1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/ADT/VectorExtras.h"
26 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
27 #include "llvm/CodeGen/CallingConvLower.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/SelectionDAG.h"
32 #include "llvm/CodeGen/SSARegMap.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/ADT/StringExtras.h"
40 static cl::opt<bool> EnableFastCC("enable-x86-fastcc", cl::Hidden,
41 cl::desc("Enable fastcc on X86"));
42 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
43 : TargetLowering(TM) {
44 Subtarget = &TM.getSubtarget<X86Subtarget>();
45 X86ScalarSSE = Subtarget->hasSSE2();
46 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
48 // Set up the TargetLowering object.
50 // X86 is weird, it always uses i8 for shift amounts and setcc results.
51 setShiftAmountType(MVT::i8);
52 setSetCCResultType(MVT::i8);
53 setSetCCResultContents(ZeroOrOneSetCCResult);
54 setSchedulingPreference(SchedulingForRegPressure);
55 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
56 setStackPointerRegisterToSaveRestore(X86StackPtr);
58 if (Subtarget->isTargetDarwin()) {
59 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
60 setUseUnderscoreSetJmp(false);
61 setUseUnderscoreLongJmp(false);
62 } else if (Subtarget->isTargetMingw()) {
63 // MS runtime is weird: it exports _setjmp, but longjmp!
64 setUseUnderscoreSetJmp(true);
65 setUseUnderscoreLongJmp(false);
67 setUseUnderscoreSetJmp(true);
68 setUseUnderscoreLongJmp(true);
71 // Add legal addressing mode scale values.
72 addLegalAddressScale(8);
73 addLegalAddressScale(4);
74 addLegalAddressScale(2);
75 // Enter the ones which require both scale + index last. These are more
77 addLegalAddressScale(9);
78 addLegalAddressScale(5);
79 addLegalAddressScale(3);
81 // Set up the register classes.
82 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
83 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
84 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
85 if (Subtarget->is64Bit())
86 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
88 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
90 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
92 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
93 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
94 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
96 if (Subtarget->is64Bit()) {
97 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
98 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
101 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
102 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
104 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
107 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
109 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
110 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
111 // SSE has no i16 to fp conversion, only i32
113 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
115 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
116 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
119 if (!Subtarget->is64Bit()) {
120 // Custom lower SINT_TO_FP and FP_TO_SINT from/to i64 in 32-bit mode.
121 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
122 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
125 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
127 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
128 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
131 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
133 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
134 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
137 // Handle FP_TO_UINT by promoting the destination to a larger signed
139 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
140 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
141 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
143 if (Subtarget->is64Bit()) {
144 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
145 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
147 if (X86ScalarSSE && !Subtarget->hasSSE3())
148 // Expand FP_TO_UINT into a select.
149 // FIXME: We would like to use a Custom expander here eventually to do
150 // the optimal thing for SSE vs. the default expansion in the legalizer.
151 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
153 // With SSE3 we can use fisttpll to convert to a signed i64.
154 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
157 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
159 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
160 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
163 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
164 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
165 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
166 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
167 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
168 if (Subtarget->is64Bit())
169 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Expand);
170 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
171 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
172 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
173 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
174 setOperationAction(ISD::FREM , MVT::f64 , Expand);
176 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
177 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
178 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
179 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
180 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
181 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
182 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
183 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
184 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
185 if (Subtarget->is64Bit()) {
186 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
187 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
188 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
191 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
192 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
194 // These should be promoted to a larger select which is supported.
195 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
196 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
197 // X86 wants to expand cmov itself.
198 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
199 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
200 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
201 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
202 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
203 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
204 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
205 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
206 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
207 if (Subtarget->is64Bit()) {
208 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
209 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
211 // X86 ret instruction may pop stack.
212 setOperationAction(ISD::RET , MVT::Other, Custom);
214 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
215 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
216 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
217 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
218 if (Subtarget->is64Bit()) {
219 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
220 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
221 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
222 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
224 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
225 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
226 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
227 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
228 // X86 wants to expand memset / memcpy itself.
229 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
230 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
232 // We don't have line number support yet.
233 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
234 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
235 // FIXME - use subtarget debug flags
236 if (!Subtarget->isTargetDarwin() &&
237 !Subtarget->isTargetELF() &&
238 !Subtarget->isTargetCygMing())
239 setOperationAction(ISD::LABEL, MVT::Other, Expand);
241 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
242 setOperationAction(ISD::VASTART , MVT::Other, Custom);
244 // Use the default implementation.
245 setOperationAction(ISD::VAARG , MVT::Other, Expand);
246 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
247 setOperationAction(ISD::VAEND , MVT::Other, Expand);
248 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
249 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
250 if (Subtarget->is64Bit())
251 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
252 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
255 // Set up the FP register classes.
256 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
257 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
259 // Use ANDPD to simulate FABS.
260 setOperationAction(ISD::FABS , MVT::f64, Custom);
261 setOperationAction(ISD::FABS , MVT::f32, Custom);
263 // Use XORP to simulate FNEG.
264 setOperationAction(ISD::FNEG , MVT::f64, Custom);
265 setOperationAction(ISD::FNEG , MVT::f32, Custom);
267 // Use ANDPD and ORPD to simulate FCOPYSIGN.
268 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
269 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
271 // We don't support sin/cos/fmod
272 setOperationAction(ISD::FSIN , MVT::f64, Expand);
273 setOperationAction(ISD::FCOS , MVT::f64, Expand);
274 setOperationAction(ISD::FREM , MVT::f64, Expand);
275 setOperationAction(ISD::FSIN , MVT::f32, Expand);
276 setOperationAction(ISD::FCOS , MVT::f32, Expand);
277 setOperationAction(ISD::FREM , MVT::f32, Expand);
279 // Expand FP immediates into loads from the stack, except for the special
281 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
282 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
283 addLegalFPImmediate(+0.0); // xorps / xorpd
285 // Set up the FP register classes.
286 addRegisterClass(MVT::f64, X86::RFPRegisterClass);
288 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
289 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
290 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
293 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
294 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
297 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
298 addLegalFPImmediate(+0.0); // FLD0
299 addLegalFPImmediate(+1.0); // FLD1
300 addLegalFPImmediate(-0.0); // FLD0/FCHS
301 addLegalFPImmediate(-1.0); // FLD1/FCHS
304 // First set operation action for all vector types to expand. Then we
305 // will selectively turn on ones that can be effectively codegen'd.
306 for (unsigned VT = (unsigned)MVT::Vector + 1;
307 VT != (unsigned)MVT::LAST_VALUETYPE; VT++) {
308 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
309 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
310 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
311 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
312 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
313 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
314 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
315 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
316 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
317 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
318 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
319 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
320 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
321 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
322 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
325 if (Subtarget->hasMMX()) {
326 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
327 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
328 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
330 // FIXME: add MMX packed arithmetics
331 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Expand);
332 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Expand);
333 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Expand);
336 if (Subtarget->hasSSE1()) {
337 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
339 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
340 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
341 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
342 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
343 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
344 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
345 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
346 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
347 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
350 if (Subtarget->hasSSE2()) {
351 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
352 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
353 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
354 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
355 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
357 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
358 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
359 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
360 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
361 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
362 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
363 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
364 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
365 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
366 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
367 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
369 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
370 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
371 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
372 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
373 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
374 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
376 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
377 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
378 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
379 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
380 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
382 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
383 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
384 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
385 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
386 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
387 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
389 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
390 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
391 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
392 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
393 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
394 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
395 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
396 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
397 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
398 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
399 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
400 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
403 // Custom lower v2i64 and v2f64 selects.
404 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
405 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
406 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
407 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
410 // We want to custom lower some of our intrinsics.
411 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
413 // We have target-specific dag combine patterns for the following nodes:
414 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
415 setTargetDAGCombine(ISD::SELECT);
417 computeRegisterProperties();
419 // FIXME: These should be based on subtarget info. Plus, the values should
420 // be smaller when we are in optimizing for size mode.
421 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
422 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
423 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
424 allowUnalignedMemoryAccesses = true; // x86 supports it!
428 //===----------------------------------------------------------------------===//
429 // Return Value Calling Convention Implementation
430 //===----------------------------------------------------------------------===//
432 /// X86_RetCC_Assign - Implement the X86 return value conventions. This returns
433 /// true if the value wasn't handled by this CC.
434 static bool X86_RetCC_Assign(unsigned ValNo, MVT::ValueType ValVT,
435 unsigned ArgFlags, CCState &State) {
436 MVT::ValueType LocVT = ValVT;
437 CCValAssign::LocInfo LocInfo = CCValAssign::Full;
439 // If this is a 32-bit value, assign to a 32-bit register if any are
441 if (LocVT == MVT::i8) {
442 static const unsigned GPR8ArgRegs[] = { X86::AL, X86::DL };
443 if (unsigned Reg = State.AllocateReg(GPR8ArgRegs, 2)) {
444 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
448 if (LocVT == MVT::i16) {
449 static const unsigned GPR16ArgRegs[] = { X86::AX, X86::DX };
450 if (unsigned Reg = State.AllocateReg(GPR16ArgRegs, 2)) {
451 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
455 if (LocVT == MVT::i32) {
456 static const unsigned GPR32ArgRegs[] = { X86::EAX, X86::EDX };
457 if (unsigned Reg = State.AllocateReg(GPR32ArgRegs, 2)) {
458 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
462 if (LocVT == MVT::i64) {
463 static const unsigned GPR64ArgRegs[] = { X86::RAX, X86::RDX };
464 if (unsigned Reg = State.AllocateReg(GPR64ArgRegs, 2)) {
465 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
469 if (MVT::isVector(LocVT)) {
470 if (unsigned Reg = State.AllocateReg(X86::XMM0)) {
471 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
475 if (LocVT == MVT::f32 || LocVT == MVT::f64) {
477 if (State.getTarget().getSubtarget<X86Subtarget>().is64Bit())
478 Reg = X86::XMM0; // FP values in X86-64 go in XMM0.
479 else if (State.getCallingConv() == CallingConv::Fast &&
480 State.getTarget().getSubtarget<X86Subtarget>().hasSSE2())
481 Reg = X86::XMM0; // FP values in X86-32 with fastcc go in XMM0.
483 Reg = X86::ST0; // FP values in X86-32 go in ST0.
485 if ((Reg = State.AllocateReg(Reg))) {
486 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
494 /// LowerRET - Lower an ISD::RET node.
495 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
496 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
498 SmallVector<CCValAssign, 16> RVLocs;
499 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
500 CCState CCInfo(CC, getTargetMachine(), RVLocs);
502 // Determine which register each value should be copied into.
503 for (unsigned i = 0; i != Op.getNumOperands() / 2; ++i) {
504 if (X86_RetCC_Assign(i, Op.getOperand(i*2+1).getValueType(),
505 cast<ConstantSDNode>(Op.getOperand(i*2+2))->getValue(),
507 assert(0 && "Unhandled result type!");
510 // If this is the first return lowered for this function, add the regs to the
511 // liveout set for the function.
512 if (DAG.getMachineFunction().liveout_empty()) {
513 for (unsigned i = 0; i != RVLocs.size(); ++i)
514 if (RVLocs[i].isRegLoc())
515 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
518 SDOperand Chain = Op.getOperand(0);
521 // Copy the result values into the output registers.
522 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
523 RVLocs[0].getLocReg() != X86::ST0) {
524 for (unsigned i = 0; i != RVLocs.size(); ++i) {
525 CCValAssign &VA = RVLocs[i];
526 assert(VA.isRegLoc() && "Can only return in registers!");
527 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
529 Flag = Chain.getValue(1);
532 // We need to handle a destination of ST0 specially, because it isn't really
534 SDOperand Value = Op.getOperand(1);
536 // If this is an FP return with ScalarSSE, we need to move the value from
537 // an XMM register onto the fp-stack.
541 // If this is a load into a scalarsse value, don't store the loaded value
542 // back to the stack, only to reload it: just replace the scalar-sse load.
543 if (ISD::isNON_EXTLoad(Value.Val) &&
544 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
545 Chain = Value.getOperand(0);
546 MemLoc = Value.getOperand(1);
548 // Spill the value to memory and reload it into top of stack.
549 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
550 MachineFunction &MF = DAG.getMachineFunction();
551 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
552 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
553 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
555 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
556 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
557 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
558 Chain = Value.getValue(1);
561 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
562 SDOperand Ops[] = { Chain, Value };
563 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
564 Flag = Chain.getValue(1);
567 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
569 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
571 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
575 /// LowerCallResult - Lower the result values of an ISD::CALL into the
576 /// appropriate copies out of appropriate physical registers. This assumes that
577 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
578 /// being lowered. The returns a SDNode with the same number of values as the
580 SDNode *X86TargetLowering::
581 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
582 unsigned CallingConv, SelectionDAG &DAG) {
583 SmallVector<SDOperand, 8> ResultVals;
585 SmallVector<CCValAssign, 16> RVLocs;
586 CCState CCInfo(CallingConv, getTargetMachine(), RVLocs);
588 for (unsigned i = 0, e = TheCall->getNumValues() - 1; i != e; ++i) {
589 if (X86_RetCC_Assign(i, TheCall->getValueType(i), 0, CCInfo))
590 assert(0 && "Unhandled result type!");
593 // Copy all of the result registers out of their specified physreg.
594 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
595 for (unsigned i = 0; i != RVLocs.size(); ++i) {
596 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
597 RVLocs[i].getValVT(), InFlag).getValue(1);
598 InFlag = Chain.getValue(2);
599 ResultVals.push_back(Chain.getValue(0));
602 // Copies from the FP stack are special, as ST0 isn't a valid register
603 // before the fp stackifier runs.
605 // Copy ST0 into an RFP register with FP_GET_RESULT.
606 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
607 SDOperand GROps[] = { Chain, InFlag };
608 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
609 Chain = RetVal.getValue(1);
610 InFlag = RetVal.getValue(2);
612 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
615 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
616 // shouldn't be necessary except that RFP cannot be live across
617 // multiple blocks. When stackifier is fixed, they can be uncoupled.
618 MachineFunction &MF = DAG.getMachineFunction();
619 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
620 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
622 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag
624 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
625 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
626 Chain = RetVal.getValue(1);
629 if (RVLocs[0].getValVT() == MVT::f32 && !X86ScalarSSE)
630 // FIXME: we would really like to remember that this FP_ROUND
631 // operation is okay to eliminate if we allow excess FP precision.
632 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
633 ResultVals.push_back(RetVal);
636 // Merge everything together with a MERGE_VALUES node.
637 ResultVals.push_back(Chain);
638 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
639 &ResultVals[0], ResultVals.size()).Val;
643 //===----------------------------------------------------------------------===//
644 // C & StdCall Calling Convention implementation
645 //===----------------------------------------------------------------------===//
646 // StdCall calling convention seems to be standard for many Windows' API
647 // routines and around. It differs from C calling convention just a little:
648 // callee should clean up the stack, not caller. Symbols should be also
649 // decorated in some fancy way :) It doesn't support any vector arguments.
651 /// AddLiveIn - This helper function adds the specified physical register to the
652 /// MachineFunction as a live in value. It also creates a corresponding virtual
654 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
655 const TargetRegisterClass *RC) {
656 assert(RC->contains(PReg) && "Not the correct regclass!");
657 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
658 MF.addLiveIn(PReg, VReg);
662 /// HowToPassArgument - Returns how an formal argument of the specified type
663 /// should be passed. If it is through stack, returns the size of the stack
664 /// slot; if it is through integer or XMM register, returns the number of
665 /// integer or XMM registers are needed.
667 HowToPassCallArgument(MVT::ValueType ObjectVT,
669 unsigned NumIntRegs, unsigned NumXMMRegs,
670 unsigned MaxNumIntRegs,
671 unsigned &ObjSize, unsigned &ObjIntRegs,
672 unsigned &ObjXMMRegs) {
677 if (MaxNumIntRegs>3) {
678 // We don't have too much registers on ia32! :)
683 default: assert(0 && "Unhandled argument type!");
685 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
691 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
697 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
703 if (ArgInReg && (NumIntRegs+2 <= MaxNumIntRegs)) {
705 } else if (ArgInReg && (NumIntRegs+1 <= MaxNumIntRegs)) {
730 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
732 unsigned NumArgs = Op.Val->getNumValues() - 1;
733 MachineFunction &MF = DAG.getMachineFunction();
734 MachineFrameInfo *MFI = MF.getFrameInfo();
735 SDOperand Root = Op.getOperand(0);
736 SmallVector<SDOperand, 8> ArgValues;
737 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
739 // Add DAG nodes to load the arguments... On entry to a function on the X86,
740 // the stack frame looks like this:
742 // [ESP] -- return address
743 // [ESP + 4] -- first argument (leftmost lexically)
744 // [ESP + 8] -- second argument, if first argument is <= 4 bytes in size
747 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
748 unsigned NumSRetBytes= 0; // How much bytes on stack used for struct return
749 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
750 unsigned NumIntRegs = 0; // Integer regs used for parameter passing
752 static const unsigned XMMArgRegs[] = {
753 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
755 static const unsigned GPRArgRegs[][3] = {
756 { X86::AL, X86::DL, X86::CL },
757 { X86::AX, X86::DX, X86::CX },
758 { X86::EAX, X86::EDX, X86::ECX }
760 static const TargetRegisterClass* GPRClasses[3] = {
761 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
764 // Handle regparm attribute
765 SmallVector<bool, 8> ArgInRegs(NumArgs, false);
766 SmallVector<bool, 8> SRetArgs(NumArgs, false);
768 for (unsigned i = 0; i<NumArgs; ++i) {
769 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3+i))->getValue();
770 ArgInRegs[i] = (Flags >> 1) & 1;
771 SRetArgs[i] = (Flags >> 2) & 1;
775 for (unsigned i = 0; i < NumArgs; ++i) {
776 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
777 unsigned ArgIncrement = 4;
778 unsigned ObjSize = 0;
779 unsigned ObjXMMRegs = 0;
780 unsigned ObjIntRegs = 0;
784 HowToPassCallArgument(ObjectVT,
786 NumIntRegs, NumXMMRegs, 3,
787 ObjSize, ObjIntRegs, ObjXMMRegs);
790 ArgIncrement = ObjSize;
792 if (ObjIntRegs || ObjXMMRegs) {
794 default: assert(0 && "Unhandled argument type!");
798 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][NumIntRegs];
799 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
800 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
809 assert(!isStdCall && "Unhandled argument type!");
810 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
811 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
814 NumIntRegs += ObjIntRegs;
815 NumXMMRegs += ObjXMMRegs;
818 // XMM arguments have to be aligned on 16-byte boundary.
820 ArgOffset = ((ArgOffset + 15) / 16) * 16;
821 // Create the SelectionDAG nodes corresponding to a load from this
823 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
824 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
825 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
827 ArgOffset += ArgIncrement; // Move on to the next argument.
829 NumSRetBytes += ArgIncrement;
832 ArgValues.push_back(ArgValue);
835 ArgValues.push_back(Root);
837 // If the function takes variable number of arguments, make a frame index for
838 // the start of the first vararg value... for expansion of llvm.va_start.
840 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
842 if (isStdCall && !isVarArg) {
843 BytesToPopOnReturn = ArgOffset; // Callee pops everything..
844 BytesCallerReserves = 0;
846 BytesToPopOnReturn = NumSRetBytes; // Callee pops hidden struct pointer.
847 BytesCallerReserves = ArgOffset;
850 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
851 ReturnAddrIndex = 0; // No return address slot generated yet.
854 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
856 // Return the new list of results.
857 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
858 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
861 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
863 SDOperand Chain = Op.getOperand(0);
864 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
865 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
866 SDOperand Callee = Op.getOperand(4);
867 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
869 static const unsigned XMMArgRegs[] = {
870 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
872 static const unsigned GPR32ArgRegs[] = {
873 X86::EAX, X86::EDX, X86::ECX
876 // Count how many bytes are to be pushed on the stack.
877 unsigned NumBytes = 0;
878 // Keep track of the number of integer regs passed so far.
879 unsigned NumIntRegs = 0;
880 // Keep track of the number of XMM regs passed so far.
881 unsigned NumXMMRegs = 0;
882 // How much bytes on stack used for struct return
883 unsigned NumSRetBytes= 0;
885 // Handle regparm attribute
886 SmallVector<bool, 8> ArgInRegs(NumOps, false);
887 SmallVector<bool, 8> SRetArgs(NumOps, false);
888 for (unsigned i = 0; i<NumOps; ++i) {
890 dyn_cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
891 ArgInRegs[i] = (Flags >> 1) & 1;
892 SRetArgs[i] = (Flags >> 2) & 1;
895 // Calculate stack frame size
896 for (unsigned i = 0; i != NumOps; ++i) {
897 SDOperand Arg = Op.getOperand(5+2*i);
898 unsigned ArgIncrement = 4;
899 unsigned ObjSize = 0;
900 unsigned ObjIntRegs = 0;
901 unsigned ObjXMMRegs = 0;
903 HowToPassCallArgument(Arg.getValueType(),
905 NumIntRegs, NumXMMRegs, 3,
906 ObjSize, ObjIntRegs, ObjXMMRegs);
908 ArgIncrement = ObjSize;
910 NumIntRegs += ObjIntRegs;
911 NumXMMRegs += ObjXMMRegs;
913 // XMM arguments have to be aligned on 16-byte boundary.
915 NumBytes = ((NumBytes + 15) / 16) * 16;
916 NumBytes += ArgIncrement;
920 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
922 // Arguments go on the stack in reverse order, as specified by the ABI.
923 unsigned ArgOffset = 0;
926 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
927 SmallVector<SDOperand, 8> MemOpChains;
928 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
929 for (unsigned i = 0; i != NumOps; ++i) {
930 SDOperand Arg = Op.getOperand(5+2*i);
931 unsigned ArgIncrement = 4;
932 unsigned ObjSize = 0;
933 unsigned ObjIntRegs = 0;
934 unsigned ObjXMMRegs = 0;
936 HowToPassCallArgument(Arg.getValueType(),
938 NumIntRegs, NumXMMRegs, 3,
939 ObjSize, ObjIntRegs, ObjXMMRegs);
942 ArgIncrement = ObjSize;
944 if (Arg.getValueType() == MVT::i8 || Arg.getValueType() == MVT::i16) {
945 // Promote the integer to 32 bits. If the input type is signed use a
946 // sign extend, otherwise use a zero extend.
947 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
949 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
950 Arg = DAG.getNode(ExtOp, MVT::i32, Arg);
953 if (ObjIntRegs || ObjXMMRegs) {
954 switch (Arg.getValueType()) {
955 default: assert(0 && "Unhandled argument type!");
957 RegsToPass.push_back(std::make_pair(GPR32ArgRegs[NumIntRegs], Arg));
965 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
969 NumIntRegs += ObjIntRegs;
970 NumXMMRegs += ObjXMMRegs;
973 // XMM arguments have to be aligned on 16-byte boundary.
975 ArgOffset = ((ArgOffset + 15) / 16) * 16;
977 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
978 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
979 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
981 ArgOffset += ArgIncrement; // Move on to the next argument.
983 NumSRetBytes += ArgIncrement;
987 // Sanity check: we haven't seen NumSRetBytes > 4
988 assert((NumSRetBytes<=4) &&
989 "Too much space for struct-return pointer requested");
991 if (!MemOpChains.empty())
992 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
993 &MemOpChains[0], MemOpChains.size());
995 // Build a sequence of copy-to-reg nodes chained together with token chain
996 // and flag operands which copy the outgoing args into registers.
998 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
999 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1001 InFlag = Chain.getValue(1);
1004 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1006 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1007 Subtarget->isPICStyleGOT()) {
1008 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1009 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1011 InFlag = Chain.getValue(1);
1014 // If the callee is a GlobalAddress node (quite common, every direct call is)
1015 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1016 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1017 // We should use extra load for direct calls to dllimported functions in
1019 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1020 getTargetMachine(), true))
1021 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1022 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1023 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1025 // Returns a chain & a flag for retval copy to use.
1026 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1027 SmallVector<SDOperand, 8> Ops;
1028 Ops.push_back(Chain);
1029 Ops.push_back(Callee);
1031 // Add argument registers to the end of the list so that they are known live
1033 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1034 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1035 RegsToPass[i].second.getValueType()));
1037 // Add an implicit use GOT pointer in EBX.
1038 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1039 Subtarget->isPICStyleGOT())
1040 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1043 Ops.push_back(InFlag);
1045 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1046 NodeTys, &Ops[0], Ops.size());
1047 InFlag = Chain.getValue(1);
1049 // Create the CALLSEQ_END node.
1050 unsigned NumBytesForCalleeToPush = 0;
1052 if (CC == CallingConv::X86_StdCall) {
1054 NumBytesForCalleeToPush = NumSRetBytes;
1056 NumBytesForCalleeToPush = NumBytes;
1058 // If this is is a call to a struct-return function, the callee
1059 // pops the hidden struct pointer, so we have to push it back.
1060 // This is common for Darwin/X86, Linux & Mingw32 targets.
1061 NumBytesForCalleeToPush = NumSRetBytes;
1064 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1066 Ops.push_back(Chain);
1067 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1068 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
1069 Ops.push_back(InFlag);
1070 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1071 InFlag = Chain.getValue(1);
1073 // Handle result values, copying them out of physregs into vregs that we
1075 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1079 //===----------------------------------------------------------------------===//
1080 // X86-64 C Calling Convention implementation
1081 //===----------------------------------------------------------------------===//
1084 /// X86_64_CCC_AssignArgument - Implement the X86-64 C Calling Convention. This
1085 /// returns true if the value was not handled by this calling convention.
1086 static bool X86_64_CCC_AssignArgument(unsigned ValNo,
1087 MVT::ValueType ArgVT, unsigned ArgFlags,
1089 MVT::ValueType LocVT = ArgVT;
1090 CCValAssign::LocInfo LocInfo = CCValAssign::Full;
1092 // Promote the integer to 32 bits. If the input type is signed use a
1093 // sign extend, otherwise use a zero extend.
1094 if (ArgVT == MVT::i8 || ArgVT == MVT::i16) {
1096 LocInfo = (ArgFlags & 1) ? CCValAssign::SExt : CCValAssign::ZExt;
1099 // If this is a 32-bit value, assign to a 32-bit register if any are
1101 if (LocVT == MVT::i32) {
1102 static const unsigned GPR32ArgRegs[] = {
1103 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
1105 if (unsigned Reg = State.AllocateReg(GPR32ArgRegs, 6)) {
1106 State.addLoc(CCValAssign::getReg(ValNo, ArgVT, Reg, LocVT, LocInfo));
1111 // If this is a 64-bit value, assign to a 64-bit register if any are
1113 if (LocVT == MVT::i64) {
1114 static const unsigned GPR64ArgRegs[] = {
1115 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1117 if (unsigned Reg = State.AllocateReg(GPR64ArgRegs, 6)) {
1118 State.addLoc(CCValAssign::getReg(ValNo, ArgVT, Reg, LocVT, LocInfo));
1123 // If this is a FP or vector type, assign to an XMM reg if any are
1125 if (MVT::isVector(LocVT) || MVT::isFloatingPoint(LocVT)) {
1126 static const unsigned XMMArgRegs[] = {
1127 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1128 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1130 if (unsigned Reg = State.AllocateReg(XMMArgRegs, 8)) {
1131 State.addLoc(CCValAssign::getReg(ValNo, ArgVT, Reg, LocVT, LocInfo));
1136 // Integer/FP values get stored in stack slots that are 8 bytes in size and
1137 // 8-byte aligned if there are no more registers to hold them.
1138 if (LocVT == MVT::i32 || LocVT == MVT::i64 ||
1139 LocVT == MVT::f32 || LocVT == MVT::f64) {
1140 unsigned Offset = State.AllocateStack(8, 8);
1141 State.addLoc(CCValAssign::getMem(ValNo, ArgVT, Offset, LocVT, LocInfo));
1145 // Vectors get 16-byte stack slots that are 16-byte aligned.
1146 if (MVT::isVector(LocVT)) {
1147 unsigned Offset = State.AllocateStack(16, 16);
1148 State.addLoc(CCValAssign::getMem(ValNo, ArgVT, Offset, LocVT, LocInfo));
1156 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1157 unsigned NumArgs = Op.Val->getNumValues() - 1;
1158 MachineFunction &MF = DAG.getMachineFunction();
1159 MachineFrameInfo *MFI = MF.getFrameInfo();
1160 SDOperand Root = Op.getOperand(0);
1161 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1163 static const unsigned GPR64ArgRegs[] = {
1164 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1166 static const unsigned XMMArgRegs[] = {
1167 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1168 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1171 SmallVector<CCValAssign, 16> ArgLocs;
1172 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
1175 for (unsigned i = 0; i != NumArgs; ++i) {
1176 MVT::ValueType ArgVT = Op.getValue(i).getValueType();
1177 unsigned ArgFlags = cast<ConstantSDNode>(Op.getOperand(3+i))->getValue();
1178 if (X86_64_CCC_AssignArgument(i, ArgVT, ArgFlags, CCInfo))
1179 assert(0 && "Unhandled argument type!");
1182 SmallVector<SDOperand, 8> ArgValues;
1183 unsigned LastVal = ~0U;
1184 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1185 CCValAssign &VA = ArgLocs[i];
1186 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1188 assert(VA.getValNo() != LastVal &&
1189 "Don't support value assigned to multiple locs yet");
1190 LastVal = VA.getValNo();
1192 if (VA.isRegLoc()) {
1193 MVT::ValueType RegVT = VA.getLocVT();
1194 TargetRegisterClass *RC;
1195 if (RegVT == MVT::i32)
1196 RC = X86::GR32RegisterClass;
1197 else if (RegVT == MVT::i64)
1198 RC = X86::GR64RegisterClass;
1199 else if (RegVT == MVT::f32)
1200 RC = X86::FR32RegisterClass;
1201 else if (RegVT == MVT::f64)
1202 RC = X86::FR64RegisterClass;
1204 assert(MVT::isVector(RegVT));
1205 RC = X86::VR128RegisterClass;
1208 SDOperand ArgValue = DAG.getCopyFromReg(Root, VA.getLocReg(), RegVT);
1209 AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1211 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1212 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1214 if (VA.getLocInfo() == CCValAssign::SExt)
1215 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1216 DAG.getValueType(VA.getValVT()));
1217 else if (VA.getLocInfo() == CCValAssign::ZExt)
1218 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1219 DAG.getValueType(VA.getValVT()));
1221 if (VA.getLocInfo() != CCValAssign::Full)
1222 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1224 ArgValues.push_back(ArgValue);
1226 assert(VA.isMemLoc());
1228 // Create the nodes corresponding to a load from this parameter slot.
1229 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1230 VA.getLocMemOffset());
1231 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1232 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
1236 unsigned StackSize = CCInfo.getNextStackOffset();
1238 // If the function takes variable number of arguments, make a frame index for
1239 // the start of the first vararg value... for expansion of llvm.va_start.
1241 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1242 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1244 // For X86-64, if there are vararg parameters that are passed via
1245 // registers, then we must store them to their spots on the stack so they
1246 // may be loaded by deferencing the result of va_next.
1247 VarArgsGPOffset = NumIntRegs * 8;
1248 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1249 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1250 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1252 // Store the integer parameter registers.
1253 SmallVector<SDOperand, 8> MemOps;
1254 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1255 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1256 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1257 for (; NumIntRegs != 6; ++NumIntRegs) {
1258 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1259 X86::GR64RegisterClass);
1260 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1261 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1262 MemOps.push_back(Store);
1263 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1264 DAG.getConstant(8, getPointerTy()));
1267 // Now store the XMM (fp + vector) parameter registers.
1268 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1269 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1270 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1271 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1272 X86::VR128RegisterClass);
1273 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1274 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1275 MemOps.push_back(Store);
1276 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1277 DAG.getConstant(16, getPointerTy()));
1279 if (!MemOps.empty())
1280 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1281 &MemOps[0], MemOps.size());
1284 ArgValues.push_back(Root);
1286 ReturnAddrIndex = 0; // No return address slot generated yet.
1287 BytesToPopOnReturn = 0; // Callee pops nothing.
1288 BytesCallerReserves = StackSize;
1290 // Return the new list of results.
1291 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1292 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1296 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1298 SDOperand Chain = Op.getOperand(0);
1299 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1300 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1301 SDOperand Callee = Op.getOperand(4);
1302 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1304 SmallVector<CCValAssign, 16> ArgLocs;
1305 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
1307 for (unsigned i = 0; i != NumOps; ++i) {
1308 MVT::ValueType ArgVT = Op.getOperand(5+2*i).getValueType();
1309 unsigned ArgFlags =cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1310 if (X86_64_CCC_AssignArgument(i, ArgVT, ArgFlags, CCInfo))
1311 assert(0 && "Unhandled argument type!");
1314 // Get a count of how many bytes are to be pushed on the stack.
1315 unsigned NumBytes = CCInfo.getNextStackOffset();
1316 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1318 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1319 SmallVector<SDOperand, 8> MemOpChains;
1323 // Walk the register/memloc assignments, inserting copies/loads.
1324 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1325 CCValAssign &VA = ArgLocs[i];
1326 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1328 // Promote the value if needed.
1329 switch (VA.getLocInfo()) {
1330 default: assert(0 && "Unknown loc info!");
1331 case CCValAssign::Full: break;
1332 case CCValAssign::SExt:
1333 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1335 case CCValAssign::ZExt:
1336 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1338 case CCValAssign::AExt:
1339 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1343 if (VA.isRegLoc()) {
1344 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1346 assert(VA.isMemLoc());
1347 if (StackPtr.Val == 0)
1348 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1349 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1350 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1351 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1355 if (!MemOpChains.empty())
1356 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1357 &MemOpChains[0], MemOpChains.size());
1359 // Build a sequence of copy-to-reg nodes chained together with token chain
1360 // and flag operands which copy the outgoing args into registers.
1362 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1363 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1365 InFlag = Chain.getValue(1);
1369 // From AMD64 ABI document:
1370 // For calls that may call functions that use varargs or stdargs
1371 // (prototype-less calls or calls to functions containing ellipsis (...) in
1372 // the declaration) %al is used as hidden argument to specify the number
1373 // of SSE registers used. The contents of %al do not need to match exactly
1374 // the number of registers, but must be an ubound on the number of SSE
1375 // registers used and is in the range 0 - 8 inclusive.
1377 // Count the number of XMM registers allocated.
1378 static const unsigned XMMArgRegs[] = {
1379 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1380 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1382 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1384 Chain = DAG.getCopyToReg(Chain, X86::AL,
1385 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1386 InFlag = Chain.getValue(1);
1389 // If the callee is a GlobalAddress node (quite common, every direct call is)
1390 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1391 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1392 // We should use extra load for direct calls to dllimported functions in
1394 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1395 getTargetMachine(), true))
1396 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1397 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1398 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1400 // Returns a chain & a flag for retval copy to use.
1401 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1402 SmallVector<SDOperand, 8> Ops;
1403 Ops.push_back(Chain);
1404 Ops.push_back(Callee);
1406 // Add argument registers to the end of the list so that they are known live
1408 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1409 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1410 RegsToPass[i].second.getValueType()));
1413 Ops.push_back(InFlag);
1415 // FIXME: Do not generate X86ISD::TAILCALL for now.
1416 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1417 NodeTys, &Ops[0], Ops.size());
1418 InFlag = Chain.getValue(1);
1420 // Returns a flag for retval copy to use.
1421 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1423 Ops.push_back(Chain);
1424 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1425 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1426 Ops.push_back(InFlag);
1427 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1428 InFlag = Chain.getValue(1);
1430 // Handle result values, copying them out of physregs into vregs that we
1432 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1435 //===----------------------------------------------------------------------===//
1436 // Fast & FastCall Calling Convention implementation
1437 //===----------------------------------------------------------------------===//
1439 // The X86 'fast' calling convention passes up to two integer arguments in
1440 // registers (an appropriate portion of EAX/EDX), passes arguments in C order,
1441 // and requires that the callee pop its arguments off the stack (allowing proper
1442 // tail calls), and has the same return value conventions as C calling convs.
1444 // This calling convention always arranges for the callee pop value to be 8n+4
1445 // bytes, which is needed for tail recursion elimination and stack alignment
1448 // Note that this can be enhanced in the future to pass fp vals in registers
1449 // (when we have a global fp allocator) and do other tricks.
1451 //===----------------------------------------------------------------------===//
1452 // The X86 'fastcall' calling convention passes up to two integer arguments in
1453 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
1454 // and requires that the callee pop its arguments off the stack (allowing proper
1455 // tail calls), and has the same return value conventions as C calling convs.
1457 // This calling convention always arranges for the callee pop value to be 8n+4
1458 // bytes, which is needed for tail recursion elimination and stack alignment
1461 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG,
1463 unsigned NumArgs = Op.Val->getNumValues()-1;
1464 MachineFunction &MF = DAG.getMachineFunction();
1465 MachineFrameInfo *MFI = MF.getFrameInfo();
1466 SDOperand Root = Op.getOperand(0);
1467 SmallVector<SDOperand, 8> ArgValues;
1469 // Add DAG nodes to load the arguments... On entry to a function the stack
1470 // frame looks like this:
1472 // [ESP] -- return address
1473 // [ESP + 4] -- first nonreg argument (leftmost lexically)
1474 // [ESP + 8] -- second nonreg argument, if 1st argument is <= 4 bytes in size
1476 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1478 // Keep track of the number of integer regs passed so far. This can be either
1479 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1481 unsigned NumIntRegs = 0;
1482 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1484 static const unsigned XMMArgRegs[] = {
1485 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1488 static const unsigned GPRArgRegs[][2][2] = {
1489 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1490 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1491 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1494 static const TargetRegisterClass* GPRClasses[3] = {
1495 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
1498 unsigned GPRInd = (isFastCall ? 1 : 0);
1499 for (unsigned i = 0; i < NumArgs; ++i) {
1500 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
1501 unsigned ArgIncrement = 4;
1502 unsigned ObjSize = 0;
1503 unsigned ObjXMMRegs = 0;
1504 unsigned ObjIntRegs = 0;
1508 HowToPassCallArgument(ObjectVT,
1509 true, // Use as much registers as possible
1510 NumIntRegs, NumXMMRegs,
1511 (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS),
1512 ObjSize, ObjIntRegs, ObjXMMRegs);
1515 ArgIncrement = ObjSize;
1517 if (ObjIntRegs || ObjXMMRegs) {
1519 default: assert(0 && "Unhandled argument type!");
1523 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][GPRInd][NumIntRegs];
1524 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
1525 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1534 assert(!isFastCall && "Unhandled argument type!");
1535 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
1536 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1540 NumIntRegs += ObjIntRegs;
1541 NumXMMRegs += ObjXMMRegs;
1544 // XMM arguments have to be aligned on 16-byte boundary.
1546 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1547 // Create the SelectionDAG nodes corresponding to a load from this
1549 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1550 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1551 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1553 ArgOffset += ArgIncrement; // Move on to the next argument.
1556 ArgValues.push_back(ArgValue);
1559 ArgValues.push_back(Root);
1561 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1562 // arguments and the arguments after the retaddr has been pushed are aligned.
1563 if ((ArgOffset & 7) == 0)
1566 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1567 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1568 ReturnAddrIndex = 0; // No return address slot generated yet.
1569 BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
1570 BytesCallerReserves = 0;
1572 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
1574 // Finally, inform the code generator which regs we return values in.
1575 switch (getValueType(MF.getFunction()->getReturnType())) {
1576 default: assert(0 && "Unknown type!");
1577 case MVT::isVoid: break;
1582 MF.addLiveOut(X86::EAX);
1585 MF.addLiveOut(X86::EAX);
1586 MF.addLiveOut(X86::EDX);
1590 MF.addLiveOut(X86::ST0);
1598 assert(!isFastCall && "Unknown result type");
1599 MF.addLiveOut(X86::XMM0);
1603 // Return the new list of results.
1604 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1605 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1608 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1610 SDOperand Chain = Op.getOperand(0);
1611 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1612 SDOperand Callee = Op.getOperand(4);
1613 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1615 // Count how many bytes are to be pushed on the stack.
1616 unsigned NumBytes = 0;
1618 // Keep track of the number of integer regs passed so far. This can be either
1619 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1621 unsigned NumIntRegs = 0;
1622 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1624 static const unsigned GPRArgRegs[][2][2] = {
1625 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1626 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1627 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1629 static const unsigned XMMArgRegs[] = {
1630 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1633 bool isFastCall = CC == CallingConv::X86_FastCall;
1634 unsigned GPRInd = isFastCall ? 1 : 0;
1635 for (unsigned i = 0; i != NumOps; ++i) {
1636 SDOperand Arg = Op.getOperand(5+2*i);
1638 switch (Arg.getValueType()) {
1639 default: assert(0 && "Unknown value type!");
1643 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1644 if (NumIntRegs < MaxNumIntRegs) {
1661 assert(!isFastCall && "Unknown value type!");
1665 // XMM arguments have to be aligned on 16-byte boundary.
1666 NumBytes = ((NumBytes + 15) / 16) * 16;
1673 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1674 // arguments and the arguments after the retaddr has been pushed are aligned.
1675 if ((NumBytes & 7) == 0)
1678 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1680 // Arguments go on the stack in reverse order, as specified by the ABI.
1681 unsigned ArgOffset = 0;
1683 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1684 SmallVector<SDOperand, 8> MemOpChains;
1685 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1686 for (unsigned i = 0; i != NumOps; ++i) {
1687 SDOperand Arg = Op.getOperand(5+2*i);
1689 switch (Arg.getValueType()) {
1690 default: assert(0 && "Unexpected ValueType for argument!");
1694 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1695 if (NumIntRegs < MaxNumIntRegs) {
1697 GPRArgRegs[Arg.getValueType()-MVT::i8][GPRInd][NumIntRegs];
1698 RegsToPass.push_back(std::make_pair(RegToUse, Arg));
1704 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1705 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1706 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1711 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1712 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1713 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1723 assert(!isFastCall && "Unexpected ValueType for argument!");
1724 if (NumXMMRegs < 4) {
1725 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1728 // XMM arguments have to be aligned on 16-byte boundary.
1729 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1730 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1731 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1732 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1739 if (!MemOpChains.empty())
1740 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1741 &MemOpChains[0], MemOpChains.size());
1743 // Build a sequence of copy-to-reg nodes chained together with token chain
1744 // and flag operands which copy the outgoing args into registers.
1746 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1747 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1749 InFlag = Chain.getValue(1);
1752 // If the callee is a GlobalAddress node (quite common, every direct call is)
1753 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1754 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1755 // We should use extra load for direct calls to dllimported functions in
1757 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1758 getTargetMachine(), true))
1759 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1760 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1761 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1763 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1765 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1766 Subtarget->isPICStyleGOT()) {
1767 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1768 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1770 InFlag = Chain.getValue(1);
1773 // Returns a chain & a flag for retval copy to use.
1774 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1775 SmallVector<SDOperand, 8> Ops;
1776 Ops.push_back(Chain);
1777 Ops.push_back(Callee);
1779 // Add argument registers to the end of the list so that they are known live
1781 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1782 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1783 RegsToPass[i].second.getValueType()));
1785 // Add an implicit use GOT pointer in EBX.
1786 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1787 Subtarget->isPICStyleGOT())
1788 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1791 Ops.push_back(InFlag);
1793 // FIXME: Do not generate X86ISD::TAILCALL for now.
1794 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1795 NodeTys, &Ops[0], Ops.size());
1796 InFlag = Chain.getValue(1);
1798 // Returns a flag for retval copy to use.
1799 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1801 Ops.push_back(Chain);
1802 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1803 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1804 Ops.push_back(InFlag);
1805 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1806 InFlag = Chain.getValue(1);
1808 // Handle result values, copying them out of physregs into vregs that we
1810 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1813 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1814 if (ReturnAddrIndex == 0) {
1815 // Set up a frame object for the return address.
1816 MachineFunction &MF = DAG.getMachineFunction();
1817 if (Subtarget->is64Bit())
1818 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1820 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1823 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1828 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1829 /// specific condition code. It returns a false if it cannot do a direct
1830 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1832 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1833 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1834 SelectionDAG &DAG) {
1835 X86CC = X86::COND_INVALID;
1837 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1838 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1839 // X > -1 -> X == 0, jump !sign.
1840 RHS = DAG.getConstant(0, RHS.getValueType());
1841 X86CC = X86::COND_NS;
1843 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1844 // X < 0 -> X == 0, jump on sign.
1845 X86CC = X86::COND_S;
1850 switch (SetCCOpcode) {
1852 case ISD::SETEQ: X86CC = X86::COND_E; break;
1853 case ISD::SETGT: X86CC = X86::COND_G; break;
1854 case ISD::SETGE: X86CC = X86::COND_GE; break;
1855 case ISD::SETLT: X86CC = X86::COND_L; break;
1856 case ISD::SETLE: X86CC = X86::COND_LE; break;
1857 case ISD::SETNE: X86CC = X86::COND_NE; break;
1858 case ISD::SETULT: X86CC = X86::COND_B; break;
1859 case ISD::SETUGT: X86CC = X86::COND_A; break;
1860 case ISD::SETULE: X86CC = X86::COND_BE; break;
1861 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1864 // On a floating point condition, the flags are set as follows:
1866 // 0 | 0 | 0 | X > Y
1867 // 0 | 0 | 1 | X < Y
1868 // 1 | 0 | 0 | X == Y
1869 // 1 | 1 | 1 | unordered
1871 switch (SetCCOpcode) {
1874 case ISD::SETEQ: X86CC = X86::COND_E; break;
1875 case ISD::SETOLT: Flip = true; // Fallthrough
1877 case ISD::SETGT: X86CC = X86::COND_A; break;
1878 case ISD::SETOLE: Flip = true; // Fallthrough
1880 case ISD::SETGE: X86CC = X86::COND_AE; break;
1881 case ISD::SETUGT: Flip = true; // Fallthrough
1883 case ISD::SETLT: X86CC = X86::COND_B; break;
1884 case ISD::SETUGE: Flip = true; // Fallthrough
1886 case ISD::SETLE: X86CC = X86::COND_BE; break;
1888 case ISD::SETNE: X86CC = X86::COND_NE; break;
1889 case ISD::SETUO: X86CC = X86::COND_P; break;
1890 case ISD::SETO: X86CC = X86::COND_NP; break;
1893 std::swap(LHS, RHS);
1896 return X86CC != X86::COND_INVALID;
1899 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1900 /// code. Current x86 isa includes the following FP cmov instructions:
1901 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1902 static bool hasFPCMov(unsigned X86CC) {
1918 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1919 /// true if Op is undef or if its value falls within the specified range (L, H].
1920 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1921 if (Op.getOpcode() == ISD::UNDEF)
1924 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1925 return (Val >= Low && Val < Hi);
1928 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1929 /// true if Op is undef or if its value equal to the specified value.
1930 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1931 if (Op.getOpcode() == ISD::UNDEF)
1933 return cast<ConstantSDNode>(Op)->getValue() == Val;
1936 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1937 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1938 bool X86::isPSHUFDMask(SDNode *N) {
1939 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1941 if (N->getNumOperands() != 4)
1944 // Check if the value doesn't reference the second vector.
1945 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1946 SDOperand Arg = N->getOperand(i);
1947 if (Arg.getOpcode() == ISD::UNDEF) continue;
1948 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1949 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
1956 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1957 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1958 bool X86::isPSHUFHWMask(SDNode *N) {
1959 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1961 if (N->getNumOperands() != 8)
1964 // Lower quadword copied in order.
1965 for (unsigned i = 0; i != 4; ++i) {
1966 SDOperand Arg = N->getOperand(i);
1967 if (Arg.getOpcode() == ISD::UNDEF) continue;
1968 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1969 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1973 // Upper quadword shuffled.
1974 for (unsigned i = 4; i != 8; ++i) {
1975 SDOperand Arg = N->getOperand(i);
1976 if (Arg.getOpcode() == ISD::UNDEF) continue;
1977 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1978 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1979 if (Val < 4 || Val > 7)
1986 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1987 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1988 bool X86::isPSHUFLWMask(SDNode *N) {
1989 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1991 if (N->getNumOperands() != 8)
1994 // Upper quadword copied in order.
1995 for (unsigned i = 4; i != 8; ++i)
1996 if (!isUndefOrEqual(N->getOperand(i), i))
1999 // Lower quadword shuffled.
2000 for (unsigned i = 0; i != 4; ++i)
2001 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2007 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2008 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2009 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2010 if (NumElems != 2 && NumElems != 4) return false;
2012 unsigned Half = NumElems / 2;
2013 for (unsigned i = 0; i < Half; ++i)
2014 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2016 for (unsigned i = Half; i < NumElems; ++i)
2017 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2023 bool X86::isSHUFPMask(SDNode *N) {
2024 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2025 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2028 /// isCommutedSHUFP - Returns true if the shuffle mask is except
2029 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2030 /// half elements to come from vector 1 (which would equal the dest.) and
2031 /// the upper half to come from vector 2.
2032 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2033 if (NumOps != 2 && NumOps != 4) return false;
2035 unsigned Half = NumOps / 2;
2036 for (unsigned i = 0; i < Half; ++i)
2037 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2039 for (unsigned i = Half; i < NumOps; ++i)
2040 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2045 static bool isCommutedSHUFP(SDNode *N) {
2046 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2047 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2050 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2051 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2052 bool X86::isMOVHLPSMask(SDNode *N) {
2053 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2055 if (N->getNumOperands() != 4)
2058 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2059 return isUndefOrEqual(N->getOperand(0), 6) &&
2060 isUndefOrEqual(N->getOperand(1), 7) &&
2061 isUndefOrEqual(N->getOperand(2), 2) &&
2062 isUndefOrEqual(N->getOperand(3), 3);
2065 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2066 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2068 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2069 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2071 if (N->getNumOperands() != 4)
2074 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2075 return isUndefOrEqual(N->getOperand(0), 2) &&
2076 isUndefOrEqual(N->getOperand(1), 3) &&
2077 isUndefOrEqual(N->getOperand(2), 2) &&
2078 isUndefOrEqual(N->getOperand(3), 3);
2081 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2082 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2083 bool X86::isMOVLPMask(SDNode *N) {
2084 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2086 unsigned NumElems = N->getNumOperands();
2087 if (NumElems != 2 && NumElems != 4)
2090 for (unsigned i = 0; i < NumElems/2; ++i)
2091 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2094 for (unsigned i = NumElems/2; i < NumElems; ++i)
2095 if (!isUndefOrEqual(N->getOperand(i), i))
2101 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2102 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2104 bool X86::isMOVHPMask(SDNode *N) {
2105 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2107 unsigned NumElems = N->getNumOperands();
2108 if (NumElems != 2 && NumElems != 4)
2111 for (unsigned i = 0; i < NumElems/2; ++i)
2112 if (!isUndefOrEqual(N->getOperand(i), i))
2115 for (unsigned i = 0; i < NumElems/2; ++i) {
2116 SDOperand Arg = N->getOperand(i + NumElems/2);
2117 if (!isUndefOrEqual(Arg, i + NumElems))
2124 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2125 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2126 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2127 bool V2IsSplat = false) {
2128 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2131 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2132 SDOperand BitI = Elts[i];
2133 SDOperand BitI1 = Elts[i+1];
2134 if (!isUndefOrEqual(BitI, j))
2137 if (isUndefOrEqual(BitI1, NumElts))
2140 if (!isUndefOrEqual(BitI1, j + NumElts))
2148 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2149 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2150 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2153 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2154 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2155 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2156 bool V2IsSplat = false) {
2157 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2160 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2161 SDOperand BitI = Elts[i];
2162 SDOperand BitI1 = Elts[i+1];
2163 if (!isUndefOrEqual(BitI, j + NumElts/2))
2166 if (isUndefOrEqual(BitI1, NumElts))
2169 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2177 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2178 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2179 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2182 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2183 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2185 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2186 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2188 unsigned NumElems = N->getNumOperands();
2189 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
2192 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2193 SDOperand BitI = N->getOperand(i);
2194 SDOperand BitI1 = N->getOperand(i+1);
2196 if (!isUndefOrEqual(BitI, j))
2198 if (!isUndefOrEqual(BitI1, j))
2205 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2206 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2207 /// MOVSD, and MOVD, i.e. setting the lowest element.
2208 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2209 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2212 if (!isUndefOrEqual(Elts[0], NumElts))
2215 for (unsigned i = 1; i < NumElts; ++i) {
2216 if (!isUndefOrEqual(Elts[i], i))
2223 bool X86::isMOVLMask(SDNode *N) {
2224 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2225 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2228 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2229 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2230 /// element of vector 2 and the other elements to come from vector 1 in order.
2231 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2232 bool V2IsSplat = false,
2233 bool V2IsUndef = false) {
2234 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2237 if (!isUndefOrEqual(Ops[0], 0))
2240 for (unsigned i = 1; i < NumOps; ++i) {
2241 SDOperand Arg = Ops[i];
2242 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2243 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2244 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2251 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2252 bool V2IsUndef = false) {
2253 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2254 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2255 V2IsSplat, V2IsUndef);
2258 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2259 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2260 bool X86::isMOVSHDUPMask(SDNode *N) {
2261 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2263 if (N->getNumOperands() != 4)
2266 // Expect 1, 1, 3, 3
2267 for (unsigned i = 0; i < 2; ++i) {
2268 SDOperand Arg = N->getOperand(i);
2269 if (Arg.getOpcode() == ISD::UNDEF) continue;
2270 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2271 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2272 if (Val != 1) return false;
2276 for (unsigned i = 2; i < 4; ++i) {
2277 SDOperand Arg = N->getOperand(i);
2278 if (Arg.getOpcode() == ISD::UNDEF) continue;
2279 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2280 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2281 if (Val != 3) return false;
2285 // Don't use movshdup if it can be done with a shufps.
2289 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2290 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2291 bool X86::isMOVSLDUPMask(SDNode *N) {
2292 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2294 if (N->getNumOperands() != 4)
2297 // Expect 0, 0, 2, 2
2298 for (unsigned i = 0; i < 2; ++i) {
2299 SDOperand Arg = N->getOperand(i);
2300 if (Arg.getOpcode() == ISD::UNDEF) continue;
2301 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2302 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2303 if (Val != 0) return false;
2307 for (unsigned i = 2; i < 4; ++i) {
2308 SDOperand Arg = N->getOperand(i);
2309 if (Arg.getOpcode() == ISD::UNDEF) continue;
2310 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2311 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2312 if (Val != 2) return false;
2316 // Don't use movshdup if it can be done with a shufps.
2320 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2321 /// a splat of a single element.
2322 static bool isSplatMask(SDNode *N) {
2323 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2325 // This is a splat operation if each element of the permute is the same, and
2326 // if the value doesn't reference the second vector.
2327 unsigned NumElems = N->getNumOperands();
2328 SDOperand ElementBase;
2330 for (; i != NumElems; ++i) {
2331 SDOperand Elt = N->getOperand(i);
2332 if (isa<ConstantSDNode>(Elt)) {
2338 if (!ElementBase.Val)
2341 for (; i != NumElems; ++i) {
2342 SDOperand Arg = N->getOperand(i);
2343 if (Arg.getOpcode() == ISD::UNDEF) continue;
2344 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2345 if (Arg != ElementBase) return false;
2348 // Make sure it is a splat of the first vector operand.
2349 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2352 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2353 /// a splat of a single element and it's a 2 or 4 element mask.
2354 bool X86::isSplatMask(SDNode *N) {
2355 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2357 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2358 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2360 return ::isSplatMask(N);
2363 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2364 /// specifies a splat of zero element.
2365 bool X86::isSplatLoMask(SDNode *N) {
2366 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2368 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2369 if (!isUndefOrEqual(N->getOperand(i), 0))
2374 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2375 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2377 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2378 unsigned NumOperands = N->getNumOperands();
2379 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2381 for (unsigned i = 0; i < NumOperands; ++i) {
2383 SDOperand Arg = N->getOperand(NumOperands-i-1);
2384 if (Arg.getOpcode() != ISD::UNDEF)
2385 Val = cast<ConstantSDNode>(Arg)->getValue();
2386 if (Val >= NumOperands) Val -= NumOperands;
2388 if (i != NumOperands - 1)
2395 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2396 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2398 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2400 // 8 nodes, but we only care about the last 4.
2401 for (unsigned i = 7; i >= 4; --i) {
2403 SDOperand Arg = N->getOperand(i);
2404 if (Arg.getOpcode() != ISD::UNDEF)
2405 Val = cast<ConstantSDNode>(Arg)->getValue();
2414 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2415 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2417 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2419 // 8 nodes, but we only care about the first 4.
2420 for (int i = 3; i >= 0; --i) {
2422 SDOperand Arg = N->getOperand(i);
2423 if (Arg.getOpcode() != ISD::UNDEF)
2424 Val = cast<ConstantSDNode>(Arg)->getValue();
2433 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2434 /// specifies a 8 element shuffle that can be broken into a pair of
2435 /// PSHUFHW and PSHUFLW.
2436 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2437 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2439 if (N->getNumOperands() != 8)
2442 // Lower quadword shuffled.
2443 for (unsigned i = 0; i != 4; ++i) {
2444 SDOperand Arg = N->getOperand(i);
2445 if (Arg.getOpcode() == ISD::UNDEF) continue;
2446 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2447 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2452 // Upper quadword shuffled.
2453 for (unsigned i = 4; i != 8; ++i) {
2454 SDOperand Arg = N->getOperand(i);
2455 if (Arg.getOpcode() == ISD::UNDEF) continue;
2456 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2457 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2458 if (Val < 4 || Val > 7)
2465 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2466 /// values in ther permute mask.
2467 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2468 SDOperand &V2, SDOperand &Mask,
2469 SelectionDAG &DAG) {
2470 MVT::ValueType VT = Op.getValueType();
2471 MVT::ValueType MaskVT = Mask.getValueType();
2472 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2473 unsigned NumElems = Mask.getNumOperands();
2474 SmallVector<SDOperand, 8> MaskVec;
2476 for (unsigned i = 0; i != NumElems; ++i) {
2477 SDOperand Arg = Mask.getOperand(i);
2478 if (Arg.getOpcode() == ISD::UNDEF) {
2479 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2482 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2483 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2485 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2487 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2491 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2492 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2495 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2496 /// match movhlps. The lower half elements should come from upper half of
2497 /// V1 (and in order), and the upper half elements should come from the upper
2498 /// half of V2 (and in order).
2499 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2500 unsigned NumElems = Mask->getNumOperands();
2503 for (unsigned i = 0, e = 2; i != e; ++i)
2504 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2506 for (unsigned i = 2; i != 4; ++i)
2507 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2512 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2513 /// is promoted to a vector.
2514 static inline bool isScalarLoadToVector(SDNode *N) {
2515 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2516 N = N->getOperand(0).Val;
2517 return ISD::isNON_EXTLoad(N);
2522 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2523 /// match movlp{s|d}. The lower half elements should come from lower half of
2524 /// V1 (and in order), and the upper half elements should come from the upper
2525 /// half of V2 (and in order). And since V1 will become the source of the
2526 /// MOVLP, it must be either a vector load or a scalar load to vector.
2527 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2528 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2530 // Is V2 is a vector load, don't do this transformation. We will try to use
2531 // load folding shufps op.
2532 if (ISD::isNON_EXTLoad(V2))
2535 unsigned NumElems = Mask->getNumOperands();
2536 if (NumElems != 2 && NumElems != 4)
2538 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2539 if (!isUndefOrEqual(Mask->getOperand(i), i))
2541 for (unsigned i = NumElems/2; i != NumElems; ++i)
2542 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2547 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2549 static bool isSplatVector(SDNode *N) {
2550 if (N->getOpcode() != ISD::BUILD_VECTOR)
2553 SDOperand SplatValue = N->getOperand(0);
2554 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2555 if (N->getOperand(i) != SplatValue)
2560 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2562 static bool isUndefShuffle(SDNode *N) {
2563 if (N->getOpcode() != ISD::BUILD_VECTOR)
2566 SDOperand V1 = N->getOperand(0);
2567 SDOperand V2 = N->getOperand(1);
2568 SDOperand Mask = N->getOperand(2);
2569 unsigned NumElems = Mask.getNumOperands();
2570 for (unsigned i = 0; i != NumElems; ++i) {
2571 SDOperand Arg = Mask.getOperand(i);
2572 if (Arg.getOpcode() != ISD::UNDEF) {
2573 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2574 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2576 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2583 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2584 /// that point to V2 points to its first element.
2585 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2586 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2588 bool Changed = false;
2589 SmallVector<SDOperand, 8> MaskVec;
2590 unsigned NumElems = Mask.getNumOperands();
2591 for (unsigned i = 0; i != NumElems; ++i) {
2592 SDOperand Arg = Mask.getOperand(i);
2593 if (Arg.getOpcode() != ISD::UNDEF) {
2594 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2595 if (Val > NumElems) {
2596 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2600 MaskVec.push_back(Arg);
2604 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2605 &MaskVec[0], MaskVec.size());
2609 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2610 /// operation of specified width.
2611 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2612 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2613 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2615 SmallVector<SDOperand, 8> MaskVec;
2616 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2617 for (unsigned i = 1; i != NumElems; ++i)
2618 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2619 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2622 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2623 /// of specified width.
2624 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2625 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2626 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2627 SmallVector<SDOperand, 8> MaskVec;
2628 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2629 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2630 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2632 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2635 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2636 /// of specified width.
2637 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2638 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2639 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2640 unsigned Half = NumElems/2;
2641 SmallVector<SDOperand, 8> MaskVec;
2642 for (unsigned i = 0; i != Half; ++i) {
2643 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2644 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2646 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2649 /// getZeroVector - Returns a vector of specified type with all zero elements.
2651 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2652 assert(MVT::isVector(VT) && "Expected a vector type");
2653 unsigned NumElems = getVectorNumElements(VT);
2654 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2655 bool isFP = MVT::isFloatingPoint(EVT);
2656 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2657 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2658 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2661 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2663 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2664 SDOperand V1 = Op.getOperand(0);
2665 SDOperand Mask = Op.getOperand(2);
2666 MVT::ValueType VT = Op.getValueType();
2667 unsigned NumElems = Mask.getNumOperands();
2668 Mask = getUnpacklMask(NumElems, DAG);
2669 while (NumElems != 4) {
2670 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2673 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2675 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2676 Mask = getZeroVector(MaskVT, DAG);
2677 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2678 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2679 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2682 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2684 static inline bool isZeroNode(SDOperand Elt) {
2685 return ((isa<ConstantSDNode>(Elt) &&
2686 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2687 (isa<ConstantFPSDNode>(Elt) &&
2688 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2691 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2692 /// vector and zero or undef vector.
2693 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2694 unsigned NumElems, unsigned Idx,
2695 bool isZero, SelectionDAG &DAG) {
2696 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2697 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2698 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2699 SDOperand Zero = DAG.getConstant(0, EVT);
2700 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2701 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2702 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2703 &MaskVec[0], MaskVec.size());
2704 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2707 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2709 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2710 unsigned NumNonZero, unsigned NumZero,
2711 SelectionDAG &DAG, TargetLowering &TLI) {
2717 for (unsigned i = 0; i < 16; ++i) {
2718 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2719 if (ThisIsNonZero && First) {
2721 V = getZeroVector(MVT::v8i16, DAG);
2723 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2728 SDOperand ThisElt(0, 0), LastElt(0, 0);
2729 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2730 if (LastIsNonZero) {
2731 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2733 if (ThisIsNonZero) {
2734 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2735 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2736 ThisElt, DAG.getConstant(8, MVT::i8));
2738 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2743 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2744 DAG.getConstant(i/2, TLI.getPointerTy()));
2748 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2751 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2753 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2754 unsigned NumNonZero, unsigned NumZero,
2755 SelectionDAG &DAG, TargetLowering &TLI) {
2761 for (unsigned i = 0; i < 8; ++i) {
2762 bool isNonZero = (NonZeros & (1 << i)) != 0;
2766 V = getZeroVector(MVT::v8i16, DAG);
2768 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2771 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2772 DAG.getConstant(i, TLI.getPointerTy()));
2780 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2781 // All zero's are handled with pxor.
2782 if (ISD::isBuildVectorAllZeros(Op.Val))
2785 // All one's are handled with pcmpeqd.
2786 if (ISD::isBuildVectorAllOnes(Op.Val))
2789 MVT::ValueType VT = Op.getValueType();
2790 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2791 unsigned EVTBits = MVT::getSizeInBits(EVT);
2793 unsigned NumElems = Op.getNumOperands();
2794 unsigned NumZero = 0;
2795 unsigned NumNonZero = 0;
2796 unsigned NonZeros = 0;
2797 std::set<SDOperand> Values;
2798 for (unsigned i = 0; i < NumElems; ++i) {
2799 SDOperand Elt = Op.getOperand(i);
2800 if (Elt.getOpcode() != ISD::UNDEF) {
2802 if (isZeroNode(Elt))
2805 NonZeros |= (1 << i);
2811 if (NumNonZero == 0)
2812 // Must be a mix of zero and undef. Return a zero vector.
2813 return getZeroVector(VT, DAG);
2815 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2816 if (Values.size() == 1)
2819 // Special case for single non-zero element.
2820 if (NumNonZero == 1) {
2821 unsigned Idx = CountTrailingZeros_32(NonZeros);
2822 SDOperand Item = Op.getOperand(Idx);
2823 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2825 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2826 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2829 if (EVTBits == 32) {
2830 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2831 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2833 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2834 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2835 SmallVector<SDOperand, 8> MaskVec;
2836 for (unsigned i = 0; i < NumElems; i++)
2837 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2838 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2839 &MaskVec[0], MaskVec.size());
2840 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2841 DAG.getNode(ISD::UNDEF, VT), Mask);
2845 // Let legalizer expand 2-wide build_vector's.
2849 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2851 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2853 if (V.Val) return V;
2856 if (EVTBits == 16) {
2857 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2859 if (V.Val) return V;
2862 // If element VT is == 32 bits, turn it into a number of shuffles.
2863 SmallVector<SDOperand, 8> V;
2865 if (NumElems == 4 && NumZero > 0) {
2866 for (unsigned i = 0; i < 4; ++i) {
2867 bool isZero = !(NonZeros & (1 << i));
2869 V[i] = getZeroVector(VT, DAG);
2871 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2874 for (unsigned i = 0; i < 2; ++i) {
2875 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2878 V[i] = V[i*2]; // Must be a zero vector.
2881 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2882 getMOVLMask(NumElems, DAG));
2885 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2886 getMOVLMask(NumElems, DAG));
2889 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2890 getUnpacklMask(NumElems, DAG));
2895 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2896 // clears the upper bits.
2897 // FIXME: we can do the same for v4f32 case when we know both parts of
2898 // the lower half come from scalar_to_vector (loadf32). We should do
2899 // that in post legalizer dag combiner with target specific hooks.
2900 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2902 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2903 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2904 SmallVector<SDOperand, 8> MaskVec;
2905 bool Reverse = (NonZeros & 0x3) == 2;
2906 for (unsigned i = 0; i < 2; ++i)
2908 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2910 MaskVec.push_back(DAG.getConstant(i, EVT));
2911 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2912 for (unsigned i = 0; i < 2; ++i)
2914 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2916 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2917 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2918 &MaskVec[0], MaskVec.size());
2919 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2922 if (Values.size() > 2) {
2923 // Expand into a number of unpckl*.
2925 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2926 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2927 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2928 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2929 for (unsigned i = 0; i < NumElems; ++i)
2930 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2932 while (NumElems != 0) {
2933 for (unsigned i = 0; i < NumElems; ++i)
2934 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
2945 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2946 SDOperand V1 = Op.getOperand(0);
2947 SDOperand V2 = Op.getOperand(1);
2948 SDOperand PermMask = Op.getOperand(2);
2949 MVT::ValueType VT = Op.getValueType();
2950 unsigned NumElems = PermMask.getNumOperands();
2951 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
2952 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
2953 bool V1IsSplat = false;
2954 bool V2IsSplat = false;
2956 if (isUndefShuffle(Op.Val))
2957 return DAG.getNode(ISD::UNDEF, VT);
2959 if (isSplatMask(PermMask.Val)) {
2960 if (NumElems <= 4) return Op;
2961 // Promote it to a v4i32 splat.
2962 return PromoteSplat(Op, DAG);
2965 if (X86::isMOVLMask(PermMask.Val))
2966 return (V1IsUndef) ? V2 : Op;
2968 if (X86::isMOVSHDUPMask(PermMask.Val) ||
2969 X86::isMOVSLDUPMask(PermMask.Val) ||
2970 X86::isMOVHLPSMask(PermMask.Val) ||
2971 X86::isMOVHPMask(PermMask.Val) ||
2972 X86::isMOVLPMask(PermMask.Val))
2975 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2976 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
2977 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2979 bool Commuted = false;
2980 V1IsSplat = isSplatVector(V1.Val);
2981 V2IsSplat = isSplatVector(V2.Val);
2982 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
2983 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2984 std::swap(V1IsSplat, V2IsSplat);
2985 std::swap(V1IsUndef, V2IsUndef);
2989 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
2990 if (V2IsUndef) return V1;
2991 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2993 // V2 is a splat, so the mask may be malformed. That is, it may point
2994 // to any V2 element. The instruction selectior won't like this. Get
2995 // a corrected mask and commute to form a proper MOVS{S|D}.
2996 SDOperand NewMask = getMOVLMask(NumElems, DAG);
2997 if (NewMask.Val != PermMask.Val)
2998 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3003 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3004 X86::isUNPCKLMask(PermMask.Val) ||
3005 X86::isUNPCKHMask(PermMask.Val))
3009 // Normalize mask so all entries that point to V2 points to its first
3010 // element then try to match unpck{h|l} again. If match, return a
3011 // new vector_shuffle with the corrected mask.
3012 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3013 if (NewMask.Val != PermMask.Val) {
3014 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3015 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3016 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3017 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3018 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3019 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3024 // Normalize the node to match x86 shuffle ops if needed
3025 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3026 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3029 // Commute is back and try unpck* again.
3030 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3031 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3032 X86::isUNPCKLMask(PermMask.Val) ||
3033 X86::isUNPCKHMask(PermMask.Val))
3037 // If VT is integer, try PSHUF* first, then SHUFP*.
3038 if (MVT::isInteger(VT)) {
3039 if (X86::isPSHUFDMask(PermMask.Val) ||
3040 X86::isPSHUFHWMask(PermMask.Val) ||
3041 X86::isPSHUFLWMask(PermMask.Val)) {
3042 if (V2.getOpcode() != ISD::UNDEF)
3043 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3044 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3048 if (X86::isSHUFPMask(PermMask.Val))
3051 // Handle v8i16 shuffle high / low shuffle node pair.
3052 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3053 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3054 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3055 SmallVector<SDOperand, 8> MaskVec;
3056 for (unsigned i = 0; i != 4; ++i)
3057 MaskVec.push_back(PermMask.getOperand(i));
3058 for (unsigned i = 4; i != 8; ++i)
3059 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3060 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3061 &MaskVec[0], MaskVec.size());
3062 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3064 for (unsigned i = 0; i != 4; ++i)
3065 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3066 for (unsigned i = 4; i != 8; ++i)
3067 MaskVec.push_back(PermMask.getOperand(i));
3068 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
3069 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3072 // Floating point cases in the other order.
3073 if (X86::isSHUFPMask(PermMask.Val))
3075 if (X86::isPSHUFDMask(PermMask.Val) ||
3076 X86::isPSHUFHWMask(PermMask.Val) ||
3077 X86::isPSHUFLWMask(PermMask.Val)) {
3078 if (V2.getOpcode() != ISD::UNDEF)
3079 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3080 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3085 if (NumElems == 4) {
3086 MVT::ValueType MaskVT = PermMask.getValueType();
3087 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
3088 SmallVector<std::pair<int, int>, 8> Locs;
3089 Locs.reserve(NumElems);
3090 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3091 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3094 // If no more than two elements come from either vector. This can be
3095 // implemented with two shuffles. First shuffle gather the elements.
3096 // The second shuffle, which takes the first shuffle as both of its
3097 // vector operands, put the elements into the right order.
3098 for (unsigned i = 0; i != NumElems; ++i) {
3099 SDOperand Elt = PermMask.getOperand(i);
3100 if (Elt.getOpcode() == ISD::UNDEF) {
3101 Locs[i] = std::make_pair(-1, -1);
3103 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3104 if (Val < NumElems) {
3105 Locs[i] = std::make_pair(0, NumLo);
3109 Locs[i] = std::make_pair(1, NumHi);
3110 if (2+NumHi < NumElems)
3111 Mask1[2+NumHi] = Elt;
3116 if (NumLo <= 2 && NumHi <= 2) {
3117 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3118 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3119 &Mask1[0], Mask1.size()));
3120 for (unsigned i = 0; i != NumElems; ++i) {
3121 if (Locs[i].first == -1)
3124 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3125 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3126 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3130 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3131 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3132 &Mask2[0], Mask2.size()));
3135 // Break it into (shuffle shuffle_hi, shuffle_lo).
3137 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3138 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3139 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3140 unsigned MaskIdx = 0;
3142 unsigned HiIdx = NumElems/2;
3143 for (unsigned i = 0; i != NumElems; ++i) {
3144 if (i == NumElems/2) {
3150 SDOperand Elt = PermMask.getOperand(i);
3151 if (Elt.getOpcode() == ISD::UNDEF) {
3152 Locs[i] = std::make_pair(-1, -1);
3153 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3154 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3155 (*MaskPtr)[LoIdx] = Elt;
3158 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3159 (*MaskPtr)[HiIdx] = Elt;
3164 SDOperand LoShuffle =
3165 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3166 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3167 &LoMask[0], LoMask.size()));
3168 SDOperand HiShuffle =
3169 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3170 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3171 &HiMask[0], HiMask.size()));
3172 SmallVector<SDOperand, 8> MaskOps;
3173 for (unsigned i = 0; i != NumElems; ++i) {
3174 if (Locs[i].first == -1) {
3175 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3177 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3178 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3181 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3182 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3183 &MaskOps[0], MaskOps.size()));
3190 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3191 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3194 MVT::ValueType VT = Op.getValueType();
3195 // TODO: handle v16i8.
3196 if (MVT::getSizeInBits(VT) == 16) {
3197 // Transform it so it match pextrw which produces a 32-bit result.
3198 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3199 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3200 Op.getOperand(0), Op.getOperand(1));
3201 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3202 DAG.getValueType(VT));
3203 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3204 } else if (MVT::getSizeInBits(VT) == 32) {
3205 SDOperand Vec = Op.getOperand(0);
3206 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3209 // SHUFPS the element to the lowest double word, then movss.
3210 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3211 SmallVector<SDOperand, 8> IdxVec;
3212 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
3213 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3214 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3215 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3216 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3217 &IdxVec[0], IdxVec.size());
3218 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3219 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3220 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3221 DAG.getConstant(0, getPointerTy()));
3222 } else if (MVT::getSizeInBits(VT) == 64) {
3223 SDOperand Vec = Op.getOperand(0);
3224 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3228 // UNPCKHPD the element to the lowest double word, then movsd.
3229 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3230 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3231 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3232 SmallVector<SDOperand, 8> IdxVec;
3233 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
3234 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3235 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3236 &IdxVec[0], IdxVec.size());
3237 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3238 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3239 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3240 DAG.getConstant(0, getPointerTy()));
3247 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3248 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3249 // as its second argument.
3250 MVT::ValueType VT = Op.getValueType();
3251 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
3252 SDOperand N0 = Op.getOperand(0);
3253 SDOperand N1 = Op.getOperand(1);
3254 SDOperand N2 = Op.getOperand(2);
3255 if (MVT::getSizeInBits(BaseVT) == 16) {
3256 if (N1.getValueType() != MVT::i32)
3257 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3258 if (N2.getValueType() != MVT::i32)
3259 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
3260 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3261 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3262 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3265 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3266 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3267 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3268 SmallVector<SDOperand, 8> MaskVec;
3269 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3270 for (unsigned i = 1; i <= 3; ++i)
3271 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3272 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3273 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3274 &MaskVec[0], MaskVec.size()));
3276 // Use two pinsrw instructions to insert a 32 bit value.
3278 if (MVT::isFloatingPoint(N1.getValueType())) {
3279 if (ISD::isNON_EXTLoad(N1.Val)) {
3280 // Just load directly from f32mem to GR32.
3281 LoadSDNode *LD = cast<LoadSDNode>(N1);
3282 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
3283 LD->getSrcValue(), LD->getSrcValueOffset());
3285 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3286 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3287 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3288 DAG.getConstant(0, getPointerTy()));
3291 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3292 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3293 DAG.getConstant(Idx, getPointerTy()));
3294 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3295 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3296 DAG.getConstant(Idx+1, getPointerTy()));
3297 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3305 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3306 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3307 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3310 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3311 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3312 // one of the above mentioned nodes. It has to be wrapped because otherwise
3313 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3314 // be used to form addressing mode. These wrapped nodes will be selected
3317 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3318 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3319 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3321 CP->getAlignment());
3322 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3323 // With PIC, the address is actually $g + Offset.
3324 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3325 !Subtarget->isPICStyleRIPRel()) {
3326 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3327 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3335 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3336 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3337 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3338 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3339 // With PIC, the address is actually $g + Offset.
3340 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3341 !Subtarget->isPICStyleRIPRel()) {
3342 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3343 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3347 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3348 // load the value at address GV, not the value of GV itself. This means that
3349 // the GlobalAddress must be in the base or index register of the address, not
3350 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3351 // The same applies for external symbols during PIC codegen
3352 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3353 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3359 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3360 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3361 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3362 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3363 // With PIC, the address is actually $g + Offset.
3364 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3365 !Subtarget->isPICStyleRIPRel()) {
3366 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3367 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3374 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3375 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3376 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3377 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3378 // With PIC, the address is actually $g + Offset.
3379 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3380 !Subtarget->isPICStyleRIPRel()) {
3381 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3382 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3389 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3390 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3391 "Not an i64 shift!");
3392 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3393 SDOperand ShOpLo = Op.getOperand(0);
3394 SDOperand ShOpHi = Op.getOperand(1);
3395 SDOperand ShAmt = Op.getOperand(2);
3396 SDOperand Tmp1 = isSRA ?
3397 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3398 DAG.getConstant(0, MVT::i32);
3400 SDOperand Tmp2, Tmp3;
3401 if (Op.getOpcode() == ISD::SHL_PARTS) {
3402 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3403 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3405 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3406 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3409 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3410 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3411 DAG.getConstant(32, MVT::i8));
3412 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
3413 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
3416 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3418 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3419 SmallVector<SDOperand, 4> Ops;
3420 if (Op.getOpcode() == ISD::SHL_PARTS) {
3421 Ops.push_back(Tmp2);
3422 Ops.push_back(Tmp3);
3424 Ops.push_back(InFlag);
3425 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3426 InFlag = Hi.getValue(1);
3429 Ops.push_back(Tmp3);
3430 Ops.push_back(Tmp1);
3432 Ops.push_back(InFlag);
3433 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3435 Ops.push_back(Tmp2);
3436 Ops.push_back(Tmp3);
3438 Ops.push_back(InFlag);
3439 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3440 InFlag = Lo.getValue(1);
3443 Ops.push_back(Tmp3);
3444 Ops.push_back(Tmp1);
3446 Ops.push_back(InFlag);
3447 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3450 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3454 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3457 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3458 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3459 Op.getOperand(0).getValueType() >= MVT::i16 &&
3460 "Unknown SINT_TO_FP to lower!");
3463 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3464 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3465 MachineFunction &MF = DAG.getMachineFunction();
3466 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3467 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3468 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3469 StackSlot, NULL, 0);
3474 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3476 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3477 SmallVector<SDOperand, 8> Ops;
3478 Ops.push_back(Chain);
3479 Ops.push_back(StackSlot);
3480 Ops.push_back(DAG.getValueType(SrcVT));
3481 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3482 Tys, &Ops[0], Ops.size());
3485 Chain = Result.getValue(1);
3486 SDOperand InFlag = Result.getValue(2);
3488 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3489 // shouldn't be necessary except that RFP cannot be live across
3490 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3491 MachineFunction &MF = DAG.getMachineFunction();
3492 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3493 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3494 Tys = DAG.getVTList(MVT::Other);
3495 SmallVector<SDOperand, 8> Ops;
3496 Ops.push_back(Chain);
3497 Ops.push_back(Result);
3498 Ops.push_back(StackSlot);
3499 Ops.push_back(DAG.getValueType(Op.getValueType()));
3500 Ops.push_back(InFlag);
3501 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3502 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3508 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3509 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3510 "Unknown FP_TO_SINT to lower!");
3511 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3513 MachineFunction &MF = DAG.getMachineFunction();
3514 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3515 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3516 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3519 switch (Op.getValueType()) {
3520 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3521 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3522 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3523 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3526 SDOperand Chain = DAG.getEntryNode();
3527 SDOperand Value = Op.getOperand(0);
3529 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3530 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3531 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3533 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3535 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3536 Chain = Value.getValue(1);
3537 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3538 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3541 // Build the FP_TO_INT*_IN_MEM
3542 SDOperand Ops[] = { Chain, Value, StackSlot };
3543 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3546 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3549 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3550 MVT::ValueType VT = Op.getValueType();
3551 const Type *OpNTy = MVT::getTypeForValueType(VT);
3552 std::vector<Constant*> CV;
3553 if (VT == MVT::f64) {
3554 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3555 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3557 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3558 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3559 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3560 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3562 Constant *CS = ConstantStruct::get(CV);
3563 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3564 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3565 SmallVector<SDOperand, 3> Ops;
3566 Ops.push_back(DAG.getEntryNode());
3567 Ops.push_back(CPIdx);
3568 Ops.push_back(DAG.getSrcValue(NULL));
3569 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3570 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3573 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3574 MVT::ValueType VT = Op.getValueType();
3575 const Type *OpNTy = MVT::getTypeForValueType(VT);
3576 std::vector<Constant*> CV;
3577 if (VT == MVT::f64) {
3578 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3579 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3581 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3582 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3583 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3584 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3586 Constant *CS = ConstantStruct::get(CV);
3587 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3588 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3589 SmallVector<SDOperand, 3> Ops;
3590 Ops.push_back(DAG.getEntryNode());
3591 Ops.push_back(CPIdx);
3592 Ops.push_back(DAG.getSrcValue(NULL));
3593 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3594 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3597 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3598 SDOperand Op0 = Op.getOperand(0);
3599 SDOperand Op1 = Op.getOperand(1);
3600 MVT::ValueType VT = Op.getValueType();
3601 MVT::ValueType SrcVT = Op1.getValueType();
3602 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3604 // If second operand is smaller, extend it first.
3605 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3606 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3610 // First get the sign bit of second operand.
3611 std::vector<Constant*> CV;
3612 if (SrcVT == MVT::f64) {
3613 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3614 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3616 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3617 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3618 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3619 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3621 Constant *CS = ConstantStruct::get(CV);
3622 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3623 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3624 SmallVector<SDOperand, 3> Ops;
3625 Ops.push_back(DAG.getEntryNode());
3626 Ops.push_back(CPIdx);
3627 Ops.push_back(DAG.getSrcValue(NULL));
3628 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3629 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3631 // Shift sign bit right or left if the two operands have different types.
3632 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3633 // Op0 is MVT::f32, Op1 is MVT::f64.
3634 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3635 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3636 DAG.getConstant(32, MVT::i32));
3637 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3638 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3639 DAG.getConstant(0, getPointerTy()));
3642 // Clear first operand sign bit.
3644 if (VT == MVT::f64) {
3645 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3646 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3648 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3649 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3650 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3651 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3653 CS = ConstantStruct::get(CV);
3654 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3655 Tys = DAG.getVTList(VT, MVT::Other);
3657 Ops.push_back(DAG.getEntryNode());
3658 Ops.push_back(CPIdx);
3659 Ops.push_back(DAG.getSrcValue(NULL));
3660 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3661 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3663 // Or the value with the sign bit.
3664 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3667 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3669 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3671 SDOperand Op0 = Op.getOperand(0);
3672 SDOperand Op1 = Op.getOperand(1);
3673 SDOperand CC = Op.getOperand(2);
3674 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3675 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3676 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3677 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3680 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3682 SDOperand Ops1[] = { Chain, Op0, Op1 };
3683 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3684 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3685 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3688 assert(isFP && "Illegal integer SetCC!");
3690 SDOperand COps[] = { Chain, Op0, Op1 };
3691 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3693 switch (SetCCOpcode) {
3694 default: assert(false && "Illegal floating point SetCC!");
3695 case ISD::SETOEQ: { // !PF & ZF
3696 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3697 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3698 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3700 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3701 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3703 case ISD::SETUNE: { // PF | !ZF
3704 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3705 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3706 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3708 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3709 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3714 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3715 bool addTest = true;
3716 SDOperand Chain = DAG.getEntryNode();
3717 SDOperand Cond = Op.getOperand(0);
3719 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3721 if (Cond.getOpcode() == ISD::SETCC)
3722 Cond = LowerSETCC(Cond, DAG, Chain);
3724 if (Cond.getOpcode() == X86ISD::SETCC) {
3725 CC = Cond.getOperand(0);
3727 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3728 // (since flag operand cannot be shared). Use it as the condition setting
3729 // operand in place of the X86ISD::SETCC.
3730 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3731 // to use a test instead of duplicating the X86ISD::CMP (for register
3732 // pressure reason)?
3733 SDOperand Cmp = Cond.getOperand(1);
3734 unsigned Opc = Cmp.getOpcode();
3735 bool IllegalFPCMov = !X86ScalarSSE &&
3736 MVT::isFloatingPoint(Op.getValueType()) &&
3737 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3738 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3740 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3741 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3747 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3748 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3749 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3752 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3753 SmallVector<SDOperand, 4> Ops;
3754 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3755 // condition is true.
3756 Ops.push_back(Op.getOperand(2));
3757 Ops.push_back(Op.getOperand(1));
3759 Ops.push_back(Cond.getValue(1));
3760 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3763 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3764 bool addTest = true;
3765 SDOperand Chain = Op.getOperand(0);
3766 SDOperand Cond = Op.getOperand(1);
3767 SDOperand Dest = Op.getOperand(2);
3769 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3771 if (Cond.getOpcode() == ISD::SETCC)
3772 Cond = LowerSETCC(Cond, DAG, Chain);
3774 if (Cond.getOpcode() == X86ISD::SETCC) {
3775 CC = Cond.getOperand(0);
3777 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3778 // (since flag operand cannot be shared). Use it as the condition setting
3779 // operand in place of the X86ISD::SETCC.
3780 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3781 // to use a test instead of duplicating the X86ISD::CMP (for register
3782 // pressure reason)?
3783 SDOperand Cmp = Cond.getOperand(1);
3784 unsigned Opc = Cmp.getOpcode();
3785 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3786 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3787 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3793 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3794 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3795 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3797 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3798 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3801 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3802 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3804 if (Subtarget->is64Bit())
3805 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3807 switch (CallingConv) {
3809 assert(0 && "Unsupported calling convention");
3810 case CallingConv::Fast:
3812 return LowerFastCCCallTo(Op, DAG, CallingConv);
3814 case CallingConv::C:
3815 case CallingConv::X86_StdCall:
3816 return LowerCCCCallTo(Op, DAG, CallingConv);
3817 case CallingConv::X86_FastCall:
3818 return LowerFastCCCallTo(Op, DAG, CallingConv);
3823 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3824 MachineFunction &MF = DAG.getMachineFunction();
3825 const Function* Fn = MF.getFunction();
3826 if (Fn->hasExternalLinkage() &&
3827 Subtarget->isTargetCygMing() &&
3828 Fn->getName() == "main")
3829 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true);
3831 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3832 if (Subtarget->is64Bit())
3833 return LowerX86_64CCCArguments(Op, DAG);
3837 assert(0 && "Unsupported calling convention");
3838 case CallingConv::Fast:
3840 return LowerFastCCArguments(Op, DAG);
3843 case CallingConv::C:
3844 return LowerCCCArguments(Op, DAG);
3845 case CallingConv::X86_StdCall:
3846 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall);
3847 return LowerCCCArguments(Op, DAG, true);
3848 case CallingConv::X86_FastCall:
3849 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall);
3850 return LowerFastCCArguments(Op, DAG, true);
3854 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3855 SDOperand InFlag(0, 0);
3856 SDOperand Chain = Op.getOperand(0);
3858 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3859 if (Align == 0) Align = 1;
3861 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3862 // If not DWORD aligned, call memset if size is less than the threshold.
3863 // It knows how to align to the right boundary first.
3864 if ((Align & 3) != 0 ||
3865 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3866 MVT::ValueType IntPtr = getPointerTy();
3867 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3868 TargetLowering::ArgListTy Args;
3869 TargetLowering::ArgListEntry Entry;
3870 Entry.Node = Op.getOperand(1);
3871 Entry.Ty = IntPtrTy;
3872 Entry.isSigned = false;
3873 Entry.isInReg = false;
3874 Entry.isSRet = false;
3875 Args.push_back(Entry);
3876 // Extend the unsigned i8 argument to be an int value for the call.
3877 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3878 Entry.Ty = IntPtrTy;
3879 Entry.isSigned = false;
3880 Entry.isInReg = false;
3881 Entry.isSRet = false;
3882 Args.push_back(Entry);
3883 Entry.Node = Op.getOperand(3);
3884 Args.push_back(Entry);
3885 std::pair<SDOperand,SDOperand> CallResult =
3886 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3887 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3888 return CallResult.second;
3893 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3894 unsigned BytesLeft = 0;
3895 bool TwoRepStos = false;
3898 uint64_t Val = ValC->getValue() & 255;
3900 // If the value is a constant, then we can potentially use larger sets.
3901 switch (Align & 3) {
3902 case 2: // WORD aligned
3905 Val = (Val << 8) | Val;
3907 case 0: // DWORD aligned
3910 Val = (Val << 8) | Val;
3911 Val = (Val << 16) | Val;
3912 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
3915 Val = (Val << 32) | Val;
3918 default: // Byte aligned
3921 Count = Op.getOperand(3);
3925 if (AVT > MVT::i8) {
3927 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3928 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3929 BytesLeft = I->getValue() % UBytes;
3931 assert(AVT >= MVT::i32 &&
3932 "Do not use rep;stos if not at least DWORD aligned");
3933 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3934 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3939 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
3941 InFlag = Chain.getValue(1);
3944 Count = Op.getOperand(3);
3945 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
3946 InFlag = Chain.getValue(1);
3949 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3951 InFlag = Chain.getValue(1);
3952 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3953 Op.getOperand(1), InFlag);
3954 InFlag = Chain.getValue(1);
3956 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3957 SmallVector<SDOperand, 8> Ops;
3958 Ops.push_back(Chain);
3959 Ops.push_back(DAG.getValueType(AVT));
3960 Ops.push_back(InFlag);
3961 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3964 InFlag = Chain.getValue(1);
3965 Count = Op.getOperand(3);
3966 MVT::ValueType CVT = Count.getValueType();
3967 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3968 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3969 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3971 InFlag = Chain.getValue(1);
3972 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3974 Ops.push_back(Chain);
3975 Ops.push_back(DAG.getValueType(MVT::i8));
3976 Ops.push_back(InFlag);
3977 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3978 } else if (BytesLeft) {
3979 // Issue stores for the last 1 - 7 bytes.
3981 unsigned Val = ValC->getValue() & 255;
3982 unsigned Offset = I->getValue() - BytesLeft;
3983 SDOperand DstAddr = Op.getOperand(1);
3984 MVT::ValueType AddrVT = DstAddr.getValueType();
3985 if (BytesLeft >= 4) {
3986 Val = (Val << 8) | Val;
3987 Val = (Val << 16) | Val;
3988 Value = DAG.getConstant(Val, MVT::i32);
3989 Chain = DAG.getStore(Chain, Value,
3990 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3991 DAG.getConstant(Offset, AddrVT)),
3996 if (BytesLeft >= 2) {
3997 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
3998 Chain = DAG.getStore(Chain, Value,
3999 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4000 DAG.getConstant(Offset, AddrVT)),
4005 if (BytesLeft == 1) {
4006 Value = DAG.getConstant(Val, MVT::i8);
4007 Chain = DAG.getStore(Chain, Value,
4008 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4009 DAG.getConstant(Offset, AddrVT)),
4017 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
4018 SDOperand Chain = Op.getOperand(0);
4020 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4021 if (Align == 0) Align = 1;
4023 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4024 // If not DWORD aligned, call memcpy if size is less than the threshold.
4025 // It knows how to align to the right boundary first.
4026 if ((Align & 3) != 0 ||
4027 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
4028 MVT::ValueType IntPtr = getPointerTy();
4029 TargetLowering::ArgListTy Args;
4030 TargetLowering::ArgListEntry Entry;
4031 Entry.Ty = getTargetData()->getIntPtrType();
4032 Entry.isSigned = false;
4033 Entry.isInReg = false;
4034 Entry.isSRet = false;
4035 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
4036 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
4037 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
4038 std::pair<SDOperand,SDOperand> CallResult =
4039 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4040 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
4041 return CallResult.second;
4046 unsigned BytesLeft = 0;
4047 bool TwoRepMovs = false;
4048 switch (Align & 3) {
4049 case 2: // WORD aligned
4052 case 0: // DWORD aligned
4054 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4057 default: // Byte aligned
4059 Count = Op.getOperand(3);
4063 if (AVT > MVT::i8) {
4065 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4066 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4067 BytesLeft = I->getValue() % UBytes;
4069 assert(AVT >= MVT::i32 &&
4070 "Do not use rep;movs if not at least DWORD aligned");
4071 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4072 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4077 SDOperand InFlag(0, 0);
4078 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4080 InFlag = Chain.getValue(1);
4081 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4082 Op.getOperand(1), InFlag);
4083 InFlag = Chain.getValue(1);
4084 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4085 Op.getOperand(2), InFlag);
4086 InFlag = Chain.getValue(1);
4088 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4089 SmallVector<SDOperand, 8> Ops;
4090 Ops.push_back(Chain);
4091 Ops.push_back(DAG.getValueType(AVT));
4092 Ops.push_back(InFlag);
4093 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4096 InFlag = Chain.getValue(1);
4097 Count = Op.getOperand(3);
4098 MVT::ValueType CVT = Count.getValueType();
4099 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4100 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4101 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4103 InFlag = Chain.getValue(1);
4104 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4106 Ops.push_back(Chain);
4107 Ops.push_back(DAG.getValueType(MVT::i8));
4108 Ops.push_back(InFlag);
4109 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4110 } else if (BytesLeft) {
4111 // Issue loads and stores for the last 1 - 7 bytes.
4112 unsigned Offset = I->getValue() - BytesLeft;
4113 SDOperand DstAddr = Op.getOperand(1);
4114 MVT::ValueType DstVT = DstAddr.getValueType();
4115 SDOperand SrcAddr = Op.getOperand(2);
4116 MVT::ValueType SrcVT = SrcAddr.getValueType();
4118 if (BytesLeft >= 4) {
4119 Value = DAG.getLoad(MVT::i32, Chain,
4120 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4121 DAG.getConstant(Offset, SrcVT)),
4123 Chain = Value.getValue(1);
4124 Chain = DAG.getStore(Chain, Value,
4125 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4126 DAG.getConstant(Offset, DstVT)),
4131 if (BytesLeft >= 2) {
4132 Value = DAG.getLoad(MVT::i16, Chain,
4133 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4134 DAG.getConstant(Offset, SrcVT)),
4136 Chain = Value.getValue(1);
4137 Chain = DAG.getStore(Chain, Value,
4138 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4139 DAG.getConstant(Offset, DstVT)),
4145 if (BytesLeft == 1) {
4146 Value = DAG.getLoad(MVT::i8, Chain,
4147 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4148 DAG.getConstant(Offset, SrcVT)),
4150 Chain = Value.getValue(1);
4151 Chain = DAG.getStore(Chain, Value,
4152 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4153 DAG.getConstant(Offset, DstVT)),
4162 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4163 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4164 SDOperand TheOp = Op.getOperand(0);
4165 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
4166 if (Subtarget->is64Bit()) {
4167 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4168 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4169 MVT::i64, Copy1.getValue(2));
4170 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4171 DAG.getConstant(32, MVT::i8));
4173 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
4176 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4177 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
4180 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4181 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4182 MVT::i32, Copy1.getValue(2));
4183 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
4184 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4185 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
4188 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4189 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4191 if (!Subtarget->is64Bit()) {
4192 // vastart just stores the address of the VarArgsFrameIndex slot into the
4193 // memory location argument.
4194 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4195 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4200 // gp_offset (0 - 6 * 8)
4201 // fp_offset (48 - 48 + 8 * 16)
4202 // overflow_arg_area (point to parameters coming in memory).
4204 SmallVector<SDOperand, 8> MemOps;
4205 SDOperand FIN = Op.getOperand(1);
4207 SDOperand Store = DAG.getStore(Op.getOperand(0),
4208 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4209 FIN, SV->getValue(), SV->getOffset());
4210 MemOps.push_back(Store);
4213 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4214 DAG.getConstant(4, getPointerTy()));
4215 Store = DAG.getStore(Op.getOperand(0),
4216 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4217 FIN, SV->getValue(), SV->getOffset());
4218 MemOps.push_back(Store);
4220 // Store ptr to overflow_arg_area
4221 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4222 DAG.getConstant(4, getPointerTy()));
4223 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4224 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4226 MemOps.push_back(Store);
4228 // Store ptr to reg_save_area.
4229 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4230 DAG.getConstant(8, getPointerTy()));
4231 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4232 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4234 MemOps.push_back(Store);
4235 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4239 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4240 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4242 default: return SDOperand(); // Don't custom lower most intrinsics.
4243 // Comparison intrinsics.
4244 case Intrinsic::x86_sse_comieq_ss:
4245 case Intrinsic::x86_sse_comilt_ss:
4246 case Intrinsic::x86_sse_comile_ss:
4247 case Intrinsic::x86_sse_comigt_ss:
4248 case Intrinsic::x86_sse_comige_ss:
4249 case Intrinsic::x86_sse_comineq_ss:
4250 case Intrinsic::x86_sse_ucomieq_ss:
4251 case Intrinsic::x86_sse_ucomilt_ss:
4252 case Intrinsic::x86_sse_ucomile_ss:
4253 case Intrinsic::x86_sse_ucomigt_ss:
4254 case Intrinsic::x86_sse_ucomige_ss:
4255 case Intrinsic::x86_sse_ucomineq_ss:
4256 case Intrinsic::x86_sse2_comieq_sd:
4257 case Intrinsic::x86_sse2_comilt_sd:
4258 case Intrinsic::x86_sse2_comile_sd:
4259 case Intrinsic::x86_sse2_comigt_sd:
4260 case Intrinsic::x86_sse2_comige_sd:
4261 case Intrinsic::x86_sse2_comineq_sd:
4262 case Intrinsic::x86_sse2_ucomieq_sd:
4263 case Intrinsic::x86_sse2_ucomilt_sd:
4264 case Intrinsic::x86_sse2_ucomile_sd:
4265 case Intrinsic::x86_sse2_ucomigt_sd:
4266 case Intrinsic::x86_sse2_ucomige_sd:
4267 case Intrinsic::x86_sse2_ucomineq_sd: {
4269 ISD::CondCode CC = ISD::SETCC_INVALID;
4272 case Intrinsic::x86_sse_comieq_ss:
4273 case Intrinsic::x86_sse2_comieq_sd:
4277 case Intrinsic::x86_sse_comilt_ss:
4278 case Intrinsic::x86_sse2_comilt_sd:
4282 case Intrinsic::x86_sse_comile_ss:
4283 case Intrinsic::x86_sse2_comile_sd:
4287 case Intrinsic::x86_sse_comigt_ss:
4288 case Intrinsic::x86_sse2_comigt_sd:
4292 case Intrinsic::x86_sse_comige_ss:
4293 case Intrinsic::x86_sse2_comige_sd:
4297 case Intrinsic::x86_sse_comineq_ss:
4298 case Intrinsic::x86_sse2_comineq_sd:
4302 case Intrinsic::x86_sse_ucomieq_ss:
4303 case Intrinsic::x86_sse2_ucomieq_sd:
4304 Opc = X86ISD::UCOMI;
4307 case Intrinsic::x86_sse_ucomilt_ss:
4308 case Intrinsic::x86_sse2_ucomilt_sd:
4309 Opc = X86ISD::UCOMI;
4312 case Intrinsic::x86_sse_ucomile_ss:
4313 case Intrinsic::x86_sse2_ucomile_sd:
4314 Opc = X86ISD::UCOMI;
4317 case Intrinsic::x86_sse_ucomigt_ss:
4318 case Intrinsic::x86_sse2_ucomigt_sd:
4319 Opc = X86ISD::UCOMI;
4322 case Intrinsic::x86_sse_ucomige_ss:
4323 case Intrinsic::x86_sse2_ucomige_sd:
4324 Opc = X86ISD::UCOMI;
4327 case Intrinsic::x86_sse_ucomineq_ss:
4328 case Intrinsic::x86_sse2_ucomineq_sd:
4329 Opc = X86ISD::UCOMI;
4335 SDOperand LHS = Op.getOperand(1);
4336 SDOperand RHS = Op.getOperand(2);
4337 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4339 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4340 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
4341 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
4342 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
4343 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
4344 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
4345 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4350 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4351 // Depths > 0 not supported yet!
4352 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4355 // Just load the return address
4356 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4357 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4360 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4361 // Depths > 0 not supported yet!
4362 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4365 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4366 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4367 DAG.getConstant(4, getPointerTy()));
4370 /// LowerOperation - Provide custom lowering hooks for some operations.
4372 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4373 switch (Op.getOpcode()) {
4374 default: assert(0 && "Should not custom lower this!");
4375 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4376 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4377 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4378 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4379 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4380 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4381 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4382 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4383 case ISD::SHL_PARTS:
4384 case ISD::SRA_PARTS:
4385 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4386 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4387 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4388 case ISD::FABS: return LowerFABS(Op, DAG);
4389 case ISD::FNEG: return LowerFNEG(Op, DAG);
4390 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4391 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
4392 case ISD::SELECT: return LowerSELECT(Op, DAG);
4393 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4394 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4395 case ISD::CALL: return LowerCALL(Op, DAG);
4396 case ISD::RET: return LowerRET(Op, DAG);
4397 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4398 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4399 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4400 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4401 case ISD::VASTART: return LowerVASTART(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);
4409 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4411 default: return NULL;
4412 case X86ISD::SHLD: return "X86ISD::SHLD";
4413 case X86ISD::SHRD: return "X86ISD::SHRD";
4414 case X86ISD::FAND: return "X86ISD::FAND";
4415 case X86ISD::FOR: return "X86ISD::FOR";
4416 case X86ISD::FXOR: return "X86ISD::FXOR";
4417 case X86ISD::FSRL: return "X86ISD::FSRL";
4418 case X86ISD::FILD: return "X86ISD::FILD";
4419 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
4420 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
4421 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
4422 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
4423 case X86ISD::FLD: return "X86ISD::FLD";
4424 case X86ISD::FST: return "X86ISD::FST";
4425 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
4426 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
4427 case X86ISD::CALL: return "X86ISD::CALL";
4428 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
4429 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
4430 case X86ISD::CMP: return "X86ISD::CMP";
4431 case X86ISD::COMI: return "X86ISD::COMI";
4432 case X86ISD::UCOMI: return "X86ISD::UCOMI";
4433 case X86ISD::SETCC: return "X86ISD::SETCC";
4434 case X86ISD::CMOV: return "X86ISD::CMOV";
4435 case X86ISD::BRCOND: return "X86ISD::BRCOND";
4436 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
4437 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
4438 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
4439 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
4440 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
4441 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
4442 case X86ISD::Wrapper: return "X86ISD::Wrapper";
4443 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4444 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4445 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4446 case X86ISD::FMAX: return "X86ISD::FMAX";
4447 case X86ISD::FMIN: return "X86ISD::FMIN";
4451 /// isLegalAddressImmediate - Return true if the integer value or
4452 /// GlobalValue can be used as the offset of the target addressing mode.
4453 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
4454 // X86 allows a sign-extended 32-bit immediate field.
4455 return (V > -(1LL << 32) && V < (1LL << 32)-1);
4458 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
4459 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement
4460 // field unless we are in small code model.
4461 if (Subtarget->is64Bit() &&
4462 getTargetMachine().getCodeModel() != CodeModel::Small)
4465 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false));
4468 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4469 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4470 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4471 /// are assumed to be legal.
4473 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4474 // Only do shuffles on 128-bit vector types for now.
4475 if (MVT::getSizeInBits(VT) == 64) return false;
4476 return (Mask.Val->getNumOperands() <= 4 ||
4477 isSplatMask(Mask.Val) ||
4478 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4479 X86::isUNPCKLMask(Mask.Val) ||
4480 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4481 X86::isUNPCKHMask(Mask.Val));
4484 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4486 SelectionDAG &DAG) const {
4487 unsigned NumElts = BVOps.size();
4488 // Only do shuffles on 128-bit vector types for now.
4489 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4490 if (NumElts == 2) return true;
4492 return (isMOVLMask(&BVOps[0], 4) ||
4493 isCommutedMOVL(&BVOps[0], 4, true) ||
4494 isSHUFPMask(&BVOps[0], 4) ||
4495 isCommutedSHUFP(&BVOps[0], 4));
4500 //===----------------------------------------------------------------------===//
4501 // X86 Scheduler Hooks
4502 //===----------------------------------------------------------------------===//
4505 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4506 MachineBasicBlock *BB) {
4507 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4508 switch (MI->getOpcode()) {
4509 default: assert(false && "Unexpected instr type to insert");
4510 case X86::CMOV_FR32:
4511 case X86::CMOV_FR64:
4512 case X86::CMOV_V4F32:
4513 case X86::CMOV_V2F64:
4514 case X86::CMOV_V2I64: {
4515 // To "insert" a SELECT_CC instruction, we actually have to insert the
4516 // diamond control-flow pattern. The incoming instruction knows the
4517 // destination vreg to set, the condition code register to branch on, the
4518 // true/false values to select between, and a branch opcode to use.
4519 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4520 ilist<MachineBasicBlock>::iterator It = BB;
4526 // cmpTY ccX, r1, r2
4528 // fallthrough --> copy0MBB
4529 MachineBasicBlock *thisMBB = BB;
4530 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4531 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4533 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4534 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4535 MachineFunction *F = BB->getParent();
4536 F->getBasicBlockList().insert(It, copy0MBB);
4537 F->getBasicBlockList().insert(It, sinkMBB);
4538 // Update machine-CFG edges by first adding all successors of the current
4539 // block to the new block which will contain the Phi node for the select.
4540 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4541 e = BB->succ_end(); i != e; ++i)
4542 sinkMBB->addSuccessor(*i);
4543 // Next, remove all successors of the current block, and add the true
4544 // and fallthrough blocks as its successors.
4545 while(!BB->succ_empty())
4546 BB->removeSuccessor(BB->succ_begin());
4547 BB->addSuccessor(copy0MBB);
4548 BB->addSuccessor(sinkMBB);
4551 // %FalseValue = ...
4552 // # fallthrough to sinkMBB
4555 // Update machine-CFG edges
4556 BB->addSuccessor(sinkMBB);
4559 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4562 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4563 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4564 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4566 delete MI; // The pseudo instruction is gone now.
4570 case X86::FP_TO_INT16_IN_MEM:
4571 case X86::FP_TO_INT32_IN_MEM:
4572 case X86::FP_TO_INT64_IN_MEM: {
4573 // Change the floating point control register to use "round towards zero"
4574 // mode when truncating to an integer value.
4575 MachineFunction *F = BB->getParent();
4576 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4577 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4579 // Load the old value of the high byte of the control word...
4581 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4582 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4584 // Set the high part to be round to zero...
4585 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4588 // Reload the modified control word now...
4589 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4591 // Restore the memory image of control word to original value
4592 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4595 // Get the X86 opcode to use.
4597 switch (MI->getOpcode()) {
4598 default: assert(0 && "illegal opcode!");
4599 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4600 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4601 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4605 MachineOperand &Op = MI->getOperand(0);
4606 if (Op.isRegister()) {
4607 AM.BaseType = X86AddressMode::RegBase;
4608 AM.Base.Reg = Op.getReg();
4610 AM.BaseType = X86AddressMode::FrameIndexBase;
4611 AM.Base.FrameIndex = Op.getFrameIndex();
4613 Op = MI->getOperand(1);
4614 if (Op.isImmediate())
4615 AM.Scale = Op.getImm();
4616 Op = MI->getOperand(2);
4617 if (Op.isImmediate())
4618 AM.IndexReg = Op.getImm();
4619 Op = MI->getOperand(3);
4620 if (Op.isGlobalAddress()) {
4621 AM.GV = Op.getGlobal();
4623 AM.Disp = Op.getImm();
4625 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4626 .addReg(MI->getOperand(4).getReg());
4628 // Reload the original control word now.
4629 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4631 delete MI; // The pseudo instruction is gone now.
4637 //===----------------------------------------------------------------------===//
4638 // X86 Optimization Hooks
4639 //===----------------------------------------------------------------------===//
4641 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4643 uint64_t &KnownZero,
4645 unsigned Depth) const {
4646 unsigned Opc = Op.getOpcode();
4647 assert((Opc >= ISD::BUILTIN_OP_END ||
4648 Opc == ISD::INTRINSIC_WO_CHAIN ||
4649 Opc == ISD::INTRINSIC_W_CHAIN ||
4650 Opc == ISD::INTRINSIC_VOID) &&
4651 "Should use MaskedValueIsZero if you don't know whether Op"
4652 " is a target node!");
4654 KnownZero = KnownOne = 0; // Don't know anything.
4658 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4663 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4664 /// element of the result of the vector shuffle.
4665 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4666 MVT::ValueType VT = N->getValueType(0);
4667 SDOperand PermMask = N->getOperand(2);
4668 unsigned NumElems = PermMask.getNumOperands();
4669 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4671 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4673 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4674 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4675 SDOperand Idx = PermMask.getOperand(i);
4676 if (Idx.getOpcode() == ISD::UNDEF)
4677 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4678 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4683 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4684 /// node is a GlobalAddress + an offset.
4685 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4686 unsigned Opc = N->getOpcode();
4687 if (Opc == X86ISD::Wrapper) {
4688 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4689 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4692 } else if (Opc == ISD::ADD) {
4693 SDOperand N1 = N->getOperand(0);
4694 SDOperand N2 = N->getOperand(1);
4695 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4696 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4698 Offset += V->getSignExtended();
4701 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4702 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4704 Offset += V->getSignExtended();
4712 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4714 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4715 MachineFrameInfo *MFI) {
4716 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4719 SDOperand Loc = N->getOperand(1);
4720 SDOperand BaseLoc = Base->getOperand(1);
4721 if (Loc.getOpcode() == ISD::FrameIndex) {
4722 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4724 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4725 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4726 int FS = MFI->getObjectSize(FI);
4727 int BFS = MFI->getObjectSize(BFI);
4728 if (FS != BFS || FS != Size) return false;
4729 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4731 GlobalValue *GV1 = NULL;
4732 GlobalValue *GV2 = NULL;
4733 int64_t Offset1 = 0;
4734 int64_t Offset2 = 0;
4735 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4736 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4737 if (isGA1 && isGA2 && GV1 == GV2)
4738 return Offset1 == (Offset2 + Dist*Size);
4744 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4745 const X86Subtarget *Subtarget) {
4748 if (isGAPlusOffset(Base, GV, Offset))
4749 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4751 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4752 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4754 // Fixed objects do not specify alignment, however the offsets are known.
4755 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4756 (MFI->getObjectOffset(BFI) % 16) == 0);
4758 return MFI->getObjectAlignment(BFI) >= 16;
4764 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4765 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4766 /// if the load addresses are consecutive, non-overlapping, and in the right
4768 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4769 const X86Subtarget *Subtarget) {
4770 MachineFunction &MF = DAG.getMachineFunction();
4771 MachineFrameInfo *MFI = MF.getFrameInfo();
4772 MVT::ValueType VT = N->getValueType(0);
4773 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4774 SDOperand PermMask = N->getOperand(2);
4775 int NumElems = (int)PermMask.getNumOperands();
4776 SDNode *Base = NULL;
4777 for (int i = 0; i < NumElems; ++i) {
4778 SDOperand Idx = PermMask.getOperand(i);
4779 if (Idx.getOpcode() == ISD::UNDEF) {
4780 if (!Base) return SDOperand();
4783 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4784 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4788 else if (!isConsecutiveLoad(Arg.Val, Base,
4789 i, MVT::getSizeInBits(EVT)/8,MFI))
4794 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4796 LoadSDNode *LD = cast<LoadSDNode>(Base);
4797 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4798 LD->getSrcValueOffset());
4800 // Just use movups, it's shorter.
4801 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
4802 SmallVector<SDOperand, 3> Ops;
4803 Ops.push_back(Base->getOperand(0));
4804 Ops.push_back(Base->getOperand(1));
4805 Ops.push_back(Base->getOperand(2));
4806 return DAG.getNode(ISD::BIT_CONVERT, VT,
4807 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
4811 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4812 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4813 const X86Subtarget *Subtarget) {
4814 SDOperand Cond = N->getOperand(0);
4816 // If we have SSE[12] support, try to form min/max nodes.
4817 if (Subtarget->hasSSE2() &&
4818 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4819 if (Cond.getOpcode() == ISD::SETCC) {
4820 // Get the LHS/RHS of the select.
4821 SDOperand LHS = N->getOperand(1);
4822 SDOperand RHS = N->getOperand(2);
4823 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4825 unsigned Opcode = 0;
4826 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4829 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4832 if (!UnsafeFPMath) break;
4834 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4836 Opcode = X86ISD::FMIN;
4839 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4842 if (!UnsafeFPMath) break;
4844 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4846 Opcode = X86ISD::FMAX;
4849 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4852 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4855 if (!UnsafeFPMath) break;
4857 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4859 Opcode = X86ISD::FMIN;
4862 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4865 if (!UnsafeFPMath) break;
4867 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4869 Opcode = X86ISD::FMAX;
4875 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4884 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4885 DAGCombinerInfo &DCI) const {
4886 SelectionDAG &DAG = DCI.DAG;
4887 switch (N->getOpcode()) {
4889 case ISD::VECTOR_SHUFFLE:
4890 return PerformShuffleCombine(N, DAG, Subtarget);
4892 return PerformSELECTCombine(N, DAG, Subtarget);
4898 //===----------------------------------------------------------------------===//
4899 // X86 Inline Assembly Support
4900 //===----------------------------------------------------------------------===//
4902 /// getConstraintType - Given a constraint letter, return the type of
4903 /// constraint it is for this target.
4904 X86TargetLowering::ConstraintType
4905 X86TargetLowering::getConstraintType(char ConstraintLetter) const {
4906 switch (ConstraintLetter) {
4915 return C_RegisterClass;
4916 default: return TargetLowering::getConstraintType(ConstraintLetter);
4920 /// isOperandValidForConstraint - Return the specified operand (possibly
4921 /// modified) if the specified SDOperand is valid for the specified target
4922 /// constraint letter, otherwise return null.
4923 SDOperand X86TargetLowering::
4924 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
4925 switch (Constraint) {
4928 // Literal immediates are always ok.
4929 if (isa<ConstantSDNode>(Op)) return Op;
4931 // If we are in non-pic codegen mode, we allow the address of a global to
4932 // be used with 'i'.
4933 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
4934 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4935 return SDOperand(0, 0);
4937 if (GA->getOpcode() != ISD::TargetGlobalAddress)
4938 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4943 // Otherwise, not valid for this mode.
4944 return SDOperand(0, 0);
4946 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
4950 std::vector<unsigned> X86TargetLowering::
4951 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4952 MVT::ValueType VT) const {
4953 if (Constraint.size() == 1) {
4954 // FIXME: not handling fp-stack yet!
4955 // FIXME: not handling MMX registers yet ('y' constraint).
4956 switch (Constraint[0]) { // GCC X86 Constraint Letters
4957 default: break; // Unknown constraint letter
4958 case 'A': // EAX/EDX
4959 if (VT == MVT::i32 || VT == MVT::i64)
4960 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
4962 case 'r': // GENERAL_REGS
4963 case 'R': // LEGACY_REGS
4964 if (VT == MVT::i64 && Subtarget->is64Bit())
4965 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
4966 X86::RSI, X86::RDI, X86::RBP, X86::RSP,
4967 X86::R8, X86::R9, X86::R10, X86::R11,
4968 X86::R12, X86::R13, X86::R14, X86::R15, 0);
4970 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
4971 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
4972 else if (VT == MVT::i16)
4973 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
4974 X86::SI, X86::DI, X86::BP, X86::SP, 0);
4975 else if (VT == MVT::i8)
4976 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
4978 case 'l': // INDEX_REGS
4980 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
4981 X86::ESI, X86::EDI, X86::EBP, 0);
4982 else if (VT == MVT::i16)
4983 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
4984 X86::SI, X86::DI, X86::BP, 0);
4985 else if (VT == MVT::i8)
4986 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4988 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
4991 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
4992 else if (VT == MVT::i16)
4993 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
4994 else if (VT == MVT::i8)
4995 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4997 case 'x': // SSE_REGS if SSE1 allowed
4998 if (Subtarget->hasSSE1())
4999 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5000 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5002 return std::vector<unsigned>();
5003 case 'Y': // SSE_REGS if SSE2 allowed
5004 if (Subtarget->hasSSE2())
5005 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5006 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5008 return std::vector<unsigned>();
5012 return std::vector<unsigned>();
5015 std::pair<unsigned, const TargetRegisterClass*>
5016 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5017 MVT::ValueType VT) const {
5018 // Use the default implementation in TargetLowering to convert the register
5019 // constraint into a member of a register class.
5020 std::pair<unsigned, const TargetRegisterClass*> Res;
5021 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5023 // Not found as a standard register?
5024 if (Res.second == 0) {
5025 // GCC calls "st(0)" just plain "st".
5026 if (StringsEqualNoCase("{st}", Constraint)) {
5027 Res.first = X86::ST0;
5028 Res.second = X86::RSTRegisterClass;
5034 // Otherwise, check to see if this is a register class of the wrong value
5035 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5036 // turn into {ax},{dx}.
5037 if (Res.second->hasType(VT))
5038 return Res; // Correct type already, nothing to do.
5040 // All of the single-register GCC register classes map their values onto
5041 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5042 // really want an 8-bit or 32-bit register, map to the appropriate register
5043 // class and return the appropriate register.
5044 if (Res.second != X86::GR16RegisterClass)
5047 if (VT == MVT::i8) {
5048 unsigned DestReg = 0;
5049 switch (Res.first) {
5051 case X86::AX: DestReg = X86::AL; break;
5052 case X86::DX: DestReg = X86::DL; break;
5053 case X86::CX: DestReg = X86::CL; break;
5054 case X86::BX: DestReg = X86::BL; break;
5057 Res.first = DestReg;
5058 Res.second = Res.second = X86::GR8RegisterClass;
5060 } else if (VT == MVT::i32) {
5061 unsigned DestReg = 0;
5062 switch (Res.first) {
5064 case X86::AX: DestReg = X86::EAX; break;
5065 case X86::DX: DestReg = X86::EDX; break;
5066 case X86::CX: DestReg = X86::ECX; break;
5067 case X86::BX: DestReg = X86::EBX; break;
5068 case X86::SI: DestReg = X86::ESI; break;
5069 case X86::DI: DestReg = X86::EDI; break;
5070 case X86::BP: DestReg = X86::EBP; break;
5071 case X86::SP: DestReg = X86::ESP; break;
5074 Res.first = DestReg;
5075 Res.second = Res.second = X86::GR32RegisterClass;
5077 } else if (VT == MVT::i64) {
5078 unsigned DestReg = 0;
5079 switch (Res.first) {
5081 case X86::AX: DestReg = X86::RAX; break;
5082 case X86::DX: DestReg = X86::RDX; break;
5083 case X86::CX: DestReg = X86::RCX; break;
5084 case X86::BX: DestReg = X86::RBX; break;
5085 case X86::SI: DestReg = X86::RSI; break;
5086 case X86::DI: DestReg = X86::RDI; break;
5087 case X86::BP: DestReg = X86::RBP; break;
5088 case X86::SP: DestReg = X86::RSP; break;
5091 Res.first = DestReg;
5092 Res.second = Res.second = X86::GR64RegisterClass;