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/Target/TargetOptions.h"
36 #include "llvm/ADT/StringExtras.h"
39 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
40 : TargetLowering(TM) {
41 Subtarget = &TM.getSubtarget<X86Subtarget>();
42 X86ScalarSSE = Subtarget->hasSSE2();
43 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
45 // Set up the TargetLowering object.
47 // X86 is weird, it always uses i8 for shift amounts and setcc results.
48 setShiftAmountType(MVT::i8);
49 setSetCCResultType(MVT::i8);
50 setSetCCResultContents(ZeroOrOneSetCCResult);
51 setSchedulingPreference(SchedulingForRegPressure);
52 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
53 setStackPointerRegisterToSaveRestore(X86StackPtr);
55 if (Subtarget->isTargetDarwin()) {
56 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
57 setUseUnderscoreSetJmp(false);
58 setUseUnderscoreLongJmp(false);
59 } else if (Subtarget->isTargetMingw()) {
60 // MS runtime is weird: it exports _setjmp, but longjmp!
61 setUseUnderscoreSetJmp(true);
62 setUseUnderscoreLongJmp(false);
64 setUseUnderscoreSetJmp(true);
65 setUseUnderscoreLongJmp(true);
68 // Set up the register classes.
69 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
70 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
71 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
72 if (Subtarget->is64Bit())
73 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
75 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
77 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
79 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
80 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
81 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
83 if (Subtarget->is64Bit()) {
84 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
85 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
88 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
89 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
91 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
94 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
96 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
97 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
98 // SSE has no i16 to fp conversion, only i32
100 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
102 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
103 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
106 if (!Subtarget->is64Bit()) {
107 // Custom lower SINT_TO_FP and FP_TO_SINT from/to i64 in 32-bit mode.
108 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
109 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
112 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
114 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
115 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
118 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
120 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
121 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
124 // Handle FP_TO_UINT by promoting the destination to a larger signed
126 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
127 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
128 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
130 if (Subtarget->is64Bit()) {
131 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
132 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
134 if (X86ScalarSSE && !Subtarget->hasSSE3())
135 // Expand FP_TO_UINT into a select.
136 // FIXME: We would like to use a Custom expander here eventually to do
137 // the optimal thing for SSE vs. the default expansion in the legalizer.
138 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
140 // With SSE3 we can use fisttpll to convert to a signed i64.
141 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
144 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
146 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
147 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
150 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
151 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
152 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
153 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
154 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
155 if (Subtarget->is64Bit())
156 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Expand);
157 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
158 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
159 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
160 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
161 setOperationAction(ISD::FREM , MVT::f64 , Expand);
163 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
164 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
165 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
166 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
167 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
168 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
169 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
170 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
171 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
172 if (Subtarget->is64Bit()) {
173 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
174 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
175 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
178 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
179 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
181 // These should be promoted to a larger select which is supported.
182 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
183 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
184 // X86 wants to expand cmov itself.
185 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
186 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
187 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
188 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
189 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
190 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
191 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
192 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
193 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
194 if (Subtarget->is64Bit()) {
195 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
196 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
198 // X86 ret instruction may pop stack.
199 setOperationAction(ISD::RET , MVT::Other, Custom);
201 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
202 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
203 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
204 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
205 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
206 if (Subtarget->is64Bit()) {
207 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
208 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
209 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
210 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
212 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
213 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
214 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
215 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
216 // X86 wants to expand memset / memcpy itself.
217 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
218 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
220 // We don't have line number support yet.
221 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
222 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
223 // FIXME - use subtarget debug flags
224 if (!Subtarget->isTargetDarwin() &&
225 !Subtarget->isTargetELF() &&
226 !Subtarget->isTargetCygMing())
227 setOperationAction(ISD::LABEL, MVT::Other, Expand);
229 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
230 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
231 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
232 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
233 if (Subtarget->is64Bit()) {
235 setExceptionPointerRegister(X86::RAX);
236 setExceptionSelectorRegister(X86::RDX);
238 setExceptionPointerRegister(X86::EAX);
239 setExceptionSelectorRegister(X86::EDX);
242 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
243 setOperationAction(ISD::VASTART , MVT::Other, Custom);
244 setOperationAction(ISD::VAARG , MVT::Other, Expand);
245 setOperationAction(ISD::VAEND , MVT::Other, Expand);
246 if (Subtarget->is64Bit())
247 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
249 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
251 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
252 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
253 if (Subtarget->is64Bit())
254 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
255 if (Subtarget->isTargetCygMing())
256 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
258 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
261 // Set up the FP register classes.
262 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
263 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
265 // Use ANDPD to simulate FABS.
266 setOperationAction(ISD::FABS , MVT::f64, Custom);
267 setOperationAction(ISD::FABS , MVT::f32, Custom);
269 // Use XORP to simulate FNEG.
270 setOperationAction(ISD::FNEG , MVT::f64, Custom);
271 setOperationAction(ISD::FNEG , MVT::f32, Custom);
273 // Use ANDPD and ORPD to simulate FCOPYSIGN.
274 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
275 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
277 // We don't support sin/cos/fmod
278 setOperationAction(ISD::FSIN , MVT::f64, Expand);
279 setOperationAction(ISD::FCOS , MVT::f64, Expand);
280 setOperationAction(ISD::FREM , MVT::f64, Expand);
281 setOperationAction(ISD::FSIN , MVT::f32, Expand);
282 setOperationAction(ISD::FCOS , MVT::f32, Expand);
283 setOperationAction(ISD::FREM , MVT::f32, Expand);
285 // Expand FP immediates into loads from the stack, except for the special
287 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
288 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
289 addLegalFPImmediate(+0.0); // xorps / xorpd
291 // Set up the FP register classes.
292 addRegisterClass(MVT::f64, X86::RFPRegisterClass);
294 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
295 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
296 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
299 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
300 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
303 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
304 addLegalFPImmediate(+0.0); // FLD0
305 addLegalFPImmediate(+1.0); // FLD1
306 addLegalFPImmediate(-0.0); // FLD0/FCHS
307 addLegalFPImmediate(-1.0); // FLD1/FCHS
310 // First set operation action for all vector types to expand. Then we
311 // will selectively turn on ones that can be effectively codegen'd.
312 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
313 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
314 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
315 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
316 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
317 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
318 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
319 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
320 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
321 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
322 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
323 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
324 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
325 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
326 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
327 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
328 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
331 if (Subtarget->hasMMX()) {
332 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
333 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
334 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
335 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
337 // FIXME: add MMX packed arithmetics
339 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
340 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
341 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
342 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
344 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
345 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
346 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
348 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
349 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
351 setOperationAction(ISD::AND, MVT::v8i8, Promote);
352 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
353 setOperationAction(ISD::AND, MVT::v4i16, Promote);
354 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
355 setOperationAction(ISD::AND, MVT::v2i32, Promote);
356 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
357 setOperationAction(ISD::AND, MVT::v1i64, Legal);
359 setOperationAction(ISD::OR, MVT::v8i8, Promote);
360 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
361 setOperationAction(ISD::OR, MVT::v4i16, Promote);
362 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
363 setOperationAction(ISD::OR, MVT::v2i32, Promote);
364 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
365 setOperationAction(ISD::OR, MVT::v1i64, Legal);
367 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
368 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
369 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
370 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
371 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
372 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
373 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
375 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
376 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
377 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
378 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
379 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
380 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
381 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
383 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
384 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
385 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
386 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
388 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
389 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
390 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
391 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
393 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
394 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
395 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Custom);
396 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
399 if (Subtarget->hasSSE1()) {
400 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
402 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
403 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
404 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
405 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
406 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
407 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
408 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
409 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
410 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
413 if (Subtarget->hasSSE2()) {
414 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
415 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
416 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
417 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
418 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
420 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
421 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
422 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
423 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
424 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
425 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
426 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
427 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
428 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
429 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
430 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
431 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
432 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
434 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
435 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
436 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
437 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
438 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
439 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
441 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
442 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
443 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
444 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
445 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
447 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
448 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
449 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
450 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
451 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
452 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
454 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
455 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
456 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
457 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
458 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
459 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
460 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
461 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
462 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
463 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
464 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
465 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
468 // Custom lower v2i64 and v2f64 selects.
469 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
470 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
471 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
472 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
475 // We want to custom lower some of our intrinsics.
476 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
478 // We have target-specific dag combine patterns for the following nodes:
479 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
480 setTargetDAGCombine(ISD::SELECT);
482 computeRegisterProperties();
484 // FIXME: These should be based on subtarget info. Plus, the values should
485 // be smaller when we are in optimizing for size mode.
486 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
487 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
488 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
489 allowUnalignedMemoryAccesses = true; // x86 supports it!
493 //===----------------------------------------------------------------------===//
494 // Return Value Calling Convention Implementation
495 //===----------------------------------------------------------------------===//
497 #include "X86GenCallingConv.inc"
499 /// LowerRET - Lower an ISD::RET node.
500 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
501 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
503 SmallVector<CCValAssign, 16> RVLocs;
504 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
505 CCState CCInfo(CC, getTargetMachine(), RVLocs);
506 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
509 // If this is the first return lowered for this function, add the regs to the
510 // liveout set for the function.
511 if (DAG.getMachineFunction().liveout_empty()) {
512 for (unsigned i = 0; i != RVLocs.size(); ++i)
513 if (RVLocs[i].isRegLoc())
514 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
517 SDOperand Chain = Op.getOperand(0);
520 // Copy the result values into the output registers.
521 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
522 RVLocs[0].getLocReg() != X86::ST0) {
523 for (unsigned i = 0; i != RVLocs.size(); ++i) {
524 CCValAssign &VA = RVLocs[i];
525 assert(VA.isRegLoc() && "Can only return in registers!");
526 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
528 Flag = Chain.getValue(1);
531 // We need to handle a destination of ST0 specially, because it isn't really
533 SDOperand Value = Op.getOperand(1);
535 // If this is an FP return with ScalarSSE, we need to move the value from
536 // an XMM register onto the fp-stack.
540 // If this is a load into a scalarsse value, don't store the loaded value
541 // back to the stack, only to reload it: just replace the scalar-sse load.
542 if (ISD::isNON_EXTLoad(Value.Val) &&
543 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
544 Chain = Value.getOperand(0);
545 MemLoc = Value.getOperand(1);
547 // Spill the value to memory and reload it into top of stack.
548 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
549 MachineFunction &MF = DAG.getMachineFunction();
550 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
551 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
552 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
554 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
555 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
556 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
557 Chain = Value.getValue(1);
560 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
561 SDOperand Ops[] = { Chain, Value };
562 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
563 Flag = Chain.getValue(1);
566 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
568 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
570 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
574 /// LowerCallResult - Lower the result values of an ISD::CALL into the
575 /// appropriate copies out of appropriate physical registers. This assumes that
576 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
577 /// being lowered. The returns a SDNode with the same number of values as the
579 SDNode *X86TargetLowering::
580 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
581 unsigned CallingConv, SelectionDAG &DAG) {
583 // Assign locations to each value returned by this call.
584 SmallVector<CCValAssign, 16> RVLocs;
585 CCState CCInfo(CallingConv, getTargetMachine(), RVLocs);
586 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
589 SmallVector<SDOperand, 8> ResultVals;
591 // Copy all of the result registers out of their specified physreg.
592 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
593 for (unsigned i = 0; i != RVLocs.size(); ++i) {
594 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
595 RVLocs[i].getValVT(), InFlag).getValue(1);
596 InFlag = Chain.getValue(2);
597 ResultVals.push_back(Chain.getValue(0));
600 // Copies from the FP stack are special, as ST0 isn't a valid register
601 // before the fp stackifier runs.
603 // Copy ST0 into an RFP register with FP_GET_RESULT.
604 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
605 SDOperand GROps[] = { Chain, InFlag };
606 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
607 Chain = RetVal.getValue(1);
608 InFlag = RetVal.getValue(2);
610 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
613 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
614 // shouldn't be necessary except that RFP cannot be live across
615 // multiple blocks. When stackifier is fixed, they can be uncoupled.
616 MachineFunction &MF = DAG.getMachineFunction();
617 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
618 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
620 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag
622 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
623 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
624 Chain = RetVal.getValue(1);
627 if (RVLocs[0].getValVT() == MVT::f32 && !X86ScalarSSE)
628 // FIXME: we would really like to remember that this FP_ROUND
629 // operation is okay to eliminate if we allow excess FP precision.
630 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
631 ResultVals.push_back(RetVal);
634 // Merge everything together with a MERGE_VALUES node.
635 ResultVals.push_back(Chain);
636 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
637 &ResultVals[0], ResultVals.size()).Val;
641 //===----------------------------------------------------------------------===//
642 // C & StdCall Calling Convention implementation
643 //===----------------------------------------------------------------------===//
644 // StdCall calling convention seems to be standard for many Windows' API
645 // routines and around. It differs from C calling convention just a little:
646 // callee should clean up the stack, not caller. Symbols should be also
647 // decorated in some fancy way :) It doesn't support any vector arguments.
649 /// AddLiveIn - This helper function adds the specified physical register to the
650 /// MachineFunction as a live in value. It also creates a corresponding virtual
652 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
653 const TargetRegisterClass *RC) {
654 assert(RC->contains(PReg) && "Not the correct regclass!");
655 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
656 MF.addLiveIn(PReg, VReg);
660 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
662 unsigned NumArgs = Op.Val->getNumValues() - 1;
663 MachineFunction &MF = DAG.getMachineFunction();
664 MachineFrameInfo *MFI = MF.getFrameInfo();
665 SDOperand Root = Op.getOperand(0);
666 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
668 // Assign locations to all of the incoming arguments.
669 SmallVector<CCValAssign, 16> ArgLocs;
670 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
672 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
674 SmallVector<SDOperand, 8> ArgValues;
675 unsigned LastVal = ~0U;
676 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
677 CCValAssign &VA = ArgLocs[i];
678 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
680 assert(VA.getValNo() != LastVal &&
681 "Don't support value assigned to multiple locs yet");
682 LastVal = VA.getValNo();
685 MVT::ValueType RegVT = VA.getLocVT();
686 TargetRegisterClass *RC;
687 if (RegVT == MVT::i32)
688 RC = X86::GR32RegisterClass;
690 assert(MVT::isVector(RegVT));
691 RC = X86::VR128RegisterClass;
694 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
695 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
697 // If this is an 8 or 16-bit value, it is really passed promoted to 32
698 // bits. Insert an assert[sz]ext to capture this, then truncate to the
700 if (VA.getLocInfo() == CCValAssign::SExt)
701 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
702 DAG.getValueType(VA.getValVT()));
703 else if (VA.getLocInfo() == CCValAssign::ZExt)
704 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
705 DAG.getValueType(VA.getValVT()));
707 if (VA.getLocInfo() != CCValAssign::Full)
708 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
710 ArgValues.push_back(ArgValue);
712 assert(VA.isMemLoc());
714 // Create the nodes corresponding to a load from this parameter slot.
715 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
716 VA.getLocMemOffset());
717 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
718 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
722 unsigned StackSize = CCInfo.getNextStackOffset();
724 ArgValues.push_back(Root);
726 // If the function takes variable number of arguments, make a frame index for
727 // the start of the first vararg value... for expansion of llvm.va_start.
729 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
731 if (isStdCall && !isVarArg) {
732 BytesToPopOnReturn = StackSize; // Callee pops everything..
733 BytesCallerReserves = 0;
735 BytesToPopOnReturn = 0; // Callee pops nothing.
737 // If this is an sret function, the return should pop the hidden pointer.
739 (cast<ConstantSDNode>(Op.getOperand(3))->getValue() &
740 ISD::ParamFlags::StructReturn))
741 BytesToPopOnReturn = 4;
743 BytesCallerReserves = StackSize;
746 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
747 ReturnAddrIndex = 0; // No return address slot generated yet.
749 MF.getInfo<X86MachineFunctionInfo>()
750 ->setBytesToPopOnReturn(BytesToPopOnReturn);
752 // Return the new list of results.
753 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
754 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
757 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
759 SDOperand Chain = Op.getOperand(0);
760 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
761 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
762 SDOperand Callee = Op.getOperand(4);
763 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
765 // Analyze operands of the call, assigning locations to each operand.
766 SmallVector<CCValAssign, 16> ArgLocs;
767 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
768 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
770 // Get a count of how many bytes are to be pushed on the stack.
771 unsigned NumBytes = CCInfo.getNextStackOffset();
773 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
775 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
776 SmallVector<SDOperand, 8> MemOpChains;
780 // Walk the register/memloc assignments, inserting copies/loads.
781 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
782 CCValAssign &VA = ArgLocs[i];
783 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
785 // Promote the value if needed.
786 switch (VA.getLocInfo()) {
787 default: assert(0 && "Unknown loc info!");
788 case CCValAssign::Full: break;
789 case CCValAssign::SExt:
790 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
792 case CCValAssign::ZExt:
793 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
795 case CCValAssign::AExt:
796 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
801 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
803 assert(VA.isMemLoc());
804 if (StackPtr.Val == 0)
805 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
806 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
807 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
808 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
812 // If the first argument is an sret pointer, remember it.
813 bool isSRet = NumOps &&
814 (cast<ConstantSDNode>(Op.getOperand(6))->getValue() &
815 ISD::ParamFlags::StructReturn);
817 if (!MemOpChains.empty())
818 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
819 &MemOpChains[0], MemOpChains.size());
821 // Build a sequence of copy-to-reg nodes chained together with token chain
822 // and flag operands which copy the outgoing args into registers.
824 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
825 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
827 InFlag = Chain.getValue(1);
830 // ELF / PIC requires GOT in the EBX register before function calls via PLT
832 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
833 Subtarget->isPICStyleGOT()) {
834 Chain = DAG.getCopyToReg(Chain, X86::EBX,
835 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
837 InFlag = Chain.getValue(1);
840 // If the callee is a GlobalAddress node (quite common, every direct call is)
841 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
842 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
843 // We should use extra load for direct calls to dllimported functions in
845 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
846 getTargetMachine(), true))
847 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
848 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
849 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
851 // Returns a chain & a flag for retval copy to use.
852 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
853 SmallVector<SDOperand, 8> Ops;
854 Ops.push_back(Chain);
855 Ops.push_back(Callee);
857 // Add argument registers to the end of the list so that they are known live
859 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
860 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
861 RegsToPass[i].second.getValueType()));
863 // Add an implicit use GOT pointer in EBX.
864 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
865 Subtarget->isPICStyleGOT())
866 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
869 Ops.push_back(InFlag);
871 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
872 NodeTys, &Ops[0], Ops.size());
873 InFlag = Chain.getValue(1);
875 // Create the CALLSEQ_END node.
876 unsigned NumBytesForCalleeToPush = 0;
878 if (CC == CallingConv::X86_StdCall) {
880 NumBytesForCalleeToPush = isSRet ? 4 : 0;
882 NumBytesForCalleeToPush = NumBytes;
884 // If this is is a call to a struct-return function, the callee
885 // pops the hidden struct pointer, so we have to push it back.
886 // This is common for Darwin/X86, Linux & Mingw32 targets.
887 NumBytesForCalleeToPush = isSRet ? 4 : 0;
890 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
892 Ops.push_back(Chain);
893 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
894 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
895 Ops.push_back(InFlag);
896 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
897 InFlag = Chain.getValue(1);
899 // Handle result values, copying them out of physregs into vregs that we
901 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
905 //===----------------------------------------------------------------------===//
906 // FastCall Calling Convention implementation
907 //===----------------------------------------------------------------------===//
909 // The X86 'fastcall' calling convention passes up to two integer arguments in
910 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
911 // and requires that the callee pop its arguments off the stack (allowing proper
912 // tail calls), and has the same return value conventions as C calling convs.
914 // This calling convention always arranges for the callee pop value to be 8n+4
915 // bytes, which is needed for tail recursion elimination and stack alignment
918 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
919 MachineFunction &MF = DAG.getMachineFunction();
920 MachineFrameInfo *MFI = MF.getFrameInfo();
921 SDOperand Root = Op.getOperand(0);
923 // Assign locations to all of the incoming arguments.
924 SmallVector<CCValAssign, 16> ArgLocs;
925 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
927 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
929 SmallVector<SDOperand, 8> ArgValues;
930 unsigned LastVal = ~0U;
931 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
932 CCValAssign &VA = ArgLocs[i];
933 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
935 assert(VA.getValNo() != LastVal &&
936 "Don't support value assigned to multiple locs yet");
937 LastVal = VA.getValNo();
940 MVT::ValueType RegVT = VA.getLocVT();
941 TargetRegisterClass *RC;
942 if (RegVT == MVT::i32)
943 RC = X86::GR32RegisterClass;
945 assert(MVT::isVector(RegVT));
946 RC = X86::VR128RegisterClass;
949 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
950 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
952 // If this is an 8 or 16-bit value, it is really passed promoted to 32
953 // bits. Insert an assert[sz]ext to capture this, then truncate to the
955 if (VA.getLocInfo() == CCValAssign::SExt)
956 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
957 DAG.getValueType(VA.getValVT()));
958 else if (VA.getLocInfo() == CCValAssign::ZExt)
959 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
960 DAG.getValueType(VA.getValVT()));
962 if (VA.getLocInfo() != CCValAssign::Full)
963 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
965 ArgValues.push_back(ArgValue);
967 assert(VA.isMemLoc());
969 // Create the nodes corresponding to a load from this parameter slot.
970 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
971 VA.getLocMemOffset());
972 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
973 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
977 ArgValues.push_back(Root);
979 unsigned StackSize = CCInfo.getNextStackOffset();
981 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
982 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
983 // arguments and the arguments after the retaddr has been pushed are aligned.
984 if ((StackSize & 7) == 0)
988 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
989 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
990 ReturnAddrIndex = 0; // No return address slot generated yet.
991 BytesToPopOnReturn = StackSize; // Callee pops all stack arguments.
992 BytesCallerReserves = 0;
994 MF.getInfo<X86MachineFunctionInfo>()
995 ->setBytesToPopOnReturn(BytesToPopOnReturn);
997 // Return the new list of results.
998 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
999 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1002 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1004 SDOperand Chain = Op.getOperand(0);
1005 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1006 SDOperand Callee = Op.getOperand(4);
1008 // Analyze operands of the call, assigning locations to each operand.
1009 SmallVector<CCValAssign, 16> ArgLocs;
1010 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
1011 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall);
1013 // Get a count of how many bytes are to be pushed on the stack.
1014 unsigned NumBytes = CCInfo.getNextStackOffset();
1016 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1017 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1018 // arguments and the arguments after the retaddr has been pushed are aligned.
1019 if ((NumBytes & 7) == 0)
1023 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1025 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1026 SmallVector<SDOperand, 8> MemOpChains;
1030 // Walk the register/memloc assignments, inserting copies/loads.
1031 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1032 CCValAssign &VA = ArgLocs[i];
1033 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1035 // Promote the value if needed.
1036 switch (VA.getLocInfo()) {
1037 default: assert(0 && "Unknown loc info!");
1038 case CCValAssign::Full: break;
1039 case CCValAssign::SExt:
1040 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1042 case CCValAssign::ZExt:
1043 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1045 case CCValAssign::AExt:
1046 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1050 if (VA.isRegLoc()) {
1051 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1053 assert(VA.isMemLoc());
1054 if (StackPtr.Val == 0)
1055 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1056 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1057 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1058 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1062 if (!MemOpChains.empty())
1063 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1064 &MemOpChains[0], MemOpChains.size());
1066 // Build a sequence of copy-to-reg nodes chained together with token chain
1067 // and flag operands which copy the outgoing args into registers.
1069 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1070 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1072 InFlag = Chain.getValue(1);
1075 // If the callee is a GlobalAddress node (quite common, every direct call is)
1076 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1077 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1078 // We should use extra load for direct calls to dllimported functions in
1080 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1081 getTargetMachine(), true))
1082 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1083 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1084 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1086 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1088 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1089 Subtarget->isPICStyleGOT()) {
1090 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1091 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1093 InFlag = Chain.getValue(1);
1096 // Returns a chain & a flag for retval copy to use.
1097 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1098 SmallVector<SDOperand, 8> Ops;
1099 Ops.push_back(Chain);
1100 Ops.push_back(Callee);
1102 // Add argument registers to the end of the list so that they are known live
1104 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1105 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1106 RegsToPass[i].second.getValueType()));
1108 // Add an implicit use GOT pointer in EBX.
1109 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1110 Subtarget->isPICStyleGOT())
1111 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1114 Ops.push_back(InFlag);
1116 // FIXME: Do not generate X86ISD::TAILCALL for now.
1117 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1118 NodeTys, &Ops[0], Ops.size());
1119 InFlag = Chain.getValue(1);
1121 // Returns a flag for retval copy to use.
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(NumBytes, 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 // X86-64 C Calling Convention implementation
1139 //===----------------------------------------------------------------------===//
1142 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1143 MachineFunction &MF = DAG.getMachineFunction();
1144 MachineFrameInfo *MFI = MF.getFrameInfo();
1145 SDOperand Root = Op.getOperand(0);
1146 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1148 static const unsigned GPR64ArgRegs[] = {
1149 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1151 static const unsigned XMMArgRegs[] = {
1152 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1153 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1157 // Assign locations to all of the incoming arguments.
1158 SmallVector<CCValAssign, 16> ArgLocs;
1159 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
1161 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1163 SmallVector<SDOperand, 8> ArgValues;
1164 unsigned LastVal = ~0U;
1165 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1166 CCValAssign &VA = ArgLocs[i];
1167 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1169 assert(VA.getValNo() != LastVal &&
1170 "Don't support value assigned to multiple locs yet");
1171 LastVal = VA.getValNo();
1173 if (VA.isRegLoc()) {
1174 MVT::ValueType RegVT = VA.getLocVT();
1175 TargetRegisterClass *RC;
1176 if (RegVT == MVT::i32)
1177 RC = X86::GR32RegisterClass;
1178 else if (RegVT == MVT::i64)
1179 RC = X86::GR64RegisterClass;
1180 else if (RegVT == MVT::f32)
1181 RC = X86::FR32RegisterClass;
1182 else if (RegVT == MVT::f64)
1183 RC = X86::FR64RegisterClass;
1185 assert(MVT::isVector(RegVT));
1186 if (MVT::getSizeInBits(RegVT) == 64)
1187 RC = X86::VR64RegisterClass;
1189 RC = X86::VR128RegisterClass;
1192 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1193 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1195 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1196 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1198 if (VA.getLocInfo() == CCValAssign::SExt)
1199 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1200 DAG.getValueType(VA.getValVT()));
1201 else if (VA.getLocInfo() == CCValAssign::ZExt)
1202 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1203 DAG.getValueType(VA.getValVT()));
1205 if (VA.getLocInfo() != CCValAssign::Full)
1206 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1208 ArgValues.push_back(ArgValue);
1210 assert(VA.isMemLoc());
1212 // Create the nodes corresponding to a load from this parameter slot.
1213 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1214 VA.getLocMemOffset());
1215 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1216 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
1220 unsigned StackSize = CCInfo.getNextStackOffset();
1222 // If the function takes variable number of arguments, make a frame index for
1223 // the start of the first vararg value... for expansion of llvm.va_start.
1225 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1226 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1228 // For X86-64, if there are vararg parameters that are passed via
1229 // registers, then we must store them to their spots on the stack so they
1230 // may be loaded by deferencing the result of va_next.
1231 VarArgsGPOffset = NumIntRegs * 8;
1232 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1233 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1234 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1236 // Store the integer parameter registers.
1237 SmallVector<SDOperand, 8> MemOps;
1238 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1239 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1240 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1241 for (; NumIntRegs != 6; ++NumIntRegs) {
1242 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1243 X86::GR64RegisterClass);
1244 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1245 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1246 MemOps.push_back(Store);
1247 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1248 DAG.getConstant(8, getPointerTy()));
1251 // Now store the XMM (fp + vector) parameter registers.
1252 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1253 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1254 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1255 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1256 X86::VR128RegisterClass);
1257 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1258 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1259 MemOps.push_back(Store);
1260 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1261 DAG.getConstant(16, getPointerTy()));
1263 if (!MemOps.empty())
1264 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1265 &MemOps[0], MemOps.size());
1268 ArgValues.push_back(Root);
1270 ReturnAddrIndex = 0; // No return address slot generated yet.
1271 BytesToPopOnReturn = 0; // Callee pops nothing.
1272 BytesCallerReserves = StackSize;
1274 // Return the new list of results.
1275 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1276 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1280 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1282 SDOperand Chain = Op.getOperand(0);
1283 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1284 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1285 SDOperand Callee = Op.getOperand(4);
1287 // Analyze operands of the call, assigning locations to each operand.
1288 SmallVector<CCValAssign, 16> ArgLocs;
1289 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
1290 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1292 // Get a count of how many bytes are to be pushed on the stack.
1293 unsigned NumBytes = CCInfo.getNextStackOffset();
1294 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1296 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1297 SmallVector<SDOperand, 8> MemOpChains;
1301 // Walk the register/memloc assignments, inserting copies/loads.
1302 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1303 CCValAssign &VA = ArgLocs[i];
1304 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1306 // Promote the value if needed.
1307 switch (VA.getLocInfo()) {
1308 default: assert(0 && "Unknown loc info!");
1309 case CCValAssign::Full: break;
1310 case CCValAssign::SExt:
1311 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1313 case CCValAssign::ZExt:
1314 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1316 case CCValAssign::AExt:
1317 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1321 if (VA.isRegLoc()) {
1322 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1324 assert(VA.isMemLoc());
1325 if (StackPtr.Val == 0)
1326 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1327 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1328 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1329 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1333 if (!MemOpChains.empty())
1334 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1335 &MemOpChains[0], MemOpChains.size());
1337 // Build a sequence of copy-to-reg nodes chained together with token chain
1338 // and flag operands which copy the outgoing args into registers.
1340 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1341 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1343 InFlag = Chain.getValue(1);
1347 // From AMD64 ABI document:
1348 // For calls that may call functions that use varargs or stdargs
1349 // (prototype-less calls or calls to functions containing ellipsis (...) in
1350 // the declaration) %al is used as hidden argument to specify the number
1351 // of SSE registers used. The contents of %al do not need to match exactly
1352 // the number of registers, but must be an ubound on the number of SSE
1353 // registers used and is in the range 0 - 8 inclusive.
1355 // Count the number of XMM registers allocated.
1356 static const unsigned XMMArgRegs[] = {
1357 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1358 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1360 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1362 Chain = DAG.getCopyToReg(Chain, X86::AL,
1363 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1364 InFlag = Chain.getValue(1);
1367 // If the callee is a GlobalAddress node (quite common, every direct call is)
1368 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1369 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1370 // We should use extra load for direct calls to dllimported functions in
1372 if (getTargetMachine().getCodeModel() != CodeModel::Large
1373 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1374 getTargetMachine(), true))
1375 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1376 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1377 if (getTargetMachine().getCodeModel() != CodeModel::Large)
1378 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1380 // Returns a chain & a flag for retval copy to use.
1381 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1382 SmallVector<SDOperand, 8> Ops;
1383 Ops.push_back(Chain);
1384 Ops.push_back(Callee);
1386 // Add argument registers to the end of the list so that they are known live
1388 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1389 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1390 RegsToPass[i].second.getValueType()));
1393 Ops.push_back(InFlag);
1395 // FIXME: Do not generate X86ISD::TAILCALL for now.
1396 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1397 NodeTys, &Ops[0], Ops.size());
1398 InFlag = Chain.getValue(1);
1400 // Returns a flag for retval copy to use.
1401 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1403 Ops.push_back(Chain);
1404 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1405 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1406 Ops.push_back(InFlag);
1407 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1408 InFlag = Chain.getValue(1);
1410 // Handle result values, copying them out of physregs into vregs that we
1412 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1416 //===----------------------------------------------------------------------===//
1417 // Other Lowering Hooks
1418 //===----------------------------------------------------------------------===//
1421 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1422 if (ReturnAddrIndex == 0) {
1423 // Set up a frame object for the return address.
1424 MachineFunction &MF = DAG.getMachineFunction();
1425 if (Subtarget->is64Bit())
1426 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1428 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1431 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1436 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1437 /// specific condition code. It returns a false if it cannot do a direct
1438 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1440 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1441 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1442 SelectionDAG &DAG) {
1443 X86CC = X86::COND_INVALID;
1445 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1446 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1447 // X > -1 -> X == 0, jump !sign.
1448 RHS = DAG.getConstant(0, RHS.getValueType());
1449 X86CC = X86::COND_NS;
1451 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1452 // X < 0 -> X == 0, jump on sign.
1453 X86CC = X86::COND_S;
1458 switch (SetCCOpcode) {
1460 case ISD::SETEQ: X86CC = X86::COND_E; break;
1461 case ISD::SETGT: X86CC = X86::COND_G; break;
1462 case ISD::SETGE: X86CC = X86::COND_GE; break;
1463 case ISD::SETLT: X86CC = X86::COND_L; break;
1464 case ISD::SETLE: X86CC = X86::COND_LE; break;
1465 case ISD::SETNE: X86CC = X86::COND_NE; break;
1466 case ISD::SETULT: X86CC = X86::COND_B; break;
1467 case ISD::SETUGT: X86CC = X86::COND_A; break;
1468 case ISD::SETULE: X86CC = X86::COND_BE; break;
1469 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1472 // On a floating point condition, the flags are set as follows:
1474 // 0 | 0 | 0 | X > Y
1475 // 0 | 0 | 1 | X < Y
1476 // 1 | 0 | 0 | X == Y
1477 // 1 | 1 | 1 | unordered
1479 switch (SetCCOpcode) {
1482 case ISD::SETEQ: X86CC = X86::COND_E; break;
1483 case ISD::SETOLT: Flip = true; // Fallthrough
1485 case ISD::SETGT: X86CC = X86::COND_A; break;
1486 case ISD::SETOLE: Flip = true; // Fallthrough
1488 case ISD::SETGE: X86CC = X86::COND_AE; break;
1489 case ISD::SETUGT: Flip = true; // Fallthrough
1491 case ISD::SETLT: X86CC = X86::COND_B; break;
1492 case ISD::SETUGE: Flip = true; // Fallthrough
1494 case ISD::SETLE: X86CC = X86::COND_BE; break;
1496 case ISD::SETNE: X86CC = X86::COND_NE; break;
1497 case ISD::SETUO: X86CC = X86::COND_P; break;
1498 case ISD::SETO: X86CC = X86::COND_NP; break;
1501 std::swap(LHS, RHS);
1504 return X86CC != X86::COND_INVALID;
1507 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1508 /// code. Current x86 isa includes the following FP cmov instructions:
1509 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1510 static bool hasFPCMov(unsigned X86CC) {
1526 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1527 /// true if Op is undef or if its value falls within the specified range (L, H].
1528 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1529 if (Op.getOpcode() == ISD::UNDEF)
1532 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1533 return (Val >= Low && Val < Hi);
1536 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1537 /// true if Op is undef or if its value equal to the specified value.
1538 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1539 if (Op.getOpcode() == ISD::UNDEF)
1541 return cast<ConstantSDNode>(Op)->getValue() == Val;
1544 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1545 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1546 bool X86::isPSHUFDMask(SDNode *N) {
1547 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1549 if (N->getNumOperands() != 4)
1552 // Check if the value doesn't reference the second vector.
1553 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1554 SDOperand Arg = N->getOperand(i);
1555 if (Arg.getOpcode() == ISD::UNDEF) continue;
1556 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1557 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
1564 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1565 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1566 bool X86::isPSHUFHWMask(SDNode *N) {
1567 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1569 if (N->getNumOperands() != 8)
1572 // Lower quadword copied in order.
1573 for (unsigned i = 0; i != 4; ++i) {
1574 SDOperand Arg = N->getOperand(i);
1575 if (Arg.getOpcode() == ISD::UNDEF) continue;
1576 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1577 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1581 // Upper quadword shuffled.
1582 for (unsigned i = 4; i != 8; ++i) {
1583 SDOperand Arg = N->getOperand(i);
1584 if (Arg.getOpcode() == ISD::UNDEF) continue;
1585 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1586 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1587 if (Val < 4 || Val > 7)
1594 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1595 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1596 bool X86::isPSHUFLWMask(SDNode *N) {
1597 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1599 if (N->getNumOperands() != 8)
1602 // Upper quadword copied in order.
1603 for (unsigned i = 4; i != 8; ++i)
1604 if (!isUndefOrEqual(N->getOperand(i), i))
1607 // Lower quadword shuffled.
1608 for (unsigned i = 0; i != 4; ++i)
1609 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1615 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1616 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1617 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
1618 if (NumElems != 2 && NumElems != 4) return false;
1620 unsigned Half = NumElems / 2;
1621 for (unsigned i = 0; i < Half; ++i)
1622 if (!isUndefOrInRange(Elems[i], 0, NumElems))
1624 for (unsigned i = Half; i < NumElems; ++i)
1625 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
1631 bool X86::isSHUFPMask(SDNode *N) {
1632 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1633 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
1636 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
1637 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
1638 /// half elements to come from vector 1 (which would equal the dest.) and
1639 /// the upper half to come from vector 2.
1640 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
1641 if (NumOps != 2 && NumOps != 4) return false;
1643 unsigned Half = NumOps / 2;
1644 for (unsigned i = 0; i < Half; ++i)
1645 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
1647 for (unsigned i = Half; i < NumOps; ++i)
1648 if (!isUndefOrInRange(Ops[i], 0, NumOps))
1653 static bool isCommutedSHUFP(SDNode *N) {
1654 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1655 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
1658 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
1659 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
1660 bool X86::isMOVHLPSMask(SDNode *N) {
1661 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1663 if (N->getNumOperands() != 4)
1666 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
1667 return isUndefOrEqual(N->getOperand(0), 6) &&
1668 isUndefOrEqual(N->getOperand(1), 7) &&
1669 isUndefOrEqual(N->getOperand(2), 2) &&
1670 isUndefOrEqual(N->getOperand(3), 3);
1673 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
1674 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
1676 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
1677 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1679 if (N->getNumOperands() != 4)
1682 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
1683 return isUndefOrEqual(N->getOperand(0), 2) &&
1684 isUndefOrEqual(N->getOperand(1), 3) &&
1685 isUndefOrEqual(N->getOperand(2), 2) &&
1686 isUndefOrEqual(N->getOperand(3), 3);
1689 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
1690 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
1691 bool X86::isMOVLPMask(SDNode *N) {
1692 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1694 unsigned NumElems = N->getNumOperands();
1695 if (NumElems != 2 && NumElems != 4)
1698 for (unsigned i = 0; i < NumElems/2; ++i)
1699 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
1702 for (unsigned i = NumElems/2; i < NumElems; ++i)
1703 if (!isUndefOrEqual(N->getOperand(i), i))
1709 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
1710 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
1712 bool X86::isMOVHPMask(SDNode *N) {
1713 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1715 unsigned NumElems = N->getNumOperands();
1716 if (NumElems != 2 && NumElems != 4)
1719 for (unsigned i = 0; i < NumElems/2; ++i)
1720 if (!isUndefOrEqual(N->getOperand(i), i))
1723 for (unsigned i = 0; i < NumElems/2; ++i) {
1724 SDOperand Arg = N->getOperand(i + NumElems/2);
1725 if (!isUndefOrEqual(Arg, i + NumElems))
1732 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
1733 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
1734 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
1735 bool V2IsSplat = false) {
1736 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1739 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1740 SDOperand BitI = Elts[i];
1741 SDOperand BitI1 = Elts[i+1];
1742 if (!isUndefOrEqual(BitI, j))
1745 if (isUndefOrEqual(BitI1, NumElts))
1748 if (!isUndefOrEqual(BitI1, j + NumElts))
1756 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
1757 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1758 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1761 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
1762 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
1763 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
1764 bool V2IsSplat = false) {
1765 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1768 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1769 SDOperand BitI = Elts[i];
1770 SDOperand BitI1 = Elts[i+1];
1771 if (!isUndefOrEqual(BitI, j + NumElts/2))
1774 if (isUndefOrEqual(BitI1, NumElts))
1777 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
1785 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
1786 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1787 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1790 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
1791 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
1793 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
1794 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1796 unsigned NumElems = N->getNumOperands();
1797 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1800 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1801 SDOperand BitI = N->getOperand(i);
1802 SDOperand BitI1 = N->getOperand(i+1);
1804 if (!isUndefOrEqual(BitI, j))
1806 if (!isUndefOrEqual(BitI1, j))
1813 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
1814 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
1816 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
1817 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1819 unsigned NumElems = N->getNumOperands();
1820 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1823 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
1824 SDOperand BitI = N->getOperand(i);
1825 SDOperand BitI1 = N->getOperand(i + 1);
1827 if (!isUndefOrEqual(BitI, j))
1829 if (!isUndefOrEqual(BitI1, j))
1836 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
1837 /// specifies a shuffle of elements that is suitable for input to MOVSS,
1838 /// MOVSD, and MOVD, i.e. setting the lowest element.
1839 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
1840 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1843 if (!isUndefOrEqual(Elts[0], NumElts))
1846 for (unsigned i = 1; i < NumElts; ++i) {
1847 if (!isUndefOrEqual(Elts[i], i))
1854 bool X86::isMOVLMask(SDNode *N) {
1855 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1856 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
1859 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
1860 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
1861 /// element of vector 2 and the other elements to come from vector 1 in order.
1862 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
1863 bool V2IsSplat = false,
1864 bool V2IsUndef = false) {
1865 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
1868 if (!isUndefOrEqual(Ops[0], 0))
1871 for (unsigned i = 1; i < NumOps; ++i) {
1872 SDOperand Arg = Ops[i];
1873 if (!(isUndefOrEqual(Arg, i+NumOps) ||
1874 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
1875 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
1882 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
1883 bool V2IsUndef = false) {
1884 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1885 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
1886 V2IsSplat, V2IsUndef);
1889 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1890 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
1891 bool X86::isMOVSHDUPMask(SDNode *N) {
1892 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1894 if (N->getNumOperands() != 4)
1897 // Expect 1, 1, 3, 3
1898 for (unsigned i = 0; i < 2; ++i) {
1899 SDOperand Arg = N->getOperand(i);
1900 if (Arg.getOpcode() == ISD::UNDEF) continue;
1901 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1902 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1903 if (Val != 1) return false;
1907 for (unsigned i = 2; i < 4; ++i) {
1908 SDOperand Arg = N->getOperand(i);
1909 if (Arg.getOpcode() == ISD::UNDEF) continue;
1910 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1911 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1912 if (Val != 3) return false;
1916 // Don't use movshdup if it can be done with a shufps.
1920 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1921 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
1922 bool X86::isMOVSLDUPMask(SDNode *N) {
1923 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1925 if (N->getNumOperands() != 4)
1928 // Expect 0, 0, 2, 2
1929 for (unsigned i = 0; i < 2; ++i) {
1930 SDOperand Arg = N->getOperand(i);
1931 if (Arg.getOpcode() == ISD::UNDEF) continue;
1932 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1933 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1934 if (Val != 0) return false;
1938 for (unsigned i = 2; i < 4; ++i) {
1939 SDOperand Arg = N->getOperand(i);
1940 if (Arg.getOpcode() == ISD::UNDEF) continue;
1941 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1942 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1943 if (Val != 2) return false;
1947 // Don't use movshdup if it can be done with a shufps.
1951 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1952 /// a splat of a single element.
1953 static bool isSplatMask(SDNode *N) {
1954 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1956 // This is a splat operation if each element of the permute is the same, and
1957 // if the value doesn't reference the second vector.
1958 unsigned NumElems = N->getNumOperands();
1959 SDOperand ElementBase;
1961 for (; i != NumElems; ++i) {
1962 SDOperand Elt = N->getOperand(i);
1963 if (isa<ConstantSDNode>(Elt)) {
1969 if (!ElementBase.Val)
1972 for (; i != NumElems; ++i) {
1973 SDOperand Arg = N->getOperand(i);
1974 if (Arg.getOpcode() == ISD::UNDEF) continue;
1975 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1976 if (Arg != ElementBase) return false;
1979 // Make sure it is a splat of the first vector operand.
1980 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
1983 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1984 /// a splat of a single element and it's a 2 or 4 element mask.
1985 bool X86::isSplatMask(SDNode *N) {
1986 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1988 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
1989 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
1991 return ::isSplatMask(N);
1994 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
1995 /// specifies a splat of zero element.
1996 bool X86::isSplatLoMask(SDNode *N) {
1997 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1999 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2000 if (!isUndefOrEqual(N->getOperand(i), 0))
2005 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2006 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2008 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2009 unsigned NumOperands = N->getNumOperands();
2010 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2012 for (unsigned i = 0; i < NumOperands; ++i) {
2014 SDOperand Arg = N->getOperand(NumOperands-i-1);
2015 if (Arg.getOpcode() != ISD::UNDEF)
2016 Val = cast<ConstantSDNode>(Arg)->getValue();
2017 if (Val >= NumOperands) Val -= NumOperands;
2019 if (i != NumOperands - 1)
2026 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2027 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2029 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2031 // 8 nodes, but we only care about the last 4.
2032 for (unsigned i = 7; i >= 4; --i) {
2034 SDOperand Arg = N->getOperand(i);
2035 if (Arg.getOpcode() != ISD::UNDEF)
2036 Val = cast<ConstantSDNode>(Arg)->getValue();
2045 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2046 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2048 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2050 // 8 nodes, but we only care about the first 4.
2051 for (int i = 3; i >= 0; --i) {
2053 SDOperand Arg = N->getOperand(i);
2054 if (Arg.getOpcode() != ISD::UNDEF)
2055 Val = cast<ConstantSDNode>(Arg)->getValue();
2064 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2065 /// specifies a 8 element shuffle that can be broken into a pair of
2066 /// PSHUFHW and PSHUFLW.
2067 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2068 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2070 if (N->getNumOperands() != 8)
2073 // Lower quadword shuffled.
2074 for (unsigned i = 0; i != 4; ++i) {
2075 SDOperand Arg = N->getOperand(i);
2076 if (Arg.getOpcode() == ISD::UNDEF) continue;
2077 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2078 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2083 // Upper quadword shuffled.
2084 for (unsigned i = 4; i != 8; ++i) {
2085 SDOperand Arg = N->getOperand(i);
2086 if (Arg.getOpcode() == ISD::UNDEF) continue;
2087 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2088 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2089 if (Val < 4 || Val > 7)
2096 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2097 /// values in ther permute mask.
2098 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2099 SDOperand &V2, SDOperand &Mask,
2100 SelectionDAG &DAG) {
2101 MVT::ValueType VT = Op.getValueType();
2102 MVT::ValueType MaskVT = Mask.getValueType();
2103 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2104 unsigned NumElems = Mask.getNumOperands();
2105 SmallVector<SDOperand, 8> MaskVec;
2107 for (unsigned i = 0; i != NumElems; ++i) {
2108 SDOperand Arg = Mask.getOperand(i);
2109 if (Arg.getOpcode() == ISD::UNDEF) {
2110 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2113 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2114 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2116 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2118 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2122 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2123 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2126 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2127 /// match movhlps. The lower half elements should come from upper half of
2128 /// V1 (and in order), and the upper half elements should come from the upper
2129 /// half of V2 (and in order).
2130 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2131 unsigned NumElems = Mask->getNumOperands();
2134 for (unsigned i = 0, e = 2; i != e; ++i)
2135 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2137 for (unsigned i = 2; i != 4; ++i)
2138 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2143 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2144 /// is promoted to a vector.
2145 static inline bool isScalarLoadToVector(SDNode *N) {
2146 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2147 N = N->getOperand(0).Val;
2148 return ISD::isNON_EXTLoad(N);
2153 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2154 /// match movlp{s|d}. The lower half elements should come from lower half of
2155 /// V1 (and in order), and the upper half elements should come from the upper
2156 /// half of V2 (and in order). And since V1 will become the source of the
2157 /// MOVLP, it must be either a vector load or a scalar load to vector.
2158 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2159 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2161 // Is V2 is a vector load, don't do this transformation. We will try to use
2162 // load folding shufps op.
2163 if (ISD::isNON_EXTLoad(V2))
2166 unsigned NumElems = Mask->getNumOperands();
2167 if (NumElems != 2 && NumElems != 4)
2169 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2170 if (!isUndefOrEqual(Mask->getOperand(i), i))
2172 for (unsigned i = NumElems/2; i != NumElems; ++i)
2173 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2178 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2180 static bool isSplatVector(SDNode *N) {
2181 if (N->getOpcode() != ISD::BUILD_VECTOR)
2184 SDOperand SplatValue = N->getOperand(0);
2185 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2186 if (N->getOperand(i) != SplatValue)
2191 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2193 static bool isUndefShuffle(SDNode *N) {
2194 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2197 SDOperand V1 = N->getOperand(0);
2198 SDOperand V2 = N->getOperand(1);
2199 SDOperand Mask = N->getOperand(2);
2200 unsigned NumElems = Mask.getNumOperands();
2201 for (unsigned i = 0; i != NumElems; ++i) {
2202 SDOperand Arg = Mask.getOperand(i);
2203 if (Arg.getOpcode() != ISD::UNDEF) {
2204 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2205 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2207 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2214 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2216 static inline bool isZeroNode(SDOperand Elt) {
2217 return ((isa<ConstantSDNode>(Elt) &&
2218 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2219 (isa<ConstantFPSDNode>(Elt) &&
2220 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2223 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2224 /// to an zero vector.
2225 static bool isZeroShuffle(SDNode *N) {
2226 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2229 SDOperand V1 = N->getOperand(0);
2230 SDOperand V2 = N->getOperand(1);
2231 SDOperand Mask = N->getOperand(2);
2232 unsigned NumElems = Mask.getNumOperands();
2233 for (unsigned i = 0; i != NumElems; ++i) {
2234 SDOperand Arg = Mask.getOperand(i);
2235 if (Arg.getOpcode() != ISD::UNDEF) {
2236 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2237 if (Idx < NumElems) {
2238 unsigned Opc = V1.Val->getOpcode();
2239 if (Opc == ISD::UNDEF)
2241 if (Opc != ISD::BUILD_VECTOR ||
2242 !isZeroNode(V1.Val->getOperand(Idx)))
2244 } else if (Idx >= NumElems) {
2245 unsigned Opc = V2.Val->getOpcode();
2246 if (Opc == ISD::UNDEF)
2248 if (Opc != ISD::BUILD_VECTOR ||
2249 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2257 /// getZeroVector - Returns a vector of specified type with all zero elements.
2259 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2260 assert(MVT::isVector(VT) && "Expected a vector type");
2261 unsigned NumElems = MVT::getVectorNumElements(VT);
2262 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2263 bool isFP = MVT::isFloatingPoint(EVT);
2264 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2265 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2266 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2269 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2270 /// that point to V2 points to its first element.
2271 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2272 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2274 bool Changed = false;
2275 SmallVector<SDOperand, 8> MaskVec;
2276 unsigned NumElems = Mask.getNumOperands();
2277 for (unsigned i = 0; i != NumElems; ++i) {
2278 SDOperand Arg = Mask.getOperand(i);
2279 if (Arg.getOpcode() != ISD::UNDEF) {
2280 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2281 if (Val > NumElems) {
2282 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2286 MaskVec.push_back(Arg);
2290 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2291 &MaskVec[0], MaskVec.size());
2295 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2296 /// operation of specified width.
2297 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2298 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2299 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2301 SmallVector<SDOperand, 8> MaskVec;
2302 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2303 for (unsigned i = 1; i != NumElems; ++i)
2304 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2305 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2308 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2309 /// of specified width.
2310 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2311 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2312 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2313 SmallVector<SDOperand, 8> MaskVec;
2314 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2315 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2316 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2318 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2321 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2322 /// of specified width.
2323 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2324 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2325 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2326 unsigned Half = NumElems/2;
2327 SmallVector<SDOperand, 8> MaskVec;
2328 for (unsigned i = 0; i != Half; ++i) {
2329 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2330 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2332 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2335 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2337 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2338 SDOperand V1 = Op.getOperand(0);
2339 SDOperand Mask = Op.getOperand(2);
2340 MVT::ValueType VT = Op.getValueType();
2341 unsigned NumElems = Mask.getNumOperands();
2342 Mask = getUnpacklMask(NumElems, DAG);
2343 while (NumElems != 4) {
2344 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2347 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2349 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2350 Mask = getZeroVector(MaskVT, DAG);
2351 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2352 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2353 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2356 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2357 /// vector of zero or undef vector.
2358 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2359 unsigned NumElems, unsigned Idx,
2360 bool isZero, SelectionDAG &DAG) {
2361 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2362 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2363 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2364 SDOperand Zero = DAG.getConstant(0, EVT);
2365 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2366 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2367 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2368 &MaskVec[0], MaskVec.size());
2369 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2372 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2374 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2375 unsigned NumNonZero, unsigned NumZero,
2376 SelectionDAG &DAG, TargetLowering &TLI) {
2382 for (unsigned i = 0; i < 16; ++i) {
2383 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2384 if (ThisIsNonZero && First) {
2386 V = getZeroVector(MVT::v8i16, DAG);
2388 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2393 SDOperand ThisElt(0, 0), LastElt(0, 0);
2394 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2395 if (LastIsNonZero) {
2396 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2398 if (ThisIsNonZero) {
2399 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2400 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2401 ThisElt, DAG.getConstant(8, MVT::i8));
2403 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2408 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2409 DAG.getConstant(i/2, TLI.getPointerTy()));
2413 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2416 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
2418 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2419 unsigned NumNonZero, unsigned NumZero,
2420 SelectionDAG &DAG, TargetLowering &TLI) {
2426 for (unsigned i = 0; i < 8; ++i) {
2427 bool isNonZero = (NonZeros & (1 << i)) != 0;
2431 V = getZeroVector(MVT::v8i16, DAG);
2433 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2436 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2437 DAG.getConstant(i, TLI.getPointerTy()));
2445 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2446 // All zero's are handled with pxor.
2447 if (ISD::isBuildVectorAllZeros(Op.Val))
2450 // All one's are handled with pcmpeqd.
2451 if (ISD::isBuildVectorAllOnes(Op.Val))
2454 MVT::ValueType VT = Op.getValueType();
2455 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2456 unsigned EVTBits = MVT::getSizeInBits(EVT);
2458 unsigned NumElems = Op.getNumOperands();
2459 unsigned NumZero = 0;
2460 unsigned NumNonZero = 0;
2461 unsigned NonZeros = 0;
2462 std::set<SDOperand> Values;
2463 for (unsigned i = 0; i < NumElems; ++i) {
2464 SDOperand Elt = Op.getOperand(i);
2465 if (Elt.getOpcode() != ISD::UNDEF) {
2467 if (isZeroNode(Elt))
2470 NonZeros |= (1 << i);
2476 if (NumNonZero == 0)
2477 // Must be a mix of zero and undef. Return a zero vector.
2478 return getZeroVector(VT, DAG);
2480 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2481 if (Values.size() == 1)
2484 // Special case for single non-zero element.
2485 if (NumNonZero == 1) {
2486 unsigned Idx = CountTrailingZeros_32(NonZeros);
2487 SDOperand Item = Op.getOperand(Idx);
2488 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2490 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2491 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2494 if (EVTBits == 32) {
2495 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2496 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2498 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2499 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2500 SmallVector<SDOperand, 8> MaskVec;
2501 for (unsigned i = 0; i < NumElems; i++)
2502 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2503 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2504 &MaskVec[0], MaskVec.size());
2505 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2506 DAG.getNode(ISD::UNDEF, VT), Mask);
2510 // Let legalizer expand 2-wide build_vectors.
2514 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2515 if (EVTBits == 8 && NumElems == 16) {
2516 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2518 if (V.Val) return V;
2521 if (EVTBits == 16 && NumElems == 8) {
2522 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2524 if (V.Val) return V;
2527 // If element VT is == 32 bits, turn it into a number of shuffles.
2528 SmallVector<SDOperand, 8> V;
2530 if (NumElems == 4 && NumZero > 0) {
2531 for (unsigned i = 0; i < 4; ++i) {
2532 bool isZero = !(NonZeros & (1 << i));
2534 V[i] = getZeroVector(VT, DAG);
2536 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2539 for (unsigned i = 0; i < 2; ++i) {
2540 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2543 V[i] = V[i*2]; // Must be a zero vector.
2546 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2547 getMOVLMask(NumElems, DAG));
2550 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2551 getMOVLMask(NumElems, DAG));
2554 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2555 getUnpacklMask(NumElems, DAG));
2560 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2561 // clears the upper bits.
2562 // FIXME: we can do the same for v4f32 case when we know both parts of
2563 // the lower half come from scalar_to_vector (loadf32). We should do
2564 // that in post legalizer dag combiner with target specific hooks.
2565 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2567 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2568 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2569 SmallVector<SDOperand, 8> MaskVec;
2570 bool Reverse = (NonZeros & 0x3) == 2;
2571 for (unsigned i = 0; i < 2; ++i)
2573 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2575 MaskVec.push_back(DAG.getConstant(i, EVT));
2576 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2577 for (unsigned i = 0; i < 2; ++i)
2579 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2581 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2582 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2583 &MaskVec[0], MaskVec.size());
2584 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2587 if (Values.size() > 2) {
2588 // Expand into a number of unpckl*.
2590 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2591 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2592 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2593 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2594 for (unsigned i = 0; i < NumElems; ++i)
2595 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2597 while (NumElems != 0) {
2598 for (unsigned i = 0; i < NumElems; ++i)
2599 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
2610 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2611 SDOperand V1 = Op.getOperand(0);
2612 SDOperand V2 = Op.getOperand(1);
2613 SDOperand PermMask = Op.getOperand(2);
2614 MVT::ValueType VT = Op.getValueType();
2615 unsigned NumElems = PermMask.getNumOperands();
2616 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
2617 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
2618 bool V1IsSplat = false;
2619 bool V2IsSplat = false;
2621 if (isUndefShuffle(Op.Val))
2622 return DAG.getNode(ISD::UNDEF, VT);
2624 if (isZeroShuffle(Op.Val))
2625 return getZeroVector(VT, DAG);
2627 if (isSplatMask(PermMask.Val)) {
2628 if (NumElems <= 4) return Op;
2629 // Promote it to a v4i32 splat.
2630 return PromoteSplat(Op, DAG);
2633 if (X86::isMOVLMask(PermMask.Val))
2634 return (V1IsUndef) ? V2 : Op;
2636 if (X86::isMOVSHDUPMask(PermMask.Val) ||
2637 X86::isMOVSLDUPMask(PermMask.Val) ||
2638 X86::isMOVHLPSMask(PermMask.Val) ||
2639 X86::isMOVHPMask(PermMask.Val) ||
2640 X86::isMOVLPMask(PermMask.Val))
2643 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2644 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
2645 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2647 bool Commuted = false;
2648 V1IsSplat = isSplatVector(V1.Val);
2649 V2IsSplat = isSplatVector(V2.Val);
2650 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
2651 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2652 std::swap(V1IsSplat, V2IsSplat);
2653 std::swap(V1IsUndef, V2IsUndef);
2657 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
2658 if (V2IsUndef) return V1;
2659 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2661 // V2 is a splat, so the mask may be malformed. That is, it may point
2662 // to any V2 element. The instruction selectior won't like this. Get
2663 // a corrected mask and commute to form a proper MOVS{S|D}.
2664 SDOperand NewMask = getMOVLMask(NumElems, DAG);
2665 if (NewMask.Val != PermMask.Val)
2666 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2671 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2672 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
2673 X86::isUNPCKLMask(PermMask.Val) ||
2674 X86::isUNPCKHMask(PermMask.Val))
2678 // Normalize mask so all entries that point to V2 points to its first
2679 // element then try to match unpck{h|l} again. If match, return a
2680 // new vector_shuffle with the corrected mask.
2681 SDOperand NewMask = NormalizeMask(PermMask, DAG);
2682 if (NewMask.Val != PermMask.Val) {
2683 if (X86::isUNPCKLMask(PermMask.Val, true)) {
2684 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
2685 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2686 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
2687 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
2688 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2693 // Normalize the node to match x86 shuffle ops if needed
2694 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
2695 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2698 // Commute is back and try unpck* again.
2699 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2700 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2701 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
2702 X86::isUNPCKLMask(PermMask.Val) ||
2703 X86::isUNPCKHMask(PermMask.Val))
2707 // If VT is integer, try PSHUF* first, then SHUFP*.
2708 if (MVT::isInteger(VT)) {
2709 if (X86::isPSHUFDMask(PermMask.Val) ||
2710 X86::isPSHUFHWMask(PermMask.Val) ||
2711 X86::isPSHUFLWMask(PermMask.Val)) {
2712 if (V2.getOpcode() != ISD::UNDEF)
2713 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2714 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2718 if (X86::isSHUFPMask(PermMask.Val) &&
2719 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
2722 // Handle v8i16 shuffle high / low shuffle node pair.
2723 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
2724 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2725 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2726 SmallVector<SDOperand, 8> MaskVec;
2727 for (unsigned i = 0; i != 4; ++i)
2728 MaskVec.push_back(PermMask.getOperand(i));
2729 for (unsigned i = 4; i != 8; ++i)
2730 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2731 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2732 &MaskVec[0], MaskVec.size());
2733 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2735 for (unsigned i = 0; i != 4; ++i)
2736 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2737 for (unsigned i = 4; i != 8; ++i)
2738 MaskVec.push_back(PermMask.getOperand(i));
2739 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
2740 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2743 // Floating point cases in the other order.
2744 if (X86::isSHUFPMask(PermMask.Val))
2746 if (X86::isPSHUFDMask(PermMask.Val) ||
2747 X86::isPSHUFHWMask(PermMask.Val) ||
2748 X86::isPSHUFLWMask(PermMask.Val)) {
2749 if (V2.getOpcode() != ISD::UNDEF)
2750 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2751 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2756 if (NumElems == 4 &&
2757 // Don't do this for MMX.
2758 MVT::getSizeInBits(VT) != 64) {
2759 MVT::ValueType MaskVT = PermMask.getValueType();
2760 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2761 SmallVector<std::pair<int, int>, 8> Locs;
2762 Locs.reserve(NumElems);
2763 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2764 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2767 // If no more than two elements come from either vector. This can be
2768 // implemented with two shuffles. First shuffle gather the elements.
2769 // The second shuffle, which takes the first shuffle as both of its
2770 // vector operands, put the elements into the right order.
2771 for (unsigned i = 0; i != NumElems; ++i) {
2772 SDOperand Elt = PermMask.getOperand(i);
2773 if (Elt.getOpcode() == ISD::UNDEF) {
2774 Locs[i] = std::make_pair(-1, -1);
2776 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
2777 if (Val < NumElems) {
2778 Locs[i] = std::make_pair(0, NumLo);
2782 Locs[i] = std::make_pair(1, NumHi);
2783 if (2+NumHi < NumElems)
2784 Mask1[2+NumHi] = Elt;
2789 if (NumLo <= 2 && NumHi <= 2) {
2790 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2791 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2792 &Mask1[0], Mask1.size()));
2793 for (unsigned i = 0; i != NumElems; ++i) {
2794 if (Locs[i].first == -1)
2797 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
2798 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
2799 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
2803 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
2804 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2805 &Mask2[0], Mask2.size()));
2808 // Break it into (shuffle shuffle_hi, shuffle_lo).
2810 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2811 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2812 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
2813 unsigned MaskIdx = 0;
2815 unsigned HiIdx = NumElems/2;
2816 for (unsigned i = 0; i != NumElems; ++i) {
2817 if (i == NumElems/2) {
2823 SDOperand Elt = PermMask.getOperand(i);
2824 if (Elt.getOpcode() == ISD::UNDEF) {
2825 Locs[i] = std::make_pair(-1, -1);
2826 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
2827 Locs[i] = std::make_pair(MaskIdx, LoIdx);
2828 (*MaskPtr)[LoIdx] = Elt;
2831 Locs[i] = std::make_pair(MaskIdx, HiIdx);
2832 (*MaskPtr)[HiIdx] = Elt;
2837 SDOperand LoShuffle =
2838 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2839 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2840 &LoMask[0], LoMask.size()));
2841 SDOperand HiShuffle =
2842 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2843 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2844 &HiMask[0], HiMask.size()));
2845 SmallVector<SDOperand, 8> MaskOps;
2846 for (unsigned i = 0; i != NumElems; ++i) {
2847 if (Locs[i].first == -1) {
2848 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
2850 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
2851 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
2854 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
2855 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2856 &MaskOps[0], MaskOps.size()));
2863 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2864 if (!isa<ConstantSDNode>(Op.getOperand(1)))
2867 MVT::ValueType VT = Op.getValueType();
2868 // TODO: handle v16i8.
2869 if (MVT::getSizeInBits(VT) == 16) {
2870 // Transform it so it match pextrw which produces a 32-bit result.
2871 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
2872 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
2873 Op.getOperand(0), Op.getOperand(1));
2874 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
2875 DAG.getValueType(VT));
2876 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
2877 } else if (MVT::getSizeInBits(VT) == 32) {
2878 SDOperand Vec = Op.getOperand(0);
2879 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2882 // SHUFPS the element to the lowest double word, then movss.
2883 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2884 SmallVector<SDOperand, 8> IdxVec;
2885 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
2886 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2887 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2888 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2889 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2890 &IdxVec[0], IdxVec.size());
2891 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2892 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2893 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2894 DAG.getConstant(0, getPointerTy()));
2895 } else if (MVT::getSizeInBits(VT) == 64) {
2896 SDOperand Vec = Op.getOperand(0);
2897 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2901 // UNPCKHPD the element to the lowest double word, then movsd.
2902 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
2903 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
2904 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2905 SmallVector<SDOperand, 8> IdxVec;
2906 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
2907 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2908 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2909 &IdxVec[0], IdxVec.size());
2910 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2911 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2912 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2913 DAG.getConstant(0, getPointerTy()));
2920 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2921 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
2922 // as its second argument.
2923 MVT::ValueType VT = Op.getValueType();
2924 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
2925 SDOperand N0 = Op.getOperand(0);
2926 SDOperand N1 = Op.getOperand(1);
2927 SDOperand N2 = Op.getOperand(2);
2928 if (MVT::getSizeInBits(BaseVT) == 16) {
2929 if (N1.getValueType() != MVT::i32)
2930 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
2931 if (N2.getValueType() != MVT::i32)
2932 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
2933 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
2934 } else if (MVT::getSizeInBits(BaseVT) == 32) {
2935 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
2938 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
2939 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2940 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2941 SmallVector<SDOperand, 8> MaskVec;
2942 MaskVec.push_back(DAG.getConstant(4, BaseVT));
2943 for (unsigned i = 1; i <= 3; ++i)
2944 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2945 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
2946 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2947 &MaskVec[0], MaskVec.size()));
2949 // Use two pinsrw instructions to insert a 32 bit value.
2951 if (MVT::isFloatingPoint(N1.getValueType())) {
2952 if (ISD::isNON_EXTLoad(N1.Val)) {
2953 // Just load directly from f32mem to GR32.
2954 LoadSDNode *LD = cast<LoadSDNode>(N1);
2955 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
2956 LD->getSrcValue(), LD->getSrcValueOffset());
2958 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
2959 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
2960 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
2961 DAG.getConstant(0, getPointerTy()));
2964 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
2965 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2966 DAG.getConstant(Idx, getPointerTy()));
2967 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
2968 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2969 DAG.getConstant(Idx+1, getPointerTy()));
2970 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
2978 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2979 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
2980 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
2983 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2984 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
2985 // one of the above mentioned nodes. It has to be wrapped because otherwise
2986 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2987 // be used to form addressing mode. These wrapped nodes will be selected
2990 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
2991 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2992 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
2994 CP->getAlignment());
2995 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2996 // With PIC, the address is actually $g + Offset.
2997 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2998 !Subtarget->isPICStyleRIPRel()) {
2999 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3000 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3008 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3009 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3010 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3011 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3012 // With PIC, the address is actually $g + Offset.
3013 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3014 !Subtarget->isPICStyleRIPRel()) {
3015 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3016 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3020 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3021 // load the value at address GV, not the value of GV itself. This means that
3022 // the GlobalAddress must be in the base or index register of the address, not
3023 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3024 // The same applies for external symbols during PIC codegen
3025 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3026 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3031 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3033 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3034 const MVT::ValueType PtrVT) {
3036 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
3037 DAG.getNode(X86ISD::GlobalBaseReg,
3039 InFlag = Chain.getValue(1);
3041 // emit leal symbol@TLSGD(,%ebx,1), %eax
3042 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3043 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3044 GA->getValueType(0),
3046 SDOperand Ops[] = { Chain, TGA, InFlag };
3047 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3048 InFlag = Result.getValue(2);
3049 Chain = Result.getValue(1);
3051 // call ___tls_get_addr. This function receives its argument in
3052 // the register EAX.
3053 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
3054 InFlag = Chain.getValue(1);
3056 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3057 SDOperand Ops1[] = { Chain,
3058 DAG.getTargetExternalSymbol("___tls_get_addr",
3060 DAG.getRegister(X86::EAX, PtrVT),
3061 DAG.getRegister(X86::EBX, PtrVT),
3063 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
3064 InFlag = Chain.getValue(1);
3066 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
3069 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
3070 // "local exec" model.
3072 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3073 const MVT::ValueType PtrVT) {
3074 // Get the Thread Pointer
3075 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
3076 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
3078 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3079 GA->getValueType(0),
3081 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
3083 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
3084 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset, NULL, 0);
3086 // The address of the thread local variable is the add of the thread
3087 // pointer with the offset of the variable.
3088 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
3092 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
3093 // TODO: implement the "local dynamic" model
3094 // TODO: implement the "initial exec"model for pic executables
3095 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
3096 "TLS not implemented for non-ELF and 64-bit targets");
3097 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3098 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
3099 // otherwise use the "Local Exec"TLS Model
3100 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
3101 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
3103 return LowerToTLSExecModel(GA, DAG, getPointerTy());
3107 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3108 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3109 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3110 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3111 // With PIC, the address is actually $g + Offset.
3112 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3113 !Subtarget->isPICStyleRIPRel()) {
3114 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3115 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3122 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3123 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3124 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3125 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3126 // With PIC, the address is actually $g + Offset.
3127 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3128 !Subtarget->isPICStyleRIPRel()) {
3129 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3130 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3137 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3138 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3139 "Not an i64 shift!");
3140 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3141 SDOperand ShOpLo = Op.getOperand(0);
3142 SDOperand ShOpHi = Op.getOperand(1);
3143 SDOperand ShAmt = Op.getOperand(2);
3144 SDOperand Tmp1 = isSRA ?
3145 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3146 DAG.getConstant(0, MVT::i32);
3148 SDOperand Tmp2, Tmp3;
3149 if (Op.getOpcode() == ISD::SHL_PARTS) {
3150 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3151 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3153 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3154 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3157 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3158 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3159 DAG.getConstant(32, MVT::i8));
3160 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
3161 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
3164 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3166 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3167 SmallVector<SDOperand, 4> Ops;
3168 if (Op.getOpcode() == ISD::SHL_PARTS) {
3169 Ops.push_back(Tmp2);
3170 Ops.push_back(Tmp3);
3172 Ops.push_back(InFlag);
3173 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3174 InFlag = Hi.getValue(1);
3177 Ops.push_back(Tmp3);
3178 Ops.push_back(Tmp1);
3180 Ops.push_back(InFlag);
3181 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3183 Ops.push_back(Tmp2);
3184 Ops.push_back(Tmp3);
3186 Ops.push_back(InFlag);
3187 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3188 InFlag = Lo.getValue(1);
3191 Ops.push_back(Tmp3);
3192 Ops.push_back(Tmp1);
3194 Ops.push_back(InFlag);
3195 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3198 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3202 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3205 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3206 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3207 Op.getOperand(0).getValueType() >= MVT::i16 &&
3208 "Unknown SINT_TO_FP to lower!");
3211 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3212 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3213 MachineFunction &MF = DAG.getMachineFunction();
3214 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3215 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3216 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3217 StackSlot, NULL, 0);
3222 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3224 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3225 SmallVector<SDOperand, 8> Ops;
3226 Ops.push_back(Chain);
3227 Ops.push_back(StackSlot);
3228 Ops.push_back(DAG.getValueType(SrcVT));
3229 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3230 Tys, &Ops[0], Ops.size());
3233 Chain = Result.getValue(1);
3234 SDOperand InFlag = Result.getValue(2);
3236 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3237 // shouldn't be necessary except that RFP cannot be live across
3238 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3239 MachineFunction &MF = DAG.getMachineFunction();
3240 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3241 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3242 Tys = DAG.getVTList(MVT::Other);
3243 SmallVector<SDOperand, 8> Ops;
3244 Ops.push_back(Chain);
3245 Ops.push_back(Result);
3246 Ops.push_back(StackSlot);
3247 Ops.push_back(DAG.getValueType(Op.getValueType()));
3248 Ops.push_back(InFlag);
3249 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3250 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3256 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3257 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3258 "Unknown FP_TO_SINT to lower!");
3259 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3261 MachineFunction &MF = DAG.getMachineFunction();
3262 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3263 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3264 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3267 switch (Op.getValueType()) {
3268 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3269 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3270 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3271 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3274 SDOperand Chain = DAG.getEntryNode();
3275 SDOperand Value = Op.getOperand(0);
3277 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3278 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3279 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3281 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3283 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3284 Chain = Value.getValue(1);
3285 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3286 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3289 // Build the FP_TO_INT*_IN_MEM
3290 SDOperand Ops[] = { Chain, Value, StackSlot };
3291 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3294 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3297 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3298 MVT::ValueType VT = Op.getValueType();
3299 const Type *OpNTy = MVT::getTypeForValueType(VT);
3300 std::vector<Constant*> CV;
3301 if (VT == MVT::f64) {
3302 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3303 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3305 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3306 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3307 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3308 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3310 Constant *CS = ConstantStruct::get(CV);
3311 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3312 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3313 SmallVector<SDOperand, 3> Ops;
3314 Ops.push_back(DAG.getEntryNode());
3315 Ops.push_back(CPIdx);
3316 Ops.push_back(DAG.getSrcValue(NULL));
3317 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3318 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3321 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3322 MVT::ValueType VT = Op.getValueType();
3323 const Type *OpNTy = MVT::getTypeForValueType(VT);
3324 std::vector<Constant*> CV;
3325 if (VT == MVT::f64) {
3326 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3327 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3329 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3330 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3331 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3332 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3334 Constant *CS = ConstantStruct::get(CV);
3335 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3336 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3337 SmallVector<SDOperand, 3> Ops;
3338 Ops.push_back(DAG.getEntryNode());
3339 Ops.push_back(CPIdx);
3340 Ops.push_back(DAG.getSrcValue(NULL));
3341 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3342 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3345 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3346 SDOperand Op0 = Op.getOperand(0);
3347 SDOperand Op1 = Op.getOperand(1);
3348 MVT::ValueType VT = Op.getValueType();
3349 MVT::ValueType SrcVT = Op1.getValueType();
3350 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3352 // If second operand is smaller, extend it first.
3353 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3354 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3358 // First get the sign bit of second operand.
3359 std::vector<Constant*> CV;
3360 if (SrcVT == MVT::f64) {
3361 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3362 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3364 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3365 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3366 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3367 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3369 Constant *CS = ConstantStruct::get(CV);
3370 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3371 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3372 SmallVector<SDOperand, 3> Ops;
3373 Ops.push_back(DAG.getEntryNode());
3374 Ops.push_back(CPIdx);
3375 Ops.push_back(DAG.getSrcValue(NULL));
3376 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3377 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3379 // Shift sign bit right or left if the two operands have different types.
3380 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3381 // Op0 is MVT::f32, Op1 is MVT::f64.
3382 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3383 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3384 DAG.getConstant(32, MVT::i32));
3385 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3386 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3387 DAG.getConstant(0, getPointerTy()));
3390 // Clear first operand sign bit.
3392 if (VT == MVT::f64) {
3393 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3394 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3396 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3397 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3398 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3399 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3401 CS = ConstantStruct::get(CV);
3402 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3403 Tys = DAG.getVTList(VT, MVT::Other);
3405 Ops.push_back(DAG.getEntryNode());
3406 Ops.push_back(CPIdx);
3407 Ops.push_back(DAG.getSrcValue(NULL));
3408 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3409 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3411 // Or the value with the sign bit.
3412 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3415 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3417 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3419 SDOperand Op0 = Op.getOperand(0);
3420 SDOperand Op1 = Op.getOperand(1);
3421 SDOperand CC = Op.getOperand(2);
3422 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3423 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3424 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3425 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3428 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3430 SDOperand Ops1[] = { Chain, Op0, Op1 };
3431 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3432 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3433 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3436 assert(isFP && "Illegal integer SetCC!");
3438 SDOperand COps[] = { Chain, Op0, Op1 };
3439 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3441 switch (SetCCOpcode) {
3442 default: assert(false && "Illegal floating point SetCC!");
3443 case ISD::SETOEQ: { // !PF & ZF
3444 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3445 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3446 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3448 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3449 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3451 case ISD::SETUNE: { // PF | !ZF
3452 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3453 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3454 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3456 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3457 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3462 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3463 bool addTest = true;
3464 SDOperand Chain = DAG.getEntryNode();
3465 SDOperand Cond = Op.getOperand(0);
3467 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3469 if (Cond.getOpcode() == ISD::SETCC)
3470 Cond = LowerSETCC(Cond, DAG, Chain);
3472 if (Cond.getOpcode() == X86ISD::SETCC) {
3473 CC = Cond.getOperand(0);
3475 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3476 // (since flag operand cannot be shared). Use it as the condition setting
3477 // operand in place of the X86ISD::SETCC.
3478 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3479 // to use a test instead of duplicating the X86ISD::CMP (for register
3480 // pressure reason)?
3481 SDOperand Cmp = Cond.getOperand(1);
3482 unsigned Opc = Cmp.getOpcode();
3483 bool IllegalFPCMov = !X86ScalarSSE &&
3484 MVT::isFloatingPoint(Op.getValueType()) &&
3485 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3486 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3488 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3489 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3495 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3496 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3497 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3500 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3501 SmallVector<SDOperand, 4> Ops;
3502 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3503 // condition is true.
3504 Ops.push_back(Op.getOperand(2));
3505 Ops.push_back(Op.getOperand(1));
3507 Ops.push_back(Cond.getValue(1));
3508 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3511 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3512 bool addTest = true;
3513 SDOperand Chain = Op.getOperand(0);
3514 SDOperand Cond = Op.getOperand(1);
3515 SDOperand Dest = Op.getOperand(2);
3517 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3519 if (Cond.getOpcode() == ISD::SETCC)
3520 Cond = LowerSETCC(Cond, DAG, Chain);
3522 if (Cond.getOpcode() == X86ISD::SETCC) {
3523 CC = Cond.getOperand(0);
3525 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3526 // (since flag operand cannot be shared). Use it as the condition setting
3527 // operand in place of the X86ISD::SETCC.
3528 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3529 // to use a test instead of duplicating the X86ISD::CMP (for register
3530 // pressure reason)?
3531 SDOperand Cmp = Cond.getOperand(1);
3532 unsigned Opc = Cmp.getOpcode();
3533 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3534 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3535 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3541 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3542 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3543 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3545 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3546 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3549 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3550 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3552 if (Subtarget->is64Bit())
3553 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3555 switch (CallingConv) {
3557 assert(0 && "Unsupported calling convention");
3558 case CallingConv::Fast:
3559 // TODO: Implement fastcc
3561 case CallingConv::C:
3562 case CallingConv::X86_StdCall:
3563 return LowerCCCCallTo(Op, DAG, CallingConv);
3564 case CallingConv::X86_FastCall:
3565 return LowerFastCCCallTo(Op, DAG, CallingConv);
3570 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
3571 // Calls to _alloca is needed to probe the stack when allocating more than 4k
3572 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
3573 // that the guard pages used by the OS virtual memory manager are allocated in
3574 // correct sequence.
3575 SDOperand X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
3576 SelectionDAG &DAG) {
3577 assert(Subtarget->isTargetCygMing() &&
3578 "This should be used only on Cygwin/Mingw targets");
3581 SDOperand Chain = Op.getOperand(0);
3582 SDOperand Size = Op.getOperand(1);
3583 // FIXME: Ensure alignment here
3585 TargetLowering::ArgListTy Args;
3586 TargetLowering::ArgListEntry Entry;
3587 MVT::ValueType IntPtr = getPointerTy();
3588 MVT::ValueType SPTy = (Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
3589 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3592 Entry.Ty = IntPtrTy;
3593 Entry.isInReg = true; // Should pass in EAX
3594 Args.push_back(Entry);
3595 std::pair<SDOperand, SDOperand> CallResult =
3596 LowerCallTo(Chain, IntPtrTy, false, false, CallingConv::C, false,
3597 DAG.getExternalSymbol("_alloca", IntPtr), Args, DAG);
3599 SDOperand SP = DAG.getCopyFromReg(CallResult.second, X86StackPtr, SPTy);
3601 std::vector<MVT::ValueType> Tys;
3602 Tys.push_back(SPTy);
3603 Tys.push_back(MVT::Other);
3604 SDOperand Ops[2] = { SP, CallResult.second };
3605 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
3609 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3610 MachineFunction &MF = DAG.getMachineFunction();
3611 const Function* Fn = MF.getFunction();
3612 if (Fn->hasExternalLinkage() &&
3613 Subtarget->isTargetCygMing() &&
3614 Fn->getName() == "main")
3615 MF.getInfo<X86MachineFunctionInfo>()->setForceFramePointer(true);
3617 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3618 if (Subtarget->is64Bit())
3619 return LowerX86_64CCCArguments(Op, DAG);
3623 assert(0 && "Unsupported calling convention");
3624 case CallingConv::Fast:
3625 // TODO: implement fastcc.
3628 case CallingConv::C:
3629 return LowerCCCArguments(Op, DAG);
3630 case CallingConv::X86_StdCall:
3631 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(StdCall);
3632 return LowerCCCArguments(Op, DAG, true);
3633 case CallingConv::X86_FastCall:
3634 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(FastCall);
3635 return LowerFastCCArguments(Op, DAG);
3639 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3640 SDOperand InFlag(0, 0);
3641 SDOperand Chain = Op.getOperand(0);
3643 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3644 if (Align == 0) Align = 1;
3646 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3647 // If not DWORD aligned, call memset if size is less than the threshold.
3648 // It knows how to align to the right boundary first.
3649 if ((Align & 3) != 0 ||
3650 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3651 MVT::ValueType IntPtr = getPointerTy();
3652 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3653 TargetLowering::ArgListTy Args;
3654 TargetLowering::ArgListEntry Entry;
3655 Entry.Node = Op.getOperand(1);
3656 Entry.Ty = IntPtrTy;
3657 Args.push_back(Entry);
3658 // Extend the unsigned i8 argument to be an int value for the call.
3659 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3660 Entry.Ty = IntPtrTy;
3661 Args.push_back(Entry);
3662 Entry.Node = Op.getOperand(3);
3663 Args.push_back(Entry);
3664 std::pair<SDOperand,SDOperand> CallResult =
3665 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3666 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3667 return CallResult.second;
3672 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3673 unsigned BytesLeft = 0;
3674 bool TwoRepStos = false;
3677 uint64_t Val = ValC->getValue() & 255;
3679 // If the value is a constant, then we can potentially use larger sets.
3680 switch (Align & 3) {
3681 case 2: // WORD aligned
3684 Val = (Val << 8) | Val;
3686 case 0: // DWORD aligned
3689 Val = (Val << 8) | Val;
3690 Val = (Val << 16) | Val;
3691 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
3694 Val = (Val << 32) | Val;
3697 default: // Byte aligned
3700 Count = Op.getOperand(3);
3704 if (AVT > MVT::i8) {
3706 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3707 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3708 BytesLeft = I->getValue() % UBytes;
3710 assert(AVT >= MVT::i32 &&
3711 "Do not use rep;stos if not at least DWORD aligned");
3712 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3713 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3718 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
3720 InFlag = Chain.getValue(1);
3723 Count = Op.getOperand(3);
3724 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
3725 InFlag = Chain.getValue(1);
3728 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3730 InFlag = Chain.getValue(1);
3731 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3732 Op.getOperand(1), InFlag);
3733 InFlag = Chain.getValue(1);
3735 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3736 SmallVector<SDOperand, 8> Ops;
3737 Ops.push_back(Chain);
3738 Ops.push_back(DAG.getValueType(AVT));
3739 Ops.push_back(InFlag);
3740 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3743 InFlag = Chain.getValue(1);
3744 Count = Op.getOperand(3);
3745 MVT::ValueType CVT = Count.getValueType();
3746 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3747 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3748 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3750 InFlag = Chain.getValue(1);
3751 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3753 Ops.push_back(Chain);
3754 Ops.push_back(DAG.getValueType(MVT::i8));
3755 Ops.push_back(InFlag);
3756 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3757 } else if (BytesLeft) {
3758 // Issue stores for the last 1 - 7 bytes.
3760 unsigned Val = ValC->getValue() & 255;
3761 unsigned Offset = I->getValue() - BytesLeft;
3762 SDOperand DstAddr = Op.getOperand(1);
3763 MVT::ValueType AddrVT = DstAddr.getValueType();
3764 if (BytesLeft >= 4) {
3765 Val = (Val << 8) | Val;
3766 Val = (Val << 16) | Val;
3767 Value = DAG.getConstant(Val, MVT::i32);
3768 Chain = DAG.getStore(Chain, Value,
3769 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3770 DAG.getConstant(Offset, AddrVT)),
3775 if (BytesLeft >= 2) {
3776 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
3777 Chain = DAG.getStore(Chain, Value,
3778 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3779 DAG.getConstant(Offset, AddrVT)),
3784 if (BytesLeft == 1) {
3785 Value = DAG.getConstant(Val, MVT::i8);
3786 Chain = DAG.getStore(Chain, Value,
3787 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3788 DAG.getConstant(Offset, AddrVT)),
3796 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
3797 SDOperand Chain = Op.getOperand(0);
3799 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3800 if (Align == 0) Align = 1;
3802 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3803 // If not DWORD aligned, call memcpy if size is less than the threshold.
3804 // It knows how to align to the right boundary first.
3805 if ((Align & 3) != 0 ||
3806 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3807 MVT::ValueType IntPtr = getPointerTy();
3808 TargetLowering::ArgListTy Args;
3809 TargetLowering::ArgListEntry Entry;
3810 Entry.Ty = getTargetData()->getIntPtrType();
3811 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
3812 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
3813 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
3814 std::pair<SDOperand,SDOperand> CallResult =
3815 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3816 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
3817 return CallResult.second;
3822 unsigned BytesLeft = 0;
3823 bool TwoRepMovs = false;
3824 switch (Align & 3) {
3825 case 2: // WORD aligned
3828 case 0: // DWORD aligned
3830 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
3833 default: // Byte aligned
3835 Count = Op.getOperand(3);
3839 if (AVT > MVT::i8) {
3841 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3842 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3843 BytesLeft = I->getValue() % UBytes;
3845 assert(AVT >= MVT::i32 &&
3846 "Do not use rep;movs if not at least DWORD aligned");
3847 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3848 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3853 SDOperand InFlag(0, 0);
3854 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3856 InFlag = Chain.getValue(1);
3857 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3858 Op.getOperand(1), InFlag);
3859 InFlag = Chain.getValue(1);
3860 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
3861 Op.getOperand(2), InFlag);
3862 InFlag = Chain.getValue(1);
3864 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3865 SmallVector<SDOperand, 8> Ops;
3866 Ops.push_back(Chain);
3867 Ops.push_back(DAG.getValueType(AVT));
3868 Ops.push_back(InFlag);
3869 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3872 InFlag = Chain.getValue(1);
3873 Count = Op.getOperand(3);
3874 MVT::ValueType CVT = Count.getValueType();
3875 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3876 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3877 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3879 InFlag = Chain.getValue(1);
3880 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3882 Ops.push_back(Chain);
3883 Ops.push_back(DAG.getValueType(MVT::i8));
3884 Ops.push_back(InFlag);
3885 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3886 } else if (BytesLeft) {
3887 // Issue loads and stores for the last 1 - 7 bytes.
3888 unsigned Offset = I->getValue() - BytesLeft;
3889 SDOperand DstAddr = Op.getOperand(1);
3890 MVT::ValueType DstVT = DstAddr.getValueType();
3891 SDOperand SrcAddr = Op.getOperand(2);
3892 MVT::ValueType SrcVT = SrcAddr.getValueType();
3894 if (BytesLeft >= 4) {
3895 Value = DAG.getLoad(MVT::i32, Chain,
3896 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3897 DAG.getConstant(Offset, SrcVT)),
3899 Chain = Value.getValue(1);
3900 Chain = DAG.getStore(Chain, Value,
3901 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3902 DAG.getConstant(Offset, DstVT)),
3907 if (BytesLeft >= 2) {
3908 Value = DAG.getLoad(MVT::i16, Chain,
3909 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3910 DAG.getConstant(Offset, SrcVT)),
3912 Chain = Value.getValue(1);
3913 Chain = DAG.getStore(Chain, Value,
3914 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3915 DAG.getConstant(Offset, DstVT)),
3921 if (BytesLeft == 1) {
3922 Value = DAG.getLoad(MVT::i8, Chain,
3923 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3924 DAG.getConstant(Offset, SrcVT)),
3926 Chain = Value.getValue(1);
3927 Chain = DAG.getStore(Chain, Value,
3928 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3929 DAG.getConstant(Offset, DstVT)),
3938 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
3939 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3940 SDOperand TheOp = Op.getOperand(0);
3941 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
3942 if (Subtarget->is64Bit()) {
3943 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
3944 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
3945 MVT::i64, Copy1.getValue(2));
3946 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
3947 DAG.getConstant(32, MVT::i8));
3949 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
3952 Tys = DAG.getVTList(MVT::i64, MVT::Other);
3953 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
3956 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
3957 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
3958 MVT::i32, Copy1.getValue(2));
3959 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
3960 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
3961 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
3964 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
3965 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
3967 if (!Subtarget->is64Bit()) {
3968 // vastart just stores the address of the VarArgsFrameIndex slot into the
3969 // memory location argument.
3970 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3971 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
3976 // gp_offset (0 - 6 * 8)
3977 // fp_offset (48 - 48 + 8 * 16)
3978 // overflow_arg_area (point to parameters coming in memory).
3980 SmallVector<SDOperand, 8> MemOps;
3981 SDOperand FIN = Op.getOperand(1);
3983 SDOperand Store = DAG.getStore(Op.getOperand(0),
3984 DAG.getConstant(VarArgsGPOffset, MVT::i32),
3985 FIN, SV->getValue(), SV->getOffset());
3986 MemOps.push_back(Store);
3989 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3990 DAG.getConstant(4, getPointerTy()));
3991 Store = DAG.getStore(Op.getOperand(0),
3992 DAG.getConstant(VarArgsFPOffset, MVT::i32),
3993 FIN, SV->getValue(), SV->getOffset());
3994 MemOps.push_back(Store);
3996 // Store ptr to overflow_arg_area
3997 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3998 DAG.getConstant(4, getPointerTy()));
3999 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4000 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4002 MemOps.push_back(Store);
4004 // Store ptr to reg_save_area.
4005 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4006 DAG.getConstant(8, getPointerTy()));
4007 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4008 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4010 MemOps.push_back(Store);
4011 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4014 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
4015 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
4016 SDOperand Chain = Op.getOperand(0);
4017 SDOperand DstPtr = Op.getOperand(1);
4018 SDOperand SrcPtr = Op.getOperand(2);
4019 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
4020 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
4022 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
4023 SrcSV->getValue(), SrcSV->getOffset());
4024 Chain = SrcPtr.getValue(1);
4025 for (unsigned i = 0; i < 3; ++i) {
4026 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
4027 SrcSV->getValue(), SrcSV->getOffset());
4028 Chain = Val.getValue(1);
4029 Chain = DAG.getStore(Chain, Val, DstPtr,
4030 DstSV->getValue(), DstSV->getOffset());
4033 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
4034 DAG.getConstant(8, getPointerTy()));
4035 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
4036 DAG.getConstant(8, getPointerTy()));
4042 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4043 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4045 default: return SDOperand(); // Don't custom lower most intrinsics.
4046 // Comparison intrinsics.
4047 case Intrinsic::x86_sse_comieq_ss:
4048 case Intrinsic::x86_sse_comilt_ss:
4049 case Intrinsic::x86_sse_comile_ss:
4050 case Intrinsic::x86_sse_comigt_ss:
4051 case Intrinsic::x86_sse_comige_ss:
4052 case Intrinsic::x86_sse_comineq_ss:
4053 case Intrinsic::x86_sse_ucomieq_ss:
4054 case Intrinsic::x86_sse_ucomilt_ss:
4055 case Intrinsic::x86_sse_ucomile_ss:
4056 case Intrinsic::x86_sse_ucomigt_ss:
4057 case Intrinsic::x86_sse_ucomige_ss:
4058 case Intrinsic::x86_sse_ucomineq_ss:
4059 case Intrinsic::x86_sse2_comieq_sd:
4060 case Intrinsic::x86_sse2_comilt_sd:
4061 case Intrinsic::x86_sse2_comile_sd:
4062 case Intrinsic::x86_sse2_comigt_sd:
4063 case Intrinsic::x86_sse2_comige_sd:
4064 case Intrinsic::x86_sse2_comineq_sd:
4065 case Intrinsic::x86_sse2_ucomieq_sd:
4066 case Intrinsic::x86_sse2_ucomilt_sd:
4067 case Intrinsic::x86_sse2_ucomile_sd:
4068 case Intrinsic::x86_sse2_ucomigt_sd:
4069 case Intrinsic::x86_sse2_ucomige_sd:
4070 case Intrinsic::x86_sse2_ucomineq_sd: {
4072 ISD::CondCode CC = ISD::SETCC_INVALID;
4075 case Intrinsic::x86_sse_comieq_ss:
4076 case Intrinsic::x86_sse2_comieq_sd:
4080 case Intrinsic::x86_sse_comilt_ss:
4081 case Intrinsic::x86_sse2_comilt_sd:
4085 case Intrinsic::x86_sse_comile_ss:
4086 case Intrinsic::x86_sse2_comile_sd:
4090 case Intrinsic::x86_sse_comigt_ss:
4091 case Intrinsic::x86_sse2_comigt_sd:
4095 case Intrinsic::x86_sse_comige_ss:
4096 case Intrinsic::x86_sse2_comige_sd:
4100 case Intrinsic::x86_sse_comineq_ss:
4101 case Intrinsic::x86_sse2_comineq_sd:
4105 case Intrinsic::x86_sse_ucomieq_ss:
4106 case Intrinsic::x86_sse2_ucomieq_sd:
4107 Opc = X86ISD::UCOMI;
4110 case Intrinsic::x86_sse_ucomilt_ss:
4111 case Intrinsic::x86_sse2_ucomilt_sd:
4112 Opc = X86ISD::UCOMI;
4115 case Intrinsic::x86_sse_ucomile_ss:
4116 case Intrinsic::x86_sse2_ucomile_sd:
4117 Opc = X86ISD::UCOMI;
4120 case Intrinsic::x86_sse_ucomigt_ss:
4121 case Intrinsic::x86_sse2_ucomigt_sd:
4122 Opc = X86ISD::UCOMI;
4125 case Intrinsic::x86_sse_ucomige_ss:
4126 case Intrinsic::x86_sse2_ucomige_sd:
4127 Opc = X86ISD::UCOMI;
4130 case Intrinsic::x86_sse_ucomineq_ss:
4131 case Intrinsic::x86_sse2_ucomineq_sd:
4132 Opc = X86ISD::UCOMI;
4138 SDOperand LHS = Op.getOperand(1);
4139 SDOperand RHS = Op.getOperand(2);
4140 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4142 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4143 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
4144 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
4145 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
4146 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
4147 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
4148 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4153 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4154 // Depths > 0 not supported yet!
4155 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4158 // Just load the return address
4159 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4160 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4163 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4164 // Depths > 0 not supported yet!
4165 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4168 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4169 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4170 DAG.getConstant(4, getPointerTy()));
4173 /// LowerOperation - Provide custom lowering hooks for some operations.
4175 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4176 switch (Op.getOpcode()) {
4177 default: assert(0 && "Should not custom lower this!");
4178 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4179 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4180 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4181 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4182 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4183 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4184 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4185 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4186 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4187 case ISD::SHL_PARTS:
4188 case ISD::SRA_PARTS:
4189 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4190 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4191 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4192 case ISD::FABS: return LowerFABS(Op, DAG);
4193 case ISD::FNEG: return LowerFNEG(Op, DAG);
4194 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4195 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
4196 case ISD::SELECT: return LowerSELECT(Op, DAG);
4197 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4198 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4199 case ISD::CALL: return LowerCALL(Op, DAG);
4200 case ISD::RET: return LowerRET(Op, DAG);
4201 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4202 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4203 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4204 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4205 case ISD::VASTART: return LowerVASTART(Op, DAG);
4206 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
4207 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4208 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4209 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4210 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
4215 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4217 default: return NULL;
4218 case X86ISD::SHLD: return "X86ISD::SHLD";
4219 case X86ISD::SHRD: return "X86ISD::SHRD";
4220 case X86ISD::FAND: return "X86ISD::FAND";
4221 case X86ISD::FOR: return "X86ISD::FOR";
4222 case X86ISD::FXOR: return "X86ISD::FXOR";
4223 case X86ISD::FSRL: return "X86ISD::FSRL";
4224 case X86ISD::FILD: return "X86ISD::FILD";
4225 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
4226 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
4227 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
4228 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
4229 case X86ISD::FLD: return "X86ISD::FLD";
4230 case X86ISD::FST: return "X86ISD::FST";
4231 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
4232 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
4233 case X86ISD::CALL: return "X86ISD::CALL";
4234 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
4235 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
4236 case X86ISD::CMP: return "X86ISD::CMP";
4237 case X86ISD::COMI: return "X86ISD::COMI";
4238 case X86ISD::UCOMI: return "X86ISD::UCOMI";
4239 case X86ISD::SETCC: return "X86ISD::SETCC";
4240 case X86ISD::CMOV: return "X86ISD::CMOV";
4241 case X86ISD::BRCOND: return "X86ISD::BRCOND";
4242 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
4243 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
4244 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
4245 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
4246 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
4247 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
4248 case X86ISD::Wrapper: return "X86ISD::Wrapper";
4249 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4250 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4251 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4252 case X86ISD::FMAX: return "X86ISD::FMAX";
4253 case X86ISD::FMIN: return "X86ISD::FMIN";
4254 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
4255 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
4259 // isLegalAddressingMode - Return true if the addressing mode represented
4260 // by AM is legal for this target, for a load/store of the specified type.
4261 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
4262 const Type *Ty) const {
4263 // X86 supports extremely general addressing modes.
4265 // X86 allows a sign-extended 32-bit immediate field as a displacement.
4266 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
4270 // X86-64 only supports addr of globals in small code model.
4271 if (Subtarget->is64Bit() &&
4272 getTargetMachine().getCodeModel() != CodeModel::Small)
4275 // We can only fold this if we don't need a load either.
4276 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
4286 // These scales always work.
4291 // These scales are formed with basereg+scalereg. Only accept if there is
4296 default: // Other stuff never works.
4304 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4305 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4306 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4307 /// are assumed to be legal.
4309 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4310 // Only do shuffles on 128-bit vector types for now.
4311 if (MVT::getSizeInBits(VT) == 64) return false;
4312 return (Mask.Val->getNumOperands() <= 4 ||
4313 isSplatMask(Mask.Val) ||
4314 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4315 X86::isUNPCKLMask(Mask.Val) ||
4316 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4317 X86::isUNPCKH_v_undef_Mask(Mask.Val) ||
4318 X86::isUNPCKHMask(Mask.Val));
4321 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4323 SelectionDAG &DAG) const {
4324 unsigned NumElts = BVOps.size();
4325 // Only do shuffles on 128-bit vector types for now.
4326 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4327 if (NumElts == 2) return true;
4329 return (isMOVLMask(&BVOps[0], 4) ||
4330 isCommutedMOVL(&BVOps[0], 4, true) ||
4331 isSHUFPMask(&BVOps[0], 4) ||
4332 isCommutedSHUFP(&BVOps[0], 4));
4337 //===----------------------------------------------------------------------===//
4338 // X86 Scheduler Hooks
4339 //===----------------------------------------------------------------------===//
4342 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4343 MachineBasicBlock *BB) {
4344 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4345 switch (MI->getOpcode()) {
4346 default: assert(false && "Unexpected instr type to insert");
4347 case X86::CMOV_FR32:
4348 case X86::CMOV_FR64:
4349 case X86::CMOV_V4F32:
4350 case X86::CMOV_V2F64:
4351 case X86::CMOV_V2I64: {
4352 // To "insert" a SELECT_CC instruction, we actually have to insert the
4353 // diamond control-flow pattern. The incoming instruction knows the
4354 // destination vreg to set, the condition code register to branch on, the
4355 // true/false values to select between, and a branch opcode to use.
4356 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4357 ilist<MachineBasicBlock>::iterator It = BB;
4363 // cmpTY ccX, r1, r2
4365 // fallthrough --> copy0MBB
4366 MachineBasicBlock *thisMBB = BB;
4367 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4368 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4370 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4371 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4372 MachineFunction *F = BB->getParent();
4373 F->getBasicBlockList().insert(It, copy0MBB);
4374 F->getBasicBlockList().insert(It, sinkMBB);
4375 // Update machine-CFG edges by first adding all successors of the current
4376 // block to the new block which will contain the Phi node for the select.
4377 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4378 e = BB->succ_end(); i != e; ++i)
4379 sinkMBB->addSuccessor(*i);
4380 // Next, remove all successors of the current block, and add the true
4381 // and fallthrough blocks as its successors.
4382 while(!BB->succ_empty())
4383 BB->removeSuccessor(BB->succ_begin());
4384 BB->addSuccessor(copy0MBB);
4385 BB->addSuccessor(sinkMBB);
4388 // %FalseValue = ...
4389 // # fallthrough to sinkMBB
4392 // Update machine-CFG edges
4393 BB->addSuccessor(sinkMBB);
4396 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4399 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4400 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4401 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4403 delete MI; // The pseudo instruction is gone now.
4407 case X86::FP_TO_INT16_IN_MEM:
4408 case X86::FP_TO_INT32_IN_MEM:
4409 case X86::FP_TO_INT64_IN_MEM: {
4410 // Change the floating point control register to use "round towards zero"
4411 // mode when truncating to an integer value.
4412 MachineFunction *F = BB->getParent();
4413 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4414 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4416 // Load the old value of the high byte of the control word...
4418 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4419 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4421 // Set the high part to be round to zero...
4422 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4425 // Reload the modified control word now...
4426 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4428 // Restore the memory image of control word to original value
4429 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4432 // Get the X86 opcode to use.
4434 switch (MI->getOpcode()) {
4435 default: assert(0 && "illegal opcode!");
4436 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4437 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4438 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4442 MachineOperand &Op = MI->getOperand(0);
4443 if (Op.isRegister()) {
4444 AM.BaseType = X86AddressMode::RegBase;
4445 AM.Base.Reg = Op.getReg();
4447 AM.BaseType = X86AddressMode::FrameIndexBase;
4448 AM.Base.FrameIndex = Op.getFrameIndex();
4450 Op = MI->getOperand(1);
4451 if (Op.isImmediate())
4452 AM.Scale = Op.getImm();
4453 Op = MI->getOperand(2);
4454 if (Op.isImmediate())
4455 AM.IndexReg = Op.getImm();
4456 Op = MI->getOperand(3);
4457 if (Op.isGlobalAddress()) {
4458 AM.GV = Op.getGlobal();
4460 AM.Disp = Op.getImm();
4462 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4463 .addReg(MI->getOperand(4).getReg());
4465 // Reload the original control word now.
4466 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4468 delete MI; // The pseudo instruction is gone now.
4474 //===----------------------------------------------------------------------===//
4475 // X86 Optimization Hooks
4476 //===----------------------------------------------------------------------===//
4478 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4480 uint64_t &KnownZero,
4482 unsigned Depth) const {
4483 unsigned Opc = Op.getOpcode();
4484 assert((Opc >= ISD::BUILTIN_OP_END ||
4485 Opc == ISD::INTRINSIC_WO_CHAIN ||
4486 Opc == ISD::INTRINSIC_W_CHAIN ||
4487 Opc == ISD::INTRINSIC_VOID) &&
4488 "Should use MaskedValueIsZero if you don't know whether Op"
4489 " is a target node!");
4491 KnownZero = KnownOne = 0; // Don't know anything.
4495 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4500 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4501 /// element of the result of the vector shuffle.
4502 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4503 MVT::ValueType VT = N->getValueType(0);
4504 SDOperand PermMask = N->getOperand(2);
4505 unsigned NumElems = PermMask.getNumOperands();
4506 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4508 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4510 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4511 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4512 SDOperand Idx = PermMask.getOperand(i);
4513 if (Idx.getOpcode() == ISD::UNDEF)
4514 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4515 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4520 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4521 /// node is a GlobalAddress + an offset.
4522 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4523 unsigned Opc = N->getOpcode();
4524 if (Opc == X86ISD::Wrapper) {
4525 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4526 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4529 } else if (Opc == ISD::ADD) {
4530 SDOperand N1 = N->getOperand(0);
4531 SDOperand N2 = N->getOperand(1);
4532 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4533 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4535 Offset += V->getSignExtended();
4538 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4539 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4541 Offset += V->getSignExtended();
4549 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4551 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4552 MachineFrameInfo *MFI) {
4553 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4556 SDOperand Loc = N->getOperand(1);
4557 SDOperand BaseLoc = Base->getOperand(1);
4558 if (Loc.getOpcode() == ISD::FrameIndex) {
4559 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4561 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4562 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4563 int FS = MFI->getObjectSize(FI);
4564 int BFS = MFI->getObjectSize(BFI);
4565 if (FS != BFS || FS != Size) return false;
4566 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4568 GlobalValue *GV1 = NULL;
4569 GlobalValue *GV2 = NULL;
4570 int64_t Offset1 = 0;
4571 int64_t Offset2 = 0;
4572 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4573 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4574 if (isGA1 && isGA2 && GV1 == GV2)
4575 return Offset1 == (Offset2 + Dist*Size);
4581 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4582 const X86Subtarget *Subtarget) {
4585 if (isGAPlusOffset(Base, GV, Offset))
4586 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4588 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4589 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4591 // Fixed objects do not specify alignment, however the offsets are known.
4592 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4593 (MFI->getObjectOffset(BFI) % 16) == 0);
4595 return MFI->getObjectAlignment(BFI) >= 16;
4601 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4602 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4603 /// if the load addresses are consecutive, non-overlapping, and in the right
4605 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4606 const X86Subtarget *Subtarget) {
4607 MachineFunction &MF = DAG.getMachineFunction();
4608 MachineFrameInfo *MFI = MF.getFrameInfo();
4609 MVT::ValueType VT = N->getValueType(0);
4610 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4611 SDOperand PermMask = N->getOperand(2);
4612 int NumElems = (int)PermMask.getNumOperands();
4613 SDNode *Base = NULL;
4614 for (int i = 0; i < NumElems; ++i) {
4615 SDOperand Idx = PermMask.getOperand(i);
4616 if (Idx.getOpcode() == ISD::UNDEF) {
4617 if (!Base) return SDOperand();
4620 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4621 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4625 else if (!isConsecutiveLoad(Arg.Val, Base,
4626 i, MVT::getSizeInBits(EVT)/8,MFI))
4631 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4633 LoadSDNode *LD = cast<LoadSDNode>(Base);
4634 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4635 LD->getSrcValueOffset());
4637 // Just use movups, it's shorter.
4638 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
4639 SmallVector<SDOperand, 3> Ops;
4640 Ops.push_back(Base->getOperand(0));
4641 Ops.push_back(Base->getOperand(1));
4642 Ops.push_back(Base->getOperand(2));
4643 return DAG.getNode(ISD::BIT_CONVERT, VT,
4644 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
4648 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4649 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4650 const X86Subtarget *Subtarget) {
4651 SDOperand Cond = N->getOperand(0);
4653 // If we have SSE[12] support, try to form min/max nodes.
4654 if (Subtarget->hasSSE2() &&
4655 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4656 if (Cond.getOpcode() == ISD::SETCC) {
4657 // Get the LHS/RHS of the select.
4658 SDOperand LHS = N->getOperand(1);
4659 SDOperand RHS = N->getOperand(2);
4660 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4662 unsigned Opcode = 0;
4663 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4666 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4669 if (!UnsafeFPMath) break;
4671 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4673 Opcode = X86ISD::FMIN;
4676 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4679 if (!UnsafeFPMath) break;
4681 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4683 Opcode = X86ISD::FMAX;
4686 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4689 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4692 if (!UnsafeFPMath) break;
4694 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4696 Opcode = X86ISD::FMIN;
4699 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4702 if (!UnsafeFPMath) break;
4704 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4706 Opcode = X86ISD::FMAX;
4712 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4721 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4722 DAGCombinerInfo &DCI) const {
4723 SelectionDAG &DAG = DCI.DAG;
4724 switch (N->getOpcode()) {
4726 case ISD::VECTOR_SHUFFLE:
4727 return PerformShuffleCombine(N, DAG, Subtarget);
4729 return PerformSELECTCombine(N, DAG, Subtarget);
4735 //===----------------------------------------------------------------------===//
4736 // X86 Inline Assembly Support
4737 //===----------------------------------------------------------------------===//
4739 /// getConstraintType - Given a constraint letter, return the type of
4740 /// constraint it is for this target.
4741 X86TargetLowering::ConstraintType
4742 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
4743 if (Constraint.size() == 1) {
4744 switch (Constraint[0]) {
4753 return C_RegisterClass;
4758 return TargetLowering::getConstraintType(Constraint);
4761 /// isOperandValidForConstraint - Return the specified operand (possibly
4762 /// modified) if the specified SDOperand is valid for the specified target
4763 /// constraint letter, otherwise return null.
4764 SDOperand X86TargetLowering::
4765 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
4766 switch (Constraint) {
4769 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4770 if (C->getValue() <= 31)
4771 return DAG.getTargetConstant(C->getValue(), Op.getValueType());
4773 return SDOperand(0,0);
4775 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4776 if (C->getValue() <= 255)
4777 return DAG.getTargetConstant(C->getValue(), Op.getValueType());
4779 return SDOperand(0,0);
4781 // Literal immediates are always ok.
4782 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op))
4783 return DAG.getTargetConstant(CST->getValue(), Op.getValueType());
4785 // If we are in non-pic codegen mode, we allow the address of a global (with
4786 // an optional displacement) to be used with 'i'.
4787 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
4790 // Match either (GA) or (GA+C)
4792 Offset = GA->getOffset();
4793 } else if (Op.getOpcode() == ISD::ADD) {
4794 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
4795 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
4797 Offset = GA->getOffset()+C->getValue();
4799 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
4800 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
4802 Offset = GA->getOffset()+C->getValue();
4809 // If addressing this global requires a load (e.g. in PIC mode), we can't
4811 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
4813 return SDOperand(0, 0);
4815 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4820 // Otherwise, not valid for this mode.
4821 return SDOperand(0, 0);
4824 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
4827 std::vector<unsigned> X86TargetLowering::
4828 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4829 MVT::ValueType VT) const {
4830 if (Constraint.size() == 1) {
4831 // FIXME: not handling fp-stack yet!
4832 switch (Constraint[0]) { // GCC X86 Constraint Letters
4833 default: break; // Unknown constraint letter
4834 case 'A': // EAX/EDX
4835 if (VT == MVT::i32 || VT == MVT::i64)
4836 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
4838 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
4841 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
4842 else if (VT == MVT::i16)
4843 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
4844 else if (VT == MVT::i8)
4845 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4850 return std::vector<unsigned>();
4853 std::pair<unsigned, const TargetRegisterClass*>
4854 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4855 MVT::ValueType VT) const {
4856 // First, see if this is a constraint that directly corresponds to an LLVM
4858 if (Constraint.size() == 1) {
4859 // GCC Constraint Letters
4860 switch (Constraint[0]) {
4862 case 'r': // GENERAL_REGS
4863 case 'R': // LEGACY_REGS
4864 case 'l': // INDEX_REGS
4865 if (VT == MVT::i64 && Subtarget->is64Bit())
4866 return std::make_pair(0U, X86::GR64RegisterClass);
4868 return std::make_pair(0U, X86::GR32RegisterClass);
4869 else if (VT == MVT::i16)
4870 return std::make_pair(0U, X86::GR16RegisterClass);
4871 else if (VT == MVT::i8)
4872 return std::make_pair(0U, X86::GR8RegisterClass);
4874 case 'y': // MMX_REGS if MMX allowed.
4875 if (!Subtarget->hasMMX()) break;
4876 return std::make_pair(0U, X86::VR64RegisterClass);
4878 case 'Y': // SSE_REGS if SSE2 allowed
4879 if (!Subtarget->hasSSE2()) break;
4881 case 'x': // SSE_REGS if SSE1 allowed
4882 if (!Subtarget->hasSSE1()) break;
4886 // Scalar SSE types.
4889 return std::make_pair(0U, X86::FR32RegisterClass);
4892 return std::make_pair(0U, X86::FR64RegisterClass);
4901 return std::make_pair(0U, X86::VR128RegisterClass);
4907 // Use the default implementation in TargetLowering to convert the register
4908 // constraint into a member of a register class.
4909 std::pair<unsigned, const TargetRegisterClass*> Res;
4910 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4912 // Not found as a standard register?
4913 if (Res.second == 0) {
4914 // GCC calls "st(0)" just plain "st".
4915 if (StringsEqualNoCase("{st}", Constraint)) {
4916 Res.first = X86::ST0;
4917 Res.second = X86::RSTRegisterClass;
4923 // Otherwise, check to see if this is a register class of the wrong value
4924 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
4925 // turn into {ax},{dx}.
4926 if (Res.second->hasType(VT))
4927 return Res; // Correct type already, nothing to do.
4929 // All of the single-register GCC register classes map their values onto
4930 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
4931 // really want an 8-bit or 32-bit register, map to the appropriate register
4932 // class and return the appropriate register.
4933 if (Res.second != X86::GR16RegisterClass)
4936 if (VT == MVT::i8) {
4937 unsigned DestReg = 0;
4938 switch (Res.first) {
4940 case X86::AX: DestReg = X86::AL; break;
4941 case X86::DX: DestReg = X86::DL; break;
4942 case X86::CX: DestReg = X86::CL; break;
4943 case X86::BX: DestReg = X86::BL; break;
4946 Res.first = DestReg;
4947 Res.second = Res.second = X86::GR8RegisterClass;
4949 } else if (VT == MVT::i32) {
4950 unsigned DestReg = 0;
4951 switch (Res.first) {
4953 case X86::AX: DestReg = X86::EAX; break;
4954 case X86::DX: DestReg = X86::EDX; break;
4955 case X86::CX: DestReg = X86::ECX; break;
4956 case X86::BX: DestReg = X86::EBX; break;
4957 case X86::SI: DestReg = X86::ESI; break;
4958 case X86::DI: DestReg = X86::EDI; break;
4959 case X86::BP: DestReg = X86::EBP; break;
4960 case X86::SP: DestReg = X86::ESP; break;
4963 Res.first = DestReg;
4964 Res.second = Res.second = X86::GR32RegisterClass;
4966 } else if (VT == MVT::i64) {
4967 unsigned DestReg = 0;
4968 switch (Res.first) {
4970 case X86::AX: DestReg = X86::RAX; break;
4971 case X86::DX: DestReg = X86::RDX; break;
4972 case X86::CX: DestReg = X86::RCX; break;
4973 case X86::BX: DestReg = X86::RBX; break;
4974 case X86::SI: DestReg = X86::RSI; break;
4975 case X86::DI: DestReg = X86::RDI; break;
4976 case X86::BP: DestReg = X86::RBP; break;
4977 case X86::SP: DestReg = X86::RSP; break;
4980 Res.first = DestReg;
4981 Res.second = Res.second = X86::GR64RegisterClass;