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/Analysis/ScalarEvolutionExpressions.h"
29 #include "llvm/CodeGen/CallingConvLower.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/ADT/SmallSet.h"
39 #include "llvm/ADT/StringExtras.h"
40 #include "llvm/ParameterAttributes.h"
43 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
44 : TargetLowering(TM) {
45 Subtarget = &TM.getSubtarget<X86Subtarget>();
46 X86ScalarSSEf64 = Subtarget->hasSSE2();
47 X86ScalarSSEf32 = Subtarget->hasSSE1();
48 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
52 RegInfo = TM.getRegisterInfo();
54 // Set up the TargetLowering object.
56 // X86 is weird, it always uses i8 for shift amounts and setcc results.
57 setShiftAmountType(MVT::i8);
58 setSetCCResultType(MVT::i8);
59 setSetCCResultContents(ZeroOrOneSetCCResult);
60 setSchedulingPreference(SchedulingForRegPressure);
61 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
62 setStackPointerRegisterToSaveRestore(X86StackPtr);
64 if (Subtarget->isTargetDarwin()) {
65 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
66 setUseUnderscoreSetJmp(false);
67 setUseUnderscoreLongJmp(false);
68 } else if (Subtarget->isTargetMingw()) {
69 // MS runtime is weird: it exports _setjmp, but longjmp!
70 setUseUnderscoreSetJmp(true);
71 setUseUnderscoreLongJmp(false);
73 setUseUnderscoreSetJmp(true);
74 setUseUnderscoreLongJmp(true);
77 // Set up the register classes.
78 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
79 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
80 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
81 if (Subtarget->is64Bit())
82 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
84 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
86 // We don't accept any truncstore of integer registers.
87 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
88 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
89 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
90 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
91 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
92 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
94 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
96 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
97 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
98 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
100 if (Subtarget->is64Bit()) {
101 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
102 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
105 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
106 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
108 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
111 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
113 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
114 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
115 // SSE has no i16 to fp conversion, only i32
116 if (X86ScalarSSEf32) {
117 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
118 // f32 and f64 cases are Legal, f80 case is not
119 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
121 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
122 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
125 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
126 // are Legal, f80 is custom lowered.
127 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
128 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
130 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
132 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
133 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
135 if (X86ScalarSSEf32) {
136 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
137 // f32 and f64 cases are Legal, f80 case is not
138 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
140 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
141 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
144 // Handle FP_TO_UINT by promoting the destination to a larger signed
146 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
147 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
148 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
150 if (Subtarget->is64Bit()) {
151 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
152 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
154 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
155 // Expand FP_TO_UINT into a select.
156 // FIXME: We would like to use a Custom expander here eventually to do
157 // the optimal thing for SSE vs. the default expansion in the legalizer.
158 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
160 // With SSE3 we can use fisttpll to convert to a signed i64.
161 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
164 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
165 if (!X86ScalarSSEf64) {
166 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
167 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
170 // Scalar integer multiply, multiply-high, divide, and remainder are
171 // lowered to use operations that produce two results, to match the
172 // available instructions. This exposes the two-result form to trivial
173 // CSE, which is able to combine x/y and x%y into a single instruction,
174 // for example. The single-result multiply instructions are introduced
175 // in X86ISelDAGToDAG.cpp, after CSE, for uses where the the high part
177 setOperationAction(ISD::MUL , MVT::i8 , Expand);
178 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
179 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
180 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
181 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
182 setOperationAction(ISD::SREM , MVT::i8 , Expand);
183 setOperationAction(ISD::UREM , MVT::i8 , Expand);
184 setOperationAction(ISD::MUL , MVT::i16 , Expand);
185 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
186 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
187 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
188 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
189 setOperationAction(ISD::SREM , MVT::i16 , Expand);
190 setOperationAction(ISD::UREM , MVT::i16 , Expand);
191 setOperationAction(ISD::MUL , MVT::i32 , Expand);
192 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
193 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
194 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
195 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
196 setOperationAction(ISD::SREM , MVT::i32 , Expand);
197 setOperationAction(ISD::UREM , MVT::i32 , Expand);
198 setOperationAction(ISD::MUL , MVT::i64 , Expand);
199 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
200 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
201 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
202 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
203 setOperationAction(ISD::SREM , MVT::i64 , Expand);
204 setOperationAction(ISD::UREM , MVT::i64 , Expand);
206 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
207 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
208 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
209 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
210 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
211 if (Subtarget->is64Bit())
212 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
213 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
214 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
215 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
216 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
217 setOperationAction(ISD::FREM , MVT::f64 , Expand);
218 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
220 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
221 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
222 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
223 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
224 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
225 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
226 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
227 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
228 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
229 if (Subtarget->is64Bit()) {
230 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
231 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
232 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
235 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
236 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
238 // These should be promoted to a larger select which is supported.
239 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
240 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
241 // X86 wants to expand cmov itself.
242 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
243 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
244 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
245 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
246 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
247 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
248 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
249 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
250 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
251 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
252 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
253 if (Subtarget->is64Bit()) {
254 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
255 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
257 // X86 ret instruction may pop stack.
258 setOperationAction(ISD::RET , MVT::Other, Custom);
259 if (!Subtarget->is64Bit())
260 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
263 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
264 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
265 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
266 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
267 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
268 if (Subtarget->is64Bit()) {
269 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
270 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
271 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
272 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
274 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
275 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
276 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
277 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
278 // X86 wants to expand memset / memcpy itself.
279 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
280 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
282 // Use the default ISD::LOCATION expansion.
283 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
284 // FIXME - use subtarget debug flags
285 if (!Subtarget->isTargetDarwin() &&
286 !Subtarget->isTargetELF() &&
287 !Subtarget->isTargetCygMing())
288 setOperationAction(ISD::LABEL, MVT::Other, Expand);
290 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
291 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
292 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
293 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
294 if (Subtarget->is64Bit()) {
296 setExceptionPointerRegister(X86::RAX);
297 setExceptionSelectorRegister(X86::RDX);
299 setExceptionPointerRegister(X86::EAX);
300 setExceptionSelectorRegister(X86::EDX);
302 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
304 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
306 setOperationAction(ISD::TRAP, MVT::Other, Legal);
308 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
309 setOperationAction(ISD::VASTART , MVT::Other, Custom);
310 setOperationAction(ISD::VAARG , MVT::Other, Expand);
311 setOperationAction(ISD::VAEND , MVT::Other, Expand);
312 if (Subtarget->is64Bit())
313 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
315 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
317 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
318 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
319 if (Subtarget->is64Bit())
320 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
321 if (Subtarget->isTargetCygMing())
322 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
324 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
326 if (X86ScalarSSEf64) {
327 // f32 and f64 use SSE.
328 // Set up the FP register classes.
329 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
330 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
332 // Use ANDPD to simulate FABS.
333 setOperationAction(ISD::FABS , MVT::f64, Custom);
334 setOperationAction(ISD::FABS , MVT::f32, Custom);
336 // Use XORP to simulate FNEG.
337 setOperationAction(ISD::FNEG , MVT::f64, Custom);
338 setOperationAction(ISD::FNEG , MVT::f32, Custom);
340 // Use ANDPD and ORPD to simulate FCOPYSIGN.
341 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
342 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
344 // We don't support sin/cos/fmod
345 setOperationAction(ISD::FSIN , MVT::f64, Expand);
346 setOperationAction(ISD::FCOS , MVT::f64, Expand);
347 setOperationAction(ISD::FREM , MVT::f64, Expand);
348 setOperationAction(ISD::FSIN , MVT::f32, Expand);
349 setOperationAction(ISD::FCOS , MVT::f32, Expand);
350 setOperationAction(ISD::FREM , MVT::f32, Expand);
352 // Expand FP immediates into loads from the stack, except for the special
354 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
355 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
356 addLegalFPImmediate(APFloat(+0.0)); // xorpd
357 addLegalFPImmediate(APFloat(+0.0f)); // xorps
359 // Floating truncations from f80 and extensions to f80 go through memory.
360 // If optimizing, we lie about this though and handle it in
361 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
363 setConvertAction(MVT::f32, MVT::f80, Expand);
364 setConvertAction(MVT::f64, MVT::f80, Expand);
365 setConvertAction(MVT::f80, MVT::f32, Expand);
366 setConvertAction(MVT::f80, MVT::f64, Expand);
368 } else if (X86ScalarSSEf32) {
369 // Use SSE for f32, x87 for f64.
370 // Set up the FP register classes.
371 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
372 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
374 // Use ANDPS to simulate FABS.
375 setOperationAction(ISD::FABS , MVT::f32, Custom);
377 // Use XORP to simulate FNEG.
378 setOperationAction(ISD::FNEG , MVT::f32, Custom);
380 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
382 // Use ANDPS and ORPS to simulate FCOPYSIGN.
383 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
384 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
386 // We don't support sin/cos/fmod
387 setOperationAction(ISD::FSIN , MVT::f32, Expand);
388 setOperationAction(ISD::FCOS , MVT::f32, Expand);
389 setOperationAction(ISD::FREM , MVT::f32, Expand);
391 // Expand FP immediates into loads from the stack, except for the special
393 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
394 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
395 addLegalFPImmediate(APFloat(+0.0f)); // xorps
396 addLegalFPImmediate(APFloat(+0.0)); // FLD0
397 addLegalFPImmediate(APFloat(+1.0)); // FLD1
398 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
399 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
401 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
402 // this though and handle it in InstructionSelectPreprocess so that
403 // dagcombine2 can hack on these.
405 setConvertAction(MVT::f32, MVT::f64, Expand);
406 setConvertAction(MVT::f32, MVT::f80, Expand);
407 setConvertAction(MVT::f80, MVT::f32, Expand);
408 setConvertAction(MVT::f64, MVT::f32, Expand);
409 // And x87->x87 truncations also.
410 setConvertAction(MVT::f80, MVT::f64, Expand);
414 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
415 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
418 // f32 and f64 in x87.
419 // Set up the FP register classes.
420 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
421 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
423 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
424 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
425 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
426 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
428 // Floating truncations go through memory. If optimizing, we lie about
429 // this though and handle it in InstructionSelectPreprocess so that
430 // dagcombine2 can hack on these.
432 setConvertAction(MVT::f80, MVT::f32, Expand);
433 setConvertAction(MVT::f64, MVT::f32, Expand);
434 setConvertAction(MVT::f80, MVT::f64, Expand);
438 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
439 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
442 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
443 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
444 addLegalFPImmediate(APFloat(+0.0)); // FLD0
445 addLegalFPImmediate(APFloat(+1.0)); // FLD1
446 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
447 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
448 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
449 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
450 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
451 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
454 // Long double always uses X87.
455 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
456 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
457 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
459 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
460 APFloat TmpFlt(+0.0);
461 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
462 addLegalFPImmediate(TmpFlt); // FLD0
464 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
465 APFloat TmpFlt2(+1.0);
466 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
467 addLegalFPImmediate(TmpFlt2); // FLD1
468 TmpFlt2.changeSign();
469 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
473 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
474 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
477 // Always use a library call for pow.
478 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
479 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
480 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
482 // First set operation action for all vector types to expand. Then we
483 // will selectively turn on ones that can be effectively codegen'd.
484 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
485 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
486 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
487 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
488 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
489 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
490 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
491 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
492 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
493 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
494 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
495 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
496 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
497 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
498 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
499 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
500 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
501 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
502 setOperationAction(ISD::FABS, (MVT::ValueType)VT, Expand);
503 setOperationAction(ISD::FSIN, (MVT::ValueType)VT, Expand);
504 setOperationAction(ISD::FCOS, (MVT::ValueType)VT, Expand);
505 setOperationAction(ISD::FREM, (MVT::ValueType)VT, Expand);
506 setOperationAction(ISD::FPOWI, (MVT::ValueType)VT, Expand);
507 setOperationAction(ISD::FSQRT, (MVT::ValueType)VT, Expand);
508 setOperationAction(ISD::FCOPYSIGN, (MVT::ValueType)VT, Expand);
509 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
510 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
511 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
512 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
513 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
514 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
515 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
516 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
517 setOperationAction(ISD::SHL, (MVT::ValueType)VT, Expand);
518 setOperationAction(ISD::SRA, (MVT::ValueType)VT, Expand);
519 setOperationAction(ISD::SRL, (MVT::ValueType)VT, Expand);
520 setOperationAction(ISD::ROTL, (MVT::ValueType)VT, Expand);
521 setOperationAction(ISD::ROTR, (MVT::ValueType)VT, Expand);
522 setOperationAction(ISD::BSWAP, (MVT::ValueType)VT, Expand);
525 if (Subtarget->hasMMX()) {
526 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
527 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
528 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
529 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
531 // FIXME: add MMX packed arithmetics
533 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
534 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
535 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
536 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
538 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
539 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
540 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
541 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
543 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
544 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
546 setOperationAction(ISD::AND, MVT::v8i8, Promote);
547 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
548 setOperationAction(ISD::AND, MVT::v4i16, Promote);
549 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
550 setOperationAction(ISD::AND, MVT::v2i32, Promote);
551 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
552 setOperationAction(ISD::AND, MVT::v1i64, Legal);
554 setOperationAction(ISD::OR, MVT::v8i8, Promote);
555 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
556 setOperationAction(ISD::OR, MVT::v4i16, Promote);
557 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
558 setOperationAction(ISD::OR, MVT::v2i32, Promote);
559 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
560 setOperationAction(ISD::OR, MVT::v1i64, Legal);
562 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
563 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
564 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
565 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
566 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
567 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
568 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
570 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
571 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
572 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
573 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
574 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
575 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
576 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
578 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
579 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
580 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
581 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
583 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
584 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
585 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
586 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
588 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
589 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
590 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Custom);
591 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
594 if (Subtarget->hasSSE1()) {
595 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
597 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
598 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
599 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
600 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
601 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
602 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
603 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
604 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
605 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
606 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
607 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
610 if (Subtarget->hasSSE2()) {
611 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
612 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
613 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
614 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
615 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
617 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
618 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
619 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
620 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
621 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
622 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
623 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
624 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
625 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
626 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
627 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
628 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
629 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
630 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
631 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
633 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
634 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
635 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
636 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
637 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
638 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
640 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
641 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
642 // Do not attempt to custom lower non-power-of-2 vectors
643 if (!isPowerOf2_32(MVT::getVectorNumElements(VT)))
645 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
646 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
647 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
649 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
650 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
651 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
652 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
653 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
654 if (Subtarget->is64Bit())
655 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
657 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
658 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
659 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
660 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
661 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
662 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
663 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
664 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
665 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
666 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
667 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
668 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
671 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
673 // Custom lower v2i64 and v2f64 selects.
674 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
675 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
676 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
677 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
680 // We want to custom lower some of our intrinsics.
681 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
683 // We have target-specific dag combine patterns for the following nodes:
684 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
685 setTargetDAGCombine(ISD::SELECT);
687 computeRegisterProperties();
689 // FIXME: These should be based on subtarget info. Plus, the values should
690 // be smaller when we are in optimizing for size mode.
691 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
692 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
693 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
694 allowUnalignedMemoryAccesses = true; // x86 supports it!
697 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
698 /// the desired ByVal argument alignment.
699 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
702 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
703 if (VTy->getBitWidth() == 128)
705 else if (VTy->getBitWidth() == 64)
708 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
709 unsigned EltAlign = 0;
710 getMaxByValAlign(ATy->getElementType(), EltAlign);
711 if (EltAlign > MaxAlign)
713 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
714 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
715 unsigned EltAlign = 0;
716 getMaxByValAlign(STy->getElementType(i), EltAlign);
717 if (EltAlign > MaxAlign)
726 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
727 /// function arguments in the caller parameter area. For X86, aggregates
728 /// that contains are placed at 16-byte boundaries while the rest are at
729 /// 4-byte boundaries.
730 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
731 if (Subtarget->is64Bit())
732 return getTargetData()->getABITypeAlignment(Ty);
734 getMaxByValAlign(Ty, Align);
738 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
740 SDOperand X86TargetLowering::getPICJumpTableRelocBase(SDOperand Table,
741 SelectionDAG &DAG) const {
742 if (usesGlobalOffsetTable())
743 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
744 if (!Subtarget->isPICStyleRIPRel())
745 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
749 //===----------------------------------------------------------------------===//
750 // Return Value Calling Convention Implementation
751 //===----------------------------------------------------------------------===//
753 #include "X86GenCallingConv.inc"
755 /// GetPossiblePreceedingTailCall - Get preceeding X86ISD::TAILCALL node if it
756 /// exists skip possible ISD:TokenFactor.
757 static SDOperand GetPossiblePreceedingTailCall(SDOperand Chain) {
758 if (Chain.getOpcode() == X86ISD::TAILCALL) {
760 } else if (Chain.getOpcode() == ISD::TokenFactor) {
761 if (Chain.getNumOperands() &&
762 Chain.getOperand(0).getOpcode() == X86ISD::TAILCALL)
763 return Chain.getOperand(0);
768 /// LowerRET - Lower an ISD::RET node.
769 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
770 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
772 SmallVector<CCValAssign, 16> RVLocs;
773 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
774 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
775 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
776 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
778 // If this is the first return lowered for this function, add the regs to the
779 // liveout set for the function.
780 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
781 for (unsigned i = 0; i != RVLocs.size(); ++i)
782 if (RVLocs[i].isRegLoc())
783 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
785 SDOperand Chain = Op.getOperand(0);
787 // Handle tail call return.
788 Chain = GetPossiblePreceedingTailCall(Chain);
789 if (Chain.getOpcode() == X86ISD::TAILCALL) {
790 SDOperand TailCall = Chain;
791 SDOperand TargetAddress = TailCall.getOperand(1);
792 SDOperand StackAdjustment = TailCall.getOperand(2);
793 assert(((TargetAddress.getOpcode() == ISD::Register &&
794 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
795 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
796 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
797 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
798 "Expecting an global address, external symbol, or register");
799 assert(StackAdjustment.getOpcode() == ISD::Constant &&
800 "Expecting a const value");
802 SmallVector<SDOperand,8> Operands;
803 Operands.push_back(Chain.getOperand(0));
804 Operands.push_back(TargetAddress);
805 Operands.push_back(StackAdjustment);
806 // Copy registers used by the call. Last operand is a flag so it is not
808 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
809 Operands.push_back(Chain.getOperand(i));
811 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
818 // Copy the result values into the output registers.
819 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
820 RVLocs[0].getLocReg() != X86::ST0) {
821 for (unsigned i = 0; i != RVLocs.size(); ++i) {
822 CCValAssign &VA = RVLocs[i];
823 assert(VA.isRegLoc() && "Can only return in registers!");
824 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
826 Flag = Chain.getValue(1);
829 // We need to handle a destination of ST0 specially, because it isn't really
831 SDOperand Value = Op.getOperand(1);
833 // an XMM register onto the fp-stack. Do this with an FP_EXTEND to f80.
834 // This will get legalized into a load/store if it can't get optimized away.
835 if (isScalarFPTypeInSSEReg(RVLocs[0].getValVT()))
836 Value = DAG.getNode(ISD::FP_EXTEND, MVT::f80, Value);
838 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
839 SDOperand Ops[] = { Chain, Value };
840 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
841 Flag = Chain.getValue(1);
844 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
846 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
848 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
852 /// LowerCallResult - Lower the result values of an ISD::CALL into the
853 /// appropriate copies out of appropriate physical registers. This assumes that
854 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
855 /// being lowered. The returns a SDNode with the same number of values as the
857 SDNode *X86TargetLowering::
858 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
859 unsigned CallingConv, SelectionDAG &DAG) {
861 // Assign locations to each value returned by this call.
862 SmallVector<CCValAssign, 16> RVLocs;
863 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
864 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
865 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
867 SmallVector<SDOperand, 8> ResultVals;
869 // Copy all of the result registers out of their specified physreg.
870 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
871 for (unsigned i = 0; i != RVLocs.size(); ++i) {
872 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
873 RVLocs[i].getValVT(), InFlag).getValue(1);
874 InFlag = Chain.getValue(2);
875 ResultVals.push_back(Chain.getValue(0));
878 // Copies from the FP stack are special, as ST0 isn't a valid register
879 // before the fp stackifier runs.
881 // Copy ST0 into an RFP register with FP_GET_RESULT. If this will end up
882 // in an SSE register, copy it out as F80 and do a truncate, otherwise use
883 // the specified value type.
884 MVT::ValueType GetResultTy = RVLocs[0].getValVT();
885 if (isScalarFPTypeInSSEReg(GetResultTy))
886 GetResultTy = MVT::f80;
887 SDVTList Tys = DAG.getVTList(GetResultTy, MVT::Other, MVT::Flag);
889 SDOperand GROps[] = { Chain, InFlag };
890 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
891 Chain = RetVal.getValue(1);
892 InFlag = RetVal.getValue(2);
894 // If we want the result in an SSE register, use an FP_TRUNCATE to get it
896 if (GetResultTy != RVLocs[0].getValVT())
897 RetVal = DAG.getNode(ISD::FP_ROUND, RVLocs[0].getValVT(), RetVal,
898 // This truncation won't change the value.
899 DAG.getIntPtrConstant(1));
901 ResultVals.push_back(RetVal);
904 // Merge everything together with a MERGE_VALUES node.
905 ResultVals.push_back(Chain);
906 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
907 &ResultVals[0], ResultVals.size()).Val;
910 /// LowerCallResultToTwo64BitRegs - Lower the result values of an x86-64
911 /// ISD::CALL where the results are known to be in two 64-bit registers,
912 /// e.g. XMM0 and XMM1. This simplify store the two values back to the
913 /// fixed stack slot allocated for StructRet.
914 SDNode *X86TargetLowering::
915 LowerCallResultToTwo64BitRegs(SDOperand Chain, SDOperand InFlag,
916 SDNode *TheCall, unsigned Reg1, unsigned Reg2,
917 MVT::ValueType VT, SelectionDAG &DAG) {
918 SDOperand RetVal1 = DAG.getCopyFromReg(Chain, Reg1, VT, InFlag);
919 Chain = RetVal1.getValue(1);
920 InFlag = RetVal1.getValue(2);
921 SDOperand RetVal2 = DAG.getCopyFromReg(Chain, Reg2, VT, InFlag);
922 Chain = RetVal2.getValue(1);
923 InFlag = RetVal2.getValue(2);
924 SDOperand FIN = TheCall->getOperand(5);
925 Chain = DAG.getStore(Chain, RetVal1, FIN, NULL, 0);
926 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
927 Chain = DAG.getStore(Chain, RetVal2, FIN, NULL, 0);
931 /// LowerCallResultToTwoX87Regs - Lower the result values of an x86-64 ISD::CALL
932 /// where the results are known to be in ST0 and ST1.
933 SDNode *X86TargetLowering::
934 LowerCallResultToTwoX87Regs(SDOperand Chain, SDOperand InFlag,
935 SDNode *TheCall, SelectionDAG &DAG) {
936 SmallVector<SDOperand, 8> ResultVals;
937 const MVT::ValueType VTs[] = { MVT::f80, MVT::f80, MVT::Other, MVT::Flag };
938 SDVTList Tys = DAG.getVTList(VTs, 4);
939 SDOperand Ops[] = { Chain, InFlag };
940 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT2, Tys, Ops, 2);
941 Chain = RetVal.getValue(2);
942 SDOperand FIN = TheCall->getOperand(5);
943 Chain = DAG.getStore(Chain, RetVal.getValue(1), FIN, NULL, 0);
944 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(16));
945 Chain = DAG.getStore(Chain, RetVal, FIN, NULL, 0);
949 //===----------------------------------------------------------------------===//
950 // C & StdCall & Fast Calling Convention implementation
951 //===----------------------------------------------------------------------===//
952 // StdCall calling convention seems to be standard for many Windows' API
953 // routines and around. It differs from C calling convention just a little:
954 // callee should clean up the stack, not caller. Symbols should be also
955 // decorated in some fancy way :) It doesn't support any vector arguments.
956 // For info on fast calling convention see Fast Calling Convention (tail call)
957 // implementation LowerX86_32FastCCCallTo.
959 /// AddLiveIn - This helper function adds the specified physical register to the
960 /// MachineFunction as a live in value. It also creates a corresponding virtual
962 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
963 const TargetRegisterClass *RC) {
964 assert(RC->contains(PReg) && "Not the correct regclass!");
965 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
966 MF.getRegInfo().addLiveIn(PReg, VReg);
970 // Determines whether a CALL node uses struct return semantics.
971 static bool CallIsStructReturn(SDOperand Op) {
972 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
976 ConstantSDNode *Flags = cast<ConstantSDNode>(Op.getOperand(6));
977 return Flags->getValue() & ISD::ParamFlags::StructReturn;
980 // Determines whether a FORMAL_ARGUMENTS node uses struct return semantics.
981 static bool ArgsAreStructReturn(SDOperand Op) {
982 unsigned NumArgs = Op.Val->getNumValues() - 1;
986 ConstantSDNode *Flags = cast<ConstantSDNode>(Op.getOperand(3));
987 return Flags->getValue() & ISD::ParamFlags::StructReturn;
990 // Determines whether a CALL or FORMAL_ARGUMENTS node requires the callee to pop
991 // its own arguments. Callee pop is necessary to support tail calls.
992 bool X86TargetLowering::IsCalleePop(SDOperand Op) {
993 bool IsVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
997 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
1000 case CallingConv::X86_StdCall:
1001 return !Subtarget->is64Bit();
1002 case CallingConv::X86_FastCall:
1003 return !Subtarget->is64Bit();
1004 case CallingConv::Fast:
1005 return PerformTailCallOpt;
1009 // Selects the correct CCAssignFn for a CALL or FORMAL_ARGUMENTS node.
1010 CCAssignFn *X86TargetLowering::CCAssignFnForNode(SDOperand Op) const {
1011 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1013 if (Subtarget->is64Bit())
1014 if (CC == CallingConv::Fast && PerformTailCallOpt)
1015 return CC_X86_64_TailCall;
1019 if (CC == CallingConv::X86_FastCall)
1020 return CC_X86_32_FastCall;
1021 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1022 return CC_X86_32_TailCall;
1027 // Selects the appropriate decoration to apply to a MachineFunction containing a
1028 // given FORMAL_ARGUMENTS node.
1030 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDOperand Op) {
1031 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1032 if (CC == CallingConv::X86_FastCall)
1034 else if (CC == CallingConv::X86_StdCall)
1040 // IsPossiblyOverwrittenArgumentOfTailCall - Check if the operand could possibly
1041 // be overwritten when lowering the outgoing arguments in a tail call. Currently
1042 // the implementation of this call is very conservative and assumes all
1043 // arguments sourcing from FORMAL_ARGUMENTS or a CopyFromReg with virtual
1044 // registers would be overwritten by direct lowering.
1045 // Possible improvement:
1046 // Check FORMAL_ARGUMENTS corresponding MERGE_VALUES for CopyFromReg nodes
1047 // indicating inreg passed arguments which also need not be lowered to a safe
1049 static bool IsPossiblyOverwrittenArgumentOfTailCall(SDOperand Op) {
1050 RegisterSDNode * OpReg = NULL;
1051 if (Op.getOpcode() == ISD::FORMAL_ARGUMENTS ||
1052 (Op.getOpcode()== ISD::CopyFromReg &&
1053 (OpReg = cast<RegisterSDNode>(Op.getOperand(1))) &&
1054 OpReg->getReg() >= MRegisterInfo::FirstVirtualRegister))
1059 // CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1060 // by "Src" to address "Dst" with size and alignment information specified by
1061 // the specific parameter attribute. The copy will be passed as a byval function
1064 CreateCopyOfByValArgument(SDOperand Src, SDOperand Dst, SDOperand Chain,
1065 unsigned Flags, SelectionDAG &DAG) {
1066 unsigned Align = 1 <<
1067 ((Flags & ISD::ParamFlags::ByValAlign) >> ISD::ParamFlags::ByValAlignOffs);
1068 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1069 ISD::ParamFlags::ByValSizeOffs;
1070 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1071 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1072 SDOperand AlwaysInline = DAG.getConstant(1, MVT::i32);
1073 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, AlignNode, AlwaysInline);
1076 SDOperand X86TargetLowering::LowerMemArgument(SDOperand Op, SelectionDAG &DAG,
1077 const CCValAssign &VA,
1078 MachineFrameInfo *MFI,
1079 SDOperand Root, unsigned i) {
1080 // Create the nodes corresponding to a load from this parameter slot.
1081 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3 + i))->getValue();
1082 bool isByVal = Flags & ISD::ParamFlags::ByVal;
1084 // FIXME: For now, all byval parameter objects are marked mutable. This
1085 // can be changed with more analysis.
1086 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1087 VA.getLocMemOffset(), !isByVal);
1088 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1091 return DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0);
1095 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
1096 MachineFunction &MF = DAG.getMachineFunction();
1097 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1099 const Function* Fn = MF.getFunction();
1100 if (Fn->hasExternalLinkage() &&
1101 Subtarget->isTargetCygMing() &&
1102 Fn->getName() == "main")
1103 FuncInfo->setForceFramePointer(true);
1105 // Decorate the function name.
1106 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1108 MachineFrameInfo *MFI = MF.getFrameInfo();
1109 SDOperand Root = Op.getOperand(0);
1110 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1111 unsigned CC = MF.getFunction()->getCallingConv();
1112 bool Is64Bit = Subtarget->is64Bit();
1114 assert(!(isVarArg && CC == CallingConv::Fast) &&
1115 "Var args not supported with calling convention fastcc");
1117 // Assign locations to all of the incoming arguments.
1118 SmallVector<CCValAssign, 16> ArgLocs;
1119 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1120 CCInfo.AnalyzeFormalArguments(Op.Val, CCAssignFnForNode(Op));
1122 SmallVector<SDOperand, 8> ArgValues;
1123 unsigned LastVal = ~0U;
1124 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1125 CCValAssign &VA = ArgLocs[i];
1126 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1128 assert(VA.getValNo() != LastVal &&
1129 "Don't support value assigned to multiple locs yet");
1130 LastVal = VA.getValNo();
1132 if (VA.isRegLoc()) {
1133 MVT::ValueType RegVT = VA.getLocVT();
1134 TargetRegisterClass *RC;
1135 if (RegVT == MVT::i32)
1136 RC = X86::GR32RegisterClass;
1137 else if (Is64Bit && RegVT == MVT::i64)
1138 RC = X86::GR64RegisterClass;
1139 else if (Is64Bit && RegVT == MVT::f32)
1140 RC = X86::FR32RegisterClass;
1141 else if (Is64Bit && RegVT == MVT::f64)
1142 RC = X86::FR64RegisterClass;
1144 assert(MVT::isVector(RegVT));
1145 if (Is64Bit && MVT::getSizeInBits(RegVT) == 64) {
1146 RC = X86::GR64RegisterClass; // MMX values are passed in GPRs.
1149 RC = X86::VR128RegisterClass;
1152 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1153 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1155 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1156 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1158 if (VA.getLocInfo() == CCValAssign::SExt)
1159 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1160 DAG.getValueType(VA.getValVT()));
1161 else if (VA.getLocInfo() == CCValAssign::ZExt)
1162 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1163 DAG.getValueType(VA.getValVT()));
1165 if (VA.getLocInfo() != CCValAssign::Full)
1166 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1168 // Handle MMX values passed in GPRs.
1169 if (Is64Bit && RegVT != VA.getLocVT() && RC == X86::GR64RegisterClass &&
1170 MVT::getSizeInBits(RegVT) == 64)
1171 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1173 ArgValues.push_back(ArgValue);
1175 assert(VA.isMemLoc());
1176 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1180 unsigned StackSize = CCInfo.getNextStackOffset();
1181 // align stack specially for tail calls
1182 if (CC == CallingConv::Fast)
1183 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1185 // If the function takes variable number of arguments, make a frame index for
1186 // the start of the first vararg value... for expansion of llvm.va_start.
1188 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1189 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1192 static const unsigned GPR64ArgRegs[] = {
1193 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1195 static const unsigned XMMArgRegs[] = {
1196 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1197 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1200 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1201 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1203 // For X86-64, if there are vararg parameters that are passed via
1204 // registers, then we must store them to their spots on the stack so they
1205 // may be loaded by deferencing the result of va_next.
1206 VarArgsGPOffset = NumIntRegs * 8;
1207 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1208 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1210 // Store the integer parameter registers.
1211 SmallVector<SDOperand, 8> MemOps;
1212 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1213 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1214 DAG.getIntPtrConstant(VarArgsGPOffset));
1215 for (; NumIntRegs != 6; ++NumIntRegs) {
1216 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1217 X86::GR64RegisterClass);
1218 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1219 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1220 MemOps.push_back(Store);
1221 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1222 DAG.getIntPtrConstant(8));
1225 // Now store the XMM (fp + vector) parameter registers.
1226 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1227 DAG.getIntPtrConstant(VarArgsFPOffset));
1228 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1229 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1230 X86::VR128RegisterClass);
1231 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1232 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1233 MemOps.push_back(Store);
1234 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1235 DAG.getIntPtrConstant(16));
1237 if (!MemOps.empty())
1238 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1239 &MemOps[0], MemOps.size());
1243 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1244 // arguments and the arguments after the retaddr has been pushed are
1246 if (!Is64Bit && CC == CallingConv::X86_FastCall &&
1247 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows() &&
1248 (StackSize & 7) == 0)
1251 ArgValues.push_back(Root);
1253 // Some CCs need callee pop.
1254 if (IsCalleePop(Op)) {
1255 BytesToPopOnReturn = StackSize; // Callee pops everything.
1256 BytesCallerReserves = 0;
1258 BytesToPopOnReturn = 0; // Callee pops nothing.
1259 // If this is an sret function, the return should pop the hidden pointer.
1260 if (!Is64Bit && ArgsAreStructReturn(Op))
1261 BytesToPopOnReturn = 4;
1262 BytesCallerReserves = StackSize;
1266 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1267 if (CC == CallingConv::X86_FastCall)
1268 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1271 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1273 // Return the new list of results.
1274 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1275 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1279 X86TargetLowering::LowerMemOpCallTo(SDOperand Op, SelectionDAG &DAG,
1280 const SDOperand &StackPtr,
1281 const CCValAssign &VA,
1284 SDOperand PtrOff = DAG.getIntPtrConstant(VA.getLocMemOffset());
1285 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1286 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1287 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1288 if (Flags & ISD::ParamFlags::ByVal) {
1289 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1291 return DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1294 /// ClassifyX86_64SRetCallReturn - Classify how to implement a x86-64
1295 /// struct return call to the specified function. X86-64 ABI specifies
1296 /// some SRet calls are actually returned in registers. Since current
1297 /// LLVM cannot represent multi-value calls, they are represent as
1298 /// calls where the results are passed in a hidden struct provided by
1299 /// the caller. This function examines the type of the struct to
1300 /// determine the correct way to implement the call.
1302 X86TargetLowering::ClassifyX86_64SRetCallReturn(const Function *Fn) {
1303 // FIXME: Disabled for now.
1304 return X86::InMemory;
1306 const PointerType *PTy = cast<PointerType>(Fn->arg_begin()->getType());
1307 const Type *RTy = PTy->getElementType();
1308 unsigned Size = getTargetData()->getABITypeSize(RTy);
1309 if (Size != 16 && Size != 32)
1310 return X86::InMemory;
1313 const StructType *STy = dyn_cast<StructType>(RTy);
1314 if (!STy) return X86::InMemory;
1315 if (STy->getNumElements() == 2 &&
1316 STy->getElementType(0) == Type::X86_FP80Ty &&
1317 STy->getElementType(1) == Type::X86_FP80Ty)
1322 for (Type::subtype_iterator I = RTy->subtype_begin(), E = RTy->subtype_end();
1324 const Type *STy = I->get();
1325 if (!STy->isFPOrFPVector()) {
1333 return X86::InGPR64;
1336 void X86TargetLowering::X86_64AnalyzeSRetCallOperands(SDNode *TheCall,
1339 unsigned NumOps = (TheCall->getNumOperands() - 5) / 2;
1340 for (unsigned i = 1; i != NumOps; ++i) {
1341 MVT::ValueType ArgVT = TheCall->getOperand(5+2*i).getValueType();
1342 SDOperand FlagOp = TheCall->getOperand(5+2*i+1);
1343 unsigned ArgFlags =cast<ConstantSDNode>(FlagOp)->getValue();
1344 if (Fn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo)) {
1345 cerr << "Call operand #" << i << " has unhandled type "
1346 << MVT::getValueTypeString(ArgVT) << "\n";
1352 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
1353 MachineFunction &MF = DAG.getMachineFunction();
1354 SDOperand Chain = Op.getOperand(0);
1355 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1356 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1357 bool IsTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0
1358 && CC == CallingConv::Fast && PerformTailCallOpt;
1359 SDOperand Callee = Op.getOperand(4);
1360 bool Is64Bit = Subtarget->is64Bit();
1361 bool IsStructRet = CallIsStructReturn(Op);
1363 assert(!(isVarArg && CC == CallingConv::Fast) &&
1364 "Var args not supported with calling convention fastcc");
1366 // Analyze operands of the call, assigning locations to each operand.
1367 SmallVector<CCValAssign, 16> ArgLocs;
1368 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1369 CCAssignFn *CCFn = CCAssignFnForNode(Op);
1371 X86::X86_64SRet SRetMethod = X86::InMemory;
1372 if (Is64Bit && IsStructRet)
1373 // FIXME: We can't figure out type of the sret structure for indirect
1374 // calls. We need to copy more information from CallSite to the ISD::CALL
1376 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1378 ClassifyX86_64SRetCallReturn(dyn_cast<Function>(G->getGlobal()));
1380 // UGLY HACK! For x86-64, some 128-bit aggregates are returns in a pair of
1381 // registers. Unfortunately, llvm does not support i128 yet so we pretend it's
1383 if (SRetMethod != X86::InMemory)
1384 X86_64AnalyzeSRetCallOperands(Op.Val, CCFn, CCInfo);
1386 CCInfo.AnalyzeCallOperands(Op.Val, CCFn);
1388 // Get a count of how many bytes are to be pushed on the stack.
1389 unsigned NumBytes = CCInfo.getNextStackOffset();
1390 if (CC == CallingConv::Fast)
1391 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1393 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1394 // arguments and the arguments after the retaddr has been pushed are aligned.
1395 if (!Is64Bit && CC == CallingConv::X86_FastCall &&
1396 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows() &&
1397 (NumBytes & 7) == 0)
1402 // Lower arguments at fp - stackoffset + fpdiff.
1403 unsigned NumBytesCallerPushed =
1404 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1405 FPDiff = NumBytesCallerPushed - NumBytes;
1407 // Set the delta of movement of the returnaddr stackslot.
1408 // But only set if delta is greater than previous delta.
1409 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1410 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1413 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes));
1415 SDOperand RetAddrFrIdx, NewRetAddrFrIdx;
1417 // Adjust the Return address stack slot.
1419 MVT::ValueType VT = Is64Bit ? MVT::i64 : MVT::i32;
1420 RetAddrFrIdx = getReturnAddressFrameIndex(DAG);
1421 // Load the "old" Return address.
1423 DAG.getLoad(VT, Chain,RetAddrFrIdx, NULL, 0);
1424 // Calculate the new stack slot for the return address.
1425 int SlotSize = Is64Bit ? 8 : 4;
1426 int NewReturnAddrFI =
1427 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1428 NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1429 Chain = SDOperand(RetAddrFrIdx.Val, 1);
1433 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1434 SmallVector<SDOperand, 8> MemOpChains;
1438 // Walk the register/memloc assignments, inserting copies/loads. For tail
1439 // calls, lower arguments which could otherwise be possibly overwritten to the
1440 // stack slot where they would go on normal function calls.
1441 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1442 CCValAssign &VA = ArgLocs[i];
1443 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1445 // Promote the value if needed.
1446 switch (VA.getLocInfo()) {
1447 default: assert(0 && "Unknown loc info!");
1448 case CCValAssign::Full: break;
1449 case CCValAssign::SExt:
1450 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1452 case CCValAssign::ZExt:
1453 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1455 case CCValAssign::AExt:
1456 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1460 if (VA.isRegLoc()) {
1461 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1463 if (!IsTailCall || IsPossiblyOverwrittenArgumentOfTailCall(Arg)) {
1464 assert(VA.isMemLoc());
1465 if (StackPtr.Val == 0)
1466 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1468 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1474 if (!MemOpChains.empty())
1475 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1476 &MemOpChains[0], MemOpChains.size());
1478 // Build a sequence of copy-to-reg nodes chained together with token chain
1479 // and flag operands which copy the outgoing args into registers.
1481 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1482 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1484 InFlag = Chain.getValue(1);
1488 InFlag = SDOperand(); // ??? Isn't this nuking the preceding loop's output?
1490 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1492 // Does not work with tail call since ebx is not restored correctly by
1493 // tailcaller. TODO: at least for x86 - verify for x86-64
1494 if (!IsTailCall && !Is64Bit &&
1495 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1496 Subtarget->isPICStyleGOT()) {
1497 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1498 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1500 InFlag = Chain.getValue(1);
1503 if (Is64Bit && isVarArg) {
1504 // From AMD64 ABI document:
1505 // For calls that may call functions that use varargs or stdargs
1506 // (prototype-less calls or calls to functions containing ellipsis (...) in
1507 // the declaration) %al is used as hidden argument to specify the number
1508 // of SSE registers used. The contents of %al do not need to match exactly
1509 // the number of registers, but must be an ubound on the number of SSE
1510 // registers used and is in the range 0 - 8 inclusive.
1512 // Count the number of XMM registers allocated.
1513 static const unsigned XMMArgRegs[] = {
1514 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1515 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1517 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1519 Chain = DAG.getCopyToReg(Chain, X86::AL,
1520 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1521 InFlag = Chain.getValue(1);
1524 // For tail calls lower the arguments to the 'real' stack slot.
1526 SmallVector<SDOperand, 8> MemOpChains2;
1529 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1530 CCValAssign &VA = ArgLocs[i];
1531 if (!VA.isRegLoc()) {
1532 assert(VA.isMemLoc());
1533 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1534 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1535 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1536 // Create frame index.
1537 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1538 uint32_t OpSize = (MVT::getSizeInBits(VA.getLocVT())+7)/8;
1539 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1540 FIN = DAG.getFrameIndex(FI, MVT::i32);
1541 SDOperand Source = Arg;
1542 if (IsPossiblyOverwrittenArgumentOfTailCall(Arg)) {
1543 // Copy from stack slots to stack slot of a tail called function. This
1544 // needs to be done because if we would lower the arguments directly
1545 // to their real stack slot we might end up overwriting each other.
1546 // Get source stack slot.
1547 Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1548 if (StackPtr.Val == 0)
1549 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1550 Source = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, Source);
1551 if ((Flags & ISD::ParamFlags::ByVal)==0)
1552 Source = DAG.getLoad(VA.getValVT(), Chain, Source, NULL, 0);
1555 if (Flags & ISD::ParamFlags::ByVal) {
1556 // Copy relative to framepointer.
1557 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1560 // Store relative to framepointer.
1561 MemOpChains2.push_back(DAG.getStore(Chain, Source, FIN, NULL, 0));
1566 if (!MemOpChains2.empty())
1567 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1568 &MemOpChains2[0], MemOpChains2.size());
1570 // Store the return address to the appropriate stack slot.
1572 Chain = DAG.getStore(Chain,RetAddrFrIdx, NewRetAddrFrIdx, NULL, 0);
1575 // If the callee is a GlobalAddress node (quite common, every direct call is)
1576 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1577 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1578 // We should use extra load for direct calls to dllimported functions in
1580 if ((IsTailCall || !Is64Bit ||
1581 getTargetMachine().getCodeModel() != CodeModel::Large)
1582 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1583 getTargetMachine(), true))
1584 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1585 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1586 if (IsTailCall || !Is64Bit ||
1587 getTargetMachine().getCodeModel() != CodeModel::Large)
1588 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1589 } else if (IsTailCall) {
1590 assert(Callee.getOpcode() == ISD::LOAD &&
1591 "Function destination must be loaded into virtual register");
1592 unsigned Opc = Is64Bit ? X86::R9 : X86::ECX;
1594 Chain = DAG.getCopyToReg(Chain,
1595 DAG.getRegister(Opc, getPointerTy()) ,
1597 Callee = DAG.getRegister(Opc, getPointerTy());
1598 // Add register as live out.
1599 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1602 // Returns a chain & a flag for retval copy to use.
1603 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1604 SmallVector<SDOperand, 8> Ops;
1607 Ops.push_back(Chain);
1608 Ops.push_back(DAG.getIntPtrConstant(NumBytes));
1609 Ops.push_back(DAG.getIntPtrConstant(0));
1611 Ops.push_back(InFlag);
1612 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1613 InFlag = Chain.getValue(1);
1615 // Returns a chain & a flag for retval copy to use.
1616 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1620 Ops.push_back(Chain);
1621 Ops.push_back(Callee);
1624 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1626 // Add an implicit use GOT pointer in EBX.
1627 if (!IsTailCall && !Is64Bit &&
1628 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1629 Subtarget->isPICStyleGOT())
1630 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1632 // Add argument registers to the end of the list so that they are known live
1634 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1635 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1636 RegsToPass[i].second.getValueType()));
1639 Ops.push_back(InFlag);
1642 assert(InFlag.Val &&
1643 "Flag must be set. Depend on flag being set in LowerRET");
1644 Chain = DAG.getNode(X86ISD::TAILCALL,
1645 Op.Val->getVTList(), &Ops[0], Ops.size());
1647 return SDOperand(Chain.Val, Op.ResNo);
1650 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1651 InFlag = Chain.getValue(1);
1653 // Create the CALLSEQ_END node.
1654 unsigned NumBytesForCalleeToPush;
1655 if (IsCalleePop(Op))
1656 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1657 else if (!Is64Bit && IsStructRet)
1658 // If this is is a call to a struct-return function, the callee
1659 // pops the hidden struct pointer, so we have to push it back.
1660 // This is common for Darwin/X86, Linux & Mingw32 targets.
1661 NumBytesForCalleeToPush = 4;
1663 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1665 // Returns a flag for retval copy to use.
1666 Chain = DAG.getCALLSEQ_END(Chain,
1667 DAG.getIntPtrConstant(NumBytes),
1668 DAG.getIntPtrConstant(NumBytesForCalleeToPush),
1670 InFlag = Chain.getValue(1);
1672 // Handle result values, copying them out of physregs into vregs that we
1674 switch (SRetMethod) {
1676 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1678 return SDOperand(LowerCallResultToTwo64BitRegs(Chain, InFlag, Op.Val,
1680 MVT::i64, DAG), Op.ResNo);
1682 return SDOperand(LowerCallResultToTwo64BitRegs(Chain, InFlag, Op.Val,
1683 X86::XMM0, X86::XMM1,
1684 MVT::f64, DAG), Op.ResNo);
1686 return SDOperand(LowerCallResultToTwoX87Regs(Chain, InFlag, Op.Val, DAG),
1692 //===----------------------------------------------------------------------===//
1693 // Fast Calling Convention (tail call) implementation
1694 //===----------------------------------------------------------------------===//
1696 // Like std call, callee cleans arguments, convention except that ECX is
1697 // reserved for storing the tail called function address. Only 2 registers are
1698 // free for argument passing (inreg). Tail call optimization is performed
1700 // * tailcallopt is enabled
1701 // * caller/callee are fastcc
1702 // * elf/pic is disabled OR
1703 // * elf/pic enabled + callee is in module + callee has
1704 // visibility protected or hidden
1705 // To keep the stack aligned according to platform abi the function
1706 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1707 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1708 // If a tail called function callee has more arguments than the caller the
1709 // caller needs to make sure that there is room to move the RETADDR to. This is
1710 // achieved by reserving an area the size of the argument delta right after the
1711 // original REtADDR, but before the saved framepointer or the spilled registers
1712 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1724 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1725 /// for a 16 byte align requirement.
1726 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1727 SelectionDAG& DAG) {
1728 if (PerformTailCallOpt) {
1729 MachineFunction &MF = DAG.getMachineFunction();
1730 const TargetMachine &TM = MF.getTarget();
1731 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1732 unsigned StackAlignment = TFI.getStackAlignment();
1733 uint64_t AlignMask = StackAlignment - 1;
1734 int64_t Offset = StackSize;
1735 unsigned SlotSize = Subtarget->is64Bit() ? 8 : 4;
1736 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1737 // Number smaller than 12 so just add the difference.
1738 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1740 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1741 Offset = ((~AlignMask) & Offset) + StackAlignment +
1742 (StackAlignment-SlotSize);
1749 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1750 /// following the call is a return. A function is eligible if caller/callee
1751 /// calling conventions match, currently only fastcc supports tail calls, and
1752 /// the function CALL is immediatly followed by a RET.
1753 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDOperand Call,
1755 SelectionDAG& DAG) const {
1756 if (!PerformTailCallOpt)
1759 // Check whether CALL node immediatly preceeds the RET node and whether the
1760 // return uses the result of the node or is a void return.
1761 unsigned NumOps = Ret.getNumOperands();
1763 (Ret.getOperand(0) == SDOperand(Call.Val,1) ||
1764 Ret.getOperand(0) == SDOperand(Call.Val,0))) ||
1766 Ret.getOperand(0) == SDOperand(Call.Val,Call.Val->getNumValues()-1) &&
1767 Ret.getOperand(1) == SDOperand(Call.Val,0))) {
1768 MachineFunction &MF = DAG.getMachineFunction();
1769 unsigned CallerCC = MF.getFunction()->getCallingConv();
1770 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1771 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1772 SDOperand Callee = Call.getOperand(4);
1773 // On elf/pic %ebx needs to be livein.
1774 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1775 !Subtarget->isPICStyleGOT())
1778 // Can only do local tail calls with PIC.
1779 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1780 return G->getGlobal()->hasHiddenVisibility()
1781 || G->getGlobal()->hasProtectedVisibility();
1788 //===----------------------------------------------------------------------===//
1789 // Other Lowering Hooks
1790 //===----------------------------------------------------------------------===//
1793 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1794 MachineFunction &MF = DAG.getMachineFunction();
1795 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1796 int ReturnAddrIndex = FuncInfo->getRAIndex();
1798 if (ReturnAddrIndex == 0) {
1799 // Set up a frame object for the return address.
1800 if (Subtarget->is64Bit())
1801 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1803 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1805 FuncInfo->setRAIndex(ReturnAddrIndex);
1808 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1813 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1814 /// specific condition code. It returns a false if it cannot do a direct
1815 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1817 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1818 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1819 SelectionDAG &DAG) {
1820 X86CC = X86::COND_INVALID;
1822 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1823 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1824 // X > -1 -> X == 0, jump !sign.
1825 RHS = DAG.getConstant(0, RHS.getValueType());
1826 X86CC = X86::COND_NS;
1828 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1829 // X < 0 -> X == 0, jump on sign.
1830 X86CC = X86::COND_S;
1832 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
1834 RHS = DAG.getConstant(0, RHS.getValueType());
1835 X86CC = X86::COND_LE;
1840 switch (SetCCOpcode) {
1842 case ISD::SETEQ: X86CC = X86::COND_E; break;
1843 case ISD::SETGT: X86CC = X86::COND_G; break;
1844 case ISD::SETGE: X86CC = X86::COND_GE; break;
1845 case ISD::SETLT: X86CC = X86::COND_L; break;
1846 case ISD::SETLE: X86CC = X86::COND_LE; break;
1847 case ISD::SETNE: X86CC = X86::COND_NE; break;
1848 case ISD::SETULT: X86CC = X86::COND_B; break;
1849 case ISD::SETUGT: X86CC = X86::COND_A; break;
1850 case ISD::SETULE: X86CC = X86::COND_BE; break;
1851 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1854 // On a floating point condition, the flags are set as follows:
1856 // 0 | 0 | 0 | X > Y
1857 // 0 | 0 | 1 | X < Y
1858 // 1 | 0 | 0 | X == Y
1859 // 1 | 1 | 1 | unordered
1861 switch (SetCCOpcode) {
1864 case ISD::SETEQ: X86CC = X86::COND_E; break;
1865 case ISD::SETOLT: Flip = true; // Fallthrough
1867 case ISD::SETGT: X86CC = X86::COND_A; break;
1868 case ISD::SETOLE: Flip = true; // Fallthrough
1870 case ISD::SETGE: X86CC = X86::COND_AE; break;
1871 case ISD::SETUGT: Flip = true; // Fallthrough
1873 case ISD::SETLT: X86CC = X86::COND_B; break;
1874 case ISD::SETUGE: Flip = true; // Fallthrough
1876 case ISD::SETLE: X86CC = X86::COND_BE; break;
1878 case ISD::SETNE: X86CC = X86::COND_NE; break;
1879 case ISD::SETUO: X86CC = X86::COND_P; break;
1880 case ISD::SETO: X86CC = X86::COND_NP; break;
1883 std::swap(LHS, RHS);
1886 return X86CC != X86::COND_INVALID;
1889 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1890 /// code. Current x86 isa includes the following FP cmov instructions:
1891 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1892 static bool hasFPCMov(unsigned X86CC) {
1908 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1909 /// true if Op is undef or if its value falls within the specified range (L, H].
1910 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1911 if (Op.getOpcode() == ISD::UNDEF)
1914 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1915 return (Val >= Low && Val < Hi);
1918 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1919 /// true if Op is undef or if its value equal to the specified value.
1920 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1921 if (Op.getOpcode() == ISD::UNDEF)
1923 return cast<ConstantSDNode>(Op)->getValue() == Val;
1926 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1927 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1928 bool X86::isPSHUFDMask(SDNode *N) {
1929 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1931 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
1934 // Check if the value doesn't reference the second vector.
1935 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1936 SDOperand Arg = N->getOperand(i);
1937 if (Arg.getOpcode() == ISD::UNDEF) continue;
1938 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1939 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
1946 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1947 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1948 bool X86::isPSHUFHWMask(SDNode *N) {
1949 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1951 if (N->getNumOperands() != 8)
1954 // Lower quadword copied in order.
1955 for (unsigned i = 0; i != 4; ++i) {
1956 SDOperand Arg = N->getOperand(i);
1957 if (Arg.getOpcode() == ISD::UNDEF) continue;
1958 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1959 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1963 // Upper quadword shuffled.
1964 for (unsigned i = 4; i != 8; ++i) {
1965 SDOperand Arg = N->getOperand(i);
1966 if (Arg.getOpcode() == ISD::UNDEF) continue;
1967 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1968 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1969 if (Val < 4 || Val > 7)
1976 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1977 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1978 bool X86::isPSHUFLWMask(SDNode *N) {
1979 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1981 if (N->getNumOperands() != 8)
1984 // Upper quadword copied in order.
1985 for (unsigned i = 4; i != 8; ++i)
1986 if (!isUndefOrEqual(N->getOperand(i), i))
1989 // Lower quadword shuffled.
1990 for (unsigned i = 0; i != 4; ++i)
1991 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1997 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1998 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1999 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2000 if (NumElems != 2 && NumElems != 4) return false;
2002 unsigned Half = NumElems / 2;
2003 for (unsigned i = 0; i < Half; ++i)
2004 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2006 for (unsigned i = Half; i < NumElems; ++i)
2007 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2013 bool X86::isSHUFPMask(SDNode *N) {
2014 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2015 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2018 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2019 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2020 /// half elements to come from vector 1 (which would equal the dest.) and
2021 /// the upper half to come from vector 2.
2022 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2023 if (NumOps != 2 && NumOps != 4) return false;
2025 unsigned Half = NumOps / 2;
2026 for (unsigned i = 0; i < Half; ++i)
2027 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2029 for (unsigned i = Half; i < NumOps; ++i)
2030 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2035 static bool isCommutedSHUFP(SDNode *N) {
2036 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2037 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2040 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2041 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2042 bool X86::isMOVHLPSMask(SDNode *N) {
2043 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2045 if (N->getNumOperands() != 4)
2048 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2049 return isUndefOrEqual(N->getOperand(0), 6) &&
2050 isUndefOrEqual(N->getOperand(1), 7) &&
2051 isUndefOrEqual(N->getOperand(2), 2) &&
2052 isUndefOrEqual(N->getOperand(3), 3);
2055 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2056 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2058 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2059 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2061 if (N->getNumOperands() != 4)
2064 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2065 return isUndefOrEqual(N->getOperand(0), 2) &&
2066 isUndefOrEqual(N->getOperand(1), 3) &&
2067 isUndefOrEqual(N->getOperand(2), 2) &&
2068 isUndefOrEqual(N->getOperand(3), 3);
2071 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2072 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2073 bool X86::isMOVLPMask(SDNode *N) {
2074 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2076 unsigned NumElems = N->getNumOperands();
2077 if (NumElems != 2 && NumElems != 4)
2080 for (unsigned i = 0; i < NumElems/2; ++i)
2081 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2084 for (unsigned i = NumElems/2; i < NumElems; ++i)
2085 if (!isUndefOrEqual(N->getOperand(i), i))
2091 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2092 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2094 bool X86::isMOVHPMask(SDNode *N) {
2095 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2097 unsigned NumElems = N->getNumOperands();
2098 if (NumElems != 2 && NumElems != 4)
2101 for (unsigned i = 0; i < NumElems/2; ++i)
2102 if (!isUndefOrEqual(N->getOperand(i), i))
2105 for (unsigned i = 0; i < NumElems/2; ++i) {
2106 SDOperand Arg = N->getOperand(i + NumElems/2);
2107 if (!isUndefOrEqual(Arg, i + NumElems))
2114 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2115 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2116 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2117 bool V2IsSplat = false) {
2118 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2121 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2122 SDOperand BitI = Elts[i];
2123 SDOperand BitI1 = Elts[i+1];
2124 if (!isUndefOrEqual(BitI, j))
2127 if (isUndefOrEqual(BitI1, NumElts))
2130 if (!isUndefOrEqual(BitI1, j + NumElts))
2138 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2139 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2140 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2143 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2144 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2145 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2146 bool V2IsSplat = false) {
2147 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2150 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2151 SDOperand BitI = Elts[i];
2152 SDOperand BitI1 = Elts[i+1];
2153 if (!isUndefOrEqual(BitI, j + NumElts/2))
2156 if (isUndefOrEqual(BitI1, NumElts))
2159 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2167 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2168 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2169 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2172 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2173 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2175 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2176 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2178 unsigned NumElems = N->getNumOperands();
2179 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2182 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2183 SDOperand BitI = N->getOperand(i);
2184 SDOperand BitI1 = N->getOperand(i+1);
2186 if (!isUndefOrEqual(BitI, j))
2188 if (!isUndefOrEqual(BitI1, j))
2195 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2196 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2198 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2199 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2201 unsigned NumElems = N->getNumOperands();
2202 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2205 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2206 SDOperand BitI = N->getOperand(i);
2207 SDOperand BitI1 = N->getOperand(i + 1);
2209 if (!isUndefOrEqual(BitI, j))
2211 if (!isUndefOrEqual(BitI1, j))
2218 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2219 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2220 /// MOVSD, and MOVD, i.e. setting the lowest element.
2221 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2222 if (NumElts != 2 && NumElts != 4)
2225 if (!isUndefOrEqual(Elts[0], NumElts))
2228 for (unsigned i = 1; i < NumElts; ++i) {
2229 if (!isUndefOrEqual(Elts[i], i))
2236 bool X86::isMOVLMask(SDNode *N) {
2237 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2238 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2241 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2242 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2243 /// element of vector 2 and the other elements to come from vector 1 in order.
2244 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2245 bool V2IsSplat = false,
2246 bool V2IsUndef = false) {
2247 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2250 if (!isUndefOrEqual(Ops[0], 0))
2253 for (unsigned i = 1; i < NumOps; ++i) {
2254 SDOperand Arg = Ops[i];
2255 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2256 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2257 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2264 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2265 bool V2IsUndef = false) {
2266 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2267 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2268 V2IsSplat, V2IsUndef);
2271 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2272 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2273 bool X86::isMOVSHDUPMask(SDNode *N) {
2274 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2276 if (N->getNumOperands() != 4)
2279 // Expect 1, 1, 3, 3
2280 for (unsigned i = 0; i < 2; ++i) {
2281 SDOperand Arg = N->getOperand(i);
2282 if (Arg.getOpcode() == ISD::UNDEF) continue;
2283 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2284 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2285 if (Val != 1) return false;
2289 for (unsigned i = 2; i < 4; ++i) {
2290 SDOperand Arg = N->getOperand(i);
2291 if (Arg.getOpcode() == ISD::UNDEF) continue;
2292 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2293 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2294 if (Val != 3) return false;
2298 // Don't use movshdup if it can be done with a shufps.
2302 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2303 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2304 bool X86::isMOVSLDUPMask(SDNode *N) {
2305 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2307 if (N->getNumOperands() != 4)
2310 // Expect 0, 0, 2, 2
2311 for (unsigned i = 0; i < 2; ++i) {
2312 SDOperand Arg = N->getOperand(i);
2313 if (Arg.getOpcode() == ISD::UNDEF) continue;
2314 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2315 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2316 if (Val != 0) return false;
2320 for (unsigned i = 2; i < 4; ++i) {
2321 SDOperand Arg = N->getOperand(i);
2322 if (Arg.getOpcode() == ISD::UNDEF) continue;
2323 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2324 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2325 if (Val != 2) return false;
2329 // Don't use movshdup if it can be done with a shufps.
2333 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2334 /// specifies a identity operation on the LHS or RHS.
2335 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2336 unsigned NumElems = N->getNumOperands();
2337 for (unsigned i = 0; i < NumElems; ++i)
2338 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2343 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2344 /// a splat of a single element.
2345 static bool isSplatMask(SDNode *N) {
2346 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2348 // This is a splat operation if each element of the permute is the same, and
2349 // if the value doesn't reference the second vector.
2350 unsigned NumElems = N->getNumOperands();
2351 SDOperand ElementBase;
2353 for (; i != NumElems; ++i) {
2354 SDOperand Elt = N->getOperand(i);
2355 if (isa<ConstantSDNode>(Elt)) {
2361 if (!ElementBase.Val)
2364 for (; i != NumElems; ++i) {
2365 SDOperand Arg = N->getOperand(i);
2366 if (Arg.getOpcode() == ISD::UNDEF) continue;
2367 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2368 if (Arg != ElementBase) return false;
2371 // Make sure it is a splat of the first vector operand.
2372 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2375 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2376 /// a splat of a single element and it's a 2 or 4 element mask.
2377 bool X86::isSplatMask(SDNode *N) {
2378 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2380 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2381 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2383 return ::isSplatMask(N);
2386 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2387 /// specifies a splat of zero element.
2388 bool X86::isSplatLoMask(SDNode *N) {
2389 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2391 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2392 if (!isUndefOrEqual(N->getOperand(i), 0))
2397 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2398 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2400 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2401 unsigned NumOperands = N->getNumOperands();
2402 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2404 for (unsigned i = 0; i < NumOperands; ++i) {
2406 SDOperand Arg = N->getOperand(NumOperands-i-1);
2407 if (Arg.getOpcode() != ISD::UNDEF)
2408 Val = cast<ConstantSDNode>(Arg)->getValue();
2409 if (Val >= NumOperands) Val -= NumOperands;
2411 if (i != NumOperands - 1)
2418 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2419 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2421 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2423 // 8 nodes, but we only care about the last 4.
2424 for (unsigned i = 7; i >= 4; --i) {
2426 SDOperand Arg = N->getOperand(i);
2427 if (Arg.getOpcode() != ISD::UNDEF)
2428 Val = cast<ConstantSDNode>(Arg)->getValue();
2437 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2438 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2440 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2442 // 8 nodes, but we only care about the first 4.
2443 for (int i = 3; i >= 0; --i) {
2445 SDOperand Arg = N->getOperand(i);
2446 if (Arg.getOpcode() != ISD::UNDEF)
2447 Val = cast<ConstantSDNode>(Arg)->getValue();
2456 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2457 /// specifies a 8 element shuffle that can be broken into a pair of
2458 /// PSHUFHW and PSHUFLW.
2459 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2460 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2462 if (N->getNumOperands() != 8)
2465 // Lower quadword shuffled.
2466 for (unsigned i = 0; i != 4; ++i) {
2467 SDOperand Arg = N->getOperand(i);
2468 if (Arg.getOpcode() == ISD::UNDEF) continue;
2469 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2470 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2475 // Upper quadword shuffled.
2476 for (unsigned i = 4; i != 8; ++i) {
2477 SDOperand Arg = N->getOperand(i);
2478 if (Arg.getOpcode() == ISD::UNDEF) continue;
2479 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2480 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2481 if (Val < 4 || Val > 7)
2488 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2489 /// values in ther permute mask.
2490 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2491 SDOperand &V2, SDOperand &Mask,
2492 SelectionDAG &DAG) {
2493 MVT::ValueType VT = Op.getValueType();
2494 MVT::ValueType MaskVT = Mask.getValueType();
2495 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2496 unsigned NumElems = Mask.getNumOperands();
2497 SmallVector<SDOperand, 8> MaskVec;
2499 for (unsigned i = 0; i != NumElems; ++i) {
2500 SDOperand Arg = Mask.getOperand(i);
2501 if (Arg.getOpcode() == ISD::UNDEF) {
2502 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2505 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2506 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2508 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2510 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2514 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2515 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2518 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2519 /// the two vector operands have swapped position.
2521 SDOperand CommuteVectorShuffleMask(SDOperand Mask, SelectionDAG &DAG) {
2522 MVT::ValueType MaskVT = Mask.getValueType();
2523 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2524 unsigned NumElems = Mask.getNumOperands();
2525 SmallVector<SDOperand, 8> MaskVec;
2526 for (unsigned i = 0; i != NumElems; ++i) {
2527 SDOperand Arg = Mask.getOperand(i);
2528 if (Arg.getOpcode() == ISD::UNDEF) {
2529 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2532 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2533 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2535 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2537 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2539 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2543 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2544 /// match movhlps. The lower half elements should come from upper half of
2545 /// V1 (and in order), and the upper half elements should come from the upper
2546 /// half of V2 (and in order).
2547 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2548 unsigned NumElems = Mask->getNumOperands();
2551 for (unsigned i = 0, e = 2; i != e; ++i)
2552 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2554 for (unsigned i = 2; i != 4; ++i)
2555 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2560 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2561 /// is promoted to a vector.
2562 static inline bool isScalarLoadToVector(SDNode *N) {
2563 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2564 N = N->getOperand(0).Val;
2565 return ISD::isNON_EXTLoad(N);
2570 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2571 /// match movlp{s|d}. The lower half elements should come from lower half of
2572 /// V1 (and in order), and the upper half elements should come from the upper
2573 /// half of V2 (and in order). And since V1 will become the source of the
2574 /// MOVLP, it must be either a vector load or a scalar load to vector.
2575 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2576 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2578 // Is V2 is a vector load, don't do this transformation. We will try to use
2579 // load folding shufps op.
2580 if (ISD::isNON_EXTLoad(V2))
2583 unsigned NumElems = Mask->getNumOperands();
2584 if (NumElems != 2 && NumElems != 4)
2586 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2587 if (!isUndefOrEqual(Mask->getOperand(i), i))
2589 for (unsigned i = NumElems/2; i != NumElems; ++i)
2590 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2595 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2597 static bool isSplatVector(SDNode *N) {
2598 if (N->getOpcode() != ISD::BUILD_VECTOR)
2601 SDOperand SplatValue = N->getOperand(0);
2602 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2603 if (N->getOperand(i) != SplatValue)
2608 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2610 static bool isUndefShuffle(SDNode *N) {
2611 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2614 SDOperand V1 = N->getOperand(0);
2615 SDOperand V2 = N->getOperand(1);
2616 SDOperand Mask = N->getOperand(2);
2617 unsigned NumElems = Mask.getNumOperands();
2618 for (unsigned i = 0; i != NumElems; ++i) {
2619 SDOperand Arg = Mask.getOperand(i);
2620 if (Arg.getOpcode() != ISD::UNDEF) {
2621 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2622 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2624 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2631 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2633 static inline bool isZeroNode(SDOperand Elt) {
2634 return ((isa<ConstantSDNode>(Elt) &&
2635 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2636 (isa<ConstantFPSDNode>(Elt) &&
2637 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2640 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2641 /// to an zero vector.
2642 static bool isZeroShuffle(SDNode *N) {
2643 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2646 SDOperand V1 = N->getOperand(0);
2647 SDOperand V2 = N->getOperand(1);
2648 SDOperand Mask = N->getOperand(2);
2649 unsigned NumElems = Mask.getNumOperands();
2650 for (unsigned i = 0; i != NumElems; ++i) {
2651 SDOperand Arg = Mask.getOperand(i);
2652 if (Arg.getOpcode() == ISD::UNDEF)
2655 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2656 if (Idx < NumElems) {
2657 unsigned Opc = V1.Val->getOpcode();
2658 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.Val))
2660 if (Opc != ISD::BUILD_VECTOR ||
2661 !isZeroNode(V1.Val->getOperand(Idx)))
2663 } else if (Idx >= NumElems) {
2664 unsigned Opc = V2.Val->getOpcode();
2665 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.Val))
2667 if (Opc != ISD::BUILD_VECTOR ||
2668 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2675 /// getZeroVector - Returns a vector of specified type with all zero elements.
2677 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2678 assert(MVT::isVector(VT) && "Expected a vector type");
2680 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2681 // type. This ensures they get CSE'd.
2682 SDOperand Cst = DAG.getTargetConstant(0, MVT::i32);
2684 if (MVT::getSizeInBits(VT) == 64) // MMX
2685 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2687 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2688 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2691 /// getOnesVector - Returns a vector of specified type with all bits set.
2693 static SDOperand getOnesVector(MVT::ValueType VT, SelectionDAG &DAG) {
2694 assert(MVT::isVector(VT) && "Expected a vector type");
2696 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2697 // type. This ensures they get CSE'd.
2698 SDOperand Cst = DAG.getTargetConstant(~0U, MVT::i32);
2700 if (MVT::getSizeInBits(VT) == 64) // MMX
2701 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2703 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2704 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2708 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2709 /// that point to V2 points to its first element.
2710 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2711 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2713 bool Changed = false;
2714 SmallVector<SDOperand, 8> MaskVec;
2715 unsigned NumElems = Mask.getNumOperands();
2716 for (unsigned i = 0; i != NumElems; ++i) {
2717 SDOperand Arg = Mask.getOperand(i);
2718 if (Arg.getOpcode() != ISD::UNDEF) {
2719 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2720 if (Val > NumElems) {
2721 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2725 MaskVec.push_back(Arg);
2729 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2730 &MaskVec[0], MaskVec.size());
2734 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2735 /// operation of specified width.
2736 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2737 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2738 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2740 SmallVector<SDOperand, 8> MaskVec;
2741 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2742 for (unsigned i = 1; i != NumElems; ++i)
2743 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2744 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2747 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2748 /// of specified width.
2749 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2750 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2751 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2752 SmallVector<SDOperand, 8> MaskVec;
2753 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2754 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2755 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2757 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2760 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2761 /// of specified width.
2762 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2763 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2764 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2765 unsigned Half = NumElems/2;
2766 SmallVector<SDOperand, 8> MaskVec;
2767 for (unsigned i = 0; i != Half; ++i) {
2768 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2769 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2771 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2774 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2776 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2777 SDOperand V1 = Op.getOperand(0);
2778 SDOperand Mask = Op.getOperand(2);
2779 MVT::ValueType VT = Op.getValueType();
2780 unsigned NumElems = Mask.getNumOperands();
2781 Mask = getUnpacklMask(NumElems, DAG);
2782 while (NumElems != 4) {
2783 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2786 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2788 Mask = getZeroVector(MVT::v4i32, DAG);
2789 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2790 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2791 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2794 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2795 /// vector of zero or undef vector. This produces a shuffle where the low
2796 /// element of V2 is swizzled into the zero/undef vector, landing at element
2797 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2798 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2799 unsigned NumElems, unsigned Idx,
2800 bool isZero, SelectionDAG &DAG) {
2801 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2802 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2803 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
2804 SmallVector<SDOperand, 16> MaskVec;
2805 for (unsigned i = 0; i != NumElems; ++i)
2806 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
2807 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
2809 MaskVec.push_back(DAG.getConstant(i, EVT));
2810 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2811 &MaskVec[0], MaskVec.size());
2812 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2815 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2817 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2818 unsigned NumNonZero, unsigned NumZero,
2819 SelectionDAG &DAG, TargetLowering &TLI) {
2825 for (unsigned i = 0; i < 16; ++i) {
2826 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2827 if (ThisIsNonZero && First) {
2829 V = getZeroVector(MVT::v8i16, DAG);
2831 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2836 SDOperand ThisElt(0, 0), LastElt(0, 0);
2837 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2838 if (LastIsNonZero) {
2839 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2841 if (ThisIsNonZero) {
2842 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2843 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2844 ThisElt, DAG.getConstant(8, MVT::i8));
2846 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2851 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2852 DAG.getIntPtrConstant(i/2));
2856 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2859 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
2861 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2862 unsigned NumNonZero, unsigned NumZero,
2863 SelectionDAG &DAG, TargetLowering &TLI) {
2869 for (unsigned i = 0; i < 8; ++i) {
2870 bool isNonZero = (NonZeros & (1 << i)) != 0;
2874 V = getZeroVector(MVT::v8i16, DAG);
2876 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2879 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2880 DAG.getIntPtrConstant(i));
2888 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2889 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
2890 if (ISD::isBuildVectorAllZeros(Op.Val) || ISD::isBuildVectorAllOnes(Op.Val)) {
2891 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
2892 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
2893 // eliminated on x86-32 hosts.
2894 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
2897 if (ISD::isBuildVectorAllOnes(Op.Val))
2898 return getOnesVector(Op.getValueType(), DAG);
2899 return getZeroVector(Op.getValueType(), DAG);
2902 MVT::ValueType VT = Op.getValueType();
2903 MVT::ValueType EVT = MVT::getVectorElementType(VT);
2904 unsigned EVTBits = MVT::getSizeInBits(EVT);
2906 unsigned NumElems = Op.getNumOperands();
2907 unsigned NumZero = 0;
2908 unsigned NumNonZero = 0;
2909 unsigned NonZeros = 0;
2910 bool HasNonImms = false;
2911 SmallSet<SDOperand, 8> Values;
2912 for (unsigned i = 0; i < NumElems; ++i) {
2913 SDOperand Elt = Op.getOperand(i);
2914 if (Elt.getOpcode() == ISD::UNDEF)
2917 if (Elt.getOpcode() != ISD::Constant &&
2918 Elt.getOpcode() != ISD::ConstantFP)
2920 if (isZeroNode(Elt))
2923 NonZeros |= (1 << i);
2928 if (NumNonZero == 0) {
2929 // All undef vector. Return an UNDEF. All zero vectors were handled above.
2930 return DAG.getNode(ISD::UNDEF, VT);
2933 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2934 if (Values.size() == 1)
2937 // Special case for single non-zero element.
2938 if (NumNonZero == 1 && NumElems <= 4) {
2939 unsigned Idx = CountTrailingZeros_32(NonZeros);
2940 SDOperand Item = Op.getOperand(Idx);
2941 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2943 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2944 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2946 else if (!HasNonImms) // Otherwise, it's better to do a constpool load.
2949 if (EVTBits == 32) {
2950 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2951 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2953 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2954 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
2955 SmallVector<SDOperand, 8> MaskVec;
2956 for (unsigned i = 0; i < NumElems; i++)
2957 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2958 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2959 &MaskVec[0], MaskVec.size());
2960 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2961 DAG.getNode(ISD::UNDEF, VT), Mask);
2965 // A vector full of immediates; various special cases are already
2966 // handled, so this is best done with a single constant-pool load.
2970 // Let legalizer expand 2-wide build_vectors.
2974 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2975 if (EVTBits == 8 && NumElems == 16) {
2976 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2978 if (V.Val) return V;
2981 if (EVTBits == 16 && NumElems == 8) {
2982 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2984 if (V.Val) return V;
2987 // If element VT is == 32 bits, turn it into a number of shuffles.
2988 SmallVector<SDOperand, 8> V;
2990 if (NumElems == 4 && NumZero > 0) {
2991 for (unsigned i = 0; i < 4; ++i) {
2992 bool isZero = !(NonZeros & (1 << i));
2994 V[i] = getZeroVector(VT, DAG);
2996 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2999 for (unsigned i = 0; i < 2; ++i) {
3000 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3003 V[i] = V[i*2]; // Must be a zero vector.
3006 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3007 getMOVLMask(NumElems, DAG));
3010 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3011 getMOVLMask(NumElems, DAG));
3014 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3015 getUnpacklMask(NumElems, DAG));
3020 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3021 // clears the upper bits.
3022 // FIXME: we can do the same for v4f32 case when we know both parts of
3023 // the lower half come from scalar_to_vector (loadf32). We should do
3024 // that in post legalizer dag combiner with target specific hooks.
3025 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3027 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3028 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3029 SmallVector<SDOperand, 8> MaskVec;
3030 bool Reverse = (NonZeros & 0x3) == 2;
3031 for (unsigned i = 0; i < 2; ++i)
3033 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3035 MaskVec.push_back(DAG.getConstant(i, EVT));
3036 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3037 for (unsigned i = 0; i < 2; ++i)
3039 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3041 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3042 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3043 &MaskVec[0], MaskVec.size());
3044 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3047 if (Values.size() > 2) {
3048 // Expand into a number of unpckl*.
3050 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3051 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3052 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3053 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3054 for (unsigned i = 0; i < NumElems; ++i)
3055 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3057 while (NumElems != 0) {
3058 for (unsigned i = 0; i < NumElems; ++i)
3059 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3070 SDOperand LowerVECTOR_SHUFFLEv8i16(SDOperand V1, SDOperand V2,
3071 SDOperand PermMask, SelectionDAG &DAG,
3072 TargetLowering &TLI) {
3074 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(8);
3075 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3076 MVT::ValueType PtrVT = TLI.getPointerTy();
3077 SmallVector<SDOperand, 8> MaskElts(PermMask.Val->op_begin(),
3078 PermMask.Val->op_end());
3080 // First record which half of which vector the low elements come from.
3081 SmallVector<unsigned, 4> LowQuad(4);
3082 for (unsigned i = 0; i < 4; ++i) {
3083 SDOperand Elt = MaskElts[i];
3084 if (Elt.getOpcode() == ISD::UNDEF)
3086 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3087 int QuadIdx = EltIdx / 4;
3090 int BestLowQuad = -1;
3091 unsigned MaxQuad = 1;
3092 for (unsigned i = 0; i < 4; ++i) {
3093 if (LowQuad[i] > MaxQuad) {
3095 MaxQuad = LowQuad[i];
3099 // Record which half of which vector the high elements come from.
3100 SmallVector<unsigned, 4> HighQuad(4);
3101 for (unsigned i = 4; i < 8; ++i) {
3102 SDOperand Elt = MaskElts[i];
3103 if (Elt.getOpcode() == ISD::UNDEF)
3105 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3106 int QuadIdx = EltIdx / 4;
3107 ++HighQuad[QuadIdx];
3109 int BestHighQuad = -1;
3111 for (unsigned i = 0; i < 4; ++i) {
3112 if (HighQuad[i] > MaxQuad) {
3114 MaxQuad = HighQuad[i];
3118 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3119 if (BestLowQuad != -1 || BestHighQuad != -1) {
3120 // First sort the 4 chunks in order using shufpd.
3121 SmallVector<SDOperand, 8> MaskVec;
3122 if (BestLowQuad != -1)
3123 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3125 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3126 if (BestHighQuad != -1)
3127 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3129 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3130 SDOperand Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3131 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3132 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3133 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3134 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3136 // Now sort high and low parts separately.
3137 BitVector InOrder(8);
3138 if (BestLowQuad != -1) {
3139 // Sort lower half in order using PSHUFLW.
3141 bool AnyOutOrder = false;
3142 for (unsigned i = 0; i != 4; ++i) {
3143 SDOperand Elt = MaskElts[i];
3144 if (Elt.getOpcode() == ISD::UNDEF) {
3145 MaskVec.push_back(Elt);
3148 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3151 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3152 // If this element is in the right place after this shuffle, then
3154 if ((int)(EltIdx / 4) == BestLowQuad)
3159 for (unsigned i = 4; i != 8; ++i)
3160 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3161 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3162 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3166 if (BestHighQuad != -1) {
3167 // Sort high half in order using PSHUFHW if possible.
3169 for (unsigned i = 0; i != 4; ++i)
3170 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3171 bool AnyOutOrder = false;
3172 for (unsigned i = 4; i != 8; ++i) {
3173 SDOperand Elt = MaskElts[i];
3174 if (Elt.getOpcode() == ISD::UNDEF) {
3175 MaskVec.push_back(Elt);
3178 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3181 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3182 // If this element is in the right place after this shuffle, then
3184 if ((int)(EltIdx / 4) == BestHighQuad)
3189 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3190 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3194 // The other elements are put in the right place using pextrw and pinsrw.
3195 for (unsigned i = 0; i != 8; ++i) {
3198 SDOperand Elt = MaskElts[i];
3199 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3202 SDOperand ExtOp = (EltIdx < 8)
3203 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3204 DAG.getConstant(EltIdx, PtrVT))
3205 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3206 DAG.getConstant(EltIdx - 8, PtrVT));
3207 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3208 DAG.getConstant(i, PtrVT));
3213 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use
3214 ///as few as possible.
3215 // First, let's find out how many elements are already in the right order.
3216 unsigned V1InOrder = 0;
3217 unsigned V1FromV1 = 0;
3218 unsigned V2InOrder = 0;
3219 unsigned V2FromV2 = 0;
3220 SmallVector<SDOperand, 8> V1Elts;
3221 SmallVector<SDOperand, 8> V2Elts;
3222 for (unsigned i = 0; i < 8; ++i) {
3223 SDOperand Elt = MaskElts[i];
3224 if (Elt.getOpcode() == ISD::UNDEF) {
3225 V1Elts.push_back(Elt);
3226 V2Elts.push_back(Elt);
3231 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3233 V1Elts.push_back(Elt);
3234 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3236 } else if (EltIdx == i+8) {
3237 V1Elts.push_back(Elt);
3238 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3240 } else if (EltIdx < 8) {
3241 V1Elts.push_back(Elt);
3244 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3249 if (V2InOrder > V1InOrder) {
3250 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3252 std::swap(V1Elts, V2Elts);
3253 std::swap(V1FromV1, V2FromV2);
3256 if ((V1FromV1 + V1InOrder) != 8) {
3257 // Some elements are from V2.
3259 // If there are elements that are from V1 but out of place,
3260 // then first sort them in place
3261 SmallVector<SDOperand, 8> MaskVec;
3262 for (unsigned i = 0; i < 8; ++i) {
3263 SDOperand Elt = V1Elts[i];
3264 if (Elt.getOpcode() == ISD::UNDEF) {
3265 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3268 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3270 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3272 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3274 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3275 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3279 for (unsigned i = 0; i < 8; ++i) {
3280 SDOperand Elt = V1Elts[i];
3281 if (Elt.getOpcode() == ISD::UNDEF)
3283 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3286 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3287 DAG.getConstant(EltIdx - 8, PtrVT));
3288 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3289 DAG.getConstant(i, PtrVT));
3293 // All elements are from V1.
3295 for (unsigned i = 0; i < 8; ++i) {
3296 SDOperand Elt = V1Elts[i];
3297 if (Elt.getOpcode() == ISD::UNDEF)
3299 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3300 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3301 DAG.getConstant(EltIdx, PtrVT));
3302 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3303 DAG.getConstant(i, PtrVT));
3309 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3310 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3311 /// done when every pair / quad of shuffle mask elements point to elements in
3312 /// the right sequence. e.g.
3313 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3315 SDOperand RewriteAsNarrowerShuffle(SDOperand V1, SDOperand V2,
3317 SDOperand PermMask, SelectionDAG &DAG,
3318 TargetLowering &TLI) {
3319 unsigned NumElems = PermMask.getNumOperands();
3320 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3321 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3322 MVT::ValueType NewVT = MaskVT;
3324 case MVT::v4f32: NewVT = MVT::v2f64; break;
3325 case MVT::v4i32: NewVT = MVT::v2i64; break;
3326 case MVT::v8i16: NewVT = MVT::v4i32; break;
3327 case MVT::v16i8: NewVT = MVT::v4i32; break;
3328 default: assert(false && "Unexpected!");
3332 if (MVT::isInteger(VT))
3336 unsigned Scale = NumElems / NewWidth;
3337 SmallVector<SDOperand, 8> MaskVec;
3338 for (unsigned i = 0; i < NumElems; i += Scale) {
3339 unsigned StartIdx = ~0U;
3340 for (unsigned j = 0; j < Scale; ++j) {
3341 SDOperand Elt = PermMask.getOperand(i+j);
3342 if (Elt.getOpcode() == ISD::UNDEF)
3344 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3345 if (StartIdx == ~0U)
3346 StartIdx = EltIdx - (EltIdx % Scale);
3347 if (EltIdx != StartIdx + j)
3350 if (StartIdx == ~0U)
3351 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
3353 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MVT::i32));
3356 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3357 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3358 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3359 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3360 &MaskVec[0], MaskVec.size()));
3364 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3365 SDOperand V1 = Op.getOperand(0);
3366 SDOperand V2 = Op.getOperand(1);
3367 SDOperand PermMask = Op.getOperand(2);
3368 MVT::ValueType VT = Op.getValueType();
3369 unsigned NumElems = PermMask.getNumOperands();
3370 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3371 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3372 bool V1IsSplat = false;
3373 bool V2IsSplat = false;
3375 if (isUndefShuffle(Op.Val))
3376 return DAG.getNode(ISD::UNDEF, VT);
3378 if (isZeroShuffle(Op.Val))
3379 return getZeroVector(VT, DAG);
3381 if (isIdentityMask(PermMask.Val))
3383 else if (isIdentityMask(PermMask.Val, true))
3386 if (isSplatMask(PermMask.Val)) {
3387 if (NumElems <= 4) return Op;
3388 // Promote it to a v4i32 splat.
3389 return PromoteSplat(Op, DAG);
3392 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3394 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3395 SDOperand NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3397 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3398 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3399 // FIXME: Figure out a cleaner way to do this.
3400 // Try to make use of movq to zero out the top part.
3401 if (ISD::isBuildVectorAllZeros(V2.Val)) {
3402 SDOperand NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3404 SDOperand NewV1 = NewOp.getOperand(0);
3405 SDOperand NewV2 = NewOp.getOperand(1);
3406 SDOperand NewMask = NewOp.getOperand(2);
3407 if (isCommutedMOVL(NewMask.Val, true, false)) {
3408 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
3409 NewOp = DAG.getNode(ISD::VECTOR_SHUFFLE, NewOp.getValueType(),
3410 NewV1, NewV2, getMOVLMask(2, DAG));
3411 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3414 } else if (ISD::isBuildVectorAllZeros(V1.Val)) {
3415 SDOperand NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3416 if (NewOp.Val && X86::isMOVLMask(NewOp.getOperand(2).Val))
3417 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3421 if (X86::isMOVLMask(PermMask.Val))
3422 return (V1IsUndef) ? V2 : Op;
3424 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3425 X86::isMOVSLDUPMask(PermMask.Val) ||
3426 X86::isMOVHLPSMask(PermMask.Val) ||
3427 X86::isMOVHPMask(PermMask.Val) ||
3428 X86::isMOVLPMask(PermMask.Val))
3431 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3432 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3433 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3435 bool Commuted = false;
3436 // FIXME: This should also accept a bitcast of a splat? Be careful, not
3437 // 1,1,1,1 -> v8i16 though.
3438 V1IsSplat = isSplatVector(V1.Val);
3439 V2IsSplat = isSplatVector(V2.Val);
3441 // Canonicalize the splat or undef, if present, to be on the RHS.
3442 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3443 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3444 std::swap(V1IsSplat, V2IsSplat);
3445 std::swap(V1IsUndef, V2IsUndef);
3449 // FIXME: Figure out a cleaner way to do this.
3450 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3451 if (V2IsUndef) return V1;
3452 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3454 // V2 is a splat, so the mask may be malformed. That is, it may point
3455 // to any V2 element. The instruction selectior won't like this. Get
3456 // a corrected mask and commute to form a proper MOVS{S|D}.
3457 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3458 if (NewMask.Val != PermMask.Val)
3459 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3464 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3465 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3466 X86::isUNPCKLMask(PermMask.Val) ||
3467 X86::isUNPCKHMask(PermMask.Val))
3471 // Normalize mask so all entries that point to V2 points to its first
3472 // element then try to match unpck{h|l} again. If match, return a
3473 // new vector_shuffle with the corrected mask.
3474 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3475 if (NewMask.Val != PermMask.Val) {
3476 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3477 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3478 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3479 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3480 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3481 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3486 // Normalize the node to match x86 shuffle ops if needed
3487 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3488 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3491 // Commute is back and try unpck* again.
3492 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3493 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3494 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3495 X86::isUNPCKLMask(PermMask.Val) ||
3496 X86::isUNPCKHMask(PermMask.Val))
3500 // If VT is integer, try PSHUF* first, then SHUFP*.
3501 if (MVT::isInteger(VT)) {
3502 // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
3503 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
3504 if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
3505 X86::isPSHUFDMask(PermMask.Val)) ||
3506 X86::isPSHUFHWMask(PermMask.Val) ||
3507 X86::isPSHUFLWMask(PermMask.Val)) {
3508 if (V2.getOpcode() != ISD::UNDEF)
3509 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3510 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3514 if (X86::isSHUFPMask(PermMask.Val) &&
3515 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
3518 // Floating point cases in the other order.
3519 if (X86::isSHUFPMask(PermMask.Val))
3521 if (X86::isPSHUFDMask(PermMask.Val) ||
3522 X86::isPSHUFHWMask(PermMask.Val) ||
3523 X86::isPSHUFLWMask(PermMask.Val)) {
3524 if (V2.getOpcode() != ISD::UNDEF)
3525 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3526 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3531 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
3532 if (VT == MVT::v8i16) {
3533 SDOperand NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
3538 // Handle all 4 wide cases with a number of shuffles.
3539 if (NumElems == 4 && MVT::getSizeInBits(VT) != 64) {
3540 // Don't do this for MMX.
3541 MVT::ValueType MaskVT = PermMask.getValueType();
3542 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3543 SmallVector<std::pair<int, int>, 8> Locs;
3544 Locs.reserve(NumElems);
3545 SmallVector<SDOperand, 8> Mask1(NumElems,
3546 DAG.getNode(ISD::UNDEF, MaskEVT));
3547 SmallVector<SDOperand, 8> Mask2(NumElems,
3548 DAG.getNode(ISD::UNDEF, MaskEVT));
3551 // If no more than two elements come from either vector. This can be
3552 // implemented with two shuffles. First shuffle gather the elements.
3553 // The second shuffle, which takes the first shuffle as both of its
3554 // vector operands, put the elements into the right order.
3555 for (unsigned i = 0; i != NumElems; ++i) {
3556 SDOperand Elt = PermMask.getOperand(i);
3557 if (Elt.getOpcode() == ISD::UNDEF) {
3558 Locs[i] = std::make_pair(-1, -1);
3560 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3561 if (Val < NumElems) {
3562 Locs[i] = std::make_pair(0, NumLo);
3566 Locs[i] = std::make_pair(1, NumHi);
3567 if (2+NumHi < NumElems)
3568 Mask1[2+NumHi] = Elt;
3573 if (NumLo <= 2 && NumHi <= 2) {
3574 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3575 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3576 &Mask1[0], Mask1.size()));
3577 for (unsigned i = 0; i != NumElems; ++i) {
3578 if (Locs[i].first == -1)
3581 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3582 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3583 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3587 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3588 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3589 &Mask2[0], Mask2.size()));
3592 // Break it into (shuffle shuffle_hi, shuffle_lo).
3594 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3595 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3596 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3597 unsigned MaskIdx = 0;
3599 unsigned HiIdx = NumElems/2;
3600 for (unsigned i = 0; i != NumElems; ++i) {
3601 if (i == NumElems/2) {
3607 SDOperand Elt = PermMask.getOperand(i);
3608 if (Elt.getOpcode() == ISD::UNDEF) {
3609 Locs[i] = std::make_pair(-1, -1);
3610 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3611 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3612 (*MaskPtr)[LoIdx] = Elt;
3615 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3616 (*MaskPtr)[HiIdx] = Elt;
3621 SDOperand LoShuffle =
3622 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3623 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3624 &LoMask[0], LoMask.size()));
3625 SDOperand HiShuffle =
3626 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3627 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3628 &HiMask[0], HiMask.size()));
3629 SmallVector<SDOperand, 8> MaskOps;
3630 for (unsigned i = 0; i != NumElems; ++i) {
3631 if (Locs[i].first == -1) {
3632 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3634 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3635 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3638 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3639 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3640 &MaskOps[0], MaskOps.size()));
3647 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3648 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3651 MVT::ValueType VT = Op.getValueType();
3652 // TODO: handle v16i8.
3653 if (MVT::getSizeInBits(VT) == 16) {
3654 SDOperand Vec = Op.getOperand(0);
3655 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3657 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
3658 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
3659 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
3661 // Transform it so it match pextrw which produces a 32-bit result.
3662 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3663 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3664 Op.getOperand(0), Op.getOperand(1));
3665 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3666 DAG.getValueType(VT));
3667 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3668 } else if (MVT::getSizeInBits(VT) == 32) {
3669 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3672 // SHUFPS the element to the lowest double word, then movss.
3673 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3674 SmallVector<SDOperand, 8> IdxVec;
3676 push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
3678 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3680 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3682 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3683 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3684 &IdxVec[0], IdxVec.size());
3685 SDOperand Vec = Op.getOperand(0);
3686 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3687 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3688 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3689 DAG.getIntPtrConstant(0));
3690 } else if (MVT::getSizeInBits(VT) == 64) {
3691 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3695 // UNPCKHPD the element to the lowest double word, then movsd.
3696 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3697 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3698 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3699 SmallVector<SDOperand, 8> IdxVec;
3700 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
3702 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3703 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3704 &IdxVec[0], IdxVec.size());
3705 SDOperand Vec = Op.getOperand(0);
3706 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3707 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3708 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3709 DAG.getIntPtrConstant(0));
3716 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3717 MVT::ValueType VT = Op.getValueType();
3718 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3722 SDOperand N0 = Op.getOperand(0);
3723 SDOperand N1 = Op.getOperand(1);
3724 SDOperand N2 = Op.getOperand(2);
3726 if (MVT::getSizeInBits(EVT) == 16) {
3727 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3728 // as its second argument.
3729 if (N1.getValueType() != MVT::i32)
3730 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3731 if (N2.getValueType() != MVT::i32)
3732 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
3733 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3739 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3740 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3741 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3744 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3745 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3746 // one of the above mentioned nodes. It has to be wrapped because otherwise
3747 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3748 // be used to form addressing mode. These wrapped nodes will be selected
3751 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3752 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3753 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3755 CP->getAlignment());
3756 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3757 // With PIC, the address is actually $g + Offset.
3758 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3759 !Subtarget->isPICStyleRIPRel()) {
3760 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3761 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3769 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3770 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3771 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3772 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3773 // With PIC, the address is actually $g + Offset.
3774 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3775 !Subtarget->isPICStyleRIPRel()) {
3776 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3777 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3781 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3782 // load the value at address GV, not the value of GV itself. This means that
3783 // the GlobalAddress must be in the base or index register of the address, not
3784 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3785 // The same applies for external symbols during PIC codegen
3786 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3787 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3792 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3794 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3795 const MVT::ValueType PtrVT) {
3797 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
3798 DAG.getNode(X86ISD::GlobalBaseReg,
3800 InFlag = Chain.getValue(1);
3802 // emit leal symbol@TLSGD(,%ebx,1), %eax
3803 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3804 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3805 GA->getValueType(0),
3807 SDOperand Ops[] = { Chain, TGA, InFlag };
3808 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3809 InFlag = Result.getValue(2);
3810 Chain = Result.getValue(1);
3812 // call ___tls_get_addr. This function receives its argument in
3813 // the register EAX.
3814 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
3815 InFlag = Chain.getValue(1);
3817 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3818 SDOperand Ops1[] = { Chain,
3819 DAG.getTargetExternalSymbol("___tls_get_addr",
3821 DAG.getRegister(X86::EAX, PtrVT),
3822 DAG.getRegister(X86::EBX, PtrVT),
3824 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
3825 InFlag = Chain.getValue(1);
3827 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
3830 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
3831 // "local exec" model.
3833 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3834 const MVT::ValueType PtrVT) {
3835 // Get the Thread Pointer
3836 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
3837 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
3839 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3840 GA->getValueType(0),
3842 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
3844 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
3845 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset, NULL, 0);
3847 // The address of the thread local variable is the add of the thread
3848 // pointer with the offset of the variable.
3849 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
3853 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
3854 // TODO: implement the "local dynamic" model
3855 // TODO: implement the "initial exec"model for pic executables
3856 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
3857 "TLS not implemented for non-ELF and 64-bit targets");
3858 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3859 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
3860 // otherwise use the "Local Exec"TLS Model
3861 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
3862 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
3864 return LowerToTLSExecModel(GA, DAG, getPointerTy());
3868 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3869 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3870 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3871 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3872 // With PIC, the address is actually $g + Offset.
3873 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3874 !Subtarget->isPICStyleRIPRel()) {
3875 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3876 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3883 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3884 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3885 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3886 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3887 // With PIC, the address is actually $g + Offset.
3888 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3889 !Subtarget->isPICStyleRIPRel()) {
3890 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3891 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3898 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
3899 /// take a 2 x i32 value to shift plus a shift amount.
3900 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3901 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3902 "Not an i64 shift!");
3903 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3904 SDOperand ShOpLo = Op.getOperand(0);
3905 SDOperand ShOpHi = Op.getOperand(1);
3906 SDOperand ShAmt = Op.getOperand(2);
3907 SDOperand Tmp1 = isSRA ?
3908 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3909 DAG.getConstant(0, MVT::i32);
3911 SDOperand Tmp2, Tmp3;
3912 if (Op.getOpcode() == ISD::SHL_PARTS) {
3913 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3914 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3916 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3917 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3920 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3921 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3922 DAG.getConstant(32, MVT::i8));
3923 SDOperand Cond = DAG.getNode(X86ISD::CMP, MVT::i32,
3924 AndNode, DAG.getConstant(0, MVT::i8));
3927 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3928 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3929 SmallVector<SDOperand, 4> Ops;
3930 if (Op.getOpcode() == ISD::SHL_PARTS) {
3931 Ops.push_back(Tmp2);
3932 Ops.push_back(Tmp3);
3934 Ops.push_back(Cond);
3935 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3938 Ops.push_back(Tmp3);
3939 Ops.push_back(Tmp1);
3941 Ops.push_back(Cond);
3942 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3944 Ops.push_back(Tmp2);
3945 Ops.push_back(Tmp3);
3947 Ops.push_back(Cond);
3948 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3951 Ops.push_back(Tmp3);
3952 Ops.push_back(Tmp1);
3954 Ops.push_back(Cond);
3955 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3958 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3962 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3965 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3966 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3967 Op.getOperand(0).getValueType() >= MVT::i16 &&
3968 "Unknown SINT_TO_FP to lower!");
3971 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3972 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3973 MachineFunction &MF = DAG.getMachineFunction();
3974 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3975 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3976 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3977 StackSlot, NULL, 0);
3979 // These are really Legal; caller falls through into that case.
3980 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
3982 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
3983 Subtarget->is64Bit())
3988 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
3990 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3992 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
3993 SmallVector<SDOperand, 8> Ops;
3994 Ops.push_back(Chain);
3995 Ops.push_back(StackSlot);
3996 Ops.push_back(DAG.getValueType(SrcVT));
3997 Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3998 Tys, &Ops[0], Ops.size());
4001 Chain = Result.getValue(1);
4002 SDOperand InFlag = Result.getValue(2);
4004 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4005 // shouldn't be necessary except that RFP cannot be live across
4006 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4007 MachineFunction &MF = DAG.getMachineFunction();
4008 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4009 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4010 Tys = DAG.getVTList(MVT::Other);
4011 SmallVector<SDOperand, 8> Ops;
4012 Ops.push_back(Chain);
4013 Ops.push_back(Result);
4014 Ops.push_back(StackSlot);
4015 Ops.push_back(DAG.getValueType(Op.getValueType()));
4016 Ops.push_back(InFlag);
4017 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4018 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
4024 std::pair<SDOperand,SDOperand> X86TargetLowering::
4025 FP_TO_SINTHelper(SDOperand Op, SelectionDAG &DAG) {
4026 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
4027 "Unknown FP_TO_SINT to lower!");
4029 // These are really Legal.
4030 if (Op.getValueType() == MVT::i32 &&
4031 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4032 return std::make_pair(SDOperand(), SDOperand());
4033 if (Subtarget->is64Bit() &&
4034 Op.getValueType() == MVT::i64 &&
4035 Op.getOperand(0).getValueType() != MVT::f80)
4036 return std::make_pair(SDOperand(), SDOperand());
4038 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4040 MachineFunction &MF = DAG.getMachineFunction();
4041 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
4042 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4043 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4045 switch (Op.getValueType()) {
4046 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4047 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4048 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4049 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4052 SDOperand Chain = DAG.getEntryNode();
4053 SDOperand Value = Op.getOperand(0);
4054 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4055 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4056 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
4057 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4059 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4061 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4062 Chain = Value.getValue(1);
4063 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4064 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4067 // Build the FP_TO_INT*_IN_MEM
4068 SDOperand Ops[] = { Chain, Value, StackSlot };
4069 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4071 return std::make_pair(FIST, StackSlot);
4074 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
4075 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(Op, DAG);
4076 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4077 if (FIST.Val == 0) return SDOperand();
4080 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4083 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4084 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(SDOperand(N, 0), DAG);
4085 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4086 if (FIST.Val == 0) return 0;
4088 // Return an i64 load from the stack slot.
4089 SDOperand Res = DAG.getLoad(MVT::i64, FIST, StackSlot, NULL, 0);
4091 // Use a MERGE_VALUES node to drop the chain result value.
4092 return DAG.getNode(ISD::MERGE_VALUES, MVT::i64, Res).Val;
4095 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
4096 MVT::ValueType VT = Op.getValueType();
4097 MVT::ValueType EltVT = VT;
4098 if (MVT::isVector(VT))
4099 EltVT = MVT::getVectorElementType(VT);
4100 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4101 std::vector<Constant*> CV;
4102 if (EltVT == MVT::f64) {
4103 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, ~(1ULL << 63))));
4107 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, ~(1U << 31))));
4113 Constant *C = ConstantVector::get(CV);
4114 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4115 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4117 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4120 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
4121 MVT::ValueType VT = Op.getValueType();
4122 MVT::ValueType EltVT = VT;
4123 unsigned EltNum = 1;
4124 if (MVT::isVector(VT)) {
4125 EltVT = MVT::getVectorElementType(VT);
4126 EltNum = MVT::getVectorNumElements(VT);
4128 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4129 std::vector<Constant*> CV;
4130 if (EltVT == MVT::f64) {
4131 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, 1ULL << 63)));
4135 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, 1U << 31)));
4141 Constant *C = ConstantVector::get(CV);
4142 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4143 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4145 if (MVT::isVector(VT)) {
4146 return DAG.getNode(ISD::BIT_CONVERT, VT,
4147 DAG.getNode(ISD::XOR, MVT::v2i64,
4148 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4149 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4151 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4155 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
4156 SDOperand Op0 = Op.getOperand(0);
4157 SDOperand Op1 = Op.getOperand(1);
4158 MVT::ValueType VT = Op.getValueType();
4159 MVT::ValueType SrcVT = Op1.getValueType();
4160 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
4162 // If second operand is smaller, extend it first.
4163 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
4164 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4166 SrcTy = MVT::getTypeForValueType(SrcVT);
4168 // And if it is bigger, shrink it first.
4169 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4170 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
4172 SrcTy = MVT::getTypeForValueType(SrcVT);
4175 // At this point the operands and the result should have the same
4176 // type, and that won't be f80 since that is not custom lowered.
4178 // First get the sign bit of second operand.
4179 std::vector<Constant*> CV;
4180 if (SrcVT == MVT::f64) {
4181 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 1ULL << 63))));
4182 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4184 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 1U << 31))));
4185 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4186 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4187 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4189 Constant *C = ConstantVector::get(CV);
4190 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4191 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx, NULL, 0,
4193 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4195 // Shift sign bit right or left if the two operands have different types.
4196 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4197 // Op0 is MVT::f32, Op1 is MVT::f64.
4198 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4199 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4200 DAG.getConstant(32, MVT::i32));
4201 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4202 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4203 DAG.getIntPtrConstant(0));
4206 // Clear first operand sign bit.
4208 if (VT == MVT::f64) {
4209 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, ~(1ULL << 63)))));
4210 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4212 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, ~(1U << 31)))));
4213 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4214 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4215 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4217 C = ConstantVector::get(CV);
4218 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4219 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4221 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4223 // Or the value with the sign bit.
4224 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4227 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
4228 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4230 SDOperand Op0 = Op.getOperand(0);
4231 SDOperand Op1 = Op.getOperand(1);
4232 SDOperand CC = Op.getOperand(2);
4233 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4234 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
4237 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4239 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4240 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4241 DAG.getConstant(X86CC, MVT::i8), Cond);
4244 assert(isFP && "Illegal integer SetCC!");
4246 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4247 switch (SetCCOpcode) {
4248 default: assert(false && "Illegal floating point SetCC!");
4249 case ISD::SETOEQ: { // !PF & ZF
4250 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4251 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4252 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4253 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4254 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4256 case ISD::SETUNE: { // PF | !ZF
4257 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4258 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4259 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4260 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4261 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4267 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
4268 bool addTest = true;
4269 SDOperand Cond = Op.getOperand(0);
4272 if (Cond.getOpcode() == ISD::SETCC)
4273 Cond = LowerSETCC(Cond, DAG);
4275 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4276 // setting operand in place of the X86ISD::SETCC.
4277 if (Cond.getOpcode() == X86ISD::SETCC) {
4278 CC = Cond.getOperand(0);
4280 SDOperand Cmp = Cond.getOperand(1);
4281 unsigned Opc = Cmp.getOpcode();
4282 MVT::ValueType VT = Op.getValueType();
4284 bool IllegalFPCMov = false;
4285 if (MVT::isFloatingPoint(VT) && !MVT::isVector(VT) &&
4286 !isScalarFPTypeInSSEReg(VT)) // FPStack?
4287 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4289 if ((Opc == X86ISD::CMP ||
4290 Opc == X86ISD::COMI ||
4291 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4298 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4299 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4302 const MVT::ValueType *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4304 SmallVector<SDOperand, 4> Ops;
4305 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4306 // condition is true.
4307 Ops.push_back(Op.getOperand(2));
4308 Ops.push_back(Op.getOperand(1));
4310 Ops.push_back(Cond);
4311 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4314 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
4315 bool addTest = true;
4316 SDOperand Chain = Op.getOperand(0);
4317 SDOperand Cond = Op.getOperand(1);
4318 SDOperand Dest = Op.getOperand(2);
4321 if (Cond.getOpcode() == ISD::SETCC)
4322 Cond = LowerSETCC(Cond, DAG);
4324 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4325 // setting operand in place of the X86ISD::SETCC.
4326 if (Cond.getOpcode() == X86ISD::SETCC) {
4327 CC = Cond.getOperand(0);
4329 SDOperand Cmp = Cond.getOperand(1);
4330 unsigned Opc = Cmp.getOpcode();
4331 if (Opc == X86ISD::CMP ||
4332 Opc == X86ISD::COMI ||
4333 Opc == X86ISD::UCOMI) {
4340 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4341 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4343 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
4344 Chain, Op.getOperand(2), CC, Cond);
4348 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
4349 // Calls to _alloca is needed to probe the stack when allocating more than 4k
4350 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
4351 // that the guard pages used by the OS virtual memory manager are allocated in
4352 // correct sequence.
4354 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
4355 SelectionDAG &DAG) {
4356 assert(Subtarget->isTargetCygMing() &&
4357 "This should be used only on Cygwin/Mingw targets");
4360 SDOperand Chain = Op.getOperand(0);
4361 SDOperand Size = Op.getOperand(1);
4362 // FIXME: Ensure alignment here
4366 MVT::ValueType IntPtr = getPointerTy();
4367 MVT::ValueType SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
4369 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
4370 Flag = Chain.getValue(1);
4372 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4373 SDOperand Ops[] = { Chain,
4374 DAG.getTargetExternalSymbol("_alloca", IntPtr),
4375 DAG.getRegister(X86::EAX, IntPtr),
4377 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
4378 Flag = Chain.getValue(1);
4380 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
4382 std::vector<MVT::ValueType> Tys;
4383 Tys.push_back(SPTy);
4384 Tys.push_back(MVT::Other);
4385 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
4386 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
4389 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4390 SDOperand InFlag(0, 0);
4391 SDOperand Chain = Op.getOperand(0);
4393 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4394 if (Align == 0) Align = 1;
4396 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4397 // If not DWORD aligned or size is more than the threshold, call memset.
4398 // The libc version is likely to be faster for these cases. It can use the
4399 // address value and run time information about the CPU.
4400 if ((Align & 3) != 0 ||
4401 (I && I->getValue() > Subtarget->getMaxInlineSizeThreshold())) {
4402 MVT::ValueType IntPtr = getPointerTy();
4403 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4404 TargetLowering::ArgListTy Args;
4405 TargetLowering::ArgListEntry Entry;
4406 Entry.Node = Op.getOperand(1);
4407 Entry.Ty = IntPtrTy;
4408 Args.push_back(Entry);
4409 // Extend the unsigned i8 argument to be an int value for the call.
4410 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4411 Entry.Ty = IntPtrTy;
4412 Args.push_back(Entry);
4413 Entry.Node = Op.getOperand(3);
4414 Args.push_back(Entry);
4415 std::pair<SDOperand,SDOperand> CallResult =
4416 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4417 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4418 return CallResult.second;
4423 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4424 unsigned BytesLeft = 0;
4425 bool TwoRepStos = false;
4428 uint64_t Val = ValC->getValue() & 255;
4430 // If the value is a constant, then we can potentially use larger sets.
4431 switch (Align & 3) {
4432 case 2: // WORD aligned
4435 Val = (Val << 8) | Val;
4437 case 0: // DWORD aligned
4440 Val = (Val << 8) | Val;
4441 Val = (Val << 16) | Val;
4442 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4445 Val = (Val << 32) | Val;
4448 default: // Byte aligned
4451 Count = Op.getOperand(3);
4455 if (AVT > MVT::i8) {
4457 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4458 Count = DAG.getIntPtrConstant(I->getValue() / UBytes);
4459 BytesLeft = I->getValue() % UBytes;
4461 assert(AVT >= MVT::i32 &&
4462 "Do not use rep;stos if not at least DWORD aligned");
4463 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4464 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4469 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4471 InFlag = Chain.getValue(1);
4474 Count = Op.getOperand(3);
4475 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4476 InFlag = Chain.getValue(1);
4479 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4481 InFlag = Chain.getValue(1);
4482 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4483 Op.getOperand(1), InFlag);
4484 InFlag = Chain.getValue(1);
4486 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4487 SmallVector<SDOperand, 8> Ops;
4488 Ops.push_back(Chain);
4489 Ops.push_back(DAG.getValueType(AVT));
4490 Ops.push_back(InFlag);
4491 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4494 InFlag = Chain.getValue(1);
4495 Count = Op.getOperand(3);
4496 MVT::ValueType CVT = Count.getValueType();
4497 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4498 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4499 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4501 InFlag = Chain.getValue(1);
4502 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4504 Ops.push_back(Chain);
4505 Ops.push_back(DAG.getValueType(MVT::i8));
4506 Ops.push_back(InFlag);
4507 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4508 } else if (BytesLeft) {
4509 // Issue stores for the last 1 - 7 bytes.
4511 unsigned Val = ValC->getValue() & 255;
4512 unsigned Offset = I->getValue() - BytesLeft;
4513 SDOperand DstAddr = Op.getOperand(1);
4514 MVT::ValueType AddrVT = DstAddr.getValueType();
4515 if (BytesLeft >= 4) {
4516 Val = (Val << 8) | Val;
4517 Val = (Val << 16) | Val;
4518 Value = DAG.getConstant(Val, MVT::i32);
4519 Chain = DAG.getStore(Chain, Value,
4520 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4521 DAG.getConstant(Offset, AddrVT)),
4526 if (BytesLeft >= 2) {
4527 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4528 Chain = DAG.getStore(Chain, Value,
4529 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4530 DAG.getConstant(Offset, AddrVT)),
4535 if (BytesLeft == 1) {
4536 Value = DAG.getConstant(Val, MVT::i8);
4537 Chain = DAG.getStore(Chain, Value,
4538 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4539 DAG.getConstant(Offset, AddrVT)),
4547 SDOperand X86TargetLowering::LowerMEMCPYInline(SDOperand Chain,
4552 SelectionDAG &DAG) {
4554 unsigned BytesLeft = 0;
4555 switch (Align & 3) {
4556 case 2: // WORD aligned
4559 case 0: // DWORD aligned
4561 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4564 default: // Byte aligned
4569 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4570 SDOperand Count = DAG.getIntPtrConstant(Size / UBytes);
4571 BytesLeft = Size % UBytes;
4573 SDOperand InFlag(0, 0);
4574 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4576 InFlag = Chain.getValue(1);
4577 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4579 InFlag = Chain.getValue(1);
4580 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4582 InFlag = Chain.getValue(1);
4584 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4585 SmallVector<SDOperand, 8> Ops;
4586 Ops.push_back(Chain);
4587 Ops.push_back(DAG.getValueType(AVT));
4588 Ops.push_back(InFlag);
4589 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4592 // Issue loads and stores for the last 1 - 7 bytes.
4593 unsigned Offset = Size - BytesLeft;
4594 SDOperand DstAddr = Dest;
4595 MVT::ValueType DstVT = DstAddr.getValueType();
4596 SDOperand SrcAddr = Source;
4597 MVT::ValueType SrcVT = SrcAddr.getValueType();
4599 if (BytesLeft >= 4) {
4600 Value = DAG.getLoad(MVT::i32, Chain,
4601 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4602 DAG.getConstant(Offset, SrcVT)),
4604 Chain = Value.getValue(1);
4605 Chain = DAG.getStore(Chain, Value,
4606 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4607 DAG.getConstant(Offset, DstVT)),
4612 if (BytesLeft >= 2) {
4613 Value = DAG.getLoad(MVT::i16, Chain,
4614 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4615 DAG.getConstant(Offset, SrcVT)),
4617 Chain = Value.getValue(1);
4618 Chain = DAG.getStore(Chain, Value,
4619 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4620 DAG.getConstant(Offset, DstVT)),
4626 if (BytesLeft == 1) {
4627 Value = DAG.getLoad(MVT::i8, Chain,
4628 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4629 DAG.getConstant(Offset, SrcVT)),
4631 Chain = Value.getValue(1);
4632 Chain = DAG.getStore(Chain, Value,
4633 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4634 DAG.getConstant(Offset, DstVT)),
4642 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
4643 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
4644 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4645 SDOperand TheChain = N->getOperand(0);
4646 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
4647 if (Subtarget->is64Bit()) {
4648 SDOperand rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4649 SDOperand rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
4650 MVT::i64, rax.getValue(2));
4651 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
4652 DAG.getConstant(32, MVT::i8));
4654 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
4657 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4658 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
4661 SDOperand eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4662 SDOperand edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
4663 MVT::i32, eax.getValue(2));
4664 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
4665 SDOperand Ops[] = { eax, edx };
4666 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
4668 // Use a MERGE_VALUES to return the value and chain.
4669 Ops[1] = edx.getValue(1);
4670 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4671 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
4674 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4675 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4677 if (!Subtarget->is64Bit()) {
4678 // vastart just stores the address of the VarArgsFrameIndex slot into the
4679 // memory location argument.
4680 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4681 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4686 // gp_offset (0 - 6 * 8)
4687 // fp_offset (48 - 48 + 8 * 16)
4688 // overflow_arg_area (point to parameters coming in memory).
4690 SmallVector<SDOperand, 8> MemOps;
4691 SDOperand FIN = Op.getOperand(1);
4693 SDOperand Store = DAG.getStore(Op.getOperand(0),
4694 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4695 FIN, SV->getValue(), SV->getOffset());
4696 MemOps.push_back(Store);
4699 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
4700 Store = DAG.getStore(Op.getOperand(0),
4701 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4702 FIN, SV->getValue(), SV->getOffset());
4703 MemOps.push_back(Store);
4705 // Store ptr to overflow_arg_area
4706 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
4707 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4708 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4710 MemOps.push_back(Store);
4712 // Store ptr to reg_save_area.
4713 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
4714 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4715 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4717 MemOps.push_back(Store);
4718 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4721 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
4722 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
4723 SDOperand Chain = Op.getOperand(0);
4724 SDOperand DstPtr = Op.getOperand(1);
4725 SDOperand SrcPtr = Op.getOperand(2);
4726 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
4727 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
4729 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
4730 SrcSV->getValue(), SrcSV->getOffset());
4731 Chain = SrcPtr.getValue(1);
4732 for (unsigned i = 0; i < 3; ++i) {
4733 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
4734 SrcSV->getValue(), SrcSV->getOffset());
4735 Chain = Val.getValue(1);
4736 Chain = DAG.getStore(Chain, Val, DstPtr,
4737 DstSV->getValue(), DstSV->getOffset());
4740 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
4741 DAG.getIntPtrConstant(8));
4742 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
4743 DAG.getIntPtrConstant(8));
4749 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4750 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4752 default: return SDOperand(); // Don't custom lower most intrinsics.
4753 // Comparison intrinsics.
4754 case Intrinsic::x86_sse_comieq_ss:
4755 case Intrinsic::x86_sse_comilt_ss:
4756 case Intrinsic::x86_sse_comile_ss:
4757 case Intrinsic::x86_sse_comigt_ss:
4758 case Intrinsic::x86_sse_comige_ss:
4759 case Intrinsic::x86_sse_comineq_ss:
4760 case Intrinsic::x86_sse_ucomieq_ss:
4761 case Intrinsic::x86_sse_ucomilt_ss:
4762 case Intrinsic::x86_sse_ucomile_ss:
4763 case Intrinsic::x86_sse_ucomigt_ss:
4764 case Intrinsic::x86_sse_ucomige_ss:
4765 case Intrinsic::x86_sse_ucomineq_ss:
4766 case Intrinsic::x86_sse2_comieq_sd:
4767 case Intrinsic::x86_sse2_comilt_sd:
4768 case Intrinsic::x86_sse2_comile_sd:
4769 case Intrinsic::x86_sse2_comigt_sd:
4770 case Intrinsic::x86_sse2_comige_sd:
4771 case Intrinsic::x86_sse2_comineq_sd:
4772 case Intrinsic::x86_sse2_ucomieq_sd:
4773 case Intrinsic::x86_sse2_ucomilt_sd:
4774 case Intrinsic::x86_sse2_ucomile_sd:
4775 case Intrinsic::x86_sse2_ucomigt_sd:
4776 case Intrinsic::x86_sse2_ucomige_sd:
4777 case Intrinsic::x86_sse2_ucomineq_sd: {
4779 ISD::CondCode CC = ISD::SETCC_INVALID;
4782 case Intrinsic::x86_sse_comieq_ss:
4783 case Intrinsic::x86_sse2_comieq_sd:
4787 case Intrinsic::x86_sse_comilt_ss:
4788 case Intrinsic::x86_sse2_comilt_sd:
4792 case Intrinsic::x86_sse_comile_ss:
4793 case Intrinsic::x86_sse2_comile_sd:
4797 case Intrinsic::x86_sse_comigt_ss:
4798 case Intrinsic::x86_sse2_comigt_sd:
4802 case Intrinsic::x86_sse_comige_ss:
4803 case Intrinsic::x86_sse2_comige_sd:
4807 case Intrinsic::x86_sse_comineq_ss:
4808 case Intrinsic::x86_sse2_comineq_sd:
4812 case Intrinsic::x86_sse_ucomieq_ss:
4813 case Intrinsic::x86_sse2_ucomieq_sd:
4814 Opc = X86ISD::UCOMI;
4817 case Intrinsic::x86_sse_ucomilt_ss:
4818 case Intrinsic::x86_sse2_ucomilt_sd:
4819 Opc = X86ISD::UCOMI;
4822 case Intrinsic::x86_sse_ucomile_ss:
4823 case Intrinsic::x86_sse2_ucomile_sd:
4824 Opc = X86ISD::UCOMI;
4827 case Intrinsic::x86_sse_ucomigt_ss:
4828 case Intrinsic::x86_sse2_ucomigt_sd:
4829 Opc = X86ISD::UCOMI;
4832 case Intrinsic::x86_sse_ucomige_ss:
4833 case Intrinsic::x86_sse2_ucomige_sd:
4834 Opc = X86ISD::UCOMI;
4837 case Intrinsic::x86_sse_ucomineq_ss:
4838 case Intrinsic::x86_sse2_ucomineq_sd:
4839 Opc = X86ISD::UCOMI;
4845 SDOperand LHS = Op.getOperand(1);
4846 SDOperand RHS = Op.getOperand(2);
4847 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4849 SDOperand Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
4850 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
4851 DAG.getConstant(X86CC, MVT::i8), Cond);
4852 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4857 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4858 // Depths > 0 not supported yet!
4859 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4862 // Just load the return address
4863 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4864 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4867 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4868 // Depths > 0 not supported yet!
4869 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4872 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4873 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4874 DAG.getIntPtrConstant(4));
4877 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
4878 SelectionDAG &DAG) {
4879 // Is not yet supported on x86-64
4880 if (Subtarget->is64Bit())
4883 return DAG.getIntPtrConstant(8);
4886 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
4888 assert(!Subtarget->is64Bit() &&
4889 "Lowering of eh_return builtin is not supported yet on x86-64");
4891 MachineFunction &MF = DAG.getMachineFunction();
4892 SDOperand Chain = Op.getOperand(0);
4893 SDOperand Offset = Op.getOperand(1);
4894 SDOperand Handler = Op.getOperand(2);
4896 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
4899 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
4900 DAG.getIntPtrConstant(-4UL));
4901 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
4902 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
4903 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
4904 MF.getRegInfo().addLiveOut(X86::ECX);
4906 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
4907 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
4910 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
4911 SelectionDAG &DAG) {
4912 SDOperand Root = Op.getOperand(0);
4913 SDOperand Trmp = Op.getOperand(1); // trampoline
4914 SDOperand FPtr = Op.getOperand(2); // nested function
4915 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
4917 SrcValueSDNode *TrmpSV = cast<SrcValueSDNode>(Op.getOperand(4));
4919 const X86InstrInfo *TII =
4920 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
4922 if (Subtarget->is64Bit()) {
4923 SDOperand OutChains[6];
4925 // Large code-model.
4927 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
4928 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
4930 const unsigned char N86R10 =
4931 ((X86RegisterInfo*)RegInfo)->getX86RegNum(X86::R10);
4932 const unsigned char N86R11 =
4933 ((X86RegisterInfo*)RegInfo)->getX86RegNum(X86::R11);
4935 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
4937 // Load the pointer to the nested function into R11.
4938 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
4939 SDOperand Addr = Trmp;
4940 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
4941 TrmpSV->getValue(), TrmpSV->getOffset());
4943 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
4944 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpSV->getValue(),
4945 TrmpSV->getOffset() + 2, false, 2);
4947 // Load the 'nest' parameter value into R10.
4948 // R10 is specified in X86CallingConv.td
4949 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
4950 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
4951 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
4952 TrmpSV->getValue(), TrmpSV->getOffset() + 10);
4954 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
4955 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
4956 TrmpSV->getOffset() + 12, false, 2);
4958 // Jump to the nested function.
4959 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
4960 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
4961 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
4962 TrmpSV->getValue(), TrmpSV->getOffset() + 20);
4964 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
4965 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
4966 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
4967 TrmpSV->getValue(), TrmpSV->getOffset() + 22);
4970 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
4971 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
4973 const Function *Func =
4974 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
4975 unsigned CC = Func->getCallingConv();
4980 assert(0 && "Unsupported calling convention");
4981 case CallingConv::C:
4982 case CallingConv::X86_StdCall: {
4983 // Pass 'nest' parameter in ECX.
4984 // Must be kept in sync with X86CallingConv.td
4987 // Check that ECX wasn't needed by an 'inreg' parameter.
4988 const FunctionType *FTy = Func->getFunctionType();
4989 const ParamAttrsList *Attrs = Func->getParamAttrs();
4991 if (Attrs && !Func->isVarArg()) {
4992 unsigned InRegCount = 0;
4995 for (FunctionType::param_iterator I = FTy->param_begin(),
4996 E = FTy->param_end(); I != E; ++I, ++Idx)
4997 if (Attrs->paramHasAttr(Idx, ParamAttr::InReg))
4998 // FIXME: should only count parameters that are lowered to integers.
4999 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
5001 if (InRegCount > 2) {
5002 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5008 case CallingConv::X86_FastCall:
5009 // Pass 'nest' parameter in EAX.
5010 // Must be kept in sync with X86CallingConv.td
5015 SDOperand OutChains[4];
5016 SDOperand Addr, Disp;
5018 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5019 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5021 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5022 const unsigned char N86Reg =
5023 ((X86RegisterInfo*)RegInfo)->getX86RegNum(NestReg);
5024 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5025 Trmp, TrmpSV->getValue(), TrmpSV->getOffset());
5027 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5028 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
5029 TrmpSV->getOffset() + 1, false, 1);
5031 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5032 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5033 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5034 TrmpSV->getValue() + 5, TrmpSV->getOffset());
5036 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5037 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpSV->getValue(),
5038 TrmpSV->getOffset() + 6, false, 1);
5041 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5042 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
5046 SDOperand X86TargetLowering::LowerFLT_ROUNDS_(SDOperand Op, SelectionDAG &DAG) {
5048 The rounding mode is in bits 11:10 of FPSR, and has the following
5055 FLT_ROUNDS, on the other hand, expects the following:
5062 To perform the conversion, we do:
5063 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5066 MachineFunction &MF = DAG.getMachineFunction();
5067 const TargetMachine &TM = MF.getTarget();
5068 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5069 unsigned StackAlignment = TFI.getStackAlignment();
5070 MVT::ValueType VT = Op.getValueType();
5072 // Save FP Control Word to stack slot
5073 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5074 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5076 SDOperand Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5077 DAG.getEntryNode(), StackSlot);
5079 // Load FP Control Word from stack slot
5080 SDOperand CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5082 // Transform as necessary
5084 DAG.getNode(ISD::SRL, MVT::i16,
5085 DAG.getNode(ISD::AND, MVT::i16,
5086 CWD, DAG.getConstant(0x800, MVT::i16)),
5087 DAG.getConstant(11, MVT::i8));
5089 DAG.getNode(ISD::SRL, MVT::i16,
5090 DAG.getNode(ISD::AND, MVT::i16,
5091 CWD, DAG.getConstant(0x400, MVT::i16)),
5092 DAG.getConstant(9, MVT::i8));
5095 DAG.getNode(ISD::AND, MVT::i16,
5096 DAG.getNode(ISD::ADD, MVT::i16,
5097 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5098 DAG.getConstant(1, MVT::i16)),
5099 DAG.getConstant(3, MVT::i16));
5102 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
5103 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5106 SDOperand X86TargetLowering::LowerCTLZ(SDOperand Op, SelectionDAG &DAG) {
5107 MVT::ValueType VT = Op.getValueType();
5108 MVT::ValueType OpVT = VT;
5109 unsigned NumBits = MVT::getSizeInBits(VT);
5111 Op = Op.getOperand(0);
5112 if (VT == MVT::i8) {
5113 // Zero extend to i32 since there is not an i8 bsr.
5115 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5118 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
5119 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5120 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
5122 // If src is zero (i.e. bsr sets ZF), returns NumBits.
5123 SmallVector<SDOperand, 4> Ops;
5125 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
5126 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5127 Ops.push_back(Op.getValue(1));
5128 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5130 // Finally xor with NumBits-1.
5131 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
5134 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5138 SDOperand X86TargetLowering::LowerCTTZ(SDOperand Op, SelectionDAG &DAG) {
5139 MVT::ValueType VT = Op.getValueType();
5140 MVT::ValueType OpVT = VT;
5141 unsigned NumBits = MVT::getSizeInBits(VT);
5143 Op = Op.getOperand(0);
5144 if (VT == MVT::i8) {
5146 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5149 // Issue a bsf (scan bits forward) which also sets EFLAGS.
5150 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5151 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
5153 // If src is zero (i.e. bsf sets ZF), returns NumBits.
5154 SmallVector<SDOperand, 4> Ops;
5156 Ops.push_back(DAG.getConstant(NumBits, OpVT));
5157 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5158 Ops.push_back(Op.getValue(1));
5159 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5162 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5166 /// LowerOperation - Provide custom lowering hooks for some operations.
5168 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
5169 switch (Op.getOpcode()) {
5170 default: assert(0 && "Should not custom lower this!");
5171 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5172 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5173 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5174 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5175 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5176 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5177 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5178 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5179 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
5180 case ISD::SHL_PARTS:
5181 case ISD::SRA_PARTS:
5182 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5183 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5184 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5185 case ISD::FABS: return LowerFABS(Op, DAG);
5186 case ISD::FNEG: return LowerFNEG(Op, DAG);
5187 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5188 case ISD::SETCC: return LowerSETCC(Op, DAG);
5189 case ISD::SELECT: return LowerSELECT(Op, DAG);
5190 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
5191 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5192 case ISD::CALL: return LowerCALL(Op, DAG);
5193 case ISD::RET: return LowerRET(Op, DAG);
5194 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
5195 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
5196 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
5197 case ISD::VASTART: return LowerVASTART(Op, DAG);
5198 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
5199 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5200 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5201 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5202 case ISD::FRAME_TO_ARGS_OFFSET:
5203 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
5204 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
5205 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
5206 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
5207 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5208 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
5209 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
5211 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
5212 case ISD::READCYCLECOUNTER:
5213 return SDOperand(ExpandREADCYCLECOUNTER(Op.Val, DAG), 0);
5217 /// ExpandOperation - Provide custom lowering hooks for expanding operations.
5218 SDNode *X86TargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
5219 switch (N->getOpcode()) {
5220 default: assert(0 && "Should not custom lower this!");
5221 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
5222 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
5226 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
5228 default: return NULL;
5229 case X86ISD::BSF: return "X86ISD::BSF";
5230 case X86ISD::BSR: return "X86ISD::BSR";
5231 case X86ISD::SHLD: return "X86ISD::SHLD";
5232 case X86ISD::SHRD: return "X86ISD::SHRD";
5233 case X86ISD::FAND: return "X86ISD::FAND";
5234 case X86ISD::FOR: return "X86ISD::FOR";
5235 case X86ISD::FXOR: return "X86ISD::FXOR";
5236 case X86ISD::FSRL: return "X86ISD::FSRL";
5237 case X86ISD::FILD: return "X86ISD::FILD";
5238 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
5239 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
5240 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
5241 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
5242 case X86ISD::FLD: return "X86ISD::FLD";
5243 case X86ISD::FST: return "X86ISD::FST";
5244 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
5245 case X86ISD::FP_GET_RESULT2: return "X86ISD::FP_GET_RESULT2";
5246 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
5247 case X86ISD::CALL: return "X86ISD::CALL";
5248 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
5249 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
5250 case X86ISD::CMP: return "X86ISD::CMP";
5251 case X86ISD::COMI: return "X86ISD::COMI";
5252 case X86ISD::UCOMI: return "X86ISD::UCOMI";
5253 case X86ISD::SETCC: return "X86ISD::SETCC";
5254 case X86ISD::CMOV: return "X86ISD::CMOV";
5255 case X86ISD::BRCOND: return "X86ISD::BRCOND";
5256 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
5257 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
5258 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
5259 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
5260 case X86ISD::Wrapper: return "X86ISD::Wrapper";
5261 case X86ISD::S2VEC: return "X86ISD::S2VEC";
5262 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
5263 case X86ISD::PINSRW: return "X86ISD::PINSRW";
5264 case X86ISD::FMAX: return "X86ISD::FMAX";
5265 case X86ISD::FMIN: return "X86ISD::FMIN";
5266 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
5267 case X86ISD::FRCP: return "X86ISD::FRCP";
5268 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
5269 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
5270 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
5271 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
5272 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
5276 // isLegalAddressingMode - Return true if the addressing mode represented
5277 // by AM is legal for this target, for a load/store of the specified type.
5278 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5279 const Type *Ty) const {
5280 // X86 supports extremely general addressing modes.
5282 // X86 allows a sign-extended 32-bit immediate field as a displacement.
5283 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
5287 // We can only fold this if we don't need an extra load.
5288 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
5291 // X86-64 only supports addr of globals in small code model.
5292 if (Subtarget->is64Bit()) {
5293 if (getTargetMachine().getCodeModel() != CodeModel::Small)
5295 // If lower 4G is not available, then we must use rip-relative addressing.
5296 if (AM.BaseOffs || AM.Scale > 1)
5307 // These scales always work.
5312 // These scales are formed with basereg+scalereg. Only accept if there is
5317 default: // Other stuff never works.
5325 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
5326 if (!Ty1->isInteger() || !Ty2->isInteger())
5328 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5329 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5330 if (NumBits1 <= NumBits2)
5332 return Subtarget->is64Bit() || NumBits1 < 64;
5335 bool X86TargetLowering::isTruncateFree(MVT::ValueType VT1,
5336 MVT::ValueType VT2) const {
5337 if (!MVT::isInteger(VT1) || !MVT::isInteger(VT2))
5339 unsigned NumBits1 = MVT::getSizeInBits(VT1);
5340 unsigned NumBits2 = MVT::getSizeInBits(VT2);
5341 if (NumBits1 <= NumBits2)
5343 return Subtarget->is64Bit() || NumBits1 < 64;
5346 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5347 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5348 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5349 /// are assumed to be legal.
5351 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
5352 // Only do shuffles on 128-bit vector types for now.
5353 if (MVT::getSizeInBits(VT) == 64) return false;
5354 return (Mask.Val->getNumOperands() <= 4 ||
5355 isIdentityMask(Mask.Val) ||
5356 isIdentityMask(Mask.Val, true) ||
5357 isSplatMask(Mask.Val) ||
5358 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
5359 X86::isUNPCKLMask(Mask.Val) ||
5360 X86::isUNPCKHMask(Mask.Val) ||
5361 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
5362 X86::isUNPCKH_v_undef_Mask(Mask.Val));
5365 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
5367 SelectionDAG &DAG) const {
5368 unsigned NumElts = BVOps.size();
5369 // Only do shuffles on 128-bit vector types for now.
5370 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
5371 if (NumElts == 2) return true;
5373 return (isMOVLMask(&BVOps[0], 4) ||
5374 isCommutedMOVL(&BVOps[0], 4, true) ||
5375 isSHUFPMask(&BVOps[0], 4) ||
5376 isCommutedSHUFP(&BVOps[0], 4));
5381 //===----------------------------------------------------------------------===//
5382 // X86 Scheduler Hooks
5383 //===----------------------------------------------------------------------===//
5386 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5387 MachineBasicBlock *BB) {
5388 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5389 switch (MI->getOpcode()) {
5390 default: assert(false && "Unexpected instr type to insert");
5391 case X86::CMOV_FR32:
5392 case X86::CMOV_FR64:
5393 case X86::CMOV_V4F32:
5394 case X86::CMOV_V2F64:
5395 case X86::CMOV_V2I64: {
5396 // To "insert" a SELECT_CC instruction, we actually have to insert the
5397 // diamond control-flow pattern. The incoming instruction knows the
5398 // destination vreg to set, the condition code register to branch on, the
5399 // true/false values to select between, and a branch opcode to use.
5400 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5401 ilist<MachineBasicBlock>::iterator It = BB;
5407 // cmpTY ccX, r1, r2
5409 // fallthrough --> copy0MBB
5410 MachineBasicBlock *thisMBB = BB;
5411 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
5412 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
5414 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
5415 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
5416 MachineFunction *F = BB->getParent();
5417 F->getBasicBlockList().insert(It, copy0MBB);
5418 F->getBasicBlockList().insert(It, sinkMBB);
5419 // Update machine-CFG edges by first adding all successors of the current
5420 // block to the new block which will contain the Phi node for the select.
5421 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
5422 e = BB->succ_end(); i != e; ++i)
5423 sinkMBB->addSuccessor(*i);
5424 // Next, remove all successors of the current block, and add the true
5425 // and fallthrough blocks as its successors.
5426 while(!BB->succ_empty())
5427 BB->removeSuccessor(BB->succ_begin());
5428 BB->addSuccessor(copy0MBB);
5429 BB->addSuccessor(sinkMBB);
5432 // %FalseValue = ...
5433 // # fallthrough to sinkMBB
5436 // Update machine-CFG edges
5437 BB->addSuccessor(sinkMBB);
5440 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5443 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
5444 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5445 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5447 delete MI; // The pseudo instruction is gone now.
5451 case X86::FP32_TO_INT16_IN_MEM:
5452 case X86::FP32_TO_INT32_IN_MEM:
5453 case X86::FP32_TO_INT64_IN_MEM:
5454 case X86::FP64_TO_INT16_IN_MEM:
5455 case X86::FP64_TO_INT32_IN_MEM:
5456 case X86::FP64_TO_INT64_IN_MEM:
5457 case X86::FP80_TO_INT16_IN_MEM:
5458 case X86::FP80_TO_INT32_IN_MEM:
5459 case X86::FP80_TO_INT64_IN_MEM: {
5460 // Change the floating point control register to use "round towards zero"
5461 // mode when truncating to an integer value.
5462 MachineFunction *F = BB->getParent();
5463 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
5464 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
5466 // Load the old value of the high byte of the control word...
5468 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
5469 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
5471 // Set the high part to be round to zero...
5472 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
5475 // Reload the modified control word now...
5476 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5478 // Restore the memory image of control word to original value
5479 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
5482 // Get the X86 opcode to use.
5484 switch (MI->getOpcode()) {
5485 default: assert(0 && "illegal opcode!");
5486 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
5487 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
5488 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
5489 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
5490 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
5491 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
5492 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
5493 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
5494 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
5498 MachineOperand &Op = MI->getOperand(0);
5499 if (Op.isRegister()) {
5500 AM.BaseType = X86AddressMode::RegBase;
5501 AM.Base.Reg = Op.getReg();
5503 AM.BaseType = X86AddressMode::FrameIndexBase;
5504 AM.Base.FrameIndex = Op.getIndex();
5506 Op = MI->getOperand(1);
5507 if (Op.isImmediate())
5508 AM.Scale = Op.getImm();
5509 Op = MI->getOperand(2);
5510 if (Op.isImmediate())
5511 AM.IndexReg = Op.getImm();
5512 Op = MI->getOperand(3);
5513 if (Op.isGlobalAddress()) {
5514 AM.GV = Op.getGlobal();
5516 AM.Disp = Op.getImm();
5518 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
5519 .addReg(MI->getOperand(4).getReg());
5521 // Reload the original control word now.
5522 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5524 delete MI; // The pseudo instruction is gone now.
5530 //===----------------------------------------------------------------------===//
5531 // X86 Optimization Hooks
5532 //===----------------------------------------------------------------------===//
5534 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
5536 uint64_t &KnownZero,
5538 const SelectionDAG &DAG,
5539 unsigned Depth) const {
5540 unsigned Opc = Op.getOpcode();
5541 assert((Opc >= ISD::BUILTIN_OP_END ||
5542 Opc == ISD::INTRINSIC_WO_CHAIN ||
5543 Opc == ISD::INTRINSIC_W_CHAIN ||
5544 Opc == ISD::INTRINSIC_VOID) &&
5545 "Should use MaskedValueIsZero if you don't know whether Op"
5546 " is a target node!");
5548 KnownZero = KnownOne = 0; // Don't know anything.
5552 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
5557 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5558 /// element of the result of the vector shuffle.
5559 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
5560 MVT::ValueType VT = N->getValueType(0);
5561 SDOperand PermMask = N->getOperand(2);
5562 unsigned NumElems = PermMask.getNumOperands();
5563 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
5565 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5567 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5568 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
5569 SDOperand Idx = PermMask.getOperand(i);
5570 if (Idx.getOpcode() == ISD::UNDEF)
5571 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5572 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
5577 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
5578 /// node is a GlobalAddress + an offset.
5579 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
5580 unsigned Opc = N->getOpcode();
5581 if (Opc == X86ISD::Wrapper) {
5582 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
5583 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
5586 } else if (Opc == ISD::ADD) {
5587 SDOperand N1 = N->getOperand(0);
5588 SDOperand N2 = N->getOperand(1);
5589 if (isGAPlusOffset(N1.Val, GA, Offset)) {
5590 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
5592 Offset += V->getSignExtended();
5595 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
5596 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
5598 Offset += V->getSignExtended();
5606 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
5608 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
5609 MachineFrameInfo *MFI) {
5610 if (N->getOperand(0).Val != Base->getOperand(0).Val)
5613 SDOperand Loc = N->getOperand(1);
5614 SDOperand BaseLoc = Base->getOperand(1);
5615 if (Loc.getOpcode() == ISD::FrameIndex) {
5616 if (BaseLoc.getOpcode() != ISD::FrameIndex)
5618 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
5619 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
5620 int FS = MFI->getObjectSize(FI);
5621 int BFS = MFI->getObjectSize(BFI);
5622 if (FS != BFS || FS != Size) return false;
5623 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
5625 GlobalValue *GV1 = NULL;
5626 GlobalValue *GV2 = NULL;
5627 int64_t Offset1 = 0;
5628 int64_t Offset2 = 0;
5629 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
5630 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
5631 if (isGA1 && isGA2 && GV1 == GV2)
5632 return Offset1 == (Offset2 + Dist*Size);
5638 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
5639 const X86Subtarget *Subtarget) {
5642 if (isGAPlusOffset(Base, GV, Offset))
5643 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
5644 // DAG combine handles the stack object case.
5649 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
5650 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
5651 /// if the load addresses are consecutive, non-overlapping, and in the right
5653 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
5654 const X86Subtarget *Subtarget) {
5655 MachineFunction &MF = DAG.getMachineFunction();
5656 MachineFrameInfo *MFI = MF.getFrameInfo();
5657 MVT::ValueType VT = N->getValueType(0);
5658 MVT::ValueType EVT = MVT::getVectorElementType(VT);
5659 SDOperand PermMask = N->getOperand(2);
5660 int NumElems = (int)PermMask.getNumOperands();
5661 SDNode *Base = NULL;
5662 for (int i = 0; i < NumElems; ++i) {
5663 SDOperand Idx = PermMask.getOperand(i);
5664 if (Idx.getOpcode() == ISD::UNDEF) {
5665 if (!Base) return SDOperand();
5668 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
5669 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
5673 else if (!isConsecutiveLoad(Arg.Val, Base,
5674 i, MVT::getSizeInBits(EVT)/8,MFI))
5679 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
5680 LoadSDNode *LD = cast<LoadSDNode>(Base);
5682 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5683 LD->getSrcValueOffset(), LD->isVolatile());
5685 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5686 LD->getSrcValueOffset(), LD->isVolatile(),
5687 LD->getAlignment());
5691 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
5692 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
5693 const X86Subtarget *Subtarget) {
5694 SDOperand Cond = N->getOperand(0);
5696 // If we have SSE[12] support, try to form min/max nodes.
5697 if (Subtarget->hasSSE2() &&
5698 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
5699 if (Cond.getOpcode() == ISD::SETCC) {
5700 // Get the LHS/RHS of the select.
5701 SDOperand LHS = N->getOperand(1);
5702 SDOperand RHS = N->getOperand(2);
5703 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
5705 unsigned Opcode = 0;
5706 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
5709 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
5712 if (!UnsafeFPMath) break;
5714 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
5716 Opcode = X86ISD::FMIN;
5719 case ISD::SETOGT: // (X > Y) ? X : Y -> max
5722 if (!UnsafeFPMath) break;
5724 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
5726 Opcode = X86ISD::FMAX;
5729 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
5732 case ISD::SETOGT: // (X > Y) ? Y : X -> min
5735 if (!UnsafeFPMath) break;
5737 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
5739 Opcode = X86ISD::FMIN;
5742 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
5745 if (!UnsafeFPMath) break;
5747 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
5749 Opcode = X86ISD::FMAX;
5755 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
5763 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
5764 /// X86ISD::FXOR nodes.
5765 static SDOperand PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
5766 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
5767 // F[X]OR(0.0, x) -> x
5768 // F[X]OR(x, 0.0) -> x
5769 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
5770 if (C->getValueAPF().isPosZero())
5771 return N->getOperand(1);
5772 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
5773 if (C->getValueAPF().isPosZero())
5774 return N->getOperand(0);
5778 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
5779 static SDOperand PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
5780 // FAND(0.0, x) -> 0.0
5781 // FAND(x, 0.0) -> 0.0
5782 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
5783 if (C->getValueAPF().isPosZero())
5784 return N->getOperand(0);
5785 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
5786 if (C->getValueAPF().isPosZero())
5787 return N->getOperand(1);
5792 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
5793 DAGCombinerInfo &DCI) const {
5794 SelectionDAG &DAG = DCI.DAG;
5795 switch (N->getOpcode()) {
5797 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, Subtarget);
5798 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
5800 case X86ISD::FOR: return PerformFORCombine(N, DAG);
5801 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
5807 //===----------------------------------------------------------------------===//
5808 // X86 Inline Assembly Support
5809 //===----------------------------------------------------------------------===//
5811 /// getConstraintType - Given a constraint letter, return the type of
5812 /// constraint it is for this target.
5813 X86TargetLowering::ConstraintType
5814 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
5815 if (Constraint.size() == 1) {
5816 switch (Constraint[0]) {
5825 return C_RegisterClass;
5830 return TargetLowering::getConstraintType(Constraint);
5833 /// LowerXConstraint - try to replace an X constraint, which matches anything,
5834 /// with another that has more specific requirements based on the type of the
5835 /// corresponding operand.
5836 void X86TargetLowering::lowerXConstraint(MVT::ValueType ConstraintVT,
5837 std::string& s) const {
5838 if (MVT::isFloatingPoint(ConstraintVT)) {
5839 if (Subtarget->hasSSE2())
5841 else if (Subtarget->hasSSE1())
5846 return TargetLowering::lowerXConstraint(ConstraintVT, s);
5849 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5850 /// vector. If it is invalid, don't add anything to Ops.
5851 void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
5853 std::vector<SDOperand>&Ops,
5854 SelectionDAG &DAG) {
5855 SDOperand Result(0, 0);
5857 switch (Constraint) {
5860 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5861 if (C->getValue() <= 31) {
5862 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5868 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5869 if (C->getValue() <= 255) {
5870 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5876 // Literal immediates are always ok.
5877 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
5878 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
5882 // If we are in non-pic codegen mode, we allow the address of a global (with
5883 // an optional displacement) to be used with 'i'.
5884 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
5887 // Match either (GA) or (GA+C)
5889 Offset = GA->getOffset();
5890 } else if (Op.getOpcode() == ISD::ADD) {
5891 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5892 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5894 Offset = GA->getOffset()+C->getValue();
5896 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5897 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5899 Offset = GA->getOffset()+C->getValue();
5906 // If addressing this global requires a load (e.g. in PIC mode), we can't
5908 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
5912 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5918 // Otherwise, not valid for this mode.
5924 Ops.push_back(Result);
5927 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5930 std::vector<unsigned> X86TargetLowering::
5931 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5932 MVT::ValueType VT) const {
5933 if (Constraint.size() == 1) {
5934 // FIXME: not handling fp-stack yet!
5935 switch (Constraint[0]) { // GCC X86 Constraint Letters
5936 default: break; // Unknown constraint letter
5937 case 'A': // EAX/EDX
5938 if (VT == MVT::i32 || VT == MVT::i64)
5939 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5941 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5944 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5945 else if (VT == MVT::i16)
5946 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5947 else if (VT == MVT::i8)
5948 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5949 else if (VT == MVT::i64)
5950 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
5955 return std::vector<unsigned>();
5958 std::pair<unsigned, const TargetRegisterClass*>
5959 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5960 MVT::ValueType VT) const {
5961 // First, see if this is a constraint that directly corresponds to an LLVM
5963 if (Constraint.size() == 1) {
5964 // GCC Constraint Letters
5965 switch (Constraint[0]) {
5967 case 'r': // GENERAL_REGS
5968 case 'R': // LEGACY_REGS
5969 case 'l': // INDEX_REGS
5970 if (VT == MVT::i64 && Subtarget->is64Bit())
5971 return std::make_pair(0U, X86::GR64RegisterClass);
5973 return std::make_pair(0U, X86::GR32RegisterClass);
5974 else if (VT == MVT::i16)
5975 return std::make_pair(0U, X86::GR16RegisterClass);
5976 else if (VT == MVT::i8)
5977 return std::make_pair(0U, X86::GR8RegisterClass);
5979 case 'y': // MMX_REGS if MMX allowed.
5980 if (!Subtarget->hasMMX()) break;
5981 return std::make_pair(0U, X86::VR64RegisterClass);
5983 case 'Y': // SSE_REGS if SSE2 allowed
5984 if (!Subtarget->hasSSE2()) break;
5986 case 'x': // SSE_REGS if SSE1 allowed
5987 if (!Subtarget->hasSSE1()) break;
5991 // Scalar SSE types.
5994 return std::make_pair(0U, X86::FR32RegisterClass);
5997 return std::make_pair(0U, X86::FR64RegisterClass);
6005 return std::make_pair(0U, X86::VR128RegisterClass);
6011 // Use the default implementation in TargetLowering to convert the register
6012 // constraint into a member of a register class.
6013 std::pair<unsigned, const TargetRegisterClass*> Res;
6014 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6016 // Not found as a standard register?
6017 if (Res.second == 0) {
6018 // GCC calls "st(0)" just plain "st".
6019 if (StringsEqualNoCase("{st}", Constraint)) {
6020 Res.first = X86::ST0;
6021 Res.second = X86::RFP80RegisterClass;
6027 // Otherwise, check to see if this is a register class of the wrong value
6028 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
6029 // turn into {ax},{dx}.
6030 if (Res.second->hasType(VT))
6031 return Res; // Correct type already, nothing to do.
6033 // All of the single-register GCC register classes map their values onto
6034 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
6035 // really want an 8-bit or 32-bit register, map to the appropriate register
6036 // class and return the appropriate register.
6037 if (Res.second != X86::GR16RegisterClass)
6040 if (VT == MVT::i8) {
6041 unsigned DestReg = 0;
6042 switch (Res.first) {
6044 case X86::AX: DestReg = X86::AL; break;
6045 case X86::DX: DestReg = X86::DL; break;
6046 case X86::CX: DestReg = X86::CL; break;
6047 case X86::BX: DestReg = X86::BL; break;
6050 Res.first = DestReg;
6051 Res.second = Res.second = X86::GR8RegisterClass;
6053 } else if (VT == MVT::i32) {
6054 unsigned DestReg = 0;
6055 switch (Res.first) {
6057 case X86::AX: DestReg = X86::EAX; break;
6058 case X86::DX: DestReg = X86::EDX; break;
6059 case X86::CX: DestReg = X86::ECX; break;
6060 case X86::BX: DestReg = X86::EBX; break;
6061 case X86::SI: DestReg = X86::ESI; break;
6062 case X86::DI: DestReg = X86::EDI; break;
6063 case X86::BP: DestReg = X86::EBP; break;
6064 case X86::SP: DestReg = X86::ESP; break;
6067 Res.first = DestReg;
6068 Res.second = Res.second = X86::GR32RegisterClass;
6070 } else if (VT == MVT::i64) {
6071 unsigned DestReg = 0;
6072 switch (Res.first) {
6074 case X86::AX: DestReg = X86::RAX; break;
6075 case X86::DX: DestReg = X86::RDX; break;
6076 case X86::CX: DestReg = X86::RCX; break;
6077 case X86::BX: DestReg = X86::RBX; break;
6078 case X86::SI: DestReg = X86::RSI; break;
6079 case X86::DI: DestReg = X86::RDI; break;
6080 case X86::BP: DestReg = X86::RBP; break;
6081 case X86::SP: DestReg = X86::RSP; break;
6084 Res.first = DestReg;
6085 Res.second = Res.second = X86::GR64RegisterClass;