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 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
623 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
624 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
625 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
626 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
628 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
629 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
630 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
631 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
633 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
634 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
636 setOperationAction(ISD::AND, MVT::v8i8, Promote);
637 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
638 setOperationAction(ISD::AND, MVT::v4i16, Promote);
639 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
640 setOperationAction(ISD::AND, MVT::v2i32, Promote);
641 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
642 setOperationAction(ISD::AND, MVT::v1i64, Legal);
644 setOperationAction(ISD::OR, MVT::v8i8, Promote);
645 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
646 setOperationAction(ISD::OR, MVT::v4i16, Promote);
647 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
648 setOperationAction(ISD::OR, MVT::v2i32, Promote);
649 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
650 setOperationAction(ISD::OR, MVT::v1i64, Legal);
652 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
653 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
654 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
655 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
656 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
657 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
658 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
660 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
661 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
662 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
663 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
664 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
665 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
666 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
669 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
670 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
671 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
674 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
675 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
676 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
678 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
679 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
680 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
682 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
684 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
685 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
686 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
687 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
689 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
690 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
692 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
693 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
694 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
695 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
696 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
700 if (!UseSoftFloat && Subtarget->hasSSE1()) {
701 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
703 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
704 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
705 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
706 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
707 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
708 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
709 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
710 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
711 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
712 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
713 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
714 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
717 if (!UseSoftFloat && Subtarget->hasSSE2()) {
718 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
720 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
721 // registers cannot be used even for integer operations.
722 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
723 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
724 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
725 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
727 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
728 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
729 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
730 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
731 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
732 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
733 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
734 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
735 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
736 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
737 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
738 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
739 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
740 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
741 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
742 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
744 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
745 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
746 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
747 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
749 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
750 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
751 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
752 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
753 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
755 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
756 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
757 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
759 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
761 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
762 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
763 EVT VT = (MVT::SimpleValueType)i;
764 // Do not attempt to custom lower non-power-of-2 vectors
765 if (!isPowerOf2_32(VT.getVectorNumElements()))
767 // Do not attempt to custom lower non-128-bit vectors
768 if (!VT.is128BitVector())
770 setOperationAction(ISD::BUILD_VECTOR,
771 VT.getSimpleVT().SimpleTy, Custom);
772 setOperationAction(ISD::VECTOR_SHUFFLE,
773 VT.getSimpleVT().SimpleTy, Custom);
774 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
775 VT.getSimpleVT().SimpleTy, Custom);
778 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
779 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
780 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
781 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
782 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
783 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
785 if (Subtarget->is64Bit()) {
786 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
787 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
790 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
791 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
792 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
795 // Do not attempt to promote non-128-bit vectors
796 if (!VT.is128BitVector()) {
800 setOperationAction(ISD::AND, SVT, Promote);
801 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
802 setOperationAction(ISD::OR, SVT, Promote);
803 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
804 setOperationAction(ISD::XOR, SVT, Promote);
805 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
806 setOperationAction(ISD::LOAD, SVT, Promote);
807 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
808 setOperationAction(ISD::SELECT, SVT, Promote);
809 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
812 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
814 // Custom lower v2i64 and v2f64 selects.
815 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
816 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
817 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
818 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
820 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
821 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
822 if (!DisableMMX && Subtarget->hasMMX()) {
823 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
824 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
828 if (Subtarget->hasSSE41()) {
829 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
830 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
831 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
832 setOperationAction(ISD::FRINT, MVT::f32, Legal);
833 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
834 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
835 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
836 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
837 setOperationAction(ISD::FRINT, MVT::f64, Legal);
838 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
840 // FIXME: Do we need to handle scalar-to-vector here?
841 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
843 // i8 and i16 vectors are custom , because the source register and source
844 // source memory operand types are not the same width. f32 vectors are
845 // custom since the immediate controlling the insert encodes additional
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
848 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
850 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
852 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
853 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
854 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
855 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
857 if (Subtarget->is64Bit()) {
858 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
859 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
863 if (Subtarget->hasSSE42()) {
864 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
867 if (!UseSoftFloat && Subtarget->hasAVX()) {
868 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
869 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
870 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
871 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
873 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
874 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
875 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
876 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
877 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
878 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
879 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
880 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
881 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
882 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
883 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
884 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
885 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
886 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
887 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
889 // Operations to consider commented out -v16i16 v32i8
890 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
891 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
892 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
893 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
894 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
895 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
896 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
897 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
898 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
899 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
900 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
901 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
902 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
903 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
905 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
906 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
907 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
908 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
910 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
911 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
912 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
913 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
914 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
916 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
917 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
918 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
919 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
920 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
921 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
924 // Not sure we want to do this since there are no 256-bit integer
927 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
928 // This includes 256-bit vectors
929 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
930 EVT VT = (MVT::SimpleValueType)i;
932 // Do not attempt to custom lower non-power-of-2 vectors
933 if (!isPowerOf2_32(VT.getVectorNumElements()))
936 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
937 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
938 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
941 if (Subtarget->is64Bit()) {
942 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
943 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
948 // Not sure we want to do this since there are no 256-bit integer
951 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
952 // Including 256-bit vectors
953 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
954 EVT VT = (MVT::SimpleValueType)i;
956 if (!VT.is256BitVector()) {
959 setOperationAction(ISD::AND, VT, Promote);
960 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
961 setOperationAction(ISD::OR, VT, Promote);
962 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
963 setOperationAction(ISD::XOR, VT, Promote);
964 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
965 setOperationAction(ISD::LOAD, VT, Promote);
966 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
967 setOperationAction(ISD::SELECT, VT, Promote);
968 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
971 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
975 // We want to custom lower some of our intrinsics.
976 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
978 // Add/Sub/Mul with overflow operations are custom lowered.
979 setOperationAction(ISD::SADDO, MVT::i32, Custom);
980 setOperationAction(ISD::UADDO, MVT::i32, Custom);
981 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
982 setOperationAction(ISD::USUBO, MVT::i32, Custom);
983 setOperationAction(ISD::SMULO, MVT::i32, Custom);
985 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
986 // handle type legalization for these operations here.
988 // FIXME: We really should do custom legalization for addition and
989 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
990 // than generic legalization for 64-bit multiplication-with-overflow, though.
991 if (Subtarget->is64Bit()) {
992 setOperationAction(ISD::SADDO, MVT::i64, Custom);
993 setOperationAction(ISD::UADDO, MVT::i64, Custom);
994 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
995 setOperationAction(ISD::USUBO, MVT::i64, Custom);
996 setOperationAction(ISD::SMULO, MVT::i64, Custom);
999 if (!Subtarget->is64Bit()) {
1000 // These libcalls are not available in 32-bit.
1001 setLibcallName(RTLIB::SHL_I128, 0);
1002 setLibcallName(RTLIB::SRL_I128, 0);
1003 setLibcallName(RTLIB::SRA_I128, 0);
1006 // We have target-specific dag combine patterns for the following nodes:
1007 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1008 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1009 setTargetDAGCombine(ISD::BUILD_VECTOR);
1010 setTargetDAGCombine(ISD::SELECT);
1011 setTargetDAGCombine(ISD::SHL);
1012 setTargetDAGCombine(ISD::SRA);
1013 setTargetDAGCombine(ISD::SRL);
1014 setTargetDAGCombine(ISD::OR);
1015 setTargetDAGCombine(ISD::STORE);
1016 setTargetDAGCombine(ISD::ZERO_EXTEND);
1017 if (Subtarget->is64Bit())
1018 setTargetDAGCombine(ISD::MUL);
1020 computeRegisterProperties();
1022 // FIXME: These should be based on subtarget info. Plus, the values should
1023 // be smaller when we are in optimizing for size mode.
1024 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1025 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1026 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1027 setPrefLoopAlignment(16);
1028 benefitFromCodePlacementOpt = true;
1032 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1037 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1038 /// the desired ByVal argument alignment.
1039 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1042 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1043 if (VTy->getBitWidth() == 128)
1045 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1046 unsigned EltAlign = 0;
1047 getMaxByValAlign(ATy->getElementType(), EltAlign);
1048 if (EltAlign > MaxAlign)
1049 MaxAlign = EltAlign;
1050 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1051 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1052 unsigned EltAlign = 0;
1053 getMaxByValAlign(STy->getElementType(i), EltAlign);
1054 if (EltAlign > MaxAlign)
1055 MaxAlign = EltAlign;
1063 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1064 /// function arguments in the caller parameter area. For X86, aggregates
1065 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1066 /// are at 4-byte boundaries.
1067 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1068 if (Subtarget->is64Bit()) {
1069 // Max of 8 and alignment of type.
1070 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1077 if (Subtarget->hasSSE1())
1078 getMaxByValAlign(Ty, Align);
1082 /// getOptimalMemOpType - Returns the target specific optimal type for load
1083 /// and store operations as a result of memset, memcpy, and memmove
1084 /// lowering. If DstAlign is zero that means it's safe to destination
1085 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1086 /// means there isn't a need to check it against alignment requirement,
1087 /// probably because the source does not need to be loaded. If
1088 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1089 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1090 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1091 /// constant so it does not need to be loaded.
1092 /// It returns EVT::Other if the type should be determined using generic
1093 /// target-independent logic.
1095 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1096 unsigned DstAlign, unsigned SrcAlign,
1097 bool NonScalarIntSafe,
1099 MachineFunction &MF) const {
1100 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1101 // linux. This is because the stack realignment code can't handle certain
1102 // cases like PR2962. This should be removed when PR2962 is fixed.
1103 const Function *F = MF.getFunction();
1104 if (NonScalarIntSafe &&
1105 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1107 (Subtarget->isUnalignedMemAccessFast() ||
1108 ((DstAlign == 0 || DstAlign >= 16) &&
1109 (SrcAlign == 0 || SrcAlign >= 16))) &&
1110 Subtarget->getStackAlignment() >= 16) {
1111 if (Subtarget->hasSSE2())
1113 if (Subtarget->hasSSE1())
1115 } else if (!MemcpyStrSrc && Size >= 8 &&
1116 !Subtarget->is64Bit() &&
1117 Subtarget->getStackAlignment() >= 8 &&
1118 Subtarget->hasSSE2()) {
1119 // Do not use f64 to lower memcpy if source is string constant. It's
1120 // better to use i32 to avoid the loads.
1124 if (Subtarget->is64Bit() && Size >= 8)
1129 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1130 /// current function. The returned value is a member of the
1131 /// MachineJumpTableInfo::JTEntryKind enum.
1132 unsigned X86TargetLowering::getJumpTableEncoding() const {
1133 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1135 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1136 Subtarget->isPICStyleGOT())
1137 return MachineJumpTableInfo::EK_Custom32;
1139 // Otherwise, use the normal jump table encoding heuristics.
1140 return TargetLowering::getJumpTableEncoding();
1143 /// getPICBaseSymbol - Return the X86-32 PIC base.
1145 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1146 MCContext &Ctx) const {
1147 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1148 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1149 Twine(MF->getFunctionNumber())+"$pb");
1154 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1155 const MachineBasicBlock *MBB,
1156 unsigned uid,MCContext &Ctx) const{
1157 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1158 Subtarget->isPICStyleGOT());
1159 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1161 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1162 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1165 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1167 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1168 SelectionDAG &DAG) const {
1169 if (!Subtarget->is64Bit())
1170 // This doesn't have DebugLoc associated with it, but is not really the
1171 // same as a Register.
1172 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1176 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1177 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1179 const MCExpr *X86TargetLowering::
1180 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1181 MCContext &Ctx) const {
1182 // X86-64 uses RIP relative addressing based on the jump table label.
1183 if (Subtarget->isPICStyleRIPRel())
1184 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1186 // Otherwise, the reference is relative to the PIC base.
1187 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1190 /// getFunctionAlignment - Return the Log2 alignment of this function.
1191 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1192 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1195 //===----------------------------------------------------------------------===//
1196 // Return Value Calling Convention Implementation
1197 //===----------------------------------------------------------------------===//
1199 #include "X86GenCallingConv.inc"
1202 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1203 const SmallVectorImpl<EVT> &OutTys,
1204 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1205 SelectionDAG &DAG) const {
1206 SmallVector<CCValAssign, 16> RVLocs;
1207 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1208 RVLocs, *DAG.getContext());
1209 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1213 X86TargetLowering::LowerReturn(SDValue Chain,
1214 CallingConv::ID CallConv, bool isVarArg,
1215 const SmallVectorImpl<ISD::OutputArg> &Outs,
1216 DebugLoc dl, SelectionDAG &DAG) const {
1217 MachineFunction &MF = DAG.getMachineFunction();
1218 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1220 SmallVector<CCValAssign, 16> RVLocs;
1221 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1222 RVLocs, *DAG.getContext());
1223 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1225 // Add the regs to the liveout set for the function.
1226 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1227 for (unsigned i = 0; i != RVLocs.size(); ++i)
1228 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1229 MRI.addLiveOut(RVLocs[i].getLocReg());
1233 SmallVector<SDValue, 6> RetOps;
1234 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1235 // Operand #1 = Bytes To Pop
1236 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1239 // Copy the result values into the output registers.
1240 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1241 CCValAssign &VA = RVLocs[i];
1242 assert(VA.isRegLoc() && "Can only return in registers!");
1243 SDValue ValToCopy = Outs[i].Val;
1245 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1246 // the RET instruction and handled by the FP Stackifier.
1247 if (VA.getLocReg() == X86::ST0 ||
1248 VA.getLocReg() == X86::ST1) {
1249 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1250 // change the value to the FP stack register class.
1251 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1252 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1253 RetOps.push_back(ValToCopy);
1254 // Don't emit a copytoreg.
1258 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1259 // which is returned in RAX / RDX.
1260 if (Subtarget->is64Bit()) {
1261 EVT ValVT = ValToCopy.getValueType();
1262 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1263 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1264 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1265 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1269 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1270 Flag = Chain.getValue(1);
1273 // The x86-64 ABI for returning structs by value requires that we copy
1274 // the sret argument into %rax for the return. We saved the argument into
1275 // a virtual register in the entry block, so now we copy the value out
1277 if (Subtarget->is64Bit() &&
1278 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1279 MachineFunction &MF = DAG.getMachineFunction();
1280 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1281 unsigned Reg = FuncInfo->getSRetReturnReg();
1283 "SRetReturnReg should have been set in LowerFormalArguments().");
1284 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1286 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1287 Flag = Chain.getValue(1);
1289 // RAX now acts like a return value.
1290 MRI.addLiveOut(X86::RAX);
1293 RetOps[0] = Chain; // Update chain.
1295 // Add the flag if we have it.
1297 RetOps.push_back(Flag);
1299 return DAG.getNode(X86ISD::RET_FLAG, dl,
1300 MVT::Other, &RetOps[0], RetOps.size());
1303 /// LowerCallResult - Lower the result values of a call into the
1304 /// appropriate copies out of appropriate physical registers.
1307 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1308 CallingConv::ID CallConv, bool isVarArg,
1309 const SmallVectorImpl<ISD::InputArg> &Ins,
1310 DebugLoc dl, SelectionDAG &DAG,
1311 SmallVectorImpl<SDValue> &InVals) const {
1313 // Assign locations to each value returned by this call.
1314 SmallVector<CCValAssign, 16> RVLocs;
1315 bool Is64Bit = Subtarget->is64Bit();
1316 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1317 RVLocs, *DAG.getContext());
1318 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1320 // Copy all of the result registers out of their specified physreg.
1321 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1322 CCValAssign &VA = RVLocs[i];
1323 EVT CopyVT = VA.getValVT();
1325 // If this is x86-64, and we disabled SSE, we can't return FP values
1326 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1327 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1328 report_fatal_error("SSE register return with SSE disabled");
1331 // If this is a call to a function that returns an fp value on the floating
1332 // point stack, but where we prefer to use the value in xmm registers, copy
1333 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1334 if ((VA.getLocReg() == X86::ST0 ||
1335 VA.getLocReg() == X86::ST1) &&
1336 isScalarFPTypeInSSEReg(VA.getValVT())) {
1341 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1342 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1343 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1344 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1345 MVT::v2i64, InFlag).getValue(1);
1346 Val = Chain.getValue(0);
1347 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1348 Val, DAG.getConstant(0, MVT::i64));
1350 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1351 MVT::i64, InFlag).getValue(1);
1352 Val = Chain.getValue(0);
1354 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1356 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1357 CopyVT, InFlag).getValue(1);
1358 Val = Chain.getValue(0);
1360 InFlag = Chain.getValue(2);
1362 if (CopyVT != VA.getValVT()) {
1363 // Round the F80 the right size, which also moves to the appropriate xmm
1365 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1366 // This truncation won't change the value.
1367 DAG.getIntPtrConstant(1));
1370 InVals.push_back(Val);
1377 //===----------------------------------------------------------------------===//
1378 // C & StdCall & Fast Calling Convention implementation
1379 //===----------------------------------------------------------------------===//
1380 // StdCall calling convention seems to be standard for many Windows' API
1381 // routines and around. It differs from C calling convention just a little:
1382 // callee should clean up the stack, not caller. Symbols should be also
1383 // decorated in some fancy way :) It doesn't support any vector arguments.
1384 // For info on fast calling convention see Fast Calling Convention (tail call)
1385 // implementation LowerX86_32FastCCCallTo.
1387 /// CallIsStructReturn - Determines whether a call uses struct return
1389 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1393 return Outs[0].Flags.isSRet();
1396 /// ArgsAreStructReturn - Determines whether a function uses struct
1397 /// return semantics.
1399 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1403 return Ins[0].Flags.isSRet();
1406 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1407 /// given CallingConvention value.
1408 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1409 if (Subtarget->is64Bit()) {
1410 if (CC == CallingConv::GHC)
1411 return CC_X86_64_GHC;
1412 else if (Subtarget->isTargetWin64())
1413 return CC_X86_Win64_C;
1418 if (CC == CallingConv::X86_FastCall)
1419 return CC_X86_32_FastCall;
1420 else if (CC == CallingConv::X86_ThisCall)
1421 return CC_X86_32_ThisCall;
1422 else if (CC == CallingConv::Fast)
1423 return CC_X86_32_FastCC;
1424 else if (CC == CallingConv::GHC)
1425 return CC_X86_32_GHC;
1430 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1431 /// by "Src" to address "Dst" with size and alignment information specified by
1432 /// the specific parameter attribute. The copy will be passed as a byval
1433 /// function parameter.
1435 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1436 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1438 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1439 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1440 /*isVolatile*/false, /*AlwaysInline=*/true,
1444 /// IsTailCallConvention - Return true if the calling convention is one that
1445 /// supports tail call optimization.
1446 static bool IsTailCallConvention(CallingConv::ID CC) {
1447 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1450 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1451 /// a tailcall target by changing its ABI.
1452 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1453 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1457 X86TargetLowering::LowerMemArgument(SDValue Chain,
1458 CallingConv::ID CallConv,
1459 const SmallVectorImpl<ISD::InputArg> &Ins,
1460 DebugLoc dl, SelectionDAG &DAG,
1461 const CCValAssign &VA,
1462 MachineFrameInfo *MFI,
1464 // Create the nodes corresponding to a load from this parameter slot.
1465 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1466 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1467 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1470 // If value is passed by pointer we have address passed instead of the value
1472 if (VA.getLocInfo() == CCValAssign::Indirect)
1473 ValVT = VA.getLocVT();
1475 ValVT = VA.getValVT();
1477 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1478 // changed with more analysis.
1479 // In case of tail call optimization mark all arguments mutable. Since they
1480 // could be overwritten by lowering of arguments in case of a tail call.
1481 if (Flags.isByVal()) {
1482 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1483 VA.getLocMemOffset(), isImmutable);
1484 return DAG.getFrameIndex(FI, getPointerTy());
1486 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1487 VA.getLocMemOffset(), isImmutable);
1488 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1489 return DAG.getLoad(ValVT, dl, Chain, FIN,
1490 PseudoSourceValue::getFixedStack(FI), 0,
1496 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1497 CallingConv::ID CallConv,
1499 const SmallVectorImpl<ISD::InputArg> &Ins,
1502 SmallVectorImpl<SDValue> &InVals)
1504 MachineFunction &MF = DAG.getMachineFunction();
1505 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1507 const Function* Fn = MF.getFunction();
1508 if (Fn->hasExternalLinkage() &&
1509 Subtarget->isTargetCygMing() &&
1510 Fn->getName() == "main")
1511 FuncInfo->setForceFramePointer(true);
1513 MachineFrameInfo *MFI = MF.getFrameInfo();
1514 bool Is64Bit = Subtarget->is64Bit();
1515 bool IsWin64 = Subtarget->isTargetWin64();
1517 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1518 "Var args not supported with calling convention fastcc or ghc");
1520 // Assign locations to all of the incoming arguments.
1521 SmallVector<CCValAssign, 16> ArgLocs;
1522 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1523 ArgLocs, *DAG.getContext());
1524 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1526 unsigned LastVal = ~0U;
1528 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1529 CCValAssign &VA = ArgLocs[i];
1530 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1532 assert(VA.getValNo() != LastVal &&
1533 "Don't support value assigned to multiple locs yet");
1534 LastVal = VA.getValNo();
1536 if (VA.isRegLoc()) {
1537 EVT RegVT = VA.getLocVT();
1538 TargetRegisterClass *RC = NULL;
1539 if (RegVT == MVT::i32)
1540 RC = X86::GR32RegisterClass;
1541 else if (Is64Bit && RegVT == MVT::i64)
1542 RC = X86::GR64RegisterClass;
1543 else if (RegVT == MVT::f32)
1544 RC = X86::FR32RegisterClass;
1545 else if (RegVT == MVT::f64)
1546 RC = X86::FR64RegisterClass;
1547 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1548 RC = X86::VR128RegisterClass;
1549 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1550 RC = X86::VR64RegisterClass;
1552 llvm_unreachable("Unknown argument type!");
1554 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1555 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1557 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1558 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1560 if (VA.getLocInfo() == CCValAssign::SExt)
1561 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1562 DAG.getValueType(VA.getValVT()));
1563 else if (VA.getLocInfo() == CCValAssign::ZExt)
1564 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1565 DAG.getValueType(VA.getValVT()));
1566 else if (VA.getLocInfo() == CCValAssign::BCvt)
1567 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1569 if (VA.isExtInLoc()) {
1570 // Handle MMX values passed in XMM regs.
1571 if (RegVT.isVector()) {
1572 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1573 ArgValue, DAG.getConstant(0, MVT::i64));
1574 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1576 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1579 assert(VA.isMemLoc());
1580 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1583 // If value is passed via pointer - do a load.
1584 if (VA.getLocInfo() == CCValAssign::Indirect)
1585 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1588 InVals.push_back(ArgValue);
1591 // The x86-64 ABI for returning structs by value requires that we copy
1592 // the sret argument into %rax for the return. Save the argument into
1593 // a virtual register so that we can access it from the return points.
1594 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1595 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1596 unsigned Reg = FuncInfo->getSRetReturnReg();
1598 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1599 FuncInfo->setSRetReturnReg(Reg);
1601 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1602 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1605 unsigned StackSize = CCInfo.getNextStackOffset();
1606 // Align stack specially for tail calls.
1607 if (FuncIsMadeTailCallSafe(CallConv))
1608 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1610 // If the function takes variable number of arguments, make a frame index for
1611 // the start of the first vararg value... for expansion of llvm.va_start.
1613 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1614 CallConv != CallingConv::X86_ThisCall)) {
1615 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1618 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1620 // FIXME: We should really autogenerate these arrays
1621 static const unsigned GPR64ArgRegsWin64[] = {
1622 X86::RCX, X86::RDX, X86::R8, X86::R9
1624 static const unsigned XMMArgRegsWin64[] = {
1625 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1627 static const unsigned GPR64ArgRegs64Bit[] = {
1628 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1630 static const unsigned XMMArgRegs64Bit[] = {
1631 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1632 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1634 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1637 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1638 GPR64ArgRegs = GPR64ArgRegsWin64;
1639 XMMArgRegs = XMMArgRegsWin64;
1641 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1642 GPR64ArgRegs = GPR64ArgRegs64Bit;
1643 XMMArgRegs = XMMArgRegs64Bit;
1645 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1647 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1650 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1651 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1652 "SSE register cannot be used when SSE is disabled!");
1653 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1654 "SSE register cannot be used when SSE is disabled!");
1655 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1656 // Kernel mode asks for SSE to be disabled, so don't push them
1658 TotalNumXMMRegs = 0;
1660 // For X86-64, if there are vararg parameters that are passed via
1661 // registers, then we must store them to their spots on the stack so they
1662 // may be loaded by deferencing the result of va_next.
1663 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1664 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1665 FuncInfo->setRegSaveFrameIndex(
1666 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1669 // Store the integer parameter registers.
1670 SmallVector<SDValue, 8> MemOps;
1671 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1673 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1674 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1675 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1676 DAG.getIntPtrConstant(Offset));
1677 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1678 X86::GR64RegisterClass);
1679 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1681 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1682 PseudoSourceValue::getFixedStack(
1683 FuncInfo->getRegSaveFrameIndex()),
1684 Offset, false, false, 0);
1685 MemOps.push_back(Store);
1689 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1690 // Now store the XMM (fp + vector) parameter registers.
1691 SmallVector<SDValue, 11> SaveXMMOps;
1692 SaveXMMOps.push_back(Chain);
1694 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1695 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1696 SaveXMMOps.push_back(ALVal);
1698 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1699 FuncInfo->getRegSaveFrameIndex()));
1700 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1701 FuncInfo->getVarArgsFPOffset()));
1703 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1704 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1705 X86::VR128RegisterClass);
1706 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1707 SaveXMMOps.push_back(Val);
1709 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1711 &SaveXMMOps[0], SaveXMMOps.size()));
1714 if (!MemOps.empty())
1715 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1716 &MemOps[0], MemOps.size());
1720 // Some CCs need callee pop.
1721 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1722 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1724 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1725 // If this is an sret function, the return should pop the hidden pointer.
1726 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1727 FuncInfo->setBytesToPopOnReturn(4);
1731 // RegSaveFrameIndex is X86-64 only.
1732 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1733 if (CallConv == CallingConv::X86_FastCall ||
1734 CallConv == CallingConv::X86_ThisCall)
1735 // fastcc functions can't have varargs.
1736 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1743 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1744 SDValue StackPtr, SDValue Arg,
1745 DebugLoc dl, SelectionDAG &DAG,
1746 const CCValAssign &VA,
1747 ISD::ArgFlagsTy Flags) const {
1748 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1749 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1750 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1751 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1752 if (Flags.isByVal()) {
1753 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1755 return DAG.getStore(Chain, dl, Arg, PtrOff,
1756 PseudoSourceValue::getStack(), LocMemOffset,
1760 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1761 /// optimization is performed and it is required.
1763 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1764 SDValue &OutRetAddr, SDValue Chain,
1765 bool IsTailCall, bool Is64Bit,
1766 int FPDiff, DebugLoc dl) const {
1767 // Adjust the Return address stack slot.
1768 EVT VT = getPointerTy();
1769 OutRetAddr = getReturnAddressFrameIndex(DAG);
1771 // Load the "old" Return address.
1772 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1773 return SDValue(OutRetAddr.getNode(), 1);
1776 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1777 /// optimization is performed and it is required (FPDiff!=0).
1779 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1780 SDValue Chain, SDValue RetAddrFrIdx,
1781 bool Is64Bit, int FPDiff, DebugLoc dl) {
1782 // Store the return address to the appropriate stack slot.
1783 if (!FPDiff) return Chain;
1784 // Calculate the new stack slot for the return address.
1785 int SlotSize = Is64Bit ? 8 : 4;
1786 int NewReturnAddrFI =
1787 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1788 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1789 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1790 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1791 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1797 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1798 CallingConv::ID CallConv, bool isVarArg,
1800 const SmallVectorImpl<ISD::OutputArg> &Outs,
1801 const SmallVectorImpl<ISD::InputArg> &Ins,
1802 DebugLoc dl, SelectionDAG &DAG,
1803 SmallVectorImpl<SDValue> &InVals) const {
1804 MachineFunction &MF = DAG.getMachineFunction();
1805 bool Is64Bit = Subtarget->is64Bit();
1806 bool IsStructRet = CallIsStructReturn(Outs);
1807 bool IsSibcall = false;
1810 // Check if it's really possible to do a tail call.
1811 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1812 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1815 // Sibcalls are automatically detected tailcalls which do not require
1817 if (!GuaranteedTailCallOpt && isTailCall)
1824 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1825 "Var args not supported with calling convention fastcc or ghc");
1827 // Analyze operands of the call, assigning locations to each operand.
1828 SmallVector<CCValAssign, 16> ArgLocs;
1829 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1830 ArgLocs, *DAG.getContext());
1831 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1833 // Get a count of how many bytes are to be pushed on the stack.
1834 unsigned NumBytes = CCInfo.getNextStackOffset();
1836 // This is a sibcall. The memory operands are available in caller's
1837 // own caller's stack.
1839 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1840 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1843 if (isTailCall && !IsSibcall) {
1844 // Lower arguments at fp - stackoffset + fpdiff.
1845 unsigned NumBytesCallerPushed =
1846 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1847 FPDiff = NumBytesCallerPushed - NumBytes;
1849 // Set the delta of movement of the returnaddr stackslot.
1850 // But only set if delta is greater than previous delta.
1851 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1852 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1856 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1858 SDValue RetAddrFrIdx;
1859 // Load return adress for tail calls.
1860 if (isTailCall && FPDiff)
1861 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1862 Is64Bit, FPDiff, dl);
1864 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1865 SmallVector<SDValue, 8> MemOpChains;
1868 // Walk the register/memloc assignments, inserting copies/loads. In the case
1869 // of tail call optimization arguments are handle later.
1870 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1871 CCValAssign &VA = ArgLocs[i];
1872 EVT RegVT = VA.getLocVT();
1873 SDValue Arg = Outs[i].Val;
1874 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1875 bool isByVal = Flags.isByVal();
1877 // Promote the value if needed.
1878 switch (VA.getLocInfo()) {
1879 default: llvm_unreachable("Unknown loc info!");
1880 case CCValAssign::Full: break;
1881 case CCValAssign::SExt:
1882 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1884 case CCValAssign::ZExt:
1885 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1887 case CCValAssign::AExt:
1888 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1889 // Special case: passing MMX values in XMM registers.
1890 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1891 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1892 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1894 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1896 case CCValAssign::BCvt:
1897 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1899 case CCValAssign::Indirect: {
1900 // Store the argument.
1901 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1902 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1903 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1904 PseudoSourceValue::getFixedStack(FI), 0,
1911 if (VA.isRegLoc()) {
1912 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1913 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1914 assert(VA.isMemLoc());
1915 if (StackPtr.getNode() == 0)
1916 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1917 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1918 dl, DAG, VA, Flags));
1922 if (!MemOpChains.empty())
1923 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1924 &MemOpChains[0], MemOpChains.size());
1926 // Build a sequence of copy-to-reg nodes chained together with token chain
1927 // and flag operands which copy the outgoing args into registers.
1929 // Tail call byval lowering might overwrite argument registers so in case of
1930 // tail call optimization the copies to registers are lowered later.
1932 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1933 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1934 RegsToPass[i].second, InFlag);
1935 InFlag = Chain.getValue(1);
1938 if (Subtarget->isPICStyleGOT()) {
1939 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1942 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1943 DAG.getNode(X86ISD::GlobalBaseReg,
1944 DebugLoc(), getPointerTy()),
1946 InFlag = Chain.getValue(1);
1948 // If we are tail calling and generating PIC/GOT style code load the
1949 // address of the callee into ECX. The value in ecx is used as target of
1950 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1951 // for tail calls on PIC/GOT architectures. Normally we would just put the
1952 // address of GOT into ebx and then call target@PLT. But for tail calls
1953 // ebx would be restored (since ebx is callee saved) before jumping to the
1956 // Note: The actual moving to ECX is done further down.
1957 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1958 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1959 !G->getGlobal()->hasProtectedVisibility())
1960 Callee = LowerGlobalAddress(Callee, DAG);
1961 else if (isa<ExternalSymbolSDNode>(Callee))
1962 Callee = LowerExternalSymbol(Callee, DAG);
1966 if (Is64Bit && isVarArg) {
1967 // From AMD64 ABI document:
1968 // For calls that may call functions that use varargs or stdargs
1969 // (prototype-less calls or calls to functions containing ellipsis (...) in
1970 // the declaration) %al is used as hidden argument to specify the number
1971 // of SSE registers used. The contents of %al do not need to match exactly
1972 // the number of registers, but must be an ubound on the number of SSE
1973 // registers used and is in the range 0 - 8 inclusive.
1975 // FIXME: Verify this on Win64
1976 // Count the number of XMM registers allocated.
1977 static const unsigned XMMArgRegs[] = {
1978 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1979 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1981 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1982 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1983 && "SSE registers cannot be used when SSE is disabled");
1985 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1986 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1987 InFlag = Chain.getValue(1);
1991 // For tail calls lower the arguments to the 'real' stack slot.
1993 // Force all the incoming stack arguments to be loaded from the stack
1994 // before any new outgoing arguments are stored to the stack, because the
1995 // outgoing stack slots may alias the incoming argument stack slots, and
1996 // the alias isn't otherwise explicit. This is slightly more conservative
1997 // than necessary, because it means that each store effectively depends
1998 // on every argument instead of just those arguments it would clobber.
1999 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2001 SmallVector<SDValue, 8> MemOpChains2;
2004 // Do not flag preceeding copytoreg stuff together with the following stuff.
2006 if (GuaranteedTailCallOpt) {
2007 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2008 CCValAssign &VA = ArgLocs[i];
2011 assert(VA.isMemLoc());
2012 SDValue Arg = Outs[i].Val;
2013 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2014 // Create frame index.
2015 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2016 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2017 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2018 FIN = DAG.getFrameIndex(FI, getPointerTy());
2020 if (Flags.isByVal()) {
2021 // Copy relative to framepointer.
2022 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2023 if (StackPtr.getNode() == 0)
2024 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2026 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2028 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2032 // Store relative to framepointer.
2033 MemOpChains2.push_back(
2034 DAG.getStore(ArgChain, dl, Arg, FIN,
2035 PseudoSourceValue::getFixedStack(FI), 0,
2041 if (!MemOpChains2.empty())
2042 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2043 &MemOpChains2[0], MemOpChains2.size());
2045 // Copy arguments to their registers.
2046 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2047 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2048 RegsToPass[i].second, InFlag);
2049 InFlag = Chain.getValue(1);
2053 // Store the return address to the appropriate stack slot.
2054 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2058 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2059 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2060 // In the 64-bit large code model, we have to make all calls
2061 // through a register, since the call instruction's 32-bit
2062 // pc-relative offset may not be large enough to hold the whole
2064 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2065 // If the callee is a GlobalAddress node (quite common, every direct call
2066 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2069 // We should use extra load for direct calls to dllimported functions in
2071 const GlobalValue *GV = G->getGlobal();
2072 if (!GV->hasDLLImportLinkage()) {
2073 unsigned char OpFlags = 0;
2075 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2076 // external symbols most go through the PLT in PIC mode. If the symbol
2077 // has hidden or protected visibility, or if it is static or local, then
2078 // we don't need to use the PLT - we can directly call it.
2079 if (Subtarget->isTargetELF() &&
2080 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2081 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2082 OpFlags = X86II::MO_PLT;
2083 } else if (Subtarget->isPICStyleStubAny() &&
2084 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2085 Subtarget->getDarwinVers() < 9) {
2086 // PC-relative references to external symbols should go through $stub,
2087 // unless we're building with the leopard linker or later, which
2088 // automatically synthesizes these stubs.
2089 OpFlags = X86II::MO_DARWIN_STUB;
2092 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2093 G->getOffset(), OpFlags);
2095 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2096 unsigned char OpFlags = 0;
2098 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2099 // symbols should go through the PLT.
2100 if (Subtarget->isTargetELF() &&
2101 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2102 OpFlags = X86II::MO_PLT;
2103 } else if (Subtarget->isPICStyleStubAny() &&
2104 Subtarget->getDarwinVers() < 9) {
2105 // PC-relative references to external symbols should go through $stub,
2106 // unless we're building with the leopard linker or later, which
2107 // automatically synthesizes these stubs.
2108 OpFlags = X86II::MO_DARWIN_STUB;
2111 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2115 // Returns a chain & a flag for retval copy to use.
2116 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2117 SmallVector<SDValue, 8> Ops;
2119 if (!IsSibcall && isTailCall) {
2120 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2121 DAG.getIntPtrConstant(0, true), InFlag);
2122 InFlag = Chain.getValue(1);
2125 Ops.push_back(Chain);
2126 Ops.push_back(Callee);
2129 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2131 // Add argument registers to the end of the list so that they are known live
2133 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2134 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2135 RegsToPass[i].second.getValueType()));
2137 // Add an implicit use GOT pointer in EBX.
2138 if (!isTailCall && Subtarget->isPICStyleGOT())
2139 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2141 // Add an implicit use of AL for x86 vararg functions.
2142 if (Is64Bit && isVarArg)
2143 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2145 if (InFlag.getNode())
2146 Ops.push_back(InFlag);
2150 //// If this is the first return lowered for this function, add the regs
2151 //// to the liveout set for the function.
2152 // This isn't right, although it's probably harmless on x86; liveouts
2153 // should be computed from returns not tail calls. Consider a void
2154 // function making a tail call to a function returning int.
2155 return DAG.getNode(X86ISD::TC_RETURN, dl,
2156 NodeTys, &Ops[0], Ops.size());
2159 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2160 InFlag = Chain.getValue(1);
2162 // Create the CALLSEQ_END node.
2163 unsigned NumBytesForCalleeToPush;
2164 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2165 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2166 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2167 // If this is a call to a struct-return function, the callee
2168 // pops the hidden struct pointer, so we have to push it back.
2169 // This is common for Darwin/X86, Linux & Mingw32 targets.
2170 NumBytesForCalleeToPush = 4;
2172 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2174 // Returns a flag for retval copy to use.
2176 Chain = DAG.getCALLSEQ_END(Chain,
2177 DAG.getIntPtrConstant(NumBytes, true),
2178 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2181 InFlag = Chain.getValue(1);
2184 // Handle result values, copying them out of physregs into vregs that we
2186 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2187 Ins, dl, DAG, InVals);
2191 //===----------------------------------------------------------------------===//
2192 // Fast Calling Convention (tail call) implementation
2193 //===----------------------------------------------------------------------===//
2195 // Like std call, callee cleans arguments, convention except that ECX is
2196 // reserved for storing the tail called function address. Only 2 registers are
2197 // free for argument passing (inreg). Tail call optimization is performed
2199 // * tailcallopt is enabled
2200 // * caller/callee are fastcc
2201 // On X86_64 architecture with GOT-style position independent code only local
2202 // (within module) calls are supported at the moment.
2203 // To keep the stack aligned according to platform abi the function
2204 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2205 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2206 // If a tail called function callee has more arguments than the caller the
2207 // caller needs to make sure that there is room to move the RETADDR to. This is
2208 // achieved by reserving an area the size of the argument delta right after the
2209 // original REtADDR, but before the saved framepointer or the spilled registers
2210 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2222 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2223 /// for a 16 byte align requirement.
2225 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2226 SelectionDAG& DAG) const {
2227 MachineFunction &MF = DAG.getMachineFunction();
2228 const TargetMachine &TM = MF.getTarget();
2229 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2230 unsigned StackAlignment = TFI.getStackAlignment();
2231 uint64_t AlignMask = StackAlignment - 1;
2232 int64_t Offset = StackSize;
2233 uint64_t SlotSize = TD->getPointerSize();
2234 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2235 // Number smaller than 12 so just add the difference.
2236 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2238 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2239 Offset = ((~AlignMask) & Offset) + StackAlignment +
2240 (StackAlignment-SlotSize);
2245 /// MatchingStackOffset - Return true if the given stack call argument is
2246 /// already available in the same position (relatively) of the caller's
2247 /// incoming argument stack.
2249 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2250 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2251 const X86InstrInfo *TII) {
2252 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2254 if (Arg.getOpcode() == ISD::CopyFromReg) {
2255 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2256 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2258 MachineInstr *Def = MRI->getVRegDef(VR);
2261 if (!Flags.isByVal()) {
2262 if (!TII->isLoadFromStackSlot(Def, FI))
2265 unsigned Opcode = Def->getOpcode();
2266 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2267 Def->getOperand(1).isFI()) {
2268 FI = Def->getOperand(1).getIndex();
2269 Bytes = Flags.getByValSize();
2273 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2274 if (Flags.isByVal())
2275 // ByVal argument is passed in as a pointer but it's now being
2276 // dereferenced. e.g.
2277 // define @foo(%struct.X* %A) {
2278 // tail call @bar(%struct.X* byval %A)
2281 SDValue Ptr = Ld->getBasePtr();
2282 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2285 FI = FINode->getIndex();
2289 assert(FI != INT_MAX);
2290 if (!MFI->isFixedObjectIndex(FI))
2292 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2295 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2296 /// for tail call optimization. Targets which want to do tail call
2297 /// optimization should implement this function.
2299 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2300 CallingConv::ID CalleeCC,
2302 bool isCalleeStructRet,
2303 bool isCallerStructRet,
2304 const SmallVectorImpl<ISD::OutputArg> &Outs,
2305 const SmallVectorImpl<ISD::InputArg> &Ins,
2306 SelectionDAG& DAG) const {
2307 if (!IsTailCallConvention(CalleeCC) &&
2308 CalleeCC != CallingConv::C)
2311 // If -tailcallopt is specified, make fastcc functions tail-callable.
2312 const MachineFunction &MF = DAG.getMachineFunction();
2313 const Function *CallerF = DAG.getMachineFunction().getFunction();
2314 CallingConv::ID CallerCC = CallerF->getCallingConv();
2315 bool CCMatch = CallerCC == CalleeCC;
2317 if (GuaranteedTailCallOpt) {
2318 if (IsTailCallConvention(CalleeCC) && CCMatch)
2323 // Look for obvious safe cases to perform tail call optimization that do not
2324 // require ABI changes. This is what gcc calls sibcall.
2326 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2327 // emit a special epilogue.
2328 if (RegInfo->needsStackRealignment(MF))
2331 // Do not sibcall optimize vararg calls unless the call site is not passing any
2333 if (isVarArg && !Outs.empty())
2336 // Also avoid sibcall optimization if either caller or callee uses struct
2337 // return semantics.
2338 if (isCalleeStructRet || isCallerStructRet)
2341 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2342 // Therefore if it's not used by the call it is not safe to optimize this into
2344 bool Unused = false;
2345 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2352 SmallVector<CCValAssign, 16> RVLocs;
2353 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2354 RVLocs, *DAG.getContext());
2355 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2356 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2357 CCValAssign &VA = RVLocs[i];
2358 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2363 // If the calling conventions do not match, then we'd better make sure the
2364 // results are returned in the same way as what the caller expects.
2366 SmallVector<CCValAssign, 16> RVLocs1;
2367 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2368 RVLocs1, *DAG.getContext());
2369 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2371 SmallVector<CCValAssign, 16> RVLocs2;
2372 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2373 RVLocs2, *DAG.getContext());
2374 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2376 if (RVLocs1.size() != RVLocs2.size())
2378 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2379 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2381 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2383 if (RVLocs1[i].isRegLoc()) {
2384 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2387 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2393 // If the callee takes no arguments then go on to check the results of the
2395 if (!Outs.empty()) {
2396 // Check if stack adjustment is needed. For now, do not do this if any
2397 // argument is passed on the stack.
2398 SmallVector<CCValAssign, 16> ArgLocs;
2399 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2400 ArgLocs, *DAG.getContext());
2401 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2402 if (CCInfo.getNextStackOffset()) {
2403 MachineFunction &MF = DAG.getMachineFunction();
2404 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2406 if (Subtarget->isTargetWin64())
2407 // Win64 ABI has additional complications.
2410 // Check if the arguments are already laid out in the right way as
2411 // the caller's fixed stack objects.
2412 MachineFrameInfo *MFI = MF.getFrameInfo();
2413 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2414 const X86InstrInfo *TII =
2415 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2416 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2417 CCValAssign &VA = ArgLocs[i];
2418 SDValue Arg = Outs[i].Val;
2419 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2420 if (VA.getLocInfo() == CCValAssign::Indirect)
2422 if (!VA.isRegLoc()) {
2423 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2430 // If the tailcall address may be in a register, then make sure it's
2431 // possible to register allocate for it. In 32-bit, the call address can
2432 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2433 // callee-saved registers are restored. In 64-bit, it's RAX, RCX, RDX, RSI,
2434 // RDI, R8, R9, R11.
2435 if (!isa<GlobalAddressSDNode>(Callee) &&
2436 !isa<ExternalSymbolSDNode>(Callee)) {
2437 unsigned Limit = Subtarget->is64Bit() ? 8 : 3;
2438 unsigned NumInRegs = 0;
2439 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2440 CCValAssign &VA = ArgLocs[i];
2441 if (VA.isRegLoc()) {
2442 if (++NumInRegs == Limit)
2453 X86TargetLowering::createFastISel(MachineFunction &mf,
2454 DenseMap<const Value *, unsigned> &vm,
2455 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2456 DenseMap<const AllocaInst *, int> &am,
2457 std::vector<std::pair<MachineInstr*, unsigned> > &pn
2459 , SmallSet<const Instruction *, 8> &cil
2462 return X86::createFastISel(mf, vm, bm, am, pn
2470 //===----------------------------------------------------------------------===//
2471 // Other Lowering Hooks
2472 //===----------------------------------------------------------------------===//
2475 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2476 MachineFunction &MF = DAG.getMachineFunction();
2477 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2478 int ReturnAddrIndex = FuncInfo->getRAIndex();
2480 if (ReturnAddrIndex == 0) {
2481 // Set up a frame object for the return address.
2482 uint64_t SlotSize = TD->getPointerSize();
2483 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2485 FuncInfo->setRAIndex(ReturnAddrIndex);
2488 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2492 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2493 bool hasSymbolicDisplacement) {
2494 // Offset should fit into 32 bit immediate field.
2495 if (!isInt<32>(Offset))
2498 // If we don't have a symbolic displacement - we don't have any extra
2500 if (!hasSymbolicDisplacement)
2503 // FIXME: Some tweaks might be needed for medium code model.
2504 if (M != CodeModel::Small && M != CodeModel::Kernel)
2507 // For small code model we assume that latest object is 16MB before end of 31
2508 // bits boundary. We may also accept pretty large negative constants knowing
2509 // that all objects are in the positive half of address space.
2510 if (M == CodeModel::Small && Offset < 16*1024*1024)
2513 // For kernel code model we know that all object resist in the negative half
2514 // of 32bits address space. We may not accept negative offsets, since they may
2515 // be just off and we may accept pretty large positive ones.
2516 if (M == CodeModel::Kernel && Offset > 0)
2522 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2523 /// specific condition code, returning the condition code and the LHS/RHS of the
2524 /// comparison to make.
2525 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2526 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2528 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2529 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2530 // X > -1 -> X == 0, jump !sign.
2531 RHS = DAG.getConstant(0, RHS.getValueType());
2532 return X86::COND_NS;
2533 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2534 // X < 0 -> X == 0, jump on sign.
2536 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2538 RHS = DAG.getConstant(0, RHS.getValueType());
2539 return X86::COND_LE;
2543 switch (SetCCOpcode) {
2544 default: llvm_unreachable("Invalid integer condition!");
2545 case ISD::SETEQ: return X86::COND_E;
2546 case ISD::SETGT: return X86::COND_G;
2547 case ISD::SETGE: return X86::COND_GE;
2548 case ISD::SETLT: return X86::COND_L;
2549 case ISD::SETLE: return X86::COND_LE;
2550 case ISD::SETNE: return X86::COND_NE;
2551 case ISD::SETULT: return X86::COND_B;
2552 case ISD::SETUGT: return X86::COND_A;
2553 case ISD::SETULE: return X86::COND_BE;
2554 case ISD::SETUGE: return X86::COND_AE;
2558 // First determine if it is required or is profitable to flip the operands.
2560 // If LHS is a foldable load, but RHS is not, flip the condition.
2561 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2562 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2563 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2564 std::swap(LHS, RHS);
2567 switch (SetCCOpcode) {
2573 std::swap(LHS, RHS);
2577 // On a floating point condition, the flags are set as follows:
2579 // 0 | 0 | 0 | X > Y
2580 // 0 | 0 | 1 | X < Y
2581 // 1 | 0 | 0 | X == Y
2582 // 1 | 1 | 1 | unordered
2583 switch (SetCCOpcode) {
2584 default: llvm_unreachable("Condcode should be pre-legalized away");
2586 case ISD::SETEQ: return X86::COND_E;
2587 case ISD::SETOLT: // flipped
2589 case ISD::SETGT: return X86::COND_A;
2590 case ISD::SETOLE: // flipped
2592 case ISD::SETGE: return X86::COND_AE;
2593 case ISD::SETUGT: // flipped
2595 case ISD::SETLT: return X86::COND_B;
2596 case ISD::SETUGE: // flipped
2598 case ISD::SETLE: return X86::COND_BE;
2600 case ISD::SETNE: return X86::COND_NE;
2601 case ISD::SETUO: return X86::COND_P;
2602 case ISD::SETO: return X86::COND_NP;
2604 case ISD::SETUNE: return X86::COND_INVALID;
2608 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2609 /// code. Current x86 isa includes the following FP cmov instructions:
2610 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2611 static bool hasFPCMov(unsigned X86CC) {
2627 /// isFPImmLegal - Returns true if the target can instruction select the
2628 /// specified FP immediate natively. If false, the legalizer will
2629 /// materialize the FP immediate as a load from a constant pool.
2630 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2631 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2632 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2638 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2639 /// the specified range (L, H].
2640 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2641 return (Val < 0) || (Val >= Low && Val < Hi);
2644 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2645 /// specified value.
2646 static bool isUndefOrEqual(int Val, int CmpVal) {
2647 if (Val < 0 || Val == CmpVal)
2652 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2653 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2654 /// the second operand.
2655 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2656 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2657 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2658 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2659 return (Mask[0] < 2 && Mask[1] < 2);
2663 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2664 SmallVector<int, 8> M;
2666 return ::isPSHUFDMask(M, N->getValueType(0));
2669 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2670 /// is suitable for input to PSHUFHW.
2671 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2672 if (VT != MVT::v8i16)
2675 // Lower quadword copied in order or undef.
2676 for (int i = 0; i != 4; ++i)
2677 if (Mask[i] >= 0 && Mask[i] != i)
2680 // Upper quadword shuffled.
2681 for (int i = 4; i != 8; ++i)
2682 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2688 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2689 SmallVector<int, 8> M;
2691 return ::isPSHUFHWMask(M, N->getValueType(0));
2694 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2695 /// is suitable for input to PSHUFLW.
2696 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2697 if (VT != MVT::v8i16)
2700 // Upper quadword copied in order.
2701 for (int i = 4; i != 8; ++i)
2702 if (Mask[i] >= 0 && Mask[i] != i)
2705 // Lower quadword shuffled.
2706 for (int i = 0; i != 4; ++i)
2713 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2714 SmallVector<int, 8> M;
2716 return ::isPSHUFLWMask(M, N->getValueType(0));
2719 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2720 /// is suitable for input to PALIGNR.
2721 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2723 int i, e = VT.getVectorNumElements();
2725 // Do not handle v2i64 / v2f64 shuffles with palignr.
2726 if (e < 4 || !hasSSSE3)
2729 for (i = 0; i != e; ++i)
2733 // All undef, not a palignr.
2737 // Determine if it's ok to perform a palignr with only the LHS, since we
2738 // don't have access to the actual shuffle elements to see if RHS is undef.
2739 bool Unary = Mask[i] < (int)e;
2740 bool NeedsUnary = false;
2742 int s = Mask[i] - i;
2744 // Check the rest of the elements to see if they are consecutive.
2745 for (++i; i != e; ++i) {
2750 Unary = Unary && (m < (int)e);
2751 NeedsUnary = NeedsUnary || (m < s);
2753 if (NeedsUnary && !Unary)
2755 if (Unary && m != ((s+i) & (e-1)))
2757 if (!Unary && m != (s+i))
2763 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2764 SmallVector<int, 8> M;
2766 return ::isPALIGNRMask(M, N->getValueType(0), true);
2769 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2770 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2771 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2772 int NumElems = VT.getVectorNumElements();
2773 if (NumElems != 2 && NumElems != 4)
2776 int Half = NumElems / 2;
2777 for (int i = 0; i < Half; ++i)
2778 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2780 for (int i = Half; i < NumElems; ++i)
2781 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2787 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2788 SmallVector<int, 8> M;
2790 return ::isSHUFPMask(M, N->getValueType(0));
2793 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2794 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2795 /// half elements to come from vector 1 (which would equal the dest.) and
2796 /// the upper half to come from vector 2.
2797 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2798 int NumElems = VT.getVectorNumElements();
2800 if (NumElems != 2 && NumElems != 4)
2803 int Half = NumElems / 2;
2804 for (int i = 0; i < Half; ++i)
2805 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2807 for (int i = Half; i < NumElems; ++i)
2808 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2813 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2814 SmallVector<int, 8> M;
2816 return isCommutedSHUFPMask(M, N->getValueType(0));
2819 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2820 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2821 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2822 if (N->getValueType(0).getVectorNumElements() != 4)
2825 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2826 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2827 isUndefOrEqual(N->getMaskElt(1), 7) &&
2828 isUndefOrEqual(N->getMaskElt(2), 2) &&
2829 isUndefOrEqual(N->getMaskElt(3), 3);
2832 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2833 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2835 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2836 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2841 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2842 isUndefOrEqual(N->getMaskElt(1), 3) &&
2843 isUndefOrEqual(N->getMaskElt(2), 2) &&
2844 isUndefOrEqual(N->getMaskElt(3), 3);
2847 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2848 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2849 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2850 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2852 if (NumElems != 2 && NumElems != 4)
2855 for (unsigned i = 0; i < NumElems/2; ++i)
2856 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2859 for (unsigned i = NumElems/2; i < NumElems; ++i)
2860 if (!isUndefOrEqual(N->getMaskElt(i), i))
2866 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2867 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2868 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2869 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2871 if (NumElems != 2 && NumElems != 4)
2874 for (unsigned i = 0; i < NumElems/2; ++i)
2875 if (!isUndefOrEqual(N->getMaskElt(i), i))
2878 for (unsigned i = 0; i < NumElems/2; ++i)
2879 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2885 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2886 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2887 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2888 bool V2IsSplat = false) {
2889 int NumElts = VT.getVectorNumElements();
2890 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2893 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2895 int BitI1 = Mask[i+1];
2896 if (!isUndefOrEqual(BitI, j))
2899 if (!isUndefOrEqual(BitI1, NumElts))
2902 if (!isUndefOrEqual(BitI1, j + NumElts))
2909 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2910 SmallVector<int, 8> M;
2912 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2915 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2916 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2917 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2918 bool V2IsSplat = false) {
2919 int NumElts = VT.getVectorNumElements();
2920 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2923 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2925 int BitI1 = Mask[i+1];
2926 if (!isUndefOrEqual(BitI, j + NumElts/2))
2929 if (isUndefOrEqual(BitI1, NumElts))
2932 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2939 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2940 SmallVector<int, 8> M;
2942 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2945 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2946 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2948 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2949 int NumElems = VT.getVectorNumElements();
2950 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2953 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2955 int BitI1 = Mask[i+1];
2956 if (!isUndefOrEqual(BitI, j))
2958 if (!isUndefOrEqual(BitI1, j))
2964 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2965 SmallVector<int, 8> M;
2967 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2970 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2971 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2973 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2974 int NumElems = VT.getVectorNumElements();
2975 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2978 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2980 int BitI1 = Mask[i+1];
2981 if (!isUndefOrEqual(BitI, j))
2983 if (!isUndefOrEqual(BitI1, j))
2989 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2990 SmallVector<int, 8> M;
2992 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2995 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2996 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2997 /// MOVSD, and MOVD, i.e. setting the lowest element.
2998 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2999 if (VT.getVectorElementType().getSizeInBits() < 32)
3002 int NumElts = VT.getVectorNumElements();
3004 if (!isUndefOrEqual(Mask[0], NumElts))
3007 for (int i = 1; i < NumElts; ++i)
3008 if (!isUndefOrEqual(Mask[i], i))
3014 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3015 SmallVector<int, 8> M;
3017 return ::isMOVLMask(M, N->getValueType(0));
3020 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3021 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3022 /// element of vector 2 and the other elements to come from vector 1 in order.
3023 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3024 bool V2IsSplat = false, bool V2IsUndef = false) {
3025 int NumOps = VT.getVectorNumElements();
3026 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3029 if (!isUndefOrEqual(Mask[0], 0))
3032 for (int i = 1; i < NumOps; ++i)
3033 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3034 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3035 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3041 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3042 bool V2IsUndef = false) {
3043 SmallVector<int, 8> M;
3045 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3048 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3049 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3050 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3051 if (N->getValueType(0).getVectorNumElements() != 4)
3054 // Expect 1, 1, 3, 3
3055 for (unsigned i = 0; i < 2; ++i) {
3056 int Elt = N->getMaskElt(i);
3057 if (Elt >= 0 && Elt != 1)
3062 for (unsigned i = 2; i < 4; ++i) {
3063 int Elt = N->getMaskElt(i);
3064 if (Elt >= 0 && Elt != 3)
3069 // Don't use movshdup if it can be done with a shufps.
3070 // FIXME: verify that matching u, u, 3, 3 is what we want.
3074 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3075 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3076 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3077 if (N->getValueType(0).getVectorNumElements() != 4)
3080 // Expect 0, 0, 2, 2
3081 for (unsigned i = 0; i < 2; ++i)
3082 if (N->getMaskElt(i) > 0)
3086 for (unsigned i = 2; i < 4; ++i) {
3087 int Elt = N->getMaskElt(i);
3088 if (Elt >= 0 && Elt != 2)
3093 // Don't use movsldup if it can be done with a shufps.
3097 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3098 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3099 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3100 int e = N->getValueType(0).getVectorNumElements() / 2;
3102 for (int i = 0; i < e; ++i)
3103 if (!isUndefOrEqual(N->getMaskElt(i), i))
3105 for (int i = 0; i < e; ++i)
3106 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3111 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3112 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3113 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3114 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3115 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3117 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3119 for (int i = 0; i < NumOperands; ++i) {
3120 int Val = SVOp->getMaskElt(NumOperands-i-1);
3121 if (Val < 0) Val = 0;
3122 if (Val >= NumOperands) Val -= NumOperands;
3124 if (i != NumOperands - 1)
3130 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3131 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3132 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3133 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3135 // 8 nodes, but we only care about the last 4.
3136 for (unsigned i = 7; i >= 4; --i) {
3137 int Val = SVOp->getMaskElt(i);
3146 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3147 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3148 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3149 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3151 // 8 nodes, but we only care about the first 4.
3152 for (int i = 3; i >= 0; --i) {
3153 int Val = SVOp->getMaskElt(i);
3162 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3163 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3164 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3165 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3166 EVT VVT = N->getValueType(0);
3167 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3171 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3172 Val = SVOp->getMaskElt(i);
3176 return (Val - i) * EltSize;
3179 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3181 bool X86::isZeroNode(SDValue Elt) {
3182 return ((isa<ConstantSDNode>(Elt) &&
3183 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3184 (isa<ConstantFPSDNode>(Elt) &&
3185 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3188 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3189 /// their permute mask.
3190 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3191 SelectionDAG &DAG) {
3192 EVT VT = SVOp->getValueType(0);
3193 unsigned NumElems = VT.getVectorNumElements();
3194 SmallVector<int, 8> MaskVec;
3196 for (unsigned i = 0; i != NumElems; ++i) {
3197 int idx = SVOp->getMaskElt(i);
3199 MaskVec.push_back(idx);
3200 else if (idx < (int)NumElems)
3201 MaskVec.push_back(idx + NumElems);
3203 MaskVec.push_back(idx - NumElems);
3205 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3206 SVOp->getOperand(0), &MaskVec[0]);
3209 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3210 /// the two vector operands have swapped position.
3211 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3212 unsigned NumElems = VT.getVectorNumElements();
3213 for (unsigned i = 0; i != NumElems; ++i) {
3217 else if (idx < (int)NumElems)
3218 Mask[i] = idx + NumElems;
3220 Mask[i] = idx - NumElems;
3224 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3225 /// match movhlps. The lower half elements should come from upper half of
3226 /// V1 (and in order), and the upper half elements should come from the upper
3227 /// half of V2 (and in order).
3228 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3229 if (Op->getValueType(0).getVectorNumElements() != 4)
3231 for (unsigned i = 0, e = 2; i != e; ++i)
3232 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3234 for (unsigned i = 2; i != 4; ++i)
3235 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3240 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3241 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3243 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3244 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3246 N = N->getOperand(0).getNode();
3247 if (!ISD::isNON_EXTLoad(N))
3250 *LD = cast<LoadSDNode>(N);
3254 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3255 /// match movlp{s|d}. The lower half elements should come from lower half of
3256 /// V1 (and in order), and the upper half elements should come from the upper
3257 /// half of V2 (and in order). And since V1 will become the source of the
3258 /// MOVLP, it must be either a vector load or a scalar load to vector.
3259 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3260 ShuffleVectorSDNode *Op) {
3261 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3263 // Is V2 is a vector load, don't do this transformation. We will try to use
3264 // load folding shufps op.
3265 if (ISD::isNON_EXTLoad(V2))
3268 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3270 if (NumElems != 2 && NumElems != 4)
3272 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3273 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3275 for (unsigned i = NumElems/2; i != NumElems; ++i)
3276 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3281 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3283 static bool isSplatVector(SDNode *N) {
3284 if (N->getOpcode() != ISD::BUILD_VECTOR)
3287 SDValue SplatValue = N->getOperand(0);
3288 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3289 if (N->getOperand(i) != SplatValue)
3294 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3295 /// to an zero vector.
3296 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3297 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3298 SDValue V1 = N->getOperand(0);
3299 SDValue V2 = N->getOperand(1);
3300 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3301 for (unsigned i = 0; i != NumElems; ++i) {
3302 int Idx = N->getMaskElt(i);
3303 if (Idx >= (int)NumElems) {
3304 unsigned Opc = V2.getOpcode();
3305 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3307 if (Opc != ISD::BUILD_VECTOR ||
3308 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3310 } else if (Idx >= 0) {
3311 unsigned Opc = V1.getOpcode();
3312 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3314 if (Opc != ISD::BUILD_VECTOR ||
3315 !X86::isZeroNode(V1.getOperand(Idx)))
3322 /// getZeroVector - Returns a vector of specified type with all zero elements.
3324 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3326 assert(VT.isVector() && "Expected a vector type");
3328 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3329 // type. This ensures they get CSE'd.
3331 if (VT.getSizeInBits() == 64) { // MMX
3332 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3333 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3334 } else if (HasSSE2) { // SSE2
3335 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3336 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3338 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3339 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3341 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3344 /// getOnesVector - Returns a vector of specified type with all bits set.
3346 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3347 assert(VT.isVector() && "Expected a vector type");
3349 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3350 // type. This ensures they get CSE'd.
3351 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3353 if (VT.getSizeInBits() == 64) // MMX
3354 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3356 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3357 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3361 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3362 /// that point to V2 points to its first element.
3363 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3364 EVT VT = SVOp->getValueType(0);
3365 unsigned NumElems = VT.getVectorNumElements();
3367 bool Changed = false;
3368 SmallVector<int, 8> MaskVec;
3369 SVOp->getMask(MaskVec);
3371 for (unsigned i = 0; i != NumElems; ++i) {
3372 if (MaskVec[i] > (int)NumElems) {
3373 MaskVec[i] = NumElems;
3378 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3379 SVOp->getOperand(1), &MaskVec[0]);
3380 return SDValue(SVOp, 0);
3383 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3384 /// operation of specified width.
3385 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3387 unsigned NumElems = VT.getVectorNumElements();
3388 SmallVector<int, 8> Mask;
3389 Mask.push_back(NumElems);
3390 for (unsigned i = 1; i != NumElems; ++i)
3392 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3395 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3396 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3398 unsigned NumElems = VT.getVectorNumElements();
3399 SmallVector<int, 8> Mask;
3400 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3402 Mask.push_back(i + NumElems);
3404 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3407 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3408 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3410 unsigned NumElems = VT.getVectorNumElements();
3411 unsigned Half = NumElems/2;
3412 SmallVector<int, 8> Mask;
3413 for (unsigned i = 0; i != Half; ++i) {
3414 Mask.push_back(i + Half);
3415 Mask.push_back(i + NumElems + Half);
3417 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3420 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3421 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3423 if (SV->getValueType(0).getVectorNumElements() <= 4)
3424 return SDValue(SV, 0);
3426 EVT PVT = MVT::v4f32;
3427 EVT VT = SV->getValueType(0);
3428 DebugLoc dl = SV->getDebugLoc();
3429 SDValue V1 = SV->getOperand(0);
3430 int NumElems = VT.getVectorNumElements();
3431 int EltNo = SV->getSplatIndex();
3433 // unpack elements to the correct location
3434 while (NumElems > 4) {
3435 if (EltNo < NumElems/2) {
3436 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3438 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3439 EltNo -= NumElems/2;
3444 // Perform the splat.
3445 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3446 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3447 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3448 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3451 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3452 /// vector of zero or undef vector. This produces a shuffle where the low
3453 /// element of V2 is swizzled into the zero/undef vector, landing at element
3454 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3455 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3456 bool isZero, bool HasSSE2,
3457 SelectionDAG &DAG) {
3458 EVT VT = V2.getValueType();
3460 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3461 unsigned NumElems = VT.getVectorNumElements();
3462 SmallVector<int, 16> MaskVec;
3463 for (unsigned i = 0; i != NumElems; ++i)
3464 // If this is the insertion idx, put the low elt of V2 here.
3465 MaskVec.push_back(i == Idx ? NumElems : i);
3466 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3469 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3470 /// a shuffle that is zero.
3472 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3473 bool Low, SelectionDAG &DAG) {
3474 unsigned NumZeros = 0;
3475 for (int i = 0; i < NumElems; ++i) {
3476 unsigned Index = Low ? i : NumElems-i-1;
3477 int Idx = SVOp->getMaskElt(Index);
3482 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3483 if (Elt.getNode() && X86::isZeroNode(Elt))
3491 /// isVectorShift - Returns true if the shuffle can be implemented as a
3492 /// logical left or right shift of a vector.
3493 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3494 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3495 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3496 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3499 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3502 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3506 bool SeenV1 = false;
3507 bool SeenV2 = false;
3508 for (unsigned i = NumZeros; i < NumElems; ++i) {
3509 unsigned Val = isLeft ? (i - NumZeros) : i;
3510 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3513 unsigned Idx = (unsigned) Idx_;
3523 if (SeenV1 && SeenV2)
3526 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3532 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3534 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3535 unsigned NumNonZero, unsigned NumZero,
3537 const TargetLowering &TLI) {
3541 DebugLoc dl = Op.getDebugLoc();
3544 for (unsigned i = 0; i < 16; ++i) {
3545 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3546 if (ThisIsNonZero && First) {
3548 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3550 V = DAG.getUNDEF(MVT::v8i16);
3555 SDValue ThisElt(0, 0), LastElt(0, 0);
3556 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3557 if (LastIsNonZero) {
3558 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3559 MVT::i16, Op.getOperand(i-1));
3561 if (ThisIsNonZero) {
3562 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3563 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3564 ThisElt, DAG.getConstant(8, MVT::i8));
3566 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3570 if (ThisElt.getNode())
3571 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3572 DAG.getIntPtrConstant(i/2));
3576 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3579 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3581 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3582 unsigned NumNonZero, unsigned NumZero,
3584 const TargetLowering &TLI) {
3588 DebugLoc dl = Op.getDebugLoc();
3591 for (unsigned i = 0; i < 8; ++i) {
3592 bool isNonZero = (NonZeros & (1 << i)) != 0;
3596 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3598 V = DAG.getUNDEF(MVT::v8i16);
3601 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3602 MVT::v8i16, V, Op.getOperand(i),
3603 DAG.getIntPtrConstant(i));
3610 /// getVShift - Return a vector logical shift node.
3612 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3613 unsigned NumBits, SelectionDAG &DAG,
3614 const TargetLowering &TLI, DebugLoc dl) {
3615 bool isMMX = VT.getSizeInBits() == 64;
3616 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3617 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3618 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3619 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3620 DAG.getNode(Opc, dl, ShVT, SrcOp,
3621 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3625 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3626 SelectionDAG &DAG) const {
3628 // Check if the scalar load can be widened into a vector load. And if
3629 // the address is "base + cst" see if the cst can be "absorbed" into
3630 // the shuffle mask.
3631 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3632 SDValue Ptr = LD->getBasePtr();
3633 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3635 EVT PVT = LD->getValueType(0);
3636 if (PVT != MVT::i32 && PVT != MVT::f32)
3641 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3642 FI = FINode->getIndex();
3644 } else if (Ptr.getOpcode() == ISD::ADD &&
3645 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3646 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3647 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3648 Offset = Ptr.getConstantOperandVal(1);
3649 Ptr = Ptr.getOperand(0);
3654 SDValue Chain = LD->getChain();
3655 // Make sure the stack object alignment is at least 16.
3656 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3657 if (DAG.InferPtrAlignment(Ptr) < 16) {
3658 if (MFI->isFixedObjectIndex(FI)) {
3659 // Can't change the alignment. FIXME: It's possible to compute
3660 // the exact stack offset and reference FI + adjust offset instead.
3661 // If someone *really* cares about this. That's the way to implement it.
3664 MFI->setObjectAlignment(FI, 16);
3668 // (Offset % 16) must be multiple of 4. Then address is then
3669 // Ptr + (Offset & ~15).
3672 if ((Offset % 16) & 3)
3674 int64_t StartOffset = Offset & ~15;
3676 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3677 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3679 int EltNo = (Offset - StartOffset) >> 2;
3680 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3681 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3682 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3684 // Canonicalize it to a v4i32 shuffle.
3685 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3686 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3687 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3688 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3694 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3695 /// vector of type 'VT', see if the elements can be replaced by a single large
3696 /// load which has the same value as a build_vector whose operands are 'elts'.
3698 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3700 /// FIXME: we'd also like to handle the case where the last elements are zero
3701 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3702 /// There's even a handy isZeroNode for that purpose.
3703 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3704 DebugLoc &dl, SelectionDAG &DAG) {
3705 EVT EltVT = VT.getVectorElementType();
3706 unsigned NumElems = Elts.size();
3708 LoadSDNode *LDBase = NULL;
3709 unsigned LastLoadedElt = -1U;
3711 // For each element in the initializer, see if we've found a load or an undef.
3712 // If we don't find an initial load element, or later load elements are
3713 // non-consecutive, bail out.
3714 for (unsigned i = 0; i < NumElems; ++i) {
3715 SDValue Elt = Elts[i];
3717 if (!Elt.getNode() ||
3718 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3721 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3723 LDBase = cast<LoadSDNode>(Elt.getNode());
3727 if (Elt.getOpcode() == ISD::UNDEF)
3730 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3731 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3736 // If we have found an entire vector of loads and undefs, then return a large
3737 // load of the entire vector width starting at the base pointer. If we found
3738 // consecutive loads for the low half, generate a vzext_load node.
3739 if (LastLoadedElt == NumElems - 1) {
3740 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3741 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3742 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3743 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3744 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3745 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3746 LDBase->isVolatile(), LDBase->isNonTemporal(),
3747 LDBase->getAlignment());
3748 } else if (NumElems == 4 && LastLoadedElt == 1) {
3749 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3750 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3751 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3752 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3758 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3759 DebugLoc dl = Op.getDebugLoc();
3760 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3761 if (ISD::isBuildVectorAllZeros(Op.getNode())
3762 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3763 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3764 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3765 // eliminated on x86-32 hosts.
3766 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3769 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3770 return getOnesVector(Op.getValueType(), DAG, dl);
3771 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3774 EVT VT = Op.getValueType();
3775 EVT ExtVT = VT.getVectorElementType();
3776 unsigned EVTBits = ExtVT.getSizeInBits();
3778 unsigned NumElems = Op.getNumOperands();
3779 unsigned NumZero = 0;
3780 unsigned NumNonZero = 0;
3781 unsigned NonZeros = 0;
3782 bool IsAllConstants = true;
3783 SmallSet<SDValue, 8> Values;
3784 for (unsigned i = 0; i < NumElems; ++i) {
3785 SDValue Elt = Op.getOperand(i);
3786 if (Elt.getOpcode() == ISD::UNDEF)
3789 if (Elt.getOpcode() != ISD::Constant &&
3790 Elt.getOpcode() != ISD::ConstantFP)
3791 IsAllConstants = false;
3792 if (X86::isZeroNode(Elt))
3795 NonZeros |= (1 << i);
3800 if (NumNonZero == 0) {
3801 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3802 return DAG.getUNDEF(VT);
3805 // Special case for single non-zero, non-undef, element.
3806 if (NumNonZero == 1) {
3807 unsigned Idx = CountTrailingZeros_32(NonZeros);
3808 SDValue Item = Op.getOperand(Idx);
3810 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3811 // the value are obviously zero, truncate the value to i32 and do the
3812 // insertion that way. Only do this if the value is non-constant or if the
3813 // value is a constant being inserted into element 0. It is cheaper to do
3814 // a constant pool load than it is to do a movd + shuffle.
3815 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3816 (!IsAllConstants || Idx == 0)) {
3817 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3818 // Handle MMX and SSE both.
3819 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3820 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3822 // Truncate the value (which may itself be a constant) to i32, and
3823 // convert it to a vector with movd (S2V+shuffle to zero extend).
3824 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3825 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3826 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3827 Subtarget->hasSSE2(), DAG);
3829 // Now we have our 32-bit value zero extended in the low element of
3830 // a vector. If Idx != 0, swizzle it into place.
3832 SmallVector<int, 4> Mask;
3833 Mask.push_back(Idx);
3834 for (unsigned i = 1; i != VecElts; ++i)
3836 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3837 DAG.getUNDEF(Item.getValueType()),
3840 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3844 // If we have a constant or non-constant insertion into the low element of
3845 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3846 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3847 // depending on what the source datatype is.
3850 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3851 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3852 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3853 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3854 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3855 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3857 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3858 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3859 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3860 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3861 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3862 Subtarget->hasSSE2(), DAG);
3863 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3867 // Is it a vector logical left shift?
3868 if (NumElems == 2 && Idx == 1 &&
3869 X86::isZeroNode(Op.getOperand(0)) &&
3870 !X86::isZeroNode(Op.getOperand(1))) {
3871 unsigned NumBits = VT.getSizeInBits();
3872 return getVShift(true, VT,
3873 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3874 VT, Op.getOperand(1)),
3875 NumBits/2, DAG, *this, dl);
3878 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3881 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3882 // is a non-constant being inserted into an element other than the low one,
3883 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3884 // movd/movss) to move this into the low element, then shuffle it into
3886 if (EVTBits == 32) {
3887 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3889 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3890 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3891 Subtarget->hasSSE2(), DAG);
3892 SmallVector<int, 8> MaskVec;
3893 for (unsigned i = 0; i < NumElems; i++)
3894 MaskVec.push_back(i == Idx ? 0 : 1);
3895 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3899 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3900 if (Values.size() == 1) {
3901 if (EVTBits == 32) {
3902 // Instead of a shuffle like this:
3903 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3904 // Check if it's possible to issue this instead.
3905 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3906 unsigned Idx = CountTrailingZeros_32(NonZeros);
3907 SDValue Item = Op.getOperand(Idx);
3908 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3909 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3914 // A vector full of immediates; various special cases are already
3915 // handled, so this is best done with a single constant-pool load.
3919 // Let legalizer expand 2-wide build_vectors.
3920 if (EVTBits == 64) {
3921 if (NumNonZero == 1) {
3922 // One half is zero or undef.
3923 unsigned Idx = CountTrailingZeros_32(NonZeros);
3924 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3925 Op.getOperand(Idx));
3926 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3927 Subtarget->hasSSE2(), DAG);
3932 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3933 if (EVTBits == 8 && NumElems == 16) {
3934 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3936 if (V.getNode()) return V;
3939 if (EVTBits == 16 && NumElems == 8) {
3940 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3942 if (V.getNode()) return V;
3945 // If element VT is == 32 bits, turn it into a number of shuffles.
3946 SmallVector<SDValue, 8> V;
3948 if (NumElems == 4 && NumZero > 0) {
3949 for (unsigned i = 0; i < 4; ++i) {
3950 bool isZero = !(NonZeros & (1 << i));
3952 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3954 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3957 for (unsigned i = 0; i < 2; ++i) {
3958 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3961 V[i] = V[i*2]; // Must be a zero vector.
3964 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3967 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3970 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3975 SmallVector<int, 8> MaskVec;
3976 bool Reverse = (NonZeros & 0x3) == 2;
3977 for (unsigned i = 0; i < 2; ++i)
3978 MaskVec.push_back(Reverse ? 1-i : i);
3979 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3980 for (unsigned i = 0; i < 2; ++i)
3981 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3982 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3985 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
3986 // Check for a build vector of consecutive loads.
3987 for (unsigned i = 0; i < NumElems; ++i)
3988 V[i] = Op.getOperand(i);
3990 // Check for elements which are consecutive loads.
3991 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
3995 // For SSE 4.1, use inserts into undef.
3996 if (getSubtarget()->hasSSE41()) {
3997 V[0] = DAG.getUNDEF(VT);
3998 for (unsigned i = 0; i < NumElems; ++i)
3999 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4000 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
4001 Op.getOperand(i), DAG.getIntPtrConstant(i));
4005 // Otherwise, expand into a number of unpckl*
4007 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4008 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4009 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4010 for (unsigned i = 0; i < NumElems; ++i)
4011 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4013 while (NumElems != 0) {
4014 for (unsigned i = 0; i < NumElems; ++i)
4015 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4024 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4025 // We support concatenate two MMX registers and place them in a MMX
4026 // register. This is better than doing a stack convert.
4027 DebugLoc dl = Op.getDebugLoc();
4028 EVT ResVT = Op.getValueType();
4029 assert(Op.getNumOperands() == 2);
4030 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4031 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4033 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4034 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4035 InVec = Op.getOperand(1);
4036 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4037 unsigned NumElts = ResVT.getVectorNumElements();
4038 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4039 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4040 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4042 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4043 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4044 Mask[0] = 0; Mask[1] = 2;
4045 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4047 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4050 // v8i16 shuffles - Prefer shuffles in the following order:
4051 // 1. [all] pshuflw, pshufhw, optional move
4052 // 2. [ssse3] 1 x pshufb
4053 // 3. [ssse3] 2 x pshufb + 1 x por
4054 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4056 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4058 const X86TargetLowering &TLI) {
4059 SDValue V1 = SVOp->getOperand(0);
4060 SDValue V2 = SVOp->getOperand(1);
4061 DebugLoc dl = SVOp->getDebugLoc();
4062 SmallVector<int, 8> MaskVals;
4064 // Determine if more than 1 of the words in each of the low and high quadwords
4065 // of the result come from the same quadword of one of the two inputs. Undef
4066 // mask values count as coming from any quadword, for better codegen.
4067 SmallVector<unsigned, 4> LoQuad(4);
4068 SmallVector<unsigned, 4> HiQuad(4);
4069 BitVector InputQuads(4);
4070 for (unsigned i = 0; i < 8; ++i) {
4071 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4072 int EltIdx = SVOp->getMaskElt(i);
4073 MaskVals.push_back(EltIdx);
4082 InputQuads.set(EltIdx / 4);
4085 int BestLoQuad = -1;
4086 unsigned MaxQuad = 1;
4087 for (unsigned i = 0; i < 4; ++i) {
4088 if (LoQuad[i] > MaxQuad) {
4090 MaxQuad = LoQuad[i];
4094 int BestHiQuad = -1;
4096 for (unsigned i = 0; i < 4; ++i) {
4097 if (HiQuad[i] > MaxQuad) {
4099 MaxQuad = HiQuad[i];
4103 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4104 // of the two input vectors, shuffle them into one input vector so only a
4105 // single pshufb instruction is necessary. If There are more than 2 input
4106 // quads, disable the next transformation since it does not help SSSE3.
4107 bool V1Used = InputQuads[0] || InputQuads[1];
4108 bool V2Used = InputQuads[2] || InputQuads[3];
4109 if (TLI.getSubtarget()->hasSSSE3()) {
4110 if (InputQuads.count() == 2 && V1Used && V2Used) {
4111 BestLoQuad = InputQuads.find_first();
4112 BestHiQuad = InputQuads.find_next(BestLoQuad);
4114 if (InputQuads.count() > 2) {
4120 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4121 // the shuffle mask. If a quad is scored as -1, that means that it contains
4122 // words from all 4 input quadwords.
4124 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4125 SmallVector<int, 8> MaskV;
4126 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4127 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4128 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4129 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4130 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4131 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4133 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4134 // source words for the shuffle, to aid later transformations.
4135 bool AllWordsInNewV = true;
4136 bool InOrder[2] = { true, true };
4137 for (unsigned i = 0; i != 8; ++i) {
4138 int idx = MaskVals[i];
4140 InOrder[i/4] = false;
4141 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4143 AllWordsInNewV = false;
4147 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4148 if (AllWordsInNewV) {
4149 for (int i = 0; i != 8; ++i) {
4150 int idx = MaskVals[i];
4153 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4154 if ((idx != i) && idx < 4)
4156 if ((idx != i) && idx > 3)
4165 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4166 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4167 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4168 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4169 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4173 // If we have SSSE3, and all words of the result are from 1 input vector,
4174 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4175 // is present, fall back to case 4.
4176 if (TLI.getSubtarget()->hasSSSE3()) {
4177 SmallVector<SDValue,16> pshufbMask;
4179 // If we have elements from both input vectors, set the high bit of the
4180 // shuffle mask element to zero out elements that come from V2 in the V1
4181 // mask, and elements that come from V1 in the V2 mask, so that the two
4182 // results can be OR'd together.
4183 bool TwoInputs = V1Used && V2Used;
4184 for (unsigned i = 0; i != 8; ++i) {
4185 int EltIdx = MaskVals[i] * 2;
4186 if (TwoInputs && (EltIdx >= 16)) {
4187 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4188 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4191 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4192 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4194 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4195 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4196 DAG.getNode(ISD::BUILD_VECTOR, dl,
4197 MVT::v16i8, &pshufbMask[0], 16));
4199 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4201 // Calculate the shuffle mask for the second input, shuffle it, and
4202 // OR it with the first shuffled input.
4204 for (unsigned i = 0; i != 8; ++i) {
4205 int EltIdx = MaskVals[i] * 2;
4207 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4208 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4211 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4212 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4214 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4215 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4216 DAG.getNode(ISD::BUILD_VECTOR, dl,
4217 MVT::v16i8, &pshufbMask[0], 16));
4218 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4219 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4222 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4223 // and update MaskVals with new element order.
4224 BitVector InOrder(8);
4225 if (BestLoQuad >= 0) {
4226 SmallVector<int, 8> MaskV;
4227 for (int i = 0; i != 4; ++i) {
4228 int idx = MaskVals[i];
4230 MaskV.push_back(-1);
4232 } else if ((idx / 4) == BestLoQuad) {
4233 MaskV.push_back(idx & 3);
4236 MaskV.push_back(-1);
4239 for (unsigned i = 4; i != 8; ++i)
4241 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4245 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4246 // and update MaskVals with the new element order.
4247 if (BestHiQuad >= 0) {
4248 SmallVector<int, 8> MaskV;
4249 for (unsigned i = 0; i != 4; ++i)
4251 for (unsigned i = 4; i != 8; ++i) {
4252 int idx = MaskVals[i];
4254 MaskV.push_back(-1);
4256 } else if ((idx / 4) == BestHiQuad) {
4257 MaskV.push_back((idx & 3) + 4);
4260 MaskV.push_back(-1);
4263 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4267 // In case BestHi & BestLo were both -1, which means each quadword has a word
4268 // from each of the four input quadwords, calculate the InOrder bitvector now
4269 // before falling through to the insert/extract cleanup.
4270 if (BestLoQuad == -1 && BestHiQuad == -1) {
4272 for (int i = 0; i != 8; ++i)
4273 if (MaskVals[i] < 0 || MaskVals[i] == i)
4277 // The other elements are put in the right place using pextrw and pinsrw.
4278 for (unsigned i = 0; i != 8; ++i) {
4281 int EltIdx = MaskVals[i];
4284 SDValue ExtOp = (EltIdx < 8)
4285 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4286 DAG.getIntPtrConstant(EltIdx))
4287 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4288 DAG.getIntPtrConstant(EltIdx - 8));
4289 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4290 DAG.getIntPtrConstant(i));
4295 // v16i8 shuffles - Prefer shuffles in the following order:
4296 // 1. [ssse3] 1 x pshufb
4297 // 2. [ssse3] 2 x pshufb + 1 x por
4298 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4300 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4302 const X86TargetLowering &TLI) {
4303 SDValue V1 = SVOp->getOperand(0);
4304 SDValue V2 = SVOp->getOperand(1);
4305 DebugLoc dl = SVOp->getDebugLoc();
4306 SmallVector<int, 16> MaskVals;
4307 SVOp->getMask(MaskVals);
4309 // If we have SSSE3, case 1 is generated when all result bytes come from
4310 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4311 // present, fall back to case 3.
4312 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4315 for (unsigned i = 0; i < 16; ++i) {
4316 int EltIdx = MaskVals[i];
4325 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4326 if (TLI.getSubtarget()->hasSSSE3()) {
4327 SmallVector<SDValue,16> pshufbMask;
4329 // If all result elements are from one input vector, then only translate
4330 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4332 // Otherwise, we have elements from both input vectors, and must zero out
4333 // elements that come from V2 in the first mask, and V1 in the second mask
4334 // so that we can OR them together.
4335 bool TwoInputs = !(V1Only || V2Only);
4336 for (unsigned i = 0; i != 16; ++i) {
4337 int EltIdx = MaskVals[i];
4338 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4339 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4342 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4344 // If all the elements are from V2, assign it to V1 and return after
4345 // building the first pshufb.
4348 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4349 DAG.getNode(ISD::BUILD_VECTOR, dl,
4350 MVT::v16i8, &pshufbMask[0], 16));
4354 // Calculate the shuffle mask for the second input, shuffle it, and
4355 // OR it with the first shuffled input.
4357 for (unsigned i = 0; i != 16; ++i) {
4358 int EltIdx = MaskVals[i];
4360 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4363 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4365 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4366 DAG.getNode(ISD::BUILD_VECTOR, dl,
4367 MVT::v16i8, &pshufbMask[0], 16));
4368 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4371 // No SSSE3 - Calculate in place words and then fix all out of place words
4372 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4373 // the 16 different words that comprise the two doublequadword input vectors.
4374 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4375 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4376 SDValue NewV = V2Only ? V2 : V1;
4377 for (int i = 0; i != 8; ++i) {
4378 int Elt0 = MaskVals[i*2];
4379 int Elt1 = MaskVals[i*2+1];
4381 // This word of the result is all undef, skip it.
4382 if (Elt0 < 0 && Elt1 < 0)
4385 // This word of the result is already in the correct place, skip it.
4386 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4388 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4391 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4392 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4395 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4396 // using a single extract together, load it and store it.
4397 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4398 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4399 DAG.getIntPtrConstant(Elt1 / 2));
4400 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4401 DAG.getIntPtrConstant(i));
4405 // If Elt1 is defined, extract it from the appropriate source. If the
4406 // source byte is not also odd, shift the extracted word left 8 bits
4407 // otherwise clear the bottom 8 bits if we need to do an or.
4409 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4410 DAG.getIntPtrConstant(Elt1 / 2));
4411 if ((Elt1 & 1) == 0)
4412 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4413 DAG.getConstant(8, TLI.getShiftAmountTy()));
4415 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4416 DAG.getConstant(0xFF00, MVT::i16));
4418 // If Elt0 is defined, extract it from the appropriate source. If the
4419 // source byte is not also even, shift the extracted word right 8 bits. If
4420 // Elt1 was also defined, OR the extracted values together before
4421 // inserting them in the result.
4423 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4424 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4425 if ((Elt0 & 1) != 0)
4426 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4427 DAG.getConstant(8, TLI.getShiftAmountTy()));
4429 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4430 DAG.getConstant(0x00FF, MVT::i16));
4431 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4434 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4435 DAG.getIntPtrConstant(i));
4437 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4440 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4441 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4442 /// done when every pair / quad of shuffle mask elements point to elements in
4443 /// the right sequence. e.g.
4444 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4446 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4448 const TargetLowering &TLI, DebugLoc dl) {
4449 EVT VT = SVOp->getValueType(0);
4450 SDValue V1 = SVOp->getOperand(0);
4451 SDValue V2 = SVOp->getOperand(1);
4452 unsigned NumElems = VT.getVectorNumElements();
4453 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4454 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4456 switch (VT.getSimpleVT().SimpleTy) {
4457 default: assert(false && "Unexpected!");
4458 case MVT::v4f32: NewVT = MVT::v2f64; break;
4459 case MVT::v4i32: NewVT = MVT::v2i64; break;
4460 case MVT::v8i16: NewVT = MVT::v4i32; break;
4461 case MVT::v16i8: NewVT = MVT::v4i32; break;
4464 if (NewWidth == 2) {
4470 int Scale = NumElems / NewWidth;
4471 SmallVector<int, 8> MaskVec;
4472 for (unsigned i = 0; i < NumElems; i += Scale) {
4474 for (int j = 0; j < Scale; ++j) {
4475 int EltIdx = SVOp->getMaskElt(i+j);
4479 StartIdx = EltIdx - (EltIdx % Scale);
4480 if (EltIdx != StartIdx + j)
4484 MaskVec.push_back(-1);
4486 MaskVec.push_back(StartIdx / Scale);
4489 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4490 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4491 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4494 /// getVZextMovL - Return a zero-extending vector move low node.
4496 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4497 SDValue SrcOp, SelectionDAG &DAG,
4498 const X86Subtarget *Subtarget, DebugLoc dl) {
4499 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4500 LoadSDNode *LD = NULL;
4501 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4502 LD = dyn_cast<LoadSDNode>(SrcOp);
4504 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4506 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4507 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4508 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4509 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4510 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4512 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4513 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4514 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4515 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4523 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4524 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4525 DAG.getNode(ISD::BIT_CONVERT, dl,
4529 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4532 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4533 SDValue V1 = SVOp->getOperand(0);
4534 SDValue V2 = SVOp->getOperand(1);
4535 DebugLoc dl = SVOp->getDebugLoc();
4536 EVT VT = SVOp->getValueType(0);
4538 SmallVector<std::pair<int, int>, 8> Locs;
4540 SmallVector<int, 8> Mask1(4U, -1);
4541 SmallVector<int, 8> PermMask;
4542 SVOp->getMask(PermMask);
4546 for (unsigned i = 0; i != 4; ++i) {
4547 int Idx = PermMask[i];
4549 Locs[i] = std::make_pair(-1, -1);
4551 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4553 Locs[i] = std::make_pair(0, NumLo);
4557 Locs[i] = std::make_pair(1, NumHi);
4559 Mask1[2+NumHi] = Idx;
4565 if (NumLo <= 2 && NumHi <= 2) {
4566 // If no more than two elements come from either vector. This can be
4567 // implemented with two shuffles. First shuffle gather the elements.
4568 // The second shuffle, which takes the first shuffle as both of its
4569 // vector operands, put the elements into the right order.
4570 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4572 SmallVector<int, 8> Mask2(4U, -1);
4574 for (unsigned i = 0; i != 4; ++i) {
4575 if (Locs[i].first == -1)
4578 unsigned Idx = (i < 2) ? 0 : 4;
4579 Idx += Locs[i].first * 2 + Locs[i].second;
4584 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4585 } else if (NumLo == 3 || NumHi == 3) {
4586 // Otherwise, we must have three elements from one vector, call it X, and
4587 // one element from the other, call it Y. First, use a shufps to build an
4588 // intermediate vector with the one element from Y and the element from X
4589 // that will be in the same half in the final destination (the indexes don't
4590 // matter). Then, use a shufps to build the final vector, taking the half
4591 // containing the element from Y from the intermediate, and the other half
4594 // Normalize it so the 3 elements come from V1.
4595 CommuteVectorShuffleMask(PermMask, VT);
4599 // Find the element from V2.
4601 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4602 int Val = PermMask[HiIndex];
4609 Mask1[0] = PermMask[HiIndex];
4611 Mask1[2] = PermMask[HiIndex^1];
4613 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4616 Mask1[0] = PermMask[0];
4617 Mask1[1] = PermMask[1];
4618 Mask1[2] = HiIndex & 1 ? 6 : 4;
4619 Mask1[3] = HiIndex & 1 ? 4 : 6;
4620 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4622 Mask1[0] = HiIndex & 1 ? 2 : 0;
4623 Mask1[1] = HiIndex & 1 ? 0 : 2;
4624 Mask1[2] = PermMask[2];
4625 Mask1[3] = PermMask[3];
4630 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4634 // Break it into (shuffle shuffle_hi, shuffle_lo).
4636 SmallVector<int,8> LoMask(4U, -1);
4637 SmallVector<int,8> HiMask(4U, -1);
4639 SmallVector<int,8> *MaskPtr = &LoMask;
4640 unsigned MaskIdx = 0;
4643 for (unsigned i = 0; i != 4; ++i) {
4650 int Idx = PermMask[i];
4652 Locs[i] = std::make_pair(-1, -1);
4653 } else if (Idx < 4) {
4654 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4655 (*MaskPtr)[LoIdx] = Idx;
4658 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4659 (*MaskPtr)[HiIdx] = Idx;
4664 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4665 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4666 SmallVector<int, 8> MaskOps;
4667 for (unsigned i = 0; i != 4; ++i) {
4668 if (Locs[i].first == -1) {
4669 MaskOps.push_back(-1);
4671 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4672 MaskOps.push_back(Idx);
4675 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4679 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4680 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4681 SDValue V1 = Op.getOperand(0);
4682 SDValue V2 = Op.getOperand(1);
4683 EVT VT = Op.getValueType();
4684 DebugLoc dl = Op.getDebugLoc();
4685 unsigned NumElems = VT.getVectorNumElements();
4686 bool isMMX = VT.getSizeInBits() == 64;
4687 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4688 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4689 bool V1IsSplat = false;
4690 bool V2IsSplat = false;
4692 if (isZeroShuffle(SVOp))
4693 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4695 // Promote splats to v4f32.
4696 if (SVOp->isSplat()) {
4697 if (isMMX || NumElems < 4)
4699 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4702 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4704 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4705 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4706 if (NewOp.getNode())
4707 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4708 LowerVECTOR_SHUFFLE(NewOp, DAG));
4709 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4710 // FIXME: Figure out a cleaner way to do this.
4711 // Try to make use of movq to zero out the top part.
4712 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4713 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4714 if (NewOp.getNode()) {
4715 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4716 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4717 DAG, Subtarget, dl);
4719 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4720 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4721 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4722 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4723 DAG, Subtarget, dl);
4727 if (X86::isPSHUFDMask(SVOp))
4730 // Check if this can be converted into a logical shift.
4731 bool isLeft = false;
4734 bool isShift = getSubtarget()->hasSSE2() &&
4735 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4736 if (isShift && ShVal.hasOneUse()) {
4737 // If the shifted value has multiple uses, it may be cheaper to use
4738 // v_set0 + movlhps or movhlps, etc.
4739 EVT EltVT = VT.getVectorElementType();
4740 ShAmt *= EltVT.getSizeInBits();
4741 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4744 if (X86::isMOVLMask(SVOp)) {
4747 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4748 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4753 // FIXME: fold these into legal mask.
4754 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4755 X86::isMOVSLDUPMask(SVOp) ||
4756 X86::isMOVHLPSMask(SVOp) ||
4757 X86::isMOVLHPSMask(SVOp) ||
4758 X86::isMOVLPMask(SVOp)))
4761 if (ShouldXformToMOVHLPS(SVOp) ||
4762 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4763 return CommuteVectorShuffle(SVOp, DAG);
4766 // No better options. Use a vshl / vsrl.
4767 EVT EltVT = VT.getVectorElementType();
4768 ShAmt *= EltVT.getSizeInBits();
4769 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4772 bool Commuted = false;
4773 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4774 // 1,1,1,1 -> v8i16 though.
4775 V1IsSplat = isSplatVector(V1.getNode());
4776 V2IsSplat = isSplatVector(V2.getNode());
4778 // Canonicalize the splat or undef, if present, to be on the RHS.
4779 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4780 Op = CommuteVectorShuffle(SVOp, DAG);
4781 SVOp = cast<ShuffleVectorSDNode>(Op);
4782 V1 = SVOp->getOperand(0);
4783 V2 = SVOp->getOperand(1);
4784 std::swap(V1IsSplat, V2IsSplat);
4785 std::swap(V1IsUndef, V2IsUndef);
4789 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4790 // Shuffling low element of v1 into undef, just return v1.
4793 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4794 // the instruction selector will not match, so get a canonical MOVL with
4795 // swapped operands to undo the commute.
4796 return getMOVL(DAG, dl, VT, V2, V1);
4799 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4800 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4801 X86::isUNPCKLMask(SVOp) ||
4802 X86::isUNPCKHMask(SVOp))
4806 // Normalize mask so all entries that point to V2 points to its first
4807 // element then try to match unpck{h|l} again. If match, return a
4808 // new vector_shuffle with the corrected mask.
4809 SDValue NewMask = NormalizeMask(SVOp, DAG);
4810 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4811 if (NSVOp != SVOp) {
4812 if (X86::isUNPCKLMask(NSVOp, true)) {
4814 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4821 // Commute is back and try unpck* again.
4822 // FIXME: this seems wrong.
4823 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4824 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4825 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4826 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4827 X86::isUNPCKLMask(NewSVOp) ||
4828 X86::isUNPCKHMask(NewSVOp))
4832 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4834 // Normalize the node to match x86 shuffle ops if needed
4835 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4836 return CommuteVectorShuffle(SVOp, DAG);
4838 // Check for legal shuffle and return?
4839 SmallVector<int, 16> PermMask;
4840 SVOp->getMask(PermMask);
4841 if (isShuffleMaskLegal(PermMask, VT))
4844 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4845 if (VT == MVT::v8i16) {
4846 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4847 if (NewOp.getNode())
4851 if (VT == MVT::v16i8) {
4852 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4853 if (NewOp.getNode())
4857 // Handle all 4 wide cases with a number of shuffles except for MMX.
4858 if (NumElems == 4 && !isMMX)
4859 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4865 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4866 SelectionDAG &DAG) const {
4867 EVT VT = Op.getValueType();
4868 DebugLoc dl = Op.getDebugLoc();
4869 if (VT.getSizeInBits() == 8) {
4870 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4871 Op.getOperand(0), Op.getOperand(1));
4872 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4873 DAG.getValueType(VT));
4874 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4875 } else if (VT.getSizeInBits() == 16) {
4876 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4877 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4879 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4880 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4881 DAG.getNode(ISD::BIT_CONVERT, dl,
4885 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, 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 == MVT::f32) {
4891 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4892 // the result back to FR32 register. It's only worth matching if the
4893 // result has a single use which is a store or a bitcast to i32. And in
4894 // the case of a store, it's not worth it if the index is a constant 0,
4895 // because a MOVSSmr can be used instead, which is smaller and faster.
4896 if (!Op.hasOneUse())
4898 SDNode *User = *Op.getNode()->use_begin();
4899 if ((User->getOpcode() != ISD::STORE ||
4900 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4901 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4902 (User->getOpcode() != ISD::BIT_CONVERT ||
4903 User->getValueType(0) != MVT::i32))
4905 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4906 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4909 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4910 } else if (VT == MVT::i32) {
4911 // ExtractPS works with constant index.
4912 if (isa<ConstantSDNode>(Op.getOperand(1)))
4920 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4921 SelectionDAG &DAG) const {
4922 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4925 if (Subtarget->hasSSE41()) {
4926 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4931 EVT VT = Op.getValueType();
4932 DebugLoc dl = Op.getDebugLoc();
4933 // TODO: handle v16i8.
4934 if (VT.getSizeInBits() == 16) {
4935 SDValue Vec = Op.getOperand(0);
4936 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4938 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4939 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4940 DAG.getNode(ISD::BIT_CONVERT, dl,
4943 // Transform it so it match pextrw which produces a 32-bit result.
4944 EVT EltVT = MVT::i32;
4945 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4946 Op.getOperand(0), Op.getOperand(1));
4947 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4948 DAG.getValueType(VT));
4949 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4950 } else if (VT.getSizeInBits() == 32) {
4951 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4955 // SHUFPS the element to the lowest double word, then movss.
4956 int Mask[4] = { Idx, -1, -1, -1 };
4957 EVT VVT = Op.getOperand(0).getValueType();
4958 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4959 DAG.getUNDEF(VVT), Mask);
4960 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4961 DAG.getIntPtrConstant(0));
4962 } else if (VT.getSizeInBits() == 64) {
4963 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4964 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4965 // to match extract_elt for f64.
4966 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4970 // UNPCKHPD the element to the lowest double word, then movsd.
4971 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4972 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4973 int Mask[2] = { 1, -1 };
4974 EVT VVT = Op.getOperand(0).getValueType();
4975 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4976 DAG.getUNDEF(VVT), Mask);
4977 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4978 DAG.getIntPtrConstant(0));
4985 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
4986 SelectionDAG &DAG) const {
4987 EVT VT = Op.getValueType();
4988 EVT EltVT = VT.getVectorElementType();
4989 DebugLoc dl = Op.getDebugLoc();
4991 SDValue N0 = Op.getOperand(0);
4992 SDValue N1 = Op.getOperand(1);
4993 SDValue N2 = Op.getOperand(2);
4995 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4996 isa<ConstantSDNode>(N2)) {
4998 if (VT == MVT::v8i16)
4999 Opc = X86ISD::PINSRW;
5000 else if (VT == MVT::v4i16)
5001 Opc = X86ISD::MMX_PINSRW;
5002 else if (VT == MVT::v16i8)
5003 Opc = X86ISD::PINSRB;
5005 Opc = X86ISD::PINSRB;
5007 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5009 if (N1.getValueType() != MVT::i32)
5010 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5011 if (N2.getValueType() != MVT::i32)
5012 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5013 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5014 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5015 // Bits [7:6] of the constant are the source select. This will always be
5016 // zero here. The DAG Combiner may combine an extract_elt index into these
5017 // bits. For example (insert (extract, 3), 2) could be matched by putting
5018 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5019 // Bits [5:4] of the constant are the destination select. This is the
5020 // value of the incoming immediate.
5021 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5022 // combine either bitwise AND or insert of float 0.0 to set these bits.
5023 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5024 // Create this as a scalar to vector..
5025 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5026 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5027 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5028 // PINSR* works with constant index.
5035 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5036 EVT VT = Op.getValueType();
5037 EVT EltVT = VT.getVectorElementType();
5039 if (Subtarget->hasSSE41())
5040 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5042 if (EltVT == MVT::i8)
5045 DebugLoc dl = Op.getDebugLoc();
5046 SDValue N0 = Op.getOperand(0);
5047 SDValue N1 = Op.getOperand(1);
5048 SDValue N2 = Op.getOperand(2);
5050 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5051 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5052 // as its second argument.
5053 if (N1.getValueType() != MVT::i32)
5054 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5055 if (N2.getValueType() != MVT::i32)
5056 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5057 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5058 dl, VT, N0, N1, N2);
5064 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5065 DebugLoc dl = Op.getDebugLoc();
5067 if (Op.getValueType() == MVT::v1i64 &&
5068 Op.getOperand(0).getValueType() == MVT::i64)
5069 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5071 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5072 EVT VT = MVT::v2i32;
5073 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5080 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5081 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5084 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5085 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5086 // one of the above mentioned nodes. It has to be wrapped because otherwise
5087 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5088 // be used to form addressing mode. These wrapped nodes will be selected
5091 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5092 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5094 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5096 unsigned char OpFlag = 0;
5097 unsigned WrapperKind = X86ISD::Wrapper;
5098 CodeModel::Model M = getTargetMachine().getCodeModel();
5100 if (Subtarget->isPICStyleRIPRel() &&
5101 (M == CodeModel::Small || M == CodeModel::Kernel))
5102 WrapperKind = X86ISD::WrapperRIP;
5103 else if (Subtarget->isPICStyleGOT())
5104 OpFlag = X86II::MO_GOTOFF;
5105 else if (Subtarget->isPICStyleStubPIC())
5106 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5108 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5110 CP->getOffset(), OpFlag);
5111 DebugLoc DL = CP->getDebugLoc();
5112 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5113 // With PIC, the address is actually $g + Offset.
5115 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5116 DAG.getNode(X86ISD::GlobalBaseReg,
5117 DebugLoc(), getPointerTy()),
5124 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5125 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5127 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5129 unsigned char OpFlag = 0;
5130 unsigned WrapperKind = X86ISD::Wrapper;
5131 CodeModel::Model M = getTargetMachine().getCodeModel();
5133 if (Subtarget->isPICStyleRIPRel() &&
5134 (M == CodeModel::Small || M == CodeModel::Kernel))
5135 WrapperKind = X86ISD::WrapperRIP;
5136 else if (Subtarget->isPICStyleGOT())
5137 OpFlag = X86II::MO_GOTOFF;
5138 else if (Subtarget->isPICStyleStubPIC())
5139 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5141 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5143 DebugLoc DL = JT->getDebugLoc();
5144 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5146 // With PIC, the address is actually $g + Offset.
5148 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5149 DAG.getNode(X86ISD::GlobalBaseReg,
5150 DebugLoc(), getPointerTy()),
5158 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5159 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5161 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5163 unsigned char OpFlag = 0;
5164 unsigned WrapperKind = X86ISD::Wrapper;
5165 CodeModel::Model M = getTargetMachine().getCodeModel();
5167 if (Subtarget->isPICStyleRIPRel() &&
5168 (M == CodeModel::Small || M == CodeModel::Kernel))
5169 WrapperKind = X86ISD::WrapperRIP;
5170 else if (Subtarget->isPICStyleGOT())
5171 OpFlag = X86II::MO_GOTOFF;
5172 else if (Subtarget->isPICStyleStubPIC())
5173 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5175 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5177 DebugLoc DL = Op.getDebugLoc();
5178 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5181 // With PIC, the address is actually $g + Offset.
5182 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5183 !Subtarget->is64Bit()) {
5184 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5185 DAG.getNode(X86ISD::GlobalBaseReg,
5186 DebugLoc(), getPointerTy()),
5194 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5195 // Create the TargetBlockAddressAddress node.
5196 unsigned char OpFlags =
5197 Subtarget->ClassifyBlockAddressReference();
5198 CodeModel::Model M = getTargetMachine().getCodeModel();
5199 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5200 DebugLoc dl = Op.getDebugLoc();
5201 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5202 /*isTarget=*/true, OpFlags);
5204 if (Subtarget->isPICStyleRIPRel() &&
5205 (M == CodeModel::Small || M == CodeModel::Kernel))
5206 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5208 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5210 // With PIC, the address is actually $g + Offset.
5211 if (isGlobalRelativeToPICBase(OpFlags)) {
5212 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5213 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5221 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5223 SelectionDAG &DAG) const {
5224 // Create the TargetGlobalAddress node, folding in the constant
5225 // offset if it is legal.
5226 unsigned char OpFlags =
5227 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5228 CodeModel::Model M = getTargetMachine().getCodeModel();
5230 if (OpFlags == X86II::MO_NO_FLAG &&
5231 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5232 // A direct static reference to a global.
5233 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5236 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5239 if (Subtarget->isPICStyleRIPRel() &&
5240 (M == CodeModel::Small || M == CodeModel::Kernel))
5241 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5243 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5245 // With PIC, the address is actually $g + Offset.
5246 if (isGlobalRelativeToPICBase(OpFlags)) {
5247 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5248 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5252 // For globals that require a load from a stub to get the address, emit the
5254 if (isGlobalStubReference(OpFlags))
5255 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5256 PseudoSourceValue::getGOT(), 0, false, false, 0);
5258 // If there was a non-zero offset that we didn't fold, create an explicit
5261 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5262 DAG.getConstant(Offset, getPointerTy()));
5268 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5269 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5270 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5271 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5275 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5276 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5277 unsigned char OperandFlags) {
5278 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5279 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5280 DebugLoc dl = GA->getDebugLoc();
5281 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5282 GA->getValueType(0),
5286 SDValue Ops[] = { Chain, TGA, *InFlag };
5287 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5289 SDValue Ops[] = { Chain, TGA };
5290 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5293 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5294 MFI->setAdjustsStack(true);
5296 SDValue Flag = Chain.getValue(1);
5297 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5300 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5302 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5305 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5306 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5307 DAG.getNode(X86ISD::GlobalBaseReg,
5308 DebugLoc(), PtrVT), InFlag);
5309 InFlag = Chain.getValue(1);
5311 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5314 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5316 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5318 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5319 X86::RAX, X86II::MO_TLSGD);
5322 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5323 // "local exec" model.
5324 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5325 const EVT PtrVT, TLSModel::Model model,
5327 DebugLoc dl = GA->getDebugLoc();
5328 // Get the Thread Pointer
5329 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5331 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5334 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5335 NULL, 0, false, false, 0);
5337 unsigned char OperandFlags = 0;
5338 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5340 unsigned WrapperKind = X86ISD::Wrapper;
5341 if (model == TLSModel::LocalExec) {
5342 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5343 } else if (is64Bit) {
5344 assert(model == TLSModel::InitialExec);
5345 OperandFlags = X86II::MO_GOTTPOFF;
5346 WrapperKind = X86ISD::WrapperRIP;
5348 assert(model == TLSModel::InitialExec);
5349 OperandFlags = X86II::MO_INDNTPOFF;
5352 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5354 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5355 GA->getOffset(), OperandFlags);
5356 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5358 if (model == TLSModel::InitialExec)
5359 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5360 PseudoSourceValue::getGOT(), 0, false, false, 0);
5362 // The address of the thread local variable is the add of the thread
5363 // pointer with the offset of the variable.
5364 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5368 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5370 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5371 const GlobalValue *GV = GA->getGlobal();
5373 if (Subtarget->isTargetELF()) {
5374 // TODO: implement the "local dynamic" model
5375 // TODO: implement the "initial exec"model for pic executables
5377 // If GV is an alias then use the aliasee for determining
5378 // thread-localness.
5379 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5380 GV = GA->resolveAliasedGlobal(false);
5382 TLSModel::Model model
5383 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5386 case TLSModel::GeneralDynamic:
5387 case TLSModel::LocalDynamic: // not implemented
5388 if (Subtarget->is64Bit())
5389 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5390 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5392 case TLSModel::InitialExec:
5393 case TLSModel::LocalExec:
5394 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5395 Subtarget->is64Bit());
5397 } else if (Subtarget->isTargetDarwin()) {
5398 // Darwin only has one model of TLS. Lower to that.
5399 unsigned char OpFlag = 0;
5400 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5401 X86ISD::WrapperRIP : X86ISD::Wrapper;
5403 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5405 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5406 !Subtarget->is64Bit();
5408 OpFlag = X86II::MO_TLVP_PIC_BASE;
5410 OpFlag = X86II::MO_TLVP;
5412 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(),
5414 GA->getOffset(), OpFlag);
5416 DebugLoc DL = Op.getDebugLoc();
5417 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5419 // With PIC32, the address is actually $g + Offset.
5421 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5422 DAG.getNode(X86ISD::GlobalBaseReg,
5423 DebugLoc(), getPointerTy()),
5426 // Lowering the machine isd will make sure everything is in the right
5428 SDValue Args[] = { Offset };
5429 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5431 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5432 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5433 MFI->setAdjustsStack(true);
5435 // And our return value (tls address) is in the standard call return value
5437 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5438 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5442 "TLS not implemented for this target.");
5444 llvm_unreachable("Unreachable");
5449 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5450 /// take a 2 x i32 value to shift plus a shift amount.
5451 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5452 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5453 EVT VT = Op.getValueType();
5454 unsigned VTBits = VT.getSizeInBits();
5455 DebugLoc dl = Op.getDebugLoc();
5456 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5457 SDValue ShOpLo = Op.getOperand(0);
5458 SDValue ShOpHi = Op.getOperand(1);
5459 SDValue ShAmt = Op.getOperand(2);
5460 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5461 DAG.getConstant(VTBits - 1, MVT::i8))
5462 : DAG.getConstant(0, VT);
5465 if (Op.getOpcode() == ISD::SHL_PARTS) {
5466 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5467 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5469 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5470 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5473 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5474 DAG.getConstant(VTBits, MVT::i8));
5475 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5476 AndNode, DAG.getConstant(0, MVT::i8));
5479 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5480 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5481 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5483 if (Op.getOpcode() == ISD::SHL_PARTS) {
5484 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5485 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5487 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5488 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5491 SDValue Ops[2] = { Lo, Hi };
5492 return DAG.getMergeValues(Ops, 2, dl);
5495 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5496 SelectionDAG &DAG) const {
5497 EVT SrcVT = Op.getOperand(0).getValueType();
5499 if (SrcVT.isVector()) {
5500 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5506 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5507 "Unknown SINT_TO_FP to lower!");
5509 // These are really Legal; return the operand so the caller accepts it as
5511 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5513 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5514 Subtarget->is64Bit()) {
5518 DebugLoc dl = Op.getDebugLoc();
5519 unsigned Size = SrcVT.getSizeInBits()/8;
5520 MachineFunction &MF = DAG.getMachineFunction();
5521 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5522 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5523 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5525 PseudoSourceValue::getFixedStack(SSFI), 0,
5527 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5530 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5532 SelectionDAG &DAG) const {
5534 DebugLoc dl = Op.getDebugLoc();
5536 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5538 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5540 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5541 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5542 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5543 Tys, Ops, array_lengthof(Ops));
5546 Chain = Result.getValue(1);
5547 SDValue InFlag = Result.getValue(2);
5549 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5550 // shouldn't be necessary except that RFP cannot be live across
5551 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5552 MachineFunction &MF = DAG.getMachineFunction();
5553 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5554 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5555 Tys = DAG.getVTList(MVT::Other);
5557 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5559 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5560 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5561 PseudoSourceValue::getFixedStack(SSFI), 0,
5568 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5569 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5570 SelectionDAG &DAG) const {
5571 // This algorithm is not obvious. Here it is in C code, more or less:
5573 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5574 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5575 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5577 // Copy ints to xmm registers.
5578 __m128i xh = _mm_cvtsi32_si128( hi );
5579 __m128i xl = _mm_cvtsi32_si128( lo );
5581 // Combine into low half of a single xmm register.
5582 __m128i x = _mm_unpacklo_epi32( xh, xl );
5586 // Merge in appropriate exponents to give the integer bits the right
5588 x = _mm_unpacklo_epi32( x, exp );
5590 // Subtract away the biases to deal with the IEEE-754 double precision
5592 d = _mm_sub_pd( (__m128d) x, bias );
5594 // All conversions up to here are exact. The correctly rounded result is
5595 // calculated using the current rounding mode using the following
5597 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5598 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5599 // store doesn't really need to be here (except
5600 // maybe to zero the other double)
5605 DebugLoc dl = Op.getDebugLoc();
5606 LLVMContext *Context = DAG.getContext();
5608 // Build some magic constants.
5609 std::vector<Constant*> CV0;
5610 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5611 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5612 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5613 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5614 Constant *C0 = ConstantVector::get(CV0);
5615 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5617 std::vector<Constant*> CV1;
5619 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5621 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5622 Constant *C1 = ConstantVector::get(CV1);
5623 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5625 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5626 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5628 DAG.getIntPtrConstant(1)));
5629 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5630 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5632 DAG.getIntPtrConstant(0)));
5633 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5634 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5635 PseudoSourceValue::getConstantPool(), 0,
5637 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5638 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5639 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5640 PseudoSourceValue::getConstantPool(), 0,
5642 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5644 // Add the halves; easiest way is to swap them into another reg first.
5645 int ShufMask[2] = { 1, -1 };
5646 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5647 DAG.getUNDEF(MVT::v2f64), ShufMask);
5648 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5649 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5650 DAG.getIntPtrConstant(0));
5653 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5654 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5655 SelectionDAG &DAG) const {
5656 DebugLoc dl = Op.getDebugLoc();
5657 // FP constant to bias correct the final result.
5658 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5661 // Load the 32-bit value into an XMM register.
5662 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5663 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5665 DAG.getIntPtrConstant(0)));
5667 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5668 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5669 DAG.getIntPtrConstant(0));
5671 // Or the load with the bias.
5672 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5673 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5674 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5676 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5677 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5678 MVT::v2f64, Bias)));
5679 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5680 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5681 DAG.getIntPtrConstant(0));
5683 // Subtract the bias.
5684 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5686 // Handle final rounding.
5687 EVT DestVT = Op.getValueType();
5689 if (DestVT.bitsLT(MVT::f64)) {
5690 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5691 DAG.getIntPtrConstant(0));
5692 } else if (DestVT.bitsGT(MVT::f64)) {
5693 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5696 // Handle final rounding.
5700 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5701 SelectionDAG &DAG) const {
5702 SDValue N0 = Op.getOperand(0);
5703 DebugLoc dl = Op.getDebugLoc();
5705 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5706 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5707 // the optimization here.
5708 if (DAG.SignBitIsZero(N0))
5709 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5711 EVT SrcVT = N0.getValueType();
5712 EVT DstVT = Op.getValueType();
5713 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5714 return LowerUINT_TO_FP_i64(Op, DAG);
5715 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5716 return LowerUINT_TO_FP_i32(Op, DAG);
5718 // Make a 64-bit buffer, and use it to build an FILD.
5719 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5720 if (SrcVT == MVT::i32) {
5721 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5722 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5723 getPointerTy(), StackSlot, WordOff);
5724 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5725 StackSlot, NULL, 0, false, false, 0);
5726 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5727 OffsetSlot, NULL, 0, false, false, 0);
5728 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5732 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5733 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5734 StackSlot, NULL, 0, false, false, 0);
5735 // For i64 source, we need to add the appropriate power of 2 if the input
5736 // was negative. This is the same as the optimization in
5737 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5738 // we must be careful to do the computation in x87 extended precision, not
5739 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5740 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5741 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5742 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5744 APInt FF(32, 0x5F800000ULL);
5746 // Check whether the sign bit is set.
5747 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5748 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5751 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5752 SDValue FudgePtr = DAG.getConstantPool(
5753 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5756 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5757 SDValue Zero = DAG.getIntPtrConstant(0);
5758 SDValue Four = DAG.getIntPtrConstant(4);
5759 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5761 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5763 // Load the value out, extending it from f32 to f80.
5764 // FIXME: Avoid the extend by constructing the right constant pool?
5765 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
5766 FudgePtr, PseudoSourceValue::getConstantPool(),
5767 0, MVT::f32, false, false, 4);
5768 // Extend everything to 80 bits to force it to be done on x87.
5769 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5770 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5773 std::pair<SDValue,SDValue> X86TargetLowering::
5774 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5775 DebugLoc dl = Op.getDebugLoc();
5777 EVT DstTy = Op.getValueType();
5780 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5784 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5785 DstTy.getSimpleVT() >= MVT::i16 &&
5786 "Unknown FP_TO_SINT to lower!");
5788 // These are really Legal.
5789 if (DstTy == MVT::i32 &&
5790 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5791 return std::make_pair(SDValue(), SDValue());
5792 if (Subtarget->is64Bit() &&
5793 DstTy == MVT::i64 &&
5794 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5795 return std::make_pair(SDValue(), SDValue());
5797 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5799 MachineFunction &MF = DAG.getMachineFunction();
5800 unsigned MemSize = DstTy.getSizeInBits()/8;
5801 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5802 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5805 switch (DstTy.getSimpleVT().SimpleTy) {
5806 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5807 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5808 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5809 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5812 SDValue Chain = DAG.getEntryNode();
5813 SDValue Value = Op.getOperand(0);
5814 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5815 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5816 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5817 PseudoSourceValue::getFixedStack(SSFI), 0,
5819 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5821 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5823 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5824 Chain = Value.getValue(1);
5825 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5826 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5829 // Build the FP_TO_INT*_IN_MEM
5830 SDValue Ops[] = { Chain, Value, StackSlot };
5831 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5833 return std::make_pair(FIST, StackSlot);
5836 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5837 SelectionDAG &DAG) const {
5838 if (Op.getValueType().isVector()) {
5839 if (Op.getValueType() == MVT::v2i32 &&
5840 Op.getOperand(0).getValueType() == MVT::v2f64) {
5846 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5847 SDValue FIST = Vals.first, StackSlot = Vals.second;
5848 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5849 if (FIST.getNode() == 0) return Op;
5852 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5853 FIST, StackSlot, NULL, 0, false, false, 0);
5856 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5857 SelectionDAG &DAG) const {
5858 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5859 SDValue FIST = Vals.first, StackSlot = Vals.second;
5860 assert(FIST.getNode() && "Unexpected failure");
5863 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5864 FIST, StackSlot, NULL, 0, false, false, 0);
5867 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5868 SelectionDAG &DAG) const {
5869 LLVMContext *Context = DAG.getContext();
5870 DebugLoc dl = Op.getDebugLoc();
5871 EVT VT = Op.getValueType();
5874 EltVT = VT.getVectorElementType();
5875 std::vector<Constant*> CV;
5876 if (EltVT == MVT::f64) {
5877 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5881 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5887 Constant *C = ConstantVector::get(CV);
5888 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5889 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5890 PseudoSourceValue::getConstantPool(), 0,
5892 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5895 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5896 LLVMContext *Context = DAG.getContext();
5897 DebugLoc dl = Op.getDebugLoc();
5898 EVT VT = Op.getValueType();
5901 EltVT = VT.getVectorElementType();
5902 std::vector<Constant*> CV;
5903 if (EltVT == MVT::f64) {
5904 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5908 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5914 Constant *C = ConstantVector::get(CV);
5915 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5916 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5917 PseudoSourceValue::getConstantPool(), 0,
5919 if (VT.isVector()) {
5920 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5921 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5922 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5924 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5926 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5930 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5931 LLVMContext *Context = DAG.getContext();
5932 SDValue Op0 = Op.getOperand(0);
5933 SDValue Op1 = Op.getOperand(1);
5934 DebugLoc dl = Op.getDebugLoc();
5935 EVT VT = Op.getValueType();
5936 EVT SrcVT = Op1.getValueType();
5938 // If second operand is smaller, extend it first.
5939 if (SrcVT.bitsLT(VT)) {
5940 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5943 // And if it is bigger, shrink it first.
5944 if (SrcVT.bitsGT(VT)) {
5945 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5949 // At this point the operands and the result should have the same
5950 // type, and that won't be f80 since that is not custom lowered.
5952 // First get the sign bit of second operand.
5953 std::vector<Constant*> CV;
5954 if (SrcVT == MVT::f64) {
5955 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5956 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5958 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5959 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5960 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5961 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5963 Constant *C = ConstantVector::get(CV);
5964 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5965 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5966 PseudoSourceValue::getConstantPool(), 0,
5968 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5970 // Shift sign bit right or left if the two operands have different types.
5971 if (SrcVT.bitsGT(VT)) {
5972 // Op0 is MVT::f32, Op1 is MVT::f64.
5973 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5974 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5975 DAG.getConstant(32, MVT::i32));
5976 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5977 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5978 DAG.getIntPtrConstant(0));
5981 // Clear first operand sign bit.
5983 if (VT == MVT::f64) {
5984 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5985 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5987 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5988 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5989 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5990 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5992 C = ConstantVector::get(CV);
5993 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5994 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5995 PseudoSourceValue::getConstantPool(), 0,
5997 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5999 // Or the value with the sign bit.
6000 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6003 /// Emit nodes that will be selected as "test Op0,Op0", or something
6005 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6006 SelectionDAG &DAG) const {
6007 DebugLoc dl = Op.getDebugLoc();
6009 // CF and OF aren't always set the way we want. Determine which
6010 // of these we need.
6011 bool NeedCF = false;
6012 bool NeedOF = false;
6015 case X86::COND_A: case X86::COND_AE:
6016 case X86::COND_B: case X86::COND_BE:
6019 case X86::COND_G: case X86::COND_GE:
6020 case X86::COND_L: case X86::COND_LE:
6021 case X86::COND_O: case X86::COND_NO:
6026 // See if we can use the EFLAGS value from the operand instead of
6027 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6028 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6029 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6030 // Emit a CMP with 0, which is the TEST pattern.
6031 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6032 DAG.getConstant(0, Op.getValueType()));
6034 unsigned Opcode = 0;
6035 unsigned NumOperands = 0;
6036 switch (Op.getNode()->getOpcode()) {
6038 // Due to an isel shortcoming, be conservative if this add is likely to be
6039 // selected as part of a load-modify-store instruction. When the root node
6040 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6041 // uses of other nodes in the match, such as the ADD in this case. This
6042 // leads to the ADD being left around and reselected, with the result being
6043 // two adds in the output. Alas, even if none our users are stores, that
6044 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6045 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6046 // climbing the DAG back to the root, and it doesn't seem to be worth the
6048 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6049 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6050 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6053 if (ConstantSDNode *C =
6054 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6055 // An add of one will be selected as an INC.
6056 if (C->getAPIntValue() == 1) {
6057 Opcode = X86ISD::INC;
6062 // An add of negative one (subtract of one) will be selected as a DEC.
6063 if (C->getAPIntValue().isAllOnesValue()) {
6064 Opcode = X86ISD::DEC;
6070 // Otherwise use a regular EFLAGS-setting add.
6071 Opcode = X86ISD::ADD;
6075 // If the primary and result isn't used, don't bother using X86ISD::AND,
6076 // because a TEST instruction will be better.
6077 bool NonFlagUse = false;
6078 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6079 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6081 unsigned UOpNo = UI.getOperandNo();
6082 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6083 // Look pass truncate.
6084 UOpNo = User->use_begin().getOperandNo();
6085 User = *User->use_begin();
6088 if (User->getOpcode() != ISD::BRCOND &&
6089 User->getOpcode() != ISD::SETCC &&
6090 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6103 // Due to the ISEL shortcoming noted above, be conservative if this op is
6104 // likely to be selected as part of a load-modify-store instruction.
6105 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6106 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6107 if (UI->getOpcode() == ISD::STORE)
6110 // Otherwise use a regular EFLAGS-setting instruction.
6111 switch (Op.getNode()->getOpcode()) {
6112 default: llvm_unreachable("unexpected operator!");
6113 case ISD::SUB: Opcode = X86ISD::SUB; break;
6114 case ISD::OR: Opcode = X86ISD::OR; break;
6115 case ISD::XOR: Opcode = X86ISD::XOR; break;
6116 case ISD::AND: Opcode = X86ISD::AND; break;
6128 return SDValue(Op.getNode(), 1);
6135 // Emit a CMP with 0, which is the TEST pattern.
6136 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6137 DAG.getConstant(0, Op.getValueType()));
6139 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6140 SmallVector<SDValue, 4> Ops;
6141 for (unsigned i = 0; i != NumOperands; ++i)
6142 Ops.push_back(Op.getOperand(i));
6144 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6145 DAG.ReplaceAllUsesWith(Op, New);
6146 return SDValue(New.getNode(), 1);
6149 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6151 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6152 SelectionDAG &DAG) const {
6153 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6154 if (C->getAPIntValue() == 0)
6155 return EmitTest(Op0, X86CC, DAG);
6157 DebugLoc dl = Op0.getDebugLoc();
6158 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6161 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6162 /// if it's possible.
6163 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6164 DebugLoc dl, SelectionDAG &DAG) const {
6165 SDValue Op0 = And.getOperand(0);
6166 SDValue Op1 = And.getOperand(1);
6167 if (Op0.getOpcode() == ISD::TRUNCATE)
6168 Op0 = Op0.getOperand(0);
6169 if (Op1.getOpcode() == ISD::TRUNCATE)
6170 Op1 = Op1.getOperand(0);
6173 if (Op1.getOpcode() == ISD::SHL)
6174 std::swap(Op0, Op1);
6175 if (Op0.getOpcode() == ISD::SHL) {
6176 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6177 if (And00C->getZExtValue() == 1) {
6178 // If we looked past a truncate, check that it's only truncating away
6180 unsigned BitWidth = Op0.getValueSizeInBits();
6181 unsigned AndBitWidth = And.getValueSizeInBits();
6182 if (BitWidth > AndBitWidth) {
6183 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6184 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6185 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6189 RHS = Op0.getOperand(1);
6191 } else if (Op1.getOpcode() == ISD::Constant) {
6192 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6193 SDValue AndLHS = Op0;
6194 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6195 LHS = AndLHS.getOperand(0);
6196 RHS = AndLHS.getOperand(1);
6200 if (LHS.getNode()) {
6201 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6202 // instruction. Since the shift amount is in-range-or-undefined, we know
6203 // that doing a bittest on the i32 value is ok. We extend to i32 because
6204 // the encoding for the i16 version is larger than the i32 version.
6205 // Also promote i16 to i32 for performance / code size reason.
6206 if (LHS.getValueType() == MVT::i8 ||
6207 LHS.getValueType() == MVT::i16)
6208 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6210 // If the operand types disagree, extend the shift amount to match. Since
6211 // BT ignores high bits (like shifts) we can use anyextend.
6212 if (LHS.getValueType() != RHS.getValueType())
6213 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6215 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6216 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6217 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6218 DAG.getConstant(Cond, MVT::i8), BT);
6224 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6225 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6226 SDValue Op0 = Op.getOperand(0);
6227 SDValue Op1 = Op.getOperand(1);
6228 DebugLoc dl = Op.getDebugLoc();
6229 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6231 // Optimize to BT if possible.
6232 // Lower (X & (1 << N)) == 0 to BT(X, N).
6233 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6234 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6235 if (Op0.getOpcode() == ISD::AND &&
6237 Op1.getOpcode() == ISD::Constant &&
6238 cast<ConstantSDNode>(Op1)->isNullValue() &&
6239 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6240 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6241 if (NewSetCC.getNode())
6245 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6246 if (Op0.getOpcode() == X86ISD::SETCC &&
6247 Op1.getOpcode() == ISD::Constant &&
6248 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6249 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6250 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6251 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6252 bool Invert = (CC == ISD::SETNE) ^
6253 cast<ConstantSDNode>(Op1)->isNullValue();
6255 CCode = X86::GetOppositeBranchCondition(CCode);
6256 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6257 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6260 bool isFP = Op1.getValueType().isFloatingPoint();
6261 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6262 if (X86CC == X86::COND_INVALID)
6265 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6267 // Use sbb x, x to materialize carry bit into a GPR.
6268 if (X86CC == X86::COND_B)
6269 return DAG.getNode(ISD::AND, dl, MVT::i8,
6270 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6271 DAG.getConstant(X86CC, MVT::i8), Cond),
6272 DAG.getConstant(1, MVT::i8));
6274 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6275 DAG.getConstant(X86CC, MVT::i8), Cond);
6278 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6280 SDValue Op0 = Op.getOperand(0);
6281 SDValue Op1 = Op.getOperand(1);
6282 SDValue CC = Op.getOperand(2);
6283 EVT VT = Op.getValueType();
6284 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6285 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6286 DebugLoc dl = Op.getDebugLoc();
6290 EVT VT0 = Op0.getValueType();
6291 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6292 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6295 switch (SetCCOpcode) {
6298 case ISD::SETEQ: SSECC = 0; break;
6300 case ISD::SETGT: Swap = true; // Fallthrough
6302 case ISD::SETOLT: SSECC = 1; break;
6304 case ISD::SETGE: Swap = true; // Fallthrough
6306 case ISD::SETOLE: SSECC = 2; break;
6307 case ISD::SETUO: SSECC = 3; break;
6309 case ISD::SETNE: SSECC = 4; break;
6310 case ISD::SETULE: Swap = true;
6311 case ISD::SETUGE: SSECC = 5; break;
6312 case ISD::SETULT: Swap = true;
6313 case ISD::SETUGT: SSECC = 6; break;
6314 case ISD::SETO: SSECC = 7; break;
6317 std::swap(Op0, Op1);
6319 // In the two special cases we can't handle, emit two comparisons.
6321 if (SetCCOpcode == ISD::SETUEQ) {
6323 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6324 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6325 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6327 else if (SetCCOpcode == ISD::SETONE) {
6329 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6330 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6331 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6333 llvm_unreachable("Illegal FP comparison");
6335 // Handle all other FP comparisons here.
6336 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6339 // We are handling one of the integer comparisons here. Since SSE only has
6340 // GT and EQ comparisons for integer, swapping operands and multiple
6341 // operations may be required for some comparisons.
6342 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6343 bool Swap = false, Invert = false, FlipSigns = false;
6345 switch (VT.getSimpleVT().SimpleTy) {
6348 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6350 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6352 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6353 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6356 switch (SetCCOpcode) {
6358 case ISD::SETNE: Invert = true;
6359 case ISD::SETEQ: Opc = EQOpc; break;
6360 case ISD::SETLT: Swap = true;
6361 case ISD::SETGT: Opc = GTOpc; break;
6362 case ISD::SETGE: Swap = true;
6363 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6364 case ISD::SETULT: Swap = true;
6365 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6366 case ISD::SETUGE: Swap = true;
6367 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6370 std::swap(Op0, Op1);
6372 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6373 // bits of the inputs before performing those operations.
6375 EVT EltVT = VT.getVectorElementType();
6376 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6378 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6379 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6381 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6382 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6385 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6387 // If the logical-not of the result is required, perform that now.
6389 Result = DAG.getNOT(dl, Result, VT);
6394 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6395 static bool isX86LogicalCmp(SDValue Op) {
6396 unsigned Opc = Op.getNode()->getOpcode();
6397 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6399 if (Op.getResNo() == 1 &&
6400 (Opc == X86ISD::ADD ||
6401 Opc == X86ISD::SUB ||
6402 Opc == X86ISD::SMUL ||
6403 Opc == X86ISD::UMUL ||
6404 Opc == X86ISD::INC ||
6405 Opc == X86ISD::DEC ||
6406 Opc == X86ISD::OR ||
6407 Opc == X86ISD::XOR ||
6408 Opc == X86ISD::AND))
6414 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6415 bool addTest = true;
6416 SDValue Cond = Op.getOperand(0);
6417 DebugLoc dl = Op.getDebugLoc();
6420 if (Cond.getOpcode() == ISD::SETCC) {
6421 SDValue NewCond = LowerSETCC(Cond, DAG);
6422 if (NewCond.getNode())
6426 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6427 SDValue Op1 = Op.getOperand(1);
6428 SDValue Op2 = Op.getOperand(2);
6429 if (Cond.getOpcode() == X86ISD::SETCC &&
6430 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6431 SDValue Cmp = Cond.getOperand(1);
6432 if (Cmp.getOpcode() == X86ISD::CMP) {
6433 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6434 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6435 ConstantSDNode *RHSC =
6436 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6437 if (N1C && N1C->isAllOnesValue() &&
6438 N2C && N2C->isNullValue() &&
6439 RHSC && RHSC->isNullValue()) {
6440 SDValue CmpOp0 = Cmp.getOperand(0);
6441 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6442 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6443 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6444 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6449 // Look pass (and (setcc_carry (cmp ...)), 1).
6450 if (Cond.getOpcode() == ISD::AND &&
6451 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6452 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6453 if (C && C->getAPIntValue() == 1)
6454 Cond = Cond.getOperand(0);
6457 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6458 // setting operand in place of the X86ISD::SETCC.
6459 if (Cond.getOpcode() == X86ISD::SETCC ||
6460 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6461 CC = Cond.getOperand(0);
6463 SDValue Cmp = Cond.getOperand(1);
6464 unsigned Opc = Cmp.getOpcode();
6465 EVT VT = Op.getValueType();
6467 bool IllegalFPCMov = false;
6468 if (VT.isFloatingPoint() && !VT.isVector() &&
6469 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6470 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6472 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6473 Opc == X86ISD::BT) { // FIXME
6480 // Look pass the truncate.
6481 if (Cond.getOpcode() == ISD::TRUNCATE)
6482 Cond = Cond.getOperand(0);
6484 // We know the result of AND is compared against zero. Try to match
6486 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6487 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6488 if (NewSetCC.getNode()) {
6489 CC = NewSetCC.getOperand(0);
6490 Cond = NewSetCC.getOperand(1);
6497 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6498 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6501 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6502 // condition is true.
6503 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6504 SDValue Ops[] = { Op2, Op1, CC, Cond };
6505 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6508 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6509 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6510 // from the AND / OR.
6511 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6512 Opc = Op.getOpcode();
6513 if (Opc != ISD::OR && Opc != ISD::AND)
6515 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6516 Op.getOperand(0).hasOneUse() &&
6517 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6518 Op.getOperand(1).hasOneUse());
6521 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6522 // 1 and that the SETCC node has a single use.
6523 static bool isXor1OfSetCC(SDValue Op) {
6524 if (Op.getOpcode() != ISD::XOR)
6526 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6527 if (N1C && N1C->getAPIntValue() == 1) {
6528 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6529 Op.getOperand(0).hasOneUse();
6534 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6535 bool addTest = true;
6536 SDValue Chain = Op.getOperand(0);
6537 SDValue Cond = Op.getOperand(1);
6538 SDValue Dest = Op.getOperand(2);
6539 DebugLoc dl = Op.getDebugLoc();
6542 if (Cond.getOpcode() == ISD::SETCC) {
6543 SDValue NewCond = LowerSETCC(Cond, DAG);
6544 if (NewCond.getNode())
6548 // FIXME: LowerXALUO doesn't handle these!!
6549 else if (Cond.getOpcode() == X86ISD::ADD ||
6550 Cond.getOpcode() == X86ISD::SUB ||
6551 Cond.getOpcode() == X86ISD::SMUL ||
6552 Cond.getOpcode() == X86ISD::UMUL)
6553 Cond = LowerXALUO(Cond, DAG);
6556 // Look pass (and (setcc_carry (cmp ...)), 1).
6557 if (Cond.getOpcode() == ISD::AND &&
6558 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6559 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6560 if (C && C->getAPIntValue() == 1)
6561 Cond = Cond.getOperand(0);
6564 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6565 // setting operand in place of the X86ISD::SETCC.
6566 if (Cond.getOpcode() == X86ISD::SETCC ||
6567 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6568 CC = Cond.getOperand(0);
6570 SDValue Cmp = Cond.getOperand(1);
6571 unsigned Opc = Cmp.getOpcode();
6572 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6573 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6577 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6581 // These can only come from an arithmetic instruction with overflow,
6582 // e.g. SADDO, UADDO.
6583 Cond = Cond.getNode()->getOperand(1);
6590 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6591 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6592 if (CondOpc == ISD::OR) {
6593 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6594 // two branches instead of an explicit OR instruction with a
6596 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6597 isX86LogicalCmp(Cmp)) {
6598 CC = Cond.getOperand(0).getOperand(0);
6599 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6600 Chain, Dest, CC, Cmp);
6601 CC = Cond.getOperand(1).getOperand(0);
6605 } else { // ISD::AND
6606 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6607 // two branches instead of an explicit AND instruction with a
6608 // separate test. However, we only do this if this block doesn't
6609 // have a fall-through edge, because this requires an explicit
6610 // jmp when the condition is false.
6611 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6612 isX86LogicalCmp(Cmp) &&
6613 Op.getNode()->hasOneUse()) {
6614 X86::CondCode CCode =
6615 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6616 CCode = X86::GetOppositeBranchCondition(CCode);
6617 CC = DAG.getConstant(CCode, MVT::i8);
6618 SDNode *User = *Op.getNode()->use_begin();
6619 // Look for an unconditional branch following this conditional branch.
6620 // We need this because we need to reverse the successors in order
6621 // to implement FCMP_OEQ.
6622 if (User->getOpcode() == ISD::BR) {
6623 SDValue FalseBB = User->getOperand(1);
6625 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6626 assert(NewBR == User);
6630 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6631 Chain, Dest, CC, Cmp);
6632 X86::CondCode CCode =
6633 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6634 CCode = X86::GetOppositeBranchCondition(CCode);
6635 CC = DAG.getConstant(CCode, MVT::i8);
6641 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6642 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6643 // It should be transformed during dag combiner except when the condition
6644 // is set by a arithmetics with overflow node.
6645 X86::CondCode CCode =
6646 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6647 CCode = X86::GetOppositeBranchCondition(CCode);
6648 CC = DAG.getConstant(CCode, MVT::i8);
6649 Cond = Cond.getOperand(0).getOperand(1);
6655 // Look pass the truncate.
6656 if (Cond.getOpcode() == ISD::TRUNCATE)
6657 Cond = Cond.getOperand(0);
6659 // We know the result of AND is compared against zero. Try to match
6661 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6662 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6663 if (NewSetCC.getNode()) {
6664 CC = NewSetCC.getOperand(0);
6665 Cond = NewSetCC.getOperand(1);
6672 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6673 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6675 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6676 Chain, Dest, CC, Cond);
6680 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6681 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6682 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6683 // that the guard pages used by the OS virtual memory manager are allocated in
6684 // correct sequence.
6686 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6687 SelectionDAG &DAG) const {
6688 assert(Subtarget->isTargetCygMing() &&
6689 "This should be used only on Cygwin/Mingw targets");
6690 DebugLoc dl = Op.getDebugLoc();
6693 SDValue Chain = Op.getOperand(0);
6694 SDValue Size = Op.getOperand(1);
6695 // FIXME: Ensure alignment here
6699 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6701 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6702 Flag = Chain.getValue(1);
6704 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6706 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6707 Flag = Chain.getValue(1);
6709 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6711 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6712 return DAG.getMergeValues(Ops1, 2, dl);
6715 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6716 MachineFunction &MF = DAG.getMachineFunction();
6717 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6719 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6720 DebugLoc dl = Op.getDebugLoc();
6722 if (!Subtarget->is64Bit()) {
6723 // vastart just stores the address of the VarArgsFrameIndex slot into the
6724 // memory location argument.
6725 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6727 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6732 // gp_offset (0 - 6 * 8)
6733 // fp_offset (48 - 48 + 8 * 16)
6734 // overflow_arg_area (point to parameters coming in memory).
6736 SmallVector<SDValue, 8> MemOps;
6737 SDValue FIN = Op.getOperand(1);
6739 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6740 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6742 FIN, SV, 0, false, false, 0);
6743 MemOps.push_back(Store);
6746 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6747 FIN, DAG.getIntPtrConstant(4));
6748 Store = DAG.getStore(Op.getOperand(0), dl,
6749 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6751 FIN, SV, 0, false, false, 0);
6752 MemOps.push_back(Store);
6754 // Store ptr to overflow_arg_area
6755 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6756 FIN, DAG.getIntPtrConstant(4));
6757 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6759 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6761 MemOps.push_back(Store);
6763 // Store ptr to reg_save_area.
6764 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6765 FIN, DAG.getIntPtrConstant(8));
6766 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6768 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6770 MemOps.push_back(Store);
6771 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6772 &MemOps[0], MemOps.size());
6775 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6776 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6777 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6779 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6783 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6784 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6785 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6786 SDValue Chain = Op.getOperand(0);
6787 SDValue DstPtr = Op.getOperand(1);
6788 SDValue SrcPtr = Op.getOperand(2);
6789 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6790 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6791 DebugLoc dl = Op.getDebugLoc();
6793 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6794 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6795 false, DstSV, 0, SrcSV, 0);
6799 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6800 DebugLoc dl = Op.getDebugLoc();
6801 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6803 default: return SDValue(); // Don't custom lower most intrinsics.
6804 // Comparison intrinsics.
6805 case Intrinsic::x86_sse_comieq_ss:
6806 case Intrinsic::x86_sse_comilt_ss:
6807 case Intrinsic::x86_sse_comile_ss:
6808 case Intrinsic::x86_sse_comigt_ss:
6809 case Intrinsic::x86_sse_comige_ss:
6810 case Intrinsic::x86_sse_comineq_ss:
6811 case Intrinsic::x86_sse_ucomieq_ss:
6812 case Intrinsic::x86_sse_ucomilt_ss:
6813 case Intrinsic::x86_sse_ucomile_ss:
6814 case Intrinsic::x86_sse_ucomigt_ss:
6815 case Intrinsic::x86_sse_ucomige_ss:
6816 case Intrinsic::x86_sse_ucomineq_ss:
6817 case Intrinsic::x86_sse2_comieq_sd:
6818 case Intrinsic::x86_sse2_comilt_sd:
6819 case Intrinsic::x86_sse2_comile_sd:
6820 case Intrinsic::x86_sse2_comigt_sd:
6821 case Intrinsic::x86_sse2_comige_sd:
6822 case Intrinsic::x86_sse2_comineq_sd:
6823 case Intrinsic::x86_sse2_ucomieq_sd:
6824 case Intrinsic::x86_sse2_ucomilt_sd:
6825 case Intrinsic::x86_sse2_ucomile_sd:
6826 case Intrinsic::x86_sse2_ucomigt_sd:
6827 case Intrinsic::x86_sse2_ucomige_sd:
6828 case Intrinsic::x86_sse2_ucomineq_sd: {
6830 ISD::CondCode CC = ISD::SETCC_INVALID;
6833 case Intrinsic::x86_sse_comieq_ss:
6834 case Intrinsic::x86_sse2_comieq_sd:
6838 case Intrinsic::x86_sse_comilt_ss:
6839 case Intrinsic::x86_sse2_comilt_sd:
6843 case Intrinsic::x86_sse_comile_ss:
6844 case Intrinsic::x86_sse2_comile_sd:
6848 case Intrinsic::x86_sse_comigt_ss:
6849 case Intrinsic::x86_sse2_comigt_sd:
6853 case Intrinsic::x86_sse_comige_ss:
6854 case Intrinsic::x86_sse2_comige_sd:
6858 case Intrinsic::x86_sse_comineq_ss:
6859 case Intrinsic::x86_sse2_comineq_sd:
6863 case Intrinsic::x86_sse_ucomieq_ss:
6864 case Intrinsic::x86_sse2_ucomieq_sd:
6865 Opc = X86ISD::UCOMI;
6868 case Intrinsic::x86_sse_ucomilt_ss:
6869 case Intrinsic::x86_sse2_ucomilt_sd:
6870 Opc = X86ISD::UCOMI;
6873 case Intrinsic::x86_sse_ucomile_ss:
6874 case Intrinsic::x86_sse2_ucomile_sd:
6875 Opc = X86ISD::UCOMI;
6878 case Intrinsic::x86_sse_ucomigt_ss:
6879 case Intrinsic::x86_sse2_ucomigt_sd:
6880 Opc = X86ISD::UCOMI;
6883 case Intrinsic::x86_sse_ucomige_ss:
6884 case Intrinsic::x86_sse2_ucomige_sd:
6885 Opc = X86ISD::UCOMI;
6888 case Intrinsic::x86_sse_ucomineq_ss:
6889 case Intrinsic::x86_sse2_ucomineq_sd:
6890 Opc = X86ISD::UCOMI;
6895 SDValue LHS = Op.getOperand(1);
6896 SDValue RHS = Op.getOperand(2);
6897 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6898 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6899 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6900 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6901 DAG.getConstant(X86CC, MVT::i8), Cond);
6902 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6904 // ptest intrinsics. The intrinsic these come from are designed to return
6905 // an integer value, not just an instruction so lower it to the ptest
6906 // pattern and a setcc for the result.
6907 case Intrinsic::x86_sse41_ptestz:
6908 case Intrinsic::x86_sse41_ptestc:
6909 case Intrinsic::x86_sse41_ptestnzc:{
6912 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6913 case Intrinsic::x86_sse41_ptestz:
6915 X86CC = X86::COND_E;
6917 case Intrinsic::x86_sse41_ptestc:
6919 X86CC = X86::COND_B;
6921 case Intrinsic::x86_sse41_ptestnzc:
6923 X86CC = X86::COND_A;
6927 SDValue LHS = Op.getOperand(1);
6928 SDValue RHS = Op.getOperand(2);
6929 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6930 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6931 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6932 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6935 // Fix vector shift instructions where the last operand is a non-immediate
6937 case Intrinsic::x86_sse2_pslli_w:
6938 case Intrinsic::x86_sse2_pslli_d:
6939 case Intrinsic::x86_sse2_pslli_q:
6940 case Intrinsic::x86_sse2_psrli_w:
6941 case Intrinsic::x86_sse2_psrli_d:
6942 case Intrinsic::x86_sse2_psrli_q:
6943 case Intrinsic::x86_sse2_psrai_w:
6944 case Intrinsic::x86_sse2_psrai_d:
6945 case Intrinsic::x86_mmx_pslli_w:
6946 case Intrinsic::x86_mmx_pslli_d:
6947 case Intrinsic::x86_mmx_pslli_q:
6948 case Intrinsic::x86_mmx_psrli_w:
6949 case Intrinsic::x86_mmx_psrli_d:
6950 case Intrinsic::x86_mmx_psrli_q:
6951 case Intrinsic::x86_mmx_psrai_w:
6952 case Intrinsic::x86_mmx_psrai_d: {
6953 SDValue ShAmt = Op.getOperand(2);
6954 if (isa<ConstantSDNode>(ShAmt))
6957 unsigned NewIntNo = 0;
6958 EVT ShAmtVT = MVT::v4i32;
6960 case Intrinsic::x86_sse2_pslli_w:
6961 NewIntNo = Intrinsic::x86_sse2_psll_w;
6963 case Intrinsic::x86_sse2_pslli_d:
6964 NewIntNo = Intrinsic::x86_sse2_psll_d;
6966 case Intrinsic::x86_sse2_pslli_q:
6967 NewIntNo = Intrinsic::x86_sse2_psll_q;
6969 case Intrinsic::x86_sse2_psrli_w:
6970 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6972 case Intrinsic::x86_sse2_psrli_d:
6973 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6975 case Intrinsic::x86_sse2_psrli_q:
6976 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6978 case Intrinsic::x86_sse2_psrai_w:
6979 NewIntNo = Intrinsic::x86_sse2_psra_w;
6981 case Intrinsic::x86_sse2_psrai_d:
6982 NewIntNo = Intrinsic::x86_sse2_psra_d;
6985 ShAmtVT = MVT::v2i32;
6987 case Intrinsic::x86_mmx_pslli_w:
6988 NewIntNo = Intrinsic::x86_mmx_psll_w;
6990 case Intrinsic::x86_mmx_pslli_d:
6991 NewIntNo = Intrinsic::x86_mmx_psll_d;
6993 case Intrinsic::x86_mmx_pslli_q:
6994 NewIntNo = Intrinsic::x86_mmx_psll_q;
6996 case Intrinsic::x86_mmx_psrli_w:
6997 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6999 case Intrinsic::x86_mmx_psrli_d:
7000 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7002 case Intrinsic::x86_mmx_psrli_q:
7003 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7005 case Intrinsic::x86_mmx_psrai_w:
7006 NewIntNo = Intrinsic::x86_mmx_psra_w;
7008 case Intrinsic::x86_mmx_psrai_d:
7009 NewIntNo = Intrinsic::x86_mmx_psra_d;
7011 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7017 // The vector shift intrinsics with scalars uses 32b shift amounts but
7018 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7022 ShOps[1] = DAG.getConstant(0, MVT::i32);
7023 if (ShAmtVT == MVT::v4i32) {
7024 ShOps[2] = DAG.getUNDEF(MVT::i32);
7025 ShOps[3] = DAG.getUNDEF(MVT::i32);
7026 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7028 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7031 EVT VT = Op.getValueType();
7032 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7033 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7034 DAG.getConstant(NewIntNo, MVT::i32),
7035 Op.getOperand(1), ShAmt);
7040 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7041 SelectionDAG &DAG) const {
7042 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7043 MFI->setReturnAddressIsTaken(true);
7045 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7046 DebugLoc dl = Op.getDebugLoc();
7049 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7051 DAG.getConstant(TD->getPointerSize(),
7052 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7053 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7054 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7056 NULL, 0, false, false, 0);
7059 // Just load the return address.
7060 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7061 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7062 RetAddrFI, NULL, 0, false, false, 0);
7065 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7066 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7067 MFI->setFrameAddressIsTaken(true);
7069 EVT VT = Op.getValueType();
7070 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7071 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7072 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7073 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7075 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7080 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7081 SelectionDAG &DAG) const {
7082 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7085 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7086 MachineFunction &MF = DAG.getMachineFunction();
7087 SDValue Chain = Op.getOperand(0);
7088 SDValue Offset = Op.getOperand(1);
7089 SDValue Handler = Op.getOperand(2);
7090 DebugLoc dl = Op.getDebugLoc();
7092 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7094 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7096 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7097 DAG.getIntPtrConstant(-TD->getPointerSize()));
7098 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7099 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7100 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7101 MF.getRegInfo().addLiveOut(StoreAddrReg);
7103 return DAG.getNode(X86ISD::EH_RETURN, dl,
7105 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7108 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7109 SelectionDAG &DAG) const {
7110 SDValue Root = Op.getOperand(0);
7111 SDValue Trmp = Op.getOperand(1); // trampoline
7112 SDValue FPtr = Op.getOperand(2); // nested function
7113 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7114 DebugLoc dl = Op.getDebugLoc();
7116 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7118 if (Subtarget->is64Bit()) {
7119 SDValue OutChains[6];
7121 // Large code-model.
7122 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7123 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7125 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7126 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7128 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7130 // Load the pointer to the nested function into R11.
7131 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7132 SDValue Addr = Trmp;
7133 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7134 Addr, TrmpAddr, 0, false, false, 0);
7136 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7137 DAG.getConstant(2, MVT::i64));
7138 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7141 // Load the 'nest' parameter value into R10.
7142 // R10 is specified in X86CallingConv.td
7143 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7144 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7145 DAG.getConstant(10, MVT::i64));
7146 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7147 Addr, TrmpAddr, 10, false, false, 0);
7149 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7150 DAG.getConstant(12, MVT::i64));
7151 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7154 // Jump to the nested function.
7155 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7156 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7157 DAG.getConstant(20, MVT::i64));
7158 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7159 Addr, TrmpAddr, 20, false, false, 0);
7161 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7162 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7163 DAG.getConstant(22, MVT::i64));
7164 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7165 TrmpAddr, 22, false, false, 0);
7168 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7169 return DAG.getMergeValues(Ops, 2, dl);
7171 const Function *Func =
7172 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7173 CallingConv::ID CC = Func->getCallingConv();
7178 llvm_unreachable("Unsupported calling convention");
7179 case CallingConv::C:
7180 case CallingConv::X86_StdCall: {
7181 // Pass 'nest' parameter in ECX.
7182 // Must be kept in sync with X86CallingConv.td
7185 // Check that ECX wasn't needed by an 'inreg' parameter.
7186 const FunctionType *FTy = Func->getFunctionType();
7187 const AttrListPtr &Attrs = Func->getAttributes();
7189 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7190 unsigned InRegCount = 0;
7193 for (FunctionType::param_iterator I = FTy->param_begin(),
7194 E = FTy->param_end(); I != E; ++I, ++Idx)
7195 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7196 // FIXME: should only count parameters that are lowered to integers.
7197 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7199 if (InRegCount > 2) {
7200 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7205 case CallingConv::X86_FastCall:
7206 case CallingConv::X86_ThisCall:
7207 case CallingConv::Fast:
7208 // Pass 'nest' parameter in EAX.
7209 // Must be kept in sync with X86CallingConv.td
7214 SDValue OutChains[4];
7217 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7218 DAG.getConstant(10, MVT::i32));
7219 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7221 // This is storing the opcode for MOV32ri.
7222 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7223 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7224 OutChains[0] = DAG.getStore(Root, dl,
7225 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7226 Trmp, TrmpAddr, 0, false, false, 0);
7228 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7229 DAG.getConstant(1, MVT::i32));
7230 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7233 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7234 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7235 DAG.getConstant(5, MVT::i32));
7236 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7237 TrmpAddr, 5, false, false, 1);
7239 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7240 DAG.getConstant(6, MVT::i32));
7241 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7245 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7246 return DAG.getMergeValues(Ops, 2, dl);
7250 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7251 SelectionDAG &DAG) const {
7253 The rounding mode is in bits 11:10 of FPSR, and has the following
7260 FLT_ROUNDS, on the other hand, expects the following:
7267 To perform the conversion, we do:
7268 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7271 MachineFunction &MF = DAG.getMachineFunction();
7272 const TargetMachine &TM = MF.getTarget();
7273 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7274 unsigned StackAlignment = TFI.getStackAlignment();
7275 EVT VT = Op.getValueType();
7276 DebugLoc dl = Op.getDebugLoc();
7278 // Save FP Control Word to stack slot
7279 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7280 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7282 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7283 DAG.getEntryNode(), StackSlot);
7285 // Load FP Control Word from stack slot
7286 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7289 // Transform as necessary
7291 DAG.getNode(ISD::SRL, dl, MVT::i16,
7292 DAG.getNode(ISD::AND, dl, MVT::i16,
7293 CWD, DAG.getConstant(0x800, MVT::i16)),
7294 DAG.getConstant(11, MVT::i8));
7296 DAG.getNode(ISD::SRL, dl, MVT::i16,
7297 DAG.getNode(ISD::AND, dl, MVT::i16,
7298 CWD, DAG.getConstant(0x400, MVT::i16)),
7299 DAG.getConstant(9, MVT::i8));
7302 DAG.getNode(ISD::AND, dl, MVT::i16,
7303 DAG.getNode(ISD::ADD, dl, MVT::i16,
7304 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7305 DAG.getConstant(1, MVT::i16)),
7306 DAG.getConstant(3, MVT::i16));
7309 return DAG.getNode((VT.getSizeInBits() < 16 ?
7310 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7313 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7314 EVT VT = Op.getValueType();
7316 unsigned NumBits = VT.getSizeInBits();
7317 DebugLoc dl = Op.getDebugLoc();
7319 Op = Op.getOperand(0);
7320 if (VT == MVT::i8) {
7321 // Zero extend to i32 since there is not an i8 bsr.
7323 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7326 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7327 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7328 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7330 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7333 DAG.getConstant(NumBits+NumBits-1, OpVT),
7334 DAG.getConstant(X86::COND_E, MVT::i8),
7337 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7339 // Finally xor with NumBits-1.
7340 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7343 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7347 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7348 EVT VT = Op.getValueType();
7350 unsigned NumBits = VT.getSizeInBits();
7351 DebugLoc dl = Op.getDebugLoc();
7353 Op = Op.getOperand(0);
7354 if (VT == MVT::i8) {
7356 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7359 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7360 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7361 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7363 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7366 DAG.getConstant(NumBits, OpVT),
7367 DAG.getConstant(X86::COND_E, MVT::i8),
7370 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7373 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7377 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7378 EVT VT = Op.getValueType();
7379 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7380 DebugLoc dl = Op.getDebugLoc();
7382 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7383 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7384 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7385 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7386 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7388 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7389 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7390 // return AloBlo + AloBhi + AhiBlo;
7392 SDValue A = Op.getOperand(0);
7393 SDValue B = Op.getOperand(1);
7395 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7396 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7397 A, DAG.getConstant(32, MVT::i32));
7398 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7399 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7400 B, DAG.getConstant(32, MVT::i32));
7401 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7402 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7404 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7405 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7407 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7408 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7410 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7411 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7412 AloBhi, DAG.getConstant(32, MVT::i32));
7413 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7414 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7415 AhiBlo, DAG.getConstant(32, MVT::i32));
7416 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7417 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7422 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7423 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7424 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7425 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7426 // has only one use.
7427 SDNode *N = Op.getNode();
7428 SDValue LHS = N->getOperand(0);
7429 SDValue RHS = N->getOperand(1);
7430 unsigned BaseOp = 0;
7432 DebugLoc dl = Op.getDebugLoc();
7434 switch (Op.getOpcode()) {
7435 default: llvm_unreachable("Unknown ovf instruction!");
7437 // A subtract of one will be selected as a INC. Note that INC doesn't
7438 // set CF, so we can't do this for UADDO.
7439 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7440 if (C->getAPIntValue() == 1) {
7441 BaseOp = X86ISD::INC;
7445 BaseOp = X86ISD::ADD;
7449 BaseOp = X86ISD::ADD;
7453 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7454 // set CF, so we can't do this for USUBO.
7455 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7456 if (C->getAPIntValue() == 1) {
7457 BaseOp = X86ISD::DEC;
7461 BaseOp = X86ISD::SUB;
7465 BaseOp = X86ISD::SUB;
7469 BaseOp = X86ISD::SMUL;
7473 BaseOp = X86ISD::UMUL;
7478 // Also sets EFLAGS.
7479 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7480 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7483 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7484 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7486 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7490 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7491 EVT T = Op.getValueType();
7492 DebugLoc dl = Op.getDebugLoc();
7495 switch(T.getSimpleVT().SimpleTy) {
7497 assert(false && "Invalid value type!");
7498 case MVT::i8: Reg = X86::AL; size = 1; break;
7499 case MVT::i16: Reg = X86::AX; size = 2; break;
7500 case MVT::i32: Reg = X86::EAX; size = 4; break;
7502 assert(Subtarget->is64Bit() && "Node not type legal!");
7503 Reg = X86::RAX; size = 8;
7506 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7507 Op.getOperand(2), SDValue());
7508 SDValue Ops[] = { cpIn.getValue(0),
7511 DAG.getTargetConstant(size, MVT::i8),
7513 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7514 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7516 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7520 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7521 SelectionDAG &DAG) const {
7522 assert(Subtarget->is64Bit() && "Result not type legalized?");
7523 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7524 SDValue TheChain = Op.getOperand(0);
7525 DebugLoc dl = Op.getDebugLoc();
7526 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7527 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7528 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7530 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7531 DAG.getConstant(32, MVT::i8));
7533 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7536 return DAG.getMergeValues(Ops, 2, dl);
7539 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7540 SelectionDAG &DAG) const {
7541 EVT SrcVT = Op.getOperand(0).getValueType();
7542 EVT DstVT = Op.getValueType();
7543 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7544 Subtarget->hasMMX() && !DisableMMX) &&
7545 "Unexpected custom BIT_CONVERT");
7546 assert((DstVT == MVT::i64 ||
7547 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7548 "Unexpected custom BIT_CONVERT");
7549 // i64 <=> MMX conversions are Legal.
7550 if (SrcVT==MVT::i64 && DstVT.isVector())
7552 if (DstVT==MVT::i64 && SrcVT.isVector())
7554 // MMX <=> MMX conversions are Legal.
7555 if (SrcVT.isVector() && DstVT.isVector())
7557 // All other conversions need to be expanded.
7560 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7561 SDNode *Node = Op.getNode();
7562 DebugLoc dl = Node->getDebugLoc();
7563 EVT T = Node->getValueType(0);
7564 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7565 DAG.getConstant(0, T), Node->getOperand(2));
7566 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7567 cast<AtomicSDNode>(Node)->getMemoryVT(),
7568 Node->getOperand(0),
7569 Node->getOperand(1), negOp,
7570 cast<AtomicSDNode>(Node)->getSrcValue(),
7571 cast<AtomicSDNode>(Node)->getAlignment());
7574 /// LowerOperation - Provide custom lowering hooks for some operations.
7576 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7577 switch (Op.getOpcode()) {
7578 default: llvm_unreachable("Should not custom lower this!");
7579 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7580 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7581 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7582 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7583 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7584 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7585 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7586 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7587 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7588 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7589 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7590 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7591 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7592 case ISD::SHL_PARTS:
7593 case ISD::SRA_PARTS:
7594 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7595 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7596 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7597 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7598 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7599 case ISD::FABS: return LowerFABS(Op, DAG);
7600 case ISD::FNEG: return LowerFNEG(Op, DAG);
7601 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7602 case ISD::SETCC: return LowerSETCC(Op, DAG);
7603 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7604 case ISD::SELECT: return LowerSELECT(Op, DAG);
7605 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7606 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7607 case ISD::VASTART: return LowerVASTART(Op, DAG);
7608 case ISD::VAARG: return LowerVAARG(Op, DAG);
7609 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7610 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7611 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7612 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7613 case ISD::FRAME_TO_ARGS_OFFSET:
7614 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7615 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7616 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7617 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7618 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7619 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7620 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7621 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7627 case ISD::UMULO: return LowerXALUO(Op, DAG);
7628 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7629 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7633 void X86TargetLowering::
7634 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7635 SelectionDAG &DAG, unsigned NewOp) const {
7636 EVT T = Node->getValueType(0);
7637 DebugLoc dl = Node->getDebugLoc();
7638 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7640 SDValue Chain = Node->getOperand(0);
7641 SDValue In1 = Node->getOperand(1);
7642 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7643 Node->getOperand(2), DAG.getIntPtrConstant(0));
7644 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7645 Node->getOperand(2), DAG.getIntPtrConstant(1));
7646 SDValue Ops[] = { Chain, In1, In2L, In2H };
7647 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7649 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7650 cast<MemSDNode>(Node)->getMemOperand());
7651 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7652 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7653 Results.push_back(Result.getValue(2));
7656 /// ReplaceNodeResults - Replace a node with an illegal result type
7657 /// with a new node built out of custom code.
7658 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7659 SmallVectorImpl<SDValue>&Results,
7660 SelectionDAG &DAG) const {
7661 DebugLoc dl = N->getDebugLoc();
7662 switch (N->getOpcode()) {
7664 assert(false && "Do not know how to custom type legalize this operation!");
7666 case ISD::FP_TO_SINT: {
7667 std::pair<SDValue,SDValue> Vals =
7668 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7669 SDValue FIST = Vals.first, StackSlot = Vals.second;
7670 if (FIST.getNode() != 0) {
7671 EVT VT = N->getValueType(0);
7672 // Return a load from the stack slot.
7673 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7678 case ISD::READCYCLECOUNTER: {
7679 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7680 SDValue TheChain = N->getOperand(0);
7681 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7682 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7684 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7686 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7687 SDValue Ops[] = { eax, edx };
7688 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7689 Results.push_back(edx.getValue(1));
7692 case ISD::ATOMIC_CMP_SWAP: {
7693 EVT T = N->getValueType(0);
7694 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7695 SDValue cpInL, cpInH;
7696 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7697 DAG.getConstant(0, MVT::i32));
7698 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7699 DAG.getConstant(1, MVT::i32));
7700 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7701 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7703 SDValue swapInL, swapInH;
7704 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7705 DAG.getConstant(0, MVT::i32));
7706 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7707 DAG.getConstant(1, MVT::i32));
7708 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7710 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7711 swapInL.getValue(1));
7712 SDValue Ops[] = { swapInH.getValue(0),
7714 swapInH.getValue(1) };
7715 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7716 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7717 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7718 MVT::i32, Result.getValue(1));
7719 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7720 MVT::i32, cpOutL.getValue(2));
7721 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7722 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7723 Results.push_back(cpOutH.getValue(1));
7726 case ISD::ATOMIC_LOAD_ADD:
7727 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7729 case ISD::ATOMIC_LOAD_AND:
7730 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7732 case ISD::ATOMIC_LOAD_NAND:
7733 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7735 case ISD::ATOMIC_LOAD_OR:
7736 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7738 case ISD::ATOMIC_LOAD_SUB:
7739 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7741 case ISD::ATOMIC_LOAD_XOR:
7742 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7744 case ISD::ATOMIC_SWAP:
7745 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7750 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7752 default: return NULL;
7753 case X86ISD::BSF: return "X86ISD::BSF";
7754 case X86ISD::BSR: return "X86ISD::BSR";
7755 case X86ISD::SHLD: return "X86ISD::SHLD";
7756 case X86ISD::SHRD: return "X86ISD::SHRD";
7757 case X86ISD::FAND: return "X86ISD::FAND";
7758 case X86ISD::FOR: return "X86ISD::FOR";
7759 case X86ISD::FXOR: return "X86ISD::FXOR";
7760 case X86ISD::FSRL: return "X86ISD::FSRL";
7761 case X86ISD::FILD: return "X86ISD::FILD";
7762 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7763 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7764 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7765 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7766 case X86ISD::FLD: return "X86ISD::FLD";
7767 case X86ISD::FST: return "X86ISD::FST";
7768 case X86ISD::CALL: return "X86ISD::CALL";
7769 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7770 case X86ISD::BT: return "X86ISD::BT";
7771 case X86ISD::CMP: return "X86ISD::CMP";
7772 case X86ISD::COMI: return "X86ISD::COMI";
7773 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7774 case X86ISD::SETCC: return "X86ISD::SETCC";
7775 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7776 case X86ISD::CMOV: return "X86ISD::CMOV";
7777 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7778 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7779 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7780 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7781 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7782 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7783 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7784 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7785 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7786 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7787 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7788 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7789 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7790 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7791 case X86ISD::FMAX: return "X86ISD::FMAX";
7792 case X86ISD::FMIN: return "X86ISD::FMIN";
7793 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7794 case X86ISD::FRCP: return "X86ISD::FRCP";
7795 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7796 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
7797 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7798 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7799 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7800 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7801 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7802 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7803 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7804 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7805 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7806 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7807 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7808 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7809 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7810 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7811 case X86ISD::VSHL: return "X86ISD::VSHL";
7812 case X86ISD::VSRL: return "X86ISD::VSRL";
7813 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7814 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7815 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7816 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7817 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7818 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7819 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7820 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7821 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7822 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7823 case X86ISD::ADD: return "X86ISD::ADD";
7824 case X86ISD::SUB: return "X86ISD::SUB";
7825 case X86ISD::SMUL: return "X86ISD::SMUL";
7826 case X86ISD::UMUL: return "X86ISD::UMUL";
7827 case X86ISD::INC: return "X86ISD::INC";
7828 case X86ISD::DEC: return "X86ISD::DEC";
7829 case X86ISD::OR: return "X86ISD::OR";
7830 case X86ISD::XOR: return "X86ISD::XOR";
7831 case X86ISD::AND: return "X86ISD::AND";
7832 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7833 case X86ISD::PTEST: return "X86ISD::PTEST";
7834 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7835 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7839 // isLegalAddressingMode - Return true if the addressing mode represented
7840 // by AM is legal for this target, for a load/store of the specified type.
7841 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7842 const Type *Ty) const {
7843 // X86 supports extremely general addressing modes.
7844 CodeModel::Model M = getTargetMachine().getCodeModel();
7846 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7847 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7852 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7854 // If a reference to this global requires an extra load, we can't fold it.
7855 if (isGlobalStubReference(GVFlags))
7858 // If BaseGV requires a register for the PIC base, we cannot also have a
7859 // BaseReg specified.
7860 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7863 // If lower 4G is not available, then we must use rip-relative addressing.
7864 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7874 // These scales always work.
7879 // These scales are formed with basereg+scalereg. Only accept if there is
7884 default: // Other stuff never works.
7892 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7893 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7895 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7896 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7897 if (NumBits1 <= NumBits2)
7902 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7903 if (!VT1.isInteger() || !VT2.isInteger())
7905 unsigned NumBits1 = VT1.getSizeInBits();
7906 unsigned NumBits2 = VT2.getSizeInBits();
7907 if (NumBits1 <= NumBits2)
7912 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7913 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7914 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7917 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7918 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7919 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7922 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7923 // i16 instructions are longer (0x66 prefix) and potentially slower.
7924 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7927 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7928 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7929 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7930 /// are assumed to be legal.
7932 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7934 // Very little shuffling can be done for 64-bit vectors right now.
7935 if (VT.getSizeInBits() == 64)
7936 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
7938 // FIXME: pshufb, blends, shifts.
7939 return (VT.getVectorNumElements() == 2 ||
7940 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7941 isMOVLMask(M, VT) ||
7942 isSHUFPMask(M, VT) ||
7943 isPSHUFDMask(M, VT) ||
7944 isPSHUFHWMask(M, VT) ||
7945 isPSHUFLWMask(M, VT) ||
7946 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7947 isUNPCKLMask(M, VT) ||
7948 isUNPCKHMask(M, VT) ||
7949 isUNPCKL_v_undef_Mask(M, VT) ||
7950 isUNPCKH_v_undef_Mask(M, VT));
7954 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7956 unsigned NumElts = VT.getVectorNumElements();
7957 // FIXME: This collection of masks seems suspect.
7960 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7961 return (isMOVLMask(Mask, VT) ||
7962 isCommutedMOVLMask(Mask, VT, true) ||
7963 isSHUFPMask(Mask, VT) ||
7964 isCommutedSHUFPMask(Mask, VT));
7969 //===----------------------------------------------------------------------===//
7970 // X86 Scheduler Hooks
7971 //===----------------------------------------------------------------------===//
7973 // private utility function
7975 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7976 MachineBasicBlock *MBB,
7984 TargetRegisterClass *RC,
7985 bool invSrc) const {
7986 // For the atomic bitwise operator, we generate
7989 // ld t1 = [bitinstr.addr]
7990 // op t2 = t1, [bitinstr.val]
7992 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7994 // fallthrough -->nextMBB
7995 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7996 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7997 MachineFunction::iterator MBBIter = MBB;
8000 /// First build the CFG
8001 MachineFunction *F = MBB->getParent();
8002 MachineBasicBlock *thisMBB = MBB;
8003 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8004 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8005 F->insert(MBBIter, newMBB);
8006 F->insert(MBBIter, nextMBB);
8008 // Move all successors to thisMBB to nextMBB
8009 nextMBB->transferSuccessors(thisMBB);
8011 // Update thisMBB to fall through to newMBB
8012 thisMBB->addSuccessor(newMBB);
8014 // newMBB jumps to itself and fall through to nextMBB
8015 newMBB->addSuccessor(nextMBB);
8016 newMBB->addSuccessor(newMBB);
8018 // Insert instructions into newMBB based on incoming instruction
8019 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8020 "unexpected number of operands");
8021 DebugLoc dl = bInstr->getDebugLoc();
8022 MachineOperand& destOper = bInstr->getOperand(0);
8023 MachineOperand* argOpers[2 + X86AddrNumOperands];
8024 int numArgs = bInstr->getNumOperands() - 1;
8025 for (int i=0; i < numArgs; ++i)
8026 argOpers[i] = &bInstr->getOperand(i+1);
8028 // x86 address has 4 operands: base, index, scale, and displacement
8029 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8030 int valArgIndx = lastAddrIndx + 1;
8032 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8033 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8034 for (int i=0; i <= lastAddrIndx; ++i)
8035 (*MIB).addOperand(*argOpers[i]);
8037 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8039 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8044 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8045 assert((argOpers[valArgIndx]->isReg() ||
8046 argOpers[valArgIndx]->isImm()) &&
8048 if (argOpers[valArgIndx]->isReg())
8049 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8051 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8053 (*MIB).addOperand(*argOpers[valArgIndx]);
8055 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8058 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8059 for (int i=0; i <= lastAddrIndx; ++i)
8060 (*MIB).addOperand(*argOpers[i]);
8062 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8063 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8064 bInstr->memoperands_end());
8066 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8070 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8072 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8076 // private utility function: 64 bit atomics on 32 bit host.
8078 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8079 MachineBasicBlock *MBB,
8084 bool invSrc) const {
8085 // For the atomic bitwise operator, we generate
8086 // thisMBB (instructions are in pairs, except cmpxchg8b)
8087 // ld t1,t2 = [bitinstr.addr]
8089 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8090 // op t5, t6 <- out1, out2, [bitinstr.val]
8091 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8092 // mov ECX, EBX <- t5, t6
8093 // mov EAX, EDX <- t1, t2
8094 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8095 // mov t3, t4 <- EAX, EDX
8097 // result in out1, out2
8098 // fallthrough -->nextMBB
8100 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8101 const unsigned LoadOpc = X86::MOV32rm;
8102 const unsigned copyOpc = X86::MOV32rr;
8103 const unsigned NotOpc = X86::NOT32r;
8104 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8105 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8106 MachineFunction::iterator MBBIter = MBB;
8109 /// First build the CFG
8110 MachineFunction *F = MBB->getParent();
8111 MachineBasicBlock *thisMBB = MBB;
8112 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8113 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8114 F->insert(MBBIter, newMBB);
8115 F->insert(MBBIter, nextMBB);
8117 // Move all successors to thisMBB to nextMBB
8118 nextMBB->transferSuccessors(thisMBB);
8120 // Update thisMBB to fall through to newMBB
8121 thisMBB->addSuccessor(newMBB);
8123 // newMBB jumps to itself and fall through to nextMBB
8124 newMBB->addSuccessor(nextMBB);
8125 newMBB->addSuccessor(newMBB);
8127 DebugLoc dl = bInstr->getDebugLoc();
8128 // Insert instructions into newMBB based on incoming instruction
8129 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8130 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8131 "unexpected number of operands");
8132 MachineOperand& dest1Oper = bInstr->getOperand(0);
8133 MachineOperand& dest2Oper = bInstr->getOperand(1);
8134 MachineOperand* argOpers[2 + X86AddrNumOperands];
8135 for (int i=0; i < 2 + X86AddrNumOperands; ++i) {
8136 argOpers[i] = &bInstr->getOperand(i+2);
8138 // We use some of the operands multiple times, so conservatively just
8139 // clear any kill flags that might be present.
8140 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8141 argOpers[i]->setIsKill(false);
8144 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8145 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8147 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8148 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8149 for (int i=0; i <= lastAddrIndx; ++i)
8150 (*MIB).addOperand(*argOpers[i]);
8151 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8152 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8153 // add 4 to displacement.
8154 for (int i=0; i <= lastAddrIndx-2; ++i)
8155 (*MIB).addOperand(*argOpers[i]);
8156 MachineOperand newOp3 = *(argOpers[3]);
8158 newOp3.setImm(newOp3.getImm()+4);
8160 newOp3.setOffset(newOp3.getOffset()+4);
8161 (*MIB).addOperand(newOp3);
8162 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8164 // t3/4 are defined later, at the bottom of the loop
8165 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8166 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8167 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8168 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8169 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8170 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8172 // The subsequent operations should be using the destination registers of
8173 //the PHI instructions.
8175 t1 = F->getRegInfo().createVirtualRegister(RC);
8176 t2 = F->getRegInfo().createVirtualRegister(RC);
8177 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8178 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8180 t1 = dest1Oper.getReg();
8181 t2 = dest2Oper.getReg();
8184 int valArgIndx = lastAddrIndx + 1;
8185 assert((argOpers[valArgIndx]->isReg() ||
8186 argOpers[valArgIndx]->isImm()) &&
8188 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8189 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8190 if (argOpers[valArgIndx]->isReg())
8191 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8193 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8194 if (regOpcL != X86::MOV32rr)
8196 (*MIB).addOperand(*argOpers[valArgIndx]);
8197 assert(argOpers[valArgIndx + 1]->isReg() ==
8198 argOpers[valArgIndx]->isReg());
8199 assert(argOpers[valArgIndx + 1]->isImm() ==
8200 argOpers[valArgIndx]->isImm());
8201 if (argOpers[valArgIndx + 1]->isReg())
8202 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8204 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8205 if (regOpcH != X86::MOV32rr)
8207 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8209 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8211 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8214 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8216 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8219 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8220 for (int i=0; i <= lastAddrIndx; ++i)
8221 (*MIB).addOperand(*argOpers[i]);
8223 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8224 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8225 bInstr->memoperands_end());
8227 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8228 MIB.addReg(X86::EAX);
8229 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8230 MIB.addReg(X86::EDX);
8233 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8235 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8239 // private utility function
8241 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8242 MachineBasicBlock *MBB,
8243 unsigned cmovOpc) const {
8244 // For the atomic min/max operator, we generate
8247 // ld t1 = [min/max.addr]
8248 // mov t2 = [min/max.val]
8250 // cmov[cond] t2 = t1
8252 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8254 // fallthrough -->nextMBB
8256 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8257 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8258 MachineFunction::iterator MBBIter = MBB;
8261 /// First build the CFG
8262 MachineFunction *F = MBB->getParent();
8263 MachineBasicBlock *thisMBB = MBB;
8264 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8265 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8266 F->insert(MBBIter, newMBB);
8267 F->insert(MBBIter, nextMBB);
8269 // Move all successors of thisMBB to nextMBB
8270 nextMBB->transferSuccessors(thisMBB);
8272 // Update thisMBB to fall through to newMBB
8273 thisMBB->addSuccessor(newMBB);
8275 // newMBB jumps to newMBB and fall through to nextMBB
8276 newMBB->addSuccessor(nextMBB);
8277 newMBB->addSuccessor(newMBB);
8279 DebugLoc dl = mInstr->getDebugLoc();
8280 // Insert instructions into newMBB based on incoming instruction
8281 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8282 "unexpected number of operands");
8283 MachineOperand& destOper = mInstr->getOperand(0);
8284 MachineOperand* argOpers[2 + X86AddrNumOperands];
8285 int numArgs = mInstr->getNumOperands() - 1;
8286 for (int i=0; i < numArgs; ++i)
8287 argOpers[i] = &mInstr->getOperand(i+1);
8289 // x86 address has 4 operands: base, index, scale, and displacement
8290 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8291 int valArgIndx = lastAddrIndx + 1;
8293 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8294 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8295 for (int i=0; i <= lastAddrIndx; ++i)
8296 (*MIB).addOperand(*argOpers[i]);
8298 // We only support register and immediate values
8299 assert((argOpers[valArgIndx]->isReg() ||
8300 argOpers[valArgIndx]->isImm()) &&
8303 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8304 if (argOpers[valArgIndx]->isReg())
8305 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8307 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8308 (*MIB).addOperand(*argOpers[valArgIndx]);
8310 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8313 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8318 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8319 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8323 // Cmp and exchange if none has modified the memory location
8324 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8325 for (int i=0; i <= lastAddrIndx; ++i)
8326 (*MIB).addOperand(*argOpers[i]);
8328 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8329 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8330 mInstr->memoperands_end());
8332 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8333 MIB.addReg(X86::EAX);
8336 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8338 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8342 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8343 // all of this code can be replaced with that in the .td file.
8345 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8346 unsigned numArgs, bool memArg) const {
8348 MachineFunction *F = BB->getParent();
8349 DebugLoc dl = MI->getDebugLoc();
8350 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8354 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8356 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8358 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8360 for (unsigned i = 0; i < numArgs; ++i) {
8361 MachineOperand &Op = MI->getOperand(i+1);
8363 if (!(Op.isReg() && Op.isImplicit()))
8367 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8370 F->DeleteMachineInstr(MI);
8376 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8378 MachineBasicBlock *MBB) const {
8379 // Emit code to save XMM registers to the stack. The ABI says that the
8380 // number of registers to save is given in %al, so it's theoretically
8381 // possible to do an indirect jump trick to avoid saving all of them,
8382 // however this code takes a simpler approach and just executes all
8383 // of the stores if %al is non-zero. It's less code, and it's probably
8384 // easier on the hardware branch predictor, and stores aren't all that
8385 // expensive anyway.
8387 // Create the new basic blocks. One block contains all the XMM stores,
8388 // and one block is the final destination regardless of whether any
8389 // stores were performed.
8390 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8391 MachineFunction *F = MBB->getParent();
8392 MachineFunction::iterator MBBIter = MBB;
8394 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8395 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8396 F->insert(MBBIter, XMMSaveMBB);
8397 F->insert(MBBIter, EndMBB);
8400 // Move any original successors of MBB to the end block.
8401 EndMBB->transferSuccessors(MBB);
8402 // The original block will now fall through to the XMM save block.
8403 MBB->addSuccessor(XMMSaveMBB);
8404 // The XMMSaveMBB will fall through to the end block.
8405 XMMSaveMBB->addSuccessor(EndMBB);
8407 // Now add the instructions.
8408 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8409 DebugLoc DL = MI->getDebugLoc();
8411 unsigned CountReg = MI->getOperand(0).getReg();
8412 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8413 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8415 if (!Subtarget->isTargetWin64()) {
8416 // If %al is 0, branch around the XMM save block.
8417 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8418 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8419 MBB->addSuccessor(EndMBB);
8422 // In the XMM save block, save all the XMM argument registers.
8423 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8424 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8425 MachineMemOperand *MMO =
8426 F->getMachineMemOperand(
8427 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8428 MachineMemOperand::MOStore, Offset,
8429 /*Size=*/16, /*Align=*/16);
8430 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8431 .addFrameIndex(RegSaveFrameIndex)
8432 .addImm(/*Scale=*/1)
8433 .addReg(/*IndexReg=*/0)
8434 .addImm(/*Disp=*/Offset)
8435 .addReg(/*Segment=*/0)
8436 .addReg(MI->getOperand(i).getReg())
8437 .addMemOperand(MMO);
8440 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8446 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8447 MachineBasicBlock *BB) const {
8448 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8449 DebugLoc DL = MI->getDebugLoc();
8451 // To "insert" a SELECT_CC instruction, we actually have to insert the
8452 // diamond control-flow pattern. The incoming instruction knows the
8453 // destination vreg to set, the condition code register to branch on, the
8454 // true/false values to select between, and a branch opcode to use.
8455 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8456 MachineFunction::iterator It = BB;
8462 // cmpTY ccX, r1, r2
8464 // fallthrough --> copy0MBB
8465 MachineBasicBlock *thisMBB = BB;
8466 MachineFunction *F = BB->getParent();
8467 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8468 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8470 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8472 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8473 F->insert(It, copy0MBB);
8474 F->insert(It, sinkMBB);
8476 // Update machine-CFG edges by first adding all successors of the current
8477 // block to the new block which will contain the Phi node for the select.
8478 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8479 E = BB->succ_end(); I != E; ++I)
8480 sinkMBB->addSuccessor(*I);
8482 // Next, remove all successors of the current block, and add the true
8483 // and fallthrough blocks as its successors.
8484 while (!BB->succ_empty())
8485 BB->removeSuccessor(BB->succ_begin());
8487 // Add the true and fallthrough blocks as its successors.
8488 BB->addSuccessor(copy0MBB);
8489 BB->addSuccessor(sinkMBB);
8491 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8492 // live into the sink and copy blocks.
8493 const MachineFunction *MF = BB->getParent();
8494 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8495 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8496 const MachineInstr *Term = BB->getFirstTerminator();
8498 for (unsigned I = 0, E = Term->getNumOperands(); I != E; ++I) {
8499 const MachineOperand &MO = Term->getOperand(I);
8500 if (!MO.isReg() || MO.isKill() || MO.isDead()) continue;
8501 unsigned Reg = MO.getReg();
8502 if (Reg != X86::EFLAGS) continue;
8503 copy0MBB->addLiveIn(Reg);
8504 sinkMBB->addLiveIn(Reg);
8508 // %FalseValue = ...
8509 // # fallthrough to sinkMBB
8510 copy0MBB->addSuccessor(sinkMBB);
8513 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8515 BuildMI(sinkMBB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8516 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8517 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8519 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8524 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8525 MachineBasicBlock *BB) const {
8526 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8527 DebugLoc DL = MI->getDebugLoc();
8528 MachineFunction *F = BB->getParent();
8530 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8531 // non-trivial part is impdef of ESP.
8532 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8535 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8536 .addExternalSymbol("_alloca")
8537 .addReg(X86::EAX, RegState::Implicit)
8538 .addReg(X86::ESP, RegState::Implicit)
8539 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8540 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8542 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8547 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8548 MachineBasicBlock *BB) const {
8549 // This is pretty easy. We're taking the value that we received from
8550 // our load from the relocation, sticking it in either RDI (x86-64)
8551 // or EAX and doing an indirect call. The return value will then
8552 // be in the normal return register.
8553 const X86InstrInfo *TII
8554 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8555 DebugLoc DL = MI->getDebugLoc();
8556 MachineFunction *F = BB->getParent();
8558 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8560 if (Subtarget->is64Bit()) {
8561 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV64rm), X86::RDI)
8563 .addImm(0).addReg(0)
8564 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8565 MI->getOperand(3).getTargetFlags())
8567 MIB = BuildMI(BB, DL, TII->get(X86::CALL64m));
8568 addDirectMem(MIB, X86::RDI).addReg(0);
8569 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8570 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV32rm), X86::EAX)
8572 .addImm(0).addReg(0)
8573 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8574 MI->getOperand(3).getTargetFlags())
8576 MIB = BuildMI(BB, DL, TII->get(X86::CALL32m));
8577 addDirectMem(MIB, X86::EAX).addReg(0);
8579 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV32rm), X86::EAX)
8580 .addReg(TII->getGlobalBaseReg(F))
8581 .addImm(0).addReg(0)
8582 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8583 MI->getOperand(3).getTargetFlags())
8585 MIB = BuildMI(BB, DL, TII->get(X86::CALL32m));
8586 addDirectMem(MIB, X86::EAX).addReg(0);
8589 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8594 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8595 MachineBasicBlock *BB) const {
8596 switch (MI->getOpcode()) {
8597 default: assert(false && "Unexpected instr type to insert");
8598 case X86::MINGW_ALLOCA:
8599 return EmitLoweredMingwAlloca(MI, BB);
8600 case X86::TLSCall_32:
8601 case X86::TLSCall_64:
8602 return EmitLoweredTLSCall(MI, BB);
8604 case X86::CMOV_V1I64:
8605 case X86::CMOV_FR32:
8606 case X86::CMOV_FR64:
8607 case X86::CMOV_V4F32:
8608 case X86::CMOV_V2F64:
8609 case X86::CMOV_V2I64:
8610 case X86::CMOV_GR16:
8611 case X86::CMOV_GR32:
8612 case X86::CMOV_RFP32:
8613 case X86::CMOV_RFP64:
8614 case X86::CMOV_RFP80:
8615 return EmitLoweredSelect(MI, BB);
8617 case X86::FP32_TO_INT16_IN_MEM:
8618 case X86::FP32_TO_INT32_IN_MEM:
8619 case X86::FP32_TO_INT64_IN_MEM:
8620 case X86::FP64_TO_INT16_IN_MEM:
8621 case X86::FP64_TO_INT32_IN_MEM:
8622 case X86::FP64_TO_INT64_IN_MEM:
8623 case X86::FP80_TO_INT16_IN_MEM:
8624 case X86::FP80_TO_INT32_IN_MEM:
8625 case X86::FP80_TO_INT64_IN_MEM: {
8626 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8627 DebugLoc DL = MI->getDebugLoc();
8629 // Change the floating point control register to use "round towards zero"
8630 // mode when truncating to an integer value.
8631 MachineFunction *F = BB->getParent();
8632 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8633 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8635 // Load the old value of the high byte of the control word...
8637 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8638 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8641 // Set the high part to be round to zero...
8642 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8645 // Reload the modified control word now...
8646 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8648 // Restore the memory image of control word to original value
8649 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8652 // Get the X86 opcode to use.
8654 switch (MI->getOpcode()) {
8655 default: llvm_unreachable("illegal opcode!");
8656 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8657 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8658 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8659 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8660 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8661 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8662 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8663 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8664 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8668 MachineOperand &Op = MI->getOperand(0);
8670 AM.BaseType = X86AddressMode::RegBase;
8671 AM.Base.Reg = Op.getReg();
8673 AM.BaseType = X86AddressMode::FrameIndexBase;
8674 AM.Base.FrameIndex = Op.getIndex();
8676 Op = MI->getOperand(1);
8678 AM.Scale = Op.getImm();
8679 Op = MI->getOperand(2);
8681 AM.IndexReg = Op.getImm();
8682 Op = MI->getOperand(3);
8683 if (Op.isGlobal()) {
8684 AM.GV = Op.getGlobal();
8686 AM.Disp = Op.getImm();
8688 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8689 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8691 // Reload the original control word now.
8692 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8694 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8697 // String/text processing lowering.
8698 case X86::PCMPISTRM128REG:
8699 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8700 case X86::PCMPISTRM128MEM:
8701 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8702 case X86::PCMPESTRM128REG:
8703 return EmitPCMP(MI, BB, 5, false /* in mem */);
8704 case X86::PCMPESTRM128MEM:
8705 return EmitPCMP(MI, BB, 5, true /* in mem */);
8708 case X86::ATOMAND32:
8709 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8710 X86::AND32ri, X86::MOV32rm,
8711 X86::LCMPXCHG32, X86::MOV32rr,
8712 X86::NOT32r, X86::EAX,
8713 X86::GR32RegisterClass);
8715 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8716 X86::OR32ri, X86::MOV32rm,
8717 X86::LCMPXCHG32, X86::MOV32rr,
8718 X86::NOT32r, X86::EAX,
8719 X86::GR32RegisterClass);
8720 case X86::ATOMXOR32:
8721 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8722 X86::XOR32ri, X86::MOV32rm,
8723 X86::LCMPXCHG32, X86::MOV32rr,
8724 X86::NOT32r, X86::EAX,
8725 X86::GR32RegisterClass);
8726 case X86::ATOMNAND32:
8727 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8728 X86::AND32ri, X86::MOV32rm,
8729 X86::LCMPXCHG32, X86::MOV32rr,
8730 X86::NOT32r, X86::EAX,
8731 X86::GR32RegisterClass, true);
8732 case X86::ATOMMIN32:
8733 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8734 case X86::ATOMMAX32:
8735 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8736 case X86::ATOMUMIN32:
8737 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8738 case X86::ATOMUMAX32:
8739 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8741 case X86::ATOMAND16:
8742 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8743 X86::AND16ri, X86::MOV16rm,
8744 X86::LCMPXCHG16, X86::MOV16rr,
8745 X86::NOT16r, X86::AX,
8746 X86::GR16RegisterClass);
8748 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8749 X86::OR16ri, X86::MOV16rm,
8750 X86::LCMPXCHG16, X86::MOV16rr,
8751 X86::NOT16r, X86::AX,
8752 X86::GR16RegisterClass);
8753 case X86::ATOMXOR16:
8754 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8755 X86::XOR16ri, X86::MOV16rm,
8756 X86::LCMPXCHG16, X86::MOV16rr,
8757 X86::NOT16r, X86::AX,
8758 X86::GR16RegisterClass);
8759 case X86::ATOMNAND16:
8760 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8761 X86::AND16ri, X86::MOV16rm,
8762 X86::LCMPXCHG16, X86::MOV16rr,
8763 X86::NOT16r, X86::AX,
8764 X86::GR16RegisterClass, true);
8765 case X86::ATOMMIN16:
8766 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8767 case X86::ATOMMAX16:
8768 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8769 case X86::ATOMUMIN16:
8770 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8771 case X86::ATOMUMAX16:
8772 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8775 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8776 X86::AND8ri, X86::MOV8rm,
8777 X86::LCMPXCHG8, X86::MOV8rr,
8778 X86::NOT8r, X86::AL,
8779 X86::GR8RegisterClass);
8781 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8782 X86::OR8ri, X86::MOV8rm,
8783 X86::LCMPXCHG8, X86::MOV8rr,
8784 X86::NOT8r, X86::AL,
8785 X86::GR8RegisterClass);
8787 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8788 X86::XOR8ri, X86::MOV8rm,
8789 X86::LCMPXCHG8, X86::MOV8rr,
8790 X86::NOT8r, X86::AL,
8791 X86::GR8RegisterClass);
8792 case X86::ATOMNAND8:
8793 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8794 X86::AND8ri, X86::MOV8rm,
8795 X86::LCMPXCHG8, X86::MOV8rr,
8796 X86::NOT8r, X86::AL,
8797 X86::GR8RegisterClass, true);
8798 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8799 // This group is for 64-bit host.
8800 case X86::ATOMAND64:
8801 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8802 X86::AND64ri32, X86::MOV64rm,
8803 X86::LCMPXCHG64, X86::MOV64rr,
8804 X86::NOT64r, X86::RAX,
8805 X86::GR64RegisterClass);
8807 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8808 X86::OR64ri32, X86::MOV64rm,
8809 X86::LCMPXCHG64, X86::MOV64rr,
8810 X86::NOT64r, X86::RAX,
8811 X86::GR64RegisterClass);
8812 case X86::ATOMXOR64:
8813 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8814 X86::XOR64ri32, X86::MOV64rm,
8815 X86::LCMPXCHG64, X86::MOV64rr,
8816 X86::NOT64r, X86::RAX,
8817 X86::GR64RegisterClass);
8818 case X86::ATOMNAND64:
8819 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8820 X86::AND64ri32, X86::MOV64rm,
8821 X86::LCMPXCHG64, X86::MOV64rr,
8822 X86::NOT64r, X86::RAX,
8823 X86::GR64RegisterClass, true);
8824 case X86::ATOMMIN64:
8825 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8826 case X86::ATOMMAX64:
8827 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8828 case X86::ATOMUMIN64:
8829 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8830 case X86::ATOMUMAX64:
8831 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8833 // This group does 64-bit operations on a 32-bit host.
8834 case X86::ATOMAND6432:
8835 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8836 X86::AND32rr, X86::AND32rr,
8837 X86::AND32ri, X86::AND32ri,
8839 case X86::ATOMOR6432:
8840 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8841 X86::OR32rr, X86::OR32rr,
8842 X86::OR32ri, X86::OR32ri,
8844 case X86::ATOMXOR6432:
8845 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8846 X86::XOR32rr, X86::XOR32rr,
8847 X86::XOR32ri, X86::XOR32ri,
8849 case X86::ATOMNAND6432:
8850 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8851 X86::AND32rr, X86::AND32rr,
8852 X86::AND32ri, X86::AND32ri,
8854 case X86::ATOMADD6432:
8855 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8856 X86::ADD32rr, X86::ADC32rr,
8857 X86::ADD32ri, X86::ADC32ri,
8859 case X86::ATOMSUB6432:
8860 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8861 X86::SUB32rr, X86::SBB32rr,
8862 X86::SUB32ri, X86::SBB32ri,
8864 case X86::ATOMSWAP6432:
8865 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8866 X86::MOV32rr, X86::MOV32rr,
8867 X86::MOV32ri, X86::MOV32ri,
8869 case X86::VASTART_SAVE_XMM_REGS:
8870 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8874 //===----------------------------------------------------------------------===//
8875 // X86 Optimization Hooks
8876 //===----------------------------------------------------------------------===//
8878 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8882 const SelectionDAG &DAG,
8883 unsigned Depth) const {
8884 unsigned Opc = Op.getOpcode();
8885 assert((Opc >= ISD::BUILTIN_OP_END ||
8886 Opc == ISD::INTRINSIC_WO_CHAIN ||
8887 Opc == ISD::INTRINSIC_W_CHAIN ||
8888 Opc == ISD::INTRINSIC_VOID) &&
8889 "Should use MaskedValueIsZero if you don't know whether Op"
8890 " is a target node!");
8892 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8904 // These nodes' second result is a boolean.
8905 if (Op.getResNo() == 0)
8909 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8910 Mask.getBitWidth() - 1);
8915 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8916 /// node is a GlobalAddress + offset.
8917 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8918 const GlobalValue* &GA,
8919 int64_t &Offset) const {
8920 if (N->getOpcode() == X86ISD::Wrapper) {
8921 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8922 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8923 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8927 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8930 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8931 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8932 /// if the load addresses are consecutive, non-overlapping, and in the right
8934 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8935 const TargetLowering &TLI) {
8936 DebugLoc dl = N->getDebugLoc();
8937 EVT VT = N->getValueType(0);
8938 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8940 if (VT.getSizeInBits() != 128)
8943 SmallVector<SDValue, 16> Elts;
8944 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8945 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8947 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8950 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8951 /// and convert it from being a bunch of shuffles and extracts to a simple
8952 /// store and scalar loads to extract the elements.
8953 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8954 const TargetLowering &TLI) {
8955 SDValue InputVector = N->getOperand(0);
8957 // Only operate on vectors of 4 elements, where the alternative shuffling
8958 // gets to be more expensive.
8959 if (InputVector.getValueType() != MVT::v4i32)
8962 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8963 // single use which is a sign-extend or zero-extend, and all elements are
8965 SmallVector<SDNode *, 4> Uses;
8966 unsigned ExtractedElements = 0;
8967 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8968 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8969 if (UI.getUse().getResNo() != InputVector.getResNo())
8972 SDNode *Extract = *UI;
8973 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8976 if (Extract->getValueType(0) != MVT::i32)
8978 if (!Extract->hasOneUse())
8980 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
8981 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
8983 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
8986 // Record which element was extracted.
8987 ExtractedElements |=
8988 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
8990 Uses.push_back(Extract);
8993 // If not all the elements were used, this may not be worthwhile.
8994 if (ExtractedElements != 15)
8997 // Ok, we've now decided to do the transformation.
8998 DebugLoc dl = InputVector.getDebugLoc();
9000 // Store the value to a temporary stack slot.
9001 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9002 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
9005 // Replace each use (extract) with a load of the appropriate element.
9006 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9007 UE = Uses.end(); UI != UE; ++UI) {
9008 SDNode *Extract = *UI;
9010 // Compute the element's address.
9011 SDValue Idx = Extract->getOperand(1);
9013 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9014 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9015 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9017 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
9020 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
9021 NULL, 0, false, false, 0);
9023 // Replace the exact with the load.
9024 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9027 // The replacement was made in place; don't return anything.
9031 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9032 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9033 const X86Subtarget *Subtarget) {
9034 DebugLoc DL = N->getDebugLoc();
9035 SDValue Cond = N->getOperand(0);
9036 // Get the LHS/RHS of the select.
9037 SDValue LHS = N->getOperand(1);
9038 SDValue RHS = N->getOperand(2);
9040 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9041 // instructions match the semantics of the common C idiom x<y?x:y but not
9042 // x<=y?x:y, because of how they handle negative zero (which can be
9043 // ignored in unsafe-math mode).
9044 if (Subtarget->hasSSE2() &&
9045 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9046 Cond.getOpcode() == ISD::SETCC) {
9047 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9049 unsigned Opcode = 0;
9050 // Check for x CC y ? x : y.
9051 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9052 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9056 // Converting this to a min would handle NaNs incorrectly, and swapping
9057 // the operands would cause it to handle comparisons between positive
9058 // and negative zero incorrectly.
9059 if (!FiniteOnlyFPMath() &&
9060 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9061 if (!UnsafeFPMath &&
9062 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9064 std::swap(LHS, RHS);
9066 Opcode = X86ISD::FMIN;
9069 // Converting this to a min would handle comparisons between positive
9070 // and negative zero incorrectly.
9071 if (!UnsafeFPMath &&
9072 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9074 Opcode = X86ISD::FMIN;
9077 // Converting this to a min would handle both negative zeros and NaNs
9078 // incorrectly, but we can swap the operands to fix both.
9079 std::swap(LHS, RHS);
9083 Opcode = X86ISD::FMIN;
9087 // Converting this to a max would handle comparisons between positive
9088 // and negative zero incorrectly.
9089 if (!UnsafeFPMath &&
9090 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9092 Opcode = X86ISD::FMAX;
9095 // Converting this to a max would handle NaNs incorrectly, and swapping
9096 // the operands would cause it to handle comparisons between positive
9097 // and negative zero incorrectly.
9098 if (!FiniteOnlyFPMath() &&
9099 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9100 if (!UnsafeFPMath &&
9101 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9103 std::swap(LHS, RHS);
9105 Opcode = X86ISD::FMAX;
9108 // Converting this to a max would handle both negative zeros and NaNs
9109 // incorrectly, but we can swap the operands to fix both.
9110 std::swap(LHS, RHS);
9114 Opcode = X86ISD::FMAX;
9117 // Check for x CC y ? y : x -- a min/max with reversed arms.
9118 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9119 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9123 // Converting this to a min would handle comparisons between positive
9124 // and negative zero incorrectly, and swapping the operands would
9125 // cause it to handle NaNs incorrectly.
9126 if (!UnsafeFPMath &&
9127 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9128 if (!FiniteOnlyFPMath() &&
9129 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9131 std::swap(LHS, RHS);
9133 Opcode = X86ISD::FMIN;
9136 // Converting this to a min would handle NaNs incorrectly.
9137 if (!UnsafeFPMath &&
9138 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9140 Opcode = X86ISD::FMIN;
9143 // Converting this to a min would handle both negative zeros and NaNs
9144 // incorrectly, but we can swap the operands to fix both.
9145 std::swap(LHS, RHS);
9149 Opcode = X86ISD::FMIN;
9153 // Converting this to a max would handle NaNs incorrectly.
9154 if (!FiniteOnlyFPMath() &&
9155 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9157 Opcode = X86ISD::FMAX;
9160 // Converting this to a max would handle comparisons between positive
9161 // and negative zero incorrectly, and swapping the operands would
9162 // cause it to handle NaNs incorrectly.
9163 if (!UnsafeFPMath &&
9164 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9165 if (!FiniteOnlyFPMath() &&
9166 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9168 std::swap(LHS, RHS);
9170 Opcode = X86ISD::FMAX;
9173 // Converting this to a max would handle both negative zeros and NaNs
9174 // incorrectly, but we can swap the operands to fix both.
9175 std::swap(LHS, RHS);
9179 Opcode = X86ISD::FMAX;
9185 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9188 // If this is a select between two integer constants, try to do some
9190 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9191 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9192 // Don't do this for crazy integer types.
9193 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9194 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9195 // so that TrueC (the true value) is larger than FalseC.
9196 bool NeedsCondInvert = false;
9198 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9199 // Efficiently invertible.
9200 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9201 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9202 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9203 NeedsCondInvert = true;
9204 std::swap(TrueC, FalseC);
9207 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9208 if (FalseC->getAPIntValue() == 0 &&
9209 TrueC->getAPIntValue().isPowerOf2()) {
9210 if (NeedsCondInvert) // Invert the condition if needed.
9211 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9212 DAG.getConstant(1, Cond.getValueType()));
9214 // Zero extend the condition if needed.
9215 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9217 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9218 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9219 DAG.getConstant(ShAmt, MVT::i8));
9222 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9223 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9224 if (NeedsCondInvert) // Invert the condition if needed.
9225 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9226 DAG.getConstant(1, Cond.getValueType()));
9228 // Zero extend the condition if needed.
9229 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9230 FalseC->getValueType(0), Cond);
9231 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9232 SDValue(FalseC, 0));
9235 // Optimize cases that will turn into an LEA instruction. This requires
9236 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9237 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9238 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9239 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9241 bool isFastMultiplier = false;
9243 switch ((unsigned char)Diff) {
9245 case 1: // result = add base, cond
9246 case 2: // result = lea base( , cond*2)
9247 case 3: // result = lea base(cond, cond*2)
9248 case 4: // result = lea base( , cond*4)
9249 case 5: // result = lea base(cond, cond*4)
9250 case 8: // result = lea base( , cond*8)
9251 case 9: // result = lea base(cond, cond*8)
9252 isFastMultiplier = true;
9257 if (isFastMultiplier) {
9258 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9259 if (NeedsCondInvert) // Invert the condition if needed.
9260 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9261 DAG.getConstant(1, Cond.getValueType()));
9263 // Zero extend the condition if needed.
9264 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9266 // Scale the condition by the difference.
9268 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9269 DAG.getConstant(Diff, Cond.getValueType()));
9271 // Add the base if non-zero.
9272 if (FalseC->getAPIntValue() != 0)
9273 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9274 SDValue(FalseC, 0));
9284 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9285 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9286 TargetLowering::DAGCombinerInfo &DCI) {
9287 DebugLoc DL = N->getDebugLoc();
9289 // If the flag operand isn't dead, don't touch this CMOV.
9290 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9293 // If this is a select between two integer constants, try to do some
9294 // optimizations. Note that the operands are ordered the opposite of SELECT
9296 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9297 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9298 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9299 // larger than FalseC (the false value).
9300 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9302 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9303 CC = X86::GetOppositeBranchCondition(CC);
9304 std::swap(TrueC, FalseC);
9307 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9308 // This is efficient for any integer data type (including i8/i16) and
9310 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9311 SDValue Cond = N->getOperand(3);
9312 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9313 DAG.getConstant(CC, MVT::i8), Cond);
9315 // Zero extend the condition if needed.
9316 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9318 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9319 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9320 DAG.getConstant(ShAmt, MVT::i8));
9321 if (N->getNumValues() == 2) // Dead flag value?
9322 return DCI.CombineTo(N, Cond, SDValue());
9326 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9327 // for any integer data type, including i8/i16.
9328 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9329 SDValue Cond = N->getOperand(3);
9330 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9331 DAG.getConstant(CC, MVT::i8), Cond);
9333 // Zero extend the condition if needed.
9334 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9335 FalseC->getValueType(0), Cond);
9336 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9337 SDValue(FalseC, 0));
9339 if (N->getNumValues() == 2) // Dead flag value?
9340 return DCI.CombineTo(N, Cond, SDValue());
9344 // Optimize cases that will turn into an LEA instruction. This requires
9345 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9346 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9347 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9348 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9350 bool isFastMultiplier = false;
9352 switch ((unsigned char)Diff) {
9354 case 1: // result = add base, cond
9355 case 2: // result = lea base( , cond*2)
9356 case 3: // result = lea base(cond, cond*2)
9357 case 4: // result = lea base( , cond*4)
9358 case 5: // result = lea base(cond, cond*4)
9359 case 8: // result = lea base( , cond*8)
9360 case 9: // result = lea base(cond, cond*8)
9361 isFastMultiplier = true;
9366 if (isFastMultiplier) {
9367 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9368 SDValue Cond = N->getOperand(3);
9369 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9370 DAG.getConstant(CC, MVT::i8), Cond);
9371 // Zero extend the condition if needed.
9372 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9374 // Scale the condition by the difference.
9376 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9377 DAG.getConstant(Diff, Cond.getValueType()));
9379 // Add the base if non-zero.
9380 if (FalseC->getAPIntValue() != 0)
9381 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9382 SDValue(FalseC, 0));
9383 if (N->getNumValues() == 2) // Dead flag value?
9384 return DCI.CombineTo(N, Cond, SDValue());
9394 /// PerformMulCombine - Optimize a single multiply with constant into two
9395 /// in order to implement it with two cheaper instructions, e.g.
9396 /// LEA + SHL, LEA + LEA.
9397 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9398 TargetLowering::DAGCombinerInfo &DCI) {
9399 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9402 EVT VT = N->getValueType(0);
9406 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9409 uint64_t MulAmt = C->getZExtValue();
9410 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9413 uint64_t MulAmt1 = 0;
9414 uint64_t MulAmt2 = 0;
9415 if ((MulAmt % 9) == 0) {
9417 MulAmt2 = MulAmt / 9;
9418 } else if ((MulAmt % 5) == 0) {
9420 MulAmt2 = MulAmt / 5;
9421 } else if ((MulAmt % 3) == 0) {
9423 MulAmt2 = MulAmt / 3;
9426 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9427 DebugLoc DL = N->getDebugLoc();
9429 if (isPowerOf2_64(MulAmt2) &&
9430 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9431 // If second multiplifer is pow2, issue it first. We want the multiply by
9432 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9434 std::swap(MulAmt1, MulAmt2);
9437 if (isPowerOf2_64(MulAmt1))
9438 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9439 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9441 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9442 DAG.getConstant(MulAmt1, VT));
9444 if (isPowerOf2_64(MulAmt2))
9445 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9446 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9448 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9449 DAG.getConstant(MulAmt2, VT));
9451 // Do not add new nodes to DAG combiner worklist.
9452 DCI.CombineTo(N, NewMul, false);
9457 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9458 SDValue N0 = N->getOperand(0);
9459 SDValue N1 = N->getOperand(1);
9460 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9461 EVT VT = N0.getValueType();
9463 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9464 // since the result of setcc_c is all zero's or all ones.
9465 if (N1C && N0.getOpcode() == ISD::AND &&
9466 N0.getOperand(1).getOpcode() == ISD::Constant) {
9467 SDValue N00 = N0.getOperand(0);
9468 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9469 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9470 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9471 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9472 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9473 APInt ShAmt = N1C->getAPIntValue();
9474 Mask = Mask.shl(ShAmt);
9476 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9477 N00, DAG.getConstant(Mask, VT));
9484 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9486 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9487 const X86Subtarget *Subtarget) {
9488 EVT VT = N->getValueType(0);
9489 if (!VT.isVector() && VT.isInteger() &&
9490 N->getOpcode() == ISD::SHL)
9491 return PerformSHLCombine(N, DAG);
9493 // On X86 with SSE2 support, we can transform this to a vector shift if
9494 // all elements are shifted by the same amount. We can't do this in legalize
9495 // because the a constant vector is typically transformed to a constant pool
9496 // so we have no knowledge of the shift amount.
9497 if (!Subtarget->hasSSE2())
9500 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9503 SDValue ShAmtOp = N->getOperand(1);
9504 EVT EltVT = VT.getVectorElementType();
9505 DebugLoc DL = N->getDebugLoc();
9506 SDValue BaseShAmt = SDValue();
9507 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9508 unsigned NumElts = VT.getVectorNumElements();
9510 for (; i != NumElts; ++i) {
9511 SDValue Arg = ShAmtOp.getOperand(i);
9512 if (Arg.getOpcode() == ISD::UNDEF) continue;
9516 for (; i != NumElts; ++i) {
9517 SDValue Arg = ShAmtOp.getOperand(i);
9518 if (Arg.getOpcode() == ISD::UNDEF) continue;
9519 if (Arg != BaseShAmt) {
9523 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9524 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9525 SDValue InVec = ShAmtOp.getOperand(0);
9526 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9527 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9529 for (; i != NumElts; ++i) {
9530 SDValue Arg = InVec.getOperand(i);
9531 if (Arg.getOpcode() == ISD::UNDEF) continue;
9535 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9536 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9537 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9538 if (C->getZExtValue() == SplatIdx)
9539 BaseShAmt = InVec.getOperand(1);
9542 if (BaseShAmt.getNode() == 0)
9543 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9544 DAG.getIntPtrConstant(0));
9548 // The shift amount is an i32.
9549 if (EltVT.bitsGT(MVT::i32))
9550 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9551 else if (EltVT.bitsLT(MVT::i32))
9552 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9554 // The shift amount is identical so we can do a vector shift.
9555 SDValue ValOp = N->getOperand(0);
9556 switch (N->getOpcode()) {
9558 llvm_unreachable("Unknown shift opcode!");
9561 if (VT == MVT::v2i64)
9562 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9563 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9565 if (VT == MVT::v4i32)
9566 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9567 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9569 if (VT == MVT::v8i16)
9570 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9571 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9575 if (VT == MVT::v4i32)
9576 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9577 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9579 if (VT == MVT::v8i16)
9580 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9581 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9585 if (VT == MVT::v2i64)
9586 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9587 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9589 if (VT == MVT::v4i32)
9590 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9591 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9593 if (VT == MVT::v8i16)
9594 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9595 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9602 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9603 TargetLowering::DAGCombinerInfo &DCI,
9604 const X86Subtarget *Subtarget) {
9605 if (DCI.isBeforeLegalizeOps())
9608 EVT VT = N->getValueType(0);
9609 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9612 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9613 SDValue N0 = N->getOperand(0);
9614 SDValue N1 = N->getOperand(1);
9615 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9617 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9619 if (!N0.hasOneUse() || !N1.hasOneUse())
9622 SDValue ShAmt0 = N0.getOperand(1);
9623 if (ShAmt0.getValueType() != MVT::i8)
9625 SDValue ShAmt1 = N1.getOperand(1);
9626 if (ShAmt1.getValueType() != MVT::i8)
9628 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9629 ShAmt0 = ShAmt0.getOperand(0);
9630 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9631 ShAmt1 = ShAmt1.getOperand(0);
9633 DebugLoc DL = N->getDebugLoc();
9634 unsigned Opc = X86ISD::SHLD;
9635 SDValue Op0 = N0.getOperand(0);
9636 SDValue Op1 = N1.getOperand(0);
9637 if (ShAmt0.getOpcode() == ISD::SUB) {
9639 std::swap(Op0, Op1);
9640 std::swap(ShAmt0, ShAmt1);
9643 unsigned Bits = VT.getSizeInBits();
9644 if (ShAmt1.getOpcode() == ISD::SUB) {
9645 SDValue Sum = ShAmt1.getOperand(0);
9646 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9647 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
9648 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
9649 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
9650 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
9651 return DAG.getNode(Opc, DL, VT,
9653 DAG.getNode(ISD::TRUNCATE, DL,
9656 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9657 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9659 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9660 return DAG.getNode(Opc, DL, VT,
9661 N0.getOperand(0), N1.getOperand(0),
9662 DAG.getNode(ISD::TRUNCATE, DL,
9669 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9670 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9671 const X86Subtarget *Subtarget) {
9672 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9673 // the FP state in cases where an emms may be missing.
9674 // A preferable solution to the general problem is to figure out the right
9675 // places to insert EMMS. This qualifies as a quick hack.
9677 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9678 StoreSDNode *St = cast<StoreSDNode>(N);
9679 EVT VT = St->getValue().getValueType();
9680 if (VT.getSizeInBits() != 64)
9683 const Function *F = DAG.getMachineFunction().getFunction();
9684 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9685 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9686 && Subtarget->hasSSE2();
9687 if ((VT.isVector() ||
9688 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9689 isa<LoadSDNode>(St->getValue()) &&
9690 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9691 St->getChain().hasOneUse() && !St->isVolatile()) {
9692 SDNode* LdVal = St->getValue().getNode();
9694 int TokenFactorIndex = -1;
9695 SmallVector<SDValue, 8> Ops;
9696 SDNode* ChainVal = St->getChain().getNode();
9697 // Must be a store of a load. We currently handle two cases: the load
9698 // is a direct child, and it's under an intervening TokenFactor. It is
9699 // possible to dig deeper under nested TokenFactors.
9700 if (ChainVal == LdVal)
9701 Ld = cast<LoadSDNode>(St->getChain());
9702 else if (St->getValue().hasOneUse() &&
9703 ChainVal->getOpcode() == ISD::TokenFactor) {
9704 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9705 if (ChainVal->getOperand(i).getNode() == LdVal) {
9706 TokenFactorIndex = i;
9707 Ld = cast<LoadSDNode>(St->getValue());
9709 Ops.push_back(ChainVal->getOperand(i));
9713 if (!Ld || !ISD::isNormalLoad(Ld))
9716 // If this is not the MMX case, i.e. we are just turning i64 load/store
9717 // into f64 load/store, avoid the transformation if there are multiple
9718 // uses of the loaded value.
9719 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9722 DebugLoc LdDL = Ld->getDebugLoc();
9723 DebugLoc StDL = N->getDebugLoc();
9724 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9725 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9727 if (Subtarget->is64Bit() || F64IsLegal) {
9728 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9729 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9730 Ld->getBasePtr(), Ld->getSrcValue(),
9731 Ld->getSrcValueOffset(), Ld->isVolatile(),
9732 Ld->isNonTemporal(), Ld->getAlignment());
9733 SDValue NewChain = NewLd.getValue(1);
9734 if (TokenFactorIndex != -1) {
9735 Ops.push_back(NewChain);
9736 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9739 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9740 St->getSrcValue(), St->getSrcValueOffset(),
9741 St->isVolatile(), St->isNonTemporal(),
9742 St->getAlignment());
9745 // Otherwise, lower to two pairs of 32-bit loads / stores.
9746 SDValue LoAddr = Ld->getBasePtr();
9747 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9748 DAG.getConstant(4, MVT::i32));
9750 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9751 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9752 Ld->isVolatile(), Ld->isNonTemporal(),
9753 Ld->getAlignment());
9754 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9755 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9756 Ld->isVolatile(), Ld->isNonTemporal(),
9757 MinAlign(Ld->getAlignment(), 4));
9759 SDValue NewChain = LoLd.getValue(1);
9760 if (TokenFactorIndex != -1) {
9761 Ops.push_back(LoLd);
9762 Ops.push_back(HiLd);
9763 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9767 LoAddr = St->getBasePtr();
9768 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9769 DAG.getConstant(4, MVT::i32));
9771 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9772 St->getSrcValue(), St->getSrcValueOffset(),
9773 St->isVolatile(), St->isNonTemporal(),
9774 St->getAlignment());
9775 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9777 St->getSrcValueOffset() + 4,
9779 St->isNonTemporal(),
9780 MinAlign(St->getAlignment(), 4));
9781 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9786 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9787 /// X86ISD::FXOR nodes.
9788 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9789 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9790 // F[X]OR(0.0, x) -> x
9791 // F[X]OR(x, 0.0) -> x
9792 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9793 if (C->getValueAPF().isPosZero())
9794 return N->getOperand(1);
9795 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9796 if (C->getValueAPF().isPosZero())
9797 return N->getOperand(0);
9801 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9802 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9803 // FAND(0.0, x) -> 0.0
9804 // FAND(x, 0.0) -> 0.0
9805 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9806 if (C->getValueAPF().isPosZero())
9807 return N->getOperand(0);
9808 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9809 if (C->getValueAPF().isPosZero())
9810 return N->getOperand(1);
9814 static SDValue PerformBTCombine(SDNode *N,
9816 TargetLowering::DAGCombinerInfo &DCI) {
9817 // BT ignores high bits in the bit index operand.
9818 SDValue Op1 = N->getOperand(1);
9819 if (Op1.hasOneUse()) {
9820 unsigned BitWidth = Op1.getValueSizeInBits();
9821 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9822 APInt KnownZero, KnownOne;
9823 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9824 !DCI.isBeforeLegalizeOps());
9825 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9826 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9827 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9828 DCI.CommitTargetLoweringOpt(TLO);
9833 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9834 SDValue Op = N->getOperand(0);
9835 if (Op.getOpcode() == ISD::BIT_CONVERT)
9836 Op = Op.getOperand(0);
9837 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9838 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9839 VT.getVectorElementType().getSizeInBits() ==
9840 OpVT.getVectorElementType().getSizeInBits()) {
9841 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9846 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9847 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9848 // (and (i32 x86isd::setcc_carry), 1)
9849 // This eliminates the zext. This transformation is necessary because
9850 // ISD::SETCC is always legalized to i8.
9851 DebugLoc dl = N->getDebugLoc();
9852 SDValue N0 = N->getOperand(0);
9853 EVT VT = N->getValueType(0);
9854 if (N0.getOpcode() == ISD::AND &&
9856 N0.getOperand(0).hasOneUse()) {
9857 SDValue N00 = N0.getOperand(0);
9858 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9860 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9861 if (!C || C->getZExtValue() != 1)
9863 return DAG.getNode(ISD::AND, dl, VT,
9864 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9865 N00.getOperand(0), N00.getOperand(1)),
9866 DAG.getConstant(1, VT));
9872 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9873 DAGCombinerInfo &DCI) const {
9874 SelectionDAG &DAG = DCI.DAG;
9875 switch (N->getOpcode()) {
9877 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9878 case ISD::EXTRACT_VECTOR_ELT:
9879 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9880 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9881 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9882 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9885 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9886 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
9887 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9889 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9890 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9891 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9892 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9893 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9899 /// isTypeDesirableForOp - Return true if the target has native support for
9900 /// the specified value type and it is 'desirable' to use the type for the
9901 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9902 /// instruction encodings are longer and some i16 instructions are slow.
9903 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9904 if (!isTypeLegal(VT))
9913 case ISD::SIGN_EXTEND:
9914 case ISD::ZERO_EXTEND:
9915 case ISD::ANY_EXTEND:
9928 static bool MayFoldLoad(SDValue Op) {
9929 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
9932 static bool MayFoldIntoStore(SDValue Op) {
9933 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
9936 /// IsDesirableToPromoteOp - This method query the target whether it is
9937 /// beneficial for dag combiner to promote the specified node. If true, it
9938 /// should return the desired promotion type by reference.
9939 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
9940 EVT VT = Op.getValueType();
9944 bool Promote = false;
9945 bool Commute = false;
9946 switch (Op.getOpcode()) {
9949 LoadSDNode *LD = cast<LoadSDNode>(Op);
9950 // If the non-extending load has a single use and it's not live out, then it
9952 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
9954 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9955 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
9956 // The only case where we'd want to promote LOAD (rather then it being
9957 // promoted as an operand is when it's only use is liveout.
9958 if (UI->getOpcode() != ISD::CopyToReg)
9965 case ISD::SIGN_EXTEND:
9966 case ISD::ZERO_EXTEND:
9967 case ISD::ANY_EXTEND:
9972 SDValue N0 = Op.getOperand(0);
9973 // Look out for (store (shl (load), x)).
9974 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
9987 SDValue N0 = Op.getOperand(0);
9988 SDValue N1 = Op.getOperand(1);
9989 if (!Commute && MayFoldLoad(N1))
9991 // Avoid disabling potential load folding opportunities.
9992 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
9994 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10004 //===----------------------------------------------------------------------===//
10005 // X86 Inline Assembly Support
10006 //===----------------------------------------------------------------------===//
10008 static bool LowerToBSwap(CallInst *CI) {
10009 // FIXME: this should verify that we are targetting a 486 or better. If not,
10010 // we will turn this bswap into something that will be lowered to logical ops
10011 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10012 // so don't worry about this.
10014 // Verify this is a simple bswap.
10015 if (CI->getNumArgOperands() != 1 ||
10016 CI->getType() != CI->getArgOperand(0)->getType() ||
10017 !CI->getType()->isIntegerTy())
10020 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10021 if (!Ty || Ty->getBitWidth() % 16 != 0)
10024 // Okay, we can do this xform, do so now.
10025 const Type *Tys[] = { Ty };
10026 Module *M = CI->getParent()->getParent()->getParent();
10027 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10029 Value *Op = CI->getArgOperand(0);
10030 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10032 CI->replaceAllUsesWith(Op);
10033 CI->eraseFromParent();
10037 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10038 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10039 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10041 std::string AsmStr = IA->getAsmString();
10043 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10044 SmallVector<StringRef, 4> AsmPieces;
10045 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10047 switch (AsmPieces.size()) {
10048 default: return false;
10050 AsmStr = AsmPieces[0];
10052 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10055 if (AsmPieces.size() == 2 &&
10056 (AsmPieces[0] == "bswap" ||
10057 AsmPieces[0] == "bswapq" ||
10058 AsmPieces[0] == "bswapl") &&
10059 (AsmPieces[1] == "$0" ||
10060 AsmPieces[1] == "${0:q}")) {
10061 // No need to check constraints, nothing other than the equivalent of
10062 // "=r,0" would be valid here.
10063 return LowerToBSwap(CI);
10065 // rorw $$8, ${0:w} --> llvm.bswap.i16
10066 if (CI->getType()->isIntegerTy(16) &&
10067 AsmPieces.size() == 3 &&
10068 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10069 AsmPieces[1] == "$$8," &&
10070 AsmPieces[2] == "${0:w}" &&
10071 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10073 const std::string &Constraints = IA->getConstraintString();
10074 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10075 std::sort(AsmPieces.begin(), AsmPieces.end());
10076 if (AsmPieces.size() == 4 &&
10077 AsmPieces[0] == "~{cc}" &&
10078 AsmPieces[1] == "~{dirflag}" &&
10079 AsmPieces[2] == "~{flags}" &&
10080 AsmPieces[3] == "~{fpsr}") {
10081 return LowerToBSwap(CI);
10086 if (CI->getType()->isIntegerTy(64) &&
10087 Constraints.size() >= 2 &&
10088 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10089 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10090 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10091 SmallVector<StringRef, 4> Words;
10092 SplitString(AsmPieces[0], Words, " \t");
10093 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10095 SplitString(AsmPieces[1], Words, " \t");
10096 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10098 SplitString(AsmPieces[2], Words, " \t,");
10099 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10100 Words[2] == "%edx") {
10101 return LowerToBSwap(CI);
10113 /// getConstraintType - Given a constraint letter, return the type of
10114 /// constraint it is for this target.
10115 X86TargetLowering::ConstraintType
10116 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10117 if (Constraint.size() == 1) {
10118 switch (Constraint[0]) {
10130 return C_RegisterClass;
10138 return TargetLowering::getConstraintType(Constraint);
10141 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10142 /// with another that has more specific requirements based on the type of the
10143 /// corresponding operand.
10144 const char *X86TargetLowering::
10145 LowerXConstraint(EVT ConstraintVT) const {
10146 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10147 // 'f' like normal targets.
10148 if (ConstraintVT.isFloatingPoint()) {
10149 if (Subtarget->hasSSE2())
10151 if (Subtarget->hasSSE1())
10155 return TargetLowering::LowerXConstraint(ConstraintVT);
10158 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10159 /// vector. If it is invalid, don't add anything to Ops.
10160 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10162 std::vector<SDValue>&Ops,
10163 SelectionDAG &DAG) const {
10164 SDValue Result(0, 0);
10166 switch (Constraint) {
10169 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10170 if (C->getZExtValue() <= 31) {
10171 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10177 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10178 if (C->getZExtValue() <= 63) {
10179 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10185 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10186 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10187 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10193 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10194 if (C->getZExtValue() <= 255) {
10195 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10201 // 32-bit signed value
10202 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10203 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10204 C->getSExtValue())) {
10205 // Widen to 64 bits here to get it sign extended.
10206 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10209 // FIXME gcc accepts some relocatable values here too, but only in certain
10210 // memory models; it's complicated.
10215 // 32-bit unsigned value
10216 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10217 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10218 C->getZExtValue())) {
10219 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10223 // FIXME gcc accepts some relocatable values here too, but only in certain
10224 // memory models; it's complicated.
10228 // Literal immediates are always ok.
10229 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10230 // Widen to 64 bits here to get it sign extended.
10231 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10235 // In any sort of PIC mode addresses need to be computed at runtime by
10236 // adding in a register or some sort of table lookup. These can't
10237 // be used as immediates.
10238 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC() ||
10239 Subtarget->isPICStyleRIPRel())
10242 // If we are in non-pic codegen mode, we allow the address of a global (with
10243 // an optional displacement) to be used with 'i'.
10244 GlobalAddressSDNode *GA = 0;
10245 int64_t Offset = 0;
10247 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10249 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10250 Offset += GA->getOffset();
10252 } else if (Op.getOpcode() == ISD::ADD) {
10253 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10254 Offset += C->getZExtValue();
10255 Op = Op.getOperand(0);
10258 } else if (Op.getOpcode() == ISD::SUB) {
10259 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10260 Offset += -C->getZExtValue();
10261 Op = Op.getOperand(0);
10266 // Otherwise, this isn't something we can handle, reject it.
10270 const GlobalValue *GV = GA->getGlobal();
10271 // If we require an extra load to get this address, as in PIC mode, we
10272 // can't accept it.
10273 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10274 getTargetMachine())))
10277 Result = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10282 if (Result.getNode()) {
10283 Ops.push_back(Result);
10286 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10289 std::vector<unsigned> X86TargetLowering::
10290 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10292 if (Constraint.size() == 1) {
10293 // FIXME: not handling fp-stack yet!
10294 switch (Constraint[0]) { // GCC X86 Constraint Letters
10295 default: break; // Unknown constraint letter
10296 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10297 if (Subtarget->is64Bit()) {
10298 if (VT == MVT::i32)
10299 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10300 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10301 X86::R10D,X86::R11D,X86::R12D,
10302 X86::R13D,X86::R14D,X86::R15D,
10303 X86::EBP, X86::ESP, 0);
10304 else if (VT == MVT::i16)
10305 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10306 X86::SI, X86::DI, X86::R8W,X86::R9W,
10307 X86::R10W,X86::R11W,X86::R12W,
10308 X86::R13W,X86::R14W,X86::R15W,
10309 X86::BP, X86::SP, 0);
10310 else if (VT == MVT::i8)
10311 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10312 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10313 X86::R10B,X86::R11B,X86::R12B,
10314 X86::R13B,X86::R14B,X86::R15B,
10315 X86::BPL, X86::SPL, 0);
10317 else if (VT == MVT::i64)
10318 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10319 X86::RSI, X86::RDI, X86::R8, X86::R9,
10320 X86::R10, X86::R11, X86::R12,
10321 X86::R13, X86::R14, X86::R15,
10322 X86::RBP, X86::RSP, 0);
10326 // 32-bit fallthrough
10327 case 'Q': // Q_REGS
10328 if (VT == MVT::i32)
10329 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10330 else if (VT == MVT::i16)
10331 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10332 else if (VT == MVT::i8)
10333 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10334 else if (VT == MVT::i64)
10335 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10340 return std::vector<unsigned>();
10343 std::pair<unsigned, const TargetRegisterClass*>
10344 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10346 // First, see if this is a constraint that directly corresponds to an LLVM
10348 if (Constraint.size() == 1) {
10349 // GCC Constraint Letters
10350 switch (Constraint[0]) {
10352 case 'r': // GENERAL_REGS
10353 case 'l': // INDEX_REGS
10355 return std::make_pair(0U, X86::GR8RegisterClass);
10356 if (VT == MVT::i16)
10357 return std::make_pair(0U, X86::GR16RegisterClass);
10358 if (VT == MVT::i32 || !Subtarget->is64Bit())
10359 return std::make_pair(0U, X86::GR32RegisterClass);
10360 return std::make_pair(0U, X86::GR64RegisterClass);
10361 case 'R': // LEGACY_REGS
10363 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10364 if (VT == MVT::i16)
10365 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10366 if (VT == MVT::i32 || !Subtarget->is64Bit())
10367 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10368 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10369 case 'f': // FP Stack registers.
10370 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10371 // value to the correct fpstack register class.
10372 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10373 return std::make_pair(0U, X86::RFP32RegisterClass);
10374 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10375 return std::make_pair(0U, X86::RFP64RegisterClass);
10376 return std::make_pair(0U, X86::RFP80RegisterClass);
10377 case 'y': // MMX_REGS if MMX allowed.
10378 if (!Subtarget->hasMMX()) break;
10379 return std::make_pair(0U, X86::VR64RegisterClass);
10380 case 'Y': // SSE_REGS if SSE2 allowed
10381 if (!Subtarget->hasSSE2()) break;
10383 case 'x': // SSE_REGS if SSE1 allowed
10384 if (!Subtarget->hasSSE1()) break;
10386 switch (VT.getSimpleVT().SimpleTy) {
10388 // Scalar SSE types.
10391 return std::make_pair(0U, X86::FR32RegisterClass);
10394 return std::make_pair(0U, X86::FR64RegisterClass);
10402 return std::make_pair(0U, X86::VR128RegisterClass);
10408 // Use the default implementation in TargetLowering to convert the register
10409 // constraint into a member of a register class.
10410 std::pair<unsigned, const TargetRegisterClass*> Res;
10411 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10413 // Not found as a standard register?
10414 if (Res.second == 0) {
10415 // Map st(0) -> st(7) -> ST0
10416 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10417 tolower(Constraint[1]) == 's' &&
10418 tolower(Constraint[2]) == 't' &&
10419 Constraint[3] == '(' &&
10420 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10421 Constraint[5] == ')' &&
10422 Constraint[6] == '}') {
10424 Res.first = X86::ST0+Constraint[4]-'0';
10425 Res.second = X86::RFP80RegisterClass;
10429 // GCC allows "st(0)" to be called just plain "st".
10430 if (StringRef("{st}").equals_lower(Constraint)) {
10431 Res.first = X86::ST0;
10432 Res.second = X86::RFP80RegisterClass;
10437 if (StringRef("{flags}").equals_lower(Constraint)) {
10438 Res.first = X86::EFLAGS;
10439 Res.second = X86::CCRRegisterClass;
10443 // 'A' means EAX + EDX.
10444 if (Constraint == "A") {
10445 Res.first = X86::EAX;
10446 Res.second = X86::GR32_ADRegisterClass;
10452 // Otherwise, check to see if this is a register class of the wrong value
10453 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10454 // turn into {ax},{dx}.
10455 if (Res.second->hasType(VT))
10456 return Res; // Correct type already, nothing to do.
10458 // All of the single-register GCC register classes map their values onto
10459 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10460 // really want an 8-bit or 32-bit register, map to the appropriate register
10461 // class and return the appropriate register.
10462 if (Res.second == X86::GR16RegisterClass) {
10463 if (VT == MVT::i8) {
10464 unsigned DestReg = 0;
10465 switch (Res.first) {
10467 case X86::AX: DestReg = X86::AL; break;
10468 case X86::DX: DestReg = X86::DL; break;
10469 case X86::CX: DestReg = X86::CL; break;
10470 case X86::BX: DestReg = X86::BL; break;
10473 Res.first = DestReg;
10474 Res.second = X86::GR8RegisterClass;
10476 } else if (VT == MVT::i32) {
10477 unsigned DestReg = 0;
10478 switch (Res.first) {
10480 case X86::AX: DestReg = X86::EAX; break;
10481 case X86::DX: DestReg = X86::EDX; break;
10482 case X86::CX: DestReg = X86::ECX; break;
10483 case X86::BX: DestReg = X86::EBX; break;
10484 case X86::SI: DestReg = X86::ESI; break;
10485 case X86::DI: DestReg = X86::EDI; break;
10486 case X86::BP: DestReg = X86::EBP; break;
10487 case X86::SP: DestReg = X86::ESP; break;
10490 Res.first = DestReg;
10491 Res.second = X86::GR32RegisterClass;
10493 } else if (VT == MVT::i64) {
10494 unsigned DestReg = 0;
10495 switch (Res.first) {
10497 case X86::AX: DestReg = X86::RAX; break;
10498 case X86::DX: DestReg = X86::RDX; break;
10499 case X86::CX: DestReg = X86::RCX; break;
10500 case X86::BX: DestReg = X86::RBX; break;
10501 case X86::SI: DestReg = X86::RSI; break;
10502 case X86::DI: DestReg = X86::RDI; break;
10503 case X86::BP: DestReg = X86::RBP; break;
10504 case X86::SP: DestReg = X86::RSP; break;
10507 Res.first = DestReg;
10508 Res.second = X86::GR64RegisterClass;
10511 } else if (Res.second == X86::FR32RegisterClass ||
10512 Res.second == X86::FR64RegisterClass ||
10513 Res.second == X86::VR128RegisterClass) {
10514 // Handle references to XMM physical registers that got mapped into the
10515 // wrong class. This can happen with constraints like {xmm0} where the
10516 // target independent register mapper will just pick the first match it can
10517 // find, ignoring the required type.
10518 if (VT == MVT::f32)
10519 Res.second = X86::FR32RegisterClass;
10520 else if (VT == MVT::f64)
10521 Res.second = X86::FR64RegisterClass;
10522 else if (X86::VR128RegisterClass->hasType(VT))
10523 Res.second = X86::VR128RegisterClass;