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 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/ADT/BitVector.h"
27 #include "llvm/ADT/VectorExtras.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/PseudoSourceValue.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Target/TargetOptions.h"
39 #include "llvm/ADT/SmallSet.h"
40 #include "llvm/ADT/StringExtras.h"
41 #include "llvm/Support/CommandLine.h"
45 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
47 // Forward declarations.
48 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
51 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
52 : TargetLowering(TM) {
53 Subtarget = &TM.getSubtarget<X86Subtarget>();
54 X86ScalarSSEf64 = Subtarget->hasSSE2();
55 X86ScalarSSEf32 = Subtarget->hasSSE1();
56 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
60 RegInfo = TM.getRegisterInfo();
63 // Set up the TargetLowering object.
65 // X86 is weird, it always uses i8 for shift amounts and setcc results.
66 setShiftAmountType(MVT::i8);
67 setBooleanContents(ZeroOrOneBooleanContent);
68 setSchedulingPreference(SchedulingForRegPressure);
69 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
70 setStackPointerRegisterToSaveRestore(X86StackPtr);
72 if (Subtarget->isTargetDarwin()) {
73 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
74 setUseUnderscoreSetJmp(false);
75 setUseUnderscoreLongJmp(false);
76 } else if (Subtarget->isTargetMingw()) {
77 // MS runtime is weird: it exports _setjmp, but longjmp!
78 setUseUnderscoreSetJmp(true);
79 setUseUnderscoreLongJmp(false);
81 setUseUnderscoreSetJmp(true);
82 setUseUnderscoreLongJmp(true);
85 // Set up the register classes.
86 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
87 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
88 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
89 if (Subtarget->is64Bit())
90 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
92 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
94 // We don't accept any truncstore of integer registers.
95 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
96 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
97 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
98 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
99 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
100 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
102 // SETOEQ and SETUNE require checking two conditions.
103 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
104 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
105 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
106 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
107 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
108 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
110 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
112 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
113 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
114 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
116 if (Subtarget->is64Bit()) {
117 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
118 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
120 if (!UseSoftFloat && !NoImplicitFloat && X86ScalarSSEf64) {
121 // We have an impenetrably clever algorithm for ui64->double only.
122 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
124 // We have faster algorithm for ui32->single only.
125 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
127 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
131 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
133 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
134 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
136 if (!UseSoftFloat && !NoImplicitFloat) {
137 // SSE has no i16 to fp conversion, only i32
138 if (X86ScalarSSEf32) {
139 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
140 // f32 and f64 cases are Legal, f80 case is not
141 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
143 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
144 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
147 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
148 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
151 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
152 // are Legal, f80 is custom lowered.
153 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
154 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
156 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
158 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
159 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
161 if (X86ScalarSSEf32) {
162 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
163 // f32 and f64 cases are Legal, f80 case is not
164 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
166 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
167 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
170 // Handle FP_TO_UINT by promoting the destination to a larger signed
172 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
173 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
174 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
176 if (Subtarget->is64Bit()) {
177 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
178 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
180 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
181 // Expand FP_TO_UINT into a select.
182 // FIXME: We would like to use a Custom expander here eventually to do
183 // the optimal thing for SSE vs. the default expansion in the legalizer.
184 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
186 // With SSE3 we can use fisttpll to convert to a signed i64.
187 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
190 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
191 if (!X86ScalarSSEf64) {
192 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
193 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
196 // Scalar integer divide and remainder are lowered to use operations that
197 // produce two results, to match the available instructions. This exposes
198 // the two-result form to trivial CSE, which is able to combine x/y and x%y
199 // into a single instruction.
201 // Scalar integer multiply-high is also lowered to use two-result
202 // operations, to match the available instructions. However, plain multiply
203 // (low) operations are left as Legal, as there are single-result
204 // instructions for this in x86. Using the two-result multiply instructions
205 // when both high and low results are needed must be arranged by dagcombine.
206 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
207 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
208 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
209 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
210 setOperationAction(ISD::SREM , MVT::i8 , Expand);
211 setOperationAction(ISD::UREM , MVT::i8 , Expand);
212 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
213 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
214 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
215 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
216 setOperationAction(ISD::SREM , MVT::i16 , Expand);
217 setOperationAction(ISD::UREM , MVT::i16 , Expand);
218 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
219 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
220 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
221 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
222 setOperationAction(ISD::SREM , MVT::i32 , Expand);
223 setOperationAction(ISD::UREM , MVT::i32 , Expand);
224 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
225 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
226 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
227 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
228 setOperationAction(ISD::SREM , MVT::i64 , Expand);
229 setOperationAction(ISD::UREM , MVT::i64 , Expand);
231 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
232 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
233 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
234 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
235 if (Subtarget->is64Bit())
236 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
237 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
238 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
239 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
240 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
241 setOperationAction(ISD::FREM , MVT::f32 , Expand);
242 setOperationAction(ISD::FREM , MVT::f64 , Expand);
243 setOperationAction(ISD::FREM , MVT::f80 , Expand);
244 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
246 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
247 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
248 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
249 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
250 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
251 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
252 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
253 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
254 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
255 if (Subtarget->is64Bit()) {
256 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
257 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
258 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
261 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
262 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
264 // These should be promoted to a larger select which is supported.
265 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
266 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
267 // X86 wants to expand cmov itself.
268 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
269 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
270 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
271 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
272 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
273 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
274 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
275 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
276 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
277 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
278 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
279 if (Subtarget->is64Bit()) {
280 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
281 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
283 // X86 ret instruction may pop stack.
284 setOperationAction(ISD::RET , MVT::Other, Custom);
285 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
288 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
289 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
290 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
291 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
292 if (Subtarget->is64Bit())
293 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
294 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
295 if (Subtarget->is64Bit()) {
296 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
297 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
298 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
299 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
301 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
302 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
303 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
304 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
305 if (Subtarget->is64Bit()) {
306 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
307 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
308 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
311 if (Subtarget->hasSSE1())
312 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
314 if (!Subtarget->hasSSE2())
315 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
317 // Expand certain atomics
318 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
319 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
320 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
321 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
323 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
324 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
325 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
326 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
328 if (!Subtarget->is64Bit()) {
329 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
330 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
331 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
332 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
333 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
334 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
335 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
338 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
339 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
340 // FIXME - use subtarget debug flags
341 if (!Subtarget->isTargetDarwin() &&
342 !Subtarget->isTargetELF() &&
343 !Subtarget->isTargetCygMing()) {
344 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
345 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
348 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
349 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
350 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
351 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
352 if (Subtarget->is64Bit()) {
353 setExceptionPointerRegister(X86::RAX);
354 setExceptionSelectorRegister(X86::RDX);
356 setExceptionPointerRegister(X86::EAX);
357 setExceptionSelectorRegister(X86::EDX);
359 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
360 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
362 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
364 setOperationAction(ISD::TRAP, MVT::Other, Legal);
366 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
367 setOperationAction(ISD::VASTART , MVT::Other, Custom);
368 setOperationAction(ISD::VAEND , MVT::Other, Expand);
369 if (Subtarget->is64Bit()) {
370 setOperationAction(ISD::VAARG , MVT::Other, Custom);
371 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
373 setOperationAction(ISD::VAARG , MVT::Other, Expand);
374 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
377 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
378 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
379 if (Subtarget->is64Bit())
380 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
381 if (Subtarget->isTargetCygMing())
382 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
384 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
386 if (!UseSoftFloat && X86ScalarSSEf64) {
387 // f32 and f64 use SSE.
388 // Set up the FP register classes.
389 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
390 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
392 // Use ANDPD to simulate FABS.
393 setOperationAction(ISD::FABS , MVT::f64, Custom);
394 setOperationAction(ISD::FABS , MVT::f32, Custom);
396 // Use XORP to simulate FNEG.
397 setOperationAction(ISD::FNEG , MVT::f64, Custom);
398 setOperationAction(ISD::FNEG , MVT::f32, Custom);
400 // Use ANDPD and ORPD to simulate FCOPYSIGN.
401 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
402 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
404 // We don't support sin/cos/fmod
405 setOperationAction(ISD::FSIN , MVT::f64, Expand);
406 setOperationAction(ISD::FCOS , MVT::f64, Expand);
407 setOperationAction(ISD::FSIN , MVT::f32, Expand);
408 setOperationAction(ISD::FCOS , MVT::f32, Expand);
410 // Expand FP immediates into loads from the stack, except for the special
412 addLegalFPImmediate(APFloat(+0.0)); // xorpd
413 addLegalFPImmediate(APFloat(+0.0f)); // xorps
415 // Floating truncations from f80 and extensions to f80 go through memory.
416 // If optimizing, we lie about this though and handle it in
417 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
419 setConvertAction(MVT::f32, MVT::f80, Expand);
420 setConvertAction(MVT::f64, MVT::f80, Expand);
421 setConvertAction(MVT::f80, MVT::f32, Expand);
422 setConvertAction(MVT::f80, MVT::f64, Expand);
424 } else if (!UseSoftFloat && X86ScalarSSEf32) {
425 // Use SSE for f32, x87 for f64.
426 // Set up the FP register classes.
427 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
428 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
430 // Use ANDPS to simulate FABS.
431 setOperationAction(ISD::FABS , MVT::f32, Custom);
433 // Use XORP to simulate FNEG.
434 setOperationAction(ISD::FNEG , MVT::f32, Custom);
436 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
438 // Use ANDPS and ORPS to simulate FCOPYSIGN.
439 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
440 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
442 // We don't support sin/cos/fmod
443 setOperationAction(ISD::FSIN , MVT::f32, Expand);
444 setOperationAction(ISD::FCOS , MVT::f32, Expand);
446 // Special cases we handle for FP constants.
447 addLegalFPImmediate(APFloat(+0.0f)); // xorps
448 addLegalFPImmediate(APFloat(+0.0)); // FLD0
449 addLegalFPImmediate(APFloat(+1.0)); // FLD1
450 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
451 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
453 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
454 // this though and handle it in InstructionSelectPreprocess so that
455 // dagcombine2 can hack on these.
457 setConvertAction(MVT::f32, MVT::f64, Expand);
458 setConvertAction(MVT::f32, MVT::f80, Expand);
459 setConvertAction(MVT::f80, MVT::f32, Expand);
460 setConvertAction(MVT::f64, MVT::f32, Expand);
461 // And x87->x87 truncations also.
462 setConvertAction(MVT::f80, MVT::f64, Expand);
466 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
467 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
469 } else if (!UseSoftFloat) {
470 // f32 and f64 in x87.
471 // Set up the FP register classes.
472 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
473 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
475 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
476 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
477 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
478 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
480 // Floating truncations go through memory. If optimizing, we lie about
481 // this though and handle it in InstructionSelectPreprocess so that
482 // dagcombine2 can hack on these.
484 setConvertAction(MVT::f80, MVT::f32, Expand);
485 setConvertAction(MVT::f64, MVT::f32, Expand);
486 setConvertAction(MVT::f80, MVT::f64, Expand);
490 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
491 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
493 addLegalFPImmediate(APFloat(+0.0)); // FLD0
494 addLegalFPImmediate(APFloat(+1.0)); // FLD1
495 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
496 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
497 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
498 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
499 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
500 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
503 // Long double always uses X87.
505 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
506 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
507 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
510 APFloat TmpFlt(+0.0);
511 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
513 addLegalFPImmediate(TmpFlt); // FLD0
515 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
516 APFloat TmpFlt2(+1.0);
517 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
519 addLegalFPImmediate(TmpFlt2); // FLD1
520 TmpFlt2.changeSign();
521 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
525 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
526 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
530 // Always use a library call for pow.
531 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
532 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
533 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
535 setOperationAction(ISD::FLOG, MVT::f80, Expand);
536 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
537 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
538 setOperationAction(ISD::FEXP, MVT::f80, Expand);
539 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
541 // First set operation action for all vector types to either promote
542 // (for widening) or expand (for scalarization). Then we will selectively
543 // turn on ones that can be effectively codegen'd.
544 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
545 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
546 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
561 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
591 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
592 // with -msoft-float, disable use of MMX as well.
593 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
594 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
595 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
596 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
597 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
598 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
600 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
601 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
602 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
603 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
605 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
606 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
607 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
608 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
610 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
611 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
613 setOperationAction(ISD::AND, MVT::v8i8, Promote);
614 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
615 setOperationAction(ISD::AND, MVT::v4i16, Promote);
616 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
617 setOperationAction(ISD::AND, MVT::v2i32, Promote);
618 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
619 setOperationAction(ISD::AND, MVT::v1i64, Legal);
621 setOperationAction(ISD::OR, MVT::v8i8, Promote);
622 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
623 setOperationAction(ISD::OR, MVT::v4i16, Promote);
624 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
625 setOperationAction(ISD::OR, MVT::v2i32, Promote);
626 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
627 setOperationAction(ISD::OR, MVT::v1i64, Legal);
629 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
630 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
631 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
632 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
633 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
634 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
635 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
637 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
638 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
639 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
640 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
641 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
642 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
643 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
644 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
645 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
647 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
648 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
649 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
650 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
651 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
653 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
654 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
655 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
656 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
658 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
659 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
660 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
661 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
663 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
665 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
666 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
667 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
668 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
669 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
670 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
673 if (!UseSoftFloat && Subtarget->hasSSE1()) {
674 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
676 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
677 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
678 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
679 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
680 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
681 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
682 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
683 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
684 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
685 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
686 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
690 if (!UseSoftFloat && Subtarget->hasSSE2()) {
691 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
693 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
694 // registers cannot be used even for integer operations.
695 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
696 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
697 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
698 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
700 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
701 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
702 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
703 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
704 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
705 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
706 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
707 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
708 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
709 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
710 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
711 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
712 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
713 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
714 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
715 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
717 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
718 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
719 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
720 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
722 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
723 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
724 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
725 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
726 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
728 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
729 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
730 MVT VT = (MVT::SimpleValueType)i;
731 // Do not attempt to custom lower non-power-of-2 vectors
732 if (!isPowerOf2_32(VT.getVectorNumElements()))
734 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
735 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
736 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
739 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
740 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
741 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
742 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
743 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
744 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
746 if (Subtarget->is64Bit()) {
747 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
748 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
751 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
752 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
753 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
754 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
755 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
756 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
757 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
758 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
759 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
760 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
761 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
762 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
765 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
767 // Custom lower v2i64 and v2f64 selects.
768 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
769 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
770 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
771 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
775 if (Subtarget->hasSSE41()) {
776 // FIXME: Do we need to handle scalar-to-vector here?
777 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
779 // i8 and i16 vectors are custom , because the source register and source
780 // source memory operand types are not the same width. f32 vectors are
781 // custom since the immediate controlling the insert encodes additional
783 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
784 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
786 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
788 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
789 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
790 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
791 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
793 if (Subtarget->is64Bit()) {
794 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
795 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
799 if (Subtarget->hasSSE42()) {
800 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
803 // We want to custom lower some of our intrinsics.
804 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
806 // Add/Sub/Mul with overflow operations are custom lowered.
807 setOperationAction(ISD::SADDO, MVT::i32, Custom);
808 setOperationAction(ISD::SADDO, MVT::i64, Custom);
809 setOperationAction(ISD::UADDO, MVT::i32, Custom);
810 setOperationAction(ISD::UADDO, MVT::i64, Custom);
811 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
812 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
813 setOperationAction(ISD::USUBO, MVT::i32, Custom);
814 setOperationAction(ISD::USUBO, MVT::i64, Custom);
815 setOperationAction(ISD::SMULO, MVT::i32, Custom);
816 setOperationAction(ISD::SMULO, MVT::i64, Custom);
817 setOperationAction(ISD::UMULO, MVT::i32, Custom);
818 setOperationAction(ISD::UMULO, MVT::i64, Custom);
820 if (!Subtarget->is64Bit()) {
821 // These libcalls are not available in 32-bit.
822 setLibcallName(RTLIB::SHL_I128, 0);
823 setLibcallName(RTLIB::SRL_I128, 0);
824 setLibcallName(RTLIB::SRA_I128, 0);
827 // We have target-specific dag combine patterns for the following nodes:
828 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
829 setTargetDAGCombine(ISD::BUILD_VECTOR);
830 setTargetDAGCombine(ISD::SELECT);
831 setTargetDAGCombine(ISD::SHL);
832 setTargetDAGCombine(ISD::SRA);
833 setTargetDAGCombine(ISD::SRL);
834 setTargetDAGCombine(ISD::STORE);
835 if (Subtarget->is64Bit())
836 setTargetDAGCombine(ISD::MUL);
838 computeRegisterProperties();
840 // FIXME: These should be based on subtarget info. Plus, the values should
841 // be smaller when we are in optimizing for size mode.
842 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
843 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
844 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
845 allowUnalignedMemoryAccesses = true; // x86 supports it!
846 setPrefLoopAlignment(16);
850 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
855 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
856 /// the desired ByVal argument alignment.
857 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
860 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
861 if (VTy->getBitWidth() == 128)
863 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
864 unsigned EltAlign = 0;
865 getMaxByValAlign(ATy->getElementType(), EltAlign);
866 if (EltAlign > MaxAlign)
868 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
869 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
870 unsigned EltAlign = 0;
871 getMaxByValAlign(STy->getElementType(i), EltAlign);
872 if (EltAlign > MaxAlign)
881 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
882 /// function arguments in the caller parameter area. For X86, aggregates
883 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
884 /// are at 4-byte boundaries.
885 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
886 if (Subtarget->is64Bit()) {
887 // Max of 8 and alignment of type.
888 unsigned TyAlign = TD->getABITypeAlignment(Ty);
895 if (Subtarget->hasSSE1())
896 getMaxByValAlign(Ty, Align);
900 /// getOptimalMemOpType - Returns the target specific optimal type for load
901 /// and store operations as a result of memset, memcpy, and memmove
902 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
905 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
906 bool isSrcConst, bool isSrcStr) const {
907 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
908 // linux. This is because the stack realignment code can't handle certain
909 // cases like PR2962. This should be removed when PR2962 is fixed.
910 if (!NoImplicitFloat && Subtarget->getStackAlignment() >= 16) {
911 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
913 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
916 if (Subtarget->is64Bit() && Size >= 8)
921 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
923 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
924 SelectionDAG &DAG) const {
925 if (usesGlobalOffsetTable())
926 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
927 if (!Subtarget->isPICStyleRIPRel())
928 // This doesn't have DebugLoc associated with it, but is not really the
929 // same as a Register.
930 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
935 //===----------------------------------------------------------------------===//
936 // Return Value Calling Convention Implementation
937 //===----------------------------------------------------------------------===//
939 #include "X86GenCallingConv.inc"
941 /// LowerRET - Lower an ISD::RET node.
942 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
943 DebugLoc dl = Op.getDebugLoc();
944 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
946 SmallVector<CCValAssign, 16> RVLocs;
947 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
948 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
949 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
950 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
952 // If this is the first return lowered for this function, add the regs to the
953 // liveout set for the function.
954 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
955 for (unsigned i = 0; i != RVLocs.size(); ++i)
956 if (RVLocs[i].isRegLoc())
957 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
959 SDValue Chain = Op.getOperand(0);
961 // Handle tail call return.
962 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
963 if (Chain.getOpcode() == X86ISD::TAILCALL) {
964 SDValue TailCall = Chain;
965 SDValue TargetAddress = TailCall.getOperand(1);
966 SDValue StackAdjustment = TailCall.getOperand(2);
967 assert(((TargetAddress.getOpcode() == ISD::Register &&
968 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
969 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
970 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
971 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
972 "Expecting an global address, external symbol, or register");
973 assert(StackAdjustment.getOpcode() == ISD::Constant &&
974 "Expecting a const value");
976 SmallVector<SDValue,8> Operands;
977 Operands.push_back(Chain.getOperand(0));
978 Operands.push_back(TargetAddress);
979 Operands.push_back(StackAdjustment);
980 // Copy registers used by the call. Last operand is a flag so it is not
982 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
983 Operands.push_back(Chain.getOperand(i));
985 return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
992 SmallVector<SDValue, 6> RetOps;
993 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
994 // Operand #1 = Bytes To Pop
995 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
997 // Copy the result values into the output registers.
998 for (unsigned i = 0; i != RVLocs.size(); ++i) {
999 CCValAssign &VA = RVLocs[i];
1000 assert(VA.isRegLoc() && "Can only return in registers!");
1001 SDValue ValToCopy = Op.getOperand(i*2+1);
1003 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1004 // the RET instruction and handled by the FP Stackifier.
1005 if (VA.getLocReg() == X86::ST0 ||
1006 VA.getLocReg() == X86::ST1) {
1007 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1008 // change the value to the FP stack register class.
1009 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1010 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1011 RetOps.push_back(ValToCopy);
1012 // Don't emit a copytoreg.
1016 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1017 // which is returned in RAX / RDX.
1018 if (Subtarget->is64Bit()) {
1019 MVT ValVT = ValToCopy.getValueType();
1020 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1021 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1022 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1023 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1027 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1028 Flag = Chain.getValue(1);
1031 // The x86-64 ABI for returning structs by value requires that we copy
1032 // the sret argument into %rax for the return. We saved the argument into
1033 // a virtual register in the entry block, so now we copy the value out
1035 if (Subtarget->is64Bit() &&
1036 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1037 MachineFunction &MF = DAG.getMachineFunction();
1038 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1039 unsigned Reg = FuncInfo->getSRetReturnReg();
1041 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1042 FuncInfo->setSRetReturnReg(Reg);
1044 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1046 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1047 Flag = Chain.getValue(1);
1050 RetOps[0] = Chain; // Update chain.
1052 // Add the flag if we have it.
1054 RetOps.push_back(Flag);
1056 return DAG.getNode(X86ISD::RET_FLAG, dl,
1057 MVT::Other, &RetOps[0], RetOps.size());
1061 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1062 /// appropriate copies out of appropriate physical registers. This assumes that
1063 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1064 /// being lowered. The returns a SDNode with the same number of values as the
1066 SDNode *X86TargetLowering::
1067 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1068 unsigned CallingConv, SelectionDAG &DAG) {
1070 DebugLoc dl = TheCall->getDebugLoc();
1071 // Assign locations to each value returned by this call.
1072 SmallVector<CCValAssign, 16> RVLocs;
1073 bool isVarArg = TheCall->isVarArg();
1074 bool Is64Bit = Subtarget->is64Bit();
1075 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
1076 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1078 SmallVector<SDValue, 8> ResultVals;
1080 // Copy all of the result registers out of their specified physreg.
1081 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1082 CCValAssign &VA = RVLocs[i];
1083 MVT CopyVT = VA.getValVT();
1085 // If this is x86-64, and we disabled SSE, we can't return FP values
1086 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1087 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1088 cerr << "SSE register return with SSE disabled\n";
1092 // If this is a call to a function that returns an fp value on the floating
1093 // point stack, but where we prefer to use the value in xmm registers, copy
1094 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1095 if ((VA.getLocReg() == X86::ST0 ||
1096 VA.getLocReg() == X86::ST1) &&
1097 isScalarFPTypeInSSEReg(VA.getValVT())) {
1102 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1103 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1104 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1105 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1106 MVT::v2i64, InFlag).getValue(1);
1107 Val = Chain.getValue(0);
1108 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1109 Val, DAG.getConstant(0, MVT::i64));
1111 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1112 MVT::i64, InFlag).getValue(1);
1113 Val = Chain.getValue(0);
1115 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1117 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1118 CopyVT, InFlag).getValue(1);
1119 Val = Chain.getValue(0);
1121 InFlag = Chain.getValue(2);
1123 if (CopyVT != VA.getValVT()) {
1124 // Round the F80 the right size, which also moves to the appropriate xmm
1126 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1127 // This truncation won't change the value.
1128 DAG.getIntPtrConstant(1));
1131 ResultVals.push_back(Val);
1134 // Merge everything together with a MERGE_VALUES node.
1135 ResultVals.push_back(Chain);
1136 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1137 &ResultVals[0], ResultVals.size()).getNode();
1141 //===----------------------------------------------------------------------===//
1142 // C & StdCall & Fast Calling Convention implementation
1143 //===----------------------------------------------------------------------===//
1144 // StdCall calling convention seems to be standard for many Windows' API
1145 // routines and around. It differs from C calling convention just a little:
1146 // callee should clean up the stack, not caller. Symbols should be also
1147 // decorated in some fancy way :) It doesn't support any vector arguments.
1148 // For info on fast calling convention see Fast Calling Convention (tail call)
1149 // implementation LowerX86_32FastCCCallTo.
1151 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1153 static bool CallIsStructReturn(CallSDNode *TheCall) {
1154 unsigned NumOps = TheCall->getNumArgs();
1158 return TheCall->getArgFlags(0).isSRet();
1161 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1162 /// return semantics.
1163 static bool ArgsAreStructReturn(SDValue Op) {
1164 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1168 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1171 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1172 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1174 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1178 switch (CallingConv) {
1181 case CallingConv::X86_StdCall:
1182 return !Subtarget->is64Bit();
1183 case CallingConv::X86_FastCall:
1184 return !Subtarget->is64Bit();
1185 case CallingConv::Fast:
1186 return PerformTailCallOpt;
1190 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1191 /// given CallingConvention value.
1192 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1193 if (Subtarget->is64Bit()) {
1194 if (Subtarget->isTargetWin64())
1195 return CC_X86_Win64_C;
1196 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1197 return CC_X86_64_TailCall;
1202 if (CC == CallingConv::X86_FastCall)
1203 return CC_X86_32_FastCall;
1204 else if (CC == CallingConv::Fast)
1205 return CC_X86_32_FastCC;
1210 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1211 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1213 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1214 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1215 if (CC == CallingConv::X86_FastCall)
1217 else if (CC == CallingConv::X86_StdCall)
1223 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1224 /// in a register before calling.
1225 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1226 return !IsTailCall && !Is64Bit &&
1227 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1228 Subtarget->isPICStyleGOT();
1231 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1232 /// address to be loaded in a register.
1234 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1235 return !Is64Bit && IsTailCall &&
1236 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1237 Subtarget->isPICStyleGOT();
1240 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1241 /// by "Src" to address "Dst" with size and alignment information specified by
1242 /// the specific parameter attribute. The copy will be passed as a byval
1243 /// function parameter.
1245 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1246 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1248 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1249 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1250 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1253 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1254 const CCValAssign &VA,
1255 MachineFrameInfo *MFI,
1257 SDValue Root, unsigned i) {
1258 // Create the nodes corresponding to a load from this parameter slot.
1259 ISD::ArgFlagsTy Flags =
1260 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1261 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1262 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1264 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1265 // changed with more analysis.
1266 // In case of tail call optimization mark all arguments mutable. Since they
1267 // could be overwritten by lowering of arguments in case of a tail call.
1268 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1269 VA.getLocMemOffset(), isImmutable);
1270 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1271 if (Flags.isByVal())
1273 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1274 PseudoSourceValue::getFixedStack(FI), 0);
1278 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1279 MachineFunction &MF = DAG.getMachineFunction();
1280 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1281 DebugLoc dl = Op.getDebugLoc();
1283 const Function* Fn = MF.getFunction();
1284 if (Fn->hasExternalLinkage() &&
1285 Subtarget->isTargetCygMing() &&
1286 Fn->getName() == "main")
1287 FuncInfo->setForceFramePointer(true);
1289 // Decorate the function name.
1290 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1292 MachineFrameInfo *MFI = MF.getFrameInfo();
1293 SDValue Root = Op.getOperand(0);
1294 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1295 unsigned CC = MF.getFunction()->getCallingConv();
1296 bool Is64Bit = Subtarget->is64Bit();
1297 bool IsWin64 = Subtarget->isTargetWin64();
1299 assert(!(isVarArg && CC == CallingConv::Fast) &&
1300 "Var args not supported with calling convention fastcc");
1302 // Assign locations to all of the incoming arguments.
1303 SmallVector<CCValAssign, 16> ArgLocs;
1304 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1305 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1307 SmallVector<SDValue, 8> ArgValues;
1308 unsigned LastVal = ~0U;
1309 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1310 CCValAssign &VA = ArgLocs[i];
1311 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1313 assert(VA.getValNo() != LastVal &&
1314 "Don't support value assigned to multiple locs yet");
1315 LastVal = VA.getValNo();
1317 if (VA.isRegLoc()) {
1318 MVT RegVT = VA.getLocVT();
1319 TargetRegisterClass *RC = NULL;
1320 if (RegVT == MVT::i32)
1321 RC = X86::GR32RegisterClass;
1322 else if (Is64Bit && RegVT == MVT::i64)
1323 RC = X86::GR64RegisterClass;
1324 else if (RegVT == MVT::f32)
1325 RC = X86::FR32RegisterClass;
1326 else if (RegVT == MVT::f64)
1327 RC = X86::FR64RegisterClass;
1328 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1329 RC = X86::VR128RegisterClass;
1330 else if (RegVT.isVector()) {
1331 assert(RegVT.getSizeInBits() == 64);
1333 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1335 // Darwin calling convention passes MMX values in either GPRs or
1336 // XMMs in x86-64. Other targets pass them in memory.
1337 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1338 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1341 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1346 assert(0 && "Unknown argument type!");
1349 unsigned Reg = DAG.getMachineFunction().addLiveIn(VA.getLocReg(), RC);
1350 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1352 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1353 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1355 if (VA.getLocInfo() == CCValAssign::SExt)
1356 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1357 DAG.getValueType(VA.getValVT()));
1358 else if (VA.getLocInfo() == CCValAssign::ZExt)
1359 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1360 DAG.getValueType(VA.getValVT()));
1362 if (VA.getLocInfo() != CCValAssign::Full)
1363 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1365 // Handle MMX values passed in GPRs.
1366 if (Is64Bit && RegVT != VA.getLocVT()) {
1367 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1368 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1369 else if (RC == X86::VR128RegisterClass) {
1370 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1371 ArgValue, DAG.getConstant(0, MVT::i64));
1372 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1376 ArgValues.push_back(ArgValue);
1378 assert(VA.isMemLoc());
1379 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1383 // The x86-64 ABI for returning structs by value requires that we copy
1384 // the sret argument into %rax for the return. Save the argument into
1385 // a virtual register so that we can access it from the return points.
1386 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1387 MachineFunction &MF = DAG.getMachineFunction();
1388 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1389 unsigned Reg = FuncInfo->getSRetReturnReg();
1391 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1392 FuncInfo->setSRetReturnReg(Reg);
1394 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1395 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1398 unsigned StackSize = CCInfo.getNextStackOffset();
1399 // align stack specially for tail calls
1400 if (PerformTailCallOpt && CC == CallingConv::Fast)
1401 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1403 // If the function takes variable number of arguments, make a frame index for
1404 // the start of the first vararg value... for expansion of llvm.va_start.
1406 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1407 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1410 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1412 // FIXME: We should really autogenerate these arrays
1413 static const unsigned GPR64ArgRegsWin64[] = {
1414 X86::RCX, X86::RDX, X86::R8, X86::R9
1416 static const unsigned XMMArgRegsWin64[] = {
1417 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1419 static const unsigned GPR64ArgRegs64Bit[] = {
1420 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1422 static const unsigned XMMArgRegs64Bit[] = {
1423 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1424 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1426 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1429 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1430 GPR64ArgRegs = GPR64ArgRegsWin64;
1431 XMMArgRegs = XMMArgRegsWin64;
1433 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1434 GPR64ArgRegs = GPR64ArgRegs64Bit;
1435 XMMArgRegs = XMMArgRegs64Bit;
1437 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1439 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1442 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1443 "SSE register cannot be used when SSE is disabled!");
1444 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloat) &&
1445 "SSE register cannot be used when SSE is disabled!");
1446 if (UseSoftFloat || NoImplicitFloat || !Subtarget->hasSSE1())
1447 // Kernel mode asks for SSE to be disabled, so don't push them
1449 TotalNumXMMRegs = 0;
1451 // For X86-64, if there are vararg parameters that are passed via
1452 // registers, then we must store them to their spots on the stack so they
1453 // may be loaded by deferencing the result of va_next.
1454 VarArgsGPOffset = NumIntRegs * 8;
1455 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1456 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1457 TotalNumXMMRegs * 16, 16);
1459 // Store the integer parameter registers.
1460 SmallVector<SDValue, 8> MemOps;
1461 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1462 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1463 DAG.getIntPtrConstant(VarArgsGPOffset));
1464 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1465 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1466 X86::GR64RegisterClass);
1467 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1469 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1470 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1471 MemOps.push_back(Store);
1472 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1473 DAG.getIntPtrConstant(8));
1476 // Now store the XMM (fp + vector) parameter registers.
1477 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1478 DAG.getIntPtrConstant(VarArgsFPOffset));
1479 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1480 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1481 X86::VR128RegisterClass);
1482 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1484 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1485 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1486 MemOps.push_back(Store);
1487 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1488 DAG.getIntPtrConstant(16));
1490 if (!MemOps.empty())
1491 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1492 &MemOps[0], MemOps.size());
1496 ArgValues.push_back(Root);
1498 // Some CCs need callee pop.
1499 if (IsCalleePop(isVarArg, CC)) {
1500 BytesToPopOnReturn = StackSize; // Callee pops everything.
1501 BytesCallerReserves = 0;
1503 BytesToPopOnReturn = 0; // Callee pops nothing.
1504 // If this is an sret function, the return should pop the hidden pointer.
1505 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1506 BytesToPopOnReturn = 4;
1507 BytesCallerReserves = StackSize;
1511 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1512 if (CC == CallingConv::X86_FastCall)
1513 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1516 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1518 // Return the new list of results.
1519 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1520 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1524 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1525 const SDValue &StackPtr,
1526 const CCValAssign &VA,
1528 SDValue Arg, ISD::ArgFlagsTy Flags) {
1529 DebugLoc dl = TheCall->getDebugLoc();
1530 unsigned LocMemOffset = VA.getLocMemOffset();
1531 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1532 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1533 if (Flags.isByVal()) {
1534 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1536 return DAG.getStore(Chain, dl, Arg, PtrOff,
1537 PseudoSourceValue::getStack(), LocMemOffset);
1540 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1541 /// optimization is performed and it is required.
1543 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1544 SDValue &OutRetAddr,
1550 if (!IsTailCall || FPDiff==0) return Chain;
1552 // Adjust the Return address stack slot.
1553 MVT VT = getPointerTy();
1554 OutRetAddr = getReturnAddressFrameIndex(DAG);
1556 // Load the "old" Return address.
1557 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1558 return SDValue(OutRetAddr.getNode(), 1);
1561 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1562 /// optimization is performed and it is required (FPDiff!=0).
1564 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1565 SDValue Chain, SDValue RetAddrFrIdx,
1566 bool Is64Bit, int FPDiff, DebugLoc dl) {
1567 // Store the return address to the appropriate stack slot.
1568 if (!FPDiff) return Chain;
1569 // Calculate the new stack slot for the return address.
1570 int SlotSize = Is64Bit ? 8 : 4;
1571 int NewReturnAddrFI =
1572 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1573 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1574 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1575 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1576 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1580 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1581 MachineFunction &MF = DAG.getMachineFunction();
1582 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1583 SDValue Chain = TheCall->getChain();
1584 unsigned CC = TheCall->getCallingConv();
1585 bool isVarArg = TheCall->isVarArg();
1586 bool IsTailCall = TheCall->isTailCall() &&
1587 CC == CallingConv::Fast && PerformTailCallOpt;
1588 SDValue Callee = TheCall->getCallee();
1589 bool Is64Bit = Subtarget->is64Bit();
1590 bool IsStructRet = CallIsStructReturn(TheCall);
1591 DebugLoc dl = TheCall->getDebugLoc();
1593 assert(!(isVarArg && CC == CallingConv::Fast) &&
1594 "Var args not supported with calling convention fastcc");
1596 // Analyze operands of the call, assigning locations to each operand.
1597 SmallVector<CCValAssign, 16> ArgLocs;
1598 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1599 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1601 // Get a count of how many bytes are to be pushed on the stack.
1602 unsigned NumBytes = CCInfo.getNextStackOffset();
1603 if (PerformTailCallOpt && CC == CallingConv::Fast)
1604 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1608 // Lower arguments at fp - stackoffset + fpdiff.
1609 unsigned NumBytesCallerPushed =
1610 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1611 FPDiff = NumBytesCallerPushed - NumBytes;
1613 // Set the delta of movement of the returnaddr stackslot.
1614 // But only set if delta is greater than previous delta.
1615 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1616 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1619 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1621 SDValue RetAddrFrIdx;
1622 // Load return adress for tail calls.
1623 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1626 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1627 SmallVector<SDValue, 8> MemOpChains;
1630 // Walk the register/memloc assignments, inserting copies/loads. In the case
1631 // of tail call optimization arguments are handle later.
1632 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1633 CCValAssign &VA = ArgLocs[i];
1634 SDValue Arg = TheCall->getArg(i);
1635 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1636 bool isByVal = Flags.isByVal();
1638 // Promote the value if needed.
1639 switch (VA.getLocInfo()) {
1640 default: assert(0 && "Unknown loc info!");
1641 case CCValAssign::Full: break;
1642 case CCValAssign::SExt:
1643 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1645 case CCValAssign::ZExt:
1646 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1648 case CCValAssign::AExt:
1649 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1653 if (VA.isRegLoc()) {
1655 MVT RegVT = VA.getLocVT();
1656 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1657 switch (VA.getLocReg()) {
1660 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1662 // Special case: passing MMX values in GPR registers.
1663 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1666 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1667 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1668 // Special case: passing MMX values in XMM registers.
1669 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1670 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1671 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1676 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1678 if (!IsTailCall || (IsTailCall && isByVal)) {
1679 assert(VA.isMemLoc());
1680 if (StackPtr.getNode() == 0)
1681 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1683 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1684 Chain, Arg, Flags));
1689 if (!MemOpChains.empty())
1690 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1691 &MemOpChains[0], MemOpChains.size());
1693 // Build a sequence of copy-to-reg nodes chained together with token chain
1694 // and flag operands which copy the outgoing args into registers.
1696 // Tail call byval lowering might overwrite argument registers so in case of
1697 // tail call optimization the copies to registers are lowered later.
1699 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1700 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1701 RegsToPass[i].second, InFlag);
1702 InFlag = Chain.getValue(1);
1705 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1707 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1708 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1709 DAG.getNode(X86ISD::GlobalBaseReg,
1710 DebugLoc::getUnknownLoc(),
1713 InFlag = Chain.getValue(1);
1715 // If we are tail calling and generating PIC/GOT style code load the address
1716 // of the callee into ecx. The value in ecx is used as target of the tail
1717 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1718 // calls on PIC/GOT architectures. Normally we would just put the address of
1719 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1720 // restored (since ebx is callee saved) before jumping to the target@PLT.
1721 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1722 // Note: The actual moving to ecx is done further down.
1723 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1724 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1725 !G->getGlobal()->hasProtectedVisibility())
1726 Callee = LowerGlobalAddress(Callee, DAG);
1727 else if (isa<ExternalSymbolSDNode>(Callee))
1728 Callee = LowerExternalSymbol(Callee,DAG);
1731 if (Is64Bit && isVarArg) {
1732 // From AMD64 ABI document:
1733 // For calls that may call functions that use varargs or stdargs
1734 // (prototype-less calls or calls to functions containing ellipsis (...) in
1735 // the declaration) %al is used as hidden argument to specify the number
1736 // of SSE registers used. The contents of %al do not need to match exactly
1737 // the number of registers, but must be an ubound on the number of SSE
1738 // registers used and is in the range 0 - 8 inclusive.
1740 // FIXME: Verify this on Win64
1741 // Count the number of XMM registers allocated.
1742 static const unsigned XMMArgRegs[] = {
1743 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1744 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1746 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1747 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1748 && "SSE registers cannot be used when SSE is disabled");
1750 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1751 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1752 InFlag = Chain.getValue(1);
1756 // For tail calls lower the arguments to the 'real' stack slot.
1758 SmallVector<SDValue, 8> MemOpChains2;
1761 // Do not flag preceeding copytoreg stuff together with the following stuff.
1763 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1764 CCValAssign &VA = ArgLocs[i];
1765 if (!VA.isRegLoc()) {
1766 assert(VA.isMemLoc());
1767 SDValue Arg = TheCall->getArg(i);
1768 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1769 // Create frame index.
1770 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1771 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1772 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1773 FIN = DAG.getFrameIndex(FI, getPointerTy());
1775 if (Flags.isByVal()) {
1776 // Copy relative to framepointer.
1777 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1778 if (StackPtr.getNode() == 0)
1779 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1781 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1783 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1786 // Store relative to framepointer.
1787 MemOpChains2.push_back(
1788 DAG.getStore(Chain, dl, Arg, FIN,
1789 PseudoSourceValue::getFixedStack(FI), 0));
1794 if (!MemOpChains2.empty())
1795 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1796 &MemOpChains2[0], MemOpChains2.size());
1798 // Copy arguments to their registers.
1799 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1800 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1801 RegsToPass[i].second, InFlag);
1802 InFlag = Chain.getValue(1);
1806 // Store the return address to the appropriate stack slot.
1807 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1811 // If the callee is a GlobalAddress node (quite common, every direct call is)
1812 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1813 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1814 // We should use extra load for direct calls to dllimported functions in
1816 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1817 getTargetMachine(), true))
1818 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy(),
1820 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1821 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1822 } else if (IsTailCall) {
1823 unsigned Opc = Is64Bit ? X86::R9 : X86::EAX;
1825 Chain = DAG.getCopyToReg(Chain, dl,
1826 DAG.getRegister(Opc, getPointerTy()),
1828 Callee = DAG.getRegister(Opc, getPointerTy());
1829 // Add register as live out.
1830 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1833 // Returns a chain & a flag for retval copy to use.
1834 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1835 SmallVector<SDValue, 8> Ops;
1838 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1839 DAG.getIntPtrConstant(0, true), InFlag);
1840 InFlag = Chain.getValue(1);
1842 // Returns a chain & a flag for retval copy to use.
1843 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1847 Ops.push_back(Chain);
1848 Ops.push_back(Callee);
1851 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1853 // Add argument registers to the end of the list so that they are known live
1855 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1856 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1857 RegsToPass[i].second.getValueType()));
1859 // Add an implicit use GOT pointer in EBX.
1860 if (!IsTailCall && !Is64Bit &&
1861 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1862 Subtarget->isPICStyleGOT())
1863 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1865 // Add an implicit use of AL for x86 vararg functions.
1866 if (Is64Bit && isVarArg)
1867 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1869 if (InFlag.getNode())
1870 Ops.push_back(InFlag);
1873 assert(InFlag.getNode() &&
1874 "Flag must be set. Depend on flag being set in LowerRET");
1875 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
1876 TheCall->getVTList(), &Ops[0], Ops.size());
1878 return SDValue(Chain.getNode(), Op.getResNo());
1881 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
1882 InFlag = Chain.getValue(1);
1884 // Create the CALLSEQ_END node.
1885 unsigned NumBytesForCalleeToPush;
1886 if (IsCalleePop(isVarArg, CC))
1887 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1888 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1889 // If this is is a call to a struct-return function, the callee
1890 // pops the hidden struct pointer, so we have to push it back.
1891 // This is common for Darwin/X86, Linux & Mingw32 targets.
1892 NumBytesForCalleeToPush = 4;
1894 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1896 // Returns a flag for retval copy to use.
1897 Chain = DAG.getCALLSEQ_END(Chain,
1898 DAG.getIntPtrConstant(NumBytes, true),
1899 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
1902 InFlag = Chain.getValue(1);
1904 // Handle result values, copying them out of physregs into vregs that we
1906 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1911 //===----------------------------------------------------------------------===//
1912 // Fast Calling Convention (tail call) implementation
1913 //===----------------------------------------------------------------------===//
1915 // Like std call, callee cleans arguments, convention except that ECX is
1916 // reserved for storing the tail called function address. Only 2 registers are
1917 // free for argument passing (inreg). Tail call optimization is performed
1919 // * tailcallopt is enabled
1920 // * caller/callee are fastcc
1921 // On X86_64 architecture with GOT-style position independent code only local
1922 // (within module) calls are supported at the moment.
1923 // To keep the stack aligned according to platform abi the function
1924 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1925 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1926 // If a tail called function callee has more arguments than the caller the
1927 // caller needs to make sure that there is room to move the RETADDR to. This is
1928 // achieved by reserving an area the size of the argument delta right after the
1929 // original REtADDR, but before the saved framepointer or the spilled registers
1930 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1942 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1943 /// for a 16 byte align requirement.
1944 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1945 SelectionDAG& DAG) {
1946 MachineFunction &MF = DAG.getMachineFunction();
1947 const TargetMachine &TM = MF.getTarget();
1948 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1949 unsigned StackAlignment = TFI.getStackAlignment();
1950 uint64_t AlignMask = StackAlignment - 1;
1951 int64_t Offset = StackSize;
1952 uint64_t SlotSize = TD->getPointerSize();
1953 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1954 // Number smaller than 12 so just add the difference.
1955 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1957 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1958 Offset = ((~AlignMask) & Offset) + StackAlignment +
1959 (StackAlignment-SlotSize);
1964 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1965 /// following the call is a return. A function is eligible if caller/callee
1966 /// calling conventions match, currently only fastcc supports tail calls, and
1967 /// the function CALL is immediatly followed by a RET.
1968 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1970 SelectionDAG& DAG) const {
1971 if (!PerformTailCallOpt)
1974 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1975 MachineFunction &MF = DAG.getMachineFunction();
1976 unsigned CallerCC = MF.getFunction()->getCallingConv();
1977 unsigned CalleeCC= TheCall->getCallingConv();
1978 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1979 SDValue Callee = TheCall->getCallee();
1980 // On x86/32Bit PIC/GOT tail calls are supported.
1981 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1982 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1985 // Can only do local tail calls (in same module, hidden or protected) on
1986 // x86_64 PIC/GOT at the moment.
1987 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1988 return G->getGlobal()->hasHiddenVisibility()
1989 || G->getGlobal()->hasProtectedVisibility();
1997 X86TargetLowering::createFastISel(MachineFunction &mf,
1998 MachineModuleInfo *mmo,
2000 DenseMap<const Value *, unsigned> &vm,
2001 DenseMap<const BasicBlock *,
2002 MachineBasicBlock *> &bm,
2003 DenseMap<const AllocaInst *, int> &am
2005 , SmallSet<Instruction*, 8> &cil
2008 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2016 //===----------------------------------------------------------------------===//
2017 // Other Lowering Hooks
2018 //===----------------------------------------------------------------------===//
2021 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2022 MachineFunction &MF = DAG.getMachineFunction();
2023 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2024 int ReturnAddrIndex = FuncInfo->getRAIndex();
2026 if (ReturnAddrIndex == 0) {
2027 // Set up a frame object for the return address.
2028 uint64_t SlotSize = TD->getPointerSize();
2029 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2030 FuncInfo->setRAIndex(ReturnAddrIndex);
2033 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2037 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2038 /// specific condition code, returning the condition code and the LHS/RHS of the
2039 /// comparison to make.
2040 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2041 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2043 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2044 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2045 // X > -1 -> X == 0, jump !sign.
2046 RHS = DAG.getConstant(0, RHS.getValueType());
2047 return X86::COND_NS;
2048 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2049 // X < 0 -> X == 0, jump on sign.
2051 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2053 RHS = DAG.getConstant(0, RHS.getValueType());
2054 return X86::COND_LE;
2058 switch (SetCCOpcode) {
2059 default: assert(0 && "Invalid integer condition!");
2060 case ISD::SETEQ: return X86::COND_E;
2061 case ISD::SETGT: return X86::COND_G;
2062 case ISD::SETGE: return X86::COND_GE;
2063 case ISD::SETLT: return X86::COND_L;
2064 case ISD::SETLE: return X86::COND_LE;
2065 case ISD::SETNE: return X86::COND_NE;
2066 case ISD::SETULT: return X86::COND_B;
2067 case ISD::SETUGT: return X86::COND_A;
2068 case ISD::SETULE: return X86::COND_BE;
2069 case ISD::SETUGE: return X86::COND_AE;
2073 // First determine if it is required or is profitable to flip the operands.
2075 // If LHS is a foldable load, but RHS is not, flip the condition.
2076 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2077 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2078 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2079 std::swap(LHS, RHS);
2082 switch (SetCCOpcode) {
2088 std::swap(LHS, RHS);
2092 // On a floating point condition, the flags are set as follows:
2094 // 0 | 0 | 0 | X > Y
2095 // 0 | 0 | 1 | X < Y
2096 // 1 | 0 | 0 | X == Y
2097 // 1 | 1 | 1 | unordered
2098 switch (SetCCOpcode) {
2099 default: assert(0 && "Condcode should be pre-legalized away");
2101 case ISD::SETEQ: return X86::COND_E;
2102 case ISD::SETOLT: // flipped
2104 case ISD::SETGT: return X86::COND_A;
2105 case ISD::SETOLE: // flipped
2107 case ISD::SETGE: return X86::COND_AE;
2108 case ISD::SETUGT: // flipped
2110 case ISD::SETLT: return X86::COND_B;
2111 case ISD::SETUGE: // flipped
2113 case ISD::SETLE: return X86::COND_BE;
2115 case ISD::SETNE: return X86::COND_NE;
2116 case ISD::SETUO: return X86::COND_P;
2117 case ISD::SETO: return X86::COND_NP;
2121 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2122 /// code. Current x86 isa includes the following FP cmov instructions:
2123 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2124 static bool hasFPCMov(unsigned X86CC) {
2140 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2141 /// the specified range (L, H].
2142 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2143 return (Val < 0) || (Val >= Low && Val < Hi);
2146 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2147 /// specified value.
2148 static bool isUndefOrEqual(int Val, int CmpVal) {
2149 if (Val < 0 || Val == CmpVal)
2154 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2155 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2156 /// the second operand.
2157 static bool isPSHUFDMask(const int *Mask, MVT VT) {
2158 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2159 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2160 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2161 return (Mask[0] < 2 && Mask[1] < 2);
2165 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2166 return ::isPSHUFDMask(N->getMask(), N->getValueType(0));
2169 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2170 /// is suitable for input to PSHUFHW.
2171 static bool isPSHUFHWMask(const int *Mask, MVT VT) {
2172 if (VT != MVT::v8i16)
2175 // Lower quadword copied in order or undef.
2176 for (int i = 0; i != 4; ++i)
2177 if (Mask[i] >= 0 && Mask[i] != i)
2180 // Upper quadword shuffled.
2181 for (int i = 4; i != 8; ++i)
2182 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2188 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2189 return ::isPSHUFHWMask(N->getMask(), N->getValueType(0));
2192 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2193 /// is suitable for input to PSHUFLW.
2194 static bool isPSHUFLWMask(const int *Mask, MVT VT) {
2195 if (VT != MVT::v8i16)
2198 // Upper quadword copied in order.
2199 for (int i = 4; i != 8; ++i)
2200 if (Mask[i] >= 0 && Mask[i] != i)
2203 // Lower quadword shuffled.
2204 for (int i = 0; i != 4; ++i)
2211 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2212 return ::isPSHUFLWMask(N->getMask(), N->getValueType(0));
2215 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2216 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2217 static bool isSHUFPMask(const int *Mask, MVT VT) {
2218 int NumElems = VT.getVectorNumElements();
2219 if (NumElems != 2 && NumElems != 4)
2222 int Half = NumElems / 2;
2223 for (int i = 0; i < Half; ++i)
2224 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2226 for (int i = Half; i < NumElems; ++i)
2227 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2233 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2234 return ::isSHUFPMask(N->getMask(), N->getValueType(0));
2237 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2238 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2239 /// half elements to come from vector 1 (which would equal the dest.) and
2240 /// the upper half to come from vector 2.
2241 static bool isCommutedSHUFPMask(const int *Mask, MVT VT) {
2242 int NumElems = VT.getVectorNumElements();
2244 if (NumElems != 2 && NumElems != 4)
2247 int Half = NumElems / 2;
2248 for (int i = 0; i < Half; ++i)
2249 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2251 for (int i = Half; i < NumElems; ++i)
2252 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2257 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2258 return isCommutedSHUFPMask(N->getMask(), N->getValueType(0));
2261 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2262 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2263 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2264 if (N->getValueType(0).getVectorNumElements() != 4)
2267 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2268 const int *Mask = N->getMask();
2269 return isUndefOrEqual(Mask[0], 6) &&
2270 isUndefOrEqual(Mask[1], 7) &&
2271 isUndefOrEqual(Mask[2], 2) &&
2272 isUndefOrEqual(Mask[3], 3);
2275 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2276 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2277 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2278 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2280 if (NumElems != 2 && NumElems != 4)
2283 const int *Mask = N->getMask();
2284 for (unsigned i = 0; i < NumElems/2; ++i)
2285 if (!isUndefOrEqual(Mask[i], i + NumElems))
2288 for (unsigned i = NumElems/2; i < NumElems; ++i)
2289 if (!isUndefOrEqual(Mask[i], i))
2295 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2296 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2298 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2299 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2301 if (NumElems != 2 && NumElems != 4)
2304 const int *Mask = N->getMask();
2305 for (unsigned i = 0; i < NumElems/2; ++i)
2306 if (!isUndefOrEqual(Mask[i], i))
2309 for (unsigned i = 0; i < NumElems/2; ++i)
2310 if (!isUndefOrEqual(Mask[i + NumElems/2], i + NumElems))
2316 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2317 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2319 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2320 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2325 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2326 const int *Mask = N->getMask();
2327 return isUndefOrEqual(Mask[0], 2) && isUndefOrEqual(Mask[1], 3) &&
2328 isUndefOrEqual(Mask[2], 2) && isUndefOrEqual(Mask[3], 3);
2331 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2332 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2333 static bool isUNPCKLMask(const int *Mask, MVT VT, bool V2IsSplat = false) {
2334 int NumElts = VT.getVectorNumElements();
2335 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2338 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2340 int BitI1 = Mask[i+1];
2341 if (!isUndefOrEqual(BitI, j))
2344 if (!isUndefOrEqual(BitI1, NumElts))
2347 if (!isUndefOrEqual(BitI1, j + NumElts))
2354 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2355 return ::isUNPCKLMask(N->getMask(), N->getValueType(0), V2IsSplat);
2358 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2359 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2360 static bool isUNPCKHMask(const int *Mask, MVT VT, bool V2IsSplat = false) {
2361 int NumElts = VT.getVectorNumElements();
2362 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2365 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2367 int BitI1 = Mask[i+1];
2368 if (!isUndefOrEqual(BitI, j + NumElts/2))
2371 if (isUndefOrEqual(BitI1, NumElts))
2374 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2381 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2382 return ::isUNPCKHMask(N->getMask(), N->getValueType(0), V2IsSplat);
2385 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2386 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2388 static bool isUNPCKL_v_undef_Mask(const int *Mask, MVT VT) {
2389 int NumElems = VT.getVectorNumElements();
2390 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2393 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2395 int BitI1 = Mask[i+1];
2396 if (!isUndefOrEqual(BitI, j))
2398 if (!isUndefOrEqual(BitI1, j))
2404 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2405 return ::isUNPCKL_v_undef_Mask(N->getMask(), N->getValueType(0));
2408 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2409 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2411 static bool isUNPCKH_v_undef_Mask(const int *Mask, MVT VT) {
2412 int NumElems = VT.getVectorNumElements();
2413 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2416 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2418 int BitI1 = Mask[i+1];
2419 if (!isUndefOrEqual(BitI, j))
2421 if (!isUndefOrEqual(BitI1, j))
2427 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2428 return ::isUNPCKH_v_undef_Mask(N->getMask(), N->getValueType(0));
2431 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2432 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2433 /// MOVSD, and MOVD, i.e. setting the lowest element.
2434 static bool isMOVLMask(const int *Mask, MVT VT) {
2435 int NumElts = VT.getVectorNumElements();
2436 if (NumElts != 2 && NumElts != 4)
2439 if (!isUndefOrEqual(Mask[0], NumElts))
2442 for (int i = 1; i < NumElts; ++i)
2443 if (!isUndefOrEqual(Mask[i], i))
2449 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2450 return ::isMOVLMask(N->getMask(), N->getValueType(0));
2453 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2454 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2455 /// element of vector 2 and the other elements to come from vector 1 in order.
2456 static bool isCommutedMOVLMask(const int *Mask, MVT VT, bool V2IsSplat = false,
2457 bool V2IsUndef = false) {
2458 int NumOps = VT.getVectorNumElements();
2459 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2462 if (!isUndefOrEqual(Mask[0], 0))
2465 for (int i = 1; i < NumOps; ++i)
2466 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2467 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2468 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2474 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2475 bool V2IsUndef = false) {
2476 return isCommutedMOVLMask(N->getMask(), N->getValueType(0), V2IsSplat,
2480 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2481 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2482 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2483 if (N->getValueType(0).getVectorNumElements() != 4)
2486 // Expect 1, 1, 3, 3
2487 const int *Mask = N->getMask();
2488 for (unsigned i = 0; i < 2; ++i)
2489 if (Mask[i] >=0 && Mask[i] != 1)
2493 for (unsigned i = 2; i < 4; ++i) {
2494 if (Mask[i] >= 0 && Mask[i] != 3)
2499 // Don't use movshdup if it can be done with a shufps.
2500 // FIXME: verify that matching u, u, 3, 3 is what we want.
2504 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2505 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2506 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2507 if (N->getValueType(0).getVectorNumElements() != 4)
2510 // Expect 0, 0, 2, 2
2511 const int *Mask = N->getMask();
2512 for (unsigned i = 0; i < 2; ++i)
2517 for (unsigned i = 2; i < 4; ++i) {
2518 if (Mask[i] >= 0 && Mask[i] != 2)
2523 // Don't use movsldup if it can be done with a shufps.
2527 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2528 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2529 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2530 int e = N->getValueType(0).getVectorNumElements() / 2;
2531 const int *Mask = N->getMask();
2533 for (int i = 0; i < e; ++i)
2534 if (!isUndefOrEqual(Mask[i], i))
2536 for (int i = 0; i < e; ++i)
2537 if (!isUndefOrEqual(Mask[e+i], i))
2542 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2543 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2545 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2546 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2547 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2548 const int *MaskP = SVOp->getMask();
2550 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2552 for (int i = 0; i < NumOperands; ++i) {
2553 int Val = MaskP[NumOperands-i-1];
2554 if (Val < 0) Val = 0;
2555 if (Val >= NumOperands) Val -= NumOperands;
2557 if (i != NumOperands - 1)
2563 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2564 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2566 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2567 const int *MaskP = cast<ShuffleVectorSDNode>(N)->getMask();
2569 // 8 nodes, but we only care about the last 4.
2570 for (unsigned i = 7; i >= 4; --i) {
2580 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2581 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2583 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2584 const int *MaskP = cast<ShuffleVectorSDNode>(N)->getMask();
2586 // 8 nodes, but we only care about the first 4.
2587 for (int i = 3; i >= 0; --i) {
2597 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2598 /// their permute mask.
2599 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2600 SelectionDAG &DAG) {
2601 MVT VT = SVOp->getValueType(0);
2602 int NumElems = VT.getVectorNumElements();
2603 const int *Mask = SVOp->getMask();
2604 SmallVector<int, 8> MaskVec;
2606 for (int i = 0; i != NumElems; ++i) {
2609 MaskVec.push_back(idx);
2610 else if (idx < NumElems)
2611 MaskVec.push_back(idx + NumElems);
2613 MaskVec.push_back(idx - NumElems);
2615 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2616 SVOp->getOperand(0), &MaskVec[0]);
2619 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2620 /// the two vector operands have swapped position.
2621 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, MVT VT) {
2622 int NumElems = VT.getVectorNumElements();
2623 for (int i = 0; i != NumElems; ++i) {
2627 else if (idx < NumElems)
2628 Mask[i] = idx + NumElems;
2630 Mask[i] = idx - NumElems;
2634 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2635 /// match movhlps. The lower half elements should come from upper half of
2636 /// V1 (and in order), and the upper half elements should come from the upper
2637 /// half of V2 (and in order).
2638 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2639 int NumElems = Op->getValueType(0).getVectorNumElements();
2640 const int *Mask = Op->getMask();
2644 for (unsigned i = 0, e = 2; i != e; ++i)
2645 if (!isUndefOrEqual(Mask[i], i+2))
2647 for (unsigned i = 2; i != 4; ++i)
2648 if (!isUndefOrEqual(Mask[i], i+4))
2653 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2654 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2656 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2657 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2659 N = N->getOperand(0).getNode();
2660 if (!ISD::isNON_EXTLoad(N))
2663 *LD = cast<LoadSDNode>(N);
2667 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2668 /// match movlp{s|d}. The lower half elements should come from lower half of
2669 /// V1 (and in order), and the upper half elements should come from the upper
2670 /// half of V2 (and in order). And since V1 will become the source of the
2671 /// MOVLP, it must be either a vector load or a scalar load to vector.
2672 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2673 ShuffleVectorSDNode *Op) {
2674 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2676 // Is V2 is a vector load, don't do this transformation. We will try to use
2677 // load folding shufps op.
2678 if (ISD::isNON_EXTLoad(V2))
2681 int NumElems = Op->getValueType(0).getVectorNumElements();
2682 const int *Mask = Op->getMask();
2684 if (NumElems != 2 && NumElems != 4)
2686 for (int i = 0, e = NumElems/2; i != e; ++i)
2687 if (!isUndefOrEqual(Mask[i], i))
2689 for (int i = NumElems/2; i != NumElems; ++i)
2690 if (!isUndefOrEqual(Mask[i], i+NumElems))
2695 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2697 static bool isSplatVector(SDNode *N) {
2698 if (N->getOpcode() != ISD::BUILD_VECTOR)
2701 SDValue SplatValue = N->getOperand(0);
2702 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2703 if (N->getOperand(i) != SplatValue)
2708 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2710 static inline bool isZeroNode(SDValue Elt) {
2711 return ((isa<ConstantSDNode>(Elt) &&
2712 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2713 (isa<ConstantFPSDNode>(Elt) &&
2714 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2717 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2718 /// to an zero vector.
2719 /// FIXME: move to dag combiner?
2720 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2721 SDValue V1 = N->getOperand(0);
2722 SDValue V2 = N->getOperand(1);
2723 const int *Mask = N->getMask();
2724 int NumElems = N->getValueType(0).getVectorNumElements();
2725 for (int i = 0; i != NumElems; ++i) {
2727 if (Idx >= NumElems) {
2728 unsigned Opc = V2.getOpcode();
2729 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2731 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V2.getOperand(Idx-NumElems)))
2733 } else if (Idx >= 0) {
2734 unsigned Opc = V1.getOpcode();
2735 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2737 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V1.getOperand(Idx)))
2744 /// getZeroVector - Returns a vector of specified type with all zero elements.
2746 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2748 assert(VT.isVector() && "Expected a vector type");
2750 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2751 // type. This ensures they get CSE'd.
2753 if (VT.getSizeInBits() == 64) { // MMX
2754 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2755 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2756 } else if (HasSSE2) { // SSE2
2757 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2758 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2760 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2761 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2763 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2766 /// getOnesVector - Returns a vector of specified type with all bits set.
2768 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2769 assert(VT.isVector() && "Expected a vector type");
2771 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2772 // type. This ensures they get CSE'd.
2773 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2775 if (VT.getSizeInBits() == 64) // MMX
2776 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2778 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2779 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2783 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2784 /// that point to V2 points to its first element.
2785 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
2786 MVT VT = SVOp->getValueType(0);
2787 int NumElems = VT.getVectorNumElements();
2788 const int *Mask = SVOp->getMask();
2790 bool Changed = false;
2791 SmallVector<int, 8> MaskVec;
2793 for (int i = 0; i != NumElems; ++i) {
2795 if (idx > NumElems) {
2799 MaskVec.push_back(idx);
2802 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
2803 SVOp->getOperand(1), &MaskVec[0]);
2804 return SDValue(SVOp, 0);
2807 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2808 /// operation of specified width.
2809 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2811 unsigned NumElems = VT.getVectorNumElements();
2812 SmallVector<int, 8> Mask;
2813 Mask.push_back(NumElems);
2814 for (unsigned i = 1; i != NumElems; ++i)
2816 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2819 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
2820 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2822 unsigned NumElems = VT.getVectorNumElements();
2823 SmallVector<int, 8> Mask;
2824 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2826 Mask.push_back(i + NumElems);
2828 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2831 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
2832 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2834 unsigned NumElems = VT.getVectorNumElements();
2835 unsigned Half = NumElems/2;
2836 SmallVector<int, 8> Mask;
2837 for (unsigned i = 0; i != Half; ++i) {
2838 Mask.push_back(i + Half);
2839 Mask.push_back(i + NumElems + Half);
2841 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2844 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2845 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
2847 if (SV->getValueType(0).getVectorNumElements() <= 4)
2848 return SDValue(SV, 0);
2850 MVT PVT = MVT::v4f32;
2851 MVT VT = SV->getValueType(0);
2852 DebugLoc dl = SV->getDebugLoc();
2853 SDValue V1 = SV->getOperand(0);
2854 int NumElems = VT.getVectorNumElements();
2855 int EltNo = SV->getSplatIndex();
2857 // unpack elements to the correct location
2858 while (NumElems > 4) {
2859 if (EltNo < NumElems/2) {
2860 V1 = getUnpackl(DAG, dl, VT, V1, V1);
2862 V1 = getUnpackh(DAG, dl, VT, V1, V1);
2863 EltNo -= NumElems/2;
2868 // Perform the splat.
2869 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
2870 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
2871 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
2872 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
2875 /// isVectorLoad - Returns true if the node is a vector load, a scalar
2876 /// load that's promoted to vector, or a load bitcasted.
2877 static bool isVectorLoad(SDValue Op) {
2878 assert(Op.getValueType().isVector() && "Expected a vector type");
2879 if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR ||
2880 Op.getOpcode() == ISD::BIT_CONVERT) {
2881 return isa<LoadSDNode>(Op.getOperand(0));
2883 return isa<LoadSDNode>(Op);
2887 /// CanonicalizeMovddup - Cannonicalize movddup shuffle to v2f64.
2889 static SDValue CanonicalizeMovddup(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
2891 // If we have sse3 and shuffle has more than one use or input is a load, then
2892 // use movddup. Otherwise, use movlhps.
2893 SDValue V1 = SV->getOperand(0);
2895 bool UseMovddup = HasSSE3 && (!SV->hasOneUse() || isVectorLoad(V1));
2896 MVT PVT = UseMovddup ? MVT::v2f64 : MVT::v4f32;
2897 MVT VT = SV->getValueType(0);
2899 return SDValue(SV, 0);
2901 DebugLoc dl = SV->getDebugLoc();
2902 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
2903 if (PVT.getVectorNumElements() == 2) {
2904 int Mask[2] = { 0, 0 };
2905 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), Mask);
2907 int Mask[4] = { 0, 1, 0, 1 };
2908 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), Mask);
2910 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
2913 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2914 /// vector of zero or undef vector. This produces a shuffle where the low
2915 /// element of V2 is swizzled into the zero/undef vector, landing at element
2916 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2917 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
2918 bool isZero, bool HasSSE2,
2919 SelectionDAG &DAG) {
2920 MVT VT = V2.getValueType();
2922 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
2923 unsigned NumElems = VT.getVectorNumElements();
2924 SmallVector<int, 16> MaskVec;
2925 for (unsigned i = 0; i != NumElems; ++i)
2926 // If this is the insertion idx, put the low elt of V2 here.
2927 MaskVec.push_back(i == Idx ? NumElems : i);
2928 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
2931 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
2932 /// a shuffle that is zero.
2934 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, const int *Mask,
2935 int NumElems, bool Low, SelectionDAG &DAG) {
2936 unsigned NumZeros = 0;
2937 for (int i = 0; i < NumElems; ++i) {
2938 unsigned Index = Low ? i : NumElems-i-1;
2939 int Idx = Mask[Index];
2944 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
2945 if (Elt.getNode() && isZeroNode(Elt))
2953 /// isVectorShift - Returns true if the shuffle can be implemented as a
2954 /// logical left or right shift of a vector.
2955 /// FIXME: split into pslldqi, psrldqi, palignr variants.
2956 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
2957 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
2958 const int *Mask = SVOp->getMask();
2959 int NumElems = SVOp->getValueType(0).getVectorNumElements();
2962 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, Mask, NumElems, true, DAG);
2965 NumZeros = getNumOfConsecutiveZeros(SVOp, Mask, NumElems, false, DAG);
2969 bool SeenV1 = false;
2970 bool SeenV2 = false;
2971 for (int i = NumZeros; i < NumElems; ++i) {
2972 int Val = isLeft ? (i - NumZeros) : i;
2973 int Idx = Mask[isLeft ? i : (i - NumZeros)];
2985 if (SeenV1 && SeenV2)
2988 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
2994 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2996 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
2997 unsigned NumNonZero, unsigned NumZero,
2998 SelectionDAG &DAG, TargetLowering &TLI) {
3002 DebugLoc dl = Op.getDebugLoc();
3005 for (unsigned i = 0; i < 16; ++i) {
3006 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3007 if (ThisIsNonZero && First) {
3009 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3011 V = DAG.getUNDEF(MVT::v8i16);
3016 SDValue ThisElt(0, 0), LastElt(0, 0);
3017 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3018 if (LastIsNonZero) {
3019 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3020 MVT::i16, Op.getOperand(i-1));
3022 if (ThisIsNonZero) {
3023 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3024 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3025 ThisElt, DAG.getConstant(8, MVT::i8));
3027 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3031 if (ThisElt.getNode())
3032 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3033 DAG.getIntPtrConstant(i/2));
3037 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3040 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3042 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3043 unsigned NumNonZero, unsigned NumZero,
3044 SelectionDAG &DAG, TargetLowering &TLI) {
3048 DebugLoc dl = Op.getDebugLoc();
3051 for (unsigned i = 0; i < 8; ++i) {
3052 bool isNonZero = (NonZeros & (1 << i)) != 0;
3056 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3058 V = DAG.getUNDEF(MVT::v8i16);
3061 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3062 MVT::v8i16, V, Op.getOperand(i),
3063 DAG.getIntPtrConstant(i));
3070 /// getVShift - Return a vector logical shift node.
3072 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3073 unsigned NumBits, SelectionDAG &DAG,
3074 const TargetLowering &TLI, DebugLoc dl) {
3075 bool isMMX = VT.getSizeInBits() == 64;
3076 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3077 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3078 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3079 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3080 DAG.getNode(Opc, dl, ShVT, SrcOp,
3081 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3085 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3086 DebugLoc dl = Op.getDebugLoc();
3087 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3088 if (ISD::isBuildVectorAllZeros(Op.getNode())
3089 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3090 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3091 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3092 // eliminated on x86-32 hosts.
3093 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3096 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3097 return getOnesVector(Op.getValueType(), DAG, dl);
3098 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3101 MVT VT = Op.getValueType();
3102 MVT EVT = VT.getVectorElementType();
3103 unsigned EVTBits = EVT.getSizeInBits();
3105 unsigned NumElems = Op.getNumOperands();
3106 unsigned NumZero = 0;
3107 unsigned NumNonZero = 0;
3108 unsigned NonZeros = 0;
3109 bool IsAllConstants = true;
3110 SmallSet<SDValue, 8> Values;
3111 for (unsigned i = 0; i < NumElems; ++i) {
3112 SDValue Elt = Op.getOperand(i);
3113 if (Elt.getOpcode() == ISD::UNDEF)
3116 if (Elt.getOpcode() != ISD::Constant &&
3117 Elt.getOpcode() != ISD::ConstantFP)
3118 IsAllConstants = false;
3119 if (isZeroNode(Elt))
3122 NonZeros |= (1 << i);
3127 if (NumNonZero == 0) {
3128 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3129 return DAG.getUNDEF(VT);
3132 // Special case for single non-zero, non-undef, element.
3133 if (NumNonZero == 1 && NumElems <= 4) {
3134 unsigned Idx = CountTrailingZeros_32(NonZeros);
3135 SDValue Item = Op.getOperand(Idx);
3137 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3138 // the value are obviously zero, truncate the value to i32 and do the
3139 // insertion that way. Only do this if the value is non-constant or if the
3140 // value is a constant being inserted into element 0. It is cheaper to do
3141 // a constant pool load than it is to do a movd + shuffle.
3142 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3143 (!IsAllConstants || Idx == 0)) {
3144 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3145 // Handle MMX and SSE both.
3146 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3147 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3149 // Truncate the value (which may itself be a constant) to i32, and
3150 // convert it to a vector with movd (S2V+shuffle to zero extend).
3151 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3152 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3153 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3154 Subtarget->hasSSE2(), DAG);
3156 // Now we have our 32-bit value zero extended in the low element of
3157 // a vector. If Idx != 0, swizzle it into place.
3159 SmallVector<int, 4> Mask;
3160 Mask.push_back(Idx);
3161 for (unsigned i = 1; i != VecElts; ++i)
3163 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3164 DAG.getUNDEF(Item.getValueType()),
3167 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3171 // If we have a constant or non-constant insertion into the low element of
3172 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3173 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3174 // depending on what the source datatype is. Because we can only get here
3175 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3177 // Don't do this for i64 values on x86-32.
3178 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3179 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3180 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3181 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3182 Subtarget->hasSSE2(), DAG);
3185 // Is it a vector logical left shift?
3186 if (NumElems == 2 && Idx == 1 &&
3187 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3188 unsigned NumBits = VT.getSizeInBits();
3189 return getVShift(true, VT,
3190 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3191 VT, Op.getOperand(1)),
3192 NumBits/2, DAG, *this, dl);
3195 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3198 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3199 // is a non-constant being inserted into an element other than the low one,
3200 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3201 // movd/movss) to move this into the low element, then shuffle it into
3203 if (EVTBits == 32) {
3204 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3206 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3207 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3208 Subtarget->hasSSE2(), DAG);
3209 SmallVector<int, 8> MaskVec;
3210 for (unsigned i = 0; i < NumElems; i++)
3211 MaskVec.push_back(i == Idx ? 0 : 1);
3212 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3216 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3217 if (Values.size() == 1)
3220 // A vector full of immediates; various special cases are already
3221 // handled, so this is best done with a single constant-pool load.
3225 // Let legalizer expand 2-wide build_vectors.
3226 if (EVTBits == 64) {
3227 if (NumNonZero == 1) {
3228 // One half is zero or undef.
3229 unsigned Idx = CountTrailingZeros_32(NonZeros);
3230 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3231 Op.getOperand(Idx));
3232 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3233 Subtarget->hasSSE2(), DAG);
3238 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3239 if (EVTBits == 8 && NumElems == 16) {
3240 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3242 if (V.getNode()) return V;
3245 if (EVTBits == 16 && NumElems == 8) {
3246 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3248 if (V.getNode()) return V;
3251 // If element VT is == 32 bits, turn it into a number of shuffles.
3252 SmallVector<SDValue, 8> V;
3254 if (NumElems == 4 && NumZero > 0) {
3255 for (unsigned i = 0; i < 4; ++i) {
3256 bool isZero = !(NonZeros & (1 << i));
3258 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3260 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3263 for (unsigned i = 0; i < 2; ++i) {
3264 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3267 V[i] = V[i*2]; // Must be a zero vector.
3270 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3273 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3276 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3281 SmallVector<int, 8> MaskVec;
3282 bool Reverse = (NonZeros & 0x3) == 2;
3283 for (unsigned i = 0; i < 2; ++i)
3284 MaskVec.push_back(Reverse ? 1-i : i);
3285 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3286 for (unsigned i = 0; i < 2; ++i)
3287 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3288 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3291 if (Values.size() > 2) {
3292 // If we have SSE 4.1, Expand into a number of inserts.
3293 if (getSubtarget()->hasSSE41()) {
3294 V[0] = DAG.getUNDEF(VT);
3295 for (unsigned i = 0; i < NumElems; ++i)
3296 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3297 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3298 Op.getOperand(i), DAG.getIntPtrConstant(i));
3301 // Expand into a number of unpckl*.
3303 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3304 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3305 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3306 for (unsigned i = 0; i < NumElems; ++i)
3307 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3309 while (NumElems != 0) {
3310 for (unsigned i = 0; i < NumElems; ++i)
3311 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3320 // v8i16 shuffles - Prefer shuffles in the following order:
3321 // 1. [all] pshuflw, pshufhw, optional move
3322 // 2. [ssse3] 1 x pshufb
3323 // 3. [ssse3] 2 x pshufb + 1 x por
3324 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3326 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3327 SelectionDAG &DAG, X86TargetLowering &TLI) {
3328 SDValue V1 = SVOp->getOperand(0);
3329 SDValue V2 = SVOp->getOperand(1);
3330 DebugLoc dl = SVOp->getDebugLoc();
3331 const int *Mask = SVOp->getMask();
3332 SmallVector<int, 8> MaskVals;
3334 // Determine if more than 1 of the words in each of the low and high quadwords
3335 // of the result come from the same quadword of one of the two inputs. Undef
3336 // mask values count as coming from any quadword, for better codegen.
3337 SmallVector<unsigned, 4> LoQuad(4);
3338 SmallVector<unsigned, 4> HiQuad(4);
3339 BitVector InputQuads(4);
3340 for (unsigned i = 0; i < 8; ++i) {
3341 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3342 int EltIdx = Mask[i];
3343 MaskVals.push_back(EltIdx);
3352 InputQuads.set(EltIdx / 4);
3355 int BestLoQuad = -1;
3356 unsigned MaxQuad = 1;
3357 for (unsigned i = 0; i < 4; ++i) {
3358 if (LoQuad[i] > MaxQuad) {
3360 MaxQuad = LoQuad[i];
3364 int BestHiQuad = -1;
3366 for (unsigned i = 0; i < 4; ++i) {
3367 if (HiQuad[i] > MaxQuad) {
3369 MaxQuad = HiQuad[i];
3373 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3374 // of the two input vectors, shuffle them into one input vector so only a
3375 // single pshufb instruction is necessary. If There are more than 2 input
3376 // quads, disable the next transformation since it does not help SSSE3.
3377 bool V1Used = InputQuads[0] || InputQuads[1];
3378 bool V2Used = InputQuads[2] || InputQuads[3];
3379 if (TLI.getSubtarget()->hasSSSE3()) {
3380 if (InputQuads.count() == 2 && V1Used && V2Used) {
3381 BestLoQuad = InputQuads.find_first();
3382 BestHiQuad = InputQuads.find_next(BestLoQuad);
3384 if (InputQuads.count() > 2) {
3390 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3391 // the shuffle mask. If a quad is scored as -1, that means that it contains
3392 // words from all 4 input quadwords.
3394 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3395 SmallVector<int, 8> MaskV;
3396 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3397 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3398 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3399 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3400 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3401 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3403 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3404 // source words for the shuffle, to aid later transformations.
3405 bool AllWordsInNewV = true;
3406 bool InOrder[2] = { true, true };
3407 for (unsigned i = 0; i != 8; ++i) {
3408 int idx = MaskVals[i];
3410 InOrder[i/4] = false;
3411 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3413 AllWordsInNewV = false;
3417 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3418 if (AllWordsInNewV) {
3419 for (int i = 0; i != 8; ++i) {
3420 int idx = MaskVals[i];
3423 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3424 if ((idx != i) && idx < 4)
3426 if ((idx != i) && idx > 3)
3435 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3436 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3437 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3438 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3439 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3443 // If we have SSSE3, and all words of the result are from 1 input vector,
3444 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3445 // is present, fall back to case 4.
3446 if (TLI.getSubtarget()->hasSSSE3()) {
3447 SmallVector<SDValue,16> pshufbMask;
3449 // If we have elements from both input vectors, set the high bit of the
3450 // shuffle mask element to zero out elements that come from V2 in the V1
3451 // mask, and elements that come from V1 in the V2 mask, so that the two
3452 // results can be OR'd together.
3453 bool TwoInputs = V1Used && V2Used;
3454 for (unsigned i = 0; i != 8; ++i) {
3455 int EltIdx = MaskVals[i] * 2;
3456 if (TwoInputs && (EltIdx >= 16)) {
3457 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3458 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3461 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3462 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3464 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3465 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3466 DAG.getNode(ISD::BUILD_VECTOR, dl,
3467 MVT::v16i8, &pshufbMask[0], 16));
3469 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3471 // Calculate the shuffle mask for the second input, shuffle it, and
3472 // OR it with the first shuffled input.
3474 for (unsigned i = 0; i != 8; ++i) {
3475 int EltIdx = MaskVals[i] * 2;
3477 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3478 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3481 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3482 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3484 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3485 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3486 DAG.getNode(ISD::BUILD_VECTOR, dl,
3487 MVT::v16i8, &pshufbMask[0], 16));
3488 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3489 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3492 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3493 // and update MaskVals with new element order.
3494 BitVector InOrder(8);
3495 if (BestLoQuad >= 0) {
3496 SmallVector<int, 8> MaskV;
3497 for (int i = 0; i != 4; ++i) {
3498 int idx = MaskVals[i];
3500 MaskV.push_back(-1);
3502 } else if ((idx / 4) == BestLoQuad) {
3503 MaskV.push_back(idx & 3);
3506 MaskV.push_back(-1);
3509 for (unsigned i = 4; i != 8; ++i)
3511 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3515 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3516 // and update MaskVals with the new element order.
3517 if (BestHiQuad >= 0) {
3518 SmallVector<int, 8> MaskV;
3519 for (unsigned i = 0; i != 4; ++i)
3521 for (unsigned i = 4; i != 8; ++i) {
3522 int idx = MaskVals[i];
3524 MaskV.push_back(-1);
3526 } else if ((idx / 4) == BestHiQuad) {
3527 MaskV.push_back((idx & 3) + 4);
3530 MaskV.push_back(-1);
3533 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3537 // In case BestHi & BestLo were both -1, which means each quadword has a word
3538 // from each of the four input quadwords, calculate the InOrder bitvector now
3539 // before falling through to the insert/extract cleanup.
3540 if (BestLoQuad == -1 && BestHiQuad == -1) {
3542 for (int i = 0; i != 8; ++i)
3543 if (MaskVals[i] < 0 || MaskVals[i] == i)
3547 // The other elements are put in the right place using pextrw and pinsrw.
3548 for (unsigned i = 0; i != 8; ++i) {
3551 int EltIdx = MaskVals[i];
3554 SDValue ExtOp = (EltIdx < 8)
3555 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3556 DAG.getIntPtrConstant(EltIdx))
3557 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3558 DAG.getIntPtrConstant(EltIdx - 8));
3559 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3560 DAG.getIntPtrConstant(i));
3565 // v16i8 shuffles - Prefer shuffles in the following order:
3566 // 1. [ssse3] 1 x pshufb
3567 // 2. [ssse3] 2 x pshufb + 1 x por
3568 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3570 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3571 SelectionDAG &DAG, X86TargetLowering &TLI) {
3572 SDValue V1 = SVOp->getOperand(0);
3573 SDValue V2 = SVOp->getOperand(1);
3574 DebugLoc dl = SVOp->getDebugLoc();
3575 const int *Mask = SVOp->getMask();
3576 SmallVector<int, 16> MaskVals;
3578 // If we have SSSE3, case 1 is generated when all result bytes come from
3579 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3580 // present, fall back to case 3.
3581 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3584 for (unsigned i = 0; i < 16; ++i) {
3585 int EltIdx = Mask[i];
3586 MaskVals.push_back(EltIdx);
3595 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3596 if (TLI.getSubtarget()->hasSSSE3()) {
3597 SmallVector<SDValue,16> pshufbMask;
3599 // If all result elements are from one input vector, then only translate
3600 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3602 // Otherwise, we have elements from both input vectors, and must zero out
3603 // elements that come from V2 in the first mask, and V1 in the second mask
3604 // so that we can OR them together.
3605 bool TwoInputs = !(V1Only || V2Only);
3606 for (unsigned i = 0; i != 16; ++i) {
3607 int EltIdx = MaskVals[i];
3608 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3609 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3612 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3614 // If all the elements are from V2, assign it to V1 and return after
3615 // building the first pshufb.
3618 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3619 DAG.getNode(ISD::BUILD_VECTOR, dl,
3620 MVT::v16i8, &pshufbMask[0], 16));
3624 // Calculate the shuffle mask for the second input, shuffle it, and
3625 // OR it with the first shuffled input.
3627 for (unsigned i = 0; i != 16; ++i) {
3628 int EltIdx = MaskVals[i];
3630 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3633 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3635 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3636 DAG.getNode(ISD::BUILD_VECTOR, dl,
3637 MVT::v16i8, &pshufbMask[0], 16));
3638 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3641 // No SSSE3 - Calculate in place words and then fix all out of place words
3642 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3643 // the 16 different words that comprise the two doublequadword input vectors.
3644 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3645 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3646 SDValue NewV = V2Only ? V2 : V1;
3647 for (int i = 0; i != 8; ++i) {
3648 int Elt0 = MaskVals[i*2];
3649 int Elt1 = MaskVals[i*2+1];
3651 // This word of the result is all undef, skip it.
3652 if (Elt0 < 0 && Elt1 < 0)
3655 // This word of the result is already in the correct place, skip it.
3656 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3658 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3661 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3662 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3665 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3666 // using a single extract together, load it and store it.
3667 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3668 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3669 DAG.getIntPtrConstant(Elt1 / 2));
3670 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3671 DAG.getIntPtrConstant(i));
3675 // If Elt1 is defined, extract it from the appropriate source. If the
3676 // source byte is not also odd, shift the extracted word left 8 bits
3677 // otherwise clear the bottom 8 bits if we need to do an or.
3679 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3680 DAG.getIntPtrConstant(Elt1 / 2));
3681 if ((Elt1 & 1) == 0)
3682 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3683 DAG.getConstant(8, TLI.getShiftAmountTy()));
3685 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3686 DAG.getConstant(0xFF00, MVT::i16));
3688 // If Elt0 is defined, extract it from the appropriate source. If the
3689 // source byte is not also even, shift the extracted word right 8 bits. If
3690 // Elt1 was also defined, OR the extracted values together before
3691 // inserting them in the result.
3693 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3694 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3695 if ((Elt0 & 1) != 0)
3696 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3697 DAG.getConstant(8, TLI.getShiftAmountTy()));
3699 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3700 DAG.getConstant(0x00FF, MVT::i16));
3701 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3704 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3705 DAG.getIntPtrConstant(i));
3707 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3710 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3711 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3712 /// done when every pair / quad of shuffle mask elements point to elements in
3713 /// the right sequence. e.g.
3714 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3716 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3718 TargetLowering &TLI, DebugLoc dl) {
3719 MVT VT = SVOp->getValueType(0);
3720 SDValue V1 = SVOp->getOperand(0);
3721 SDValue V2 = SVOp->getOperand(1);
3722 const int *PermMask = SVOp->getMask();
3723 unsigned NumElems = VT.getVectorNumElements();
3724 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3725 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3726 MVT MaskEltVT = MaskVT.getVectorElementType();
3728 switch (VT.getSimpleVT()) {
3729 default: assert(false && "Unexpected!");
3730 case MVT::v4f32: NewVT = MVT::v2f64; break;
3731 case MVT::v4i32: NewVT = MVT::v2i64; break;
3732 case MVT::v8i16: NewVT = MVT::v4i32; break;
3733 case MVT::v16i8: NewVT = MVT::v4i32; break;
3736 if (NewWidth == 2) {
3742 int Scale = NumElems / NewWidth;
3743 SmallVector<int, 8> MaskVec;
3744 for (unsigned i = 0; i < NumElems; i += Scale) {
3746 for (int j = 0; j < Scale; ++j) {
3747 int EltIdx = PermMask[i+j];
3751 StartIdx = EltIdx - (EltIdx % Scale);
3752 if (EltIdx != StartIdx + j)
3756 MaskVec.push_back(-1);
3758 MaskVec.push_back(StartIdx / Scale);
3761 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3762 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3763 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3766 /// getVZextMovL - Return a zero-extending vector move low node.
3768 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3769 SDValue SrcOp, SelectionDAG &DAG,
3770 const X86Subtarget *Subtarget, DebugLoc dl) {
3771 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3772 LoadSDNode *LD = NULL;
3773 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3774 LD = dyn_cast<LoadSDNode>(SrcOp);
3776 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3778 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3779 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3780 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3781 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3782 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3784 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3785 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3786 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3787 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3795 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3796 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3797 DAG.getNode(ISD::BIT_CONVERT, dl,
3801 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3804 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3805 SDValue V1 = SVOp->getOperand(0);
3806 SDValue V2 = SVOp->getOperand(1);
3807 DebugLoc dl = SVOp->getDebugLoc();
3808 MVT VT = SVOp->getValueType(0);
3809 const int *PermMaskPtr = SVOp->getMask();
3811 SmallVector<std::pair<int, int>, 8> Locs;
3813 SmallVector<int, 8> Mask1(4U, -1);
3814 SmallVector<int, 8> PermMask;
3816 for (unsigned i = 0; i != 8; ++i)
3817 PermMask.push_back(PermMaskPtr[i]);
3821 for (unsigned i = 0; i != 4; ++i) {
3822 int Idx = PermMask[i];
3824 Locs[i] = std::make_pair(-1, -1);
3826 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
3828 Locs[i] = std::make_pair(0, NumLo);
3832 Locs[i] = std::make_pair(1, NumHi);
3834 Mask1[2+NumHi] = Idx;
3840 if (NumLo <= 2 && NumHi <= 2) {
3841 // If no more than two elements come from either vector. This can be
3842 // implemented with two shuffles. First shuffle gather the elements.
3843 // The second shuffle, which takes the first shuffle as both of its
3844 // vector operands, put the elements into the right order.
3845 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3847 SmallVector<int, 8> Mask2(4U, -1);
3849 for (unsigned i = 0; i != 4; ++i) {
3850 if (Locs[i].first == -1)
3853 unsigned Idx = (i < 2) ? 0 : 4;
3854 Idx += Locs[i].first * 2 + Locs[i].second;
3859 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
3860 } else if (NumLo == 3 || NumHi == 3) {
3861 // Otherwise, we must have three elements from one vector, call it X, and
3862 // one element from the other, call it Y. First, use a shufps to build an
3863 // intermediate vector with the one element from Y and the element from X
3864 // that will be in the same half in the final destination (the indexes don't
3865 // matter). Then, use a shufps to build the final vector, taking the half
3866 // containing the element from Y from the intermediate, and the other half
3869 // Normalize it so the 3 elements come from V1.
3870 CommuteVectorShuffleMask(PermMask, VT);
3874 // Find the element from V2.
3876 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3877 int Val = PermMask[HiIndex];
3884 Mask1[0] = PermMask[HiIndex];
3886 Mask1[2] = PermMask[HiIndex^1];
3888 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3891 Mask1[0] = PermMask[0];
3892 Mask1[1] = PermMask[1];
3893 Mask1[2] = HiIndex & 1 ? 6 : 4;
3894 Mask1[3] = HiIndex & 1 ? 4 : 6;
3895 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3897 Mask1[0] = HiIndex & 1 ? 2 : 0;
3898 Mask1[1] = HiIndex & 1 ? 0 : 2;
3899 Mask1[2] = PermMask[2];
3900 Mask1[3] = PermMask[3];
3905 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
3909 // Break it into (shuffle shuffle_hi, shuffle_lo).
3911 SmallVector<int,8> LoMask(4U, -1);
3912 SmallVector<int,8> HiMask(4U, -1);
3914 SmallVector<int,8> *MaskPtr = &LoMask;
3915 unsigned MaskIdx = 0;
3918 for (unsigned i = 0; i != 4; ++i) {
3925 int Idx = PermMask[i];
3927 Locs[i] = std::make_pair(-1, -1);
3928 } else if (Idx < 4) {
3929 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3930 (*MaskPtr)[LoIdx] = Idx;
3933 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3934 (*MaskPtr)[HiIdx] = Idx;
3939 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
3940 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
3941 SmallVector<int, 8> MaskOps;
3942 for (unsigned i = 0; i != 4; ++i) {
3943 if (Locs[i].first == -1) {
3944 MaskOps.push_back(-1);
3946 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
3947 MaskOps.push_back(Idx);
3950 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
3954 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3955 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
3956 SDValue V1 = Op.getOperand(0);
3957 SDValue V2 = Op.getOperand(1);
3958 MVT VT = Op.getValueType();
3959 DebugLoc dl = Op.getDebugLoc();
3960 const int *PermMask = cast<ShuffleVectorSDNode>(Op.getNode())->getMask();
3961 unsigned NumElems = VT.getVectorNumElements();
3962 bool isMMX = VT.getSizeInBits() == 64;
3963 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3964 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3965 bool V1IsSplat = false;
3966 bool V2IsSplat = false;
3968 if (isZeroShuffle(SVOp))
3969 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3971 // Canonicalize movddup shuffles.
3972 if (V2IsUndef && Subtarget->hasSSE2() && VT.getSizeInBits() == 128 &&
3973 X86::isMOVDDUPMask(SVOp))
3974 return CanonicalizeMovddup(SVOp, DAG, Subtarget->hasSSE3());
3976 // Promote splats to v4f32.
3977 if (SVOp->isSplat()) {
3978 if (isMMX || NumElems < 4)
3980 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
3983 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3985 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3986 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
3987 if (NewOp.getNode())
3988 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3989 LowerVECTOR_SHUFFLE(NewOp, DAG));
3990 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3991 // FIXME: Figure out a cleaner way to do this.
3992 // Try to make use of movq to zero out the top part.
3993 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
3994 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
3995 if (NewOp.getNode()) {
3996 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
3997 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
3998 DAG, Subtarget, dl);
4000 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4001 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4002 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4003 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4004 DAG, Subtarget, dl);
4008 if (X86::isPSHUFDMask(SVOp))
4011 // Check if this can be converted into a logical shift.
4012 bool isLeft = false;
4015 bool isShift = getSubtarget()->hasSSE2() &&
4016 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4017 if (isShift && ShVal.hasOneUse()) {
4018 // If the shifted value has multiple uses, it may be cheaper to use
4019 // v_set0 + movlhps or movhlps, etc.
4020 MVT EVT = VT.getVectorElementType();
4021 ShAmt *= EVT.getSizeInBits();
4022 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4025 if (X86::isMOVLMask(SVOp)) {
4028 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4029 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4034 // FIXME: fold these into legal mask.
4035 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4036 X86::isMOVSLDUPMask(SVOp) ||
4037 X86::isMOVHLPSMask(SVOp) ||
4038 X86::isMOVHPMask(SVOp) ||
4039 X86::isMOVLPMask(SVOp)))
4042 if (ShouldXformToMOVHLPS(SVOp) ||
4043 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4044 return CommuteVectorShuffle(SVOp, DAG);
4047 // No better options. Use a vshl / vsrl.
4048 MVT EVT = VT.getVectorElementType();
4049 ShAmt *= EVT.getSizeInBits();
4050 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4053 bool Commuted = false;
4054 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4055 // 1,1,1,1 -> v8i16 though.
4056 V1IsSplat = isSplatVector(V1.getNode());
4057 V2IsSplat = isSplatVector(V2.getNode());
4059 // Canonicalize the splat or undef, if present, to be on the RHS.
4060 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4061 Op = CommuteVectorShuffle(SVOp, DAG);
4062 SVOp = cast<ShuffleVectorSDNode>(Op);
4063 V1 = SVOp->getOperand(0);
4064 V2 = SVOp->getOperand(1);
4065 std::swap(V1IsSplat, V2IsSplat);
4066 std::swap(V1IsUndef, V2IsUndef);
4070 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4071 // Shuffling low element of v1 into undef, just return v1.
4074 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4075 // the instruction selector will not match, so get a canonical MOVL with
4076 // swapped operands to undo the commute.
4077 return getMOVL(DAG, dl, VT, V2, V1);
4080 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4081 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4082 X86::isUNPCKLMask(SVOp) ||
4083 X86::isUNPCKHMask(SVOp))
4087 // Normalize mask so all entries that point to V2 points to its first
4088 // element then try to match unpck{h|l} again. If match, return a
4089 // new vector_shuffle with the corrected mask.
4090 SDValue NewMask = NormalizeMask(SVOp, DAG);
4091 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4092 if (NSVOp != SVOp) {
4093 if (X86::isUNPCKLMask(NSVOp, true)) {
4095 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4102 // Commute is back and try unpck* again.
4103 // FIXME: this seems wrong.
4104 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4105 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4106 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4107 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4108 X86::isUNPCKLMask(NewSVOp) ||
4109 X86::isUNPCKHMask(NewSVOp))
4113 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4115 // Normalize the node to match x86 shuffle ops if needed
4116 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4117 return CommuteVectorShuffle(SVOp, DAG);
4119 // Check for legal shuffle and return?
4120 if (isShuffleMaskLegal(PermMask, VT))
4123 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4124 if (VT == MVT::v8i16) {
4125 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4126 if (NewOp.getNode())
4130 if (VT == MVT::v16i8) {
4131 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4132 if (NewOp.getNode())
4136 // Handle all 4 wide cases with a number of shuffles except for MMX.
4137 if (NumElems == 4 && !isMMX)
4138 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4144 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4145 SelectionDAG &DAG) {
4146 MVT VT = Op.getValueType();
4147 DebugLoc dl = Op.getDebugLoc();
4148 if (VT.getSizeInBits() == 8) {
4149 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4150 Op.getOperand(0), Op.getOperand(1));
4151 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4152 DAG.getValueType(VT));
4153 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4154 } else if (VT.getSizeInBits() == 16) {
4155 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4156 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4158 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4159 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4160 DAG.getNode(ISD::BIT_CONVERT, dl,
4164 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4165 Op.getOperand(0), Op.getOperand(1));
4166 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4167 DAG.getValueType(VT));
4168 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4169 } else if (VT == MVT::f32) {
4170 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4171 // the result back to FR32 register. It's only worth matching if the
4172 // result has a single use which is a store or a bitcast to i32. And in
4173 // the case of a store, it's not worth it if the index is a constant 0,
4174 // because a MOVSSmr can be used instead, which is smaller and faster.
4175 if (!Op.hasOneUse())
4177 SDNode *User = *Op.getNode()->use_begin();
4178 if ((User->getOpcode() != ISD::STORE ||
4179 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4180 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4181 (User->getOpcode() != ISD::BIT_CONVERT ||
4182 User->getValueType(0) != MVT::i32))
4184 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4185 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4188 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4189 } else if (VT == MVT::i32) {
4190 // ExtractPS works with constant index.
4191 if (isa<ConstantSDNode>(Op.getOperand(1)))
4199 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4200 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4203 if (Subtarget->hasSSE41()) {
4204 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4209 MVT VT = Op.getValueType();
4210 DebugLoc dl = Op.getDebugLoc();
4211 // TODO: handle v16i8.
4212 if (VT.getSizeInBits() == 16) {
4213 SDValue Vec = Op.getOperand(0);
4214 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4216 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4217 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4218 DAG.getNode(ISD::BIT_CONVERT, dl,
4221 // Transform it so it match pextrw which produces a 32-bit result.
4222 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4223 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4224 Op.getOperand(0), Op.getOperand(1));
4225 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4226 DAG.getValueType(VT));
4227 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4228 } else if (VT.getSizeInBits() == 32) {
4229 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4233 // SHUFPS the element to the lowest double word, then movss.
4234 int Mask[4] = { Idx, -1, -1, -1 };
4235 MVT VVT = Op.getOperand(0).getValueType();
4236 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4237 DAG.getUNDEF(VVT), Mask);
4238 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4239 DAG.getIntPtrConstant(0));
4240 } else if (VT.getSizeInBits() == 64) {
4241 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4242 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4243 // to match extract_elt for f64.
4244 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4248 // UNPCKHPD the element to the lowest double word, then movsd.
4249 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4250 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4251 int Mask[2] = { 1, -1 };
4252 MVT VVT = Op.getOperand(0).getValueType();
4253 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4254 DAG.getUNDEF(VVT), Mask);
4255 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4256 DAG.getIntPtrConstant(0));
4263 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4264 MVT VT = Op.getValueType();
4265 MVT EVT = VT.getVectorElementType();
4266 DebugLoc dl = Op.getDebugLoc();
4268 SDValue N0 = Op.getOperand(0);
4269 SDValue N1 = Op.getOperand(1);
4270 SDValue N2 = Op.getOperand(2);
4272 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4273 isa<ConstantSDNode>(N2)) {
4274 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4276 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4278 if (N1.getValueType() != MVT::i32)
4279 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4280 if (N2.getValueType() != MVT::i32)
4281 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4282 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4283 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4284 // Bits [7:6] of the constant are the source select. This will always be
4285 // zero here. The DAG Combiner may combine an extract_elt index into these
4286 // bits. For example (insert (extract, 3), 2) could be matched by putting
4287 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4288 // Bits [5:4] of the constant are the destination select. This is the
4289 // value of the incoming immediate.
4290 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4291 // combine either bitwise AND or insert of float 0.0 to set these bits.
4292 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4293 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4294 } else if (EVT == MVT::i32) {
4295 // InsertPS works with constant index.
4296 if (isa<ConstantSDNode>(N2))
4303 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4304 MVT VT = Op.getValueType();
4305 MVT EVT = VT.getVectorElementType();
4307 if (Subtarget->hasSSE41())
4308 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4313 DebugLoc dl = Op.getDebugLoc();
4314 SDValue N0 = Op.getOperand(0);
4315 SDValue N1 = Op.getOperand(1);
4316 SDValue N2 = Op.getOperand(2);
4318 if (EVT.getSizeInBits() == 16) {
4319 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4320 // as its second argument.
4321 if (N1.getValueType() != MVT::i32)
4322 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4323 if (N2.getValueType() != MVT::i32)
4324 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4325 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4331 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4332 DebugLoc dl = Op.getDebugLoc();
4333 if (Op.getValueType() == MVT::v2f32)
4334 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4335 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4336 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4337 Op.getOperand(0))));
4339 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4340 MVT VT = MVT::v2i32;
4341 switch (Op.getValueType().getSimpleVT()) {
4348 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4349 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4352 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4353 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4354 // one of the above mentioned nodes. It has to be wrapped because otherwise
4355 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4356 // be used to form addressing mode. These wrapped nodes will be selected
4359 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4360 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4361 // FIXME there isn't really any debug info here, should come from the parent
4362 DebugLoc dl = CP->getDebugLoc();
4363 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4364 CP->getAlignment());
4365 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4366 // With PIC, the address is actually $g + Offset.
4367 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4368 !Subtarget->isPICStyleRIPRel()) {
4369 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4370 DAG.getNode(X86ISD::GlobalBaseReg,
4371 DebugLoc::getUnknownLoc(),
4380 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4382 SelectionDAG &DAG) const {
4383 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4384 bool ExtraLoadRequired =
4385 Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false);
4387 // Create the TargetGlobalAddress node, folding in the constant
4388 // offset if it is legal.
4390 if (!IsPic && !ExtraLoadRequired && isInt32(Offset)) {
4391 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4394 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0);
4395 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4397 // With PIC, the address is actually $g + Offset.
4398 if (IsPic && !Subtarget->isPICStyleRIPRel()) {
4399 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4400 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4404 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4405 // load the value at address GV, not the value of GV itself. This means that
4406 // the GlobalAddress must be in the base or index register of the address, not
4407 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4408 // The same applies for external symbols during PIC codegen
4409 if (ExtraLoadRequired)
4410 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4411 PseudoSourceValue::getGOT(), 0);
4413 // If there was a non-zero offset that we didn't fold, create an explicit
4416 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4417 DAG.getConstant(Offset, getPointerTy()));
4423 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4424 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4425 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4426 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4430 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4432 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4433 DebugLoc dl = GA->getDebugLoc();
4434 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4435 GA->getValueType(0),
4438 SDValue Ops[] = { Chain, TGA, *InFlag };
4439 return DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4441 SDValue Ops[] = { Chain, TGA };
4442 return DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4446 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4448 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4451 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4452 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4453 DAG.getNode(X86ISD::GlobalBaseReg,
4454 DebugLoc::getUnknownLoc(),
4456 InFlag = Chain.getValue(1);
4458 Chain = GetTLSADDR(DAG, Chain, GA, &InFlag);
4459 InFlag = Chain.getValue(1);
4461 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4462 SDValue Ops1[] = { Chain,
4463 DAG.getTargetExternalSymbol("___tls_get_addr",
4465 DAG.getRegister(X86::EAX, PtrVT),
4466 DAG.getRegister(X86::EBX, PtrVT),
4468 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops1, 5);
4469 InFlag = Chain.getValue(1);
4471 return DAG.getCopyFromReg(Chain, dl, X86::EAX, PtrVT, InFlag);
4474 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4476 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4478 SDValue InFlag, Chain;
4479 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4481 Chain = GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL);
4482 InFlag = Chain.getValue(1);
4484 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4485 SDValue Ops1[] = { Chain,
4486 DAG.getTargetExternalSymbol("__tls_get_addr",
4488 DAG.getRegister(X86::RDI, PtrVT),
4490 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops1, 4);
4491 InFlag = Chain.getValue(1);
4493 return DAG.getCopyFromReg(Chain, dl, X86::RAX, PtrVT, InFlag);
4496 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4497 // "local exec" model.
4498 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4499 const MVT PtrVT, TLSModel::Model model,
4501 DebugLoc dl = GA->getDebugLoc();
4502 // Get the Thread Pointer
4503 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4504 DebugLoc::getUnknownLoc(), PtrVT,
4505 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4508 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4511 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4513 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4514 GA->getValueType(0),
4516 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
4518 if (model == TLSModel::InitialExec)
4519 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4520 PseudoSourceValue::getGOT(), 0);
4522 // The address of the thread local variable is the add of the thread
4523 // pointer with the offset of the variable.
4524 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4528 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4529 // TODO: implement the "local dynamic" model
4530 // TODO: implement the "initial exec"model for pic executables
4531 assert(Subtarget->isTargetELF() &&
4532 "TLS not implemented for non-ELF targets");
4533 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4534 GlobalValue *GV = GA->getGlobal();
4535 TLSModel::Model model =
4536 getTLSModel (GV, getTargetMachine().getRelocationModel());
4537 if (Subtarget->is64Bit()) {
4539 case TLSModel::GeneralDynamic:
4540 case TLSModel::LocalDynamic: // not implemented
4541 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4543 case TLSModel::InitialExec:
4544 case TLSModel::LocalExec:
4545 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model, true);
4549 case TLSModel::GeneralDynamic:
4550 case TLSModel::LocalDynamic: // not implemented
4551 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4553 case TLSModel::InitialExec:
4554 case TLSModel::LocalExec:
4555 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model, false);
4558 assert(0 && "Unreachable");
4563 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4564 // FIXME there isn't really any debug info here
4565 DebugLoc dl = Op.getDebugLoc();
4566 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4567 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4568 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4569 // With PIC, the address is actually $g + Offset.
4570 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4571 !Subtarget->isPICStyleRIPRel()) {
4572 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4573 DAG.getNode(X86ISD::GlobalBaseReg,
4574 DebugLoc::getUnknownLoc(),
4582 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4583 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4584 // FIXME there isn't really any debug into here
4585 DebugLoc dl = JT->getDebugLoc();
4586 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4587 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4588 // With PIC, the address is actually $g + Offset.
4589 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4590 !Subtarget->isPICStyleRIPRel()) {
4591 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4592 DAG.getNode(X86ISD::GlobalBaseReg,
4593 DebugLoc::getUnknownLoc(),
4601 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4602 /// take a 2 x i32 value to shift plus a shift amount.
4603 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4604 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4605 MVT VT = Op.getValueType();
4606 unsigned VTBits = VT.getSizeInBits();
4607 DebugLoc dl = Op.getDebugLoc();
4608 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4609 SDValue ShOpLo = Op.getOperand(0);
4610 SDValue ShOpHi = Op.getOperand(1);
4611 SDValue ShAmt = Op.getOperand(2);
4612 SDValue Tmp1 = isSRA ?
4613 DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4614 DAG.getConstant(VTBits - 1, MVT::i8)) :
4615 DAG.getConstant(0, VT);
4618 if (Op.getOpcode() == ISD::SHL_PARTS) {
4619 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4620 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4622 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4623 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4626 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4627 DAG.getConstant(VTBits, MVT::i8));
4628 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4629 AndNode, DAG.getConstant(0, MVT::i8));
4632 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4633 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4634 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4636 if (Op.getOpcode() == ISD::SHL_PARTS) {
4637 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4638 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4640 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4641 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4644 SDValue Ops[2] = { Lo, Hi };
4645 return DAG.getMergeValues(Ops, 2, dl);
4648 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4649 MVT SrcVT = Op.getOperand(0).getValueType();
4650 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4651 "Unknown SINT_TO_FP to lower!");
4653 // These are really Legal; caller falls through into that case.
4654 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4656 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4657 Subtarget->is64Bit())
4660 DebugLoc dl = Op.getDebugLoc();
4661 unsigned Size = SrcVT.getSizeInBits()/8;
4662 MachineFunction &MF = DAG.getMachineFunction();
4663 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4664 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4665 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4667 PseudoSourceValue::getFixedStack(SSFI), 0);
4671 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4673 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4675 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4676 SmallVector<SDValue, 8> Ops;
4677 Ops.push_back(Chain);
4678 Ops.push_back(StackSlot);
4679 Ops.push_back(DAG.getValueType(SrcVT));
4680 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4681 Tys, &Ops[0], Ops.size());
4684 Chain = Result.getValue(1);
4685 SDValue InFlag = Result.getValue(2);
4687 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4688 // shouldn't be necessary except that RFP cannot be live across
4689 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4690 MachineFunction &MF = DAG.getMachineFunction();
4691 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4692 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4693 Tys = DAG.getVTList(MVT::Other);
4694 SmallVector<SDValue, 8> Ops;
4695 Ops.push_back(Chain);
4696 Ops.push_back(Result);
4697 Ops.push_back(StackSlot);
4698 Ops.push_back(DAG.getValueType(Op.getValueType()));
4699 Ops.push_back(InFlag);
4700 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4701 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4702 PseudoSourceValue::getFixedStack(SSFI), 0);
4708 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4709 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4710 // This algorithm is not obvious. Here it is in C code, more or less:
4712 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4713 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4714 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4716 // Copy ints to xmm registers.
4717 __m128i xh = _mm_cvtsi32_si128( hi );
4718 __m128i xl = _mm_cvtsi32_si128( lo );
4720 // Combine into low half of a single xmm register.
4721 __m128i x = _mm_unpacklo_epi32( xh, xl );
4725 // Merge in appropriate exponents to give the integer bits the right
4727 x = _mm_unpacklo_epi32( x, exp );
4729 // Subtract away the biases to deal with the IEEE-754 double precision
4731 d = _mm_sub_pd( (__m128d) x, bias );
4733 // All conversions up to here are exact. The correctly rounded result is
4734 // calculated using the current rounding mode using the following
4736 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4737 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4738 // store doesn't really need to be here (except
4739 // maybe to zero the other double)
4744 DebugLoc dl = Op.getDebugLoc();
4746 // Build some magic constants.
4747 std::vector<Constant*> CV0;
4748 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
4749 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
4750 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4751 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4752 Constant *C0 = ConstantVector::get(CV0);
4753 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
4755 std::vector<Constant*> CV1;
4756 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
4757 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
4758 Constant *C1 = ConstantVector::get(CV1);
4759 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
4761 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4762 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4764 DAG.getIntPtrConstant(1)));
4765 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4766 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4768 DAG.getIntPtrConstant(0)));
4769 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
4770 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4771 PseudoSourceValue::getConstantPool(), 0,
4773 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
4774 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
4775 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
4776 PseudoSourceValue::getConstantPool(), 0,
4778 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
4780 // Add the halves; easiest way is to swap them into another reg first.
4781 int ShufMask[2] = { 1, -1 };
4782 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
4783 DAG.getUNDEF(MVT::v2f64), ShufMask);
4784 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
4785 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
4786 DAG.getIntPtrConstant(0));
4789 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
4790 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
4791 DebugLoc dl = Op.getDebugLoc();
4792 // FP constant to bias correct the final result.
4793 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
4796 // Load the 32-bit value into an XMM register.
4797 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4798 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4800 DAG.getIntPtrConstant(0)));
4802 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4803 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
4804 DAG.getIntPtrConstant(0));
4806 // Or the load with the bias.
4807 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
4808 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4809 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4811 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4812 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4813 MVT::v2f64, Bias)));
4814 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4815 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
4816 DAG.getIntPtrConstant(0));
4818 // Subtract the bias.
4819 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
4821 // Handle final rounding.
4822 MVT DestVT = Op.getValueType();
4824 if (DestVT.bitsLT(MVT::f64)) {
4825 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
4826 DAG.getIntPtrConstant(0));
4827 } else if (DestVT.bitsGT(MVT::f64)) {
4828 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
4831 // Handle final rounding.
4835 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4836 SDValue N0 = Op.getOperand(0);
4837 DebugLoc dl = Op.getDebugLoc();
4839 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
4840 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
4841 // the optimization here.
4842 if (DAG.SignBitIsZero(N0))
4843 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
4845 MVT SrcVT = N0.getValueType();
4846 if (SrcVT == MVT::i64) {
4847 // We only handle SSE2 f64 target here; caller can handle the rest.
4848 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
4851 return LowerUINT_TO_FP_i64(Op, DAG);
4852 } else if (SrcVT == MVT::i32) {
4853 return LowerUINT_TO_FP_i32(Op, DAG);
4856 assert(0 && "Unknown UINT_TO_FP to lower!");
4860 std::pair<SDValue,SDValue> X86TargetLowering::
4861 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
4862 DebugLoc dl = Op.getDebugLoc();
4863 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
4864 Op.getValueType().getSimpleVT() >= MVT::i16 &&
4865 "Unknown FP_TO_SINT to lower!");
4867 // These are really Legal.
4868 if (Op.getValueType() == MVT::i32 &&
4869 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4870 return std::make_pair(SDValue(), SDValue());
4871 if (Subtarget->is64Bit() &&
4872 Op.getValueType() == MVT::i64 &&
4873 Op.getOperand(0).getValueType() != MVT::f80)
4874 return std::make_pair(SDValue(), SDValue());
4876 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4878 MachineFunction &MF = DAG.getMachineFunction();
4879 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
4880 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4881 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4883 switch (Op.getValueType().getSimpleVT()) {
4884 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4885 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4886 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4887 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4890 SDValue Chain = DAG.getEntryNode();
4891 SDValue Value = Op.getOperand(0);
4892 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4893 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4894 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
4895 PseudoSourceValue::getFixedStack(SSFI), 0);
4896 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4898 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4900 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
4901 Chain = Value.getValue(1);
4902 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4903 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4906 // Build the FP_TO_INT*_IN_MEM
4907 SDValue Ops[] = { Chain, Value, StackSlot };
4908 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
4910 return std::make_pair(FIST, StackSlot);
4913 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
4914 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
4915 SDValue FIST = Vals.first, StackSlot = Vals.second;
4916 if (FIST.getNode() == 0) return SDValue();
4919 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
4920 FIST, StackSlot, NULL, 0);
4923 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
4924 DebugLoc dl = Op.getDebugLoc();
4925 MVT VT = Op.getValueType();
4928 EltVT = VT.getVectorElementType();
4929 std::vector<Constant*> CV;
4930 if (EltVT == MVT::f64) {
4931 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4935 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4941 Constant *C = ConstantVector::get(CV);
4942 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
4943 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
4944 PseudoSourceValue::getConstantPool(), 0,
4946 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
4949 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
4950 DebugLoc dl = Op.getDebugLoc();
4951 MVT VT = Op.getValueType();
4953 unsigned EltNum = 1;
4954 if (VT.isVector()) {
4955 EltVT = VT.getVectorElementType();
4956 EltNum = VT.getVectorNumElements();
4958 std::vector<Constant*> CV;
4959 if (EltVT == MVT::f64) {
4960 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
4964 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
4970 Constant *C = ConstantVector::get(CV);
4971 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
4972 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
4973 PseudoSourceValue::getConstantPool(), 0,
4975 if (VT.isVector()) {
4976 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4977 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
4978 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4980 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
4982 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
4986 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
4987 SDValue Op0 = Op.getOperand(0);
4988 SDValue Op1 = Op.getOperand(1);
4989 DebugLoc dl = Op.getDebugLoc();
4990 MVT VT = Op.getValueType();
4991 MVT SrcVT = Op1.getValueType();
4993 // If second operand is smaller, extend it first.
4994 if (SrcVT.bitsLT(VT)) {
4995 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
4998 // And if it is bigger, shrink it first.
4999 if (SrcVT.bitsGT(VT)) {
5000 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5004 // At this point the operands and the result should have the same
5005 // type, and that won't be f80 since that is not custom lowered.
5007 // First get the sign bit of second operand.
5008 std::vector<Constant*> CV;
5009 if (SrcVT == MVT::f64) {
5010 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
5011 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5013 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
5014 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5015 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5016 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5018 Constant *C = ConstantVector::get(CV);
5019 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5020 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5021 PseudoSourceValue::getConstantPool(), 0,
5023 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5025 // Shift sign bit right or left if the two operands have different types.
5026 if (SrcVT.bitsGT(VT)) {
5027 // Op0 is MVT::f32, Op1 is MVT::f64.
5028 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5029 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5030 DAG.getConstant(32, MVT::i32));
5031 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5032 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5033 DAG.getIntPtrConstant(0));
5036 // Clear first operand sign bit.
5038 if (VT == MVT::f64) {
5039 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5040 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5042 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5043 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5044 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5045 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5047 C = ConstantVector::get(CV);
5048 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5049 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5050 PseudoSourceValue::getConstantPool(), 0,
5052 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5054 // Or the value with the sign bit.
5055 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5058 /// Emit nodes that will be selected as "test Op0,Op0", or something
5060 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5061 SelectionDAG &DAG) {
5062 DebugLoc dl = Op.getDebugLoc();
5064 // CF and OF aren't always set the way we want. Determine which
5065 // of these we need.
5066 bool NeedCF = false;
5067 bool NeedOF = false;
5069 case X86::COND_A: case X86::COND_AE:
5070 case X86::COND_B: case X86::COND_BE:
5073 case X86::COND_G: case X86::COND_GE:
5074 case X86::COND_L: case X86::COND_LE:
5075 case X86::COND_O: case X86::COND_NO:
5081 // See if we can use the EFLAGS value from the operand instead of
5082 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5083 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5084 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5085 unsigned Opcode = 0;
5086 unsigned NumOperands = 0;
5087 switch (Op.getNode()->getOpcode()) {
5089 // Due to an isel shortcoming, be conservative if this add is likely to
5090 // be selected as part of a load-modify-store instruction. When the root
5091 // node in a match is a store, isel doesn't know how to remap non-chain
5092 // non-flag uses of other nodes in the match, such as the ADD in this
5093 // case. This leads to the ADD being left around and reselected, with
5094 // the result being two adds in the output.
5095 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5096 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5097 if (UI->getOpcode() == ISD::STORE)
5099 if (ConstantSDNode *C =
5100 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5101 // An add of one will be selected as an INC.
5102 if (C->getAPIntValue() == 1) {
5103 Opcode = X86ISD::INC;
5107 // An add of negative one (subtract of one) will be selected as a DEC.
5108 if (C->getAPIntValue().isAllOnesValue()) {
5109 Opcode = X86ISD::DEC;
5114 // Otherwise use a regular EFLAGS-setting add.
5115 Opcode = X86ISD::ADD;
5119 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5120 // likely to be selected as part of a load-modify-store instruction.
5121 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5122 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5123 if (UI->getOpcode() == ISD::STORE)
5125 // Otherwise use a regular EFLAGS-setting sub.
5126 Opcode = X86ISD::SUB;
5133 return SDValue(Op.getNode(), 1);
5139 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5140 SmallVector<SDValue, 4> Ops;
5141 for (unsigned i = 0; i != NumOperands; ++i)
5142 Ops.push_back(Op.getOperand(i));
5143 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5144 DAG.ReplaceAllUsesWith(Op, New);
5145 return SDValue(New.getNode(), 1);
5149 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5150 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5151 DAG.getConstant(0, Op.getValueType()));
5154 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5156 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5157 SelectionDAG &DAG) {
5158 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5159 if (C->getAPIntValue() == 0)
5160 return EmitTest(Op0, X86CC, DAG);
5162 DebugLoc dl = Op0.getDebugLoc();
5163 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5166 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5167 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5168 SDValue Op0 = Op.getOperand(0);
5169 SDValue Op1 = Op.getOperand(1);
5170 DebugLoc dl = Op.getDebugLoc();
5171 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5173 // Lower (X & (1 << N)) == 0 to BT(X, N).
5174 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5175 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5176 if (Op0.getOpcode() == ISD::AND &&
5178 Op1.getOpcode() == ISD::Constant &&
5179 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5180 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5182 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5183 if (ConstantSDNode *Op010C =
5184 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5185 if (Op010C->getZExtValue() == 1) {
5186 LHS = Op0.getOperand(0);
5187 RHS = Op0.getOperand(1).getOperand(1);
5189 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5190 if (ConstantSDNode *Op000C =
5191 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5192 if (Op000C->getZExtValue() == 1) {
5193 LHS = Op0.getOperand(1);
5194 RHS = Op0.getOperand(0).getOperand(1);
5196 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5197 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5198 SDValue AndLHS = Op0.getOperand(0);
5199 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5200 LHS = AndLHS.getOperand(0);
5201 RHS = AndLHS.getOperand(1);
5205 if (LHS.getNode()) {
5206 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5207 // instruction. Since the shift amount is in-range-or-undefined, we know
5208 // that doing a bittest on the i16 value is ok. We extend to i32 because
5209 // the encoding for the i16 version is larger than the i32 version.
5210 if (LHS.getValueType() == MVT::i8)
5211 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5213 // If the operand types disagree, extend the shift amount to match. Since
5214 // BT ignores high bits (like shifts) we can use anyextend.
5215 if (LHS.getValueType() != RHS.getValueType())
5216 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5218 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5219 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5220 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5221 DAG.getConstant(Cond, MVT::i8), BT);
5225 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5226 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5228 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5229 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5230 DAG.getConstant(X86CC, MVT::i8), Cond);
5233 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5235 SDValue Op0 = Op.getOperand(0);
5236 SDValue Op1 = Op.getOperand(1);
5237 SDValue CC = Op.getOperand(2);
5238 MVT VT = Op.getValueType();
5239 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5240 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5241 DebugLoc dl = Op.getDebugLoc();
5245 MVT VT0 = Op0.getValueType();
5246 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5247 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5250 switch (SetCCOpcode) {
5253 case ISD::SETEQ: SSECC = 0; break;
5255 case ISD::SETGT: Swap = true; // Fallthrough
5257 case ISD::SETOLT: SSECC = 1; break;
5259 case ISD::SETGE: Swap = true; // Fallthrough
5261 case ISD::SETOLE: SSECC = 2; break;
5262 case ISD::SETUO: SSECC = 3; break;
5264 case ISD::SETNE: SSECC = 4; break;
5265 case ISD::SETULE: Swap = true;
5266 case ISD::SETUGE: SSECC = 5; break;
5267 case ISD::SETULT: Swap = true;
5268 case ISD::SETUGT: SSECC = 6; break;
5269 case ISD::SETO: SSECC = 7; break;
5272 std::swap(Op0, Op1);
5274 // In the two special cases we can't handle, emit two comparisons.
5276 if (SetCCOpcode == ISD::SETUEQ) {
5278 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5279 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5280 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5282 else if (SetCCOpcode == ISD::SETONE) {
5284 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5285 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5286 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5288 assert(0 && "Illegal FP comparison");
5290 // Handle all other FP comparisons here.
5291 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5294 // We are handling one of the integer comparisons here. Since SSE only has
5295 // GT and EQ comparisons for integer, swapping operands and multiple
5296 // operations may be required for some comparisons.
5297 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5298 bool Swap = false, Invert = false, FlipSigns = false;
5300 switch (VT.getSimpleVT()) {
5302 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5303 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5304 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5305 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5308 switch (SetCCOpcode) {
5310 case ISD::SETNE: Invert = true;
5311 case ISD::SETEQ: Opc = EQOpc; break;
5312 case ISD::SETLT: Swap = true;
5313 case ISD::SETGT: Opc = GTOpc; break;
5314 case ISD::SETGE: Swap = true;
5315 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5316 case ISD::SETULT: Swap = true;
5317 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5318 case ISD::SETUGE: Swap = true;
5319 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5322 std::swap(Op0, Op1);
5324 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5325 // bits of the inputs before performing those operations.
5327 MVT EltVT = VT.getVectorElementType();
5328 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5330 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5331 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5333 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5334 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5337 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5339 // If the logical-not of the result is required, perform that now.
5341 Result = DAG.getNOT(dl, Result, VT);
5346 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5347 static bool isX86LogicalCmp(SDValue Op) {
5348 unsigned Opc = Op.getNode()->getOpcode();
5349 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5351 if (Op.getResNo() == 1 &&
5352 (Opc == X86ISD::ADD ||
5353 Opc == X86ISD::SUB ||
5354 Opc == X86ISD::SMUL ||
5355 Opc == X86ISD::UMUL ||
5356 Opc == X86ISD::INC ||
5357 Opc == X86ISD::DEC))
5363 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5364 bool addTest = true;
5365 SDValue Cond = Op.getOperand(0);
5366 DebugLoc dl = Op.getDebugLoc();
5369 if (Cond.getOpcode() == ISD::SETCC)
5370 Cond = LowerSETCC(Cond, DAG);
5372 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5373 // setting operand in place of the X86ISD::SETCC.
5374 if (Cond.getOpcode() == X86ISD::SETCC) {
5375 CC = Cond.getOperand(0);
5377 SDValue Cmp = Cond.getOperand(1);
5378 unsigned Opc = Cmp.getOpcode();
5379 MVT VT = Op.getValueType();
5381 bool IllegalFPCMov = false;
5382 if (VT.isFloatingPoint() && !VT.isVector() &&
5383 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5384 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5386 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5387 Opc == X86ISD::BT) { // FIXME
5394 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5395 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5398 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5399 SmallVector<SDValue, 4> Ops;
5400 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5401 // condition is true.
5402 Ops.push_back(Op.getOperand(2));
5403 Ops.push_back(Op.getOperand(1));
5405 Ops.push_back(Cond);
5406 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5409 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5410 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5411 // from the AND / OR.
5412 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5413 Opc = Op.getOpcode();
5414 if (Opc != ISD::OR && Opc != ISD::AND)
5416 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5417 Op.getOperand(0).hasOneUse() &&
5418 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5419 Op.getOperand(1).hasOneUse());
5422 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5423 // 1 and that the SETCC node has a single use.
5424 static bool isXor1OfSetCC(SDValue Op) {
5425 if (Op.getOpcode() != ISD::XOR)
5427 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5428 if (N1C && N1C->getAPIntValue() == 1) {
5429 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5430 Op.getOperand(0).hasOneUse();
5435 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5436 bool addTest = true;
5437 SDValue Chain = Op.getOperand(0);
5438 SDValue Cond = Op.getOperand(1);
5439 SDValue Dest = Op.getOperand(2);
5440 DebugLoc dl = Op.getDebugLoc();
5443 if (Cond.getOpcode() == ISD::SETCC)
5444 Cond = LowerSETCC(Cond, DAG);
5446 // FIXME: LowerXALUO doesn't handle these!!
5447 else if (Cond.getOpcode() == X86ISD::ADD ||
5448 Cond.getOpcode() == X86ISD::SUB ||
5449 Cond.getOpcode() == X86ISD::SMUL ||
5450 Cond.getOpcode() == X86ISD::UMUL)
5451 Cond = LowerXALUO(Cond, DAG);
5454 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5455 // setting operand in place of the X86ISD::SETCC.
5456 if (Cond.getOpcode() == X86ISD::SETCC) {
5457 CC = Cond.getOperand(0);
5459 SDValue Cmp = Cond.getOperand(1);
5460 unsigned Opc = Cmp.getOpcode();
5461 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5462 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5466 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5470 // These can only come from an arithmetic instruction with overflow,
5471 // e.g. SADDO, UADDO.
5472 Cond = Cond.getNode()->getOperand(1);
5479 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5480 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5481 if (CondOpc == ISD::OR) {
5482 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5483 // two branches instead of an explicit OR instruction with a
5485 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5486 isX86LogicalCmp(Cmp)) {
5487 CC = Cond.getOperand(0).getOperand(0);
5488 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5489 Chain, Dest, CC, Cmp);
5490 CC = Cond.getOperand(1).getOperand(0);
5494 } else { // ISD::AND
5495 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5496 // two branches instead of an explicit AND instruction with a
5497 // separate test. However, we only do this if this block doesn't
5498 // have a fall-through edge, because this requires an explicit
5499 // jmp when the condition is false.
5500 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5501 isX86LogicalCmp(Cmp) &&
5502 Op.getNode()->hasOneUse()) {
5503 X86::CondCode CCode =
5504 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5505 CCode = X86::GetOppositeBranchCondition(CCode);
5506 CC = DAG.getConstant(CCode, MVT::i8);
5507 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5508 // Look for an unconditional branch following this conditional branch.
5509 // We need this because we need to reverse the successors in order
5510 // to implement FCMP_OEQ.
5511 if (User.getOpcode() == ISD::BR) {
5512 SDValue FalseBB = User.getOperand(1);
5514 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5515 assert(NewBR == User);
5518 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5519 Chain, Dest, CC, Cmp);
5520 X86::CondCode CCode =
5521 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5522 CCode = X86::GetOppositeBranchCondition(CCode);
5523 CC = DAG.getConstant(CCode, MVT::i8);
5529 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5530 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5531 // It should be transformed during dag combiner except when the condition
5532 // is set by a arithmetics with overflow node.
5533 X86::CondCode CCode =
5534 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5535 CCode = X86::GetOppositeBranchCondition(CCode);
5536 CC = DAG.getConstant(CCode, MVT::i8);
5537 Cond = Cond.getOperand(0).getOperand(1);
5543 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5544 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5546 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5547 Chain, Dest, CC, Cond);
5551 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5552 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5553 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5554 // that the guard pages used by the OS virtual memory manager are allocated in
5555 // correct sequence.
5557 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5558 SelectionDAG &DAG) {
5559 assert(Subtarget->isTargetCygMing() &&
5560 "This should be used only on Cygwin/Mingw targets");
5561 DebugLoc dl = Op.getDebugLoc();
5564 SDValue Chain = Op.getOperand(0);
5565 SDValue Size = Op.getOperand(1);
5566 // FIXME: Ensure alignment here
5570 MVT IntPtr = getPointerTy();
5571 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5573 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5575 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5576 Flag = Chain.getValue(1);
5578 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5579 SDValue Ops[] = { Chain,
5580 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5581 DAG.getRegister(X86::EAX, IntPtr),
5582 DAG.getRegister(X86StackPtr, SPTy),
5584 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5585 Flag = Chain.getValue(1);
5587 Chain = DAG.getCALLSEQ_END(Chain,
5588 DAG.getIntPtrConstant(0, true),
5589 DAG.getIntPtrConstant(0, true),
5592 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5594 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5595 return DAG.getMergeValues(Ops1, 2, dl);
5599 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5601 SDValue Dst, SDValue Src,
5602 SDValue Size, unsigned Align,
5604 uint64_t DstSVOff) {
5605 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5607 // If not DWORD aligned or size is more than the threshold, call the library.
5608 // The libc version is likely to be faster for these cases. It can use the
5609 // address value and run time information about the CPU.
5610 if ((Align & 3) != 0 ||
5612 ConstantSize->getZExtValue() >
5613 getSubtarget()->getMaxInlineSizeThreshold()) {
5614 SDValue InFlag(0, 0);
5616 // Check to see if there is a specialized entry-point for memory zeroing.
5617 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5619 if (const char *bzeroEntry = V &&
5620 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5621 MVT IntPtr = getPointerTy();
5622 const Type *IntPtrTy = TD->getIntPtrType();
5623 TargetLowering::ArgListTy Args;
5624 TargetLowering::ArgListEntry Entry;
5626 Entry.Ty = IntPtrTy;
5627 Args.push_back(Entry);
5629 Args.push_back(Entry);
5630 std::pair<SDValue,SDValue> CallResult =
5631 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5632 CallingConv::C, false,
5633 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5634 return CallResult.second;
5637 // Otherwise have the target-independent code call memset.
5641 uint64_t SizeVal = ConstantSize->getZExtValue();
5642 SDValue InFlag(0, 0);
5645 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5646 unsigned BytesLeft = 0;
5647 bool TwoRepStos = false;
5650 uint64_t Val = ValC->getZExtValue() & 255;
5652 // If the value is a constant, then we can potentially use larger sets.
5653 switch (Align & 3) {
5654 case 2: // WORD aligned
5657 Val = (Val << 8) | Val;
5659 case 0: // DWORD aligned
5662 Val = (Val << 8) | Val;
5663 Val = (Val << 16) | Val;
5664 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5667 Val = (Val << 32) | Val;
5670 default: // Byte aligned
5673 Count = DAG.getIntPtrConstant(SizeVal);
5677 if (AVT.bitsGT(MVT::i8)) {
5678 unsigned UBytes = AVT.getSizeInBits() / 8;
5679 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5680 BytesLeft = SizeVal % UBytes;
5683 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5685 InFlag = Chain.getValue(1);
5688 Count = DAG.getIntPtrConstant(SizeVal);
5689 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5690 InFlag = Chain.getValue(1);
5693 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5696 InFlag = Chain.getValue(1);
5697 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5700 InFlag = Chain.getValue(1);
5702 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5703 SmallVector<SDValue, 8> Ops;
5704 Ops.push_back(Chain);
5705 Ops.push_back(DAG.getValueType(AVT));
5706 Ops.push_back(InFlag);
5707 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5710 InFlag = Chain.getValue(1);
5712 MVT CVT = Count.getValueType();
5713 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5714 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5715 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
5718 InFlag = Chain.getValue(1);
5719 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5721 Ops.push_back(Chain);
5722 Ops.push_back(DAG.getValueType(MVT::i8));
5723 Ops.push_back(InFlag);
5724 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5725 } else if (BytesLeft) {
5726 // Handle the last 1 - 7 bytes.
5727 unsigned Offset = SizeVal - BytesLeft;
5728 MVT AddrVT = Dst.getValueType();
5729 MVT SizeVT = Size.getValueType();
5731 Chain = DAG.getMemset(Chain, dl,
5732 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
5733 DAG.getConstant(Offset, AddrVT)),
5735 DAG.getConstant(BytesLeft, SizeVT),
5736 Align, DstSV, DstSVOff + Offset);
5739 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5744 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
5745 SDValue Chain, SDValue Dst, SDValue Src,
5746 SDValue Size, unsigned Align,
5748 const Value *DstSV, uint64_t DstSVOff,
5749 const Value *SrcSV, uint64_t SrcSVOff) {
5750 // This requires the copy size to be a constant, preferrably
5751 // within a subtarget-specific limit.
5752 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5755 uint64_t SizeVal = ConstantSize->getZExtValue();
5756 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5759 /// If not DWORD aligned, call the library.
5760 if ((Align & 3) != 0)
5765 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5768 unsigned UBytes = AVT.getSizeInBits() / 8;
5769 unsigned CountVal = SizeVal / UBytes;
5770 SDValue Count = DAG.getIntPtrConstant(CountVal);
5771 unsigned BytesLeft = SizeVal % UBytes;
5773 SDValue InFlag(0, 0);
5774 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5777 InFlag = Chain.getValue(1);
5778 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5781 InFlag = Chain.getValue(1);
5782 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
5785 InFlag = Chain.getValue(1);
5787 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5788 SmallVector<SDValue, 8> Ops;
5789 Ops.push_back(Chain);
5790 Ops.push_back(DAG.getValueType(AVT));
5791 Ops.push_back(InFlag);
5792 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
5794 SmallVector<SDValue, 4> Results;
5795 Results.push_back(RepMovs);
5797 // Handle the last 1 - 7 bytes.
5798 unsigned Offset = SizeVal - BytesLeft;
5799 MVT DstVT = Dst.getValueType();
5800 MVT SrcVT = Src.getValueType();
5801 MVT SizeVT = Size.getValueType();
5802 Results.push_back(DAG.getMemcpy(Chain, dl,
5803 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
5804 DAG.getConstant(Offset, DstVT)),
5805 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
5806 DAG.getConstant(Offset, SrcVT)),
5807 DAG.getConstant(BytesLeft, SizeVT),
5808 Align, AlwaysInline,
5809 DstSV, DstSVOff + Offset,
5810 SrcSV, SrcSVOff + Offset));
5813 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5814 &Results[0], Results.size());
5817 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5818 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5819 DebugLoc dl = Op.getDebugLoc();
5821 if (!Subtarget->is64Bit()) {
5822 // vastart just stores the address of the VarArgsFrameIndex slot into the
5823 // memory location argument.
5824 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5825 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
5829 // gp_offset (0 - 6 * 8)
5830 // fp_offset (48 - 48 + 8 * 16)
5831 // overflow_arg_area (point to parameters coming in memory).
5833 SmallVector<SDValue, 8> MemOps;
5834 SDValue FIN = Op.getOperand(1);
5836 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
5837 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5839 MemOps.push_back(Store);
5842 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5843 FIN, DAG.getIntPtrConstant(4));
5844 Store = DAG.getStore(Op.getOperand(0), dl,
5845 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5847 MemOps.push_back(Store);
5849 // Store ptr to overflow_arg_area
5850 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5851 FIN, DAG.getIntPtrConstant(4));
5852 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5853 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
5854 MemOps.push_back(Store);
5856 // Store ptr to reg_save_area.
5857 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5858 FIN, DAG.getIntPtrConstant(8));
5859 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5860 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
5861 MemOps.push_back(Store);
5862 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5863 &MemOps[0], MemOps.size());
5866 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5867 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5868 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5869 SDValue Chain = Op.getOperand(0);
5870 SDValue SrcPtr = Op.getOperand(1);
5871 SDValue SrcSV = Op.getOperand(2);
5873 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5878 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5879 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5880 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5881 SDValue Chain = Op.getOperand(0);
5882 SDValue DstPtr = Op.getOperand(1);
5883 SDValue SrcPtr = Op.getOperand(2);
5884 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5885 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5886 DebugLoc dl = Op.getDebugLoc();
5888 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
5889 DAG.getIntPtrConstant(24), 8, false,
5890 DstSV, 0, SrcSV, 0);
5894 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5895 DebugLoc dl = Op.getDebugLoc();
5896 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5898 default: return SDValue(); // Don't custom lower most intrinsics.
5899 // Comparison intrinsics.
5900 case Intrinsic::x86_sse_comieq_ss:
5901 case Intrinsic::x86_sse_comilt_ss:
5902 case Intrinsic::x86_sse_comile_ss:
5903 case Intrinsic::x86_sse_comigt_ss:
5904 case Intrinsic::x86_sse_comige_ss:
5905 case Intrinsic::x86_sse_comineq_ss:
5906 case Intrinsic::x86_sse_ucomieq_ss:
5907 case Intrinsic::x86_sse_ucomilt_ss:
5908 case Intrinsic::x86_sse_ucomile_ss:
5909 case Intrinsic::x86_sse_ucomigt_ss:
5910 case Intrinsic::x86_sse_ucomige_ss:
5911 case Intrinsic::x86_sse_ucomineq_ss:
5912 case Intrinsic::x86_sse2_comieq_sd:
5913 case Intrinsic::x86_sse2_comilt_sd:
5914 case Intrinsic::x86_sse2_comile_sd:
5915 case Intrinsic::x86_sse2_comigt_sd:
5916 case Intrinsic::x86_sse2_comige_sd:
5917 case Intrinsic::x86_sse2_comineq_sd:
5918 case Intrinsic::x86_sse2_ucomieq_sd:
5919 case Intrinsic::x86_sse2_ucomilt_sd:
5920 case Intrinsic::x86_sse2_ucomile_sd:
5921 case Intrinsic::x86_sse2_ucomigt_sd:
5922 case Intrinsic::x86_sse2_ucomige_sd:
5923 case Intrinsic::x86_sse2_ucomineq_sd: {
5925 ISD::CondCode CC = ISD::SETCC_INVALID;
5928 case Intrinsic::x86_sse_comieq_ss:
5929 case Intrinsic::x86_sse2_comieq_sd:
5933 case Intrinsic::x86_sse_comilt_ss:
5934 case Intrinsic::x86_sse2_comilt_sd:
5938 case Intrinsic::x86_sse_comile_ss:
5939 case Intrinsic::x86_sse2_comile_sd:
5943 case Intrinsic::x86_sse_comigt_ss:
5944 case Intrinsic::x86_sse2_comigt_sd:
5948 case Intrinsic::x86_sse_comige_ss:
5949 case Intrinsic::x86_sse2_comige_sd:
5953 case Intrinsic::x86_sse_comineq_ss:
5954 case Intrinsic::x86_sse2_comineq_sd:
5958 case Intrinsic::x86_sse_ucomieq_ss:
5959 case Intrinsic::x86_sse2_ucomieq_sd:
5960 Opc = X86ISD::UCOMI;
5963 case Intrinsic::x86_sse_ucomilt_ss:
5964 case Intrinsic::x86_sse2_ucomilt_sd:
5965 Opc = X86ISD::UCOMI;
5968 case Intrinsic::x86_sse_ucomile_ss:
5969 case Intrinsic::x86_sse2_ucomile_sd:
5970 Opc = X86ISD::UCOMI;
5973 case Intrinsic::x86_sse_ucomigt_ss:
5974 case Intrinsic::x86_sse2_ucomigt_sd:
5975 Opc = X86ISD::UCOMI;
5978 case Intrinsic::x86_sse_ucomige_ss:
5979 case Intrinsic::x86_sse2_ucomige_sd:
5980 Opc = X86ISD::UCOMI;
5983 case Intrinsic::x86_sse_ucomineq_ss:
5984 case Intrinsic::x86_sse2_ucomineq_sd:
5985 Opc = X86ISD::UCOMI;
5990 SDValue LHS = Op.getOperand(1);
5991 SDValue RHS = Op.getOperand(2);
5992 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
5993 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
5994 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5995 DAG.getConstant(X86CC, MVT::i8), Cond);
5996 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
5999 // Fix vector shift instructions where the last operand is a non-immediate
6001 case Intrinsic::x86_sse2_pslli_w:
6002 case Intrinsic::x86_sse2_pslli_d:
6003 case Intrinsic::x86_sse2_pslli_q:
6004 case Intrinsic::x86_sse2_psrli_w:
6005 case Intrinsic::x86_sse2_psrli_d:
6006 case Intrinsic::x86_sse2_psrli_q:
6007 case Intrinsic::x86_sse2_psrai_w:
6008 case Intrinsic::x86_sse2_psrai_d:
6009 case Intrinsic::x86_mmx_pslli_w:
6010 case Intrinsic::x86_mmx_pslli_d:
6011 case Intrinsic::x86_mmx_pslli_q:
6012 case Intrinsic::x86_mmx_psrli_w:
6013 case Intrinsic::x86_mmx_psrli_d:
6014 case Intrinsic::x86_mmx_psrli_q:
6015 case Intrinsic::x86_mmx_psrai_w:
6016 case Intrinsic::x86_mmx_psrai_d: {
6017 SDValue ShAmt = Op.getOperand(2);
6018 if (isa<ConstantSDNode>(ShAmt))
6021 unsigned NewIntNo = 0;
6022 MVT ShAmtVT = MVT::v4i32;
6024 case Intrinsic::x86_sse2_pslli_w:
6025 NewIntNo = Intrinsic::x86_sse2_psll_w;
6027 case Intrinsic::x86_sse2_pslli_d:
6028 NewIntNo = Intrinsic::x86_sse2_psll_d;
6030 case Intrinsic::x86_sse2_pslli_q:
6031 NewIntNo = Intrinsic::x86_sse2_psll_q;
6033 case Intrinsic::x86_sse2_psrli_w:
6034 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6036 case Intrinsic::x86_sse2_psrli_d:
6037 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6039 case Intrinsic::x86_sse2_psrli_q:
6040 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6042 case Intrinsic::x86_sse2_psrai_w:
6043 NewIntNo = Intrinsic::x86_sse2_psra_w;
6045 case Intrinsic::x86_sse2_psrai_d:
6046 NewIntNo = Intrinsic::x86_sse2_psra_d;
6049 ShAmtVT = MVT::v2i32;
6051 case Intrinsic::x86_mmx_pslli_w:
6052 NewIntNo = Intrinsic::x86_mmx_psll_w;
6054 case Intrinsic::x86_mmx_pslli_d:
6055 NewIntNo = Intrinsic::x86_mmx_psll_d;
6057 case Intrinsic::x86_mmx_pslli_q:
6058 NewIntNo = Intrinsic::x86_mmx_psll_q;
6060 case Intrinsic::x86_mmx_psrli_w:
6061 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6063 case Intrinsic::x86_mmx_psrli_d:
6064 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6066 case Intrinsic::x86_mmx_psrli_q:
6067 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6069 case Intrinsic::x86_mmx_psrai_w:
6070 NewIntNo = Intrinsic::x86_mmx_psra_w;
6072 case Intrinsic::x86_mmx_psrai_d:
6073 NewIntNo = Intrinsic::x86_mmx_psra_d;
6075 default: abort(); // Can't reach here.
6080 MVT VT = Op.getValueType();
6081 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6082 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6083 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6084 DAG.getConstant(NewIntNo, MVT::i32),
6085 Op.getOperand(1), ShAmt);
6090 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6091 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6092 DebugLoc dl = Op.getDebugLoc();
6095 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6097 DAG.getConstant(TD->getPointerSize(),
6098 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6099 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6100 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6105 // Just load the return address.
6106 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6107 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6108 RetAddrFI, NULL, 0);
6111 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6112 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6113 MFI->setFrameAddressIsTaken(true);
6114 MVT VT = Op.getValueType();
6115 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6116 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6117 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6118 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6120 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6124 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6125 SelectionDAG &DAG) {
6126 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6129 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6131 MachineFunction &MF = DAG.getMachineFunction();
6132 SDValue Chain = Op.getOperand(0);
6133 SDValue Offset = Op.getOperand(1);
6134 SDValue Handler = Op.getOperand(2);
6135 DebugLoc dl = Op.getDebugLoc();
6137 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6139 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6141 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6142 DAG.getIntPtrConstant(-TD->getPointerSize()));
6143 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6144 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6145 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6146 MF.getRegInfo().addLiveOut(StoreAddrReg);
6148 return DAG.getNode(X86ISD::EH_RETURN, dl,
6150 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6153 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6154 SelectionDAG &DAG) {
6155 SDValue Root = Op.getOperand(0);
6156 SDValue Trmp = Op.getOperand(1); // trampoline
6157 SDValue FPtr = Op.getOperand(2); // nested function
6158 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6159 DebugLoc dl = Op.getDebugLoc();
6161 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6163 const X86InstrInfo *TII =
6164 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6166 if (Subtarget->is64Bit()) {
6167 SDValue OutChains[6];
6169 // Large code-model.
6171 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6172 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6174 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6175 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6177 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6179 // Load the pointer to the nested function into R11.
6180 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6181 SDValue Addr = Trmp;
6182 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6185 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6186 DAG.getConstant(2, MVT::i64));
6187 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6189 // Load the 'nest' parameter value into R10.
6190 // R10 is specified in X86CallingConv.td
6191 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6192 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6193 DAG.getConstant(10, MVT::i64));
6194 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6195 Addr, TrmpAddr, 10);
6197 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6198 DAG.getConstant(12, MVT::i64));
6199 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6201 // Jump to the nested function.
6202 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6203 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6204 DAG.getConstant(20, MVT::i64));
6205 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6206 Addr, TrmpAddr, 20);
6208 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6209 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6210 DAG.getConstant(22, MVT::i64));
6211 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6215 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6216 return DAG.getMergeValues(Ops, 2, dl);
6218 const Function *Func =
6219 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6220 unsigned CC = Func->getCallingConv();
6225 assert(0 && "Unsupported calling convention");
6226 case CallingConv::C:
6227 case CallingConv::X86_StdCall: {
6228 // Pass 'nest' parameter in ECX.
6229 // Must be kept in sync with X86CallingConv.td
6232 // Check that ECX wasn't needed by an 'inreg' parameter.
6233 const FunctionType *FTy = Func->getFunctionType();
6234 const AttrListPtr &Attrs = Func->getAttributes();
6236 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6237 unsigned InRegCount = 0;
6240 for (FunctionType::param_iterator I = FTy->param_begin(),
6241 E = FTy->param_end(); I != E; ++I, ++Idx)
6242 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6243 // FIXME: should only count parameters that are lowered to integers.
6244 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6246 if (InRegCount > 2) {
6247 cerr << "Nest register in use - reduce number of inreg parameters!\n";
6253 case CallingConv::X86_FastCall:
6254 case CallingConv::Fast:
6255 // Pass 'nest' parameter in EAX.
6256 // Must be kept in sync with X86CallingConv.td
6261 SDValue OutChains[4];
6264 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6265 DAG.getConstant(10, MVT::i32));
6266 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6268 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6269 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6270 OutChains[0] = DAG.getStore(Root, dl,
6271 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6274 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6275 DAG.getConstant(1, MVT::i32));
6276 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6278 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6279 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6280 DAG.getConstant(5, MVT::i32));
6281 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6282 TrmpAddr, 5, false, 1);
6284 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6285 DAG.getConstant(6, MVT::i32));
6286 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6289 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6290 return DAG.getMergeValues(Ops, 2, dl);
6294 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6296 The rounding mode is in bits 11:10 of FPSR, and has the following
6303 FLT_ROUNDS, on the other hand, expects the following:
6310 To perform the conversion, we do:
6311 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6314 MachineFunction &MF = DAG.getMachineFunction();
6315 const TargetMachine &TM = MF.getTarget();
6316 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6317 unsigned StackAlignment = TFI.getStackAlignment();
6318 MVT VT = Op.getValueType();
6319 DebugLoc dl = Op.getDebugLoc();
6321 // Save FP Control Word to stack slot
6322 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6323 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6325 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6326 DAG.getEntryNode(), StackSlot);
6328 // Load FP Control Word from stack slot
6329 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6331 // Transform as necessary
6333 DAG.getNode(ISD::SRL, dl, MVT::i16,
6334 DAG.getNode(ISD::AND, dl, MVT::i16,
6335 CWD, DAG.getConstant(0x800, MVT::i16)),
6336 DAG.getConstant(11, MVT::i8));
6338 DAG.getNode(ISD::SRL, dl, MVT::i16,
6339 DAG.getNode(ISD::AND, dl, MVT::i16,
6340 CWD, DAG.getConstant(0x400, MVT::i16)),
6341 DAG.getConstant(9, MVT::i8));
6344 DAG.getNode(ISD::AND, dl, MVT::i16,
6345 DAG.getNode(ISD::ADD, dl, MVT::i16,
6346 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6347 DAG.getConstant(1, MVT::i16)),
6348 DAG.getConstant(3, MVT::i16));
6351 return DAG.getNode((VT.getSizeInBits() < 16 ?
6352 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6355 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6356 MVT VT = Op.getValueType();
6358 unsigned NumBits = VT.getSizeInBits();
6359 DebugLoc dl = Op.getDebugLoc();
6361 Op = Op.getOperand(0);
6362 if (VT == MVT::i8) {
6363 // Zero extend to i32 since there is not an i8 bsr.
6365 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6368 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6369 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6370 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6372 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6373 SmallVector<SDValue, 4> Ops;
6375 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6376 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6377 Ops.push_back(Op.getValue(1));
6378 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6380 // Finally xor with NumBits-1.
6381 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6384 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6388 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6389 MVT VT = Op.getValueType();
6391 unsigned NumBits = VT.getSizeInBits();
6392 DebugLoc dl = Op.getDebugLoc();
6394 Op = Op.getOperand(0);
6395 if (VT == MVT::i8) {
6397 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6400 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6401 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6402 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6404 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6405 SmallVector<SDValue, 4> Ops;
6407 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6408 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6409 Ops.push_back(Op.getValue(1));
6410 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6413 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6417 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6418 MVT VT = Op.getValueType();
6419 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6420 DebugLoc dl = Op.getDebugLoc();
6422 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6423 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6424 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6425 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6426 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6428 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6429 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6430 // return AloBlo + AloBhi + AhiBlo;
6432 SDValue A = Op.getOperand(0);
6433 SDValue B = Op.getOperand(1);
6435 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6436 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6437 A, DAG.getConstant(32, MVT::i32));
6438 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6439 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6440 B, DAG.getConstant(32, MVT::i32));
6441 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6442 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6444 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6445 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6447 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6448 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6450 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6451 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6452 AloBhi, DAG.getConstant(32, MVT::i32));
6453 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6454 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6455 AhiBlo, DAG.getConstant(32, MVT::i32));
6456 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6457 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6462 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6463 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6464 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6465 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6466 // has only one use.
6467 SDNode *N = Op.getNode();
6468 SDValue LHS = N->getOperand(0);
6469 SDValue RHS = N->getOperand(1);
6470 unsigned BaseOp = 0;
6472 DebugLoc dl = Op.getDebugLoc();
6474 switch (Op.getOpcode()) {
6475 default: assert(0 && "Unknown ovf instruction!");
6477 // A subtract of one will be selected as a INC. Note that INC doesn't
6478 // set CF, so we can't do this for UADDO.
6479 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6480 if (C->getAPIntValue() == 1) {
6481 BaseOp = X86ISD::INC;
6485 BaseOp = X86ISD::ADD;
6489 BaseOp = X86ISD::ADD;
6493 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6494 // set CF, so we can't do this for USUBO.
6495 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6496 if (C->getAPIntValue() == 1) {
6497 BaseOp = X86ISD::DEC;
6501 BaseOp = X86ISD::SUB;
6505 BaseOp = X86ISD::SUB;
6509 BaseOp = X86ISD::SMUL;
6513 BaseOp = X86ISD::UMUL;
6518 // Also sets EFLAGS.
6519 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6520 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6523 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6524 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6526 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6530 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6531 MVT T = Op.getValueType();
6532 DebugLoc dl = Op.getDebugLoc();
6535 switch(T.getSimpleVT()) {
6537 assert(false && "Invalid value type!");
6538 case MVT::i8: Reg = X86::AL; size = 1; break;
6539 case MVT::i16: Reg = X86::AX; size = 2; break;
6540 case MVT::i32: Reg = X86::EAX; size = 4; break;
6542 assert(Subtarget->is64Bit() && "Node not type legal!");
6543 Reg = X86::RAX; size = 8;
6546 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6547 Op.getOperand(2), SDValue());
6548 SDValue Ops[] = { cpIn.getValue(0),
6551 DAG.getTargetConstant(size, MVT::i8),
6553 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6554 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6556 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6560 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6561 SelectionDAG &DAG) {
6562 assert(Subtarget->is64Bit() && "Result not type legalized?");
6563 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6564 SDValue TheChain = Op.getOperand(0);
6565 DebugLoc dl = Op.getDebugLoc();
6566 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6567 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6568 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6570 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6571 DAG.getConstant(32, MVT::i8));
6573 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6576 return DAG.getMergeValues(Ops, 2, dl);
6579 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6580 SDNode *Node = Op.getNode();
6581 DebugLoc dl = Node->getDebugLoc();
6582 MVT T = Node->getValueType(0);
6583 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6584 DAG.getConstant(0, T), Node->getOperand(2));
6585 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6586 cast<AtomicSDNode>(Node)->getMemoryVT(),
6587 Node->getOperand(0),
6588 Node->getOperand(1), negOp,
6589 cast<AtomicSDNode>(Node)->getSrcValue(),
6590 cast<AtomicSDNode>(Node)->getAlignment());
6593 /// LowerOperation - Provide custom lowering hooks for some operations.
6595 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6596 switch (Op.getOpcode()) {
6597 default: assert(0 && "Should not custom lower this!");
6598 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6599 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6600 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6601 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6602 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6603 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6604 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6605 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6606 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6607 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6608 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6609 case ISD::SHL_PARTS:
6610 case ISD::SRA_PARTS:
6611 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6612 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6613 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6614 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6615 case ISD::FABS: return LowerFABS(Op, DAG);
6616 case ISD::FNEG: return LowerFNEG(Op, DAG);
6617 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6618 case ISD::SETCC: return LowerSETCC(Op, DAG);
6619 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6620 case ISD::SELECT: return LowerSELECT(Op, DAG);
6621 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6622 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6623 case ISD::CALL: return LowerCALL(Op, DAG);
6624 case ISD::RET: return LowerRET(Op, DAG);
6625 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6626 case ISD::VASTART: return LowerVASTART(Op, DAG);
6627 case ISD::VAARG: return LowerVAARG(Op, DAG);
6628 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6629 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6630 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6631 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6632 case ISD::FRAME_TO_ARGS_OFFSET:
6633 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6634 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6635 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6636 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6637 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6638 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6639 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6640 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6646 case ISD::UMULO: return LowerXALUO(Op, DAG);
6647 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6651 void X86TargetLowering::
6652 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6653 SelectionDAG &DAG, unsigned NewOp) {
6654 MVT T = Node->getValueType(0);
6655 DebugLoc dl = Node->getDebugLoc();
6656 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6658 SDValue Chain = Node->getOperand(0);
6659 SDValue In1 = Node->getOperand(1);
6660 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6661 Node->getOperand(2), DAG.getIntPtrConstant(0));
6662 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6663 Node->getOperand(2), DAG.getIntPtrConstant(1));
6664 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6665 // have a MemOperand. Pass the info through as a normal operand.
6666 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6667 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6668 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6669 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
6670 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6671 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6672 Results.push_back(Result.getValue(2));
6675 /// ReplaceNodeResults - Replace a node with an illegal result type
6676 /// with a new node built out of custom code.
6677 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6678 SmallVectorImpl<SDValue>&Results,
6679 SelectionDAG &DAG) {
6680 DebugLoc dl = N->getDebugLoc();
6681 switch (N->getOpcode()) {
6683 assert(false && "Do not know how to custom type legalize this operation!");
6685 case ISD::FP_TO_SINT: {
6686 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
6687 SDValue FIST = Vals.first, StackSlot = Vals.second;
6688 if (FIST.getNode() != 0) {
6689 MVT VT = N->getValueType(0);
6690 // Return a load from the stack slot.
6691 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
6695 case ISD::READCYCLECOUNTER: {
6696 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6697 SDValue TheChain = N->getOperand(0);
6698 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6699 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
6701 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
6703 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6704 SDValue Ops[] = { eax, edx };
6705 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
6706 Results.push_back(edx.getValue(1));
6709 case ISD::ATOMIC_CMP_SWAP: {
6710 MVT T = N->getValueType(0);
6711 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6712 SDValue cpInL, cpInH;
6713 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6714 DAG.getConstant(0, MVT::i32));
6715 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6716 DAG.getConstant(1, MVT::i32));
6717 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
6718 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
6720 SDValue swapInL, swapInH;
6721 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6722 DAG.getConstant(0, MVT::i32));
6723 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6724 DAG.getConstant(1, MVT::i32));
6725 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
6727 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
6728 swapInL.getValue(1));
6729 SDValue Ops[] = { swapInH.getValue(0),
6731 swapInH.getValue(1) };
6732 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6733 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
6734 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
6735 MVT::i32, Result.getValue(1));
6736 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
6737 MVT::i32, cpOutL.getValue(2));
6738 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6739 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6740 Results.push_back(cpOutH.getValue(1));
6743 case ISD::ATOMIC_LOAD_ADD:
6744 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
6746 case ISD::ATOMIC_LOAD_AND:
6747 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
6749 case ISD::ATOMIC_LOAD_NAND:
6750 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
6752 case ISD::ATOMIC_LOAD_OR:
6753 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
6755 case ISD::ATOMIC_LOAD_SUB:
6756 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
6758 case ISD::ATOMIC_LOAD_XOR:
6759 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
6761 case ISD::ATOMIC_SWAP:
6762 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
6767 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6769 default: return NULL;
6770 case X86ISD::BSF: return "X86ISD::BSF";
6771 case X86ISD::BSR: return "X86ISD::BSR";
6772 case X86ISD::SHLD: return "X86ISD::SHLD";
6773 case X86ISD::SHRD: return "X86ISD::SHRD";
6774 case X86ISD::FAND: return "X86ISD::FAND";
6775 case X86ISD::FOR: return "X86ISD::FOR";
6776 case X86ISD::FXOR: return "X86ISD::FXOR";
6777 case X86ISD::FSRL: return "X86ISD::FSRL";
6778 case X86ISD::FILD: return "X86ISD::FILD";
6779 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6780 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6781 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6782 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6783 case X86ISD::FLD: return "X86ISD::FLD";
6784 case X86ISD::FST: return "X86ISD::FST";
6785 case X86ISD::CALL: return "X86ISD::CALL";
6786 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6787 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6788 case X86ISD::BT: return "X86ISD::BT";
6789 case X86ISD::CMP: return "X86ISD::CMP";
6790 case X86ISD::COMI: return "X86ISD::COMI";
6791 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6792 case X86ISD::SETCC: return "X86ISD::SETCC";
6793 case X86ISD::CMOV: return "X86ISD::CMOV";
6794 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6795 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6796 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6797 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6798 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6799 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6800 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6801 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6802 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6803 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6804 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6805 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
6806 case X86ISD::FMAX: return "X86ISD::FMAX";
6807 case X86ISD::FMIN: return "X86ISD::FMIN";
6808 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6809 case X86ISD::FRCP: return "X86ISD::FRCP";
6810 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6811 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
6812 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6813 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6814 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6815 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6816 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6817 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
6818 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
6819 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
6820 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
6821 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
6822 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
6823 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6824 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6825 case X86ISD::VSHL: return "X86ISD::VSHL";
6826 case X86ISD::VSRL: return "X86ISD::VSRL";
6827 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6828 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6829 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6830 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6831 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6832 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6833 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6834 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6835 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6836 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6837 case X86ISD::ADD: return "X86ISD::ADD";
6838 case X86ISD::SUB: return "X86ISD::SUB";
6839 case X86ISD::SMUL: return "X86ISD::SMUL";
6840 case X86ISD::UMUL: return "X86ISD::UMUL";
6841 case X86ISD::INC: return "X86ISD::INC";
6842 case X86ISD::DEC: return "X86ISD::DEC";
6843 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
6847 // isLegalAddressingMode - Return true if the addressing mode represented
6848 // by AM is legal for this target, for a load/store of the specified type.
6849 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6850 const Type *Ty) const {
6851 // X86 supports extremely general addressing modes.
6853 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6854 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6858 // We can only fold this if we don't need an extra load.
6859 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6861 // If BaseGV requires a register, we cannot also have a BaseReg.
6862 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine(), false) &&
6866 // X86-64 only supports addr of globals in small code model.
6867 if (Subtarget->is64Bit()) {
6868 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6870 // If lower 4G is not available, then we must use rip-relative addressing.
6871 if (AM.BaseOffs || AM.Scale > 1)
6882 // These scales always work.
6887 // These scales are formed with basereg+scalereg. Only accept if there is
6892 default: // Other stuff never works.
6900 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6901 if (!Ty1->isInteger() || !Ty2->isInteger())
6903 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6904 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6905 if (NumBits1 <= NumBits2)
6907 return Subtarget->is64Bit() || NumBits1 < 64;
6910 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6911 if (!VT1.isInteger() || !VT2.isInteger())
6913 unsigned NumBits1 = VT1.getSizeInBits();
6914 unsigned NumBits2 = VT2.getSizeInBits();
6915 if (NumBits1 <= NumBits2)
6917 return Subtarget->is64Bit() || NumBits1 < 64;
6920 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
6921 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
6922 return Ty1 == Type::Int32Ty && Ty2 == Type::Int64Ty && Subtarget->is64Bit();
6925 bool X86TargetLowering::isZExtFree(MVT VT1, MVT VT2) const {
6926 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
6927 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
6930 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6931 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6932 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6933 /// are assumed to be legal.
6935 X86TargetLowering::isShuffleMaskLegal(const int *Mask, MVT VT) const {
6936 // Only do shuffles on 128-bit vector types for now.
6937 if (VT.getSizeInBits() == 64)
6940 // FIXME: pshufb, blends, palignr, shifts.
6941 return (VT.getVectorNumElements() == 2 ||
6942 ShuffleVectorSDNode::isSplatMask(Mask, VT) ||
6943 isMOVLMask(Mask, VT) ||
6944 isSHUFPMask(Mask, VT) ||
6945 isPSHUFDMask(Mask, VT) ||
6946 isPSHUFHWMask(Mask, VT) ||
6947 isPSHUFLWMask(Mask, VT) ||
6948 isUNPCKLMask(Mask, VT) ||
6949 isUNPCKHMask(Mask, VT) ||
6950 isUNPCKL_v_undef_Mask(Mask, VT) ||
6951 isUNPCKH_v_undef_Mask(Mask, VT));
6955 X86TargetLowering::isVectorClearMaskLegal(const int *Mask, MVT VT) const {
6956 unsigned NumElts = VT.getVectorNumElements();
6957 // FIXME: This collection of masks seems suspect.
6960 if (NumElts == 4 && VT.getSizeInBits() == 128) {
6961 return (isMOVLMask(Mask, VT) ||
6962 isCommutedMOVLMask(Mask, VT, true) ||
6963 isSHUFPMask(Mask, VT) ||
6964 isCommutedSHUFPMask(Mask, VT));
6969 //===----------------------------------------------------------------------===//
6970 // X86 Scheduler Hooks
6971 //===----------------------------------------------------------------------===//
6973 // private utility function
6975 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
6976 MachineBasicBlock *MBB,
6984 TargetRegisterClass *RC,
6985 bool invSrc) const {
6986 // For the atomic bitwise operator, we generate
6989 // ld t1 = [bitinstr.addr]
6990 // op t2 = t1, [bitinstr.val]
6992 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6994 // fallthrough -->nextMBB
6995 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6996 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6997 MachineFunction::iterator MBBIter = MBB;
7000 /// First build the CFG
7001 MachineFunction *F = MBB->getParent();
7002 MachineBasicBlock *thisMBB = MBB;
7003 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7004 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7005 F->insert(MBBIter, newMBB);
7006 F->insert(MBBIter, nextMBB);
7008 // Move all successors to thisMBB to nextMBB
7009 nextMBB->transferSuccessors(thisMBB);
7011 // Update thisMBB to fall through to newMBB
7012 thisMBB->addSuccessor(newMBB);
7014 // newMBB jumps to itself and fall through to nextMBB
7015 newMBB->addSuccessor(nextMBB);
7016 newMBB->addSuccessor(newMBB);
7018 // Insert instructions into newMBB based on incoming instruction
7019 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7020 "unexpected number of operands");
7021 DebugLoc dl = bInstr->getDebugLoc();
7022 MachineOperand& destOper = bInstr->getOperand(0);
7023 MachineOperand* argOpers[2 + X86AddrNumOperands];
7024 int numArgs = bInstr->getNumOperands() - 1;
7025 for (int i=0; i < numArgs; ++i)
7026 argOpers[i] = &bInstr->getOperand(i+1);
7028 // x86 address has 4 operands: base, index, scale, and displacement
7029 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7030 int valArgIndx = lastAddrIndx + 1;
7032 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7033 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7034 for (int i=0; i <= lastAddrIndx; ++i)
7035 (*MIB).addOperand(*argOpers[i]);
7037 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7039 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7044 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7045 assert((argOpers[valArgIndx]->isReg() ||
7046 argOpers[valArgIndx]->isImm()) &&
7048 if (argOpers[valArgIndx]->isReg())
7049 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7051 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7053 (*MIB).addOperand(*argOpers[valArgIndx]);
7055 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7058 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7059 for (int i=0; i <= lastAddrIndx; ++i)
7060 (*MIB).addOperand(*argOpers[i]);
7062 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7063 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7065 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7069 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7071 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7075 // private utility function: 64 bit atomics on 32 bit host.
7077 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7078 MachineBasicBlock *MBB,
7083 bool invSrc) const {
7084 // For the atomic bitwise operator, we generate
7085 // thisMBB (instructions are in pairs, except cmpxchg8b)
7086 // ld t1,t2 = [bitinstr.addr]
7088 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7089 // op t5, t6 <- out1, out2, [bitinstr.val]
7090 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7091 // mov ECX, EBX <- t5, t6
7092 // mov EAX, EDX <- t1, t2
7093 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7094 // mov t3, t4 <- EAX, EDX
7096 // result in out1, out2
7097 // fallthrough -->nextMBB
7099 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7100 const unsigned LoadOpc = X86::MOV32rm;
7101 const unsigned copyOpc = X86::MOV32rr;
7102 const unsigned NotOpc = X86::NOT32r;
7103 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7104 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7105 MachineFunction::iterator MBBIter = MBB;
7108 /// First build the CFG
7109 MachineFunction *F = MBB->getParent();
7110 MachineBasicBlock *thisMBB = MBB;
7111 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7112 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7113 F->insert(MBBIter, newMBB);
7114 F->insert(MBBIter, nextMBB);
7116 // Move all successors to thisMBB to nextMBB
7117 nextMBB->transferSuccessors(thisMBB);
7119 // Update thisMBB to fall through to newMBB
7120 thisMBB->addSuccessor(newMBB);
7122 // newMBB jumps to itself and fall through to nextMBB
7123 newMBB->addSuccessor(nextMBB);
7124 newMBB->addSuccessor(newMBB);
7126 DebugLoc dl = bInstr->getDebugLoc();
7127 // Insert instructions into newMBB based on incoming instruction
7128 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7129 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7130 "unexpected number of operands");
7131 MachineOperand& dest1Oper = bInstr->getOperand(0);
7132 MachineOperand& dest2Oper = bInstr->getOperand(1);
7133 MachineOperand* argOpers[2 + X86AddrNumOperands];
7134 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7135 argOpers[i] = &bInstr->getOperand(i+2);
7137 // x86 address has 4 operands: base, index, scale, and displacement
7138 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7140 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7141 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7142 for (int i=0; i <= lastAddrIndx; ++i)
7143 (*MIB).addOperand(*argOpers[i]);
7144 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7145 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7146 // add 4 to displacement.
7147 for (int i=0; i <= lastAddrIndx-2; ++i)
7148 (*MIB).addOperand(*argOpers[i]);
7149 MachineOperand newOp3 = *(argOpers[3]);
7151 newOp3.setImm(newOp3.getImm()+4);
7153 newOp3.setOffset(newOp3.getOffset()+4);
7154 (*MIB).addOperand(newOp3);
7155 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7157 // t3/4 are defined later, at the bottom of the loop
7158 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7159 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7160 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7161 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7162 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7163 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7165 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7166 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7168 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7169 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7175 int valArgIndx = lastAddrIndx + 1;
7176 assert((argOpers[valArgIndx]->isReg() ||
7177 argOpers[valArgIndx]->isImm()) &&
7179 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7180 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7181 if (argOpers[valArgIndx]->isReg())
7182 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7184 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7185 if (regOpcL != X86::MOV32rr)
7187 (*MIB).addOperand(*argOpers[valArgIndx]);
7188 assert(argOpers[valArgIndx + 1]->isReg() ==
7189 argOpers[valArgIndx]->isReg());
7190 assert(argOpers[valArgIndx + 1]->isImm() ==
7191 argOpers[valArgIndx]->isImm());
7192 if (argOpers[valArgIndx + 1]->isReg())
7193 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7195 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7196 if (regOpcH != X86::MOV32rr)
7198 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7200 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7202 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7205 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7207 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7210 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7211 for (int i=0; i <= lastAddrIndx; ++i)
7212 (*MIB).addOperand(*argOpers[i]);
7214 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7215 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7217 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7218 MIB.addReg(X86::EAX);
7219 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7220 MIB.addReg(X86::EDX);
7223 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7225 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7229 // private utility function
7231 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7232 MachineBasicBlock *MBB,
7233 unsigned cmovOpc) const {
7234 // For the atomic min/max operator, we generate
7237 // ld t1 = [min/max.addr]
7238 // mov t2 = [min/max.val]
7240 // cmov[cond] t2 = t1
7242 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7244 // fallthrough -->nextMBB
7246 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7247 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7248 MachineFunction::iterator MBBIter = MBB;
7251 /// First build the CFG
7252 MachineFunction *F = MBB->getParent();
7253 MachineBasicBlock *thisMBB = MBB;
7254 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7255 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7256 F->insert(MBBIter, newMBB);
7257 F->insert(MBBIter, nextMBB);
7259 // Move all successors to thisMBB to nextMBB
7260 nextMBB->transferSuccessors(thisMBB);
7262 // Update thisMBB to fall through to newMBB
7263 thisMBB->addSuccessor(newMBB);
7265 // newMBB jumps to newMBB and fall through to nextMBB
7266 newMBB->addSuccessor(nextMBB);
7267 newMBB->addSuccessor(newMBB);
7269 DebugLoc dl = mInstr->getDebugLoc();
7270 // Insert instructions into newMBB based on incoming instruction
7271 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7272 "unexpected number of operands");
7273 MachineOperand& destOper = mInstr->getOperand(0);
7274 MachineOperand* argOpers[2 + X86AddrNumOperands];
7275 int numArgs = mInstr->getNumOperands() - 1;
7276 for (int i=0; i < numArgs; ++i)
7277 argOpers[i] = &mInstr->getOperand(i+1);
7279 // x86 address has 4 operands: base, index, scale, and displacement
7280 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7281 int valArgIndx = lastAddrIndx + 1;
7283 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7284 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7285 for (int i=0; i <= lastAddrIndx; ++i)
7286 (*MIB).addOperand(*argOpers[i]);
7288 // We only support register and immediate values
7289 assert((argOpers[valArgIndx]->isReg() ||
7290 argOpers[valArgIndx]->isImm()) &&
7293 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7294 if (argOpers[valArgIndx]->isReg())
7295 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7297 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7298 (*MIB).addOperand(*argOpers[valArgIndx]);
7300 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7303 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7308 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7309 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7313 // Cmp and exchange if none has modified the memory location
7314 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7315 for (int i=0; i <= lastAddrIndx; ++i)
7316 (*MIB).addOperand(*argOpers[i]);
7318 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7319 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7321 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7322 MIB.addReg(X86::EAX);
7325 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7327 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7333 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7334 MachineBasicBlock *BB) const {
7335 DebugLoc dl = MI->getDebugLoc();
7336 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7337 switch (MI->getOpcode()) {
7338 default: assert(false && "Unexpected instr type to insert");
7339 case X86::CMOV_V1I64:
7340 case X86::CMOV_FR32:
7341 case X86::CMOV_FR64:
7342 case X86::CMOV_V4F32:
7343 case X86::CMOV_V2F64:
7344 case X86::CMOV_V2I64: {
7345 // To "insert" a SELECT_CC instruction, we actually have to insert the
7346 // diamond control-flow pattern. The incoming instruction knows the
7347 // destination vreg to set, the condition code register to branch on, the
7348 // true/false values to select between, and a branch opcode to use.
7349 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7350 MachineFunction::iterator It = BB;
7356 // cmpTY ccX, r1, r2
7358 // fallthrough --> copy0MBB
7359 MachineBasicBlock *thisMBB = BB;
7360 MachineFunction *F = BB->getParent();
7361 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7362 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7364 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7365 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7366 F->insert(It, copy0MBB);
7367 F->insert(It, sinkMBB);
7368 // Update machine-CFG edges by transferring all successors of the current
7369 // block to the new block which will contain the Phi node for the select.
7370 sinkMBB->transferSuccessors(BB);
7372 // Add the true and fallthrough blocks as its successors.
7373 BB->addSuccessor(copy0MBB);
7374 BB->addSuccessor(sinkMBB);
7377 // %FalseValue = ...
7378 // # fallthrough to sinkMBB
7381 // Update machine-CFG edges
7382 BB->addSuccessor(sinkMBB);
7385 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7388 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7389 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7390 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7392 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7396 case X86::FP32_TO_INT16_IN_MEM:
7397 case X86::FP32_TO_INT32_IN_MEM:
7398 case X86::FP32_TO_INT64_IN_MEM:
7399 case X86::FP64_TO_INT16_IN_MEM:
7400 case X86::FP64_TO_INT32_IN_MEM:
7401 case X86::FP64_TO_INT64_IN_MEM:
7402 case X86::FP80_TO_INT16_IN_MEM:
7403 case X86::FP80_TO_INT32_IN_MEM:
7404 case X86::FP80_TO_INT64_IN_MEM: {
7405 // Change the floating point control register to use "round towards zero"
7406 // mode when truncating to an integer value.
7407 MachineFunction *F = BB->getParent();
7408 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7409 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7411 // Load the old value of the high byte of the control word...
7413 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7414 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7417 // Set the high part to be round to zero...
7418 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7421 // Reload the modified control word now...
7422 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7424 // Restore the memory image of control word to original value
7425 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7428 // Get the X86 opcode to use.
7430 switch (MI->getOpcode()) {
7431 default: assert(0 && "illegal opcode!");
7432 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7433 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7434 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7435 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7436 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7437 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7438 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7439 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7440 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7444 MachineOperand &Op = MI->getOperand(0);
7446 AM.BaseType = X86AddressMode::RegBase;
7447 AM.Base.Reg = Op.getReg();
7449 AM.BaseType = X86AddressMode::FrameIndexBase;
7450 AM.Base.FrameIndex = Op.getIndex();
7452 Op = MI->getOperand(1);
7454 AM.Scale = Op.getImm();
7455 Op = MI->getOperand(2);
7457 AM.IndexReg = Op.getImm();
7458 Op = MI->getOperand(3);
7459 if (Op.isGlobal()) {
7460 AM.GV = Op.getGlobal();
7462 AM.Disp = Op.getImm();
7464 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7465 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7467 // Reload the original control word now.
7468 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7470 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7473 case X86::ATOMAND32:
7474 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7475 X86::AND32ri, X86::MOV32rm,
7476 X86::LCMPXCHG32, X86::MOV32rr,
7477 X86::NOT32r, X86::EAX,
7478 X86::GR32RegisterClass);
7480 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7481 X86::OR32ri, X86::MOV32rm,
7482 X86::LCMPXCHG32, X86::MOV32rr,
7483 X86::NOT32r, X86::EAX,
7484 X86::GR32RegisterClass);
7485 case X86::ATOMXOR32:
7486 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7487 X86::XOR32ri, X86::MOV32rm,
7488 X86::LCMPXCHG32, X86::MOV32rr,
7489 X86::NOT32r, X86::EAX,
7490 X86::GR32RegisterClass);
7491 case X86::ATOMNAND32:
7492 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7493 X86::AND32ri, X86::MOV32rm,
7494 X86::LCMPXCHG32, X86::MOV32rr,
7495 X86::NOT32r, X86::EAX,
7496 X86::GR32RegisterClass, true);
7497 case X86::ATOMMIN32:
7498 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7499 case X86::ATOMMAX32:
7500 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7501 case X86::ATOMUMIN32:
7502 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7503 case X86::ATOMUMAX32:
7504 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7506 case X86::ATOMAND16:
7507 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7508 X86::AND16ri, X86::MOV16rm,
7509 X86::LCMPXCHG16, X86::MOV16rr,
7510 X86::NOT16r, X86::AX,
7511 X86::GR16RegisterClass);
7513 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7514 X86::OR16ri, X86::MOV16rm,
7515 X86::LCMPXCHG16, X86::MOV16rr,
7516 X86::NOT16r, X86::AX,
7517 X86::GR16RegisterClass);
7518 case X86::ATOMXOR16:
7519 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7520 X86::XOR16ri, X86::MOV16rm,
7521 X86::LCMPXCHG16, X86::MOV16rr,
7522 X86::NOT16r, X86::AX,
7523 X86::GR16RegisterClass);
7524 case X86::ATOMNAND16:
7525 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7526 X86::AND16ri, X86::MOV16rm,
7527 X86::LCMPXCHG16, X86::MOV16rr,
7528 X86::NOT16r, X86::AX,
7529 X86::GR16RegisterClass, true);
7530 case X86::ATOMMIN16:
7531 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7532 case X86::ATOMMAX16:
7533 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7534 case X86::ATOMUMIN16:
7535 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7536 case X86::ATOMUMAX16:
7537 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7540 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7541 X86::AND8ri, X86::MOV8rm,
7542 X86::LCMPXCHG8, X86::MOV8rr,
7543 X86::NOT8r, X86::AL,
7544 X86::GR8RegisterClass);
7546 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7547 X86::OR8ri, X86::MOV8rm,
7548 X86::LCMPXCHG8, X86::MOV8rr,
7549 X86::NOT8r, X86::AL,
7550 X86::GR8RegisterClass);
7552 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7553 X86::XOR8ri, X86::MOV8rm,
7554 X86::LCMPXCHG8, X86::MOV8rr,
7555 X86::NOT8r, X86::AL,
7556 X86::GR8RegisterClass);
7557 case X86::ATOMNAND8:
7558 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7559 X86::AND8ri, X86::MOV8rm,
7560 X86::LCMPXCHG8, X86::MOV8rr,
7561 X86::NOT8r, X86::AL,
7562 X86::GR8RegisterClass, true);
7563 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7564 // This group is for 64-bit host.
7565 case X86::ATOMAND64:
7566 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7567 X86::AND64ri32, X86::MOV64rm,
7568 X86::LCMPXCHG64, X86::MOV64rr,
7569 X86::NOT64r, X86::RAX,
7570 X86::GR64RegisterClass);
7572 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7573 X86::OR64ri32, X86::MOV64rm,
7574 X86::LCMPXCHG64, X86::MOV64rr,
7575 X86::NOT64r, X86::RAX,
7576 X86::GR64RegisterClass);
7577 case X86::ATOMXOR64:
7578 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7579 X86::XOR64ri32, X86::MOV64rm,
7580 X86::LCMPXCHG64, X86::MOV64rr,
7581 X86::NOT64r, X86::RAX,
7582 X86::GR64RegisterClass);
7583 case X86::ATOMNAND64:
7584 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7585 X86::AND64ri32, X86::MOV64rm,
7586 X86::LCMPXCHG64, X86::MOV64rr,
7587 X86::NOT64r, X86::RAX,
7588 X86::GR64RegisterClass, true);
7589 case X86::ATOMMIN64:
7590 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7591 case X86::ATOMMAX64:
7592 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7593 case X86::ATOMUMIN64:
7594 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7595 case X86::ATOMUMAX64:
7596 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7598 // This group does 64-bit operations on a 32-bit host.
7599 case X86::ATOMAND6432:
7600 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7601 X86::AND32rr, X86::AND32rr,
7602 X86::AND32ri, X86::AND32ri,
7604 case X86::ATOMOR6432:
7605 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7606 X86::OR32rr, X86::OR32rr,
7607 X86::OR32ri, X86::OR32ri,
7609 case X86::ATOMXOR6432:
7610 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7611 X86::XOR32rr, X86::XOR32rr,
7612 X86::XOR32ri, X86::XOR32ri,
7614 case X86::ATOMNAND6432:
7615 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7616 X86::AND32rr, X86::AND32rr,
7617 X86::AND32ri, X86::AND32ri,
7619 case X86::ATOMADD6432:
7620 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7621 X86::ADD32rr, X86::ADC32rr,
7622 X86::ADD32ri, X86::ADC32ri,
7624 case X86::ATOMSUB6432:
7625 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7626 X86::SUB32rr, X86::SBB32rr,
7627 X86::SUB32ri, X86::SBB32ri,
7629 case X86::ATOMSWAP6432:
7630 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7631 X86::MOV32rr, X86::MOV32rr,
7632 X86::MOV32ri, X86::MOV32ri,
7637 //===----------------------------------------------------------------------===//
7638 // X86 Optimization Hooks
7639 //===----------------------------------------------------------------------===//
7641 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7645 const SelectionDAG &DAG,
7646 unsigned Depth) const {
7647 unsigned Opc = Op.getOpcode();
7648 assert((Opc >= ISD::BUILTIN_OP_END ||
7649 Opc == ISD::INTRINSIC_WO_CHAIN ||
7650 Opc == ISD::INTRINSIC_W_CHAIN ||
7651 Opc == ISD::INTRINSIC_VOID) &&
7652 "Should use MaskedValueIsZero if you don't know whether Op"
7653 " is a target node!");
7655 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7664 // These nodes' second result is a boolean.
7665 if (Op.getResNo() == 0)
7669 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7670 Mask.getBitWidth() - 1);
7675 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7676 /// node is a GlobalAddress + offset.
7677 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7678 GlobalValue* &GA, int64_t &Offset) const{
7679 if (N->getOpcode() == X86ISD::Wrapper) {
7680 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7681 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7682 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7686 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7689 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7690 const TargetLowering &TLI) {
7693 if (TLI.isGAPlusOffset(Base, GV, Offset))
7694 return (GV->getAlignment() >= N && (Offset % N) == 0);
7695 // DAG combine handles the stack object case.
7699 static bool EltsFromConsecutiveLoads(SDNode *N, const int *PermMask,
7700 unsigned NumElems, MVT EVT,
7702 SelectionDAG &DAG, MachineFrameInfo *MFI,
7703 const TargetLowering &TLI) {
7705 for (unsigned i = 0; i < NumElems; ++i) {
7706 if (PermMask[i] < 0) {
7712 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7713 if (!Elt.getNode() ||
7714 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7717 Base = Elt.getNode();
7718 if (Base->getOpcode() == ISD::UNDEF)
7722 if (Elt.getOpcode() == ISD::UNDEF)
7725 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
7726 EVT.getSizeInBits()/8, i, MFI))
7732 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7733 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7734 /// if the load addresses are consecutive, non-overlapping, and in the right
7735 /// order. In the case of v2i64, it will see if it can rewrite the
7736 /// shuffle to be an appropriate build vector so it can take advantage of
7737 // performBuildVectorCombine.
7738 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7739 const TargetLowering &TLI) {
7740 DebugLoc dl = N->getDebugLoc();
7741 MVT VT = N->getValueType(0);
7742 MVT EVT = VT.getVectorElementType();
7743 const int *PermMask = cast<ShuffleVectorSDNode>(N)->getMask();
7744 unsigned NumElems = VT.getVectorNumElements();
7746 // For x86-32 machines, if we see an insert and then a shuffle in a v2i64
7747 // where the upper half is 0, it is advantageous to rewrite it as a build
7748 // vector of (0, val) so it can use movq.
7749 if (VT == MVT::v2i64) {
7751 In[0] = N->getOperand(0);
7752 In[1] = N->getOperand(1);
7753 unsigned Idx0 = PermMask[0];
7754 unsigned Idx1 = PermMask[1];
7755 // FIXME: can we take advantage of undef index?
7756 if (PermMask[0] >= 0 && PermMask[1] >= 0 &&
7757 In[Idx0/2].getOpcode() == ISD::INSERT_VECTOR_ELT &&
7758 In[Idx1/2].getOpcode() == ISD::BUILD_VECTOR) {
7759 ConstantSDNode* InsertVecIdx =
7760 dyn_cast<ConstantSDNode>(In[Idx0/2].getOperand(2));
7762 InsertVecIdx->getZExtValue() == (Idx0 % 2) &&
7763 isZeroNode(In[Idx1/2].getOperand(Idx1 % 2))) {
7764 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
7765 In[Idx0/2].getOperand(1),
7766 In[Idx1/2].getOperand(Idx1 % 2));
7771 // Try to combine a vector_shuffle into a 128-bit load.
7772 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7773 SDNode *Base = NULL;
7774 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
7778 LoadSDNode *LD = cast<LoadSDNode>(Base);
7779 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
7780 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7781 LD->getSrcValue(), LD->getSrcValueOffset(),
7783 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7784 LD->getSrcValue(), LD->getSrcValueOffset(),
7785 LD->isVolatile(), LD->getAlignment());
7788 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
7789 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
7790 TargetLowering::DAGCombinerInfo &DCI,
7791 const X86Subtarget *Subtarget,
7792 const TargetLowering &TLI) {
7793 unsigned NumOps = N->getNumOperands();
7794 DebugLoc dl = N->getDebugLoc();
7796 // Ignore single operand BUILD_VECTOR.
7800 MVT VT = N->getValueType(0);
7801 MVT EVT = VT.getVectorElementType();
7802 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
7803 // We are looking for load i64 and zero extend. We want to transform
7804 // it before legalizer has a chance to expand it. Also look for i64
7805 // BUILD_PAIR bit casted to f64.
7807 // This must be an insertion into a zero vector.
7808 SDValue HighElt = N->getOperand(1);
7809 if (!isZeroNode(HighElt))
7812 // Value must be a load.
7813 SDNode *Base = N->getOperand(0).getNode();
7814 if (!isa<LoadSDNode>(Base)) {
7815 if (Base->getOpcode() != ISD::BIT_CONVERT)
7817 Base = Base->getOperand(0).getNode();
7818 if (!isa<LoadSDNode>(Base))
7822 // Transform it into VZEXT_LOAD addr.
7823 LoadSDNode *LD = cast<LoadSDNode>(Base);
7825 // Load must not be an extload.
7826 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
7829 // Load type should legal type so we don't have to legalize it.
7830 if (!TLI.isTypeLegal(VT))
7833 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
7834 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
7835 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
7836 TargetLowering::TargetLoweringOpt TLO(DAG);
7837 TLO.CombineTo(SDValue(Base, 1), ResNode.getValue(1));
7838 DCI.CommitTargetLoweringOpt(TLO);
7842 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
7843 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
7844 const X86Subtarget *Subtarget) {
7845 DebugLoc DL = N->getDebugLoc();
7846 SDValue Cond = N->getOperand(0);
7847 // Get the LHS/RHS of the select.
7848 SDValue LHS = N->getOperand(1);
7849 SDValue RHS = N->getOperand(2);
7851 // If we have SSE[12] support, try to form min/max nodes.
7852 if (Subtarget->hasSSE2() &&
7853 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
7854 Cond.getOpcode() == ISD::SETCC) {
7855 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
7857 unsigned Opcode = 0;
7858 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
7861 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
7864 if (!UnsafeFPMath) break;
7866 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
7868 Opcode = X86ISD::FMIN;
7871 case ISD::SETOGT: // (X > Y) ? X : Y -> max
7874 if (!UnsafeFPMath) break;
7876 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
7878 Opcode = X86ISD::FMAX;
7881 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
7884 case ISD::SETOGT: // (X > Y) ? Y : X -> min
7887 if (!UnsafeFPMath) break;
7889 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
7891 Opcode = X86ISD::FMIN;
7894 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
7897 if (!UnsafeFPMath) break;
7899 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
7901 Opcode = X86ISD::FMAX;
7907 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
7910 // If this is a select between two integer constants, try to do some
7912 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
7913 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
7914 // Don't do this for crazy integer types.
7915 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
7916 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
7917 // so that TrueC (the true value) is larger than FalseC.
7918 bool NeedsCondInvert = false;
7920 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
7921 // Efficiently invertible.
7922 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
7923 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
7924 isa<ConstantSDNode>(Cond.getOperand(1))))) {
7925 NeedsCondInvert = true;
7926 std::swap(TrueC, FalseC);
7929 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
7930 if (FalseC->getAPIntValue() == 0 &&
7931 TrueC->getAPIntValue().isPowerOf2()) {
7932 if (NeedsCondInvert) // Invert the condition if needed.
7933 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
7934 DAG.getConstant(1, Cond.getValueType()));
7936 // Zero extend the condition if needed.
7937 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
7939 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
7940 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
7941 DAG.getConstant(ShAmt, MVT::i8));
7944 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
7945 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
7946 if (NeedsCondInvert) // Invert the condition if needed.
7947 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
7948 DAG.getConstant(1, Cond.getValueType()));
7950 // Zero extend the condition if needed.
7951 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
7952 FalseC->getValueType(0), Cond);
7953 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
7954 SDValue(FalseC, 0));
7957 // Optimize cases that will turn into an LEA instruction. This requires
7958 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
7959 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
7960 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
7961 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
7963 bool isFastMultiplier = false;
7965 switch ((unsigned char)Diff) {
7967 case 1: // result = add base, cond
7968 case 2: // result = lea base( , cond*2)
7969 case 3: // result = lea base(cond, cond*2)
7970 case 4: // result = lea base( , cond*4)
7971 case 5: // result = lea base(cond, cond*4)
7972 case 8: // result = lea base( , cond*8)
7973 case 9: // result = lea base(cond, cond*8)
7974 isFastMultiplier = true;
7979 if (isFastMultiplier) {
7980 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
7981 if (NeedsCondInvert) // Invert the condition if needed.
7982 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
7983 DAG.getConstant(1, Cond.getValueType()));
7985 // Zero extend the condition if needed.
7986 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
7988 // Scale the condition by the difference.
7990 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
7991 DAG.getConstant(Diff, Cond.getValueType()));
7993 // Add the base if non-zero.
7994 if (FalseC->getAPIntValue() != 0)
7995 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
7996 SDValue(FalseC, 0));
8006 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8007 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8008 TargetLowering::DAGCombinerInfo &DCI) {
8009 DebugLoc DL = N->getDebugLoc();
8011 // If the flag operand isn't dead, don't touch this CMOV.
8012 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8015 // If this is a select between two integer constants, try to do some
8016 // optimizations. Note that the operands are ordered the opposite of SELECT
8018 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8019 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8020 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8021 // larger than FalseC (the false value).
8022 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8024 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8025 CC = X86::GetOppositeBranchCondition(CC);
8026 std::swap(TrueC, FalseC);
8029 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8030 // This is efficient for any integer data type (including i8/i16) and
8032 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8033 SDValue Cond = N->getOperand(3);
8034 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8035 DAG.getConstant(CC, MVT::i8), Cond);
8037 // Zero extend the condition if needed.
8038 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8040 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8041 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8042 DAG.getConstant(ShAmt, MVT::i8));
8043 if (N->getNumValues() == 2) // Dead flag value?
8044 return DCI.CombineTo(N, Cond, SDValue());
8048 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8049 // for any integer data type, including i8/i16.
8050 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8051 SDValue Cond = N->getOperand(3);
8052 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8053 DAG.getConstant(CC, MVT::i8), Cond);
8055 // Zero extend the condition if needed.
8056 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8057 FalseC->getValueType(0), Cond);
8058 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8059 SDValue(FalseC, 0));
8061 if (N->getNumValues() == 2) // Dead flag value?
8062 return DCI.CombineTo(N, Cond, SDValue());
8066 // Optimize cases that will turn into an LEA instruction. This requires
8067 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8068 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8069 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8070 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8072 bool isFastMultiplier = false;
8074 switch ((unsigned char)Diff) {
8076 case 1: // result = add base, cond
8077 case 2: // result = lea base( , cond*2)
8078 case 3: // result = lea base(cond, cond*2)
8079 case 4: // result = lea base( , cond*4)
8080 case 5: // result = lea base(cond, cond*4)
8081 case 8: // result = lea base( , cond*8)
8082 case 9: // result = lea base(cond, cond*8)
8083 isFastMultiplier = true;
8088 if (isFastMultiplier) {
8089 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8090 SDValue Cond = N->getOperand(3);
8091 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8092 DAG.getConstant(CC, MVT::i8), Cond);
8093 // Zero extend the condition if needed.
8094 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8096 // Scale the condition by the difference.
8098 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8099 DAG.getConstant(Diff, Cond.getValueType()));
8101 // Add the base if non-zero.
8102 if (FalseC->getAPIntValue() != 0)
8103 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8104 SDValue(FalseC, 0));
8105 if (N->getNumValues() == 2) // Dead flag value?
8106 return DCI.CombineTo(N, Cond, SDValue());
8116 /// PerformMulCombine - Optimize a single multiply with constant into two
8117 /// in order to implement it with two cheaper instructions, e.g.
8118 /// LEA + SHL, LEA + LEA.
8119 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8120 TargetLowering::DAGCombinerInfo &DCI) {
8121 if (DAG.getMachineFunction().
8122 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8125 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8128 MVT VT = N->getValueType(0);
8132 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8135 uint64_t MulAmt = C->getZExtValue();
8136 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8139 uint64_t MulAmt1 = 0;
8140 uint64_t MulAmt2 = 0;
8141 if ((MulAmt % 9) == 0) {
8143 MulAmt2 = MulAmt / 9;
8144 } else if ((MulAmt % 5) == 0) {
8146 MulAmt2 = MulAmt / 5;
8147 } else if ((MulAmt % 3) == 0) {
8149 MulAmt2 = MulAmt / 3;
8152 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8153 DebugLoc DL = N->getDebugLoc();
8155 if (isPowerOf2_64(MulAmt2) &&
8156 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8157 // If second multiplifer is pow2, issue it first. We want the multiply by
8158 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8160 std::swap(MulAmt1, MulAmt2);
8163 if (isPowerOf2_64(MulAmt1))
8164 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8165 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8167 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8168 DAG.getConstant(MulAmt1, VT));
8170 if (isPowerOf2_64(MulAmt2))
8171 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8172 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8174 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8175 DAG.getConstant(MulAmt2, VT));
8177 // Do not add new nodes to DAG combiner worklist.
8178 DCI.CombineTo(N, NewMul, false);
8184 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8186 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8187 const X86Subtarget *Subtarget) {
8188 // On X86 with SSE2 support, we can transform this to a vector shift if
8189 // all elements are shifted by the same amount. We can't do this in legalize
8190 // because the a constant vector is typically transformed to a constant pool
8191 // so we have no knowledge of the shift amount.
8192 if (!Subtarget->hasSSE2())
8195 MVT VT = N->getValueType(0);
8196 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8199 SDValue ShAmtOp = N->getOperand(1);
8200 MVT EltVT = VT.getVectorElementType();
8201 DebugLoc DL = N->getDebugLoc();
8203 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8204 unsigned NumElts = VT.getVectorNumElements();
8206 for (; i != NumElts; ++i) {
8207 SDValue Arg = ShAmtOp.getOperand(i);
8208 if (Arg.getOpcode() == ISD::UNDEF) continue;
8212 for (; i != NumElts; ++i) {
8213 SDValue Arg = ShAmtOp.getOperand(i);
8214 if (Arg.getOpcode() == ISD::UNDEF) continue;
8215 if (Arg != BaseShAmt) {
8219 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8220 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8221 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8222 DAG.getIntPtrConstant(0));
8226 if (EltVT.bitsGT(MVT::i32))
8227 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8228 else if (EltVT.bitsLT(MVT::i32))
8229 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8231 // The shift amount is identical so we can do a vector shift.
8232 SDValue ValOp = N->getOperand(0);
8233 switch (N->getOpcode()) {
8235 assert(0 && "Unknown shift opcode!");
8238 if (VT == MVT::v2i64)
8239 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8240 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8242 if (VT == MVT::v4i32)
8243 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8244 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8246 if (VT == MVT::v8i16)
8247 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8248 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8252 if (VT == MVT::v4i32)
8253 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8254 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8256 if (VT == MVT::v8i16)
8257 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8258 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8262 if (VT == MVT::v2i64)
8263 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8264 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8266 if (VT == MVT::v4i32)
8267 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8268 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8270 if (VT == MVT::v8i16)
8271 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8272 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8279 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8280 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8281 const X86Subtarget *Subtarget) {
8282 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8283 // the FP state in cases where an emms may be missing.
8284 // A preferable solution to the general problem is to figure out the right
8285 // places to insert EMMS. This qualifies as a quick hack.
8287 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8288 StoreSDNode *St = cast<StoreSDNode>(N);
8289 MVT VT = St->getValue().getValueType();
8290 if (VT.getSizeInBits() != 64)
8293 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloat && Subtarget->hasSSE2();
8294 if ((VT.isVector() ||
8295 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8296 isa<LoadSDNode>(St->getValue()) &&
8297 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8298 St->getChain().hasOneUse() && !St->isVolatile()) {
8299 SDNode* LdVal = St->getValue().getNode();
8301 int TokenFactorIndex = -1;
8302 SmallVector<SDValue, 8> Ops;
8303 SDNode* ChainVal = St->getChain().getNode();
8304 // Must be a store of a load. We currently handle two cases: the load
8305 // is a direct child, and it's under an intervening TokenFactor. It is
8306 // possible to dig deeper under nested TokenFactors.
8307 if (ChainVal == LdVal)
8308 Ld = cast<LoadSDNode>(St->getChain());
8309 else if (St->getValue().hasOneUse() &&
8310 ChainVal->getOpcode() == ISD::TokenFactor) {
8311 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8312 if (ChainVal->getOperand(i).getNode() == LdVal) {
8313 TokenFactorIndex = i;
8314 Ld = cast<LoadSDNode>(St->getValue());
8316 Ops.push_back(ChainVal->getOperand(i));
8320 if (!Ld || !ISD::isNormalLoad(Ld))
8323 // If this is not the MMX case, i.e. we are just turning i64 load/store
8324 // into f64 load/store, avoid the transformation if there are multiple
8325 // uses of the loaded value.
8326 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8329 DebugLoc LdDL = Ld->getDebugLoc();
8330 DebugLoc StDL = N->getDebugLoc();
8331 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8332 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8334 if (Subtarget->is64Bit() || F64IsLegal) {
8335 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8336 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8337 Ld->getBasePtr(), Ld->getSrcValue(),
8338 Ld->getSrcValueOffset(), Ld->isVolatile(),
8339 Ld->getAlignment());
8340 SDValue NewChain = NewLd.getValue(1);
8341 if (TokenFactorIndex != -1) {
8342 Ops.push_back(NewChain);
8343 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8346 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8347 St->getSrcValue(), St->getSrcValueOffset(),
8348 St->isVolatile(), St->getAlignment());
8351 // Otherwise, lower to two pairs of 32-bit loads / stores.
8352 SDValue LoAddr = Ld->getBasePtr();
8353 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8354 DAG.getConstant(4, MVT::i32));
8356 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8357 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8358 Ld->isVolatile(), Ld->getAlignment());
8359 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8360 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8362 MinAlign(Ld->getAlignment(), 4));
8364 SDValue NewChain = LoLd.getValue(1);
8365 if (TokenFactorIndex != -1) {
8366 Ops.push_back(LoLd);
8367 Ops.push_back(HiLd);
8368 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8372 LoAddr = St->getBasePtr();
8373 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8374 DAG.getConstant(4, MVT::i32));
8376 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8377 St->getSrcValue(), St->getSrcValueOffset(),
8378 St->isVolatile(), St->getAlignment());
8379 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8381 St->getSrcValueOffset() + 4,
8383 MinAlign(St->getAlignment(), 4));
8384 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8389 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8390 /// X86ISD::FXOR nodes.
8391 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8392 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8393 // F[X]OR(0.0, x) -> x
8394 // F[X]OR(x, 0.0) -> x
8395 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8396 if (C->getValueAPF().isPosZero())
8397 return N->getOperand(1);
8398 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8399 if (C->getValueAPF().isPosZero())
8400 return N->getOperand(0);
8404 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8405 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8406 // FAND(0.0, x) -> 0.0
8407 // FAND(x, 0.0) -> 0.0
8408 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8409 if (C->getValueAPF().isPosZero())
8410 return N->getOperand(0);
8411 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8412 if (C->getValueAPF().isPosZero())
8413 return N->getOperand(1);
8417 static SDValue PerformBTCombine(SDNode *N,
8419 TargetLowering::DAGCombinerInfo &DCI) {
8420 // BT ignores high bits in the bit index operand.
8421 SDValue Op1 = N->getOperand(1);
8422 if (Op1.hasOneUse()) {
8423 unsigned BitWidth = Op1.getValueSizeInBits();
8424 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8425 APInt KnownZero, KnownOne;
8426 TargetLowering::TargetLoweringOpt TLO(DAG);
8427 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8428 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8429 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8430 DCI.CommitTargetLoweringOpt(TLO);
8435 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8436 DAGCombinerInfo &DCI) const {
8437 SelectionDAG &DAG = DCI.DAG;
8438 switch (N->getOpcode()) {
8440 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8441 case ISD::BUILD_VECTOR:
8442 return PerformBuildVectorCombine(N, DAG, DCI, Subtarget, *this);
8443 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8444 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8445 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8448 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8449 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8451 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8452 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8453 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8459 //===----------------------------------------------------------------------===//
8460 // X86 Inline Assembly Support
8461 //===----------------------------------------------------------------------===//
8463 /// getConstraintType - Given a constraint letter, return the type of
8464 /// constraint it is for this target.
8465 X86TargetLowering::ConstraintType
8466 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8467 if (Constraint.size() == 1) {
8468 switch (Constraint[0]) {
8480 return C_RegisterClass;
8488 return TargetLowering::getConstraintType(Constraint);
8491 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8492 /// with another that has more specific requirements based on the type of the
8493 /// corresponding operand.
8494 const char *X86TargetLowering::
8495 LowerXConstraint(MVT ConstraintVT) const {
8496 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8497 // 'f' like normal targets.
8498 if (ConstraintVT.isFloatingPoint()) {
8499 if (Subtarget->hasSSE2())
8501 if (Subtarget->hasSSE1())
8505 return TargetLowering::LowerXConstraint(ConstraintVT);
8508 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8509 /// vector. If it is invalid, don't add anything to Ops.
8510 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8513 std::vector<SDValue>&Ops,
8514 SelectionDAG &DAG) const {
8515 SDValue Result(0, 0);
8517 switch (Constraint) {
8520 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8521 if (C->getZExtValue() <= 31) {
8522 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8528 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8529 if (C->getZExtValue() <= 63) {
8530 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8536 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8537 if (C->getZExtValue() <= 255) {
8538 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8544 // 32-bit signed value
8545 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8546 const ConstantInt *CI = C->getConstantIntValue();
8547 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8548 // Widen to 64 bits here to get it sign extended.
8549 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8552 // FIXME gcc accepts some relocatable values here too, but only in certain
8553 // memory models; it's complicated.
8558 // 32-bit unsigned value
8559 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8560 const ConstantInt *CI = C->getConstantIntValue();
8561 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8562 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8566 // FIXME gcc accepts some relocatable values here too, but only in certain
8567 // memory models; it's complicated.
8571 // Literal immediates are always ok.
8572 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8573 // Widen to 64 bits here to get it sign extended.
8574 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8578 // If we are in non-pic codegen mode, we allow the address of a global (with
8579 // an optional displacement) to be used with 'i'.
8580 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
8583 // Match either (GA) or (GA+C)
8585 Offset = GA->getOffset();
8586 } else if (Op.getOpcode() == ISD::ADD) {
8587 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8588 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
8590 Offset = GA->getOffset()+C->getZExtValue();
8592 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8593 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
8595 Offset = GA->getOffset()+C->getZExtValue();
8603 Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(),
8606 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
8612 // Otherwise, not valid for this mode.
8617 if (Result.getNode()) {
8618 Ops.push_back(Result);
8621 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8625 std::vector<unsigned> X86TargetLowering::
8626 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8628 if (Constraint.size() == 1) {
8629 // FIXME: not handling fp-stack yet!
8630 switch (Constraint[0]) { // GCC X86 Constraint Letters
8631 default: break; // Unknown constraint letter
8632 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
8635 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
8636 else if (VT == MVT::i16)
8637 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
8638 else if (VT == MVT::i8)
8639 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
8640 else if (VT == MVT::i64)
8641 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
8646 return std::vector<unsigned>();
8649 std::pair<unsigned, const TargetRegisterClass*>
8650 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8652 // First, see if this is a constraint that directly corresponds to an LLVM
8654 if (Constraint.size() == 1) {
8655 // GCC Constraint Letters
8656 switch (Constraint[0]) {
8658 case 'r': // GENERAL_REGS
8659 case 'R': // LEGACY_REGS
8660 case 'l': // INDEX_REGS
8662 return std::make_pair(0U, X86::GR8RegisterClass);
8664 return std::make_pair(0U, X86::GR16RegisterClass);
8665 if (VT == MVT::i32 || !Subtarget->is64Bit())
8666 return std::make_pair(0U, X86::GR32RegisterClass);
8667 return std::make_pair(0U, X86::GR64RegisterClass);
8668 case 'f': // FP Stack registers.
8669 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
8670 // value to the correct fpstack register class.
8671 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
8672 return std::make_pair(0U, X86::RFP32RegisterClass);
8673 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
8674 return std::make_pair(0U, X86::RFP64RegisterClass);
8675 return std::make_pair(0U, X86::RFP80RegisterClass);
8676 case 'y': // MMX_REGS if MMX allowed.
8677 if (!Subtarget->hasMMX()) break;
8678 return std::make_pair(0U, X86::VR64RegisterClass);
8679 case 'Y': // SSE_REGS if SSE2 allowed
8680 if (!Subtarget->hasSSE2()) break;
8682 case 'x': // SSE_REGS if SSE1 allowed
8683 if (!Subtarget->hasSSE1()) break;
8685 switch (VT.getSimpleVT()) {
8687 // Scalar SSE types.
8690 return std::make_pair(0U, X86::FR32RegisterClass);
8693 return std::make_pair(0U, X86::FR64RegisterClass);
8701 return std::make_pair(0U, X86::VR128RegisterClass);
8707 // Use the default implementation in TargetLowering to convert the register
8708 // constraint into a member of a register class.
8709 std::pair<unsigned, const TargetRegisterClass*> Res;
8710 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8712 // Not found as a standard register?
8713 if (Res.second == 0) {
8714 // GCC calls "st(0)" just plain "st".
8715 if (StringsEqualNoCase("{st}", Constraint)) {
8716 Res.first = X86::ST0;
8717 Res.second = X86::RFP80RegisterClass;
8719 // 'A' means EAX + EDX.
8720 if (Constraint == "A") {
8721 Res.first = X86::EAX;
8722 Res.second = X86::GRADRegisterClass;
8727 // Otherwise, check to see if this is a register class of the wrong value
8728 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
8729 // turn into {ax},{dx}.
8730 if (Res.second->hasType(VT))
8731 return Res; // Correct type already, nothing to do.
8733 // All of the single-register GCC register classes map their values onto
8734 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
8735 // really want an 8-bit or 32-bit register, map to the appropriate register
8736 // class and return the appropriate register.
8737 if (Res.second == X86::GR16RegisterClass) {
8738 if (VT == MVT::i8) {
8739 unsigned DestReg = 0;
8740 switch (Res.first) {
8742 case X86::AX: DestReg = X86::AL; break;
8743 case X86::DX: DestReg = X86::DL; break;
8744 case X86::CX: DestReg = X86::CL; break;
8745 case X86::BX: DestReg = X86::BL; break;
8748 Res.first = DestReg;
8749 Res.second = X86::GR8RegisterClass;
8751 } else if (VT == MVT::i32) {
8752 unsigned DestReg = 0;
8753 switch (Res.first) {
8755 case X86::AX: DestReg = X86::EAX; break;
8756 case X86::DX: DestReg = X86::EDX; break;
8757 case X86::CX: DestReg = X86::ECX; break;
8758 case X86::BX: DestReg = X86::EBX; break;
8759 case X86::SI: DestReg = X86::ESI; break;
8760 case X86::DI: DestReg = X86::EDI; break;
8761 case X86::BP: DestReg = X86::EBP; break;
8762 case X86::SP: DestReg = X86::ESP; break;
8765 Res.first = DestReg;
8766 Res.second = X86::GR32RegisterClass;
8768 } else if (VT == MVT::i64) {
8769 unsigned DestReg = 0;
8770 switch (Res.first) {
8772 case X86::AX: DestReg = X86::RAX; break;
8773 case X86::DX: DestReg = X86::RDX; break;
8774 case X86::CX: DestReg = X86::RCX; break;
8775 case X86::BX: DestReg = X86::RBX; break;
8776 case X86::SI: DestReg = X86::RSI; break;
8777 case X86::DI: DestReg = X86::RDI; break;
8778 case X86::BP: DestReg = X86::RBP; break;
8779 case X86::SP: DestReg = X86::RSP; break;
8782 Res.first = DestReg;
8783 Res.second = X86::GR64RegisterClass;
8786 } else if (Res.second == X86::FR32RegisterClass ||
8787 Res.second == X86::FR64RegisterClass ||
8788 Res.second == X86::VR128RegisterClass) {
8789 // Handle references to XMM physical registers that got mapped into the
8790 // wrong class. This can happen with constraints like {xmm0} where the
8791 // target independent register mapper will just pick the first match it can
8792 // find, ignoring the required type.
8794 Res.second = X86::FR32RegisterClass;
8795 else if (VT == MVT::f64)
8796 Res.second = X86::FR64RegisterClass;
8797 else if (X86::VR128RegisterClass->hasType(VT))
8798 Res.second = X86::VR128RegisterClass;
8804 //===----------------------------------------------------------------------===//
8805 // X86 Widen vector type
8806 //===----------------------------------------------------------------------===//
8808 /// getWidenVectorType: given a vector type, returns the type to widen
8809 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
8810 /// If there is no vector type that we want to widen to, returns MVT::Other
8811 /// When and where to widen is target dependent based on the cost of
8812 /// scalarizing vs using the wider vector type.
8814 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
8815 assert(VT.isVector());
8816 if (isTypeLegal(VT))
8819 // TODO: In computeRegisterProperty, we can compute the list of legal vector
8820 // type based on element type. This would speed up our search (though
8821 // it may not be worth it since the size of the list is relatively
8823 MVT EltVT = VT.getVectorElementType();
8824 unsigned NElts = VT.getVectorNumElements();
8826 // On X86, it make sense to widen any vector wider than 1
8830 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
8831 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
8832 MVT SVT = (MVT::SimpleValueType)nVT;
8834 if (isTypeLegal(SVT) &&
8835 SVT.getVectorElementType() == EltVT &&
8836 SVT.getVectorNumElements() > NElts)