1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner and is distributed under
6 // the University of Illinois Open Source 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/VectorExtras.h"
27 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/SelectionDAG.h"
33 #include "llvm/CodeGen/SSARegMap.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/ADT/StringExtras.h"
39 #include "llvm/ParameterAttributes.h"
42 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
43 : TargetLowering(TM) {
44 Subtarget = &TM.getSubtarget<X86Subtarget>();
45 X86ScalarSSEf64 = Subtarget->hasSSE2();
46 X86ScalarSSEf32 = Subtarget->hasSSE1();
47 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
50 RegInfo = TM.getRegisterInfo();
52 // Set up the TargetLowering object.
54 // X86 is weird, it always uses i8 for shift amounts and setcc results.
55 setShiftAmountType(MVT::i8);
56 setSetCCResultType(MVT::i8);
57 setSetCCResultContents(ZeroOrOneSetCCResult);
58 setSchedulingPreference(SchedulingForRegPressure);
59 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
60 setStackPointerRegisterToSaveRestore(X86StackPtr);
62 if (Subtarget->isTargetDarwin()) {
63 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
64 setUseUnderscoreSetJmp(false);
65 setUseUnderscoreLongJmp(false);
66 } else if (Subtarget->isTargetMingw()) {
67 // MS runtime is weird: it exports _setjmp, but longjmp!
68 setUseUnderscoreSetJmp(true);
69 setUseUnderscoreLongJmp(false);
71 setUseUnderscoreSetJmp(true);
72 setUseUnderscoreLongJmp(true);
75 // Set up the register classes.
76 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
77 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
78 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
79 if (Subtarget->is64Bit())
80 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
82 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
84 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
86 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
87 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
88 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
90 if (Subtarget->is64Bit()) {
91 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
92 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
95 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
96 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
98 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
101 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
103 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
104 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
105 // SSE has no i16 to fp conversion, only i32
106 if (X86ScalarSSEf32) {
107 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
108 // f32 and f64 cases are Legal, f80 case is not
109 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
111 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
112 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
115 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
116 // are Legal, f80 is custom lowered.
117 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
118 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
120 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
122 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
123 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
125 if (X86ScalarSSEf32) {
126 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
127 // f32 and f64 cases are Legal, f80 case is not
128 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
130 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
131 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
134 // Handle FP_TO_UINT by promoting the destination to a larger signed
136 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
137 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
138 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
140 if (Subtarget->is64Bit()) {
141 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
142 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
144 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
145 // Expand FP_TO_UINT into a select.
146 // FIXME: We would like to use a Custom expander here eventually to do
147 // the optimal thing for SSE vs. the default expansion in the legalizer.
148 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
150 // With SSE3 we can use fisttpll to convert to a signed i64.
151 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
154 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
155 if (!X86ScalarSSEf64) {
156 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
157 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
160 // Scalar integer multiply, multiply-high, divide, and remainder are
161 // lowered to use operations that produce two results, to match the
162 // available instructions. This exposes the two-result form to trivial
163 // CSE, which is able to combine x/y and x%y into a single instruction,
164 // for example. The single-result multiply instructions are introduced
165 // in X86ISelDAGToDAG.cpp, after CSE, for uses where the the high part
167 setOperationAction(ISD::MUL , MVT::i8 , Expand);
168 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
169 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
170 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
171 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
172 setOperationAction(ISD::SREM , MVT::i8 , Expand);
173 setOperationAction(ISD::UREM , MVT::i8 , Expand);
174 setOperationAction(ISD::MUL , MVT::i16 , Expand);
175 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
176 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
177 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
178 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
179 setOperationAction(ISD::SREM , MVT::i16 , Expand);
180 setOperationAction(ISD::UREM , MVT::i16 , Expand);
181 setOperationAction(ISD::MUL , MVT::i32 , Expand);
182 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
183 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
184 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
185 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
186 setOperationAction(ISD::SREM , MVT::i32 , Expand);
187 setOperationAction(ISD::UREM , MVT::i32 , Expand);
188 setOperationAction(ISD::MUL , MVT::i64 , Expand);
189 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
190 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
191 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
192 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
193 setOperationAction(ISD::SREM , MVT::i64 , Expand);
194 setOperationAction(ISD::UREM , MVT::i64 , Expand);
196 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
197 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
198 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
199 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
200 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
201 if (Subtarget->is64Bit())
202 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
203 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
204 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
205 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
206 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
207 setOperationAction(ISD::FREM , MVT::f64 , Expand);
209 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
210 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
211 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
212 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
213 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
214 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
215 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
216 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
217 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
218 if (Subtarget->is64Bit()) {
219 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
220 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
221 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
224 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
225 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
227 // These should be promoted to a larger select which is supported.
228 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
229 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
230 // X86 wants to expand cmov itself.
231 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
232 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
233 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
234 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
235 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
236 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
237 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
238 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
239 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
240 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
241 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
242 if (Subtarget->is64Bit()) {
243 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
244 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
246 // X86 ret instruction may pop stack.
247 setOperationAction(ISD::RET , MVT::Other, Custom);
248 if (!Subtarget->is64Bit())
249 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
252 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
253 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
254 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
255 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
256 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
257 if (Subtarget->is64Bit()) {
258 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
259 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
260 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
261 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
263 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
264 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
265 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
266 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
267 // X86 wants to expand memset / memcpy itself.
268 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
269 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
271 // Use the default ISD::LOCATION expansion.
272 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
273 // FIXME - use subtarget debug flags
274 if (!Subtarget->isTargetDarwin() &&
275 !Subtarget->isTargetELF() &&
276 !Subtarget->isTargetCygMing())
277 setOperationAction(ISD::LABEL, MVT::Other, Expand);
279 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
280 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
281 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
282 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
283 if (Subtarget->is64Bit()) {
285 setExceptionPointerRegister(X86::RAX);
286 setExceptionSelectorRegister(X86::RDX);
288 setExceptionPointerRegister(X86::EAX);
289 setExceptionSelectorRegister(X86::EDX);
291 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
293 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
295 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
296 setOperationAction(ISD::VASTART , MVT::Other, Custom);
297 setOperationAction(ISD::VAARG , MVT::Other, Expand);
298 setOperationAction(ISD::VAEND , MVT::Other, Expand);
299 if (Subtarget->is64Bit())
300 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
302 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
304 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
305 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
306 if (Subtarget->is64Bit())
307 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
308 if (Subtarget->isTargetCygMing())
309 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
311 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
313 if (X86ScalarSSEf64) {
314 // f32 and f64 use SSE.
315 // Set up the FP register classes.
316 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
317 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
319 // Use ANDPD to simulate FABS.
320 setOperationAction(ISD::FABS , MVT::f64, Custom);
321 setOperationAction(ISD::FABS , MVT::f32, Custom);
323 // Use XORP to simulate FNEG.
324 setOperationAction(ISD::FNEG , MVT::f64, Custom);
325 setOperationAction(ISD::FNEG , MVT::f32, Custom);
327 // Use ANDPD and ORPD to simulate FCOPYSIGN.
328 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
329 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
331 // We don't support sin/cos/fmod
332 setOperationAction(ISD::FSIN , MVT::f64, Expand);
333 setOperationAction(ISD::FCOS , MVT::f64, Expand);
334 setOperationAction(ISD::FREM , MVT::f64, Expand);
335 setOperationAction(ISD::FSIN , MVT::f32, Expand);
336 setOperationAction(ISD::FCOS , MVT::f32, Expand);
337 setOperationAction(ISD::FREM , MVT::f32, Expand);
339 // Expand FP immediates into loads from the stack, except for the special
341 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
342 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
343 addLegalFPImmediate(APFloat(+0.0)); // xorpd
344 addLegalFPImmediate(APFloat(+0.0f)); // xorps
346 // Conversions to long double (in X87) go through memory.
347 setConvertAction(MVT::f32, MVT::f80, Expand);
348 setConvertAction(MVT::f64, MVT::f80, Expand);
350 // Conversions from long double (in X87) go through memory.
351 setConvertAction(MVT::f80, MVT::f32, Expand);
352 setConvertAction(MVT::f80, MVT::f64, Expand);
353 } else if (X86ScalarSSEf32) {
354 // Use SSE for f32, x87 for f64.
355 // Set up the FP register classes.
356 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
357 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
359 // Use ANDPS to simulate FABS.
360 setOperationAction(ISD::FABS , MVT::f32, Custom);
362 // Use XORP to simulate FNEG.
363 setOperationAction(ISD::FNEG , MVT::f32, Custom);
365 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
367 // Use ANDPS and ORPS to simulate FCOPYSIGN.
368 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
369 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
371 // We don't support sin/cos/fmod
372 setOperationAction(ISD::FSIN , MVT::f32, Expand);
373 setOperationAction(ISD::FCOS , MVT::f32, Expand);
374 setOperationAction(ISD::FREM , MVT::f32, Expand);
376 // Expand FP immediates into loads from the stack, except for the special
378 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
379 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
380 addLegalFPImmediate(APFloat(+0.0f)); // xorps
381 addLegalFPImmediate(APFloat(+0.0)); // FLD0
382 addLegalFPImmediate(APFloat(+1.0)); // FLD1
383 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
384 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
386 // SSE->x87 conversions go through memory.
387 setConvertAction(MVT::f32, MVT::f64, Expand);
388 setConvertAction(MVT::f32, MVT::f80, Expand);
390 // x87->SSE truncations need to go through memory.
391 setConvertAction(MVT::f80, MVT::f32, Expand);
392 setConvertAction(MVT::f64, MVT::f32, Expand);
393 // And x87->x87 truncations also.
394 setConvertAction(MVT::f80, MVT::f64, Expand);
397 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
398 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
401 // f32 and f64 in x87.
402 // Set up the FP register classes.
403 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
404 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
406 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
407 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
408 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
409 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
411 // Floating truncations need to go through memory.
412 setConvertAction(MVT::f80, MVT::f32, Expand);
413 setConvertAction(MVT::f64, MVT::f32, Expand);
414 setConvertAction(MVT::f80, MVT::f64, Expand);
417 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
418 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
421 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
422 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
423 addLegalFPImmediate(APFloat(+0.0)); // FLD0
424 addLegalFPImmediate(APFloat(+1.0)); // FLD1
425 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
426 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
427 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
428 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
429 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
430 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
433 // Long double always uses X87.
434 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
435 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
436 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
437 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
439 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
440 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
443 // Always use a library call for pow.
444 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
445 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
446 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
448 // First set operation action for all vector types to expand. Then we
449 // will selectively turn on ones that can be effectively codegen'd.
450 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
451 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
452 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
453 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
454 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
455 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
456 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
457 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
458 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
459 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
460 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
461 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
462 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
463 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
464 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
465 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
466 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
467 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
468 setOperationAction(ISD::FABS, (MVT::ValueType)VT, Expand);
469 setOperationAction(ISD::FSIN, (MVT::ValueType)VT, Expand);
470 setOperationAction(ISD::FCOS, (MVT::ValueType)VT, Expand);
471 setOperationAction(ISD::FREM, (MVT::ValueType)VT, Expand);
472 setOperationAction(ISD::FPOWI, (MVT::ValueType)VT, Expand);
473 setOperationAction(ISD::FSQRT, (MVT::ValueType)VT, Expand);
474 setOperationAction(ISD::FCOPYSIGN, (MVT::ValueType)VT, Expand);
475 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
476 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
477 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
478 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
479 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
480 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
481 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
482 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
485 if (Subtarget->hasMMX()) {
486 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
487 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
488 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
489 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
491 // FIXME: add MMX packed arithmetics
493 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
494 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
495 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
496 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
498 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
499 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
500 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
501 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
503 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
504 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
506 setOperationAction(ISD::AND, MVT::v8i8, Promote);
507 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
508 setOperationAction(ISD::AND, MVT::v4i16, Promote);
509 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
510 setOperationAction(ISD::AND, MVT::v2i32, Promote);
511 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
512 setOperationAction(ISD::AND, MVT::v1i64, Legal);
514 setOperationAction(ISD::OR, MVT::v8i8, Promote);
515 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
516 setOperationAction(ISD::OR, MVT::v4i16, Promote);
517 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
518 setOperationAction(ISD::OR, MVT::v2i32, Promote);
519 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
520 setOperationAction(ISD::OR, MVT::v1i64, Legal);
522 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
523 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
524 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
525 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
526 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
527 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
528 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
530 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
531 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
532 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
533 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
534 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
535 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
536 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
538 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
539 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
540 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
541 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
543 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
544 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
545 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
546 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
548 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
549 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
550 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Custom);
551 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
554 if (Subtarget->hasSSE1()) {
555 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
557 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
558 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
559 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
560 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
561 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
562 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
563 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
564 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
565 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
566 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
567 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
570 if (Subtarget->hasSSE2()) {
571 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
572 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
573 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
574 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
575 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
577 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
578 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
579 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
580 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
581 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
582 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
583 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
584 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
585 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
586 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
587 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
588 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
589 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
590 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
591 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
593 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
594 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
595 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
596 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
597 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
598 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
600 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
601 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
602 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
603 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
604 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
606 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
607 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
608 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
609 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
610 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
611 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
613 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
614 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
615 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
616 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
617 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
618 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
619 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
620 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
621 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
622 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
623 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
624 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
627 // Custom lower v2i64 and v2f64 selects.
628 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
629 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
630 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
631 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
634 // We want to custom lower some of our intrinsics.
635 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
637 // We have target-specific dag combine patterns for the following nodes:
638 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
639 setTargetDAGCombine(ISD::SELECT);
641 computeRegisterProperties();
643 // FIXME: These should be based on subtarget info. Plus, the values should
644 // be smaller when we are in optimizing for size mode.
645 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
646 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
647 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
648 allowUnalignedMemoryAccesses = true; // x86 supports it!
652 //===----------------------------------------------------------------------===//
653 // Return Value Calling Convention Implementation
654 //===----------------------------------------------------------------------===//
656 #include "X86GenCallingConv.inc"
658 /// GetPossiblePreceedingTailCall - Get preceeding X86ISD::TAILCALL node if it
659 /// exists skip possible ISD:TokenFactor.
660 static SDOperand GetPossiblePreceedingTailCall(SDOperand Chain) {
661 if (Chain.getOpcode()==X86ISD::TAILCALL) {
663 } else if (Chain.getOpcode()==ISD::TokenFactor) {
664 if (Chain.getNumOperands() &&
665 Chain.getOperand(0).getOpcode()==X86ISD::TAILCALL)
666 return Chain.getOperand(0);
671 /// LowerRET - Lower an ISD::RET node.
672 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
673 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
675 SmallVector<CCValAssign, 16> RVLocs;
676 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
677 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
678 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
679 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
681 // If this is the first return lowered for this function, add the regs to the
682 // liveout set for the function.
683 if (DAG.getMachineFunction().liveout_empty()) {
684 for (unsigned i = 0; i != RVLocs.size(); ++i)
685 if (RVLocs[i].isRegLoc())
686 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
688 SDOperand Chain = Op.getOperand(0);
690 // Handle tail call return.
691 Chain = GetPossiblePreceedingTailCall(Chain);
692 if (Chain.getOpcode() == X86ISD::TAILCALL) {
693 SDOperand TailCall = Chain;
694 SDOperand TargetAddress = TailCall.getOperand(1);
695 SDOperand StackAdjustment = TailCall.getOperand(2);
696 assert ( ((TargetAddress.getOpcode() == ISD::Register &&
697 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
698 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
699 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
700 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
701 "Expecting an global address, external symbol, or register");
702 assert( StackAdjustment.getOpcode() == ISD::Constant &&
703 "Expecting a const value");
705 SmallVector<SDOperand,8> Operands;
706 Operands.push_back(Chain.getOperand(0));
707 Operands.push_back(TargetAddress);
708 Operands.push_back(StackAdjustment);
709 // Copy registers used by the call. Last operand is a flag so it is not
711 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
712 Operands.push_back(Chain.getOperand(i));
714 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
721 // Copy the result values into the output registers.
722 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
723 RVLocs[0].getLocReg() != X86::ST0) {
724 for (unsigned i = 0; i != RVLocs.size(); ++i) {
725 CCValAssign &VA = RVLocs[i];
726 assert(VA.isRegLoc() && "Can only return in registers!");
727 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
729 Flag = Chain.getValue(1);
732 // We need to handle a destination of ST0 specially, because it isn't really
734 SDOperand Value = Op.getOperand(1);
736 // If this is an FP return with ScalarSSE, we need to move the value from
737 // an XMM register onto the fp-stack.
738 if ((X86ScalarSSEf32 && RVLocs[0].getValVT()==MVT::f32) ||
739 (X86ScalarSSEf64 && RVLocs[0].getValVT()==MVT::f64)) {
742 // If this is a load into a scalarsse value, don't store the loaded value
743 // back to the stack, only to reload it: just replace the scalar-sse load.
744 if (ISD::isNON_EXTLoad(Value.Val) &&
745 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
746 Chain = Value.getOperand(0);
747 MemLoc = Value.getOperand(1);
749 // Spill the value to memory and reload it into top of stack.
750 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
751 MachineFunction &MF = DAG.getMachineFunction();
752 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
753 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
754 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
756 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other);
757 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
758 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
759 Chain = Value.getValue(1);
762 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
763 SDOperand Ops[] = { Chain, Value };
764 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
765 Flag = Chain.getValue(1);
768 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
770 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
772 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
776 /// LowerCallResult - Lower the result values of an ISD::CALL into the
777 /// appropriate copies out of appropriate physical registers. This assumes that
778 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
779 /// being lowered. The returns a SDNode with the same number of values as the
781 SDNode *X86TargetLowering::
782 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
783 unsigned CallingConv, SelectionDAG &DAG) {
785 // Assign locations to each value returned by this call.
786 SmallVector<CCValAssign, 16> RVLocs;
787 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
788 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
789 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
792 SmallVector<SDOperand, 8> ResultVals;
794 // Copy all of the result registers out of their specified physreg.
795 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
796 for (unsigned i = 0; i != RVLocs.size(); ++i) {
797 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
798 RVLocs[i].getValVT(), InFlag).getValue(1);
799 InFlag = Chain.getValue(2);
800 ResultVals.push_back(Chain.getValue(0));
803 // Copies from the FP stack are special, as ST0 isn't a valid register
804 // before the fp stackifier runs.
806 // Copy ST0 into an RFP register with FP_GET_RESULT.
807 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other, MVT::Flag);
808 SDOperand GROps[] = { Chain, InFlag };
809 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
810 Chain = RetVal.getValue(1);
811 InFlag = RetVal.getValue(2);
813 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
815 if ((X86ScalarSSEf32 && RVLocs[0].getValVT() == MVT::f32) ||
816 (X86ScalarSSEf64 && RVLocs[0].getValVT() == MVT::f64)) {
817 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
818 // shouldn't be necessary except that RFP cannot be live across
819 // multiple blocks. When stackifier is fixed, they can be uncoupled.
820 MachineFunction &MF = DAG.getMachineFunction();
821 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
822 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
824 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag
826 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
827 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
828 Chain = RetVal.getValue(1);
830 ResultVals.push_back(RetVal);
833 // Merge everything together with a MERGE_VALUES node.
834 ResultVals.push_back(Chain);
835 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
836 &ResultVals[0], ResultVals.size()).Val;
840 //===----------------------------------------------------------------------===//
841 // C & StdCall & Fast Calling Convention implementation
842 //===----------------------------------------------------------------------===//
843 // StdCall calling convention seems to be standard for many Windows' API
844 // routines and around. It differs from C calling convention just a little:
845 // callee should clean up the stack, not caller. Symbols should be also
846 // decorated in some fancy way :) It doesn't support any vector arguments.
847 // For info on fast calling convention see Fast Calling Convention (tail call)
848 // implementation LowerX86_32FastCCCallTo.
850 /// AddLiveIn - This helper function adds the specified physical register to the
851 /// MachineFunction as a live in value. It also creates a corresponding virtual
853 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
854 const TargetRegisterClass *RC) {
855 assert(RC->contains(PReg) && "Not the correct regclass!");
856 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
857 MF.addLiveIn(PReg, VReg);
861 // align stack arguments according to platform alignment needed for tail calls
862 unsigned GetAlignedArgumentStackSize(unsigned StackSize, SelectionDAG& DAG);
864 SDOperand X86TargetLowering::LowerMemArgument(SDOperand Op, SelectionDAG &DAG,
865 const CCValAssign &VA,
866 MachineFrameInfo *MFI,
867 SDOperand Root, unsigned i) {
868 // Create the nodes corresponding to a load from this parameter slot.
869 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
870 VA.getLocMemOffset());
871 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
873 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3 + i))->getValue();
875 if (Flags & ISD::ParamFlags::ByVal)
878 return DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0);
881 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
883 unsigned NumArgs = Op.Val->getNumValues() - 1;
884 MachineFunction &MF = DAG.getMachineFunction();
885 MachineFrameInfo *MFI = MF.getFrameInfo();
886 SDOperand Root = Op.getOperand(0);
887 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
888 unsigned CC = MF.getFunction()->getCallingConv();
889 // Assign locations to all of the incoming arguments.
890 SmallVector<CCValAssign, 16> ArgLocs;
891 CCState CCInfo(CC, isVarArg,
892 getTargetMachine(), ArgLocs);
893 // Check for possible tail call calling convention.
894 if (CC == CallingConv::Fast && PerformTailCallOpt)
895 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_TailCall);
897 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
899 SmallVector<SDOperand, 8> ArgValues;
900 unsigned LastVal = ~0U;
901 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
902 CCValAssign &VA = ArgLocs[i];
903 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
905 assert(VA.getValNo() != LastVal &&
906 "Don't support value assigned to multiple locs yet");
907 LastVal = VA.getValNo();
910 MVT::ValueType RegVT = VA.getLocVT();
911 TargetRegisterClass *RC;
912 if (RegVT == MVT::i32)
913 RC = X86::GR32RegisterClass;
915 assert(MVT::isVector(RegVT));
916 RC = X86::VR128RegisterClass;
919 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
920 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
922 // If this is an 8 or 16-bit value, it is really passed promoted to 32
923 // bits. Insert an assert[sz]ext to capture this, then truncate to the
925 if (VA.getLocInfo() == CCValAssign::SExt)
926 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
927 DAG.getValueType(VA.getValVT()));
928 else if (VA.getLocInfo() == CCValAssign::ZExt)
929 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
930 DAG.getValueType(VA.getValVT()));
932 if (VA.getLocInfo() != CCValAssign::Full)
933 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
935 ArgValues.push_back(ArgValue);
937 assert(VA.isMemLoc());
938 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
942 unsigned StackSize = CCInfo.getNextStackOffset();
943 // align stack specially for tail calls
944 if (CC==CallingConv::Fast)
945 StackSize = GetAlignedArgumentStackSize(StackSize,DAG);
947 ArgValues.push_back(Root);
949 // If the function takes variable number of arguments, make a frame index for
950 // the start of the first vararg value... for expansion of llvm.va_start.
952 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
954 // Tail call calling convention (CallingConv::Fast) does not support varargs.
955 assert( !(isVarArg && CC == CallingConv::Fast) &&
956 "CallingConv::Fast does not support varargs.");
958 if (isStdCall && !isVarArg &&
959 (CC==CallingConv::Fast && PerformTailCallOpt || CC!=CallingConv::Fast)) {
960 BytesToPopOnReturn = StackSize; // Callee pops everything..
961 BytesCallerReserves = 0;
963 BytesToPopOnReturn = 0; // Callee pops nothing.
965 // If this is an sret function, the return should pop the hidden pointer.
967 (cast<ConstantSDNode>(Op.getOperand(3))->getValue() &
968 ISD::ParamFlags::StructReturn))
969 BytesToPopOnReturn = 4;
971 BytesCallerReserves = StackSize;
974 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
976 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
977 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
979 // Return the new list of results.
980 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
981 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
984 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
986 SDOperand Chain = Op.getOperand(0);
987 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
988 SDOperand Callee = Op.getOperand(4);
989 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
991 // Analyze operands of the call, assigning locations to each operand.
992 SmallVector<CCValAssign, 16> ArgLocs;
993 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
994 if(CC==CallingConv::Fast && PerformTailCallOpt)
995 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_TailCall);
997 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
999 // Get a count of how many bytes are to be pushed on the stack.
1000 unsigned NumBytes = CCInfo.getNextStackOffset();
1001 if (CC==CallingConv::Fast)
1002 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1004 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1006 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1007 SmallVector<SDOperand, 8> MemOpChains;
1011 // Walk the register/memloc assignments, inserting copies/loads.
1012 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1013 CCValAssign &VA = ArgLocs[i];
1014 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1016 // Promote the value if needed.
1017 switch (VA.getLocInfo()) {
1018 default: assert(0 && "Unknown loc info!");
1019 case CCValAssign::Full: break;
1020 case CCValAssign::SExt:
1021 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1023 case CCValAssign::ZExt:
1024 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1026 case CCValAssign::AExt:
1027 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1031 if (VA.isRegLoc()) {
1032 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1034 assert(VA.isMemLoc());
1035 if (StackPtr.Val == 0)
1036 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1038 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1043 // If the first argument is an sret pointer, remember it.
1044 bool isSRet = NumOps &&
1045 (cast<ConstantSDNode>(Op.getOperand(6))->getValue() &
1046 ISD::ParamFlags::StructReturn);
1048 if (!MemOpChains.empty())
1049 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1050 &MemOpChains[0], MemOpChains.size());
1052 // Build a sequence of copy-to-reg nodes chained together with token chain
1053 // and flag operands which copy the outgoing args into registers.
1055 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1056 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1058 InFlag = Chain.getValue(1);
1061 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1063 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1064 Subtarget->isPICStyleGOT()) {
1065 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1066 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1068 InFlag = Chain.getValue(1);
1071 // If the callee is a GlobalAddress node (quite common, every direct call is)
1072 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1073 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1074 // We should use extra load for direct calls to dllimported functions in
1076 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1077 getTargetMachine(), true))
1078 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1079 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1080 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1082 // Returns a chain & a flag for retval copy to use.
1083 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1084 SmallVector<SDOperand, 8> Ops;
1085 Ops.push_back(Chain);
1086 Ops.push_back(Callee);
1088 // Add argument registers to the end of the list so that they are known live
1090 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1091 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1092 RegsToPass[i].second.getValueType()));
1094 // Add an implicit use GOT pointer in EBX.
1095 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1096 Subtarget->isPICStyleGOT())
1097 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1100 Ops.push_back(InFlag);
1102 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1103 InFlag = Chain.getValue(1);
1105 // Create the CALLSEQ_END node.
1106 unsigned NumBytesForCalleeToPush = 0;
1108 if (CC == CallingConv::X86_StdCall ||
1109 (CC == CallingConv::Fast && PerformTailCallOpt)) {
1111 NumBytesForCalleeToPush = isSRet ? 4 : 0;
1113 NumBytesForCalleeToPush = NumBytes;
1114 assert(!(isVarArg && CC==CallingConv::Fast) &&
1115 "CallingConv::Fast does not support varargs.");
1117 // If this is is a call to a struct-return function, the callee
1118 // pops the hidden struct pointer, so we have to push it back.
1119 // This is common for Darwin/X86, Linux & Mingw32 targets.
1120 NumBytesForCalleeToPush = isSRet ? 4 : 0;
1123 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1125 Ops.push_back(Chain);
1126 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1127 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
1128 Ops.push_back(InFlag);
1129 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1130 InFlag = Chain.getValue(1);
1132 // Handle result values, copying them out of physregs into vregs that we
1134 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1138 //===----------------------------------------------------------------------===//
1139 // FastCall Calling Convention implementation
1140 //===----------------------------------------------------------------------===//
1142 // The X86 'fastcall' calling convention passes up to two integer arguments in
1143 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
1144 // and requires that the callee pop its arguments off the stack (allowing proper
1145 // tail calls), and has the same return value conventions as C calling convs.
1147 // This calling convention always arranges for the callee pop value to be 8n+4
1148 // bytes, which is needed for tail recursion elimination and stack alignment
1151 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
1152 MachineFunction &MF = DAG.getMachineFunction();
1153 MachineFrameInfo *MFI = MF.getFrameInfo();
1154 SDOperand Root = Op.getOperand(0);
1155 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1157 // Assign locations to all of the incoming arguments.
1158 SmallVector<CCValAssign, 16> ArgLocs;
1159 CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
1160 getTargetMachine(), ArgLocs);
1161 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
1163 SmallVector<SDOperand, 8> ArgValues;
1164 unsigned LastVal = ~0U;
1165 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1166 CCValAssign &VA = ArgLocs[i];
1167 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1169 assert(VA.getValNo() != LastVal &&
1170 "Don't support value assigned to multiple locs yet");
1171 LastVal = VA.getValNo();
1173 if (VA.isRegLoc()) {
1174 MVT::ValueType RegVT = VA.getLocVT();
1175 TargetRegisterClass *RC;
1176 if (RegVT == MVT::i32)
1177 RC = X86::GR32RegisterClass;
1179 assert(MVT::isVector(RegVT));
1180 RC = X86::VR128RegisterClass;
1183 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1184 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1186 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1187 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1189 if (VA.getLocInfo() == CCValAssign::SExt)
1190 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1191 DAG.getValueType(VA.getValVT()));
1192 else if (VA.getLocInfo() == CCValAssign::ZExt)
1193 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1194 DAG.getValueType(VA.getValVT()));
1196 if (VA.getLocInfo() != CCValAssign::Full)
1197 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1199 ArgValues.push_back(ArgValue);
1201 assert(VA.isMemLoc());
1202 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1206 ArgValues.push_back(Root);
1208 unsigned StackSize = CCInfo.getNextStackOffset();
1210 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1211 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1212 // arguments and the arguments after the retaddr has been pushed are
1214 if ((StackSize & 7) == 0)
1218 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1219 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1220 BytesToPopOnReturn = StackSize; // Callee pops all stack arguments.
1221 BytesCallerReserves = 0;
1223 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1224 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1226 // Return the new list of results.
1227 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1228 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1232 X86TargetLowering::LowerMemOpCallTo(SDOperand Op, SelectionDAG &DAG,
1233 const SDOperand &StackPtr,
1234 const CCValAssign &VA,
1237 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1238 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1239 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1240 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1241 if (Flags & ISD::ParamFlags::ByVal) {
1242 unsigned Align = 1 << ((Flags & ISD::ParamFlags::ByValAlign) >>
1243 ISD::ParamFlags::ByValAlignOffs);
1245 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1246 ISD::ParamFlags::ByValSizeOffs;
1248 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1249 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1250 SDOperand AlwaysInline = DAG.getConstant(1, MVT::i1);
1252 return DAG.getMemcpy(Chain, PtrOff, Arg, SizeNode, AlignNode,
1255 return DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1259 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1261 SDOperand Chain = Op.getOperand(0);
1262 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1263 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1264 SDOperand Callee = Op.getOperand(4);
1266 // Analyze operands of the call, assigning locations to each operand.
1267 SmallVector<CCValAssign, 16> ArgLocs;
1268 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1269 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall);
1271 // Get a count of how many bytes are to be pushed on the stack.
1272 unsigned NumBytes = CCInfo.getNextStackOffset();
1274 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1275 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1276 // arguments and the arguments after the retaddr has been pushed are
1278 if ((NumBytes & 7) == 0)
1282 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1284 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1285 SmallVector<SDOperand, 8> MemOpChains;
1289 // Walk the register/memloc assignments, inserting copies/loads.
1290 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1291 CCValAssign &VA = ArgLocs[i];
1292 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1294 // Promote the value if needed.
1295 switch (VA.getLocInfo()) {
1296 default: assert(0 && "Unknown loc info!");
1297 case CCValAssign::Full: break;
1298 case CCValAssign::SExt:
1299 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1301 case CCValAssign::ZExt:
1302 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1304 case CCValAssign::AExt:
1305 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1309 if (VA.isRegLoc()) {
1310 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1312 assert(VA.isMemLoc());
1313 if (StackPtr.Val == 0)
1314 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1316 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1321 if (!MemOpChains.empty())
1322 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1323 &MemOpChains[0], MemOpChains.size());
1325 // Build a sequence of copy-to-reg nodes chained together with token chain
1326 // and flag operands which copy the outgoing args into registers.
1328 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1329 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1331 InFlag = Chain.getValue(1);
1334 // If the callee is a GlobalAddress node (quite common, every direct call is)
1335 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1336 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1337 // We should use extra load for direct calls to dllimported functions in
1339 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1340 getTargetMachine(), true))
1341 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1342 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1343 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1345 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1347 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1348 Subtarget->isPICStyleGOT()) {
1349 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1350 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1352 InFlag = Chain.getValue(1);
1355 // Returns a chain & a flag for retval copy to use.
1356 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1357 SmallVector<SDOperand, 8> Ops;
1358 Ops.push_back(Chain);
1359 Ops.push_back(Callee);
1361 // Add argument registers to the end of the list so that they are known live
1363 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1364 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1365 RegsToPass[i].second.getValueType()));
1367 // Add an implicit use GOT pointer in EBX.
1368 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1369 Subtarget->isPICStyleGOT())
1370 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1373 Ops.push_back(InFlag);
1375 assert(isTailCall==false && "no tail call here");
1376 Chain = DAG.getNode(X86ISD::CALL,
1377 NodeTys, &Ops[0], Ops.size());
1378 InFlag = Chain.getValue(1);
1380 // Returns a flag for retval copy to use.
1381 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1383 Ops.push_back(Chain);
1384 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1385 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1386 Ops.push_back(InFlag);
1387 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1388 InFlag = Chain.getValue(1);
1390 // Handle result values, copying them out of physregs into vregs that we
1392 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1395 //===----------------------------------------------------------------------===//
1396 // Fast Calling Convention (tail call) implementation
1397 //===----------------------------------------------------------------------===//
1399 // Like std call, callee cleans arguments, convention except that ECX is
1400 // reserved for storing the tail called function address. Only 2 registers are
1401 // free for argument passing (inreg). Tail call optimization is performed
1403 // * tailcallopt is enabled
1404 // * caller/callee are fastcc
1405 // * elf/pic is disabled OR
1406 // * elf/pic enabled + callee is in module + callee has
1407 // visibility protected or hidden
1408 // To keep the stack aligned according to platform abi the function
1409 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1410 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1411 // If a tail called function callee has more arguments than the caller the
1412 // caller needs to make sure that there is room to move the RETADDR to. This is
1413 // achieved by reserving an area the size of the argument delta right after the
1414 // original REtADDR, but before the saved framepointer or the spilled registers
1415 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1427 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1428 /// for a 16 byte align requirement.
1429 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1430 SelectionDAG& DAG) {
1431 if (PerformTailCallOpt) {
1432 MachineFunction &MF = DAG.getMachineFunction();
1433 const TargetMachine &TM = MF.getTarget();
1434 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1435 unsigned StackAlignment = TFI.getStackAlignment();
1436 uint64_t AlignMask = StackAlignment - 1;
1437 int64_t Offset = StackSize;
1438 unsigned SlotSize = Subtarget->is64Bit() ? 8 : 4;
1439 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1440 // Number smaller than 12 so just add the difference.
1441 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1443 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1444 Offset = ((~AlignMask) & Offset) + StackAlignment +
1445 (StackAlignment-SlotSize);
1452 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1453 // following the call is a return. A function is eligible if caller/callee
1454 // calling conventions match, currently only fastcc supports tail calls, and the
1455 // function CALL is immediatly followed by a RET.
1456 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDOperand Call,
1458 SelectionDAG& DAG) const {
1459 bool IsEligible = false;
1461 // Check whether CALL node immediatly preceeds the RET node and whether the
1462 // return uses the result of the node or is a void return.
1463 if ((Ret.getNumOperands() == 1 &&
1464 (Ret.getOperand(0)== SDOperand(Call.Val,1) ||
1465 Ret.getOperand(0)== SDOperand(Call.Val,0))) ||
1466 (Ret.getOperand(0)== SDOperand(Call.Val,Call.Val->getNumValues()-1) &&
1467 Ret.getOperand(1)== SDOperand(Call.Val,0))) {
1468 MachineFunction &MF = DAG.getMachineFunction();
1469 unsigned CallerCC = MF.getFunction()->getCallingConv();
1470 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1471 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1472 SDOperand Callee = Call.getOperand(4);
1473 // On elf/pic %ebx needs to be livein.
1474 if(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1475 Subtarget->isPICStyleGOT()) {
1476 // Can only do local tail calls with PIC.
1477 GlobalValue * GV = 0;
1478 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1480 (GV = G->getGlobal()) &&
1481 (GV->hasHiddenVisibility() || GV->hasProtectedVisibility()))
1491 SDOperand X86TargetLowering::LowerX86_TailCallTo(SDOperand Op,
1494 SDOperand Chain = Op.getOperand(0);
1495 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1496 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1497 SDOperand Callee = Op.getOperand(4);
1498 bool is64Bit = Subtarget->is64Bit();
1500 assert(isTailCall && PerformTailCallOpt && "Should only emit tail calls.");
1502 // Analyze operands of the call, assigning locations to each operand.
1503 SmallVector<CCValAssign, 16> ArgLocs;
1504 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1506 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_TailCall);
1508 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_TailCall);
1511 // Lower arguments at fp - stackoffset + fpdiff.
1512 MachineFunction &MF = DAG.getMachineFunction();
1514 unsigned NumBytesToBePushed =
1515 GetAlignedArgumentStackSize(CCInfo.getNextStackOffset(), DAG);
1517 unsigned NumBytesCallerPushed =
1518 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1519 int FPDiff = NumBytesCallerPushed - NumBytesToBePushed;
1521 // Set the delta of movement of the returnaddr stackslot.
1522 // But only set if delta is greater than previous delta.
1523 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1524 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1527 getCALLSEQ_START(Chain, DAG.getConstant(NumBytesToBePushed, getPointerTy()));
1529 // Adjust the Return address stack slot.
1530 SDOperand RetAddrFrIdx, NewRetAddrFrIdx;
1532 MVT::ValueType VT = is64Bit ? MVT::i64 : MVT::i32;
1533 RetAddrFrIdx = getReturnAddressFrameIndex(DAG);
1534 // Load the "old" Return address.
1536 DAG.getLoad(VT, Chain,RetAddrFrIdx, NULL, 0);
1537 // Calculate the new stack slot for the return address.
1538 int SlotSize = is64Bit ? 8 : 4;
1539 int NewReturnAddrFI =
1540 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1541 NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1542 Chain = SDOperand(RetAddrFrIdx.Val, 1);
1545 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1546 SmallVector<SDOperand, 8> MemOpChains;
1547 SmallVector<SDOperand, 8> MemOpChains2;
1548 SDOperand FramePtr, StackPtr;
1553 // Walk the register/memloc assignments, inserting copies/loads. Lower
1554 // arguments first to the stack slot where they would normally - in case of a
1555 // normal function call - be.
1556 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1557 CCValAssign &VA = ArgLocs[i];
1558 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1560 // Promote the value if needed.
1561 switch (VA.getLocInfo()) {
1562 default: assert(0 && "Unknown loc info!");
1563 case CCValAssign::Full: break;
1564 case CCValAssign::SExt:
1565 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1567 case CCValAssign::ZExt:
1568 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1570 case CCValAssign::AExt:
1571 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1575 if (VA.isRegLoc()) {
1576 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1578 assert(VA.isMemLoc());
1579 if (StackPtr.Val == 0)
1580 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1582 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1587 if (!MemOpChains.empty())
1588 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1589 &MemOpChains[0], MemOpChains.size());
1591 // Build a sequence of copy-to-reg nodes chained together with token chain
1592 // and flag operands which copy the outgoing args into registers.
1594 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1595 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1597 InFlag = Chain.getValue(1);
1599 InFlag = SDOperand();
1601 // Copy from stack slots to stack slot of a tail called function. This needs
1602 // to be done because if we would lower the arguments directly to their real
1603 // stack slot we might end up overwriting each other.
1604 // TODO: To make this more efficient (sometimes saving a store/load) we could
1605 // analyse the arguments and emit this store/load/store sequence only for
1606 // arguments which would be overwritten otherwise.
1607 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1608 CCValAssign &VA = ArgLocs[i];
1609 if (!VA.isRegLoc()) {
1610 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1611 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1613 // Get source stack slot.
1614 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1615 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1616 // Create frame index.
1617 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1618 uint32_t OpSize = (MVT::getSizeInBits(VA.getLocVT())+7)/8;
1619 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1620 FIN = DAG.getFrameIndex(FI, MVT::i32);
1621 if (Flags & ISD::ParamFlags::ByVal) {
1622 // Copy relative to framepointer.
1623 unsigned Align = 1 << ((Flags & ISD::ParamFlags::ByValAlign) >>
1624 ISD::ParamFlags::ByValAlignOffs);
1626 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1627 ISD::ParamFlags::ByValSizeOffs;
1629 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1630 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1631 // Copy relative to framepointer.
1632 MemOpChains2.push_back(DAG.getNode(ISD::MEMCPY, MVT::Other, Chain, FIN,
1633 PtrOff, SizeNode, AlignNode));
1635 SDOperand LoadedArg = DAG.getLoad(VA.getValVT(), Chain, PtrOff, NULL,0);
1636 // Store relative to framepointer.
1637 MemOpChains2.push_back(DAG.getStore(Chain, LoadedArg, FIN, NULL, 0));
1642 if (!MemOpChains2.empty())
1643 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1644 &MemOpChains2[0], MemOpChains.size());
1646 // Store the return address to the appropriate stack slot.
1648 Chain = DAG.getStore(Chain,RetAddrFrIdx, NewRetAddrFrIdx, NULL, 0);
1650 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1652 // Does not work with tail call since ebx is not restored correctly by
1653 // tailcaller. TODO: at least for x86 - verify for x86-64
1655 // If the callee is a GlobalAddress node (quite common, every direct call is)
1656 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1657 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1658 // We should use extra load for direct calls to dllimported functions in
1660 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1661 getTargetMachine(), true))
1662 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1663 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1664 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1666 assert(Callee.getOpcode() == ISD::LOAD &&
1667 "Function destination must be loaded into virtual register");
1668 unsigned Opc = is64Bit ? X86::R9 : X86::ECX;
1670 Chain = DAG.getCopyToReg(Chain,
1671 DAG.getRegister(Opc, getPointerTy()) ,
1673 Callee = DAG.getRegister(Opc, getPointerTy());
1674 // Add register as live out.
1675 DAG.getMachineFunction().addLiveOut(Opc);
1678 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1679 SmallVector<SDOperand, 8> Ops;
1681 Ops.push_back(Chain);
1682 Ops.push_back(DAG.getConstant(NumBytesToBePushed, getPointerTy()));
1683 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1685 Ops.push_back(InFlag);
1686 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1687 InFlag = Chain.getValue(1);
1689 // Returns a chain & a flag for retval copy to use.
1690 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1692 Ops.push_back(Chain);
1693 Ops.push_back(Callee);
1694 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1695 // Add argument registers to the end of the list so that they are known live
1697 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1698 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1699 RegsToPass[i].second.getValueType()));
1701 Ops.push_back(InFlag);
1702 assert(InFlag.Val &&
1703 "Flag must be set. Depend on flag being set in LowerRET");
1704 Chain = DAG.getNode(X86ISD::TAILCALL,
1705 Op.Val->getVTList(), &Ops[0], Ops.size());
1707 return SDOperand(Chain.Val, Op.ResNo);
1710 //===----------------------------------------------------------------------===//
1711 // X86-64 C Calling Convention implementation
1712 //===----------------------------------------------------------------------===//
1715 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1716 MachineFunction &MF = DAG.getMachineFunction();
1717 MachineFrameInfo *MFI = MF.getFrameInfo();
1718 SDOperand Root = Op.getOperand(0);
1719 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1720 unsigned CC= MF.getFunction()->getCallingConv();
1722 static const unsigned GPR64ArgRegs[] = {
1723 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1725 static const unsigned XMMArgRegs[] = {
1726 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1727 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1731 // Assign locations to all of the incoming arguments.
1732 SmallVector<CCValAssign, 16> ArgLocs;
1733 CCState CCInfo(CC, isVarArg,
1734 getTargetMachine(), ArgLocs);
1735 if (CC == CallingConv::Fast && PerformTailCallOpt)
1736 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_TailCall);
1738 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1740 SmallVector<SDOperand, 8> ArgValues;
1741 unsigned LastVal = ~0U;
1742 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1743 CCValAssign &VA = ArgLocs[i];
1744 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1746 assert(VA.getValNo() != LastVal &&
1747 "Don't support value assigned to multiple locs yet");
1748 LastVal = VA.getValNo();
1750 if (VA.isRegLoc()) {
1751 MVT::ValueType RegVT = VA.getLocVT();
1752 TargetRegisterClass *RC;
1753 if (RegVT == MVT::i32)
1754 RC = X86::GR32RegisterClass;
1755 else if (RegVT == MVT::i64)
1756 RC = X86::GR64RegisterClass;
1757 else if (RegVT == MVT::f32)
1758 RC = X86::FR32RegisterClass;
1759 else if (RegVT == MVT::f64)
1760 RC = X86::FR64RegisterClass;
1762 assert(MVT::isVector(RegVT));
1763 if (MVT::getSizeInBits(RegVT) == 64) {
1764 RC = X86::GR64RegisterClass; // MMX values are passed in GPRs.
1767 RC = X86::VR128RegisterClass;
1770 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1771 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1773 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1774 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1776 if (VA.getLocInfo() == CCValAssign::SExt)
1777 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1778 DAG.getValueType(VA.getValVT()));
1779 else if (VA.getLocInfo() == CCValAssign::ZExt)
1780 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1781 DAG.getValueType(VA.getValVT()));
1783 if (VA.getLocInfo() != CCValAssign::Full)
1784 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1786 // Handle MMX values passed in GPRs.
1787 if (RegVT != VA.getLocVT() && RC == X86::GR64RegisterClass &&
1788 MVT::getSizeInBits(RegVT) == 64)
1789 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1791 ArgValues.push_back(ArgValue);
1793 assert(VA.isMemLoc());
1794 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1798 unsigned StackSize = CCInfo.getNextStackOffset();
1799 if (CC==CallingConv::Fast)
1800 StackSize =GetAlignedArgumentStackSize(StackSize, DAG);
1802 // If the function takes variable number of arguments, make a frame index for
1803 // the start of the first vararg value... for expansion of llvm.va_start.
1805 assert(CC!=CallingConv::Fast
1806 && "Var arg not supported with calling convention fastcc");
1807 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1808 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1810 // For X86-64, if there are vararg parameters that are passed via
1811 // registers, then we must store them to their spots on the stack so they
1812 // may be loaded by deferencing the result of va_next.
1813 VarArgsGPOffset = NumIntRegs * 8;
1814 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1815 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1816 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1818 // Store the integer parameter registers.
1819 SmallVector<SDOperand, 8> MemOps;
1820 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1821 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1822 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1823 for (; NumIntRegs != 6; ++NumIntRegs) {
1824 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1825 X86::GR64RegisterClass);
1826 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1827 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1828 MemOps.push_back(Store);
1829 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1830 DAG.getConstant(8, getPointerTy()));
1833 // Now store the XMM (fp + vector) parameter registers.
1834 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1835 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1836 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1837 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1838 X86::VR128RegisterClass);
1839 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1840 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1841 MemOps.push_back(Store);
1842 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1843 DAG.getConstant(16, getPointerTy()));
1845 if (!MemOps.empty())
1846 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1847 &MemOps[0], MemOps.size());
1850 ArgValues.push_back(Root);
1851 // Tail call convention (fastcc) needs callee pop.
1852 if (CC == CallingConv::Fast && PerformTailCallOpt) {
1853 BytesToPopOnReturn = StackSize; // Callee pops everything.
1854 BytesCallerReserves = 0;
1856 BytesToPopOnReturn = 0; // Callee pops nothing.
1857 BytesCallerReserves = StackSize;
1859 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1860 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1862 // Return the new list of results.
1863 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1864 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1868 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1870 SDOperand Chain = Op.getOperand(0);
1871 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1872 SDOperand Callee = Op.getOperand(4);
1874 // Analyze operands of the call, assigning locations to each operand.
1875 SmallVector<CCValAssign, 16> ArgLocs;
1876 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1877 if (CC==CallingConv::Fast && PerformTailCallOpt)
1878 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_TailCall);
1880 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1882 // Get a count of how many bytes are to be pushed on the stack.
1883 unsigned NumBytes = CCInfo.getNextStackOffset();
1884 if (CC == CallingConv::Fast)
1885 NumBytes = GetAlignedArgumentStackSize(NumBytes,DAG);
1887 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1889 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1890 SmallVector<SDOperand, 8> MemOpChains;
1894 // Walk the register/memloc assignments, inserting copies/loads.
1895 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1896 CCValAssign &VA = ArgLocs[i];
1897 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1899 // Promote the value if needed.
1900 switch (VA.getLocInfo()) {
1901 default: assert(0 && "Unknown loc info!");
1902 case CCValAssign::Full: break;
1903 case CCValAssign::SExt:
1904 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1906 case CCValAssign::ZExt:
1907 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1909 case CCValAssign::AExt:
1910 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1914 if (VA.isRegLoc()) {
1915 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1917 assert(VA.isMemLoc());
1918 if (StackPtr.Val == 0)
1919 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1921 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1926 if (!MemOpChains.empty())
1927 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1928 &MemOpChains[0], MemOpChains.size());
1930 // Build a sequence of copy-to-reg nodes chained together with token chain
1931 // and flag operands which copy the outgoing args into registers.
1933 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1934 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1936 InFlag = Chain.getValue(1);
1940 assert ( CallingConv::Fast != CC &&
1941 "Var args not supported with calling convention fastcc");
1943 // From AMD64 ABI document:
1944 // For calls that may call functions that use varargs or stdargs
1945 // (prototype-less calls or calls to functions containing ellipsis (...) in
1946 // the declaration) %al is used as hidden argument to specify the number
1947 // of SSE registers used. The contents of %al do not need to match exactly
1948 // the number of registers, but must be an ubound on the number of SSE
1949 // registers used and is in the range 0 - 8 inclusive.
1951 // Count the number of XMM registers allocated.
1952 static const unsigned XMMArgRegs[] = {
1953 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1954 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1956 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1958 Chain = DAG.getCopyToReg(Chain, X86::AL,
1959 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1960 InFlag = Chain.getValue(1);
1963 // If the callee is a GlobalAddress node (quite common, every direct call is)
1964 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1965 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1966 // We should use extra load for direct calls to dllimported functions in
1968 if (getTargetMachine().getCodeModel() != CodeModel::Large
1969 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1970 getTargetMachine(), true))
1971 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1972 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1973 if (getTargetMachine().getCodeModel() != CodeModel::Large)
1974 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1976 // Returns a chain & a flag for retval copy to use.
1977 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1978 SmallVector<SDOperand, 8> Ops;
1979 Ops.push_back(Chain);
1980 Ops.push_back(Callee);
1982 // Add argument registers to the end of the list so that they are known live
1984 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1985 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1986 RegsToPass[i].second.getValueType()));
1989 Ops.push_back(InFlag);
1991 Chain = DAG.getNode(X86ISD::CALL,
1992 NodeTys, &Ops[0], Ops.size());
1993 InFlag = Chain.getValue(1);
1994 int NumBytesForCalleeToPush = 0;
1995 if (CC==CallingConv::Fast && PerformTailCallOpt) {
1996 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1998 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2000 // Returns a flag for retval copy to use.
2001 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2003 Ops.push_back(Chain);
2004 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
2005 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
2006 Ops.push_back(InFlag);
2007 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
2008 InFlag = Chain.getValue(1);
2010 // Handle result values, copying them out of physregs into vregs that we
2012 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
2016 //===----------------------------------------------------------------------===//
2017 // Other Lowering Hooks
2018 //===----------------------------------------------------------------------===//
2021 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2022 MachineFunction &MF = DAG.getMachineFunction();
2023 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2024 int ReturnAddrIndex = FuncInfo->getRAIndex();
2026 if (ReturnAddrIndex == 0) {
2027 // Set up a frame object for the return address.
2028 if (Subtarget->is64Bit())
2029 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
2031 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
2033 FuncInfo->setRAIndex(ReturnAddrIndex);
2036 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2041 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
2042 /// specific condition code. It returns a false if it cannot do a direct
2043 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
2045 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2046 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
2047 SelectionDAG &DAG) {
2048 X86CC = X86::COND_INVALID;
2050 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2051 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2052 // X > -1 -> X == 0, jump !sign.
2053 RHS = DAG.getConstant(0, RHS.getValueType());
2054 X86CC = X86::COND_NS;
2056 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2057 // X < 0 -> X == 0, jump on sign.
2058 X86CC = X86::COND_S;
2060 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
2062 RHS = DAG.getConstant(0, RHS.getValueType());
2063 X86CC = X86::COND_LE;
2068 switch (SetCCOpcode) {
2070 case ISD::SETEQ: X86CC = X86::COND_E; break;
2071 case ISD::SETGT: X86CC = X86::COND_G; break;
2072 case ISD::SETGE: X86CC = X86::COND_GE; break;
2073 case ISD::SETLT: X86CC = X86::COND_L; break;
2074 case ISD::SETLE: X86CC = X86::COND_LE; break;
2075 case ISD::SETNE: X86CC = X86::COND_NE; break;
2076 case ISD::SETULT: X86CC = X86::COND_B; break;
2077 case ISD::SETUGT: X86CC = X86::COND_A; break;
2078 case ISD::SETULE: X86CC = X86::COND_BE; break;
2079 case ISD::SETUGE: X86CC = X86::COND_AE; break;
2082 // On a floating point condition, the flags are set as follows:
2084 // 0 | 0 | 0 | X > Y
2085 // 0 | 0 | 1 | X < Y
2086 // 1 | 0 | 0 | X == Y
2087 // 1 | 1 | 1 | unordered
2089 switch (SetCCOpcode) {
2092 case ISD::SETEQ: X86CC = X86::COND_E; break;
2093 case ISD::SETOLT: Flip = true; // Fallthrough
2095 case ISD::SETGT: X86CC = X86::COND_A; break;
2096 case ISD::SETOLE: Flip = true; // Fallthrough
2098 case ISD::SETGE: X86CC = X86::COND_AE; break;
2099 case ISD::SETUGT: Flip = true; // Fallthrough
2101 case ISD::SETLT: X86CC = X86::COND_B; break;
2102 case ISD::SETUGE: Flip = true; // Fallthrough
2104 case ISD::SETLE: X86CC = X86::COND_BE; break;
2106 case ISD::SETNE: X86CC = X86::COND_NE; break;
2107 case ISD::SETUO: X86CC = X86::COND_P; break;
2108 case ISD::SETO: X86CC = X86::COND_NP; break;
2111 std::swap(LHS, RHS);
2114 return X86CC != X86::COND_INVALID;
2117 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2118 /// code. Current x86 isa includes the following FP cmov instructions:
2119 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2120 static bool hasFPCMov(unsigned X86CC) {
2136 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2137 /// true if Op is undef or if its value falls within the specified range (L, H].
2138 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
2139 if (Op.getOpcode() == ISD::UNDEF)
2142 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2143 return (Val >= Low && Val < Hi);
2146 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2147 /// true if Op is undef or if its value equal to the specified value.
2148 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
2149 if (Op.getOpcode() == ISD::UNDEF)
2151 return cast<ConstantSDNode>(Op)->getValue() == Val;
2154 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2155 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2156 bool X86::isPSHUFDMask(SDNode *N) {
2157 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2159 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2162 // Check if the value doesn't reference the second vector.
2163 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2164 SDOperand Arg = N->getOperand(i);
2165 if (Arg.getOpcode() == ISD::UNDEF) continue;
2166 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2167 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
2174 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2175 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2176 bool X86::isPSHUFHWMask(SDNode *N) {
2177 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2179 if (N->getNumOperands() != 8)
2182 // Lower quadword copied in order.
2183 for (unsigned i = 0; i != 4; ++i) {
2184 SDOperand Arg = N->getOperand(i);
2185 if (Arg.getOpcode() == ISD::UNDEF) continue;
2186 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2187 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2191 // Upper quadword shuffled.
2192 for (unsigned i = 4; i != 8; ++i) {
2193 SDOperand Arg = N->getOperand(i);
2194 if (Arg.getOpcode() == ISD::UNDEF) continue;
2195 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2196 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2197 if (Val < 4 || Val > 7)
2204 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2205 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2206 bool X86::isPSHUFLWMask(SDNode *N) {
2207 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2209 if (N->getNumOperands() != 8)
2212 // Upper quadword copied in order.
2213 for (unsigned i = 4; i != 8; ++i)
2214 if (!isUndefOrEqual(N->getOperand(i), i))
2217 // Lower quadword shuffled.
2218 for (unsigned i = 0; i != 4; ++i)
2219 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2225 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2226 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2227 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2228 if (NumElems != 2 && NumElems != 4) return false;
2230 unsigned Half = NumElems / 2;
2231 for (unsigned i = 0; i < Half; ++i)
2232 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2234 for (unsigned i = Half; i < NumElems; ++i)
2235 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2241 bool X86::isSHUFPMask(SDNode *N) {
2242 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2243 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2246 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2247 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2248 /// half elements to come from vector 1 (which would equal the dest.) and
2249 /// the upper half to come from vector 2.
2250 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2251 if (NumOps != 2 && NumOps != 4) return false;
2253 unsigned Half = NumOps / 2;
2254 for (unsigned i = 0; i < Half; ++i)
2255 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2257 for (unsigned i = Half; i < NumOps; ++i)
2258 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2263 static bool isCommutedSHUFP(SDNode *N) {
2264 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2265 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2268 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2269 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2270 bool X86::isMOVHLPSMask(SDNode *N) {
2271 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2273 if (N->getNumOperands() != 4)
2276 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2277 return isUndefOrEqual(N->getOperand(0), 6) &&
2278 isUndefOrEqual(N->getOperand(1), 7) &&
2279 isUndefOrEqual(N->getOperand(2), 2) &&
2280 isUndefOrEqual(N->getOperand(3), 3);
2283 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2284 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2286 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2287 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2289 if (N->getNumOperands() != 4)
2292 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2293 return isUndefOrEqual(N->getOperand(0), 2) &&
2294 isUndefOrEqual(N->getOperand(1), 3) &&
2295 isUndefOrEqual(N->getOperand(2), 2) &&
2296 isUndefOrEqual(N->getOperand(3), 3);
2299 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2300 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2301 bool X86::isMOVLPMask(SDNode *N) {
2302 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2304 unsigned NumElems = N->getNumOperands();
2305 if (NumElems != 2 && NumElems != 4)
2308 for (unsigned i = 0; i < NumElems/2; ++i)
2309 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2312 for (unsigned i = NumElems/2; i < NumElems; ++i)
2313 if (!isUndefOrEqual(N->getOperand(i), i))
2319 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2320 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2322 bool X86::isMOVHPMask(SDNode *N) {
2323 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2325 unsigned NumElems = N->getNumOperands();
2326 if (NumElems != 2 && NumElems != 4)
2329 for (unsigned i = 0; i < NumElems/2; ++i)
2330 if (!isUndefOrEqual(N->getOperand(i), i))
2333 for (unsigned i = 0; i < NumElems/2; ++i) {
2334 SDOperand Arg = N->getOperand(i + NumElems/2);
2335 if (!isUndefOrEqual(Arg, i + NumElems))
2342 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2343 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2344 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2345 bool V2IsSplat = false) {
2346 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2349 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2350 SDOperand BitI = Elts[i];
2351 SDOperand BitI1 = Elts[i+1];
2352 if (!isUndefOrEqual(BitI, j))
2355 if (isUndefOrEqual(BitI1, NumElts))
2358 if (!isUndefOrEqual(BitI1, j + NumElts))
2366 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2367 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2368 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2371 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2372 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2373 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2374 bool V2IsSplat = false) {
2375 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2378 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2379 SDOperand BitI = Elts[i];
2380 SDOperand BitI1 = Elts[i+1];
2381 if (!isUndefOrEqual(BitI, j + NumElts/2))
2384 if (isUndefOrEqual(BitI1, NumElts))
2387 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2395 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2396 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2397 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2400 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2401 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2403 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2404 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2406 unsigned NumElems = N->getNumOperands();
2407 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2410 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2411 SDOperand BitI = N->getOperand(i);
2412 SDOperand BitI1 = N->getOperand(i+1);
2414 if (!isUndefOrEqual(BitI, j))
2416 if (!isUndefOrEqual(BitI1, j))
2423 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2424 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2426 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2427 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2429 unsigned NumElems = N->getNumOperands();
2430 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2433 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2434 SDOperand BitI = N->getOperand(i);
2435 SDOperand BitI1 = N->getOperand(i + 1);
2437 if (!isUndefOrEqual(BitI, j))
2439 if (!isUndefOrEqual(BitI1, j))
2446 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2447 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2448 /// MOVSD, and MOVD, i.e. setting the lowest element.
2449 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2450 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2453 if (!isUndefOrEqual(Elts[0], NumElts))
2456 for (unsigned i = 1; i < NumElts; ++i) {
2457 if (!isUndefOrEqual(Elts[i], i))
2464 bool X86::isMOVLMask(SDNode *N) {
2465 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2466 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2469 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2470 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2471 /// element of vector 2 and the other elements to come from vector 1 in order.
2472 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2473 bool V2IsSplat = false,
2474 bool V2IsUndef = false) {
2475 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2478 if (!isUndefOrEqual(Ops[0], 0))
2481 for (unsigned i = 1; i < NumOps; ++i) {
2482 SDOperand Arg = Ops[i];
2483 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2484 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2485 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2492 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2493 bool V2IsUndef = false) {
2494 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2495 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2496 V2IsSplat, V2IsUndef);
2499 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2500 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2501 bool X86::isMOVSHDUPMask(SDNode *N) {
2502 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2504 if (N->getNumOperands() != 4)
2507 // Expect 1, 1, 3, 3
2508 for (unsigned i = 0; i < 2; ++i) {
2509 SDOperand Arg = N->getOperand(i);
2510 if (Arg.getOpcode() == ISD::UNDEF) continue;
2511 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2512 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2513 if (Val != 1) return false;
2517 for (unsigned i = 2; i < 4; ++i) {
2518 SDOperand Arg = N->getOperand(i);
2519 if (Arg.getOpcode() == ISD::UNDEF) continue;
2520 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2521 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2522 if (Val != 3) return false;
2526 // Don't use movshdup if it can be done with a shufps.
2530 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2531 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2532 bool X86::isMOVSLDUPMask(SDNode *N) {
2533 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2535 if (N->getNumOperands() != 4)
2538 // Expect 0, 0, 2, 2
2539 for (unsigned i = 0; i < 2; ++i) {
2540 SDOperand Arg = N->getOperand(i);
2541 if (Arg.getOpcode() == ISD::UNDEF) continue;
2542 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2543 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2544 if (Val != 0) return false;
2548 for (unsigned i = 2; i < 4; ++i) {
2549 SDOperand Arg = N->getOperand(i);
2550 if (Arg.getOpcode() == ISD::UNDEF) continue;
2551 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2552 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2553 if (Val != 2) return false;
2557 // Don't use movshdup if it can be done with a shufps.
2561 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2562 /// specifies a identity operation on the LHS or RHS.
2563 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2564 unsigned NumElems = N->getNumOperands();
2565 for (unsigned i = 0; i < NumElems; ++i)
2566 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2571 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2572 /// a splat of a single element.
2573 static bool isSplatMask(SDNode *N) {
2574 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2576 // This is a splat operation if each element of the permute is the same, and
2577 // if the value doesn't reference the second vector.
2578 unsigned NumElems = N->getNumOperands();
2579 SDOperand ElementBase;
2581 for (; i != NumElems; ++i) {
2582 SDOperand Elt = N->getOperand(i);
2583 if (isa<ConstantSDNode>(Elt)) {
2589 if (!ElementBase.Val)
2592 for (; i != NumElems; ++i) {
2593 SDOperand Arg = N->getOperand(i);
2594 if (Arg.getOpcode() == ISD::UNDEF) continue;
2595 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2596 if (Arg != ElementBase) return false;
2599 // Make sure it is a splat of the first vector operand.
2600 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2603 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2604 /// a splat of a single element and it's a 2 or 4 element mask.
2605 bool X86::isSplatMask(SDNode *N) {
2606 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2608 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2609 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2611 return ::isSplatMask(N);
2614 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2615 /// specifies a splat of zero element.
2616 bool X86::isSplatLoMask(SDNode *N) {
2617 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2619 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2620 if (!isUndefOrEqual(N->getOperand(i), 0))
2625 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2626 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2628 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2629 unsigned NumOperands = N->getNumOperands();
2630 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2632 for (unsigned i = 0; i < NumOperands; ++i) {
2634 SDOperand Arg = N->getOperand(NumOperands-i-1);
2635 if (Arg.getOpcode() != ISD::UNDEF)
2636 Val = cast<ConstantSDNode>(Arg)->getValue();
2637 if (Val >= NumOperands) Val -= NumOperands;
2639 if (i != NumOperands - 1)
2646 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2647 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2649 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2651 // 8 nodes, but we only care about the last 4.
2652 for (unsigned i = 7; i >= 4; --i) {
2654 SDOperand Arg = N->getOperand(i);
2655 if (Arg.getOpcode() != ISD::UNDEF)
2656 Val = cast<ConstantSDNode>(Arg)->getValue();
2665 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2666 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2668 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2670 // 8 nodes, but we only care about the first 4.
2671 for (int i = 3; i >= 0; --i) {
2673 SDOperand Arg = N->getOperand(i);
2674 if (Arg.getOpcode() != ISD::UNDEF)
2675 Val = cast<ConstantSDNode>(Arg)->getValue();
2684 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2685 /// specifies a 8 element shuffle that can be broken into a pair of
2686 /// PSHUFHW and PSHUFLW.
2687 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2688 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2690 if (N->getNumOperands() != 8)
2693 // Lower quadword shuffled.
2694 for (unsigned i = 0; i != 4; ++i) {
2695 SDOperand Arg = N->getOperand(i);
2696 if (Arg.getOpcode() == ISD::UNDEF) continue;
2697 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2698 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2703 // Upper quadword shuffled.
2704 for (unsigned i = 4; i != 8; ++i) {
2705 SDOperand Arg = N->getOperand(i);
2706 if (Arg.getOpcode() == ISD::UNDEF) continue;
2707 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2708 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2709 if (Val < 4 || Val > 7)
2716 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2717 /// values in ther permute mask.
2718 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2719 SDOperand &V2, SDOperand &Mask,
2720 SelectionDAG &DAG) {
2721 MVT::ValueType VT = Op.getValueType();
2722 MVT::ValueType MaskVT = Mask.getValueType();
2723 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2724 unsigned NumElems = Mask.getNumOperands();
2725 SmallVector<SDOperand, 8> MaskVec;
2727 for (unsigned i = 0; i != NumElems; ++i) {
2728 SDOperand Arg = Mask.getOperand(i);
2729 if (Arg.getOpcode() == ISD::UNDEF) {
2730 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2733 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2734 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2736 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2738 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2742 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2743 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2746 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2747 /// match movhlps. The lower half elements should come from upper half of
2748 /// V1 (and in order), and the upper half elements should come from the upper
2749 /// half of V2 (and in order).
2750 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2751 unsigned NumElems = Mask->getNumOperands();
2754 for (unsigned i = 0, e = 2; i != e; ++i)
2755 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2757 for (unsigned i = 2; i != 4; ++i)
2758 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2763 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2764 /// is promoted to a vector.
2765 static inline bool isScalarLoadToVector(SDNode *N) {
2766 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2767 N = N->getOperand(0).Val;
2768 return ISD::isNON_EXTLoad(N);
2773 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2774 /// match movlp{s|d}. The lower half elements should come from lower half of
2775 /// V1 (and in order), and the upper half elements should come from the upper
2776 /// half of V2 (and in order). And since V1 will become the source of the
2777 /// MOVLP, it must be either a vector load or a scalar load to vector.
2778 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2779 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2781 // Is V2 is a vector load, don't do this transformation. We will try to use
2782 // load folding shufps op.
2783 if (ISD::isNON_EXTLoad(V2))
2786 unsigned NumElems = Mask->getNumOperands();
2787 if (NumElems != 2 && NumElems != 4)
2789 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2790 if (!isUndefOrEqual(Mask->getOperand(i), i))
2792 for (unsigned i = NumElems/2; i != NumElems; ++i)
2793 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2798 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2800 static bool isSplatVector(SDNode *N) {
2801 if (N->getOpcode() != ISD::BUILD_VECTOR)
2804 SDOperand SplatValue = N->getOperand(0);
2805 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2806 if (N->getOperand(i) != SplatValue)
2811 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2813 static bool isUndefShuffle(SDNode *N) {
2814 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2817 SDOperand V1 = N->getOperand(0);
2818 SDOperand V2 = N->getOperand(1);
2819 SDOperand Mask = N->getOperand(2);
2820 unsigned NumElems = Mask.getNumOperands();
2821 for (unsigned i = 0; i != NumElems; ++i) {
2822 SDOperand Arg = Mask.getOperand(i);
2823 if (Arg.getOpcode() != ISD::UNDEF) {
2824 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2825 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2827 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2834 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2836 static inline bool isZeroNode(SDOperand Elt) {
2837 return ((isa<ConstantSDNode>(Elt) &&
2838 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2839 (isa<ConstantFPSDNode>(Elt) &&
2840 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2843 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2844 /// to an zero vector.
2845 static bool isZeroShuffle(SDNode *N) {
2846 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2849 SDOperand V1 = N->getOperand(0);
2850 SDOperand V2 = N->getOperand(1);
2851 SDOperand Mask = N->getOperand(2);
2852 unsigned NumElems = Mask.getNumOperands();
2853 for (unsigned i = 0; i != NumElems; ++i) {
2854 SDOperand Arg = Mask.getOperand(i);
2855 if (Arg.getOpcode() != ISD::UNDEF) {
2856 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2857 if (Idx < NumElems) {
2858 unsigned Opc = V1.Val->getOpcode();
2859 if (Opc == ISD::UNDEF)
2861 if (Opc != ISD::BUILD_VECTOR ||
2862 !isZeroNode(V1.Val->getOperand(Idx)))
2864 } else if (Idx >= NumElems) {
2865 unsigned Opc = V2.Val->getOpcode();
2866 if (Opc == ISD::UNDEF)
2868 if (Opc != ISD::BUILD_VECTOR ||
2869 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2877 /// getZeroVector - Returns a vector of specified type with all zero elements.
2879 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2880 assert(MVT::isVector(VT) && "Expected a vector type");
2881 unsigned NumElems = MVT::getVectorNumElements(VT);
2882 MVT::ValueType EVT = MVT::getVectorElementType(VT);
2883 bool isFP = MVT::isFloatingPoint(EVT);
2884 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2885 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2886 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2889 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2890 /// that point to V2 points to its first element.
2891 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2892 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2894 bool Changed = false;
2895 SmallVector<SDOperand, 8> MaskVec;
2896 unsigned NumElems = Mask.getNumOperands();
2897 for (unsigned i = 0; i != NumElems; ++i) {
2898 SDOperand Arg = Mask.getOperand(i);
2899 if (Arg.getOpcode() != ISD::UNDEF) {
2900 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2901 if (Val > NumElems) {
2902 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2906 MaskVec.push_back(Arg);
2910 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2911 &MaskVec[0], MaskVec.size());
2915 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2916 /// operation of specified width.
2917 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2918 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2919 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2921 SmallVector<SDOperand, 8> MaskVec;
2922 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2923 for (unsigned i = 1; i != NumElems; ++i)
2924 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2925 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2928 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2929 /// of specified width.
2930 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2931 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2932 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2933 SmallVector<SDOperand, 8> MaskVec;
2934 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2935 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2936 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2938 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2941 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2942 /// of specified width.
2943 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2944 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2945 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2946 unsigned Half = NumElems/2;
2947 SmallVector<SDOperand, 8> MaskVec;
2948 for (unsigned i = 0; i != Half; ++i) {
2949 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2950 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2952 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2955 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2957 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2958 SDOperand V1 = Op.getOperand(0);
2959 SDOperand Mask = Op.getOperand(2);
2960 MVT::ValueType VT = Op.getValueType();
2961 unsigned NumElems = Mask.getNumOperands();
2962 Mask = getUnpacklMask(NumElems, DAG);
2963 while (NumElems != 4) {
2964 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2967 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2969 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2970 Mask = getZeroVector(MaskVT, DAG);
2971 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2972 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2973 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2976 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2977 /// vector of zero or undef vector.
2978 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2979 unsigned NumElems, unsigned Idx,
2980 bool isZero, SelectionDAG &DAG) {
2981 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2982 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2983 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
2984 SDOperand Zero = DAG.getConstant(0, EVT);
2985 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2986 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2987 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2988 &MaskVec[0], MaskVec.size());
2989 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2992 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2994 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2995 unsigned NumNonZero, unsigned NumZero,
2996 SelectionDAG &DAG, TargetLowering &TLI) {
3002 for (unsigned i = 0; i < 16; ++i) {
3003 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3004 if (ThisIsNonZero && First) {
3006 V = getZeroVector(MVT::v8i16, DAG);
3008 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3013 SDOperand ThisElt(0, 0), LastElt(0, 0);
3014 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3015 if (LastIsNonZero) {
3016 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3018 if (ThisIsNonZero) {
3019 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3020 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3021 ThisElt, DAG.getConstant(8, MVT::i8));
3023 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3028 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3029 DAG.getConstant(i/2, TLI.getPointerTy()));
3033 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3036 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3038 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
3039 unsigned NumNonZero, unsigned NumZero,
3040 SelectionDAG &DAG, TargetLowering &TLI) {
3046 for (unsigned i = 0; i < 8; ++i) {
3047 bool isNonZero = (NonZeros & (1 << i)) != 0;
3051 V = getZeroVector(MVT::v8i16, DAG);
3053 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3056 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3057 DAG.getConstant(i, TLI.getPointerTy()));
3065 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3066 // All zero's are handled with pxor.
3067 if (ISD::isBuildVectorAllZeros(Op.Val))
3070 // All one's are handled with pcmpeqd.
3071 if (ISD::isBuildVectorAllOnes(Op.Val))
3074 MVT::ValueType VT = Op.getValueType();
3075 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3076 unsigned EVTBits = MVT::getSizeInBits(EVT);
3078 unsigned NumElems = Op.getNumOperands();
3079 unsigned NumZero = 0;
3080 unsigned NumNonZero = 0;
3081 unsigned NonZeros = 0;
3082 unsigned NumNonZeroImms = 0;
3083 std::set<SDOperand> Values;
3084 for (unsigned i = 0; i < NumElems; ++i) {
3085 SDOperand Elt = Op.getOperand(i);
3086 if (Elt.getOpcode() != ISD::UNDEF) {
3088 if (isZeroNode(Elt))
3091 NonZeros |= (1 << i);
3093 if (Elt.getOpcode() == ISD::Constant ||
3094 Elt.getOpcode() == ISD::ConstantFP)
3100 if (NumNonZero == 0) {
3102 // All undef vector. Return an UNDEF.
3103 return DAG.getNode(ISD::UNDEF, VT);
3105 // A mix of zero and undef. Return a zero vector.
3106 return getZeroVector(VT, DAG);
3109 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3110 if (Values.size() == 1)
3113 // Special case for single non-zero element.
3114 if (NumNonZero == 1) {
3115 unsigned Idx = CountTrailingZeros_32(NonZeros);
3116 SDOperand Item = Op.getOperand(Idx);
3117 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3119 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3120 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
3123 if (EVTBits == 32) {
3124 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3125 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
3127 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3128 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3129 SmallVector<SDOperand, 8> MaskVec;
3130 for (unsigned i = 0; i < NumElems; i++)
3131 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3132 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3133 &MaskVec[0], MaskVec.size());
3134 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3135 DAG.getNode(ISD::UNDEF, VT), Mask);
3139 // A vector full of immediates; various special cases are already
3140 // handled, so this is best done with a single constant-pool load.
3141 if (NumNonZero == NumNonZeroImms)
3144 // Let legalizer expand 2-wide build_vectors.
3148 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3149 if (EVTBits == 8 && NumElems == 16) {
3150 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3152 if (V.Val) return V;
3155 if (EVTBits == 16 && NumElems == 8) {
3156 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3158 if (V.Val) return V;
3161 // If element VT is == 32 bits, turn it into a number of shuffles.
3162 SmallVector<SDOperand, 8> V;
3164 if (NumElems == 4 && NumZero > 0) {
3165 for (unsigned i = 0; i < 4; ++i) {
3166 bool isZero = !(NonZeros & (1 << i));
3168 V[i] = getZeroVector(VT, DAG);
3170 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3173 for (unsigned i = 0; i < 2; ++i) {
3174 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3177 V[i] = V[i*2]; // Must be a zero vector.
3180 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3181 getMOVLMask(NumElems, DAG));
3184 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3185 getMOVLMask(NumElems, DAG));
3188 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3189 getUnpacklMask(NumElems, DAG));
3194 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3195 // clears the upper bits.
3196 // FIXME: we can do the same for v4f32 case when we know both parts of
3197 // the lower half come from scalar_to_vector (loadf32). We should do
3198 // that in post legalizer dag combiner with target specific hooks.
3199 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3201 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3202 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3203 SmallVector<SDOperand, 8> MaskVec;
3204 bool Reverse = (NonZeros & 0x3) == 2;
3205 for (unsigned i = 0; i < 2; ++i)
3207 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3209 MaskVec.push_back(DAG.getConstant(i, EVT));
3210 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3211 for (unsigned i = 0; i < 2; ++i)
3213 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3215 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3216 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3217 &MaskVec[0], MaskVec.size());
3218 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3221 if (Values.size() > 2) {
3222 // Expand into a number of unpckl*.
3224 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3225 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3226 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3227 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3228 for (unsigned i = 0; i < NumElems; ++i)
3229 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3231 while (NumElems != 0) {
3232 for (unsigned i = 0; i < NumElems; ++i)
3233 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3244 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3245 SDOperand V1 = Op.getOperand(0);
3246 SDOperand V2 = Op.getOperand(1);
3247 SDOperand PermMask = Op.getOperand(2);
3248 MVT::ValueType VT = Op.getValueType();
3249 unsigned NumElems = PermMask.getNumOperands();
3250 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3251 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3252 bool V1IsSplat = false;
3253 bool V2IsSplat = false;
3255 if (isUndefShuffle(Op.Val))
3256 return DAG.getNode(ISD::UNDEF, VT);
3258 if (isZeroShuffle(Op.Val))
3259 return getZeroVector(VT, DAG);
3261 if (isIdentityMask(PermMask.Val))
3263 else if (isIdentityMask(PermMask.Val, true))
3266 if (isSplatMask(PermMask.Val)) {
3267 if (NumElems <= 4) return Op;
3268 // Promote it to a v4i32 splat.
3269 return PromoteSplat(Op, DAG);
3272 if (X86::isMOVLMask(PermMask.Val))
3273 return (V1IsUndef) ? V2 : Op;
3275 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3276 X86::isMOVSLDUPMask(PermMask.Val) ||
3277 X86::isMOVHLPSMask(PermMask.Val) ||
3278 X86::isMOVHPMask(PermMask.Val) ||
3279 X86::isMOVLPMask(PermMask.Val))
3282 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3283 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3284 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3286 bool Commuted = false;
3287 V1IsSplat = isSplatVector(V1.Val);
3288 V2IsSplat = isSplatVector(V2.Val);
3289 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3290 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3291 std::swap(V1IsSplat, V2IsSplat);
3292 std::swap(V1IsUndef, V2IsUndef);
3296 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3297 if (V2IsUndef) return V1;
3298 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3300 // V2 is a splat, so the mask may be malformed. That is, it may point
3301 // to any V2 element. The instruction selectior won't like this. Get
3302 // a corrected mask and commute to form a proper MOVS{S|D}.
3303 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3304 if (NewMask.Val != PermMask.Val)
3305 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3310 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3311 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3312 X86::isUNPCKLMask(PermMask.Val) ||
3313 X86::isUNPCKHMask(PermMask.Val))
3317 // Normalize mask so all entries that point to V2 points to its first
3318 // element then try to match unpck{h|l} again. If match, return a
3319 // new vector_shuffle with the corrected mask.
3320 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3321 if (NewMask.Val != PermMask.Val) {
3322 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3323 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3324 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3325 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3326 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3327 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3332 // Normalize the node to match x86 shuffle ops if needed
3333 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3334 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3337 // Commute is back and try unpck* again.
3338 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3339 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3340 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3341 X86::isUNPCKLMask(PermMask.Val) ||
3342 X86::isUNPCKHMask(PermMask.Val))
3346 // If VT is integer, try PSHUF* first, then SHUFP*.
3347 if (MVT::isInteger(VT)) {
3348 // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
3349 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
3350 if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
3351 X86::isPSHUFDMask(PermMask.Val)) ||
3352 X86::isPSHUFHWMask(PermMask.Val) ||
3353 X86::isPSHUFLWMask(PermMask.Val)) {
3354 if (V2.getOpcode() != ISD::UNDEF)
3355 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3356 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3360 if (X86::isSHUFPMask(PermMask.Val) &&
3361 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
3364 // Handle v8i16 shuffle high / low shuffle node pair.
3365 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3366 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3367 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3368 SmallVector<SDOperand, 8> MaskVec;
3369 for (unsigned i = 0; i != 4; ++i)
3370 MaskVec.push_back(PermMask.getOperand(i));
3371 for (unsigned i = 4; i != 8; ++i)
3372 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3373 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3374 &MaskVec[0], MaskVec.size());
3375 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3377 for (unsigned i = 0; i != 4; ++i)
3378 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3379 for (unsigned i = 4; i != 8; ++i)
3380 MaskVec.push_back(PermMask.getOperand(i));
3381 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
3382 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3385 // Floating point cases in the other order.
3386 if (X86::isSHUFPMask(PermMask.Val))
3388 if (X86::isPSHUFDMask(PermMask.Val) ||
3389 X86::isPSHUFHWMask(PermMask.Val) ||
3390 X86::isPSHUFLWMask(PermMask.Val)) {
3391 if (V2.getOpcode() != ISD::UNDEF)
3392 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3393 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3398 if (NumElems == 4 &&
3399 // Don't do this for MMX.
3400 MVT::getSizeInBits(VT) != 64) {
3401 MVT::ValueType MaskVT = PermMask.getValueType();
3402 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3403 SmallVector<std::pair<int, int>, 8> Locs;
3404 Locs.reserve(NumElems);
3405 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3406 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3409 // If no more than two elements come from either vector. This can be
3410 // implemented with two shuffles. First shuffle gather the elements.
3411 // The second shuffle, which takes the first shuffle as both of its
3412 // vector operands, put the elements into the right order.
3413 for (unsigned i = 0; i != NumElems; ++i) {
3414 SDOperand Elt = PermMask.getOperand(i);
3415 if (Elt.getOpcode() == ISD::UNDEF) {
3416 Locs[i] = std::make_pair(-1, -1);
3418 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3419 if (Val < NumElems) {
3420 Locs[i] = std::make_pair(0, NumLo);
3424 Locs[i] = std::make_pair(1, NumHi);
3425 if (2+NumHi < NumElems)
3426 Mask1[2+NumHi] = Elt;
3431 if (NumLo <= 2 && NumHi <= 2) {
3432 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3433 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3434 &Mask1[0], Mask1.size()));
3435 for (unsigned i = 0; i != NumElems; ++i) {
3436 if (Locs[i].first == -1)
3439 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3440 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3441 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3445 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3446 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3447 &Mask2[0], Mask2.size()));
3450 // Break it into (shuffle shuffle_hi, shuffle_lo).
3452 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3453 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3454 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3455 unsigned MaskIdx = 0;
3457 unsigned HiIdx = NumElems/2;
3458 for (unsigned i = 0; i != NumElems; ++i) {
3459 if (i == NumElems/2) {
3465 SDOperand Elt = PermMask.getOperand(i);
3466 if (Elt.getOpcode() == ISD::UNDEF) {
3467 Locs[i] = std::make_pair(-1, -1);
3468 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3469 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3470 (*MaskPtr)[LoIdx] = Elt;
3473 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3474 (*MaskPtr)[HiIdx] = Elt;
3479 SDOperand LoShuffle =
3480 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3481 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3482 &LoMask[0], LoMask.size()));
3483 SDOperand HiShuffle =
3484 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3485 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3486 &HiMask[0], HiMask.size()));
3487 SmallVector<SDOperand, 8> MaskOps;
3488 for (unsigned i = 0; i != NumElems; ++i) {
3489 if (Locs[i].first == -1) {
3490 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3492 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3493 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3496 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3497 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3498 &MaskOps[0], MaskOps.size()));
3505 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3506 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3509 MVT::ValueType VT = Op.getValueType();
3510 // TODO: handle v16i8.
3511 if (MVT::getSizeInBits(VT) == 16) {
3512 // Transform it so it match pextrw which produces a 32-bit result.
3513 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3514 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3515 Op.getOperand(0), Op.getOperand(1));
3516 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3517 DAG.getValueType(VT));
3518 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3519 } else if (MVT::getSizeInBits(VT) == 32) {
3520 SDOperand Vec = Op.getOperand(0);
3521 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3524 // SHUFPS the element to the lowest double word, then movss.
3525 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3526 SmallVector<SDOperand, 8> IdxVec;
3528 push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
3530 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3532 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3534 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3535 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3536 &IdxVec[0], IdxVec.size());
3537 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3538 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3539 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3540 DAG.getConstant(0, getPointerTy()));
3541 } else if (MVT::getSizeInBits(VT) == 64) {
3542 SDOperand Vec = Op.getOperand(0);
3543 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3547 // UNPCKHPD the element to the lowest double word, then movsd.
3548 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3549 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3550 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3551 SmallVector<SDOperand, 8> IdxVec;
3552 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
3554 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3555 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3556 &IdxVec[0], IdxVec.size());
3557 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3558 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3559 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3560 DAG.getConstant(0, getPointerTy()));
3567 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3568 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3569 // as its second argument.
3570 MVT::ValueType VT = Op.getValueType();
3571 MVT::ValueType BaseVT = MVT::getVectorElementType(VT);
3572 SDOperand N0 = Op.getOperand(0);
3573 SDOperand N1 = Op.getOperand(1);
3574 SDOperand N2 = Op.getOperand(2);
3575 if (MVT::getSizeInBits(BaseVT) == 16) {
3576 if (N1.getValueType() != MVT::i32)
3577 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3578 if (N2.getValueType() != MVT::i32)
3579 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(),getPointerTy());
3580 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3581 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3582 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3585 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3586 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3587 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3588 SmallVector<SDOperand, 8> MaskVec;
3589 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3590 for (unsigned i = 1; i <= 3; ++i)
3591 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3592 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3593 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3594 &MaskVec[0], MaskVec.size()));
3596 // Use two pinsrw instructions to insert a 32 bit value.
3598 if (MVT::isFloatingPoint(N1.getValueType())) {
3599 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3600 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3601 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3602 DAG.getConstant(0, getPointerTy()));
3604 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3605 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3606 DAG.getConstant(Idx, getPointerTy()));
3607 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3608 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3609 DAG.getConstant(Idx+1, getPointerTy()));
3610 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3618 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3619 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3620 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3623 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3624 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3625 // one of the above mentioned nodes. It has to be wrapped because otherwise
3626 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3627 // be used to form addressing mode. These wrapped nodes will be selected
3630 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3631 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3632 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3634 CP->getAlignment());
3635 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3636 // With PIC, the address is actually $g + Offset.
3637 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3638 !Subtarget->isPICStyleRIPRel()) {
3639 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3640 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3648 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3649 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3650 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3651 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3652 // With PIC, the address is actually $g + Offset.
3653 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3654 !Subtarget->isPICStyleRIPRel()) {
3655 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3656 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3660 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3661 // load the value at address GV, not the value of GV itself. This means that
3662 // the GlobalAddress must be in the base or index register of the address, not
3663 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3664 // The same applies for external symbols during PIC codegen
3665 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3666 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3671 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3673 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3674 const MVT::ValueType PtrVT) {
3676 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
3677 DAG.getNode(X86ISD::GlobalBaseReg,
3679 InFlag = Chain.getValue(1);
3681 // emit leal symbol@TLSGD(,%ebx,1), %eax
3682 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3683 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3684 GA->getValueType(0),
3686 SDOperand Ops[] = { Chain, TGA, InFlag };
3687 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3688 InFlag = Result.getValue(2);
3689 Chain = Result.getValue(1);
3691 // call ___tls_get_addr. This function receives its argument in
3692 // the register EAX.
3693 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
3694 InFlag = Chain.getValue(1);
3696 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3697 SDOperand Ops1[] = { Chain,
3698 DAG.getTargetExternalSymbol("___tls_get_addr",
3700 DAG.getRegister(X86::EAX, PtrVT),
3701 DAG.getRegister(X86::EBX, PtrVT),
3703 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
3704 InFlag = Chain.getValue(1);
3706 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
3709 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
3710 // "local exec" model.
3712 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3713 const MVT::ValueType PtrVT) {
3714 // Get the Thread Pointer
3715 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
3716 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
3718 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3719 GA->getValueType(0),
3721 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
3723 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
3724 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset, NULL, 0);
3726 // The address of the thread local variable is the add of the thread
3727 // pointer with the offset of the variable.
3728 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
3732 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
3733 // TODO: implement the "local dynamic" model
3734 // TODO: implement the "initial exec"model for pic executables
3735 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
3736 "TLS not implemented for non-ELF and 64-bit targets");
3737 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3738 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
3739 // otherwise use the "Local Exec"TLS Model
3740 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
3741 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
3743 return LowerToTLSExecModel(GA, DAG, getPointerTy());
3747 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3748 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3749 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3750 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3751 // With PIC, the address is actually $g + Offset.
3752 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3753 !Subtarget->isPICStyleRIPRel()) {
3754 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3755 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3762 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3763 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3764 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3765 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3766 // With PIC, the address is actually $g + Offset.
3767 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3768 !Subtarget->isPICStyleRIPRel()) {
3769 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3770 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3777 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
3778 /// take a 2 x i32 value to shift plus a shift amount.
3779 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3780 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3781 "Not an i64 shift!");
3782 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3783 SDOperand ShOpLo = Op.getOperand(0);
3784 SDOperand ShOpHi = Op.getOperand(1);
3785 SDOperand ShAmt = Op.getOperand(2);
3786 SDOperand Tmp1 = isSRA ?
3787 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3788 DAG.getConstant(0, MVT::i32);
3790 SDOperand Tmp2, Tmp3;
3791 if (Op.getOpcode() == ISD::SHL_PARTS) {
3792 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3793 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3795 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3796 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3799 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3800 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3801 DAG.getConstant(32, MVT::i8));
3802 SDOperand Cond = DAG.getNode(X86ISD::CMP, MVT::i32,
3803 AndNode, DAG.getConstant(0, MVT::i8));
3806 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3807 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3808 SmallVector<SDOperand, 4> Ops;
3809 if (Op.getOpcode() == ISD::SHL_PARTS) {
3810 Ops.push_back(Tmp2);
3811 Ops.push_back(Tmp3);
3813 Ops.push_back(Cond);
3814 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3817 Ops.push_back(Tmp3);
3818 Ops.push_back(Tmp1);
3820 Ops.push_back(Cond);
3821 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3823 Ops.push_back(Tmp2);
3824 Ops.push_back(Tmp3);
3826 Ops.push_back(Cond);
3827 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3830 Ops.push_back(Tmp3);
3831 Ops.push_back(Tmp1);
3833 Ops.push_back(Cond);
3834 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3837 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3841 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3844 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3845 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3846 Op.getOperand(0).getValueType() >= MVT::i16 &&
3847 "Unknown SINT_TO_FP to lower!");
3850 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3851 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3852 MachineFunction &MF = DAG.getMachineFunction();
3853 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3854 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3855 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3856 StackSlot, NULL, 0);
3858 // These are really Legal; caller falls through into that case.
3859 if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f32 && X86ScalarSSEf32)
3861 if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f64 && X86ScalarSSEf64)
3863 if (SrcVT==MVT::i64 && Op.getValueType() != MVT::f80 &&
3864 Subtarget->is64Bit())
3869 bool useSSE = (X86ScalarSSEf32 && Op.getValueType() == MVT::f32) ||
3870 (X86ScalarSSEf64 && Op.getValueType() == MVT::f64);
3872 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3874 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
3875 SmallVector<SDOperand, 8> Ops;
3876 Ops.push_back(Chain);
3877 Ops.push_back(StackSlot);
3878 Ops.push_back(DAG.getValueType(SrcVT));
3879 Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3880 Tys, &Ops[0], Ops.size());
3883 Chain = Result.getValue(1);
3884 SDOperand InFlag = Result.getValue(2);
3886 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3887 // shouldn't be necessary except that RFP cannot be live across
3888 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3889 MachineFunction &MF = DAG.getMachineFunction();
3890 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3891 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3892 Tys = DAG.getVTList(MVT::Other);
3893 SmallVector<SDOperand, 8> Ops;
3894 Ops.push_back(Chain);
3895 Ops.push_back(Result);
3896 Ops.push_back(StackSlot);
3897 Ops.push_back(DAG.getValueType(Op.getValueType()));
3898 Ops.push_back(InFlag);
3899 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3900 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3906 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3907 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3908 "Unknown FP_TO_SINT to lower!");
3911 // These are really Legal.
3912 if (Op.getValueType() == MVT::i32 &&
3913 X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32)
3915 if (Op.getValueType() == MVT::i32 &&
3916 X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)
3918 if (Subtarget->is64Bit() &&
3919 Op.getValueType() == MVT::i64 &&
3920 Op.getOperand(0).getValueType() != MVT::f80)
3923 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3925 MachineFunction &MF = DAG.getMachineFunction();
3926 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3927 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3928 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3930 switch (Op.getValueType()) {
3931 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3932 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3933 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3934 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3937 SDOperand Chain = DAG.getEntryNode();
3938 SDOperand Value = Op.getOperand(0);
3939 if ((X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32) ||
3940 (X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)) {
3941 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3942 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3943 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
3945 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3947 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3948 Chain = Value.getValue(1);
3949 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3950 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3953 // Build the FP_TO_INT*_IN_MEM
3954 SDOperand Ops[] = { Chain, Value, StackSlot };
3955 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3957 // Load the result. If this is an i64 load on an x86-32 host, expand the
3959 if (Op.getValueType() != MVT::i64 || Subtarget->is64Bit())
3960 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3962 SDOperand Lo = DAG.getLoad(MVT::i32, FIST, StackSlot, NULL, 0);
3963 StackSlot = DAG.getNode(ISD::ADD, StackSlot.getValueType(), StackSlot,
3964 DAG.getConstant(StackSlot.getValueType(), 4));
3965 SDOperand Hi = DAG.getLoad(MVT::i32, FIST, StackSlot, NULL, 0);
3968 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
3971 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3972 MVT::ValueType VT = Op.getValueType();
3973 MVT::ValueType EltVT = VT;
3974 if (MVT::isVector(VT))
3975 EltVT = MVT::getVectorElementType(VT);
3976 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
3977 std::vector<Constant*> CV;
3978 if (EltVT == MVT::f64) {
3979 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, ~(1ULL << 63))));
3983 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, ~(1U << 31))));
3989 Constant *C = ConstantVector::get(CV);
3990 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
3991 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
3993 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3996 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3997 MVT::ValueType VT = Op.getValueType();
3998 MVT::ValueType EltVT = VT;
3999 unsigned EltNum = 1;
4000 if (MVT::isVector(VT)) {
4001 EltVT = MVT::getVectorElementType(VT);
4002 EltNum = MVT::getVectorNumElements(VT);
4004 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4005 std::vector<Constant*> CV;
4006 if (EltVT == MVT::f64) {
4007 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, 1ULL << 63)));
4011 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, 1U << 31)));
4017 Constant *C = ConstantVector::get(CV);
4018 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4019 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4021 if (MVT::isVector(VT)) {
4022 return DAG.getNode(ISD::BIT_CONVERT, VT,
4023 DAG.getNode(ISD::XOR, MVT::v2i64,
4024 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4025 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4027 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4031 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
4032 SDOperand Op0 = Op.getOperand(0);
4033 SDOperand Op1 = Op.getOperand(1);
4034 MVT::ValueType VT = Op.getValueType();
4035 MVT::ValueType SrcVT = Op1.getValueType();
4036 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
4038 // If second operand is smaller, extend it first.
4039 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
4040 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4042 SrcTy = MVT::getTypeForValueType(SrcVT);
4044 // And if it is bigger, shrink it first.
4045 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4046 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1);
4048 SrcTy = MVT::getTypeForValueType(SrcVT);
4051 // At this point the operands and the result should have the same
4052 // type, and that won't be f80 since that is not custom lowered.
4054 // First get the sign bit of second operand.
4055 std::vector<Constant*> CV;
4056 if (SrcVT == MVT::f64) {
4057 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 1ULL << 63))));
4058 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4060 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 1U << 31))));
4061 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4062 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4063 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4065 Constant *C = ConstantVector::get(CV);
4066 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4067 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx, NULL, 0,
4069 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4071 // Shift sign bit right or left if the two operands have different types.
4072 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4073 // Op0 is MVT::f32, Op1 is MVT::f64.
4074 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4075 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4076 DAG.getConstant(32, MVT::i32));
4077 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4078 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4079 DAG.getConstant(0, getPointerTy()));
4082 // Clear first operand sign bit.
4084 if (VT == MVT::f64) {
4085 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, ~(1ULL << 63)))));
4086 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4088 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, ~(1U << 31)))));
4089 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4090 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4091 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4093 C = ConstantVector::get(CV);
4094 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4095 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4097 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4099 // Or the value with the sign bit.
4100 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4103 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
4104 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4106 SDOperand Op0 = Op.getOperand(0);
4107 SDOperand Op1 = Op.getOperand(1);
4108 SDOperand CC = Op.getOperand(2);
4109 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4110 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
4113 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4115 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4116 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4117 DAG.getConstant(X86CC, MVT::i8), Cond);
4120 assert(isFP && "Illegal integer SetCC!");
4122 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4123 switch (SetCCOpcode) {
4124 default: assert(false && "Illegal floating point SetCC!");
4125 case ISD::SETOEQ: { // !PF & ZF
4126 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4127 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4128 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4129 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4130 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4132 case ISD::SETUNE: { // PF | !ZF
4133 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4134 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4135 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4136 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4137 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4143 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
4144 bool addTest = true;
4145 SDOperand Cond = Op.getOperand(0);
4148 if (Cond.getOpcode() == ISD::SETCC)
4149 Cond = LowerSETCC(Cond, DAG);
4151 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4152 // setting operand in place of the X86ISD::SETCC.
4153 if (Cond.getOpcode() == X86ISD::SETCC) {
4154 CC = Cond.getOperand(0);
4156 SDOperand Cmp = Cond.getOperand(1);
4157 unsigned Opc = Cmp.getOpcode();
4158 MVT::ValueType VT = Op.getValueType();
4159 bool IllegalFPCMov = false;
4160 if (VT == MVT::f32 && !X86ScalarSSEf32)
4161 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4162 else if (VT == MVT::f64 && !X86ScalarSSEf64)
4163 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4164 else if (VT == MVT::f80)
4165 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4166 if ((Opc == X86ISD::CMP ||
4167 Opc == X86ISD::COMI ||
4168 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4175 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4176 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4179 const MVT::ValueType *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4181 SmallVector<SDOperand, 4> Ops;
4182 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4183 // condition is true.
4184 Ops.push_back(Op.getOperand(2));
4185 Ops.push_back(Op.getOperand(1));
4187 Ops.push_back(Cond);
4188 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4191 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
4192 bool addTest = true;
4193 SDOperand Chain = Op.getOperand(0);
4194 SDOperand Cond = Op.getOperand(1);
4195 SDOperand Dest = Op.getOperand(2);
4198 if (Cond.getOpcode() == ISD::SETCC)
4199 Cond = LowerSETCC(Cond, DAG);
4201 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4202 // setting operand in place of the X86ISD::SETCC.
4203 if (Cond.getOpcode() == X86ISD::SETCC) {
4204 CC = Cond.getOperand(0);
4206 SDOperand Cmp = Cond.getOperand(1);
4207 unsigned Opc = Cmp.getOpcode();
4208 if (Opc == X86ISD::CMP ||
4209 Opc == X86ISD::COMI ||
4210 Opc == X86ISD::UCOMI) {
4217 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4218 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4220 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
4221 Chain, Op.getOperand(2), CC, Cond);
4224 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
4225 unsigned CallingConv = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4226 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
4228 if (Subtarget->is64Bit())
4229 if(CallingConv==CallingConv::Fast && isTailCall && PerformTailCallOpt)
4230 return LowerX86_TailCallTo(Op, DAG, CallingConv);
4232 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
4234 switch (CallingConv) {
4236 assert(0 && "Unsupported calling convention");
4237 case CallingConv::Fast:
4238 if (isTailCall && PerformTailCallOpt)
4239 return LowerX86_TailCallTo(Op, DAG, CallingConv);
4241 return LowerCCCCallTo(Op,DAG, CallingConv);
4242 case CallingConv::C:
4243 case CallingConv::X86_StdCall:
4244 return LowerCCCCallTo(Op, DAG, CallingConv);
4245 case CallingConv::X86_FastCall:
4246 return LowerFastCCCallTo(Op, DAG, CallingConv);
4251 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
4252 // Calls to _alloca is needed to probe the stack when allocating more than 4k
4253 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
4254 // that the guard pages used by the OS virtual memory manager are allocated in
4255 // correct sequence.
4257 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
4258 SelectionDAG &DAG) {
4259 assert(Subtarget->isTargetCygMing() &&
4260 "This should be used only on Cygwin/Mingw targets");
4263 SDOperand Chain = Op.getOperand(0);
4264 SDOperand Size = Op.getOperand(1);
4265 // FIXME: Ensure alignment here
4269 MVT::ValueType IntPtr = getPointerTy();
4270 MVT::ValueType SPTy = (Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
4272 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
4273 Flag = Chain.getValue(1);
4275 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4276 SDOperand Ops[] = { Chain,
4277 DAG.getTargetExternalSymbol("_alloca", IntPtr),
4278 DAG.getRegister(X86::EAX, IntPtr),
4280 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
4281 Flag = Chain.getValue(1);
4283 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
4285 std::vector<MVT::ValueType> Tys;
4286 Tys.push_back(SPTy);
4287 Tys.push_back(MVT::Other);
4288 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
4289 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
4293 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
4294 MachineFunction &MF = DAG.getMachineFunction();
4295 const Function* Fn = MF.getFunction();
4296 if (Fn->hasExternalLinkage() &&
4297 Subtarget->isTargetCygMing() &&
4298 Fn->getName() == "main")
4299 MF.getInfo<X86MachineFunctionInfo>()->setForceFramePointer(true);
4301 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4302 if (Subtarget->is64Bit())
4303 return LowerX86_64CCCArguments(Op, DAG);
4307 assert(0 && "Unsupported calling convention");
4308 case CallingConv::Fast:
4309 return LowerCCCArguments(Op,DAG, true);
4311 case CallingConv::C:
4312 return LowerCCCArguments(Op, DAG);
4313 case CallingConv::X86_StdCall:
4314 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(StdCall);
4315 return LowerCCCArguments(Op, DAG, true);
4316 case CallingConv::X86_FastCall:
4317 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(FastCall);
4318 return LowerFastCCArguments(Op, DAG);
4322 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4323 SDOperand InFlag(0, 0);
4324 SDOperand Chain = Op.getOperand(0);
4326 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4327 if (Align == 0) Align = 1;
4329 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4330 // If not DWORD aligned or size is more than the threshold, call memset.
4331 // The libc version is likely to be faster for these cases. It can use the
4332 // address value and run time information about the CPU.
4333 if ((Align & 3) != 0 ||
4334 (I && I->getValue() > Subtarget->getMinRepStrSizeThreshold())) {
4335 MVT::ValueType IntPtr = getPointerTy();
4336 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4337 TargetLowering::ArgListTy Args;
4338 TargetLowering::ArgListEntry Entry;
4339 Entry.Node = Op.getOperand(1);
4340 Entry.Ty = IntPtrTy;
4341 Args.push_back(Entry);
4342 // Extend the unsigned i8 argument to be an int value for the call.
4343 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4344 Entry.Ty = IntPtrTy;
4345 Args.push_back(Entry);
4346 Entry.Node = Op.getOperand(3);
4347 Args.push_back(Entry);
4348 std::pair<SDOperand,SDOperand> CallResult =
4349 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4350 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4351 return CallResult.second;
4356 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4357 unsigned BytesLeft = 0;
4358 bool TwoRepStos = false;
4361 uint64_t Val = ValC->getValue() & 255;
4363 // If the value is a constant, then we can potentially use larger sets.
4364 switch (Align & 3) {
4365 case 2: // WORD aligned
4368 Val = (Val << 8) | Val;
4370 case 0: // DWORD aligned
4373 Val = (Val << 8) | Val;
4374 Val = (Val << 16) | Val;
4375 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4378 Val = (Val << 32) | Val;
4381 default: // Byte aligned
4384 Count = Op.getOperand(3);
4388 if (AVT > MVT::i8) {
4390 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4391 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4392 BytesLeft = I->getValue() % UBytes;
4394 assert(AVT >= MVT::i32 &&
4395 "Do not use rep;stos if not at least DWORD aligned");
4396 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4397 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4402 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4404 InFlag = Chain.getValue(1);
4407 Count = Op.getOperand(3);
4408 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4409 InFlag = Chain.getValue(1);
4412 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4414 InFlag = Chain.getValue(1);
4415 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4416 Op.getOperand(1), InFlag);
4417 InFlag = Chain.getValue(1);
4419 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4420 SmallVector<SDOperand, 8> Ops;
4421 Ops.push_back(Chain);
4422 Ops.push_back(DAG.getValueType(AVT));
4423 Ops.push_back(InFlag);
4424 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4427 InFlag = Chain.getValue(1);
4428 Count = Op.getOperand(3);
4429 MVT::ValueType CVT = Count.getValueType();
4430 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4431 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4432 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4434 InFlag = Chain.getValue(1);
4435 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4437 Ops.push_back(Chain);
4438 Ops.push_back(DAG.getValueType(MVT::i8));
4439 Ops.push_back(InFlag);
4440 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4441 } else if (BytesLeft) {
4442 // Issue stores for the last 1 - 7 bytes.
4444 unsigned Val = ValC->getValue() & 255;
4445 unsigned Offset = I->getValue() - BytesLeft;
4446 SDOperand DstAddr = Op.getOperand(1);
4447 MVT::ValueType AddrVT = DstAddr.getValueType();
4448 if (BytesLeft >= 4) {
4449 Val = (Val << 8) | Val;
4450 Val = (Val << 16) | Val;
4451 Value = DAG.getConstant(Val, MVT::i32);
4452 Chain = DAG.getStore(Chain, Value,
4453 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4454 DAG.getConstant(Offset, AddrVT)),
4459 if (BytesLeft >= 2) {
4460 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4461 Chain = DAG.getStore(Chain, Value,
4462 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4463 DAG.getConstant(Offset, AddrVT)),
4468 if (BytesLeft == 1) {
4469 Value = DAG.getConstant(Val, MVT::i8);
4470 Chain = DAG.getStore(Chain, Value,
4471 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4472 DAG.getConstant(Offset, AddrVT)),
4480 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
4481 SDOperand ChainOp = Op.getOperand(0);
4482 SDOperand DestOp = Op.getOperand(1);
4483 SDOperand SourceOp = Op.getOperand(2);
4484 SDOperand CountOp = Op.getOperand(3);
4485 SDOperand AlignOp = Op.getOperand(4);
4486 SDOperand AlwaysInlineOp = Op.getOperand(5);
4488 bool AlwaysInline = (bool)cast<ConstantSDNode>(AlwaysInlineOp)->getValue();
4489 unsigned Align = (unsigned)cast<ConstantSDNode>(AlignOp)->getValue();
4490 if (Align == 0) Align = 1;
4492 // If size is unknown, call memcpy.
4493 ConstantSDNode *I = dyn_cast<ConstantSDNode>(CountOp);
4495 assert(!AlwaysInline && "Cannot inline copy of unknown size");
4496 return LowerMEMCPYCall(ChainOp, DestOp, SourceOp, CountOp, DAG);
4498 unsigned Size = I->getValue();
4501 return LowerMEMCPYInline(ChainOp, DestOp, SourceOp, Size, Align, DAG);
4503 // The libc version is likely to be faster for the following cases. It can
4504 // use the address value and run time information about the CPU.
4505 // With glibc 2.6.1 on a core 2, coping an array of 100M longs was 30% faster
4507 // If not DWORD aligned, call memcpy.
4508 if ((Align & 3) != 0)
4509 return LowerMEMCPYCall(ChainOp, DestOp, SourceOp, CountOp, DAG);
4511 // If size is more than the threshold, call memcpy.
4512 if (Size > Subtarget->getMinRepStrSizeThreshold())
4513 return LowerMEMCPYCall(ChainOp, DestOp, SourceOp, CountOp, DAG);
4515 return LowerMEMCPYInline(ChainOp, DestOp, SourceOp, Size, Align, DAG);
4518 SDOperand X86TargetLowering::LowerMEMCPYCall(SDOperand Chain,
4522 SelectionDAG &DAG) {
4523 MVT::ValueType IntPtr = getPointerTy();
4524 TargetLowering::ArgListTy Args;
4525 TargetLowering::ArgListEntry Entry;
4526 Entry.Ty = getTargetData()->getIntPtrType();
4527 Entry.Node = Dest; Args.push_back(Entry);
4528 Entry.Node = Source; Args.push_back(Entry);
4529 Entry.Node = Count; Args.push_back(Entry);
4530 std::pair<SDOperand,SDOperand> CallResult =
4531 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4532 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
4533 return CallResult.second;
4536 SDOperand X86TargetLowering::LowerMEMCPYInline(SDOperand Chain,
4541 SelectionDAG &DAG) {
4543 unsigned BytesLeft = 0;
4544 switch (Align & 3) {
4545 case 2: // WORD aligned
4548 case 0: // DWORD aligned
4550 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4553 default: // Byte aligned
4558 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4559 SDOperand Count = DAG.getConstant(Size / UBytes, getPointerTy());
4560 BytesLeft = Size % UBytes;
4562 SDOperand InFlag(0, 0);
4563 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4565 InFlag = Chain.getValue(1);
4566 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4568 InFlag = Chain.getValue(1);
4569 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4571 InFlag = Chain.getValue(1);
4573 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4574 SmallVector<SDOperand, 8> Ops;
4575 Ops.push_back(Chain);
4576 Ops.push_back(DAG.getValueType(AVT));
4577 Ops.push_back(InFlag);
4578 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4581 // Issue loads and stores for the last 1 - 7 bytes.
4582 unsigned Offset = Size - BytesLeft;
4583 SDOperand DstAddr = Dest;
4584 MVT::ValueType DstVT = DstAddr.getValueType();
4585 SDOperand SrcAddr = Source;
4586 MVT::ValueType SrcVT = SrcAddr.getValueType();
4588 if (BytesLeft >= 4) {
4589 Value = DAG.getLoad(MVT::i32, Chain,
4590 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4591 DAG.getConstant(Offset, SrcVT)),
4593 Chain = Value.getValue(1);
4594 Chain = DAG.getStore(Chain, Value,
4595 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4596 DAG.getConstant(Offset, DstVT)),
4601 if (BytesLeft >= 2) {
4602 Value = DAG.getLoad(MVT::i16, Chain,
4603 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4604 DAG.getConstant(Offset, SrcVT)),
4606 Chain = Value.getValue(1);
4607 Chain = DAG.getStore(Chain, Value,
4608 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4609 DAG.getConstant(Offset, DstVT)),
4615 if (BytesLeft == 1) {
4616 Value = DAG.getLoad(MVT::i8, Chain,
4617 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4618 DAG.getConstant(Offset, SrcVT)),
4620 Chain = Value.getValue(1);
4621 Chain = DAG.getStore(Chain, Value,
4622 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4623 DAG.getConstant(Offset, DstVT)),
4632 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4633 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4634 SDOperand TheOp = Op.getOperand(0);
4635 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
4636 if (Subtarget->is64Bit()) {
4638 DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4639 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4640 MVT::i64, Copy1.getValue(2));
4641 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4642 DAG.getConstant(32, MVT::i8));
4644 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
4647 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4648 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
4651 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4652 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4653 MVT::i32, Copy1.getValue(2));
4654 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
4655 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4656 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
4659 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4660 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4662 if (!Subtarget->is64Bit()) {
4663 // vastart just stores the address of the VarArgsFrameIndex slot into the
4664 // memory location argument.
4665 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4666 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4671 // gp_offset (0 - 6 * 8)
4672 // fp_offset (48 - 48 + 8 * 16)
4673 // overflow_arg_area (point to parameters coming in memory).
4675 SmallVector<SDOperand, 8> MemOps;
4676 SDOperand FIN = Op.getOperand(1);
4678 SDOperand Store = DAG.getStore(Op.getOperand(0),
4679 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4680 FIN, SV->getValue(), SV->getOffset());
4681 MemOps.push_back(Store);
4684 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4685 DAG.getConstant(4, getPointerTy()));
4686 Store = DAG.getStore(Op.getOperand(0),
4687 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4688 FIN, SV->getValue(), SV->getOffset());
4689 MemOps.push_back(Store);
4691 // Store ptr to overflow_arg_area
4692 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4693 DAG.getConstant(4, getPointerTy()));
4694 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4695 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4697 MemOps.push_back(Store);
4699 // Store ptr to reg_save_area.
4700 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4701 DAG.getConstant(8, getPointerTy()));
4702 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4703 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4705 MemOps.push_back(Store);
4706 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4709 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
4710 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
4711 SDOperand Chain = Op.getOperand(0);
4712 SDOperand DstPtr = Op.getOperand(1);
4713 SDOperand SrcPtr = Op.getOperand(2);
4714 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
4715 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
4717 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
4718 SrcSV->getValue(), SrcSV->getOffset());
4719 Chain = SrcPtr.getValue(1);
4720 for (unsigned i = 0; i < 3; ++i) {
4721 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
4722 SrcSV->getValue(), SrcSV->getOffset());
4723 Chain = Val.getValue(1);
4724 Chain = DAG.getStore(Chain, Val, DstPtr,
4725 DstSV->getValue(), DstSV->getOffset());
4728 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
4729 DAG.getConstant(8, getPointerTy()));
4730 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
4731 DAG.getConstant(8, getPointerTy()));
4737 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4738 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4740 default: return SDOperand(); // Don't custom lower most intrinsics.
4741 // Comparison intrinsics.
4742 case Intrinsic::x86_sse_comieq_ss:
4743 case Intrinsic::x86_sse_comilt_ss:
4744 case Intrinsic::x86_sse_comile_ss:
4745 case Intrinsic::x86_sse_comigt_ss:
4746 case Intrinsic::x86_sse_comige_ss:
4747 case Intrinsic::x86_sse_comineq_ss:
4748 case Intrinsic::x86_sse_ucomieq_ss:
4749 case Intrinsic::x86_sse_ucomilt_ss:
4750 case Intrinsic::x86_sse_ucomile_ss:
4751 case Intrinsic::x86_sse_ucomigt_ss:
4752 case Intrinsic::x86_sse_ucomige_ss:
4753 case Intrinsic::x86_sse_ucomineq_ss:
4754 case Intrinsic::x86_sse2_comieq_sd:
4755 case Intrinsic::x86_sse2_comilt_sd:
4756 case Intrinsic::x86_sse2_comile_sd:
4757 case Intrinsic::x86_sse2_comigt_sd:
4758 case Intrinsic::x86_sse2_comige_sd:
4759 case Intrinsic::x86_sse2_comineq_sd:
4760 case Intrinsic::x86_sse2_ucomieq_sd:
4761 case Intrinsic::x86_sse2_ucomilt_sd:
4762 case Intrinsic::x86_sse2_ucomile_sd:
4763 case Intrinsic::x86_sse2_ucomigt_sd:
4764 case Intrinsic::x86_sse2_ucomige_sd:
4765 case Intrinsic::x86_sse2_ucomineq_sd: {
4767 ISD::CondCode CC = ISD::SETCC_INVALID;
4770 case Intrinsic::x86_sse_comieq_ss:
4771 case Intrinsic::x86_sse2_comieq_sd:
4775 case Intrinsic::x86_sse_comilt_ss:
4776 case Intrinsic::x86_sse2_comilt_sd:
4780 case Intrinsic::x86_sse_comile_ss:
4781 case Intrinsic::x86_sse2_comile_sd:
4785 case Intrinsic::x86_sse_comigt_ss:
4786 case Intrinsic::x86_sse2_comigt_sd:
4790 case Intrinsic::x86_sse_comige_ss:
4791 case Intrinsic::x86_sse2_comige_sd:
4795 case Intrinsic::x86_sse_comineq_ss:
4796 case Intrinsic::x86_sse2_comineq_sd:
4800 case Intrinsic::x86_sse_ucomieq_ss:
4801 case Intrinsic::x86_sse2_ucomieq_sd:
4802 Opc = X86ISD::UCOMI;
4805 case Intrinsic::x86_sse_ucomilt_ss:
4806 case Intrinsic::x86_sse2_ucomilt_sd:
4807 Opc = X86ISD::UCOMI;
4810 case Intrinsic::x86_sse_ucomile_ss:
4811 case Intrinsic::x86_sse2_ucomile_sd:
4812 Opc = X86ISD::UCOMI;
4815 case Intrinsic::x86_sse_ucomigt_ss:
4816 case Intrinsic::x86_sse2_ucomigt_sd:
4817 Opc = X86ISD::UCOMI;
4820 case Intrinsic::x86_sse_ucomige_ss:
4821 case Intrinsic::x86_sse2_ucomige_sd:
4822 Opc = X86ISD::UCOMI;
4825 case Intrinsic::x86_sse_ucomineq_ss:
4826 case Intrinsic::x86_sse2_ucomineq_sd:
4827 Opc = X86ISD::UCOMI;
4833 SDOperand LHS = Op.getOperand(1);
4834 SDOperand RHS = Op.getOperand(2);
4835 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4837 SDOperand Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
4838 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
4839 DAG.getConstant(X86CC, MVT::i8), Cond);
4840 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4845 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4846 // Depths > 0 not supported yet!
4847 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4850 // Just load the return address
4851 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4852 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4855 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4856 // Depths > 0 not supported yet!
4857 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4860 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4861 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4862 DAG.getConstant(4, getPointerTy()));
4865 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
4866 SelectionDAG &DAG) {
4867 // Is not yet supported on x86-64
4868 if (Subtarget->is64Bit())
4871 return DAG.getConstant(8, getPointerTy());
4874 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
4876 assert(!Subtarget->is64Bit() &&
4877 "Lowering of eh_return builtin is not supported yet on x86-64");
4879 MachineFunction &MF = DAG.getMachineFunction();
4880 SDOperand Chain = Op.getOperand(0);
4881 SDOperand Offset = Op.getOperand(1);
4882 SDOperand Handler = Op.getOperand(2);
4884 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
4887 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
4888 DAG.getConstant(-4UL, getPointerTy()));
4889 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
4890 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
4891 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
4892 MF.addLiveOut(X86::ECX);
4894 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
4895 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
4898 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
4899 SelectionDAG &DAG) {
4900 SDOperand Root = Op.getOperand(0);
4901 SDOperand Trmp = Op.getOperand(1); // trampoline
4902 SDOperand FPtr = Op.getOperand(2); // nested function
4903 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
4905 SrcValueSDNode *TrmpSV = cast<SrcValueSDNode>(Op.getOperand(4));
4907 if (Subtarget->is64Bit()) {
4908 return SDOperand(); // not yet supported
4910 Function *Func = (Function *)
4911 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
4912 unsigned CC = Func->getCallingConv();
4917 assert(0 && "Unsupported calling convention");
4918 case CallingConv::C:
4919 case CallingConv::X86_StdCall: {
4920 // Pass 'nest' parameter in ECX.
4921 // Must be kept in sync with X86CallingConv.td
4924 // Check that ECX wasn't needed by an 'inreg' parameter.
4925 const FunctionType *FTy = Func->getFunctionType();
4926 const ParamAttrsList *Attrs = FTy->getParamAttrs();
4928 if (Attrs && !Func->isVarArg()) {
4929 unsigned InRegCount = 0;
4932 for (FunctionType::param_iterator I = FTy->param_begin(),
4933 E = FTy->param_end(); I != E; ++I, ++Idx)
4934 if (Attrs->paramHasAttr(Idx, ParamAttr::InReg))
4935 // FIXME: should only count parameters that are lowered to integers.
4936 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
4938 if (InRegCount > 2) {
4939 cerr << "Nest register in use - reduce number of inreg parameters!\n";
4945 case CallingConv::X86_FastCall:
4946 // Pass 'nest' parameter in EAX.
4947 // Must be kept in sync with X86CallingConv.td
4952 const X86InstrInfo *TII =
4953 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
4955 SDOperand OutChains[4];
4956 SDOperand Addr, Disp;
4958 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
4959 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
4961 unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
4962 unsigned char N86Reg = ((X86RegisterInfo&)RegInfo).getX86RegNum(NestReg);
4963 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
4964 Trmp, TrmpSV->getValue(), TrmpSV->getOffset());
4966 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
4967 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
4968 TrmpSV->getOffset() + 1, false, 1);
4970 unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
4971 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
4972 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
4973 TrmpSV->getValue() + 5, TrmpSV->getOffset());
4975 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
4976 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpSV->getValue(),
4977 TrmpSV->getOffset() + 6, false, 1);
4980 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
4981 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
4985 /// LowerOperation - Provide custom lowering hooks for some operations.
4987 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4988 switch (Op.getOpcode()) {
4989 default: assert(0 && "Should not custom lower this!");
4990 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4991 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4992 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4993 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4994 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4995 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4996 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4997 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4998 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4999 case ISD::SHL_PARTS:
5000 case ISD::SRA_PARTS:
5001 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5002 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5003 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5004 case ISD::FABS: return LowerFABS(Op, DAG);
5005 case ISD::FNEG: return LowerFNEG(Op, DAG);
5006 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5007 case ISD::SETCC: return LowerSETCC(Op, DAG);
5008 case ISD::SELECT: return LowerSELECT(Op, DAG);
5009 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
5010 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5011 case ISD::CALL: return LowerCALL(Op, DAG);
5012 case ISD::RET: return LowerRET(Op, DAG);
5013 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
5014 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
5015 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
5016 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
5017 case ISD::VASTART: return LowerVASTART(Op, DAG);
5018 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
5019 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5020 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5021 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5022 case ISD::FRAME_TO_ARGS_OFFSET:
5023 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
5024 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
5025 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
5026 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
5031 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
5033 default: return NULL;
5034 case X86ISD::SHLD: return "X86ISD::SHLD";
5035 case X86ISD::SHRD: return "X86ISD::SHRD";
5036 case X86ISD::FAND: return "X86ISD::FAND";
5037 case X86ISD::FOR: return "X86ISD::FOR";
5038 case X86ISD::FXOR: return "X86ISD::FXOR";
5039 case X86ISD::FSRL: return "X86ISD::FSRL";
5040 case X86ISD::FILD: return "X86ISD::FILD";
5041 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
5042 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
5043 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
5044 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
5045 case X86ISD::FLD: return "X86ISD::FLD";
5046 case X86ISD::FST: return "X86ISD::FST";
5047 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
5048 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
5049 case X86ISD::CALL: return "X86ISD::CALL";
5050 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
5051 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
5052 case X86ISD::CMP: return "X86ISD::CMP";
5053 case X86ISD::COMI: return "X86ISD::COMI";
5054 case X86ISD::UCOMI: return "X86ISD::UCOMI";
5055 case X86ISD::SETCC: return "X86ISD::SETCC";
5056 case X86ISD::CMOV: return "X86ISD::CMOV";
5057 case X86ISD::BRCOND: return "X86ISD::BRCOND";
5058 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
5059 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
5060 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
5061 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
5062 case X86ISD::Wrapper: return "X86ISD::Wrapper";
5063 case X86ISD::S2VEC: return "X86ISD::S2VEC";
5064 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
5065 case X86ISD::PINSRW: return "X86ISD::PINSRW";
5066 case X86ISD::FMAX: return "X86ISD::FMAX";
5067 case X86ISD::FMIN: return "X86ISD::FMIN";
5068 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
5069 case X86ISD::FRCP: return "X86ISD::FRCP";
5070 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
5071 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
5072 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
5073 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
5077 // isLegalAddressingMode - Return true if the addressing mode represented
5078 // by AM is legal for this target, for a load/store of the specified type.
5079 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5080 const Type *Ty) const {
5081 // X86 supports extremely general addressing modes.
5083 // X86 allows a sign-extended 32-bit immediate field as a displacement.
5084 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
5088 // We can only fold this if we don't need an extra load.
5089 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
5092 // X86-64 only supports addr of globals in small code model.
5093 if (Subtarget->is64Bit()) {
5094 if (getTargetMachine().getCodeModel() != CodeModel::Small)
5096 // If lower 4G is not available, then we must use rip-relative addressing.
5097 if (AM.BaseOffs || AM.Scale > 1)
5108 // These scales always work.
5113 // These scales are formed with basereg+scalereg. Only accept if there is
5118 default: // Other stuff never works.
5126 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
5127 if (!Ty1->isInteger() || !Ty2->isInteger())
5129 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5130 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5131 if (NumBits1 <= NumBits2)
5133 return Subtarget->is64Bit() || NumBits1 < 64;
5136 bool X86TargetLowering::isTruncateFree(MVT::ValueType VT1,
5137 MVT::ValueType VT2) const {
5138 if (!MVT::isInteger(VT1) || !MVT::isInteger(VT2))
5140 unsigned NumBits1 = MVT::getSizeInBits(VT1);
5141 unsigned NumBits2 = MVT::getSizeInBits(VT2);
5142 if (NumBits1 <= NumBits2)
5144 return Subtarget->is64Bit() || NumBits1 < 64;
5147 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5148 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5149 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5150 /// are assumed to be legal.
5152 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
5153 // Only do shuffles on 128-bit vector types for now.
5154 if (MVT::getSizeInBits(VT) == 64) return false;
5155 return (Mask.Val->getNumOperands() <= 4 ||
5156 isIdentityMask(Mask.Val) ||
5157 isIdentityMask(Mask.Val, true) ||
5158 isSplatMask(Mask.Val) ||
5159 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
5160 X86::isUNPCKLMask(Mask.Val) ||
5161 X86::isUNPCKHMask(Mask.Val) ||
5162 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
5163 X86::isUNPCKH_v_undef_Mask(Mask.Val));
5166 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
5168 SelectionDAG &DAG) const {
5169 unsigned NumElts = BVOps.size();
5170 // Only do shuffles on 128-bit vector types for now.
5171 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
5172 if (NumElts == 2) return true;
5174 return (isMOVLMask(&BVOps[0], 4) ||
5175 isCommutedMOVL(&BVOps[0], 4, true) ||
5176 isSHUFPMask(&BVOps[0], 4) ||
5177 isCommutedSHUFP(&BVOps[0], 4));
5182 //===----------------------------------------------------------------------===//
5183 // X86 Scheduler Hooks
5184 //===----------------------------------------------------------------------===//
5187 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
5188 MachineBasicBlock *BB) {
5189 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5190 switch (MI->getOpcode()) {
5191 default: assert(false && "Unexpected instr type to insert");
5192 case X86::CMOV_FR32:
5193 case X86::CMOV_FR64:
5194 case X86::CMOV_V4F32:
5195 case X86::CMOV_V2F64:
5196 case X86::CMOV_V2I64: {
5197 // To "insert" a SELECT_CC instruction, we actually have to insert the
5198 // diamond control-flow pattern. The incoming instruction knows the
5199 // destination vreg to set, the condition code register to branch on, the
5200 // true/false values to select between, and a branch opcode to use.
5201 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5202 ilist<MachineBasicBlock>::iterator It = BB;
5208 // cmpTY ccX, r1, r2
5210 // fallthrough --> copy0MBB
5211 MachineBasicBlock *thisMBB = BB;
5212 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
5213 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
5215 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
5216 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
5217 MachineFunction *F = BB->getParent();
5218 F->getBasicBlockList().insert(It, copy0MBB);
5219 F->getBasicBlockList().insert(It, sinkMBB);
5220 // Update machine-CFG edges by first adding all successors of the current
5221 // block to the new block which will contain the Phi node for the select.
5222 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
5223 e = BB->succ_end(); i != e; ++i)
5224 sinkMBB->addSuccessor(*i);
5225 // Next, remove all successors of the current block, and add the true
5226 // and fallthrough blocks as its successors.
5227 while(!BB->succ_empty())
5228 BB->removeSuccessor(BB->succ_begin());
5229 BB->addSuccessor(copy0MBB);
5230 BB->addSuccessor(sinkMBB);
5233 // %FalseValue = ...
5234 // # fallthrough to sinkMBB
5237 // Update machine-CFG edges
5238 BB->addSuccessor(sinkMBB);
5241 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5244 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
5245 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5246 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5248 delete MI; // The pseudo instruction is gone now.
5252 case X86::FP32_TO_INT16_IN_MEM:
5253 case X86::FP32_TO_INT32_IN_MEM:
5254 case X86::FP32_TO_INT64_IN_MEM:
5255 case X86::FP64_TO_INT16_IN_MEM:
5256 case X86::FP64_TO_INT32_IN_MEM:
5257 case X86::FP64_TO_INT64_IN_MEM:
5258 case X86::FP80_TO_INT16_IN_MEM:
5259 case X86::FP80_TO_INT32_IN_MEM:
5260 case X86::FP80_TO_INT64_IN_MEM: {
5261 // Change the floating point control register to use "round towards zero"
5262 // mode when truncating to an integer value.
5263 MachineFunction *F = BB->getParent();
5264 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
5265 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
5267 // Load the old value of the high byte of the control word...
5269 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
5270 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
5272 // Set the high part to be round to zero...
5273 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
5276 // Reload the modified control word now...
5277 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5279 // Restore the memory image of control word to original value
5280 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
5283 // Get the X86 opcode to use.
5285 switch (MI->getOpcode()) {
5286 default: assert(0 && "illegal opcode!");
5287 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
5288 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
5289 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
5290 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
5291 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
5292 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
5293 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
5294 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
5295 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
5299 MachineOperand &Op = MI->getOperand(0);
5300 if (Op.isRegister()) {
5301 AM.BaseType = X86AddressMode::RegBase;
5302 AM.Base.Reg = Op.getReg();
5304 AM.BaseType = X86AddressMode::FrameIndexBase;
5305 AM.Base.FrameIndex = Op.getFrameIndex();
5307 Op = MI->getOperand(1);
5308 if (Op.isImmediate())
5309 AM.Scale = Op.getImm();
5310 Op = MI->getOperand(2);
5311 if (Op.isImmediate())
5312 AM.IndexReg = Op.getImm();
5313 Op = MI->getOperand(3);
5314 if (Op.isGlobalAddress()) {
5315 AM.GV = Op.getGlobal();
5317 AM.Disp = Op.getImm();
5319 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
5320 .addReg(MI->getOperand(4).getReg());
5322 // Reload the original control word now.
5323 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5325 delete MI; // The pseudo instruction is gone now.
5331 //===----------------------------------------------------------------------===//
5332 // X86 Optimization Hooks
5333 //===----------------------------------------------------------------------===//
5335 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
5337 uint64_t &KnownZero,
5339 const SelectionDAG &DAG,
5340 unsigned Depth) const {
5341 unsigned Opc = Op.getOpcode();
5342 assert((Opc >= ISD::BUILTIN_OP_END ||
5343 Opc == ISD::INTRINSIC_WO_CHAIN ||
5344 Opc == ISD::INTRINSIC_W_CHAIN ||
5345 Opc == ISD::INTRINSIC_VOID) &&
5346 "Should use MaskedValueIsZero if you don't know whether Op"
5347 " is a target node!");
5349 KnownZero = KnownOne = 0; // Don't know anything.
5353 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
5358 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5359 /// element of the result of the vector shuffle.
5360 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
5361 MVT::ValueType VT = N->getValueType(0);
5362 SDOperand PermMask = N->getOperand(2);
5363 unsigned NumElems = PermMask.getNumOperands();
5364 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
5366 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5368 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5369 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
5370 SDOperand Idx = PermMask.getOperand(i);
5371 if (Idx.getOpcode() == ISD::UNDEF)
5372 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5373 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
5378 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
5379 /// node is a GlobalAddress + an offset.
5380 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
5381 unsigned Opc = N->getOpcode();
5382 if (Opc == X86ISD::Wrapper) {
5383 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
5384 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
5387 } else if (Opc == ISD::ADD) {
5388 SDOperand N1 = N->getOperand(0);
5389 SDOperand N2 = N->getOperand(1);
5390 if (isGAPlusOffset(N1.Val, GA, Offset)) {
5391 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
5393 Offset += V->getSignExtended();
5396 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
5397 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
5399 Offset += V->getSignExtended();
5407 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
5409 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
5410 MachineFrameInfo *MFI) {
5411 if (N->getOperand(0).Val != Base->getOperand(0).Val)
5414 SDOperand Loc = N->getOperand(1);
5415 SDOperand BaseLoc = Base->getOperand(1);
5416 if (Loc.getOpcode() == ISD::FrameIndex) {
5417 if (BaseLoc.getOpcode() != ISD::FrameIndex)
5419 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
5420 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
5421 int FS = MFI->getObjectSize(FI);
5422 int BFS = MFI->getObjectSize(BFI);
5423 if (FS != BFS || FS != Size) return false;
5424 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
5426 GlobalValue *GV1 = NULL;
5427 GlobalValue *GV2 = NULL;
5428 int64_t Offset1 = 0;
5429 int64_t Offset2 = 0;
5430 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
5431 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
5432 if (isGA1 && isGA2 && GV1 == GV2)
5433 return Offset1 == (Offset2 + Dist*Size);
5439 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
5440 const X86Subtarget *Subtarget) {
5443 if (isGAPlusOffset(Base, GV, Offset))
5444 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
5446 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
5447 int BFI = cast<FrameIndexSDNode>(Base)->getIndex();
5449 // Fixed objects do not specify alignment, however the offsets are known.
5450 return ((Subtarget->getStackAlignment() % 16) == 0 &&
5451 (MFI->getObjectOffset(BFI) % 16) == 0);
5453 return MFI->getObjectAlignment(BFI) >= 16;
5459 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
5460 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
5461 /// if the load addresses are consecutive, non-overlapping, and in the right
5463 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
5464 const X86Subtarget *Subtarget) {
5465 MachineFunction &MF = DAG.getMachineFunction();
5466 MachineFrameInfo *MFI = MF.getFrameInfo();
5467 MVT::ValueType VT = N->getValueType(0);
5468 MVT::ValueType EVT = MVT::getVectorElementType(VT);
5469 SDOperand PermMask = N->getOperand(2);
5470 int NumElems = (int)PermMask.getNumOperands();
5471 SDNode *Base = NULL;
5472 for (int i = 0; i < NumElems; ++i) {
5473 SDOperand Idx = PermMask.getOperand(i);
5474 if (Idx.getOpcode() == ISD::UNDEF) {
5475 if (!Base) return SDOperand();
5478 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
5479 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
5483 else if (!isConsecutiveLoad(Arg.Val, Base,
5484 i, MVT::getSizeInBits(EVT)/8,MFI))
5489 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
5490 LoadSDNode *LD = cast<LoadSDNode>(Base);
5492 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5493 LD->getSrcValueOffset(), LD->isVolatile());
5495 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5496 LD->getSrcValueOffset(), LD->isVolatile(),
5497 LD->getAlignment());
5501 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
5502 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
5503 const X86Subtarget *Subtarget) {
5504 SDOperand Cond = N->getOperand(0);
5506 // If we have SSE[12] support, try to form min/max nodes.
5507 if (Subtarget->hasSSE2() &&
5508 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
5509 if (Cond.getOpcode() == ISD::SETCC) {
5510 // Get the LHS/RHS of the select.
5511 SDOperand LHS = N->getOperand(1);
5512 SDOperand RHS = N->getOperand(2);
5513 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
5515 unsigned Opcode = 0;
5516 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
5519 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
5522 if (!UnsafeFPMath) break;
5524 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
5526 Opcode = X86ISD::FMIN;
5529 case ISD::SETOGT: // (X > Y) ? X : Y -> max
5532 if (!UnsafeFPMath) break;
5534 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
5536 Opcode = X86ISD::FMAX;
5539 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
5542 case ISD::SETOGT: // (X > Y) ? Y : X -> min
5545 if (!UnsafeFPMath) break;
5547 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
5549 Opcode = X86ISD::FMIN;
5552 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
5555 if (!UnsafeFPMath) break;
5557 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
5559 Opcode = X86ISD::FMAX;
5565 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
5574 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
5575 DAGCombinerInfo &DCI) const {
5576 SelectionDAG &DAG = DCI.DAG;
5577 switch (N->getOpcode()) {
5579 case ISD::VECTOR_SHUFFLE:
5580 return PerformShuffleCombine(N, DAG, Subtarget);
5582 return PerformSELECTCombine(N, DAG, Subtarget);
5588 //===----------------------------------------------------------------------===//
5589 // X86 Inline Assembly Support
5590 //===----------------------------------------------------------------------===//
5592 /// getConstraintType - Given a constraint letter, return the type of
5593 /// constraint it is for this target.
5594 X86TargetLowering::ConstraintType
5595 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
5596 if (Constraint.size() == 1) {
5597 switch (Constraint[0]) {
5606 return C_RegisterClass;
5611 return TargetLowering::getConstraintType(Constraint);
5614 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5615 /// vector. If it is invalid, don't add anything to Ops.
5616 void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
5618 std::vector<SDOperand>&Ops,
5619 SelectionDAG &DAG) {
5620 SDOperand Result(0, 0);
5622 switch (Constraint) {
5625 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5626 if (C->getValue() <= 31) {
5627 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5633 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5634 if (C->getValue() <= 255) {
5635 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5641 // Literal immediates are always ok.
5642 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
5643 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
5647 // If we are in non-pic codegen mode, we allow the address of a global (with
5648 // an optional displacement) to be used with 'i'.
5649 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
5652 // Match either (GA) or (GA+C)
5654 Offset = GA->getOffset();
5655 } else if (Op.getOpcode() == ISD::ADD) {
5656 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5657 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5659 Offset = GA->getOffset()+C->getValue();
5661 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5662 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5664 Offset = GA->getOffset()+C->getValue();
5671 // If addressing this global requires a load (e.g. in PIC mode), we can't
5673 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
5677 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5683 // Otherwise, not valid for this mode.
5689 Ops.push_back(Result);
5692 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5695 std::vector<unsigned> X86TargetLowering::
5696 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5697 MVT::ValueType VT) const {
5698 if (Constraint.size() == 1) {
5699 // FIXME: not handling fp-stack yet!
5700 switch (Constraint[0]) { // GCC X86 Constraint Letters
5701 default: break; // Unknown constraint letter
5702 case 'A': // EAX/EDX
5703 if (VT == MVT::i32 || VT == MVT::i64)
5704 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5706 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5709 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5710 else if (VT == MVT::i16)
5711 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5712 else if (VT == MVT::i8)
5713 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5718 return std::vector<unsigned>();
5721 std::pair<unsigned, const TargetRegisterClass*>
5722 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5723 MVT::ValueType VT) const {
5724 // First, see if this is a constraint that directly corresponds to an LLVM
5726 if (Constraint.size() == 1) {
5727 // GCC Constraint Letters
5728 switch (Constraint[0]) {
5730 case 'r': // GENERAL_REGS
5731 case 'R': // LEGACY_REGS
5732 case 'l': // INDEX_REGS
5733 if (VT == MVT::i64 && Subtarget->is64Bit())
5734 return std::make_pair(0U, X86::GR64RegisterClass);
5736 return std::make_pair(0U, X86::GR32RegisterClass);
5737 else if (VT == MVT::i16)
5738 return std::make_pair(0U, X86::GR16RegisterClass);
5739 else if (VT == MVT::i8)
5740 return std::make_pair(0U, X86::GR8RegisterClass);
5742 case 'y': // MMX_REGS if MMX allowed.
5743 if (!Subtarget->hasMMX()) break;
5744 return std::make_pair(0U, X86::VR64RegisterClass);
5746 case 'Y': // SSE_REGS if SSE2 allowed
5747 if (!Subtarget->hasSSE2()) break;
5749 case 'x': // SSE_REGS if SSE1 allowed
5750 if (!Subtarget->hasSSE1()) break;
5754 // Scalar SSE types.
5757 return std::make_pair(0U, X86::FR32RegisterClass);
5760 return std::make_pair(0U, X86::FR64RegisterClass);
5768 return std::make_pair(0U, X86::VR128RegisterClass);
5774 // Use the default implementation in TargetLowering to convert the register
5775 // constraint into a member of a register class.
5776 std::pair<unsigned, const TargetRegisterClass*> Res;
5777 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5779 // Not found as a standard register?
5780 if (Res.second == 0) {
5781 // GCC calls "st(0)" just plain "st".
5782 if (StringsEqualNoCase("{st}", Constraint)) {
5783 Res.first = X86::ST0;
5784 Res.second = X86::RFP80RegisterClass;
5790 // Otherwise, check to see if this is a register class of the wrong value
5791 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5792 // turn into {ax},{dx}.
5793 if (Res.second->hasType(VT))
5794 return Res; // Correct type already, nothing to do.
5796 // All of the single-register GCC register classes map their values onto
5797 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5798 // really want an 8-bit or 32-bit register, map to the appropriate register
5799 // class and return the appropriate register.
5800 if (Res.second != X86::GR16RegisterClass)
5803 if (VT == MVT::i8) {
5804 unsigned DestReg = 0;
5805 switch (Res.first) {
5807 case X86::AX: DestReg = X86::AL; break;
5808 case X86::DX: DestReg = X86::DL; break;
5809 case X86::CX: DestReg = X86::CL; break;
5810 case X86::BX: DestReg = X86::BL; break;
5813 Res.first = DestReg;
5814 Res.second = Res.second = X86::GR8RegisterClass;
5816 } else if (VT == MVT::i32) {
5817 unsigned DestReg = 0;
5818 switch (Res.first) {
5820 case X86::AX: DestReg = X86::EAX; break;
5821 case X86::DX: DestReg = X86::EDX; break;
5822 case X86::CX: DestReg = X86::ECX; break;
5823 case X86::BX: DestReg = X86::EBX; break;
5824 case X86::SI: DestReg = X86::ESI; break;
5825 case X86::DI: DestReg = X86::EDI; break;
5826 case X86::BP: DestReg = X86::EBP; break;
5827 case X86::SP: DestReg = X86::ESP; break;
5830 Res.first = DestReg;
5831 Res.second = Res.second = X86::GR32RegisterClass;
5833 } else if (VT == MVT::i64) {
5834 unsigned DestReg = 0;
5835 switch (Res.first) {
5837 case X86::AX: DestReg = X86::RAX; break;
5838 case X86::DX: DestReg = X86::RDX; break;
5839 case X86::CX: DestReg = X86::RCX; break;
5840 case X86::BX: DestReg = X86::RBX; break;
5841 case X86::SI: DestReg = X86::RSI; break;
5842 case X86::DI: DestReg = X86::RDI; break;
5843 case X86::BP: DestReg = X86::RBP; break;
5844 case X86::SP: DestReg = X86::RSP; break;
5847 Res.first = DestReg;
5848 Res.second = Res.second = X86::GR64RegisterClass;