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::Vector + 1;
313 VT != (unsigned)MVT::LAST_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 RC = X86::VR128RegisterClass;
1189 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1190 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1192 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1193 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1195 if (VA.getLocInfo() == CCValAssign::SExt)
1196 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1197 DAG.getValueType(VA.getValVT()));
1198 else if (VA.getLocInfo() == CCValAssign::ZExt)
1199 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1200 DAG.getValueType(VA.getValVT()));
1202 if (VA.getLocInfo() != CCValAssign::Full)
1203 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1205 ArgValues.push_back(ArgValue);
1207 assert(VA.isMemLoc());
1209 // Create the nodes corresponding to a load from this parameter slot.
1210 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1211 VA.getLocMemOffset());
1212 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1213 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
1217 unsigned StackSize = CCInfo.getNextStackOffset();
1219 // If the function takes variable number of arguments, make a frame index for
1220 // the start of the first vararg value... for expansion of llvm.va_start.
1222 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1223 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1225 // For X86-64, if there are vararg parameters that are passed via
1226 // registers, then we must store them to their spots on the stack so they
1227 // may be loaded by deferencing the result of va_next.
1228 VarArgsGPOffset = NumIntRegs * 8;
1229 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1230 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1231 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1233 // Store the integer parameter registers.
1234 SmallVector<SDOperand, 8> MemOps;
1235 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1236 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1237 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1238 for (; NumIntRegs != 6; ++NumIntRegs) {
1239 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1240 X86::GR64RegisterClass);
1241 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1242 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1243 MemOps.push_back(Store);
1244 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1245 DAG.getConstant(8, getPointerTy()));
1248 // Now store the XMM (fp + vector) parameter registers.
1249 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1250 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1251 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1252 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1253 X86::VR128RegisterClass);
1254 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1255 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1256 MemOps.push_back(Store);
1257 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1258 DAG.getConstant(16, getPointerTy()));
1260 if (!MemOps.empty())
1261 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1262 &MemOps[0], MemOps.size());
1265 ArgValues.push_back(Root);
1267 ReturnAddrIndex = 0; // No return address slot generated yet.
1268 BytesToPopOnReturn = 0; // Callee pops nothing.
1269 BytesCallerReserves = StackSize;
1271 // Return the new list of results.
1272 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1273 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1277 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1279 SDOperand Chain = Op.getOperand(0);
1280 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1281 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1282 SDOperand Callee = Op.getOperand(4);
1284 // Analyze operands of the call, assigning locations to each operand.
1285 SmallVector<CCValAssign, 16> ArgLocs;
1286 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
1287 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1289 // Get a count of how many bytes are to be pushed on the stack.
1290 unsigned NumBytes = CCInfo.getNextStackOffset();
1291 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1293 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1294 SmallVector<SDOperand, 8> MemOpChains;
1298 // Walk the register/memloc assignments, inserting copies/loads.
1299 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1300 CCValAssign &VA = ArgLocs[i];
1301 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1303 // Promote the value if needed.
1304 switch (VA.getLocInfo()) {
1305 default: assert(0 && "Unknown loc info!");
1306 case CCValAssign::Full: break;
1307 case CCValAssign::SExt:
1308 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1310 case CCValAssign::ZExt:
1311 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1313 case CCValAssign::AExt:
1314 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1318 if (VA.isRegLoc()) {
1319 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1321 assert(VA.isMemLoc());
1322 if (StackPtr.Val == 0)
1323 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1324 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1325 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1326 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1330 if (!MemOpChains.empty())
1331 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1332 &MemOpChains[0], MemOpChains.size());
1334 // Build a sequence of copy-to-reg nodes chained together with token chain
1335 // and flag operands which copy the outgoing args into registers.
1337 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1338 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1340 InFlag = Chain.getValue(1);
1344 // From AMD64 ABI document:
1345 // For calls that may call functions that use varargs or stdargs
1346 // (prototype-less calls or calls to functions containing ellipsis (...) in
1347 // the declaration) %al is used as hidden argument to specify the number
1348 // of SSE registers used. The contents of %al do not need to match exactly
1349 // the number of registers, but must be an ubound on the number of SSE
1350 // registers used and is in the range 0 - 8 inclusive.
1352 // Count the number of XMM registers allocated.
1353 static const unsigned XMMArgRegs[] = {
1354 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1355 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1357 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1359 Chain = DAG.getCopyToReg(Chain, X86::AL,
1360 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1361 InFlag = Chain.getValue(1);
1364 // If the callee is a GlobalAddress node (quite common, every direct call is)
1365 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1366 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1367 // We should use extra load for direct calls to dllimported functions in
1369 if (getTargetMachine().getCodeModel() != CodeModel::Large
1370 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1371 getTargetMachine(), true))
1372 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1373 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1374 if (getTargetMachine().getCodeModel() != CodeModel::Large)
1375 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1377 // Returns a chain & a flag for retval copy to use.
1378 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1379 SmallVector<SDOperand, 8> Ops;
1380 Ops.push_back(Chain);
1381 Ops.push_back(Callee);
1383 // Add argument registers to the end of the list so that they are known live
1385 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1386 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1387 RegsToPass[i].second.getValueType()));
1390 Ops.push_back(InFlag);
1392 // FIXME: Do not generate X86ISD::TAILCALL for now.
1393 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1394 NodeTys, &Ops[0], Ops.size());
1395 InFlag = Chain.getValue(1);
1397 // Returns a flag for retval copy to use.
1398 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1400 Ops.push_back(Chain);
1401 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1402 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1403 Ops.push_back(InFlag);
1404 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1405 InFlag = Chain.getValue(1);
1407 // Handle result values, copying them out of physregs into vregs that we
1409 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1413 //===----------------------------------------------------------------------===//
1414 // Other Lowering Hooks
1415 //===----------------------------------------------------------------------===//
1418 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1419 if (ReturnAddrIndex == 0) {
1420 // Set up a frame object for the return address.
1421 MachineFunction &MF = DAG.getMachineFunction();
1422 if (Subtarget->is64Bit())
1423 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1425 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1428 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1433 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1434 /// specific condition code. It returns a false if it cannot do a direct
1435 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1437 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1438 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1439 SelectionDAG &DAG) {
1440 X86CC = X86::COND_INVALID;
1442 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1443 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1444 // X > -1 -> X == 0, jump !sign.
1445 RHS = DAG.getConstant(0, RHS.getValueType());
1446 X86CC = X86::COND_NS;
1448 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1449 // X < 0 -> X == 0, jump on sign.
1450 X86CC = X86::COND_S;
1455 switch (SetCCOpcode) {
1457 case ISD::SETEQ: X86CC = X86::COND_E; break;
1458 case ISD::SETGT: X86CC = X86::COND_G; break;
1459 case ISD::SETGE: X86CC = X86::COND_GE; break;
1460 case ISD::SETLT: X86CC = X86::COND_L; break;
1461 case ISD::SETLE: X86CC = X86::COND_LE; break;
1462 case ISD::SETNE: X86CC = X86::COND_NE; break;
1463 case ISD::SETULT: X86CC = X86::COND_B; break;
1464 case ISD::SETUGT: X86CC = X86::COND_A; break;
1465 case ISD::SETULE: X86CC = X86::COND_BE; break;
1466 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1469 // On a floating point condition, the flags are set as follows:
1471 // 0 | 0 | 0 | X > Y
1472 // 0 | 0 | 1 | X < Y
1473 // 1 | 0 | 0 | X == Y
1474 // 1 | 1 | 1 | unordered
1476 switch (SetCCOpcode) {
1479 case ISD::SETEQ: X86CC = X86::COND_E; break;
1480 case ISD::SETOLT: Flip = true; // Fallthrough
1482 case ISD::SETGT: X86CC = X86::COND_A; break;
1483 case ISD::SETOLE: Flip = true; // Fallthrough
1485 case ISD::SETGE: X86CC = X86::COND_AE; break;
1486 case ISD::SETUGT: Flip = true; // Fallthrough
1488 case ISD::SETLT: X86CC = X86::COND_B; break;
1489 case ISD::SETUGE: Flip = true; // Fallthrough
1491 case ISD::SETLE: X86CC = X86::COND_BE; break;
1493 case ISD::SETNE: X86CC = X86::COND_NE; break;
1494 case ISD::SETUO: X86CC = X86::COND_P; break;
1495 case ISD::SETO: X86CC = X86::COND_NP; break;
1498 std::swap(LHS, RHS);
1501 return X86CC != X86::COND_INVALID;
1504 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1505 /// code. Current x86 isa includes the following FP cmov instructions:
1506 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1507 static bool hasFPCMov(unsigned X86CC) {
1523 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1524 /// true if Op is undef or if its value falls within the specified range (L, H].
1525 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1526 if (Op.getOpcode() == ISD::UNDEF)
1529 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1530 return (Val >= Low && Val < Hi);
1533 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1534 /// true if Op is undef or if its value equal to the specified value.
1535 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1536 if (Op.getOpcode() == ISD::UNDEF)
1538 return cast<ConstantSDNode>(Op)->getValue() == Val;
1541 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1542 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1543 bool X86::isPSHUFDMask(SDNode *N) {
1544 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1546 if (N->getNumOperands() != 4)
1549 // Check if the value doesn't reference the second vector.
1550 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1551 SDOperand Arg = N->getOperand(i);
1552 if (Arg.getOpcode() == ISD::UNDEF) continue;
1553 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1554 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
1561 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1562 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1563 bool X86::isPSHUFHWMask(SDNode *N) {
1564 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1566 if (N->getNumOperands() != 8)
1569 // Lower quadword copied in order.
1570 for (unsigned i = 0; i != 4; ++i) {
1571 SDOperand Arg = N->getOperand(i);
1572 if (Arg.getOpcode() == ISD::UNDEF) continue;
1573 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1574 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1578 // Upper quadword shuffled.
1579 for (unsigned i = 4; i != 8; ++i) {
1580 SDOperand Arg = N->getOperand(i);
1581 if (Arg.getOpcode() == ISD::UNDEF) continue;
1582 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1583 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1584 if (Val < 4 || Val > 7)
1591 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1592 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1593 bool X86::isPSHUFLWMask(SDNode *N) {
1594 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1596 if (N->getNumOperands() != 8)
1599 // Upper quadword copied in order.
1600 for (unsigned i = 4; i != 8; ++i)
1601 if (!isUndefOrEqual(N->getOperand(i), i))
1604 // Lower quadword shuffled.
1605 for (unsigned i = 0; i != 4; ++i)
1606 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1612 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1613 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1614 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
1615 if (NumElems != 2 && NumElems != 4) return false;
1617 unsigned Half = NumElems / 2;
1618 for (unsigned i = 0; i < Half; ++i)
1619 if (!isUndefOrInRange(Elems[i], 0, NumElems))
1621 for (unsigned i = Half; i < NumElems; ++i)
1622 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
1628 bool X86::isSHUFPMask(SDNode *N) {
1629 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1630 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
1633 /// isCommutedSHUFP - Returns true if the shuffle mask is except
1634 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
1635 /// half elements to come from vector 1 (which would equal the dest.) and
1636 /// the upper half to come from vector 2.
1637 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
1638 if (NumOps != 2 && NumOps != 4) return false;
1640 unsigned Half = NumOps / 2;
1641 for (unsigned i = 0; i < Half; ++i)
1642 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
1644 for (unsigned i = Half; i < NumOps; ++i)
1645 if (!isUndefOrInRange(Ops[i], 0, NumOps))
1650 static bool isCommutedSHUFP(SDNode *N) {
1651 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1652 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
1655 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
1656 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
1657 bool X86::isMOVHLPSMask(SDNode *N) {
1658 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1660 if (N->getNumOperands() != 4)
1663 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
1664 return isUndefOrEqual(N->getOperand(0), 6) &&
1665 isUndefOrEqual(N->getOperand(1), 7) &&
1666 isUndefOrEqual(N->getOperand(2), 2) &&
1667 isUndefOrEqual(N->getOperand(3), 3);
1670 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
1671 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
1673 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
1674 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1676 if (N->getNumOperands() != 4)
1679 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
1680 return isUndefOrEqual(N->getOperand(0), 2) &&
1681 isUndefOrEqual(N->getOperand(1), 3) &&
1682 isUndefOrEqual(N->getOperand(2), 2) &&
1683 isUndefOrEqual(N->getOperand(3), 3);
1686 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
1687 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
1688 bool X86::isMOVLPMask(SDNode *N) {
1689 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1691 unsigned NumElems = N->getNumOperands();
1692 if (NumElems != 2 && NumElems != 4)
1695 for (unsigned i = 0; i < NumElems/2; ++i)
1696 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
1699 for (unsigned i = NumElems/2; i < NumElems; ++i)
1700 if (!isUndefOrEqual(N->getOperand(i), i))
1706 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
1707 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
1709 bool X86::isMOVHPMask(SDNode *N) {
1710 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1712 unsigned NumElems = N->getNumOperands();
1713 if (NumElems != 2 && NumElems != 4)
1716 for (unsigned i = 0; i < NumElems/2; ++i)
1717 if (!isUndefOrEqual(N->getOperand(i), i))
1720 for (unsigned i = 0; i < NumElems/2; ++i) {
1721 SDOperand Arg = N->getOperand(i + NumElems/2);
1722 if (!isUndefOrEqual(Arg, i + NumElems))
1729 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
1730 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
1731 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
1732 bool V2IsSplat = false) {
1733 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1736 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1737 SDOperand BitI = Elts[i];
1738 SDOperand BitI1 = Elts[i+1];
1739 if (!isUndefOrEqual(BitI, j))
1742 if (isUndefOrEqual(BitI1, NumElts))
1745 if (!isUndefOrEqual(BitI1, j + NumElts))
1753 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
1754 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1755 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1758 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
1759 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
1760 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
1761 bool V2IsSplat = false) {
1762 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1765 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1766 SDOperand BitI = Elts[i];
1767 SDOperand BitI1 = Elts[i+1];
1768 if (!isUndefOrEqual(BitI, j + NumElts/2))
1771 if (isUndefOrEqual(BitI1, NumElts))
1774 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
1782 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
1783 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1784 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1787 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
1788 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
1790 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
1791 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1793 unsigned NumElems = N->getNumOperands();
1794 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1797 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1798 SDOperand BitI = N->getOperand(i);
1799 SDOperand BitI1 = N->getOperand(i+1);
1801 if (!isUndefOrEqual(BitI, j))
1803 if (!isUndefOrEqual(BitI1, j))
1810 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
1811 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
1813 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
1814 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1816 unsigned NumElems = N->getNumOperands();
1817 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1820 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
1821 SDOperand BitI = N->getOperand(i);
1822 SDOperand BitI1 = N->getOperand(i + 1);
1824 if (!isUndefOrEqual(BitI, j))
1826 if (!isUndefOrEqual(BitI1, j))
1833 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
1834 /// specifies a shuffle of elements that is suitable for input to MOVSS,
1835 /// MOVSD, and MOVD, i.e. setting the lowest element.
1836 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
1837 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1840 if (!isUndefOrEqual(Elts[0], NumElts))
1843 for (unsigned i = 1; i < NumElts; ++i) {
1844 if (!isUndefOrEqual(Elts[i], i))
1851 bool X86::isMOVLMask(SDNode *N) {
1852 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1853 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
1856 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
1857 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
1858 /// element of vector 2 and the other elements to come from vector 1 in order.
1859 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
1860 bool V2IsSplat = false,
1861 bool V2IsUndef = false) {
1862 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
1865 if (!isUndefOrEqual(Ops[0], 0))
1868 for (unsigned i = 1; i < NumOps; ++i) {
1869 SDOperand Arg = Ops[i];
1870 if (!(isUndefOrEqual(Arg, i+NumOps) ||
1871 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
1872 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
1879 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
1880 bool V2IsUndef = false) {
1881 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1882 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
1883 V2IsSplat, V2IsUndef);
1886 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1887 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
1888 bool X86::isMOVSHDUPMask(SDNode *N) {
1889 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1891 if (N->getNumOperands() != 4)
1894 // Expect 1, 1, 3, 3
1895 for (unsigned i = 0; i < 2; ++i) {
1896 SDOperand Arg = N->getOperand(i);
1897 if (Arg.getOpcode() == ISD::UNDEF) continue;
1898 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1899 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1900 if (Val != 1) return false;
1904 for (unsigned i = 2; i < 4; ++i) {
1905 SDOperand Arg = N->getOperand(i);
1906 if (Arg.getOpcode() == ISD::UNDEF) continue;
1907 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1908 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1909 if (Val != 3) return false;
1913 // Don't use movshdup if it can be done with a shufps.
1917 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1918 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
1919 bool X86::isMOVSLDUPMask(SDNode *N) {
1920 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1922 if (N->getNumOperands() != 4)
1925 // Expect 0, 0, 2, 2
1926 for (unsigned i = 0; i < 2; ++i) {
1927 SDOperand Arg = N->getOperand(i);
1928 if (Arg.getOpcode() == ISD::UNDEF) continue;
1929 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1930 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1931 if (Val != 0) return false;
1935 for (unsigned i = 2; i < 4; ++i) {
1936 SDOperand Arg = N->getOperand(i);
1937 if (Arg.getOpcode() == ISD::UNDEF) continue;
1938 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1939 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1940 if (Val != 2) return false;
1944 // Don't use movshdup if it can be done with a shufps.
1948 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1949 /// a splat of a single element.
1950 static bool isSplatMask(SDNode *N) {
1951 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1953 // This is a splat operation if each element of the permute is the same, and
1954 // if the value doesn't reference the second vector.
1955 unsigned NumElems = N->getNumOperands();
1956 SDOperand ElementBase;
1958 for (; i != NumElems; ++i) {
1959 SDOperand Elt = N->getOperand(i);
1960 if (isa<ConstantSDNode>(Elt)) {
1966 if (!ElementBase.Val)
1969 for (; i != NumElems; ++i) {
1970 SDOperand Arg = N->getOperand(i);
1971 if (Arg.getOpcode() == ISD::UNDEF) continue;
1972 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1973 if (Arg != ElementBase) return false;
1976 // Make sure it is a splat of the first vector operand.
1977 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
1980 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1981 /// a splat of a single element and it's a 2 or 4 element mask.
1982 bool X86::isSplatMask(SDNode *N) {
1983 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1985 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
1986 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
1988 return ::isSplatMask(N);
1991 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
1992 /// specifies a splat of zero element.
1993 bool X86::isSplatLoMask(SDNode *N) {
1994 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1996 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
1997 if (!isUndefOrEqual(N->getOperand(i), 0))
2002 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2003 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2005 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2006 unsigned NumOperands = N->getNumOperands();
2007 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2009 for (unsigned i = 0; i < NumOperands; ++i) {
2011 SDOperand Arg = N->getOperand(NumOperands-i-1);
2012 if (Arg.getOpcode() != ISD::UNDEF)
2013 Val = cast<ConstantSDNode>(Arg)->getValue();
2014 if (Val >= NumOperands) Val -= NumOperands;
2016 if (i != NumOperands - 1)
2023 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2024 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2026 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2028 // 8 nodes, but we only care about the last 4.
2029 for (unsigned i = 7; i >= 4; --i) {
2031 SDOperand Arg = N->getOperand(i);
2032 if (Arg.getOpcode() != ISD::UNDEF)
2033 Val = cast<ConstantSDNode>(Arg)->getValue();
2042 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2043 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2045 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2047 // 8 nodes, but we only care about the first 4.
2048 for (int i = 3; i >= 0; --i) {
2050 SDOperand Arg = N->getOperand(i);
2051 if (Arg.getOpcode() != ISD::UNDEF)
2052 Val = cast<ConstantSDNode>(Arg)->getValue();
2061 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2062 /// specifies a 8 element shuffle that can be broken into a pair of
2063 /// PSHUFHW and PSHUFLW.
2064 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2065 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2067 if (N->getNumOperands() != 8)
2070 // Lower quadword shuffled.
2071 for (unsigned i = 0; i != 4; ++i) {
2072 SDOperand Arg = N->getOperand(i);
2073 if (Arg.getOpcode() == ISD::UNDEF) continue;
2074 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2075 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2080 // Upper quadword shuffled.
2081 for (unsigned i = 4; i != 8; ++i) {
2082 SDOperand Arg = N->getOperand(i);
2083 if (Arg.getOpcode() == ISD::UNDEF) continue;
2084 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2085 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2086 if (Val < 4 || Val > 7)
2093 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2094 /// values in ther permute mask.
2095 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2096 SDOperand &V2, SDOperand &Mask,
2097 SelectionDAG &DAG) {
2098 MVT::ValueType VT = Op.getValueType();
2099 MVT::ValueType MaskVT = Mask.getValueType();
2100 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2101 unsigned NumElems = Mask.getNumOperands();
2102 SmallVector<SDOperand, 8> MaskVec;
2104 for (unsigned i = 0; i != NumElems; ++i) {
2105 SDOperand Arg = Mask.getOperand(i);
2106 if (Arg.getOpcode() == ISD::UNDEF) {
2107 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2110 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2111 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2113 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2115 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2119 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2120 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2123 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2124 /// match movhlps. The lower half elements should come from upper half of
2125 /// V1 (and in order), and the upper half elements should come from the upper
2126 /// half of V2 (and in order).
2127 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2128 unsigned NumElems = Mask->getNumOperands();
2131 for (unsigned i = 0, e = 2; i != e; ++i)
2132 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2134 for (unsigned i = 2; i != 4; ++i)
2135 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2140 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2141 /// is promoted to a vector.
2142 static inline bool isScalarLoadToVector(SDNode *N) {
2143 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2144 N = N->getOperand(0).Val;
2145 return ISD::isNON_EXTLoad(N);
2150 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2151 /// match movlp{s|d}. The lower half elements should come from lower half of
2152 /// V1 (and in order), and the upper half elements should come from the upper
2153 /// half of V2 (and in order). And since V1 will become the source of the
2154 /// MOVLP, it must be either a vector load or a scalar load to vector.
2155 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2156 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2158 // Is V2 is a vector load, don't do this transformation. We will try to use
2159 // load folding shufps op.
2160 if (ISD::isNON_EXTLoad(V2))
2163 unsigned NumElems = Mask->getNumOperands();
2164 if (NumElems != 2 && NumElems != 4)
2166 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2167 if (!isUndefOrEqual(Mask->getOperand(i), i))
2169 for (unsigned i = NumElems/2; i != NumElems; ++i)
2170 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2175 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2177 static bool isSplatVector(SDNode *N) {
2178 if (N->getOpcode() != ISD::BUILD_VECTOR)
2181 SDOperand SplatValue = N->getOperand(0);
2182 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2183 if (N->getOperand(i) != SplatValue)
2188 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2190 static bool isUndefShuffle(SDNode *N) {
2191 if (N->getOpcode() != ISD::BUILD_VECTOR)
2194 SDOperand V1 = N->getOperand(0);
2195 SDOperand V2 = N->getOperand(1);
2196 SDOperand Mask = N->getOperand(2);
2197 unsigned NumElems = Mask.getNumOperands();
2198 for (unsigned i = 0; i != NumElems; ++i) {
2199 SDOperand Arg = Mask.getOperand(i);
2200 if (Arg.getOpcode() != ISD::UNDEF) {
2201 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2202 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2204 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2211 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2212 /// that point to V2 points to its first element.
2213 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2214 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2216 bool Changed = false;
2217 SmallVector<SDOperand, 8> MaskVec;
2218 unsigned NumElems = Mask.getNumOperands();
2219 for (unsigned i = 0; i != NumElems; ++i) {
2220 SDOperand Arg = Mask.getOperand(i);
2221 if (Arg.getOpcode() != ISD::UNDEF) {
2222 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2223 if (Val > NumElems) {
2224 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2228 MaskVec.push_back(Arg);
2232 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2233 &MaskVec[0], MaskVec.size());
2237 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2238 /// operation of specified width.
2239 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2240 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2241 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2243 SmallVector<SDOperand, 8> MaskVec;
2244 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2245 for (unsigned i = 1; i != NumElems; ++i)
2246 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2247 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2250 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2251 /// of specified width.
2252 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2253 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2254 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2255 SmallVector<SDOperand, 8> MaskVec;
2256 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2257 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2258 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2260 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2263 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2264 /// of specified width.
2265 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2266 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2267 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2268 unsigned Half = NumElems/2;
2269 SmallVector<SDOperand, 8> MaskVec;
2270 for (unsigned i = 0; i != Half; ++i) {
2271 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2272 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2274 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2277 /// getZeroVector - Returns a vector of specified type with all zero elements.
2279 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2280 assert(MVT::isVector(VT) && "Expected a vector type");
2281 unsigned NumElems = getVectorNumElements(VT);
2282 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2283 bool isFP = MVT::isFloatingPoint(EVT);
2284 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2285 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2286 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2289 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2291 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2292 SDOperand V1 = Op.getOperand(0);
2293 SDOperand Mask = Op.getOperand(2);
2294 MVT::ValueType VT = Op.getValueType();
2295 unsigned NumElems = Mask.getNumOperands();
2296 Mask = getUnpacklMask(NumElems, DAG);
2297 while (NumElems != 4) {
2298 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2301 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2303 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2304 Mask = getZeroVector(MaskVT, DAG);
2305 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2306 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2307 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2310 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2312 static inline bool isZeroNode(SDOperand Elt) {
2313 return ((isa<ConstantSDNode>(Elt) &&
2314 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2315 (isa<ConstantFPSDNode>(Elt) &&
2316 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2319 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2320 /// vector and zero or undef vector.
2321 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2322 unsigned NumElems, unsigned Idx,
2323 bool isZero, SelectionDAG &DAG) {
2324 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2325 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2326 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2327 SDOperand Zero = DAG.getConstant(0, EVT);
2328 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2329 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2330 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2331 &MaskVec[0], MaskVec.size());
2332 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2335 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2337 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2338 unsigned NumNonZero, unsigned NumZero,
2339 SelectionDAG &DAG, TargetLowering &TLI) {
2345 for (unsigned i = 0; i < 16; ++i) {
2346 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2347 if (ThisIsNonZero && First) {
2349 V = getZeroVector(MVT::v8i16, DAG);
2351 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2356 SDOperand ThisElt(0, 0), LastElt(0, 0);
2357 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2358 if (LastIsNonZero) {
2359 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2361 if (ThisIsNonZero) {
2362 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2363 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2364 ThisElt, DAG.getConstant(8, MVT::i8));
2366 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2371 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2372 DAG.getConstant(i/2, TLI.getPointerTy()));
2376 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2379 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
2381 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2382 unsigned NumNonZero, unsigned NumZero,
2383 SelectionDAG &DAG, TargetLowering &TLI) {
2389 for (unsigned i = 0; i < 8; ++i) {
2390 bool isNonZero = (NonZeros & (1 << i)) != 0;
2394 V = getZeroVector(MVT::v8i16, DAG);
2396 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2399 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2400 DAG.getConstant(i, TLI.getPointerTy()));
2408 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2409 // All zero's are handled with pxor.
2410 if (ISD::isBuildVectorAllZeros(Op.Val))
2413 // All one's are handled with pcmpeqd.
2414 if (ISD::isBuildVectorAllOnes(Op.Val))
2417 MVT::ValueType VT = Op.getValueType();
2418 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2419 unsigned EVTBits = MVT::getSizeInBits(EVT);
2421 unsigned NumElems = Op.getNumOperands();
2422 unsigned NumZero = 0;
2423 unsigned NumNonZero = 0;
2424 unsigned NonZeros = 0;
2425 std::set<SDOperand> Values;
2426 for (unsigned i = 0; i < NumElems; ++i) {
2427 SDOperand Elt = Op.getOperand(i);
2428 if (Elt.getOpcode() != ISD::UNDEF) {
2430 if (isZeroNode(Elt))
2433 NonZeros |= (1 << i);
2439 if (NumNonZero == 0)
2440 // Must be a mix of zero and undef. Return a zero vector.
2441 return getZeroVector(VT, DAG);
2443 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2444 if (Values.size() == 1)
2447 // Special case for single non-zero element.
2448 if (NumNonZero == 1) {
2449 unsigned Idx = CountTrailingZeros_32(NonZeros);
2450 SDOperand Item = Op.getOperand(Idx);
2451 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2453 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2454 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2457 if (EVTBits == 32) {
2458 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2459 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2461 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2462 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2463 SmallVector<SDOperand, 8> MaskVec;
2464 for (unsigned i = 0; i < NumElems; i++)
2465 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2466 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2467 &MaskVec[0], MaskVec.size());
2468 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2469 DAG.getNode(ISD::UNDEF, VT), Mask);
2473 // Let legalizer expand 2-wide build_vectors.
2477 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2478 if (EVTBits == 8 && NumElems == 16) {
2479 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2481 if (V.Val) return V;
2484 if (EVTBits == 16 && NumElems == 8) {
2485 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2487 if (V.Val) return V;
2490 // If element VT is == 32 bits, turn it into a number of shuffles.
2491 SmallVector<SDOperand, 8> V;
2493 if (NumElems == 4 && NumZero > 0) {
2494 for (unsigned i = 0; i < 4; ++i) {
2495 bool isZero = !(NonZeros & (1 << i));
2497 V[i] = getZeroVector(VT, DAG);
2499 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2502 for (unsigned i = 0; i < 2; ++i) {
2503 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2506 V[i] = V[i*2]; // Must be a zero vector.
2509 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2510 getMOVLMask(NumElems, DAG));
2513 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2514 getMOVLMask(NumElems, DAG));
2517 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2518 getUnpacklMask(NumElems, DAG));
2523 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2524 // clears the upper bits.
2525 // FIXME: we can do the same for v4f32 case when we know both parts of
2526 // the lower half come from scalar_to_vector (loadf32). We should do
2527 // that in post legalizer dag combiner with target specific hooks.
2528 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2530 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2531 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2532 SmallVector<SDOperand, 8> MaskVec;
2533 bool Reverse = (NonZeros & 0x3) == 2;
2534 for (unsigned i = 0; i < 2; ++i)
2536 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2538 MaskVec.push_back(DAG.getConstant(i, EVT));
2539 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2540 for (unsigned i = 0; i < 2; ++i)
2542 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2544 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2545 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2546 &MaskVec[0], MaskVec.size());
2547 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2550 if (Values.size() > 2) {
2551 // Expand into a number of unpckl*.
2553 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2554 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2555 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2556 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2557 for (unsigned i = 0; i < NumElems; ++i)
2558 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2560 while (NumElems != 0) {
2561 for (unsigned i = 0; i < NumElems; ++i)
2562 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
2573 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2574 SDOperand V1 = Op.getOperand(0);
2575 SDOperand V2 = Op.getOperand(1);
2576 SDOperand PermMask = Op.getOperand(2);
2577 MVT::ValueType VT = Op.getValueType();
2578 unsigned NumElems = PermMask.getNumOperands();
2579 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
2580 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
2581 bool V1IsSplat = false;
2582 bool V2IsSplat = false;
2584 if (isUndefShuffle(Op.Val))
2585 return DAG.getNode(ISD::UNDEF, VT);
2587 if (isSplatMask(PermMask.Val)) {
2588 if (NumElems <= 4) return Op;
2589 // Promote it to a v4i32 splat.
2590 return PromoteSplat(Op, DAG);
2593 if (X86::isMOVLMask(PermMask.Val))
2594 return (V1IsUndef) ? V2 : Op;
2596 if (X86::isMOVSHDUPMask(PermMask.Val) ||
2597 X86::isMOVSLDUPMask(PermMask.Val) ||
2598 X86::isMOVHLPSMask(PermMask.Val) ||
2599 X86::isMOVHPMask(PermMask.Val) ||
2600 X86::isMOVLPMask(PermMask.Val))
2603 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2604 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
2605 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2607 bool Commuted = false;
2608 V1IsSplat = isSplatVector(V1.Val);
2609 V2IsSplat = isSplatVector(V2.Val);
2610 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
2611 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2612 std::swap(V1IsSplat, V2IsSplat);
2613 std::swap(V1IsUndef, V2IsUndef);
2617 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
2618 if (V2IsUndef) return V1;
2619 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2621 // V2 is a splat, so the mask may be malformed. That is, it may point
2622 // to any V2 element. The instruction selectior won't like this. Get
2623 // a corrected mask and commute to form a proper MOVS{S|D}.
2624 SDOperand NewMask = getMOVLMask(NumElems, DAG);
2625 if (NewMask.Val != PermMask.Val)
2626 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2631 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2632 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
2633 X86::isUNPCKLMask(PermMask.Val) ||
2634 X86::isUNPCKHMask(PermMask.Val))
2638 // Normalize mask so all entries that point to V2 points to its first
2639 // element then try to match unpck{h|l} again. If match, return a
2640 // new vector_shuffle with the corrected mask.
2641 SDOperand NewMask = NormalizeMask(PermMask, DAG);
2642 if (NewMask.Val != PermMask.Val) {
2643 if (X86::isUNPCKLMask(PermMask.Val, true)) {
2644 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
2645 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2646 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
2647 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
2648 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2653 // Normalize the node to match x86 shuffle ops if needed
2654 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
2655 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2658 // Commute is back and try unpck* again.
2659 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2660 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2661 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
2662 X86::isUNPCKLMask(PermMask.Val) ||
2663 X86::isUNPCKHMask(PermMask.Val))
2667 // If VT is integer, try PSHUF* first, then SHUFP*.
2668 if (MVT::isInteger(VT)) {
2669 if (X86::isPSHUFDMask(PermMask.Val) ||
2670 X86::isPSHUFHWMask(PermMask.Val) ||
2671 X86::isPSHUFLWMask(PermMask.Val)) {
2672 if (V2.getOpcode() != ISD::UNDEF)
2673 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2674 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2678 if (X86::isSHUFPMask(PermMask.Val) &&
2679 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
2682 // Handle v8i16 shuffle high / low shuffle node pair.
2683 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
2684 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2685 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2686 SmallVector<SDOperand, 8> MaskVec;
2687 for (unsigned i = 0; i != 4; ++i)
2688 MaskVec.push_back(PermMask.getOperand(i));
2689 for (unsigned i = 4; i != 8; ++i)
2690 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2691 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2692 &MaskVec[0], MaskVec.size());
2693 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2695 for (unsigned i = 0; i != 4; ++i)
2696 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2697 for (unsigned i = 4; i != 8; ++i)
2698 MaskVec.push_back(PermMask.getOperand(i));
2699 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
2700 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2703 // Floating point cases in the other order.
2704 if (X86::isSHUFPMask(PermMask.Val))
2706 if (X86::isPSHUFDMask(PermMask.Val) ||
2707 X86::isPSHUFHWMask(PermMask.Val) ||
2708 X86::isPSHUFLWMask(PermMask.Val)) {
2709 if (V2.getOpcode() != ISD::UNDEF)
2710 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2711 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2716 if (NumElems == 4 &&
2717 // Don't do this for MMX.
2718 MVT::getSizeInBits(VT) != 64) {
2719 MVT::ValueType MaskVT = PermMask.getValueType();
2720 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2721 SmallVector<std::pair<int, int>, 8> Locs;
2722 Locs.reserve(NumElems);
2723 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2724 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2727 // If no more than two elements come from either vector. This can be
2728 // implemented with two shuffles. First shuffle gather the elements.
2729 // The second shuffle, which takes the first shuffle as both of its
2730 // vector operands, put the elements into the right order.
2731 for (unsigned i = 0; i != NumElems; ++i) {
2732 SDOperand Elt = PermMask.getOperand(i);
2733 if (Elt.getOpcode() == ISD::UNDEF) {
2734 Locs[i] = std::make_pair(-1, -1);
2736 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
2737 if (Val < NumElems) {
2738 Locs[i] = std::make_pair(0, NumLo);
2742 Locs[i] = std::make_pair(1, NumHi);
2743 if (2+NumHi < NumElems)
2744 Mask1[2+NumHi] = Elt;
2749 if (NumLo <= 2 && NumHi <= 2) {
2750 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2751 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2752 &Mask1[0], Mask1.size()));
2753 for (unsigned i = 0; i != NumElems; ++i) {
2754 if (Locs[i].first == -1)
2757 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
2758 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
2759 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
2763 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
2764 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2765 &Mask2[0], Mask2.size()));
2768 // Break it into (shuffle shuffle_hi, shuffle_lo).
2770 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2771 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2772 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
2773 unsigned MaskIdx = 0;
2775 unsigned HiIdx = NumElems/2;
2776 for (unsigned i = 0; i != NumElems; ++i) {
2777 if (i == NumElems/2) {
2783 SDOperand Elt = PermMask.getOperand(i);
2784 if (Elt.getOpcode() == ISD::UNDEF) {
2785 Locs[i] = std::make_pair(-1, -1);
2786 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
2787 Locs[i] = std::make_pair(MaskIdx, LoIdx);
2788 (*MaskPtr)[LoIdx] = Elt;
2791 Locs[i] = std::make_pair(MaskIdx, HiIdx);
2792 (*MaskPtr)[HiIdx] = Elt;
2797 SDOperand LoShuffle =
2798 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2799 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2800 &LoMask[0], LoMask.size()));
2801 SDOperand HiShuffle =
2802 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2803 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2804 &HiMask[0], HiMask.size()));
2805 SmallVector<SDOperand, 8> MaskOps;
2806 for (unsigned i = 0; i != NumElems; ++i) {
2807 if (Locs[i].first == -1) {
2808 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
2810 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
2811 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
2814 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
2815 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2816 &MaskOps[0], MaskOps.size()));
2823 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2824 if (!isa<ConstantSDNode>(Op.getOperand(1)))
2827 MVT::ValueType VT = Op.getValueType();
2828 // TODO: handle v16i8.
2829 if (MVT::getSizeInBits(VT) == 16) {
2830 // Transform it so it match pextrw which produces a 32-bit result.
2831 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
2832 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
2833 Op.getOperand(0), Op.getOperand(1));
2834 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
2835 DAG.getValueType(VT));
2836 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
2837 } else if (MVT::getSizeInBits(VT) == 32) {
2838 SDOperand Vec = Op.getOperand(0);
2839 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2842 // SHUFPS the element to the lowest double word, then movss.
2843 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2844 SmallVector<SDOperand, 8> IdxVec;
2845 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
2846 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2847 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2848 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2849 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2850 &IdxVec[0], IdxVec.size());
2851 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2852 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2853 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2854 DAG.getConstant(0, getPointerTy()));
2855 } else if (MVT::getSizeInBits(VT) == 64) {
2856 SDOperand Vec = Op.getOperand(0);
2857 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2861 // UNPCKHPD the element to the lowest double word, then movsd.
2862 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
2863 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
2864 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2865 SmallVector<SDOperand, 8> IdxVec;
2866 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
2867 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2868 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2869 &IdxVec[0], IdxVec.size());
2870 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2871 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2872 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2873 DAG.getConstant(0, getPointerTy()));
2880 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2881 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
2882 // as its second argument.
2883 MVT::ValueType VT = Op.getValueType();
2884 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
2885 SDOperand N0 = Op.getOperand(0);
2886 SDOperand N1 = Op.getOperand(1);
2887 SDOperand N2 = Op.getOperand(2);
2888 if (MVT::getSizeInBits(BaseVT) == 16) {
2889 if (N1.getValueType() != MVT::i32)
2890 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
2891 if (N2.getValueType() != MVT::i32)
2892 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
2893 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
2894 } else if (MVT::getSizeInBits(BaseVT) == 32) {
2895 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
2898 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
2899 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2900 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2901 SmallVector<SDOperand, 8> MaskVec;
2902 MaskVec.push_back(DAG.getConstant(4, BaseVT));
2903 for (unsigned i = 1; i <= 3; ++i)
2904 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2905 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
2906 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2907 &MaskVec[0], MaskVec.size()));
2909 // Use two pinsrw instructions to insert a 32 bit value.
2911 if (MVT::isFloatingPoint(N1.getValueType())) {
2912 if (ISD::isNON_EXTLoad(N1.Val)) {
2913 // Just load directly from f32mem to GR32.
2914 LoadSDNode *LD = cast<LoadSDNode>(N1);
2915 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
2916 LD->getSrcValue(), LD->getSrcValueOffset());
2918 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
2919 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
2920 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
2921 DAG.getConstant(0, getPointerTy()));
2924 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
2925 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2926 DAG.getConstant(Idx, getPointerTy()));
2927 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
2928 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2929 DAG.getConstant(Idx+1, getPointerTy()));
2930 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
2938 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2939 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
2940 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
2943 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2944 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
2945 // one of the above mentioned nodes. It has to be wrapped because otherwise
2946 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2947 // be used to form addressing mode. These wrapped nodes will be selected
2950 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
2951 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2952 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
2954 CP->getAlignment());
2955 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2956 // With PIC, the address is actually $g + Offset.
2957 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2958 !Subtarget->isPICStyleRIPRel()) {
2959 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2960 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2968 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
2969 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2970 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
2971 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2972 // With PIC, the address is actually $g + Offset.
2973 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2974 !Subtarget->isPICStyleRIPRel()) {
2975 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2976 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2980 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
2981 // load the value at address GV, not the value of GV itself. This means that
2982 // the GlobalAddress must be in the base or index register of the address, not
2983 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
2984 // The same applies for external symbols during PIC codegen
2985 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
2986 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
2991 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2993 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
2994 const MVT::ValueType PtrVT) {
2996 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
2997 DAG.getNode(X86ISD::GlobalBaseReg,
2999 InFlag = Chain.getValue(1);
3001 // emit leal symbol@TLSGD(,%ebx,1), %eax
3002 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3003 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3004 GA->getValueType(0),
3006 SDOperand Ops[] = { Chain, TGA, InFlag };
3007 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3008 InFlag = Result.getValue(2);
3009 Chain = Result.getValue(1);
3011 // call ___tls_get_addr. This function receives its argument in
3012 // the register EAX.
3013 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
3014 InFlag = Chain.getValue(1);
3016 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3017 SDOperand Ops1[] = { Chain,
3018 DAG.getTargetExternalSymbol("___tls_get_addr",
3020 DAG.getRegister(X86::EAX, PtrVT),
3021 DAG.getRegister(X86::EBX, PtrVT),
3023 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
3024 InFlag = Chain.getValue(1);
3026 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
3029 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
3030 // "local exec" model.
3032 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3033 const MVT::ValueType PtrVT) {
3034 // Get the Thread Pointer
3035 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
3036 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
3038 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3039 GA->getValueType(0),
3041 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
3043 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
3044 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset, NULL, 0);
3046 // The address of the thread local variable is the add of the thread
3047 // pointer with the offset of the variable.
3048 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
3052 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
3053 // TODO: implement the "local dynamic" model
3054 // TODO: implement the "initial exec"model for pic executables
3055 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
3056 "TLS not implemented for non-ELF and 64-bit targets");
3057 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3058 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
3059 // otherwise use the "Local Exec"TLS Model
3060 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
3061 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
3063 return LowerToTLSExecModel(GA, DAG, getPointerTy());
3067 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3068 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3069 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3070 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3071 // With PIC, the address is actually $g + Offset.
3072 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3073 !Subtarget->isPICStyleRIPRel()) {
3074 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3075 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3082 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3083 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3084 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3085 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3086 // With PIC, the address is actually $g + Offset.
3087 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3088 !Subtarget->isPICStyleRIPRel()) {
3089 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3090 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3097 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3098 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3099 "Not an i64 shift!");
3100 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3101 SDOperand ShOpLo = Op.getOperand(0);
3102 SDOperand ShOpHi = Op.getOperand(1);
3103 SDOperand ShAmt = Op.getOperand(2);
3104 SDOperand Tmp1 = isSRA ?
3105 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3106 DAG.getConstant(0, MVT::i32);
3108 SDOperand Tmp2, Tmp3;
3109 if (Op.getOpcode() == ISD::SHL_PARTS) {
3110 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3111 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3113 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3114 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3117 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3118 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3119 DAG.getConstant(32, MVT::i8));
3120 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
3121 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
3124 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3126 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3127 SmallVector<SDOperand, 4> Ops;
3128 if (Op.getOpcode() == ISD::SHL_PARTS) {
3129 Ops.push_back(Tmp2);
3130 Ops.push_back(Tmp3);
3132 Ops.push_back(InFlag);
3133 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3134 InFlag = Hi.getValue(1);
3137 Ops.push_back(Tmp3);
3138 Ops.push_back(Tmp1);
3140 Ops.push_back(InFlag);
3141 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3143 Ops.push_back(Tmp2);
3144 Ops.push_back(Tmp3);
3146 Ops.push_back(InFlag);
3147 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3148 InFlag = Lo.getValue(1);
3151 Ops.push_back(Tmp3);
3152 Ops.push_back(Tmp1);
3154 Ops.push_back(InFlag);
3155 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3158 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3162 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3165 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3166 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3167 Op.getOperand(0).getValueType() >= MVT::i16 &&
3168 "Unknown SINT_TO_FP to lower!");
3171 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3172 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3173 MachineFunction &MF = DAG.getMachineFunction();
3174 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3175 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3176 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3177 StackSlot, NULL, 0);
3182 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3184 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3185 SmallVector<SDOperand, 8> Ops;
3186 Ops.push_back(Chain);
3187 Ops.push_back(StackSlot);
3188 Ops.push_back(DAG.getValueType(SrcVT));
3189 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3190 Tys, &Ops[0], Ops.size());
3193 Chain = Result.getValue(1);
3194 SDOperand InFlag = Result.getValue(2);
3196 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3197 // shouldn't be necessary except that RFP cannot be live across
3198 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3199 MachineFunction &MF = DAG.getMachineFunction();
3200 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3201 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3202 Tys = DAG.getVTList(MVT::Other);
3203 SmallVector<SDOperand, 8> Ops;
3204 Ops.push_back(Chain);
3205 Ops.push_back(Result);
3206 Ops.push_back(StackSlot);
3207 Ops.push_back(DAG.getValueType(Op.getValueType()));
3208 Ops.push_back(InFlag);
3209 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3210 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3216 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3217 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3218 "Unknown FP_TO_SINT to lower!");
3219 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3221 MachineFunction &MF = DAG.getMachineFunction();
3222 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3223 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3224 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3227 switch (Op.getValueType()) {
3228 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3229 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3230 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3231 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3234 SDOperand Chain = DAG.getEntryNode();
3235 SDOperand Value = Op.getOperand(0);
3237 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3238 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3239 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3241 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3243 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3244 Chain = Value.getValue(1);
3245 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3246 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3249 // Build the FP_TO_INT*_IN_MEM
3250 SDOperand Ops[] = { Chain, Value, StackSlot };
3251 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3254 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3257 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3258 MVT::ValueType VT = Op.getValueType();
3259 const Type *OpNTy = MVT::getTypeForValueType(VT);
3260 std::vector<Constant*> CV;
3261 if (VT == MVT::f64) {
3262 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3263 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3265 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3266 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3267 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3268 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3270 Constant *CS = ConstantStruct::get(CV);
3271 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3272 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3273 SmallVector<SDOperand, 3> Ops;
3274 Ops.push_back(DAG.getEntryNode());
3275 Ops.push_back(CPIdx);
3276 Ops.push_back(DAG.getSrcValue(NULL));
3277 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3278 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3281 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3282 MVT::ValueType VT = Op.getValueType();
3283 const Type *OpNTy = MVT::getTypeForValueType(VT);
3284 std::vector<Constant*> CV;
3285 if (VT == MVT::f64) {
3286 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3287 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3289 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3290 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3291 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3292 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3294 Constant *CS = ConstantStruct::get(CV);
3295 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3296 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3297 SmallVector<SDOperand, 3> Ops;
3298 Ops.push_back(DAG.getEntryNode());
3299 Ops.push_back(CPIdx);
3300 Ops.push_back(DAG.getSrcValue(NULL));
3301 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3302 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3305 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3306 SDOperand Op0 = Op.getOperand(0);
3307 SDOperand Op1 = Op.getOperand(1);
3308 MVT::ValueType VT = Op.getValueType();
3309 MVT::ValueType SrcVT = Op1.getValueType();
3310 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3312 // If second operand is smaller, extend it first.
3313 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3314 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3318 // First get the sign bit of second operand.
3319 std::vector<Constant*> CV;
3320 if (SrcVT == MVT::f64) {
3321 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3322 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3324 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3325 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3326 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3327 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3329 Constant *CS = ConstantStruct::get(CV);
3330 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3331 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3332 SmallVector<SDOperand, 3> Ops;
3333 Ops.push_back(DAG.getEntryNode());
3334 Ops.push_back(CPIdx);
3335 Ops.push_back(DAG.getSrcValue(NULL));
3336 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3337 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3339 // Shift sign bit right or left if the two operands have different types.
3340 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3341 // Op0 is MVT::f32, Op1 is MVT::f64.
3342 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3343 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3344 DAG.getConstant(32, MVT::i32));
3345 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3346 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3347 DAG.getConstant(0, getPointerTy()));
3350 // Clear first operand sign bit.
3352 if (VT == MVT::f64) {
3353 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3354 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3356 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3357 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3358 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3359 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3361 CS = ConstantStruct::get(CV);
3362 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3363 Tys = DAG.getVTList(VT, MVT::Other);
3365 Ops.push_back(DAG.getEntryNode());
3366 Ops.push_back(CPIdx);
3367 Ops.push_back(DAG.getSrcValue(NULL));
3368 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3369 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3371 // Or the value with the sign bit.
3372 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3375 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3377 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3379 SDOperand Op0 = Op.getOperand(0);
3380 SDOperand Op1 = Op.getOperand(1);
3381 SDOperand CC = Op.getOperand(2);
3382 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3383 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3384 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3385 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3388 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3390 SDOperand Ops1[] = { Chain, Op0, Op1 };
3391 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3392 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3393 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3396 assert(isFP && "Illegal integer SetCC!");
3398 SDOperand COps[] = { Chain, Op0, Op1 };
3399 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3401 switch (SetCCOpcode) {
3402 default: assert(false && "Illegal floating point SetCC!");
3403 case ISD::SETOEQ: { // !PF & ZF
3404 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3405 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3406 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3408 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3409 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3411 case ISD::SETUNE: { // PF | !ZF
3412 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3413 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3414 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3416 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3417 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3422 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3423 bool addTest = true;
3424 SDOperand Chain = DAG.getEntryNode();
3425 SDOperand Cond = Op.getOperand(0);
3427 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3429 if (Cond.getOpcode() == ISD::SETCC)
3430 Cond = LowerSETCC(Cond, DAG, Chain);
3432 if (Cond.getOpcode() == X86ISD::SETCC) {
3433 CC = Cond.getOperand(0);
3435 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3436 // (since flag operand cannot be shared). Use it as the condition setting
3437 // operand in place of the X86ISD::SETCC.
3438 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3439 // to use a test instead of duplicating the X86ISD::CMP (for register
3440 // pressure reason)?
3441 SDOperand Cmp = Cond.getOperand(1);
3442 unsigned Opc = Cmp.getOpcode();
3443 bool IllegalFPCMov = !X86ScalarSSE &&
3444 MVT::isFloatingPoint(Op.getValueType()) &&
3445 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3446 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3448 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3449 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3455 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3456 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3457 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3460 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3461 SmallVector<SDOperand, 4> Ops;
3462 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3463 // condition is true.
3464 Ops.push_back(Op.getOperand(2));
3465 Ops.push_back(Op.getOperand(1));
3467 Ops.push_back(Cond.getValue(1));
3468 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3471 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3472 bool addTest = true;
3473 SDOperand Chain = Op.getOperand(0);
3474 SDOperand Cond = Op.getOperand(1);
3475 SDOperand Dest = Op.getOperand(2);
3477 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3479 if (Cond.getOpcode() == ISD::SETCC)
3480 Cond = LowerSETCC(Cond, DAG, Chain);
3482 if (Cond.getOpcode() == X86ISD::SETCC) {
3483 CC = Cond.getOperand(0);
3485 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3486 // (since flag operand cannot be shared). Use it as the condition setting
3487 // operand in place of the X86ISD::SETCC.
3488 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3489 // to use a test instead of duplicating the X86ISD::CMP (for register
3490 // pressure reason)?
3491 SDOperand Cmp = Cond.getOperand(1);
3492 unsigned Opc = Cmp.getOpcode();
3493 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3494 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3495 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3501 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3502 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3503 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3505 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3506 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3509 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3510 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3512 if (Subtarget->is64Bit())
3513 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3515 switch (CallingConv) {
3517 assert(0 && "Unsupported calling convention");
3518 case CallingConv::Fast:
3519 // TODO: Implement fastcc
3521 case CallingConv::C:
3522 case CallingConv::X86_StdCall:
3523 return LowerCCCCallTo(Op, DAG, CallingConv);
3524 case CallingConv::X86_FastCall:
3525 return LowerFastCCCallTo(Op, DAG, CallingConv);
3530 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
3531 // Calls to _alloca is needed to probe the stack when allocating more than 4k
3532 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
3533 // that the guard pages used by the OS virtual memory manager are allocated in
3534 // correct sequence.
3535 SDOperand X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
3536 SelectionDAG &DAG) {
3537 assert(Subtarget->isTargetCygMing() &&
3538 "This should be used only on Cygwin/Mingw targets");
3541 SDOperand Chain = Op.getOperand(0);
3542 SDOperand Size = Op.getOperand(1);
3543 // FIXME: Ensure alignment here
3545 TargetLowering::ArgListTy Args;
3546 TargetLowering::ArgListEntry Entry;
3547 MVT::ValueType IntPtr = getPointerTy();
3548 MVT::ValueType SPTy = (Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
3549 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3552 Entry.Ty = IntPtrTy;
3553 Entry.isInReg = true; // Should pass in EAX
3554 Args.push_back(Entry);
3555 std::pair<SDOperand, SDOperand> CallResult =
3556 LowerCallTo(Chain, IntPtrTy, false, false, CallingConv::C, false,
3557 DAG.getExternalSymbol("_alloca", IntPtr), Args, DAG);
3559 SDOperand SP = DAG.getCopyFromReg(CallResult.second, X86StackPtr, SPTy);
3561 std::vector<MVT::ValueType> Tys;
3562 Tys.push_back(SPTy);
3563 Tys.push_back(MVT::Other);
3564 SDOperand Ops[2] = { SP, CallResult.second };
3565 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
3569 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3570 MachineFunction &MF = DAG.getMachineFunction();
3571 const Function* Fn = MF.getFunction();
3572 if (Fn->hasExternalLinkage() &&
3573 Subtarget->isTargetCygMing() &&
3574 Fn->getName() == "main")
3575 MF.getInfo<X86MachineFunctionInfo>()->setForceFramePointer(true);
3577 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3578 if (Subtarget->is64Bit())
3579 return LowerX86_64CCCArguments(Op, DAG);
3583 assert(0 && "Unsupported calling convention");
3584 case CallingConv::Fast:
3585 // TODO: implement fastcc.
3588 case CallingConv::C:
3589 return LowerCCCArguments(Op, DAG);
3590 case CallingConv::X86_StdCall:
3591 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(StdCall);
3592 return LowerCCCArguments(Op, DAG, true);
3593 case CallingConv::X86_FastCall:
3594 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(FastCall);
3595 return LowerFastCCArguments(Op, DAG);
3599 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3600 SDOperand InFlag(0, 0);
3601 SDOperand Chain = Op.getOperand(0);
3603 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3604 if (Align == 0) Align = 1;
3606 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3607 // If not DWORD aligned, call memset if size is less than the threshold.
3608 // It knows how to align to the right boundary first.
3609 if ((Align & 3) != 0 ||
3610 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3611 MVT::ValueType IntPtr = getPointerTy();
3612 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3613 TargetLowering::ArgListTy Args;
3614 TargetLowering::ArgListEntry Entry;
3615 Entry.Node = Op.getOperand(1);
3616 Entry.Ty = IntPtrTy;
3617 Args.push_back(Entry);
3618 // Extend the unsigned i8 argument to be an int value for the call.
3619 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3620 Entry.Ty = IntPtrTy;
3621 Args.push_back(Entry);
3622 Entry.Node = Op.getOperand(3);
3623 Args.push_back(Entry);
3624 std::pair<SDOperand,SDOperand> CallResult =
3625 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3626 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3627 return CallResult.second;
3632 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3633 unsigned BytesLeft = 0;
3634 bool TwoRepStos = false;
3637 uint64_t Val = ValC->getValue() & 255;
3639 // If the value is a constant, then we can potentially use larger sets.
3640 switch (Align & 3) {
3641 case 2: // WORD aligned
3644 Val = (Val << 8) | Val;
3646 case 0: // DWORD aligned
3649 Val = (Val << 8) | Val;
3650 Val = (Val << 16) | Val;
3651 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
3654 Val = (Val << 32) | Val;
3657 default: // Byte aligned
3660 Count = Op.getOperand(3);
3664 if (AVT > MVT::i8) {
3666 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3667 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3668 BytesLeft = I->getValue() % UBytes;
3670 assert(AVT >= MVT::i32 &&
3671 "Do not use rep;stos if not at least DWORD aligned");
3672 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3673 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3678 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
3680 InFlag = Chain.getValue(1);
3683 Count = Op.getOperand(3);
3684 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
3685 InFlag = Chain.getValue(1);
3688 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3690 InFlag = Chain.getValue(1);
3691 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3692 Op.getOperand(1), InFlag);
3693 InFlag = Chain.getValue(1);
3695 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3696 SmallVector<SDOperand, 8> Ops;
3697 Ops.push_back(Chain);
3698 Ops.push_back(DAG.getValueType(AVT));
3699 Ops.push_back(InFlag);
3700 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3703 InFlag = Chain.getValue(1);
3704 Count = Op.getOperand(3);
3705 MVT::ValueType CVT = Count.getValueType();
3706 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3707 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3708 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3710 InFlag = Chain.getValue(1);
3711 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3713 Ops.push_back(Chain);
3714 Ops.push_back(DAG.getValueType(MVT::i8));
3715 Ops.push_back(InFlag);
3716 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3717 } else if (BytesLeft) {
3718 // Issue stores for the last 1 - 7 bytes.
3720 unsigned Val = ValC->getValue() & 255;
3721 unsigned Offset = I->getValue() - BytesLeft;
3722 SDOperand DstAddr = Op.getOperand(1);
3723 MVT::ValueType AddrVT = DstAddr.getValueType();
3724 if (BytesLeft >= 4) {
3725 Val = (Val << 8) | Val;
3726 Val = (Val << 16) | Val;
3727 Value = DAG.getConstant(Val, MVT::i32);
3728 Chain = DAG.getStore(Chain, Value,
3729 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3730 DAG.getConstant(Offset, AddrVT)),
3735 if (BytesLeft >= 2) {
3736 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
3737 Chain = DAG.getStore(Chain, Value,
3738 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3739 DAG.getConstant(Offset, AddrVT)),
3744 if (BytesLeft == 1) {
3745 Value = DAG.getConstant(Val, MVT::i8);
3746 Chain = DAG.getStore(Chain, Value,
3747 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3748 DAG.getConstant(Offset, AddrVT)),
3756 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
3757 SDOperand Chain = Op.getOperand(0);
3759 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3760 if (Align == 0) Align = 1;
3762 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3763 // If not DWORD aligned, call memcpy if size is less than the threshold.
3764 // It knows how to align to the right boundary first.
3765 if ((Align & 3) != 0 ||
3766 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3767 MVT::ValueType IntPtr = getPointerTy();
3768 TargetLowering::ArgListTy Args;
3769 TargetLowering::ArgListEntry Entry;
3770 Entry.Ty = getTargetData()->getIntPtrType();
3771 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
3772 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
3773 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
3774 std::pair<SDOperand,SDOperand> CallResult =
3775 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3776 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
3777 return CallResult.second;
3782 unsigned BytesLeft = 0;
3783 bool TwoRepMovs = false;
3784 switch (Align & 3) {
3785 case 2: // WORD aligned
3788 case 0: // DWORD aligned
3790 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
3793 default: // Byte aligned
3795 Count = Op.getOperand(3);
3799 if (AVT > MVT::i8) {
3801 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3802 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3803 BytesLeft = I->getValue() % UBytes;
3805 assert(AVT >= MVT::i32 &&
3806 "Do not use rep;movs if not at least DWORD aligned");
3807 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3808 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3813 SDOperand InFlag(0, 0);
3814 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3816 InFlag = Chain.getValue(1);
3817 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3818 Op.getOperand(1), InFlag);
3819 InFlag = Chain.getValue(1);
3820 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
3821 Op.getOperand(2), InFlag);
3822 InFlag = Chain.getValue(1);
3824 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3825 SmallVector<SDOperand, 8> Ops;
3826 Ops.push_back(Chain);
3827 Ops.push_back(DAG.getValueType(AVT));
3828 Ops.push_back(InFlag);
3829 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3832 InFlag = Chain.getValue(1);
3833 Count = Op.getOperand(3);
3834 MVT::ValueType CVT = Count.getValueType();
3835 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3836 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3837 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3839 InFlag = Chain.getValue(1);
3840 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3842 Ops.push_back(Chain);
3843 Ops.push_back(DAG.getValueType(MVT::i8));
3844 Ops.push_back(InFlag);
3845 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3846 } else if (BytesLeft) {
3847 // Issue loads and stores for the last 1 - 7 bytes.
3848 unsigned Offset = I->getValue() - BytesLeft;
3849 SDOperand DstAddr = Op.getOperand(1);
3850 MVT::ValueType DstVT = DstAddr.getValueType();
3851 SDOperand SrcAddr = Op.getOperand(2);
3852 MVT::ValueType SrcVT = SrcAddr.getValueType();
3854 if (BytesLeft >= 4) {
3855 Value = DAG.getLoad(MVT::i32, Chain,
3856 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3857 DAG.getConstant(Offset, SrcVT)),
3859 Chain = Value.getValue(1);
3860 Chain = DAG.getStore(Chain, Value,
3861 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3862 DAG.getConstant(Offset, DstVT)),
3867 if (BytesLeft >= 2) {
3868 Value = DAG.getLoad(MVT::i16, Chain,
3869 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3870 DAG.getConstant(Offset, SrcVT)),
3872 Chain = Value.getValue(1);
3873 Chain = DAG.getStore(Chain, Value,
3874 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3875 DAG.getConstant(Offset, DstVT)),
3881 if (BytesLeft == 1) {
3882 Value = DAG.getLoad(MVT::i8, Chain,
3883 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3884 DAG.getConstant(Offset, SrcVT)),
3886 Chain = Value.getValue(1);
3887 Chain = DAG.getStore(Chain, Value,
3888 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3889 DAG.getConstant(Offset, DstVT)),
3898 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
3899 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3900 SDOperand TheOp = Op.getOperand(0);
3901 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
3902 if (Subtarget->is64Bit()) {
3903 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
3904 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
3905 MVT::i64, Copy1.getValue(2));
3906 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
3907 DAG.getConstant(32, MVT::i8));
3909 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
3912 Tys = DAG.getVTList(MVT::i64, MVT::Other);
3913 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
3916 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
3917 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
3918 MVT::i32, Copy1.getValue(2));
3919 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
3920 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
3921 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
3924 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
3925 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
3927 if (!Subtarget->is64Bit()) {
3928 // vastart just stores the address of the VarArgsFrameIndex slot into the
3929 // memory location argument.
3930 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3931 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
3936 // gp_offset (0 - 6 * 8)
3937 // fp_offset (48 - 48 + 8 * 16)
3938 // overflow_arg_area (point to parameters coming in memory).
3940 SmallVector<SDOperand, 8> MemOps;
3941 SDOperand FIN = Op.getOperand(1);
3943 SDOperand Store = DAG.getStore(Op.getOperand(0),
3944 DAG.getConstant(VarArgsGPOffset, MVT::i32),
3945 FIN, SV->getValue(), SV->getOffset());
3946 MemOps.push_back(Store);
3949 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3950 DAG.getConstant(4, getPointerTy()));
3951 Store = DAG.getStore(Op.getOperand(0),
3952 DAG.getConstant(VarArgsFPOffset, MVT::i32),
3953 FIN, SV->getValue(), SV->getOffset());
3954 MemOps.push_back(Store);
3956 // Store ptr to overflow_arg_area
3957 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3958 DAG.getConstant(4, getPointerTy()));
3959 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3960 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
3962 MemOps.push_back(Store);
3964 // Store ptr to reg_save_area.
3965 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3966 DAG.getConstant(8, getPointerTy()));
3967 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
3968 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
3970 MemOps.push_back(Store);
3971 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
3974 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
3975 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
3976 SDOperand Chain = Op.getOperand(0);
3977 SDOperand DstPtr = Op.getOperand(1);
3978 SDOperand SrcPtr = Op.getOperand(2);
3979 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
3980 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
3982 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
3983 SrcSV->getValue(), SrcSV->getOffset());
3984 Chain = SrcPtr.getValue(1);
3985 for (unsigned i = 0; i < 3; ++i) {
3986 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
3987 SrcSV->getValue(), SrcSV->getOffset());
3988 Chain = Val.getValue(1);
3989 Chain = DAG.getStore(Chain, Val, DstPtr,
3990 DstSV->getValue(), DstSV->getOffset());
3993 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
3994 DAG.getConstant(8, getPointerTy()));
3995 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
3996 DAG.getConstant(8, getPointerTy()));
4002 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4003 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4005 default: return SDOperand(); // Don't custom lower most intrinsics.
4006 // Comparison intrinsics.
4007 case Intrinsic::x86_sse_comieq_ss:
4008 case Intrinsic::x86_sse_comilt_ss:
4009 case Intrinsic::x86_sse_comile_ss:
4010 case Intrinsic::x86_sse_comigt_ss:
4011 case Intrinsic::x86_sse_comige_ss:
4012 case Intrinsic::x86_sse_comineq_ss:
4013 case Intrinsic::x86_sse_ucomieq_ss:
4014 case Intrinsic::x86_sse_ucomilt_ss:
4015 case Intrinsic::x86_sse_ucomile_ss:
4016 case Intrinsic::x86_sse_ucomigt_ss:
4017 case Intrinsic::x86_sse_ucomige_ss:
4018 case Intrinsic::x86_sse_ucomineq_ss:
4019 case Intrinsic::x86_sse2_comieq_sd:
4020 case Intrinsic::x86_sse2_comilt_sd:
4021 case Intrinsic::x86_sse2_comile_sd:
4022 case Intrinsic::x86_sse2_comigt_sd:
4023 case Intrinsic::x86_sse2_comige_sd:
4024 case Intrinsic::x86_sse2_comineq_sd:
4025 case Intrinsic::x86_sse2_ucomieq_sd:
4026 case Intrinsic::x86_sse2_ucomilt_sd:
4027 case Intrinsic::x86_sse2_ucomile_sd:
4028 case Intrinsic::x86_sse2_ucomigt_sd:
4029 case Intrinsic::x86_sse2_ucomige_sd:
4030 case Intrinsic::x86_sse2_ucomineq_sd: {
4032 ISD::CondCode CC = ISD::SETCC_INVALID;
4035 case Intrinsic::x86_sse_comieq_ss:
4036 case Intrinsic::x86_sse2_comieq_sd:
4040 case Intrinsic::x86_sse_comilt_ss:
4041 case Intrinsic::x86_sse2_comilt_sd:
4045 case Intrinsic::x86_sse_comile_ss:
4046 case Intrinsic::x86_sse2_comile_sd:
4050 case Intrinsic::x86_sse_comigt_ss:
4051 case Intrinsic::x86_sse2_comigt_sd:
4055 case Intrinsic::x86_sse_comige_ss:
4056 case Intrinsic::x86_sse2_comige_sd:
4060 case Intrinsic::x86_sse_comineq_ss:
4061 case Intrinsic::x86_sse2_comineq_sd:
4065 case Intrinsic::x86_sse_ucomieq_ss:
4066 case Intrinsic::x86_sse2_ucomieq_sd:
4067 Opc = X86ISD::UCOMI;
4070 case Intrinsic::x86_sse_ucomilt_ss:
4071 case Intrinsic::x86_sse2_ucomilt_sd:
4072 Opc = X86ISD::UCOMI;
4075 case Intrinsic::x86_sse_ucomile_ss:
4076 case Intrinsic::x86_sse2_ucomile_sd:
4077 Opc = X86ISD::UCOMI;
4080 case Intrinsic::x86_sse_ucomigt_ss:
4081 case Intrinsic::x86_sse2_ucomigt_sd:
4082 Opc = X86ISD::UCOMI;
4085 case Intrinsic::x86_sse_ucomige_ss:
4086 case Intrinsic::x86_sse2_ucomige_sd:
4087 Opc = X86ISD::UCOMI;
4090 case Intrinsic::x86_sse_ucomineq_ss:
4091 case Intrinsic::x86_sse2_ucomineq_sd:
4092 Opc = X86ISD::UCOMI;
4098 SDOperand LHS = Op.getOperand(1);
4099 SDOperand RHS = Op.getOperand(2);
4100 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4102 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4103 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
4104 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
4105 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
4106 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
4107 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
4108 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4113 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4114 // Depths > 0 not supported yet!
4115 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4118 // Just load the return address
4119 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4120 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4123 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4124 // Depths > 0 not supported yet!
4125 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4128 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4129 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4130 DAG.getConstant(4, getPointerTy()));
4133 /// LowerOperation - Provide custom lowering hooks for some operations.
4135 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4136 switch (Op.getOpcode()) {
4137 default: assert(0 && "Should not custom lower this!");
4138 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4139 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4140 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4141 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4142 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4143 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4144 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4145 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4146 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4147 case ISD::SHL_PARTS:
4148 case ISD::SRA_PARTS:
4149 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4150 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4151 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4152 case ISD::FABS: return LowerFABS(Op, DAG);
4153 case ISD::FNEG: return LowerFNEG(Op, DAG);
4154 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4155 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
4156 case ISD::SELECT: return LowerSELECT(Op, DAG);
4157 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4158 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4159 case ISD::CALL: return LowerCALL(Op, DAG);
4160 case ISD::RET: return LowerRET(Op, DAG);
4161 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4162 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4163 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4164 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4165 case ISD::VASTART: return LowerVASTART(Op, DAG);
4166 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
4167 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4168 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4169 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4170 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
4175 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4177 default: return NULL;
4178 case X86ISD::SHLD: return "X86ISD::SHLD";
4179 case X86ISD::SHRD: return "X86ISD::SHRD";
4180 case X86ISD::FAND: return "X86ISD::FAND";
4181 case X86ISD::FOR: return "X86ISD::FOR";
4182 case X86ISD::FXOR: return "X86ISD::FXOR";
4183 case X86ISD::FSRL: return "X86ISD::FSRL";
4184 case X86ISD::FILD: return "X86ISD::FILD";
4185 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
4186 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
4187 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
4188 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
4189 case X86ISD::FLD: return "X86ISD::FLD";
4190 case X86ISD::FST: return "X86ISD::FST";
4191 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
4192 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
4193 case X86ISD::CALL: return "X86ISD::CALL";
4194 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
4195 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
4196 case X86ISD::CMP: return "X86ISD::CMP";
4197 case X86ISD::COMI: return "X86ISD::COMI";
4198 case X86ISD::UCOMI: return "X86ISD::UCOMI";
4199 case X86ISD::SETCC: return "X86ISD::SETCC";
4200 case X86ISD::CMOV: return "X86ISD::CMOV";
4201 case X86ISD::BRCOND: return "X86ISD::BRCOND";
4202 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
4203 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
4204 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
4205 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
4206 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
4207 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
4208 case X86ISD::Wrapper: return "X86ISD::Wrapper";
4209 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4210 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4211 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4212 case X86ISD::FMAX: return "X86ISD::FMAX";
4213 case X86ISD::FMIN: return "X86ISD::FMIN";
4214 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
4215 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
4219 // isLegalAddressingMode - Return true if the addressing mode represented
4220 // by AM is legal for this target, for a load/store of the specified type.
4221 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
4222 const Type *Ty) const {
4223 // X86 supports extremely general addressing modes.
4225 // X86 allows a sign-extended 32-bit immediate field as a displacement.
4226 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
4230 // X86-64 only supports addr of globals in small code model.
4231 if (Subtarget->is64Bit() &&
4232 getTargetMachine().getCodeModel() != CodeModel::Small)
4235 // We can only fold this if we don't need a load either.
4236 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
4246 // These scales always work.
4251 // These scales are formed with basereg+scalereg. Only accept if there is
4256 default: // Other stuff never works.
4264 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4265 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4266 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4267 /// are assumed to be legal.
4269 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4270 // Only do shuffles on 128-bit vector types for now.
4271 if (MVT::getSizeInBits(VT) == 64) return false;
4272 return (Mask.Val->getNumOperands() <= 4 ||
4273 isSplatMask(Mask.Val) ||
4274 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4275 X86::isUNPCKLMask(Mask.Val) ||
4276 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4277 X86::isUNPCKH_v_undef_Mask(Mask.Val) ||
4278 X86::isUNPCKHMask(Mask.Val));
4281 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4283 SelectionDAG &DAG) const {
4284 unsigned NumElts = BVOps.size();
4285 // Only do shuffles on 128-bit vector types for now.
4286 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4287 if (NumElts == 2) return true;
4289 return (isMOVLMask(&BVOps[0], 4) ||
4290 isCommutedMOVL(&BVOps[0], 4, true) ||
4291 isSHUFPMask(&BVOps[0], 4) ||
4292 isCommutedSHUFP(&BVOps[0], 4));
4297 //===----------------------------------------------------------------------===//
4298 // X86 Scheduler Hooks
4299 //===----------------------------------------------------------------------===//
4302 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4303 MachineBasicBlock *BB) {
4304 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4305 switch (MI->getOpcode()) {
4306 default: assert(false && "Unexpected instr type to insert");
4307 case X86::CMOV_FR32:
4308 case X86::CMOV_FR64:
4309 case X86::CMOV_V4F32:
4310 case X86::CMOV_V2F64:
4311 case X86::CMOV_V2I64: {
4312 // To "insert" a SELECT_CC instruction, we actually have to insert the
4313 // diamond control-flow pattern. The incoming instruction knows the
4314 // destination vreg to set, the condition code register to branch on, the
4315 // true/false values to select between, and a branch opcode to use.
4316 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4317 ilist<MachineBasicBlock>::iterator It = BB;
4323 // cmpTY ccX, r1, r2
4325 // fallthrough --> copy0MBB
4326 MachineBasicBlock *thisMBB = BB;
4327 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4328 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4330 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4331 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4332 MachineFunction *F = BB->getParent();
4333 F->getBasicBlockList().insert(It, copy0MBB);
4334 F->getBasicBlockList().insert(It, sinkMBB);
4335 // Update machine-CFG edges by first adding all successors of the current
4336 // block to the new block which will contain the Phi node for the select.
4337 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4338 e = BB->succ_end(); i != e; ++i)
4339 sinkMBB->addSuccessor(*i);
4340 // Next, remove all successors of the current block, and add the true
4341 // and fallthrough blocks as its successors.
4342 while(!BB->succ_empty())
4343 BB->removeSuccessor(BB->succ_begin());
4344 BB->addSuccessor(copy0MBB);
4345 BB->addSuccessor(sinkMBB);
4348 // %FalseValue = ...
4349 // # fallthrough to sinkMBB
4352 // Update machine-CFG edges
4353 BB->addSuccessor(sinkMBB);
4356 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4359 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4360 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4361 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4363 delete MI; // The pseudo instruction is gone now.
4367 case X86::FP_TO_INT16_IN_MEM:
4368 case X86::FP_TO_INT32_IN_MEM:
4369 case X86::FP_TO_INT64_IN_MEM: {
4370 // Change the floating point control register to use "round towards zero"
4371 // mode when truncating to an integer value.
4372 MachineFunction *F = BB->getParent();
4373 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4374 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4376 // Load the old value of the high byte of the control word...
4378 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4379 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4381 // Set the high part to be round to zero...
4382 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4385 // Reload the modified control word now...
4386 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4388 // Restore the memory image of control word to original value
4389 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4392 // Get the X86 opcode to use.
4394 switch (MI->getOpcode()) {
4395 default: assert(0 && "illegal opcode!");
4396 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4397 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4398 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4402 MachineOperand &Op = MI->getOperand(0);
4403 if (Op.isRegister()) {
4404 AM.BaseType = X86AddressMode::RegBase;
4405 AM.Base.Reg = Op.getReg();
4407 AM.BaseType = X86AddressMode::FrameIndexBase;
4408 AM.Base.FrameIndex = Op.getFrameIndex();
4410 Op = MI->getOperand(1);
4411 if (Op.isImmediate())
4412 AM.Scale = Op.getImm();
4413 Op = MI->getOperand(2);
4414 if (Op.isImmediate())
4415 AM.IndexReg = Op.getImm();
4416 Op = MI->getOperand(3);
4417 if (Op.isGlobalAddress()) {
4418 AM.GV = Op.getGlobal();
4420 AM.Disp = Op.getImm();
4422 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4423 .addReg(MI->getOperand(4).getReg());
4425 // Reload the original control word now.
4426 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4428 delete MI; // The pseudo instruction is gone now.
4434 //===----------------------------------------------------------------------===//
4435 // X86 Optimization Hooks
4436 //===----------------------------------------------------------------------===//
4438 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4440 uint64_t &KnownZero,
4442 unsigned Depth) const {
4443 unsigned Opc = Op.getOpcode();
4444 assert((Opc >= ISD::BUILTIN_OP_END ||
4445 Opc == ISD::INTRINSIC_WO_CHAIN ||
4446 Opc == ISD::INTRINSIC_W_CHAIN ||
4447 Opc == ISD::INTRINSIC_VOID) &&
4448 "Should use MaskedValueIsZero if you don't know whether Op"
4449 " is a target node!");
4451 KnownZero = KnownOne = 0; // Don't know anything.
4455 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4460 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4461 /// element of the result of the vector shuffle.
4462 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4463 MVT::ValueType VT = N->getValueType(0);
4464 SDOperand PermMask = N->getOperand(2);
4465 unsigned NumElems = PermMask.getNumOperands();
4466 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4468 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4470 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4471 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4472 SDOperand Idx = PermMask.getOperand(i);
4473 if (Idx.getOpcode() == ISD::UNDEF)
4474 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4475 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4480 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4481 /// node is a GlobalAddress + an offset.
4482 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4483 unsigned Opc = N->getOpcode();
4484 if (Opc == X86ISD::Wrapper) {
4485 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4486 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4489 } else if (Opc == ISD::ADD) {
4490 SDOperand N1 = N->getOperand(0);
4491 SDOperand N2 = N->getOperand(1);
4492 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4493 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4495 Offset += V->getSignExtended();
4498 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4499 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4501 Offset += V->getSignExtended();
4509 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4511 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4512 MachineFrameInfo *MFI) {
4513 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4516 SDOperand Loc = N->getOperand(1);
4517 SDOperand BaseLoc = Base->getOperand(1);
4518 if (Loc.getOpcode() == ISD::FrameIndex) {
4519 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4521 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4522 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4523 int FS = MFI->getObjectSize(FI);
4524 int BFS = MFI->getObjectSize(BFI);
4525 if (FS != BFS || FS != Size) return false;
4526 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4528 GlobalValue *GV1 = NULL;
4529 GlobalValue *GV2 = NULL;
4530 int64_t Offset1 = 0;
4531 int64_t Offset2 = 0;
4532 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4533 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4534 if (isGA1 && isGA2 && GV1 == GV2)
4535 return Offset1 == (Offset2 + Dist*Size);
4541 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4542 const X86Subtarget *Subtarget) {
4545 if (isGAPlusOffset(Base, GV, Offset))
4546 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4548 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4549 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4551 // Fixed objects do not specify alignment, however the offsets are known.
4552 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4553 (MFI->getObjectOffset(BFI) % 16) == 0);
4555 return MFI->getObjectAlignment(BFI) >= 16;
4561 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4562 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4563 /// if the load addresses are consecutive, non-overlapping, and in the right
4565 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4566 const X86Subtarget *Subtarget) {
4567 MachineFunction &MF = DAG.getMachineFunction();
4568 MachineFrameInfo *MFI = MF.getFrameInfo();
4569 MVT::ValueType VT = N->getValueType(0);
4570 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4571 SDOperand PermMask = N->getOperand(2);
4572 int NumElems = (int)PermMask.getNumOperands();
4573 SDNode *Base = NULL;
4574 for (int i = 0; i < NumElems; ++i) {
4575 SDOperand Idx = PermMask.getOperand(i);
4576 if (Idx.getOpcode() == ISD::UNDEF) {
4577 if (!Base) return SDOperand();
4580 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4581 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4585 else if (!isConsecutiveLoad(Arg.Val, Base,
4586 i, MVT::getSizeInBits(EVT)/8,MFI))
4591 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4593 LoadSDNode *LD = cast<LoadSDNode>(Base);
4594 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4595 LD->getSrcValueOffset());
4597 // Just use movups, it's shorter.
4598 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
4599 SmallVector<SDOperand, 3> Ops;
4600 Ops.push_back(Base->getOperand(0));
4601 Ops.push_back(Base->getOperand(1));
4602 Ops.push_back(Base->getOperand(2));
4603 return DAG.getNode(ISD::BIT_CONVERT, VT,
4604 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
4608 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4609 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4610 const X86Subtarget *Subtarget) {
4611 SDOperand Cond = N->getOperand(0);
4613 // If we have SSE[12] support, try to form min/max nodes.
4614 if (Subtarget->hasSSE2() &&
4615 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4616 if (Cond.getOpcode() == ISD::SETCC) {
4617 // Get the LHS/RHS of the select.
4618 SDOperand LHS = N->getOperand(1);
4619 SDOperand RHS = N->getOperand(2);
4620 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4622 unsigned Opcode = 0;
4623 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4626 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4629 if (!UnsafeFPMath) break;
4631 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4633 Opcode = X86ISD::FMIN;
4636 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4639 if (!UnsafeFPMath) break;
4641 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4643 Opcode = X86ISD::FMAX;
4646 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4649 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4652 if (!UnsafeFPMath) break;
4654 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4656 Opcode = X86ISD::FMIN;
4659 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4662 if (!UnsafeFPMath) break;
4664 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4666 Opcode = X86ISD::FMAX;
4672 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4681 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4682 DAGCombinerInfo &DCI) const {
4683 SelectionDAG &DAG = DCI.DAG;
4684 switch (N->getOpcode()) {
4686 case ISD::VECTOR_SHUFFLE:
4687 return PerformShuffleCombine(N, DAG, Subtarget);
4689 return PerformSELECTCombine(N, DAG, Subtarget);
4695 //===----------------------------------------------------------------------===//
4696 // X86 Inline Assembly Support
4697 //===----------------------------------------------------------------------===//
4699 /// getConstraintType - Given a constraint letter, return the type of
4700 /// constraint it is for this target.
4701 X86TargetLowering::ConstraintType
4702 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
4703 if (Constraint.size() == 1) {
4704 switch (Constraint[0]) {
4713 return C_RegisterClass;
4718 return TargetLowering::getConstraintType(Constraint);
4721 /// isOperandValidForConstraint - Return the specified operand (possibly
4722 /// modified) if the specified SDOperand is valid for the specified target
4723 /// constraint letter, otherwise return null.
4724 SDOperand X86TargetLowering::
4725 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
4726 switch (Constraint) {
4729 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4730 if (C->getValue() <= 31)
4731 return DAG.getTargetConstant(C->getValue(), Op.getValueType());
4733 return SDOperand(0,0);
4735 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4736 if (C->getValue() <= 255)
4737 return DAG.getTargetConstant(C->getValue(), Op.getValueType());
4739 return SDOperand(0,0);
4741 // Literal immediates are always ok.
4742 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op))
4743 return DAG.getTargetConstant(CST->getValue(), Op.getValueType());
4745 // If we are in non-pic codegen mode, we allow the address of a global (with
4746 // an optional displacement) to be used with 'i'.
4747 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
4750 // Match either (GA) or (GA+C)
4752 Offset = GA->getOffset();
4753 } else if (Op.getOpcode() == ISD::ADD) {
4754 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
4755 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
4757 Offset = GA->getOffset()+C->getValue();
4759 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
4760 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
4762 Offset = GA->getOffset()+C->getValue();
4769 // If addressing this global requires a load (e.g. in PIC mode), we can't
4771 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
4773 return SDOperand(0, 0);
4775 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4780 // Otherwise, not valid for this mode.
4781 return SDOperand(0, 0);
4784 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
4787 std::vector<unsigned> X86TargetLowering::
4788 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4789 MVT::ValueType VT) const {
4790 if (Constraint.size() == 1) {
4791 // FIXME: not handling fp-stack yet!
4792 switch (Constraint[0]) { // GCC X86 Constraint Letters
4793 default: break; // Unknown constraint letter
4794 case 'A': // EAX/EDX
4795 if (VT == MVT::i32 || VT == MVT::i64)
4796 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
4798 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
4801 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
4802 else if (VT == MVT::i16)
4803 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
4804 else if (VT == MVT::i8)
4805 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4810 return std::vector<unsigned>();
4813 std::pair<unsigned, const TargetRegisterClass*>
4814 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4815 MVT::ValueType VT) const {
4816 // First, see if this is a constraint that directly corresponds to an LLVM
4818 if (Constraint.size() == 1) {
4819 // GCC Constraint Letters
4820 switch (Constraint[0]) {
4822 case 'r': // GENERAL_REGS
4823 case 'R': // LEGACY_REGS
4824 case 'l': // INDEX_REGS
4825 if (VT == MVT::i64 && Subtarget->is64Bit())
4826 return std::make_pair(0U, X86::GR64RegisterClass);
4828 return std::make_pair(0U, X86::GR32RegisterClass);
4829 else if (VT == MVT::i16)
4830 return std::make_pair(0U, X86::GR16RegisterClass);
4831 else if (VT == MVT::i8)
4832 return std::make_pair(0U, X86::GR8RegisterClass);
4834 case 'y': // MMX_REGS if MMX allowed.
4835 if (!Subtarget->hasMMX()) break;
4836 return std::make_pair(0U, X86::VR64RegisterClass);
4838 case 'Y': // SSE_REGS if SSE2 allowed
4839 if (!Subtarget->hasSSE2()) break;
4841 case 'x': // SSE_REGS if SSE1 allowed
4842 if (!Subtarget->hasSSE1()) break;
4846 // Scalar SSE types.
4849 return std::make_pair(0U, X86::FR32RegisterClass);
4852 return std::make_pair(0U, X86::FR64RegisterClass);
4861 return std::make_pair(0U, X86::VR128RegisterClass);
4867 // Use the default implementation in TargetLowering to convert the register
4868 // constraint into a member of a register class.
4869 std::pair<unsigned, const TargetRegisterClass*> Res;
4870 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4872 // Not found as a standard register?
4873 if (Res.second == 0) {
4874 // GCC calls "st(0)" just plain "st".
4875 if (StringsEqualNoCase("{st}", Constraint)) {
4876 Res.first = X86::ST0;
4877 Res.second = X86::RSTRegisterClass;
4883 // Otherwise, check to see if this is a register class of the wrong value
4884 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
4885 // turn into {ax},{dx}.
4886 if (Res.second->hasType(VT))
4887 return Res; // Correct type already, nothing to do.
4889 // All of the single-register GCC register classes map their values onto
4890 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
4891 // really want an 8-bit or 32-bit register, map to the appropriate register
4892 // class and return the appropriate register.
4893 if (Res.second != X86::GR16RegisterClass)
4896 if (VT == MVT::i8) {
4897 unsigned DestReg = 0;
4898 switch (Res.first) {
4900 case X86::AX: DestReg = X86::AL; break;
4901 case X86::DX: DestReg = X86::DL; break;
4902 case X86::CX: DestReg = X86::CL; break;
4903 case X86::BX: DestReg = X86::BL; break;
4906 Res.first = DestReg;
4907 Res.second = Res.second = X86::GR8RegisterClass;
4909 } else if (VT == MVT::i32) {
4910 unsigned DestReg = 0;
4911 switch (Res.first) {
4913 case X86::AX: DestReg = X86::EAX; break;
4914 case X86::DX: DestReg = X86::EDX; break;
4915 case X86::CX: DestReg = X86::ECX; break;
4916 case X86::BX: DestReg = X86::EBX; break;
4917 case X86::SI: DestReg = X86::ESI; break;
4918 case X86::DI: DestReg = X86::EDI; break;
4919 case X86::BP: DestReg = X86::EBP; break;
4920 case X86::SP: DestReg = X86::ESP; break;
4923 Res.first = DestReg;
4924 Res.second = Res.second = X86::GR32RegisterClass;
4926 } else if (VT == MVT::i64) {
4927 unsigned DestReg = 0;
4928 switch (Res.first) {
4930 case X86::AX: DestReg = X86::RAX; break;
4931 case X86::DX: DestReg = X86::RDX; break;
4932 case X86::CX: DestReg = X86::RCX; break;
4933 case X86::BX: DestReg = X86::RBX; break;
4934 case X86::SI: DestReg = X86::RSI; break;
4935 case X86::DI: DestReg = X86::RDI; break;
4936 case X86::BP: DestReg = X86::RBP; break;
4937 case X86::SP: DestReg = X86::RSP; break;
4940 Res.first = DestReg;
4941 Res.second = Res.second = X86::GR64RegisterClass;