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);
502 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
503 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
505 setOperationAction(ISD::AND, MVT::v8i8, Promote);
506 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
507 setOperationAction(ISD::AND, MVT::v4i16, Promote);
508 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
509 setOperationAction(ISD::AND, MVT::v2i32, Promote);
510 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
511 setOperationAction(ISD::AND, MVT::v1i64, Legal);
513 setOperationAction(ISD::OR, MVT::v8i8, Promote);
514 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
515 setOperationAction(ISD::OR, MVT::v4i16, Promote);
516 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
517 setOperationAction(ISD::OR, MVT::v2i32, Promote);
518 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
519 setOperationAction(ISD::OR, MVT::v1i64, Legal);
521 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
522 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
523 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
524 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
525 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
526 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
527 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
529 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
530 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
531 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
532 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
533 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
534 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
535 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
537 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
538 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
539 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
540 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
542 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
543 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
544 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
545 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
547 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
548 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
549 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Custom);
550 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
553 if (Subtarget->hasSSE1()) {
554 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
556 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
557 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
558 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
559 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
560 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
561 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
562 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
563 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
564 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
565 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
566 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
569 if (Subtarget->hasSSE2()) {
570 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
571 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
572 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
573 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
574 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
576 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
577 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
578 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
579 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
580 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
581 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
582 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
583 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
584 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
585 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
586 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
587 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
588 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
589 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
590 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
592 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
593 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
594 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
595 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
596 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
597 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
599 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
600 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
601 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
602 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
603 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
605 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
606 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
607 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
608 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
609 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
610 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
612 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
613 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
614 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
615 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
616 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
617 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
618 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
619 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
620 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
621 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
622 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
623 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
626 // Custom lower v2i64 and v2f64 selects.
627 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
628 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
629 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
630 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
633 // We want to custom lower some of our intrinsics.
634 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
636 // We have target-specific dag combine patterns for the following nodes:
637 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
638 setTargetDAGCombine(ISD::SELECT);
640 computeRegisterProperties();
642 // FIXME: These should be based on subtarget info. Plus, the values should
643 // be smaller when we are in optimizing for size mode.
644 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
645 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
646 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
647 allowUnalignedMemoryAccesses = true; // x86 supports it!
651 //===----------------------------------------------------------------------===//
652 // Return Value Calling Convention Implementation
653 //===----------------------------------------------------------------------===//
655 #include "X86GenCallingConv.inc"
657 /// GetPossiblePreceedingTailCall - Get preceeding X86ISD::TAILCALL node if it
658 /// exists skip possible ISD:TokenFactor.
659 static SDOperand GetPossiblePreceedingTailCall(SDOperand Chain) {
660 if (Chain.getOpcode()==X86ISD::TAILCALL) {
662 } else if (Chain.getOpcode()==ISD::TokenFactor) {
663 if (Chain.getNumOperands() &&
664 Chain.getOperand(0).getOpcode()==X86ISD::TAILCALL)
665 return Chain.getOperand(0);
670 /// LowerRET - Lower an ISD::RET node.
671 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
672 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
674 SmallVector<CCValAssign, 16> RVLocs;
675 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
676 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
677 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
678 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
680 // If this is the first return lowered for this function, add the regs to the
681 // liveout set for the function.
682 if (DAG.getMachineFunction().liveout_empty()) {
683 for (unsigned i = 0; i != RVLocs.size(); ++i)
684 if (RVLocs[i].isRegLoc())
685 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
687 SDOperand Chain = Op.getOperand(0);
689 // Handle tail call return.
690 Chain = GetPossiblePreceedingTailCall(Chain);
691 if (Chain.getOpcode() == X86ISD::TAILCALL) {
692 SDOperand TailCall = Chain;
693 SDOperand TargetAddress = TailCall.getOperand(1);
694 SDOperand StackAdjustment = TailCall.getOperand(2);
695 assert ( ((TargetAddress.getOpcode() == ISD::Register &&
696 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
697 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
698 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
699 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
700 "Expecting an global address, external symbol, or register");
701 assert( StackAdjustment.getOpcode() == ISD::Constant &&
702 "Expecting a const value");
704 SmallVector<SDOperand,8> Operands;
705 Operands.push_back(Chain.getOperand(0));
706 Operands.push_back(TargetAddress);
707 Operands.push_back(StackAdjustment);
708 // Copy registers used by the call. Last operand is a flag so it is not
710 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
711 Operands.push_back(Chain.getOperand(i));
713 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
720 // Copy the result values into the output registers.
721 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
722 RVLocs[0].getLocReg() != X86::ST0) {
723 for (unsigned i = 0; i != RVLocs.size(); ++i) {
724 CCValAssign &VA = RVLocs[i];
725 assert(VA.isRegLoc() && "Can only return in registers!");
726 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
728 Flag = Chain.getValue(1);
731 // We need to handle a destination of ST0 specially, because it isn't really
733 SDOperand Value = Op.getOperand(1);
735 // If this is an FP return with ScalarSSE, we need to move the value from
736 // an XMM register onto the fp-stack.
737 if ((X86ScalarSSEf32 && RVLocs[0].getValVT()==MVT::f32) ||
738 (X86ScalarSSEf64 && RVLocs[0].getValVT()==MVT::f64)) {
741 // If this is a load into a scalarsse value, don't store the loaded value
742 // back to the stack, only to reload it: just replace the scalar-sse load.
743 if (ISD::isNON_EXTLoad(Value.Val) &&
744 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
745 Chain = Value.getOperand(0);
746 MemLoc = Value.getOperand(1);
748 // Spill the value to memory and reload it into top of stack.
749 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
750 MachineFunction &MF = DAG.getMachineFunction();
751 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
752 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
753 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
755 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other);
756 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
757 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
758 Chain = Value.getValue(1);
761 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
762 SDOperand Ops[] = { Chain, Value };
763 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
764 Flag = Chain.getValue(1);
767 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
769 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
771 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
775 /// LowerCallResult - Lower the result values of an ISD::CALL into the
776 /// appropriate copies out of appropriate physical registers. This assumes that
777 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
778 /// being lowered. The returns a SDNode with the same number of values as the
780 SDNode *X86TargetLowering::
781 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
782 unsigned CallingConv, SelectionDAG &DAG) {
784 // Assign locations to each value returned by this call.
785 SmallVector<CCValAssign, 16> RVLocs;
786 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
787 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
788 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
791 SmallVector<SDOperand, 8> ResultVals;
793 // Copy all of the result registers out of their specified physreg.
794 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
795 for (unsigned i = 0; i != RVLocs.size(); ++i) {
796 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
797 RVLocs[i].getValVT(), InFlag).getValue(1);
798 InFlag = Chain.getValue(2);
799 ResultVals.push_back(Chain.getValue(0));
802 // Copies from the FP stack are special, as ST0 isn't a valid register
803 // before the fp stackifier runs.
805 // Copy ST0 into an RFP register with FP_GET_RESULT.
806 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other, MVT::Flag);
807 SDOperand GROps[] = { Chain, InFlag };
808 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
809 Chain = RetVal.getValue(1);
810 InFlag = RetVal.getValue(2);
812 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
814 if ((X86ScalarSSEf32 && RVLocs[0].getValVT() == MVT::f32) ||
815 (X86ScalarSSEf64 && RVLocs[0].getValVT() == MVT::f64)) {
816 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
817 // shouldn't be necessary except that RFP cannot be live across
818 // multiple blocks. When stackifier is fixed, they can be uncoupled.
819 MachineFunction &MF = DAG.getMachineFunction();
820 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
821 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
823 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag
825 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
826 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
827 Chain = RetVal.getValue(1);
829 ResultVals.push_back(RetVal);
832 // Merge everything together with a MERGE_VALUES node.
833 ResultVals.push_back(Chain);
834 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
835 &ResultVals[0], ResultVals.size()).Val;
839 //===----------------------------------------------------------------------===//
840 // C & StdCall & Fast Calling Convention implementation
841 //===----------------------------------------------------------------------===//
842 // StdCall calling convention seems to be standard for many Windows' API
843 // routines and around. It differs from C calling convention just a little:
844 // callee should clean up the stack, not caller. Symbols should be also
845 // decorated in some fancy way :) It doesn't support any vector arguments.
846 // For info on fast calling convention see Fast Calling Convention (tail call)
847 // implementation LowerX86_32FastCCCallTo.
849 /// AddLiveIn - This helper function adds the specified physical register to the
850 /// MachineFunction as a live in value. It also creates a corresponding virtual
852 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
853 const TargetRegisterClass *RC) {
854 assert(RC->contains(PReg) && "Not the correct regclass!");
855 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
856 MF.addLiveIn(PReg, VReg);
860 // align stack arguments according to platform alignment needed for tail calls
861 unsigned GetAlignedArgumentStackSize(unsigned StackSize, SelectionDAG& DAG);
863 SDOperand X86TargetLowering::LowerMemArgument(SDOperand Op, SelectionDAG &DAG,
864 const CCValAssign &VA,
865 MachineFrameInfo *MFI,
866 SDOperand Root, unsigned i) {
867 // Create the nodes corresponding to a load from this parameter slot.
868 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
869 VA.getLocMemOffset());
870 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
872 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3 + i))->getValue();
874 if (Flags & ISD::ParamFlags::ByVal)
877 return DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0);
880 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
882 unsigned NumArgs = Op.Val->getNumValues() - 1;
883 MachineFunction &MF = DAG.getMachineFunction();
884 MachineFrameInfo *MFI = MF.getFrameInfo();
885 SDOperand Root = Op.getOperand(0);
886 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
887 unsigned CC = MF.getFunction()->getCallingConv();
888 // Assign locations to all of the incoming arguments.
889 SmallVector<CCValAssign, 16> ArgLocs;
890 CCState CCInfo(CC, isVarArg,
891 getTargetMachine(), ArgLocs);
892 // Check for possible tail call calling convention.
893 if (CC == CallingConv::Fast && PerformTailCallOpt)
894 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_TailCall);
896 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
898 SmallVector<SDOperand, 8> ArgValues;
899 unsigned LastVal = ~0U;
900 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
901 CCValAssign &VA = ArgLocs[i];
902 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
904 assert(VA.getValNo() != LastVal &&
905 "Don't support value assigned to multiple locs yet");
906 LastVal = VA.getValNo();
909 MVT::ValueType RegVT = VA.getLocVT();
910 TargetRegisterClass *RC;
911 if (RegVT == MVT::i32)
912 RC = X86::GR32RegisterClass;
914 assert(MVT::isVector(RegVT));
915 RC = X86::VR128RegisterClass;
918 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
919 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
921 // If this is an 8 or 16-bit value, it is really passed promoted to 32
922 // bits. Insert an assert[sz]ext to capture this, then truncate to the
924 if (VA.getLocInfo() == CCValAssign::SExt)
925 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
926 DAG.getValueType(VA.getValVT()));
927 else if (VA.getLocInfo() == CCValAssign::ZExt)
928 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
929 DAG.getValueType(VA.getValVT()));
931 if (VA.getLocInfo() != CCValAssign::Full)
932 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
934 ArgValues.push_back(ArgValue);
936 assert(VA.isMemLoc());
937 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
941 unsigned StackSize = CCInfo.getNextStackOffset();
942 // align stack specially for tail calls
943 if (CC==CallingConv::Fast)
944 StackSize = GetAlignedArgumentStackSize(StackSize,DAG);
946 ArgValues.push_back(Root);
948 // If the function takes variable number of arguments, make a frame index for
949 // the start of the first vararg value... for expansion of llvm.va_start.
951 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
953 // Tail call calling convention (CallingConv::Fast) does not support varargs.
954 assert( !(isVarArg && CC == CallingConv::Fast) &&
955 "CallingConv::Fast does not support varargs.");
957 if (isStdCall && !isVarArg &&
958 (CC==CallingConv::Fast && PerformTailCallOpt || CC!=CallingConv::Fast)) {
959 BytesToPopOnReturn = StackSize; // Callee pops everything..
960 BytesCallerReserves = 0;
962 BytesToPopOnReturn = 0; // Callee pops nothing.
964 // If this is an sret function, the return should pop the hidden pointer.
966 (cast<ConstantSDNode>(Op.getOperand(3))->getValue() &
967 ISD::ParamFlags::StructReturn))
968 BytesToPopOnReturn = 4;
970 BytesCallerReserves = StackSize;
973 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
975 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
976 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
978 // Return the new list of results.
979 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
980 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
983 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
985 SDOperand Chain = Op.getOperand(0);
986 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
987 SDOperand Callee = Op.getOperand(4);
988 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
990 // Analyze operands of the call, assigning locations to each operand.
991 SmallVector<CCValAssign, 16> ArgLocs;
992 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
993 if(CC==CallingConv::Fast && PerformTailCallOpt)
994 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_TailCall);
996 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
998 // Get a count of how many bytes are to be pushed on the stack.
999 unsigned NumBytes = CCInfo.getNextStackOffset();
1000 if (CC==CallingConv::Fast)
1001 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1003 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1005 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1006 SmallVector<SDOperand, 8> MemOpChains;
1010 // Walk the register/memloc assignments, inserting copies/loads.
1011 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1012 CCValAssign &VA = ArgLocs[i];
1013 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1015 // Promote the value if needed.
1016 switch (VA.getLocInfo()) {
1017 default: assert(0 && "Unknown loc info!");
1018 case CCValAssign::Full: break;
1019 case CCValAssign::SExt:
1020 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1022 case CCValAssign::ZExt:
1023 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1025 case CCValAssign::AExt:
1026 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1030 if (VA.isRegLoc()) {
1031 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1033 assert(VA.isMemLoc());
1034 if (StackPtr.Val == 0)
1035 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1037 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1042 // If the first argument is an sret pointer, remember it.
1043 bool isSRet = NumOps &&
1044 (cast<ConstantSDNode>(Op.getOperand(6))->getValue() &
1045 ISD::ParamFlags::StructReturn);
1047 if (!MemOpChains.empty())
1048 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1049 &MemOpChains[0], MemOpChains.size());
1051 // Build a sequence of copy-to-reg nodes chained together with token chain
1052 // and flag operands which copy the outgoing args into registers.
1054 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1055 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1057 InFlag = Chain.getValue(1);
1060 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1062 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1063 Subtarget->isPICStyleGOT()) {
1064 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1065 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1067 InFlag = Chain.getValue(1);
1070 // If the callee is a GlobalAddress node (quite common, every direct call is)
1071 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1072 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1073 // We should use extra load for direct calls to dllimported functions in
1075 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1076 getTargetMachine(), true))
1077 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1078 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1079 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1081 // Returns a chain & a flag for retval copy to use.
1082 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1083 SmallVector<SDOperand, 8> Ops;
1084 Ops.push_back(Chain);
1085 Ops.push_back(Callee);
1087 // Add argument registers to the end of the list so that they are known live
1089 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1090 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1091 RegsToPass[i].second.getValueType()));
1093 // Add an implicit use GOT pointer in EBX.
1094 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1095 Subtarget->isPICStyleGOT())
1096 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1099 Ops.push_back(InFlag);
1101 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1102 InFlag = Chain.getValue(1);
1104 // Create the CALLSEQ_END node.
1105 unsigned NumBytesForCalleeToPush = 0;
1107 if (CC == CallingConv::X86_StdCall ||
1108 (CC == CallingConv::Fast && PerformTailCallOpt)) {
1110 NumBytesForCalleeToPush = isSRet ? 4 : 0;
1112 NumBytesForCalleeToPush = NumBytes;
1113 assert(!(isVarArg && CC==CallingConv::Fast) &&
1114 "CallingConv::Fast does not support varargs.");
1116 // If this is is a call to a struct-return function, the callee
1117 // pops the hidden struct pointer, so we have to push it back.
1118 // This is common for Darwin/X86, Linux & Mingw32 targets.
1119 NumBytesForCalleeToPush = isSRet ? 4 : 0;
1122 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1124 Ops.push_back(Chain);
1125 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1126 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
1127 Ops.push_back(InFlag);
1128 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1129 InFlag = Chain.getValue(1);
1131 // Handle result values, copying them out of physregs into vregs that we
1133 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1137 //===----------------------------------------------------------------------===//
1138 // FastCall Calling Convention implementation
1139 //===----------------------------------------------------------------------===//
1141 // The X86 'fastcall' calling convention passes up to two integer arguments in
1142 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
1143 // and requires that the callee pop its arguments off the stack (allowing proper
1144 // tail calls), and has the same return value conventions as C calling convs.
1146 // This calling convention always arranges for the callee pop value to be 8n+4
1147 // bytes, which is needed for tail recursion elimination and stack alignment
1150 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
1151 MachineFunction &MF = DAG.getMachineFunction();
1152 MachineFrameInfo *MFI = MF.getFrameInfo();
1153 SDOperand Root = Op.getOperand(0);
1154 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1156 // Assign locations to all of the incoming arguments.
1157 SmallVector<CCValAssign, 16> ArgLocs;
1158 CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
1159 getTargetMachine(), ArgLocs);
1160 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
1162 SmallVector<SDOperand, 8> ArgValues;
1163 unsigned LastVal = ~0U;
1164 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1165 CCValAssign &VA = ArgLocs[i];
1166 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1168 assert(VA.getValNo() != LastVal &&
1169 "Don't support value assigned to multiple locs yet");
1170 LastVal = VA.getValNo();
1172 if (VA.isRegLoc()) {
1173 MVT::ValueType RegVT = VA.getLocVT();
1174 TargetRegisterClass *RC;
1175 if (RegVT == MVT::i32)
1176 RC = X86::GR32RegisterClass;
1178 assert(MVT::isVector(RegVT));
1179 RC = X86::VR128RegisterClass;
1182 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1183 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1185 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1186 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1188 if (VA.getLocInfo() == CCValAssign::SExt)
1189 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1190 DAG.getValueType(VA.getValVT()));
1191 else if (VA.getLocInfo() == CCValAssign::ZExt)
1192 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1193 DAG.getValueType(VA.getValVT()));
1195 if (VA.getLocInfo() != CCValAssign::Full)
1196 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1198 ArgValues.push_back(ArgValue);
1200 assert(VA.isMemLoc());
1201 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1205 ArgValues.push_back(Root);
1207 unsigned StackSize = CCInfo.getNextStackOffset();
1209 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1210 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1211 // arguments and the arguments after the retaddr has been pushed are
1213 if ((StackSize & 7) == 0)
1217 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1218 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1219 BytesToPopOnReturn = StackSize; // Callee pops all stack arguments.
1220 BytesCallerReserves = 0;
1222 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1223 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1225 // Return the new list of results.
1226 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1227 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1231 X86TargetLowering::LowerMemOpCallTo(SDOperand Op, SelectionDAG &DAG,
1232 const SDOperand &StackPtr,
1233 const CCValAssign &VA,
1236 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1237 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1238 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1239 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1240 if (Flags & ISD::ParamFlags::ByVal) {
1241 unsigned Align = 1 << ((Flags & ISD::ParamFlags::ByValAlign) >>
1242 ISD::ParamFlags::ByValAlignOffs);
1244 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1245 ISD::ParamFlags::ByValSizeOffs;
1247 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1248 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1250 return DAG.getNode(ISD::MEMCPY, MVT::Other, Chain, PtrOff, Arg, SizeNode,
1253 return DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1257 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1259 SDOperand Chain = Op.getOperand(0);
1260 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1261 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1262 SDOperand Callee = Op.getOperand(4);
1264 // Analyze operands of the call, assigning locations to each operand.
1265 SmallVector<CCValAssign, 16> ArgLocs;
1266 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1267 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall);
1269 // Get a count of how many bytes are to be pushed on the stack.
1270 unsigned NumBytes = CCInfo.getNextStackOffset();
1272 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1273 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1274 // arguments and the arguments after the retaddr has been pushed are
1276 if ((NumBytes & 7) == 0)
1280 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1282 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1283 SmallVector<SDOperand, 8> MemOpChains;
1287 // Walk the register/memloc assignments, inserting copies/loads.
1288 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1289 CCValAssign &VA = ArgLocs[i];
1290 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1292 // Promote the value if needed.
1293 switch (VA.getLocInfo()) {
1294 default: assert(0 && "Unknown loc info!");
1295 case CCValAssign::Full: break;
1296 case CCValAssign::SExt:
1297 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1299 case CCValAssign::ZExt:
1300 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1302 case CCValAssign::AExt:
1303 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1307 if (VA.isRegLoc()) {
1308 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1310 assert(VA.isMemLoc());
1311 if (StackPtr.Val == 0)
1312 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1314 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1319 if (!MemOpChains.empty())
1320 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1321 &MemOpChains[0], MemOpChains.size());
1323 // Build a sequence of copy-to-reg nodes chained together with token chain
1324 // and flag operands which copy the outgoing args into registers.
1326 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1327 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1329 InFlag = Chain.getValue(1);
1332 // If the callee is a GlobalAddress node (quite common, every direct call is)
1333 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1334 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1335 // We should use extra load for direct calls to dllimported functions in
1337 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1338 getTargetMachine(), true))
1339 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1340 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1341 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1343 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1345 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1346 Subtarget->isPICStyleGOT()) {
1347 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1348 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1350 InFlag = Chain.getValue(1);
1353 // Returns a chain & a flag for retval copy to use.
1354 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1355 SmallVector<SDOperand, 8> Ops;
1356 Ops.push_back(Chain);
1357 Ops.push_back(Callee);
1359 // Add argument registers to the end of the list so that they are known live
1361 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1362 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1363 RegsToPass[i].second.getValueType()));
1365 // Add an implicit use GOT pointer in EBX.
1366 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1367 Subtarget->isPICStyleGOT())
1368 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1371 Ops.push_back(InFlag);
1373 assert(isTailCall==false && "no tail call here");
1374 Chain = DAG.getNode(X86ISD::CALL,
1375 NodeTys, &Ops[0], Ops.size());
1376 InFlag = Chain.getValue(1);
1378 // Returns a flag for retval copy to use.
1379 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1381 Ops.push_back(Chain);
1382 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1383 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1384 Ops.push_back(InFlag);
1385 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1386 InFlag = Chain.getValue(1);
1388 // Handle result values, copying them out of physregs into vregs that we
1390 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1393 //===----------------------------------------------------------------------===//
1394 // Fast Calling Convention (tail call) implementation
1395 //===----------------------------------------------------------------------===//
1397 // Like std call, callee cleans arguments, convention except that ECX is
1398 // reserved for storing the tail called function address. Only 2 registers are
1399 // free for argument passing (inreg). Tail call optimization is performed
1401 // * tailcallopt is enabled
1402 // * caller/callee are fastcc
1403 // * elf/pic is disabled OR
1404 // * elf/pic enabled + callee is in module + callee has
1405 // visibility protected or hidden
1406 // To keep the stack aligned according to platform abi the function
1407 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1408 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1409 // If a tail called function callee has more arguments than the caller the
1410 // caller needs to make sure that there is room to move the RETADDR to. This is
1411 // achieved by reserving an area the size of the argument delta right after the
1412 // original REtADDR, but before the saved framepointer or the spilled registers
1413 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1425 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1426 /// for a 16 byte align requirement.
1427 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1428 SelectionDAG& DAG) {
1429 if (PerformTailCallOpt) {
1430 MachineFunction &MF = DAG.getMachineFunction();
1431 const TargetMachine &TM = MF.getTarget();
1432 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1433 unsigned StackAlignment = TFI.getStackAlignment();
1434 uint64_t AlignMask = StackAlignment - 1;
1435 int64_t Offset = StackSize;
1436 unsigned SlotSize = Subtarget->is64Bit() ? 8 : 4;
1437 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1438 // Number smaller than 12 so just add the difference.
1439 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1441 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1442 Offset = ((~AlignMask) & Offset) + StackAlignment +
1443 (StackAlignment-SlotSize);
1450 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1451 // following the call is a return. A function is eligible if caller/callee
1452 // calling conventions match, currently only fastcc supports tail calls, and the
1453 // function CALL is immediatly followed by a RET.
1454 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDOperand Call,
1456 SelectionDAG& DAG) const {
1457 bool IsEligible = false;
1459 // Check whether CALL node immediatly preceeds the RET node and whether the
1460 // return uses the result of the node or is a void return.
1461 if ((Ret.getNumOperands() == 1 &&
1462 (Ret.getOperand(0)== SDOperand(Call.Val,1) ||
1463 Ret.getOperand(0)== SDOperand(Call.Val,0))) ||
1464 (Ret.getOperand(0)== SDOperand(Call.Val,Call.Val->getNumValues()-1) &&
1465 Ret.getOperand(1)== SDOperand(Call.Val,0))) {
1466 MachineFunction &MF = DAG.getMachineFunction();
1467 unsigned CallerCC = MF.getFunction()->getCallingConv();
1468 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1469 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1470 SDOperand Callee = Call.getOperand(4);
1471 // On elf/pic %ebx needs to be livein.
1472 if(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1473 Subtarget->isPICStyleGOT()) {
1474 // Can only do local tail calls with PIC.
1475 GlobalValue * GV = 0;
1476 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1478 (GV = G->getGlobal()) &&
1479 (GV->hasHiddenVisibility() || GV->hasProtectedVisibility()))
1489 SDOperand X86TargetLowering::LowerX86_TailCallTo(SDOperand Op,
1492 SDOperand Chain = Op.getOperand(0);
1493 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1494 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1495 SDOperand Callee = Op.getOperand(4);
1496 bool is64Bit = Subtarget->is64Bit();
1498 assert(isTailCall && PerformTailCallOpt && "Should only emit tail calls.");
1500 // Analyze operands of the call, assigning locations to each operand.
1501 SmallVector<CCValAssign, 16> ArgLocs;
1502 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1504 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_TailCall);
1506 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_TailCall);
1509 // Lower arguments at fp - stackoffset + fpdiff.
1510 MachineFunction &MF = DAG.getMachineFunction();
1512 unsigned NumBytesToBePushed =
1513 GetAlignedArgumentStackSize(CCInfo.getNextStackOffset(), DAG);
1515 unsigned NumBytesCallerPushed =
1516 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1517 int FPDiff = NumBytesCallerPushed - NumBytesToBePushed;
1519 // Set the delta of movement of the returnaddr stackslot.
1520 // But only set if delta is greater than previous delta.
1521 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1522 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1525 getCALLSEQ_START(Chain, DAG.getConstant(NumBytesToBePushed, getPointerTy()));
1527 // Adjust the Return address stack slot.
1528 SDOperand RetAddrFrIdx, NewRetAddrFrIdx;
1530 MVT::ValueType VT = is64Bit ? MVT::i64 : MVT::i32;
1531 RetAddrFrIdx = getReturnAddressFrameIndex(DAG);
1532 // Load the "old" Return address.
1534 DAG.getLoad(VT, Chain,RetAddrFrIdx, NULL, 0);
1535 // Calculate the new stack slot for the return address.
1536 int SlotSize = is64Bit ? 8 : 4;
1537 int NewReturnAddrFI =
1538 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1539 NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1540 Chain = SDOperand(RetAddrFrIdx.Val, 1);
1543 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1544 SmallVector<SDOperand, 8> MemOpChains;
1545 SmallVector<SDOperand, 8> MemOpChains2;
1546 SDOperand FramePtr, StackPtr;
1551 // Walk the register/memloc assignments, inserting copies/loads. Lower
1552 // arguments first to the stack slot where they would normally - in case of a
1553 // normal function call - be.
1554 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1555 CCValAssign &VA = ArgLocs[i];
1556 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1558 // Promote the value if needed.
1559 switch (VA.getLocInfo()) {
1560 default: assert(0 && "Unknown loc info!");
1561 case CCValAssign::Full: break;
1562 case CCValAssign::SExt:
1563 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1565 case CCValAssign::ZExt:
1566 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1568 case CCValAssign::AExt:
1569 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1573 if (VA.isRegLoc()) {
1574 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1576 assert(VA.isMemLoc());
1577 if (StackPtr.Val == 0)
1578 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1580 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1585 if (!MemOpChains.empty())
1586 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1587 &MemOpChains[0], MemOpChains.size());
1589 // Build a sequence of copy-to-reg nodes chained together with token chain
1590 // and flag operands which copy the outgoing args into registers.
1592 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1593 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1595 InFlag = Chain.getValue(1);
1597 InFlag = SDOperand();
1599 // Copy from stack slots to stack slot of a tail called function. This needs
1600 // to be done because if we would lower the arguments directly to their real
1601 // stack slot we might end up overwriting each other.
1602 // TODO: To make this more efficient (sometimes saving a store/load) we could
1603 // analyse the arguments and emit this store/load/store sequence only for
1604 // arguments which would be overwritten otherwise.
1605 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1606 CCValAssign &VA = ArgLocs[i];
1607 if (!VA.isRegLoc()) {
1608 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1609 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1611 // Get source stack slot.
1612 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1613 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1614 // Create frame index.
1615 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1616 uint32_t OpSize = (MVT::getSizeInBits(VA.getLocVT())+7)/8;
1617 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1618 FIN = DAG.getFrameIndex(FI, MVT::i32);
1619 if (Flags & ISD::ParamFlags::ByVal) {
1620 // Copy relative to framepointer.
1621 unsigned Align = 1 << ((Flags & ISD::ParamFlags::ByValAlign) >>
1622 ISD::ParamFlags::ByValAlignOffs);
1624 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1625 ISD::ParamFlags::ByValSizeOffs;
1627 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1628 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1629 // Copy relative to framepointer.
1630 MemOpChains2.push_back(DAG.getNode(ISD::MEMCPY, MVT::Other, Chain, FIN,
1631 PtrOff, SizeNode, AlignNode));
1633 SDOperand LoadedArg = DAG.getLoad(VA.getValVT(), Chain, PtrOff, NULL,0);
1634 // Store relative to framepointer.
1635 MemOpChains2.push_back(DAG.getStore(Chain, LoadedArg, FIN, NULL, 0));
1640 if (!MemOpChains2.empty())
1641 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1642 &MemOpChains2[0], MemOpChains.size());
1644 // Store the return address to the appropriate stack slot.
1646 Chain = DAG.getStore(Chain,RetAddrFrIdx, NewRetAddrFrIdx, NULL, 0);
1648 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1650 // Does not work with tail call since ebx is not restored correctly by
1651 // tailcaller. TODO: at least for x86 - verify for x86-64
1653 // If the callee is a GlobalAddress node (quite common, every direct call is)
1654 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1655 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1656 // We should use extra load for direct calls to dllimported functions in
1658 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1659 getTargetMachine(), true))
1660 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1661 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1662 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1664 assert(Callee.getOpcode() == ISD::LOAD &&
1665 "Function destination must be loaded into virtual register");
1666 unsigned Opc = is64Bit ? X86::R9 : X86::ECX;
1668 Chain = DAG.getCopyToReg(Chain,
1669 DAG.getRegister(Opc, getPointerTy()) ,
1671 Callee = DAG.getRegister(Opc, getPointerTy());
1672 // Add register as live out.
1673 DAG.getMachineFunction().addLiveOut(Opc);
1676 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1677 SmallVector<SDOperand, 8> Ops;
1679 Ops.push_back(Chain);
1680 Ops.push_back(DAG.getConstant(NumBytesToBePushed, getPointerTy()));
1681 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1683 Ops.push_back(InFlag);
1684 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1685 InFlag = Chain.getValue(1);
1687 // Returns a chain & a flag for retval copy to use.
1688 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1690 Ops.push_back(Chain);
1691 Ops.push_back(Callee);
1692 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1693 // Add argument registers to the end of the list so that they are known live
1695 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1696 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1697 RegsToPass[i].second.getValueType()));
1699 Ops.push_back(InFlag);
1700 assert(InFlag.Val &&
1701 "Flag must be set. Depend on flag being set in LowerRET");
1702 Chain = DAG.getNode(X86ISD::TAILCALL,
1703 Op.Val->getVTList(), &Ops[0], Ops.size());
1705 return SDOperand(Chain.Val, Op.ResNo);
1708 //===----------------------------------------------------------------------===//
1709 // X86-64 C Calling Convention implementation
1710 //===----------------------------------------------------------------------===//
1713 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1714 MachineFunction &MF = DAG.getMachineFunction();
1715 MachineFrameInfo *MFI = MF.getFrameInfo();
1716 SDOperand Root = Op.getOperand(0);
1717 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1718 unsigned CC= MF.getFunction()->getCallingConv();
1720 static const unsigned GPR64ArgRegs[] = {
1721 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1723 static const unsigned XMMArgRegs[] = {
1724 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1725 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1729 // Assign locations to all of the incoming arguments.
1730 SmallVector<CCValAssign, 16> ArgLocs;
1731 CCState CCInfo(CC, isVarArg,
1732 getTargetMachine(), ArgLocs);
1733 if (CC == CallingConv::Fast && PerformTailCallOpt)
1734 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_TailCall);
1736 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1738 SmallVector<SDOperand, 8> ArgValues;
1739 unsigned LastVal = ~0U;
1740 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1741 CCValAssign &VA = ArgLocs[i];
1742 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1744 assert(VA.getValNo() != LastVal &&
1745 "Don't support value assigned to multiple locs yet");
1746 LastVal = VA.getValNo();
1748 if (VA.isRegLoc()) {
1749 MVT::ValueType RegVT = VA.getLocVT();
1750 TargetRegisterClass *RC;
1751 if (RegVT == MVT::i32)
1752 RC = X86::GR32RegisterClass;
1753 else if (RegVT == MVT::i64)
1754 RC = X86::GR64RegisterClass;
1755 else if (RegVT == MVT::f32)
1756 RC = X86::FR32RegisterClass;
1757 else if (RegVT == MVT::f64)
1758 RC = X86::FR64RegisterClass;
1760 assert(MVT::isVector(RegVT));
1761 if (MVT::getSizeInBits(RegVT) == 64) {
1762 RC = X86::GR64RegisterClass; // MMX values are passed in GPRs.
1765 RC = X86::VR128RegisterClass;
1768 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1769 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1771 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1772 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1774 if (VA.getLocInfo() == CCValAssign::SExt)
1775 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1776 DAG.getValueType(VA.getValVT()));
1777 else if (VA.getLocInfo() == CCValAssign::ZExt)
1778 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1779 DAG.getValueType(VA.getValVT()));
1781 if (VA.getLocInfo() != CCValAssign::Full)
1782 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1784 // Handle MMX values passed in GPRs.
1785 if (RegVT != VA.getLocVT() && RC == X86::GR64RegisterClass &&
1786 MVT::getSizeInBits(RegVT) == 64)
1787 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1789 ArgValues.push_back(ArgValue);
1791 assert(VA.isMemLoc());
1792 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1796 unsigned StackSize = CCInfo.getNextStackOffset();
1797 if (CC==CallingConv::Fast)
1798 StackSize =GetAlignedArgumentStackSize(StackSize, DAG);
1800 // If the function takes variable number of arguments, make a frame index for
1801 // the start of the first vararg value... for expansion of llvm.va_start.
1803 assert(CC!=CallingConv::Fast
1804 && "Var arg not supported with calling convention fastcc");
1805 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1806 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1808 // For X86-64, if there are vararg parameters that are passed via
1809 // registers, then we must store them to their spots on the stack so they
1810 // may be loaded by deferencing the result of va_next.
1811 VarArgsGPOffset = NumIntRegs * 8;
1812 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1813 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1814 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1816 // Store the integer parameter registers.
1817 SmallVector<SDOperand, 8> MemOps;
1818 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1819 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1820 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1821 for (; NumIntRegs != 6; ++NumIntRegs) {
1822 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1823 X86::GR64RegisterClass);
1824 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1825 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1826 MemOps.push_back(Store);
1827 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1828 DAG.getConstant(8, getPointerTy()));
1831 // Now store the XMM (fp + vector) parameter registers.
1832 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1833 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1834 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1835 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1836 X86::VR128RegisterClass);
1837 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1838 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1839 MemOps.push_back(Store);
1840 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1841 DAG.getConstant(16, getPointerTy()));
1843 if (!MemOps.empty())
1844 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1845 &MemOps[0], MemOps.size());
1848 ArgValues.push_back(Root);
1849 // Tail call convention (fastcc) needs callee pop.
1850 if (CC == CallingConv::Fast && PerformTailCallOpt) {
1851 BytesToPopOnReturn = StackSize; // Callee pops everything.
1852 BytesCallerReserves = 0;
1854 BytesToPopOnReturn = 0; // Callee pops nothing.
1855 BytesCallerReserves = StackSize;
1857 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1858 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1860 // Return the new list of results.
1861 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1862 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1866 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1868 SDOperand Chain = Op.getOperand(0);
1869 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1870 SDOperand Callee = Op.getOperand(4);
1872 // Analyze operands of the call, assigning locations to each operand.
1873 SmallVector<CCValAssign, 16> ArgLocs;
1874 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1875 if (CC==CallingConv::Fast && PerformTailCallOpt)
1876 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_TailCall);
1878 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1880 // Get a count of how many bytes are to be pushed on the stack.
1881 unsigned NumBytes = CCInfo.getNextStackOffset();
1882 if (CC == CallingConv::Fast)
1883 NumBytes = GetAlignedArgumentStackSize(NumBytes,DAG);
1885 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1887 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1888 SmallVector<SDOperand, 8> MemOpChains;
1892 // Walk the register/memloc assignments, inserting copies/loads.
1893 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1894 CCValAssign &VA = ArgLocs[i];
1895 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1897 // Promote the value if needed.
1898 switch (VA.getLocInfo()) {
1899 default: assert(0 && "Unknown loc info!");
1900 case CCValAssign::Full: break;
1901 case CCValAssign::SExt:
1902 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1904 case CCValAssign::ZExt:
1905 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1907 case CCValAssign::AExt:
1908 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1912 if (VA.isRegLoc()) {
1913 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1915 assert(VA.isMemLoc());
1916 if (StackPtr.Val == 0)
1917 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1919 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1924 if (!MemOpChains.empty())
1925 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1926 &MemOpChains[0], MemOpChains.size());
1928 // Build a sequence of copy-to-reg nodes chained together with token chain
1929 // and flag operands which copy the outgoing args into registers.
1931 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1932 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1934 InFlag = Chain.getValue(1);
1938 assert ( CallingConv::Fast != CC &&
1939 "Var args not supported with calling convention fastcc");
1941 // From AMD64 ABI document:
1942 // For calls that may call functions that use varargs or stdargs
1943 // (prototype-less calls or calls to functions containing ellipsis (...) in
1944 // the declaration) %al is used as hidden argument to specify the number
1945 // of SSE registers used. The contents of %al do not need to match exactly
1946 // the number of registers, but must be an ubound on the number of SSE
1947 // registers used and is in the range 0 - 8 inclusive.
1949 // Count the number of XMM registers allocated.
1950 static const unsigned XMMArgRegs[] = {
1951 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1952 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1954 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1956 Chain = DAG.getCopyToReg(Chain, X86::AL,
1957 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1958 InFlag = Chain.getValue(1);
1961 // If the callee is a GlobalAddress node (quite common, every direct call is)
1962 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1963 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1964 // We should use extra load for direct calls to dllimported functions in
1966 if (getTargetMachine().getCodeModel() != CodeModel::Large
1967 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1968 getTargetMachine(), true))
1969 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1970 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1971 if (getTargetMachine().getCodeModel() != CodeModel::Large)
1972 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1974 // Returns a chain & a flag for retval copy to use.
1975 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1976 SmallVector<SDOperand, 8> Ops;
1977 Ops.push_back(Chain);
1978 Ops.push_back(Callee);
1980 // Add argument registers to the end of the list so that they are known live
1982 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1983 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1984 RegsToPass[i].second.getValueType()));
1987 Ops.push_back(InFlag);
1989 Chain = DAG.getNode(X86ISD::CALL,
1990 NodeTys, &Ops[0], Ops.size());
1991 InFlag = Chain.getValue(1);
1992 int NumBytesForCalleeToPush = 0;
1993 if (CC==CallingConv::Fast && PerformTailCallOpt) {
1994 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1996 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1998 // Returns a flag for retval copy to use.
1999 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2001 Ops.push_back(Chain);
2002 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
2003 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
2004 Ops.push_back(InFlag);
2005 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
2006 InFlag = Chain.getValue(1);
2008 // Handle result values, copying them out of physregs into vregs that we
2010 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
2014 //===----------------------------------------------------------------------===//
2015 // Other Lowering Hooks
2016 //===----------------------------------------------------------------------===//
2019 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2020 MachineFunction &MF = DAG.getMachineFunction();
2021 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2022 int ReturnAddrIndex = FuncInfo->getRAIndex();
2024 if (ReturnAddrIndex == 0) {
2025 // Set up a frame object for the return address.
2026 if (Subtarget->is64Bit())
2027 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
2029 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
2031 FuncInfo->setRAIndex(ReturnAddrIndex);
2034 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2039 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
2040 /// specific condition code. It returns a false if it cannot do a direct
2041 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
2043 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2044 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
2045 SelectionDAG &DAG) {
2046 X86CC = X86::COND_INVALID;
2048 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2049 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2050 // X > -1 -> X == 0, jump !sign.
2051 RHS = DAG.getConstant(0, RHS.getValueType());
2052 X86CC = X86::COND_NS;
2054 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2055 // X < 0 -> X == 0, jump on sign.
2056 X86CC = X86::COND_S;
2058 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
2060 RHS = DAG.getConstant(0, RHS.getValueType());
2061 X86CC = X86::COND_LE;
2066 switch (SetCCOpcode) {
2068 case ISD::SETEQ: X86CC = X86::COND_E; break;
2069 case ISD::SETGT: X86CC = X86::COND_G; break;
2070 case ISD::SETGE: X86CC = X86::COND_GE; break;
2071 case ISD::SETLT: X86CC = X86::COND_L; break;
2072 case ISD::SETLE: X86CC = X86::COND_LE; break;
2073 case ISD::SETNE: X86CC = X86::COND_NE; break;
2074 case ISD::SETULT: X86CC = X86::COND_B; break;
2075 case ISD::SETUGT: X86CC = X86::COND_A; break;
2076 case ISD::SETULE: X86CC = X86::COND_BE; break;
2077 case ISD::SETUGE: X86CC = X86::COND_AE; break;
2080 // On a floating point condition, the flags are set as follows:
2082 // 0 | 0 | 0 | X > Y
2083 // 0 | 0 | 1 | X < Y
2084 // 1 | 0 | 0 | X == Y
2085 // 1 | 1 | 1 | unordered
2087 switch (SetCCOpcode) {
2090 case ISD::SETEQ: X86CC = X86::COND_E; break;
2091 case ISD::SETOLT: Flip = true; // Fallthrough
2093 case ISD::SETGT: X86CC = X86::COND_A; break;
2094 case ISD::SETOLE: Flip = true; // Fallthrough
2096 case ISD::SETGE: X86CC = X86::COND_AE; break;
2097 case ISD::SETUGT: Flip = true; // Fallthrough
2099 case ISD::SETLT: X86CC = X86::COND_B; break;
2100 case ISD::SETUGE: Flip = true; // Fallthrough
2102 case ISD::SETLE: X86CC = X86::COND_BE; break;
2104 case ISD::SETNE: X86CC = X86::COND_NE; break;
2105 case ISD::SETUO: X86CC = X86::COND_P; break;
2106 case ISD::SETO: X86CC = X86::COND_NP; break;
2109 std::swap(LHS, RHS);
2112 return X86CC != X86::COND_INVALID;
2115 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2116 /// code. Current x86 isa includes the following FP cmov instructions:
2117 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2118 static bool hasFPCMov(unsigned X86CC) {
2134 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2135 /// true if Op is undef or if its value falls within the specified range (L, H].
2136 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
2137 if (Op.getOpcode() == ISD::UNDEF)
2140 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2141 return (Val >= Low && Val < Hi);
2144 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2145 /// true if Op is undef or if its value equal to the specified value.
2146 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
2147 if (Op.getOpcode() == ISD::UNDEF)
2149 return cast<ConstantSDNode>(Op)->getValue() == Val;
2152 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2153 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2154 bool X86::isPSHUFDMask(SDNode *N) {
2155 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2157 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2160 // Check if the value doesn't reference the second vector.
2161 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2162 SDOperand Arg = N->getOperand(i);
2163 if (Arg.getOpcode() == ISD::UNDEF) continue;
2164 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2165 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
2172 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2173 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2174 bool X86::isPSHUFHWMask(SDNode *N) {
2175 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2177 if (N->getNumOperands() != 8)
2180 // Lower quadword copied in order.
2181 for (unsigned i = 0; i != 4; ++i) {
2182 SDOperand Arg = N->getOperand(i);
2183 if (Arg.getOpcode() == ISD::UNDEF) continue;
2184 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2185 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2189 // Upper quadword shuffled.
2190 for (unsigned i = 4; i != 8; ++i) {
2191 SDOperand Arg = N->getOperand(i);
2192 if (Arg.getOpcode() == ISD::UNDEF) continue;
2193 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2194 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2195 if (Val < 4 || Val > 7)
2202 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2203 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2204 bool X86::isPSHUFLWMask(SDNode *N) {
2205 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2207 if (N->getNumOperands() != 8)
2210 // Upper quadword copied in order.
2211 for (unsigned i = 4; i != 8; ++i)
2212 if (!isUndefOrEqual(N->getOperand(i), i))
2215 // Lower quadword shuffled.
2216 for (unsigned i = 0; i != 4; ++i)
2217 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2223 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2224 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2225 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2226 if (NumElems != 2 && NumElems != 4) return false;
2228 unsigned Half = NumElems / 2;
2229 for (unsigned i = 0; i < Half; ++i)
2230 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2232 for (unsigned i = Half; i < NumElems; ++i)
2233 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2239 bool X86::isSHUFPMask(SDNode *N) {
2240 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2241 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2244 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2245 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2246 /// half elements to come from vector 1 (which would equal the dest.) and
2247 /// the upper half to come from vector 2.
2248 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2249 if (NumOps != 2 && NumOps != 4) return false;
2251 unsigned Half = NumOps / 2;
2252 for (unsigned i = 0; i < Half; ++i)
2253 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2255 for (unsigned i = Half; i < NumOps; ++i)
2256 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2261 static bool isCommutedSHUFP(SDNode *N) {
2262 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2263 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2266 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2267 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2268 bool X86::isMOVHLPSMask(SDNode *N) {
2269 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2271 if (N->getNumOperands() != 4)
2274 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2275 return isUndefOrEqual(N->getOperand(0), 6) &&
2276 isUndefOrEqual(N->getOperand(1), 7) &&
2277 isUndefOrEqual(N->getOperand(2), 2) &&
2278 isUndefOrEqual(N->getOperand(3), 3);
2281 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2282 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2284 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2285 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2287 if (N->getNumOperands() != 4)
2290 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2291 return isUndefOrEqual(N->getOperand(0), 2) &&
2292 isUndefOrEqual(N->getOperand(1), 3) &&
2293 isUndefOrEqual(N->getOperand(2), 2) &&
2294 isUndefOrEqual(N->getOperand(3), 3);
2297 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2298 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2299 bool X86::isMOVLPMask(SDNode *N) {
2300 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2302 unsigned NumElems = N->getNumOperands();
2303 if (NumElems != 2 && NumElems != 4)
2306 for (unsigned i = 0; i < NumElems/2; ++i)
2307 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2310 for (unsigned i = NumElems/2; i < NumElems; ++i)
2311 if (!isUndefOrEqual(N->getOperand(i), i))
2317 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2318 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2320 bool X86::isMOVHPMask(SDNode *N) {
2321 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2323 unsigned NumElems = N->getNumOperands();
2324 if (NumElems != 2 && NumElems != 4)
2327 for (unsigned i = 0; i < NumElems/2; ++i)
2328 if (!isUndefOrEqual(N->getOperand(i), i))
2331 for (unsigned i = 0; i < NumElems/2; ++i) {
2332 SDOperand Arg = N->getOperand(i + NumElems/2);
2333 if (!isUndefOrEqual(Arg, i + NumElems))
2340 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2341 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2342 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2343 bool V2IsSplat = false) {
2344 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2347 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2348 SDOperand BitI = Elts[i];
2349 SDOperand BitI1 = Elts[i+1];
2350 if (!isUndefOrEqual(BitI, j))
2353 if (isUndefOrEqual(BitI1, NumElts))
2356 if (!isUndefOrEqual(BitI1, j + NumElts))
2364 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2365 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2366 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2369 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2370 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2371 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2372 bool V2IsSplat = false) {
2373 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2376 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2377 SDOperand BitI = Elts[i];
2378 SDOperand BitI1 = Elts[i+1];
2379 if (!isUndefOrEqual(BitI, j + NumElts/2))
2382 if (isUndefOrEqual(BitI1, NumElts))
2385 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2393 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2394 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2395 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2398 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2399 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2401 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2402 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2404 unsigned NumElems = N->getNumOperands();
2405 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2408 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2409 SDOperand BitI = N->getOperand(i);
2410 SDOperand BitI1 = N->getOperand(i+1);
2412 if (!isUndefOrEqual(BitI, j))
2414 if (!isUndefOrEqual(BitI1, j))
2421 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2422 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2424 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2425 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2427 unsigned NumElems = N->getNumOperands();
2428 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2431 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2432 SDOperand BitI = N->getOperand(i);
2433 SDOperand BitI1 = N->getOperand(i + 1);
2435 if (!isUndefOrEqual(BitI, j))
2437 if (!isUndefOrEqual(BitI1, j))
2444 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2445 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2446 /// MOVSD, and MOVD, i.e. setting the lowest element.
2447 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2448 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2451 if (!isUndefOrEqual(Elts[0], NumElts))
2454 for (unsigned i = 1; i < NumElts; ++i) {
2455 if (!isUndefOrEqual(Elts[i], i))
2462 bool X86::isMOVLMask(SDNode *N) {
2463 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2464 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2467 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2468 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2469 /// element of vector 2 and the other elements to come from vector 1 in order.
2470 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2471 bool V2IsSplat = false,
2472 bool V2IsUndef = false) {
2473 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2476 if (!isUndefOrEqual(Ops[0], 0))
2479 for (unsigned i = 1; i < NumOps; ++i) {
2480 SDOperand Arg = Ops[i];
2481 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2482 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2483 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2490 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2491 bool V2IsUndef = false) {
2492 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2493 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2494 V2IsSplat, V2IsUndef);
2497 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2498 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2499 bool X86::isMOVSHDUPMask(SDNode *N) {
2500 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2502 if (N->getNumOperands() != 4)
2505 // Expect 1, 1, 3, 3
2506 for (unsigned i = 0; i < 2; ++i) {
2507 SDOperand Arg = N->getOperand(i);
2508 if (Arg.getOpcode() == ISD::UNDEF) continue;
2509 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2510 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2511 if (Val != 1) return false;
2515 for (unsigned i = 2; i < 4; ++i) {
2516 SDOperand Arg = N->getOperand(i);
2517 if (Arg.getOpcode() == ISD::UNDEF) continue;
2518 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2519 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2520 if (Val != 3) return false;
2524 // Don't use movshdup if it can be done with a shufps.
2528 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2529 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2530 bool X86::isMOVSLDUPMask(SDNode *N) {
2531 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2533 if (N->getNumOperands() != 4)
2536 // Expect 0, 0, 2, 2
2537 for (unsigned i = 0; i < 2; ++i) {
2538 SDOperand Arg = N->getOperand(i);
2539 if (Arg.getOpcode() == ISD::UNDEF) continue;
2540 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2541 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2542 if (Val != 0) return false;
2546 for (unsigned i = 2; i < 4; ++i) {
2547 SDOperand Arg = N->getOperand(i);
2548 if (Arg.getOpcode() == ISD::UNDEF) continue;
2549 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2550 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2551 if (Val != 2) return false;
2555 // Don't use movshdup if it can be done with a shufps.
2559 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2560 /// specifies a identity operation on the LHS or RHS.
2561 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2562 unsigned NumElems = N->getNumOperands();
2563 for (unsigned i = 0; i < NumElems; ++i)
2564 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2569 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2570 /// a splat of a single element.
2571 static bool isSplatMask(SDNode *N) {
2572 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2574 // This is a splat operation if each element of the permute is the same, and
2575 // if the value doesn't reference the second vector.
2576 unsigned NumElems = N->getNumOperands();
2577 SDOperand ElementBase;
2579 for (; i != NumElems; ++i) {
2580 SDOperand Elt = N->getOperand(i);
2581 if (isa<ConstantSDNode>(Elt)) {
2587 if (!ElementBase.Val)
2590 for (; i != NumElems; ++i) {
2591 SDOperand Arg = N->getOperand(i);
2592 if (Arg.getOpcode() == ISD::UNDEF) continue;
2593 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2594 if (Arg != ElementBase) return false;
2597 // Make sure it is a splat of the first vector operand.
2598 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2601 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2602 /// a splat of a single element and it's a 2 or 4 element mask.
2603 bool X86::isSplatMask(SDNode *N) {
2604 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2606 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2607 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2609 return ::isSplatMask(N);
2612 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2613 /// specifies a splat of zero element.
2614 bool X86::isSplatLoMask(SDNode *N) {
2615 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2617 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2618 if (!isUndefOrEqual(N->getOperand(i), 0))
2623 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2624 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2626 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2627 unsigned NumOperands = N->getNumOperands();
2628 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2630 for (unsigned i = 0; i < NumOperands; ++i) {
2632 SDOperand Arg = N->getOperand(NumOperands-i-1);
2633 if (Arg.getOpcode() != ISD::UNDEF)
2634 Val = cast<ConstantSDNode>(Arg)->getValue();
2635 if (Val >= NumOperands) Val -= NumOperands;
2637 if (i != NumOperands - 1)
2644 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2645 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2647 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2649 // 8 nodes, but we only care about the last 4.
2650 for (unsigned i = 7; i >= 4; --i) {
2652 SDOperand Arg = N->getOperand(i);
2653 if (Arg.getOpcode() != ISD::UNDEF)
2654 Val = cast<ConstantSDNode>(Arg)->getValue();
2663 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2664 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2666 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2668 // 8 nodes, but we only care about the first 4.
2669 for (int i = 3; i >= 0; --i) {
2671 SDOperand Arg = N->getOperand(i);
2672 if (Arg.getOpcode() != ISD::UNDEF)
2673 Val = cast<ConstantSDNode>(Arg)->getValue();
2682 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2683 /// specifies a 8 element shuffle that can be broken into a pair of
2684 /// PSHUFHW and PSHUFLW.
2685 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2686 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2688 if (N->getNumOperands() != 8)
2691 // Lower quadword shuffled.
2692 for (unsigned i = 0; i != 4; ++i) {
2693 SDOperand Arg = N->getOperand(i);
2694 if (Arg.getOpcode() == ISD::UNDEF) continue;
2695 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2696 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2701 // Upper quadword shuffled.
2702 for (unsigned i = 4; i != 8; ++i) {
2703 SDOperand Arg = N->getOperand(i);
2704 if (Arg.getOpcode() == ISD::UNDEF) continue;
2705 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2706 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2707 if (Val < 4 || Val > 7)
2714 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2715 /// values in ther permute mask.
2716 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2717 SDOperand &V2, SDOperand &Mask,
2718 SelectionDAG &DAG) {
2719 MVT::ValueType VT = Op.getValueType();
2720 MVT::ValueType MaskVT = Mask.getValueType();
2721 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2722 unsigned NumElems = Mask.getNumOperands();
2723 SmallVector<SDOperand, 8> MaskVec;
2725 for (unsigned i = 0; i != NumElems; ++i) {
2726 SDOperand Arg = Mask.getOperand(i);
2727 if (Arg.getOpcode() == ISD::UNDEF) {
2728 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2731 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2732 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2734 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2736 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2740 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2741 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2744 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2745 /// match movhlps. The lower half elements should come from upper half of
2746 /// V1 (and in order), and the upper half elements should come from the upper
2747 /// half of V2 (and in order).
2748 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2749 unsigned NumElems = Mask->getNumOperands();
2752 for (unsigned i = 0, e = 2; i != e; ++i)
2753 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2755 for (unsigned i = 2; i != 4; ++i)
2756 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2761 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2762 /// is promoted to a vector.
2763 static inline bool isScalarLoadToVector(SDNode *N) {
2764 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2765 N = N->getOperand(0).Val;
2766 return ISD::isNON_EXTLoad(N);
2771 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2772 /// match movlp{s|d}. The lower half elements should come from lower half of
2773 /// V1 (and in order), and the upper half elements should come from the upper
2774 /// half of V2 (and in order). And since V1 will become the source of the
2775 /// MOVLP, it must be either a vector load or a scalar load to vector.
2776 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2777 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2779 // Is V2 is a vector load, don't do this transformation. We will try to use
2780 // load folding shufps op.
2781 if (ISD::isNON_EXTLoad(V2))
2784 unsigned NumElems = Mask->getNumOperands();
2785 if (NumElems != 2 && NumElems != 4)
2787 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2788 if (!isUndefOrEqual(Mask->getOperand(i), i))
2790 for (unsigned i = NumElems/2; i != NumElems; ++i)
2791 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2796 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2798 static bool isSplatVector(SDNode *N) {
2799 if (N->getOpcode() != ISD::BUILD_VECTOR)
2802 SDOperand SplatValue = N->getOperand(0);
2803 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2804 if (N->getOperand(i) != SplatValue)
2809 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2811 static bool isUndefShuffle(SDNode *N) {
2812 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2815 SDOperand V1 = N->getOperand(0);
2816 SDOperand V2 = N->getOperand(1);
2817 SDOperand Mask = N->getOperand(2);
2818 unsigned NumElems = Mask.getNumOperands();
2819 for (unsigned i = 0; i != NumElems; ++i) {
2820 SDOperand Arg = Mask.getOperand(i);
2821 if (Arg.getOpcode() != ISD::UNDEF) {
2822 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2823 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2825 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2832 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2834 static inline bool isZeroNode(SDOperand Elt) {
2835 return ((isa<ConstantSDNode>(Elt) &&
2836 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2837 (isa<ConstantFPSDNode>(Elt) &&
2838 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2841 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2842 /// to an zero vector.
2843 static bool isZeroShuffle(SDNode *N) {
2844 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2847 SDOperand V1 = N->getOperand(0);
2848 SDOperand V2 = N->getOperand(1);
2849 SDOperand Mask = N->getOperand(2);
2850 unsigned NumElems = Mask.getNumOperands();
2851 for (unsigned i = 0; i != NumElems; ++i) {
2852 SDOperand Arg = Mask.getOperand(i);
2853 if (Arg.getOpcode() != ISD::UNDEF) {
2854 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2855 if (Idx < NumElems) {
2856 unsigned Opc = V1.Val->getOpcode();
2857 if (Opc == ISD::UNDEF)
2859 if (Opc != ISD::BUILD_VECTOR ||
2860 !isZeroNode(V1.Val->getOperand(Idx)))
2862 } else if (Idx >= NumElems) {
2863 unsigned Opc = V2.Val->getOpcode();
2864 if (Opc == ISD::UNDEF)
2866 if (Opc != ISD::BUILD_VECTOR ||
2867 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2875 /// getZeroVector - Returns a vector of specified type with all zero elements.
2877 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2878 assert(MVT::isVector(VT) && "Expected a vector type");
2879 unsigned NumElems = MVT::getVectorNumElements(VT);
2880 MVT::ValueType EVT = MVT::getVectorElementType(VT);
2881 bool isFP = MVT::isFloatingPoint(EVT);
2882 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2883 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2884 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2887 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2888 /// that point to V2 points to its first element.
2889 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2890 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2892 bool Changed = false;
2893 SmallVector<SDOperand, 8> MaskVec;
2894 unsigned NumElems = Mask.getNumOperands();
2895 for (unsigned i = 0; i != NumElems; ++i) {
2896 SDOperand Arg = Mask.getOperand(i);
2897 if (Arg.getOpcode() != ISD::UNDEF) {
2898 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2899 if (Val > NumElems) {
2900 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2904 MaskVec.push_back(Arg);
2908 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2909 &MaskVec[0], MaskVec.size());
2913 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2914 /// operation of specified width.
2915 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2916 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2917 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2919 SmallVector<SDOperand, 8> MaskVec;
2920 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2921 for (unsigned i = 1; i != NumElems; ++i)
2922 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2923 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2926 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2927 /// of specified width.
2928 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2929 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2930 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2931 SmallVector<SDOperand, 8> MaskVec;
2932 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2933 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2934 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2936 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2939 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2940 /// of specified width.
2941 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2942 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2943 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2944 unsigned Half = NumElems/2;
2945 SmallVector<SDOperand, 8> MaskVec;
2946 for (unsigned i = 0; i != Half; ++i) {
2947 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2948 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2950 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2953 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2955 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2956 SDOperand V1 = Op.getOperand(0);
2957 SDOperand Mask = Op.getOperand(2);
2958 MVT::ValueType VT = Op.getValueType();
2959 unsigned NumElems = Mask.getNumOperands();
2960 Mask = getUnpacklMask(NumElems, DAG);
2961 while (NumElems != 4) {
2962 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2965 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2967 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2968 Mask = getZeroVector(MaskVT, DAG);
2969 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2970 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2971 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2974 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2975 /// vector of zero or undef vector.
2976 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2977 unsigned NumElems, unsigned Idx,
2978 bool isZero, SelectionDAG &DAG) {
2979 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2980 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2981 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
2982 SDOperand Zero = DAG.getConstant(0, EVT);
2983 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2984 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2985 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2986 &MaskVec[0], MaskVec.size());
2987 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2990 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2992 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2993 unsigned NumNonZero, unsigned NumZero,
2994 SelectionDAG &DAG, TargetLowering &TLI) {
3000 for (unsigned i = 0; i < 16; ++i) {
3001 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3002 if (ThisIsNonZero && First) {
3004 V = getZeroVector(MVT::v8i16, DAG);
3006 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3011 SDOperand ThisElt(0, 0), LastElt(0, 0);
3012 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3013 if (LastIsNonZero) {
3014 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3016 if (ThisIsNonZero) {
3017 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3018 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3019 ThisElt, DAG.getConstant(8, MVT::i8));
3021 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3026 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3027 DAG.getConstant(i/2, TLI.getPointerTy()));
3031 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3034 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3036 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
3037 unsigned NumNonZero, unsigned NumZero,
3038 SelectionDAG &DAG, TargetLowering &TLI) {
3044 for (unsigned i = 0; i < 8; ++i) {
3045 bool isNonZero = (NonZeros & (1 << i)) != 0;
3049 V = getZeroVector(MVT::v8i16, DAG);
3051 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3054 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3055 DAG.getConstant(i, TLI.getPointerTy()));
3063 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3064 // All zero's are handled with pxor.
3065 if (ISD::isBuildVectorAllZeros(Op.Val))
3068 // All one's are handled with pcmpeqd.
3069 if (ISD::isBuildVectorAllOnes(Op.Val))
3072 MVT::ValueType VT = Op.getValueType();
3073 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3074 unsigned EVTBits = MVT::getSizeInBits(EVT);
3076 unsigned NumElems = Op.getNumOperands();
3077 unsigned NumZero = 0;
3078 unsigned NumNonZero = 0;
3079 unsigned NonZeros = 0;
3080 unsigned NumNonZeroImms = 0;
3081 std::set<SDOperand> Values;
3082 for (unsigned i = 0; i < NumElems; ++i) {
3083 SDOperand Elt = Op.getOperand(i);
3084 if (Elt.getOpcode() != ISD::UNDEF) {
3086 if (isZeroNode(Elt))
3089 NonZeros |= (1 << i);
3091 if (Elt.getOpcode() == ISD::Constant ||
3092 Elt.getOpcode() == ISD::ConstantFP)
3098 if (NumNonZero == 0) {
3100 // All undef vector. Return an UNDEF.
3101 return DAG.getNode(ISD::UNDEF, VT);
3103 // A mix of zero and undef. Return a zero vector.
3104 return getZeroVector(VT, DAG);
3107 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3108 if (Values.size() == 1)
3111 // Special case for single non-zero element.
3112 if (NumNonZero == 1) {
3113 unsigned Idx = CountTrailingZeros_32(NonZeros);
3114 SDOperand Item = Op.getOperand(Idx);
3115 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3117 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3118 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
3121 if (EVTBits == 32) {
3122 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3123 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
3125 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3126 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3127 SmallVector<SDOperand, 8> MaskVec;
3128 for (unsigned i = 0; i < NumElems; i++)
3129 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3130 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3131 &MaskVec[0], MaskVec.size());
3132 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3133 DAG.getNode(ISD::UNDEF, VT), Mask);
3137 // A vector full of immediates; various special cases are already
3138 // handled, so this is best done with a single constant-pool load.
3139 if (NumNonZero == NumNonZeroImms)
3142 // Let legalizer expand 2-wide build_vectors.
3146 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3147 if (EVTBits == 8 && NumElems == 16) {
3148 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3150 if (V.Val) return V;
3153 if (EVTBits == 16 && NumElems == 8) {
3154 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3156 if (V.Val) return V;
3159 // If element VT is == 32 bits, turn it into a number of shuffles.
3160 SmallVector<SDOperand, 8> V;
3162 if (NumElems == 4 && NumZero > 0) {
3163 for (unsigned i = 0; i < 4; ++i) {
3164 bool isZero = !(NonZeros & (1 << i));
3166 V[i] = getZeroVector(VT, DAG);
3168 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3171 for (unsigned i = 0; i < 2; ++i) {
3172 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3175 V[i] = V[i*2]; // Must be a zero vector.
3178 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3179 getMOVLMask(NumElems, DAG));
3182 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3183 getMOVLMask(NumElems, DAG));
3186 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3187 getUnpacklMask(NumElems, DAG));
3192 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3193 // clears the upper bits.
3194 // FIXME: we can do the same for v4f32 case when we know both parts of
3195 // the lower half come from scalar_to_vector (loadf32). We should do
3196 // that in post legalizer dag combiner with target specific hooks.
3197 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3199 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3200 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3201 SmallVector<SDOperand, 8> MaskVec;
3202 bool Reverse = (NonZeros & 0x3) == 2;
3203 for (unsigned i = 0; i < 2; ++i)
3205 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3207 MaskVec.push_back(DAG.getConstant(i, EVT));
3208 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3209 for (unsigned i = 0; i < 2; ++i)
3211 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3213 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3214 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3215 &MaskVec[0], MaskVec.size());
3216 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3219 if (Values.size() > 2) {
3220 // Expand into a number of unpckl*.
3222 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3223 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3224 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3225 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3226 for (unsigned i = 0; i < NumElems; ++i)
3227 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3229 while (NumElems != 0) {
3230 for (unsigned i = 0; i < NumElems; ++i)
3231 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3242 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3243 SDOperand V1 = Op.getOperand(0);
3244 SDOperand V2 = Op.getOperand(1);
3245 SDOperand PermMask = Op.getOperand(2);
3246 MVT::ValueType VT = Op.getValueType();
3247 unsigned NumElems = PermMask.getNumOperands();
3248 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3249 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3250 bool V1IsSplat = false;
3251 bool V2IsSplat = false;
3253 if (isUndefShuffle(Op.Val))
3254 return DAG.getNode(ISD::UNDEF, VT);
3256 if (isZeroShuffle(Op.Val))
3257 return getZeroVector(VT, DAG);
3259 if (isIdentityMask(PermMask.Val))
3261 else if (isIdentityMask(PermMask.Val, true))
3264 if (isSplatMask(PermMask.Val)) {
3265 if (NumElems <= 4) return Op;
3266 // Promote it to a v4i32 splat.
3267 return PromoteSplat(Op, DAG);
3270 if (X86::isMOVLMask(PermMask.Val))
3271 return (V1IsUndef) ? V2 : Op;
3273 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3274 X86::isMOVSLDUPMask(PermMask.Val) ||
3275 X86::isMOVHLPSMask(PermMask.Val) ||
3276 X86::isMOVHPMask(PermMask.Val) ||
3277 X86::isMOVLPMask(PermMask.Val))
3280 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3281 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3282 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3284 bool Commuted = false;
3285 V1IsSplat = isSplatVector(V1.Val);
3286 V2IsSplat = isSplatVector(V2.Val);
3287 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3288 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3289 std::swap(V1IsSplat, V2IsSplat);
3290 std::swap(V1IsUndef, V2IsUndef);
3294 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3295 if (V2IsUndef) return V1;
3296 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3298 // V2 is a splat, so the mask may be malformed. That is, it may point
3299 // to any V2 element. The instruction selectior won't like this. Get
3300 // a corrected mask and commute to form a proper MOVS{S|D}.
3301 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3302 if (NewMask.Val != PermMask.Val)
3303 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3308 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3309 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3310 X86::isUNPCKLMask(PermMask.Val) ||
3311 X86::isUNPCKHMask(PermMask.Val))
3315 // Normalize mask so all entries that point to V2 points to its first
3316 // element then try to match unpck{h|l} again. If match, return a
3317 // new vector_shuffle with the corrected mask.
3318 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3319 if (NewMask.Val != PermMask.Val) {
3320 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3321 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3322 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3323 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3324 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3325 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3330 // Normalize the node to match x86 shuffle ops if needed
3331 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3332 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3335 // Commute is back and try unpck* again.
3336 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3337 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3338 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3339 X86::isUNPCKLMask(PermMask.Val) ||
3340 X86::isUNPCKHMask(PermMask.Val))
3344 // If VT is integer, try PSHUF* first, then SHUFP*.
3345 if (MVT::isInteger(VT)) {
3346 // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
3347 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
3348 if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
3349 X86::isPSHUFDMask(PermMask.Val)) ||
3350 X86::isPSHUFHWMask(PermMask.Val) ||
3351 X86::isPSHUFLWMask(PermMask.Val)) {
3352 if (V2.getOpcode() != ISD::UNDEF)
3353 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3354 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3358 if (X86::isSHUFPMask(PermMask.Val) &&
3359 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
3362 // Handle v8i16 shuffle high / low shuffle node pair.
3363 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3364 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3365 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3366 SmallVector<SDOperand, 8> MaskVec;
3367 for (unsigned i = 0; i != 4; ++i)
3368 MaskVec.push_back(PermMask.getOperand(i));
3369 for (unsigned i = 4; i != 8; ++i)
3370 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3371 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3372 &MaskVec[0], MaskVec.size());
3373 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3375 for (unsigned i = 0; i != 4; ++i)
3376 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3377 for (unsigned i = 4; i != 8; ++i)
3378 MaskVec.push_back(PermMask.getOperand(i));
3379 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
3380 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3383 // Floating point cases in the other order.
3384 if (X86::isSHUFPMask(PermMask.Val))
3386 if (X86::isPSHUFDMask(PermMask.Val) ||
3387 X86::isPSHUFHWMask(PermMask.Val) ||
3388 X86::isPSHUFLWMask(PermMask.Val)) {
3389 if (V2.getOpcode() != ISD::UNDEF)
3390 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3391 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3396 if (NumElems == 4 &&
3397 // Don't do this for MMX.
3398 MVT::getSizeInBits(VT) != 64) {
3399 MVT::ValueType MaskVT = PermMask.getValueType();
3400 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3401 SmallVector<std::pair<int, int>, 8> Locs;
3402 Locs.reserve(NumElems);
3403 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3404 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3407 // If no more than two elements come from either vector. This can be
3408 // implemented with two shuffles. First shuffle gather the elements.
3409 // The second shuffle, which takes the first shuffle as both of its
3410 // vector operands, put the elements into the right order.
3411 for (unsigned i = 0; i != NumElems; ++i) {
3412 SDOperand Elt = PermMask.getOperand(i);
3413 if (Elt.getOpcode() == ISD::UNDEF) {
3414 Locs[i] = std::make_pair(-1, -1);
3416 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3417 if (Val < NumElems) {
3418 Locs[i] = std::make_pair(0, NumLo);
3422 Locs[i] = std::make_pair(1, NumHi);
3423 if (2+NumHi < NumElems)
3424 Mask1[2+NumHi] = Elt;
3429 if (NumLo <= 2 && NumHi <= 2) {
3430 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3431 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3432 &Mask1[0], Mask1.size()));
3433 for (unsigned i = 0; i != NumElems; ++i) {
3434 if (Locs[i].first == -1)
3437 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3438 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3439 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3443 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3444 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3445 &Mask2[0], Mask2.size()));
3448 // Break it into (shuffle shuffle_hi, shuffle_lo).
3450 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3451 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3452 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3453 unsigned MaskIdx = 0;
3455 unsigned HiIdx = NumElems/2;
3456 for (unsigned i = 0; i != NumElems; ++i) {
3457 if (i == NumElems/2) {
3463 SDOperand Elt = PermMask.getOperand(i);
3464 if (Elt.getOpcode() == ISD::UNDEF) {
3465 Locs[i] = std::make_pair(-1, -1);
3466 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3467 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3468 (*MaskPtr)[LoIdx] = Elt;
3471 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3472 (*MaskPtr)[HiIdx] = Elt;
3477 SDOperand LoShuffle =
3478 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3479 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3480 &LoMask[0], LoMask.size()));
3481 SDOperand HiShuffle =
3482 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3483 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3484 &HiMask[0], HiMask.size()));
3485 SmallVector<SDOperand, 8> MaskOps;
3486 for (unsigned i = 0; i != NumElems; ++i) {
3487 if (Locs[i].first == -1) {
3488 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3490 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3491 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3494 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3495 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3496 &MaskOps[0], MaskOps.size()));
3503 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3504 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3507 MVT::ValueType VT = Op.getValueType();
3508 // TODO: handle v16i8.
3509 if (MVT::getSizeInBits(VT) == 16) {
3510 // Transform it so it match pextrw which produces a 32-bit result.
3511 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3512 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3513 Op.getOperand(0), Op.getOperand(1));
3514 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3515 DAG.getValueType(VT));
3516 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3517 } else if (MVT::getSizeInBits(VT) == 32) {
3518 SDOperand Vec = Op.getOperand(0);
3519 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3522 // SHUFPS the element to the lowest double word, then movss.
3523 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3524 SmallVector<SDOperand, 8> IdxVec;
3526 push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
3528 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3530 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3532 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3533 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3534 &IdxVec[0], IdxVec.size());
3535 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3536 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3537 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3538 DAG.getConstant(0, getPointerTy()));
3539 } else if (MVT::getSizeInBits(VT) == 64) {
3540 SDOperand Vec = Op.getOperand(0);
3541 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3545 // UNPCKHPD the element to the lowest double word, then movsd.
3546 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3547 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3548 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3549 SmallVector<SDOperand, 8> IdxVec;
3550 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
3552 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3553 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3554 &IdxVec[0], IdxVec.size());
3555 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3556 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3557 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3558 DAG.getConstant(0, getPointerTy()));
3565 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3566 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3567 // as its second argument.
3568 MVT::ValueType VT = Op.getValueType();
3569 MVT::ValueType BaseVT = MVT::getVectorElementType(VT);
3570 SDOperand N0 = Op.getOperand(0);
3571 SDOperand N1 = Op.getOperand(1);
3572 SDOperand N2 = Op.getOperand(2);
3573 if (MVT::getSizeInBits(BaseVT) == 16) {
3574 if (N1.getValueType() != MVT::i32)
3575 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3576 if (N2.getValueType() != MVT::i32)
3577 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(),getPointerTy());
3578 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3579 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3580 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3583 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3584 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3585 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3586 SmallVector<SDOperand, 8> MaskVec;
3587 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3588 for (unsigned i = 1; i <= 3; ++i)
3589 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3590 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3591 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3592 &MaskVec[0], MaskVec.size()));
3594 // Use two pinsrw instructions to insert a 32 bit value.
3596 if (MVT::isFloatingPoint(N1.getValueType())) {
3597 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3598 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3599 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3600 DAG.getConstant(0, getPointerTy()));
3602 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3603 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3604 DAG.getConstant(Idx, getPointerTy()));
3605 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3606 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3607 DAG.getConstant(Idx+1, getPointerTy()));
3608 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3616 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3617 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3618 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3621 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3622 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3623 // one of the above mentioned nodes. It has to be wrapped because otherwise
3624 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3625 // be used to form addressing mode. These wrapped nodes will be selected
3628 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3629 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3630 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3632 CP->getAlignment());
3633 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3634 // With PIC, the address is actually $g + Offset.
3635 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3636 !Subtarget->isPICStyleRIPRel()) {
3637 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3638 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3646 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3647 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3648 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3649 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3650 // With PIC, the address is actually $g + Offset.
3651 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3652 !Subtarget->isPICStyleRIPRel()) {
3653 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3654 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3658 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3659 // load the value at address GV, not the value of GV itself. This means that
3660 // the GlobalAddress must be in the base or index register of the address, not
3661 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3662 // The same applies for external symbols during PIC codegen
3663 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3664 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3669 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3671 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3672 const MVT::ValueType PtrVT) {
3674 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
3675 DAG.getNode(X86ISD::GlobalBaseReg,
3677 InFlag = Chain.getValue(1);
3679 // emit leal symbol@TLSGD(,%ebx,1), %eax
3680 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3681 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3682 GA->getValueType(0),
3684 SDOperand Ops[] = { Chain, TGA, InFlag };
3685 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3686 InFlag = Result.getValue(2);
3687 Chain = Result.getValue(1);
3689 // call ___tls_get_addr. This function receives its argument in
3690 // the register EAX.
3691 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
3692 InFlag = Chain.getValue(1);
3694 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3695 SDOperand Ops1[] = { Chain,
3696 DAG.getTargetExternalSymbol("___tls_get_addr",
3698 DAG.getRegister(X86::EAX, PtrVT),
3699 DAG.getRegister(X86::EBX, PtrVT),
3701 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
3702 InFlag = Chain.getValue(1);
3704 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
3707 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
3708 // "local exec" model.
3710 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3711 const MVT::ValueType PtrVT) {
3712 // Get the Thread Pointer
3713 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
3714 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
3716 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3717 GA->getValueType(0),
3719 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
3721 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
3722 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset, NULL, 0);
3724 // The address of the thread local variable is the add of the thread
3725 // pointer with the offset of the variable.
3726 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
3730 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
3731 // TODO: implement the "local dynamic" model
3732 // TODO: implement the "initial exec"model for pic executables
3733 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
3734 "TLS not implemented for non-ELF and 64-bit targets");
3735 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3736 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
3737 // otherwise use the "Local Exec"TLS Model
3738 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
3739 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
3741 return LowerToTLSExecModel(GA, DAG, getPointerTy());
3745 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3746 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3747 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3748 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3749 // With PIC, the address is actually $g + Offset.
3750 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3751 !Subtarget->isPICStyleRIPRel()) {
3752 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3753 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3760 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3761 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3762 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3763 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3764 // With PIC, the address is actually $g + Offset.
3765 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3766 !Subtarget->isPICStyleRIPRel()) {
3767 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3768 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3775 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
3776 /// take a 2 x i32 value to shift plus a shift amount.
3777 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3778 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3779 "Not an i64 shift!");
3780 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3781 SDOperand ShOpLo = Op.getOperand(0);
3782 SDOperand ShOpHi = Op.getOperand(1);
3783 SDOperand ShAmt = Op.getOperand(2);
3784 SDOperand Tmp1 = isSRA ?
3785 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3786 DAG.getConstant(0, MVT::i32);
3788 SDOperand Tmp2, Tmp3;
3789 if (Op.getOpcode() == ISD::SHL_PARTS) {
3790 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3791 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3793 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3794 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3797 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3798 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3799 DAG.getConstant(32, MVT::i8));
3800 SDOperand Cond = DAG.getNode(X86ISD::CMP, MVT::i32,
3801 AndNode, DAG.getConstant(0, MVT::i8));
3804 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3805 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3806 SmallVector<SDOperand, 4> Ops;
3807 if (Op.getOpcode() == ISD::SHL_PARTS) {
3808 Ops.push_back(Tmp2);
3809 Ops.push_back(Tmp3);
3811 Ops.push_back(Cond);
3812 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3815 Ops.push_back(Tmp3);
3816 Ops.push_back(Tmp1);
3818 Ops.push_back(Cond);
3819 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3821 Ops.push_back(Tmp2);
3822 Ops.push_back(Tmp3);
3824 Ops.push_back(Cond);
3825 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3828 Ops.push_back(Tmp3);
3829 Ops.push_back(Tmp1);
3831 Ops.push_back(Cond);
3832 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3835 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3839 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3842 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3843 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3844 Op.getOperand(0).getValueType() >= MVT::i16 &&
3845 "Unknown SINT_TO_FP to lower!");
3848 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3849 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3850 MachineFunction &MF = DAG.getMachineFunction();
3851 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3852 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3853 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3854 StackSlot, NULL, 0);
3856 // These are really Legal; caller falls through into that case.
3857 if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f32 && X86ScalarSSEf32)
3859 if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f64 && X86ScalarSSEf64)
3861 if (SrcVT==MVT::i64 && Op.getValueType() != MVT::f80 &&
3862 Subtarget->is64Bit())
3867 bool useSSE = (X86ScalarSSEf32 && Op.getValueType() == MVT::f32) ||
3868 (X86ScalarSSEf64 && Op.getValueType() == MVT::f64);
3870 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3872 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
3873 SmallVector<SDOperand, 8> Ops;
3874 Ops.push_back(Chain);
3875 Ops.push_back(StackSlot);
3876 Ops.push_back(DAG.getValueType(SrcVT));
3877 Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3878 Tys, &Ops[0], Ops.size());
3881 Chain = Result.getValue(1);
3882 SDOperand InFlag = Result.getValue(2);
3884 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3885 // shouldn't be necessary except that RFP cannot be live across
3886 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3887 MachineFunction &MF = DAG.getMachineFunction();
3888 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3889 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3890 Tys = DAG.getVTList(MVT::Other);
3891 SmallVector<SDOperand, 8> Ops;
3892 Ops.push_back(Chain);
3893 Ops.push_back(Result);
3894 Ops.push_back(StackSlot);
3895 Ops.push_back(DAG.getValueType(Op.getValueType()));
3896 Ops.push_back(InFlag);
3897 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3898 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3904 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3905 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3906 "Unknown FP_TO_SINT to lower!");
3909 // These are really Legal.
3910 if (Op.getValueType() == MVT::i32 &&
3911 X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32)
3913 if (Op.getValueType() == MVT::i32 &&
3914 X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)
3916 if (Subtarget->is64Bit() &&
3917 Op.getValueType() == MVT::i64 &&
3918 Op.getOperand(0).getValueType() != MVT::f80)
3921 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3923 MachineFunction &MF = DAG.getMachineFunction();
3924 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3925 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3926 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3928 switch (Op.getValueType()) {
3929 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3930 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3931 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3932 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3935 SDOperand Chain = DAG.getEntryNode();
3936 SDOperand Value = Op.getOperand(0);
3937 if ((X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32) ||
3938 (X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)) {
3939 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3940 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3941 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
3943 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3945 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3946 Chain = Value.getValue(1);
3947 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3948 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3951 // Build the FP_TO_INT*_IN_MEM
3952 SDOperand Ops[] = { Chain, Value, StackSlot };
3953 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3955 // Load the result. If this is an i64 load on an x86-32 host, expand the
3957 if (Op.getValueType() != MVT::i64 || Subtarget->is64Bit())
3958 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3960 SDOperand Lo = DAG.getLoad(MVT::i32, FIST, StackSlot, NULL, 0);
3961 StackSlot = DAG.getNode(ISD::ADD, StackSlot.getValueType(), StackSlot,
3962 DAG.getConstant(StackSlot.getValueType(), 4));
3963 SDOperand Hi = DAG.getLoad(MVT::i32, FIST, StackSlot, NULL, 0);
3966 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
3969 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3970 MVT::ValueType VT = Op.getValueType();
3971 MVT::ValueType EltVT = VT;
3972 if (MVT::isVector(VT))
3973 EltVT = MVT::getVectorElementType(VT);
3974 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
3975 std::vector<Constant*> CV;
3976 if (EltVT == MVT::f64) {
3977 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, ~(1ULL << 63))));
3981 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, ~(1U << 31))));
3987 Constant *C = ConstantVector::get(CV);
3988 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
3989 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
3991 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3994 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3995 MVT::ValueType VT = Op.getValueType();
3996 MVT::ValueType EltVT = VT;
3997 unsigned EltNum = 1;
3998 if (MVT::isVector(VT)) {
3999 EltVT = MVT::getVectorElementType(VT);
4000 EltNum = MVT::getVectorNumElements(VT);
4002 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4003 std::vector<Constant*> CV;
4004 if (EltVT == MVT::f64) {
4005 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, 1ULL << 63)));
4009 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, 1U << 31)));
4015 Constant *C = ConstantVector::get(CV);
4016 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4017 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4019 if (MVT::isVector(VT)) {
4020 return DAG.getNode(ISD::BIT_CONVERT, VT,
4021 DAG.getNode(ISD::XOR, MVT::v2i64,
4022 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4023 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4025 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4029 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
4030 SDOperand Op0 = Op.getOperand(0);
4031 SDOperand Op1 = Op.getOperand(1);
4032 MVT::ValueType VT = Op.getValueType();
4033 MVT::ValueType SrcVT = Op1.getValueType();
4034 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
4036 // If second operand is smaller, extend it first.
4037 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
4038 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4040 SrcTy = MVT::getTypeForValueType(SrcVT);
4043 // First get the sign bit of second operand.
4044 std::vector<Constant*> CV;
4045 if (SrcVT == MVT::f64) {
4046 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 1ULL << 63))));
4047 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4049 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 1U << 31))));
4050 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4051 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4052 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4054 Constant *C = ConstantVector::get(CV);
4055 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4056 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx, NULL, 0,
4058 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4060 // Shift sign bit right or left if the two operands have different types.
4061 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4062 // Op0 is MVT::f32, Op1 is MVT::f64.
4063 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4064 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4065 DAG.getConstant(32, MVT::i32));
4066 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4067 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4068 DAG.getConstant(0, getPointerTy()));
4071 // Clear first operand sign bit.
4073 if (VT == MVT::f64) {
4074 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, ~(1ULL << 63)))));
4075 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4077 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, ~(1U << 31)))));
4078 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4079 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4080 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4082 C = ConstantVector::get(CV);
4083 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4084 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4086 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4088 // Or the value with the sign bit.
4089 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4092 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
4093 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4095 SDOperand Op0 = Op.getOperand(0);
4096 SDOperand Op1 = Op.getOperand(1);
4097 SDOperand CC = Op.getOperand(2);
4098 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4099 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
4102 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4104 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4105 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4106 DAG.getConstant(X86CC, MVT::i8), Cond);
4109 assert(isFP && "Illegal integer SetCC!");
4111 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4112 switch (SetCCOpcode) {
4113 default: assert(false && "Illegal floating point SetCC!");
4114 case ISD::SETOEQ: { // !PF & ZF
4115 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4116 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4117 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4118 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4119 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4121 case ISD::SETUNE: { // PF | !ZF
4122 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4123 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4124 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4125 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4126 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4132 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
4133 bool addTest = true;
4134 SDOperand Cond = Op.getOperand(0);
4137 if (Cond.getOpcode() == ISD::SETCC)
4138 Cond = LowerSETCC(Cond, DAG);
4140 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4141 // setting operand in place of the X86ISD::SETCC.
4142 if (Cond.getOpcode() == X86ISD::SETCC) {
4143 CC = Cond.getOperand(0);
4145 SDOperand Cmp = Cond.getOperand(1);
4146 unsigned Opc = Cmp.getOpcode();
4147 MVT::ValueType VT = Op.getValueType();
4148 bool IllegalFPCMov = false;
4149 if (VT == MVT::f32 && !X86ScalarSSEf32)
4150 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4151 else if (VT == MVT::f64 && !X86ScalarSSEf64)
4152 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4153 else if (VT == MVT::f80)
4154 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4155 if ((Opc == X86ISD::CMP ||
4156 Opc == X86ISD::COMI ||
4157 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4164 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4165 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4168 const MVT::ValueType *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4170 SmallVector<SDOperand, 4> Ops;
4171 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4172 // condition is true.
4173 Ops.push_back(Op.getOperand(2));
4174 Ops.push_back(Op.getOperand(1));
4176 Ops.push_back(Cond);
4177 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4180 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
4181 bool addTest = true;
4182 SDOperand Chain = Op.getOperand(0);
4183 SDOperand Cond = Op.getOperand(1);
4184 SDOperand Dest = Op.getOperand(2);
4187 if (Cond.getOpcode() == ISD::SETCC)
4188 Cond = LowerSETCC(Cond, DAG);
4190 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4191 // setting operand in place of the X86ISD::SETCC.
4192 if (Cond.getOpcode() == X86ISD::SETCC) {
4193 CC = Cond.getOperand(0);
4195 SDOperand Cmp = Cond.getOperand(1);
4196 unsigned Opc = Cmp.getOpcode();
4197 if (Opc == X86ISD::CMP ||
4198 Opc == X86ISD::COMI ||
4199 Opc == X86ISD::UCOMI) {
4206 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4207 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4209 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
4210 Chain, Op.getOperand(2), CC, Cond);
4213 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
4214 unsigned CallingConv = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4215 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
4217 if (Subtarget->is64Bit())
4218 if(CallingConv==CallingConv::Fast && isTailCall && PerformTailCallOpt)
4219 return LowerX86_TailCallTo(Op, DAG, CallingConv);
4221 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
4223 switch (CallingConv) {
4225 assert(0 && "Unsupported calling convention");
4226 case CallingConv::Fast:
4227 if (isTailCall && PerformTailCallOpt)
4228 return LowerX86_TailCallTo(Op, DAG, CallingConv);
4230 return LowerCCCCallTo(Op,DAG, CallingConv);
4231 case CallingConv::C:
4232 case CallingConv::X86_StdCall:
4233 return LowerCCCCallTo(Op, DAG, CallingConv);
4234 case CallingConv::X86_FastCall:
4235 return LowerFastCCCallTo(Op, DAG, CallingConv);
4240 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
4241 // Calls to _alloca is needed to probe the stack when allocating more than 4k
4242 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
4243 // that the guard pages used by the OS virtual memory manager are allocated in
4244 // correct sequence.
4246 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
4247 SelectionDAG &DAG) {
4248 assert(Subtarget->isTargetCygMing() &&
4249 "This should be used only on Cygwin/Mingw targets");
4252 SDOperand Chain = Op.getOperand(0);
4253 SDOperand Size = Op.getOperand(1);
4254 // FIXME: Ensure alignment here
4258 MVT::ValueType IntPtr = getPointerTy();
4259 MVT::ValueType SPTy = (Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
4261 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
4262 Flag = Chain.getValue(1);
4264 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4265 SDOperand Ops[] = { Chain,
4266 DAG.getTargetExternalSymbol("_alloca", IntPtr),
4267 DAG.getRegister(X86::EAX, IntPtr),
4269 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
4270 Flag = Chain.getValue(1);
4272 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
4274 std::vector<MVT::ValueType> Tys;
4275 Tys.push_back(SPTy);
4276 Tys.push_back(MVT::Other);
4277 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
4278 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
4282 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
4283 MachineFunction &MF = DAG.getMachineFunction();
4284 const Function* Fn = MF.getFunction();
4285 if (Fn->hasExternalLinkage() &&
4286 Subtarget->isTargetCygMing() &&
4287 Fn->getName() == "main")
4288 MF.getInfo<X86MachineFunctionInfo>()->setForceFramePointer(true);
4290 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4291 if (Subtarget->is64Bit())
4292 return LowerX86_64CCCArguments(Op, DAG);
4296 assert(0 && "Unsupported calling convention");
4297 case CallingConv::Fast:
4298 return LowerCCCArguments(Op,DAG, true);
4300 case CallingConv::C:
4301 return LowerCCCArguments(Op, DAG);
4302 case CallingConv::X86_StdCall:
4303 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(StdCall);
4304 return LowerCCCArguments(Op, DAG, true);
4305 case CallingConv::X86_FastCall:
4306 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(FastCall);
4307 return LowerFastCCArguments(Op, DAG);
4311 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4312 SDOperand InFlag(0, 0);
4313 SDOperand Chain = Op.getOperand(0);
4315 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4316 if (Align == 0) Align = 1;
4318 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4319 // If not DWORD aligned or size is more than the threshold, call memset.
4320 // The libc version is likely to be faster for these cases. It can use the
4321 // address value and run time information about the CPU.
4322 if ((Align & 3) != 0 ||
4323 (I && I->getValue() > Subtarget->getMinRepStrSizeThreshold())) {
4324 MVT::ValueType IntPtr = getPointerTy();
4325 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4326 TargetLowering::ArgListTy Args;
4327 TargetLowering::ArgListEntry Entry;
4328 Entry.Node = Op.getOperand(1);
4329 Entry.Ty = IntPtrTy;
4330 Args.push_back(Entry);
4331 // Extend the unsigned i8 argument to be an int value for the call.
4332 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4333 Entry.Ty = IntPtrTy;
4334 Args.push_back(Entry);
4335 Entry.Node = Op.getOperand(3);
4336 Args.push_back(Entry);
4337 std::pair<SDOperand,SDOperand> CallResult =
4338 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4339 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4340 return CallResult.second;
4345 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4346 unsigned BytesLeft = 0;
4347 bool TwoRepStos = false;
4350 uint64_t Val = ValC->getValue() & 255;
4352 // If the value is a constant, then we can potentially use larger sets.
4353 switch (Align & 3) {
4354 case 2: // WORD aligned
4357 Val = (Val << 8) | Val;
4359 case 0: // DWORD aligned
4362 Val = (Val << 8) | Val;
4363 Val = (Val << 16) | Val;
4364 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4367 Val = (Val << 32) | Val;
4370 default: // Byte aligned
4373 Count = Op.getOperand(3);
4377 if (AVT > MVT::i8) {
4379 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4380 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4381 BytesLeft = I->getValue() % UBytes;
4383 assert(AVT >= MVT::i32 &&
4384 "Do not use rep;stos if not at least DWORD aligned");
4385 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4386 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4391 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4393 InFlag = Chain.getValue(1);
4396 Count = Op.getOperand(3);
4397 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4398 InFlag = Chain.getValue(1);
4401 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4403 InFlag = Chain.getValue(1);
4404 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4405 Op.getOperand(1), InFlag);
4406 InFlag = Chain.getValue(1);
4408 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4409 SmallVector<SDOperand, 8> Ops;
4410 Ops.push_back(Chain);
4411 Ops.push_back(DAG.getValueType(AVT));
4412 Ops.push_back(InFlag);
4413 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4416 InFlag = Chain.getValue(1);
4417 Count = Op.getOperand(3);
4418 MVT::ValueType CVT = Count.getValueType();
4419 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4420 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4421 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4423 InFlag = Chain.getValue(1);
4424 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4426 Ops.push_back(Chain);
4427 Ops.push_back(DAG.getValueType(MVT::i8));
4428 Ops.push_back(InFlag);
4429 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4430 } else if (BytesLeft) {
4431 // Issue stores for the last 1 - 7 bytes.
4433 unsigned Val = ValC->getValue() & 255;
4434 unsigned Offset = I->getValue() - BytesLeft;
4435 SDOperand DstAddr = Op.getOperand(1);
4436 MVT::ValueType AddrVT = DstAddr.getValueType();
4437 if (BytesLeft >= 4) {
4438 Val = (Val << 8) | Val;
4439 Val = (Val << 16) | Val;
4440 Value = DAG.getConstant(Val, MVT::i32);
4441 Chain = DAG.getStore(Chain, Value,
4442 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4443 DAG.getConstant(Offset, AddrVT)),
4448 if (BytesLeft >= 2) {
4449 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4450 Chain = DAG.getStore(Chain, Value,
4451 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4452 DAG.getConstant(Offset, AddrVT)),
4457 if (BytesLeft == 1) {
4458 Value = DAG.getConstant(Val, MVT::i8);
4459 Chain = DAG.getStore(Chain, Value,
4460 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4461 DAG.getConstant(Offset, AddrVT)),
4469 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
4470 SDOperand ChainOp = Op.getOperand(0);
4471 SDOperand DestOp = Op.getOperand(1);
4472 SDOperand SourceOp = Op.getOperand(2);
4473 SDOperand CountOp = Op.getOperand(3);
4474 SDOperand AlignOp = Op.getOperand(4);
4475 unsigned Align = (unsigned)cast<ConstantSDNode>(AlignOp)->getValue();
4476 if (Align == 0) Align = 1;
4478 // The libc version is likely to be faster for the following cases. It can
4479 // use the address value and run time information about the CPU.
4480 // With glibc 2.6.1 on a core 2, coping an array of 100M longs was 30% faster
4482 // If not DWORD aligned, call memcpy.
4483 if ((Align & 3) != 0)
4484 return LowerMEMCPYCall(ChainOp, DestOp, SourceOp, CountOp, DAG);
4486 // If size is unknown, call memcpy.
4487 ConstantSDNode *I = dyn_cast<ConstantSDNode>(CountOp);
4489 return LowerMEMCPYCall(ChainOp, DestOp, SourceOp, CountOp, DAG);
4491 // If size is more than the threshold, call memcpy.
4492 unsigned Size = I->getValue();
4493 if (Size > Subtarget->getMinRepStrSizeThreshold())
4494 return LowerMEMCPYCall(ChainOp, DestOp, SourceOp, CountOp, DAG);
4496 return LowerMEMCPYInline(ChainOp, DestOp, SourceOp, Size, Align, DAG);
4499 SDOperand X86TargetLowering::LowerMEMCPYCall(SDOperand Chain,
4503 SelectionDAG &DAG) {
4504 MVT::ValueType IntPtr = getPointerTy();
4505 TargetLowering::ArgListTy Args;
4506 TargetLowering::ArgListEntry Entry;
4507 Entry.Ty = getTargetData()->getIntPtrType();
4508 Entry.Node = Dest; Args.push_back(Entry);
4509 Entry.Node = Source; Args.push_back(Entry);
4510 Entry.Node = Count; Args.push_back(Entry);
4511 std::pair<SDOperand,SDOperand> CallResult =
4512 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4513 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
4514 return CallResult.second;
4517 SDOperand X86TargetLowering::LowerMEMCPYInline(SDOperand Chain,
4522 SelectionDAG &DAG) {
4524 unsigned BytesLeft = 0;
4525 switch (Align & 3) {
4526 case 2: // WORD aligned
4529 case 0: // DWORD aligned
4531 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4534 default: // Byte aligned
4539 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4540 SDOperand Count = DAG.getConstant(Size / UBytes, getPointerTy());
4541 BytesLeft = Size % UBytes;
4543 SDOperand InFlag(0, 0);
4544 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4546 InFlag = Chain.getValue(1);
4547 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4549 InFlag = Chain.getValue(1);
4550 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4552 InFlag = Chain.getValue(1);
4554 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4555 SmallVector<SDOperand, 8> Ops;
4556 Ops.push_back(Chain);
4557 Ops.push_back(DAG.getValueType(AVT));
4558 Ops.push_back(InFlag);
4559 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4562 // Issue loads and stores for the last 1 - 7 bytes.
4563 unsigned Offset = Size - BytesLeft;
4564 SDOperand DstAddr = Dest;
4565 MVT::ValueType DstVT = DstAddr.getValueType();
4566 SDOperand SrcAddr = Source;
4567 MVT::ValueType SrcVT = SrcAddr.getValueType();
4569 if (BytesLeft >= 4) {
4570 Value = DAG.getLoad(MVT::i32, Chain,
4571 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4572 DAG.getConstant(Offset, SrcVT)),
4574 Chain = Value.getValue(1);
4575 Chain = DAG.getStore(Chain, Value,
4576 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4577 DAG.getConstant(Offset, DstVT)),
4582 if (BytesLeft >= 2) {
4583 Value = DAG.getLoad(MVT::i16, Chain,
4584 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4585 DAG.getConstant(Offset, SrcVT)),
4587 Chain = Value.getValue(1);
4588 Chain = DAG.getStore(Chain, Value,
4589 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4590 DAG.getConstant(Offset, DstVT)),
4596 if (BytesLeft == 1) {
4597 Value = DAG.getLoad(MVT::i8, Chain,
4598 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4599 DAG.getConstant(Offset, SrcVT)),
4601 Chain = Value.getValue(1);
4602 Chain = DAG.getStore(Chain, Value,
4603 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4604 DAG.getConstant(Offset, DstVT)),
4613 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4614 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4615 SDOperand TheOp = Op.getOperand(0);
4616 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
4617 if (Subtarget->is64Bit()) {
4619 DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4620 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4621 MVT::i64, Copy1.getValue(2));
4622 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4623 DAG.getConstant(32, MVT::i8));
4625 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
4628 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4629 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
4632 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4633 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4634 MVT::i32, Copy1.getValue(2));
4635 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
4636 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4637 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
4640 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4641 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4643 if (!Subtarget->is64Bit()) {
4644 // vastart just stores the address of the VarArgsFrameIndex slot into the
4645 // memory location argument.
4646 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4647 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4652 // gp_offset (0 - 6 * 8)
4653 // fp_offset (48 - 48 + 8 * 16)
4654 // overflow_arg_area (point to parameters coming in memory).
4656 SmallVector<SDOperand, 8> MemOps;
4657 SDOperand FIN = Op.getOperand(1);
4659 SDOperand Store = DAG.getStore(Op.getOperand(0),
4660 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4661 FIN, SV->getValue(), SV->getOffset());
4662 MemOps.push_back(Store);
4665 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4666 DAG.getConstant(4, getPointerTy()));
4667 Store = DAG.getStore(Op.getOperand(0),
4668 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4669 FIN, SV->getValue(), SV->getOffset());
4670 MemOps.push_back(Store);
4672 // Store ptr to overflow_arg_area
4673 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4674 DAG.getConstant(4, getPointerTy()));
4675 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4676 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4678 MemOps.push_back(Store);
4680 // Store ptr to reg_save_area.
4681 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4682 DAG.getConstant(8, getPointerTy()));
4683 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4684 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4686 MemOps.push_back(Store);
4687 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4690 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
4691 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
4692 SDOperand Chain = Op.getOperand(0);
4693 SDOperand DstPtr = Op.getOperand(1);
4694 SDOperand SrcPtr = Op.getOperand(2);
4695 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
4696 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
4698 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
4699 SrcSV->getValue(), SrcSV->getOffset());
4700 Chain = SrcPtr.getValue(1);
4701 for (unsigned i = 0; i < 3; ++i) {
4702 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
4703 SrcSV->getValue(), SrcSV->getOffset());
4704 Chain = Val.getValue(1);
4705 Chain = DAG.getStore(Chain, Val, DstPtr,
4706 DstSV->getValue(), DstSV->getOffset());
4709 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
4710 DAG.getConstant(8, getPointerTy()));
4711 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
4712 DAG.getConstant(8, getPointerTy()));
4718 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4719 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4721 default: return SDOperand(); // Don't custom lower most intrinsics.
4722 // Comparison intrinsics.
4723 case Intrinsic::x86_sse_comieq_ss:
4724 case Intrinsic::x86_sse_comilt_ss:
4725 case Intrinsic::x86_sse_comile_ss:
4726 case Intrinsic::x86_sse_comigt_ss:
4727 case Intrinsic::x86_sse_comige_ss:
4728 case Intrinsic::x86_sse_comineq_ss:
4729 case Intrinsic::x86_sse_ucomieq_ss:
4730 case Intrinsic::x86_sse_ucomilt_ss:
4731 case Intrinsic::x86_sse_ucomile_ss:
4732 case Intrinsic::x86_sse_ucomigt_ss:
4733 case Intrinsic::x86_sse_ucomige_ss:
4734 case Intrinsic::x86_sse_ucomineq_ss:
4735 case Intrinsic::x86_sse2_comieq_sd:
4736 case Intrinsic::x86_sse2_comilt_sd:
4737 case Intrinsic::x86_sse2_comile_sd:
4738 case Intrinsic::x86_sse2_comigt_sd:
4739 case Intrinsic::x86_sse2_comige_sd:
4740 case Intrinsic::x86_sse2_comineq_sd:
4741 case Intrinsic::x86_sse2_ucomieq_sd:
4742 case Intrinsic::x86_sse2_ucomilt_sd:
4743 case Intrinsic::x86_sse2_ucomile_sd:
4744 case Intrinsic::x86_sse2_ucomigt_sd:
4745 case Intrinsic::x86_sse2_ucomige_sd:
4746 case Intrinsic::x86_sse2_ucomineq_sd: {
4748 ISD::CondCode CC = ISD::SETCC_INVALID;
4751 case Intrinsic::x86_sse_comieq_ss:
4752 case Intrinsic::x86_sse2_comieq_sd:
4756 case Intrinsic::x86_sse_comilt_ss:
4757 case Intrinsic::x86_sse2_comilt_sd:
4761 case Intrinsic::x86_sse_comile_ss:
4762 case Intrinsic::x86_sse2_comile_sd:
4766 case Intrinsic::x86_sse_comigt_ss:
4767 case Intrinsic::x86_sse2_comigt_sd:
4771 case Intrinsic::x86_sse_comige_ss:
4772 case Intrinsic::x86_sse2_comige_sd:
4776 case Intrinsic::x86_sse_comineq_ss:
4777 case Intrinsic::x86_sse2_comineq_sd:
4781 case Intrinsic::x86_sse_ucomieq_ss:
4782 case Intrinsic::x86_sse2_ucomieq_sd:
4783 Opc = X86ISD::UCOMI;
4786 case Intrinsic::x86_sse_ucomilt_ss:
4787 case Intrinsic::x86_sse2_ucomilt_sd:
4788 Opc = X86ISD::UCOMI;
4791 case Intrinsic::x86_sse_ucomile_ss:
4792 case Intrinsic::x86_sse2_ucomile_sd:
4793 Opc = X86ISD::UCOMI;
4796 case Intrinsic::x86_sse_ucomigt_ss:
4797 case Intrinsic::x86_sse2_ucomigt_sd:
4798 Opc = X86ISD::UCOMI;
4801 case Intrinsic::x86_sse_ucomige_ss:
4802 case Intrinsic::x86_sse2_ucomige_sd:
4803 Opc = X86ISD::UCOMI;
4806 case Intrinsic::x86_sse_ucomineq_ss:
4807 case Intrinsic::x86_sse2_ucomineq_sd:
4808 Opc = X86ISD::UCOMI;
4814 SDOperand LHS = Op.getOperand(1);
4815 SDOperand RHS = Op.getOperand(2);
4816 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4818 SDOperand Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
4819 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
4820 DAG.getConstant(X86CC, MVT::i8), Cond);
4821 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4826 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4827 // Depths > 0 not supported yet!
4828 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4831 // Just load the return address
4832 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4833 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4836 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4837 // Depths > 0 not supported yet!
4838 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4841 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4842 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4843 DAG.getConstant(4, getPointerTy()));
4846 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
4847 SelectionDAG &DAG) {
4848 // Is not yet supported on x86-64
4849 if (Subtarget->is64Bit())
4852 return DAG.getConstant(8, getPointerTy());
4855 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
4857 assert(!Subtarget->is64Bit() &&
4858 "Lowering of eh_return builtin is not supported yet on x86-64");
4860 MachineFunction &MF = DAG.getMachineFunction();
4861 SDOperand Chain = Op.getOperand(0);
4862 SDOperand Offset = Op.getOperand(1);
4863 SDOperand Handler = Op.getOperand(2);
4865 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
4868 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
4869 DAG.getConstant(-4UL, getPointerTy()));
4870 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
4871 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
4872 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
4873 MF.addLiveOut(X86::ECX);
4875 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
4876 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
4879 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
4880 SelectionDAG &DAG) {
4881 SDOperand Root = Op.getOperand(0);
4882 SDOperand Trmp = Op.getOperand(1); // trampoline
4883 SDOperand FPtr = Op.getOperand(2); // nested function
4884 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
4886 SrcValueSDNode *TrmpSV = cast<SrcValueSDNode>(Op.getOperand(4));
4888 if (Subtarget->is64Bit()) {
4889 return SDOperand(); // not yet supported
4891 Function *Func = (Function *)
4892 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
4893 unsigned CC = Func->getCallingConv();
4898 assert(0 && "Unsupported calling convention");
4899 case CallingConv::C:
4900 case CallingConv::X86_StdCall: {
4901 // Pass 'nest' parameter in ECX.
4902 // Must be kept in sync with X86CallingConv.td
4905 // Check that ECX wasn't needed by an 'inreg' parameter.
4906 const FunctionType *FTy = Func->getFunctionType();
4907 const ParamAttrsList *Attrs = FTy->getParamAttrs();
4909 if (Attrs && !Func->isVarArg()) {
4910 unsigned InRegCount = 0;
4913 for (FunctionType::param_iterator I = FTy->param_begin(),
4914 E = FTy->param_end(); I != E; ++I, ++Idx)
4915 if (Attrs->paramHasAttr(Idx, ParamAttr::InReg))
4916 // FIXME: should only count parameters that are lowered to integers.
4917 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
4919 if (InRegCount > 2) {
4920 cerr << "Nest register in use - reduce number of inreg parameters!\n";
4926 case CallingConv::X86_FastCall:
4927 // Pass 'nest' parameter in EAX.
4928 // Must be kept in sync with X86CallingConv.td
4933 const X86InstrInfo *TII =
4934 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
4936 SDOperand OutChains[4];
4937 SDOperand Addr, Disp;
4939 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
4940 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
4942 unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
4943 unsigned char N86Reg = ((X86RegisterInfo&)RegInfo).getX86RegNum(NestReg);
4944 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
4945 Trmp, TrmpSV->getValue(), TrmpSV->getOffset());
4947 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
4948 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
4949 TrmpSV->getOffset() + 1, false, 1);
4951 unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
4952 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
4953 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
4954 TrmpSV->getValue() + 5, TrmpSV->getOffset());
4956 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
4957 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpSV->getValue(),
4958 TrmpSV->getOffset() + 6, false, 1);
4961 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
4962 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
4966 /// LowerOperation - Provide custom lowering hooks for some operations.
4968 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4969 switch (Op.getOpcode()) {
4970 default: assert(0 && "Should not custom lower this!");
4971 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4972 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4973 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4974 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4975 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4976 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4977 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4978 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4979 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4980 case ISD::SHL_PARTS:
4981 case ISD::SRA_PARTS:
4982 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4983 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4984 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4985 case ISD::FABS: return LowerFABS(Op, DAG);
4986 case ISD::FNEG: return LowerFNEG(Op, DAG);
4987 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4988 case ISD::SETCC: return LowerSETCC(Op, DAG);
4989 case ISD::SELECT: return LowerSELECT(Op, DAG);
4990 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4991 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4992 case ISD::CALL: return LowerCALL(Op, DAG);
4993 case ISD::RET: return LowerRET(Op, DAG);
4994 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4995 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4996 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4997 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4998 case ISD::VASTART: return LowerVASTART(Op, DAG);
4999 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
5000 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5001 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5002 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5003 case ISD::FRAME_TO_ARGS_OFFSET:
5004 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
5005 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
5006 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
5007 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
5012 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
5014 default: return NULL;
5015 case X86ISD::SHLD: return "X86ISD::SHLD";
5016 case X86ISD::SHRD: return "X86ISD::SHRD";
5017 case X86ISD::FAND: return "X86ISD::FAND";
5018 case X86ISD::FOR: return "X86ISD::FOR";
5019 case X86ISD::FXOR: return "X86ISD::FXOR";
5020 case X86ISD::FSRL: return "X86ISD::FSRL";
5021 case X86ISD::FILD: return "X86ISD::FILD";
5022 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
5023 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
5024 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
5025 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
5026 case X86ISD::FLD: return "X86ISD::FLD";
5027 case X86ISD::FST: return "X86ISD::FST";
5028 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
5029 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
5030 case X86ISD::CALL: return "X86ISD::CALL";
5031 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
5032 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
5033 case X86ISD::CMP: return "X86ISD::CMP";
5034 case X86ISD::COMI: return "X86ISD::COMI";
5035 case X86ISD::UCOMI: return "X86ISD::UCOMI";
5036 case X86ISD::SETCC: return "X86ISD::SETCC";
5037 case X86ISD::CMOV: return "X86ISD::CMOV";
5038 case X86ISD::BRCOND: return "X86ISD::BRCOND";
5039 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
5040 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
5041 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
5042 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
5043 case X86ISD::Wrapper: return "X86ISD::Wrapper";
5044 case X86ISD::S2VEC: return "X86ISD::S2VEC";
5045 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
5046 case X86ISD::PINSRW: return "X86ISD::PINSRW";
5047 case X86ISD::FMAX: return "X86ISD::FMAX";
5048 case X86ISD::FMIN: return "X86ISD::FMIN";
5049 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
5050 case X86ISD::FRCP: return "X86ISD::FRCP";
5051 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
5052 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
5053 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
5054 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
5058 // isLegalAddressingMode - Return true if the addressing mode represented
5059 // by AM is legal for this target, for a load/store of the specified type.
5060 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5061 const Type *Ty) const {
5062 // X86 supports extremely general addressing modes.
5064 // X86 allows a sign-extended 32-bit immediate field as a displacement.
5065 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
5069 // We can only fold this if we don't need an extra load.
5070 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
5073 // X86-64 only supports addr of globals in small code model.
5074 if (Subtarget->is64Bit()) {
5075 if (getTargetMachine().getCodeModel() != CodeModel::Small)
5077 // If lower 4G is not available, then we must use rip-relative addressing.
5078 if (AM.BaseOffs || AM.Scale > 1)
5089 // These scales always work.
5094 // These scales are formed with basereg+scalereg. Only accept if there is
5099 default: // Other stuff never works.
5107 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5108 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5109 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5110 /// are assumed to be legal.
5112 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
5113 // Only do shuffles on 128-bit vector types for now.
5114 if (MVT::getSizeInBits(VT) == 64) return false;
5115 return (Mask.Val->getNumOperands() <= 4 ||
5116 isIdentityMask(Mask.Val) ||
5117 isIdentityMask(Mask.Val, true) ||
5118 isSplatMask(Mask.Val) ||
5119 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
5120 X86::isUNPCKLMask(Mask.Val) ||
5121 X86::isUNPCKHMask(Mask.Val) ||
5122 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
5123 X86::isUNPCKH_v_undef_Mask(Mask.Val));
5126 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
5128 SelectionDAG &DAG) const {
5129 unsigned NumElts = BVOps.size();
5130 // Only do shuffles on 128-bit vector types for now.
5131 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
5132 if (NumElts == 2) return true;
5134 return (isMOVLMask(&BVOps[0], 4) ||
5135 isCommutedMOVL(&BVOps[0], 4, true) ||
5136 isSHUFPMask(&BVOps[0], 4) ||
5137 isCommutedSHUFP(&BVOps[0], 4));
5142 //===----------------------------------------------------------------------===//
5143 // X86 Scheduler Hooks
5144 //===----------------------------------------------------------------------===//
5147 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
5148 MachineBasicBlock *BB) {
5149 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5150 switch (MI->getOpcode()) {
5151 default: assert(false && "Unexpected instr type to insert");
5152 case X86::CMOV_FR32:
5153 case X86::CMOV_FR64:
5154 case X86::CMOV_V4F32:
5155 case X86::CMOV_V2F64:
5156 case X86::CMOV_V2I64: {
5157 // To "insert" a SELECT_CC instruction, we actually have to insert the
5158 // diamond control-flow pattern. The incoming instruction knows the
5159 // destination vreg to set, the condition code register to branch on, the
5160 // true/false values to select between, and a branch opcode to use.
5161 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5162 ilist<MachineBasicBlock>::iterator It = BB;
5168 // cmpTY ccX, r1, r2
5170 // fallthrough --> copy0MBB
5171 MachineBasicBlock *thisMBB = BB;
5172 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
5173 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
5175 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
5176 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
5177 MachineFunction *F = BB->getParent();
5178 F->getBasicBlockList().insert(It, copy0MBB);
5179 F->getBasicBlockList().insert(It, sinkMBB);
5180 // Update machine-CFG edges by first adding all successors of the current
5181 // block to the new block which will contain the Phi node for the select.
5182 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
5183 e = BB->succ_end(); i != e; ++i)
5184 sinkMBB->addSuccessor(*i);
5185 // Next, remove all successors of the current block, and add the true
5186 // and fallthrough blocks as its successors.
5187 while(!BB->succ_empty())
5188 BB->removeSuccessor(BB->succ_begin());
5189 BB->addSuccessor(copy0MBB);
5190 BB->addSuccessor(sinkMBB);
5193 // %FalseValue = ...
5194 // # fallthrough to sinkMBB
5197 // Update machine-CFG edges
5198 BB->addSuccessor(sinkMBB);
5201 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5204 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
5205 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5206 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5208 delete MI; // The pseudo instruction is gone now.
5212 case X86::FP32_TO_INT16_IN_MEM:
5213 case X86::FP32_TO_INT32_IN_MEM:
5214 case X86::FP32_TO_INT64_IN_MEM:
5215 case X86::FP64_TO_INT16_IN_MEM:
5216 case X86::FP64_TO_INT32_IN_MEM:
5217 case X86::FP64_TO_INT64_IN_MEM:
5218 case X86::FP80_TO_INT16_IN_MEM:
5219 case X86::FP80_TO_INT32_IN_MEM:
5220 case X86::FP80_TO_INT64_IN_MEM: {
5221 // Change the floating point control register to use "round towards zero"
5222 // mode when truncating to an integer value.
5223 MachineFunction *F = BB->getParent();
5224 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
5225 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
5227 // Load the old value of the high byte of the control word...
5229 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
5230 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
5232 // Set the high part to be round to zero...
5233 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
5236 // Reload the modified control word now...
5237 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5239 // Restore the memory image of control word to original value
5240 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
5243 // Get the X86 opcode to use.
5245 switch (MI->getOpcode()) {
5246 default: assert(0 && "illegal opcode!");
5247 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
5248 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
5249 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
5250 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
5251 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
5252 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
5253 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
5254 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
5255 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
5259 MachineOperand &Op = MI->getOperand(0);
5260 if (Op.isRegister()) {
5261 AM.BaseType = X86AddressMode::RegBase;
5262 AM.Base.Reg = Op.getReg();
5264 AM.BaseType = X86AddressMode::FrameIndexBase;
5265 AM.Base.FrameIndex = Op.getFrameIndex();
5267 Op = MI->getOperand(1);
5268 if (Op.isImmediate())
5269 AM.Scale = Op.getImm();
5270 Op = MI->getOperand(2);
5271 if (Op.isImmediate())
5272 AM.IndexReg = Op.getImm();
5273 Op = MI->getOperand(3);
5274 if (Op.isGlobalAddress()) {
5275 AM.GV = Op.getGlobal();
5277 AM.Disp = Op.getImm();
5279 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
5280 .addReg(MI->getOperand(4).getReg());
5282 // Reload the original control word now.
5283 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5285 delete MI; // The pseudo instruction is gone now.
5291 //===----------------------------------------------------------------------===//
5292 // X86 Optimization Hooks
5293 //===----------------------------------------------------------------------===//
5295 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
5297 uint64_t &KnownZero,
5299 const SelectionDAG &DAG,
5300 unsigned Depth) const {
5301 unsigned Opc = Op.getOpcode();
5302 assert((Opc >= ISD::BUILTIN_OP_END ||
5303 Opc == ISD::INTRINSIC_WO_CHAIN ||
5304 Opc == ISD::INTRINSIC_W_CHAIN ||
5305 Opc == ISD::INTRINSIC_VOID) &&
5306 "Should use MaskedValueIsZero if you don't know whether Op"
5307 " is a target node!");
5309 KnownZero = KnownOne = 0; // Don't know anything.
5313 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
5318 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5319 /// element of the result of the vector shuffle.
5320 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
5321 MVT::ValueType VT = N->getValueType(0);
5322 SDOperand PermMask = N->getOperand(2);
5323 unsigned NumElems = PermMask.getNumOperands();
5324 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
5326 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5328 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5329 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
5330 SDOperand Idx = PermMask.getOperand(i);
5331 if (Idx.getOpcode() == ISD::UNDEF)
5332 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5333 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
5338 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
5339 /// node is a GlobalAddress + an offset.
5340 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
5341 unsigned Opc = N->getOpcode();
5342 if (Opc == X86ISD::Wrapper) {
5343 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
5344 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
5347 } else if (Opc == ISD::ADD) {
5348 SDOperand N1 = N->getOperand(0);
5349 SDOperand N2 = N->getOperand(1);
5350 if (isGAPlusOffset(N1.Val, GA, Offset)) {
5351 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
5353 Offset += V->getSignExtended();
5356 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
5357 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
5359 Offset += V->getSignExtended();
5367 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
5369 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
5370 MachineFrameInfo *MFI) {
5371 if (N->getOperand(0).Val != Base->getOperand(0).Val)
5374 SDOperand Loc = N->getOperand(1);
5375 SDOperand BaseLoc = Base->getOperand(1);
5376 if (Loc.getOpcode() == ISD::FrameIndex) {
5377 if (BaseLoc.getOpcode() != ISD::FrameIndex)
5379 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
5380 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
5381 int FS = MFI->getObjectSize(FI);
5382 int BFS = MFI->getObjectSize(BFI);
5383 if (FS != BFS || FS != Size) return false;
5384 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
5386 GlobalValue *GV1 = NULL;
5387 GlobalValue *GV2 = NULL;
5388 int64_t Offset1 = 0;
5389 int64_t Offset2 = 0;
5390 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
5391 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
5392 if (isGA1 && isGA2 && GV1 == GV2)
5393 return Offset1 == (Offset2 + Dist*Size);
5399 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
5400 const X86Subtarget *Subtarget) {
5403 if (isGAPlusOffset(Base, GV, Offset))
5404 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
5406 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
5407 int BFI = cast<FrameIndexSDNode>(Base)->getIndex();
5409 // Fixed objects do not specify alignment, however the offsets are known.
5410 return ((Subtarget->getStackAlignment() % 16) == 0 &&
5411 (MFI->getObjectOffset(BFI) % 16) == 0);
5413 return MFI->getObjectAlignment(BFI) >= 16;
5419 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
5420 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
5421 /// if the load addresses are consecutive, non-overlapping, and in the right
5423 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
5424 const X86Subtarget *Subtarget) {
5425 MachineFunction &MF = DAG.getMachineFunction();
5426 MachineFrameInfo *MFI = MF.getFrameInfo();
5427 MVT::ValueType VT = N->getValueType(0);
5428 MVT::ValueType EVT = MVT::getVectorElementType(VT);
5429 SDOperand PermMask = N->getOperand(2);
5430 int NumElems = (int)PermMask.getNumOperands();
5431 SDNode *Base = NULL;
5432 for (int i = 0; i < NumElems; ++i) {
5433 SDOperand Idx = PermMask.getOperand(i);
5434 if (Idx.getOpcode() == ISD::UNDEF) {
5435 if (!Base) return SDOperand();
5438 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
5439 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
5443 else if (!isConsecutiveLoad(Arg.Val, Base,
5444 i, MVT::getSizeInBits(EVT)/8,MFI))
5449 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
5450 LoadSDNode *LD = cast<LoadSDNode>(Base);
5452 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5453 LD->getSrcValueOffset(), LD->isVolatile());
5455 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5456 LD->getSrcValueOffset(), LD->isVolatile(),
5457 LD->getAlignment());
5461 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
5462 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
5463 const X86Subtarget *Subtarget) {
5464 SDOperand Cond = N->getOperand(0);
5466 // If we have SSE[12] support, try to form min/max nodes.
5467 if (Subtarget->hasSSE2() &&
5468 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
5469 if (Cond.getOpcode() == ISD::SETCC) {
5470 // Get the LHS/RHS of the select.
5471 SDOperand LHS = N->getOperand(1);
5472 SDOperand RHS = N->getOperand(2);
5473 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
5475 unsigned Opcode = 0;
5476 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
5479 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
5482 if (!UnsafeFPMath) break;
5484 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
5486 Opcode = X86ISD::FMIN;
5489 case ISD::SETOGT: // (X > Y) ? X : Y -> max
5492 if (!UnsafeFPMath) break;
5494 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
5496 Opcode = X86ISD::FMAX;
5499 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
5502 case ISD::SETOGT: // (X > Y) ? Y : X -> min
5505 if (!UnsafeFPMath) break;
5507 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
5509 Opcode = X86ISD::FMIN;
5512 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
5515 if (!UnsafeFPMath) break;
5517 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
5519 Opcode = X86ISD::FMAX;
5525 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
5534 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
5535 DAGCombinerInfo &DCI) const {
5536 SelectionDAG &DAG = DCI.DAG;
5537 switch (N->getOpcode()) {
5539 case ISD::VECTOR_SHUFFLE:
5540 return PerformShuffleCombine(N, DAG, Subtarget);
5542 return PerformSELECTCombine(N, DAG, Subtarget);
5548 //===----------------------------------------------------------------------===//
5549 // X86 Inline Assembly Support
5550 //===----------------------------------------------------------------------===//
5552 /// getConstraintType - Given a constraint letter, return the type of
5553 /// constraint it is for this target.
5554 X86TargetLowering::ConstraintType
5555 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
5556 if (Constraint.size() == 1) {
5557 switch (Constraint[0]) {
5566 return C_RegisterClass;
5571 return TargetLowering::getConstraintType(Constraint);
5574 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5575 /// vector. If it is invalid, don't add anything to Ops.
5576 void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
5578 std::vector<SDOperand>&Ops,
5579 SelectionDAG &DAG) {
5580 SDOperand Result(0, 0);
5582 switch (Constraint) {
5585 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5586 if (C->getValue() <= 31) {
5587 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5593 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5594 if (C->getValue() <= 255) {
5595 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5601 // Literal immediates are always ok.
5602 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
5603 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
5607 // If we are in non-pic codegen mode, we allow the address of a global (with
5608 // an optional displacement) to be used with 'i'.
5609 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
5612 // Match either (GA) or (GA+C)
5614 Offset = GA->getOffset();
5615 } else if (Op.getOpcode() == ISD::ADD) {
5616 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5617 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5619 Offset = GA->getOffset()+C->getValue();
5621 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5622 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5624 Offset = GA->getOffset()+C->getValue();
5631 // If addressing this global requires a load (e.g. in PIC mode), we can't
5633 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
5637 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5643 // Otherwise, not valid for this mode.
5649 Ops.push_back(Result);
5652 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5655 std::vector<unsigned> X86TargetLowering::
5656 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5657 MVT::ValueType VT) const {
5658 if (Constraint.size() == 1) {
5659 // FIXME: not handling fp-stack yet!
5660 switch (Constraint[0]) { // GCC X86 Constraint Letters
5661 default: break; // Unknown constraint letter
5662 case 'A': // EAX/EDX
5663 if (VT == MVT::i32 || VT == MVT::i64)
5664 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5666 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5669 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5670 else if (VT == MVT::i16)
5671 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5672 else if (VT == MVT::i8)
5673 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5678 return std::vector<unsigned>();
5681 std::pair<unsigned, const TargetRegisterClass*>
5682 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5683 MVT::ValueType VT) const {
5684 // First, see if this is a constraint that directly corresponds to an LLVM
5686 if (Constraint.size() == 1) {
5687 // GCC Constraint Letters
5688 switch (Constraint[0]) {
5690 case 'r': // GENERAL_REGS
5691 case 'R': // LEGACY_REGS
5692 case 'l': // INDEX_REGS
5693 if (VT == MVT::i64 && Subtarget->is64Bit())
5694 return std::make_pair(0U, X86::GR64RegisterClass);
5696 return std::make_pair(0U, X86::GR32RegisterClass);
5697 else if (VT == MVT::i16)
5698 return std::make_pair(0U, X86::GR16RegisterClass);
5699 else if (VT == MVT::i8)
5700 return std::make_pair(0U, X86::GR8RegisterClass);
5702 case 'y': // MMX_REGS if MMX allowed.
5703 if (!Subtarget->hasMMX()) break;
5704 return std::make_pair(0U, X86::VR64RegisterClass);
5706 case 'Y': // SSE_REGS if SSE2 allowed
5707 if (!Subtarget->hasSSE2()) break;
5709 case 'x': // SSE_REGS if SSE1 allowed
5710 if (!Subtarget->hasSSE1()) break;
5714 // Scalar SSE types.
5717 return std::make_pair(0U, X86::FR32RegisterClass);
5720 return std::make_pair(0U, X86::FR64RegisterClass);
5728 return std::make_pair(0U, X86::VR128RegisterClass);
5734 // Use the default implementation in TargetLowering to convert the register
5735 // constraint into a member of a register class.
5736 std::pair<unsigned, const TargetRegisterClass*> Res;
5737 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5739 // Not found as a standard register?
5740 if (Res.second == 0) {
5741 // GCC calls "st(0)" just plain "st".
5742 if (StringsEqualNoCase("{st}", Constraint)) {
5743 Res.first = X86::ST0;
5744 Res.second = X86::RFP80RegisterClass;
5750 // Otherwise, check to see if this is a register class of the wrong value
5751 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5752 // turn into {ax},{dx}.
5753 if (Res.second->hasType(VT))
5754 return Res; // Correct type already, nothing to do.
5756 // All of the single-register GCC register classes map their values onto
5757 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5758 // really want an 8-bit or 32-bit register, map to the appropriate register
5759 // class and return the appropriate register.
5760 if (Res.second != X86::GR16RegisterClass)
5763 if (VT == MVT::i8) {
5764 unsigned DestReg = 0;
5765 switch (Res.first) {
5767 case X86::AX: DestReg = X86::AL; break;
5768 case X86::DX: DestReg = X86::DL; break;
5769 case X86::CX: DestReg = X86::CL; break;
5770 case X86::BX: DestReg = X86::BL; break;
5773 Res.first = DestReg;
5774 Res.second = Res.second = X86::GR8RegisterClass;
5776 } else if (VT == MVT::i32) {
5777 unsigned DestReg = 0;
5778 switch (Res.first) {
5780 case X86::AX: DestReg = X86::EAX; break;
5781 case X86::DX: DestReg = X86::EDX; break;
5782 case X86::CX: DestReg = X86::ECX; break;
5783 case X86::BX: DestReg = X86::EBX; break;
5784 case X86::SI: DestReg = X86::ESI; break;
5785 case X86::DI: DestReg = X86::EDI; break;
5786 case X86::BP: DestReg = X86::EBP; break;
5787 case X86::SP: DestReg = X86::ESP; break;
5790 Res.first = DestReg;
5791 Res.second = Res.second = X86::GR32RegisterClass;
5793 } else if (VT == MVT::i64) {
5794 unsigned DestReg = 0;
5795 switch (Res.first) {
5797 case X86::AX: DestReg = X86::RAX; break;
5798 case X86::DX: DestReg = X86::RDX; break;
5799 case X86::CX: DestReg = X86::RCX; break;
5800 case X86::BX: DestReg = X86::RBX; break;
5801 case X86::SI: DestReg = X86::RSI; break;
5802 case X86::DI: DestReg = X86::RDI; break;
5803 case X86::BP: DestReg = X86::RBP; break;
5804 case X86::SP: DestReg = X86::RSP; break;
5807 Res.first = DestReg;
5808 Res.second = Res.second = X86::GR64RegisterClass;