794 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
795 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
796 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
797 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
798 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
800 if (Subtarget->is64Bit()) {
801 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
802 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
805 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
806 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
807 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
810 // Do not attempt to promote non-128-bit vectors
811 if (!VT.is128BitVector()) {
815 setOperationAction(ISD::AND, SVT, Promote);
816 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
817 setOperationAction(ISD::OR, SVT, Promote);
818 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
819 setOperationAction(ISD::XOR, SVT, Promote);
820 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
821 setOperationAction(ISD::LOAD, SVT, Promote);
822 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
823 setOperationAction(ISD::SELECT, SVT, Promote);
824 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
827 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
829 // Custom lower v2i64 and v2f64 selects.
830 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
831 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
832 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
833 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
835 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
836 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
837 if (!DisableMMX && Subtarget->hasMMX()) {
838 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
839 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
843 if (Subtarget->hasSSE41()) {
844 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
845 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
846 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
847 setOperationAction(ISD::FRINT, MVT::f32, Legal);
848 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
849 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
850 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
851 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
852 setOperationAction(ISD::FRINT, MVT::f64, Legal);
853 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
855 // FIXME: Do we need to handle scalar-to-vector here?
856 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
858 // i8 and i16 vectors are custom , because the source register and source
859 // source memory operand types are not the same width. f32 vectors are
860 // custom since the immediate controlling the insert encodes additional
862 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
863 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
864 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
865 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
867 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
868 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
869 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
870 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
872 if (Subtarget->is64Bit()) {
873 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
874 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
878 if (Subtarget->hasSSE42()) {
879 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
882 if (!UseSoftFloat && Subtarget->hasAVX()) {
883 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
884 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
885 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
886 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
888 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
889 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
890 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
891 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
892 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
893 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
894 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
895 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
896 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
897 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
898 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
899 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
900 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
901 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
902 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
904 // Operations to consider commented out -v16i16 v32i8
905 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
906 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
907 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
908 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
909 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
910 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
911 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
912 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
913 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
914 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
915 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
916 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
917 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
918 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
920 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
921 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
922 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
923 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
925 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
926 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
927 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
928 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
929 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
931 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
932 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
933 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
934 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
935 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
936 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
939 // Not sure we want to do this since there are no 256-bit integer
942 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
943 // This includes 256-bit vectors
944 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
945 EVT VT = (MVT::SimpleValueType)i;
947 // Do not attempt to custom lower non-power-of-2 vectors
948 if (!isPowerOf2_32(VT.getVectorNumElements()))
951 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
952 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
953 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
956 if (Subtarget->is64Bit()) {
957 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
958 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
963 // Not sure we want to do this since there are no 256-bit integer
966 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
967 // Including 256-bit vectors
968 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
969 EVT VT = (MVT::SimpleValueType)i;
971 if (!VT.is256BitVector()) {
974 setOperationAction(ISD::AND, VT, Promote);
975 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
976 setOperationAction(ISD::OR, VT, Promote);
977 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
978 setOperationAction(ISD::XOR, VT, Promote);
979 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
980 setOperationAction(ISD::LOAD, VT, Promote);
981 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
982 setOperationAction(ISD::SELECT, VT, Promote);
983 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
986 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
990 // We want to custom lower some of our intrinsics.
991 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
993 // Add/Sub/Mul with overflow operations are custom lowered.
994 setOperationAction(ISD::SADDO, MVT::i32, Custom);
995 setOperationAction(ISD::UADDO, MVT::i32, Custom);
996 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
997 setOperationAction(ISD::USUBO, MVT::i32, Custom);
998 setOperationAction(ISD::SMULO, MVT::i32, Custom);
1000 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1001 // handle type legalization for these operations here.
1003 // FIXME: We really should do custom legalization for addition and
1004 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1005 // than generic legalization for 64-bit multiplication-with-overflow, though.
1006 if (Subtarget->is64Bit()) {
1007 setOperationAction(ISD::SADDO, MVT::i64, Custom);
1008 setOperationAction(ISD::UADDO, MVT::i64, Custom);
1009 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
1010 setOperationAction(ISD::USUBO, MVT::i64, Custom);
1011 setOperationAction(ISD::SMULO, MVT::i64, Custom);
1014 if (!Subtarget->is64Bit()) {
1015 // These libcalls are not available in 32-bit.
1016 setLibcallName(RTLIB::SHL_I128, 0);
1017 setLibcallName(RTLIB::SRL_I128, 0);
1018 setLibcallName(RTLIB::SRA_I128, 0);
1021 // We have target-specific dag combine patterns for the following nodes:
1022 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1023 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1024 setTargetDAGCombine(ISD::BUILD_VECTOR);
1025 setTargetDAGCombine(ISD::SELECT);
1026 setTargetDAGCombine(ISD::SHL);
1027 setTargetDAGCombine(ISD::SRA);
1028 setTargetDAGCombine(ISD::SRL);
1029 setTargetDAGCombine(ISD::OR);
1030 setTargetDAGCombine(ISD::STORE);
1031 setTargetDAGCombine(ISD::ZERO_EXTEND);
1032 if (Subtarget->is64Bit())
1033 setTargetDAGCombine(ISD::MUL);
1035 computeRegisterProperties();
1037 // FIXME: These should be based on subtarget info. Plus, the values should
1038 // be smaller when we are in optimizing for size mode.
1039 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1040 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1041 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1042 setPrefLoopAlignment(16);
1043 benefitFromCodePlacementOpt = true;
1047 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1052 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1053 /// the desired ByVal argument alignment.
1054 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1057 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1058 if (VTy->getBitWidth() == 128)
1060 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1061 unsigned EltAlign = 0;
1062 getMaxByValAlign(ATy->getElementType(), EltAlign);
1063 if (EltAlign > MaxAlign)
1064 MaxAlign = EltAlign;
1065 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1066 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1067 unsigned EltAlign = 0;
1068 getMaxByValAlign(STy->getElementType(i), EltAlign);
1069 if (EltAlign > MaxAlign)
1070 MaxAlign = EltAlign;
1078 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1079 /// function arguments in the caller parameter area. For X86, aggregates
1080 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1081 /// are at 4-byte boundaries.
1082 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1083 if (Subtarget->is64Bit()) {
1084 // Max of 8 and alignment of type.
1085 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1092 if (Subtarget->hasSSE1())
1093 getMaxByValAlign(Ty, Align);
1097 /// getOptimalMemOpType - Returns the target specific optimal type for load
1098 /// and store operations as a result of memset, memcpy, and memmove
1099 /// lowering. If DstAlign is zero that means it's safe to destination
1100 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1101 /// means there isn't a need to check it against alignment requirement,
1102 /// probably because the source does not need to be loaded. If
1103 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1104 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1105 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1106 /// constant so it does not need to be loaded.
1107 /// It returns EVT::Other if the type should be determined using generic
1108 /// target-independent logic.
1110 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1111 unsigned DstAlign, unsigned SrcAlign,
1112 bool NonScalarIntSafe,
1114 MachineFunction &MF) const {
1115 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1116 // linux. This is because the stack realignment code can't handle certain
1117 // cases like PR2962. This should be removed when PR2962 is fixed.
1118 const Function *F = MF.getFunction();
1119 if (NonScalarIntSafe &&
1120 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1122 (Subtarget->isUnalignedMemAccessFast() ||
1123 ((DstAlign == 0 || DstAlign >= 16) &&
1124 (SrcAlign == 0 || SrcAlign >= 16))) &&
1125 Subtarget->getStackAlignment() >= 16) {
1126 if (Subtarget->hasSSE2())
1128 if (Subtarget->hasSSE1())
1130 } else if (!MemcpyStrSrc && Size >= 8 &&
1131 !Subtarget->is64Bit() &&
1132 Subtarget->getStackAlignment() >= 8 &&
1133 Subtarget->hasSSE2()) {
1134 // Do not use f64 to lower memcpy if source is string constant. It's
1135 // better to use i32 to avoid the loads.
1139 if (Subtarget->is64Bit() && Size >= 8)
1144 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1145 /// current function. The returned value is a member of the
1146 /// MachineJumpTableInfo::JTEntryKind enum.
1147 unsigned X86TargetLowering::getJumpTableEncoding() const {
1148 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1150 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1151 Subtarget->isPICStyleGOT())
1152 return MachineJumpTableInfo::EK_Custom32;
1154 // Otherwise, use the normal jump table encoding heuristics.
1155 return TargetLowering::getJumpTableEncoding();
1158 /// getPICBaseSymbol - Return the X86-32 PIC base.
1160 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1161 MCContext &Ctx) const {
1162 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1163 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1164 Twine(MF->getFunctionNumber())+"$pb");
1169 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1170 const MachineBasicBlock *MBB,
1171 unsigned uid,MCContext &Ctx) const{
1172 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1173 Subtarget->isPICStyleGOT());
1174 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1176 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1177 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1180 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1182 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1183 SelectionDAG &DAG) const {
1184 if (!Subtarget->is64Bit())
1185 // This doesn't have DebugLoc associated with it, but is not really the
1186 // same as a Register.
1187 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1191 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1192 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1194 const MCExpr *X86TargetLowering::
1195 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1196 MCContext &Ctx) const {
1197 // X86-64 uses RIP relative addressing based on the jump table label.
1198 if (Subtarget->isPICStyleRIPRel())
1199 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1201 // Otherwise, the reference is relative to the PIC base.
1202 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1205 /// getFunctionAlignment - Return the Log2 alignment of this function.
1206 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1207 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1210 //===----------------------------------------------------------------------===//
1211 // Return Value Calling Convention Implementation
1212 //===----------------------------------------------------------------------===//
1214 #include "X86GenCallingConv.inc"
1217 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1218 const SmallVectorImpl<EVT> &OutTys,
1219 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1220 SelectionDAG &DAG) const {
1221 SmallVector<CCValAssign, 16> RVLocs;
1222 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1223 RVLocs, *DAG.getContext());
1224 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1228 X86TargetLowering::LowerReturn(SDValue Chain,
1229 CallingConv::ID CallConv, bool isVarArg,
1230 const SmallVectorImpl<ISD::OutputArg> &Outs,
1231 DebugLoc dl, SelectionDAG &DAG) const {
1232 MachineFunction &MF = DAG.getMachineFunction();
1233 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1235 SmallVector<CCValAssign, 16> RVLocs;
1236 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1237 RVLocs, *DAG.getContext());
1238 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1240 // Add the regs to the liveout set for the function.
1241 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1242 for (unsigned i = 0; i != RVLocs.size(); ++i)
1243 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1244 MRI.addLiveOut(RVLocs[i].getLocReg());
1248 SmallVector<SDValue, 6> RetOps;
1249 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1250 // Operand #1 = Bytes To Pop
1251 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1254 // Copy the result values into the output registers.
1255 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1256 CCValAssign &VA = RVLocs[i];
1257 assert(VA.isRegLoc() && "Can only return in registers!");
1258 SDValue ValToCopy = Outs[i].Val;
1260 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1261 // the RET instruction and handled by the FP Stackifier.
1262 if (VA.getLocReg() == X86::ST0 ||
1263 VA.getLocReg() == X86::ST1) {
1264 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1265 // change the value to the FP stack register class.
1266 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1267 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1268 RetOps.push_back(ValToCopy);
1269 // Don't emit a copytoreg.
1273 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1274 // which is returned in RAX / RDX.
1275 if (Subtarget->is64Bit()) {
1276 EVT ValVT = ValToCopy.getValueType();
1277 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1278 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1279 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1280 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1284 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1285 Flag = Chain.getValue(1);
1288 // The x86-64 ABI for returning structs by value requires that we copy
1289 // the sret argument into %rax for the return. We saved the argument into
1290 // a virtual register in the entry block, so now we copy the value out
1292 if (Subtarget->is64Bit() &&
1293 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1294 MachineFunction &MF = DAG.getMachineFunction();
1295 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1296 unsigned Reg = FuncInfo->getSRetReturnReg();
1298 "SRetReturnReg should have been set in LowerFormalArguments().");
1299 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1301 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1302 Flag = Chain.getValue(1);
1304 // RAX now acts like a return value.
1305 MRI.addLiveOut(X86::RAX);
1308 RetOps[0] = Chain; // Update chain.
1310 // Add the flag if we have it.
1312 RetOps.push_back(Flag);
1314 return DAG.getNode(X86ISD::RET_FLAG, dl,
1315 MVT::Other, &RetOps[0], RetOps.size());
1318 /// LowerCallResult - Lower the result values of a call into the
1319 /// appropriate copies out of appropriate physical registers.
1322 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1323 CallingConv::ID CallConv, bool isVarArg,
1324 const SmallVectorImpl<ISD::InputArg> &Ins,
1325 DebugLoc dl, SelectionDAG &DAG,
1326 SmallVectorImpl<SDValue> &InVals) const {
1328 // Assign locations to each value returned by this call.
1329 SmallVector<CCValAssign, 16> RVLocs;
1330 bool Is64Bit = Subtarget->is64Bit();
1331 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1332 RVLocs, *DAG.getContext());
1333 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1335 // Copy all of the result registers out of their specified physreg.
1336 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1337 CCValAssign &VA = RVLocs[i];
1338 EVT CopyVT = VA.getValVT();
1340 // If this is x86-64, and we disabled SSE, we can't return FP values
1341 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1342 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1343 report_fatal_error("SSE register return with SSE disabled");
1346 // If this is a call to a function that returns an fp value on the floating
1347 // point stack, but where we prefer to use the value in xmm registers, copy
1348 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1349 if ((VA.getLocReg() == X86::ST0 ||
1350 VA.getLocReg() == X86::ST1) &&
1351 isScalarFPTypeInSSEReg(VA.getValVT())) {
1356 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1357 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1358 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1359 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1360 MVT::v2i64, InFlag).getValue(1);
1361 Val = Chain.getValue(0);
1362 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1363 Val, DAG.getConstant(0, MVT::i64));
1365 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1366 MVT::i64, InFlag).getValue(1);
1367 Val = Chain.getValue(0);
1369 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1371 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1372 CopyVT, InFlag).getValue(1);
1373 Val = Chain.getValue(0);
1375 InFlag = Chain.getValue(2);
1377 if (CopyVT != VA.getValVT()) {
1378 // Round the F80 the right size, which also moves to the appropriate xmm
1380 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1381 // This truncation won't change the value.
1382 DAG.getIntPtrConstant(1));
1385 InVals.push_back(Val);
1392 //===----------------------------------------------------------------------===//
1393 // C & StdCall & Fast Calling Convention implementation
1394 //===----------------------------------------------------------------------===//
1395 // StdCall calling convention seems to be standard for many Windows' API
1396 // routines and around. It differs from C calling convention just a little:
1397 // callee should clean up the stack, not caller. Symbols should be also
1398 // decorated in some fancy way :) It doesn't support any vector arguments.
1399 // For info on fast calling convention see Fast Calling Convention (tail call)
1400 // implementation LowerX86_32FastCCCallTo.
1402 /// CallIsStructReturn - Determines whether a call uses struct return
1404 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1408 return Outs[0].Flags.isSRet();
1411 /// ArgsAreStructReturn - Determines whether a function uses struct
1412 /// return semantics.
1414 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1418 return Ins[0].Flags.isSRet();
1421 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1422 /// given CallingConvention value.
1423 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1424 if (Subtarget->is64Bit()) {
1425 if (CC == CallingConv::GHC)
1426 return CC_X86_64_GHC;
1427 else if (Subtarget->isTargetWin64())
1428 return CC_X86_Win64_C;
1433 if (CC == CallingConv::X86_FastCall)
1434 return CC_X86_32_FastCall;
1435 else if (CC == CallingConv::X86_ThisCall)
1436 return CC_X86_32_ThisCall;
1437 else if (CC == CallingConv::Fast)
1438 return CC_X86_32_FastCC;
1439 else if (CC == CallingConv::GHC)
1440 return CC_X86_32_GHC;
1445 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1446 /// by "Src" to address "Dst" with size and alignment information specified by
1447 /// the specific parameter attribute. The copy will be passed as a byval
1448 /// function parameter.
1450 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1451 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1453 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1454 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1455 /*isVolatile*/false, /*AlwaysInline=*/true,
1459 /// IsTailCallConvention - Return true if the calling convention is one that
1460 /// supports tail call optimization.
1461 static bool IsTailCallConvention(CallingConv::ID CC) {
1462 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1465 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1466 /// a tailcall target by changing its ABI.
1467 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1468 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1472 X86TargetLowering::LowerMemArgument(SDValue Chain,
1473 CallingConv::ID CallConv,
1474 const SmallVectorImpl<ISD::InputArg> &Ins,
1475 DebugLoc dl, SelectionDAG &DAG,
1476 const CCValAssign &VA,
1477 MachineFrameInfo *MFI,
1479 // Create the nodes corresponding to a load from this parameter slot.
1480 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1481 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1482 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1485 // If value is passed by pointer we have address passed instead of the value
1487 if (VA.getLocInfo() == CCValAssign::Indirect)
1488 ValVT = VA.getLocVT();
1490 ValVT = VA.getValVT();
1492 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1493 // changed with more analysis.
1494 // In case of tail call optimization mark all arguments mutable. Since they
1495 // could be overwritten by lowering of arguments in case of a tail call.
1496 if (Flags.isByVal()) {
1497 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1498 VA.getLocMemOffset(), isImmutable);
1499 return DAG.getFrameIndex(FI, getPointerTy());
1501 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1502 VA.getLocMemOffset(), isImmutable);
1503 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1504 return DAG.getLoad(ValVT, dl, Chain, FIN,
1505 PseudoSourceValue::getFixedStack(FI), 0,
1511 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1512 CallingConv::ID CallConv,
1514 const SmallVectorImpl<ISD::InputArg> &Ins,
1517 SmallVectorImpl<SDValue> &InVals)
1519 MachineFunction &MF = DAG.getMachineFunction();
1520 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1522 const Function* Fn = MF.getFunction();
1523 if (Fn->hasExternalLinkage() &&
1524 Subtarget->isTargetCygMing() &&
1525 Fn->getName() == "main")
1526 FuncInfo->setForceFramePointer(true);
1528 MachineFrameInfo *MFI = MF.getFrameInfo();
1529 bool Is64Bit = Subtarget->is64Bit();
1530 bool IsWin64 = Subtarget->isTargetWin64();
1532 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1533 "Var args not supported with calling convention fastcc or ghc");
1535 // Assign locations to all of the incoming arguments.
1536 SmallVector<CCValAssign, 16> ArgLocs;
1537 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1538 ArgLocs, *DAG.getContext());
1539 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1541 unsigned LastVal = ~0U;
1543 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1544 CCValAssign &VA = ArgLocs[i];
1545 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1547 assert(VA.getValNo() != LastVal &&
1548 "Don't support value assigned to multiple locs yet");
1549 LastVal = VA.getValNo();
1551 if (VA.isRegLoc()) {
1552 EVT RegVT = VA.getLocVT();
1553 TargetRegisterClass *RC = NULL;
1554 if (RegVT == MVT::i32)
1555 RC = X86::GR32RegisterClass;
1556 else if (Is64Bit && RegVT == MVT::i64)
1557 RC = X86::GR64RegisterClass;
1558 else if (RegVT == MVT::f32)
1559 RC = X86::FR32RegisterClass;
1560 else if (RegVT == MVT::f64)
1561 RC = X86::FR64RegisterClass;
1562 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1563 RC = X86::VR128RegisterClass;
1564 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1565 RC = X86::VR64RegisterClass;
1567 llvm_unreachable("Unknown argument type!");
1569 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1570 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1572 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1573 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1575 if (VA.getLocInfo() == CCValAssign::SExt)
1576 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1577 DAG.getValueType(VA.getValVT()));
1578 else if (VA.getLocInfo() == CCValAssign::ZExt)
1579 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1580 DAG.getValueType(VA.getValVT()));
1581 else if (VA.getLocInfo() == CCValAssign::BCvt)
1582 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1584 if (VA.isExtInLoc()) {
1585 // Handle MMX values passed in XMM regs.
1586 if (RegVT.isVector()) {
1587 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1588 ArgValue, DAG.getConstant(0, MVT::i64));
1589 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1591 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1594 assert(VA.isMemLoc());
1595 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1598 // If value is passed via pointer - do a load.
1599 if (VA.getLocInfo() == CCValAssign::Indirect)
1600 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1603 InVals.push_back(ArgValue);
1606 // The x86-64 ABI for returning structs by value requires that we copy
1607 // the sret argument into %rax for the return. Save the argument into
1608 // a virtual register so that we can access it from the return points.
1609 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1610 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1611 unsigned Reg = FuncInfo->getSRetReturnReg();
1613 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1614 FuncInfo->setSRetReturnReg(Reg);
1616 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1617 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1620 unsigned StackSize = CCInfo.getNextStackOffset();
1621 // Align stack specially for tail calls.
1622 if (FuncIsMadeTailCallSafe(CallConv))
1623 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1625 // If the function takes variable number of arguments, make a frame index for
1626 // the start of the first vararg value... for expansion of llvm.va_start.
1628 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1629 CallConv != CallingConv::X86_ThisCall)) {
1630 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1633 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1635 // FIXME: We should really autogenerate these arrays
1636 static const unsigned GPR64ArgRegsWin64[] = {
1637 X86::RCX, X86::RDX, X86::R8, X86::R9
1639 static const unsigned XMMArgRegsWin64[] = {
1640 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1642 static const unsigned GPR64ArgRegs64Bit[] = {
1643 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1645 static const unsigned XMMArgRegs64Bit[] = {
1646 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1647 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1649 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1652 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1653 GPR64ArgRegs = GPR64ArgRegsWin64;
1654 XMMArgRegs = XMMArgRegsWin64;
1656 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1657 GPR64ArgRegs = GPR64ArgRegs64Bit;
1658 XMMArgRegs = XMMArgRegs64Bit;
1660 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1662 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1665 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1666 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1667 "SSE register cannot be used when SSE is disabled!");
1668 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1669 "SSE register cannot be used when SSE is disabled!");
1670 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1671 // Kernel mode asks for SSE to be disabled, so don't push them
1673 TotalNumXMMRegs = 0;
1675 // For X86-64, if there are vararg parameters that are passed via
1676 // registers, then we must store them to their spots on the stack so they
1677 // may be loaded by deferencing the result of va_next.
1678 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1679 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1680 FuncInfo->setRegSaveFrameIndex(
1681 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1684 // Store the integer parameter registers.
1685 SmallVector<SDValue, 8> MemOps;
1686 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1688 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1689 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1690 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1691 DAG.getIntPtrConstant(Offset));
1692 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1693 X86::GR64RegisterClass);
1694 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1696 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1697 PseudoSourceValue::getFixedStack(
1698 FuncInfo->getRegSaveFrameIndex()),
1699 Offset, false, false, 0);
1700 MemOps.push_back(Store);
1704 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1705 // Now store the XMM (fp + vector) parameter registers.
1706 SmallVector<SDValue, 11> SaveXMMOps;
1707 SaveXMMOps.push_back(Chain);
1709 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1710 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1711 SaveXMMOps.push_back(ALVal);
1713 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1714 FuncInfo->getRegSaveFrameIndex()));
1715 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1716 FuncInfo->getVarArgsFPOffset()));
1718 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1719 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1720 X86::VR128RegisterClass);
1721 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1722 SaveXMMOps.push_back(Val);
1724 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1726 &SaveXMMOps[0], SaveXMMOps.size()));
1729 if (!MemOps.empty())
1730 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1731 &MemOps[0], MemOps.size());
1735 // Some CCs need callee pop.
1736 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1737 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1739 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1740 // If this is an sret function, the return should pop the hidden pointer.
1741 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1742 FuncInfo->setBytesToPopOnReturn(4);
1746 // RegSaveFrameIndex is X86-64 only.
1747 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1748 if (CallConv == CallingConv::X86_FastCall ||
1749 CallConv == CallingConv::X86_ThisCall)
1750 // fastcc functions can't have varargs.
1751 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1758 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1759 SDValue StackPtr, SDValue Arg,
1760 DebugLoc dl, SelectionDAG &DAG,
1761 const CCValAssign &VA,
1762 ISD::ArgFlagsTy Flags) const {
1763 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1764 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1765 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1766 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1767 if (Flags.isByVal()) {
1768 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1770 return DAG.getStore(Chain, dl, Arg, PtrOff,
1771 PseudoSourceValue::getStack(), LocMemOffset,
1775 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1776 /// optimization is performed and it is required.
1778 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1779 SDValue &OutRetAddr, SDValue Chain,
1780 bool IsTailCall, bool Is64Bit,
1781 int FPDiff, DebugLoc dl) const {
1782 // Adjust the Return address stack slot.
1783 EVT VT = getPointerTy();
1784 OutRetAddr = getReturnAddressFrameIndex(DAG);
1786 // Load the "old" Return address.
1787 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1788 return SDValue(OutRetAddr.getNode(), 1);
1791 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1792 /// optimization is performed and it is required (FPDiff!=0).
1794 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1795 SDValue Chain, SDValue RetAddrFrIdx,
1796 bool Is64Bit, int FPDiff, DebugLoc dl) {
1797 // Store the return address to the appropriate stack slot.
1798 if (!FPDiff) return Chain;
1799 // Calculate the new stack slot for the return address.
1800 int SlotSize = Is64Bit ? 8 : 4;
1801 int NewReturnAddrFI =
1802 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1803 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1804 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1805 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1806 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1812 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1813 CallingConv::ID CallConv, bool isVarArg,
1815 const SmallVectorImpl<ISD::OutputArg> &Outs,
1816 const SmallVectorImpl<ISD::InputArg> &Ins,
1817 DebugLoc dl, SelectionDAG &DAG,
1818 SmallVectorImpl<SDValue> &InVals) const {
1819 MachineFunction &MF = DAG.getMachineFunction();
1820 bool Is64Bit = Subtarget->is64Bit();
1821 bool IsStructRet = CallIsStructReturn(Outs);
1822 bool IsSibcall = false;
1825 // Check if it's really possible to do a tail call.
1826 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1827 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1830 // Sibcalls are automatically detected tailcalls which do not require
1832 if (!GuaranteedTailCallOpt && isTailCall)
1839 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1840 "Var args not supported with calling convention fastcc or ghc");
1842 // Analyze operands of the call, assigning locations to each operand.
1843 SmallVector<CCValAssign, 16> ArgLocs;
1844 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1845 ArgLocs, *DAG.getContext());
1846 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1848 // Get a count of how many bytes are to be pushed on the stack.
1849 unsigned NumBytes = CCInfo.getNextStackOffset();
1851 // This is a sibcall. The memory operands are available in caller's
1852 // own caller's stack.
1854 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1855 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1858 if (isTailCall && !IsSibcall) {
1859 // Lower arguments at fp - stackoffset + fpdiff.
1860 unsigned NumBytesCallerPushed =
1861 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1862 FPDiff = NumBytesCallerPushed - NumBytes;
1864 // Set the delta of movement of the returnaddr stackslot.
1865 // But only set if delta is greater than previous delta.
1866 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1867 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1871 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1873 SDValue RetAddrFrIdx;
1874 // Load return adress for tail calls.
1875 if (isTailCall && FPDiff)
1876 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1877 Is64Bit, FPDiff, dl);
1879 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1880 SmallVector<SDValue, 8> MemOpChains;
1883 // Walk the register/memloc assignments, inserting copies/loads. In the case
1884 // of tail call optimization arguments are handle later.
1885 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1886 CCValAssign &VA = ArgLocs[i];
1887 EVT RegVT = VA.getLocVT();
1888 SDValue Arg = Outs[i].Val;
1889 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1890 bool isByVal = Flags.isByVal();
1892 // Promote the value if needed.
1893 switch (VA.getLocInfo()) {
1894 default: llvm_unreachable("Unknown loc info!");
1895 case CCValAssign::Full: break;
1896 case CCValAssign::SExt:
1897 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1899 case CCValAssign::ZExt:
1900 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1902 case CCValAssign::AExt:
1903 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1904 // Special case: passing MMX values in XMM registers.
1905 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1906 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1907 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1909 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1911 case CCValAssign::BCvt:
1912 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1914 case CCValAssign::Indirect: {
1915 // Store the argument.
1916 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1917 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1918 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1919 PseudoSourceValue::getFixedStack(FI), 0,
1926 if (VA.isRegLoc()) {
1927 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1928 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1929 assert(VA.isMemLoc());
1930 if (StackPtr.getNode() == 0)
1931 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1932 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1933 dl, DAG, VA, Flags));
1937 if (!MemOpChains.empty())
1938 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1939 &MemOpChains[0], MemOpChains.size());
1941 // Build a sequence of copy-to-reg nodes chained together with token chain
1942 // and flag operands which copy the outgoing args into registers.
1944 // Tail call byval lowering might overwrite argument registers so in case of
1945 // tail call optimization the copies to registers are lowered later.
1947 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1948 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1949 RegsToPass[i].second, InFlag);
1950 InFlag = Chain.getValue(1);
1953 if (Subtarget->isPICStyleGOT()) {
1954 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1957 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1958 DAG.getNode(X86ISD::GlobalBaseReg,
1959 DebugLoc(), getPointerTy()),
1961 InFlag = Chain.getValue(1);
1963 // If we are tail calling and generating PIC/GOT style code load the
1964 // address of the callee into ECX. The value in ecx is used as target of
1965 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1966 // for tail calls on PIC/GOT architectures. Normally we would just put the
1967 // address of GOT into ebx and then call target@PLT. But for tail calls
1968 // ebx would be restored (since ebx is callee saved) before jumping to the
1971 // Note: The actual moving to ECX is done further down.
1972 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1973 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1974 !G->getGlobal()->hasProtectedVisibility())
1975 Callee = LowerGlobalAddress(Callee, DAG);
1976 else if (isa<ExternalSymbolSDNode>(Callee))
1977 Callee = LowerExternalSymbol(Callee, DAG);
1981 if (Is64Bit && isVarArg) {
1982 // From AMD64 ABI document:
1983 // For calls that may call functions that use varargs or stdargs
1984 // (prototype-less calls or calls to functions containing ellipsis (...) in
1985 // the declaration) %al is used as hidden argument to specify the number
1986 // of SSE registers used. The contents of %al do not need to match exactly
1987 // the number of registers, but must be an ubound on the number of SSE
1988 // registers used and is in the range 0 - 8 inclusive.
1990 // FIXME: Verify this on Win64
1991 // Count the number of XMM registers allocated.
1992 static const unsigned XMMArgRegs[] = {
1993 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1994 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1996 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1997 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1998 && "SSE registers cannot be used when SSE is disabled");
2000 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2001 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2002 InFlag = Chain.getValue(1);
2006 // For tail calls lower the arguments to the 'real' stack slot.
2008 // Force all the incoming stack arguments to be loaded from the stack
2009 // before any new outgoing arguments are stored to the stack, because the
2010 // outgoing stack slots may alias the incoming argument stack slots, and
2011 // the alias isn't otherwise explicit. This is slightly more conservative
2012 // than necessary, because it means that each store effectively depends
2013 // on every argument instead of just those arguments it would clobber.
2014 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2016 SmallVector<SDValue, 8> MemOpChains2;
2019 // Do not flag preceeding copytoreg stuff together with the following stuff.
2021 if (GuaranteedTailCallOpt) {
2022 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2023 CCValAssign &VA = ArgLocs[i];
2026 assert(VA.isMemLoc());
2027 SDValue Arg = Outs[i].Val;
2028 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2029 // Create frame index.
2030 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2031 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2032 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2033 FIN = DAG.getFrameIndex(FI, getPointerTy());
2035 if (Flags.isByVal()) {
2036 // Copy relative to framepointer.
2037 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2038 if (StackPtr.getNode() == 0)
2039 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2041 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2043 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2047 // Store relative to framepointer.
2048 MemOpChains2.push_back(
2049 DAG.getStore(ArgChain, dl, Arg, FIN,
2050 PseudoSourceValue::getFixedStack(FI), 0,
2056 if (!MemOpChains2.empty())
2057 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2058 &MemOpChains2[0], MemOpChains2.size());
2060 // Copy arguments to their registers.
2061 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2062 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2063 RegsToPass[i].second, InFlag);
2064 InFlag = Chain.getValue(1);
2068 // Store the return address to the appropriate stack slot.
2069 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2073 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2074 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2075 // In the 64-bit large code model, we have to make all calls
2076 // through a register, since the call instruction's 32-bit
2077 // pc-relative offset may not be large enough to hold the whole
2079 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2080 // If the callee is a GlobalAddress node (quite common, every direct call
2081 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2084 // We should use extra load for direct calls to dllimported functions in
2086 const GlobalValue *GV = G->getGlobal();
2087 if (!GV->hasDLLImportLinkage()) {
2088 unsigned char OpFlags = 0;
2090 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2091 // external symbols most go through the PLT in PIC mode. If the symbol
2092 // has hidden or protected visibility, or if it is static or local, then
2093 // we don't need to use the PLT - we can directly call it.
2094 if (Subtarget->isTargetELF() &&
2095 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2096 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2097 OpFlags = X86II::MO_PLT;
2098 } else if (Subtarget->isPICStyleStubAny() &&
2099 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2100 Subtarget->getDarwinVers() < 9) {
2101 // PC-relative references to external symbols should go through $stub,
2102 // unless we're building with the leopard linker or later, which
2103 // automatically synthesizes these stubs.
2104 OpFlags = X86II::MO_DARWIN_STUB;
2107 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2108 G->getOffset(), OpFlags);
2110 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2111 unsigned char OpFlags = 0;
2113 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2114 // symbols should go through the PLT.
2115 if (Subtarget->isTargetELF() &&
2116 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2117 OpFlags = X86II::MO_PLT;
2118 } else if (Subtarget->isPICStyleStubAny() &&
2119 Subtarget->getDarwinVers() < 9) {
2120 // PC-relative references to external symbols should go through $stub,
2121 // unless we're building with the leopard linker or later, which
2122 // automatically synthesizes these stubs.
2123 OpFlags = X86II::MO_DARWIN_STUB;
2126 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2130 // Returns a chain & a flag for retval copy to use.
2131 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2132 SmallVector<SDValue, 8> Ops;
2134 if (!IsSibcall && isTailCall) {
2135 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2136 DAG.getIntPtrConstant(0, true), InFlag);
2137 InFlag = Chain.getValue(1);
2140 Ops.push_back(Chain);
2141 Ops.push_back(Callee);
2144 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2146 // Add argument registers to the end of the list so that they are known live
2148 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2149 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2150 RegsToPass[i].second.getValueType()));
2152 // Add an implicit use GOT pointer in EBX.
2153 if (!isTailCall && Subtarget->isPICStyleGOT())
2154 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2156 // Add an implicit use of AL for x86 vararg functions.
2157 if (Is64Bit && isVarArg)
2158 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2160 if (InFlag.getNode())
2161 Ops.push_back(InFlag);
2165 //// If this is the first return lowered for this function, add the regs
2166 //// to the liveout set for the function.
2167 // This isn't right, although it's probably harmless on x86; liveouts
2168 // should be computed from returns not tail calls. Consider a void
2169 // function making a tail call to a function returning int.
2170 return DAG.getNode(X86ISD::TC_RETURN, dl,
2171 NodeTys, &Ops[0], Ops.size());
2174 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2175 InFlag = Chain.getValue(1);
2177 // Create the CALLSEQ_END node.
2178 unsigned NumBytesForCalleeToPush;
2179 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2180 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2181 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2182 // If this is a call to a struct-return function, the callee
2183 // pops the hidden struct pointer, so we have to push it back.
2184 // This is common for Darwin/X86, Linux & Mingw32 targets.
2185 NumBytesForCalleeToPush = 4;
2187 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2189 // Returns a flag for retval copy to use.
2191 Chain = DAG.getCALLSEQ_END(Chain,
2192 DAG.getIntPtrConstant(NumBytes, true),
2193 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2196 InFlag = Chain.getValue(1);
2199 // Handle result values, copying them out of physregs into vregs that we
2201 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2202 Ins, dl, DAG, InVals);
2206 //===----------------------------------------------------------------------===//
2207 // Fast Calling Convention (tail call) implementation
2208 //===----------------------------------------------------------------------===//
2210 // Like std call, callee cleans arguments, convention except that ECX is
2211 // reserved for storing the tail called function address. Only 2 registers are
2212 // free for argument passing (inreg). Tail call optimization is performed
2214 // * tailcallopt is enabled
2215 // * caller/callee are fastcc
2216 // On X86_64 architecture with GOT-style position independent code only local
2217 // (within module) calls are supported at the moment.
2218 // To keep the stack aligned according to platform abi the function
2219 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2220 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2221 // If a tail called function callee has more arguments than the caller the
2222 // caller needs to make sure that there is room to move the RETADDR to. This is
2223 // achieved by reserving an area the size of the argument delta right after the
2224 // original REtADDR, but before the saved framepointer or the spilled registers
2225 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2237 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2238 /// for a 16 byte align requirement.
2240 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2241 SelectionDAG& DAG) const {
2242 MachineFunction &MF = DAG.getMachineFunction();
2243 const TargetMachine &TM = MF.getTarget();
2244 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2245 unsigned StackAlignment = TFI.getStackAlignment();
2246 uint64_t AlignMask = StackAlignment - 1;
2247 int64_t Offset = StackSize;
2248 uint64_t SlotSize = TD->getPointerSize();
2249 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2250 // Number smaller than 12 so just add the difference.
2251 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2253 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2254 Offset = ((~AlignMask) & Offset) + StackAlignment +
2255 (StackAlignment-SlotSize);
2260 /// MatchingStackOffset - Return true if the given stack call argument is
2261 /// already available in the same position (relatively) of the caller's
2262 /// incoming argument stack.
2264 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2265 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2266 const X86InstrInfo *TII) {
2267 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2269 if (Arg.getOpcode() == ISD::CopyFromReg) {
2270 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2271 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2273 MachineInstr *Def = MRI->getVRegDef(VR);
2276 if (!Flags.isByVal()) {
2277 if (!TII->isLoadFromStackSlot(Def, FI))
2280 unsigned Opcode = Def->getOpcode();
2281 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2282 Def->getOperand(1).isFI()) {
2283 FI = Def->getOperand(1).getIndex();
2284 Bytes = Flags.getByValSize();
2288 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2289 if (Flags.isByVal())
2290 // ByVal argument is passed in as a pointer but it's now being
2291 // dereferenced. e.g.
2292 // define @foo(%struct.X* %A) {
2293 // tail call @bar(%struct.X* byval %A)
2296 SDValue Ptr = Ld->getBasePtr();
2297 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2300 FI = FINode->getIndex();
2304 assert(FI != INT_MAX);
2305 if (!MFI->isFixedObjectIndex(FI))
2307 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2310 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2311 /// for tail call optimization. Targets which want to do tail call
2312 /// optimization should implement this function.
2314 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2315 CallingConv::ID CalleeCC,
2317 bool isCalleeStructRet,
2318 bool isCallerStructRet,
2319 const SmallVectorImpl<ISD::OutputArg> &Outs,
2320 const SmallVectorImpl<ISD::InputArg> &Ins,
2321 SelectionDAG& DAG) const {
2322 if (!IsTailCallConvention(CalleeCC) &&
2323 CalleeCC != CallingConv::C)
2326 // If -tailcallopt is specified, make fastcc functions tail-callable.
2327 const MachineFunction &MF = DAG.getMachineFunction();
2328 const Function *CallerF = DAG.getMachineFunction().getFunction();
2329 CallingConv::ID CallerCC = CallerF->getCallingConv();
2330 bool CCMatch = CallerCC == CalleeCC;
2332 if (GuaranteedTailCallOpt) {
2333 if (IsTailCallConvention(CalleeCC) && CCMatch)
2338 // Look for obvious safe cases to perform tail call optimization that do not
2339 // require ABI changes. This is what gcc calls sibcall.
2341 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2342 // emit a special epilogue.
2343 if (RegInfo->needsStackRealignment(MF))
2346 // Do not sibcall optimize vararg calls unless the call site is not passing any
2348 if (isVarArg && !Outs.empty())
2351 // Also avoid sibcall optimization if either caller or callee uses struct
2352 // return semantics.
2353 if (isCalleeStructRet || isCallerStructRet)
2356 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2357 // Therefore if it's not used by the call it is not safe to optimize this into
2359 bool Unused = false;
2360 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2367 SmallVector<CCValAssign, 16> RVLocs;
2368 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2369 RVLocs, *DAG.getContext());
2370 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2371 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2372 CCValAssign &VA = RVLocs[i];
2373 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2378 // If the calling conventions do not match, then we'd better make sure the
2379 // results are returned in the same way as what the caller expects.
2381 SmallVector<CCValAssign, 16> RVLocs1;
2382 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2383 RVLocs1, *DAG.getContext());
2384 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2386 SmallVector<CCValAssign, 16> RVLocs2;
2387 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2388 RVLocs2, *DAG.getContext());
2389 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2391 if (RVLocs1.size() != RVLocs2.size())
2393 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2394 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2396 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2398 if (RVLocs1[i].isRegLoc()) {
2399 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2402 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2408 // If the callee takes no arguments then go on to check the results of the
2410 if (!Outs.empty()) {
2411 // Check if stack adjustment is needed. For now, do not do this if any
2412 // argument is passed on the stack.
2413 SmallVector<CCValAssign, 16> ArgLocs;
2414 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2415 ArgLocs, *DAG.getContext());
2416 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2417 if (CCInfo.getNextStackOffset()) {
2418 MachineFunction &MF = DAG.getMachineFunction();
2419 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2421 if (Subtarget->isTargetWin64())
2422 // Win64 ABI has additional complications.
2425 // Check if the arguments are already laid out in the right way as
2426 // the caller's fixed stack objects.
2427 MachineFrameInfo *MFI = MF.getFrameInfo();
2428 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2429 const X86InstrInfo *TII =
2430 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2431 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2432 CCValAssign &VA = ArgLocs[i];
2433 SDValue Arg = Outs[i].Val;
2434 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2435 if (VA.getLocInfo() == CCValAssign::Indirect)
2437 if (!VA.isRegLoc()) {
2438 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2445 // If the tailcall address may be in a register, then make sure it's
2446 // possible to register allocate for it. In 32-bit, the call address can
2447 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2448 // callee-saved registers are restored. In 64-bit, it's RAX, RCX, RDX, RSI,
2449 // RDI, R8, R9, R11.
2450 if (!isa<GlobalAddressSDNode>(Callee) &&
2451 !isa<ExternalSymbolSDNode>(Callee)) {
2452 unsigned Limit = Subtarget->is64Bit() ? 8 : 3;
2453 unsigned NumInRegs = 0;
2454 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2455 CCValAssign &VA = ArgLocs[i];
2456 if (VA.isRegLoc()) {
2457 if (++NumInRegs == Limit)
2468 X86TargetLowering::createFastISel(MachineFunction &mf,
2469 DenseMap<const Value *, unsigned> &vm,
2470 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2471 DenseMap<const AllocaInst *, int> &am,
2472 std::vector<std::pair<MachineInstr*, unsigned> > &pn
2474 , SmallSet<const Instruction *, 8> &cil
2477 return X86::createFastISel(mf, vm, bm, am, pn
2485 //===----------------------------------------------------------------------===//
2486 // Other Lowering Hooks
2487 //===----------------------------------------------------------------------===//
2490 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2491 MachineFunction &MF = DAG.getMachineFunction();
2492 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2493 int ReturnAddrIndex = FuncInfo->getRAIndex();
2495 if (ReturnAddrIndex == 0) {
2496 // Set up a frame object for the return address.
2497 uint64_t SlotSize = TD->getPointerSize();
2498 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2500 FuncInfo->setRAIndex(ReturnAddrIndex);
2503 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2507 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2508 bool hasSymbolicDisplacement) {
2509 // Offset should fit into 32 bit immediate field.
2510 if (!isInt<32>(Offset))
2513 // If we don't have a symbolic displacement - we don't have any extra
2515 if (!hasSymbolicDisplacement)
2518 // FIXME: Some tweaks might be needed for medium code model.
2519 if (M != CodeModel::Small && M != CodeModel::Kernel)
2522 // For small code model we assume that latest object is 16MB before end of 31
2523 // bits boundary. We may also accept pretty large negative constants knowing
2524 // that all objects are in the positive half of address space.
2525 if (M == CodeModel::Small && Offset < 16*1024*1024)
2528 // For kernel code model we know that all object resist in the negative half
2529 // of 32bits address space. We may not accept negative offsets, since they may
2530 // be just off and we may accept pretty large positive ones.
2531 if (M == CodeModel::Kernel && Offset > 0)
2537 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2538 /// specific condition code, returning the condition code and the LHS/RHS of the
2539 /// comparison to make.
2540 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2541 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2543 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2544 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2545 // X > -1 -> X == 0, jump !sign.
2546 RHS = DAG.getConstant(0, RHS.getValueType());
2547 return X86::COND_NS;
2548 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2549 // X < 0 -> X == 0, jump on sign.
2551 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2553 RHS = DAG.getConstant(0, RHS.getValueType());
2554 return X86::COND_LE;
2558 switch (SetCCOpcode) {
2559 default: llvm_unreachable("Invalid integer condition!");
2560 case ISD::SETEQ: return X86::COND_E;
2561 case ISD::SETGT: return X86::COND_G;
2562 case ISD::SETGE: return X86::COND_GE;
2563 case ISD::SETLT: return X86::COND_L;
2564 case ISD::SETLE: return X86::COND_LE;
2565 case ISD::SETNE: return X86::COND_NE;
2566 case ISD::SETULT: return X86::COND_B;
2567 case ISD::SETUGT: return X86::COND_A;
2568 case ISD::SETULE: return X86::COND_BE;
2569 case ISD::SETUGE: return X86::COND_AE;
2573 // First determine if it is required or is profitable to flip the operands.
2575 // If LHS is a foldable load, but RHS is not, flip the condition.
2576 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2577 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2578 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2579 std::swap(LHS, RHS);
2582 switch (SetCCOpcode) {
2588 std::swap(LHS, RHS);
2592 // On a floating point condition, the flags are set as follows:
2594 // 0 | 0 | 0 | X > Y
2595 // 0 | 0 | 1 | X < Y
2596 // 1 | 0 | 0 | X == Y
2597 // 1 | 1 | 1 | unordered
2598 switch (SetCCOpcode) {
2599 default: llvm_unreachable("Condcode should be pre-legalized away");
2601 case ISD::SETEQ: return X86::COND_E;
2602 case ISD::SETOLT: // flipped
2604 case ISD::SETGT: return X86::COND_A;
2605 case ISD::SETOLE: // flipped
2607 case ISD::SETGE: return X86::COND_AE;
2608 case ISD::SETUGT: // flipped
2610 case ISD::SETLT: return X86::COND_B;
2611 case ISD::SETUGE: // flipped
2613 case ISD::SETLE: return X86::COND_BE;
2615 case ISD::SETNE: return X86::COND_NE;
2616 case ISD::SETUO: return X86::COND_P;
2617 case ISD::SETO: return X86::COND_NP;
2619 case ISD::SETUNE: return X86::COND_INVALID;
2623 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2624 /// code. Current x86 isa includes the following FP cmov instructions:
2625 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2626 static bool hasFPCMov(unsigned X86CC) {
2642 /// isFPImmLegal - Returns true if the target can instruction select the
2643 /// specified FP immediate natively. If false, the legalizer will
2644 /// materialize the FP immediate as a load from a constant pool.
2645 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2646 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2647 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2653 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2654 /// the specified range (L, H].
2655 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2656 return (Val < 0) || (Val >= Low && Val < Hi);
2659 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2660 /// specified value.
2661 static bool isUndefOrEqual(int Val, int CmpVal) {
2662 if (Val < 0 || Val == CmpVal)
2667 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2668 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2669 /// the second operand.
2670 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2671 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2672 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2673 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2674 return (Mask[0] < 2 && Mask[1] < 2);
2678 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2679 SmallVector<int, 8> M;
2681 return ::isPSHUFDMask(M, N->getValueType(0));
2684 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2685 /// is suitable for input to PSHUFHW.
2686 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2687 if (VT != MVT::v8i16)
2690 // Lower quadword copied in order or undef.
2691 for (int i = 0; i != 4; ++i)
2692 if (Mask[i] >= 0 && Mask[i] != i)
2695 // Upper quadword shuffled.
2696 for (int i = 4; i != 8; ++i)
2697 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2703 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2704 SmallVector<int, 8> M;
2706 return ::isPSHUFHWMask(M, N->getValueType(0));
2709 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2710 /// is suitable for input to PSHUFLW.
2711 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2712 if (VT != MVT::v8i16)
2715 // Upper quadword copied in order.
2716 for (int i = 4; i != 8; ++i)
2717 if (Mask[i] >= 0 && Mask[i] != i)
2720 // Lower quadword shuffled.
2721 for (int i = 0; i != 4; ++i)
2728 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2729 SmallVector<int, 8> M;
2731 return ::isPSHUFLWMask(M, N->getValueType(0));
2734 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2735 /// is suitable for input to PALIGNR.
2736 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2738 int i, e = VT.getVectorNumElements();
2740 // Do not handle v2i64 / v2f64 shuffles with palignr.
2741 if (e < 4 || !hasSSSE3)
2744 for (i = 0; i != e; ++i)
2748 // All undef, not a palignr.
2752 // Determine if it's ok to perform a palignr with only the LHS, since we
2753 // don't have access to the actual shuffle elements to see if RHS is undef.
2754 bool Unary = Mask[i] < (int)e;
2755 bool NeedsUnary = false;
2757 int s = Mask[i] - i;
2759 // Check the rest of the elements to see if they are consecutive.
2760 for (++i; i != e; ++i) {
2765 Unary = Unary && (m < (int)e);
2766 NeedsUnary = NeedsUnary || (m < s);
2768 if (NeedsUnary && !Unary)
2770 if (Unary && m != ((s+i) & (e-1)))
2772 if (!Unary && m != (s+i))
2778 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2779 SmallVector<int, 8> M;
2781 return ::isPALIGNRMask(M, N->getValueType(0), true);
2784 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2785 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2786 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2787 int NumElems = VT.getVectorNumElements();
2788 if (NumElems != 2 && NumElems != 4)
2791 int Half = NumElems / 2;
2792 for (int i = 0; i < Half; ++i)
2793 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2795 for (int i = Half; i < NumElems; ++i)
2796 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2802 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2803 SmallVector<int, 8> M;
2805 return ::isSHUFPMask(M, N->getValueType(0));
2808 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2809 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2810 /// half elements to come from vector 1 (which would equal the dest.) and
2811 /// the upper half to come from vector 2.
2812 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2813 int NumElems = VT.getVectorNumElements();
2815 if (NumElems != 2 && NumElems != 4)
2818 int Half = NumElems / 2;
2819 for (int i = 0; i < Half; ++i)
2820 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2822 for (int i = Half; i < NumElems; ++i)
2823 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2828 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2829 SmallVector<int, 8> M;
2831 return isCommutedSHUFPMask(M, N->getValueType(0));
2834 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2835 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2836 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2837 if (N->getValueType(0).getVectorNumElements() != 4)
2840 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2841 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2842 isUndefOrEqual(N->getMaskElt(1), 7) &&
2843 isUndefOrEqual(N->getMaskElt(2), 2) &&
2844 isUndefOrEqual(N->getMaskElt(3), 3);
2847 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2848 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2850 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2851 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2856 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2857 isUndefOrEqual(N->getMaskElt(1), 3) &&
2858 isUndefOrEqual(N->getMaskElt(2), 2) &&
2859 isUndefOrEqual(N->getMaskElt(3), 3);
2862 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2863 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2864 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2865 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2867 if (NumElems != 2 && NumElems != 4)
2870 for (unsigned i = 0; i < NumElems/2; ++i)
2871 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2874 for (unsigned i = NumElems/2; i < NumElems; ++i)
2875 if (!isUndefOrEqual(N->getMaskElt(i), i))
2881 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2882 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2883 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2884 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2886 if (NumElems != 2 && NumElems != 4)
2889 for (unsigned i = 0; i < NumElems/2; ++i)
2890 if (!isUndefOrEqual(N->getMaskElt(i), i))
2893 for (unsigned i = 0; i < NumElems/2; ++i)
2894 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2900 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2901 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2902 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2903 bool V2IsSplat = false) {
2904 int NumElts = VT.getVectorNumElements();
2905 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2908 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2910 int BitI1 = Mask[i+1];
2911 if (!isUndefOrEqual(BitI, j))
2914 if (!isUndefOrEqual(BitI1, NumElts))
2917 if (!isUndefOrEqual(BitI1, j + NumElts))
2924 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2925 SmallVector<int, 8> M;
2927 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2930 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2931 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2932 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2933 bool V2IsSplat = false) {
2934 int NumElts = VT.getVectorNumElements();
2935 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2938 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2940 int BitI1 = Mask[i+1];
2941 if (!isUndefOrEqual(BitI, j + NumElts/2))
2944 if (isUndefOrEqual(BitI1, NumElts))
2947 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2954 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2955 SmallVector<int, 8> M;
2957 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2960 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2961 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2963 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2964 int NumElems = VT.getVectorNumElements();
2965 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2968 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2970 int BitI1 = Mask[i+1];
2971 if (!isUndefOrEqual(BitI, j))
2973 if (!isUndefOrEqual(BitI1, j))
2979 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2980 SmallVector<int, 8> M;
2982 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2985 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2986 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2988 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2989 int NumElems = VT.getVectorNumElements();
2990 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2993 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2995 int BitI1 = Mask[i+1];
2996 if (!isUndefOrEqual(BitI, j))
2998 if (!isUndefOrEqual(BitI1, j))
3004 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3005 SmallVector<int, 8> M;
3007 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3010 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3011 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3012 /// MOVSD, and MOVD, i.e. setting the lowest element.
3013 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3014 if (VT.getVectorElementType().getSizeInBits() < 32)
3017 int NumElts = VT.getVectorNumElements();
3019 if (!isUndefOrEqual(Mask[0], NumElts))
3022 for (int i = 1; i < NumElts; ++i)
3023 if (!isUndefOrEqual(Mask[i], i))
3029 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3030 SmallVector<int, 8> M;
3032 return ::isMOVLMask(M, N->getValueType(0));
3035 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3036 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3037 /// element of vector 2 and the other elements to come from vector 1 in order.
3038 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3039 bool V2IsSplat = false, bool V2IsUndef = false) {
3040 int NumOps = VT.getVectorNumElements();
3041 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3044 if (!isUndefOrEqual(Mask[0], 0))
3047 for (int i = 1; i < NumOps; ++i)
3048 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3049 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3050 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3056 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3057 bool V2IsUndef = false) {
3058 SmallVector<int, 8> M;
3060 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3063 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3064 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3065 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3066 if (N->getValueType(0).getVectorNumElements() != 4)
3069 // Expect 1, 1, 3, 3
3070 for (unsigned i = 0; i < 2; ++i) {
3071 int Elt = N->getMaskElt(i);
3072 if (Elt >= 0 && Elt != 1)
3077 for (unsigned i = 2; i < 4; ++i) {
3078 int Elt = N->getMaskElt(i);
3079 if (Elt >= 0 && Elt != 3)
3084 // Don't use movshdup if it can be done with a shufps.
3085 // FIXME: verify that matching u, u, 3, 3 is what we want.
3089 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3090 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3091 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3092 if (N->getValueType(0).getVectorNumElements() != 4)
3095 // Expect 0, 0, 2, 2
3096 for (unsigned i = 0; i < 2; ++i)
3097 if (N->getMaskElt(i) > 0)
3101 for (unsigned i = 2; i < 4; ++i) {
3102 int Elt = N->getMaskElt(i);
3103 if (Elt >= 0 && Elt != 2)
3108 // Don't use movsldup if it can be done with a shufps.
3112 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3113 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3114 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3115 int e = N->getValueType(0).getVectorNumElements() / 2;
3117 for (int i = 0; i < e; ++i)
3118 if (!isUndefOrEqual(N->getMaskElt(i), i))
3120 for (int i = 0; i < e; ++i)
3121 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3126 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3127 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3128 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3129 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3130 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3132 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3134 for (int i = 0; i < NumOperands; ++i) {
3135 int Val = SVOp->getMaskElt(NumOperands-i-1);
3136 if (Val < 0) Val = 0;
3137 if (Val >= NumOperands) Val -= NumOperands;
3139 if (i != NumOperands - 1)
3145 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3146 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3147 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3148 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3150 // 8 nodes, but we only care about the last 4.
3151 for (unsigned i = 7; i >= 4; --i) {
3152 int Val = SVOp->getMaskElt(i);
3161 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3162 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3163 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3164 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3166 // 8 nodes, but we only care about the first 4.
3167 for (int i = 3; i >= 0; --i) {
3168 int Val = SVOp->getMaskElt(i);
3177 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3178 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3179 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3180 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3181 EVT VVT = N->getValueType(0);
3182 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3186 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3187 Val = SVOp->getMaskElt(i);
3191 return (Val - i) * EltSize;
3194 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3196 bool X86::isZeroNode(SDValue Elt) {
3197 return ((isa<ConstantSDNode>(Elt) &&
3198 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3199 (isa<ConstantFPSDNode>(Elt) &&
3200 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3203 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3204 /// their permute mask.
3205 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3206 SelectionDAG &DAG) {
3207 EVT VT = SVOp->getValueType(0);
3208 unsigned NumElems = VT.getVectorNumElements();
3209 SmallVector<int, 8> MaskVec;
3211 for (unsigned i = 0; i != NumElems; ++i) {
3212 int idx = SVOp->getMaskElt(i);
3214 MaskVec.push_back(idx);
3215 else if (idx < (int)NumElems)
3216 MaskVec.push_back(idx + NumElems);
3218 MaskVec.push_back(idx - NumElems);
3220 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3221 SVOp->getOperand(0), &MaskVec[0]);
3224 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3225 /// the two vector operands have swapped position.
3226 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3227 unsigned NumElems = VT.getVectorNumElements();
3228 for (unsigned i = 0; i != NumElems; ++i) {
3232 else if (idx < (int)NumElems)
3233 Mask[i] = idx + NumElems;
3235 Mask[i] = idx - NumElems;
3239 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3240 /// match movhlps. The lower half elements should come from upper half of
3241 /// V1 (and in order), and the upper half elements should come from the upper
3242 /// half of V2 (and in order).
3243 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3244 if (Op->getValueType(0).getVectorNumElements() != 4)
3246 for (unsigned i = 0, e = 2; i != e; ++i)
3247 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3249 for (unsigned i = 2; i != 4; ++i)
3250 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3255 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3256 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3258 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3259 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3261 N = N->getOperand(0).getNode();
3262 if (!ISD::isNON_EXTLoad(N))
3265 *LD = cast<LoadSDNode>(N);
3269 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3270 /// match movlp{s|d}. The lower half elements should come from lower half of
3271 /// V1 (and in order), and the upper half elements should come from the upper
3272 /// half of V2 (and in order). And since V1 will become the source of the
3273 /// MOVLP, it must be either a vector load or a scalar load to vector.
3274 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3275 ShuffleVectorSDNode *Op) {
3276 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3278 // Is V2 is a vector load, don't do this transformation. We will try to use
3279 // load folding shufps op.
3280 if (ISD::isNON_EXTLoad(V2))
3283 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3285 if (NumElems != 2 && NumElems != 4)
3287 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3288 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3290 for (unsigned i = NumElems/2; i != NumElems; ++i)
3291 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3296 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3298 static bool isSplatVector(SDNode *N) {
3299 if (N->getOpcode() != ISD::BUILD_VECTOR)
3302 SDValue SplatValue = N->getOperand(0);
3303 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3304 if (N->getOperand(i) != SplatValue)
3309 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3310 /// to an zero vector.
3311 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3312 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3313 SDValue V1 = N->getOperand(0);
3314 SDValue V2 = N->getOperand(1);
3315 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3316 for (unsigned i = 0; i != NumElems; ++i) {
3317 int Idx = N->getMaskElt(i);
3318 if (Idx >= (int)NumElems) {
3319 unsigned Opc = V2.getOpcode();
3320 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3322 if (Opc != ISD::BUILD_VECTOR ||
3323 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3325 } else if (Idx >= 0) {
3326 unsigned Opc = V1.getOpcode();
3327 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3329 if (Opc != ISD::BUILD_VECTOR ||
3330 !X86::isZeroNode(V1.getOperand(Idx)))
3337 /// getZeroVector - Returns a vector of specified type with all zero elements.
3339 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3341 assert(VT.isVector() && "Expected a vector type");
3343 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3344 // type. This ensures they get CSE'd.
3346 if (VT.getSizeInBits() == 64) { // MMX
3347 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3348 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3349 } else if (HasSSE2) { // SSE2
3350 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3351 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3353 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3354 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3356 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3359 /// getOnesVector - Returns a vector of specified type with all bits set.
3361 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3362 assert(VT.isVector() && "Expected a vector type");
3364 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3365 // type. This ensures they get CSE'd.
3366 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3368 if (VT.getSizeInBits() == 64) // MMX
3369 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3371 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3372 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3376 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3377 /// that point to V2 points to its first element.
3378 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3379 EVT VT = SVOp->getValueType(0);
3380 unsigned NumElems = VT.getVectorNumElements();
3382 bool Changed = false;
3383 SmallVector<int, 8> MaskVec;
3384 SVOp->getMask(MaskVec);
3386 for (unsigned i = 0; i != NumElems; ++i) {
3387 if (MaskVec[i] > (int)NumElems) {
3388 MaskVec[i] = NumElems;
3393 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3394 SVOp->getOperand(1), &MaskVec[0]);
3395 return SDValue(SVOp, 0);
3398 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3399 /// operation of specified width.
3400 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3402 unsigned NumElems = VT.getVectorNumElements();
3403 SmallVector<int, 8> Mask;
3404 Mask.push_back(NumElems);
3405 for (unsigned i = 1; i != NumElems; ++i)
3407 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3410 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3411 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3413 unsigned NumElems = VT.getVectorNumElements();
3414 SmallVector<int, 8> Mask;
3415 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3417 Mask.push_back(i + NumElems);
3419 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3422 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3423 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3425 unsigned NumElems = VT.getVectorNumElements();
3426 unsigned Half = NumElems/2;
3427 SmallVector<int, 8> Mask;
3428 for (unsigned i = 0; i != Half; ++i) {
3429 Mask.push_back(i + Half);
3430 Mask.push_back(i + NumElems + Half);
3432 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3435 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3436 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3438 if (SV->getValueType(0).getVectorNumElements() <= 4)
3439 return SDValue(SV, 0);
3441 EVT PVT = MVT::v4f32;
3442 EVT VT = SV->getValueType(0);
3443 DebugLoc dl = SV->getDebugLoc();
3444 SDValue V1 = SV->getOperand(0);
3445 int NumElems = VT.getVectorNumElements();
3446 int EltNo = SV->getSplatIndex();
3448 // unpack elements to the correct location
3449 while (NumElems > 4) {
3450 if (EltNo < NumElems/2) {
3451 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3453 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3454 EltNo -= NumElems/2;
3459 // Perform the splat.
3460 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3461 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3462 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3463 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3466 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3467 /// vector of zero or undef vector. This produces a shuffle where the low
3468 /// element of V2 is swizzled into the zero/undef vector, landing at element
3469 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3470 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3471 bool isZero, bool HasSSE2,
3472 SelectionDAG &DAG) {
3473 EVT VT = V2.getValueType();
3475 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3476 unsigned NumElems = VT.getVectorNumElements();
3477 SmallVector<int, 16> MaskVec;
3478 for (unsigned i = 0; i != NumElems; ++i)
3479 // If this is the insertion idx, put the low elt of V2 here.
3480 MaskVec.push_back(i == Idx ? NumElems : i);
3481 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3484 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3485 /// a shuffle that is zero.
3487 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3488 bool Low, SelectionDAG &DAG) {
3489 unsigned NumZeros = 0;
3490 for (int i = 0; i < NumElems; ++i) {
3491 unsigned Index = Low ? i : NumElems-i-1;
3492 int Idx = SVOp->getMaskElt(Index);
3497 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3498 if (Elt.getNode() && X86::isZeroNode(Elt))
3506 /// isVectorShift - Returns true if the shuffle can be implemented as a
3507 /// logical left or right shift of a vector.
3508 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3509 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3510 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3511 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3514 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3517 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3521 bool SeenV1 = false;
3522 bool SeenV2 = false;
3523 for (unsigned i = NumZeros; i < NumElems; ++i) {
3524 unsigned Val = isLeft ? (i - NumZeros) : i;
3525 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3528 unsigned Idx = (unsigned) Idx_;
3538 if (SeenV1 && SeenV2)
3541 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3547 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3549 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3550 unsigned NumNonZero, unsigned NumZero,
3552 const TargetLowering &TLI) {
3556 DebugLoc dl = Op.getDebugLoc();
3559 for (unsigned i = 0; i < 16; ++i) {
3560 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3561 if (ThisIsNonZero && First) {
3563 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3565 V = DAG.getUNDEF(MVT::v8i16);
3570 SDValue ThisElt(0, 0), LastElt(0, 0);
3571 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3572 if (LastIsNonZero) {
3573 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3574 MVT::i16, Op.getOperand(i-1));
3576 if (ThisIsNonZero) {
3577 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3578 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3579 ThisElt, DAG.getConstant(8, MVT::i8));
3581 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3585 if (ThisElt.getNode())
3586 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3587 DAG.getIntPtrConstant(i/2));
3591 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3594 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3596 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3597 unsigned NumNonZero, unsigned NumZero,
3599 const TargetLowering &TLI) {
3603 DebugLoc dl = Op.getDebugLoc();
3606 for (unsigned i = 0; i < 8; ++i) {
3607 bool isNonZero = (NonZeros & (1 << i)) != 0;
3611 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3613 V = DAG.getUNDEF(MVT::v8i16);
3616 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3617 MVT::v8i16, V, Op.getOperand(i),
3618 DAG.getIntPtrConstant(i));
3625 /// getVShift - Return a vector logical shift node.
3627 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3628 unsigned NumBits, SelectionDAG &DAG,
3629 const TargetLowering &TLI, DebugLoc dl) {
3630 bool isMMX = VT.getSizeInBits() == 64;
3631 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3632 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3633 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3634 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3635 DAG.getNode(Opc, dl, ShVT, SrcOp,
3636 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3640 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3641 SelectionDAG &DAG) const {
3643 // Check if the scalar load can be widened into a vector load. And if
3644 // the address is "base + cst" see if the cst can be "absorbed" into
3645 // the shuffle mask.
3646 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3647 SDValue Ptr = LD->getBasePtr();
3648 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3650 EVT PVT = LD->getValueType(0);
3651 if (PVT != MVT::i32 && PVT != MVT::f32)
3656 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3657 FI = FINode->getIndex();
3659 } else if (Ptr.getOpcode() == ISD::ADD &&
3660 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3661 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3662 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3663 Offset = Ptr.getConstantOperandVal(1);
3664 Ptr = Ptr.getOperand(0);
3669 SDValue Chain = LD->getChain();
3670 // Make sure the stack object alignment is at least 16.
3671 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3672 if (DAG.InferPtrAlignment(Ptr) < 16) {
3673 if (MFI->isFixedObjectIndex(FI)) {
3674 // Can't change the alignment. FIXME: It's possible to compute
3675 // the exact stack offset and reference FI + adjust offset instead.
3676 // If someone *really* cares about this. That's the way to implement it.
3679 MFI->setObjectAlignment(FI, 16);
3683 // (Offset % 16) must be multiple of 4. Then address is then
3684 // Ptr + (Offset & ~15).
3687 if ((Offset % 16) & 3)
3689 int64_t StartOffset = Offset & ~15;
3691 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3692 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3694 int EltNo = (Offset - StartOffset) >> 2;
3695 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3696 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3697 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3699 // Canonicalize it to a v4i32 shuffle.
3700 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3701 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3702 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3703 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3709 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3710 /// vector of type 'VT', see if the elements can be replaced by a single large
3711 /// load which has the same value as a build_vector whose operands are 'elts'.
3713 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3715 /// FIXME: we'd also like to handle the case where the last elements are zero
3716 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3717 /// There's even a handy isZeroNode for that purpose.
3718 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3719 DebugLoc &dl, SelectionDAG &DAG) {
3720 EVT EltVT = VT.getVectorElementType();
3721 unsigned NumElems = Elts.size();
3723 LoadSDNode *LDBase = NULL;
3724 unsigned LastLoadedElt = -1U;
3726 // For each element in the initializer, see if we've found a load or an undef.
3727 // If we don't find an initial load element, or later load elements are
3728 // non-consecutive, bail out.
3729 for (unsigned i = 0; i < NumElems; ++i) {
3730 SDValue Elt = Elts[i];
3732 if (!Elt.getNode() ||
3733 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3736 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3738 LDBase = cast<LoadSDNode>(Elt.getNode());
3742 if (Elt.getOpcode() == ISD::UNDEF)
3745 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3746 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3751 // If we have found an entire vector of loads and undefs, then return a large
3752 // load of the entire vector width starting at the base pointer. If we found
3753 // consecutive loads for the low half, generate a vzext_load node.
3754 if (LastLoadedElt == NumElems - 1) {
3755 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3756 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3757 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3758 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3759 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3760 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3761 LDBase->isVolatile(), LDBase->isNonTemporal(),
3762 LDBase->getAlignment());
3763 } else if (NumElems == 4 && LastLoadedElt == 1) {
3764 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3765 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3766 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3767 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3773 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3774 DebugLoc dl = Op.getDebugLoc();
3775 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3776 if (ISD::isBuildVectorAllZeros(Op.getNode())
3777 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3778 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3779 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3780 // eliminated on x86-32 hosts.
3781 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3784 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3785 return getOnesVector(Op.getValueType(), DAG, dl);
3786 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3789 EVT VT = Op.getValueType();
3790 EVT ExtVT = VT.getVectorElementType();
3791 unsigned EVTBits = ExtVT.getSizeInBits();
3793 unsigned NumElems = Op.getNumOperands();
3794 unsigned NumZero = 0;
3795 unsigned NumNonZero = 0;
3796 unsigned NonZeros = 0;
3797 bool IsAllConstants = true;
3798 SmallSet<SDValue, 8> Values;
3799 for (unsigned i = 0; i < NumElems; ++i) {
3800 SDValue Elt = Op.getOperand(i);
3801 if (Elt.getOpcode() == ISD::UNDEF)
3804 if (Elt.getOpcode() != ISD::Constant &&
3805 Elt.getOpcode() != ISD::ConstantFP)
3806 IsAllConstants = false;
3807 if (X86::isZeroNode(Elt))
3810 NonZeros |= (1 << i);
3815 if (NumNonZero == 0) {
3816 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3817 return DAG.getUNDEF(VT);
3820 // Special case for single non-zero, non-undef, element.
3821 if (NumNonZero == 1) {
3822 unsigned Idx = CountTrailingZeros_32(NonZeros);
3823 SDValue Item = Op.getOperand(Idx);
3825 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3826 // the value are obviously zero, truncate the value to i32 and do the
3827 // insertion that way. Only do this if the value is non-constant or if the
3828 // value is a constant being inserted into element 0. It is cheaper to do
3829 // a constant pool load than it is to do a movd + shuffle.
3830 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3831 (!IsAllConstants || Idx == 0)) {
3832 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3833 // Handle MMX and SSE both.
3834 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3835 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3837 // Truncate the value (which may itself be a constant) to i32, and
3838 // convert it to a vector with movd (S2V+shuffle to zero extend).
3839 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3840 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3841 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3842 Subtarget->hasSSE2(), DAG);
3844 // Now we have our 32-bit value zero extended in the low element of
3845 // a vector. If Idx != 0, swizzle it into place.
3847 SmallVector<int, 4> Mask;
3848 Mask.push_back(Idx);
3849 for (unsigned i = 1; i != VecElts; ++i)
3851 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3852 DAG.getUNDEF(Item.getValueType()),
3855 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3859 // If we have a constant or non-constant insertion into the low element of
3860 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3861 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3862 // depending on what the source datatype is.
3865 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3866 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3867 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3868 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3869 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3870 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3872 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3873 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3874 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3875 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3876 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3877 Subtarget->hasSSE2(), DAG);
3878 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3882 // Is it a vector logical left shift?
3883 if (NumElems == 2 && Idx == 1 &&
3884 X86::isZeroNode(Op.getOperand(0)) &&
3885 !X86::isZeroNode(Op.getOperand(1))) {
3886 unsigned NumBits = VT.getSizeInBits();
3887 return getVShift(true, VT,
3888 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3889 VT, Op.getOperand(1)),
3890 NumBits/2, DAG, *this, dl);
3893 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3896 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3897 // is a non-constant being inserted into an element other than the low one,
3898 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3899 // movd/movss) to move this into the low element, then shuffle it into
3901 if (EVTBits == 32) {
3902 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3904 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3905 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3906 Subtarget->hasSSE2(), DAG);
3907 SmallVector<int, 8> MaskVec;
3908 for (unsigned i = 0; i < NumElems; i++)
3909 MaskVec.push_back(i == Idx ? 0 : 1);
3910 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3914 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3915 if (Values.size() == 1) {
3916 if (EVTBits == 32) {
3917 // Instead of a shuffle like this:
3918 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3919 // Check if it's possible to issue this instead.
3920 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3921 unsigned Idx = CountTrailingZeros_32(NonZeros);
3922 SDValue Item = Op.getOperand(Idx);
3923 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3924 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3929 // A vector full of immediates; various special cases are already
3930 // handled, so this is best done with a single constant-pool load.
3934 // Let legalizer expand 2-wide build_vectors.
3935 if (EVTBits == 64) {
3936 if (NumNonZero == 1) {
3937 // One half is zero or undef.
3938 unsigned Idx = CountTrailingZeros_32(NonZeros);
3939 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3940 Op.getOperand(Idx));
3941 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3942 Subtarget->hasSSE2(), DAG);
3947 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3948 if (EVTBits == 8 && NumElems == 16) {
3949 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3951 if (V.getNode()) return V;
3954 if (EVTBits == 16 && NumElems == 8) {
3955 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3957 if (V.getNode()) return V;
3960 // If element VT is == 32 bits, turn it into a number of shuffles.
3961 SmallVector<SDValue, 8> V;
3963 if (NumElems == 4 && NumZero > 0) {
3964 for (unsigned i = 0; i < 4; ++i) {
3965 bool isZero = !(NonZeros & (1 << i));
3967 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3969 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3972 for (unsigned i = 0; i < 2; ++i) {
3973 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3976 V[i] = V[i*2]; // Must be a zero vector.
3979 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3982 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3985 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3990 SmallVector<int, 8> MaskVec;
3991 bool Reverse = (NonZeros & 0x3) == 2;
3992 for (unsigned i = 0; i < 2; ++i)
3993 MaskVec.push_back(Reverse ? 1-i : i);
3994 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3995 for (unsigned i = 0; i < 2; ++i)
3996 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3997 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4000 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4001 // Check for a build vector of consecutive loads.
4002 for (unsigned i = 0; i < NumElems; ++i)
4003 V[i] = Op.getOperand(i);
4005 // Check for elements which are consecutive loads.
4006 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4010 // For SSE 4.1, use inserts into undef.
4011 if (getSubtarget()->hasSSE41()) {
4012 V[0] = DAG.getUNDEF(VT);
4013 for (unsigned i = 0; i < NumElems; ++i)
4014 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4015 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
4016 Op.getOperand(i), DAG.getIntPtrConstant(i));
4020 // Otherwise, expand into a number of unpckl*
4022 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4023 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4024 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4025 for (unsigned i = 0; i < NumElems; ++i)
4026 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4028 while (NumElems != 0) {
4029 for (unsigned i = 0; i < NumElems; ++i)
4030 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4039 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4040 // We support concatenate two MMX registers and place them in a MMX
4041 // register. This is better than doing a stack convert.
4042 DebugLoc dl = Op.getDebugLoc();
4043 EVT ResVT = Op.getValueType();
4044 assert(Op.getNumOperands() == 2);
4045 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4046 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4048 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4049 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4050 InVec = Op.getOperand(1);
4051 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4052 unsigned NumElts = ResVT.getVectorNumElements();
4053 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4054 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4055 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4057 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4058 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4059 Mask[0] = 0; Mask[1] = 2;
4060 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4062 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4065 // v8i16 shuffles - Prefer shuffles in the following order:
4066 // 1. [all] pshuflw, pshufhw, optional move
4067 // 2. [ssse3] 1 x pshufb
4068 // 3. [ssse3] 2 x pshufb + 1 x por
4069 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4071 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4073 const X86TargetLowering &TLI) {
4074 SDValue V1 = SVOp->getOperand(0);
4075 SDValue V2 = SVOp->getOperand(1);
4076 DebugLoc dl = SVOp->getDebugLoc();
4077 SmallVector<int, 8> MaskVals;
4079 // Determine if more than 1 of the words in each of the low and high quadwords
4080 // of the result come from the same quadword of one of the two inputs. Undef
4081 // mask values count as coming from any quadword, for better codegen.
4082 SmallVector<unsigned, 4> LoQuad(4);
4083 SmallVector<unsigned, 4> HiQuad(4);
4084 BitVector InputQuads(4);
4085 for (unsigned i = 0; i < 8; ++i) {
4086 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4087 int EltIdx = SVOp->getMaskElt(i);
4088 MaskVals.push_back(EltIdx);
4097 InputQuads.set(EltIdx / 4);
4100 int BestLoQuad = -1;
4101 unsigned MaxQuad = 1;
4102 for (unsigned i = 0; i < 4; ++i) {
4103 if (LoQuad[i] > MaxQuad) {
4105 MaxQuad = LoQuad[i];
4109 int BestHiQuad = -1;
4111 for (unsigned i = 0; i < 4; ++i) {
4112 if (HiQuad[i] > MaxQuad) {
4114 MaxQuad = HiQuad[i];
4118 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4119 // of the two input vectors, shuffle them into one input vector so only a
4120 // single pshufb instruction is necessary. If There are more than 2 input
4121 // quads, disable the next transformation since it does not help SSSE3.
4122 bool V1Used = InputQuads[0] || InputQuads[1];
4123 bool V2Used = InputQuads[2] || InputQuads[3];
4124 if (TLI.getSubtarget()->hasSSSE3()) {
4125 if (InputQuads.count() == 2 && V1Used && V2Used) {
4126 BestLoQuad = InputQuads.find_first();
4127 BestHiQuad = InputQuads.find_next(BestLoQuad);
4129 if (InputQuads.count() > 2) {
4135 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4136 // the shuffle mask. If a quad is scored as -1, that means that it contains
4137 // words from all 4 input quadwords.
4139 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4140 SmallVector<int, 8> MaskV;
4141 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4142 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4143 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4144 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4145 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4146 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4148 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4149 // source words for the shuffle, to aid later transformations.
4150 bool AllWordsInNewV = true;
4151 bool InOrder[2] = { true, true };
4152 for (unsigned i = 0; i != 8; ++i) {
4153 int idx = MaskVals[i];
4155 InOrder[i/4] = false;
4156 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4158 AllWordsInNewV = false;
4162 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4163 if (AllWordsInNewV) {
4164 for (int i = 0; i != 8; ++i) {
4165 int idx = MaskVals[i];
4168 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4169 if ((idx != i) && idx < 4)
4171 if ((idx != i) && idx > 3)
4180 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4181 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4182 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4183 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4184 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4188 // If we have SSSE3, and all words of the result are from 1 input vector,
4189 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4190 // is present, fall back to case 4.
4191 if (TLI.getSubtarget()->hasSSSE3()) {
4192 SmallVector<SDValue,16> pshufbMask;
4194 // If we have elements from both input vectors, set the high bit of the
4195 // shuffle mask element to zero out elements that come from V2 in the V1
4196 // mask, and elements that come from V1 in the V2 mask, so that the two
4197 // results can be OR'd together.
4198 bool TwoInputs = V1Used && V2Used;
4199 for (unsigned i = 0; i != 8; ++i) {
4200 int EltIdx = MaskVals[i] * 2;
4201 if (TwoInputs && (EltIdx >= 16)) {
4202 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4203 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4206 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4207 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4209 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4210 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4211 DAG.getNode(ISD::BUILD_VECTOR, dl,
4212 MVT::v16i8, &pshufbMask[0], 16));
4214 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4216 // Calculate the shuffle mask for the second input, shuffle it, and
4217 // OR it with the first shuffled input.
4219 for (unsigned i = 0; i != 8; ++i) {
4220 int EltIdx = MaskVals[i] * 2;
4222 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4223 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4226 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4227 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4229 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4230 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4231 DAG.getNode(ISD::BUILD_VECTOR, dl,
4232 MVT::v16i8, &pshufbMask[0], 16));
4233 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4234 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4237 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4238 // and update MaskVals with new element order.
4239 BitVector InOrder(8);
4240 if (BestLoQuad >= 0) {
4241 SmallVector<int, 8> MaskV;
4242 for (int i = 0; i != 4; ++i) {
4243 int idx = MaskVals[i];
4245 MaskV.push_back(-1);
4247 } else if ((idx / 4) == BestLoQuad) {
4248 MaskV.push_back(idx & 3);
4251 MaskV.push_back(-1);
4254 for (unsigned i = 4; i != 8; ++i)
4256 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4260 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4261 // and update MaskVals with the new element order.
4262 if (BestHiQuad >= 0) {
4263 SmallVector<int, 8> MaskV;
4264 for (unsigned i = 0; i != 4; ++i)
4266 for (unsigned i = 4; i != 8; ++i) {
4267 int idx = MaskVals[i];
4269 MaskV.push_back(-1);
4271 } else if ((idx / 4) == BestHiQuad) {
4272 MaskV.push_back((idx & 3) + 4);
4275 MaskV.push_back(-1);
4278 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4282 // In case BestHi & BestLo were both -1, which means each quadword has a word
4283 // from each of the four input quadwords, calculate the InOrder bitvector now
4284 // before falling through to the insert/extract cleanup.
4285 if (BestLoQuad == -1 && BestHiQuad == -1) {
4287 for (int i = 0; i != 8; ++i)
4288 if (MaskVals[i] < 0 || MaskVals[i] == i)
4292 // The other elements are put in the right place using pextrw and pinsrw.
4293 for (unsigned i = 0; i != 8; ++i) {
4296 int EltIdx = MaskVals[i];
4299 SDValue ExtOp = (EltIdx < 8)
4300 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4301 DAG.getIntPtrConstant(EltIdx))
4302 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4303 DAG.getIntPtrConstant(EltIdx - 8));
4304 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4305 DAG.getIntPtrConstant(i));
4310 // v16i8 shuffles - Prefer shuffles in the following order:
4311 // 1. [ssse3] 1 x pshufb
4312 // 2. [ssse3] 2 x pshufb + 1 x por
4313 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4315 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4317 const X86TargetLowering &TLI) {
4318 SDValue V1 = SVOp->getOperand(0);
4319 SDValue V2 = SVOp->getOperand(1);
4320 DebugLoc dl = SVOp->getDebugLoc();
4321 SmallVector<int, 16> MaskVals;
4322 SVOp->getMask(MaskVals);
4324 // If we have SSSE3, case 1 is generated when all result bytes come from
4325 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4326 // present, fall back to case 3.
4327 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4330 for (unsigned i = 0; i < 16; ++i) {
4331 int EltIdx = MaskVals[i];
4340 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4341 if (TLI.getSubtarget()->hasSSSE3()) {
4342 SmallVector<SDValue,16> pshufbMask;
4344 // If all result elements are from one input vector, then only translate
4345 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4347 // Otherwise, we have elements from both input vectors, and must zero out
4348 // elements that come from V2 in the first mask, and V1 in the second mask
4349 // so that we can OR them together.
4350 bool TwoInputs = !(V1Only || V2Only);
4351 for (unsigned i = 0; i != 16; ++i) {
4352 int EltIdx = MaskVals[i];
4353 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4354 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4357 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4359 // If all the elements are from V2, assign it to V1 and return after
4360 // building the first pshufb.
4363 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4364 DAG.getNode(ISD::BUILD_VECTOR, dl,
4365 MVT::v16i8, &pshufbMask[0], 16));
4369 // Calculate the shuffle mask for the second input, shuffle it, and
4370 // OR it with the first shuffled input.
4372 for (unsigned i = 0; i != 16; ++i) {
4373 int EltIdx = MaskVals[i];
4375 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4378 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4380 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4381 DAG.getNode(ISD::BUILD_VECTOR, dl,
4382 MVT::v16i8, &pshufbMask[0], 16));
4383 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4386 // No SSSE3 - Calculate in place words and then fix all out of place words
4387 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4388 // the 16 different words that comprise the two doublequadword input vectors.
4389 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4390 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4391 SDValue NewV = V2Only ? V2 : V1;
4392 for (int i = 0; i != 8; ++i) {
4393 int Elt0 = MaskVals[i*2];
4394 int Elt1 = MaskVals[i*2+1];
4396 // This word of the result is all undef, skip it.
4397 if (Elt0 < 0 && Elt1 < 0)
4400 // This word of the result is already in the correct place, skip it.
4401 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4403 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4406 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4407 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4410 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4411 // using a single extract together, load it and store it.
4412 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4413 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4414 DAG.getIntPtrConstant(Elt1 / 2));
4415 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4416 DAG.getIntPtrConstant(i));
4420 // If Elt1 is defined, extract it from the appropriate source. If the
4421 // source byte is not also odd, shift the extracted word left 8 bits
4422 // otherwise clear the bottom 8 bits if we need to do an or.
4424 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4425 DAG.getIntPtrConstant(Elt1 / 2));
4426 if ((Elt1 & 1) == 0)
4427 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4428 DAG.getConstant(8, TLI.getShiftAmountTy()));
4430 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4431 DAG.getConstant(0xFF00, MVT::i16));
4433 // If Elt0 is defined, extract it from the appropriate source. If the
4434 // source byte is not also even, shift the extracted word right 8 bits. If
4435 // Elt1 was also defined, OR the extracted values together before
4436 // inserting them in the result.
4438 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4439 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4440 if ((Elt0 & 1) != 0)
4441 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4442 DAG.getConstant(8, TLI.getShiftAmountTy()));
4444 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4445 DAG.getConstant(0x00FF, MVT::i16));
4446 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4449 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4450 DAG.getIntPtrConstant(i));
4452 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4455 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4456 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4457 /// done when every pair / quad of shuffle mask elements point to elements in
4458 /// the right sequence. e.g.
4459 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4461 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4463 const TargetLowering &TLI, DebugLoc dl) {
4464 EVT VT = SVOp->getValueType(0);
4465 SDValue V1 = SVOp->getOperand(0);
4466 SDValue V2 = SVOp->getOperand(1);
4467 unsigned NumElems = VT.getVectorNumElements();
4468 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4469 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4471 switch (VT.getSimpleVT().SimpleTy) {
4472 default: assert(false && "Unexpected!");
4473 case MVT::v4f32: NewVT = MVT::v2f64; break;
4474 case MVT::v4i32: NewVT = MVT::v2i64; break;
4475 case MVT::v8i16: NewVT = MVT::v4i32; break;
4476 case MVT::v16i8: NewVT = MVT::v4i32; break;
4479 if (NewWidth == 2) {
4485 int Scale = NumElems / NewWidth;
4486 SmallVector<int, 8> MaskVec;
4487 for (unsigned i = 0; i < NumElems; i += Scale) {
4489 for (int j = 0; j < Scale; ++j) {
4490 int EltIdx = SVOp->getMaskElt(i+j);
4494 StartIdx = EltIdx - (EltIdx % Scale);
4495 if (EltIdx != StartIdx + j)
4499 MaskVec.push_back(-1);
4501 MaskVec.push_back(StartIdx / Scale);
4504 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4505 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4506 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4509 /// getVZextMovL - Return a zero-extending vector move low node.
4511 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4512 SDValue SrcOp, SelectionDAG &DAG,
4513 const X86Subtarget *Subtarget, DebugLoc dl) {
4514 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4515 LoadSDNode *LD = NULL;
4516 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4517 LD = dyn_cast<LoadSDNode>(SrcOp);
4519 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4521 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4522 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4523 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4524 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4525 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4527 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4528 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4529 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4530 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4538 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4539 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4540 DAG.getNode(ISD::BIT_CONVERT, dl,
4544 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4547 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4548 SDValue V1 = SVOp->getOperand(0);
4549 SDValue V2 = SVOp->getOperand(1);
4550 DebugLoc dl = SVOp->getDebugLoc();
4551 EVT VT = SVOp->getValueType(0);
4553 SmallVector<std::pair<int, int>, 8> Locs;
4555 SmallVector<int, 8> Mask1(4U, -1);
4556 SmallVector<int, 8> PermMask;
4557 SVOp->getMask(PermMask);
4561 for (unsigned i = 0; i != 4; ++i) {
4562 int Idx = PermMask[i];
4564 Locs[i] = std::make_pair(-1, -1);
4566 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4568 Locs[i] = std::make_pair(0, NumLo);
4572 Locs[i] = std::make_pair(1, NumHi);
4574 Mask1[2+NumHi] = Idx;
4580 if (NumLo <= 2 && NumHi <= 2) {
4581 // If no more than two elements come from either vector. This can be
4582 // implemented with two shuffles. First shuffle gather the elements.
4583 // The second shuffle, which takes the first shuffle as both of its
4584 // vector operands, put the elements into the right order.
4585 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4587 SmallVector<int, 8> Mask2(4U, -1);
4589 for (unsigned i = 0; i != 4; ++i) {
4590 if (Locs[i].first == -1)
4593 unsigned Idx = (i < 2) ? 0 : 4;
4594 Idx += Locs[i].first * 2 + Locs[i].second;
4599 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4600 } else if (NumLo == 3 || NumHi == 3) {
4601 // Otherwise, we must have three elements from one vector, call it X, and
4602 // one element from the other, call it Y. First, use a shufps to build an
4603 // intermediate vector with the one element from Y and the element from X
4604 // that will be in the same half in the final destination (the indexes don't
4605 // matter). Then, use a shufps to build the final vector, taking the half
4606 // containing the element from Y from the intermediate, and the other half
4609 // Normalize it so the 3 elements come from V1.
4610 CommuteVectorShuffleMask(PermMask, VT);
4614 // Find the element from V2.
4616 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4617 int Val = PermMask[HiIndex];
4624 Mask1[0] = PermMask[HiIndex];
4626 Mask1[2] = PermMask[HiIndex^1];
4628 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4631 Mask1[0] = PermMask[0];
4632 Mask1[1] = PermMask[1];
4633 Mask1[2] = HiIndex & 1 ? 6 : 4;
4634 Mask1[3] = HiIndex & 1 ? 4 : 6;
4635 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4637 Mask1[0] = HiIndex & 1 ? 2 : 0;
4638 Mask1[1] = HiIndex & 1 ? 0 : 2;
4639 Mask1[2] = PermMask[2];
4640 Mask1[3] = PermMask[3];
4645 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4649 // Break it into (shuffle shuffle_hi, shuffle_lo).
4651 SmallVector<int,8> LoMask(4U, -1);
4652 SmallVector<int,8> HiMask(4U, -1);
4654 SmallVector<int,8> *MaskPtr = &LoMask;
4655 unsigned MaskIdx = 0;
4658 for (unsigned i = 0; i != 4; ++i) {
4665 int Idx = PermMask[i];
4667 Locs[i] = std::make_pair(-1, -1);
4668 } else if (Idx < 4) {
4669 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4670 (*MaskPtr)[LoIdx] = Idx;
4673 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4674 (*MaskPtr)[HiIdx] = Idx;
4679 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4680 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4681 SmallVector<int, 8> MaskOps;
4682 for (unsigned i = 0; i != 4; ++i) {
4683 if (Locs[i].first == -1) {
4684 MaskOps.push_back(-1);
4686 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4687 MaskOps.push_back(Idx);
4690 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4694 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4695 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4696 SDValue V1 = Op.getOperand(0);
4697 SDValue V2 = Op.getOperand(1);
4698 EVT VT = Op.getValueType();
4699 DebugLoc dl = Op.getDebugLoc();
4700 unsigned NumElems = VT.getVectorNumElements();
4701 bool isMMX = VT.getSizeInBits() == 64;
4702 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4703 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4704 bool V1IsSplat = false;
4705 bool V2IsSplat = false;
4707 if (isZeroShuffle(SVOp))
4708 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4710 // Promote splats to v4f32.
4711 if (SVOp->isSplat()) {
4712 if (isMMX || NumElems < 4)
4714 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4717 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4719 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4720 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4721 if (NewOp.getNode())
4722 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4723 LowerVECTOR_SHUFFLE(NewOp, DAG));
4724 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4725 // FIXME: Figure out a cleaner way to do this.
4726 // Try to make use of movq to zero out the top part.
4727 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4728 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4729 if (NewOp.getNode()) {
4730 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4731 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4732 DAG, Subtarget, dl);
4734 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4735 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4736 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4737 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4738 DAG, Subtarget, dl);
4742 if (X86::isPSHUFDMask(SVOp))
4745 // Check if this can be converted into a logical shift.
4746 bool isLeft = false;
4749 bool isShift = getSubtarget()->hasSSE2() &&
4750 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4751 if (isShift && ShVal.hasOneUse()) {
4752 // If the shifted value has multiple uses, it may be cheaper to use
4753 // v_set0 + movlhps or movhlps, etc.
4754 EVT EltVT = VT.getVectorElementType();
4755 ShAmt *= EltVT.getSizeInBits();
4756 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4759 if (X86::isMOVLMask(SVOp)) {
4762 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4763 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4768 // FIXME: fold these into legal mask.
4769 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4770 X86::isMOVSLDUPMask(SVOp) ||
4771 X86::isMOVHLPSMask(SVOp) ||
4772 X86::isMOVLHPSMask(SVOp) ||
4773 X86::isMOVLPMask(SVOp)))
4776 if (ShouldXformToMOVHLPS(SVOp) ||
4777 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4778 return CommuteVectorShuffle(SVOp, DAG);
4781 // No better options. Use a vshl / vsrl.
4782 EVT EltVT = VT.getVectorElementType();
4783 ShAmt *= EltVT.getSizeInBits();
4784 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4787 bool Commuted = false;
4788 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4789 // 1,1,1,1 -> v8i16 though.
4790 V1IsSplat = isSplatVector(V1.getNode());
4791 V2IsSplat = isSplatVector(V2.getNode());
4793 // Canonicalize the splat or undef, if present, to be on the RHS.
4794 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4795 Op = CommuteVectorShuffle(SVOp, DAG);
4796 SVOp = cast<ShuffleVectorSDNode>(Op);
4797 V1 = SVOp->getOperand(0);
4798 V2 = SVOp->getOperand(1);
4799 std::swap(V1IsSplat, V2IsSplat);
4800 std::swap(V1IsUndef, V2IsUndef);
4804 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4805 // Shuffling low element of v1 into undef, just return v1.
4808 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4809 // the instruction selector will not match, so get a canonical MOVL with
4810 // swapped operands to undo the commute.
4811 return getMOVL(DAG, dl, VT, V2, V1);
4814 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4815 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4816 X86::isUNPCKLMask(SVOp) ||
4817 X86::isUNPCKHMask(SVOp))
4821 // Normalize mask so all entries that point to V2 points to its first
4822 // element then try to match unpck{h|l} again. If match, return a
4823 // new vector_shuffle with the corrected mask.
4824 SDValue NewMask = NormalizeMask(SVOp, DAG);
4825 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4826 if (NSVOp != SVOp) {
4827 if (X86::isUNPCKLMask(NSVOp, true)) {
4829 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4836 // Commute is back and try unpck* again.
4837 // FIXME: this seems wrong.
4838 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4839 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4840 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4841 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4842 X86::isUNPCKLMask(NewSVOp) ||
4843 X86::isUNPCKHMask(NewSVOp))
4847 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4849 // Normalize the node to match x86 shuffle ops if needed
4850 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4851 return CommuteVectorShuffle(SVOp, DAG);
4853 // Check for legal shuffle and return?
4854 SmallVector<int, 16> PermMask;
4855 SVOp->getMask(PermMask);
4856 if (isShuffleMaskLegal(PermMask, VT))
4859 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4860 if (VT == MVT::v8i16) {
4861 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4862 if (NewOp.getNode())
4866 if (VT == MVT::v16i8) {
4867 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4868 if (NewOp.getNode())
4872 // Handle all 4 wide cases with a number of shuffles except for MMX.
4873 if (NumElems == 4 && !isMMX)
4874 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4880 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4881 SelectionDAG &DAG) const {
4882 EVT VT = Op.getValueType();
4883 DebugLoc dl = Op.getDebugLoc();
4884 if (VT.getSizeInBits() == 8) {
4885 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4886 Op.getOperand(0), Op.getOperand(1));
4887 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4888 DAG.getValueType(VT));
4889 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4890 } else if (VT.getSizeInBits() == 16) {
4891 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4892 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4894 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4895 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4896 DAG.getNode(ISD::BIT_CONVERT, dl,
4900 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4901 Op.getOperand(0), Op.getOperand(1));
4902 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4903 DAG.getValueType(VT));
4904 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4905 } else if (VT == MVT::f32) {
4906 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4907 // the result back to FR32 register. It's only worth matching if the
4908 // result has a single use which is a store or a bitcast to i32. And in
4909 // the case of a store, it's not worth it if the index is a constant 0,
4910 // because a MOVSSmr can be used instead, which is smaller and faster.
4911 if (!Op.hasOneUse())
4913 SDNode *User = *Op.getNode()->use_begin();
4914 if ((User->getOpcode() != ISD::STORE ||
4915 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4916 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4917 (User->getOpcode() != ISD::BIT_CONVERT ||
4918 User->getValueType(0) != MVT::i32))
4920 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4921 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4924 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4925 } else if (VT == MVT::i32) {
4926 // ExtractPS works with constant index.
4927 if (isa<ConstantSDNode>(Op.getOperand(1)))
4935 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4936 SelectionDAG &DAG) const {
4937 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4940 if (Subtarget->hasSSE41()) {
4941 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4946 EVT VT = Op.getValueType();
4947 DebugLoc dl = Op.getDebugLoc();
4948 // TODO: handle v16i8.
4949 if (VT.getSizeInBits() == 16) {
4950 SDValue Vec = Op.getOperand(0);
4951 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4953 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4954 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4955 DAG.getNode(ISD::BIT_CONVERT, dl,
4958 // Transform it so it match pextrw which produces a 32-bit result.
4959 EVT EltVT = MVT::i32;
4960 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4961 Op.getOperand(0), Op.getOperand(1));
4962 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4963 DAG.getValueType(VT));
4964 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4965 } else if (VT.getSizeInBits() == 32) {
4966 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4970 // SHUFPS the element to the lowest double word, then movss.
4971 int Mask[4] = { Idx, -1, -1, -1 };
4972 EVT VVT = Op.getOperand(0).getValueType();
4973 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4974 DAG.getUNDEF(VVT), Mask);
4975 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4976 DAG.getIntPtrConstant(0));
4977 } else if (VT.getSizeInBits() == 64) {
4978 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4979 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4980 // to match extract_elt for f64.
4981 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4985 // UNPCKHPD the element to the lowest double word, then movsd.
4986 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4987 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4988 int Mask[2] = { 1, -1 };
4989 EVT VVT = Op.getOperand(0).getValueType();
4990 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4991 DAG.getUNDEF(VVT), Mask);
4992 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4993 DAG.getIntPtrConstant(0));
5000 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5001 SelectionDAG &DAG) const {
5002 EVT VT = Op.getValueType();
5003 EVT EltVT = VT.getVectorElementType();
5004 DebugLoc dl = Op.getDebugLoc();
5006 SDValue N0 = Op.getOperand(0);
5007 SDValue N1 = Op.getOperand(1);
5008 SDValue N2 = Op.getOperand(2);
5010 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5011 isa<ConstantSDNode>(N2)) {
5013 if (VT == MVT::v8i16)
5014 Opc = X86ISD::PINSRW;
5015 else if (VT == MVT::v4i16)
5016 Opc = X86ISD::MMX_PINSRW;
5017 else if (VT == MVT::v16i8)
5018 Opc = X86ISD::PINSRB;
5020 Opc = X86ISD::PINSRB;
5022 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5024 if (N1.getValueType() != MVT::i32)
5025 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5026 if (N2.getValueType() != MVT::i32)
5027 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5028 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5029 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5030 // Bits [7:6] of the constant are the source select. This will always be
5031 // zero here. The DAG Combiner may combine an extract_elt index into these
5032 // bits. For example (insert (extract, 3), 2) could be matched by putting
5033 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5034 // Bits [5:4] of the constant are the destination select. This is the
5035 // value of the incoming immediate.
5036 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5037 // combine either bitwise AND or insert of float 0.0 to set these bits.
5038 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5039 // Create this as a scalar to vector..
5040 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5041 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5042 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5043 // PINSR* works with constant index.
5050 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5051 EVT VT = Op.getValueType();
5052 EVT EltVT = VT.getVectorElementType();
5054 if (Subtarget->hasSSE41())
5055 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5057 if (EltVT == MVT::i8)
5060 DebugLoc dl = Op.getDebugLoc();
5061 SDValue N0 = Op.getOperand(0);
5062 SDValue N1 = Op.getOperand(1);
5063 SDValue N2 = Op.getOperand(2);
5065 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5066 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5067 // as its second argument.
5068 if (N1.getValueType() != MVT::i32)
5069 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5070 if (N2.getValueType() != MVT::i32)
5071 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5072 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5073 dl, VT, N0, N1, N2);
5079 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5080 DebugLoc dl = Op.getDebugLoc();
5081 if (Op.getValueType() == MVT::v2f32)
5082 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
5083 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
5084 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
5085 Op.getOperand(0))));
5087 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
5088 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5090 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5091 EVT VT = MVT::v2i32;
5092 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5099 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5100 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5103 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5104 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5105 // one of the above mentioned nodes. It has to be wrapped because otherwise
5106 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5107 // be used to form addressing mode. These wrapped nodes will be selected
5110 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5111 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5113 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5115 unsigned char OpFlag = 0;
5116 unsigned WrapperKind = X86ISD::Wrapper;
5117 CodeModel::Model M = getTargetMachine().getCodeModel();
5119 if (Subtarget->isPICStyleRIPRel() &&
5120 (M == CodeModel::Small || M == CodeModel::Kernel))
5121 WrapperKind = X86ISD::WrapperRIP;
5122 else if (Subtarget->isPICStyleGOT())
5123 OpFlag = X86II::MO_GOTOFF;
5124 else if (Subtarget->isPICStyleStubPIC())
5125 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5127 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5129 CP->getOffset(), OpFlag);
5130 DebugLoc DL = CP->getDebugLoc();
5131 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5132 // With PIC, the address is actually $g + Offset.
5134 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5135 DAG.getNode(X86ISD::GlobalBaseReg,
5136 DebugLoc(), getPointerTy()),
5143 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5144 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5146 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5148 unsigned char OpFlag = 0;
5149 unsigned WrapperKind = X86ISD::Wrapper;
5150 CodeModel::Model M = getTargetMachine().getCodeModel();
5152 if (Subtarget->isPICStyleRIPRel() &&
5153 (M == CodeModel::Small || M == CodeModel::Kernel))
5154 WrapperKind = X86ISD::WrapperRIP;
5155 else if (Subtarget->isPICStyleGOT())
5156 OpFlag = X86II::MO_GOTOFF;
5157 else if (Subtarget->isPICStyleStubPIC())
5158 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5160 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5162 DebugLoc DL = JT->getDebugLoc();
5163 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5165 // With PIC, the address is actually $g + Offset.
5167 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5168 DAG.getNode(X86ISD::GlobalBaseReg,
5169 DebugLoc(), getPointerTy()),
5177 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5178 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5180 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5182 unsigned char OpFlag = 0;
5183 unsigned WrapperKind = X86ISD::Wrapper;
5184 CodeModel::Model M = getTargetMachine().getCodeModel();
5186 if (Subtarget->isPICStyleRIPRel() &&
5187 (M == CodeModel::Small || M == CodeModel::Kernel))
5188 WrapperKind = X86ISD::WrapperRIP;
5189 else if (Subtarget->isPICStyleGOT())
5190 OpFlag = X86II::MO_GOTOFF;
5191 else if (Subtarget->isPICStyleStubPIC())
5192 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5194 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5196 DebugLoc DL = Op.getDebugLoc();
5197 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5200 // With PIC, the address is actually $g + Offset.
5201 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5202 !Subtarget->is64Bit()) {
5203 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5204 DAG.getNode(X86ISD::GlobalBaseReg,
5205 DebugLoc(), getPointerTy()),
5213 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5214 // Create the TargetBlockAddressAddress node.
5215 unsigned char OpFlags =
5216 Subtarget->ClassifyBlockAddressReference();
5217 CodeModel::Model M = getTargetMachine().getCodeModel();
5218 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5219 DebugLoc dl = Op.getDebugLoc();
5220 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5221 /*isTarget=*/true, OpFlags);
5223 if (Subtarget->isPICStyleRIPRel() &&
5224 (M == CodeModel::Small || M == CodeModel::Kernel))
5225 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5227 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5229 // With PIC, the address is actually $g + Offset.
5230 if (isGlobalRelativeToPICBase(OpFlags)) {
5231 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5232 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5240 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5242 SelectionDAG &DAG) const {
5243 // Create the TargetGlobalAddress node, folding in the constant
5244 // offset if it is legal.
5245 unsigned char OpFlags =
5246 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5247 CodeModel::Model M = getTargetMachine().getCodeModel();
5249 if (OpFlags == X86II::MO_NO_FLAG &&
5250 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5251 // A direct static reference to a global.
5252 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5255 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5258 if (Subtarget->isPICStyleRIPRel() &&
5259 (M == CodeModel::Small || M == CodeModel::Kernel))
5260 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5262 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5264 // With PIC, the address is actually $g + Offset.
5265 if (isGlobalRelativeToPICBase(OpFlags)) {
5266 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5267 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5271 // For globals that require a load from a stub to get the address, emit the
5273 if (isGlobalStubReference(OpFlags))
5274 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5275 PseudoSourceValue::getGOT(), 0, false, false, 0);
5277 // If there was a non-zero offset that we didn't fold, create an explicit
5280 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5281 DAG.getConstant(Offset, getPointerTy()));
5287 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5288 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5289 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5290 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5294 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5295 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5296 unsigned char OperandFlags) {
5297 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5298 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5299 DebugLoc dl = GA->getDebugLoc();
5300 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5301 GA->getValueType(0),
5305 SDValue Ops[] = { Chain, TGA, *InFlag };
5306 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5308 SDValue Ops[] = { Chain, TGA };
5309 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5312 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5313 MFI->setAdjustsStack(true);
5315 SDValue Flag = Chain.getValue(1);
5316 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5319 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5321 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5324 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5325 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5326 DAG.getNode(X86ISD::GlobalBaseReg,
5327 DebugLoc(), PtrVT), InFlag);
5328 InFlag = Chain.getValue(1);
5330 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5333 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5335 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5337 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5338 X86::RAX, X86II::MO_TLSGD);
5341 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5342 // "local exec" model.
5343 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5344 const EVT PtrVT, TLSModel::Model model,
5346 DebugLoc dl = GA->getDebugLoc();
5347 // Get the Thread Pointer
5348 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5350 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5353 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5354 NULL, 0, false, false, 0);
5356 unsigned char OperandFlags = 0;
5357 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5359 unsigned WrapperKind = X86ISD::Wrapper;
5360 if (model == TLSModel::LocalExec) {
5361 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5362 } else if (is64Bit) {
5363 assert(model == TLSModel::InitialExec);
5364 OperandFlags = X86II::MO_GOTTPOFF;
5365 WrapperKind = X86ISD::WrapperRIP;
5367 assert(model == TLSModel::InitialExec);
5368 OperandFlags = X86II::MO_INDNTPOFF;
5371 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5373 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5374 GA->getOffset(), OperandFlags);
5375 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5377 if (model == TLSModel::InitialExec)
5378 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5379 PseudoSourceValue::getGOT(), 0, false, false, 0);
5381 // The address of the thread local variable is the add of the thread
5382 // pointer with the offset of the variable.
5383 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5387 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5389 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5390 const GlobalValue *GV = GA->getGlobal();
5392 if (Subtarget->isTargetELF()) {
5393 // TODO: implement the "local dynamic" model
5394 // TODO: implement the "initial exec"model for pic executables
5396 // If GV is an alias then use the aliasee for determining
5397 // thread-localness.
5398 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5399 GV = GA->resolveAliasedGlobal(false);
5401 TLSModel::Model model
5402 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5405 case TLSModel::GeneralDynamic:
5406 case TLSModel::LocalDynamic: // not implemented
5407 if (Subtarget->is64Bit())
5408 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5409 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5411 case TLSModel::InitialExec:
5412 case TLSModel::LocalExec:
5413 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5414 Subtarget->is64Bit());
5416 } else if (Subtarget->isTargetDarwin()) {
5417 // Darwin only has one model of TLS. Lower to that.
5418 unsigned char OpFlag = 0;
5419 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5420 X86ISD::WrapperRIP : X86ISD::Wrapper;
5422 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5424 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5425 !Subtarget->is64Bit();
5427 OpFlag = X86II::MO_TLVP_PIC_BASE;
5429 OpFlag = X86II::MO_TLVP;
5431 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(),
5433 GA->getOffset(), OpFlag);
5435 DebugLoc DL = Op.getDebugLoc();
5436 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5438 // With PIC32, the address is actually $g + Offset.
5440 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5441 DAG.getNode(X86ISD::GlobalBaseReg,
5442 DebugLoc(), getPointerTy()),
5445 // Lowering the machine isd will make sure everything is in the right
5447 SDValue Args[] = { Offset };
5448 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5450 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5451 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5452 MFI->setAdjustsStack(true);
5454 // And our return value (tls address) is in the standard call return value
5456 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5457 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5461 "TLS not implemented for this target.");
5463 llvm_unreachable("Unreachable");
5468 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5469 /// take a 2 x i32 value to shift plus a shift amount.
5470 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5471 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5472 EVT VT = Op.getValueType();
5473 unsigned VTBits = VT.getSizeInBits();
5474 DebugLoc dl = Op.getDebugLoc();
5475 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5476 SDValue ShOpLo = Op.getOperand(0);
5477 SDValue ShOpHi = Op.getOperand(1);
5478 SDValue ShAmt = Op.getOperand(2);
5479 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5480 DAG.getConstant(VTBits - 1, MVT::i8))
5481 : DAG.getConstant(0, VT);
5484 if (Op.getOpcode() == ISD::SHL_PARTS) {
5485 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5486 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5488 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5489 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5492 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5493 DAG.getConstant(VTBits, MVT::i8));
5494 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5495 AndNode, DAG.getConstant(0, MVT::i8));
5498 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5499 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5500 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5502 if (Op.getOpcode() == ISD::SHL_PARTS) {
5503 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5504 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5506 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5507 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5510 SDValue Ops[2] = { Lo, Hi };
5511 return DAG.getMergeValues(Ops, 2, dl);
5514 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5515 SelectionDAG &DAG) const {
5516 EVT SrcVT = Op.getOperand(0).getValueType();
5518 if (SrcVT.isVector()) {
5519 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5525 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5526 "Unknown SINT_TO_FP to lower!");
5528 // These are really Legal; return the operand so the caller accepts it as
5530 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5532 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5533 Subtarget->is64Bit()) {
5537 DebugLoc dl = Op.getDebugLoc();
5538 unsigned Size = SrcVT.getSizeInBits()/8;
5539 MachineFunction &MF = DAG.getMachineFunction();
5540 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5541 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5542 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5544 PseudoSourceValue::getFixedStack(SSFI), 0,
5546 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5549 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5551 SelectionDAG &DAG) const {
5553 DebugLoc dl = Op.getDebugLoc();
5555 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5557 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5559 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5560 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5561 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5562 Tys, Ops, array_lengthof(Ops));
5565 Chain = Result.getValue(1);
5566 SDValue InFlag = Result.getValue(2);
5568 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5569 // shouldn't be necessary except that RFP cannot be live across
5570 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5571 MachineFunction &MF = DAG.getMachineFunction();
5572 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5573 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5574 Tys = DAG.getVTList(MVT::Other);
5576 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5578 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5579 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5580 PseudoSourceValue::getFixedStack(SSFI), 0,
5587 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5588 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5589 SelectionDAG &DAG) const {
5590 // This algorithm is not obvious. Here it is in C code, more or less:
5592 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5593 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5594 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5596 // Copy ints to xmm registers.
5597 __m128i xh = _mm_cvtsi32_si128( hi );
5598 __m128i xl = _mm_cvtsi32_si128( lo );
5600 // Combine into low half of a single xmm register.
5601 __m128i x = _mm_unpacklo_epi32( xh, xl );
5605 // Merge in appropriate exponents to give the integer bits the right
5607 x = _mm_unpacklo_epi32( x, exp );
5609 // Subtract away the biases to deal with the IEEE-754 double precision
5611 d = _mm_sub_pd( (__m128d) x, bias );
5613 // All conversions up to here are exact. The correctly rounded result is
5614 // calculated using the current rounding mode using the following
5616 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5617 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5618 // store doesn't really need to be here (except
5619 // maybe to zero the other double)
5624 DebugLoc dl = Op.getDebugLoc();
5625 LLVMContext *Context = DAG.getContext();
5627 // Build some magic constants.
5628 std::vector<Constant*> CV0;
5629 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5630 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5631 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5632 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5633 Constant *C0 = ConstantVector::get(CV0);
5634 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5636 std::vector<Constant*> CV1;
5638 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5640 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5641 Constant *C1 = ConstantVector::get(CV1);
5642 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5644 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5645 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5647 DAG.getIntPtrConstant(1)));
5648 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5649 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5651 DAG.getIntPtrConstant(0)));
5652 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5653 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5654 PseudoSourceValue::getConstantPool(), 0,
5656 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5657 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5658 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5659 PseudoSourceValue::getConstantPool(), 0,
5661 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5663 // Add the halves; easiest way is to swap them into another reg first.
5664 int ShufMask[2] = { 1, -1 };
5665 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5666 DAG.getUNDEF(MVT::v2f64), ShufMask);
5667 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5668 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5669 DAG.getIntPtrConstant(0));
5672 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5673 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5674 SelectionDAG &DAG) const {
5675 DebugLoc dl = Op.getDebugLoc();
5676 // FP constant to bias correct the final result.
5677 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5680 // Load the 32-bit value into an XMM register.
5681 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5682 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5684 DAG.getIntPtrConstant(0)));
5686 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5687 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5688 DAG.getIntPtrConstant(0));
5690 // Or the load with the bias.
5691 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5692 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5693 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5695 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5696 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5697 MVT::v2f64, Bias)));
5698 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5699 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5700 DAG.getIntPtrConstant(0));
5702 // Subtract the bias.
5703 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5705 // Handle final rounding.
5706 EVT DestVT = Op.getValueType();
5708 if (DestVT.bitsLT(MVT::f64)) {
5709 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5710 DAG.getIntPtrConstant(0));
5711 } else if (DestVT.bitsGT(MVT::f64)) {
5712 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5715 // Handle final rounding.
5719 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5720 SelectionDAG &DAG) const {
5721 SDValue N0 = Op.getOperand(0);
5722 DebugLoc dl = Op.getDebugLoc();
5724 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5725 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5726 // the optimization here.
5727 if (DAG.SignBitIsZero(N0))
5728 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5730 EVT SrcVT = N0.getValueType();
5731 EVT DstVT = Op.getValueType();
5732 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5733 return LowerUINT_TO_FP_i64(Op, DAG);
5734 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5735 return LowerUINT_TO_FP_i32(Op, DAG);
5737 // Make a 64-bit buffer, and use it to build an FILD.
5738 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5739 if (SrcVT == MVT::i32) {
5740 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5741 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5742 getPointerTy(), StackSlot, WordOff);
5743 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5744 StackSlot, NULL, 0, false, false, 0);
5745 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5746 OffsetSlot, NULL, 0, false, false, 0);
5747 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5751 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5752 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5753 StackSlot, NULL, 0, false, false, 0);
5754 // For i64 source, we need to add the appropriate power of 2 if the input
5755 // was negative. This is the same as the optimization in
5756 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5757 // we must be careful to do the computation in x87 extended precision, not
5758 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5759 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5760 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5761 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5763 APInt FF(32, 0x5F800000ULL);
5765 // Check whether the sign bit is set.
5766 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5767 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5770 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5771 SDValue FudgePtr = DAG.getConstantPool(
5772 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5775 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5776 SDValue Zero = DAG.getIntPtrConstant(0);
5777 SDValue Four = DAG.getIntPtrConstant(4);
5778 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5780 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5782 // Load the value out, extending it from f32 to f80.
5783 // FIXME: Avoid the extend by constructing the right constant pool?
5784 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
5785 FudgePtr, PseudoSourceValue::getConstantPool(),
5786 0, MVT::f32, false, false, 4);
5787 // Extend everything to 80 bits to force it to be done on x87.
5788 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5789 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5792 std::pair<SDValue,SDValue> X86TargetLowering::
5793 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5794 DebugLoc dl = Op.getDebugLoc();
5796 EVT DstTy = Op.getValueType();
5799 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5803 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5804 DstTy.getSimpleVT() >= MVT::i16 &&
5805 "Unknown FP_TO_SINT to lower!");
5807 // These are really Legal.
5808 if (DstTy == MVT::i32 &&
5809 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5810 return std::make_pair(SDValue(), SDValue());
5811 if (Subtarget->is64Bit() &&
5812 DstTy == MVT::i64 &&
5813 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5814 return std::make_pair(SDValue(), SDValue());
5816 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5818 MachineFunction &MF = DAG.getMachineFunction();
5819 unsigned MemSize = DstTy.getSizeInBits()/8;
5820 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5821 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5824 switch (DstTy.getSimpleVT().SimpleTy) {
5825 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5826 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5827 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5828 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5831 SDValue Chain = DAG.getEntryNode();
5832 SDValue Value = Op.getOperand(0);
5833 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5834 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5835 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5836 PseudoSourceValue::getFixedStack(SSFI), 0,
5838 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5840 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5842 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5843 Chain = Value.getValue(1);
5844 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5845 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5848 // Build the FP_TO_INT*_IN_MEM
5849 SDValue Ops[] = { Chain, Value, StackSlot };
5850 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5852 return std::make_pair(FIST, StackSlot);
5855 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5856 SelectionDAG &DAG) const {
5857 if (Op.getValueType().isVector()) {
5858 if (Op.getValueType() == MVT::v2i32 &&
5859 Op.getOperand(0).getValueType() == MVT::v2f64) {
5865 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5866 SDValue FIST = Vals.first, StackSlot = Vals.second;
5867 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5868 if (FIST.getNode() == 0) return Op;
5871 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5872 FIST, StackSlot, NULL, 0, false, false, 0);
5875 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5876 SelectionDAG &DAG) const {
5877 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5878 SDValue FIST = Vals.first, StackSlot = Vals.second;
5879 assert(FIST.getNode() && "Unexpected failure");
5882 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5883 FIST, StackSlot, NULL, 0, false, false, 0);
5886 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5887 SelectionDAG &DAG) const {
5888 LLVMContext *Context = DAG.getContext();
5889 DebugLoc dl = Op.getDebugLoc();
5890 EVT VT = Op.getValueType();
5893 EltVT = VT.getVectorElementType();
5894 std::vector<Constant*> CV;
5895 if (EltVT == MVT::f64) {
5896 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5900 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5906 Constant *C = ConstantVector::get(CV);
5907 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5908 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5909 PseudoSourceValue::getConstantPool(), 0,
5911 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5914 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5915 LLVMContext *Context = DAG.getContext();
5916 DebugLoc dl = Op.getDebugLoc();
5917 EVT VT = Op.getValueType();
5920 EltVT = VT.getVectorElementType();
5921 std::vector<Constant*> CV;
5922 if (EltVT == MVT::f64) {
5923 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5927 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5933 Constant *C = ConstantVector::get(CV);
5934 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5935 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5936 PseudoSourceValue::getConstantPool(), 0,
5938 if (VT.isVector()) {
5939 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5940 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5941 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5943 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5945 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5949 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5950 LLVMContext *Context = DAG.getContext();
5951 SDValue Op0 = Op.getOperand(0);
5952 SDValue Op1 = Op.getOperand(1);
5953 DebugLoc dl = Op.getDebugLoc();
5954 EVT VT = Op.getValueType();
5955 EVT SrcVT = Op1.getValueType();
5957 // If second operand is smaller, extend it first.
5958 if (SrcVT.bitsLT(VT)) {
5959 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5962 // And if it is bigger, shrink it first.
5963 if (SrcVT.bitsGT(VT)) {
5964 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5968 // At this point the operands and the result should have the same
5969 // type, and that won't be f80 since that is not custom lowered.
5971 // First get the sign bit of second operand.
5972 std::vector<Constant*> CV;
5973 if (SrcVT == MVT::f64) {
5974 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5975 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5977 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5978 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5979 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5980 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5982 Constant *C = ConstantVector::get(CV);
5983 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5984 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5985 PseudoSourceValue::getConstantPool(), 0,
5987 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5989 // Shift sign bit right or left if the two operands have different types.
5990 if (SrcVT.bitsGT(VT)) {
5991 // Op0 is MVT::f32, Op1 is MVT::f64.
5992 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5993 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5994 DAG.getConstant(32, MVT::i32));
5995 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5996 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5997 DAG.getIntPtrConstant(0));
6000 // Clear first operand sign bit.
6002 if (VT == MVT::f64) {
6003 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6004 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6006 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6007 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6008 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6009 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6011 C = ConstantVector::get(CV);
6012 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6013 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6014 PseudoSourceValue::getConstantPool(), 0,
6016 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6018 // Or the value with the sign bit.
6019 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6022 /// Emit nodes that will be selected as "test Op0,Op0", or something
6024 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6025 SelectionDAG &DAG) const {
6026 DebugLoc dl = Op.getDebugLoc();
6028 // CF and OF aren't always set the way we want. Determine which
6029 // of these we need.
6030 bool NeedCF = false;
6031 bool NeedOF = false;
6034 case X86::COND_A: case X86::COND_AE:
6035 case X86::COND_B: case X86::COND_BE:
6038 case X86::COND_G: case X86::COND_GE:
6039 case X86::COND_L: case X86::COND_LE:
6040 case X86::COND_O: case X86::COND_NO:
6045 // See if we can use the EFLAGS value from the operand instead of
6046 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6047 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6048 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6049 // Emit a CMP with 0, which is the TEST pattern.
6050 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6051 DAG.getConstant(0, Op.getValueType()));
6053 unsigned Opcode = 0;
6054 unsigned NumOperands = 0;
6055 switch (Op.getNode()->getOpcode()) {
6057 // Due to an isel shortcoming, be conservative if this add is likely to be
6058 // selected as part of a load-modify-store instruction. When the root node
6059 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6060 // uses of other nodes in the match, such as the ADD in this case. This
6061 // leads to the ADD being left around and reselected, with the result being
6062 // two adds in the output. Alas, even if none our users are stores, that
6063 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6064 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6065 // climbing the DAG back to the root, and it doesn't seem to be worth the
6067 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6068 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6069 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6072 if (ConstantSDNode *C =
6073 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6074 // An add of one will be selected as an INC.
6075 if (C->getAPIntValue() == 1) {
6076 Opcode = X86ISD::INC;
6081 // An add of negative one (subtract of one) will be selected as a DEC.
6082 if (C->getAPIntValue().isAllOnesValue()) {
6083 Opcode = X86ISD::DEC;
6089 // Otherwise use a regular EFLAGS-setting add.
6090 Opcode = X86ISD::ADD;
6094 // If the primary and result isn't used, don't bother using X86ISD::AND,
6095 // because a TEST instruction will be better.
6096 bool NonFlagUse = false;
6097 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6098 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6100 unsigned UOpNo = UI.getOperandNo();
6101 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6102 // Look pass truncate.
6103 UOpNo = User->use_begin().getOperandNo();
6104 User = *User->use_begin();
6107 if (User->getOpcode() != ISD::BRCOND &&
6108 User->getOpcode() != ISD::SETCC &&
6109 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6122 // Due to the ISEL shortcoming noted above, be conservative if this op is
6123 // likely to be selected as part of a load-modify-store instruction.
6124 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6125 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6126 if (UI->getOpcode() == ISD::STORE)
6129 // Otherwise use a regular EFLAGS-setting instruction.
6130 switch (Op.getNode()->getOpcode()) {
6131 default: llvm_unreachable("unexpected operator!");
6132 case ISD::SUB: Opcode = X86ISD::SUB; break;
6133 case ISD::OR: Opcode = X86ISD::OR; break;
6134 case ISD::XOR: Opcode = X86ISD::XOR; break;
6135 case ISD::AND: Opcode = X86ISD::AND; break;
6147 return SDValue(Op.getNode(), 1);
6154 // Emit a CMP with 0, which is the TEST pattern.
6155 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6156 DAG.getConstant(0, Op.getValueType()));
6158 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6159 SmallVector<SDValue, 4> Ops;
6160 for (unsigned i = 0; i != NumOperands; ++i)
6161 Ops.push_back(Op.getOperand(i));
6163 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6164 DAG.ReplaceAllUsesWith(Op, New);
6165 return SDValue(New.getNode(), 1);
6168 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6170 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6171 SelectionDAG &DAG) const {
6172 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6173 if (C->getAPIntValue() == 0)
6174 return EmitTest(Op0, X86CC, DAG);
6176 DebugLoc dl = Op0.getDebugLoc();
6177 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6180 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6181 /// if it's possible.
6182 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6183 DebugLoc dl, SelectionDAG &DAG) const {
6184 SDValue Op0 = And.getOperand(0);
6185 SDValue Op1 = And.getOperand(1);
6186 if (Op0.getOpcode() == ISD::TRUNCATE)
6187 Op0 = Op0.getOperand(0);
6188 if (Op1.getOpcode() == ISD::TRUNCATE)
6189 Op1 = Op1.getOperand(0);
6192 if (Op1.getOpcode() == ISD::SHL)
6193 std::swap(Op0, Op1);
6194 if (Op0.getOpcode() == ISD::SHL) {
6195 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6196 if (And00C->getZExtValue() == 1) {
6197 // If we looked past a truncate, check that it's only truncating away
6199 unsigned BitWidth = Op0.getValueSizeInBits();
6200 unsigned AndBitWidth = And.getValueSizeInBits();
6201 if (BitWidth > AndBitWidth) {
6202 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6203 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6204 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6208 RHS = Op0.getOperand(1);
6210 } else if (Op1.getOpcode() == ISD::Constant) {
6211 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6212 SDValue AndLHS = Op0;
6213 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6214 LHS = AndLHS.getOperand(0);
6215 RHS = AndLHS.getOperand(1);
6219 if (LHS.getNode()) {
6220 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6221 // instruction. Since the shift amount is in-range-or-undefined, we know
6222 // that doing a bittest on the i32 value is ok. We extend to i32 because
6223 // the encoding for the i16 version is larger than the i32 version.
6224 // Also promote i16 to i32 for performance / code size reason.
6225 if (LHS.getValueType() == MVT::i8 ||
6226 LHS.getValueType() == MVT::i16)
6227 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6229 // If the operand types disagree, extend the shift amount to match. Since
6230 // BT ignores high bits (like shifts) we can use anyextend.
6231 if (LHS.getValueType() != RHS.getValueType())
6232 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6234 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6235 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6236 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6237 DAG.getConstant(Cond, MVT::i8), BT);
6243 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6244 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6245 SDValue Op0 = Op.getOperand(0);
6246 SDValue Op1 = Op.getOperand(1);
6247 DebugLoc dl = Op.getDebugLoc();
6248 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6250 // Optimize to BT if possible.
6251 // Lower (X & (1 << N)) == 0 to BT(X, N).
6252 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6253 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6254 if (Op0.getOpcode() == ISD::AND &&
6256 Op1.getOpcode() == ISD::Constant &&
6257 cast<ConstantSDNode>(Op1)->isNullValue() &&
6258 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6259 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6260 if (NewSetCC.getNode())
6264 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6265 if (Op0.getOpcode() == X86ISD::SETCC &&
6266 Op1.getOpcode() == ISD::Constant &&
6267 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6268 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6269 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6270 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6271 bool Invert = (CC == ISD::SETNE) ^
6272 cast<ConstantSDNode>(Op1)->isNullValue();
6274 CCode = X86::GetOppositeBranchCondition(CCode);
6275 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6276 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6279 bool isFP = Op1.getValueType().isFloatingPoint();
6280 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6281 if (X86CC == X86::COND_INVALID)
6284 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6286 // Use sbb x, x to materialize carry bit into a GPR.
6287 if (X86CC == X86::COND_B)
6288 return DAG.getNode(ISD::AND, dl, MVT::i8,
6289 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6290 DAG.getConstant(X86CC, MVT::i8), Cond),
6291 DAG.getConstant(1, MVT::i8));
6293 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6294 DAG.getConstant(X86CC, MVT::i8), Cond);
6297 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6299 SDValue Op0 = Op.getOperand(0);
6300 SDValue Op1 = Op.getOperand(1);
6301 SDValue CC = Op.getOperand(2);
6302 EVT VT = Op.getValueType();
6303 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6304 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6305 DebugLoc dl = Op.getDebugLoc();
6309 EVT VT0 = Op0.getValueType();
6310 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6311 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6314 switch (SetCCOpcode) {
6317 case ISD::SETEQ: SSECC = 0; break;
6319 case ISD::SETGT: Swap = true; // Fallthrough
6321 case ISD::SETOLT: SSECC = 1; break;
6323 case ISD::SETGE: Swap = true; // Fallthrough
6325 case ISD::SETOLE: SSECC = 2; break;
6326 case ISD::SETUO: SSECC = 3; break;
6328 case ISD::SETNE: SSECC = 4; break;
6329 case ISD::SETULE: Swap = true;
6330 case ISD::SETUGE: SSECC = 5; break;
6331 case ISD::SETULT: Swap = true;
6332 case ISD::SETUGT: SSECC = 6; break;
6333 case ISD::SETO: SSECC = 7; break;
6336 std::swap(Op0, Op1);
6338 // In the two special cases we can't handle, emit two comparisons.
6340 if (SetCCOpcode == ISD::SETUEQ) {
6342 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6343 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6344 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6346 else if (SetCCOpcode == ISD::SETONE) {
6348 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6349 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6350 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6352 llvm_unreachable("Illegal FP comparison");
6354 // Handle all other FP comparisons here.
6355 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6358 // We are handling one of the integer comparisons here. Since SSE only has
6359 // GT and EQ comparisons for integer, swapping operands and multiple
6360 // operations may be required for some comparisons.
6361 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6362 bool Swap = false, Invert = false, FlipSigns = false;
6364 switch (VT.getSimpleVT().SimpleTy) {
6367 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6369 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6371 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6372 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6375 switch (SetCCOpcode) {
6377 case ISD::SETNE: Invert = true;
6378 case ISD::SETEQ: Opc = EQOpc; break;
6379 case ISD::SETLT: Swap = true;
6380 case ISD::SETGT: Opc = GTOpc; break;
6381 case ISD::SETGE: Swap = true;
6382 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6383 case ISD::SETULT: Swap = true;
6384 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6385 case ISD::SETUGE: Swap = true;
6386 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6389 std::swap(Op0, Op1);
6391 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6392 // bits of the inputs before performing those operations.
6394 EVT EltVT = VT.getVectorElementType();
6395 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6397 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6398 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6400 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6401 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6404 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6406 // If the logical-not of the result is required, perform that now.
6408 Result = DAG.getNOT(dl, Result, VT);
6413 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6414 static bool isX86LogicalCmp(SDValue Op) {
6415 unsigned Opc = Op.getNode()->getOpcode();
6416 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6418 if (Op.getResNo() == 1 &&
6419 (Opc == X86ISD::ADD ||
6420 Opc == X86ISD::SUB ||
6421 Opc == X86ISD::SMUL ||
6422 Opc == X86ISD::UMUL ||
6423 Opc == X86ISD::INC ||
6424 Opc == X86ISD::DEC ||
6425 Opc == X86ISD::OR ||
6426 Opc == X86ISD::XOR ||
6427 Opc == X86ISD::AND))
6433 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6434 bool addTest = true;
6435 SDValue Cond = Op.getOperand(0);
6436 DebugLoc dl = Op.getDebugLoc();
6439 if (Cond.getOpcode() == ISD::SETCC) {
6440 SDValue NewCond = LowerSETCC(Cond, DAG);
6441 if (NewCond.getNode())
6445 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6446 SDValue Op1 = Op.getOperand(1);
6447 SDValue Op2 = Op.getOperand(2);
6448 if (Cond.getOpcode() == X86ISD::SETCC &&
6449 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6450 SDValue Cmp = Cond.getOperand(1);
6451 if (Cmp.getOpcode() == X86ISD::CMP) {
6452 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6453 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6454 ConstantSDNode *RHSC =
6455 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6456 if (N1C && N1C->isAllOnesValue() &&
6457 N2C && N2C->isNullValue() &&
6458 RHSC && RHSC->isNullValue()) {
6459 SDValue CmpOp0 = Cmp.getOperand(0);
6460 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6461 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6462 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6463 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6468 // Look pass (and (setcc_carry (cmp ...)), 1).
6469 if (Cond.getOpcode() == ISD::AND &&
6470 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6471 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6472 if (C && C->getAPIntValue() == 1)
6473 Cond = Cond.getOperand(0);
6476 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6477 // setting operand in place of the X86ISD::SETCC.
6478 if (Cond.getOpcode() == X86ISD::SETCC ||
6479 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6480 CC = Cond.getOperand(0);
6482 SDValue Cmp = Cond.getOperand(1);
6483 unsigned Opc = Cmp.getOpcode();
6484 EVT VT = Op.getValueType();
6486 bool IllegalFPCMov = false;
6487 if (VT.isFloatingPoint() && !VT.isVector() &&
6488 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6489 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6491 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6492 Opc == X86ISD::BT) { // FIXME
6499 // Look pass the truncate.
6500 if (Cond.getOpcode() == ISD::TRUNCATE)
6501 Cond = Cond.getOperand(0);
6503 // We know the result of AND is compared against zero. Try to match
6505 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6506 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6507 if (NewSetCC.getNode()) {
6508 CC = NewSetCC.getOperand(0);
6509 Cond = NewSetCC.getOperand(1);
6516 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6517 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6520 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6521 // condition is true.
6522 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6523 SDValue Ops[] = { Op2, Op1, CC, Cond };
6524 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6527 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6528 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6529 // from the AND / OR.
6530 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6531 Opc = Op.getOpcode();
6532 if (Opc != ISD::OR && Opc != ISD::AND)
6534 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6535 Op.getOperand(0).hasOneUse() &&
6536 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6537 Op.getOperand(1).hasOneUse());
6540 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6541 // 1 and that the SETCC node has a single use.
6542 static bool isXor1OfSetCC(SDValue Op) {
6543 if (Op.getOpcode() != ISD::XOR)
6545 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6546 if (N1C && N1C->getAPIntValue() == 1) {
6547 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6548 Op.getOperand(0).hasOneUse();
6553 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6554 bool addTest = true;
6555 SDValue Chain = Op.getOperand(0);
6556 SDValue Cond = Op.getOperand(1);
6557 SDValue Dest = Op.getOperand(2);
6558 DebugLoc dl = Op.getDebugLoc();
6561 if (Cond.getOpcode() == ISD::SETCC) {
6562 SDValue NewCond = LowerSETCC(Cond, DAG);
6563 if (NewCond.getNode())
6567 // FIXME: LowerXALUO doesn't handle these!!
6568 else if (Cond.getOpcode() == X86ISD::ADD ||
6569 Cond.getOpcode() == X86ISD::SUB ||
6570 Cond.getOpcode() == X86ISD::SMUL ||
6571 Cond.getOpcode() == X86ISD::UMUL)
6572 Cond = LowerXALUO(Cond, DAG);
6575 // Look pass (and (setcc_carry (cmp ...)), 1).
6576 if (Cond.getOpcode() == ISD::AND &&
6577 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6578 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6579 if (C && C->getAPIntValue() == 1)
6580 Cond = Cond.getOperand(0);
6583 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6584 // setting operand in place of the X86ISD::SETCC.
6585 if (Cond.getOpcode() == X86ISD::SETCC ||
6586 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6587 CC = Cond.getOperand(0);
6589 SDValue Cmp = Cond.getOperand(1);
6590 unsigned Opc = Cmp.getOpcode();
6591 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6592 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6596 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6600 // These can only come from an arithmetic instruction with overflow,
6601 // e.g. SADDO, UADDO.
6602 Cond = Cond.getNode()->getOperand(1);
6609 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6610 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6611 if (CondOpc == ISD::OR) {
6612 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6613 // two branches instead of an explicit OR instruction with a
6615 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6616 isX86LogicalCmp(Cmp)) {
6617 CC = Cond.getOperand(0).getOperand(0);
6618 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6619 Chain, Dest, CC, Cmp);
6620 CC = Cond.getOperand(1).getOperand(0);
6624 } else { // ISD::AND
6625 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6626 // two branches instead of an explicit AND instruction with a
6627 // separate test. However, we only do this if this block doesn't
6628 // have a fall-through edge, because this requires an explicit
6629 // jmp when the condition is false.
6630 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6631 isX86LogicalCmp(Cmp) &&
6632 Op.getNode()->hasOneUse()) {
6633 X86::CondCode CCode =
6634 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6635 CCode = X86::GetOppositeBranchCondition(CCode);
6636 CC = DAG.getConstant(CCode, MVT::i8);
6637 SDNode *User = *Op.getNode()->use_begin();
6638 // Look for an unconditional branch following this conditional branch.
6639 // We need this because we need to reverse the successors in order
6640 // to implement FCMP_OEQ.
6641 if (User->getOpcode() == ISD::BR) {
6642 SDValue FalseBB = User->getOperand(1);
6644 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6645 assert(NewBR == User);
6649 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6650 Chain, Dest, CC, Cmp);
6651 X86::CondCode CCode =
6652 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6653 CCode = X86::GetOppositeBranchCondition(CCode);
6654 CC = DAG.getConstant(CCode, MVT::i8);
6660 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6661 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6662 // It should be transformed during dag combiner except when the condition
6663 // is set by a arithmetics with overflow node.
6664 X86::CondCode CCode =
6665 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6666 CCode = X86::GetOppositeBranchCondition(CCode);
6667 CC = DAG.getConstant(CCode, MVT::i8);
6668 Cond = Cond.getOperand(0).getOperand(1);
6674 // Look pass the truncate.
6675 if (Cond.getOpcode() == ISD::TRUNCATE)
6676 Cond = Cond.getOperand(0);
6678 // We know the result of AND is compared against zero. Try to match
6680 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6681 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6682 if (NewSetCC.getNode()) {
6683 CC = NewSetCC.getOperand(0);
6684 Cond = NewSetCC.getOperand(1);
6691 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6692 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6694 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6695 Chain, Dest, CC, Cond);
6699 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6700 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6701 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6702 // that the guard pages used by the OS virtual memory manager are allocated in
6703 // correct sequence.
6705 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6706 SelectionDAG &DAG) const {
6707 assert(Subtarget->isTargetCygMing() &&
6708 "This should be used only on Cygwin/Mingw targets");
6709 DebugLoc dl = Op.getDebugLoc();
6712 SDValue Chain = Op.getOperand(0);
6713 SDValue Size = Op.getOperand(1);
6714 // FIXME: Ensure alignment here
6718 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6720 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6721 Flag = Chain.getValue(1);
6723 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6725 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6726 Flag = Chain.getValue(1);
6728 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6730 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6731 return DAG.getMergeValues(Ops1, 2, dl);
6734 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6735 MachineFunction &MF = DAG.getMachineFunction();
6736 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6738 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6739 DebugLoc dl = Op.getDebugLoc();
6741 if (!Subtarget->is64Bit()) {
6742 // vastart just stores the address of the VarArgsFrameIndex slot into the
6743 // memory location argument.
6744 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6746 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6751 // gp_offset (0 - 6 * 8)
6752 // fp_offset (48 - 48 + 8 * 16)
6753 // overflow_arg_area (point to parameters coming in memory).
6755 SmallVector<SDValue, 8> MemOps;
6756 SDValue FIN = Op.getOperand(1);
6758 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6759 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6761 FIN, SV, 0, false, false, 0);
6762 MemOps.push_back(Store);
6765 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6766 FIN, DAG.getIntPtrConstant(4));
6767 Store = DAG.getStore(Op.getOperand(0), dl,
6768 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6770 FIN, SV, 0, false, false, 0);
6771 MemOps.push_back(Store);
6773 // Store ptr to overflow_arg_area
6774 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6775 FIN, DAG.getIntPtrConstant(4));
6776 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6778 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6780 MemOps.push_back(Store);
6782 // Store ptr to reg_save_area.
6783 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6784 FIN, DAG.getIntPtrConstant(8));
6785 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6787 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6789 MemOps.push_back(Store);
6790 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6791 &MemOps[0], MemOps.size());
6794 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6795 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6796 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6798 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6802 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6803 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6804 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6805 SDValue Chain = Op.getOperand(0);
6806 SDValue DstPtr = Op.getOperand(1);
6807 SDValue SrcPtr = Op.getOperand(2);
6808 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6809 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6810 DebugLoc dl = Op.getDebugLoc();
6812 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6813 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6814 false, DstSV, 0, SrcSV, 0);
6818 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6819 DebugLoc dl = Op.getDebugLoc();
6820 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6822 default: return SDValue(); // Don't custom lower most intrinsics.
6823 // Comparison intrinsics.
6824 case Intrinsic::x86_sse_comieq_ss:
6825 case Intrinsic::x86_sse_comilt_ss:
6826 case Intrinsic::x86_sse_comile_ss:
6827 case Intrinsic::x86_sse_comigt_ss:
6828 case Intrinsic::x86_sse_comige_ss:
6829 case Intrinsic::x86_sse_comineq_ss:
6830 case Intrinsic::x86_sse_ucomieq_ss:
6831 case Intrinsic::x86_sse_ucomilt_ss:
6832 case Intrinsic::x86_sse_ucomile_ss:
6833 case Intrinsic::x86_sse_ucomigt_ss:
6834 case Intrinsic::x86_sse_ucomige_ss:
6835 case Intrinsic::x86_sse_ucomineq_ss:
6836 case Intrinsic::x86_sse2_comieq_sd:
6837 case Intrinsic::x86_sse2_comilt_sd:
6838 case Intrinsic::x86_sse2_comile_sd:
6839 case Intrinsic::x86_sse2_comigt_sd:
6840 case Intrinsic::x86_sse2_comige_sd:
6841 case Intrinsic::x86_sse2_comineq_sd:
6842 case Intrinsic::x86_sse2_ucomieq_sd:
6843 case Intrinsic::x86_sse2_ucomilt_sd:
6844 case Intrinsic::x86_sse2_ucomile_sd:
6845 case Intrinsic::x86_sse2_ucomigt_sd:
6846 case Intrinsic::x86_sse2_ucomige_sd:
6847 case Intrinsic::x86_sse2_ucomineq_sd: {
6849 ISD::CondCode CC = ISD::SETCC_INVALID;
6852 case Intrinsic::x86_sse_comieq_ss:
6853 case Intrinsic::x86_sse2_comieq_sd:
6857 case Intrinsic::x86_sse_comilt_ss:
6858 case Intrinsic::x86_sse2_comilt_sd:
6862 case Intrinsic::x86_sse_comile_ss:
6863 case Intrinsic::x86_sse2_comile_sd:
6867 case Intrinsic::x86_sse_comigt_ss:
6868 case Intrinsic::x86_sse2_comigt_sd:
6872 case Intrinsic::x86_sse_comige_ss:
6873 case Intrinsic::x86_sse2_comige_sd:
6877 case Intrinsic::x86_sse_comineq_ss:
6878 case Intrinsic::x86_sse2_comineq_sd:
6882 case Intrinsic::x86_sse_ucomieq_ss:
6883 case Intrinsic::x86_sse2_ucomieq_sd:
6884 Opc = X86ISD::UCOMI;
6887 case Intrinsic::x86_sse_ucomilt_ss:
6888 case Intrinsic::x86_sse2_ucomilt_sd:
6889 Opc = X86ISD::UCOMI;
6892 case Intrinsic::x86_sse_ucomile_ss:
6893 case Intrinsic::x86_sse2_ucomile_sd:
6894 Opc = X86ISD::UCOMI;
6897 case Intrinsic::x86_sse_ucomigt_ss:
6898 case Intrinsic::x86_sse2_ucomigt_sd:
6899 Opc = X86ISD::UCOMI;
6902 case Intrinsic::x86_sse_ucomige_ss:
6903 case Intrinsic::x86_sse2_ucomige_sd:
6904 Opc = X86ISD::UCOMI;
6907 case Intrinsic::x86_sse_ucomineq_ss:
6908 case Intrinsic::x86_sse2_ucomineq_sd:
6909 Opc = X86ISD::UCOMI;
6914 SDValue LHS = Op.getOperand(1);
6915 SDValue RHS = Op.getOperand(2);
6916 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6917 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6918 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6919 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6920 DAG.getConstant(X86CC, MVT::i8), Cond);
6921 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6923 // ptest intrinsics. The intrinsic these come from are designed to return
6924 // an integer value, not just an instruction so lower it to the ptest
6925 // pattern and a setcc for the result.
6926 case Intrinsic::x86_sse41_ptestz:
6927 case Intrinsic::x86_sse41_ptestc:
6928 case Intrinsic::x86_sse41_ptestnzc:{
6931 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6932 case Intrinsic::x86_sse41_ptestz:
6934 X86CC = X86::COND_E;
6936 case Intrinsic::x86_sse41_ptestc:
6938 X86CC = X86::COND_B;
6940 case Intrinsic::x86_sse41_ptestnzc:
6942 X86CC = X86::COND_A;
6946 SDValue LHS = Op.getOperand(1);
6947 SDValue RHS = Op.getOperand(2);
6948 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6949 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6950 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6951 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6954 // Fix vector shift instructions where the last operand is a non-immediate
6956 case Intrinsic::x86_sse2_pslli_w:
6957 case Intrinsic::x86_sse2_pslli_d:
6958 case Intrinsic::x86_sse2_pslli_q:
6959 case Intrinsic::x86_sse2_psrli_w:
6960 case Intrinsic::x86_sse2_psrli_d:
6961 case Intrinsic::x86_sse2_psrli_q:
6962 case Intrinsic::x86_sse2_psrai_w:
6963 case Intrinsic::x86_sse2_psrai_d:
6964 case Intrinsic::x86_mmx_pslli_w:
6965 case Intrinsic::x86_mmx_pslli_d:
6966 case Intrinsic::x86_mmx_pslli_q:
6967 case Intrinsic::x86_mmx_psrli_w:
6968 case Intrinsic::x86_mmx_psrli_d:
6969 case Intrinsic::x86_mmx_psrli_q:
6970 case Intrinsic::x86_mmx_psrai_w:
6971 case Intrinsic::x86_mmx_psrai_d: {
6972 SDValue ShAmt = Op.getOperand(2);
6973 if (isa<ConstantSDNode>(ShAmt))
6976 unsigned NewIntNo = 0;
6977 EVT ShAmtVT = MVT::v4i32;
6979 case Intrinsic::x86_sse2_pslli_w:
6980 NewIntNo = Intrinsic::x86_sse2_psll_w;
6982 case Intrinsic::x86_sse2_pslli_d:
6983 NewIntNo = Intrinsic::x86_sse2_psll_d;
6985 case Intrinsic::x86_sse2_pslli_q:
6986 NewIntNo = Intrinsic::x86_sse2_psll_q;
6988 case Intrinsic::x86_sse2_psrli_w:
6989 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6991 case Intrinsic::x86_sse2_psrli_d:
6992 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6994 case Intrinsic::x86_sse2_psrli_q:
6995 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6997 case Intrinsic::x86_sse2_psrai_w:
6998 NewIntNo = Intrinsic::x86_sse2_psra_w;
7000 case Intrinsic::x86_sse2_psrai_d:
7001 NewIntNo = Intrinsic::x86_sse2_psra_d;
7004 ShAmtVT = MVT::v2i32;
7006 case Intrinsic::x86_mmx_pslli_w:
7007 NewIntNo = Intrinsic::x86_mmx_psll_w;
7009 case Intrinsic::x86_mmx_pslli_d:
7010 NewIntNo = Intrinsic::x86_mmx_psll_d;
7012 case Intrinsic::x86_mmx_pslli_q:
7013 NewIntNo = Intrinsic::x86_mmx_psll_q;
7015 case Intrinsic::x86_mmx_psrli_w:
7016 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7018 case Intrinsic::x86_mmx_psrli_d:
7019 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7021 case Intrinsic::x86_mmx_psrli_q:
7022 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7024 case Intrinsic::x86_mmx_psrai_w:
7025 NewIntNo = Intrinsic::x86_mmx_psra_w;
7027 case Intrinsic::x86_mmx_psrai_d:
7028 NewIntNo = Intrinsic::x86_mmx_psra_d;
7030 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7036 // The vector shift intrinsics with scalars uses 32b shift amounts but
7037 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7041 ShOps[1] = DAG.getConstant(0, MVT::i32);
7042 if (ShAmtVT == MVT::v4i32) {
7043 ShOps[2] = DAG.getUNDEF(MVT::i32);
7044 ShOps[3] = DAG.getUNDEF(MVT::i32);
7045 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7047 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7050 EVT VT = Op.getValueType();
7051 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7052 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7053 DAG.getConstant(NewIntNo, MVT::i32),
7054 Op.getOperand(1), ShAmt);
7059 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7060 SelectionDAG &DAG) const {
7061 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7062 MFI->setReturnAddressIsTaken(true);
7064 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7065 DebugLoc dl = Op.getDebugLoc();
7068 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7070 DAG.getConstant(TD->getPointerSize(),
7071 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7072 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7073 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7075 NULL, 0, false, false, 0);
7078 // Just load the return address.
7079 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7080 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7081 RetAddrFI, NULL, 0, false, false, 0);
7084 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7085 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7086 MFI->setFrameAddressIsTaken(true);
7088 EVT VT = Op.getValueType();
7089 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7090 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7091 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7092 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7094 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7099 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7100 SelectionDAG &DAG) const {
7101 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7104 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7105 MachineFunction &MF = DAG.getMachineFunction();
7106 SDValue Chain = Op.getOperand(0);
7107 SDValue Offset = Op.getOperand(1);
7108 SDValue Handler = Op.getOperand(2);
7109 DebugLoc dl = Op.getDebugLoc();
7111 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7113 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7115 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7116 DAG.getIntPtrConstant(-TD->getPointerSize()));
7117 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7118 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7119 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7120 MF.getRegInfo().addLiveOut(StoreAddrReg);
7122 return DAG.getNode(X86ISD::EH_RETURN, dl,
7124 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7127 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7128 SelectionDAG &DAG) const {
7129 SDValue Root = Op.getOperand(0);
7130 SDValue Trmp = Op.getOperand(1); // trampoline
7131 SDValue FPtr = Op.getOperand(2); // nested function
7132 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7133 DebugLoc dl = Op.getDebugLoc();
7135 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7137 if (Subtarget->is64Bit()) {
7138 SDValue OutChains[6];
7140 // Large code-model.
7141 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7142 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7144 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7145 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7147 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7149 // Load the pointer to the nested function into R11.
7150 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7151 SDValue Addr = Trmp;
7152 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7153 Addr, TrmpAddr, 0, false, false, 0);
7155 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7156 DAG.getConstant(2, MVT::i64));
7157 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7160 // Load the 'nest' parameter value into R10.
7161 // R10 is specified in X86CallingConv.td
7162 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7163 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7164 DAG.getConstant(10, MVT::i64));
7165 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7166 Addr, TrmpAddr, 10, false, false, 0);
7168 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7169 DAG.getConstant(12, MVT::i64));
7170 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7173 // Jump to the nested function.
7174 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7175 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7176 DAG.getConstant(20, MVT::i64));
7177 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7178 Addr, TrmpAddr, 20, false, false, 0);
7180 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7181 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7182 DAG.getConstant(22, MVT::i64));
7183 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7184 TrmpAddr, 22, false, false, 0);
7187 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7188 return DAG.getMergeValues(Ops, 2, dl);
7190 const Function *Func =
7191 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7192 CallingConv::ID CC = Func->getCallingConv();
7197 llvm_unreachable("Unsupported calling convention");
7198 case CallingConv::C:
7199 case CallingConv::X86_StdCall: {
7200 // Pass 'nest' parameter in ECX.
7201 // Must be kept in sync with X86CallingConv.td
7204 // Check that ECX wasn't needed by an 'inreg' parameter.
7205 const FunctionType *FTy = Func->getFunctionType();
7206 const AttrListPtr &Attrs = Func->getAttributes();
7208 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7209 unsigned InRegCount = 0;
7212 for (FunctionType::param_iterator I = FTy->param_begin(),
7213 E = FTy->param_end(); I != E; ++I, ++Idx)
7214 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7215 // FIXME: should only count parameters that are lowered to integers.
7216 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7218 if (InRegCount > 2) {
7219 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7224 case CallingConv::X86_FastCall:
7225 case CallingConv::X86_ThisCall:
7226 case CallingConv::Fast:
7227 // Pass 'nest' parameter in EAX.
7228 // Must be kept in sync with X86CallingConv.td
7233 SDValue OutChains[4];
7236 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7237 DAG.getConstant(10, MVT::i32));
7238 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7240 // This is storing the opcode for MOV32ri.
7241 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7242 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7243 OutChains[0] = DAG.getStore(Root, dl,
7244 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7245 Trmp, TrmpAddr, 0, false, false, 0);
7247 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7248 DAG.getConstant(1, MVT::i32));
7249 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7252 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7253 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7254 DAG.getConstant(5, MVT::i32));
7255 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7256 TrmpAddr, 5, false, false, 1);
7258 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7259 DAG.getConstant(6, MVT::i32));
7260 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7264 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7265 return DAG.getMergeValues(Ops, 2, dl);
7269 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7270 SelectionDAG &DAG) const {
7272 The rounding mode is in bits 11:10 of FPSR, and has the following
7279 FLT_ROUNDS, on the other hand, expects the following:
7286 To perform the conversion, we do:
7287 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7290 MachineFunction &MF = DAG.getMachineFunction();
7291 const TargetMachine &TM = MF.getTarget();
7292 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7293 unsigned StackAlignment = TFI.getStackAlignment();
7294 EVT VT = Op.getValueType();
7295 DebugLoc dl = Op.getDebugLoc();
7297 // Save FP Control Word to stack slot
7298 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7299 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7301 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7302 DAG.getEntryNode(), StackSlot);
7304 // Load FP Control Word from stack slot
7305 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7308 // Transform as necessary
7310 DAG.getNode(ISD::SRL, dl, MVT::i16,
7311 DAG.getNode(ISD::AND, dl, MVT::i16,
7312 CWD, DAG.getConstant(0x800, MVT::i16)),
7313 DAG.getConstant(11, MVT::i8));
7315 DAG.getNode(ISD::SRL, dl, MVT::i16,
7316 DAG.getNode(ISD::AND, dl, MVT::i16,
7317 CWD, DAG.getConstant(0x400, MVT::i16)),
7318 DAG.getConstant(9, MVT::i8));
7321 DAG.getNode(ISD::AND, dl, MVT::i16,
7322 DAG.getNode(ISD::ADD, dl, MVT::i16,
7323 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7324 DAG.getConstant(1, MVT::i16)),
7325 DAG.getConstant(3, MVT::i16));
7328 return DAG.getNode((VT.getSizeInBits() < 16 ?
7329 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7332 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7333 EVT VT = Op.getValueType();
7335 unsigned NumBits = VT.getSizeInBits();
7336 DebugLoc dl = Op.getDebugLoc();
7338 Op = Op.getOperand(0);
7339 if (VT == MVT::i8) {
7340 // Zero extend to i32 since there is not an i8 bsr.
7342 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7345 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7346 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7347 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7349 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7352 DAG.getConstant(NumBits+NumBits-1, OpVT),
7353 DAG.getConstant(X86::COND_E, MVT::i8),
7356 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7358 // Finally xor with NumBits-1.
7359 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7362 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7366 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7367 EVT VT = Op.getValueType();
7369 unsigned NumBits = VT.getSizeInBits();
7370 DebugLoc dl = Op.getDebugLoc();
7372 Op = Op.getOperand(0);
7373 if (VT == MVT::i8) {
7375 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7378 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7379 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7380 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7382 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7385 DAG.getConstant(NumBits, OpVT),
7386 DAG.getConstant(X86::COND_E, MVT::i8),
7389 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7392 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7396 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7397 EVT VT = Op.getValueType();
7398 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7399 DebugLoc dl = Op.getDebugLoc();
7401 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7402 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7403 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7404 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7405 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7407 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7408 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7409 // return AloBlo + AloBhi + AhiBlo;
7411 SDValue A = Op.getOperand(0);
7412 SDValue B = Op.getOperand(1);
7414 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7415 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7416 A, DAG.getConstant(32, MVT::i32));
7417 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7418 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7419 B, DAG.getConstant(32, MVT::i32));
7420 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7421 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7423 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7424 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7426 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7427 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7429 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7430 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7431 AloBhi, DAG.getConstant(32, MVT::i32));
7432 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7433 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7434 AhiBlo, DAG.getConstant(32, MVT::i32));
7435 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7436 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7441 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7442 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7443 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7444 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7445 // has only one use.
7446 SDNode *N = Op.getNode();
7447 SDValue LHS = N->getOperand(0);
7448 SDValue RHS = N->getOperand(1);
7449 unsigned BaseOp = 0;
7451 DebugLoc dl = Op.getDebugLoc();
7453 switch (Op.getOpcode()) {
7454 default: llvm_unreachable("Unknown ovf instruction!");
7456 // A subtract of one will be selected as a INC. Note that INC doesn't
7457 // set CF, so we can't do this for UADDO.
7458 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7459 if (C->getAPIntValue() == 1) {
7460 BaseOp = X86ISD::INC;
7464 BaseOp = X86ISD::ADD;
7468 BaseOp = X86ISD::ADD;
7472 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7473 // set CF, so we can't do this for USUBO.
7474 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7475 if (C->getAPIntValue() == 1) {
7476 BaseOp = X86ISD::DEC;
7480 BaseOp = X86ISD::SUB;
7484 BaseOp = X86ISD::SUB;
7488 BaseOp = X86ISD::SMUL;
7492 BaseOp = X86ISD::UMUL;
7497 // Also sets EFLAGS.
7498 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7499 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7502 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7503 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7505 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7509 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7510 EVT T = Op.getValueType();
7511 DebugLoc dl = Op.getDebugLoc();
7514 switch(T.getSimpleVT().SimpleTy) {
7516 assert(false && "Invalid value type!");
7517 case MVT::i8: Reg = X86::AL; size = 1; break;
7518 case MVT::i16: Reg = X86::AX; size = 2; break;
7519 case MVT::i32: Reg = X86::EAX; size = 4; break;
7521 assert(Subtarget->is64Bit() && "Node not type legal!");
7522 Reg = X86::RAX; size = 8;
7525 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7526 Op.getOperand(2), SDValue());
7527 SDValue Ops[] = { cpIn.getValue(0),
7530 DAG.getTargetConstant(size, MVT::i8),
7532 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7533 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7535 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7539 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7540 SelectionDAG &DAG) const {
7541 assert(Subtarget->is64Bit() && "Result not type legalized?");
7542 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7543 SDValue TheChain = Op.getOperand(0);
7544 DebugLoc dl = Op.getDebugLoc();
7545 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7546 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7547 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7549 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7550 DAG.getConstant(32, MVT::i8));
7552 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7555 return DAG.getMergeValues(Ops, 2, dl);
7558 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7559 SelectionDAG &DAG) const {
7560 EVT SrcVT = Op.getOperand(0).getValueType();
7561 EVT DstVT = Op.getValueType();
7562 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7563 Subtarget->hasMMX() && !DisableMMX) &&
7564 "Unexpected custom BIT_CONVERT");
7565 assert((DstVT == MVT::i64 ||
7566 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7567 "Unexpected custom BIT_CONVERT");
7568 // i64 <=> MMX conversions are Legal.
7569 if (SrcVT==MVT::i64 && DstVT.isVector())
7571 if (DstVT==MVT::i64 && SrcVT.isVector())
7573 // MMX <=> MMX conversions are Legal.
7574 if (SrcVT.isVector() && DstVT.isVector())
7576 // All other conversions need to be expanded.
7579 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7580 SDNode *Node = Op.getNode();
7581 DebugLoc dl = Node->getDebugLoc();
7582 EVT T = Node->getValueType(0);
7583 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7584 DAG.getConstant(0, T), Node->getOperand(2));
7585 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7586 cast<AtomicSDNode>(Node)->getMemoryVT(),
7587 Node->getOperand(0),
7588 Node->getOperand(1), negOp,
7589 cast<AtomicSDNode>(Node)->getSrcValue(),
7590 cast<AtomicSDNode>(Node)->getAlignment());
7593 /// LowerOperation - Provide custom lowering hooks for some operations.
7595 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7596 switch (Op.getOpcode()) {
7597 default: llvm_unreachable("Should not custom lower this!");
7598 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7599 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7600 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7601 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7602 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7603 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7604 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7605 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7606 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7607 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7608 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7609 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7610 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7611 case ISD::SHL_PARTS:
7612 case ISD::SRA_PARTS:
7613 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7614 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7615 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7616 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7617 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7618 case ISD::FABS: return LowerFABS(Op, DAG);
7619 case ISD::FNEG: return LowerFNEG(Op, DAG);
7620 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7621 case ISD::SETCC: return LowerSETCC(Op, DAG);
7622 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7623 case ISD::SELECT: return LowerSELECT(Op, DAG);
7624 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7625 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7626 case ISD::VASTART: return LowerVASTART(Op, DAG);
7627 case ISD::VAARG: return LowerVAARG(Op, DAG);
7628 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7629 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7630 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7631 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7632 case ISD::FRAME_TO_ARGS_OFFSET:
7633 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7634 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7635 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7636 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7637 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7638 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7639 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7640 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7646 case ISD::UMULO: return LowerXALUO(Op, DAG);
7647 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7648 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7652 void X86TargetLowering::
7653 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7654 SelectionDAG &DAG, unsigned NewOp) const {
7655 EVT T = Node->getValueType(0);
7656 DebugLoc dl = Node->getDebugLoc();
7657 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7659 SDValue Chain = Node->getOperand(0);
7660 SDValue In1 = Node->getOperand(1);
7661 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7662 Node->getOperand(2), DAG.getIntPtrConstant(0));
7663 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7664 Node->getOperand(2), DAG.getIntPtrConstant(1));
7665 SDValue Ops[] = { Chain, In1, In2L, In2H };
7666 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7668 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7669 cast<MemSDNode>(Node)->getMemOperand());
7670 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7671 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7672 Results.push_back(Result.getValue(2));
7675 /// ReplaceNodeResults - Replace a node with an illegal result type
7676 /// with a new node built out of custom code.
7677 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7678 SmallVectorImpl<SDValue>&Results,
7679 SelectionDAG &DAG) const {
7680 DebugLoc dl = N->getDebugLoc();
7681 switch (N->getOpcode()) {
7683 assert(false && "Do not know how to custom type legalize this operation!");
7685 case ISD::FP_TO_SINT: {
7686 std::pair<SDValue,SDValue> Vals =
7687 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7688 SDValue FIST = Vals.first, StackSlot = Vals.second;
7689 if (FIST.getNode() != 0) {
7690 EVT VT = N->getValueType(0);
7691 // Return a load from the stack slot.
7692 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7697 case ISD::READCYCLECOUNTER: {
7698 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7699 SDValue TheChain = N->getOperand(0);
7700 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7701 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7703 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7705 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7706 SDValue Ops[] = { eax, edx };
7707 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7708 Results.push_back(edx.getValue(1));
7711 case ISD::ATOMIC_CMP_SWAP: {
7712 EVT T = N->getValueType(0);
7713 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7714 SDValue cpInL, cpInH;
7715 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7716 DAG.getConstant(0, MVT::i32));
7717 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7718 DAG.getConstant(1, MVT::i32));
7719 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7720 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7722 SDValue swapInL, swapInH;
7723 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7724 DAG.getConstant(0, MVT::i32));
7725 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7726 DAG.getConstant(1, MVT::i32));
7727 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7729 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7730 swapInL.getValue(1));
7731 SDValue Ops[] = { swapInH.getValue(0),
7733 swapInH.getValue(1) };
7734 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7735 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7736 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7737 MVT::i32, Result.getValue(1));
7738 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7739 MVT::i32, cpOutL.getValue(2));
7740 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7741 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7742 Results.push_back(cpOutH.getValue(1));
7745 case ISD::ATOMIC_LOAD_ADD:
7746 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7748 case ISD::ATOMIC_LOAD_AND:
7749 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7751 case ISD::ATOMIC_LOAD_NAND:
7752 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7754 case ISD::ATOMIC_LOAD_OR:
7755 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7757 case ISD::ATOMIC_LOAD_SUB:
7758 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7760 case ISD::ATOMIC_LOAD_XOR:
7761 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7763 case ISD::ATOMIC_SWAP:
7764 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7769 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7771 default: return NULL;
7772 case X86ISD::BSF: return "X86ISD::BSF";
7773 case X86ISD::BSR: return "X86ISD::BSR";
7774 case X86ISD::SHLD: return "X86ISD::SHLD";
7775 case X86ISD::SHRD: return "X86ISD::SHRD";
7776 case X86ISD::FAND: return "X86ISD::FAND";
7777 case X86ISD::FOR: return "X86ISD::FOR";
7778 case X86ISD::FXOR: return "X86ISD::FXOR";
7779 case X86ISD::FSRL: return "X86ISD::FSRL";
7780 case X86ISD::FILD: return "X86ISD::FILD";
7781 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7782 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7783 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7784 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7785 case X86ISD::FLD: return "X86ISD::FLD";
7786 case X86ISD::FST: return "X86ISD::FST";
7787 case X86ISD::CALL: return "X86ISD::CALL";
7788 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7789 case X86ISD::BT: return "X86ISD::BT";
7790 case X86ISD::CMP: return "X86ISD::CMP";
7791 case X86ISD::COMI: return "X86ISD::COMI";
7792 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7793 case X86ISD::SETCC: return "X86ISD::SETCC";
7794 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7795 case X86ISD::CMOV: return "X86ISD::CMOV";
7796 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7797 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7798 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7799 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7800 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7801 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7802 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7803 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7804 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7805 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7806 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7807 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7808 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7809 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7810 case X86ISD::FMAX: return "X86ISD::FMAX";
7811 case X86ISD::FMIN: return "X86ISD::FMIN";
7812 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7813 case X86ISD::FRCP: return "X86ISD::FRCP";
7814 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7815 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
7816 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7817 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7818 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7819 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7820 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7821 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7822 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7823 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7824 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7825 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7826 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7827 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7828 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7829 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7830 case X86ISD::VSHL: return "X86ISD::VSHL";
7831 case X86ISD::VSRL: return "X86ISD::VSRL";
7832 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7833 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7834 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7835 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7836 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7837 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7838 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7839 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7840 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7841 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7842 case X86ISD::ADD: return "X86ISD::ADD";
7843 case X86ISD::SUB: return "X86ISD::SUB";
7844 case X86ISD::SMUL: return "X86ISD::SMUL";
7845 case X86ISD::UMUL: return "X86ISD::UMUL";
7846 case X86ISD::INC: return "X86ISD::INC";
7847 case X86ISD::DEC: return "X86ISD::DEC";
7848 case X86ISD::OR: return "X86ISD::OR";
7849 case X86ISD::XOR: return "X86ISD::XOR";
7850 case X86ISD::AND: return "X86ISD::AND";
7851 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7852 case X86ISD::PTEST: return "X86ISD::PTEST";
7853 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7854 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7858 // isLegalAddressingMode - Return true if the addressing mode represented
7859 // by AM is legal for this target, for a load/store of the specified type.
7860 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7861 const Type *Ty) const {
7862 // X86 supports extremely general addressing modes.
7863 CodeModel::Model M = getTargetMachine().getCodeModel();
7865 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7866 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7871 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7873 // If a reference to this global requires an extra load, we can't fold it.
7874 if (isGlobalStubReference(GVFlags))
7877 // If BaseGV requires a register for the PIC base, we cannot also have a
7878 // BaseReg specified.
7879 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7882 // If lower 4G is not available, then we must use rip-relative addressing.
7883 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7893 // These scales always work.
7898 // These scales are formed with basereg+scalereg. Only accept if there is
7903 default: // Other stuff never works.
7911 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7912 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7914 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7915 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7916 if (NumBits1 <= NumBits2)
7921 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7922 if (!VT1.isInteger() || !VT2.isInteger())
7924 unsigned NumBits1 = VT1.getSizeInBits();
7925 unsigned NumBits2 = VT2.getSizeInBits();
7926 if (NumBits1 <= NumBits2)
7931 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7932 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7933 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7936 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7937 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7938 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7941 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7942 // i16 instructions are longer (0x66 prefix) and potentially slower.
7943 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7946 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7947 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7948 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7949 /// are assumed to be legal.
7951 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7953 // Very little shuffling can be done for 64-bit vectors right now.
7954 if (VT.getSizeInBits() == 64)
7955 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
7957 // FIXME: pshufb, blends, shifts.
7958 return (VT.getVectorNumElements() == 2 ||
7959 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7960 isMOVLMask(M, VT) ||
7961 isSHUFPMask(M, VT) ||
7962 isPSHUFDMask(M, VT) ||
7963 isPSHUFHWMask(M, VT) ||
7964 isPSHUFLWMask(M, VT) ||
7965 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7966 isUNPCKLMask(M, VT) ||
7967 isUNPCKHMask(M, VT) ||
7968 isUNPCKL_v_undef_Mask(M, VT) ||
7969 isUNPCKH_v_undef_Mask(M, VT));
7973 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7975 unsigned NumElts = VT.getVectorNumElements();
7976 // FIXME: This collection of masks seems suspect.
7979 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7980 return (isMOVLMask(Mask, VT) ||
7981 isCommutedMOVLMask(Mask, VT, true) ||
7982 isSHUFPMask(Mask, VT) ||
7983 isCommutedSHUFPMask(Mask, VT));
7988 //===----------------------------------------------------------------------===//
7989 // X86 Scheduler Hooks
7990 //===----------------------------------------------------------------------===//
7992 // private utility function
7994 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7995 MachineBasicBlock *MBB,
8003 TargetRegisterClass *RC,
8004 bool invSrc) const {
8005 // For the atomic bitwise operator, we generate
8008 // ld t1 = [bitinstr.addr]
8009 // op t2 = t1, [bitinstr.val]
8011 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8013 // fallthrough -->nextMBB
8014 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8015 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8016 MachineFunction::iterator MBBIter = MBB;
8019 /// First build the CFG
8020 MachineFunction *F = MBB->getParent();
8021 MachineBasicBlock *thisMBB = MBB;
8022 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8023 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8024 F->insert(MBBIter, newMBB);
8025 F->insert(MBBIter, nextMBB);
8027 // Move all successors to thisMBB to nextMBB
8028 nextMBB->transferSuccessors(thisMBB);
8030 // Update thisMBB to fall through to newMBB
8031 thisMBB->addSuccessor(newMBB);
8033 // newMBB jumps to itself and fall through to nextMBB
8034 newMBB->addSuccessor(nextMBB);
8035 newMBB->addSuccessor(newMBB);
8037 // Insert instructions into newMBB based on incoming instruction
8038 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8039 "unexpected number of operands");
8040 DebugLoc dl = bInstr->getDebugLoc();
8041 MachineOperand& destOper = bInstr->getOperand(0);
8042 MachineOperand* argOpers[2 + X86AddrNumOperands];
8043 int numArgs = bInstr->getNumOperands() - 1;
8044 for (int i=0; i < numArgs; ++i)
8045 argOpers[i] = &bInstr->getOperand(i+1);
8047 // x86 address has 4 operands: base, index, scale, and displacement
8048 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8049 int valArgIndx = lastAddrIndx + 1;
8051 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8052 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8053 for (int i=0; i <= lastAddrIndx; ++i)
8054 (*MIB).addOperand(*argOpers[i]);
8056 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8058 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8063 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8064 assert((argOpers[valArgIndx]->isReg() ||
8065 argOpers[valArgIndx]->isImm()) &&
8067 if (argOpers[valArgIndx]->isReg())
8068 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8070 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8072 (*MIB).addOperand(*argOpers[valArgIndx]);
8074 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8077 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8078 for (int i=0; i <= lastAddrIndx; ++i)
8079 (*MIB).addOperand(*argOpers[i]);
8081 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8082 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8083 bInstr->memoperands_end());
8085 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8089 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8091 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8095 // private utility function: 64 bit atomics on 32 bit host.
8097 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8098 MachineBasicBlock *MBB,
8103 bool invSrc) const {
8104 // For the atomic bitwise operator, we generate
8105 // thisMBB (instructions are in pairs, except cmpxchg8b)
8106 // ld t1,t2 = [bitinstr.addr]
8108 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8109 // op t5, t6 <- out1, out2, [bitinstr.val]
8110 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8111 // mov ECX, EBX <- t5, t6
8112 // mov EAX, EDX <- t1, t2
8113 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8114 // mov t3, t4 <- EAX, EDX
8116 // result in out1, out2
8117 // fallthrough -->nextMBB
8119 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8120 const unsigned LoadOpc = X86::MOV32rm;
8121 const unsigned copyOpc = X86::MOV32rr;
8122 const unsigned NotOpc = X86::NOT32r;
8123 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8124 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8125 MachineFunction::iterator MBBIter = MBB;
8128 /// First build the CFG
8129 MachineFunction *F = MBB->getParent();
8130 MachineBasicBlock *thisMBB = MBB;
8131 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8132 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8133 F->insert(MBBIter, newMBB);
8134 F->insert(MBBIter, nextMBB);
8136 // Move all successors to thisMBB to nextMBB
8137 nextMBB->transferSuccessors(thisMBB);
8139 // Update thisMBB to fall through to newMBB
8140 thisMBB->addSuccessor(newMBB);
8142 // newMBB jumps to itself and fall through to nextMBB
8143 newMBB->addSuccessor(nextMBB);
8144 newMBB->addSuccessor(newMBB);
8146 DebugLoc dl = bInstr->getDebugLoc();
8147 // Insert instructions into newMBB based on incoming instruction
8148 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8149 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8150 "unexpected number of operands");
8151 MachineOperand& dest1Oper = bInstr->getOperand(0);
8152 MachineOperand& dest2Oper = bInstr->getOperand(1);
8153 MachineOperand* argOpers[2 + X86AddrNumOperands];
8154 for (int i=0; i < 2 + X86AddrNumOperands; ++i) {
8155 argOpers[i] = &bInstr->getOperand(i+2);
8157 // We use some of the operands multiple times, so conservatively just
8158 // clear any kill flags that might be present.
8159 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8160 argOpers[i]->setIsKill(false);
8163 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8164 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8166 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8167 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8168 for (int i=0; i <= lastAddrIndx; ++i)
8169 (*MIB).addOperand(*argOpers[i]);
8170 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8171 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8172 // add 4 to displacement.
8173 for (int i=0; i <= lastAddrIndx-2; ++i)
8174 (*MIB).addOperand(*argOpers[i]);
8175 MachineOperand newOp3 = *(argOpers[3]);
8177 newOp3.setImm(newOp3.getImm()+4);
8179 newOp3.setOffset(newOp3.getOffset()+4);
8180 (*MIB).addOperand(newOp3);
8181 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8183 // t3/4 are defined later, at the bottom of the loop
8184 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8185 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8186 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8187 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8188 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8189 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8191 // The subsequent operations should be using the destination registers of
8192 //the PHI instructions.
8194 t1 = F->getRegInfo().createVirtualRegister(RC);
8195 t2 = F->getRegInfo().createVirtualRegister(RC);
8196 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8197 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8199 t1 = dest1Oper.getReg();
8200 t2 = dest2Oper.getReg();
8203 int valArgIndx = lastAddrIndx + 1;
8204 assert((argOpers[valArgIndx]->isReg() ||
8205 argOpers[valArgIndx]->isImm()) &&
8207 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8208 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8209 if (argOpers[valArgIndx]->isReg())
8210 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8212 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8213 if (regOpcL != X86::MOV32rr)
8215 (*MIB).addOperand(*argOpers[valArgIndx]);
8216 assert(argOpers[valArgIndx + 1]->isReg() ==
8217 argOpers[valArgIndx]->isReg());
8218 assert(argOpers[valArgIndx + 1]->isImm() ==
8219 argOpers[valArgIndx]->isImm());
8220 if (argOpers[valArgIndx + 1]->isReg())
8221 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8223 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8224 if (regOpcH != X86::MOV32rr)
8226 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8228 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8230 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8233 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8235 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8238 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8239 for (int i=0; i <= lastAddrIndx; ++i)
8240 (*MIB).addOperand(*argOpers[i]);
8242 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8243 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8244 bInstr->memoperands_end());
8246 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8247 MIB.addReg(X86::EAX);
8248 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8249 MIB.addReg(X86::EDX);
8252 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8254 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8258 // private utility function
8260 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8261 MachineBasicBlock *MBB,
8262 unsigned cmovOpc) const {
8263 // For the atomic min/max operator, we generate
8266 // ld t1 = [min/max.addr]
8267 // mov t2 = [min/max.val]
8269 // cmov[cond] t2 = t1
8271 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8273 // fallthrough -->nextMBB
8275 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8276 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8277 MachineFunction::iterator MBBIter = MBB;
8280 /// First build the CFG
8281 MachineFunction *F = MBB->getParent();
8282 MachineBasicBlock *thisMBB = MBB;
8283 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8284 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8285 F->insert(MBBIter, newMBB);
8286 F->insert(MBBIter, nextMBB);
8288 // Move all successors of thisMBB to nextMBB
8289 nextMBB->transferSuccessors(thisMBB);
8291 // Update thisMBB to fall through to newMBB
8292 thisMBB->addSuccessor(newMBB);
8294 // newMBB jumps to newMBB and fall through to nextMBB
8295 newMBB->addSuccessor(nextMBB);
8296 newMBB->addSuccessor(newMBB);
8298 DebugLoc dl = mInstr->getDebugLoc();
8299 // Insert instructions into newMBB based on incoming instruction
8300 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8301 "unexpected number of operands");
8302 MachineOperand& destOper = mInstr->getOperand(0);
8303 MachineOperand* argOpers[2 + X86AddrNumOperands];
8304 int numArgs = mInstr->getNumOperands() - 1;
8305 for (int i=0; i < numArgs; ++i)
8306 argOpers[i] = &mInstr->getOperand(i+1);
8308 // x86 address has 4 operands: base, index, scale, and displacement
8309 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8310 int valArgIndx = lastAddrIndx + 1;
8312 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8313 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8314 for (int i=0; i <= lastAddrIndx; ++i)
8315 (*MIB).addOperand(*argOpers[i]);
8317 // We only support register and immediate values
8318 assert((argOpers[valArgIndx]->isReg() ||
8319 argOpers[valArgIndx]->isImm()) &&
8322 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8323 if (argOpers[valArgIndx]->isReg())
8324 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8326 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8327 (*MIB).addOperand(*argOpers[valArgIndx]);
8329 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8332 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8337 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8338 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8342 // Cmp and exchange if none has modified the memory location
8343 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8344 for (int i=0; i <= lastAddrIndx; ++i)
8345 (*MIB).addOperand(*argOpers[i]);
8347 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8348 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8349 mInstr->memoperands_end());
8351 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8352 MIB.addReg(X86::EAX);
8355 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8357 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8361 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8362 // all of this code can be replaced with that in the .td file.
8364 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8365 unsigned numArgs, bool memArg) const {
8367 MachineFunction *F = BB->getParent();
8368 DebugLoc dl = MI->getDebugLoc();
8369 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8373 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8375 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8377 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8379 for (unsigned i = 0; i < numArgs; ++i) {
8380 MachineOperand &Op = MI->getOperand(i+1);
8382 if (!(Op.isReg() && Op.isImplicit()))
8386 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8389 F->DeleteMachineInstr(MI);
8395 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8397 MachineBasicBlock *MBB) const {
8398 // Emit code to save XMM registers to the stack. The ABI says that the
8399 // number of registers to save is given in %al, so it's theoretically
8400 // possible to do an indirect jump trick to avoid saving all of them,
8401 // however this code takes a simpler approach and just executes all
8402 // of the stores if %al is non-zero. It's less code, and it's probably
8403 // easier on the hardware branch predictor, and stores aren't all that
8404 // expensive anyway.
8406 // Create the new basic blocks. One block contains all the XMM stores,
8407 // and one block is the final destination regardless of whether any
8408 // stores were performed.
8409 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8410 MachineFunction *F = MBB->getParent();
8411 MachineFunction::iterator MBBIter = MBB;
8413 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8414 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8415 F->insert(MBBIter, XMMSaveMBB);
8416 F->insert(MBBIter, EndMBB);
8419 // Move any original successors of MBB to the end block.
8420 EndMBB->transferSuccessors(MBB);
8421 // The original block will now fall through to the XMM save block.
8422 MBB->addSuccessor(XMMSaveMBB);
8423 // The XMMSaveMBB will fall through to the end block.
8424 XMMSaveMBB->addSuccessor(EndMBB);
8426 // Now add the instructions.
8427 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8428 DebugLoc DL = MI->getDebugLoc();
8430 unsigned CountReg = MI->getOperand(0).getReg();
8431 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8432 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8434 if (!Subtarget->isTargetWin64()) {
8435 // If %al is 0, branch around the XMM save block.
8436 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8437 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8438 MBB->addSuccessor(EndMBB);
8441 // In the XMM save block, save all the XMM argument registers.
8442 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8443 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8444 MachineMemOperand *MMO =
8445 F->getMachineMemOperand(
8446 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8447 MachineMemOperand::MOStore, Offset,
8448 /*Size=*/16, /*Align=*/16);
8449 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8450 .addFrameIndex(RegSaveFrameIndex)
8451 .addImm(/*Scale=*/1)
8452 .addReg(/*IndexReg=*/0)
8453 .addImm(/*Disp=*/Offset)
8454 .addReg(/*Segment=*/0)
8455 .addReg(MI->getOperand(i).getReg())
8456 .addMemOperand(MMO);
8459 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8465 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8466 MachineBasicBlock *BB) const {
8467 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8468 DebugLoc DL = MI->getDebugLoc();
8470 // To "insert" a SELECT_CC instruction, we actually have to insert the
8471 // diamond control-flow pattern. The incoming instruction knows the
8472 // destination vreg to set, the condition code register to branch on, the
8473 // true/false values to select between, and a branch opcode to use.
8474 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8475 MachineFunction::iterator It = BB;
8481 // cmpTY ccX, r1, r2
8483 // fallthrough --> copy0MBB
8484 MachineBasicBlock *thisMBB = BB;
8485 MachineFunction *F = BB->getParent();
8486 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8487 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8489 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8491 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8492 F->insert(It, copy0MBB);
8493 F->insert(It, sinkMBB);
8495 // Update machine-CFG edges by first adding all successors of the current
8496 // block to the new block which will contain the Phi node for the select.
8497 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8498 E = BB->succ_end(); I != E; ++I)
8499 sinkMBB->addSuccessor(*I);
8501 // Next, remove all successors of the current block, and add the true
8502 // and fallthrough blocks as its successors.
8503 while (!BB->succ_empty())
8504 BB->removeSuccessor(BB->succ_begin());
8506 // Add the true and fallthrough blocks as its successors.
8507 BB->addSuccessor(copy0MBB);
8508 BB->addSuccessor(sinkMBB);
8510 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8511 // live into the sink and copy blocks.
8512 const MachineFunction *MF = BB->getParent();
8513 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8514 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8515 const MachineInstr *Term = BB->getFirstTerminator();
8517 for (unsigned I = 0, E = Term->getNumOperands(); I != E; ++I) {
8518 const MachineOperand &MO = Term->getOperand(I);
8519 if (!MO.isReg() || MO.isKill() || MO.isDead()) continue;
8520 unsigned Reg = MO.getReg();
8521 if (Reg != X86::EFLAGS) continue;
8522 copy0MBB->addLiveIn(Reg);
8523 sinkMBB->addLiveIn(Reg);
8527 // %FalseValue = ...
8528 // # fallthrough to sinkMBB
8529 copy0MBB->addSuccessor(sinkMBB);
8532 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8534 BuildMI(sinkMBB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8535 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8536 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8538 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8543 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8544 MachineBasicBlock *BB) const {
8545 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8546 DebugLoc DL = MI->getDebugLoc();
8547 MachineFunction *F = BB->getParent();
8549 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8550 // non-trivial part is impdef of ESP.
8551 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8554 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8555 .addExternalSymbol("_alloca")
8556 .addReg(X86::EAX, RegState::Implicit)
8557 .addReg(X86::ESP, RegState::Implicit)
8558 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8559 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8561 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8566 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8567 MachineBasicBlock *BB) const {
8568 // This is pretty easy. We're taking the value that we received from
8569 // our load from the relocation, sticking it in either RDI (x86-64)
8570 // or EAX and doing an indirect call. The return value will then
8571 // be in the normal return register.
8572 const X86InstrInfo *TII
8573 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8574 DebugLoc DL = MI->getDebugLoc();
8575 MachineFunction *F = BB->getParent();
8577 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8579 if (Subtarget->is64Bit()) {
8580 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV64rm), X86::RDI)
8582 .addImm(0).addReg(0)
8583 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8584 MI->getOperand(3).getTargetFlags())
8586 MIB = BuildMI(BB, DL, TII->get(X86::CALL64m));
8587 addDirectMem(MIB, X86::RDI).addReg(0);
8588 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8589 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV32rm), X86::EAX)
8591 .addImm(0).addReg(0)
8592 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8593 MI->getOperand(3).getTargetFlags())
8595 MIB = BuildMI(BB, DL, TII->get(X86::CALL32m));
8596 addDirectMem(MIB, X86::EAX).addReg(0);
8598 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV32rm), X86::EAX)
8599 .addReg(TII->getGlobalBaseReg(F))
8600 .addImm(0).addReg(0)
8601 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8602 MI->getOperand(3).getTargetFlags())
8604 MIB = BuildMI(BB, DL, TII->get(X86::CALL32m));
8605 addDirectMem(MIB, X86::EAX).addReg(0);
8608 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8613 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8614 MachineBasicBlock *BB) const {
8615 switch (MI->getOpcode()) {
8616 default: assert(false && "Unexpected instr type to insert");
8617 case X86::MINGW_ALLOCA:
8618 return EmitLoweredMingwAlloca(MI, BB);
8619 case X86::TLSCall_32:
8620 case X86::TLSCall_64:
8621 return EmitLoweredTLSCall(MI, BB);
8623 case X86::CMOV_V1I64:
8624 case X86::CMOV_FR32:
8625 case X86::CMOV_FR64:
8626 case X86::CMOV_V4F32:
8627 case X86::CMOV_V2F64:
8628 case X86::CMOV_V2I64:
8629 case X86::CMOV_GR16:
8630 case X86::CMOV_GR32:
8631 case X86::CMOV_RFP32:
8632 case X86::CMOV_RFP64:
8633 case X86::CMOV_RFP80:
8634 return EmitLoweredSelect(MI, BB);
8636 case X86::FP32_TO_INT16_IN_MEM:
8637 case X86::FP32_TO_INT32_IN_MEM:
8638 case X86::FP32_TO_INT64_IN_MEM:
8639 case X86::FP64_TO_INT16_IN_MEM:
8640 case X86::FP64_TO_INT32_IN_MEM:
8641 case X86::FP64_TO_INT64_IN_MEM:
8642 case X86::FP80_TO_INT16_IN_MEM:
8643 case X86::FP80_TO_INT32_IN_MEM:
8644 case X86::FP80_TO_INT64_IN_MEM: {
8645 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8646 DebugLoc DL = MI->getDebugLoc();
8648 // Change the floating point control register to use "round towards zero"
8649 // mode when truncating to an integer value.
8650 MachineFunction *F = BB->getParent();
8651 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8652 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8654 // Load the old value of the high byte of the control word...
8656 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8657 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8660 // Set the high part to be round to zero...
8661 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8664 // Reload the modified control word now...
8665 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8667 // Restore the memory image of control word to original value
8668 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8671 // Get the X86 opcode to use.
8673 switch (MI->getOpcode()) {
8674 default: llvm_unreachable("illegal opcode!");
8675 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8676 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8677 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8678 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8679 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8680 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8681 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8682 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8683 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8687 MachineOperand &Op = MI->getOperand(0);
8689 AM.BaseType = X86AddressMode::RegBase;
8690 AM.Base.Reg = Op.getReg();
8692 AM.BaseType = X86AddressMode::FrameIndexBase;
8693 AM.Base.FrameIndex = Op.getIndex();
8695 Op = MI->getOperand(1);
8697 AM.Scale = Op.getImm();
8698 Op = MI->getOperand(2);
8700 AM.IndexReg = Op.getImm();
8701 Op = MI->getOperand(3);
8702 if (Op.isGlobal()) {
8703 AM.GV = Op.getGlobal();
8705 AM.Disp = Op.getImm();
8707 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8708 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8710 // Reload the original control word now.
8711 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8713 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8716 // String/text processing lowering.
8717 case X86::PCMPISTRM128REG:
8718 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8719 case X86::PCMPISTRM128MEM:
8720 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8721 case X86::PCMPESTRM128REG:
8722 return EmitPCMP(MI, BB, 5, false /* in mem */);
8723 case X86::PCMPESTRM128MEM:
8724 return EmitPCMP(MI, BB, 5, true /* in mem */);
8727 case X86::ATOMAND32:
8728 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8729 X86::AND32ri, X86::MOV32rm,
8730 X86::LCMPXCHG32, X86::MOV32rr,
8731 X86::NOT32r, X86::EAX,
8732 X86::GR32RegisterClass);
8734 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8735 X86::OR32ri, X86::MOV32rm,
8736 X86::LCMPXCHG32, X86::MOV32rr,
8737 X86::NOT32r, X86::EAX,
8738 X86::GR32RegisterClass);
8739 case X86::ATOMXOR32:
8740 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8741 X86::XOR32ri, X86::MOV32rm,
8742 X86::LCMPXCHG32, X86::MOV32rr,
8743 X86::NOT32r, X86::EAX,
8744 X86::GR32RegisterClass);
8745 case X86::ATOMNAND32:
8746 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8747 X86::AND32ri, X86::MOV32rm,
8748 X86::LCMPXCHG32, X86::MOV32rr,
8749 X86::NOT32r, X86::EAX,
8750 X86::GR32RegisterClass, true);
8751 case X86::ATOMMIN32:
8752 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8753 case X86::ATOMMAX32:
8754 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8755 case X86::ATOMUMIN32:
8756 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8757 case X86::ATOMUMAX32:
8758 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8760 case X86::ATOMAND16:
8761 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8762 X86::AND16ri, X86::MOV16rm,
8763 X86::LCMPXCHG16, X86::MOV16rr,
8764 X86::NOT16r, X86::AX,
8765 X86::GR16RegisterClass);
8767 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8768 X86::OR16ri, X86::MOV16rm,
8769 X86::LCMPXCHG16, X86::MOV16rr,
8770 X86::NOT16r, X86::AX,
8771 X86::GR16RegisterClass);
8772 case X86::ATOMXOR16:
8773 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8774 X86::XOR16ri, X86::MOV16rm,
8775 X86::LCMPXCHG16, X86::MOV16rr,
8776 X86::NOT16r, X86::AX,
8777 X86::GR16RegisterClass);
8778 case X86::ATOMNAND16:
8779 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8780 X86::AND16ri, X86::MOV16rm,
8781 X86::LCMPXCHG16, X86::MOV16rr,
8782 X86::NOT16r, X86::AX,
8783 X86::GR16RegisterClass, true);
8784 case X86::ATOMMIN16:
8785 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8786 case X86::ATOMMAX16:
8787 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8788 case X86::ATOMUMIN16:
8789 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8790 case X86::ATOMUMAX16:
8791 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8794 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8795 X86::AND8ri, X86::MOV8rm,
8796 X86::LCMPXCHG8, X86::MOV8rr,
8797 X86::NOT8r, X86::AL,
8798 X86::GR8RegisterClass);
8800 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8801 X86::OR8ri, X86::MOV8rm,
8802 X86::LCMPXCHG8, X86::MOV8rr,
8803 X86::NOT8r, X86::AL,
8804 X86::GR8RegisterClass);
8806 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8807 X86::XOR8ri, X86::MOV8rm,
8808 X86::LCMPXCHG8, X86::MOV8rr,
8809 X86::NOT8r, X86::AL,
8810 X86::GR8RegisterClass);
8811 case X86::ATOMNAND8:
8812 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8813 X86::AND8ri, X86::MOV8rm,
8814 X86::LCMPXCHG8, X86::MOV8rr,
8815 X86::NOT8r, X86::AL,
8816 X86::GR8RegisterClass, true);
8817 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8818 // This group is for 64-bit host.
8819 case X86::ATOMAND64:
8820 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8821 X86::AND64ri32, X86::MOV64rm,
8822 X86::LCMPXCHG64, X86::MOV64rr,
8823 X86::NOT64r, X86::RAX,
8824 X86::GR64RegisterClass);
8826 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8827 X86::OR64ri32, X86::MOV64rm,
8828 X86::LCMPXCHG64, X86::MOV64rr,
8829 X86::NOT64r, X86::RAX,
8830 X86::GR64RegisterClass);
8831 case X86::ATOMXOR64:
8832 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8833 X86::XOR64ri32, X86::MOV64rm,
8834 X86::LCMPXCHG64, X86::MOV64rr,
8835 X86::NOT64r, X86::RAX,
8836 X86::GR64RegisterClass);
8837 case X86::ATOMNAND64:
8838 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8839 X86::AND64ri32, X86::MOV64rm,
8840 X86::LCMPXCHG64, X86::MOV64rr,
8841 X86::NOT64r, X86::RAX,
8842 X86::GR64RegisterClass, true);
8843 case X86::ATOMMIN64:
8844 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8845 case X86::ATOMMAX64:
8846 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8847 case X86::ATOMUMIN64:
8848 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8849 case X86::ATOMUMAX64:
8850 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8852 // This group does 64-bit operations on a 32-bit host.
8853 case X86::ATOMAND6432:
8854 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8855 X86::AND32rr, X86::AND32rr,
8856 X86::AND32ri, X86::AND32ri,
8858 case X86::ATOMOR6432:
8859 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8860 X86::OR32rr, X86::OR32rr,
8861 X86::OR32ri, X86::OR32ri,
8863 case X86::ATOMXOR6432:
8864 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8865 X86::XOR32rr, X86::XOR32rr,
8866 X86::XOR32ri, X86::XOR32ri,
8868 case X86::ATOMNAND6432:
8869 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8870 X86::AND32rr, X86::AND32rr,
8871 X86::AND32ri, X86::AND32ri,
8873 case X86::ATOMADD6432:
8874 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8875 X86::ADD32rr, X86::ADC32rr,
8876 X86::ADD32ri, X86::ADC32ri,
8878 case X86::ATOMSUB6432:
8879 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8880 X86::SUB32rr, X86::SBB32rr,
8881 X86::SUB32ri, X86::SBB32ri,
8883 case X86::ATOMSWAP6432:
8884 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8885 X86::MOV32rr, X86::MOV32rr,
8886 X86::MOV32ri, X86::MOV32ri,
8888 case X86::VASTART_SAVE_XMM_REGS:
8889 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8893 //===----------------------------------------------------------------------===//
8894 // X86 Optimization Hooks
8895 //===----------------------------------------------------------------------===//
8897 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8901 const SelectionDAG &DAG,
8902 unsigned Depth) const {
8903 unsigned Opc = Op.getOpcode();
8904 assert((Opc >= ISD::BUILTIN_OP_END ||
8905 Opc == ISD::INTRINSIC_WO_CHAIN ||
8906 Opc == ISD::INTRINSIC_W_CHAIN ||
8907 Opc == ISD::INTRINSIC_VOID) &&
8908 "Should use MaskedValueIsZero if you don't know whether Op"
8909 " is a target node!");
8911 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8923 // These nodes' second result is a boolean.
8924 if (Op.getResNo() == 0)
8928 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8929 Mask.getBitWidth() - 1);
8934 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8935 /// node is a GlobalAddress + offset.
8936 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8937 const GlobalValue* &GA,
8938 int64_t &Offset) const {
8939 if (N->getOpcode() == X86ISD::Wrapper) {
8940 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8941 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8942 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8946 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8949 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8950 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8951 /// if the load addresses are consecutive, non-overlapping, and in the right
8953 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8954 const TargetLowering &TLI) {
8955 DebugLoc dl = N->getDebugLoc();
8956 EVT VT = N->getValueType(0);
8957 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8959 if (VT.getSizeInBits() != 128)
8962 SmallVector<SDValue, 16> Elts;
8963 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8964 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8966 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8969 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8970 /// and convert it from being a bunch of shuffles and extracts to a simple
8971 /// store and scalar loads to extract the elements.
8972 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8973 const TargetLowering &TLI) {
8974 SDValue InputVector = N->getOperand(0);
8976 // Only operate on vectors of 4 elements, where the alternative shuffling
8977 // gets to be more expensive.
8978 if (InputVector.getValueType() != MVT::v4i32)
8981 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8982 // single use which is a sign-extend or zero-extend, and all elements are
8984 SmallVector<SDNode *, 4> Uses;
8985 unsigned ExtractedElements = 0;
8986 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8987 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8988 if (UI.getUse().getResNo() != InputVector.getResNo())
8991 SDNode *Extract = *UI;
8992 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8995 if (Extract->getValueType(0) != MVT::i32)
8997 if (!Extract->hasOneUse())
8999 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9000 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9002 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9005 // Record which element was extracted.
9006 ExtractedElements |=
9007 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9009 Uses.push_back(Extract);
9012 // If not all the elements were used, this may not be worthwhile.
9013 if (ExtractedElements != 15)
9016 // Ok, we've now decided to do the transformation.
9017 DebugLoc dl = InputVector.getDebugLoc();
9019 // Store the value to a temporary stack slot.
9020 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9021 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
9024 // Replace each use (extract) with a load of the appropriate element.
9025 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9026 UE = Uses.end(); UI != UE; ++UI) {
9027 SDNode *Extract = *UI;
9029 // Compute the element's address.
9030 SDValue Idx = Extract->getOperand(1);
9032 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9033 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9034 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9036 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
9039 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
9040 NULL, 0, false, false, 0);
9042 // Replace the exact with the load.
9043 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9046 // The replacement was made in place; don't return anything.
9050 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9051 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9052 const X86Subtarget *Subtarget) {
9053 DebugLoc DL = N->getDebugLoc();
9054 SDValue Cond = N->getOperand(0);
9055 // Get the LHS/RHS of the select.
9056 SDValue LHS = N->getOperand(1);
9057 SDValue RHS = N->getOperand(2);
9059 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9060 // instructions match the semantics of the common C idiom x<y?x:y but not
9061 // x<=y?x:y, because of how they handle negative zero (which can be
9062 // ignored in unsafe-math mode).
9063 if (Subtarget->hasSSE2() &&
9064 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9065 Cond.getOpcode() == ISD::SETCC) {
9066 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9068 unsigned Opcode = 0;
9069 // Check for x CC y ? x : y.
9070 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9071 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9075 // Converting this to a min would handle NaNs incorrectly, and swapping
9076 // the operands would cause it to handle comparisons between positive
9077 // and negative zero incorrectly.
9078 if (!FiniteOnlyFPMath() &&
9079 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9080 if (!UnsafeFPMath &&
9081 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9083 std::swap(LHS, RHS);
9085 Opcode = X86ISD::FMIN;
9088 // Converting this to a min would handle comparisons between positive
9089 // and negative zero incorrectly.
9090 if (!UnsafeFPMath &&
9091 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9093 Opcode = X86ISD::FMIN;
9096 // Converting this to a min would handle both negative zeros and NaNs
9097 // incorrectly, but we can swap the operands to fix both.
9098 std::swap(LHS, RHS);
9102 Opcode = X86ISD::FMIN;
9106 // Converting this to a max would handle comparisons between positive
9107 // and negative zero incorrectly.
9108 if (!UnsafeFPMath &&
9109 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9111 Opcode = X86ISD::FMAX;
9114 // Converting this to a max would handle NaNs incorrectly, and swapping
9115 // the operands would cause it to handle comparisons between positive
9116 // and negative zero incorrectly.
9117 if (!FiniteOnlyFPMath() &&
9118 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9119 if (!UnsafeFPMath &&
9120 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9122 std::swap(LHS, RHS);
9124 Opcode = X86ISD::FMAX;
9127 // Converting this to a max would handle both negative zeros and NaNs
9128 // incorrectly, but we can swap the operands to fix both.
9129 std::swap(LHS, RHS);
9133 Opcode = X86ISD::FMAX;
9136 // Check for x CC y ? y : x -- a min/max with reversed arms.
9137 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9138 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9142 // Converting this to a min would handle comparisons between positive
9143 // and negative zero incorrectly, and swapping the operands would
9144 // cause it to handle NaNs incorrectly.
9145 if (!UnsafeFPMath &&
9146 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9147 if (!FiniteOnlyFPMath() &&
9148 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9150 std::swap(LHS, RHS);
9152 Opcode = X86ISD::FMIN;
9155 // Converting this to a min would handle NaNs incorrectly.
9156 if (!UnsafeFPMath &&
9157 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9159 Opcode = X86ISD::FMIN;
9162 // Converting this to a min would handle both negative zeros and NaNs
9163 // incorrectly, but we can swap the operands to fix both.
9164 std::swap(LHS, RHS);
9168 Opcode = X86ISD::FMIN;
9172 // Converting this to a max would handle NaNs incorrectly.
9173 if (!FiniteOnlyFPMath() &&
9174 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9176 Opcode = X86ISD::FMAX;
9179 // Converting this to a max would handle comparisons between positive
9180 // and negative zero incorrectly, and swapping the operands would
9181 // cause it to handle NaNs incorrectly.
9182 if (!UnsafeFPMath &&
9183 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9184 if (!FiniteOnlyFPMath() &&
9185 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9187 std::swap(LHS, RHS);
9189 Opcode = X86ISD::FMAX;
9192 // Converting this to a max would handle both negative zeros and NaNs
9193 // incorrectly, but we can swap the operands to fix both.
9194 std::swap(LHS, RHS);
9198 Opcode = X86ISD::FMAX;
9204 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9207 // If this is a select between two integer constants, try to do some
9209 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9210 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9211 // Don't do this for crazy integer types.
9212 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9213 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9214 // so that TrueC (the true value) is larger than FalseC.
9215 bool NeedsCondInvert = false;
9217 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9218 // Efficiently invertible.
9219 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9220 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9221 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9222 NeedsCondInvert = true;
9223 std::swap(TrueC, FalseC);
9226 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9227 if (FalseC->getAPIntValue() == 0 &&
9228 TrueC->getAPIntValue().isPowerOf2()) {
9229 if (NeedsCondInvert) // Invert the condition if needed.
9230 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9231 DAG.getConstant(1, Cond.getValueType()));
9233 // Zero extend the condition if needed.
9234 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9236 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9237 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9238 DAG.getConstant(ShAmt, MVT::i8));
9241 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9242 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9243 if (NeedsCondInvert) // Invert the condition if needed.
9244 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9245 DAG.getConstant(1, Cond.getValueType()));
9247 // Zero extend the condition if needed.
9248 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9249 FalseC->getValueType(0), Cond);
9250 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9251 SDValue(FalseC, 0));
9254 // Optimize cases that will turn into an LEA instruction. This requires
9255 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9256 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9257 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9258 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9260 bool isFastMultiplier = false;
9262 switch ((unsigned char)Diff) {
9264 case 1: // result = add base, cond
9265 case 2: // result = lea base( , cond*2)
9266 case 3: // result = lea base(cond, cond*2)
9267 case 4: // result = lea base( , cond*4)
9268 case 5: // result = lea base(cond, cond*4)
9269 case 8: // result = lea base( , cond*8)
9270 case 9: // result = lea base(cond, cond*8)
9271 isFastMultiplier = true;
9276 if (isFastMultiplier) {
9277 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9278 if (NeedsCondInvert) // Invert the condition if needed.
9279 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9280 DAG.getConstant(1, Cond.getValueType()));
9282 // Zero extend the condition if needed.
9283 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9285 // Scale the condition by the difference.
9287 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9288 DAG.getConstant(Diff, Cond.getValueType()));
9290 // Add the base if non-zero.
9291 if (FalseC->getAPIntValue() != 0)
9292 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9293 SDValue(FalseC, 0));
9303 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9304 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9305 TargetLowering::DAGCombinerInfo &DCI) {
9306 DebugLoc DL = N->getDebugLoc();
9308 // If the flag operand isn't dead, don't touch this CMOV.
9309 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9312 // If this is a select between two integer constants, try to do some
9313 // optimizations. Note that the operands are ordered the opposite of SELECT
9315 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9316 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9317 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9318 // larger than FalseC (the false value).
9319 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9321 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9322 CC = X86::GetOppositeBranchCondition(CC);
9323 std::swap(TrueC, FalseC);
9326 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9327 // This is efficient for any integer data type (including i8/i16) and
9329 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9330 SDValue Cond = N->getOperand(3);
9331 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9332 DAG.getConstant(CC, MVT::i8), Cond);
9334 // Zero extend the condition if needed.
9335 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9337 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9338 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9339 DAG.getConstant(ShAmt, MVT::i8));
9340 if (N->getNumValues() == 2) // Dead flag value?
9341 return DCI.CombineTo(N, Cond, SDValue());
9345 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9346 // for any integer data type, including i8/i16.
9347 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9348 SDValue Cond = N->getOperand(3);
9349 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9350 DAG.getConstant(CC, MVT::i8), Cond);
9352 // Zero extend the condition if needed.
9353 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9354 FalseC->getValueType(0), Cond);
9355 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9356 SDValue(FalseC, 0));
9358 if (N->getNumValues() == 2) // Dead flag value?
9359 return DCI.CombineTo(N, Cond, SDValue());
9363 // Optimize cases that will turn into an LEA instruction. This requires
9364 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9365 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9366 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9367 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9369 bool isFastMultiplier = false;
9371 switch ((unsigned char)Diff) {
9373 case 1: // result = add base, cond
9374 case 2: // result = lea base( , cond*2)
9375 case 3: // result = lea base(cond, cond*2)
9376 case 4: // result = lea base( , cond*4)
9377 case 5: // result = lea base(cond, cond*4)
9378 case 8: // result = lea base( , cond*8)
9379 case 9: // result = lea base(cond, cond*8)
9380 isFastMultiplier = true;
9385 if (isFastMultiplier) {
9386 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9387 SDValue Cond = N->getOperand(3);
9388 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9389 DAG.getConstant(CC, MVT::i8), Cond);
9390 // Zero extend the condition if needed.
9391 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9393 // Scale the condition by the difference.
9395 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9396 DAG.getConstant(Diff, Cond.getValueType()));
9398 // Add the base if non-zero.
9399 if (FalseC->getAPIntValue() != 0)
9400 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9401 SDValue(FalseC, 0));
9402 if (N->getNumValues() == 2) // Dead flag value?
9403 return DCI.CombineTo(N, Cond, SDValue());
9413 /// PerformMulCombine - Optimize a single multiply with constant into two
9414 /// in order to implement it with two cheaper instructions, e.g.
9415 /// LEA + SHL, LEA + LEA.
9416 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9417 TargetLowering::DAGCombinerInfo &DCI) {
9418 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9421 EVT VT = N->getValueType(0);
9425 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9428 uint64_t MulAmt = C->getZExtValue();
9429 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9432 uint64_t MulAmt1 = 0;
9433 uint64_t MulAmt2 = 0;
9434 if ((MulAmt % 9) == 0) {
9436 MulAmt2 = MulAmt / 9;
9437 } else if ((MulAmt % 5) == 0) {
9439 MulAmt2 = MulAmt / 5;
9440 } else if ((MulAmt % 3) == 0) {
9442 MulAmt2 = MulAmt / 3;
9445 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9446 DebugLoc DL = N->getDebugLoc();
9448 if (isPowerOf2_64(MulAmt2) &&
9449 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9450 // If second multiplifer is pow2, issue it first. We want the multiply by
9451 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9453 std::swap(MulAmt1, MulAmt2);
9456 if (isPowerOf2_64(MulAmt1))
9457 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9458 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9460 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9461 DAG.getConstant(MulAmt1, VT));
9463 if (isPowerOf2_64(MulAmt2))
9464 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9465 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9467 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9468 DAG.getConstant(MulAmt2, VT));
9470 // Do not add new nodes to DAG combiner worklist.
9471 DCI.CombineTo(N, NewMul, false);
9476 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9477 SDValue N0 = N->getOperand(0);
9478 SDValue N1 = N->getOperand(1);
9479 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9480 EVT VT = N0.getValueType();
9482 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9483 // since the result of setcc_c is all zero's or all ones.
9484 if (N1C && N0.getOpcode() == ISD::AND &&
9485 N0.getOperand(1).getOpcode() == ISD::Constant) {
9486 SDValue N00 = N0.getOperand(0);
9487 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9488 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9489 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9490 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9491 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9492 APInt ShAmt = N1C->getAPIntValue();
9493 Mask = Mask.shl(ShAmt);
9495 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9496 N00, DAG.getConstant(Mask, VT));
9503 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9505 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9506 const X86Subtarget *Subtarget) {
9507 EVT VT = N->getValueType(0);
9508 if (!VT.isVector() && VT.isInteger() &&
9509 N->getOpcode() == ISD::SHL)
9510 return PerformSHLCombine(N, DAG);
9512 // On X86 with SSE2 support, we can transform this to a vector shift if
9513 // all elements are shifted by the same amount. We can't do this in legalize
9514 // because the a constant vector is typically transformed to a constant pool
9515 // so we have no knowledge of the shift amount.
9516 if (!Subtarget->hasSSE2())
9519 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9522 SDValue ShAmtOp = N->getOperand(1);
9523 EVT EltVT = VT.getVectorElementType();
9524 DebugLoc DL = N->getDebugLoc();
9525 SDValue BaseShAmt = SDValue();
9526 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9527 unsigned NumElts = VT.getVectorNumElements();
9529 for (; i != NumElts; ++i) {
9530 SDValue Arg = ShAmtOp.getOperand(i);
9531 if (Arg.getOpcode() == ISD::UNDEF) continue;
9535 for (; i != NumElts; ++i) {
9536 SDValue Arg = ShAmtOp.getOperand(i);
9537 if (Arg.getOpcode() == ISD::UNDEF) continue;
9538 if (Arg != BaseShAmt) {
9542 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9543 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9544 SDValue InVec = ShAmtOp.getOperand(0);
9545 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9546 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9548 for (; i != NumElts; ++i) {
9549 SDValue Arg = InVec.getOperand(i);
9550 if (Arg.getOpcode() == ISD::UNDEF) continue;
9554 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9555 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9556 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9557 if (C->getZExtValue() == SplatIdx)
9558 BaseShAmt = InVec.getOperand(1);
9561 if (BaseShAmt.getNode() == 0)
9562 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9563 DAG.getIntPtrConstant(0));
9567 // The shift amount is an i32.
9568 if (EltVT.bitsGT(MVT::i32))
9569 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9570 else if (EltVT.bitsLT(MVT::i32))
9571 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9573 // The shift amount is identical so we can do a vector shift.
9574 SDValue ValOp = N->getOperand(0);
9575 switch (N->getOpcode()) {
9577 llvm_unreachable("Unknown shift opcode!");
9580 if (VT == MVT::v2i64)
9581 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9582 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9584 if (VT == MVT::v4i32)
9585 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9586 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9588 if (VT == MVT::v8i16)
9589 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9590 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9594 if (VT == MVT::v4i32)
9595 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9596 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9598 if (VT == MVT::v8i16)
9599 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9600 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9604 if (VT == MVT::v2i64)
9605 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9606 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9608 if (VT == MVT::v4i32)
9609 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9610 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9612 if (VT == MVT::v8i16)
9613 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9614 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9621 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9622 TargetLowering::DAGCombinerInfo &DCI,
9623 const X86Subtarget *Subtarget) {
9624 if (DCI.isBeforeLegalizeOps())
9627 EVT VT = N->getValueType(0);
9628 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9631 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9632 SDValue N0 = N->getOperand(0);
9633 SDValue N1 = N->getOperand(1);
9634 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9636 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9638 if (!N0.hasOneUse() || !N1.hasOneUse())
9641 SDValue ShAmt0 = N0.getOperand(1);
9642 if (ShAmt0.getValueType() != MVT::i8)
9644 SDValue ShAmt1 = N1.getOperand(1);
9645 if (ShAmt1.getValueType() != MVT::i8)
9647 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9648 ShAmt0 = ShAmt0.getOperand(0);
9649 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9650 ShAmt1 = ShAmt1.getOperand(0);
9652 DebugLoc DL = N->getDebugLoc();
9653 unsigned Opc = X86ISD::SHLD;
9654 SDValue Op0 = N0.getOperand(0);
9655 SDValue Op1 = N1.getOperand(0);
9656 if (ShAmt0.getOpcode() == ISD::SUB) {
9658 std::swap(Op0, Op1);
9659 std::swap(ShAmt0, ShAmt1);
9662 unsigned Bits = VT.getSizeInBits();
9663 if (ShAmt1.getOpcode() == ISD::SUB) {
9664 SDValue Sum = ShAmt1.getOperand(0);
9665 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9666 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
9667 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
9668 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
9669 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
9670 return DAG.getNode(Opc, DL, VT,
9672 DAG.getNode(ISD::TRUNCATE, DL,
9675 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9676 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9678 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9679 return DAG.getNode(Opc, DL, VT,
9680 N0.getOperand(0), N1.getOperand(0),
9681 DAG.getNode(ISD::TRUNCATE, DL,
9688 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9689 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9690 const X86Subtarget *Subtarget) {
9691 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9692 // the FP state in cases where an emms may be missing.
9693 // A preferable solution to the general problem is to figure out the right
9694 // places to insert EMMS. This qualifies as a quick hack.
9696 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9697 StoreSDNode *St = cast<StoreSDNode>(N);
9698 EVT VT = St->getValue().getValueType();
9699 if (VT.getSizeInBits() != 64)
9702 const Function *F = DAG.getMachineFunction().getFunction();
9703 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9704 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9705 && Subtarget->hasSSE2();
9706 if ((VT.isVector() ||
9707 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9708 isa<LoadSDNode>(St->getValue()) &&
9709 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9710 St->getChain().hasOneUse() && !St->isVolatile()) {
9711 SDNode* LdVal = St->getValue().getNode();
9713 int TokenFactorIndex = -1;
9714 SmallVector<SDValue, 8> Ops;
9715 SDNode* ChainVal = St->getChain().getNode();
9716 // Must be a store of a load. We currently handle two cases: the load
9717 // is a direct child, and it's under an intervening TokenFactor. It is
9718 // possible to dig deeper under nested TokenFactors.
9719 if (ChainVal == LdVal)
9720 Ld = cast<LoadSDNode>(St->getChain());
9721 else if (St->getValue().hasOneUse() &&
9722 ChainVal->getOpcode() == ISD::TokenFactor) {
9723 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9724 if (ChainVal->getOperand(i).getNode() == LdVal) {
9725 TokenFactorIndex = i;
9726 Ld = cast<LoadSDNode>(St->getValue());
9728 Ops.push_back(ChainVal->getOperand(i));
9732 if (!Ld || !ISD::isNormalLoad(Ld))
9735 // If this is not the MMX case, i.e. we are just turning i64 load/store
9736 // into f64 load/store, avoid the transformation if there are multiple
9737 // uses of the loaded value.
9738 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9741 DebugLoc LdDL = Ld->getDebugLoc();
9742 DebugLoc StDL = N->getDebugLoc();
9743 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9744 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9746 if (Subtarget->is64Bit() || F64IsLegal) {
9747 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9748 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9749 Ld->getBasePtr(), Ld->getSrcValue(),
9750 Ld->getSrcValueOffset(), Ld->isVolatile(),
9751 Ld->isNonTemporal(), Ld->getAlignment());
9752 SDValue NewChain = NewLd.getValue(1);
9753 if (TokenFactorIndex != -1) {
9754 Ops.push_back(NewChain);
9755 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9758 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9759 St->getSrcValue(), St->getSrcValueOffset(),
9760 St->isVolatile(), St->isNonTemporal(),
9761 St->getAlignment());
9764 // Otherwise, lower to two pairs of 32-bit loads / stores.
9765 SDValue LoAddr = Ld->getBasePtr();
9766 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9767 DAG.getConstant(4, MVT::i32));
9769 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9770 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9771 Ld->isVolatile(), Ld->isNonTemporal(),
9772 Ld->getAlignment());
9773 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9774 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9775 Ld->isVolatile(), Ld->isNonTemporal(),
9776 MinAlign(Ld->getAlignment(), 4));
9778 SDValue NewChain = LoLd.getValue(1);
9779 if (TokenFactorIndex != -1) {
9780 Ops.push_back(LoLd);
9781 Ops.push_back(HiLd);
9782 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9786 LoAddr = St->getBasePtr();
9787 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9788 DAG.getConstant(4, MVT::i32));
9790 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9791 St->getSrcValue(), St->getSrcValueOffset(),
9792 St->isVolatile(), St->isNonTemporal(),
9793 St->getAlignment());
9794 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9796 St->getSrcValueOffset() + 4,
9798 St->isNonTemporal(),
9799 MinAlign(St->getAlignment(), 4));
9800 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9805 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9806 /// X86ISD::FXOR nodes.
9807 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9808 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9809 // F[X]OR(0.0, x) -> x
9810 // F[X]OR(x, 0.0) -> x
9811 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9812 if (C->getValueAPF().isPosZero())
9813 return N->getOperand(1);
9814 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9815 if (C->getValueAPF().isPosZero())
9816 return N->getOperand(0);
9820 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9821 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9822 // FAND(0.0, x) -> 0.0
9823 // FAND(x, 0.0) -> 0.0
9824 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9825 if (C->getValueAPF().isPosZero())
9826 return N->getOperand(0);
9827 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9828 if (C->getValueAPF().isPosZero())
9829 return N->getOperand(1);
9833 static SDValue PerformBTCombine(SDNode *N,
9835 TargetLowering::DAGCombinerInfo &DCI) {
9836 // BT ignores high bits in the bit index operand.
9837 SDValue Op1 = N->getOperand(1);
9838 if (Op1.hasOneUse()) {
9839 unsigned BitWidth = Op1.getValueSizeInBits();
9840 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9841 APInt KnownZero, KnownOne;
9842 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9843 !DCI.isBeforeLegalizeOps());
9844 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9845 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9846 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9847 DCI.CommitTargetLoweringOpt(TLO);
9852 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9853 SDValue Op = N->getOperand(0);
9854 if (Op.getOpcode() == ISD::BIT_CONVERT)
9855 Op = Op.getOperand(0);
9856 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9857 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9858 VT.getVectorElementType().getSizeInBits() ==
9859 OpVT.getVectorElementType().getSizeInBits()) {
9860 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9865 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9866 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9867 // (and (i32 x86isd::setcc_carry), 1)
9868 // This eliminates the zext. This transformation is necessary because
9869 // ISD::SETCC is always legalized to i8.
9870 DebugLoc dl = N->getDebugLoc();
9871 SDValue N0 = N->getOperand(0);
9872 EVT VT = N->getValueType(0);
9873 if (N0.getOpcode() == ISD::AND &&
9875 N0.getOperand(0).hasOneUse()) {
9876 SDValue N00 = N0.getOperand(0);
9877 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9879 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9880 if (!C || C->getZExtValue() != 1)
9882 return DAG.getNode(ISD::AND, dl, VT,
9883 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9884 N00.getOperand(0), N00.getOperand(1)),
9885 DAG.getConstant(1, VT));
9891 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9892 DAGCombinerInfo &DCI) const {
9893 SelectionDAG &DAG = DCI.DAG;
9894 switch (N->getOpcode()) {
9896 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9897 case ISD::EXTRACT_VECTOR_ELT:
9898 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9899 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9900 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9901 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9904 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9905 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
9906 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9908 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9909 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9910 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9911 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9912 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9918 /// isTypeDesirableForOp - Return true if the target has native support for
9919 /// the specified value type and it is 'desirable' to use the type for the
9920 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9921 /// instruction encodings are longer and some i16 instructions are slow.
9922 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9923 if (!isTypeLegal(VT))
9932 case ISD::SIGN_EXTEND:
9933 case ISD::ZERO_EXTEND:
9934 case ISD::ANY_EXTEND:
9947 static bool MayFoldLoad(SDValue Op) {
9948 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
9951 static bool MayFoldIntoStore(SDValue Op) {
9952 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
9955 /// IsDesirableToPromoteOp - This method query the target whether it is
9956 /// beneficial for dag combiner to promote the specified node. If true, it
9957 /// should return the desired promotion type by reference.
9958 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
9959 EVT VT = Op.getValueType();
9963 bool Promote = false;
9964 bool Commute = false;
9965 switch (Op.getOpcode()) {
9968 LoadSDNode *LD = cast<LoadSDNode>(Op);
9969 // If the non-extending load has a single use and it's not live out, then it
9971 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
9973 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9974 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
9975 // The only case where we'd want to promote LOAD (rather then it being
9976 // promoted as an operand is when it's only use is liveout.
9977 if (UI->getOpcode() != ISD::CopyToReg)
9984 case ISD::SIGN_EXTEND:
9985 case ISD::ZERO_EXTEND:
9986 case ISD::ANY_EXTEND:
9991 SDValue N0 = Op.getOperand(0);
9992 // Look out for (store (shl (load), x)).
9993 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10006 SDValue N0 = Op.getOperand(0);
10007 SDValue N1 = Op.getOperand(1);
10008 if (!Commute && MayFoldLoad(N1))
10010 // Avoid disabling potential load folding opportunities.
10011 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10013 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10023 //===----------------------------------------------------------------------===//
10024 // X86 Inline Assembly Support
10025 //===----------------------------------------------------------------------===//
10027 static bool LowerToBSwap(CallInst *CI) {
10028 // FIXME: this should verify that we are targetting a 486 or better. If not,
10029 // we will turn this bswap into something that will be lowered to logical ops
10030 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10031 // so don't worry about this.
10033 // Verify this is a simple bswap.
10034 if (CI->getNumArgOperands() != 1 ||
10035 CI->getType() != CI->getArgOperand(0)->getType() ||
10036 !CI->getType()->isIntegerTy())
10039 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10040 if (!Ty || Ty->getBitWidth() % 16 != 0)
10043 // Okay, we can do this xform, do so now.
10044 const Type *Tys[] = { Ty };
10045 Module *M = CI->getParent()->getParent()->getParent();
10046 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10048 Value *Op = CI->getArgOperand(0);
10049 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10051 CI->replaceAllUsesWith(Op);
10052 CI->eraseFromParent();
10056 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10057 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10058 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10060 std::string AsmStr = IA->getAsmString();
10062 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10063 SmallVector<StringRef, 4> AsmPieces;
10064 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10066 switch (AsmPieces.size()) {
10067 default: return false;
10069 AsmStr = AsmPieces[0];
10071 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10074 if (AsmPieces.size() == 2 &&
10075 (AsmPieces[0] == "bswap" ||
10076 AsmPieces[0] == "bswapq" ||
10077 AsmPieces[0] == "bswapl") &&
10078 (AsmPieces[1] == "$0" ||
10079 AsmPieces[1] == "${0:q}")) {
10080 // No need to check constraints, nothing other than the equivalent of
10081 // "=r,0" would be valid here.
10082 return LowerToBSwap(CI);
10084 // rorw $$8, ${0:w} --> llvm.bswap.i16
10085 if (CI->getType()->isIntegerTy(16) &&
10086 AsmPieces.size() == 3 &&
10087 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10088 AsmPieces[1] == "$$8," &&
10089 AsmPieces[2] == "${0:w}" &&
10090 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10092 const std::string &Constraints = IA->getConstraintString();
10093 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10094 std::sort(AsmPieces.begin(), AsmPieces.end());
10095 if (AsmPieces.size() == 4 &&
10096 AsmPieces[0] == "~{cc}" &&
10097 AsmPieces[1] == "~{dirflag}" &&
10098 AsmPieces[2] == "~{flags}" &&
10099 AsmPieces[3] == "~{fpsr}") {
10100 return LowerToBSwap(CI);
10105 if (CI->getType()->isIntegerTy(64) &&
10106 Constraints.size() >= 2 &&
10107 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10108 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10109 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10110 SmallVector<StringRef, 4> Words;
10111 SplitString(AsmPieces[0], Words, " \t");
10112 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10114 SplitString(AsmPieces[1], Words, " \t");
10115 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10117 SplitString(AsmPieces[2], Words, " \t,");
10118 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10119 Words[2] == "%edx") {
10120 return LowerToBSwap(CI);
10132 /// getConstraintType - Given a constraint letter, return the type of
10133 /// constraint it is for this target.
10134 X86TargetLowering::ConstraintType
10135 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10136 if (Constraint.size() == 1) {
10137 switch (Constraint[0]) {
10149 return C_RegisterClass;
10157 return TargetLowering::getConstraintType(Constraint);
10160 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10161 /// with another that has more specific requirements based on the type of the
10162 /// corresponding operand.
10163 const char *X86TargetLowering::
10164 LowerXConstraint(EVT ConstraintVT) const {
10165 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10166 // 'f' like normal targets.
10167 if (ConstraintVT.isFloatingPoint()) {
10168 if (Subtarget->hasSSE2())
10170 if (Subtarget->hasSSE1())
10174 return TargetLowering::LowerXConstraint(ConstraintVT);
10177 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10178 /// vector. If it is invalid, don't add anything to Ops.
10179 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10181 std::vector<SDValue>&Ops,
10182 SelectionDAG &DAG) const {
10183 SDValue Result(0, 0);
10185 switch (Constraint) {
10188 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10189 if (C->getZExtValue() <= 31) {
10190 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10196 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10197 if (C->getZExtValue() <= 63) {
10198 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10204 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10205 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10206 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10212 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10213 if (C->getZExtValue() <= 255) {
10214 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10220 // 32-bit signed value
10221 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10222 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10223 C->getSExtValue())) {
10224 // Widen to 64 bits here to get it sign extended.
10225 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10228 // FIXME gcc accepts some relocatable values here too, but only in certain
10229 // memory models; it's complicated.
10234 // 32-bit unsigned value
10235 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10236 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10237 C->getZExtValue())) {
10238 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10242 // FIXME gcc accepts some relocatable values here too, but only in certain
10243 // memory models; it's complicated.
10247 // Literal immediates are always ok.
10248 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10249 // Widen to 64 bits here to get it sign extended.
10250 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10254 // In any sort of PIC mode addresses need to be computed at runtime by
10255 // adding in a register or some sort of table lookup. These can't
10256 // be used as immediates.
10257 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC() ||
10258 Subtarget->isPICStyleRIPRel())
10261 // If we are in non-pic codegen mode, we allow the address of a global (with
10262 // an optional displacement) to be used with 'i'.
10263 GlobalAddressSDNode *GA = 0;
10264 int64_t Offset = 0;
10266 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10268 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10269 Offset += GA->getOffset();
10271 } else if (Op.getOpcode() == ISD::ADD) {
10272 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10273 Offset += C->getZExtValue();
10274 Op = Op.getOperand(0);
10277 } else if (Op.getOpcode() == ISD::SUB) {
10278 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10279 Offset += -C->getZExtValue();
10280 Op = Op.getOperand(0);
10285 // Otherwise, this isn't something we can handle, reject it.
10289 const GlobalValue *GV = GA->getGlobal();
10290 // If we require an extra load to get this address, as in PIC mode, we
10291 // can't accept it.
10292 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10293 getTargetMachine())))
10296 Result = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10301 if (Result.getNode()) {
10302 Ops.push_back(Result);
10305 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10308 std::vector<unsigned> X86TargetLowering::
10309 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10311 if (Constraint.size() == 1) {
10312 // FIXME: not handling fp-stack yet!
10313 switch (Constraint[0]) { // GCC X86 Constraint Letters
10314 default: break; // Unknown constraint letter
10315 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10316 if (Subtarget->is64Bit()) {
10317 if (VT == MVT::i32)
10318 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10319 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10320 X86::R10D,X86::R11D,X86::R12D,
10321 X86::R13D,X86::R14D,X86::R15D,
10322 X86::EBP, X86::ESP, 0);
10323 else if (VT == MVT::i16)
10324 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10325 X86::SI, X86::DI, X86::R8W,X86::R9W,
10326 X86::R10W,X86::R11W,X86::R12W,
10327 X86::R13W,X86::R14W,X86::R15W,
10328 X86::BP, X86::SP, 0);
10329 else if (VT == MVT::i8)
10330 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10331 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10332 X86::R10B,X86::R11B,X86::R12B,
10333 X86::R13B,X86::R14B,X86::R15B,
10334 X86::BPL, X86::SPL, 0);
10336 else if (VT == MVT::i64)
10337 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10338 X86::RSI, X86::RDI, X86::R8, X86::R9,
10339 X86::R10, X86::R11, X86::R12,
10340 X86::R13, X86::R14, X86::R15,
10341 X86::RBP, X86::RSP, 0);
10345 // 32-bit fallthrough
10346 case 'Q': // Q_REGS
10347 if (VT == MVT::i32)
10348 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10349 else if (VT == MVT::i16)
10350 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10351 else if (VT == MVT::i8)
10352 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10353 else if (VT == MVT::i64)
10354 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10359 return std::vector<unsigned>();
10362 std::pair<unsigned, const TargetRegisterClass*>
10363 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10365 // First, see if this is a constraint that directly corresponds to an LLVM
10367 if (Constraint.size() == 1) {
10368 // GCC Constraint Letters
10369 switch (Constraint[0]) {
10371 case 'r': // GENERAL_REGS
10372 case 'l': // INDEX_REGS
10374 return std::make_pair(0U, X86::GR8RegisterClass);
10375 if (VT == MVT::i16)
10376 return std::make_pair(0U, X86::GR16RegisterClass);
10377 if (VT == MVT::i32 || !Subtarget->is64Bit())
10378 return std::make_pair(0U, X86::GR32RegisterClass);
10379 return std::make_pair(0U, X86::GR64RegisterClass);
10380 case 'R': // LEGACY_REGS
10382 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10383 if (VT == MVT::i16)
10384 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10385 if (VT == MVT::i32 || !Subtarget->is64Bit())
10386 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10387 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10388 case 'f': // FP Stack registers.
10389 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10390 // value to the correct fpstack register class.
10391 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10392 return std::make_pair(0U, X86::RFP32RegisterClass);
10393 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10394 return std::make_pair(0U, X86::RFP64RegisterClass);
10395 return std::make_pair(0U, X86::RFP80RegisterClass);
10396 case 'y': // MMX_REGS if MMX allowed.
10397 if (!Subtarget->hasMMX()) break;
10398 return std::make_pair(0U, X86::VR64RegisterClass);
10399 case 'Y': // SSE_REGS if SSE2 allowed
10400 if (!Subtarget->hasSSE2()) break;
10402 case 'x': // SSE_REGS if SSE1 allowed
10403 if (!Subtarget->hasSSE1()) break;
10405 switch (VT.getSimpleVT().SimpleTy) {
10407 // Scalar SSE types.
10410 return std::make_pair(0U, X86::FR32RegisterClass);
10413 return std::make_pair(0U, X86::FR64RegisterClass);
10421 return std::make_pair(0U, X86::VR128RegisterClass);
10427 // Use the default implementation in TargetLowering to convert the register
10428 // constraint into a member of a register class.
10429 std::pair<unsigned, const TargetRegisterClass*> Res;
10430 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10432 // Not found as a standard register?
10433 if (Res.second == 0) {
10434 // Map st(0) -> st(7) -> ST0
10435 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10436 tolower(Constraint[1]) == 's' &&
10437 tolower(Constraint[2]) == 't' &&
10438 Constraint[3] == '(' &&
10439 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10440 Constraint[5] == ')' &&
10441 Constraint[6] == '}') {
10443 Res.first = X86::ST0+Constraint[4]-'0';
10444 Res.second = X86::RFP80RegisterClass;
10448 // GCC allows "st(0)" to be called just plain "st".
10449 if (StringRef("{st}").equals_lower(Constraint)) {
10450 Res.first = X86::ST0;
10451 Res.second = X86::RFP80RegisterClass;
10456 if (StringRef("{flags}").equals_lower(Constraint)) {
10457 Res.first = X86::EFLAGS;
10458 Res.second = X86::CCRRegisterClass;
10462 // 'A' means EAX + EDX.
10463 if (Constraint == "A") {
10464 Res.first = X86::EAX;
10465 Res.second = X86::GR32_ADRegisterClass;
10471 // Otherwise, check to see if this is a register class of the wrong value
10472 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10473 // turn into {ax},{dx}.
10474 if (Res.second->hasType(VT))
10475 return Res; // Correct type already, nothing to do.
10477 // All of the single-register GCC register classes map their values onto
10478 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10479 // really want an 8-bit or 32-bit register, map to the appropriate register
10480 // class and return the appropriate register.
10481 if (Res.second == X86::GR16RegisterClass) {
10482 if (VT == MVT::i8) {
10483 unsigned DestReg = 0;
10484 switch (Res.first) {
10486 case X86::AX: DestReg = X86::AL; break;
10487 case X86::DX: DestReg = X86::DL; break;
10488 case X86::CX: DestReg = X86::CL; break;
10489 case X86::BX: DestReg = X86::BL; break;
10492 Res.first = DestReg;
10493 Res.second = X86::GR8RegisterClass;
10495 } else if (VT == MVT::i32) {
10496 unsigned DestReg = 0;
10497 switch (Res.first) {
10499 case X86::AX: DestReg = X86::EAX; break;
10500 case X86::DX: DestReg = X86::EDX; break;
10501 case X86::CX: DestReg = X86::ECX; break;
10502 case X86::BX: DestReg = X86::EBX; break;
10503 case X86::SI: DestReg = X86::ESI; break;
10504 case X86::DI: DestReg = X86::EDI; break;
10505 case X86::BP: DestReg = X86::EBP; break;
10506 case X86::SP: DestReg = X86::ESP; break;
10509 Res.first = DestReg;
10510 Res.second = X86::GR32RegisterClass;
10512 } else if (VT == MVT::i64) {
10513 unsigned DestReg = 0;
10514 switch (Res.first) {
10516 case X86::AX: DestReg = X86::RAX; break;
10517 case X86::DX: DestReg = X86::RDX; break;
10518 case X86::CX: DestReg = X86::RCX; break;
10519 case X86::BX: DestReg = X86::RBX; break;
10520 case X86::SI: DestReg = X86::RSI; break;
10521 case X86::DI: DestReg = X86::RDI; break;
10522 case X86::BP: DestReg = X86::RBP; break;
10523 case X86::SP: DestReg = X86::RSP; break;
10526 Res.first = DestReg;
10527 Res.second = X86::GR64RegisterClass;
10530 } else if (Res.second == X86::FR32RegisterClass ||
10531 Res.second == X86::FR64RegisterClass ||
10532 Res.second == X86::VR128RegisterClass) {
10533 // Handle references to XMM physical registers that got mapped into the
10534 // wrong class. This can happen with constraints like {xmm0} where the
10535 // target independent register mapper will just pick the first match it can
10536 // find, ignoring the required type.
10537 if (VT == MVT::f32)
10538 Res.second = X86::FR32RegisterClass;
10539 else if (VT == MVT::f64)
10540 Res.second = X86::FR64RegisterClass;
10541 else if (X86::VR128RegisterClass->hasType(VT))
10542 Res.second = X86::VR128RegisterClass;