1 //===-- X86ISelLowering.h - X86 DAG Lowering Interface ----------*- C++ -*-===//
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 "X86TargetMachine.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/Function.h"
22 #include "llvm/Intrinsics.h"
23 #include "llvm/ADT/VectorExtras.h"
24 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/SSARegMap.h"
30 #include "llvm/Support/MathExtras.h"
31 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/Support/CommandLine.h"
36 static cl::opt<bool> EnableFastCC("enable-x86-fastcc", cl::Hidden,
37 cl::desc("Enable fastcc on X86"));
39 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
40 : TargetLowering(TM) {
41 Subtarget = &TM.getSubtarget<X86Subtarget>();
42 X86ScalarSSE = Subtarget->hasSSE2();
44 // Set up the TargetLowering object.
46 // X86 is weird, it always uses i8 for shift amounts and setcc results.
47 setShiftAmountType(MVT::i8);
48 setSetCCResultType(MVT::i8);
49 setSetCCResultContents(ZeroOrOneSetCCResult);
50 setSchedulingPreference(SchedulingForRegPressure);
51 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
52 setStackPointerRegisterToSaveRestore(X86::ESP);
54 if (!Subtarget->isTargetDarwin())
55 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
56 setUseUnderscoreSetJmpLongJmp(true);
58 // Add legal addressing mode scale values.
59 addLegalAddressScale(8);
60 addLegalAddressScale(4);
61 addLegalAddressScale(2);
62 // Enter the ones which require both scale + index last. These are more
64 addLegalAddressScale(9);
65 addLegalAddressScale(5);
66 addLegalAddressScale(3);
68 // Set up the register classes.
69 addRegisterClass(MVT::i8, X86::R8RegisterClass);
70 addRegisterClass(MVT::i16, X86::R16RegisterClass);
71 addRegisterClass(MVT::i32, X86::R32RegisterClass);
73 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
75 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
76 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
77 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
80 // No SSE i64 SINT_TO_FP, so expand i32 UINT_TO_FP instead.
81 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
83 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
85 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
87 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
88 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
89 // SSE has no i16 to fp conversion, only i32
91 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
93 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
94 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
97 // We can handle SINT_TO_FP and FP_TO_SINT from/to i64 even though i64
99 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
100 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
102 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
104 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
105 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
108 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
110 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
111 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
114 // Handle FP_TO_UINT by promoting the destination to a larger signed
116 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
117 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
118 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
120 if (X86ScalarSSE && !Subtarget->hasSSE3())
121 // Expand FP_TO_UINT into a select.
122 // FIXME: We would like to use a Custom expander here eventually to do
123 // the optimal thing for SSE vs. the default expansion in the legalizer.
124 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
126 // With SSE3 we can use fisttpll to convert to a signed i64.
127 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
129 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
130 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
132 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
133 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
134 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
135 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
136 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
137 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
138 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
139 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
140 setOperationAction(ISD::SEXTLOAD , MVT::i1 , Expand);
141 setOperationAction(ISD::FREM , MVT::f64 , Expand);
142 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
143 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
144 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
145 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
146 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
147 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
148 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
149 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
150 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
151 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
152 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
154 // These should be promoted to a larger select which is supported.
155 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
156 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
158 // X86 wants to expand cmov itself.
159 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
160 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
161 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
162 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
163 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
164 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
165 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
166 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
167 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
168 // X86 ret instruction may pop stack.
169 setOperationAction(ISD::RET , MVT::Other, Custom);
171 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
172 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
173 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
174 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
175 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
176 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
177 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
178 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
179 // X86 wants to expand memset / memcpy itself.
180 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
181 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
183 // We don't have line number support yet.
184 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
185 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
186 // FIXME - use subtarget debug flags
187 if (!Subtarget->isTargetDarwin())
188 setOperationAction(ISD::DEBUG_LABEL, MVT::Other, Expand);
190 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
191 setOperationAction(ISD::VASTART , MVT::Other, Custom);
193 // Use the default implementation.
194 setOperationAction(ISD::VAARG , MVT::Other, Expand);
195 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
196 setOperationAction(ISD::VAEND , MVT::Other, Expand);
197 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
198 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
199 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
201 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
202 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
205 // Set up the FP register classes.
206 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
207 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
209 // SSE has no load+extend ops
210 setOperationAction(ISD::EXTLOAD, MVT::f32, Expand);
211 setOperationAction(ISD::ZEXTLOAD, MVT::f32, Expand);
213 // Use ANDPD to simulate FABS.
214 setOperationAction(ISD::FABS , MVT::f64, Custom);
215 setOperationAction(ISD::FABS , MVT::f32, Custom);
217 // Use XORP to simulate FNEG.
218 setOperationAction(ISD::FNEG , MVT::f64, Custom);
219 setOperationAction(ISD::FNEG , MVT::f32, Custom);
221 // We don't support sin/cos/fmod
222 setOperationAction(ISD::FSIN , MVT::f64, Expand);
223 setOperationAction(ISD::FCOS , MVT::f64, Expand);
224 setOperationAction(ISD::FREM , MVT::f64, Expand);
225 setOperationAction(ISD::FSIN , MVT::f32, Expand);
226 setOperationAction(ISD::FCOS , MVT::f32, Expand);
227 setOperationAction(ISD::FREM , MVT::f32, Expand);
229 // Expand FP immediates into loads from the stack, except for the special
231 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
232 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
233 addLegalFPImmediate(+0.0); // xorps / xorpd
235 // Set up the FP register classes.
236 addRegisterClass(MVT::f64, X86::RFPRegisterClass);
238 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
241 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
242 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
245 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
246 addLegalFPImmediate(+0.0); // FLD0
247 addLegalFPImmediate(+1.0); // FLD1
248 addLegalFPImmediate(-0.0); // FLD0/FCHS
249 addLegalFPImmediate(-1.0); // FLD1/FCHS
252 // First set operation action for all vector types to expand. Then we
253 // will selectively turn on ones that can be effectively codegen'd.
254 for (unsigned VT = (unsigned)MVT::Vector + 1;
255 VT != (unsigned)MVT::LAST_VALUETYPE; VT++) {
256 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
257 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
258 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
259 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
260 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
261 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
262 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
265 if (Subtarget->hasMMX()) {
266 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
267 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
268 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
270 // FIXME: add MMX packed arithmetics
271 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Expand);
272 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Expand);
273 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Expand);
276 if (Subtarget->hasSSE1()) {
277 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
279 setOperationAction(ISD::AND, MVT::v4f32, Legal);
280 setOperationAction(ISD::OR, MVT::v4f32, Legal);
281 setOperationAction(ISD::XOR, MVT::v4f32, Legal);
282 setOperationAction(ISD::ADD, MVT::v4f32, Legal);
283 setOperationAction(ISD::SUB, MVT::v4f32, Legal);
284 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
285 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
286 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
287 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
288 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
289 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
292 if (Subtarget->hasSSE2()) {
293 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
294 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
295 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
296 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
297 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
299 setOperationAction(ISD::ADD, MVT::v2f64, Legal);
300 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
301 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
302 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
303 setOperationAction(ISD::SUB, MVT::v2f64, Legal);
304 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
305 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
306 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
307 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
308 setOperationAction(ISD::MUL, MVT::v2f64, Legal);
310 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
311 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
312 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
313 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
314 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
315 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
317 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
318 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
319 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
320 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
321 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
323 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
324 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
325 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
326 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
327 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
328 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
330 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
331 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
332 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
333 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
334 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
335 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
336 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
337 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
338 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
339 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
340 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
341 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
344 // Custom lower v2i64 and v2f64 selects.
345 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
346 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
347 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
348 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
351 // We want to custom lower some of our intrinsics.
352 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
354 computeRegisterProperties();
356 // FIXME: These should be based on subtarget info. Plus, the values should
357 // be smaller when we are in optimizing for size mode.
358 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
359 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
360 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
361 allowUnalignedMemoryAccesses = true; // x86 supports it!
364 std::vector<SDOperand>
365 X86TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
366 if (F.getCallingConv() == CallingConv::Fast && EnableFastCC)
367 return LowerFastCCArguments(F, DAG);
368 return LowerCCCArguments(F, DAG);
371 std::pair<SDOperand, SDOperand>
372 X86TargetLowering::LowerCallTo(SDOperand Chain, const Type *RetTy,
373 bool isVarArg, unsigned CallingConv,
375 SDOperand Callee, ArgListTy &Args,
377 assert((!isVarArg || CallingConv == CallingConv::C) &&
378 "Only C takes varargs!");
380 // If the callee is a GlobalAddress node (quite common, every direct call is)
381 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
382 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
383 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
384 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
385 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
387 if (CallingConv == CallingConv::Fast && EnableFastCC)
388 return LowerFastCCCallTo(Chain, RetTy, isTailCall, Callee, Args, DAG);
389 return LowerCCCCallTo(Chain, RetTy, isVarArg, isTailCall, Callee, Args, DAG);
392 //===----------------------------------------------------------------------===//
393 // C Calling Convention implementation
394 //===----------------------------------------------------------------------===//
396 std::vector<SDOperand>
397 X86TargetLowering::LowerCCCArguments(Function &F, SelectionDAG &DAG) {
398 std::vector<SDOperand> ArgValues;
400 MachineFunction &MF = DAG.getMachineFunction();
401 MachineFrameInfo *MFI = MF.getFrameInfo();
403 // Add DAG nodes to load the arguments... On entry to a function on the X86,
404 // the stack frame looks like this:
406 // [ESP] -- return address
407 // [ESP + 4] -- first argument (leftmost lexically)
408 // [ESP + 8] -- second argument, if first argument is four bytes in size
411 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
412 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
413 MVT::ValueType ObjectVT = getValueType(I->getType());
414 unsigned ArgIncrement = 4;
417 default: assert(0 && "Unhandled argument type!");
419 case MVT::i8: ObjSize = 1; break;
420 case MVT::i16: ObjSize = 2; break;
421 case MVT::i32: ObjSize = 4; break;
422 case MVT::i64: ObjSize = ArgIncrement = 8; break;
423 case MVT::f32: ObjSize = 4; break;
424 case MVT::f64: ObjSize = ArgIncrement = 8; break;
426 // Create the frame index object for this incoming parameter...
427 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
429 // Create the SelectionDAG nodes corresponding to a load from this parameter
430 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
432 // Don't codegen dead arguments. FIXME: remove this check when we can nuke
436 ArgValue = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
437 DAG.getSrcValue(NULL));
439 if (MVT::isInteger(ObjectVT))
440 ArgValue = DAG.getConstant(0, ObjectVT);
442 ArgValue = DAG.getConstantFP(0, ObjectVT);
444 ArgValues.push_back(ArgValue);
446 ArgOffset += ArgIncrement; // Move on to the next argument...
449 // If the function takes variable number of arguments, make a frame index for
450 // the start of the first vararg value... for expansion of llvm.va_start.
452 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
453 ReturnAddrIndex = 0; // No return address slot generated yet.
454 BytesToPopOnReturn = 0; // Callee pops nothing.
455 BytesCallerReserves = ArgOffset;
459 std::pair<SDOperand, SDOperand>
460 X86TargetLowering::LowerCCCCallTo(SDOperand Chain, const Type *RetTy,
461 bool isVarArg, bool isTailCall,
462 SDOperand Callee, ArgListTy &Args,
464 // Count how many bytes are to be pushed on the stack.
465 unsigned NumBytes = 0;
469 Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(0, getPointerTy()));
471 for (unsigned i = 0, e = Args.size(); i != e; ++i)
472 switch (getValueType(Args[i].second)) {
473 default: assert(0 && "Unknown value type!");
487 Chain = DAG.getCALLSEQ_START(Chain,
488 DAG.getConstant(NumBytes, getPointerTy()));
490 // Arguments go on the stack in reverse order, as specified by the ABI.
491 unsigned ArgOffset = 0;
492 SDOperand StackPtr = DAG.getRegister(X86::ESP, MVT::i32);
493 std::vector<SDOperand> Stores;
495 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
496 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
497 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
499 switch (getValueType(Args[i].second)) {
500 default: assert(0 && "Unexpected ValueType for argument!");
504 // Promote the integer to 32 bits. If the input type is signed use a
505 // sign extend, otherwise use a zero extend.
506 if (Args[i].second->isSigned())
507 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
509 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
514 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
515 Args[i].first, PtrOff,
516 DAG.getSrcValue(NULL)));
521 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
522 Args[i].first, PtrOff,
523 DAG.getSrcValue(NULL)));
528 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
531 std::vector<MVT::ValueType> RetVals;
532 MVT::ValueType RetTyVT = getValueType(RetTy);
533 RetVals.push_back(MVT::Other);
535 // The result values produced have to be legal. Promote the result.
537 case MVT::isVoid: break;
539 RetVals.push_back(RetTyVT);
544 RetVals.push_back(MVT::i32);
548 RetVals.push_back(MVT::f32);
550 RetVals.push_back(MVT::f64);
553 RetVals.push_back(MVT::i32);
554 RetVals.push_back(MVT::i32);
558 std::vector<MVT::ValueType> NodeTys;
559 NodeTys.push_back(MVT::Other); // Returns a chain
560 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
561 std::vector<SDOperand> Ops;
562 Ops.push_back(Chain);
563 Ops.push_back(Callee);
565 // FIXME: Do not generate X86ISD::TAILCALL for now.
566 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops);
567 SDOperand InFlag = Chain.getValue(1);
570 NodeTys.push_back(MVT::Other); // Returns a chain
571 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
573 Ops.push_back(Chain);
574 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
575 Ops.push_back(DAG.getConstant(0, getPointerTy()));
576 Ops.push_back(InFlag);
577 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, Ops);
578 InFlag = Chain.getValue(1);
581 if (RetTyVT != MVT::isVoid) {
583 default: assert(0 && "Unknown value type to return!");
586 RetVal = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag);
587 Chain = RetVal.getValue(1);
588 if (RetTyVT == MVT::i1)
589 RetVal = DAG.getNode(ISD::TRUNCATE, MVT::i1, RetVal);
592 RetVal = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag);
593 Chain = RetVal.getValue(1);
596 RetVal = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag);
597 Chain = RetVal.getValue(1);
600 SDOperand Lo = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag);
601 SDOperand Hi = DAG.getCopyFromReg(Lo.getValue(1), X86::EDX, MVT::i32,
603 RetVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
604 Chain = Hi.getValue(1);
609 std::vector<MVT::ValueType> Tys;
610 Tys.push_back(MVT::f64);
611 Tys.push_back(MVT::Other);
612 Tys.push_back(MVT::Flag);
613 std::vector<SDOperand> Ops;
614 Ops.push_back(Chain);
615 Ops.push_back(InFlag);
616 RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, Ops);
617 Chain = RetVal.getValue(1);
618 InFlag = RetVal.getValue(2);
620 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
621 // shouldn't be necessary except that RFP cannot be live across
622 // multiple blocks. When stackifier is fixed, they can be uncoupled.
623 MachineFunction &MF = DAG.getMachineFunction();
624 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
625 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
627 Tys.push_back(MVT::Other);
629 Ops.push_back(Chain);
630 Ops.push_back(RetVal);
631 Ops.push_back(StackSlot);
632 Ops.push_back(DAG.getValueType(RetTyVT));
633 Ops.push_back(InFlag);
634 Chain = DAG.getNode(X86ISD::FST, Tys, Ops);
635 RetVal = DAG.getLoad(RetTyVT, Chain, StackSlot,
636 DAG.getSrcValue(NULL));
637 Chain = RetVal.getValue(1);
640 if (RetTyVT == MVT::f32 && !X86ScalarSSE)
641 // FIXME: we would really like to remember that this FP_ROUND
642 // operation is okay to eliminate if we allow excess FP precision.
643 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
649 return std::make_pair(RetVal, Chain);
652 //===----------------------------------------------------------------------===//
653 // Fast Calling Convention implementation
654 //===----------------------------------------------------------------------===//
656 // The X86 'fast' calling convention passes up to two integer arguments in
657 // registers (an appropriate portion of EAX/EDX), passes arguments in C order,
658 // and requires that the callee pop its arguments off the stack (allowing proper
659 // tail calls), and has the same return value conventions as C calling convs.
661 // This calling convention always arranges for the callee pop value to be 8n+4
662 // bytes, which is needed for tail recursion elimination and stack alignment
665 // Note that this can be enhanced in the future to pass fp vals in registers
666 // (when we have a global fp allocator) and do other tricks.
669 /// AddLiveIn - This helper function adds the specified physical register to the
670 /// MachineFunction as a live in value. It also creates a corresponding virtual
672 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
673 TargetRegisterClass *RC) {
674 assert(RC->contains(PReg) && "Not the correct regclass!");
675 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
676 MF.addLiveIn(PReg, VReg);
680 // FASTCC_NUM_INT_ARGS_INREGS - This is the max number of integer arguments
681 // to pass in registers. 0 is none, 1 is is "use EAX", 2 is "use EAX and
682 // EDX". Anything more is illegal.
684 // FIXME: The linscan register allocator currently has problem with
685 // coalescing. At the time of this writing, whenever it decides to coalesce
686 // a physreg with a virtreg, this increases the size of the physreg's live
687 // range, and the live range cannot ever be reduced. This causes problems if
688 // too many physregs are coaleced with virtregs, which can cause the register
689 // allocator to wedge itself.
691 // This code triggers this problem more often if we pass args in registers,
692 // so disable it until this is fixed.
694 // NOTE: this isn't marked const, so that GCC doesn't emit annoying warnings
695 // about code being dead.
697 static unsigned FASTCC_NUM_INT_ARGS_INREGS = 0;
700 std::vector<SDOperand>
701 X86TargetLowering::LowerFastCCArguments(Function &F, SelectionDAG &DAG) {
702 std::vector<SDOperand> ArgValues;
704 MachineFunction &MF = DAG.getMachineFunction();
705 MachineFrameInfo *MFI = MF.getFrameInfo();
707 // Add DAG nodes to load the arguments... On entry to a function the stack
708 // frame looks like this:
710 // [ESP] -- return address
711 // [ESP + 4] -- first nonreg argument (leftmost lexically)
712 // [ESP + 8] -- second nonreg argument, if first argument is 4 bytes in size
714 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
716 // Keep track of the number of integer regs passed so far. This can be either
717 // 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
719 unsigned NumIntRegs = 0;
721 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
722 MVT::ValueType ObjectVT = getValueType(I->getType());
723 unsigned ArgIncrement = 4;
724 unsigned ObjSize = 0;
728 default: assert(0 && "Unhandled argument type!");
731 if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
732 if (!I->use_empty()) {
733 unsigned VReg = AddLiveIn(MF, NumIntRegs ? X86::DL : X86::AL,
734 X86::R8RegisterClass);
735 ArgValue = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i8);
736 DAG.setRoot(ArgValue.getValue(1));
737 if (ObjectVT == MVT::i1)
738 // FIXME: Should insert a assertzext here.
739 ArgValue = DAG.getNode(ISD::TRUNCATE, MVT::i1, ArgValue);
748 if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
749 if (!I->use_empty()) {
750 unsigned VReg = AddLiveIn(MF, NumIntRegs ? X86::DX : X86::AX,
751 X86::R16RegisterClass);
752 ArgValue = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i16);
753 DAG.setRoot(ArgValue.getValue(1));
761 if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
762 if (!I->use_empty()) {
763 unsigned VReg = AddLiveIn(MF, NumIntRegs ? X86::EDX : X86::EAX,
764 X86::R32RegisterClass);
765 ArgValue = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
766 DAG.setRoot(ArgValue.getValue(1));
774 if (NumIntRegs+2 <= FASTCC_NUM_INT_ARGS_INREGS) {
775 if (!I->use_empty()) {
776 unsigned BotReg = AddLiveIn(MF, X86::EAX, X86::R32RegisterClass);
777 unsigned TopReg = AddLiveIn(MF, X86::EDX, X86::R32RegisterClass);
779 SDOperand Low = DAG.getCopyFromReg(DAG.getRoot(), BotReg, MVT::i32);
780 SDOperand Hi = DAG.getCopyFromReg(Low.getValue(1), TopReg, MVT::i32);
781 DAG.setRoot(Hi.getValue(1));
783 ArgValue = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Low, Hi);
787 } else if (NumIntRegs+1 <= FASTCC_NUM_INT_ARGS_INREGS) {
788 if (!I->use_empty()) {
789 unsigned BotReg = AddLiveIn(MF, X86::EDX, X86::R32RegisterClass);
790 SDOperand Low = DAG.getCopyFromReg(DAG.getRoot(), BotReg, MVT::i32);
791 DAG.setRoot(Low.getValue(1));
793 // Load the high part from memory.
794 // Create the frame index object for this incoming parameter...
795 int FI = MFI->CreateFixedObject(4, ArgOffset);
796 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
797 SDOperand Hi = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
798 DAG.getSrcValue(NULL));
799 ArgValue = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Low, Hi);
802 NumIntRegs = FASTCC_NUM_INT_ARGS_INREGS;
805 ObjSize = ArgIncrement = 8;
807 case MVT::f32: ObjSize = 4; break;
808 case MVT::f64: ObjSize = ArgIncrement = 8; break;
811 // Don't codegen dead arguments. FIXME: remove this check when we can nuke
813 if (ObjSize && !I->use_empty()) {
814 // Create the frame index object for this incoming parameter...
815 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
817 // Create the SelectionDAG nodes corresponding to a load from this
819 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
821 ArgValue = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
822 DAG.getSrcValue(NULL));
823 } else if (ArgValue.Val == 0) {
824 if (MVT::isInteger(ObjectVT))
825 ArgValue = DAG.getConstant(0, ObjectVT);
827 ArgValue = DAG.getConstantFP(0, ObjectVT);
829 ArgValues.push_back(ArgValue);
832 ArgOffset += ArgIncrement; // Move on to the next argument.
835 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
836 // arguments and the arguments after the retaddr has been pushed are aligned.
837 if ((ArgOffset & 7) == 0)
840 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
841 ReturnAddrIndex = 0; // No return address slot generated yet.
842 BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
843 BytesCallerReserves = 0;
845 // Finally, inform the code generator which regs we return values in.
846 switch (getValueType(F.getReturnType())) {
847 default: assert(0 && "Unknown type!");
848 case MVT::isVoid: break;
853 MF.addLiveOut(X86::EAX);
856 MF.addLiveOut(X86::EAX);
857 MF.addLiveOut(X86::EDX);
861 MF.addLiveOut(X86::ST0);
867 std::pair<SDOperand, SDOperand>
868 X86TargetLowering::LowerFastCCCallTo(SDOperand Chain, const Type *RetTy,
869 bool isTailCall, SDOperand Callee,
870 ArgListTy &Args, SelectionDAG &DAG) {
871 // Count how many bytes are to be pushed on the stack.
872 unsigned NumBytes = 0;
874 // Keep track of the number of integer regs passed so far. This can be either
875 // 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
877 unsigned NumIntRegs = 0;
879 for (unsigned i = 0, e = Args.size(); i != e; ++i)
880 switch (getValueType(Args[i].second)) {
881 default: assert(0 && "Unknown value type!");
886 if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
895 if (NumIntRegs+2 <= FASTCC_NUM_INT_ARGS_INREGS) {
898 } else if (NumIntRegs+1 <= FASTCC_NUM_INT_ARGS_INREGS) {
899 NumIntRegs = FASTCC_NUM_INT_ARGS_INREGS;
910 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
911 // arguments and the arguments after the retaddr has been pushed are aligned.
912 if ((NumBytes & 7) == 0)
915 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
917 // Arguments go on the stack in reverse order, as specified by the ABI.
918 unsigned ArgOffset = 0;
919 SDOperand StackPtr = DAG.getRegister(X86::ESP, MVT::i32);
921 std::vector<SDOperand> Stores;
922 std::vector<SDOperand> RegValuesToPass;
923 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
924 switch (getValueType(Args[i].second)) {
925 default: assert(0 && "Unexpected ValueType for argument!");
927 Args[i].first = DAG.getNode(ISD::ANY_EXTEND, MVT::i8, Args[i].first);
932 if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
933 RegValuesToPass.push_back(Args[i].first);
939 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
940 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
941 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
942 Args[i].first, PtrOff,
943 DAG.getSrcValue(NULL)));
948 // Can pass (at least) part of it in regs?
949 if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
950 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
951 Args[i].first, DAG.getConstant(1, MVT::i32));
952 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
953 Args[i].first, DAG.getConstant(0, MVT::i32));
954 RegValuesToPass.push_back(Lo);
957 // Pass both parts in regs?
958 if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
959 RegValuesToPass.push_back(Hi);
962 // Pass the high part in memory.
963 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
964 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
965 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
966 Hi, PtrOff, DAG.getSrcValue(NULL)));
973 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
974 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
975 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
976 Args[i].first, PtrOff,
977 DAG.getSrcValue(NULL)));
983 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
985 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
986 // arguments and the arguments after the retaddr has been pushed are aligned.
987 if ((ArgOffset & 7) == 0)
990 std::vector<MVT::ValueType> RetVals;
991 MVT::ValueType RetTyVT = getValueType(RetTy);
993 RetVals.push_back(MVT::Other);
995 // The result values produced have to be legal. Promote the result.
997 case MVT::isVoid: break;
999 RetVals.push_back(RetTyVT);
1004 RetVals.push_back(MVT::i32);
1008 RetVals.push_back(MVT::f32);
1010 RetVals.push_back(MVT::f64);
1013 RetVals.push_back(MVT::i32);
1014 RetVals.push_back(MVT::i32);
1018 // Build a sequence of copy-to-reg nodes chained together with token chain
1019 // and flag operands which copy the outgoing args into registers.
1021 for (unsigned i = 0, e = RegValuesToPass.size(); i != e; ++i) {
1023 SDOperand RegToPass = RegValuesToPass[i];
1024 switch (RegToPass.getValueType()) {
1025 default: assert(0 && "Bad thing to pass in regs");
1027 CCReg = (i == 0) ? X86::AL : X86::DL;
1030 CCReg = (i == 0) ? X86::AX : X86::DX;
1033 CCReg = (i == 0) ? X86::EAX : X86::EDX;
1037 Chain = DAG.getCopyToReg(Chain, CCReg, RegToPass, InFlag);
1038 InFlag = Chain.getValue(1);
1041 std::vector<MVT::ValueType> NodeTys;
1042 NodeTys.push_back(MVT::Other); // Returns a chain
1043 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1044 std::vector<SDOperand> Ops;
1045 Ops.push_back(Chain);
1046 Ops.push_back(Callee);
1048 Ops.push_back(InFlag);
1050 // FIXME: Do not generate X86ISD::TAILCALL for now.
1051 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops);
1052 InFlag = Chain.getValue(1);
1055 NodeTys.push_back(MVT::Other); // Returns a chain
1056 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1058 Ops.push_back(Chain);
1059 Ops.push_back(DAG.getConstant(ArgOffset, getPointerTy()));
1060 Ops.push_back(DAG.getConstant(ArgOffset, getPointerTy()));
1061 Ops.push_back(InFlag);
1062 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, Ops);
1063 InFlag = Chain.getValue(1);
1066 if (RetTyVT != MVT::isVoid) {
1068 default: assert(0 && "Unknown value type to return!");
1071 RetVal = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag);
1072 Chain = RetVal.getValue(1);
1073 if (RetTyVT == MVT::i1)
1074 RetVal = DAG.getNode(ISD::TRUNCATE, MVT::i1, RetVal);
1077 RetVal = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag);
1078 Chain = RetVal.getValue(1);
1081 RetVal = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag);
1082 Chain = RetVal.getValue(1);
1085 SDOperand Lo = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag);
1086 SDOperand Hi = DAG.getCopyFromReg(Lo.getValue(1), X86::EDX, MVT::i32,
1088 RetVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
1089 Chain = Hi.getValue(1);
1094 std::vector<MVT::ValueType> Tys;
1095 Tys.push_back(MVT::f64);
1096 Tys.push_back(MVT::Other);
1097 Tys.push_back(MVT::Flag);
1098 std::vector<SDOperand> Ops;
1099 Ops.push_back(Chain);
1100 Ops.push_back(InFlag);
1101 RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, Ops);
1102 Chain = RetVal.getValue(1);
1103 InFlag = RetVal.getValue(2);
1105 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
1106 // shouldn't be necessary except that RFP cannot be live across
1107 // multiple blocks. When stackifier is fixed, they can be uncoupled.
1108 MachineFunction &MF = DAG.getMachineFunction();
1109 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
1110 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
1112 Tys.push_back(MVT::Other);
1114 Ops.push_back(Chain);
1115 Ops.push_back(RetVal);
1116 Ops.push_back(StackSlot);
1117 Ops.push_back(DAG.getValueType(RetTyVT));
1118 Ops.push_back(InFlag);
1119 Chain = DAG.getNode(X86ISD::FST, Tys, Ops);
1120 RetVal = DAG.getLoad(RetTyVT, Chain, StackSlot,
1121 DAG.getSrcValue(NULL));
1122 Chain = RetVal.getValue(1);
1125 if (RetTyVT == MVT::f32 && !X86ScalarSSE)
1126 // FIXME: we would really like to remember that this FP_ROUND
1127 // operation is okay to eliminate if we allow excess FP precision.
1128 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
1134 return std::make_pair(RetVal, Chain);
1137 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1138 if (ReturnAddrIndex == 0) {
1139 // Set up a frame object for the return address.
1140 MachineFunction &MF = DAG.getMachineFunction();
1141 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1144 return DAG.getFrameIndex(ReturnAddrIndex, MVT::i32);
1149 std::pair<SDOperand, SDOperand> X86TargetLowering::
1150 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
1151 SelectionDAG &DAG) {
1153 if (Depth) // Depths > 0 not supported yet!
1154 Result = DAG.getConstant(0, getPointerTy());
1156 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
1157 if (!isFrameAddress)
1158 // Just load the return address
1159 Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(), RetAddrFI,
1160 DAG.getSrcValue(NULL));
1162 Result = DAG.getNode(ISD::SUB, MVT::i32, RetAddrFI,
1163 DAG.getConstant(4, MVT::i32));
1165 return std::make_pair(Result, Chain);
1168 /// getCondBrOpcodeForX86CC - Returns the X86 conditional branch opcode
1169 /// which corresponds to the condition code.
1170 static unsigned getCondBrOpcodeForX86CC(unsigned X86CC) {
1172 default: assert(0 && "Unknown X86 conditional code!");
1173 case X86ISD::COND_A: return X86::JA;
1174 case X86ISD::COND_AE: return X86::JAE;
1175 case X86ISD::COND_B: return X86::JB;
1176 case X86ISD::COND_BE: return X86::JBE;
1177 case X86ISD::COND_E: return X86::JE;
1178 case X86ISD::COND_G: return X86::JG;
1179 case X86ISD::COND_GE: return X86::JGE;
1180 case X86ISD::COND_L: return X86::JL;
1181 case X86ISD::COND_LE: return X86::JLE;
1182 case X86ISD::COND_NE: return X86::JNE;
1183 case X86ISD::COND_NO: return X86::JNO;
1184 case X86ISD::COND_NP: return X86::JNP;
1185 case X86ISD::COND_NS: return X86::JNS;
1186 case X86ISD::COND_O: return X86::JO;
1187 case X86ISD::COND_P: return X86::JP;
1188 case X86ISD::COND_S: return X86::JS;
1192 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1193 /// specific condition code. It returns a false if it cannot do a direct
1194 /// translation. X86CC is the translated CondCode. Flip is set to true if the
1195 /// the order of comparison operands should be flipped.
1196 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1197 unsigned &X86CC, bool &Flip) {
1199 X86CC = X86ISD::COND_INVALID;
1201 switch (SetCCOpcode) {
1203 case ISD::SETEQ: X86CC = X86ISD::COND_E; break;
1204 case ISD::SETGT: X86CC = X86ISD::COND_G; break;
1205 case ISD::SETGE: X86CC = X86ISD::COND_GE; break;
1206 case ISD::SETLT: X86CC = X86ISD::COND_L; break;
1207 case ISD::SETLE: X86CC = X86ISD::COND_LE; break;
1208 case ISD::SETNE: X86CC = X86ISD::COND_NE; break;
1209 case ISD::SETULT: X86CC = X86ISD::COND_B; break;
1210 case ISD::SETUGT: X86CC = X86ISD::COND_A; break;
1211 case ISD::SETULE: X86CC = X86ISD::COND_BE; break;
1212 case ISD::SETUGE: X86CC = X86ISD::COND_AE; break;
1215 // On a floating point condition, the flags are set as follows:
1217 // 0 | 0 | 0 | X > Y
1218 // 0 | 0 | 1 | X < Y
1219 // 1 | 0 | 0 | X == Y
1220 // 1 | 1 | 1 | unordered
1221 switch (SetCCOpcode) {
1224 case ISD::SETEQ: X86CC = X86ISD::COND_E; break;
1225 case ISD::SETOLT: Flip = true; // Fallthrough
1227 case ISD::SETGT: X86CC = X86ISD::COND_A; break;
1228 case ISD::SETOLE: Flip = true; // Fallthrough
1230 case ISD::SETGE: X86CC = X86ISD::COND_AE; break;
1231 case ISD::SETUGT: Flip = true; // Fallthrough
1233 case ISD::SETLT: X86CC = X86ISD::COND_B; break;
1234 case ISD::SETUGE: Flip = true; // Fallthrough
1236 case ISD::SETLE: X86CC = X86ISD::COND_BE; break;
1238 case ISD::SETNE: X86CC = X86ISD::COND_NE; break;
1239 case ISD::SETUO: X86CC = X86ISD::COND_P; break;
1240 case ISD::SETO: X86CC = X86ISD::COND_NP; break;
1244 return X86CC != X86ISD::COND_INVALID;
1247 static bool translateX86CC(SDOperand CC, bool isFP, unsigned &X86CC,
1249 return translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC, Flip);
1252 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1253 /// code. Current x86 isa includes the following FP cmov instructions:
1254 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1255 static bool hasFPCMov(unsigned X86CC) {
1259 case X86ISD::COND_B:
1260 case X86ISD::COND_BE:
1261 case X86ISD::COND_E:
1262 case X86ISD::COND_P:
1263 case X86ISD::COND_A:
1264 case X86ISD::COND_AE:
1265 case X86ISD::COND_NE:
1266 case X86ISD::COND_NP:
1272 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
1273 MachineBasicBlock *BB) {
1274 switch (MI->getOpcode()) {
1275 default: assert(false && "Unexpected instr type to insert");
1276 case X86::CMOV_FR32:
1277 case X86::CMOV_FR64:
1278 case X86::CMOV_V4F32:
1279 case X86::CMOV_V2F64:
1280 case X86::CMOV_V2I64: {
1281 // To "insert" a SELECT_CC instruction, we actually have to insert the
1282 // diamond control-flow pattern. The incoming instruction knows the
1283 // destination vreg to set, the condition code register to branch on, the
1284 // true/false values to select between, and a branch opcode to use.
1285 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1286 ilist<MachineBasicBlock>::iterator It = BB;
1292 // cmpTY ccX, r1, r2
1294 // fallthrough --> copy0MBB
1295 MachineBasicBlock *thisMBB = BB;
1296 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1297 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1298 unsigned Opc = getCondBrOpcodeForX86CC(MI->getOperand(3).getImmedValue());
1299 BuildMI(BB, Opc, 1).addMBB(sinkMBB);
1300 MachineFunction *F = BB->getParent();
1301 F->getBasicBlockList().insert(It, copy0MBB);
1302 F->getBasicBlockList().insert(It, sinkMBB);
1303 // Update machine-CFG edges by first adding all successors of the current
1304 // block to the new block which will contain the Phi node for the select.
1305 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
1306 e = BB->succ_end(); i != e; ++i)
1307 sinkMBB->addSuccessor(*i);
1308 // Next, remove all successors of the current block, and add the true
1309 // and fallthrough blocks as its successors.
1310 while(!BB->succ_empty())
1311 BB->removeSuccessor(BB->succ_begin());
1312 BB->addSuccessor(copy0MBB);
1313 BB->addSuccessor(sinkMBB);
1316 // %FalseValue = ...
1317 // # fallthrough to sinkMBB
1320 // Update machine-CFG edges
1321 BB->addSuccessor(sinkMBB);
1324 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1327 BuildMI(BB, X86::PHI, 4, MI->getOperand(0).getReg())
1328 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
1329 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
1331 delete MI; // The pseudo instruction is gone now.
1335 case X86::FP_TO_INT16_IN_MEM:
1336 case X86::FP_TO_INT32_IN_MEM:
1337 case X86::FP_TO_INT64_IN_MEM: {
1338 // Change the floating point control register to use "round towards zero"
1339 // mode when truncating to an integer value.
1340 MachineFunction *F = BB->getParent();
1341 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
1342 addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
1344 // Load the old value of the high byte of the control word...
1346 F->getSSARegMap()->createVirtualRegister(X86::R16RegisterClass);
1347 addFrameReference(BuildMI(BB, X86::MOV16rm, 4, OldCW), CWFrameIdx);
1349 // Set the high part to be round to zero...
1350 addFrameReference(BuildMI(BB, X86::MOV16mi, 5), CWFrameIdx).addImm(0xC7F);
1352 // Reload the modified control word now...
1353 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
1355 // Restore the memory image of control word to original value
1356 addFrameReference(BuildMI(BB, X86::MOV16mr, 5), CWFrameIdx).addReg(OldCW);
1358 // Get the X86 opcode to use.
1360 switch (MI->getOpcode()) {
1361 default: assert(0 && "illegal opcode!");
1362 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
1363 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
1364 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
1368 MachineOperand &Op = MI->getOperand(0);
1369 if (Op.isRegister()) {
1370 AM.BaseType = X86AddressMode::RegBase;
1371 AM.Base.Reg = Op.getReg();
1373 AM.BaseType = X86AddressMode::FrameIndexBase;
1374 AM.Base.FrameIndex = Op.getFrameIndex();
1376 Op = MI->getOperand(1);
1377 if (Op.isImmediate())
1378 AM.Scale = Op.getImmedValue();
1379 Op = MI->getOperand(2);
1380 if (Op.isImmediate())
1381 AM.IndexReg = Op.getImmedValue();
1382 Op = MI->getOperand(3);
1383 if (Op.isGlobalAddress()) {
1384 AM.GV = Op.getGlobal();
1386 AM.Disp = Op.getImmedValue();
1388 addFullAddress(BuildMI(BB, Opc, 5), AM).addReg(MI->getOperand(4).getReg());
1390 // Reload the original control word now.
1391 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
1393 delete MI; // The pseudo instruction is gone now.
1400 //===----------------------------------------------------------------------===//
1401 // X86 Custom Lowering Hooks
1402 //===----------------------------------------------------------------------===//
1404 /// DarwinGVRequiresExtraLoad - true if accessing the GV requires an extra
1405 /// load. For Darwin, external and weak symbols are indirect, loading the value
1406 /// at address GV rather then the value of GV itself. This means that the
1407 /// GlobalAddress must be in the base or index register of the address, not the
1408 /// GV offset field.
1409 static bool DarwinGVRequiresExtraLoad(GlobalValue *GV) {
1410 return (GV->hasWeakLinkage() || GV->hasLinkOnceLinkage() ||
1411 (GV->isExternal() && !GV->hasNotBeenReadFromBytecode()));
1414 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1415 /// true if Op is undef or if its value falls within the specified range (L, H].
1416 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1417 if (Op.getOpcode() == ISD::UNDEF)
1420 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1421 return (Val >= Low && Val < Hi);
1424 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1425 /// true if Op is undef or if its value equal to the specified value.
1426 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1427 if (Op.getOpcode() == ISD::UNDEF)
1429 return cast<ConstantSDNode>(Op)->getValue() == Val;
1432 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1433 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1434 bool X86::isPSHUFDMask(SDNode *N) {
1435 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1437 if (N->getNumOperands() != 4)
1440 // Check if the value doesn't reference the second vector.
1441 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1442 SDOperand Arg = N->getOperand(i);
1443 if (Arg.getOpcode() == ISD::UNDEF) continue;
1444 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1445 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
1452 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1453 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1454 bool X86::isPSHUFHWMask(SDNode *N) {
1455 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1457 if (N->getNumOperands() != 8)
1460 // Lower quadword copied in order.
1461 for (unsigned i = 0; i != 4; ++i) {
1462 SDOperand Arg = N->getOperand(i);
1463 if (Arg.getOpcode() == ISD::UNDEF) continue;
1464 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1465 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1469 // Upper quadword shuffled.
1470 for (unsigned i = 4; i != 8; ++i) {
1471 SDOperand Arg = N->getOperand(i);
1472 if (Arg.getOpcode() == ISD::UNDEF) continue;
1473 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1474 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1475 if (Val < 4 || Val > 7)
1482 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1483 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1484 bool X86::isPSHUFLWMask(SDNode *N) {
1485 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1487 if (N->getNumOperands() != 8)
1490 // Upper quadword copied in order.
1491 for (unsigned i = 4; i != 8; ++i)
1492 if (!isUndefOrEqual(N->getOperand(i), i))
1495 // Lower quadword shuffled.
1496 for (unsigned i = 0; i != 4; ++i)
1497 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1503 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1504 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1505 static bool isSHUFPMask(std::vector<SDOperand> &N) {
1506 unsigned NumElems = N.size();
1507 if (NumElems != 2 && NumElems != 4) return false;
1509 unsigned Half = NumElems / 2;
1510 for (unsigned i = 0; i < Half; ++i)
1511 if (!isUndefOrInRange(N[i], 0, NumElems))
1513 for (unsigned i = Half; i < NumElems; ++i)
1514 if (!isUndefOrInRange(N[i], NumElems, NumElems*2))
1520 bool X86::isSHUFPMask(SDNode *N) {
1521 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1522 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
1523 return ::isSHUFPMask(Ops);
1526 /// isCommutedSHUFP - Returns true if the shuffle mask is except
1527 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
1528 /// half elements to come from vector 1 (which would equal the dest.) and
1529 /// the upper half to come from vector 2.
1530 static bool isCommutedSHUFP(std::vector<SDOperand> &Ops) {
1531 unsigned NumElems = Ops.size();
1532 if (NumElems != 2 && NumElems != 4) return false;
1534 unsigned Half = NumElems / 2;
1535 for (unsigned i = 0; i < Half; ++i)
1536 if (!isUndefOrInRange(Ops[i], NumElems, NumElems*2))
1538 for (unsigned i = Half; i < NumElems; ++i)
1539 if (!isUndefOrInRange(Ops[i], 0, NumElems))
1544 static bool isCommutedSHUFP(SDNode *N) {
1545 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1546 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
1547 return isCommutedSHUFP(Ops);
1550 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
1551 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
1552 bool X86::isMOVHLPSMask(SDNode *N) {
1553 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1555 if (N->getNumOperands() != 4)
1558 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
1559 return isUndefOrEqual(N->getOperand(0), 6) &&
1560 isUndefOrEqual(N->getOperand(1), 7) &&
1561 isUndefOrEqual(N->getOperand(2), 2) &&
1562 isUndefOrEqual(N->getOperand(3), 3);
1565 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
1566 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
1567 bool X86::isMOVLPMask(SDNode *N) {
1568 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1570 unsigned NumElems = N->getNumOperands();
1571 if (NumElems != 2 && NumElems != 4)
1574 for (unsigned i = 0; i < NumElems/2; ++i)
1575 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
1578 for (unsigned i = NumElems/2; i < NumElems; ++i)
1579 if (!isUndefOrEqual(N->getOperand(i), i))
1585 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
1586 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
1588 bool X86::isMOVHPMask(SDNode *N) {
1589 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1591 unsigned NumElems = N->getNumOperands();
1592 if (NumElems != 2 && NumElems != 4)
1595 for (unsigned i = 0; i < NumElems/2; ++i)
1596 if (!isUndefOrEqual(N->getOperand(i), i))
1599 for (unsigned i = 0; i < NumElems/2; ++i) {
1600 SDOperand Arg = N->getOperand(i + NumElems/2);
1601 if (!isUndefOrEqual(Arg, i + NumElems))
1608 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
1609 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
1610 bool static isUNPCKLMask(std::vector<SDOperand> &N, bool V2IsSplat = false) {
1611 unsigned NumElems = N.size();
1612 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1615 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1616 SDOperand BitI = N[i];
1617 SDOperand BitI1 = N[i+1];
1618 if (!isUndefOrEqual(BitI, j))
1621 if (isUndefOrEqual(BitI1, NumElems))
1624 if (!isUndefOrEqual(BitI1, j + NumElems))
1632 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
1633 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1634 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
1635 return ::isUNPCKLMask(Ops, V2IsSplat);
1638 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
1639 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
1640 bool static isUNPCKHMask(std::vector<SDOperand> &N, bool V2IsSplat = false) {
1641 unsigned NumElems = N.size();
1642 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1645 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1646 SDOperand BitI = N[i];
1647 SDOperand BitI1 = N[i+1];
1648 if (!isUndefOrEqual(BitI, j + NumElems/2))
1651 if (isUndefOrEqual(BitI1, NumElems))
1654 if (!isUndefOrEqual(BitI1, j + NumElems/2 + NumElems))
1662 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
1663 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1664 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
1665 return ::isUNPCKHMask(Ops, V2IsSplat);
1668 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
1669 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
1671 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
1672 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1674 unsigned NumElems = N->getNumOperands();
1675 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
1678 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1679 SDOperand BitI = N->getOperand(i);
1680 SDOperand BitI1 = N->getOperand(i+1);
1682 if (!isUndefOrEqual(BitI, j))
1684 if (!isUndefOrEqual(BitI1, j))
1691 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
1692 /// specifies a shuffle of elements that is suitable for input to MOVSS,
1693 /// MOVSD, and MOVD, i.e. setting the lowest element.
1694 static bool isMOVLMask(std::vector<SDOperand> &N) {
1695 unsigned NumElems = N.size();
1696 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1699 if (!isUndefOrEqual(N[0], NumElems))
1702 for (unsigned i = 1; i < NumElems; ++i) {
1703 SDOperand Arg = N[i];
1704 if (!isUndefOrEqual(Arg, i))
1711 bool X86::isMOVLMask(SDNode *N) {
1712 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1713 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
1714 return ::isMOVLMask(Ops);
1717 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
1718 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
1719 /// element of vector 2 and the other elements to come from vector 1 in order.
1720 static bool isCommutedMOVL(std::vector<SDOperand> &Ops, bool V2IsSplat = false) {
1721 unsigned NumElems = Ops.size();
1722 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1725 if (!isUndefOrEqual(Ops[0], 0))
1728 for (unsigned i = 1; i < NumElems; ++i) {
1729 SDOperand Arg = Ops[i];
1731 if (!isUndefOrEqual(Arg, NumElems))
1734 if (!isUndefOrEqual(Arg, i+NumElems))
1742 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false) {
1743 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1744 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
1745 return isCommutedMOVL(Ops, V2IsSplat);
1748 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1749 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
1750 bool X86::isMOVSHDUPMask(SDNode *N) {
1751 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1753 if (N->getNumOperands() != 4)
1756 // Expect 1, 1, 3, 3
1757 for (unsigned i = 0; i < 2; ++i) {
1758 SDOperand Arg = N->getOperand(i);
1759 if (Arg.getOpcode() == ISD::UNDEF) continue;
1760 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1761 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1762 if (Val != 1) return false;
1766 for (unsigned i = 2; i < 4; ++i) {
1767 SDOperand Arg = N->getOperand(i);
1768 if (Arg.getOpcode() == ISD::UNDEF) continue;
1769 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1770 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1771 if (Val != 3) return false;
1775 // Don't use movshdup if it can be done with a shufps.
1779 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1780 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
1781 bool X86::isMOVSLDUPMask(SDNode *N) {
1782 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1784 if (N->getNumOperands() != 4)
1787 // Expect 0, 0, 2, 2
1788 for (unsigned i = 0; i < 2; ++i) {
1789 SDOperand Arg = N->getOperand(i);
1790 if (Arg.getOpcode() == ISD::UNDEF) continue;
1791 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1792 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1793 if (Val != 0) return false;
1797 for (unsigned i = 2; i < 4; ++i) {
1798 SDOperand Arg = N->getOperand(i);
1799 if (Arg.getOpcode() == ISD::UNDEF) continue;
1800 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1801 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1802 if (Val != 2) return false;
1806 // Don't use movshdup if it can be done with a shufps.
1810 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1811 /// a splat of a single element.
1812 static bool isSplatMask(SDNode *N) {
1813 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1815 // This is a splat operation if each element of the permute is the same, and
1816 // if the value doesn't reference the second vector.
1817 unsigned NumElems = N->getNumOperands();
1818 SDOperand ElementBase;
1820 for (; i != NumElems; ++i) {
1821 SDOperand Elt = N->getOperand(i);
1822 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt)) {
1828 if (!ElementBase.Val)
1831 for (; i != NumElems; ++i) {
1832 SDOperand Arg = N->getOperand(i);
1833 if (Arg.getOpcode() == ISD::UNDEF) continue;
1834 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1835 if (Arg != ElementBase) return false;
1838 // Make sure it is a splat of the first vector operand.
1839 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
1842 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1843 /// a splat of a single element and it's a 2 or 4 element mask.
1844 bool X86::isSplatMask(SDNode *N) {
1845 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1847 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
1848 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
1850 return ::isSplatMask(N);
1853 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
1854 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
1856 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
1857 unsigned NumOperands = N->getNumOperands();
1858 unsigned Shift = (NumOperands == 4) ? 2 : 1;
1860 for (unsigned i = 0; i < NumOperands; ++i) {
1862 SDOperand Arg = N->getOperand(NumOperands-i-1);
1863 if (Arg.getOpcode() != ISD::UNDEF)
1864 Val = cast<ConstantSDNode>(Arg)->getValue();
1865 if (Val >= NumOperands) Val -= NumOperands;
1867 if (i != NumOperands - 1)
1874 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
1875 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
1877 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
1879 // 8 nodes, but we only care about the last 4.
1880 for (unsigned i = 7; i >= 4; --i) {
1882 SDOperand Arg = N->getOperand(i);
1883 if (Arg.getOpcode() != ISD::UNDEF)
1884 Val = cast<ConstantSDNode>(Arg)->getValue();
1893 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
1894 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
1896 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
1898 // 8 nodes, but we only care about the first 4.
1899 for (int i = 3; i >= 0; --i) {
1901 SDOperand Arg = N->getOperand(i);
1902 if (Arg.getOpcode() != ISD::UNDEF)
1903 Val = cast<ConstantSDNode>(Arg)->getValue();
1912 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
1913 /// specifies a 8 element shuffle that can be broken into a pair of
1914 /// PSHUFHW and PSHUFLW.
1915 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
1916 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1918 if (N->getNumOperands() != 8)
1921 // Lower quadword shuffled.
1922 for (unsigned i = 0; i != 4; ++i) {
1923 SDOperand Arg = N->getOperand(i);
1924 if (Arg.getOpcode() == ISD::UNDEF) continue;
1925 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1926 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1931 // Upper quadword shuffled.
1932 for (unsigned i = 4; i != 8; ++i) {
1933 SDOperand Arg = N->getOperand(i);
1934 if (Arg.getOpcode() == ISD::UNDEF) continue;
1935 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1936 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1937 if (Val < 4 || Val > 7)
1944 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
1945 /// values in ther permute mask.
1946 static SDOperand CommuteVectorShuffle(SDOperand Op, SelectionDAG &DAG) {
1947 SDOperand V1 = Op.getOperand(0);
1948 SDOperand V2 = Op.getOperand(1);
1949 SDOperand Mask = Op.getOperand(2);
1950 MVT::ValueType VT = Op.getValueType();
1951 MVT::ValueType MaskVT = Mask.getValueType();
1952 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
1953 unsigned NumElems = Mask.getNumOperands();
1954 std::vector<SDOperand> MaskVec;
1956 for (unsigned i = 0; i != NumElems; ++i) {
1957 SDOperand Arg = Mask.getOperand(i);
1958 if (Arg.getOpcode() == ISD::UNDEF) {
1959 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
1962 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1963 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1965 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
1967 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
1970 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
1971 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V2, V1, Mask);
1974 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
1975 /// match movhlps. The lower half elements should come from upper half of
1976 /// V1 (and in order), and the upper half elements should come from the upper
1977 /// half of V2 (and in order).
1978 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
1979 unsigned NumElems = Mask->getNumOperands();
1982 for (unsigned i = 0, e = 2; i != e; ++i)
1983 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
1985 for (unsigned i = 2; i != 4; ++i)
1986 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
1991 /// isScalarLoadToVector - Returns true if the node is a scalar load that
1992 /// is promoted to a vector.
1993 static inline bool isScalarLoadToVector(SDNode *N) {
1994 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
1995 N = N->getOperand(0).Val;
1996 return (N->getOpcode() == ISD::LOAD);
2001 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2002 /// match movlp{s|d}. The lower half elements should come from lower half of
2003 /// V1 (and in order), and the upper half elements should come from the upper
2004 /// half of V2 (and in order). And since V1 will become the source of the
2005 /// MOVLP, it must be either a vector load or a scalar load to vector.
2006 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *Mask) {
2007 if (V1->getOpcode() != ISD::LOAD && !isScalarLoadToVector(V1))
2010 unsigned NumElems = Mask->getNumOperands();
2011 if (NumElems != 2 && NumElems != 4)
2013 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2014 if (!isUndefOrEqual(Mask->getOperand(i), i))
2016 for (unsigned i = NumElems/2; i != NumElems; ++i)
2017 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2022 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2024 static bool isSplatVector(SDNode *N) {
2025 if (N->getOpcode() != ISD::BUILD_VECTOR)
2028 SDOperand SplatValue = N->getOperand(0);
2029 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2030 if (N->getOperand(i) != SplatValue)
2035 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2036 /// that point to V2 points to its first element.
2037 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2038 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2040 bool Changed = false;
2041 std::vector<SDOperand> MaskVec;
2042 unsigned NumElems = Mask.getNumOperands();
2043 for (unsigned i = 0; i != NumElems; ++i) {
2044 SDOperand Arg = Mask.getOperand(i);
2045 if (Arg.getOpcode() != ISD::UNDEF) {
2046 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2047 if (Val > NumElems) {
2048 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2052 MaskVec.push_back(Arg);
2056 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(), MaskVec);
2060 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2061 /// operation of specified width.
2062 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2063 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2064 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2066 std::vector<SDOperand> MaskVec;
2067 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2068 for (unsigned i = 1; i != NumElems; ++i)
2069 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2070 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2073 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2074 /// of specified width.
2075 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2076 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2077 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2078 std::vector<SDOperand> MaskVec;
2079 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2080 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2081 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2083 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2086 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2087 /// of specified width.
2088 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2089 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2090 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2091 unsigned Half = NumElems/2;
2092 std::vector<SDOperand> MaskVec;
2093 for (unsigned i = 0; i != Half; ++i) {
2094 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2095 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2097 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2100 /// getZeroVector - Returns a vector of specified type with all zero elements.
2102 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2103 assert(MVT::isVector(VT) && "Expected a vector type");
2104 unsigned NumElems = getVectorNumElements(VT);
2105 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2106 bool isFP = MVT::isFloatingPoint(EVT);
2107 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2108 std::vector<SDOperand> ZeroVec(NumElems, Zero);
2109 return DAG.getNode(ISD::BUILD_VECTOR, VT, ZeroVec);
2112 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2114 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2115 SDOperand V1 = Op.getOperand(0);
2116 SDOperand Mask = Op.getOperand(2);
2117 MVT::ValueType VT = Op.getValueType();
2118 unsigned NumElems = Mask.getNumOperands();
2119 Mask = getUnpacklMask(NumElems, DAG);
2120 while (NumElems != 4) {
2121 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2124 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2126 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2127 Mask = getZeroVector(MaskVT, DAG);
2128 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2129 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2130 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2133 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2135 static inline bool isZeroNode(SDOperand Elt) {
2136 return ((isa<ConstantSDNode>(Elt) &&
2137 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2138 (isa<ConstantFPSDNode>(Elt) &&
2139 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2142 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2143 /// vector and zero or undef vector.
2144 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2145 unsigned NumElems, unsigned Idx,
2146 bool isZero, SelectionDAG &DAG) {
2147 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2148 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2149 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2150 SDOperand Zero = DAG.getConstant(0, EVT);
2151 std::vector<SDOperand> MaskVec(NumElems, Zero);
2152 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2153 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2154 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2157 /// LowerOperation - Provide custom lowering hooks for some operations.
2159 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
2160 switch (Op.getOpcode()) {
2161 default: assert(0 && "Should not custom lower this!");
2162 case ISD::SHL_PARTS:
2163 case ISD::SRA_PARTS:
2164 case ISD::SRL_PARTS: {
2165 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2166 "Not an i64 shift!");
2167 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
2168 SDOperand ShOpLo = Op.getOperand(0);
2169 SDOperand ShOpHi = Op.getOperand(1);
2170 SDOperand ShAmt = Op.getOperand(2);
2171 SDOperand Tmp1 = isSRA ? DAG.getNode(ISD::SRA, MVT::i32, ShOpHi,
2172 DAG.getConstant(31, MVT::i8))
2173 : DAG.getConstant(0, MVT::i32);
2175 SDOperand Tmp2, Tmp3;
2176 if (Op.getOpcode() == ISD::SHL_PARTS) {
2177 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
2178 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
2180 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
2181 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
2184 SDOperand InFlag = DAG.getNode(X86ISD::TEST, MVT::Flag,
2185 ShAmt, DAG.getConstant(32, MVT::i8));
2188 SDOperand CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
2190 std::vector<MVT::ValueType> Tys;
2191 Tys.push_back(MVT::i32);
2192 Tys.push_back(MVT::Flag);
2193 std::vector<SDOperand> Ops;
2194 if (Op.getOpcode() == ISD::SHL_PARTS) {
2195 Ops.push_back(Tmp2);
2196 Ops.push_back(Tmp3);
2198 Ops.push_back(InFlag);
2199 Hi = DAG.getNode(X86ISD::CMOV, Tys, Ops);
2200 InFlag = Hi.getValue(1);
2203 Ops.push_back(Tmp3);
2204 Ops.push_back(Tmp1);
2206 Ops.push_back(InFlag);
2207 Lo = DAG.getNode(X86ISD::CMOV, Tys, Ops);
2209 Ops.push_back(Tmp2);
2210 Ops.push_back(Tmp3);
2212 Ops.push_back(InFlag);
2213 Lo = DAG.getNode(X86ISD::CMOV, Tys, Ops);
2214 InFlag = Lo.getValue(1);
2217 Ops.push_back(Tmp3);
2218 Ops.push_back(Tmp1);
2220 Ops.push_back(InFlag);
2221 Hi = DAG.getNode(X86ISD::CMOV, Tys, Ops);
2225 Tys.push_back(MVT::i32);
2226 Tys.push_back(MVT::i32);
2230 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops);
2232 case ISD::SINT_TO_FP: {
2233 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
2234 Op.getOperand(0).getValueType() >= MVT::i16 &&
2235 "Unknown SINT_TO_FP to lower!");
2238 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
2239 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
2240 MachineFunction &MF = DAG.getMachineFunction();
2241 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
2242 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
2243 SDOperand Chain = DAG.getNode(ISD::STORE, MVT::Other,
2244 DAG.getEntryNode(), Op.getOperand(0),
2245 StackSlot, DAG.getSrcValue(NULL));
2248 std::vector<MVT::ValueType> Tys;
2249 Tys.push_back(MVT::f64);
2250 Tys.push_back(MVT::Other);
2251 if (X86ScalarSSE) Tys.push_back(MVT::Flag);
2252 std::vector<SDOperand> Ops;
2253 Ops.push_back(Chain);
2254 Ops.push_back(StackSlot);
2255 Ops.push_back(DAG.getValueType(SrcVT));
2256 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
2260 Chain = Result.getValue(1);
2261 SDOperand InFlag = Result.getValue(2);
2263 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
2264 // shouldn't be necessary except that RFP cannot be live across
2265 // multiple blocks. When stackifier is fixed, they can be uncoupled.
2266 MachineFunction &MF = DAG.getMachineFunction();
2267 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
2268 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
2269 std::vector<MVT::ValueType> Tys;
2270 Tys.push_back(MVT::Other);
2271 std::vector<SDOperand> Ops;
2272 Ops.push_back(Chain);
2273 Ops.push_back(Result);
2274 Ops.push_back(StackSlot);
2275 Ops.push_back(DAG.getValueType(Op.getValueType()));
2276 Ops.push_back(InFlag);
2277 Chain = DAG.getNode(X86ISD::FST, Tys, Ops);
2278 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
2279 DAG.getSrcValue(NULL));
2284 case ISD::FP_TO_SINT: {
2285 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
2286 "Unknown FP_TO_SINT to lower!");
2287 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
2289 MachineFunction &MF = DAG.getMachineFunction();
2290 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
2291 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
2292 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
2295 switch (Op.getValueType()) {
2296 default: assert(0 && "Invalid FP_TO_SINT to lower!");
2297 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
2298 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
2299 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
2302 SDOperand Chain = DAG.getEntryNode();
2303 SDOperand Value = Op.getOperand(0);
2305 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
2306 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value, StackSlot,
2307 DAG.getSrcValue(0));
2308 std::vector<MVT::ValueType> Tys;
2309 Tys.push_back(MVT::f64);
2310 Tys.push_back(MVT::Other);
2311 std::vector<SDOperand> Ops;
2312 Ops.push_back(Chain);
2313 Ops.push_back(StackSlot);
2314 Ops.push_back(DAG.getValueType(Op.getOperand(0).getValueType()));
2315 Value = DAG.getNode(X86ISD::FLD, Tys, Ops);
2316 Chain = Value.getValue(1);
2317 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
2318 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
2321 // Build the FP_TO_INT*_IN_MEM
2322 std::vector<SDOperand> Ops;
2323 Ops.push_back(Chain);
2324 Ops.push_back(Value);
2325 Ops.push_back(StackSlot);
2326 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops);
2329 return DAG.getLoad(Op.getValueType(), FIST, StackSlot,
2330 DAG.getSrcValue(NULL));
2332 case ISD::READCYCLECOUNTER: {
2333 std::vector<MVT::ValueType> Tys;
2334 Tys.push_back(MVT::Other);
2335 Tys.push_back(MVT::Flag);
2336 std::vector<SDOperand> Ops;
2337 Ops.push_back(Op.getOperand(0));
2338 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, Ops);
2340 Ops.push_back(DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1)));
2341 Ops.push_back(DAG.getCopyFromReg(Ops[0].getValue(1), X86::EDX,
2342 MVT::i32, Ops[0].getValue(2)));
2343 Ops.push_back(Ops[1].getValue(1));
2344 Tys[0] = Tys[1] = MVT::i32;
2345 Tys.push_back(MVT::Other);
2346 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops);
2349 MVT::ValueType VT = Op.getValueType();
2350 const Type *OpNTy = MVT::getTypeForValueType(VT);
2351 std::vector<Constant*> CV;
2352 if (VT == MVT::f64) {
2353 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
2354 CV.push_back(ConstantFP::get(OpNTy, 0.0));
2356 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
2357 CV.push_back(ConstantFP::get(OpNTy, 0.0));
2358 CV.push_back(ConstantFP::get(OpNTy, 0.0));
2359 CV.push_back(ConstantFP::get(OpNTy, 0.0));
2361 Constant *CS = ConstantStruct::get(CV);
2362 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
2364 = DAG.getNode(X86ISD::LOAD_PACK,
2365 VT, DAG.getEntryNode(), CPIdx, DAG.getSrcValue(NULL));
2366 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
2369 MVT::ValueType VT = Op.getValueType();
2370 const Type *OpNTy = MVT::getTypeForValueType(VT);
2371 std::vector<Constant*> CV;
2372 if (VT == MVT::f64) {
2373 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
2374 CV.push_back(ConstantFP::get(OpNTy, 0.0));
2376 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
2377 CV.push_back(ConstantFP::get(OpNTy, 0.0));
2378 CV.push_back(ConstantFP::get(OpNTy, 0.0));
2379 CV.push_back(ConstantFP::get(OpNTy, 0.0));
2381 Constant *CS = ConstantStruct::get(CV);
2382 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
2384 = DAG.getNode(X86ISD::LOAD_PACK,
2385 VT, DAG.getEntryNode(), CPIdx, DAG.getSrcValue(NULL));
2386 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
2389 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
2391 SDOperand CC = Op.getOperand(2);
2392 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2393 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
2396 if (translateX86CC(CC, isFP, X86CC, Flip)) {
2398 Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
2399 Op.getOperand(1), Op.getOperand(0));
2401 Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
2402 Op.getOperand(0), Op.getOperand(1));
2403 return DAG.getNode(X86ISD::SETCC, MVT::i8,
2404 DAG.getConstant(X86CC, MVT::i8), Cond);
2406 assert(isFP && "Illegal integer SetCC!");
2408 Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
2409 Op.getOperand(0), Op.getOperand(1));
2410 std::vector<MVT::ValueType> Tys;
2411 std::vector<SDOperand> Ops;
2412 switch (SetCCOpcode) {
2413 default: assert(false && "Illegal floating point SetCC!");
2414 case ISD::SETOEQ: { // !PF & ZF
2415 Tys.push_back(MVT::i8);
2416 Tys.push_back(MVT::Flag);
2417 Ops.push_back(DAG.getConstant(X86ISD::COND_NP, MVT::i8));
2418 Ops.push_back(Cond);
2419 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
2420 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
2421 DAG.getConstant(X86ISD::COND_E, MVT::i8),
2423 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
2425 case ISD::SETUNE: { // PF | !ZF
2426 Tys.push_back(MVT::i8);
2427 Tys.push_back(MVT::Flag);
2428 Ops.push_back(DAG.getConstant(X86ISD::COND_P, MVT::i8));
2429 Ops.push_back(Cond);
2430 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
2431 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
2432 DAG.getConstant(X86ISD::COND_NE, MVT::i8),
2434 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
2440 MVT::ValueType VT = Op.getValueType();
2441 bool isFPStack = MVT::isFloatingPoint(VT) && !X86ScalarSSE;
2442 bool addTest = false;
2443 SDOperand Op0 = Op.getOperand(0);
2445 if (Op0.getOpcode() == ISD::SETCC)
2446 Op0 = LowerOperation(Op0, DAG);
2448 if (Op0.getOpcode() == X86ISD::SETCC) {
2449 // If condition flag is set by a X86ISD::CMP, then make a copy of it
2450 // (since flag operand cannot be shared). If the X86ISD::SETCC does not
2451 // have another use it will be eliminated.
2452 // If the X86ISD::SETCC has more than one use, then it's probably better
2453 // to use a test instead of duplicating the X86ISD::CMP (for register
2454 // pressure reason).
2455 unsigned CmpOpc = Op0.getOperand(1).getOpcode();
2456 if (CmpOpc == X86ISD::CMP || CmpOpc == X86ISD::COMI ||
2457 CmpOpc == X86ISD::UCOMI) {
2458 if (!Op0.hasOneUse()) {
2459 std::vector<MVT::ValueType> Tys;
2460 for (unsigned i = 0; i < Op0.Val->getNumValues(); ++i)
2461 Tys.push_back(Op0.Val->getValueType(i));
2462 std::vector<SDOperand> Ops;
2463 for (unsigned i = 0; i < Op0.getNumOperands(); ++i)
2464 Ops.push_back(Op0.getOperand(i));
2465 Op0 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
2468 CC = Op0.getOperand(0);
2469 Cond = Op0.getOperand(1);
2470 // Make a copy as flag result cannot be used by more than one.
2471 Cond = DAG.getNode(CmpOpc, MVT::Flag,
2472 Cond.getOperand(0), Cond.getOperand(1));
2474 isFPStack && !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
2481 CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
2482 Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Op0, Op0);
2485 std::vector<MVT::ValueType> Tys;
2486 Tys.push_back(Op.getValueType());
2487 Tys.push_back(MVT::Flag);
2488 std::vector<SDOperand> Ops;
2489 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
2490 // condition is true.
2491 Ops.push_back(Op.getOperand(2));
2492 Ops.push_back(Op.getOperand(1));
2494 Ops.push_back(Cond);
2495 return DAG.getNode(X86ISD::CMOV, Tys, Ops);
2498 bool addTest = false;
2499 SDOperand Cond = Op.getOperand(1);
2500 SDOperand Dest = Op.getOperand(2);
2502 if (Cond.getOpcode() == ISD::SETCC)
2503 Cond = LowerOperation(Cond, DAG);
2505 if (Cond.getOpcode() == X86ISD::SETCC) {
2506 // If condition flag is set by a X86ISD::CMP, then make a copy of it
2507 // (since flag operand cannot be shared). If the X86ISD::SETCC does not
2508 // have another use it will be eliminated.
2509 // If the X86ISD::SETCC has more than one use, then it's probably better
2510 // to use a test instead of duplicating the X86ISD::CMP (for register
2511 // pressure reason).
2512 unsigned CmpOpc = Cond.getOperand(1).getOpcode();
2513 if (CmpOpc == X86ISD::CMP || CmpOpc == X86ISD::COMI ||
2514 CmpOpc == X86ISD::UCOMI) {
2515 if (!Cond.hasOneUse()) {
2516 std::vector<MVT::ValueType> Tys;
2517 for (unsigned i = 0; i < Cond.Val->getNumValues(); ++i)
2518 Tys.push_back(Cond.Val->getValueType(i));
2519 std::vector<SDOperand> Ops;
2520 for (unsigned i = 0; i < Cond.getNumOperands(); ++i)
2521 Ops.push_back(Cond.getOperand(i));
2522 Cond = DAG.getNode(X86ISD::SETCC, Tys, Ops);
2525 CC = Cond.getOperand(0);
2526 Cond = Cond.getOperand(1);
2527 // Make a copy as flag result cannot be used by more than one.
2528 Cond = DAG.getNode(CmpOpc, MVT::Flag,
2529 Cond.getOperand(0), Cond.getOperand(1));
2536 CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
2537 Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Cond, Cond);
2539 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
2540 Op.getOperand(0), Op.getOperand(2), CC, Cond);
2543 SDOperand InFlag(0, 0);
2544 SDOperand Chain = Op.getOperand(0);
2546 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
2547 if (Align == 0) Align = 1;
2549 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
2550 // If not DWORD aligned, call memset if size is less than the threshold.
2551 // It knows how to align to the right boundary first.
2552 if ((Align & 3) != 0 ||
2553 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
2554 MVT::ValueType IntPtr = getPointerTy();
2555 const Type *IntPtrTy = getTargetData().getIntPtrType();
2556 std::vector<std::pair<SDOperand, const Type*> > Args;
2557 Args.push_back(std::make_pair(Op.getOperand(1), IntPtrTy));
2558 // Extend the ubyte argument to be an int value for the call.
2559 SDOperand Val = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
2560 Args.push_back(std::make_pair(Val, IntPtrTy));
2561 Args.push_back(std::make_pair(Op.getOperand(3), IntPtrTy));
2562 std::pair<SDOperand,SDOperand> CallResult =
2563 LowerCallTo(Chain, Type::VoidTy, false, CallingConv::C, false,
2564 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
2565 return CallResult.second;
2570 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2571 unsigned BytesLeft = 0;
2572 bool TwoRepStos = false;
2575 unsigned Val = ValC->getValue() & 255;
2577 // If the value is a constant, then we can potentially use larger sets.
2578 switch (Align & 3) {
2579 case 2: // WORD aligned
2581 Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
2582 BytesLeft = I->getValue() % 2;
2583 Val = (Val << 8) | Val;
2586 case 0: // DWORD aligned
2589 Count = DAG.getConstant(I->getValue() / 4, MVT::i32);
2590 BytesLeft = I->getValue() % 4;
2592 Count = DAG.getNode(ISD::SRL, MVT::i32, Op.getOperand(3),
2593 DAG.getConstant(2, MVT::i8));
2596 Val = (Val << 8) | Val;
2597 Val = (Val << 16) | Val;
2600 default: // Byte aligned
2602 Count = Op.getOperand(3);
2607 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
2609 InFlag = Chain.getValue(1);
2612 Count = Op.getOperand(3);
2613 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
2614 InFlag = Chain.getValue(1);
2617 Chain = DAG.getCopyToReg(Chain, X86::ECX, Count, InFlag);
2618 InFlag = Chain.getValue(1);
2619 Chain = DAG.getCopyToReg(Chain, X86::EDI, Op.getOperand(1), InFlag);
2620 InFlag = Chain.getValue(1);
2622 std::vector<MVT::ValueType> Tys;
2623 Tys.push_back(MVT::Other);
2624 Tys.push_back(MVT::Flag);
2625 std::vector<SDOperand> Ops;
2626 Ops.push_back(Chain);
2627 Ops.push_back(DAG.getValueType(AVT));
2628 Ops.push_back(InFlag);
2629 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, Ops);
2632 InFlag = Chain.getValue(1);
2633 Count = Op.getOperand(3);
2634 MVT::ValueType CVT = Count.getValueType();
2635 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
2636 DAG.getConstant(3, CVT));
2637 Chain = DAG.getCopyToReg(Chain, X86::ECX, Left, InFlag);
2638 InFlag = Chain.getValue(1);
2640 Tys.push_back(MVT::Other);
2641 Tys.push_back(MVT::Flag);
2643 Ops.push_back(Chain);
2644 Ops.push_back(DAG.getValueType(MVT::i8));
2645 Ops.push_back(InFlag);
2646 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, Ops);
2647 } else if (BytesLeft) {
2648 // Issue stores for the last 1 - 3 bytes.
2650 unsigned Val = ValC->getValue() & 255;
2651 unsigned Offset = I->getValue() - BytesLeft;
2652 SDOperand DstAddr = Op.getOperand(1);
2653 MVT::ValueType AddrVT = DstAddr.getValueType();
2654 if (BytesLeft >= 2) {
2655 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
2656 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
2657 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
2658 DAG.getConstant(Offset, AddrVT)),
2659 DAG.getSrcValue(NULL));
2664 if (BytesLeft == 1) {
2665 Value = DAG.getConstant(Val, MVT::i8);
2666 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
2667 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
2668 DAG.getConstant(Offset, AddrVT)),
2669 DAG.getSrcValue(NULL));
2676 SDOperand Chain = Op.getOperand(0);
2678 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
2679 if (Align == 0) Align = 1;
2681 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
2682 // If not DWORD aligned, call memcpy if size is less than the threshold.
2683 // It knows how to align to the right boundary first.
2684 if ((Align & 3) != 0 ||
2685 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
2686 MVT::ValueType IntPtr = getPointerTy();
2687 const Type *IntPtrTy = getTargetData().getIntPtrType();
2688 std::vector<std::pair<SDOperand, const Type*> > Args;
2689 Args.push_back(std::make_pair(Op.getOperand(1), IntPtrTy));
2690 Args.push_back(std::make_pair(Op.getOperand(2), IntPtrTy));
2691 Args.push_back(std::make_pair(Op.getOperand(3), IntPtrTy));
2692 std::pair<SDOperand,SDOperand> CallResult =
2693 LowerCallTo(Chain, Type::VoidTy, false, CallingConv::C, false,
2694 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
2695 return CallResult.second;
2700 unsigned BytesLeft = 0;
2701 bool TwoRepMovs = false;
2702 switch (Align & 3) {
2703 case 2: // WORD aligned
2705 Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
2706 BytesLeft = I->getValue() % 2;
2708 case 0: // DWORD aligned
2711 Count = DAG.getConstant(I->getValue() / 4, MVT::i32);
2712 BytesLeft = I->getValue() % 4;
2714 Count = DAG.getNode(ISD::SRL, MVT::i32, Op.getOperand(3),
2715 DAG.getConstant(2, MVT::i8));
2719 default: // Byte aligned
2721 Count = Op.getOperand(3);
2725 SDOperand InFlag(0, 0);
2726 Chain = DAG.getCopyToReg(Chain, X86::ECX, Count, InFlag);
2727 InFlag = Chain.getValue(1);
2728 Chain = DAG.getCopyToReg(Chain, X86::EDI, Op.getOperand(1), InFlag);
2729 InFlag = Chain.getValue(1);
2730 Chain = DAG.getCopyToReg(Chain, X86::ESI, Op.getOperand(2), InFlag);
2731 InFlag = Chain.getValue(1);
2733 std::vector<MVT::ValueType> Tys;
2734 Tys.push_back(MVT::Other);
2735 Tys.push_back(MVT::Flag);
2736 std::vector<SDOperand> Ops;
2737 Ops.push_back(Chain);
2738 Ops.push_back(DAG.getValueType(AVT));
2739 Ops.push_back(InFlag);
2740 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, Ops);
2743 InFlag = Chain.getValue(1);
2744 Count = Op.getOperand(3);
2745 MVT::ValueType CVT = Count.getValueType();
2746 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
2747 DAG.getConstant(3, CVT));
2748 Chain = DAG.getCopyToReg(Chain, X86::ECX, Left, InFlag);
2749 InFlag = Chain.getValue(1);
2751 Tys.push_back(MVT::Other);
2752 Tys.push_back(MVT::Flag);
2754 Ops.push_back(Chain);
2755 Ops.push_back(DAG.getValueType(MVT::i8));
2756 Ops.push_back(InFlag);
2757 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, Ops);
2758 } else if (BytesLeft) {
2759 // Issue loads and stores for the last 1 - 3 bytes.
2760 unsigned Offset = I->getValue() - BytesLeft;
2761 SDOperand DstAddr = Op.getOperand(1);
2762 MVT::ValueType DstVT = DstAddr.getValueType();
2763 SDOperand SrcAddr = Op.getOperand(2);
2764 MVT::ValueType SrcVT = SrcAddr.getValueType();
2766 if (BytesLeft >= 2) {
2767 Value = DAG.getLoad(MVT::i16, Chain,
2768 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
2769 DAG.getConstant(Offset, SrcVT)),
2770 DAG.getSrcValue(NULL));
2771 Chain = Value.getValue(1);
2772 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
2773 DAG.getNode(ISD::ADD, DstVT, DstAddr,
2774 DAG.getConstant(Offset, DstVT)),
2775 DAG.getSrcValue(NULL));
2780 if (BytesLeft == 1) {
2781 Value = DAG.getLoad(MVT::i8, Chain,
2782 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
2783 DAG.getConstant(Offset, SrcVT)),
2784 DAG.getSrcValue(NULL));
2785 Chain = Value.getValue(1);
2786 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
2787 DAG.getNode(ISD::ADD, DstVT, DstAddr,
2788 DAG.getConstant(Offset, DstVT)),
2789 DAG.getSrcValue(NULL));
2796 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2797 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
2798 // one of the above mentioned nodes. It has to be wrapped because otherwise
2799 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2800 // be used to form addressing mode. These wrapped nodes will be selected
2802 case ISD::ConstantPool: {
2803 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2804 SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
2805 DAG.getTargetConstantPool(CP->get(), getPointerTy(),
2806 CP->getAlignment()));
2807 if (Subtarget->isTargetDarwin()) {
2808 // With PIC, the address is actually $g + Offset.
2809 if (getTargetMachine().getRelocationModel() == Reloc::PIC)
2810 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2811 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), Result);
2816 case ISD::JumpTable: {
2817 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
2818 SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
2819 DAG.getTargetJumpTable(JT->getIndex(),
2821 if (Subtarget->isTargetDarwin()) {
2822 // With PIC, the address is actually $g + Offset.
2823 if (getTargetMachine().getRelocationModel() == Reloc::PIC)
2824 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2825 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), Result);
2830 case ISD::GlobalAddress: {
2831 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2832 SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
2833 DAG.getTargetGlobalAddress(GV, getPointerTy()));
2834 if (Subtarget->isTargetDarwin()) {
2835 // With PIC, the address is actually $g + Offset.
2836 if (getTargetMachine().getRelocationModel() == Reloc::PIC)
2837 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2838 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), Result);
2840 // For Darwin, external and weak symbols are indirect, so we want to load
2841 // the value at address GV, not the value of GV itself. This means that
2842 // the GlobalAddress must be in the base or index register of the address,
2843 // not the GV offset field.
2844 if (getTargetMachine().getRelocationModel() != Reloc::Static &&
2845 DarwinGVRequiresExtraLoad(GV))
2846 Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(),
2847 Result, DAG.getSrcValue(NULL));
2852 case ISD::ExternalSymbol: {
2853 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
2854 SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
2855 DAG.getTargetExternalSymbol(Sym, getPointerTy()));
2856 if (Subtarget->isTargetDarwin()) {
2857 // With PIC, the address is actually $g + Offset.
2858 if (getTargetMachine().getRelocationModel() == Reloc::PIC)
2859 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2860 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), Result);
2865 case ISD::VASTART: {
2866 // vastart just stores the address of the VarArgsFrameIndex slot into the
2867 // memory location argument.
2868 // FIXME: Replace MVT::i32 with PointerTy
2869 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
2870 return DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0), FR,
2871 Op.getOperand(1), Op.getOperand(2));
2876 switch(Op.getNumOperands()) {
2878 assert(0 && "Do not know how to return this many arguments!");
2880 case 1: // ret void.
2881 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Op.getOperand(0),
2882 DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
2884 MVT::ValueType ArgVT = Op.getOperand(1).getValueType();
2886 if (MVT::isVector(ArgVT)) {
2887 // Integer or FP vector result -> XMM0.
2888 if (DAG.getMachineFunction().liveout_empty())
2889 DAG.getMachineFunction().addLiveOut(X86::XMM0);
2890 Copy = DAG.getCopyToReg(Op.getOperand(0), X86::XMM0, Op.getOperand(1),
2892 } else if (MVT::isInteger(ArgVT)) {
2893 // Integer result -> EAX
2894 if (DAG.getMachineFunction().liveout_empty())
2895 DAG.getMachineFunction().addLiveOut(X86::EAX);
2897 Copy = DAG.getCopyToReg(Op.getOperand(0), X86::EAX, Op.getOperand(1),
2899 } else if (!X86ScalarSSE) {
2900 // FP return with fp-stack value.
2901 if (DAG.getMachineFunction().liveout_empty())
2902 DAG.getMachineFunction().addLiveOut(X86::ST0);
2904 std::vector<MVT::ValueType> Tys;
2905 Tys.push_back(MVT::Other);
2906 Tys.push_back(MVT::Flag);
2907 std::vector<SDOperand> Ops;
2908 Ops.push_back(Op.getOperand(0));
2909 Ops.push_back(Op.getOperand(1));
2910 Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops);
2912 // FP return with ScalarSSE (return on fp-stack).
2913 if (DAG.getMachineFunction().liveout_empty())
2914 DAG.getMachineFunction().addLiveOut(X86::ST0);
2917 SDOperand Chain = Op.getOperand(0);
2918 SDOperand Value = Op.getOperand(1);
2920 if (Value.getOpcode() == ISD::LOAD &&
2921 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
2922 Chain = Value.getOperand(0);
2923 MemLoc = Value.getOperand(1);
2925 // Spill the value to memory and reload it into top of stack.
2926 unsigned Size = MVT::getSizeInBits(ArgVT)/8;
2927 MachineFunction &MF = DAG.getMachineFunction();
2928 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
2929 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
2930 Chain = DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0),
2931 Value, MemLoc, DAG.getSrcValue(0));
2933 std::vector<MVT::ValueType> Tys;
2934 Tys.push_back(MVT::f64);
2935 Tys.push_back(MVT::Other);
2936 std::vector<SDOperand> Ops;
2937 Ops.push_back(Chain);
2938 Ops.push_back(MemLoc);
2939 Ops.push_back(DAG.getValueType(ArgVT));
2940 Copy = DAG.getNode(X86ISD::FLD, Tys, Ops);
2942 Tys.push_back(MVT::Other);
2943 Tys.push_back(MVT::Flag);
2945 Ops.push_back(Copy.getValue(1));
2946 Ops.push_back(Copy);
2947 Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops);
2952 if (DAG.getMachineFunction().liveout_empty()) {
2953 DAG.getMachineFunction().addLiveOut(X86::EAX);
2954 DAG.getMachineFunction().addLiveOut(X86::EDX);
2957 Copy = DAG.getCopyToReg(Op.getOperand(0), X86::EDX, Op.getOperand(2),
2959 Copy = DAG.getCopyToReg(Copy, X86::EAX,Op.getOperand(1),Copy.getValue(1));
2962 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other,
2963 Copy, DAG.getConstant(getBytesToPopOnReturn(), MVT::i16),
2966 case ISD::SCALAR_TO_VECTOR: {
2967 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
2968 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
2970 case ISD::VECTOR_SHUFFLE: {
2971 SDOperand V1 = Op.getOperand(0);
2972 SDOperand V2 = Op.getOperand(1);
2973 SDOperand PermMask = Op.getOperand(2);
2974 MVT::ValueType VT = Op.getValueType();
2975 unsigned NumElems = PermMask.getNumOperands();
2977 if (isSplatMask(PermMask.Val)) {
2978 if (NumElems <= 4) return Op;
2979 // Promote it to a v4i32 splat.
2980 return PromoteSplat(Op, DAG);
2983 if (X86::isMOVLMask(PermMask.Val) ||
2984 X86::isMOVSHDUPMask(PermMask.Val) ||
2985 X86::isMOVSLDUPMask(PermMask.Val) ||
2986 X86::isMOVHLPSMask(PermMask.Val) ||
2987 X86::isMOVHPMask(PermMask.Val) ||
2988 X86::isMOVLPMask(PermMask.Val))
2991 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2992 ShouldXformToMOVLP(V1.Val, PermMask.Val))
2993 return CommuteVectorShuffle(Op, DAG);
2995 bool V1IsSplat = isSplatVector(V1.Val) || V1.getOpcode() == ISD::UNDEF;
2996 bool V2IsSplat = isSplatVector(V2.Val) || V2.getOpcode() == ISD::UNDEF;
2997 if (V1IsSplat && !V2IsSplat) {
2998 Op = CommuteVectorShuffle(Op, DAG);
2999 V1 = Op.getOperand(0);
3000 V2 = Op.getOperand(1);
3001 PermMask = Op.getOperand(2);
3005 if (isCommutedMOVL(PermMask.Val, V2IsSplat)) {
3006 Op = CommuteVectorShuffle(Op, DAG);
3007 V1 = Op.getOperand(0);
3008 V2 = Op.getOperand(1);
3009 PermMask = Op.getOperand(2);
3011 // V2 is a splat, so the mask may be malformed. That is, it may point
3012 // to any V2 element. The instruction selectior won't like this. Get
3013 // a corrected mask and commute to form a proper MOVS{S|D}.
3014 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3015 if (NewMask.Val != PermMask.Val)
3016 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3021 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3022 X86::isUNPCKLMask(PermMask.Val) ||
3023 X86::isUNPCKHMask(PermMask.Val))
3027 // Normalize mask so all entries that point to V2 points to its first
3028 // element then try to match unpck{h|l} again. If match, return a
3029 // new vector_shuffle with the corrected mask.
3030 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3031 if (NewMask.Val != PermMask.Val) {
3032 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3033 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3034 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3035 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3036 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3037 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3042 // Normalize the node to match x86 shuffle ops if needed
3043 if (V2.getOpcode() != ISD::UNDEF)
3044 if (isCommutedSHUFP(PermMask.Val)) {
3045 Op = CommuteVectorShuffle(Op, DAG);
3046 V1 = Op.getOperand(0);
3047 V2 = Op.getOperand(1);
3048 PermMask = Op.getOperand(2);
3051 // If VT is integer, try PSHUF* first, then SHUFP*.
3052 if (MVT::isInteger(VT)) {
3053 if (X86::isPSHUFDMask(PermMask.Val) ||
3054 X86::isPSHUFHWMask(PermMask.Val) ||
3055 X86::isPSHUFLWMask(PermMask.Val)) {
3056 if (V2.getOpcode() != ISD::UNDEF)
3057 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3058 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3062 if (X86::isSHUFPMask(PermMask.Val))
3065 // Handle v8i16 shuffle high / low shuffle node pair.
3066 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3067 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3068 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3069 std::vector<SDOperand> MaskVec;
3070 for (unsigned i = 0; i != 4; ++i)
3071 MaskVec.push_back(PermMask.getOperand(i));
3072 for (unsigned i = 4; i != 8; ++i)
3073 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3074 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
3075 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3077 for (unsigned i = 0; i != 4; ++i)
3078 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3079 for (unsigned i = 4; i != 8; ++i)
3080 MaskVec.push_back(PermMask.getOperand(i));
3081 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
3082 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3085 // Floating point cases in the other order.
3086 if (X86::isSHUFPMask(PermMask.Val))
3088 if (X86::isPSHUFDMask(PermMask.Val) ||
3089 X86::isPSHUFHWMask(PermMask.Val) ||
3090 X86::isPSHUFLWMask(PermMask.Val)) {
3091 if (V2.getOpcode() != ISD::UNDEF)
3092 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3093 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3098 if (NumElems == 4) {
3099 // Break it into (shuffle shuffle_hi, shuffle_lo).
3100 MVT::ValueType MaskVT = PermMask.getValueType();
3101 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
3102 std::map<unsigned, std::pair<int, int> > Locs;
3103 std::vector<SDOperand> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3104 std::vector<SDOperand> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3105 std::vector<SDOperand> *MaskPtr = &LoMask;
3106 unsigned MaskIdx = 0;
3108 unsigned HiIdx = NumElems/2;
3109 for (unsigned i = 0; i != NumElems; ++i) {
3110 if (i == NumElems/2) {
3116 SDOperand Elt = PermMask.getOperand(i);
3117 if (Elt.getOpcode() == ISD::UNDEF) {
3118 Locs[i] = std::make_pair(-1, -1);
3119 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3120 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3121 (*MaskPtr)[LoIdx] = Elt;
3124 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3125 (*MaskPtr)[HiIdx] = Elt;
3130 SDOperand LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3131 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, LoMask));
3132 SDOperand HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3133 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, HiMask));
3134 std::vector<SDOperand> MaskOps;
3135 for (unsigned i = 0; i != NumElems; ++i) {
3136 if (Locs[i].first == -1) {
3137 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3139 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3140 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3143 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3144 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskOps));
3149 case ISD::BUILD_VECTOR: {
3150 // All one's are handled with pcmpeqd.
3151 if (ISD::isBuildVectorAllOnes(Op.Val))
3154 unsigned NumElems = Op.getNumOperands();
3158 unsigned Half = NumElems/2;
3159 MVT::ValueType VT = Op.getValueType();
3160 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
3161 unsigned NumZero = 0;
3162 unsigned NonZeros = 0;
3163 std::set<SDOperand> Values;
3164 for (unsigned i = 0; i < NumElems; ++i) {
3165 SDOperand Elt = Op.getOperand(i);
3167 if (isZeroNode(Elt))
3169 else if (Elt.getOpcode() != ISD::UNDEF)
3170 NonZeros |= (1 << i);
3173 unsigned NumNonZero = CountPopulation_32(NonZeros);
3174 if (NumNonZero == 0)
3177 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3178 if (Values.size() == 1)
3181 // If element VT is >= 32 bits, turn it into a number of shuffles.
3182 if (NumNonZero == 1) {
3183 unsigned Idx = CountTrailingZeros_32(NonZeros);
3184 SDOperand Item = Op.getOperand(Idx);
3185 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3187 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3188 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
3191 if (MVT::getSizeInBits(EVT) >= 32) {
3192 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3193 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
3195 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3196 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
3197 std::vector<SDOperand> MaskVec;
3198 for (unsigned i = 0; i < NumElems; i++)
3199 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3200 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
3201 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3202 DAG.getNode(ISD::UNDEF, VT), Mask);
3206 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3207 if (MVT::getSizeInBits(EVT) <= 16) {
3208 if (NumNonZero <= Half) {
3211 for (unsigned i = 0; i < NumNonZero; ++i) {
3212 unsigned Idx = CountTrailingZeros_32(NonZeros);
3213 NonZeros ^= (1 << Idx);
3214 SDOperand Item = Op.getOperand(Idx);
3217 V = getZeroVector(MVT::v8i16, DAG);
3219 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3221 if (EVT == MVT::i8) {
3222 Item = DAG.getNode(ISD::ANY_EXTEND, MVT::i16, Item);
3224 Item = DAG.getNode(ISD::SHL, MVT::i16,
3225 Item, DAG.getConstant(8, MVT::i8));
3228 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Item,
3229 DAG.getConstant(Idx, MVT::i32));
3233 V = DAG.getNode(ISD::BIT_CONVERT, VT, V);
3238 std::vector<SDOperand> V(NumElems);
3239 if (NumElems == 4 && NumZero > 0) {
3240 for (unsigned i = 0; i < 4; ++i) {
3241 bool isZero = !(NonZeros & (1 << i));
3243 V[i] = getZeroVector(VT, DAG);
3245 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3248 for (unsigned i = 0; i < 2; ++i) {
3249 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3252 V[i] = V[i*2]; // Must be a zero vector.
3255 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3256 getMOVLMask(NumElems, DAG));
3259 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3260 getMOVLMask(NumElems, DAG));
3263 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3264 getUnpacklMask(NumElems, DAG));
3269 // Take advantage of the fact R32 to VR128 scalar_to_vector (i.e. movd)
3270 // clears the upper bits.
3271 // FIXME: we can do the same for v4f32 case when we know both parts of
3272 // the lower half come from scalar_to_vector (loadf32). We should do
3273 // that in post legalizer dag combiner with target specific hooks.
3274 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3276 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3277 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
3278 std::vector<SDOperand> MaskVec;
3279 bool Reverse = (NonZeros & 0x3) == 2;
3280 for (unsigned i = 0; i < 2; ++i)
3282 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3284 MaskVec.push_back(DAG.getConstant(i, EVT));
3285 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3286 for (unsigned i = 0; i < 2; ++i)
3288 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3290 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3291 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
3292 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3295 if (Values.size() > 2) {
3296 // Expand into a number of unpckl*.
3298 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3299 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3300 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3301 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3302 for (unsigned i = 0; i < NumElems; ++i)
3303 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3305 while (NumElems != 0) {
3306 for (unsigned i = 0; i < NumElems; ++i)
3307 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3316 case ISD::EXTRACT_VECTOR_ELT: {
3317 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3320 MVT::ValueType VT = Op.getValueType();
3321 // TODO: handle v16i8.
3322 if (MVT::getSizeInBits(VT) == 16) {
3323 // Transform it so it match pextrw which produces a 32-bit result.
3324 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3325 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3326 Op.getOperand(0), Op.getOperand(1));
3327 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3328 DAG.getValueType(VT));
3329 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3330 } else if (MVT::getSizeInBits(VT) == 32) {
3331 SDOperand Vec = Op.getOperand(0);
3332 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3336 // SHUFPS the element to the lowest double word, then movss.
3337 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3338 SDOperand IdxNode = DAG.getConstant((Idx < 2) ? Idx : Idx+4,
3339 MVT::getVectorBaseType(MaskVT));
3340 std::vector<SDOperand> IdxVec;
3341 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
3342 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3343 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3344 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3345 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, IdxVec);
3346 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3348 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3349 DAG.getConstant(0, MVT::i32));
3350 } else if (MVT::getSizeInBits(VT) == 64) {
3351 SDOperand Vec = Op.getOperand(0);
3352 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3356 // UNPCKHPD the element to the lowest double word, then movsd.
3357 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3358 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3359 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3360 std::vector<SDOperand> IdxVec;
3361 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
3362 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3363 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, IdxVec);
3364 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3365 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3366 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3367 DAG.getConstant(0, MVT::i32));
3372 case ISD::INSERT_VECTOR_ELT: {
3373 // Transform it so it match pinsrw which expects a 16-bit value in a R32
3374 // as its second argument.
3375 MVT::ValueType VT = Op.getValueType();
3376 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
3377 SDOperand N0 = Op.getOperand(0);
3378 SDOperand N1 = Op.getOperand(1);
3379 SDOperand N2 = Op.getOperand(2);
3380 if (MVT::getSizeInBits(BaseVT) == 16) {
3381 if (N1.getValueType() != MVT::i32)
3382 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3383 if (N2.getValueType() != MVT::i32)
3384 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
3385 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3386 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3387 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3390 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3391 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3392 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3393 std::vector<SDOperand> MaskVec;
3394 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3395 for (unsigned i = 1; i <= 3; ++i)
3396 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3397 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3398 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec));
3400 // Use two pinsrw instructions to insert a 32 bit value.
3402 if (MVT::isFloatingPoint(N1.getValueType())) {
3403 if (N1.getOpcode() == ISD::LOAD) {
3404 // Just load directly from f32mem to R32.
3405 N1 = DAG.getLoad(MVT::i32, N1.getOperand(0), N1.getOperand(1),
3408 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3409 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3410 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3411 DAG.getConstant(0, MVT::i32));
3414 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3415 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3416 DAG.getConstant(Idx, MVT::i32));
3417 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3418 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3419 DAG.getConstant(Idx+1, MVT::i32));
3420 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3426 case ISD::INTRINSIC_WO_CHAIN: {
3427 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
3429 default: return SDOperand(); // Don't custom lower most intrinsics.
3430 // Comparison intrinsics.
3431 case Intrinsic::x86_sse_comieq_ss:
3432 case Intrinsic::x86_sse_comilt_ss:
3433 case Intrinsic::x86_sse_comile_ss:
3434 case Intrinsic::x86_sse_comigt_ss:
3435 case Intrinsic::x86_sse_comige_ss:
3436 case Intrinsic::x86_sse_comineq_ss:
3437 case Intrinsic::x86_sse_ucomieq_ss:
3438 case Intrinsic::x86_sse_ucomilt_ss:
3439 case Intrinsic::x86_sse_ucomile_ss:
3440 case Intrinsic::x86_sse_ucomigt_ss:
3441 case Intrinsic::x86_sse_ucomige_ss:
3442 case Intrinsic::x86_sse_ucomineq_ss:
3443 case Intrinsic::x86_sse2_comieq_sd:
3444 case Intrinsic::x86_sse2_comilt_sd:
3445 case Intrinsic::x86_sse2_comile_sd:
3446 case Intrinsic::x86_sse2_comigt_sd:
3447 case Intrinsic::x86_sse2_comige_sd:
3448 case Intrinsic::x86_sse2_comineq_sd:
3449 case Intrinsic::x86_sse2_ucomieq_sd:
3450 case Intrinsic::x86_sse2_ucomilt_sd:
3451 case Intrinsic::x86_sse2_ucomile_sd:
3452 case Intrinsic::x86_sse2_ucomigt_sd:
3453 case Intrinsic::x86_sse2_ucomige_sd:
3454 case Intrinsic::x86_sse2_ucomineq_sd: {
3456 ISD::CondCode CC = ISD::SETCC_INVALID;
3459 case Intrinsic::x86_sse_comieq_ss:
3460 case Intrinsic::x86_sse2_comieq_sd:
3464 case Intrinsic::x86_sse_comilt_ss:
3465 case Intrinsic::x86_sse2_comilt_sd:
3469 case Intrinsic::x86_sse_comile_ss:
3470 case Intrinsic::x86_sse2_comile_sd:
3474 case Intrinsic::x86_sse_comigt_ss:
3475 case Intrinsic::x86_sse2_comigt_sd:
3479 case Intrinsic::x86_sse_comige_ss:
3480 case Intrinsic::x86_sse2_comige_sd:
3484 case Intrinsic::x86_sse_comineq_ss:
3485 case Intrinsic::x86_sse2_comineq_sd:
3489 case Intrinsic::x86_sse_ucomieq_ss:
3490 case Intrinsic::x86_sse2_ucomieq_sd:
3491 Opc = X86ISD::UCOMI;
3494 case Intrinsic::x86_sse_ucomilt_ss:
3495 case Intrinsic::x86_sse2_ucomilt_sd:
3496 Opc = X86ISD::UCOMI;
3499 case Intrinsic::x86_sse_ucomile_ss:
3500 case Intrinsic::x86_sse2_ucomile_sd:
3501 Opc = X86ISD::UCOMI;
3504 case Intrinsic::x86_sse_ucomigt_ss:
3505 case Intrinsic::x86_sse2_ucomigt_sd:
3506 Opc = X86ISD::UCOMI;
3509 case Intrinsic::x86_sse_ucomige_ss:
3510 case Intrinsic::x86_sse2_ucomige_sd:
3511 Opc = X86ISD::UCOMI;
3514 case Intrinsic::x86_sse_ucomineq_ss:
3515 case Intrinsic::x86_sse2_ucomineq_sd:
3516 Opc = X86ISD::UCOMI;
3522 translateX86CC(CC, true, X86CC, Flip);
3523 SDOperand Cond = DAG.getNode(Opc, MVT::Flag, Op.getOperand(Flip?2:1),
3524 Op.getOperand(Flip?1:2));
3525 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
3526 DAG.getConstant(X86CC, MVT::i8), Cond);
3527 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
3534 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
3536 default: return NULL;
3537 case X86ISD::SHLD: return "X86ISD::SHLD";
3538 case X86ISD::SHRD: return "X86ISD::SHRD";
3539 case X86ISD::FAND: return "X86ISD::FAND";
3540 case X86ISD::FXOR: return "X86ISD::FXOR";
3541 case X86ISD::FILD: return "X86ISD::FILD";
3542 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
3543 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
3544 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
3545 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
3546 case X86ISD::FLD: return "X86ISD::FLD";
3547 case X86ISD::FST: return "X86ISD::FST";
3548 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
3549 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
3550 case X86ISD::CALL: return "X86ISD::CALL";
3551 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
3552 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
3553 case X86ISD::CMP: return "X86ISD::CMP";
3554 case X86ISD::TEST: return "X86ISD::TEST";
3555 case X86ISD::COMI: return "X86ISD::COMI";
3556 case X86ISD::UCOMI: return "X86ISD::UCOMI";
3557 case X86ISD::SETCC: return "X86ISD::SETCC";
3558 case X86ISD::CMOV: return "X86ISD::CMOV";
3559 case X86ISD::BRCOND: return "X86ISD::BRCOND";
3560 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
3561 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
3562 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
3563 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
3564 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
3565 case X86ISD::Wrapper: return "X86ISD::Wrapper";
3566 case X86ISD::S2VEC: return "X86ISD::S2VEC";
3567 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
3568 case X86ISD::PINSRW: return "X86ISD::PINSRW";
3572 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
3574 uint64_t &KnownZero,
3576 unsigned Depth) const {
3577 unsigned Opc = Op.getOpcode();
3578 assert((Opc >= ISD::BUILTIN_OP_END ||
3579 Opc == ISD::INTRINSIC_WO_CHAIN ||
3580 Opc == ISD::INTRINSIC_W_CHAIN ||
3581 Opc == ISD::INTRINSIC_VOID) &&
3582 "Should use MaskedValueIsZero if you don't know whether Op"
3583 " is a target node!");
3585 KnownZero = KnownOne = 0; // Don't know anything.
3589 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
3594 std::vector<unsigned> X86TargetLowering::
3595 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3596 MVT::ValueType VT) const {
3597 if (Constraint.size() == 1) {
3598 // FIXME: not handling fp-stack yet!
3599 // FIXME: not handling MMX registers yet ('y' constraint).
3600 switch (Constraint[0]) { // GCC X86 Constraint Letters
3601 default: break; // Unknown constriant letter
3602 case 'r': // GENERAL_REGS
3603 case 'R': // LEGACY_REGS
3604 return make_vector<unsigned>(X86::EAX, X86::EBX, X86::ECX, X86::EDX,
3605 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
3606 case 'l': // INDEX_REGS
3607 return make_vector<unsigned>(X86::EAX, X86::EBX, X86::ECX, X86::EDX,
3608 X86::ESI, X86::EDI, X86::EBP, 0);
3609 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
3611 return make_vector<unsigned>(X86::EAX, X86::EBX, X86::ECX, X86::EDX, 0);
3612 case 'x': // SSE_REGS if SSE1 allowed
3613 if (Subtarget->hasSSE1())
3614 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3615 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
3617 return std::vector<unsigned>();
3618 case 'Y': // SSE_REGS if SSE2 allowed
3619 if (Subtarget->hasSSE2())
3620 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3621 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
3623 return std::vector<unsigned>();
3627 return std::vector<unsigned>();
3630 /// isLegalAddressImmediate - Return true if the integer value or
3631 /// GlobalValue can be used as the offset of the target addressing mode.
3632 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
3633 // X86 allows a sign-extended 32-bit immediate field.
3634 return (V > -(1LL << 32) && V < (1LL << 32)-1);
3637 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
3638 if (Subtarget->isTargetDarwin()) {
3639 Reloc::Model RModel = getTargetMachine().getRelocationModel();
3640 if (RModel == Reloc::Static)
3642 else if (RModel == Reloc::DynamicNoPIC)
3643 return !DarwinGVRequiresExtraLoad(GV);
3650 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3651 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3652 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3653 /// are assumed to be legal.
3655 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
3656 // Only do shuffles on 128-bit vector types for now.
3657 if (MVT::getSizeInBits(VT) == 64) return false;
3658 return (Mask.Val->getNumOperands() <= 4 ||
3659 isSplatMask(Mask.Val) ||
3660 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
3661 X86::isUNPCKLMask(Mask.Val) ||
3662 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
3663 X86::isUNPCKHMask(Mask.Val));
3666 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
3668 SelectionDAG &DAG) const {
3669 unsigned NumElts = BVOps.size();
3670 // Only do shuffles on 128-bit vector types for now.
3671 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
3672 if (NumElts == 2) return true;
3674 return (isMOVLMask(BVOps) || isCommutedMOVL(BVOps, true) ||
3675 isSHUFPMask(BVOps) || isCommutedSHUFP(BVOps));