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 "X86CodeEmitter.h"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86MachineFunctionInfo.h"
20 #include "X86TargetMachine.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalVariable.h"
25 #include "llvm/Function.h"
26 #include "llvm/Intrinsics.h"
27 #include "llvm/ADT/VectorExtras.h"
28 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
29 #include "llvm/CodeGen/CallingConvLower.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/SelectionDAG.h"
34 #include "llvm/CodeGen/SSARegMap.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Target/TargetOptions.h"
37 #include "llvm/ADT/StringExtras.h"
38 #include "llvm/ParameterAttributes.h"
41 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
42 : TargetLowering(TM) {
43 Subtarget = &TM.getSubtarget<X86Subtarget>();
44 X86ScalarSSE = Subtarget->hasSSE2();
45 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
47 RegInfo = TM.getRegisterInfo();
49 // Set up the TargetLowering object.
51 // X86 is weird, it always uses i8 for shift amounts and setcc results.
52 setShiftAmountType(MVT::i8);
53 setSetCCResultType(MVT::i8);
54 setSetCCResultContents(ZeroOrOneSetCCResult);
55 setSchedulingPreference(SchedulingForRegPressure);
56 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
57 setStackPointerRegisterToSaveRestore(X86StackPtr);
59 if (Subtarget->isTargetDarwin()) {
60 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
61 setUseUnderscoreSetJmp(false);
62 setUseUnderscoreLongJmp(false);
63 } else if (Subtarget->isTargetMingw()) {
64 // MS runtime is weird: it exports _setjmp, but longjmp!
65 setUseUnderscoreSetJmp(true);
66 setUseUnderscoreLongJmp(false);
68 setUseUnderscoreSetJmp(true);
69 setUseUnderscoreLongJmp(true);
72 // Set up the register classes.
73 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
74 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
75 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
76 if (Subtarget->is64Bit())
77 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
79 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
81 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
83 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
84 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
85 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
87 if (Subtarget->is64Bit()) {
88 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
89 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
92 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
93 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
95 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
98 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
100 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
101 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
102 // SSE has no i16 to fp conversion, only i32
104 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
106 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
107 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
110 if (!Subtarget->is64Bit()) {
111 // Custom lower SINT_TO_FP and FP_TO_SINT from/to i64 in 32-bit mode.
112 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
113 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
116 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
118 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
119 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
122 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
124 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
125 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
128 // Handle FP_TO_UINT by promoting the destination to a larger signed
130 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
131 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
132 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
134 if (Subtarget->is64Bit()) {
135 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
136 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
138 if (X86ScalarSSE && !Subtarget->hasSSE3())
139 // Expand FP_TO_UINT into a select.
140 // FIXME: We would like to use a Custom expander here eventually to do
141 // the optimal thing for SSE vs. the default expansion in the legalizer.
142 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
144 // With SSE3 we can use fisttpll to convert to a signed i64.
145 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
148 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
150 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
151 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
154 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
155 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
156 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
157 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
158 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
159 if (Subtarget->is64Bit())
160 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Expand);
161 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
162 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
163 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
164 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
165 setOperationAction(ISD::FREM , MVT::f64 , Expand);
167 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
168 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
169 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
170 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
171 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
172 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
173 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
174 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
175 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
176 if (Subtarget->is64Bit()) {
177 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
178 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
179 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
182 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
183 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
185 // These should be promoted to a larger select which is supported.
186 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
187 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
188 // X86 wants to expand cmov itself.
189 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
190 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
191 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
192 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
193 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
194 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
195 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
196 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
197 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
198 if (Subtarget->is64Bit()) {
199 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
200 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
202 // X86 ret instruction may pop stack.
203 setOperationAction(ISD::RET , MVT::Other, Custom);
204 if (!Subtarget->is64Bit())
205 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
208 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
209 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
210 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
211 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
212 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
213 if (Subtarget->is64Bit()) {
214 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
215 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
216 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
217 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
219 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
220 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
221 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
222 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
223 // X86 wants to expand memset / memcpy itself.
224 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
225 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
227 // We don't have line number support yet.
228 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
229 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
230 // FIXME - use subtarget debug flags
231 if (!Subtarget->isTargetDarwin() &&
232 !Subtarget->isTargetELF() &&
233 !Subtarget->isTargetCygMing())
234 setOperationAction(ISD::LABEL, MVT::Other, Expand);
236 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
237 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
238 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
239 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
240 if (Subtarget->is64Bit()) {
242 setExceptionPointerRegister(X86::RAX);
243 setExceptionSelectorRegister(X86::RDX);
245 setExceptionPointerRegister(X86::EAX);
246 setExceptionSelectorRegister(X86::EDX);
249 setOperationAction(ISD::ADJUST_TRAMP, MVT::i32, Expand);
250 setOperationAction(ISD::ADJUST_TRAMP, MVT::i64, Expand);
251 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
253 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
254 setOperationAction(ISD::VASTART , MVT::Other, Custom);
255 setOperationAction(ISD::VAARG , MVT::Other, Expand);
256 setOperationAction(ISD::VAEND , MVT::Other, Expand);
257 if (Subtarget->is64Bit())
258 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
260 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
262 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
263 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
264 if (Subtarget->is64Bit())
265 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
266 if (Subtarget->isTargetCygMing())
267 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
269 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
272 // Set up the FP register classes.
273 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
274 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
276 // Use ANDPD to simulate FABS.
277 setOperationAction(ISD::FABS , MVT::f64, Custom);
278 setOperationAction(ISD::FABS , MVT::f32, Custom);
280 // Use XORP to simulate FNEG.
281 setOperationAction(ISD::FNEG , MVT::f64, Custom);
282 setOperationAction(ISD::FNEG , MVT::f32, Custom);
284 // Use ANDPD and ORPD to simulate FCOPYSIGN.
285 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
286 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
288 // We don't support sin/cos/fmod
289 setOperationAction(ISD::FSIN , MVT::f64, Expand);
290 setOperationAction(ISD::FCOS , MVT::f64, Expand);
291 setOperationAction(ISD::FREM , MVT::f64, Expand);
292 setOperationAction(ISD::FSIN , MVT::f32, Expand);
293 setOperationAction(ISD::FCOS , MVT::f32, Expand);
294 setOperationAction(ISD::FREM , MVT::f32, Expand);
296 // Expand FP immediates into loads from the stack, except for the special
298 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
299 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
300 addLegalFPImmediate(+0.0); // xorps / xorpd
302 // Conversions to long double (in X87) go through memory.
303 setConvertAction(MVT::f32, MVT::f80, Expand);
304 setConvertAction(MVT::f64, MVT::f80, Expand);
306 // Conversions from long double (in X87) go through memory.
307 setConvertAction(MVT::f80, MVT::f32, Expand);
308 setConvertAction(MVT::f80, MVT::f64, Expand);
310 // Set up the FP register classes.
311 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
312 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
314 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
315 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
316 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
317 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
319 // Floating truncations need to go through memory.
320 setConvertAction(MVT::f80, MVT::f32, Expand);
321 setConvertAction(MVT::f64, MVT::f32, Expand);
322 setConvertAction(MVT::f80, MVT::f64, Expand);
325 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
326 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
329 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
330 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
331 addLegalFPImmediate(+0.0); // FLD0
332 addLegalFPImmediate(+1.0); // FLD1
333 addLegalFPImmediate(-0.0); // FLD0/FCHS
334 addLegalFPImmediate(-1.0); // FLD1/FCHS
337 // Long double always uses X87.
338 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
340 // First set operation action for all vector types to expand. Then we
341 // will selectively turn on ones that can be effectively codegen'd.
342 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
343 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
344 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
345 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
346 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
347 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
348 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
349 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
350 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
351 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
352 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
353 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
354 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
355 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
356 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
357 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
358 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
359 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
360 setOperationAction(ISD::FABS, (MVT::ValueType)VT, Expand);
361 setOperationAction(ISD::FSIN, (MVT::ValueType)VT, Expand);
362 setOperationAction(ISD::FCOS, (MVT::ValueType)VT, Expand);
363 setOperationAction(ISD::FREM, (MVT::ValueType)VT, Expand);
364 setOperationAction(ISD::FPOWI, (MVT::ValueType)VT, Expand);
365 setOperationAction(ISD::FSQRT, (MVT::ValueType)VT, Expand);
366 setOperationAction(ISD::FCOPYSIGN, (MVT::ValueType)VT, Expand);
369 if (Subtarget->hasMMX()) {
370 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
371 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
372 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
373 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
375 // FIXME: add MMX packed arithmetics
377 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
378 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
379 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
380 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
382 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
383 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
384 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
386 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
387 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
389 setOperationAction(ISD::AND, MVT::v8i8, Promote);
390 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
391 setOperationAction(ISD::AND, MVT::v4i16, Promote);
392 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
393 setOperationAction(ISD::AND, MVT::v2i32, Promote);
394 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
395 setOperationAction(ISD::AND, MVT::v1i64, Legal);
397 setOperationAction(ISD::OR, MVT::v8i8, Promote);
398 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
399 setOperationAction(ISD::OR, MVT::v4i16, Promote);
400 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
401 setOperationAction(ISD::OR, MVT::v2i32, Promote);
402 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
403 setOperationAction(ISD::OR, MVT::v1i64, Legal);
405 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
406 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
407 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
408 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
409 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
410 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
411 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
413 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
414 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
415 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
416 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
417 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
418 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
419 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
421 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
422 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
423 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
424 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
426 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
427 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
428 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
429 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
431 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
432 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
433 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Custom);
434 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
437 if (Subtarget->hasSSE1()) {
438 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
440 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
441 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
442 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
443 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
444 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
445 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
446 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
447 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
448 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
449 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
450 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
453 if (Subtarget->hasSSE2()) {
454 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
455 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
456 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
457 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
458 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
460 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
461 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
462 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
463 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
464 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
465 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
466 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
467 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
468 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
469 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
470 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
471 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
472 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
473 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
474 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
476 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
477 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
478 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
479 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
480 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
481 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
483 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
484 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
485 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
486 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
487 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
489 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
490 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
491 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
492 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
493 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
494 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
496 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
497 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
498 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
499 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
500 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
501 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
502 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
503 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
504 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
505 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
506 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
507 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
510 // Custom lower v2i64 and v2f64 selects.
511 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
512 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
513 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
514 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
517 // We want to custom lower some of our intrinsics.
518 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
520 // We have target-specific dag combine patterns for the following nodes:
521 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
522 setTargetDAGCombine(ISD::SELECT);
524 computeRegisterProperties();
526 // FIXME: These should be based on subtarget info. Plus, the values should
527 // be smaller when we are in optimizing for size mode.
528 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
529 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
530 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
531 allowUnalignedMemoryAccesses = true; // x86 supports it!
535 //===----------------------------------------------------------------------===//
536 // Return Value Calling Convention Implementation
537 //===----------------------------------------------------------------------===//
539 #include "X86GenCallingConv.inc"
541 /// LowerRET - Lower an ISD::RET node.
542 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
543 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
545 SmallVector<CCValAssign, 16> RVLocs;
546 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
547 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
548 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
549 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
552 // If this is the first return lowered for this function, add the regs to the
553 // liveout set for the function.
554 if (DAG.getMachineFunction().liveout_empty()) {
555 for (unsigned i = 0; i != RVLocs.size(); ++i)
556 if (RVLocs[i].isRegLoc())
557 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
560 SDOperand Chain = Op.getOperand(0);
563 // Copy the result values into the output registers.
564 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
565 RVLocs[0].getLocReg() != X86::ST0) {
566 for (unsigned i = 0; i != RVLocs.size(); ++i) {
567 CCValAssign &VA = RVLocs[i];
568 assert(VA.isRegLoc() && "Can only return in registers!");
569 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
571 Flag = Chain.getValue(1);
574 // We need to handle a destination of ST0 specially, because it isn't really
576 SDOperand Value = Op.getOperand(1);
578 // If this is an FP return with ScalarSSE, we need to move the value from
579 // an XMM register onto the fp-stack.
583 // If this is a load into a scalarsse value, don't store the loaded value
584 // back to the stack, only to reload it: just replace the scalar-sse load.
585 if (ISD::isNON_EXTLoad(Value.Val) &&
586 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
587 Chain = Value.getOperand(0);
588 MemLoc = Value.getOperand(1);
590 // Spill the value to memory and reload it into top of stack.
591 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
592 MachineFunction &MF = DAG.getMachineFunction();
593 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
594 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
595 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
597 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other);
598 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
599 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
600 Chain = Value.getValue(1);
603 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
604 SDOperand Ops[] = { Chain, Value };
605 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
606 Flag = Chain.getValue(1);
609 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
611 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
613 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
617 /// LowerCallResult - Lower the result values of an ISD::CALL into the
618 /// appropriate copies out of appropriate physical registers. This assumes that
619 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
620 /// being lowered. The returns a SDNode with the same number of values as the
622 SDNode *X86TargetLowering::
623 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
624 unsigned CallingConv, SelectionDAG &DAG) {
626 // Assign locations to each value returned by this call.
627 SmallVector<CCValAssign, 16> RVLocs;
628 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
629 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
630 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
633 SmallVector<SDOperand, 8> ResultVals;
635 // Copy all of the result registers out of their specified physreg.
636 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
637 for (unsigned i = 0; i != RVLocs.size(); ++i) {
638 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
639 RVLocs[i].getValVT(), InFlag).getValue(1);
640 InFlag = Chain.getValue(2);
641 ResultVals.push_back(Chain.getValue(0));
644 // Copies from the FP stack are special, as ST0 isn't a valid register
645 // before the fp stackifier runs.
647 // Copy ST0 into an RFP register with FP_GET_RESULT.
648 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other, MVT::Flag);
649 SDOperand GROps[] = { Chain, InFlag };
650 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
651 Chain = RetVal.getValue(1);
652 InFlag = RetVal.getValue(2);
654 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
657 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
658 // shouldn't be necessary except that RFP cannot be live across
659 // multiple blocks. When stackifier is fixed, they can be uncoupled.
660 MachineFunction &MF = DAG.getMachineFunction();
661 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
662 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
664 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag
666 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
667 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
668 Chain = RetVal.getValue(1);
670 ResultVals.push_back(RetVal);
673 // Merge everything together with a MERGE_VALUES node.
674 ResultVals.push_back(Chain);
675 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
676 &ResultVals[0], ResultVals.size()).Val;
680 //===----------------------------------------------------------------------===//
681 // C & StdCall Calling Convention implementation
682 //===----------------------------------------------------------------------===//
683 // StdCall calling convention seems to be standard for many Windows' API
684 // routines and around. It differs from C calling convention just a little:
685 // callee should clean up the stack, not caller. Symbols should be also
686 // decorated in some fancy way :) It doesn't support any vector arguments.
688 /// AddLiveIn - This helper function adds the specified physical register to the
689 /// MachineFunction as a live in value. It also creates a corresponding virtual
691 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
692 const TargetRegisterClass *RC) {
693 assert(RC->contains(PReg) && "Not the correct regclass!");
694 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
695 MF.addLiveIn(PReg, VReg);
699 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
701 unsigned NumArgs = Op.Val->getNumValues() - 1;
702 MachineFunction &MF = DAG.getMachineFunction();
703 MachineFrameInfo *MFI = MF.getFrameInfo();
704 SDOperand Root = Op.getOperand(0);
705 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
707 // Assign locations to all of the incoming arguments.
708 SmallVector<CCValAssign, 16> ArgLocs;
709 CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
710 getTargetMachine(), ArgLocs);
711 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
713 SmallVector<SDOperand, 8> ArgValues;
714 unsigned LastVal = ~0U;
715 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
716 CCValAssign &VA = ArgLocs[i];
717 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
719 assert(VA.getValNo() != LastVal &&
720 "Don't support value assigned to multiple locs yet");
721 LastVal = VA.getValNo();
724 MVT::ValueType RegVT = VA.getLocVT();
725 TargetRegisterClass *RC;
726 if (RegVT == MVT::i32)
727 RC = X86::GR32RegisterClass;
729 assert(MVT::isVector(RegVT));
730 RC = X86::VR128RegisterClass;
733 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
734 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
736 // If this is an 8 or 16-bit value, it is really passed promoted to 32
737 // bits. Insert an assert[sz]ext to capture this, then truncate to the
739 if (VA.getLocInfo() == CCValAssign::SExt)
740 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
741 DAG.getValueType(VA.getValVT()));
742 else if (VA.getLocInfo() == CCValAssign::ZExt)
743 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
744 DAG.getValueType(VA.getValVT()));
746 if (VA.getLocInfo() != CCValAssign::Full)
747 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
749 ArgValues.push_back(ArgValue);
751 assert(VA.isMemLoc());
753 // Create the nodes corresponding to a load from this parameter slot.
754 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
755 VA.getLocMemOffset());
756 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
757 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
761 unsigned StackSize = CCInfo.getNextStackOffset();
763 ArgValues.push_back(Root);
765 // If the function takes variable number of arguments, make a frame index for
766 // the start of the first vararg value... for expansion of llvm.va_start.
768 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
770 if (isStdCall && !isVarArg) {
771 BytesToPopOnReturn = StackSize; // Callee pops everything..
772 BytesCallerReserves = 0;
774 BytesToPopOnReturn = 0; // Callee pops nothing.
776 // If this is an sret function, the return should pop the hidden pointer.
778 (cast<ConstantSDNode>(Op.getOperand(3))->getValue() &
779 ISD::ParamFlags::StructReturn))
780 BytesToPopOnReturn = 4;
782 BytesCallerReserves = StackSize;
785 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
786 ReturnAddrIndex = 0; // No return address slot generated yet.
788 MF.getInfo<X86MachineFunctionInfo>()
789 ->setBytesToPopOnReturn(BytesToPopOnReturn);
791 // Return the new list of results.
792 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
793 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
796 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
798 SDOperand Chain = Op.getOperand(0);
799 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
800 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
801 SDOperand Callee = Op.getOperand(4);
802 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
804 // Analyze operands of the call, assigning locations to each operand.
805 SmallVector<CCValAssign, 16> ArgLocs;
806 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
807 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
809 // Get a count of how many bytes are to be pushed on the stack.
810 unsigned NumBytes = CCInfo.getNextStackOffset();
812 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
814 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
815 SmallVector<SDOperand, 8> MemOpChains;
819 // Walk the register/memloc assignments, inserting copies/loads.
820 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
821 CCValAssign &VA = ArgLocs[i];
822 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
824 // Promote the value if needed.
825 switch (VA.getLocInfo()) {
826 default: assert(0 && "Unknown loc info!");
827 case CCValAssign::Full: break;
828 case CCValAssign::SExt:
829 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
831 case CCValAssign::ZExt:
832 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
834 case CCValAssign::AExt:
835 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
840 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
842 assert(VA.isMemLoc());
843 if (StackPtr.Val == 0)
844 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
845 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
846 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
847 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
851 // If the first argument is an sret pointer, remember it.
852 bool isSRet = NumOps &&
853 (cast<ConstantSDNode>(Op.getOperand(6))->getValue() &
854 ISD::ParamFlags::StructReturn);
856 if (!MemOpChains.empty())
857 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
858 &MemOpChains[0], MemOpChains.size());
860 // Build a sequence of copy-to-reg nodes chained together with token chain
861 // and flag operands which copy the outgoing args into registers.
863 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
864 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
866 InFlag = Chain.getValue(1);
869 // ELF / PIC requires GOT in the EBX register before function calls via PLT
871 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
872 Subtarget->isPICStyleGOT()) {
873 Chain = DAG.getCopyToReg(Chain, X86::EBX,
874 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
876 InFlag = Chain.getValue(1);
879 // If the callee is a GlobalAddress node (quite common, every direct call is)
880 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
881 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
882 // We should use extra load for direct calls to dllimported functions in
884 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
885 getTargetMachine(), true))
886 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
887 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
888 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
890 // Returns a chain & a flag for retval copy to use.
891 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
892 SmallVector<SDOperand, 8> Ops;
893 Ops.push_back(Chain);
894 Ops.push_back(Callee);
896 // Add argument registers to the end of the list so that they are known live
898 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
899 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
900 RegsToPass[i].second.getValueType()));
902 // Add an implicit use GOT pointer in EBX.
903 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
904 Subtarget->isPICStyleGOT())
905 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
908 Ops.push_back(InFlag);
910 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
911 NodeTys, &Ops[0], Ops.size());
912 InFlag = Chain.getValue(1);
914 // Create the CALLSEQ_END node.
915 unsigned NumBytesForCalleeToPush = 0;
917 if (CC == CallingConv::X86_StdCall) {
919 NumBytesForCalleeToPush = isSRet ? 4 : 0;
921 NumBytesForCalleeToPush = NumBytes;
923 // If this is is a call to a struct-return function, the callee
924 // pops the hidden struct pointer, so we have to push it back.
925 // This is common for Darwin/X86, Linux & Mingw32 targets.
926 NumBytesForCalleeToPush = isSRet ? 4 : 0;
929 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
931 Ops.push_back(Chain);
932 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
933 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
934 Ops.push_back(InFlag);
935 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
936 InFlag = Chain.getValue(1);
938 // Handle result values, copying them out of physregs into vregs that we
940 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
944 //===----------------------------------------------------------------------===//
945 // FastCall Calling Convention implementation
946 //===----------------------------------------------------------------------===//
948 // The X86 'fastcall' calling convention passes up to two integer arguments in
949 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
950 // and requires that the callee pop its arguments off the stack (allowing proper
951 // tail calls), and has the same return value conventions as C calling convs.
953 // This calling convention always arranges for the callee pop value to be 8n+4
954 // bytes, which is needed for tail recursion elimination and stack alignment
957 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
958 MachineFunction &MF = DAG.getMachineFunction();
959 MachineFrameInfo *MFI = MF.getFrameInfo();
960 SDOperand Root = Op.getOperand(0);
961 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
963 // Assign locations to all of the incoming arguments.
964 SmallVector<CCValAssign, 16> ArgLocs;
965 CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
966 getTargetMachine(), ArgLocs);
967 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
969 SmallVector<SDOperand, 8> ArgValues;
970 unsigned LastVal = ~0U;
971 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
972 CCValAssign &VA = ArgLocs[i];
973 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
975 assert(VA.getValNo() != LastVal &&
976 "Don't support value assigned to multiple locs yet");
977 LastVal = VA.getValNo();
980 MVT::ValueType RegVT = VA.getLocVT();
981 TargetRegisterClass *RC;
982 if (RegVT == MVT::i32)
983 RC = X86::GR32RegisterClass;
985 assert(MVT::isVector(RegVT));
986 RC = X86::VR128RegisterClass;
989 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
990 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
992 // If this is an 8 or 16-bit value, it is really passed promoted to 32
993 // bits. Insert an assert[sz]ext to capture this, then truncate to the
995 if (VA.getLocInfo() == CCValAssign::SExt)
996 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
997 DAG.getValueType(VA.getValVT()));
998 else if (VA.getLocInfo() == CCValAssign::ZExt)
999 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1000 DAG.getValueType(VA.getValVT()));
1002 if (VA.getLocInfo() != CCValAssign::Full)
1003 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1005 ArgValues.push_back(ArgValue);
1007 assert(VA.isMemLoc());
1009 // Create the nodes corresponding to a load from this parameter slot.
1010 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1011 VA.getLocMemOffset());
1012 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1013 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
1017 ArgValues.push_back(Root);
1019 unsigned StackSize = CCInfo.getNextStackOffset();
1021 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1022 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1023 // arguments and the arguments after the retaddr has been pushed are aligned.
1024 if ((StackSize & 7) == 0)
1028 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1029 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1030 ReturnAddrIndex = 0; // No return address slot generated yet.
1031 BytesToPopOnReturn = StackSize; // Callee pops all stack arguments.
1032 BytesCallerReserves = 0;
1034 MF.getInfo<X86MachineFunctionInfo>()
1035 ->setBytesToPopOnReturn(BytesToPopOnReturn);
1037 // Return the new list of results.
1038 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1039 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1042 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1044 SDOperand Chain = Op.getOperand(0);
1045 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1046 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1047 SDOperand Callee = Op.getOperand(4);
1049 // Analyze operands of the call, assigning locations to each operand.
1050 SmallVector<CCValAssign, 16> ArgLocs;
1051 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1052 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall);
1054 // Get a count of how many bytes are to be pushed on the stack.
1055 unsigned NumBytes = CCInfo.getNextStackOffset();
1057 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1058 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1059 // arguments and the arguments after the retaddr has been pushed are aligned.
1060 if ((NumBytes & 7) == 0)
1064 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1066 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1067 SmallVector<SDOperand, 8> MemOpChains;
1071 // Walk the register/memloc assignments, inserting copies/loads.
1072 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1073 CCValAssign &VA = ArgLocs[i];
1074 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1076 // Promote the value if needed.
1077 switch (VA.getLocInfo()) {
1078 default: assert(0 && "Unknown loc info!");
1079 case CCValAssign::Full: break;
1080 case CCValAssign::SExt:
1081 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1083 case CCValAssign::ZExt:
1084 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1086 case CCValAssign::AExt:
1087 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1091 if (VA.isRegLoc()) {
1092 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1094 assert(VA.isMemLoc());
1095 if (StackPtr.Val == 0)
1096 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1097 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1098 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1099 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1103 if (!MemOpChains.empty())
1104 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1105 &MemOpChains[0], MemOpChains.size());
1107 // Build a sequence of copy-to-reg nodes chained together with token chain
1108 // and flag operands which copy the outgoing args into registers.
1110 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1111 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1113 InFlag = Chain.getValue(1);
1116 // If the callee is a GlobalAddress node (quite common, every direct call is)
1117 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1118 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1119 // We should use extra load for direct calls to dllimported functions in
1121 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1122 getTargetMachine(), true))
1123 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1124 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1125 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1127 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1129 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1130 Subtarget->isPICStyleGOT()) {
1131 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1132 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1134 InFlag = Chain.getValue(1);
1137 // Returns a chain & a flag for retval copy to use.
1138 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1139 SmallVector<SDOperand, 8> Ops;
1140 Ops.push_back(Chain);
1141 Ops.push_back(Callee);
1143 // Add argument registers to the end of the list so that they are known live
1145 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1146 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1147 RegsToPass[i].second.getValueType()));
1149 // Add an implicit use GOT pointer in EBX.
1150 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1151 Subtarget->isPICStyleGOT())
1152 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1155 Ops.push_back(InFlag);
1157 // FIXME: Do not generate X86ISD::TAILCALL for now.
1158 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1159 NodeTys, &Ops[0], Ops.size());
1160 InFlag = Chain.getValue(1);
1162 // Returns a flag for retval copy to use.
1163 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1165 Ops.push_back(Chain);
1166 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1167 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1168 Ops.push_back(InFlag);
1169 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1170 InFlag = Chain.getValue(1);
1172 // Handle result values, copying them out of physregs into vregs that we
1174 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1178 //===----------------------------------------------------------------------===//
1179 // X86-64 C Calling Convention implementation
1180 //===----------------------------------------------------------------------===//
1183 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1184 MachineFunction &MF = DAG.getMachineFunction();
1185 MachineFrameInfo *MFI = MF.getFrameInfo();
1186 SDOperand Root = Op.getOperand(0);
1187 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1189 static const unsigned GPR64ArgRegs[] = {
1190 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1192 static const unsigned XMMArgRegs[] = {
1193 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1194 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1198 // Assign locations to all of the incoming arguments.
1199 SmallVector<CCValAssign, 16> ArgLocs;
1200 CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
1201 getTargetMachine(), ArgLocs);
1202 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1204 SmallVector<SDOperand, 8> ArgValues;
1205 unsigned LastVal = ~0U;
1206 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1207 CCValAssign &VA = ArgLocs[i];
1208 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1210 assert(VA.getValNo() != LastVal &&
1211 "Don't support value assigned to multiple locs yet");
1212 LastVal = VA.getValNo();
1214 if (VA.isRegLoc()) {
1215 MVT::ValueType RegVT = VA.getLocVT();
1216 TargetRegisterClass *RC;
1217 if (RegVT == MVT::i32)
1218 RC = X86::GR32RegisterClass;
1219 else if (RegVT == MVT::i64)
1220 RC = X86::GR64RegisterClass;
1221 else if (RegVT == MVT::f32)
1222 RC = X86::FR32RegisterClass;
1223 else if (RegVT == MVT::f64)
1224 RC = X86::FR64RegisterClass;
1226 assert(MVT::isVector(RegVT));
1227 if (MVT::getSizeInBits(RegVT) == 64) {
1228 RC = X86::GR64RegisterClass; // MMX values are passed in GPRs.
1231 RC = X86::VR128RegisterClass;
1234 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1235 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1237 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1238 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1240 if (VA.getLocInfo() == CCValAssign::SExt)
1241 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1242 DAG.getValueType(VA.getValVT()));
1243 else if (VA.getLocInfo() == CCValAssign::ZExt)
1244 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1245 DAG.getValueType(VA.getValVT()));
1247 if (VA.getLocInfo() != CCValAssign::Full)
1248 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1250 // Handle MMX values passed in GPRs.
1251 if (RegVT != VA.getLocVT() && RC == X86::GR64RegisterClass &&
1252 MVT::getSizeInBits(RegVT) == 64)
1253 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1255 ArgValues.push_back(ArgValue);
1257 assert(VA.isMemLoc());
1259 // Create the nodes corresponding to a load from this parameter slot.
1260 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1261 VA.getLocMemOffset());
1262 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1263 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
1267 unsigned StackSize = CCInfo.getNextStackOffset();
1269 // If the function takes variable number of arguments, make a frame index for
1270 // the start of the first vararg value... for expansion of llvm.va_start.
1272 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1273 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1275 // For X86-64, if there are vararg parameters that are passed via
1276 // registers, then we must store them to their spots on the stack so they
1277 // may be loaded by deferencing the result of va_next.
1278 VarArgsGPOffset = NumIntRegs * 8;
1279 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1280 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1281 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1283 // Store the integer parameter registers.
1284 SmallVector<SDOperand, 8> MemOps;
1285 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1286 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1287 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1288 for (; NumIntRegs != 6; ++NumIntRegs) {
1289 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1290 X86::GR64RegisterClass);
1291 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1292 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1293 MemOps.push_back(Store);
1294 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1295 DAG.getConstant(8, getPointerTy()));
1298 // Now store the XMM (fp + vector) parameter registers.
1299 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1300 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1301 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1302 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1303 X86::VR128RegisterClass);
1304 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1305 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1306 MemOps.push_back(Store);
1307 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1308 DAG.getConstant(16, getPointerTy()));
1310 if (!MemOps.empty())
1311 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1312 &MemOps[0], MemOps.size());
1315 ArgValues.push_back(Root);
1317 ReturnAddrIndex = 0; // No return address slot generated yet.
1318 BytesToPopOnReturn = 0; // Callee pops nothing.
1319 BytesCallerReserves = StackSize;
1321 // Return the new list of results.
1322 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1323 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1327 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1329 SDOperand Chain = Op.getOperand(0);
1330 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1331 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1332 SDOperand Callee = Op.getOperand(4);
1334 // Analyze operands of the call, assigning locations to each operand.
1335 SmallVector<CCValAssign, 16> ArgLocs;
1336 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1337 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1339 // Get a count of how many bytes are to be pushed on the stack.
1340 unsigned NumBytes = CCInfo.getNextStackOffset();
1341 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1343 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1344 SmallVector<SDOperand, 8> MemOpChains;
1348 // Walk the register/memloc assignments, inserting copies/loads.
1349 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1350 CCValAssign &VA = ArgLocs[i];
1351 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1353 // Promote the value if needed.
1354 switch (VA.getLocInfo()) {
1355 default: assert(0 && "Unknown loc info!");
1356 case CCValAssign::Full: break;
1357 case CCValAssign::SExt:
1358 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1360 case CCValAssign::ZExt:
1361 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1363 case CCValAssign::AExt:
1364 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1368 if (VA.isRegLoc()) {
1369 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1371 assert(VA.isMemLoc());
1372 if (StackPtr.Val == 0)
1373 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1374 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1375 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1376 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1380 if (!MemOpChains.empty())
1381 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1382 &MemOpChains[0], MemOpChains.size());
1384 // Build a sequence of copy-to-reg nodes chained together with token chain
1385 // and flag operands which copy the outgoing args into registers.
1387 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1388 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1390 InFlag = Chain.getValue(1);
1394 // From AMD64 ABI document:
1395 // For calls that may call functions that use varargs or stdargs
1396 // (prototype-less calls or calls to functions containing ellipsis (...) in
1397 // the declaration) %al is used as hidden argument to specify the number
1398 // of SSE registers used. The contents of %al do not need to match exactly
1399 // the number of registers, but must be an ubound on the number of SSE
1400 // registers used and is in the range 0 - 8 inclusive.
1402 // Count the number of XMM registers allocated.
1403 static const unsigned XMMArgRegs[] = {
1404 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1405 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1407 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1409 Chain = DAG.getCopyToReg(Chain, X86::AL,
1410 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1411 InFlag = Chain.getValue(1);
1414 // If the callee is a GlobalAddress node (quite common, every direct call is)
1415 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1416 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1417 // We should use extra load for direct calls to dllimported functions in
1419 if (getTargetMachine().getCodeModel() != CodeModel::Large
1420 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1421 getTargetMachine(), true))
1422 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1423 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1424 if (getTargetMachine().getCodeModel() != CodeModel::Large)
1425 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1427 // Returns a chain & a flag for retval copy to use.
1428 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1429 SmallVector<SDOperand, 8> Ops;
1430 Ops.push_back(Chain);
1431 Ops.push_back(Callee);
1433 // Add argument registers to the end of the list so that they are known live
1435 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1436 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1437 RegsToPass[i].second.getValueType()));
1440 Ops.push_back(InFlag);
1442 // FIXME: Do not generate X86ISD::TAILCALL for now.
1443 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1444 NodeTys, &Ops[0], Ops.size());
1445 InFlag = Chain.getValue(1);
1447 // Returns a flag for retval copy to use.
1448 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1450 Ops.push_back(Chain);
1451 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1452 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1453 Ops.push_back(InFlag);
1454 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1455 InFlag = Chain.getValue(1);
1457 // Handle result values, copying them out of physregs into vregs that we
1459 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1463 //===----------------------------------------------------------------------===//
1464 // Other Lowering Hooks
1465 //===----------------------------------------------------------------------===//
1468 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1469 if (ReturnAddrIndex == 0) {
1470 // Set up a frame object for the return address.
1471 MachineFunction &MF = DAG.getMachineFunction();
1472 if (Subtarget->is64Bit())
1473 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1475 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1478 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1483 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1484 /// specific condition code. It returns a false if it cannot do a direct
1485 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1487 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1488 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1489 SelectionDAG &DAG) {
1490 X86CC = X86::COND_INVALID;
1492 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1493 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1494 // X > -1 -> X == 0, jump !sign.
1495 RHS = DAG.getConstant(0, RHS.getValueType());
1496 X86CC = X86::COND_NS;
1498 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1499 // X < 0 -> X == 0, jump on sign.
1500 X86CC = X86::COND_S;
1505 switch (SetCCOpcode) {
1507 case ISD::SETEQ: X86CC = X86::COND_E; break;
1508 case ISD::SETGT: X86CC = X86::COND_G; break;
1509 case ISD::SETGE: X86CC = X86::COND_GE; break;
1510 case ISD::SETLT: X86CC = X86::COND_L; break;
1511 case ISD::SETLE: X86CC = X86::COND_LE; break;
1512 case ISD::SETNE: X86CC = X86::COND_NE; break;
1513 case ISD::SETULT: X86CC = X86::COND_B; break;
1514 case ISD::SETUGT: X86CC = X86::COND_A; break;
1515 case ISD::SETULE: X86CC = X86::COND_BE; break;
1516 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1519 // On a floating point condition, the flags are set as follows:
1521 // 0 | 0 | 0 | X > Y
1522 // 0 | 0 | 1 | X < Y
1523 // 1 | 0 | 0 | X == Y
1524 // 1 | 1 | 1 | unordered
1526 switch (SetCCOpcode) {
1529 case ISD::SETEQ: X86CC = X86::COND_E; break;
1530 case ISD::SETOLT: Flip = true; // Fallthrough
1532 case ISD::SETGT: X86CC = X86::COND_A; break;
1533 case ISD::SETOLE: Flip = true; // Fallthrough
1535 case ISD::SETGE: X86CC = X86::COND_AE; break;
1536 case ISD::SETUGT: Flip = true; // Fallthrough
1538 case ISD::SETLT: X86CC = X86::COND_B; break;
1539 case ISD::SETUGE: Flip = true; // Fallthrough
1541 case ISD::SETLE: X86CC = X86::COND_BE; break;
1543 case ISD::SETNE: X86CC = X86::COND_NE; break;
1544 case ISD::SETUO: X86CC = X86::COND_P; break;
1545 case ISD::SETO: X86CC = X86::COND_NP; break;
1548 std::swap(LHS, RHS);
1551 return X86CC != X86::COND_INVALID;
1554 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1555 /// code. Current x86 isa includes the following FP cmov instructions:
1556 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1557 static bool hasFPCMov(unsigned X86CC) {
1573 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1574 /// true if Op is undef or if its value falls within the specified range (L, H].
1575 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1576 if (Op.getOpcode() == ISD::UNDEF)
1579 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1580 return (Val >= Low && Val < Hi);
1583 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1584 /// true if Op is undef or if its value equal to the specified value.
1585 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1586 if (Op.getOpcode() == ISD::UNDEF)
1588 return cast<ConstantSDNode>(Op)->getValue() == Val;
1591 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1592 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1593 bool X86::isPSHUFDMask(SDNode *N) {
1594 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1596 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
1599 // Check if the value doesn't reference the second vector.
1600 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1601 SDOperand Arg = N->getOperand(i);
1602 if (Arg.getOpcode() == ISD::UNDEF) continue;
1603 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1604 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
1611 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1612 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1613 bool X86::isPSHUFHWMask(SDNode *N) {
1614 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1616 if (N->getNumOperands() != 8)
1619 // Lower quadword copied in order.
1620 for (unsigned i = 0; i != 4; ++i) {
1621 SDOperand Arg = N->getOperand(i);
1622 if (Arg.getOpcode() == ISD::UNDEF) continue;
1623 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1624 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1628 // Upper quadword shuffled.
1629 for (unsigned i = 4; i != 8; ++i) {
1630 SDOperand Arg = N->getOperand(i);
1631 if (Arg.getOpcode() == ISD::UNDEF) continue;
1632 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1633 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1634 if (Val < 4 || Val > 7)
1641 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1642 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1643 bool X86::isPSHUFLWMask(SDNode *N) {
1644 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1646 if (N->getNumOperands() != 8)
1649 // Upper quadword copied in order.
1650 for (unsigned i = 4; i != 8; ++i)
1651 if (!isUndefOrEqual(N->getOperand(i), i))
1654 // Lower quadword shuffled.
1655 for (unsigned i = 0; i != 4; ++i)
1656 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1662 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1663 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1664 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
1665 if (NumElems != 2 && NumElems != 4) return false;
1667 unsigned Half = NumElems / 2;
1668 for (unsigned i = 0; i < Half; ++i)
1669 if (!isUndefOrInRange(Elems[i], 0, NumElems))
1671 for (unsigned i = Half; i < NumElems; ++i)
1672 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
1678 bool X86::isSHUFPMask(SDNode *N) {
1679 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1680 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
1683 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
1684 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
1685 /// half elements to come from vector 1 (which would equal the dest.) and
1686 /// the upper half to come from vector 2.
1687 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
1688 if (NumOps != 2 && NumOps != 4) return false;
1690 unsigned Half = NumOps / 2;
1691 for (unsigned i = 0; i < Half; ++i)
1692 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
1694 for (unsigned i = Half; i < NumOps; ++i)
1695 if (!isUndefOrInRange(Ops[i], 0, NumOps))
1700 static bool isCommutedSHUFP(SDNode *N) {
1701 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1702 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
1705 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
1706 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
1707 bool X86::isMOVHLPSMask(SDNode *N) {
1708 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1710 if (N->getNumOperands() != 4)
1713 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
1714 return isUndefOrEqual(N->getOperand(0), 6) &&
1715 isUndefOrEqual(N->getOperand(1), 7) &&
1716 isUndefOrEqual(N->getOperand(2), 2) &&
1717 isUndefOrEqual(N->getOperand(3), 3);
1720 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
1721 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
1723 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
1724 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1726 if (N->getNumOperands() != 4)
1729 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
1730 return isUndefOrEqual(N->getOperand(0), 2) &&
1731 isUndefOrEqual(N->getOperand(1), 3) &&
1732 isUndefOrEqual(N->getOperand(2), 2) &&
1733 isUndefOrEqual(N->getOperand(3), 3);
1736 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
1737 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
1738 bool X86::isMOVLPMask(SDNode *N) {
1739 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1741 unsigned NumElems = N->getNumOperands();
1742 if (NumElems != 2 && NumElems != 4)
1745 for (unsigned i = 0; i < NumElems/2; ++i)
1746 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
1749 for (unsigned i = NumElems/2; i < NumElems; ++i)
1750 if (!isUndefOrEqual(N->getOperand(i), i))
1756 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
1757 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
1759 bool X86::isMOVHPMask(SDNode *N) {
1760 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1762 unsigned NumElems = N->getNumOperands();
1763 if (NumElems != 2 && NumElems != 4)
1766 for (unsigned i = 0; i < NumElems/2; ++i)
1767 if (!isUndefOrEqual(N->getOperand(i), i))
1770 for (unsigned i = 0; i < NumElems/2; ++i) {
1771 SDOperand Arg = N->getOperand(i + NumElems/2);
1772 if (!isUndefOrEqual(Arg, i + NumElems))
1779 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
1780 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
1781 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
1782 bool V2IsSplat = false) {
1783 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1786 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1787 SDOperand BitI = Elts[i];
1788 SDOperand BitI1 = Elts[i+1];
1789 if (!isUndefOrEqual(BitI, j))
1792 if (isUndefOrEqual(BitI1, NumElts))
1795 if (!isUndefOrEqual(BitI1, j + NumElts))
1803 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
1804 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1805 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1808 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
1809 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
1810 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
1811 bool V2IsSplat = false) {
1812 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1815 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1816 SDOperand BitI = Elts[i];
1817 SDOperand BitI1 = Elts[i+1];
1818 if (!isUndefOrEqual(BitI, j + NumElts/2))
1821 if (isUndefOrEqual(BitI1, NumElts))
1824 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
1832 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
1833 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1834 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1837 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
1838 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
1840 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
1841 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1843 unsigned NumElems = N->getNumOperands();
1844 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1847 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1848 SDOperand BitI = N->getOperand(i);
1849 SDOperand BitI1 = N->getOperand(i+1);
1851 if (!isUndefOrEqual(BitI, j))
1853 if (!isUndefOrEqual(BitI1, j))
1860 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
1861 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
1863 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
1864 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1866 unsigned NumElems = N->getNumOperands();
1867 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1870 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
1871 SDOperand BitI = N->getOperand(i);
1872 SDOperand BitI1 = N->getOperand(i + 1);
1874 if (!isUndefOrEqual(BitI, j))
1876 if (!isUndefOrEqual(BitI1, j))
1883 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
1884 /// specifies a shuffle of elements that is suitable for input to MOVSS,
1885 /// MOVSD, and MOVD, i.e. setting the lowest element.
1886 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
1887 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1890 if (!isUndefOrEqual(Elts[0], NumElts))
1893 for (unsigned i = 1; i < NumElts; ++i) {
1894 if (!isUndefOrEqual(Elts[i], i))
1901 bool X86::isMOVLMask(SDNode *N) {
1902 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1903 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
1906 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
1907 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
1908 /// element of vector 2 and the other elements to come from vector 1 in order.
1909 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
1910 bool V2IsSplat = false,
1911 bool V2IsUndef = false) {
1912 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
1915 if (!isUndefOrEqual(Ops[0], 0))
1918 for (unsigned i = 1; i < NumOps; ++i) {
1919 SDOperand Arg = Ops[i];
1920 if (!(isUndefOrEqual(Arg, i+NumOps) ||
1921 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
1922 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
1929 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
1930 bool V2IsUndef = false) {
1931 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1932 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
1933 V2IsSplat, V2IsUndef);
1936 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1937 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
1938 bool X86::isMOVSHDUPMask(SDNode *N) {
1939 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1941 if (N->getNumOperands() != 4)
1944 // Expect 1, 1, 3, 3
1945 for (unsigned i = 0; i < 2; ++i) {
1946 SDOperand Arg = N->getOperand(i);
1947 if (Arg.getOpcode() == ISD::UNDEF) continue;
1948 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1949 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1950 if (Val != 1) return false;
1954 for (unsigned i = 2; i < 4; ++i) {
1955 SDOperand Arg = N->getOperand(i);
1956 if (Arg.getOpcode() == ISD::UNDEF) continue;
1957 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1958 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1959 if (Val != 3) return false;
1963 // Don't use movshdup if it can be done with a shufps.
1967 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1968 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
1969 bool X86::isMOVSLDUPMask(SDNode *N) {
1970 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1972 if (N->getNumOperands() != 4)
1975 // Expect 0, 0, 2, 2
1976 for (unsigned i = 0; i < 2; ++i) {
1977 SDOperand Arg = N->getOperand(i);
1978 if (Arg.getOpcode() == ISD::UNDEF) continue;
1979 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1980 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1981 if (Val != 0) return false;
1985 for (unsigned i = 2; i < 4; ++i) {
1986 SDOperand Arg = N->getOperand(i);
1987 if (Arg.getOpcode() == ISD::UNDEF) continue;
1988 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1989 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1990 if (Val != 2) return false;
1994 // Don't use movshdup if it can be done with a shufps.
1998 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
1999 /// specifies a identity operation on the LHS or RHS.
2000 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2001 unsigned NumElems = N->getNumOperands();
2002 for (unsigned i = 0; i < NumElems; ++i)
2003 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2008 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2009 /// a splat of a single element.
2010 static bool isSplatMask(SDNode *N) {
2011 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2013 // This is a splat operation if each element of the permute is the same, and
2014 // if the value doesn't reference the second vector.
2015 unsigned NumElems = N->getNumOperands();
2016 SDOperand ElementBase;
2018 for (; i != NumElems; ++i) {
2019 SDOperand Elt = N->getOperand(i);
2020 if (isa<ConstantSDNode>(Elt)) {
2026 if (!ElementBase.Val)
2029 for (; i != NumElems; ++i) {
2030 SDOperand Arg = N->getOperand(i);
2031 if (Arg.getOpcode() == ISD::UNDEF) continue;
2032 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2033 if (Arg != ElementBase) return false;
2036 // Make sure it is a splat of the first vector operand.
2037 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2040 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2041 /// a splat of a single element and it's a 2 or 4 element mask.
2042 bool X86::isSplatMask(SDNode *N) {
2043 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2045 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2046 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2048 return ::isSplatMask(N);
2051 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2052 /// specifies a splat of zero element.
2053 bool X86::isSplatLoMask(SDNode *N) {
2054 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2056 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2057 if (!isUndefOrEqual(N->getOperand(i), 0))
2062 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2063 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2065 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2066 unsigned NumOperands = N->getNumOperands();
2067 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2069 for (unsigned i = 0; i < NumOperands; ++i) {
2071 SDOperand Arg = N->getOperand(NumOperands-i-1);
2072 if (Arg.getOpcode() != ISD::UNDEF)
2073 Val = cast<ConstantSDNode>(Arg)->getValue();
2074 if (Val >= NumOperands) Val -= NumOperands;
2076 if (i != NumOperands - 1)
2083 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2084 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2086 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2088 // 8 nodes, but we only care about the last 4.
2089 for (unsigned i = 7; i >= 4; --i) {
2091 SDOperand Arg = N->getOperand(i);
2092 if (Arg.getOpcode() != ISD::UNDEF)
2093 Val = cast<ConstantSDNode>(Arg)->getValue();
2102 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2103 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2105 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2107 // 8 nodes, but we only care about the first 4.
2108 for (int i = 3; i >= 0; --i) {
2110 SDOperand Arg = N->getOperand(i);
2111 if (Arg.getOpcode() != ISD::UNDEF)
2112 Val = cast<ConstantSDNode>(Arg)->getValue();
2121 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2122 /// specifies a 8 element shuffle that can be broken into a pair of
2123 /// PSHUFHW and PSHUFLW.
2124 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2125 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2127 if (N->getNumOperands() != 8)
2130 // Lower quadword shuffled.
2131 for (unsigned i = 0; i != 4; ++i) {
2132 SDOperand Arg = N->getOperand(i);
2133 if (Arg.getOpcode() == ISD::UNDEF) continue;
2134 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2135 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2140 // Upper quadword shuffled.
2141 for (unsigned i = 4; i != 8; ++i) {
2142 SDOperand Arg = N->getOperand(i);
2143 if (Arg.getOpcode() == ISD::UNDEF) continue;
2144 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2145 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2146 if (Val < 4 || Val > 7)
2153 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2154 /// values in ther permute mask.
2155 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2156 SDOperand &V2, SDOperand &Mask,
2157 SelectionDAG &DAG) {
2158 MVT::ValueType VT = Op.getValueType();
2159 MVT::ValueType MaskVT = Mask.getValueType();
2160 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2161 unsigned NumElems = Mask.getNumOperands();
2162 SmallVector<SDOperand, 8> MaskVec;
2164 for (unsigned i = 0; i != NumElems; ++i) {
2165 SDOperand Arg = Mask.getOperand(i);
2166 if (Arg.getOpcode() == ISD::UNDEF) {
2167 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2170 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2171 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2173 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2175 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2179 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2180 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2183 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2184 /// match movhlps. The lower half elements should come from upper half of
2185 /// V1 (and in order), and the upper half elements should come from the upper
2186 /// half of V2 (and in order).
2187 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2188 unsigned NumElems = Mask->getNumOperands();
2191 for (unsigned i = 0, e = 2; i != e; ++i)
2192 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2194 for (unsigned i = 2; i != 4; ++i)
2195 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2200 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2201 /// is promoted to a vector.
2202 static inline bool isScalarLoadToVector(SDNode *N) {
2203 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2204 N = N->getOperand(0).Val;
2205 return ISD::isNON_EXTLoad(N);
2210 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2211 /// match movlp{s|d}. The lower half elements should come from lower half of
2212 /// V1 (and in order), and the upper half elements should come from the upper
2213 /// half of V2 (and in order). And since V1 will become the source of the
2214 /// MOVLP, it must be either a vector load or a scalar load to vector.
2215 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2216 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2218 // Is V2 is a vector load, don't do this transformation. We will try to use
2219 // load folding shufps op.
2220 if (ISD::isNON_EXTLoad(V2))
2223 unsigned NumElems = Mask->getNumOperands();
2224 if (NumElems != 2 && NumElems != 4)
2226 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2227 if (!isUndefOrEqual(Mask->getOperand(i), i))
2229 for (unsigned i = NumElems/2; i != NumElems; ++i)
2230 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2235 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2237 static bool isSplatVector(SDNode *N) {
2238 if (N->getOpcode() != ISD::BUILD_VECTOR)
2241 SDOperand SplatValue = N->getOperand(0);
2242 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2243 if (N->getOperand(i) != SplatValue)
2248 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2250 static bool isUndefShuffle(SDNode *N) {
2251 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2254 SDOperand V1 = N->getOperand(0);
2255 SDOperand V2 = N->getOperand(1);
2256 SDOperand Mask = N->getOperand(2);
2257 unsigned NumElems = Mask.getNumOperands();
2258 for (unsigned i = 0; i != NumElems; ++i) {
2259 SDOperand Arg = Mask.getOperand(i);
2260 if (Arg.getOpcode() != ISD::UNDEF) {
2261 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2262 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2264 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2271 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2273 static inline bool isZeroNode(SDOperand Elt) {
2274 return ((isa<ConstantSDNode>(Elt) &&
2275 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2276 (isa<ConstantFPSDNode>(Elt) &&
2277 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2280 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2281 /// to an zero vector.
2282 static bool isZeroShuffle(SDNode *N) {
2283 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2286 SDOperand V1 = N->getOperand(0);
2287 SDOperand V2 = N->getOperand(1);
2288 SDOperand Mask = N->getOperand(2);
2289 unsigned NumElems = Mask.getNumOperands();
2290 for (unsigned i = 0; i != NumElems; ++i) {
2291 SDOperand Arg = Mask.getOperand(i);
2292 if (Arg.getOpcode() != ISD::UNDEF) {
2293 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2294 if (Idx < NumElems) {
2295 unsigned Opc = V1.Val->getOpcode();
2296 if (Opc == ISD::UNDEF)
2298 if (Opc != ISD::BUILD_VECTOR ||
2299 !isZeroNode(V1.Val->getOperand(Idx)))
2301 } else if (Idx >= NumElems) {
2302 unsigned Opc = V2.Val->getOpcode();
2303 if (Opc == ISD::UNDEF)
2305 if (Opc != ISD::BUILD_VECTOR ||
2306 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2314 /// getZeroVector - Returns a vector of specified type with all zero elements.
2316 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2317 assert(MVT::isVector(VT) && "Expected a vector type");
2318 unsigned NumElems = MVT::getVectorNumElements(VT);
2319 MVT::ValueType EVT = MVT::getVectorElementType(VT);
2320 bool isFP = MVT::isFloatingPoint(EVT);
2321 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2322 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2323 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2326 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2327 /// that point to V2 points to its first element.
2328 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2329 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2331 bool Changed = false;
2332 SmallVector<SDOperand, 8> MaskVec;
2333 unsigned NumElems = Mask.getNumOperands();
2334 for (unsigned i = 0; i != NumElems; ++i) {
2335 SDOperand Arg = Mask.getOperand(i);
2336 if (Arg.getOpcode() != ISD::UNDEF) {
2337 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2338 if (Val > NumElems) {
2339 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2343 MaskVec.push_back(Arg);
2347 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2348 &MaskVec[0], MaskVec.size());
2352 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2353 /// operation of specified width.
2354 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2355 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2356 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2358 SmallVector<SDOperand, 8> MaskVec;
2359 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2360 for (unsigned i = 1; i != NumElems; ++i)
2361 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2362 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2365 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2366 /// of specified width.
2367 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2368 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2369 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2370 SmallVector<SDOperand, 8> MaskVec;
2371 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2372 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2373 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2375 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2378 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2379 /// of specified width.
2380 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2381 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2382 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2383 unsigned Half = NumElems/2;
2384 SmallVector<SDOperand, 8> MaskVec;
2385 for (unsigned i = 0; i != Half; ++i) {
2386 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2387 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2389 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2392 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2394 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2395 SDOperand V1 = Op.getOperand(0);
2396 SDOperand Mask = Op.getOperand(2);
2397 MVT::ValueType VT = Op.getValueType();
2398 unsigned NumElems = Mask.getNumOperands();
2399 Mask = getUnpacklMask(NumElems, DAG);
2400 while (NumElems != 4) {
2401 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2404 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2406 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2407 Mask = getZeroVector(MaskVT, DAG);
2408 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2409 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2410 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2413 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2414 /// vector of zero or undef vector.
2415 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2416 unsigned NumElems, unsigned Idx,
2417 bool isZero, SelectionDAG &DAG) {
2418 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2419 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2420 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
2421 SDOperand Zero = DAG.getConstant(0, EVT);
2422 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2423 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2424 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2425 &MaskVec[0], MaskVec.size());
2426 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2429 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2431 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2432 unsigned NumNonZero, unsigned NumZero,
2433 SelectionDAG &DAG, TargetLowering &TLI) {
2439 for (unsigned i = 0; i < 16; ++i) {
2440 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2441 if (ThisIsNonZero && First) {
2443 V = getZeroVector(MVT::v8i16, DAG);
2445 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2450 SDOperand ThisElt(0, 0), LastElt(0, 0);
2451 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2452 if (LastIsNonZero) {
2453 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2455 if (ThisIsNonZero) {
2456 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2457 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2458 ThisElt, DAG.getConstant(8, MVT::i8));
2460 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2465 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2466 DAG.getConstant(i/2, TLI.getPointerTy()));
2470 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2473 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
2475 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2476 unsigned NumNonZero, unsigned NumZero,
2477 SelectionDAG &DAG, TargetLowering &TLI) {
2483 for (unsigned i = 0; i < 8; ++i) {
2484 bool isNonZero = (NonZeros & (1 << i)) != 0;
2488 V = getZeroVector(MVT::v8i16, DAG);
2490 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2493 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2494 DAG.getConstant(i, TLI.getPointerTy()));
2502 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2503 // All zero's are handled with pxor.
2504 if (ISD::isBuildVectorAllZeros(Op.Val))
2507 // All one's are handled with pcmpeqd.
2508 if (ISD::isBuildVectorAllOnes(Op.Val))
2511 MVT::ValueType VT = Op.getValueType();
2512 MVT::ValueType EVT = MVT::getVectorElementType(VT);
2513 unsigned EVTBits = MVT::getSizeInBits(EVT);
2515 unsigned NumElems = Op.getNumOperands();
2516 unsigned NumZero = 0;
2517 unsigned NumNonZero = 0;
2518 unsigned NonZeros = 0;
2519 unsigned NumNonZeroImms = 0;
2520 std::set<SDOperand> Values;
2521 for (unsigned i = 0; i < NumElems; ++i) {
2522 SDOperand Elt = Op.getOperand(i);
2523 if (Elt.getOpcode() != ISD::UNDEF) {
2525 if (isZeroNode(Elt))
2528 NonZeros |= (1 << i);
2530 if (Elt.getOpcode() == ISD::Constant ||
2531 Elt.getOpcode() == ISD::ConstantFP)
2537 if (NumNonZero == 0) {
2539 // All undef vector. Return an UNDEF.
2540 return DAG.getNode(ISD::UNDEF, VT);
2542 // A mix of zero and undef. Return a zero vector.
2543 return getZeroVector(VT, DAG);
2546 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2547 if (Values.size() == 1)
2550 // Special case for single non-zero element.
2551 if (NumNonZero == 1) {
2552 unsigned Idx = CountTrailingZeros_32(NonZeros);
2553 SDOperand Item = Op.getOperand(Idx);
2554 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2556 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2557 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2560 if (EVTBits == 32) {
2561 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2562 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2564 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2565 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
2566 SmallVector<SDOperand, 8> MaskVec;
2567 for (unsigned i = 0; i < NumElems; i++)
2568 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2569 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2570 &MaskVec[0], MaskVec.size());
2571 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2572 DAG.getNode(ISD::UNDEF, VT), Mask);
2576 // A vector full of immediates; various special cases are already
2577 // handled, so this is best done with a single constant-pool load.
2578 if (NumNonZero == NumNonZeroImms)
2581 // Let legalizer expand 2-wide build_vectors.
2585 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2586 if (EVTBits == 8 && NumElems == 16) {
2587 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2589 if (V.Val) return V;
2592 if (EVTBits == 16 && NumElems == 8) {
2593 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2595 if (V.Val) return V;
2598 // If element VT is == 32 bits, turn it into a number of shuffles.
2599 SmallVector<SDOperand, 8> V;
2601 if (NumElems == 4 && NumZero > 0) {
2602 for (unsigned i = 0; i < 4; ++i) {
2603 bool isZero = !(NonZeros & (1 << i));
2605 V[i] = getZeroVector(VT, DAG);
2607 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2610 for (unsigned i = 0; i < 2; ++i) {
2611 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2614 V[i] = V[i*2]; // Must be a zero vector.
2617 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2618 getMOVLMask(NumElems, DAG));
2621 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2622 getMOVLMask(NumElems, DAG));
2625 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2626 getUnpacklMask(NumElems, DAG));
2631 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2632 // clears the upper bits.
2633 // FIXME: we can do the same for v4f32 case when we know both parts of
2634 // the lower half come from scalar_to_vector (loadf32). We should do
2635 // that in post legalizer dag combiner with target specific hooks.
2636 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2638 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2639 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
2640 SmallVector<SDOperand, 8> MaskVec;
2641 bool Reverse = (NonZeros & 0x3) == 2;
2642 for (unsigned i = 0; i < 2; ++i)
2644 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2646 MaskVec.push_back(DAG.getConstant(i, EVT));
2647 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2648 for (unsigned i = 0; i < 2; ++i)
2650 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2652 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2653 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2654 &MaskVec[0], MaskVec.size());
2655 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2658 if (Values.size() > 2) {
2659 // Expand into a number of unpckl*.
2661 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2662 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2663 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2664 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2665 for (unsigned i = 0; i < NumElems; ++i)
2666 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2668 while (NumElems != 0) {
2669 for (unsigned i = 0; i < NumElems; ++i)
2670 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
2681 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2682 SDOperand V1 = Op.getOperand(0);
2683 SDOperand V2 = Op.getOperand(1);
2684 SDOperand PermMask = Op.getOperand(2);
2685 MVT::ValueType VT = Op.getValueType();
2686 unsigned NumElems = PermMask.getNumOperands();
2687 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
2688 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
2689 bool V1IsSplat = false;
2690 bool V2IsSplat = false;
2692 if (isUndefShuffle(Op.Val))
2693 return DAG.getNode(ISD::UNDEF, VT);
2695 if (isZeroShuffle(Op.Val))
2696 return getZeroVector(VT, DAG);
2698 if (isIdentityMask(PermMask.Val))
2700 else if (isIdentityMask(PermMask.Val, true))
2703 if (isSplatMask(PermMask.Val)) {
2704 if (NumElems <= 4) return Op;
2705 // Promote it to a v4i32 splat.
2706 return PromoteSplat(Op, DAG);
2709 if (X86::isMOVLMask(PermMask.Val))
2710 return (V1IsUndef) ? V2 : Op;
2712 if (X86::isMOVSHDUPMask(PermMask.Val) ||
2713 X86::isMOVSLDUPMask(PermMask.Val) ||
2714 X86::isMOVHLPSMask(PermMask.Val) ||
2715 X86::isMOVHPMask(PermMask.Val) ||
2716 X86::isMOVLPMask(PermMask.Val))
2719 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2720 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
2721 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2723 bool Commuted = false;
2724 V1IsSplat = isSplatVector(V1.Val);
2725 V2IsSplat = isSplatVector(V2.Val);
2726 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
2727 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2728 std::swap(V1IsSplat, V2IsSplat);
2729 std::swap(V1IsUndef, V2IsUndef);
2733 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
2734 if (V2IsUndef) return V1;
2735 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2737 // V2 is a splat, so the mask may be malformed. That is, it may point
2738 // to any V2 element. The instruction selectior won't like this. Get
2739 // a corrected mask and commute to form a proper MOVS{S|D}.
2740 SDOperand NewMask = getMOVLMask(NumElems, DAG);
2741 if (NewMask.Val != PermMask.Val)
2742 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2747 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2748 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
2749 X86::isUNPCKLMask(PermMask.Val) ||
2750 X86::isUNPCKHMask(PermMask.Val))
2754 // Normalize mask so all entries that point to V2 points to its first
2755 // element then try to match unpck{h|l} again. If match, return a
2756 // new vector_shuffle with the corrected mask.
2757 SDOperand NewMask = NormalizeMask(PermMask, DAG);
2758 if (NewMask.Val != PermMask.Val) {
2759 if (X86::isUNPCKLMask(PermMask.Val, true)) {
2760 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
2761 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2762 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
2763 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
2764 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2769 // Normalize the node to match x86 shuffle ops if needed
2770 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
2771 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2774 // Commute is back and try unpck* again.
2775 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2776 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2777 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
2778 X86::isUNPCKLMask(PermMask.Val) ||
2779 X86::isUNPCKHMask(PermMask.Val))
2783 // If VT is integer, try PSHUF* first, then SHUFP*.
2784 if (MVT::isInteger(VT)) {
2785 // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
2786 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
2787 if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
2788 X86::isPSHUFDMask(PermMask.Val)) ||
2789 X86::isPSHUFHWMask(PermMask.Val) ||
2790 X86::isPSHUFLWMask(PermMask.Val)) {
2791 if (V2.getOpcode() != ISD::UNDEF)
2792 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2793 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2797 if (X86::isSHUFPMask(PermMask.Val) &&
2798 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
2801 // Handle v8i16 shuffle high / low shuffle node pair.
2802 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
2803 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2804 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2805 SmallVector<SDOperand, 8> MaskVec;
2806 for (unsigned i = 0; i != 4; ++i)
2807 MaskVec.push_back(PermMask.getOperand(i));
2808 for (unsigned i = 4; i != 8; ++i)
2809 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2810 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2811 &MaskVec[0], MaskVec.size());
2812 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2814 for (unsigned i = 0; i != 4; ++i)
2815 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2816 for (unsigned i = 4; i != 8; ++i)
2817 MaskVec.push_back(PermMask.getOperand(i));
2818 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
2819 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2822 // Floating point cases in the other order.
2823 if (X86::isSHUFPMask(PermMask.Val))
2825 if (X86::isPSHUFDMask(PermMask.Val) ||
2826 X86::isPSHUFHWMask(PermMask.Val) ||
2827 X86::isPSHUFLWMask(PermMask.Val)) {
2828 if (V2.getOpcode() != ISD::UNDEF)
2829 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2830 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2835 if (NumElems == 4 &&
2836 // Don't do this for MMX.
2837 MVT::getSizeInBits(VT) != 64) {
2838 MVT::ValueType MaskVT = PermMask.getValueType();
2839 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
2840 SmallVector<std::pair<int, int>, 8> Locs;
2841 Locs.reserve(NumElems);
2842 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2843 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2846 // If no more than two elements come from either vector. This can be
2847 // implemented with two shuffles. First shuffle gather the elements.
2848 // The second shuffle, which takes the first shuffle as both of its
2849 // vector operands, put the elements into the right order.
2850 for (unsigned i = 0; i != NumElems; ++i) {
2851 SDOperand Elt = PermMask.getOperand(i);
2852 if (Elt.getOpcode() == ISD::UNDEF) {
2853 Locs[i] = std::make_pair(-1, -1);
2855 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
2856 if (Val < NumElems) {
2857 Locs[i] = std::make_pair(0, NumLo);
2861 Locs[i] = std::make_pair(1, NumHi);
2862 if (2+NumHi < NumElems)
2863 Mask1[2+NumHi] = Elt;
2868 if (NumLo <= 2 && NumHi <= 2) {
2869 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2870 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2871 &Mask1[0], Mask1.size()));
2872 for (unsigned i = 0; i != NumElems; ++i) {
2873 if (Locs[i].first == -1)
2876 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
2877 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
2878 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
2882 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
2883 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2884 &Mask2[0], Mask2.size()));
2887 // Break it into (shuffle shuffle_hi, shuffle_lo).
2889 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2890 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2891 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
2892 unsigned MaskIdx = 0;
2894 unsigned HiIdx = NumElems/2;
2895 for (unsigned i = 0; i != NumElems; ++i) {
2896 if (i == NumElems/2) {
2902 SDOperand Elt = PermMask.getOperand(i);
2903 if (Elt.getOpcode() == ISD::UNDEF) {
2904 Locs[i] = std::make_pair(-1, -1);
2905 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
2906 Locs[i] = std::make_pair(MaskIdx, LoIdx);
2907 (*MaskPtr)[LoIdx] = Elt;
2910 Locs[i] = std::make_pair(MaskIdx, HiIdx);
2911 (*MaskPtr)[HiIdx] = Elt;
2916 SDOperand LoShuffle =
2917 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2918 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2919 &LoMask[0], LoMask.size()));
2920 SDOperand HiShuffle =
2921 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2922 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2923 &HiMask[0], HiMask.size()));
2924 SmallVector<SDOperand, 8> MaskOps;
2925 for (unsigned i = 0; i != NumElems; ++i) {
2926 if (Locs[i].first == -1) {
2927 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
2929 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
2930 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
2933 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
2934 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2935 &MaskOps[0], MaskOps.size()));
2942 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2943 if (!isa<ConstantSDNode>(Op.getOperand(1)))
2946 MVT::ValueType VT = Op.getValueType();
2947 // TODO: handle v16i8.
2948 if (MVT::getSizeInBits(VT) == 16) {
2949 // Transform it so it match pextrw which produces a 32-bit result.
2950 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
2951 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
2952 Op.getOperand(0), Op.getOperand(1));
2953 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
2954 DAG.getValueType(VT));
2955 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
2956 } else if (MVT::getSizeInBits(VT) == 32) {
2957 SDOperand Vec = Op.getOperand(0);
2958 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2961 // SHUFPS the element to the lowest double word, then movss.
2962 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2963 SmallVector<SDOperand, 8> IdxVec;
2964 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
2965 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
2966 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
2967 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
2968 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2969 &IdxVec[0], IdxVec.size());
2970 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2971 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2972 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2973 DAG.getConstant(0, getPointerTy()));
2974 } else if (MVT::getSizeInBits(VT) == 64) {
2975 SDOperand Vec = Op.getOperand(0);
2976 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2980 // UNPCKHPD the element to the lowest double word, then movsd.
2981 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
2982 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
2983 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2984 SmallVector<SDOperand, 8> IdxVec;
2985 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
2986 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
2987 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2988 &IdxVec[0], IdxVec.size());
2989 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2990 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2991 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2992 DAG.getConstant(0, getPointerTy()));
2999 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3000 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3001 // as its second argument.
3002 MVT::ValueType VT = Op.getValueType();
3003 MVT::ValueType BaseVT = MVT::getVectorElementType(VT);
3004 SDOperand N0 = Op.getOperand(0);
3005 SDOperand N1 = Op.getOperand(1);
3006 SDOperand N2 = Op.getOperand(2);
3007 if (MVT::getSizeInBits(BaseVT) == 16) {
3008 if (N1.getValueType() != MVT::i32)
3009 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3010 if (N2.getValueType() != MVT::i32)
3011 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(),getPointerTy());
3012 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3013 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3014 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3017 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3018 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3019 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3020 SmallVector<SDOperand, 8> MaskVec;
3021 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3022 for (unsigned i = 1; i <= 3; ++i)
3023 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3024 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3025 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3026 &MaskVec[0], MaskVec.size()));
3028 // Use two pinsrw instructions to insert a 32 bit value.
3030 if (MVT::isFloatingPoint(N1.getValueType())) {
3031 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3032 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3033 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3034 DAG.getConstant(0, getPointerTy()));
3036 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3037 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3038 DAG.getConstant(Idx, getPointerTy()));
3039 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3040 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3041 DAG.getConstant(Idx+1, getPointerTy()));
3042 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3050 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3051 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3052 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3055 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3056 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3057 // one of the above mentioned nodes. It has to be wrapped because otherwise
3058 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3059 // be used to form addressing mode. These wrapped nodes will be selected
3062 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3063 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3064 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3066 CP->getAlignment());
3067 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3068 // With PIC, the address is actually $g + Offset.
3069 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3070 !Subtarget->isPICStyleRIPRel()) {
3071 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3072 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3080 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3081 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3082 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3083 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3084 // With PIC, the address is actually $g + Offset.
3085 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3086 !Subtarget->isPICStyleRIPRel()) {
3087 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3088 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3092 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3093 // load the value at address GV, not the value of GV itself. This means that
3094 // the GlobalAddress must be in the base or index register of the address, not
3095 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3096 // The same applies for external symbols during PIC codegen
3097 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3098 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3103 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3105 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3106 const MVT::ValueType PtrVT) {
3108 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
3109 DAG.getNode(X86ISD::GlobalBaseReg,
3111 InFlag = Chain.getValue(1);
3113 // emit leal symbol@TLSGD(,%ebx,1), %eax
3114 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3115 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3116 GA->getValueType(0),
3118 SDOperand Ops[] = { Chain, TGA, InFlag };
3119 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3120 InFlag = Result.getValue(2);
3121 Chain = Result.getValue(1);
3123 // call ___tls_get_addr. This function receives its argument in
3124 // the register EAX.
3125 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
3126 InFlag = Chain.getValue(1);
3128 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3129 SDOperand Ops1[] = { Chain,
3130 DAG.getTargetExternalSymbol("___tls_get_addr",
3132 DAG.getRegister(X86::EAX, PtrVT),
3133 DAG.getRegister(X86::EBX, PtrVT),
3135 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
3136 InFlag = Chain.getValue(1);
3138 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
3141 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
3142 // "local exec" model.
3144 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3145 const MVT::ValueType PtrVT) {
3146 // Get the Thread Pointer
3147 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
3148 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
3150 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3151 GA->getValueType(0),
3153 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
3155 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
3156 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset, NULL, 0);
3158 // The address of the thread local variable is the add of the thread
3159 // pointer with the offset of the variable.
3160 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
3164 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
3165 // TODO: implement the "local dynamic" model
3166 // TODO: implement the "initial exec"model for pic executables
3167 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
3168 "TLS not implemented for non-ELF and 64-bit targets");
3169 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3170 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
3171 // otherwise use the "Local Exec"TLS Model
3172 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
3173 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
3175 return LowerToTLSExecModel(GA, DAG, getPointerTy());
3179 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3180 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3181 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3182 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3183 // With PIC, the address is actually $g + Offset.
3184 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3185 !Subtarget->isPICStyleRIPRel()) {
3186 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3187 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3194 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3195 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3196 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3197 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3198 // With PIC, the address is actually $g + Offset.
3199 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3200 !Subtarget->isPICStyleRIPRel()) {
3201 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3202 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3209 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3210 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3211 "Not an i64 shift!");
3212 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3213 SDOperand ShOpLo = Op.getOperand(0);
3214 SDOperand ShOpHi = Op.getOperand(1);
3215 SDOperand ShAmt = Op.getOperand(2);
3216 SDOperand Tmp1 = isSRA ?
3217 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3218 DAG.getConstant(0, MVT::i32);
3220 SDOperand Tmp2, Tmp3;
3221 if (Op.getOpcode() == ISD::SHL_PARTS) {
3222 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3223 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3225 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3226 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3229 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3230 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3231 DAG.getConstant(32, MVT::i8));
3232 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
3233 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
3236 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3238 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3239 SmallVector<SDOperand, 4> Ops;
3240 if (Op.getOpcode() == ISD::SHL_PARTS) {
3241 Ops.push_back(Tmp2);
3242 Ops.push_back(Tmp3);
3244 Ops.push_back(InFlag);
3245 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3246 InFlag = Hi.getValue(1);
3249 Ops.push_back(Tmp3);
3250 Ops.push_back(Tmp1);
3252 Ops.push_back(InFlag);
3253 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3255 Ops.push_back(Tmp2);
3256 Ops.push_back(Tmp3);
3258 Ops.push_back(InFlag);
3259 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3260 InFlag = Lo.getValue(1);
3263 Ops.push_back(Tmp3);
3264 Ops.push_back(Tmp1);
3266 Ops.push_back(InFlag);
3267 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3270 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3274 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3277 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3278 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3279 Op.getOperand(0).getValueType() >= MVT::i16 &&
3280 "Unknown SINT_TO_FP to lower!");
3283 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3284 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3285 MachineFunction &MF = DAG.getMachineFunction();
3286 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3287 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3288 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3289 StackSlot, NULL, 0);
3294 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3296 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
3297 SmallVector<SDOperand, 8> Ops;
3298 Ops.push_back(Chain);
3299 Ops.push_back(StackSlot);
3300 Ops.push_back(DAG.getValueType(SrcVT));
3301 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3302 Tys, &Ops[0], Ops.size());
3305 Chain = Result.getValue(1);
3306 SDOperand InFlag = Result.getValue(2);
3308 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3309 // shouldn't be necessary except that RFP cannot be live across
3310 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3311 MachineFunction &MF = DAG.getMachineFunction();
3312 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3313 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3314 Tys = DAG.getVTList(MVT::Other);
3315 SmallVector<SDOperand, 8> Ops;
3316 Ops.push_back(Chain);
3317 Ops.push_back(Result);
3318 Ops.push_back(StackSlot);
3319 Ops.push_back(DAG.getValueType(Op.getValueType()));
3320 Ops.push_back(InFlag);
3321 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3322 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3328 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3329 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3330 "Unknown FP_TO_SINT to lower!");
3331 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3333 MachineFunction &MF = DAG.getMachineFunction();
3334 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3335 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3336 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3339 switch (Op.getValueType()) {
3340 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3341 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3342 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3343 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3346 SDOperand Chain = DAG.getEntryNode();
3347 SDOperand Value = Op.getOperand(0);
3349 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3350 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3351 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
3353 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3355 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3356 Chain = Value.getValue(1);
3357 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3358 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3361 // Build the FP_TO_INT*_IN_MEM
3362 SDOperand Ops[] = { Chain, Value, StackSlot };
3363 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3366 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3369 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3370 MVT::ValueType VT = Op.getValueType();
3371 MVT::ValueType EltVT = VT;
3372 if (MVT::isVector(VT))
3373 EltVT = MVT::getVectorElementType(VT);
3374 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
3375 std::vector<Constant*> CV;
3376 if (EltVT == MVT::f64) {
3377 Constant *C = ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63)));
3381 Constant *C = ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31)));
3387 Constant *C = ConstantVector::get(CV);
3388 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
3389 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
3391 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3394 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3395 MVT::ValueType VT = Op.getValueType();
3396 MVT::ValueType EltVT = VT;
3397 unsigned EltNum = 1;
3398 if (MVT::isVector(VT)) {
3399 EltVT = MVT::getVectorElementType(VT);
3400 EltNum = MVT::getVectorNumElements(VT);
3402 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
3403 std::vector<Constant*> CV;
3404 if (EltVT == MVT::f64) {
3405 Constant *C = ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63));
3409 Constant *C = ConstantFP::get(OpNTy, BitsToFloat(1U << 31));
3415 Constant *C = ConstantVector::get(CV);
3416 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
3417 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
3419 if (MVT::isVector(VT)) {
3420 return DAG.getNode(ISD::BIT_CONVERT, VT,
3421 DAG.getNode(ISD::XOR, MVT::v2i64,
3422 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
3423 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
3425 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3429 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3430 SDOperand Op0 = Op.getOperand(0);
3431 SDOperand Op1 = Op.getOperand(1);
3432 MVT::ValueType VT = Op.getValueType();
3433 MVT::ValueType SrcVT = Op1.getValueType();
3434 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3436 // If second operand is smaller, extend it first.
3437 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3438 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3442 // First get the sign bit of second operand.
3443 std::vector<Constant*> CV;
3444 if (SrcVT == MVT::f64) {
3445 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3446 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3448 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3449 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3450 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3451 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3453 Constant *C = ConstantVector::get(CV);
3454 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
3455 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx, NULL, 0,
3457 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3459 // Shift sign bit right or left if the two operands have different types.
3460 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3461 // Op0 is MVT::f32, Op1 is MVT::f64.
3462 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3463 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3464 DAG.getConstant(32, MVT::i32));
3465 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3466 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3467 DAG.getConstant(0, getPointerTy()));
3470 // Clear first operand sign bit.
3472 if (VT == MVT::f64) {
3473 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3474 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3476 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3477 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3478 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3479 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3481 C = ConstantVector::get(CV);
3482 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
3483 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
3485 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3487 // Or the value with the sign bit.
3488 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3491 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3493 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3495 SDOperand Op0 = Op.getOperand(0);
3496 SDOperand Op1 = Op.getOperand(1);
3497 SDOperand CC = Op.getOperand(2);
3498 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3499 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3500 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3501 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3504 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3506 SDOperand Ops1[] = { Chain, Op0, Op1 };
3507 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3508 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3509 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3512 assert(isFP && "Illegal integer SetCC!");
3514 SDOperand COps[] = { Chain, Op0, Op1 };
3515 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3517 switch (SetCCOpcode) {
3518 default: assert(false && "Illegal floating point SetCC!");
3519 case ISD::SETOEQ: { // !PF & ZF
3520 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3521 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3522 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3524 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3525 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3527 case ISD::SETUNE: { // PF | !ZF
3528 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3529 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3530 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3532 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3533 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3538 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3539 bool addTest = true;
3540 SDOperand Chain = DAG.getEntryNode();
3541 SDOperand Cond = Op.getOperand(0);
3543 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3545 if (Cond.getOpcode() == ISD::SETCC)
3546 Cond = LowerSETCC(Cond, DAG, Chain);
3548 if (Cond.getOpcode() == X86ISD::SETCC) {
3549 CC = Cond.getOperand(0);
3551 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3552 // (since flag operand cannot be shared). Use it as the condition setting
3553 // operand in place of the X86ISD::SETCC.
3554 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3555 // to use a test instead of duplicating the X86ISD::CMP (for register
3556 // pressure reason)?
3557 SDOperand Cmp = Cond.getOperand(1);
3558 unsigned Opc = Cmp.getOpcode();
3559 bool IllegalFPCMov = !X86ScalarSSE &&
3560 MVT::isFloatingPoint(Op.getValueType()) &&
3561 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3562 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3564 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3565 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3571 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3572 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3573 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3576 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3577 SmallVector<SDOperand, 4> Ops;
3578 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3579 // condition is true.
3580 Ops.push_back(Op.getOperand(2));
3581 Ops.push_back(Op.getOperand(1));
3583 Ops.push_back(Cond.getValue(1));
3584 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3587 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3588 bool addTest = true;
3589 SDOperand Chain = Op.getOperand(0);
3590 SDOperand Cond = Op.getOperand(1);
3591 SDOperand Dest = Op.getOperand(2);
3593 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3595 if (Cond.getOpcode() == ISD::SETCC)
3596 Cond = LowerSETCC(Cond, DAG, Chain);
3598 if (Cond.getOpcode() == X86ISD::SETCC) {
3599 CC = Cond.getOperand(0);
3601 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3602 // (since flag operand cannot be shared). Use it as the condition setting
3603 // operand in place of the X86ISD::SETCC.
3604 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3605 // to use a test instead of duplicating the X86ISD::CMP (for register
3606 // pressure reason)?
3607 SDOperand Cmp = Cond.getOperand(1);
3608 unsigned Opc = Cmp.getOpcode();
3609 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3610 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3611 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3617 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3618 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3619 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3621 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3622 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3625 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3626 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3628 if (Subtarget->is64Bit())
3629 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3631 switch (CallingConv) {
3633 assert(0 && "Unsupported calling convention");
3634 case CallingConv::Fast:
3635 // TODO: Implement fastcc
3637 case CallingConv::C:
3638 case CallingConv::X86_StdCall:
3639 return LowerCCCCallTo(Op, DAG, CallingConv);
3640 case CallingConv::X86_FastCall:
3641 return LowerFastCCCallTo(Op, DAG, CallingConv);
3646 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
3647 // Calls to _alloca is needed to probe the stack when allocating more than 4k
3648 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
3649 // that the guard pages used by the OS virtual memory manager are allocated in
3650 // correct sequence.
3652 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
3653 SelectionDAG &DAG) {
3654 assert(Subtarget->isTargetCygMing() &&
3655 "This should be used only on Cygwin/Mingw targets");
3658 SDOperand Chain = Op.getOperand(0);
3659 SDOperand Size = Op.getOperand(1);
3660 // FIXME: Ensure alignment here
3664 MVT::ValueType IntPtr = getPointerTy();
3665 MVT::ValueType SPTy = (Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
3667 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
3668 Flag = Chain.getValue(1);
3670 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3671 SDOperand Ops[] = { Chain,
3672 DAG.getTargetExternalSymbol("_alloca", IntPtr),
3673 DAG.getRegister(X86::EAX, IntPtr),
3675 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
3676 Flag = Chain.getValue(1);
3678 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
3680 std::vector<MVT::ValueType> Tys;
3681 Tys.push_back(SPTy);
3682 Tys.push_back(MVT::Other);
3683 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
3684 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
3688 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3689 MachineFunction &MF = DAG.getMachineFunction();
3690 const Function* Fn = MF.getFunction();
3691 if (Fn->hasExternalLinkage() &&
3692 Subtarget->isTargetCygMing() &&
3693 Fn->getName() == "main")
3694 MF.getInfo<X86MachineFunctionInfo>()->setForceFramePointer(true);
3696 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3697 if (Subtarget->is64Bit())
3698 return LowerX86_64CCCArguments(Op, DAG);
3702 assert(0 && "Unsupported calling convention");
3703 case CallingConv::Fast:
3704 // TODO: implement fastcc.
3707 case CallingConv::C:
3708 return LowerCCCArguments(Op, DAG);
3709 case CallingConv::X86_StdCall:
3710 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(StdCall);
3711 return LowerCCCArguments(Op, DAG, true);
3712 case CallingConv::X86_FastCall:
3713 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(FastCall);
3714 return LowerFastCCArguments(Op, DAG);
3718 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3719 SDOperand InFlag(0, 0);
3720 SDOperand Chain = Op.getOperand(0);
3722 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3723 if (Align == 0) Align = 1;
3725 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3726 // If not DWORD aligned, call memset if size is less than the threshold.
3727 // It knows how to align to the right boundary first.
3728 if ((Align & 3) != 0 ||
3729 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3730 MVT::ValueType IntPtr = getPointerTy();
3731 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3732 TargetLowering::ArgListTy Args;
3733 TargetLowering::ArgListEntry Entry;
3734 Entry.Node = Op.getOperand(1);
3735 Entry.Ty = IntPtrTy;
3736 Args.push_back(Entry);
3737 // Extend the unsigned i8 argument to be an int value for the call.
3738 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3739 Entry.Ty = IntPtrTy;
3740 Args.push_back(Entry);
3741 Entry.Node = Op.getOperand(3);
3742 Args.push_back(Entry);
3743 std::pair<SDOperand,SDOperand> CallResult =
3744 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3745 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3746 return CallResult.second;
3751 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3752 unsigned BytesLeft = 0;
3753 bool TwoRepStos = false;
3756 uint64_t Val = ValC->getValue() & 255;
3758 // If the value is a constant, then we can potentially use larger sets.
3759 switch (Align & 3) {
3760 case 2: // WORD aligned
3763 Val = (Val << 8) | Val;
3765 case 0: // DWORD aligned
3768 Val = (Val << 8) | Val;
3769 Val = (Val << 16) | Val;
3770 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
3773 Val = (Val << 32) | Val;
3776 default: // Byte aligned
3779 Count = Op.getOperand(3);
3783 if (AVT > MVT::i8) {
3785 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3786 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3787 BytesLeft = I->getValue() % UBytes;
3789 assert(AVT >= MVT::i32 &&
3790 "Do not use rep;stos if not at least DWORD aligned");
3791 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3792 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3797 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
3799 InFlag = Chain.getValue(1);
3802 Count = Op.getOperand(3);
3803 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
3804 InFlag = Chain.getValue(1);
3807 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3809 InFlag = Chain.getValue(1);
3810 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3811 Op.getOperand(1), InFlag);
3812 InFlag = Chain.getValue(1);
3814 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3815 SmallVector<SDOperand, 8> Ops;
3816 Ops.push_back(Chain);
3817 Ops.push_back(DAG.getValueType(AVT));
3818 Ops.push_back(InFlag);
3819 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3822 InFlag = Chain.getValue(1);
3823 Count = Op.getOperand(3);
3824 MVT::ValueType CVT = Count.getValueType();
3825 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3826 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3827 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3829 InFlag = Chain.getValue(1);
3830 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3832 Ops.push_back(Chain);
3833 Ops.push_back(DAG.getValueType(MVT::i8));
3834 Ops.push_back(InFlag);
3835 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3836 } else if (BytesLeft) {
3837 // Issue stores for the last 1 - 7 bytes.
3839 unsigned Val = ValC->getValue() & 255;
3840 unsigned Offset = I->getValue() - BytesLeft;
3841 SDOperand DstAddr = Op.getOperand(1);
3842 MVT::ValueType AddrVT = DstAddr.getValueType();
3843 if (BytesLeft >= 4) {
3844 Val = (Val << 8) | Val;
3845 Val = (Val << 16) | Val;
3846 Value = DAG.getConstant(Val, MVT::i32);
3847 Chain = DAG.getStore(Chain, Value,
3848 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3849 DAG.getConstant(Offset, AddrVT)),
3854 if (BytesLeft >= 2) {
3855 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
3856 Chain = DAG.getStore(Chain, Value,
3857 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3858 DAG.getConstant(Offset, AddrVT)),
3863 if (BytesLeft == 1) {
3864 Value = DAG.getConstant(Val, MVT::i8);
3865 Chain = DAG.getStore(Chain, Value,
3866 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3867 DAG.getConstant(Offset, AddrVT)),
3875 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
3876 SDOperand Chain = Op.getOperand(0);
3878 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3879 if (Align == 0) Align = 1;
3881 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3882 // If not DWORD aligned, call memcpy if size is less than the threshold.
3883 // It knows how to align to the right boundary first.
3884 if ((Align & 3) != 0 ||
3885 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3886 MVT::ValueType IntPtr = getPointerTy();
3887 TargetLowering::ArgListTy Args;
3888 TargetLowering::ArgListEntry Entry;
3889 Entry.Ty = getTargetData()->getIntPtrType();
3890 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
3891 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
3892 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
3893 std::pair<SDOperand,SDOperand> CallResult =
3894 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3895 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
3896 return CallResult.second;
3901 unsigned BytesLeft = 0;
3902 bool TwoRepMovs = false;
3903 switch (Align & 3) {
3904 case 2: // WORD aligned
3907 case 0: // DWORD aligned
3909 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
3912 default: // Byte aligned
3914 Count = Op.getOperand(3);
3918 if (AVT > MVT::i8) {
3920 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3921 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3922 BytesLeft = I->getValue() % UBytes;
3924 assert(AVT >= MVT::i32 &&
3925 "Do not use rep;movs if not at least DWORD aligned");
3926 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3927 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3932 SDOperand InFlag(0, 0);
3933 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3935 InFlag = Chain.getValue(1);
3936 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3937 Op.getOperand(1), InFlag);
3938 InFlag = Chain.getValue(1);
3939 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
3940 Op.getOperand(2), InFlag);
3941 InFlag = Chain.getValue(1);
3943 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3944 SmallVector<SDOperand, 8> Ops;
3945 Ops.push_back(Chain);
3946 Ops.push_back(DAG.getValueType(AVT));
3947 Ops.push_back(InFlag);
3948 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3951 InFlag = Chain.getValue(1);
3952 Count = Op.getOperand(3);
3953 MVT::ValueType CVT = Count.getValueType();
3954 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3955 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3956 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3958 InFlag = Chain.getValue(1);
3959 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3961 Ops.push_back(Chain);
3962 Ops.push_back(DAG.getValueType(MVT::i8));
3963 Ops.push_back(InFlag);
3964 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3965 } else if (BytesLeft) {
3966 // Issue loads and stores for the last 1 - 7 bytes.
3967 unsigned Offset = I->getValue() - BytesLeft;
3968 SDOperand DstAddr = Op.getOperand(1);
3969 MVT::ValueType DstVT = DstAddr.getValueType();
3970 SDOperand SrcAddr = Op.getOperand(2);
3971 MVT::ValueType SrcVT = SrcAddr.getValueType();
3973 if (BytesLeft >= 4) {
3974 Value = DAG.getLoad(MVT::i32, Chain,
3975 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3976 DAG.getConstant(Offset, SrcVT)),
3978 Chain = Value.getValue(1);
3979 Chain = DAG.getStore(Chain, Value,
3980 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3981 DAG.getConstant(Offset, DstVT)),
3986 if (BytesLeft >= 2) {
3987 Value = DAG.getLoad(MVT::i16, Chain,
3988 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3989 DAG.getConstant(Offset, SrcVT)),
3991 Chain = Value.getValue(1);
3992 Chain = DAG.getStore(Chain, Value,
3993 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3994 DAG.getConstant(Offset, DstVT)),
4000 if (BytesLeft == 1) {
4001 Value = DAG.getLoad(MVT::i8, Chain,
4002 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4003 DAG.getConstant(Offset, SrcVT)),
4005 Chain = Value.getValue(1);
4006 Chain = DAG.getStore(Chain, Value,
4007 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4008 DAG.getConstant(Offset, DstVT)),
4017 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4018 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4019 SDOperand TheOp = Op.getOperand(0);
4020 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
4021 if (Subtarget->is64Bit()) {
4022 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4023 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4024 MVT::i64, Copy1.getValue(2));
4025 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4026 DAG.getConstant(32, MVT::i8));
4028 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
4031 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4032 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
4035 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4036 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4037 MVT::i32, Copy1.getValue(2));
4038 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
4039 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4040 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
4043 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4044 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4046 if (!Subtarget->is64Bit()) {
4047 // vastart just stores the address of the VarArgsFrameIndex slot into the
4048 // memory location argument.
4049 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4050 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4055 // gp_offset (0 - 6 * 8)
4056 // fp_offset (48 - 48 + 8 * 16)
4057 // overflow_arg_area (point to parameters coming in memory).
4059 SmallVector<SDOperand, 8> MemOps;
4060 SDOperand FIN = Op.getOperand(1);
4062 SDOperand Store = DAG.getStore(Op.getOperand(0),
4063 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4064 FIN, SV->getValue(), SV->getOffset());
4065 MemOps.push_back(Store);
4068 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4069 DAG.getConstant(4, getPointerTy()));
4070 Store = DAG.getStore(Op.getOperand(0),
4071 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4072 FIN, SV->getValue(), SV->getOffset());
4073 MemOps.push_back(Store);
4075 // Store ptr to overflow_arg_area
4076 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4077 DAG.getConstant(4, getPointerTy()));
4078 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4079 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4081 MemOps.push_back(Store);
4083 // Store ptr to reg_save_area.
4084 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4085 DAG.getConstant(8, getPointerTy()));
4086 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4087 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4089 MemOps.push_back(Store);
4090 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4093 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
4094 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
4095 SDOperand Chain = Op.getOperand(0);
4096 SDOperand DstPtr = Op.getOperand(1);
4097 SDOperand SrcPtr = Op.getOperand(2);
4098 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
4099 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
4101 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
4102 SrcSV->getValue(), SrcSV->getOffset());
4103 Chain = SrcPtr.getValue(1);
4104 for (unsigned i = 0; i < 3; ++i) {
4105 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
4106 SrcSV->getValue(), SrcSV->getOffset());
4107 Chain = Val.getValue(1);
4108 Chain = DAG.getStore(Chain, Val, DstPtr,
4109 DstSV->getValue(), DstSV->getOffset());
4112 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
4113 DAG.getConstant(8, getPointerTy()));
4114 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
4115 DAG.getConstant(8, getPointerTy()));
4121 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4122 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4124 default: return SDOperand(); // Don't custom lower most intrinsics.
4125 // Comparison intrinsics.
4126 case Intrinsic::x86_sse_comieq_ss:
4127 case Intrinsic::x86_sse_comilt_ss:
4128 case Intrinsic::x86_sse_comile_ss:
4129 case Intrinsic::x86_sse_comigt_ss:
4130 case Intrinsic::x86_sse_comige_ss:
4131 case Intrinsic::x86_sse_comineq_ss:
4132 case Intrinsic::x86_sse_ucomieq_ss:
4133 case Intrinsic::x86_sse_ucomilt_ss:
4134 case Intrinsic::x86_sse_ucomile_ss:
4135 case Intrinsic::x86_sse_ucomigt_ss:
4136 case Intrinsic::x86_sse_ucomige_ss:
4137 case Intrinsic::x86_sse_ucomineq_ss:
4138 case Intrinsic::x86_sse2_comieq_sd:
4139 case Intrinsic::x86_sse2_comilt_sd:
4140 case Intrinsic::x86_sse2_comile_sd:
4141 case Intrinsic::x86_sse2_comigt_sd:
4142 case Intrinsic::x86_sse2_comige_sd:
4143 case Intrinsic::x86_sse2_comineq_sd:
4144 case Intrinsic::x86_sse2_ucomieq_sd:
4145 case Intrinsic::x86_sse2_ucomilt_sd:
4146 case Intrinsic::x86_sse2_ucomile_sd:
4147 case Intrinsic::x86_sse2_ucomigt_sd:
4148 case Intrinsic::x86_sse2_ucomige_sd:
4149 case Intrinsic::x86_sse2_ucomineq_sd: {
4151 ISD::CondCode CC = ISD::SETCC_INVALID;
4154 case Intrinsic::x86_sse_comieq_ss:
4155 case Intrinsic::x86_sse2_comieq_sd:
4159 case Intrinsic::x86_sse_comilt_ss:
4160 case Intrinsic::x86_sse2_comilt_sd:
4164 case Intrinsic::x86_sse_comile_ss:
4165 case Intrinsic::x86_sse2_comile_sd:
4169 case Intrinsic::x86_sse_comigt_ss:
4170 case Intrinsic::x86_sse2_comigt_sd:
4174 case Intrinsic::x86_sse_comige_ss:
4175 case Intrinsic::x86_sse2_comige_sd:
4179 case Intrinsic::x86_sse_comineq_ss:
4180 case Intrinsic::x86_sse2_comineq_sd:
4184 case Intrinsic::x86_sse_ucomieq_ss:
4185 case Intrinsic::x86_sse2_ucomieq_sd:
4186 Opc = X86ISD::UCOMI;
4189 case Intrinsic::x86_sse_ucomilt_ss:
4190 case Intrinsic::x86_sse2_ucomilt_sd:
4191 Opc = X86ISD::UCOMI;
4194 case Intrinsic::x86_sse_ucomile_ss:
4195 case Intrinsic::x86_sse2_ucomile_sd:
4196 Opc = X86ISD::UCOMI;
4199 case Intrinsic::x86_sse_ucomigt_ss:
4200 case Intrinsic::x86_sse2_ucomigt_sd:
4201 Opc = X86ISD::UCOMI;
4204 case Intrinsic::x86_sse_ucomige_ss:
4205 case Intrinsic::x86_sse2_ucomige_sd:
4206 Opc = X86ISD::UCOMI;
4209 case Intrinsic::x86_sse_ucomineq_ss:
4210 case Intrinsic::x86_sse2_ucomineq_sd:
4211 Opc = X86ISD::UCOMI;
4217 SDOperand LHS = Op.getOperand(1);
4218 SDOperand RHS = Op.getOperand(2);
4219 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4221 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4222 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
4223 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
4224 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
4225 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
4226 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
4227 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4232 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4233 // Depths > 0 not supported yet!
4234 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4237 // Just load the return address
4238 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4239 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4242 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4243 // Depths > 0 not supported yet!
4244 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4247 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4248 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4249 DAG.getConstant(4, getPointerTy()));
4252 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
4253 SelectionDAG &DAG) {
4254 // Is not yet supported on x86-64
4255 if (Subtarget->is64Bit())
4258 return DAG.getConstant(8, getPointerTy());
4261 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
4263 assert(!Subtarget->is64Bit() &&
4264 "Lowering of eh_return builtin is not supported yet on x86-64");
4266 MachineFunction &MF = DAG.getMachineFunction();
4267 SDOperand Chain = Op.getOperand(0);
4268 SDOperand Offset = Op.getOperand(1);
4269 SDOperand Handler = Op.getOperand(2);
4271 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
4274 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
4275 DAG.getConstant(-4UL, getPointerTy()));
4276 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
4277 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
4278 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
4279 MF.addLiveOut(X86::ECX);
4281 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
4282 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
4285 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
4286 SelectionDAG &DAG) {
4287 SDOperand Root = Op.getOperand(0);
4288 SDOperand Trmp = Op.getOperand(1); // trampoline
4289 SDOperand FPtr = Op.getOperand(2); // nested function
4290 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
4292 SrcValueSDNode *TrmpSV = cast<SrcValueSDNode>(Op.getOperand(4));
4294 if (Subtarget->is64Bit()) {
4295 return SDOperand(); // not yet supported
4297 Function *Func = (Function *)
4298 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
4299 unsigned CC = Func->getCallingConv();
4300 unsigned char NestReg;
4304 assert(0 && "Unsupported calling convention");
4305 case CallingConv::C:
4306 case CallingConv::Fast:
4307 case CallingConv::X86_StdCall: {
4308 // Pass 'nest' parameter in ECX.
4309 // Must be kept in sync with X86CallingConv.td
4312 // Check that ECX wasn't needed by an 'inreg' parameter.
4313 const FunctionType *FTy = Func->getFunctionType();
4314 const ParamAttrsList *Attrs = FTy->getParamAttrs();
4316 if (Attrs && !Func->isVarArg()) {
4317 unsigned InRegCount = 0;
4320 for (FunctionType::param_iterator I = FTy->param_begin(),
4321 E = FTy->param_end(); I != E; ++I, ++Idx)
4322 if (Attrs->paramHasAttr(Idx, ParamAttr::InReg))
4323 // FIXME: should only count parameters that are lowered to integers.
4324 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
4326 if (InRegCount > 2) {
4327 cerr << "Nest register in use - reduce number of inreg parameters!\n";
4333 case CallingConv::X86_FastCall:
4334 // Pass 'nest' parameter in EAX.
4335 // Must be kept in sync with X86CallingConv.td
4340 SDOperand OutChains[4];
4341 SDOperand Addr, Disp;
4343 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
4344 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
4346 const unsigned char MOV32ri = 0xB8;
4347 const unsigned char JMP = 0xE9;
4349 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|NestReg, MVT::i8),
4350 Trmp, TrmpSV->getValue(), TrmpSV->getOffset());
4352 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
4353 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
4354 TrmpSV->getOffset() + 1, false, 1);
4356 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
4357 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
4358 TrmpSV->getValue() + 5, TrmpSV->getOffset());
4360 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
4361 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpSV->getValue(),
4362 TrmpSV->getOffset() + 6, false, 1);
4364 return DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4);
4368 /// LowerOperation - Provide custom lowering hooks for some operations.
4370 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4371 switch (Op.getOpcode()) {
4372 default: assert(0 && "Should not custom lower this!");
4373 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4374 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4375 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4376 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4377 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4378 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4379 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4380 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4381 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4382 case ISD::SHL_PARTS:
4383 case ISD::SRA_PARTS:
4384 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4385 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4386 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4387 case ISD::FABS: return LowerFABS(Op, DAG);
4388 case ISD::FNEG: return LowerFNEG(Op, DAG);
4389 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4390 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
4391 case ISD::SELECT: return LowerSELECT(Op, DAG);
4392 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4393 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4394 case ISD::CALL: return LowerCALL(Op, DAG);
4395 case ISD::RET: return LowerRET(Op, DAG);
4396 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4397 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4398 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4399 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4400 case ISD::VASTART: return LowerVASTART(Op, DAG);
4401 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
4402 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4403 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4404 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4405 case ISD::FRAME_TO_ARGS_OFFSET:
4406 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
4407 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
4408 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
4409 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
4414 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4416 default: return NULL;
4417 case X86ISD::SHLD: return "X86ISD::SHLD";
4418 case X86ISD::SHRD: return "X86ISD::SHRD";
4419 case X86ISD::FAND: return "X86ISD::FAND";
4420 case X86ISD::FOR: return "X86ISD::FOR";
4421 case X86ISD::FXOR: return "X86ISD::FXOR";
4422 case X86ISD::FSRL: return "X86ISD::FSRL";
4423 case X86ISD::FILD: return "X86ISD::FILD";
4424 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
4425 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
4426 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
4427 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
4428 case X86ISD::FLD: return "X86ISD::FLD";
4429 case X86ISD::FST: return "X86ISD::FST";
4430 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
4431 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
4432 case X86ISD::CALL: return "X86ISD::CALL";
4433 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
4434 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
4435 case X86ISD::CMP: return "X86ISD::CMP";
4436 case X86ISD::COMI: return "X86ISD::COMI";
4437 case X86ISD::UCOMI: return "X86ISD::UCOMI";
4438 case X86ISD::SETCC: return "X86ISD::SETCC";
4439 case X86ISD::CMOV: return "X86ISD::CMOV";
4440 case X86ISD::BRCOND: return "X86ISD::BRCOND";
4441 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
4442 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
4443 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
4444 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
4445 case X86ISD::Wrapper: return "X86ISD::Wrapper";
4446 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4447 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4448 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4449 case X86ISD::FMAX: return "X86ISD::FMAX";
4450 case X86ISD::FMIN: return "X86ISD::FMIN";
4451 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
4452 case X86ISD::FRCP: return "X86ISD::FRCP";
4453 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
4454 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
4455 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
4459 // isLegalAddressingMode - Return true if the addressing mode represented
4460 // by AM is legal for this target, for a load/store of the specified type.
4461 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
4462 const Type *Ty) const {
4463 // X86 supports extremely general addressing modes.
4465 // X86 allows a sign-extended 32-bit immediate field as a displacement.
4466 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
4470 // We can only fold this if we don't need an extra load.
4471 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
4474 // X86-64 only supports addr of globals in small code model.
4475 if (Subtarget->is64Bit()) {
4476 if (getTargetMachine().getCodeModel() != CodeModel::Small)
4478 // If lower 4G is not available, then we must use rip-relative addressing.
4479 if (AM.BaseOffs || AM.Scale > 1)
4490 // These scales always work.
4495 // These scales are formed with basereg+scalereg. Only accept if there is
4500 default: // Other stuff never works.
4508 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4509 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4510 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4511 /// are assumed to be legal.
4513 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4514 // Only do shuffles on 128-bit vector types for now.
4515 if (MVT::getSizeInBits(VT) == 64) return false;
4516 return (Mask.Val->getNumOperands() <= 4 ||
4517 isIdentityMask(Mask.Val) ||
4518 isIdentityMask(Mask.Val, true) ||
4519 isSplatMask(Mask.Val) ||
4520 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4521 X86::isUNPCKLMask(Mask.Val) ||
4522 X86::isUNPCKHMask(Mask.Val) ||
4523 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4524 X86::isUNPCKH_v_undef_Mask(Mask.Val));
4527 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4529 SelectionDAG &DAG) const {
4530 unsigned NumElts = BVOps.size();
4531 // Only do shuffles on 128-bit vector types for now.
4532 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4533 if (NumElts == 2) return true;
4535 return (isMOVLMask(&BVOps[0], 4) ||
4536 isCommutedMOVL(&BVOps[0], 4, true) ||
4537 isSHUFPMask(&BVOps[0], 4) ||
4538 isCommutedSHUFP(&BVOps[0], 4));
4543 //===----------------------------------------------------------------------===//
4544 // X86 Scheduler Hooks
4545 //===----------------------------------------------------------------------===//
4548 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4549 MachineBasicBlock *BB) {
4550 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4551 switch (MI->getOpcode()) {
4552 default: assert(false && "Unexpected instr type to insert");
4553 case X86::CMOV_FR32:
4554 case X86::CMOV_FR64:
4555 case X86::CMOV_V4F32:
4556 case X86::CMOV_V2F64:
4557 case X86::CMOV_V2I64: {
4558 // To "insert" a SELECT_CC instruction, we actually have to insert the
4559 // diamond control-flow pattern. The incoming instruction knows the
4560 // destination vreg to set, the condition code register to branch on, the
4561 // true/false values to select between, and a branch opcode to use.
4562 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4563 ilist<MachineBasicBlock>::iterator It = BB;
4569 // cmpTY ccX, r1, r2
4571 // fallthrough --> copy0MBB
4572 MachineBasicBlock *thisMBB = BB;
4573 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4574 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4576 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4577 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4578 MachineFunction *F = BB->getParent();
4579 F->getBasicBlockList().insert(It, copy0MBB);
4580 F->getBasicBlockList().insert(It, sinkMBB);
4581 // Update machine-CFG edges by first adding all successors of the current
4582 // block to the new block which will contain the Phi node for the select.
4583 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4584 e = BB->succ_end(); i != e; ++i)
4585 sinkMBB->addSuccessor(*i);
4586 // Next, remove all successors of the current block, and add the true
4587 // and fallthrough blocks as its successors.
4588 while(!BB->succ_empty())
4589 BB->removeSuccessor(BB->succ_begin());
4590 BB->addSuccessor(copy0MBB);
4591 BB->addSuccessor(sinkMBB);
4594 // %FalseValue = ...
4595 // # fallthrough to sinkMBB
4598 // Update machine-CFG edges
4599 BB->addSuccessor(sinkMBB);
4602 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4605 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4606 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4607 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4609 delete MI; // The pseudo instruction is gone now.
4613 case X86::FP32_TO_INT16_IN_MEM:
4614 case X86::FP32_TO_INT32_IN_MEM:
4615 case X86::FP32_TO_INT64_IN_MEM:
4616 case X86::FP64_TO_INT16_IN_MEM:
4617 case X86::FP64_TO_INT32_IN_MEM:
4618 case X86::FP64_TO_INT64_IN_MEM:
4619 case X86::FP80_TO_INT16_IN_MEM:
4620 case X86::FP80_TO_INT32_IN_MEM:
4621 case X86::FP80_TO_INT64_IN_MEM: {
4622 // Change the floating point control register to use "round towards zero"
4623 // mode when truncating to an integer value.
4624 MachineFunction *F = BB->getParent();
4625 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4626 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4628 // Load the old value of the high byte of the control word...
4630 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4631 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4633 // Set the high part to be round to zero...
4634 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4637 // Reload the modified control word now...
4638 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4640 // Restore the memory image of control word to original value
4641 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4644 // Get the X86 opcode to use.
4646 switch (MI->getOpcode()) {
4647 default: assert(0 && "illegal opcode!");
4648 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
4649 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
4650 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
4651 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
4652 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
4653 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
4654 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
4655 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
4656 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
4660 MachineOperand &Op = MI->getOperand(0);
4661 if (Op.isRegister()) {
4662 AM.BaseType = X86AddressMode::RegBase;
4663 AM.Base.Reg = Op.getReg();
4665 AM.BaseType = X86AddressMode::FrameIndexBase;
4666 AM.Base.FrameIndex = Op.getFrameIndex();
4668 Op = MI->getOperand(1);
4669 if (Op.isImmediate())
4670 AM.Scale = Op.getImm();
4671 Op = MI->getOperand(2);
4672 if (Op.isImmediate())
4673 AM.IndexReg = Op.getImm();
4674 Op = MI->getOperand(3);
4675 if (Op.isGlobalAddress()) {
4676 AM.GV = Op.getGlobal();
4678 AM.Disp = Op.getImm();
4680 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4681 .addReg(MI->getOperand(4).getReg());
4683 // Reload the original control word now.
4684 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4686 delete MI; // The pseudo instruction is gone now.
4692 //===----------------------------------------------------------------------===//
4693 // X86 Optimization Hooks
4694 //===----------------------------------------------------------------------===//
4696 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4698 uint64_t &KnownZero,
4700 const SelectionDAG &DAG,
4701 unsigned Depth) const {
4702 unsigned Opc = Op.getOpcode();
4703 assert((Opc >= ISD::BUILTIN_OP_END ||
4704 Opc == ISD::INTRINSIC_WO_CHAIN ||
4705 Opc == ISD::INTRINSIC_W_CHAIN ||
4706 Opc == ISD::INTRINSIC_VOID) &&
4707 "Should use MaskedValueIsZero if you don't know whether Op"
4708 " is a target node!");
4710 KnownZero = KnownOne = 0; // Don't know anything.
4714 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4719 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4720 /// element of the result of the vector shuffle.
4721 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4722 MVT::ValueType VT = N->getValueType(0);
4723 SDOperand PermMask = N->getOperand(2);
4724 unsigned NumElems = PermMask.getNumOperands();
4725 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4727 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4729 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
4730 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4731 SDOperand Idx = PermMask.getOperand(i);
4732 if (Idx.getOpcode() == ISD::UNDEF)
4733 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
4734 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4739 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4740 /// node is a GlobalAddress + an offset.
4741 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4742 unsigned Opc = N->getOpcode();
4743 if (Opc == X86ISD::Wrapper) {
4744 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4745 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4748 } else if (Opc == ISD::ADD) {
4749 SDOperand N1 = N->getOperand(0);
4750 SDOperand N2 = N->getOperand(1);
4751 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4752 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4754 Offset += V->getSignExtended();
4757 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4758 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4760 Offset += V->getSignExtended();
4768 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4770 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4771 MachineFrameInfo *MFI) {
4772 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4775 SDOperand Loc = N->getOperand(1);
4776 SDOperand BaseLoc = Base->getOperand(1);
4777 if (Loc.getOpcode() == ISD::FrameIndex) {
4778 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4780 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
4781 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4782 int FS = MFI->getObjectSize(FI);
4783 int BFS = MFI->getObjectSize(BFI);
4784 if (FS != BFS || FS != Size) return false;
4785 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4787 GlobalValue *GV1 = NULL;
4788 GlobalValue *GV2 = NULL;
4789 int64_t Offset1 = 0;
4790 int64_t Offset2 = 0;
4791 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4792 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4793 if (isGA1 && isGA2 && GV1 == GV2)
4794 return Offset1 == (Offset2 + Dist*Size);
4800 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4801 const X86Subtarget *Subtarget) {
4804 if (isGAPlusOffset(Base, GV, Offset))
4805 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4807 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4808 int BFI = cast<FrameIndexSDNode>(Base)->getIndex();
4810 // Fixed objects do not specify alignment, however the offsets are known.
4811 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4812 (MFI->getObjectOffset(BFI) % 16) == 0);
4814 return MFI->getObjectAlignment(BFI) >= 16;
4820 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4821 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4822 /// if the load addresses are consecutive, non-overlapping, and in the right
4824 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4825 const X86Subtarget *Subtarget) {
4826 MachineFunction &MF = DAG.getMachineFunction();
4827 MachineFrameInfo *MFI = MF.getFrameInfo();
4828 MVT::ValueType VT = N->getValueType(0);
4829 MVT::ValueType EVT = MVT::getVectorElementType(VT);
4830 SDOperand PermMask = N->getOperand(2);
4831 int NumElems = (int)PermMask.getNumOperands();
4832 SDNode *Base = NULL;
4833 for (int i = 0; i < NumElems; ++i) {
4834 SDOperand Idx = PermMask.getOperand(i);
4835 if (Idx.getOpcode() == ISD::UNDEF) {
4836 if (!Base) return SDOperand();
4839 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4840 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4844 else if (!isConsecutiveLoad(Arg.Val, Base,
4845 i, MVT::getSizeInBits(EVT)/8,MFI))
4850 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4851 LoadSDNode *LD = cast<LoadSDNode>(Base);
4853 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4854 LD->getSrcValueOffset(), LD->isVolatile());
4856 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4857 LD->getSrcValueOffset(), LD->isVolatile(),
4858 LD->getAlignment());
4862 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4863 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4864 const X86Subtarget *Subtarget) {
4865 SDOperand Cond = N->getOperand(0);
4867 // If we have SSE[12] support, try to form min/max nodes.
4868 if (Subtarget->hasSSE2() &&
4869 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4870 if (Cond.getOpcode() == ISD::SETCC) {
4871 // Get the LHS/RHS of the select.
4872 SDOperand LHS = N->getOperand(1);
4873 SDOperand RHS = N->getOperand(2);
4874 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4876 unsigned Opcode = 0;
4877 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4880 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4883 if (!UnsafeFPMath) break;
4885 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4887 Opcode = X86ISD::FMIN;
4890 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4893 if (!UnsafeFPMath) break;
4895 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4897 Opcode = X86ISD::FMAX;
4900 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4903 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4906 if (!UnsafeFPMath) break;
4908 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4910 Opcode = X86ISD::FMIN;
4913 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4916 if (!UnsafeFPMath) break;
4918 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4920 Opcode = X86ISD::FMAX;
4926 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4935 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4936 DAGCombinerInfo &DCI) const {
4937 SelectionDAG &DAG = DCI.DAG;
4938 switch (N->getOpcode()) {
4940 case ISD::VECTOR_SHUFFLE:
4941 return PerformShuffleCombine(N, DAG, Subtarget);
4943 return PerformSELECTCombine(N, DAG, Subtarget);
4949 //===----------------------------------------------------------------------===//
4950 // X86 Inline Assembly Support
4951 //===----------------------------------------------------------------------===//
4953 /// getConstraintType - Given a constraint letter, return the type of
4954 /// constraint it is for this target.
4955 X86TargetLowering::ConstraintType
4956 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
4957 if (Constraint.size() == 1) {
4958 switch (Constraint[0]) {
4967 return C_RegisterClass;
4972 return TargetLowering::getConstraintType(Constraint);
4975 /// isOperandValidForConstraint - Return the specified operand (possibly
4976 /// modified) if the specified SDOperand is valid for the specified target
4977 /// constraint letter, otherwise return null.
4978 SDOperand X86TargetLowering::
4979 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
4980 switch (Constraint) {
4983 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4984 if (C->getValue() <= 31)
4985 return DAG.getTargetConstant(C->getValue(), Op.getValueType());
4987 return SDOperand(0,0);
4989 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4990 if (C->getValue() <= 255)
4991 return DAG.getTargetConstant(C->getValue(), Op.getValueType());
4993 return SDOperand(0,0);
4995 // Literal immediates are always ok.
4996 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op))
4997 return DAG.getTargetConstant(CST->getValue(), Op.getValueType());
4999 // If we are in non-pic codegen mode, we allow the address of a global (with
5000 // an optional displacement) to be used with 'i'.
5001 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
5004 // Match either (GA) or (GA+C)
5006 Offset = GA->getOffset();
5007 } else if (Op.getOpcode() == ISD::ADD) {
5008 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5009 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5011 Offset = GA->getOffset()+C->getValue();
5013 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5014 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5016 Offset = GA->getOffset()+C->getValue();
5023 // If addressing this global requires a load (e.g. in PIC mode), we can't
5025 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
5027 return SDOperand(0, 0);
5029 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5034 // Otherwise, not valid for this mode.
5035 return SDOperand(0, 0);
5038 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
5041 std::vector<unsigned> X86TargetLowering::
5042 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5043 MVT::ValueType VT) const {
5044 if (Constraint.size() == 1) {
5045 // FIXME: not handling fp-stack yet!
5046 switch (Constraint[0]) { // GCC X86 Constraint Letters
5047 default: break; // Unknown constraint letter
5048 case 'A': // EAX/EDX
5049 if (VT == MVT::i32 || VT == MVT::i64)
5050 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5052 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5055 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5056 else if (VT == MVT::i16)
5057 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5058 else if (VT == MVT::i8)
5059 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5064 return std::vector<unsigned>();
5067 std::pair<unsigned, const TargetRegisterClass*>
5068 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5069 MVT::ValueType VT) const {
5070 // First, see if this is a constraint that directly corresponds to an LLVM
5072 if (Constraint.size() == 1) {
5073 // GCC Constraint Letters
5074 switch (Constraint[0]) {
5076 case 'r': // GENERAL_REGS
5077 case 'R': // LEGACY_REGS
5078 case 'l': // INDEX_REGS
5079 if (VT == MVT::i64 && Subtarget->is64Bit())
5080 return std::make_pair(0U, X86::GR64RegisterClass);
5082 return std::make_pair(0U, X86::GR32RegisterClass);
5083 else if (VT == MVT::i16)
5084 return std::make_pair(0U, X86::GR16RegisterClass);
5085 else if (VT == MVT::i8)
5086 return std::make_pair(0U, X86::GR8RegisterClass);
5088 case 'y': // MMX_REGS if MMX allowed.
5089 if (!Subtarget->hasMMX()) break;
5090 return std::make_pair(0U, X86::VR64RegisterClass);
5092 case 'Y': // SSE_REGS if SSE2 allowed
5093 if (!Subtarget->hasSSE2()) break;
5095 case 'x': // SSE_REGS if SSE1 allowed
5096 if (!Subtarget->hasSSE1()) break;
5100 // Scalar SSE types.
5103 return std::make_pair(0U, X86::FR32RegisterClass);
5106 return std::make_pair(0U, X86::FR64RegisterClass);
5114 return std::make_pair(0U, X86::VR128RegisterClass);
5120 // Use the default implementation in TargetLowering to convert the register
5121 // constraint into a member of a register class.
5122 std::pair<unsigned, const TargetRegisterClass*> Res;
5123 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5125 // Not found as a standard register?
5126 if (Res.second == 0) {
5127 // GCC calls "st(0)" just plain "st".
5128 if (StringsEqualNoCase("{st}", Constraint)) {
5129 Res.first = X86::ST0;
5130 Res.second = X86::RSTRegisterClass;
5136 // Otherwise, check to see if this is a register class of the wrong value
5137 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5138 // turn into {ax},{dx}.
5139 if (Res.second->hasType(VT))
5140 return Res; // Correct type already, nothing to do.
5142 // All of the single-register GCC register classes map their values onto
5143 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5144 // really want an 8-bit or 32-bit register, map to the appropriate register
5145 // class and return the appropriate register.
5146 if (Res.second != X86::GR16RegisterClass)
5149 if (VT == MVT::i8) {
5150 unsigned DestReg = 0;
5151 switch (Res.first) {
5153 case X86::AX: DestReg = X86::AL; break;
5154 case X86::DX: DestReg = X86::DL; break;
5155 case X86::CX: DestReg = X86::CL; break;
5156 case X86::BX: DestReg = X86::BL; break;
5159 Res.first = DestReg;
5160 Res.second = Res.second = X86::GR8RegisterClass;
5162 } else if (VT == MVT::i32) {
5163 unsigned DestReg = 0;
5164 switch (Res.first) {
5166 case X86::AX: DestReg = X86::EAX; break;
5167 case X86::DX: DestReg = X86::EDX; break;
5168 case X86::CX: DestReg = X86::ECX; break;
5169 case X86::BX: DestReg = X86::EBX; break;
5170 case X86::SI: DestReg = X86::ESI; break;
5171 case X86::DI: DestReg = X86::EDI; break;
5172 case X86::BP: DestReg = X86::EBP; break;
5173 case X86::SP: DestReg = X86::ESP; break;
5176 Res.first = DestReg;
5177 Res.second = Res.second = X86::GR32RegisterClass;
5179 } else if (VT == MVT::i64) {
5180 unsigned DestReg = 0;
5181 switch (Res.first) {
5183 case X86::AX: DestReg = X86::RAX; break;
5184 case X86::DX: DestReg = X86::RDX; break;
5185 case X86::CX: DestReg = X86::RCX; break;
5186 case X86::BX: DestReg = X86::RBX; break;
5187 case X86::SI: DestReg = X86::RSI; break;
5188 case X86::DI: DestReg = X86::RDI; break;
5189 case X86::BP: DestReg = X86::RBP; break;
5190 case X86::SP: DestReg = X86::RSP; break;
5193 Res.first = DestReg;
5194 Res.second = Res.second = X86::GR64RegisterClass;