1 //===-- X86ISelPattern.cpp - A pattern matching inst selector for X86 -----===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a pattern matching instruction selector for X86.
12 //===----------------------------------------------------------------------===//
15 #include "X86InstrBuilder.h"
16 #include "X86RegisterInfo.h"
17 #include "X86Subtarget.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/Constants.h"
20 #include "llvm/Instructions.h"
21 #include "llvm/Function.h"
22 #include "llvm/CodeGen/MachineConstantPool.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/SelectionDAG.h"
26 #include "llvm/CodeGen/SelectionDAGISel.h"
27 #include "llvm/CodeGen/SSARegMap.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Target/TargetLowering.h"
30 #include "llvm/Target/TargetMachine.h"
31 #include "llvm/Target/TargetOptions.h"
32 #include "llvm/Support/CFG.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/ADT/Statistic.h"
40 #include "llvm/Support/CommandLine.h"
41 static cl::opt<bool> EnableFastCC("enable-x86-fastcc", cl::Hidden,
42 cl::desc("Enable fastcc on X86"));
45 // X86 Specific DAG Nodes
48 // Start the numbering where the builtin ops leave off.
49 FIRST_NUMBER = ISD::BUILTIN_OP_END,
51 /// FILD64m - This instruction implements SINT_TO_FP with a
52 /// 64-bit source in memory and a FP reg result. This corresponds to
53 /// the X86::FILD64m instruction. It has two inputs (token chain and
54 /// address) and two outputs (FP value and token chain).
57 /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the
58 /// integer destination in memory and a FP reg source. This corresponds
59 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
60 /// has two inputs (token chain and address) and two outputs (FP value and
66 /// CALL/TAILCALL - These operations represent an abstract X86 call
67 /// instruction, which includes a bunch of information. In particular the
68 /// operands of these node are:
70 /// #0 - The incoming token chain
72 /// #2 - The number of arg bytes the caller pushes on the stack.
73 /// #3 - The number of arg bytes the callee pops off the stack.
74 /// #4 - The value to pass in AL/AX/EAX (optional)
75 /// #5 - The value to pass in DL/DX/EDX (optional)
77 /// The result values of these nodes are:
79 /// #0 - The outgoing token chain
80 /// #1 - The first register result value (optional)
81 /// #2 - The second register result value (optional)
83 /// The CALL vs TAILCALL distinction boils down to whether the callee is
84 /// known not to modify the caller's stack frame, as is standard with
92 //===----------------------------------------------------------------------===//
93 // X86TargetLowering - X86 Implementation of the TargetLowering interface
95 class X86TargetLowering : public TargetLowering {
96 int VarArgsFrameIndex; // FrameIndex for start of varargs area.
97 int ReturnAddrIndex; // FrameIndex for return slot.
98 int BytesToPopOnReturn; // Number of arg bytes ret should pop.
99 int BytesCallerReserves; // Number of arg bytes caller makes.
101 X86TargetLowering(TargetMachine &TM) : TargetLowering(TM) {
102 // Set up the TargetLowering object.
104 // X86 is weird, it always uses i8 for shift amounts and setcc results.
105 setShiftAmountType(MVT::i8);
106 setSetCCResultType(MVT::i8);
107 setSetCCResultContents(ZeroOrOneSetCCResult);
108 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
110 // Set up the register classes.
111 // FIXME: Eliminate these two classes when legalize can handle promotions
113 addRegisterClass(MVT::i1, X86::R8RegisterClass);
114 addRegisterClass(MVT::i8, X86::R8RegisterClass);
115 addRegisterClass(MVT::i16, X86::R16RegisterClass);
116 addRegisterClass(MVT::i32, X86::R32RegisterClass);
118 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
120 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
121 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
122 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
123 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
125 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
127 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
128 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
131 // We can handle SINT_TO_FP and FP_TO_SINT from/TO i64 even though i64
133 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
134 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
135 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
136 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
139 // Handle FP_TO_UINT by promoting the destination to a larger signed
141 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
142 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
143 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
146 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
148 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
150 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
151 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
152 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
154 setOperationAction(ISD::BRCONDTWOWAY , MVT::Other, Expand);
155 setOperationAction(ISD::BRTWOWAY_CC , MVT::Other, Expand);
156 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
157 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
158 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
159 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
160 setOperationAction(ISD::SEXTLOAD , MVT::i1 , Expand);
161 setOperationAction(ISD::FREM , MVT::f64 , Expand);
162 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
163 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
164 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
165 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
166 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
167 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
168 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
169 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
170 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
172 setOperationAction(ISD::READIO , MVT::i1 , Expand);
173 setOperationAction(ISD::READIO , MVT::i8 , Expand);
174 setOperationAction(ISD::READIO , MVT::i16 , Expand);
175 setOperationAction(ISD::READIO , MVT::i32 , Expand);
176 setOperationAction(ISD::WRITEIO , MVT::i1 , Expand);
177 setOperationAction(ISD::WRITEIO , MVT::i8 , Expand);
178 setOperationAction(ISD::WRITEIO , MVT::i16 , Expand);
179 setOperationAction(ISD::WRITEIO , MVT::i32 , Expand);
181 // These should be promoted to a larger select which is supported.
182 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
183 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
186 // Set up the FP register classes.
187 addRegisterClass(MVT::f32, X86::RXMMRegisterClass);
188 addRegisterClass(MVT::f64, X86::RXMMRegisterClass);
190 // SSE has no load+extend ops
191 setOperationAction(ISD::EXTLOAD, MVT::f32, Expand);
192 setOperationAction(ISD::ZEXTLOAD, MVT::f32, Expand);
194 // SSE has no i16 to fp conversion, only i32
195 setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
196 setOperationAction(ISD::FP_TO_SINT, MVT::i16, Promote);
198 // Expand FP_TO_UINT into a select.
199 // FIXME: We would like to use a Custom expander here eventually to do
200 // the optimal thing for SSE vs. the default expansion in the legalizer.
201 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
203 // We don't support sin/cos/sqrt/fmod
204 setOperationAction(ISD::FSIN , MVT::f64, Expand);
205 setOperationAction(ISD::FCOS , MVT::f64, Expand);
206 setOperationAction(ISD::FABS , MVT::f64, Expand);
207 setOperationAction(ISD::FNEG , MVT::f64, Expand);
208 setOperationAction(ISD::FREM , MVT::f64, Expand);
209 setOperationAction(ISD::FSIN , MVT::f32, Expand);
210 setOperationAction(ISD::FCOS , MVT::f32, Expand);
211 setOperationAction(ISD::FABS , MVT::f32, Expand);
212 setOperationAction(ISD::FNEG , MVT::f32, Expand);
213 setOperationAction(ISD::FREM , MVT::f32, Expand);
215 addLegalFPImmediate(+0.0); // xorps / xorpd
217 // Set up the FP register classes.
218 addRegisterClass(MVT::f64, X86::RFPRegisterClass);
221 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
222 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
225 addLegalFPImmediate(+0.0); // FLD0
226 addLegalFPImmediate(+1.0); // FLD1
227 addLegalFPImmediate(-0.0); // FLD0/FCHS
228 addLegalFPImmediate(-1.0); // FLD1/FCHS
230 computeRegisterProperties();
232 maxStoresPerMemSet = 8; // For %llvm.memset -> sequence of stores
233 maxStoresPerMemCpy = 8; // For %llvm.memcpy -> sequence of stores
234 maxStoresPerMemMove = 8; // For %llvm.memmove -> sequence of stores
235 allowUnalignedMemoryAccesses = true; // x86 supports it!
238 // Return the number of bytes that a function should pop when it returns (in
239 // addition to the space used by the return address).
241 unsigned getBytesToPopOnReturn() const { return BytesToPopOnReturn; }
243 // Return the number of bytes that the caller reserves for arguments passed
245 unsigned getBytesCallerReserves() const { return BytesCallerReserves; }
247 /// LowerOperation - Provide custom lowering hooks for some operations.
249 virtual SDOperand LowerOperation(SDOperand Op, SelectionDAG &DAG);
251 /// LowerArguments - This hook must be implemented to indicate how we should
252 /// lower the arguments for the specified function, into the specified DAG.
253 virtual std::vector<SDOperand>
254 LowerArguments(Function &F, SelectionDAG &DAG);
256 /// LowerCallTo - This hook lowers an abstract call to a function into an
258 virtual std::pair<SDOperand, SDOperand>
259 LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, unsigned CC,
260 bool isTailCall, SDOperand Callee, ArgListTy &Args,
263 virtual SDOperand LowerVAStart(SDOperand Chain, SDOperand VAListP,
264 Value *VAListV, SelectionDAG &DAG);
265 virtual std::pair<SDOperand,SDOperand>
266 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
267 const Type *ArgTy, SelectionDAG &DAG);
269 virtual std::pair<SDOperand, SDOperand>
270 LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
273 SDOperand getReturnAddressFrameIndex(SelectionDAG &DAG);
276 // C Calling Convention implementation.
277 std::vector<SDOperand> LowerCCCArguments(Function &F, SelectionDAG &DAG);
278 std::pair<SDOperand, SDOperand>
279 LowerCCCCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg,
281 SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG);
283 // Fast Calling Convention implementation.
284 std::vector<SDOperand> LowerFastCCArguments(Function &F, SelectionDAG &DAG);
285 std::pair<SDOperand, SDOperand>
286 LowerFastCCCallTo(SDOperand Chain, const Type *RetTy, bool isTailCall,
287 SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG);
291 std::vector<SDOperand>
292 X86TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
293 if (F.getCallingConv() == CallingConv::Fast && EnableFastCC)
294 return LowerFastCCArguments(F, DAG);
295 return LowerCCCArguments(F, DAG);
298 std::pair<SDOperand, SDOperand>
299 X86TargetLowering::LowerCallTo(SDOperand Chain, const Type *RetTy,
300 bool isVarArg, unsigned CallingConv,
302 SDOperand Callee, ArgListTy &Args,
304 assert((!isVarArg || CallingConv == CallingConv::C) &&
305 "Only C takes varargs!");
306 if (CallingConv == CallingConv::Fast && EnableFastCC)
307 return LowerFastCCCallTo(Chain, RetTy, isTailCall, Callee, Args, DAG);
308 return LowerCCCCallTo(Chain, RetTy, isVarArg, isTailCall, Callee, Args, DAG);
311 //===----------------------------------------------------------------------===//
312 // C Calling Convention implementation
313 //===----------------------------------------------------------------------===//
315 std::vector<SDOperand>
316 X86TargetLowering::LowerCCCArguments(Function &F, SelectionDAG &DAG) {
317 std::vector<SDOperand> ArgValues;
319 MachineFunction &MF = DAG.getMachineFunction();
320 MachineFrameInfo *MFI = MF.getFrameInfo();
322 // Add DAG nodes to load the arguments... On entry to a function on the X86,
323 // the stack frame looks like this:
325 // [ESP] -- return address
326 // [ESP + 4] -- first argument (leftmost lexically)
327 // [ESP + 8] -- second argument, if first argument is four bytes in size
330 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
331 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
332 MVT::ValueType ObjectVT = getValueType(I->getType());
333 unsigned ArgIncrement = 4;
336 default: assert(0 && "Unhandled argument type!");
338 case MVT::i8: ObjSize = 1; break;
339 case MVT::i16: ObjSize = 2; break;
340 case MVT::i32: ObjSize = 4; break;
341 case MVT::i64: ObjSize = ArgIncrement = 8; break;
342 case MVT::f32: ObjSize = 4; break;
343 case MVT::f64: ObjSize = ArgIncrement = 8; break;
345 // Create the frame index object for this incoming parameter...
346 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
348 // Create the SelectionDAG nodes corresponding to a load from this parameter
349 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
351 // Don't codegen dead arguments. FIXME: remove this check when we can nuke
355 ArgValue = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
356 DAG.getSrcValue(NULL));
358 if (MVT::isInteger(ObjectVT))
359 ArgValue = DAG.getConstant(0, ObjectVT);
361 ArgValue = DAG.getConstantFP(0, ObjectVT);
363 ArgValues.push_back(ArgValue);
365 ArgOffset += ArgIncrement; // Move on to the next argument...
368 // If the function takes variable number of arguments, make a frame index for
369 // the start of the first vararg value... for expansion of llvm.va_start.
371 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
372 ReturnAddrIndex = 0; // No return address slot generated yet.
373 BytesToPopOnReturn = 0; // Callee pops nothing.
374 BytesCallerReserves = ArgOffset;
376 // Finally, inform the code generator which regs we return values in.
377 switch (getValueType(F.getReturnType())) {
378 default: assert(0 && "Unknown type!");
379 case MVT::isVoid: break;
384 MF.addLiveOut(X86::EAX);
387 MF.addLiveOut(X86::EAX);
388 MF.addLiveOut(X86::EDX);
392 MF.addLiveOut(X86::ST0);
398 std::pair<SDOperand, SDOperand>
399 X86TargetLowering::LowerCCCCallTo(SDOperand Chain, const Type *RetTy,
400 bool isVarArg, bool isTailCall,
401 SDOperand Callee, ArgListTy &Args,
403 // Count how many bytes are to be pushed on the stack.
404 unsigned NumBytes = 0;
408 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
409 DAG.getConstant(0, getPointerTy()));
411 for (unsigned i = 0, e = Args.size(); i != e; ++i)
412 switch (getValueType(Args[i].second)) {
413 default: assert(0 && "Unknown value type!");
427 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
428 DAG.getConstant(NumBytes, getPointerTy()));
430 // Arguments go on the stack in reverse order, as specified by the ABI.
431 unsigned ArgOffset = 0;
432 SDOperand StackPtr = DAG.getCopyFromReg(DAG.getEntryNode(),
434 std::vector<SDOperand> Stores;
436 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
437 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
438 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
440 switch (getValueType(Args[i].second)) {
441 default: assert(0 && "Unexpected ValueType for argument!");
445 // Promote the integer to 32 bits. If the input type is signed use a
446 // sign extend, otherwise use a zero extend.
447 if (Args[i].second->isSigned())
448 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
450 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
455 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
456 Args[i].first, PtrOff,
457 DAG.getSrcValue(NULL)));
462 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
463 Args[i].first, PtrOff,
464 DAG.getSrcValue(NULL)));
469 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
472 std::vector<MVT::ValueType> RetVals;
473 MVT::ValueType RetTyVT = getValueType(RetTy);
474 RetVals.push_back(MVT::Other);
476 // The result values produced have to be legal. Promote the result.
478 case MVT::isVoid: break;
480 RetVals.push_back(RetTyVT);
485 RetVals.push_back(MVT::i32);
489 RetVals.push_back(MVT::f32);
491 RetVals.push_back(MVT::f64);
494 RetVals.push_back(MVT::i32);
495 RetVals.push_back(MVT::i32);
498 std::vector<SDOperand> Ops;
499 Ops.push_back(Chain);
500 Ops.push_back(Callee);
501 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
502 Ops.push_back(DAG.getConstant(0, getPointerTy()));
503 SDOperand TheCall = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
505 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, TheCall);
509 case MVT::isVoid: break;
511 ResultVal = TheCall.getValue(1);
516 ResultVal = DAG.getNode(ISD::TRUNCATE, RetTyVT, TheCall.getValue(1));
519 // FIXME: we would really like to remember that this FP_ROUND operation is
520 // okay to eliminate if we allow excess FP precision.
521 ResultVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, TheCall.getValue(1));
524 ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, TheCall.getValue(1),
525 TheCall.getValue(2));
529 return std::make_pair(ResultVal, Chain);
533 X86TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
534 Value *VAListV, SelectionDAG &DAG) {
535 // vastart just stores the address of the VarArgsFrameIndex slot.
536 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
537 return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR, VAListP,
538 DAG.getSrcValue(VAListV));
542 std::pair<SDOperand,SDOperand>
543 X86TargetLowering::LowerVAArg(SDOperand Chain, SDOperand VAListP,
544 Value *VAListV, const Type *ArgTy,
546 MVT::ValueType ArgVT = getValueType(ArgTy);
547 SDOperand Val = DAG.getLoad(MVT::i32, Chain,
548 VAListP, DAG.getSrcValue(VAListV));
549 SDOperand Result = DAG.getLoad(ArgVT, Chain, Val,
550 DAG.getSrcValue(NULL));
552 if (ArgVT == MVT::i32)
555 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
556 "Other types should have been promoted for varargs!");
559 Val = DAG.getNode(ISD::ADD, Val.getValueType(), Val,
560 DAG.getConstant(Amt, Val.getValueType()));
561 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
562 Val, VAListP, DAG.getSrcValue(VAListV));
563 return std::make_pair(Result, Chain);
566 //===----------------------------------------------------------------------===//
567 // Fast Calling Convention implementation
568 //===----------------------------------------------------------------------===//
570 // The X86 'fast' calling convention passes up to two integer arguments in
571 // registers (an appropriate portion of EAX/EDX), passes arguments in C order,
572 // and requires that the callee pop its arguments off the stack (allowing proper
573 // tail calls), and has the same return value conventions as C calling convs.
575 // This calling convention always arranges for the callee pop value to be 8n+4
576 // bytes, which is needed for tail recursion elimination and stack alignment
579 // Note that this can be enhanced in the future to pass fp vals in registers
580 // (when we have a global fp allocator) and do other tricks.
583 /// AddLiveIn - This helper function adds the specified physical register to the
584 /// MachineFunction as a live in value. It also creates a corresponding virtual
586 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
587 TargetRegisterClass *RC) {
588 assert(RC->contains(PReg) && "Not the correct regclass!");
589 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
590 MF.addLiveIn(PReg, VReg);
595 std::vector<SDOperand>
596 X86TargetLowering::LowerFastCCArguments(Function &F, SelectionDAG &DAG) {
597 std::vector<SDOperand> ArgValues;
599 MachineFunction &MF = DAG.getMachineFunction();
600 MachineFrameInfo *MFI = MF.getFrameInfo();
602 // Add DAG nodes to load the arguments... On entry to a function the stack
603 // frame looks like this:
605 // [ESP] -- return address
606 // [ESP + 4] -- first nonreg argument (leftmost lexically)
607 // [ESP + 8] -- second nonreg argument, if first argument is 4 bytes in size
609 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
611 // Keep track of the number of integer regs passed so far. This can be either
612 // 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
614 unsigned NumIntRegs = 0;
616 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
617 MVT::ValueType ObjectVT = getValueType(I->getType());
618 unsigned ArgIncrement = 4;
619 unsigned ObjSize = 0;
623 default: assert(0 && "Unhandled argument type!");
626 if (NumIntRegs < 2) {
627 if (!I->use_empty()) {
628 unsigned VReg = AddLiveIn(MF, NumIntRegs ? X86::DL : X86::AL,
629 X86::R8RegisterClass);
630 ArgValue = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i8);
631 DAG.setRoot(ArgValue.getValue(1));
640 if (NumIntRegs < 2) {
641 if (!I->use_empty()) {
642 unsigned VReg = AddLiveIn(MF, NumIntRegs ? X86::DX : X86::AX,
643 X86::R16RegisterClass);
644 ArgValue = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i16);
645 DAG.setRoot(ArgValue.getValue(1));
653 if (NumIntRegs < 2) {
654 if (!I->use_empty()) {
655 unsigned VReg = AddLiveIn(MF,NumIntRegs ? X86::EDX : X86::EAX,
656 X86::R32RegisterClass);
657 ArgValue = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
658 DAG.setRoot(ArgValue.getValue(1));
666 if (NumIntRegs == 0) {
667 if (!I->use_empty()) {
668 unsigned BotReg = AddLiveIn(MF, X86::EAX, X86::R32RegisterClass);
669 unsigned TopReg = AddLiveIn(MF, X86::EDX, X86::R32RegisterClass);
671 SDOperand Low = DAG.getCopyFromReg(DAG.getRoot(), BotReg, MVT::i32);
672 SDOperand Hi = DAG.getCopyFromReg(Low.getValue(1), TopReg, MVT::i32);
673 DAG.setRoot(Hi.getValue(1));
675 ArgValue = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Low, Hi);
679 } else if (NumIntRegs == 1) {
680 if (!I->use_empty()) {
681 unsigned BotReg = AddLiveIn(MF, X86::EDX, X86::R32RegisterClass);
682 SDOperand Low = DAG.getCopyFromReg(DAG.getRoot(), BotReg, MVT::i32);
683 DAG.setRoot(Low.getValue(1));
685 // Load the high part from memory.
686 // Create the frame index object for this incoming parameter...
687 int FI = MFI->CreateFixedObject(4, ArgOffset);
688 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
689 SDOperand Hi = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
690 DAG.getSrcValue(NULL));
691 ArgValue = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Low, Hi);
697 ObjSize = ArgIncrement = 8;
699 case MVT::f32: ObjSize = 4; break;
700 case MVT::f64: ObjSize = ArgIncrement = 8; break;
703 // Don't codegen dead arguments. FIXME: remove this check when we can nuke
705 if (ObjSize && !I->use_empty()) {
706 // Create the frame index object for this incoming parameter...
707 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
709 // Create the SelectionDAG nodes corresponding to a load from this
711 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
713 ArgValue = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
714 DAG.getSrcValue(NULL));
715 } else if (ArgValue.Val == 0) {
716 if (MVT::isInteger(ObjectVT))
717 ArgValue = DAG.getConstant(0, ObjectVT);
719 ArgValue = DAG.getConstantFP(0, ObjectVT);
721 ArgValues.push_back(ArgValue);
724 ArgOffset += ArgIncrement; // Move on to the next argument.
727 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
728 // arguments and the arguments after the retaddr has been pushed are aligned.
729 if ((ArgOffset & 7) == 0)
732 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
733 ReturnAddrIndex = 0; // No return address slot generated yet.
734 BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
735 BytesCallerReserves = 0;
737 // Finally, inform the code generator which regs we return values in.
738 switch (getValueType(F.getReturnType())) {
739 default: assert(0 && "Unknown type!");
740 case MVT::isVoid: break;
745 MF.addLiveOut(X86::EAX);
748 MF.addLiveOut(X86::EAX);
749 MF.addLiveOut(X86::EDX);
753 MF.addLiveOut(X86::ST0);
759 std::pair<SDOperand, SDOperand>
760 X86TargetLowering::LowerFastCCCallTo(SDOperand Chain, const Type *RetTy,
761 bool isTailCall, SDOperand Callee,
762 ArgListTy &Args, SelectionDAG &DAG) {
763 // Count how many bytes are to be pushed on the stack.
764 unsigned NumBytes = 0;
766 // Keep track of the number of integer regs passed so far. This can be either
767 // 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
769 unsigned NumIntRegs = 0;
771 for (unsigned i = 0, e = Args.size(); i != e; ++i)
772 switch (getValueType(Args[i].second)) {
773 default: assert(0 && "Unknown value type!");
778 if (NumIntRegs < 2) {
787 if (NumIntRegs == 0) {
790 } else if (NumIntRegs == 1) {
802 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
803 // arguments and the arguments after the retaddr has been pushed are aligned.
804 if ((NumBytes & 7) == 0)
807 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
808 DAG.getConstant(NumBytes, getPointerTy()));
810 // Arguments go on the stack in reverse order, as specified by the ABI.
811 unsigned ArgOffset = 0;
812 SDOperand StackPtr = DAG.getCopyFromReg(DAG.getEntryNode(),
815 std::vector<SDOperand> Stores;
816 std::vector<SDOperand> RegValuesToPass;
817 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
818 switch (getValueType(Args[i].second)) {
819 default: assert(0 && "Unexpected ValueType for argument!");
824 if (NumIntRegs < 2) {
825 RegValuesToPass.push_back(Args[i].first);
831 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
832 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
833 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
834 Args[i].first, PtrOff,
835 DAG.getSrcValue(NULL)));
840 if (NumIntRegs < 2) { // Can pass part of it in regs?
841 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
842 Args[i].first, DAG.getConstant(1, MVT::i32));
843 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
844 Args[i].first, DAG.getConstant(0, MVT::i32));
845 RegValuesToPass.push_back(Lo);
847 if (NumIntRegs < 2) { // Pass both parts in regs?
848 RegValuesToPass.push_back(Hi);
851 // Pass the high part in memory.
852 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
853 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
854 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
855 Hi, PtrOff, DAG.getSrcValue(NULL)));
862 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
863 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
864 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
865 Args[i].first, PtrOff,
866 DAG.getSrcValue(NULL)));
872 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
874 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
875 // arguments and the arguments after the retaddr has been pushed are aligned.
876 if ((ArgOffset & 7) == 0)
879 std::vector<MVT::ValueType> RetVals;
880 MVT::ValueType RetTyVT = getValueType(RetTy);
882 RetVals.push_back(MVT::Other);
884 // The result values produced have to be legal. Promote the result.
886 case MVT::isVoid: break;
888 RetVals.push_back(RetTyVT);
893 RetVals.push_back(MVT::i32);
897 RetVals.push_back(MVT::f32);
899 RetVals.push_back(MVT::f64);
902 RetVals.push_back(MVT::i32);
903 RetVals.push_back(MVT::i32);
907 std::vector<SDOperand> Ops;
908 Ops.push_back(Chain);
909 Ops.push_back(Callee);
910 Ops.push_back(DAG.getConstant(ArgOffset, getPointerTy()));
911 // Callee pops all arg values on the stack.
912 Ops.push_back(DAG.getConstant(ArgOffset, getPointerTy()));
914 // Pass register arguments as needed.
915 Ops.insert(Ops.end(), RegValuesToPass.begin(), RegValuesToPass.end());
917 SDOperand TheCall = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
919 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, TheCall);
923 case MVT::isVoid: break;
925 ResultVal = TheCall.getValue(1);
930 ResultVal = DAG.getNode(ISD::TRUNCATE, RetTyVT, TheCall.getValue(1));
933 // FIXME: we would really like to remember that this FP_ROUND operation is
934 // okay to eliminate if we allow excess FP precision.
935 ResultVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, TheCall.getValue(1));
938 ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, TheCall.getValue(1),
939 TheCall.getValue(2));
943 return std::make_pair(ResultVal, Chain);
946 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
947 if (ReturnAddrIndex == 0) {
948 // Set up a frame object for the return address.
949 MachineFunction &MF = DAG.getMachineFunction();
950 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
953 return DAG.getFrameIndex(ReturnAddrIndex, MVT::i32);
958 std::pair<SDOperand, SDOperand> X86TargetLowering::
959 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
962 if (Depth) // Depths > 0 not supported yet!
963 Result = DAG.getConstant(0, getPointerTy());
965 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
967 // Just load the return address
968 Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(), RetAddrFI,
969 DAG.getSrcValue(NULL));
971 Result = DAG.getNode(ISD::SUB, MVT::i32, RetAddrFI,
972 DAG.getConstant(4, MVT::i32));
974 return std::make_pair(Result, Chain);
977 //===----------------------------------------------------------------------===//
978 // X86 Custom Lowering Hooks
979 //===----------------------------------------------------------------------===//
981 /// LowerOperation - Provide custom lowering hooks for some operations.
983 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
984 switch (Op.getOpcode()) {
985 default: assert(0 && "Should not custom lower this!");
986 case ISD::SINT_TO_FP: {
987 assert(Op.getValueType() == MVT::f64 &&
988 Op.getOperand(0).getValueType() == MVT::i64 &&
989 "Unknown SINT_TO_FP to lower!");
990 // We lower sint64->FP into a store to a temporary stack slot, followed by a
992 MachineFunction &MF = DAG.getMachineFunction();
993 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
994 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
995 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, DAG.getEntryNode(),
996 Op.getOperand(0), StackSlot, DAG.getSrcValue(NULL));
997 std::vector<MVT::ValueType> RTs;
998 RTs.push_back(MVT::f64);
999 RTs.push_back(MVT::Other);
1000 std::vector<SDOperand> Ops;
1001 Ops.push_back(Store);
1002 Ops.push_back(StackSlot);
1003 return DAG.getNode(X86ISD::FILD64m, RTs, Ops);
1005 case ISD::FP_TO_SINT: {
1006 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
1007 Op.getOperand(0).getValueType() == MVT::f64 &&
1008 "Unknown FP_TO_SINT to lower!");
1009 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
1011 MachineFunction &MF = DAG.getMachineFunction();
1012 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
1013 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
1014 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
1017 switch (Op.getValueType()) {
1018 default: assert(0 && "Invalid FP_TO_SINT to lower!");
1019 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
1020 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
1021 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
1024 // Build the FP_TO_INT*_IN_MEM
1025 std::vector<SDOperand> Ops;
1026 Ops.push_back(DAG.getEntryNode());
1027 Ops.push_back(Op.getOperand(0));
1028 Ops.push_back(StackSlot);
1029 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops);
1032 return DAG.getLoad(Op.getValueType(), FIST, StackSlot,
1033 DAG.getSrcValue(NULL));
1039 //===----------------------------------------------------------------------===//
1040 // Pattern Matcher Implementation
1041 //===----------------------------------------------------------------------===//
1044 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
1045 /// SDOperand's instead of register numbers for the leaves of the matched
1047 struct X86ISelAddressMode {
1053 struct { // This is really a union, discriminated by BaseType!
1063 X86ISelAddressMode()
1064 : BaseType(RegBase), Scale(1), IndexReg(), Disp(), GV(0) {
1072 NumFPKill("x86-codegen", "Number of FP_REG_KILL instructions added");
1074 //===--------------------------------------------------------------------===//
1075 /// ISel - X86 specific code to select X86 machine instructions for
1076 /// SelectionDAG operations.
1078 class ISel : public SelectionDAGISel {
1079 /// ContainsFPCode - Every instruction we select that uses or defines a FP
1080 /// register should set this to true.
1081 bool ContainsFPCode;
1083 /// X86Lowering - This object fully describes how to lower LLVM code to an
1084 /// X86-specific SelectionDAG.
1085 X86TargetLowering X86Lowering;
1087 /// RegPressureMap - This keeps an approximate count of the number of
1088 /// registers required to evaluate each node in the graph.
1089 std::map<SDNode*, unsigned> RegPressureMap;
1091 /// ExprMap - As shared expressions are codegen'd, we keep track of which
1092 /// vreg the value is produced in, so we only emit one copy of each compiled
1094 std::map<SDOperand, unsigned> ExprMap;
1096 /// TheDAG - The DAG being selected during Select* operations.
1097 SelectionDAG *TheDAG;
1099 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
1100 /// make the right decision when generating code for different targets.
1101 const X86Subtarget *Subtarget;
1103 ISel(TargetMachine &TM) : SelectionDAGISel(X86Lowering), X86Lowering(TM) {
1104 Subtarget = &TM.getSubtarget<X86Subtarget>();
1107 virtual const char *getPassName() const {
1108 return "X86 Pattern Instruction Selection";
1111 unsigned getRegPressure(SDOperand O) {
1112 return RegPressureMap[O.Val];
1114 unsigned ComputeRegPressure(SDOperand O);
1116 /// InstructionSelectBasicBlock - This callback is invoked by
1117 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
1118 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
1120 virtual void EmitFunctionEntryCode(Function &Fn, MachineFunction &MF);
1122 bool isFoldableLoad(SDOperand Op, SDOperand OtherOp,
1123 bool FloatPromoteOk = false);
1124 void EmitFoldedLoad(SDOperand Op, X86AddressMode &AM);
1125 bool TryToFoldLoadOpStore(SDNode *Node);
1126 bool EmitOrOpOp(SDOperand Op1, SDOperand Op2, unsigned DestReg);
1127 void EmitCMP(SDOperand LHS, SDOperand RHS, bool isOnlyUse);
1128 bool EmitBranchCC(MachineBasicBlock *Dest, SDOperand Chain, SDOperand Cond);
1129 void EmitSelectCC(SDOperand Cond, SDOperand True, SDOperand False,
1130 MVT::ValueType SVT, unsigned RDest);
1131 unsigned SelectExpr(SDOperand N);
1133 X86AddressMode SelectAddrExprs(const X86ISelAddressMode &IAM);
1134 bool MatchAddress(SDOperand N, X86ISelAddressMode &AM);
1135 void SelectAddress(SDOperand N, X86AddressMode &AM);
1136 bool EmitPotentialTailCall(SDNode *Node);
1137 void EmitFastCCToFastCCTailCall(SDNode *TailCallNode);
1138 void Select(SDOperand N);
1142 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
1143 /// the main function.
1144 static void EmitSpecialCodeForMain(MachineBasicBlock *BB,
1145 MachineFrameInfo *MFI) {
1146 // Switch the FPU to 64-bit precision mode for better compatibility and speed.
1147 int CWFrameIdx = MFI->CreateStackObject(2, 2);
1148 addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
1150 // Set the high part to be 64-bit precision.
1151 addFrameReference(BuildMI(BB, X86::MOV8mi, 5),
1152 CWFrameIdx, 1).addImm(2);
1154 // Reload the modified control word now.
1155 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
1158 void ISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
1159 // If this is main, emit special code for main.
1160 MachineBasicBlock *BB = MF.begin();
1161 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
1162 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
1166 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
1167 /// when it has created a SelectionDAG for us to codegen.
1168 void ISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
1169 // While we're doing this, keep track of whether we see any FP code for
1170 // FP_REG_KILL insertion.
1171 ContainsFPCode = false;
1172 MachineFunction *MF = BB->getParent();
1174 // Scan the PHI nodes that already are inserted into this basic block. If any
1175 // of them is a PHI of a floating point value, we need to insert an
1177 SSARegMap *RegMap = MF->getSSARegMap();
1178 if (BB != MF->begin())
1179 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
1181 assert(I->getOpcode() == X86::PHI &&
1182 "Isn't just PHI nodes?");
1183 if (RegMap->getRegClass(I->getOperand(0).getReg()) ==
1184 X86::RFPRegisterClass) {
1185 ContainsFPCode = true;
1190 // Compute the RegPressureMap, which is an approximation for the number of
1191 // registers required to compute each node.
1192 ComputeRegPressure(DAG.getRoot());
1196 // Codegen the basic block.
1197 Select(DAG.getRoot());
1201 // Finally, look at all of the successors of this block. If any contain a PHI
1202 // node of FP type, we need to insert an FP_REG_KILL in this block.
1203 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1204 E = BB->succ_end(); SI != E && !ContainsFPCode; ++SI)
1205 for (MachineBasicBlock::iterator I = (*SI)->begin(), E = (*SI)->end();
1206 I != E && I->getOpcode() == X86::PHI; ++I) {
1207 if (RegMap->getRegClass(I->getOperand(0).getReg()) ==
1208 X86::RFPRegisterClass) {
1209 ContainsFPCode = true;
1214 // Final check, check LLVM BB's that are successors to the LLVM BB
1215 // corresponding to BB for FP PHI nodes.
1216 const BasicBlock *LLVMBB = BB->getBasicBlock();
1218 if (!ContainsFPCode)
1219 for (succ_const_iterator SI = succ_begin(LLVMBB), E = succ_end(LLVMBB);
1220 SI != E && !ContainsFPCode; ++SI)
1221 for (BasicBlock::const_iterator II = SI->begin();
1222 (PN = dyn_cast<PHINode>(II)); ++II)
1223 if (PN->getType()->isFloatingPoint()) {
1224 ContainsFPCode = true;
1229 // Insert FP_REG_KILL instructions into basic blocks that need them. This
1230 // only occurs due to the floating point stackifier not being aggressive
1231 // enough to handle arbitrary global stackification.
1233 // Currently we insert an FP_REG_KILL instruction into each block that uses or
1234 // defines a floating point virtual register.
1236 // When the global register allocators (like linear scan) finally update live
1237 // variable analysis, we can keep floating point values in registers across
1238 // basic blocks. This will be a huge win, but we are waiting on the global
1239 // allocators before we can do this.
1241 if (ContainsFPCode) {
1242 BuildMI(*BB, BB->getFirstTerminator(), X86::FP_REG_KILL, 0);
1246 // Clear state used for selection.
1248 RegPressureMap.clear();
1252 // ComputeRegPressure - Compute the RegPressureMap, which is an approximation
1253 // for the number of registers required to compute each node. This is basically
1254 // computing a generalized form of the Sethi-Ullman number for each node.
1255 unsigned ISel::ComputeRegPressure(SDOperand O) {
1257 unsigned &Result = RegPressureMap[N];
1258 if (Result) return Result;
1260 // FIXME: Should operations like CALL (which clobber lots o regs) have a
1261 // higher fixed cost??
1263 if (N->getNumOperands() == 0) {
1266 unsigned MaxRegUse = 0;
1267 unsigned NumExtraMaxRegUsers = 0;
1268 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1270 if (N->getOperand(i).getOpcode() == ISD::Constant)
1273 Regs = ComputeRegPressure(N->getOperand(i));
1274 if (Regs > MaxRegUse) {
1276 NumExtraMaxRegUsers = 0;
1277 } else if (Regs == MaxRegUse &&
1278 N->getOperand(i).getValueType() != MVT::Other) {
1279 ++NumExtraMaxRegUsers;
1283 if (O.getOpcode() != ISD::TokenFactor)
1284 Result = MaxRegUse+NumExtraMaxRegUsers;
1286 Result = MaxRegUse == 1 ? 0 : MaxRegUse-1;
1289 //std::cerr << " WEIGHT: " << Result << " "; N->dump(); std::cerr << "\n";
1293 /// NodeTransitivelyUsesValue - Return true if N or any of its uses uses Op.
1294 /// The DAG cannot have cycles in it, by definition, so the visited set is not
1295 /// needed to prevent infinite loops. The DAG CAN, however, have unbounded
1296 /// reuse, so it prevents exponential cases.
1298 static bool NodeTransitivelyUsesValue(SDOperand N, SDOperand Op,
1299 std::set<SDNode*> &Visited) {
1300 if (N == Op) return true; // Found it.
1301 SDNode *Node = N.Val;
1302 if (Node->getNumOperands() == 0 || // Leaf?
1303 Node->getNodeDepth() <= Op.getNodeDepth()) return false; // Can't find it?
1304 if (!Visited.insert(Node).second) return false; // Already visited?
1306 // Recurse for the first N-1 operands.
1307 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
1308 if (NodeTransitivelyUsesValue(Node->getOperand(i), Op, Visited))
1311 // Tail recurse for the last operand.
1312 return NodeTransitivelyUsesValue(Node->getOperand(0), Op, Visited);
1315 X86AddressMode ISel::SelectAddrExprs(const X86ISelAddressMode &IAM) {
1316 X86AddressMode Result;
1318 // If we need to emit two register operands, emit the one with the highest
1319 // register pressure first.
1320 if (IAM.BaseType == X86ISelAddressMode::RegBase &&
1321 IAM.Base.Reg.Val && IAM.IndexReg.Val) {
1322 bool EmitBaseThenIndex;
1323 if (getRegPressure(IAM.Base.Reg) > getRegPressure(IAM.IndexReg)) {
1324 std::set<SDNode*> Visited;
1325 EmitBaseThenIndex = true;
1326 // If Base ends up pointing to Index, we must emit index first. This is
1327 // because of the way we fold loads, we may end up doing bad things with
1329 if (NodeTransitivelyUsesValue(IAM.Base.Reg, IAM.IndexReg, Visited))
1330 EmitBaseThenIndex = false;
1332 std::set<SDNode*> Visited;
1333 EmitBaseThenIndex = false;
1334 // If Base ends up pointing to Index, we must emit index first. This is
1335 // because of the way we fold loads, we may end up doing bad things with
1337 if (NodeTransitivelyUsesValue(IAM.IndexReg, IAM.Base.Reg, Visited))
1338 EmitBaseThenIndex = true;
1341 if (EmitBaseThenIndex) {
1342 Result.Base.Reg = SelectExpr(IAM.Base.Reg);
1343 Result.IndexReg = SelectExpr(IAM.IndexReg);
1345 Result.IndexReg = SelectExpr(IAM.IndexReg);
1346 Result.Base.Reg = SelectExpr(IAM.Base.Reg);
1349 } else if (IAM.BaseType == X86ISelAddressMode::RegBase && IAM.Base.Reg.Val) {
1350 Result.Base.Reg = SelectExpr(IAM.Base.Reg);
1351 } else if (IAM.IndexReg.Val) {
1352 Result.IndexReg = SelectExpr(IAM.IndexReg);
1355 switch (IAM.BaseType) {
1356 case X86ISelAddressMode::RegBase:
1357 Result.BaseType = X86AddressMode::RegBase;
1359 case X86ISelAddressMode::FrameIndexBase:
1360 Result.BaseType = X86AddressMode::FrameIndexBase;
1361 Result.Base.FrameIndex = IAM.Base.FrameIndex;
1364 assert(0 && "Unknown base type!");
1367 Result.Scale = IAM.Scale;
1368 Result.Disp = IAM.Disp;
1373 /// SelectAddress - Pattern match the maximal addressing mode for this node and
1374 /// emit all of the leaf registers.
1375 void ISel::SelectAddress(SDOperand N, X86AddressMode &AM) {
1376 X86ISelAddressMode IAM;
1377 MatchAddress(N, IAM);
1378 AM = SelectAddrExprs(IAM);
1381 /// MatchAddress - Add the specified node to the specified addressing mode,
1382 /// returning true if it cannot be done. This just pattern matches for the
1383 /// addressing mode, it does not cause any code to be emitted. For that, use
1385 bool ISel::MatchAddress(SDOperand N, X86ISelAddressMode &AM) {
1386 switch (N.getOpcode()) {
1388 case ISD::FrameIndex:
1389 if (AM.BaseType == X86ISelAddressMode::RegBase && AM.Base.Reg.Val == 0) {
1390 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
1391 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
1395 case ISD::GlobalAddress:
1397 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
1398 // For Darwin, external and weak symbols are indirect, so we want to load
1399 // the value at address GV, not the value of GV itself. This means that
1400 // the GlobalAddress must be in the base or index register of the address,
1401 // not the GV offset field.
1402 if (Subtarget->getIndirectExternAndWeakGlobals() &&
1403 (GV->hasWeakLinkage() || GV->isExternal())) {
1412 AM.Disp += cast<ConstantSDNode>(N)->getValue();
1415 // We might have folded the load into this shift, so don't regen the value
1417 if (ExprMap.count(N)) break;
1419 if (AM.IndexReg.Val == 0 && AM.Scale == 1)
1420 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1))) {
1421 unsigned Val = CN->getValue();
1422 if (Val == 1 || Val == 2 || Val == 3) {
1423 AM.Scale = 1 << Val;
1424 SDOperand ShVal = N.Val->getOperand(0);
1426 // Okay, we know that we have a scale by now. However, if the scaled
1427 // value is an add of something and a constant, we can fold the
1428 // constant into the disp field here.
1429 if (ShVal.Val->getOpcode() == ISD::ADD && ShVal.hasOneUse() &&
1430 isa<ConstantSDNode>(ShVal.Val->getOperand(1))) {
1431 AM.IndexReg = ShVal.Val->getOperand(0);
1432 ConstantSDNode *AddVal =
1433 cast<ConstantSDNode>(ShVal.Val->getOperand(1));
1434 AM.Disp += AddVal->getValue() << Val;
1436 AM.IndexReg = ShVal;
1443 // We might have folded the load into this mul, so don't regen the value if
1445 if (ExprMap.count(N)) break;
1447 // X*[3,5,9] -> X+X*[2,4,8]
1448 if (AM.IndexReg.Val == 0 && AM.BaseType == X86ISelAddressMode::RegBase &&
1449 AM.Base.Reg.Val == 0)
1450 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1)))
1451 if (CN->getValue() == 3 || CN->getValue() == 5 || CN->getValue() == 9) {
1452 AM.Scale = unsigned(CN->getValue())-1;
1454 SDOperand MulVal = N.Val->getOperand(0);
1457 // Okay, we know that we have a scale by now. However, if the scaled
1458 // value is an add of something and a constant, we can fold the
1459 // constant into the disp field here.
1460 if (MulVal.Val->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
1461 isa<ConstantSDNode>(MulVal.Val->getOperand(1))) {
1462 Reg = MulVal.Val->getOperand(0);
1463 ConstantSDNode *AddVal =
1464 cast<ConstantSDNode>(MulVal.Val->getOperand(1));
1465 AM.Disp += AddVal->getValue() * CN->getValue();
1467 Reg = N.Val->getOperand(0);
1470 AM.IndexReg = AM.Base.Reg = Reg;
1476 // We might have folded the load into this mul, so don't regen the value if
1478 if (ExprMap.count(N)) break;
1480 X86ISelAddressMode Backup = AM;
1481 if (!MatchAddress(N.Val->getOperand(0), AM) &&
1482 !MatchAddress(N.Val->getOperand(1), AM))
1485 if (!MatchAddress(N.Val->getOperand(1), AM) &&
1486 !MatchAddress(N.Val->getOperand(0), AM))
1493 // Is the base register already occupied?
1494 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.Val) {
1495 // If so, check to see if the scale index register is set.
1496 if (AM.IndexReg.Val == 0) {
1502 // Otherwise, we cannot select it.
1506 // Default, generate it as a register.
1507 AM.BaseType = X86ISelAddressMode::RegBase;
1512 /// Emit2SetCCsAndLogical - Emit the following sequence of instructions,
1513 /// assuming that the temporary registers are in the 8-bit register class.
1517 /// DestReg = logicalop Tmp1, Tmp2
1519 static void Emit2SetCCsAndLogical(MachineBasicBlock *BB, unsigned SetCC1,
1520 unsigned SetCC2, unsigned LogicalOp,
1522 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
1523 unsigned Tmp1 = RegMap->createVirtualRegister(X86::R8RegisterClass);
1524 unsigned Tmp2 = RegMap->createVirtualRegister(X86::R8RegisterClass);
1525 BuildMI(BB, SetCC1, 0, Tmp1);
1526 BuildMI(BB, SetCC2, 0, Tmp2);
1527 BuildMI(BB, LogicalOp, 2, DestReg).addReg(Tmp1).addReg(Tmp2);
1530 /// EmitSetCC - Emit the code to set the specified 8-bit register to 1 if the
1531 /// condition codes match the specified SetCCOpcode. Note that some conditions
1532 /// require multiple instructions to generate the correct value.
1533 static void EmitSetCC(MachineBasicBlock *BB, unsigned DestReg,
1534 ISD::CondCode SetCCOpcode, bool isFP) {
1537 switch (SetCCOpcode) {
1538 default: assert(0 && "Illegal integer SetCC!");
1539 case ISD::SETEQ: Opc = X86::SETEr; break;
1540 case ISD::SETGT: Opc = X86::SETGr; break;
1541 case ISD::SETGE: Opc = X86::SETGEr; break;
1542 case ISD::SETLT: Opc = X86::SETLr; break;
1543 case ISD::SETLE: Opc = X86::SETLEr; break;
1544 case ISD::SETNE: Opc = X86::SETNEr; break;
1545 case ISD::SETULT: Opc = X86::SETBr; break;
1546 case ISD::SETUGT: Opc = X86::SETAr; break;
1547 case ISD::SETULE: Opc = X86::SETBEr; break;
1548 case ISD::SETUGE: Opc = X86::SETAEr; break;
1551 // On a floating point condition, the flags are set as follows:
1553 // 0 | 0 | 0 | X > Y
1554 // 0 | 0 | 1 | X < Y
1555 // 1 | 0 | 0 | X == Y
1556 // 1 | 1 | 1 | unordered
1558 switch (SetCCOpcode) {
1559 default: assert(0 && "Invalid FP setcc!");
1562 Opc = X86::SETEr; // True if ZF = 1
1566 Opc = X86::SETAr; // True if CF = 0 and ZF = 0
1570 Opc = X86::SETAEr; // True if CF = 0
1574 Opc = X86::SETBr; // True if CF = 1
1578 Opc = X86::SETBEr; // True if CF = 1 or ZF = 1
1582 Opc = X86::SETNEr; // True if ZF = 0
1585 Opc = X86::SETPr; // True if PF = 1
1588 Opc = X86::SETNPr; // True if PF = 0
1590 case ISD::SETOEQ: // !PF & ZF
1591 Emit2SetCCsAndLogical(BB, X86::SETNPr, X86::SETEr, X86::AND8rr, DestReg);
1593 case ISD::SETOLT: // !PF & CF
1594 Emit2SetCCsAndLogical(BB, X86::SETNPr, X86::SETBr, X86::AND8rr, DestReg);
1596 case ISD::SETOLE: // !PF & (CF || ZF)
1597 Emit2SetCCsAndLogical(BB, X86::SETNPr, X86::SETBEr, X86::AND8rr, DestReg);
1599 case ISD::SETUGT: // PF | (!ZF & !CF)
1600 Emit2SetCCsAndLogical(BB, X86::SETPr, X86::SETAr, X86::OR8rr, DestReg);
1602 case ISD::SETUGE: // PF | !CF
1603 Emit2SetCCsAndLogical(BB, X86::SETPr, X86::SETAEr, X86::OR8rr, DestReg);
1605 case ISD::SETUNE: // PF | !ZF
1606 Emit2SetCCsAndLogical(BB, X86::SETPr, X86::SETNEr, X86::OR8rr, DestReg);
1610 BuildMI(BB, Opc, 0, DestReg);
1614 /// EmitBranchCC - Emit code into BB that arranges for control to transfer to
1615 /// the Dest block if the Cond condition is true. If we cannot fold this
1616 /// condition into the branch, return true.
1618 bool ISel::EmitBranchCC(MachineBasicBlock *Dest, SDOperand Chain,
1620 // FIXME: Evaluate whether it would be good to emit code like (X < Y) | (A >
1621 // B) using two conditional branches instead of one condbr, two setcc's, and
1623 if ((Cond.getOpcode() == ISD::OR ||
1624 Cond.getOpcode() == ISD::AND) && Cond.Val->hasOneUse()) {
1625 // And and or set the flags for us, so there is no need to emit a TST of the
1626 // result. It is only safe to do this if there is only a single use of the
1627 // AND/OR though, otherwise we don't know it will be emitted here.
1630 BuildMI(BB, X86::JNE, 1).addMBB(Dest);
1634 // Codegen br not C -> JE.
1635 if (Cond.getOpcode() == ISD::XOR)
1636 if (ConstantSDNode *NC = dyn_cast<ConstantSDNode>(Cond.Val->getOperand(1)))
1637 if (NC->isAllOnesValue()) {
1639 if (getRegPressure(Chain) > getRegPressure(Cond)) {
1641 CondR = SelectExpr(Cond.Val->getOperand(0));
1643 CondR = SelectExpr(Cond.Val->getOperand(0));
1646 BuildMI(BB, X86::TEST8rr, 2).addReg(CondR).addReg(CondR);
1647 BuildMI(BB, X86::JE, 1).addMBB(Dest);
1651 if (Cond.getOpcode() != ISD::SETCC)
1652 return true; // Can only handle simple setcc's so far.
1653 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
1657 // Handle integer conditions first.
1658 if (MVT::isInteger(Cond.getOperand(0).getValueType())) {
1660 default: assert(0 && "Illegal integer SetCC!");
1661 case ISD::SETEQ: Opc = X86::JE; break;
1662 case ISD::SETGT: Opc = X86::JG; break;
1663 case ISD::SETGE: Opc = X86::JGE; break;
1664 case ISD::SETLT: Opc = X86::JL; break;
1665 case ISD::SETLE: Opc = X86::JLE; break;
1666 case ISD::SETNE: Opc = X86::JNE; break;
1667 case ISD::SETULT: Opc = X86::JB; break;
1668 case ISD::SETUGT: Opc = X86::JA; break;
1669 case ISD::SETULE: Opc = X86::JBE; break;
1670 case ISD::SETUGE: Opc = X86::JAE; break;
1673 EmitCMP(Cond.getOperand(0), Cond.getOperand(1), Cond.hasOneUse());
1674 BuildMI(BB, Opc, 1).addMBB(Dest);
1678 unsigned Opc2 = 0; // Second branch if needed.
1680 // On a floating point condition, the flags are set as follows:
1682 // 0 | 0 | 0 | X > Y
1683 // 0 | 0 | 1 | X < Y
1684 // 1 | 0 | 0 | X == Y
1685 // 1 | 1 | 1 | unordered
1688 default: assert(0 && "Invalid FP setcc!");
1690 case ISD::SETEQ: Opc = X86::JE; break; // True if ZF = 1
1692 case ISD::SETGT: Opc = X86::JA; break; // True if CF = 0 and ZF = 0
1694 case ISD::SETGE: Opc = X86::JAE; break; // True if CF = 0
1696 case ISD::SETLT: Opc = X86::JB; break; // True if CF = 1
1698 case ISD::SETLE: Opc = X86::JBE; break; // True if CF = 1 or ZF = 1
1700 case ISD::SETNE: Opc = X86::JNE; break; // True if ZF = 0
1701 case ISD::SETUO: Opc = X86::JP; break; // True if PF = 1
1702 case ISD::SETO: Opc = X86::JNP; break; // True if PF = 0
1703 case ISD::SETUGT: // PF = 1 | (ZF = 0 & CF = 0)
1704 Opc = X86::JA; // ZF = 0 & CF = 0
1705 Opc2 = X86::JP; // PF = 1
1707 case ISD::SETUGE: // PF = 1 | CF = 0
1708 Opc = X86::JAE; // CF = 0
1709 Opc2 = X86::JP; // PF = 1
1711 case ISD::SETUNE: // PF = 1 | ZF = 0
1712 Opc = X86::JNE; // ZF = 0
1713 Opc2 = X86::JP; // PF = 1
1715 case ISD::SETOEQ: // PF = 0 & ZF = 1
1718 return true; // FIXME: Emit more efficient code for this branch.
1719 case ISD::SETOLT: // PF = 0 & CF = 1
1722 return true; // FIXME: Emit more efficient code for this branch.
1723 case ISD::SETOLE: // PF = 0 & (CF = 1 || ZF = 1)
1724 //X86::JNP, X86::JBE
1726 return true; // FIXME: Emit more efficient code for this branch.
1730 EmitCMP(Cond.getOperand(0), Cond.getOperand(1), Cond.hasOneUse());
1731 BuildMI(BB, Opc, 1).addMBB(Dest);
1733 BuildMI(BB, Opc2, 1).addMBB(Dest);
1737 /// EmitSelectCC - Emit code into BB that performs a select operation between
1738 /// the two registers RTrue and RFalse, generating a result into RDest.
1740 void ISel::EmitSelectCC(SDOperand Cond, SDOperand True, SDOperand False,
1741 MVT::ValueType SVT, unsigned RDest) {
1742 unsigned RTrue, RFalse;
1744 EQ, NE, LT, LE, GT, GE, B, BE, A, AE, P, NP,
1746 } CondCode = NOT_SET;
1748 static const unsigned CMOVTAB16[] = {
1749 X86::CMOVE16rr, X86::CMOVNE16rr, X86::CMOVL16rr, X86::CMOVLE16rr,
1750 X86::CMOVG16rr, X86::CMOVGE16rr, X86::CMOVB16rr, X86::CMOVBE16rr,
1751 X86::CMOVA16rr, X86::CMOVAE16rr, X86::CMOVP16rr, X86::CMOVNP16rr,
1753 static const unsigned CMOVTAB32[] = {
1754 X86::CMOVE32rr, X86::CMOVNE32rr, X86::CMOVL32rr, X86::CMOVLE32rr,
1755 X86::CMOVG32rr, X86::CMOVGE32rr, X86::CMOVB32rr, X86::CMOVBE32rr,
1756 X86::CMOVA32rr, X86::CMOVAE32rr, X86::CMOVP32rr, X86::CMOVNP32rr,
1758 static const unsigned CMOVTABFP[] = {
1759 X86::FCMOVE , X86::FCMOVNE, /*missing*/0, /*missing*/0,
1760 /*missing*/0, /*missing*/0, X86::FCMOVB , X86::FCMOVBE,
1761 X86::FCMOVA , X86::FCMOVAE, X86::FCMOVP , X86::FCMOVNP
1763 static const int SSE_CMOVTAB[] = {
1764 /*CMPEQ*/ 0, /*CMPNEQ*/ 4, /*missing*/ 0, /*missing*/ 0,
1765 /*missing*/ 0, /*missing*/ 0, /*CMPLT*/ 1, /*CMPLE*/ 2,
1766 /*CMPNLE*/ 6, /*CMPNLT*/ 5, /*CMPUNORD*/ 3, /*CMPORD*/ 7
1769 if (Cond.getOpcode() == ISD::SETCC) {
1770 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
1771 if (MVT::isInteger(Cond.getOperand(0).getValueType())) {
1773 default: assert(0 && "Unknown integer comparison!");
1774 case ISD::SETEQ: CondCode = EQ; break;
1775 case ISD::SETGT: CondCode = GT; break;
1776 case ISD::SETGE: CondCode = GE; break;
1777 case ISD::SETLT: CondCode = LT; break;
1778 case ISD::SETLE: CondCode = LE; break;
1779 case ISD::SETNE: CondCode = NE; break;
1780 case ISD::SETULT: CondCode = B; break;
1781 case ISD::SETUGT: CondCode = A; break;
1782 case ISD::SETULE: CondCode = BE; break;
1783 case ISD::SETUGE: CondCode = AE; break;
1786 // On a floating point condition, the flags are set as follows:
1788 // 0 | 0 | 0 | X > Y
1789 // 0 | 0 | 1 | X < Y
1790 // 1 | 0 | 0 | X == Y
1791 // 1 | 1 | 1 | unordered
1794 default: assert(0 && "Unknown FP comparison!");
1796 case ISD::SETEQ: CondCode = EQ; break; // True if ZF = 1
1798 case ISD::SETGT: CondCode = A; break; // True if CF = 0 and ZF = 0
1800 case ISD::SETGE: CondCode = AE; break; // True if CF = 0
1802 case ISD::SETLT: CondCode = B; break; // True if CF = 1
1804 case ISD::SETLE: CondCode = BE; break; // True if CF = 1 or ZF = 1
1806 case ISD::SETNE: CondCode = NE; break; // True if ZF = 0
1807 case ISD::SETUO: CondCode = P; break; // True if PF = 1
1808 case ISD::SETO: CondCode = NP; break; // True if PF = 0
1809 case ISD::SETUGT: // PF = 1 | (ZF = 0 & CF = 0)
1810 case ISD::SETUGE: // PF = 1 | CF = 0
1811 case ISD::SETUNE: // PF = 1 | ZF = 0
1812 case ISD::SETOEQ: // PF = 0 & ZF = 1
1813 case ISD::SETOLT: // PF = 0 & CF = 1
1814 case ISD::SETOLE: // PF = 0 & (CF = 1 || ZF = 1)
1815 // We cannot emit this comparison as a single cmov.
1821 // There's no SSE equivalent of FCMOVE. For cases where we set a condition
1822 // code above and one of the results of the select is +0.0, then we can fake
1823 // it up through a clever AND with mask. Otherwise, we will fall through to
1824 // the code below that will use a PHI node to select the right value.
1825 if (X86ScalarSSE && (SVT == MVT::f32 || SVT == MVT::f64)) {
1826 if (Cond.getOperand(0).getValueType() == SVT &&
1827 NOT_SET != CondCode) {
1828 ConstantFPSDNode *CT = dyn_cast<ConstantFPSDNode>(True);
1829 ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(False);
1830 bool TrueZero = CT && CT->isExactlyValue(0.0);
1831 bool FalseZero = CF && CF->isExactlyValue(0.0);
1832 if (TrueZero || FalseZero) {
1833 SDOperand LHS = Cond.getOperand(0);
1834 SDOperand RHS = Cond.getOperand(1);
1836 // Select the two halves of the condition
1837 unsigned RLHS, RRHS;
1838 if (getRegPressure(LHS) > getRegPressure(RHS)) {
1839 RLHS = SelectExpr(LHS);
1840 RRHS = SelectExpr(RHS);
1842 RRHS = SelectExpr(RHS);
1843 RLHS = SelectExpr(LHS);
1846 // Emit the comparison and generate a mask from it
1847 unsigned MaskReg = MakeReg(SVT);
1848 unsigned Opc = (SVT == MVT::f32) ? X86::CMPSSrr : X86::CMPSDrr;
1849 BuildMI(BB, Opc, 3, MaskReg).addReg(RLHS).addReg(RRHS)
1850 .addImm(SSE_CMOVTAB[CondCode]);
1853 RFalse = SelectExpr(False);
1854 Opc = (SVT == MVT::f32) ? X86::ANDNPSrr : X86::ANDNPDrr;
1855 BuildMI(BB, Opc, 2, RDest).addReg(MaskReg).addReg(RFalse);
1857 RTrue = SelectExpr(True);
1858 Opc = (SVT == MVT::f32) ? X86::ANDPSrr : X86::ANDPDrr;
1859 BuildMI(BB, Opc, 2, RDest).addReg(MaskReg).addReg(RTrue);
1867 // Select the true and false values for use in both the SSE PHI case, and the
1868 // integer or x87 cmov cases below.
1869 if (getRegPressure(True) > getRegPressure(False)) {
1870 RTrue = SelectExpr(True);
1871 RFalse = SelectExpr(False);
1873 RFalse = SelectExpr(False);
1874 RTrue = SelectExpr(True);
1877 // Since there's no SSE equivalent of FCMOVE, and we couldn't generate an
1878 // AND with mask, we'll have to do the normal RISC thing and generate a PHI
1879 // node to select between the true and false values.
1880 if (X86ScalarSSE && (SVT == MVT::f32 || SVT == MVT::f64)) {
1881 // FIXME: emit a direct compare and branch rather than setting a cond reg
1883 unsigned CondReg = SelectExpr(Cond);
1884 BuildMI(BB, X86::TEST8rr, 2).addReg(CondReg).addReg(CondReg);
1886 // Create an iterator with which to insert the MBB for copying the false
1887 // value and the MBB to hold the PHI instruction for this SetCC.
1888 MachineBasicBlock *thisMBB = BB;
1889 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1890 ilist<MachineBasicBlock>::iterator It = BB;
1896 // cmpTY ccX, r1, r2
1898 // fallthrough --> copy0MBB
1899 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1900 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1901 BuildMI(BB, X86::JNE, 1).addMBB(sinkMBB);
1902 MachineFunction *F = BB->getParent();
1903 F->getBasicBlockList().insert(It, copy0MBB);
1904 F->getBasicBlockList().insert(It, sinkMBB);
1905 // Update machine-CFG edges
1906 BB->addSuccessor(copy0MBB);
1907 BB->addSuccessor(sinkMBB);
1910 // %FalseValue = ...
1911 // # fallthrough to sinkMBB
1913 // Update machine-CFG edges
1914 BB->addSuccessor(sinkMBB);
1917 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1920 BuildMI(BB, X86::PHI, 4, RDest).addReg(RFalse)
1921 .addMBB(copy0MBB).addReg(RTrue).addMBB(thisMBB);
1926 if (CondCode != NOT_SET) {
1928 default: assert(0 && "Cannot select this type!");
1929 case MVT::i16: Opc = CMOVTAB16[CondCode]; break;
1930 case MVT::i32: Opc = CMOVTAB32[CondCode]; break;
1931 case MVT::f64: Opc = CMOVTABFP[CondCode]; break;
1935 // Finally, if we weren't able to fold this, just emit the condition and test
1937 if (CondCode == NOT_SET || Opc == 0) {
1938 // Get the condition into the zero flag.
1939 unsigned CondReg = SelectExpr(Cond);
1940 BuildMI(BB, X86::TEST8rr, 2).addReg(CondReg).addReg(CondReg);
1943 default: assert(0 && "Cannot select this type!");
1944 case MVT::i16: Opc = X86::CMOVE16rr; break;
1945 case MVT::i32: Opc = X86::CMOVE32rr; break;
1946 case MVT::f64: Opc = X86::FCMOVE; break;
1949 // FIXME: CMP R, 0 -> TEST R, R
1950 EmitCMP(Cond.getOperand(0), Cond.getOperand(1), Cond.Val->hasOneUse());
1951 std::swap(RTrue, RFalse);
1953 BuildMI(BB, Opc, 2, RDest).addReg(RTrue).addReg(RFalse);
1956 void ISel::EmitCMP(SDOperand LHS, SDOperand RHS, bool HasOneUse) {
1958 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(RHS)) {
1960 if (HasOneUse && isFoldableLoad(LHS, RHS)) {
1961 switch (RHS.getValueType()) {
1964 case MVT::i8: Opc = X86::CMP8mi; break;
1965 case MVT::i16: Opc = X86::CMP16mi; break;
1966 case MVT::i32: Opc = X86::CMP32mi; break;
1970 EmitFoldedLoad(LHS, AM);
1971 addFullAddress(BuildMI(BB, Opc, 5), AM).addImm(CN->getValue());
1976 switch (RHS.getValueType()) {
1979 case MVT::i8: Opc = X86::CMP8ri; break;
1980 case MVT::i16: Opc = X86::CMP16ri; break;
1981 case MVT::i32: Opc = X86::CMP32ri; break;
1984 unsigned Tmp1 = SelectExpr(LHS);
1985 BuildMI(BB, Opc, 2).addReg(Tmp1).addImm(CN->getValue());
1988 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(RHS)) {
1989 if (!X86ScalarSSE && (CN->isExactlyValue(+0.0) ||
1990 CN->isExactlyValue(-0.0))) {
1991 unsigned Reg = SelectExpr(LHS);
1992 BuildMI(BB, X86::FTST, 1).addReg(Reg);
1993 BuildMI(BB, X86::FNSTSW8r, 0);
1994 BuildMI(BB, X86::SAHF, 1);
2000 if (HasOneUse && isFoldableLoad(LHS, RHS)) {
2001 switch (RHS.getValueType()) {
2004 case MVT::i8: Opc = X86::CMP8mr; break;
2005 case MVT::i16: Opc = X86::CMP16mr; break;
2006 case MVT::i32: Opc = X86::CMP32mr; break;
2010 EmitFoldedLoad(LHS, AM);
2011 unsigned Reg = SelectExpr(RHS);
2012 addFullAddress(BuildMI(BB, Opc, 5), AM).addReg(Reg);
2017 switch (LHS.getValueType()) {
2018 default: assert(0 && "Cannot compare this value!");
2020 case MVT::i8: Opc = X86::CMP8rr; break;
2021 case MVT::i16: Opc = X86::CMP16rr; break;
2022 case MVT::i32: Opc = X86::CMP32rr; break;
2023 case MVT::f32: Opc = X86::UCOMISSrr; break;
2024 case MVT::f64: Opc = X86ScalarSSE ? X86::UCOMISDrr : X86::FUCOMIr; break;
2026 unsigned Tmp1, Tmp2;
2027 if (getRegPressure(LHS) > getRegPressure(RHS)) {
2028 Tmp1 = SelectExpr(LHS);
2029 Tmp2 = SelectExpr(RHS);
2031 Tmp2 = SelectExpr(RHS);
2032 Tmp1 = SelectExpr(LHS);
2034 BuildMI(BB, Opc, 2).addReg(Tmp1).addReg(Tmp2);
2037 /// isFoldableLoad - Return true if this is a load instruction that can safely
2038 /// be folded into an operation that uses it.
2039 bool ISel::isFoldableLoad(SDOperand Op, SDOperand OtherOp, bool FloatPromoteOk){
2040 if (Op.getOpcode() == ISD::LOAD) {
2041 // FIXME: currently can't fold constant pool indexes.
2042 if (isa<ConstantPoolSDNode>(Op.getOperand(1)))
2044 } else if (FloatPromoteOk && Op.getOpcode() == ISD::EXTLOAD &&
2045 cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::f32) {
2046 // FIXME: currently can't fold constant pool indexes.
2047 if (isa<ConstantPoolSDNode>(Op.getOperand(1)))
2053 // If this load has already been emitted, we clearly can't fold it.
2054 assert(Op.ResNo == 0 && "Not a use of the value of the load?");
2055 if (ExprMap.count(Op.getValue(1))) return false;
2056 assert(!ExprMap.count(Op.getValue(0)) && "Value in map but not token chain?");
2057 assert(!ExprMap.count(Op.getValue(1))&&"Token lowered but value not in map?");
2059 // If there is not just one use of its value, we cannot fold.
2060 if (!Op.Val->hasNUsesOfValue(1, 0)) return false;
2062 // Finally, we cannot fold the load into the operation if this would induce a
2063 // cycle into the resultant dag. To check for this, see if OtherOp (the other
2064 // operand of the operation we are folding the load into) can possible use the
2065 // chain node defined by the load.
2066 if (OtherOp.Val && !Op.Val->hasNUsesOfValue(0, 1)) { // Has uses of chain?
2067 std::set<SDNode*> Visited;
2068 if (NodeTransitivelyUsesValue(OtherOp, Op.getValue(1), Visited))
2075 /// EmitFoldedLoad - Ensure that the arguments of the load are code generated,
2076 /// and compute the address being loaded into AM.
2077 void ISel::EmitFoldedLoad(SDOperand Op, X86AddressMode &AM) {
2078 SDOperand Chain = Op.getOperand(0);
2079 SDOperand Address = Op.getOperand(1);
2081 if (getRegPressure(Chain) > getRegPressure(Address)) {
2083 SelectAddress(Address, AM);
2085 SelectAddress(Address, AM);
2089 // The chain for this load is now lowered.
2090 assert(ExprMap.count(SDOperand(Op.Val, 1)) == 0 &&
2091 "Load emitted more than once?");
2092 if (!ExprMap.insert(std::make_pair(Op.getValue(1), 1)).second)
2093 assert(0 && "Load emitted more than once!");
2096 // EmitOrOpOp - Pattern match the expression (Op1|Op2), where we know that op1
2097 // and op2 are i8/i16/i32 values with one use each (the or). If we can form a
2098 // SHLD or SHRD, emit the instruction (generating the value into DestReg) and
2100 bool ISel::EmitOrOpOp(SDOperand Op1, SDOperand Op2, unsigned DestReg) {
2101 if (Op1.getOpcode() == ISD::SHL && Op2.getOpcode() == ISD::SRL) {
2103 } else if (Op2.getOpcode() == ISD::SHL && Op1.getOpcode() == ISD::SRL) {
2104 std::swap(Op1, Op2); // Op1 is the SHL now.
2106 return false; // No match
2109 SDOperand ShlVal = Op1.getOperand(0);
2110 SDOperand ShlAmt = Op1.getOperand(1);
2111 SDOperand ShrVal = Op2.getOperand(0);
2112 SDOperand ShrAmt = Op2.getOperand(1);
2114 unsigned RegSize = MVT::getSizeInBits(Op1.getValueType());
2116 // Find out if ShrAmt = 32-ShlAmt or ShlAmt = 32-ShrAmt.
2117 if (ShlAmt.getOpcode() == ISD::SUB && ShlAmt.getOperand(1) == ShrAmt)
2118 if (ConstantSDNode *SubCST = dyn_cast<ConstantSDNode>(ShlAmt.getOperand(0)))
2119 if (SubCST->getValue() == RegSize) {
2120 // (A >> ShrAmt) | (A << (32-ShrAmt)) ==> ROR A, ShrAmt
2121 // (A >> ShrAmt) | (B << (32-ShrAmt)) ==> SHRD A, B, ShrAmt
2122 if (ShrVal == ShlVal) {
2123 unsigned Reg, ShAmt;
2124 if (getRegPressure(ShrVal) > getRegPressure(ShrAmt)) {
2125 Reg = SelectExpr(ShrVal);
2126 ShAmt = SelectExpr(ShrAmt);
2128 ShAmt = SelectExpr(ShrAmt);
2129 Reg = SelectExpr(ShrVal);
2131 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(ShAmt);
2132 unsigned Opc = RegSize == 8 ? X86::ROR8rCL :
2133 (RegSize == 16 ? X86::ROR16rCL : X86::ROR32rCL);
2134 BuildMI(BB, Opc, 1, DestReg).addReg(Reg);
2136 } else if (RegSize != 8) {
2137 unsigned AReg, BReg;
2138 if (getRegPressure(ShlVal) > getRegPressure(ShrVal)) {
2139 BReg = SelectExpr(ShlVal);
2140 AReg = SelectExpr(ShrVal);
2142 AReg = SelectExpr(ShrVal);
2143 BReg = SelectExpr(ShlVal);
2145 unsigned ShAmt = SelectExpr(ShrAmt);
2146 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(ShAmt);
2147 unsigned Opc = RegSize == 16 ? X86::SHRD16rrCL : X86::SHRD32rrCL;
2148 BuildMI(BB, Opc, 2, DestReg).addReg(AReg).addReg(BReg);
2153 if (ShrAmt.getOpcode() == ISD::SUB && ShrAmt.getOperand(1) == ShlAmt)
2154 if (ConstantSDNode *SubCST = dyn_cast<ConstantSDNode>(ShrAmt.getOperand(0)))
2155 if (SubCST->getValue() == RegSize) {
2156 // (A << ShlAmt) | (A >> (32-ShlAmt)) ==> ROL A, ShrAmt
2157 // (A << ShlAmt) | (B >> (32-ShlAmt)) ==> SHLD A, B, ShrAmt
2158 if (ShrVal == ShlVal) {
2159 unsigned Reg, ShAmt;
2160 if (getRegPressure(ShrVal) > getRegPressure(ShlAmt)) {
2161 Reg = SelectExpr(ShrVal);
2162 ShAmt = SelectExpr(ShlAmt);
2164 ShAmt = SelectExpr(ShlAmt);
2165 Reg = SelectExpr(ShrVal);
2167 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(ShAmt);
2168 unsigned Opc = RegSize == 8 ? X86::ROL8rCL :
2169 (RegSize == 16 ? X86::ROL16rCL : X86::ROL32rCL);
2170 BuildMI(BB, Opc, 1, DestReg).addReg(Reg);
2172 } else if (RegSize != 8) {
2173 unsigned AReg, BReg;
2174 if (getRegPressure(ShlVal) > getRegPressure(ShrVal)) {
2175 AReg = SelectExpr(ShlVal);
2176 BReg = SelectExpr(ShrVal);
2178 BReg = SelectExpr(ShrVal);
2179 AReg = SelectExpr(ShlVal);
2181 unsigned ShAmt = SelectExpr(ShlAmt);
2182 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(ShAmt);
2183 unsigned Opc = RegSize == 16 ? X86::SHLD16rrCL : X86::SHLD32rrCL;
2184 BuildMI(BB, Opc, 2, DestReg).addReg(AReg).addReg(BReg);
2189 if (ConstantSDNode *ShrCst = dyn_cast<ConstantSDNode>(ShrAmt))
2190 if (ConstantSDNode *ShlCst = dyn_cast<ConstantSDNode>(ShlAmt))
2191 if (ShrCst->getValue() < RegSize && ShlCst->getValue() < RegSize)
2192 if (ShrCst->getValue() == RegSize-ShlCst->getValue()) {
2193 // (A >> 5) | (A << 27) --> ROR A, 5
2194 // (A >> 5) | (B << 27) --> SHRD A, B, 5
2195 if (ShrVal == ShlVal) {
2196 unsigned Reg = SelectExpr(ShrVal);
2197 unsigned Opc = RegSize == 8 ? X86::ROR8ri :
2198 (RegSize == 16 ? X86::ROR16ri : X86::ROR32ri);
2199 BuildMI(BB, Opc, 2, DestReg).addReg(Reg).addImm(ShrCst->getValue());
2201 } else if (RegSize != 8) {
2202 unsigned AReg, BReg;
2203 if (getRegPressure(ShlVal) > getRegPressure(ShrVal)) {
2204 BReg = SelectExpr(ShlVal);
2205 AReg = SelectExpr(ShrVal);
2207 AReg = SelectExpr(ShrVal);
2208 BReg = SelectExpr(ShlVal);
2210 unsigned Opc = RegSize == 16 ? X86::SHRD16rri8 : X86::SHRD32rri8;
2211 BuildMI(BB, Opc, 3, DestReg).addReg(AReg).addReg(BReg)
2212 .addImm(ShrCst->getValue());
2220 unsigned ISel::SelectExpr(SDOperand N) {
2222 unsigned Tmp1 = 0, Tmp2 = 0, Tmp3 = 0, Opc = 0;
2223 SDNode *Node = N.Val;
2226 if (Node->getOpcode() == ISD::CopyFromReg) {
2227 unsigned Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
2228 // Just use the specified register as our input if we can.
2229 if (MRegisterInfo::isVirtualRegister(Reg) || Reg == X86::ESP)
2233 unsigned &Reg = ExprMap[N];
2234 if (Reg) return Reg;
2236 switch (N.getOpcode()) {
2238 Reg = Result = (N.getValueType() != MVT::Other) ?
2239 MakeReg(N.getValueType()) : 1;
2241 case X86ISD::TAILCALL:
2243 // If this is a call instruction, make sure to prepare ALL of the result
2244 // values as well as the chain.
2245 ExprMap[N.getValue(0)] = 1;
2246 if (Node->getNumValues() > 1) {
2247 Result = MakeReg(Node->getValueType(1));
2248 ExprMap[N.getValue(1)] = Result;
2249 for (unsigned i = 2, e = Node->getNumValues(); i != e; ++i)
2250 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
2255 case ISD::ADD_PARTS:
2256 case ISD::SUB_PARTS:
2257 case ISD::SHL_PARTS:
2258 case ISD::SRL_PARTS:
2259 case ISD::SRA_PARTS:
2260 Result = MakeReg(Node->getValueType(0));
2261 ExprMap[N.getValue(0)] = Result;
2262 for (unsigned i = 1, e = N.Val->getNumValues(); i != e; ++i)
2263 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
2267 switch (N.getOpcode()) {
2270 assert(0 && "Node not handled!\n");
2271 case ISD::FP_EXTEND:
2272 assert(X86ScalarSSE && "Scalar SSE FP must be enabled to use f32");
2273 Tmp1 = SelectExpr(N.getOperand(0));
2274 BuildMI(BB, X86::CVTSS2SDrr, 1, Result).addReg(Tmp1);
2277 assert(X86ScalarSSE && "Scalar SSE FP must be enabled to use f32");
2278 Tmp1 = SelectExpr(N.getOperand(0));
2279 BuildMI(BB, X86::CVTSD2SSrr, 1, Result).addReg(Tmp1);
2281 case ISD::CopyFromReg:
2282 Select(N.getOperand(0));
2284 Reg = Result = ExprMap[N.getValue(0)] =
2285 MakeReg(N.getValue(0).getValueType());
2287 Tmp1 = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
2288 switch (Node->getValueType(0)) {
2289 default: assert(0 && "Cannot CopyFromReg this!");
2292 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(Tmp1);
2295 BuildMI(BB, X86::MOV16rr, 1, Result).addReg(Tmp1);
2298 BuildMI(BB, X86::MOV32rr, 1, Result).addReg(Tmp1);
2302 case ISD::FrameIndex:
2303 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
2304 addFrameReference(BuildMI(BB, X86::LEA32r, 4, Result), (int)Tmp1);
2306 case ISD::ConstantPool:
2307 Tmp1 = BB->getParent()->getConstantPool()->
2308 getConstantPoolIndex(cast<ConstantPoolSDNode>(N)->get());
2309 addConstantPoolReference(BuildMI(BB, X86::LEA32r, 4, Result), Tmp1);
2311 case ISD::ConstantFP:
2313 assert(cast<ConstantFPSDNode>(N)->isExactlyValue(+0.0) &&
2314 "SSE only supports +0.0");
2315 Opc = (N.getValueType() == MVT::f32) ? X86::FLD0SS : X86::FLD0SD;
2316 BuildMI(BB, Opc, 0, Result);
2319 ContainsFPCode = true;
2320 Tmp1 = Result; // Intermediate Register
2321 if (cast<ConstantFPSDNode>(N)->getValue() < 0.0 ||
2322 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
2323 Tmp1 = MakeReg(MVT::f64);
2325 if (cast<ConstantFPSDNode>(N)->isExactlyValue(+0.0) ||
2326 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
2327 BuildMI(BB, X86::FLD0, 0, Tmp1);
2328 else if (cast<ConstantFPSDNode>(N)->isExactlyValue(+1.0) ||
2329 cast<ConstantFPSDNode>(N)->isExactlyValue(-1.0))
2330 BuildMI(BB, X86::FLD1, 0, Tmp1);
2332 assert(0 && "Unexpected constant!");
2334 BuildMI(BB, X86::FCHS, 1, Result).addReg(Tmp1);
2337 switch (N.getValueType()) {
2338 default: assert(0 && "Cannot use constants of this type!");
2340 case MVT::i8: Opc = X86::MOV8ri; break;
2341 case MVT::i16: Opc = X86::MOV16ri; break;
2342 case MVT::i32: Opc = X86::MOV32ri; break;
2344 BuildMI(BB, Opc, 1,Result).addImm(cast<ConstantSDNode>(N)->getValue());
2347 if (Node->getValueType(0) == MVT::f64) {
2348 // FIXME: SHOULD TEACH STACKIFIER ABOUT UNDEF VALUES!
2349 BuildMI(BB, X86::FLD0, 0, Result);
2351 BuildMI(BB, X86::IMPLICIT_DEF, 0, Result);
2354 case ISD::GlobalAddress: {
2355 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
2356 // For Darwin, external and weak symbols are indirect, so we want to load
2357 // the value at address GV, not the value of GV itself.
2358 if (Subtarget->getIndirectExternAndWeakGlobals() &&
2359 (GV->hasWeakLinkage() || GV->isExternal())) {
2360 BuildMI(BB, X86::MOV32rm, 4, Result).addReg(0).addZImm(1).addReg(0)
2361 .addGlobalAddress(GV, false, 0);
2363 BuildMI(BB, X86::MOV32ri, 1, Result).addGlobalAddress(GV);
2367 case ISD::ExternalSymbol: {
2368 const char *Sym = cast<ExternalSymbolSDNode>(N)->getSymbol();
2369 BuildMI(BB, X86::MOV32ri, 1, Result).addExternalSymbol(Sym);
2372 case ISD::ANY_EXTEND: // treat any extend like zext
2373 case ISD::ZERO_EXTEND: {
2374 int DestIs16 = N.getValueType() == MVT::i16;
2375 int SrcIs16 = N.getOperand(0).getValueType() == MVT::i16;
2377 // FIXME: This hack is here for zero extension casts from bool to i8. This
2378 // would not be needed if bools were promoted by Legalize.
2379 if (N.getValueType() == MVT::i8) {
2380 Tmp1 = SelectExpr(N.getOperand(0));
2381 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(Tmp1);
2385 if (isFoldableLoad(N.getOperand(0), SDOperand())) {
2386 static const unsigned Opc[3] = {
2387 X86::MOVZX32rm8, X86::MOVZX32rm16, X86::MOVZX16rm8
2391 EmitFoldedLoad(N.getOperand(0), AM);
2392 addFullAddress(BuildMI(BB, Opc[SrcIs16+DestIs16*2], 4, Result), AM);
2397 static const unsigned Opc[3] = {
2398 X86::MOVZX32rr8, X86::MOVZX32rr16, X86::MOVZX16rr8
2400 Tmp1 = SelectExpr(N.getOperand(0));
2401 BuildMI(BB, Opc[SrcIs16+DestIs16*2], 1, Result).addReg(Tmp1);
2404 case ISD::SIGN_EXTEND: {
2405 int DestIs16 = N.getValueType() == MVT::i16;
2406 int SrcIs16 = N.getOperand(0).getValueType() == MVT::i16;
2408 // FIXME: Legalize should promote bools to i8!
2409 assert(N.getOperand(0).getValueType() != MVT::i1 &&
2410 "Sign extend from bool not implemented!");
2412 if (isFoldableLoad(N.getOperand(0), SDOperand())) {
2413 static const unsigned Opc[3] = {
2414 X86::MOVSX32rm8, X86::MOVSX32rm16, X86::MOVSX16rm8
2418 EmitFoldedLoad(N.getOperand(0), AM);
2419 addFullAddress(BuildMI(BB, Opc[SrcIs16+DestIs16*2], 4, Result), AM);
2423 static const unsigned Opc[3] = {
2424 X86::MOVSX32rr8, X86::MOVSX32rr16, X86::MOVSX16rr8
2426 Tmp1 = SelectExpr(N.getOperand(0));
2427 BuildMI(BB, Opc[SrcIs16+DestIs16*2], 1, Result).addReg(Tmp1);
2431 // Fold TRUNCATE (LOAD P) into a smaller load from P.
2432 // FIXME: This should be performed by the DAGCombiner.
2433 if (isFoldableLoad(N.getOperand(0), SDOperand())) {
2434 switch (N.getValueType()) {
2435 default: assert(0 && "Unknown truncate!");
2437 case MVT::i8: Opc = X86::MOV8rm; break;
2438 case MVT::i16: Opc = X86::MOV16rm; break;
2441 EmitFoldedLoad(N.getOperand(0), AM);
2442 addFullAddress(BuildMI(BB, Opc, 4, Result), AM);
2446 // Handle cast of LARGER int to SMALLER int using a move to EAX followed by
2447 // a move out of AX or AL.
2448 switch (N.getOperand(0).getValueType()) {
2449 default: assert(0 && "Unknown truncate!");
2450 case MVT::i8: Tmp2 = X86::AL; Opc = X86::MOV8rr; break;
2451 case MVT::i16: Tmp2 = X86::AX; Opc = X86::MOV16rr; break;
2452 case MVT::i32: Tmp2 = X86::EAX; Opc = X86::MOV32rr; break;
2454 Tmp1 = SelectExpr(N.getOperand(0));
2455 BuildMI(BB, Opc, 1, Tmp2).addReg(Tmp1);
2457 switch (N.getValueType()) {
2458 default: assert(0 && "Unknown truncate!");
2460 case MVT::i8: Tmp2 = X86::AL; Opc = X86::MOV8rr; break;
2461 case MVT::i16: Tmp2 = X86::AX; Opc = X86::MOV16rr; break;
2463 BuildMI(BB, Opc, 1, Result).addReg(Tmp2);
2466 case ISD::SINT_TO_FP: {
2467 Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
2468 unsigned PromoteOpcode = 0;
2470 // We can handle any sint to fp with the direct sse conversion instructions.
2472 Opc = (N.getValueType() == MVT::f64) ? X86::CVTSI2SDrr : X86::CVTSI2SSrr;
2473 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
2477 ContainsFPCode = true;
2479 // Spill the integer to memory and reload it from there.
2480 MVT::ValueType SrcTy = N.getOperand(0).getValueType();
2481 unsigned Size = MVT::getSizeInBits(SrcTy)/8;
2482 MachineFunction *F = BB->getParent();
2483 int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, Size);
2487 addFrameReference(BuildMI(BB, X86::MOV32mr, 5), FrameIdx).addReg(Tmp1);
2488 addFrameReference(BuildMI(BB, X86::FILD32m, 5, Result), FrameIdx);
2491 addFrameReference(BuildMI(BB, X86::MOV16mr, 5), FrameIdx).addReg(Tmp1);
2492 addFrameReference(BuildMI(BB, X86::FILD16m, 5, Result), FrameIdx);
2494 default: break; // No promotion required.
2498 case ISD::FP_TO_SINT:
2499 Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
2501 // If the target supports SSE2 and is performing FP operations in SSE regs
2502 // instead of the FP stack, then we can use the efficient CVTSS2SI and
2503 // CVTSD2SI instructions.
2504 assert(X86ScalarSSE);
2505 if (MVT::f32 == N.getOperand(0).getValueType()) {
2506 BuildMI(BB, X86::CVTTSS2SIrr, 1, Result).addReg(Tmp1);
2507 } else if (MVT::f64 == N.getOperand(0).getValueType()) {
2508 BuildMI(BB, X86::CVTTSD2SIrr, 1, Result).addReg(Tmp1);
2510 assert(0 && "Not an f32 or f64?");
2517 Op0 = N.getOperand(0);
2518 Op1 = N.getOperand(1);
2520 if (isFoldableLoad(Op0, Op1, true)) {
2521 std::swap(Op0, Op1);
2525 if (isFoldableLoad(Op1, Op0, true)) {
2527 switch (N.getValueType()) {
2528 default: assert(0 && "Cannot add this type!");
2530 case MVT::i8: Opc = X86::ADD8rm; break;
2531 case MVT::i16: Opc = X86::ADD16rm; break;
2532 case MVT::i32: Opc = X86::ADD32rm; break;
2533 case MVT::f32: Opc = X86::ADDSSrm; break;
2535 // For F64, handle promoted load operations (from F32) as well!
2537 assert(Op1.getOpcode() == ISD::LOAD && "SSE load not promoted");
2540 Opc = Op1.getOpcode() == ISD::LOAD ? X86::FADD64m : X86::FADD32m;
2545 EmitFoldedLoad(Op1, AM);
2546 Tmp1 = SelectExpr(Op0);
2547 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
2551 // See if we can codegen this as an LEA to fold operations together.
2552 if (N.getValueType() == MVT::i32) {
2554 X86ISelAddressMode AM;
2555 MatchAddress(N, AM);
2556 ExprMap[N] = Result;
2558 // If this is not just an add, emit the LEA. For a simple add (like
2559 // reg+reg or reg+imm), we just emit an add. It might be a good idea to
2560 // leave this as LEA, then peephole it to 'ADD' after two address elim
2562 if (AM.Scale != 1 || AM.BaseType == X86ISelAddressMode::FrameIndexBase||
2563 AM.GV || (AM.Base.Reg.Val && AM.IndexReg.Val && AM.Disp)) {
2564 X86AddressMode XAM = SelectAddrExprs(AM);
2565 addFullAddress(BuildMI(BB, X86::LEA32r, 4, Result), XAM);
2570 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1)) {
2572 if (CN->getValue() == 1) { // add X, 1 -> inc X
2573 switch (N.getValueType()) {
2574 default: assert(0 && "Cannot integer add this type!");
2575 case MVT::i8: Opc = X86::INC8r; break;
2576 case MVT::i16: Opc = X86::INC16r; break;
2577 case MVT::i32: Opc = X86::INC32r; break;
2579 } else if (CN->isAllOnesValue()) { // add X, -1 -> dec X
2580 switch (N.getValueType()) {
2581 default: assert(0 && "Cannot integer add this type!");
2582 case MVT::i8: Opc = X86::DEC8r; break;
2583 case MVT::i16: Opc = X86::DEC16r; break;
2584 case MVT::i32: Opc = X86::DEC32r; break;
2589 Tmp1 = SelectExpr(Op0);
2590 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
2594 switch (N.getValueType()) {
2595 default: assert(0 && "Cannot add this type!");
2596 case MVT::i8: Opc = X86::ADD8ri; break;
2597 case MVT::i16: Opc = X86::ADD16ri; break;
2598 case MVT::i32: Opc = X86::ADD32ri; break;
2601 Tmp1 = SelectExpr(Op0);
2602 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
2607 switch (N.getValueType()) {
2608 default: assert(0 && "Cannot add this type!");
2609 case MVT::i8: Opc = X86::ADD8rr; break;
2610 case MVT::i16: Opc = X86::ADD16rr; break;
2611 case MVT::i32: Opc = X86::ADD32rr; break;
2612 case MVT::f32: Opc = X86::ADDSSrr; break;
2613 case MVT::f64: Opc = X86ScalarSSE ? X86::ADDSDrr : X86::FpADD; break;
2616 if (getRegPressure(Op0) > getRegPressure(Op1)) {
2617 Tmp1 = SelectExpr(Op0);
2618 Tmp2 = SelectExpr(Op1);
2620 Tmp2 = SelectExpr(Op1);
2621 Tmp1 = SelectExpr(Op0);
2624 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
2628 Tmp1 = SelectExpr(Node->getOperand(0));
2630 Opc = (N.getValueType() == MVT::f32) ? X86::SQRTSSrr : X86::SQRTSDrr;
2631 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
2633 BuildMI(BB, X86::FSQRT, 1, Result).addReg(Tmp1);
2638 // Once we can spill 16 byte constants into the constant pool, we can
2639 // implement SSE equivalents of FABS and FCHS.
2644 assert(N.getValueType()==MVT::f64 && "Illegal type for this operation");
2645 Tmp1 = SelectExpr(Node->getOperand(0));
2646 switch (N.getOpcode()) {
2647 default: assert(0 && "Unreachable!");
2648 case ISD::FABS: BuildMI(BB, X86::FABS, 1, Result).addReg(Tmp1); break;
2649 case ISD::FNEG: BuildMI(BB, X86::FCHS, 1, Result).addReg(Tmp1); break;
2650 case ISD::FSIN: BuildMI(BB, X86::FSIN, 1, Result).addReg(Tmp1); break;
2651 case ISD::FCOS: BuildMI(BB, X86::FCOS, 1, Result).addReg(Tmp1); break;
2656 switch (N.getValueType()) {
2657 default: assert(0 && "Unsupported VT!");
2658 case MVT::i8: Tmp2 = X86::MUL8r; break;
2659 case MVT::i16: Tmp2 = X86::MUL16r; break;
2660 case MVT::i32: Tmp2 = X86::MUL32r; break;
2664 unsigned MovOpc, LowReg, HiReg;
2665 switch (N.getValueType()) {
2666 default: assert(0 && "Unsupported VT!");
2668 MovOpc = X86::MOV8rr;
2674 MovOpc = X86::MOV16rr;
2680 MovOpc = X86::MOV32rr;
2686 if (Node->getOpcode() != ISD::MULHS)
2687 Opc = Tmp2; // Get the MULHU opcode.
2689 Op0 = Node->getOperand(0);
2690 Op1 = Node->getOperand(1);
2691 if (getRegPressure(Op0) > getRegPressure(Op1)) {
2692 Tmp1 = SelectExpr(Op0);
2693 Tmp2 = SelectExpr(Op1);
2695 Tmp2 = SelectExpr(Op1);
2696 Tmp1 = SelectExpr(Op0);
2699 // FIXME: Implement folding of loads into the memory operands here!
2700 BuildMI(BB, MovOpc, 1, LowReg).addReg(Tmp1);
2701 BuildMI(BB, Opc, 1).addReg(Tmp2);
2702 BuildMI(BB, MovOpc, 1, Result).addReg(HiReg);
2713 static const unsigned SUBTab[] = {
2714 X86::SUB8ri, X86::SUB16ri, X86::SUB32ri, 0, 0,
2715 X86::SUB8rm, X86::SUB16rm, X86::SUB32rm, X86::FSUB32m, X86::FSUB64m,
2716 X86::SUB8rr, X86::SUB16rr, X86::SUB32rr, X86::FpSUB , X86::FpSUB,
2718 static const unsigned SSE_SUBTab[] = {
2719 X86::SUB8ri, X86::SUB16ri, X86::SUB32ri, 0, 0,
2720 X86::SUB8rm, X86::SUB16rm, X86::SUB32rm, X86::SUBSSrm, X86::SUBSDrm,
2721 X86::SUB8rr, X86::SUB16rr, X86::SUB32rr, X86::SUBSSrr, X86::SUBSDrr,
2723 static const unsigned MULTab[] = {
2724 0, X86::IMUL16rri, X86::IMUL32rri, 0, 0,
2725 0, X86::IMUL16rm , X86::IMUL32rm, X86::FMUL32m, X86::FMUL64m,
2726 0, X86::IMUL16rr , X86::IMUL32rr, X86::FpMUL , X86::FpMUL,
2728 static const unsigned SSE_MULTab[] = {
2729 0, X86::IMUL16rri, X86::IMUL32rri, 0, 0,
2730 0, X86::IMUL16rm , X86::IMUL32rm, X86::MULSSrm, X86::MULSDrm,
2731 0, X86::IMUL16rr , X86::IMUL32rr, X86::MULSSrr, X86::MULSDrr,
2733 static const unsigned ANDTab[] = {
2734 X86::AND8ri, X86::AND16ri, X86::AND32ri, 0, 0,
2735 X86::AND8rm, X86::AND16rm, X86::AND32rm, 0, 0,
2736 X86::AND8rr, X86::AND16rr, X86::AND32rr, 0, 0,
2738 static const unsigned ORTab[] = {
2739 X86::OR8ri, X86::OR16ri, X86::OR32ri, 0, 0,
2740 X86::OR8rm, X86::OR16rm, X86::OR32rm, 0, 0,
2741 X86::OR8rr, X86::OR16rr, X86::OR32rr, 0, 0,
2743 static const unsigned XORTab[] = {
2744 X86::XOR8ri, X86::XOR16ri, X86::XOR32ri, 0, 0,
2745 X86::XOR8rm, X86::XOR16rm, X86::XOR32rm, 0, 0,
2746 X86::XOR8rr, X86::XOR16rr, X86::XOR32rr, 0, 0,
2749 Op0 = Node->getOperand(0);
2750 Op1 = Node->getOperand(1);
2752 if (Node->getOpcode() == ISD::OR && Op0.hasOneUse() && Op1.hasOneUse())
2753 if (EmitOrOpOp(Op0, Op1, Result)) // Match SHLD, SHRD, and rotates.
2756 if (Node->getOpcode() == ISD::SUB)
2757 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(0)))
2758 if (CN->isNullValue()) { // 0 - N -> neg N
2759 switch (N.getValueType()) {
2760 default: assert(0 && "Cannot sub this type!");
2762 case MVT::i8: Opc = X86::NEG8r; break;
2763 case MVT::i16: Opc = X86::NEG16r; break;
2764 case MVT::i32: Opc = X86::NEG32r; break;
2766 Tmp1 = SelectExpr(N.getOperand(1));
2767 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
2771 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1)) {
2772 if (CN->isAllOnesValue() && Node->getOpcode() == ISD::XOR) {
2774 switch (N.getValueType()) {
2775 default: assert(0 && "Cannot add this type!");
2776 case MVT::i1: break; // Not supported, don't invert upper bits!
2777 case MVT::i8: Opc = X86::NOT8r; break;
2778 case MVT::i16: Opc = X86::NOT16r; break;
2779 case MVT::i32: Opc = X86::NOT32r; break;
2782 Tmp1 = SelectExpr(Op0);
2783 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
2788 // Fold common multiplies into LEA instructions.
2789 if (Node->getOpcode() == ISD::MUL && N.getValueType() == MVT::i32) {
2790 switch ((int)CN->getValue()) {
2795 // Remove N from exprmap so SelectAddress doesn't get confused.
2798 SelectAddress(N, AM);
2799 // Restore it to the map.
2800 ExprMap[N] = Result;
2801 addFullAddress(BuildMI(BB, X86::LEA32r, 4, Result), AM);
2806 switch (N.getValueType()) {
2807 default: assert(0 && "Cannot xor this type!");
2809 case MVT::i8: Opc = 0; break;
2810 case MVT::i16: Opc = 1; break;
2811 case MVT::i32: Opc = 2; break;
2813 switch (Node->getOpcode()) {
2814 default: assert(0 && "Unreachable!");
2816 case ISD::SUB: Opc = X86ScalarSSE ? SSE_SUBTab[Opc] : SUBTab[Opc]; break;
2818 case ISD::MUL: Opc = X86ScalarSSE ? SSE_MULTab[Opc] : MULTab[Opc]; break;
2819 case ISD::AND: Opc = ANDTab[Opc]; break;
2820 case ISD::OR: Opc = ORTab[Opc]; break;
2821 case ISD::XOR: Opc = XORTab[Opc]; break;
2823 if (Opc) { // Can't fold MUL:i8 R, imm
2824 Tmp1 = SelectExpr(Op0);
2825 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
2830 if (isFoldableLoad(Op0, Op1, true))
2831 if (Node->getOpcode() != ISD::SUB && Node->getOpcode() != ISD::FSUB) {
2832 std::swap(Op0, Op1);
2835 // For FP, emit 'reverse' subract, with a memory operand.
2836 if (N.getValueType() == MVT::f64 && !X86ScalarSSE) {
2837 if (Op0.getOpcode() == ISD::EXTLOAD)
2838 Opc = X86::FSUBR32m;
2840 Opc = X86::FSUBR64m;
2843 EmitFoldedLoad(Op0, AM);
2844 Tmp1 = SelectExpr(Op1);
2845 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
2850 if (isFoldableLoad(Op1, Op0, true)) {
2852 switch (N.getValueType()) {
2853 default: assert(0 && "Cannot operate on this type!");
2855 case MVT::i8: Opc = 5; break;
2856 case MVT::i16: Opc = 6; break;
2857 case MVT::i32: Opc = 7; break;
2858 case MVT::f32: Opc = 8; break;
2859 // For F64, handle promoted load operations (from F32) as well!
2861 assert((!X86ScalarSSE || Op1.getOpcode() == ISD::LOAD) &&
2862 "SSE load should have been promoted");
2863 Opc = Op1.getOpcode() == ISD::LOAD ? 9 : 8; break;
2865 switch (Node->getOpcode()) {
2866 default: assert(0 && "Unreachable!");
2868 case ISD::SUB: Opc = X86ScalarSSE ? SSE_SUBTab[Opc] : SUBTab[Opc]; break;
2870 case ISD::MUL: Opc = X86ScalarSSE ? SSE_MULTab[Opc] : MULTab[Opc]; break;
2871 case ISD::AND: Opc = ANDTab[Opc]; break;
2872 case ISD::OR: Opc = ORTab[Opc]; break;
2873 case ISD::XOR: Opc = XORTab[Opc]; break;
2877 EmitFoldedLoad(Op1, AM);
2878 Tmp1 = SelectExpr(Op0);
2880 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
2882 assert(Node->getOpcode() == ISD::MUL &&
2883 N.getValueType() == MVT::i8 && "Unexpected situation!");
2884 // Must use the MUL instruction, which forces use of AL.
2885 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(Tmp1);
2886 addFullAddress(BuildMI(BB, X86::MUL8m, 1), AM);
2887 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
2892 if (getRegPressure(Op0) > getRegPressure(Op1)) {
2893 Tmp1 = SelectExpr(Op0);
2894 Tmp2 = SelectExpr(Op1);
2896 Tmp2 = SelectExpr(Op1);
2897 Tmp1 = SelectExpr(Op0);
2900 switch (N.getValueType()) {
2901 default: assert(0 && "Cannot add this type!");
2903 case MVT::i8: Opc = 10; break;
2904 case MVT::i16: Opc = 11; break;
2905 case MVT::i32: Opc = 12; break;
2906 case MVT::f32: Opc = 13; break;
2907 case MVT::f64: Opc = 14; break;
2909 switch (Node->getOpcode()) {
2910 default: assert(0 && "Unreachable!");
2912 case ISD::SUB: Opc = X86ScalarSSE ? SSE_SUBTab[Opc] : SUBTab[Opc]; break;
2914 case ISD::MUL: Opc = X86ScalarSSE ? SSE_MULTab[Opc] : MULTab[Opc]; break;
2915 case ISD::AND: Opc = ANDTab[Opc]; break;
2916 case ISD::OR: Opc = ORTab[Opc]; break;
2917 case ISD::XOR: Opc = XORTab[Opc]; break;
2920 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
2922 assert(Node->getOpcode() == ISD::MUL &&
2923 N.getValueType() == MVT::i8 && "Unexpected situation!");
2924 // Must use the MUL instruction, which forces use of AL.
2925 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(Tmp1);
2926 BuildMI(BB, X86::MUL8r, 1).addReg(Tmp2);
2927 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
2931 case ISD::ADD_PARTS:
2932 case ISD::SUB_PARTS: {
2933 assert(N.getNumOperands() == 4 && N.getValueType() == MVT::i32 &&
2934 "Not an i64 add/sub!");
2935 // Emit all of the operands.
2936 std::vector<unsigned> InVals;
2937 for (unsigned i = 0, e = N.getNumOperands(); i != e; ++i)
2938 InVals.push_back(SelectExpr(N.getOperand(i)));
2939 if (N.getOpcode() == ISD::ADD_PARTS) {
2940 BuildMI(BB, X86::ADD32rr, 2, Result).addReg(InVals[0]).addReg(InVals[2]);
2941 BuildMI(BB, X86::ADC32rr,2,Result+1).addReg(InVals[1]).addReg(InVals[3]);
2943 BuildMI(BB, X86::SUB32rr, 2, Result).addReg(InVals[0]).addReg(InVals[2]);
2944 BuildMI(BB, X86::SBB32rr, 2,Result+1).addReg(InVals[1]).addReg(InVals[3]);
2946 return Result+N.ResNo;
2949 case ISD::SHL_PARTS:
2950 case ISD::SRA_PARTS:
2951 case ISD::SRL_PARTS: {
2952 assert(N.getNumOperands() == 3 && N.getValueType() == MVT::i32 &&
2953 "Not an i64 shift!");
2954 unsigned ShiftOpLo = SelectExpr(N.getOperand(0));
2955 unsigned ShiftOpHi = SelectExpr(N.getOperand(1));
2956 unsigned TmpReg = MakeReg(MVT::i32);
2957 if (N.getOpcode() == ISD::SRA_PARTS) {
2958 // If this is a SHR of a Long, then we need to do funny sign extension
2959 // stuff. TmpReg gets the value to use as the high-part if we are
2960 // shifting more than 32 bits.
2961 BuildMI(BB, X86::SAR32ri, 2, TmpReg).addReg(ShiftOpHi).addImm(31);
2963 // Other shifts use a fixed zero value if the shift is more than 32 bits.
2964 BuildMI(BB, X86::MOV32ri, 1, TmpReg).addImm(0);
2967 // Initialize CL with the shift amount.
2968 unsigned ShiftAmountReg = SelectExpr(N.getOperand(2));
2969 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(ShiftAmountReg);
2971 unsigned TmpReg2 = MakeReg(MVT::i32);
2972 unsigned TmpReg3 = MakeReg(MVT::i32);
2973 if (N.getOpcode() == ISD::SHL_PARTS) {
2974 // TmpReg2 = shld inHi, inLo
2975 BuildMI(BB, X86::SHLD32rrCL, 2,TmpReg2).addReg(ShiftOpHi)
2977 // TmpReg3 = shl inLo, CL
2978 BuildMI(BB, X86::SHL32rCL, 1, TmpReg3).addReg(ShiftOpLo);
2980 // Set the flags to indicate whether the shift was by more than 32 bits.
2981 BuildMI(BB, X86::TEST8ri, 2).addReg(X86::CL).addImm(32);
2983 // DestHi = (>32) ? TmpReg3 : TmpReg2;
2984 BuildMI(BB, X86::CMOVNE32rr, 2,
2985 Result+1).addReg(TmpReg2).addReg(TmpReg3);
2986 // DestLo = (>32) ? TmpReg : TmpReg3;
2987 BuildMI(BB, X86::CMOVNE32rr, 2,
2988 Result).addReg(TmpReg3).addReg(TmpReg);
2990 // TmpReg2 = shrd inLo, inHi
2991 BuildMI(BB, X86::SHRD32rrCL,2,TmpReg2).addReg(ShiftOpLo)
2993 // TmpReg3 = s[ah]r inHi, CL
2994 BuildMI(BB, N.getOpcode() == ISD::SRA_PARTS ? X86::SAR32rCL
2995 : X86::SHR32rCL, 1, TmpReg3)
2998 // Set the flags to indicate whether the shift was by more than 32 bits.
2999 BuildMI(BB, X86::TEST8ri, 2).addReg(X86::CL).addImm(32);
3001 // DestLo = (>32) ? TmpReg3 : TmpReg2;
3002 BuildMI(BB, X86::CMOVNE32rr, 2,
3003 Result).addReg(TmpReg2).addReg(TmpReg3);
3005 // DestHi = (>32) ? TmpReg : TmpReg3;
3006 BuildMI(BB, X86::CMOVNE32rr, 2,
3007 Result+1).addReg(TmpReg3).addReg(TmpReg);
3009 return Result+N.ResNo;
3013 EmitSelectCC(N.getOperand(0), N.getOperand(1), N.getOperand(2),
3014 N.getValueType(), Result);
3023 assert((N.getOpcode() != ISD::SREM || MVT::isInteger(N.getValueType())) &&
3024 "We don't support this operator!");
3026 if (N.getOpcode() == ISD::SDIV || N.getOpcode() == ISD::FDIV) {
3027 // We can fold loads into FpDIVs, but not really into any others.
3028 if (N.getValueType() == MVT::f64 && !X86ScalarSSE) {
3029 // Check for reversed and unreversed DIV.
3030 if (isFoldableLoad(N.getOperand(0), N.getOperand(1), true)) {
3031 if (N.getOperand(0).getOpcode() == ISD::EXTLOAD)
3032 Opc = X86::FDIVR32m;
3034 Opc = X86::FDIVR64m;
3036 EmitFoldedLoad(N.getOperand(0), AM);
3037 Tmp1 = SelectExpr(N.getOperand(1));
3038 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
3040 } else if (isFoldableLoad(N.getOperand(1), N.getOperand(0), true) &&
3041 N.getOperand(1).getOpcode() == ISD::LOAD) {
3042 if (N.getOperand(1).getOpcode() == ISD::EXTLOAD)
3047 EmitFoldedLoad(N.getOperand(1), AM);
3048 Tmp1 = SelectExpr(N.getOperand(0));
3049 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
3054 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
3055 // FIXME: These special cases should be handled by the lowering impl!
3056 unsigned RHS = CN->getValue();
3062 if (RHS && (RHS & (RHS-1)) == 0) { // Signed division by power of 2?
3063 unsigned Log = Log2_32(RHS);
3064 unsigned SAROpc, SHROpc, ADDOpc, NEGOpc;
3065 switch (N.getValueType()) {
3066 default: assert("Unknown type to signed divide!");
3068 SAROpc = X86::SAR8ri;
3069 SHROpc = X86::SHR8ri;
3070 ADDOpc = X86::ADD8rr;
3071 NEGOpc = X86::NEG8r;
3074 SAROpc = X86::SAR16ri;
3075 SHROpc = X86::SHR16ri;
3076 ADDOpc = X86::ADD16rr;
3077 NEGOpc = X86::NEG16r;
3080 SAROpc = X86::SAR32ri;
3081 SHROpc = X86::SHR32ri;
3082 ADDOpc = X86::ADD32rr;
3083 NEGOpc = X86::NEG32r;
3086 unsigned RegSize = MVT::getSizeInBits(N.getValueType());
3087 Tmp1 = SelectExpr(N.getOperand(0));
3090 TmpReg = MakeReg(N.getValueType());
3091 BuildMI(BB, SAROpc, 2, TmpReg).addReg(Tmp1).addImm(Log-1);
3095 unsigned TmpReg2 = MakeReg(N.getValueType());
3096 BuildMI(BB, SHROpc, 2, TmpReg2).addReg(TmpReg).addImm(RegSize-Log);
3097 unsigned TmpReg3 = MakeReg(N.getValueType());
3098 BuildMI(BB, ADDOpc, 2, TmpReg3).addReg(Tmp1).addReg(TmpReg2);
3100 unsigned TmpReg4 = isNeg ? MakeReg(N.getValueType()) : Result;
3101 BuildMI(BB, SAROpc, 2, TmpReg4).addReg(TmpReg3).addImm(Log);
3103 BuildMI(BB, NEGOpc, 1, Result).addReg(TmpReg4);
3109 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
3110 Tmp1 = SelectExpr(N.getOperand(0));
3111 Tmp2 = SelectExpr(N.getOperand(1));
3113 Tmp2 = SelectExpr(N.getOperand(1));
3114 Tmp1 = SelectExpr(N.getOperand(0));
3117 bool isSigned = N.getOpcode() == ISD::SDIV || N.getOpcode() == ISD::SREM;
3118 bool isDiv = N.getOpcode() == ISD::SDIV || N.getOpcode() == ISD::UDIV;
3119 unsigned LoReg, HiReg, DivOpcode, MovOpcode, ClrOpcode, SExtOpcode;
3120 switch (N.getValueType()) {
3121 default: assert(0 && "Cannot sdiv this type!");
3123 DivOpcode = isSigned ? X86::IDIV8r : X86::DIV8r;
3126 MovOpcode = X86::MOV8rr;
3127 ClrOpcode = X86::MOV8ri;
3128 SExtOpcode = X86::CBW;
3131 DivOpcode = isSigned ? X86::IDIV16r : X86::DIV16r;
3134 MovOpcode = X86::MOV16rr;
3135 ClrOpcode = X86::MOV16ri;
3136 SExtOpcode = X86::CWD;
3139 DivOpcode = isSigned ? X86::IDIV32r : X86::DIV32r;
3142 MovOpcode = X86::MOV32rr;
3143 ClrOpcode = X86::MOV32ri;
3144 SExtOpcode = X86::CDQ;
3147 BuildMI(BB, X86::DIVSSrr, 2, Result).addReg(Tmp1).addReg(Tmp2);
3150 Opc = X86ScalarSSE ? X86::DIVSDrr : X86::FpDIV;
3151 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
3155 // Set up the low part.
3156 BuildMI(BB, MovOpcode, 1, LoReg).addReg(Tmp1);
3159 // Sign extend the low part into the high part.
3160 BuildMI(BB, SExtOpcode, 0);
3162 // Zero out the high part, effectively zero extending the input.
3163 BuildMI(BB, ClrOpcode, 1, HiReg).addImm(0);
3166 // Emit the DIV/IDIV instruction.
3167 BuildMI(BB, DivOpcode, 1).addReg(Tmp2);
3169 // Get the result of the divide or rem.
3170 BuildMI(BB, MovOpcode, 1, Result).addReg(isDiv ? LoReg : HiReg);
3175 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
3176 if (CN->getValue() == 1) { // X = SHL Y, 1 -> X = ADD Y, Y
3177 switch (N.getValueType()) {
3178 default: assert(0 && "Cannot shift this type!");
3179 case MVT::i8: Opc = X86::ADD8rr; break;
3180 case MVT::i16: Opc = X86::ADD16rr; break;
3181 case MVT::i32: Opc = X86::ADD32rr; break;
3183 Tmp1 = SelectExpr(N.getOperand(0));
3184 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp1);
3188 switch (N.getValueType()) {
3189 default: assert(0 && "Cannot shift this type!");
3190 case MVT::i8: Opc = X86::SHL8ri; break;
3191 case MVT::i16: Opc = X86::SHL16ri; break;
3192 case MVT::i32: Opc = X86::SHL32ri; break;
3194 Tmp1 = SelectExpr(N.getOperand(0));
3195 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
3199 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
3200 Tmp1 = SelectExpr(N.getOperand(0));
3201 Tmp2 = SelectExpr(N.getOperand(1));
3203 Tmp2 = SelectExpr(N.getOperand(1));
3204 Tmp1 = SelectExpr(N.getOperand(0));
3207 switch (N.getValueType()) {
3208 default: assert(0 && "Cannot shift this type!");
3209 case MVT::i8 : Opc = X86::SHL8rCL; break;
3210 case MVT::i16: Opc = X86::SHL16rCL; break;
3211 case MVT::i32: Opc = X86::SHL32rCL; break;
3213 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(Tmp2);
3214 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
3217 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
3218 switch (N.getValueType()) {
3219 default: assert(0 && "Cannot shift this type!");
3220 case MVT::i8: Opc = X86::SHR8ri; break;
3221 case MVT::i16: Opc = X86::SHR16ri; break;
3222 case MVT::i32: Opc = X86::SHR32ri; break;
3224 Tmp1 = SelectExpr(N.getOperand(0));
3225 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
3229 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
3230 Tmp1 = SelectExpr(N.getOperand(0));
3231 Tmp2 = SelectExpr(N.getOperand(1));
3233 Tmp2 = SelectExpr(N.getOperand(1));
3234 Tmp1 = SelectExpr(N.getOperand(0));
3237 switch (N.getValueType()) {
3238 default: assert(0 && "Cannot shift this type!");
3239 case MVT::i8 : Opc = X86::SHR8rCL; break;
3240 case MVT::i16: Opc = X86::SHR16rCL; break;
3241 case MVT::i32: Opc = X86::SHR32rCL; break;
3243 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(Tmp2);
3244 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
3247 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
3248 switch (N.getValueType()) {
3249 default: assert(0 && "Cannot shift this type!");
3250 case MVT::i8: Opc = X86::SAR8ri; break;
3251 case MVT::i16: Opc = X86::SAR16ri; break;
3252 case MVT::i32: Opc = X86::SAR32ri; break;
3254 Tmp1 = SelectExpr(N.getOperand(0));
3255 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
3259 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
3260 Tmp1 = SelectExpr(N.getOperand(0));
3261 Tmp2 = SelectExpr(N.getOperand(1));
3263 Tmp2 = SelectExpr(N.getOperand(1));
3264 Tmp1 = SelectExpr(N.getOperand(0));
3267 switch (N.getValueType()) {
3268 default: assert(0 && "Cannot shift this type!");
3269 case MVT::i8 : Opc = X86::SAR8rCL; break;
3270 case MVT::i16: Opc = X86::SAR16rCL; break;
3271 case MVT::i32: Opc = X86::SAR32rCL; break;
3273 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(Tmp2);
3274 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
3278 EmitCMP(N.getOperand(0), N.getOperand(1), Node->hasOneUse());
3279 EmitSetCC(BB, Result, cast<CondCodeSDNode>(N.getOperand(2))->get(),
3280 MVT::isFloatingPoint(N.getOperand(1).getValueType()));
3283 // Make sure we generate both values.
3284 if (Result != 1) { // Generate the token
3285 if (!ExprMap.insert(std::make_pair(N.getValue(1), 1)).second)
3286 assert(0 && "Load already emitted!?");
3288 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
3290 switch (Node->getValueType(0)) {
3291 default: assert(0 && "Cannot load this type!");
3293 case MVT::i8: Opc = X86::MOV8rm; break;
3294 case MVT::i16: Opc = X86::MOV16rm; break;
3295 case MVT::i32: Opc = X86::MOV32rm; break;
3296 case MVT::f32: Opc = X86::MOVSSrm; break;
3302 ContainsFPCode = true;
3307 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N.getOperand(1))){
3308 unsigned CPIdx = BB->getParent()->getConstantPool()->
3309 getConstantPoolIndex(CP->get());
3310 Select(N.getOperand(0));
3311 addConstantPoolReference(BuildMI(BB, Opc, 4, Result), CPIdx);
3315 SDOperand Chain = N.getOperand(0);
3316 SDOperand Address = N.getOperand(1);
3317 if (getRegPressure(Chain) > getRegPressure(Address)) {
3319 SelectAddress(Address, AM);
3321 SelectAddress(Address, AM);
3325 addFullAddress(BuildMI(BB, Opc, 4, Result), AM);
3328 case X86ISD::FILD64m:
3329 // Make sure we generate both values.
3330 assert(Result != 1 && N.getValueType() == MVT::f64);
3331 if (!ExprMap.insert(std::make_pair(N.getValue(1), 1)).second)
3332 assert(0 && "Load already emitted!?");
3337 SDOperand Chain = N.getOperand(0);
3338 SDOperand Address = N.getOperand(1);
3339 if (getRegPressure(Chain) > getRegPressure(Address)) {
3341 SelectAddress(Address, AM);
3343 SelectAddress(Address, AM);
3347 addFullAddress(BuildMI(BB, X86::FILD64m, 4, Result), AM);
3351 case ISD::EXTLOAD: // Arbitrarily codegen extloads as MOVZX*
3352 case ISD::ZEXTLOAD: {
3353 // Make sure we generate both values.
3355 ExprMap[N.getValue(1)] = 1; // Generate the token
3357 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
3359 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N.getOperand(1)))
3360 if (Node->getValueType(0) == MVT::f64) {
3361 assert(cast<VTSDNode>(Node->getOperand(3))->getVT() == MVT::f32 &&
3363 unsigned CPIdx = BB->getParent()->getConstantPool()->
3364 getConstantPoolIndex(CP->get());
3366 addConstantPoolReference(BuildMI(BB, X86::FLD32m, 4, Result), CPIdx);
3371 if (getRegPressure(Node->getOperand(0)) >
3372 getRegPressure(Node->getOperand(1))) {
3373 Select(Node->getOperand(0)); // chain
3374 SelectAddress(Node->getOperand(1), AM);
3376 SelectAddress(Node->getOperand(1), AM);
3377 Select(Node->getOperand(0)); // chain
3380 switch (Node->getValueType(0)) {
3381 default: assert(0 && "Unknown type to sign extend to.");
3383 assert(cast<VTSDNode>(Node->getOperand(3))->getVT() == MVT::f32 &&
3385 addFullAddress(BuildMI(BB, X86::FLD32m, 5, Result), AM);
3388 switch (cast<VTSDNode>(Node->getOperand(3))->getVT()) {
3390 assert(0 && "Bad zero extend!");
3393 addFullAddress(BuildMI(BB, X86::MOVZX32rm8, 5, Result), AM);
3396 addFullAddress(BuildMI(BB, X86::MOVZX32rm16, 5, Result), AM);
3401 assert(cast<VTSDNode>(Node->getOperand(3))->getVT() <= MVT::i8 &&
3402 "Bad zero extend!");
3403 addFullAddress(BuildMI(BB, X86::MOVSX16rm8, 5, Result), AM);
3406 assert(cast<VTSDNode>(Node->getOperand(3))->getVT() == MVT::i1 &&
3407 "Bad zero extend!");
3408 addFullAddress(BuildMI(BB, X86::MOV8rm, 5, Result), AM);
3413 case ISD::SEXTLOAD: {
3414 // Make sure we generate both values.
3416 ExprMap[N.getValue(1)] = 1; // Generate the token
3418 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
3421 if (getRegPressure(Node->getOperand(0)) >
3422 getRegPressure(Node->getOperand(1))) {
3423 Select(Node->getOperand(0)); // chain
3424 SelectAddress(Node->getOperand(1), AM);
3426 SelectAddress(Node->getOperand(1), AM);
3427 Select(Node->getOperand(0)); // chain
3430 switch (Node->getValueType(0)) {
3431 case MVT::i8: assert(0 && "Cannot sign extend from bool!");
3432 default: assert(0 && "Unknown type to sign extend to.");
3434 switch (cast<VTSDNode>(Node->getOperand(3))->getVT()) {
3436 case MVT::i1: assert(0 && "Cannot sign extend from bool!");
3438 addFullAddress(BuildMI(BB, X86::MOVSX32rm8, 5, Result), AM);
3441 addFullAddress(BuildMI(BB, X86::MOVSX32rm16, 5, Result), AM);
3446 assert(cast<VTSDNode>(Node->getOperand(3))->getVT() == MVT::i8 &&
3447 "Cannot sign extend from bool!");
3448 addFullAddress(BuildMI(BB, X86::MOVSX16rm8, 5, Result), AM);
3454 case ISD::DYNAMIC_STACKALLOC:
3455 // Generate both result values.
3457 ExprMap[N.getValue(1)] = 1; // Generate the token
3459 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
3461 // FIXME: We are currently ignoring the requested alignment for handling
3462 // greater than the stack alignment. This will need to be revisited at some
3463 // point. Align = N.getOperand(2);
3465 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
3466 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
3467 std::cerr << "Cannot allocate stack object with greater alignment than"
3468 << " the stack alignment yet!";
3472 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
3473 Select(N.getOperand(0));
3474 BuildMI(BB, X86::SUB32ri, 2, X86::ESP).addReg(X86::ESP)
3475 .addImm(CN->getValue());
3477 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
3478 Select(N.getOperand(0));
3479 Tmp1 = SelectExpr(N.getOperand(1));
3481 Tmp1 = SelectExpr(N.getOperand(1));
3482 Select(N.getOperand(0));
3485 // Subtract size from stack pointer, thereby allocating some space.
3486 BuildMI(BB, X86::SUB32rr, 2, X86::ESP).addReg(X86::ESP).addReg(Tmp1);
3489 // Put a pointer to the space into the result register, by copying the stack
3491 BuildMI(BB, X86::MOV32rr, 1, Result).addReg(X86::ESP);
3494 case X86ISD::TAILCALL:
3495 case X86ISD::CALL: {
3496 // The chain for this call is now lowered.
3497 ExprMap.insert(std::make_pair(N.getValue(0), 1));
3499 bool isDirect = isa<GlobalAddressSDNode>(N.getOperand(1)) ||
3500 isa<ExternalSymbolSDNode>(N.getOperand(1));
3501 unsigned Callee = 0;
3503 Select(N.getOperand(0));
3505 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
3506 Select(N.getOperand(0));
3507 Callee = SelectExpr(N.getOperand(1));
3509 Callee = SelectExpr(N.getOperand(1));
3510 Select(N.getOperand(0));
3514 // If this call has values to pass in registers, do so now.
3515 if (Node->getNumOperands() > 4) {
3516 // The first value is passed in (a part of) EAX, the second in EDX.
3517 unsigned RegOp1 = SelectExpr(N.getOperand(4));
3519 Node->getNumOperands() > 5 ? SelectExpr(N.getOperand(5)) : 0;
3521 switch (N.getOperand(4).getValueType()) {
3522 default: assert(0 && "Bad thing to pass in regs");
3524 case MVT::i8: BuildMI(BB, X86::MOV8rr , 1,X86::AL).addReg(RegOp1); break;
3525 case MVT::i16: BuildMI(BB, X86::MOV16rr, 1,X86::AX).addReg(RegOp1); break;
3526 case MVT::i32: BuildMI(BB, X86::MOV32rr, 1,X86::EAX).addReg(RegOp1);break;
3529 switch (N.getOperand(5).getValueType()) {
3530 default: assert(0 && "Bad thing to pass in regs");
3533 BuildMI(BB, X86::MOV8rr , 1, X86::DL).addReg(RegOp2);
3536 BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(RegOp2);
3539 BuildMI(BB, X86::MOV32rr, 1, X86::EDX).addReg(RegOp2);
3544 if (GlobalAddressSDNode *GASD =
3545 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
3546 BuildMI(BB, X86::CALLpcrel32, 1).addGlobalAddress(GASD->getGlobal(),true);
3547 } else if (ExternalSymbolSDNode *ESSDN =
3548 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
3549 BuildMI(BB, X86::CALLpcrel32,
3550 1).addExternalSymbol(ESSDN->getSymbol(), true);
3552 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
3553 Select(N.getOperand(0));
3554 Tmp1 = SelectExpr(N.getOperand(1));
3556 Tmp1 = SelectExpr(N.getOperand(1));
3557 Select(N.getOperand(0));
3560 BuildMI(BB, X86::CALL32r, 1).addReg(Tmp1);
3563 // Get caller stack amount and amount the callee added to the stack pointer.
3564 Tmp1 = cast<ConstantSDNode>(N.getOperand(2))->getValue();
3565 Tmp2 = cast<ConstantSDNode>(N.getOperand(3))->getValue();
3566 BuildMI(BB, X86::ADJCALLSTACKUP, 2).addImm(Tmp1).addImm(Tmp2);
3568 if (Node->getNumValues() != 1)
3569 switch (Node->getValueType(1)) {
3570 default: assert(0 && "Unknown value type for call result!");
3571 case MVT::Other: return 1;
3574 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
3577 BuildMI(BB, X86::MOV16rr, 1, Result).addReg(X86::AX);
3580 BuildMI(BB, X86::MOV32rr, 1, Result).addReg(X86::EAX);
3581 if (Node->getNumValues() == 3 && Node->getValueType(2) == MVT::i32)
3582 BuildMI(BB, X86::MOV32rr, 1, Result+1).addReg(X86::EDX);
3584 case MVT::f64: // Floating-point return values live in %ST(0)
3586 ContainsFPCode = true;
3587 BuildMI(BB, X86::FpGETRESULT, 1, X86::FP0);
3589 unsigned Size = MVT::getSizeInBits(MVT::f64)/8;
3590 MachineFunction *F = BB->getParent();
3591 int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, Size);
3592 addFrameReference(BuildMI(BB, X86::FST64m, 5), FrameIdx).addReg(X86::FP0);
3593 addFrameReference(BuildMI(BB, X86::MOVSDrm, 4, Result), FrameIdx);
3596 ContainsFPCode = true;
3597 BuildMI(BB, X86::FpGETRESULT, 1, Result);
3601 return Result+N.ResNo-1;
3604 // First, determine that the size of the operand falls within the acceptable
3605 // range for this architecture.
3607 if (Node->getOperand(1).getValueType() != MVT::i16) {
3608 std::cerr << "llvm.readport: Address size is not 16 bits\n";
3612 // Make sure we generate both values.
3613 if (Result != 1) { // Generate the token
3614 if (!ExprMap.insert(std::make_pair(N.getValue(1), 1)).second)
3615 assert(0 && "readport already emitted!?");
3617 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
3619 Select(Node->getOperand(0)); // Select the chain.
3621 // If the port is a single-byte constant, use the immediate form.
3622 if (ConstantSDNode *Port = dyn_cast<ConstantSDNode>(Node->getOperand(1)))
3623 if ((Port->getValue() & 255) == Port->getValue()) {
3624 switch (Node->getValueType(0)) {
3626 BuildMI(BB, X86::IN8ri, 1).addImm(Port->getValue());
3627 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
3630 BuildMI(BB, X86::IN16ri, 1).addImm(Port->getValue());
3631 BuildMI(BB, X86::MOV16rr, 1, Result).addReg(X86::AX);
3634 BuildMI(BB, X86::IN32ri, 1).addImm(Port->getValue());
3635 BuildMI(BB, X86::MOV32rr, 1, Result).addReg(X86::EAX);
3641 // Now, move the I/O port address into the DX register and use the IN
3642 // instruction to get the input data.
3644 Tmp1 = SelectExpr(Node->getOperand(1));
3645 BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(Tmp1);
3646 switch (Node->getValueType(0)) {
3648 BuildMI(BB, X86::IN8rr, 0);
3649 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
3652 BuildMI(BB, X86::IN16rr, 0);
3653 BuildMI(BB, X86::MOV16rr, 1, Result).addReg(X86::AX);
3656 BuildMI(BB, X86::IN32rr, 0);
3657 BuildMI(BB, X86::MOV32rr, 1, Result).addReg(X86::EAX);
3660 std::cerr << "Cannot do input on this data type";
3669 /// TryToFoldLoadOpStore - Given a store node, try to fold together a
3670 /// load/op/store instruction. If successful return true.
3671 bool ISel::TryToFoldLoadOpStore(SDNode *Node) {
3672 assert(Node->getOpcode() == ISD::STORE && "Can only do this for stores!");
3673 SDOperand Chain = Node->getOperand(0);
3674 SDOperand StVal = Node->getOperand(1);
3675 SDOperand StPtr = Node->getOperand(2);
3677 // The chain has to be a load, the stored value must be an integer binary
3678 // operation with one use.
3679 if (!StVal.Val->hasOneUse() || StVal.Val->getNumOperands() != 2 ||
3680 MVT::isFloatingPoint(StVal.getValueType()))
3683 // Token chain must either be a factor node or the load to fold.
3684 if (Chain.getOpcode() != ISD::LOAD && Chain.getOpcode() != ISD::TokenFactor)
3689 // Check to see if there is a load from the same pointer that we're storing
3690 // to in either operand of the binop.
3691 if (StVal.getOperand(0).getOpcode() == ISD::LOAD &&
3692 StVal.getOperand(0).getOperand(1) == StPtr)
3693 TheLoad = StVal.getOperand(0);
3694 else if (StVal.getOperand(1).getOpcode() == ISD::LOAD &&
3695 StVal.getOperand(1).getOperand(1) == StPtr)
3696 TheLoad = StVal.getOperand(1);
3698 return false; // No matching load operand.
3700 // We can only fold the load if there are no intervening side-effecting
3701 // operations. This means that the store uses the load as its token chain, or
3702 // there are only token factor nodes in between the store and load.
3703 if (Chain != TheLoad.getValue(1)) {
3704 // Okay, the other option is that we have a store referring to (possibly
3705 // nested) token factor nodes. For now, just try peeking through one level
3706 // of token factors to see if this is the case.
3707 bool ChainOk = false;
3708 if (Chain.getOpcode() == ISD::TokenFactor) {
3709 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
3710 if (Chain.getOperand(i) == TheLoad.getValue(1)) {
3716 if (!ChainOk) return false;
3719 if (TheLoad.getOperand(1) != StPtr)
3722 // Make sure that one of the operands of the binop is the load, and that the
3723 // load folds into the binop.
3724 if (((StVal.getOperand(0) != TheLoad ||
3725 !isFoldableLoad(TheLoad, StVal.getOperand(1))) &&
3726 (StVal.getOperand(1) != TheLoad ||
3727 !isFoldableLoad(TheLoad, StVal.getOperand(0)))))
3730 // Finally, check to see if this is one of the ops we can handle!
3731 static const unsigned ADDTAB[] = {
3732 X86::ADD8mi, X86::ADD16mi, X86::ADD32mi,
3733 X86::ADD8mr, X86::ADD16mr, X86::ADD32mr,
3735 static const unsigned SUBTAB[] = {
3736 X86::SUB8mi, X86::SUB16mi, X86::SUB32mi,
3737 X86::SUB8mr, X86::SUB16mr, X86::SUB32mr,
3739 static const unsigned ANDTAB[] = {
3740 X86::AND8mi, X86::AND16mi, X86::AND32mi,
3741 X86::AND8mr, X86::AND16mr, X86::AND32mr,
3743 static const unsigned ORTAB[] = {
3744 X86::OR8mi, X86::OR16mi, X86::OR32mi,
3745 X86::OR8mr, X86::OR16mr, X86::OR32mr,
3747 static const unsigned XORTAB[] = {
3748 X86::XOR8mi, X86::XOR16mi, X86::XOR32mi,
3749 X86::XOR8mr, X86::XOR16mr, X86::XOR32mr,
3751 static const unsigned SHLTAB[] = {
3752 X86::SHL8mi, X86::SHL16mi, X86::SHL32mi,
3753 /*Have to put the reg in CL*/0, 0, 0,
3755 static const unsigned SARTAB[] = {
3756 X86::SAR8mi, X86::SAR16mi, X86::SAR32mi,
3757 /*Have to put the reg in CL*/0, 0, 0,
3759 static const unsigned SHRTAB[] = {
3760 X86::SHR8mi, X86::SHR16mi, X86::SHR32mi,
3761 /*Have to put the reg in CL*/0, 0, 0,
3764 const unsigned *TabPtr = 0;
3765 switch (StVal.getOpcode()) {
3767 std::cerr << "CANNOT [mem] op= val: ";
3768 StVal.Val->dump(); std::cerr << "\n";
3776 case ISD::UREM: return false;
3778 case ISD::ADD: TabPtr = ADDTAB; break;
3779 case ISD::SUB: TabPtr = SUBTAB; break;
3780 case ISD::AND: TabPtr = ANDTAB; break;
3781 case ISD:: OR: TabPtr = ORTAB; break;
3782 case ISD::XOR: TabPtr = XORTAB; break;
3783 case ISD::SHL: TabPtr = SHLTAB; break;
3784 case ISD::SRA: TabPtr = SARTAB; break;
3785 case ISD::SRL: TabPtr = SHRTAB; break;
3788 // Handle: [mem] op= CST
3789 SDOperand Op0 = StVal.getOperand(0);
3790 SDOperand Op1 = StVal.getOperand(1);
3792 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1)) {
3793 switch (Op0.getValueType()) { // Use Op0's type because of shifts.
3796 case MVT::i8: Opc = TabPtr[0]; break;
3797 case MVT::i16: Opc = TabPtr[1]; break;
3798 case MVT::i32: Opc = TabPtr[2]; break;
3802 if (!ExprMap.insert(std::make_pair(TheLoad.getValue(1), 1)).second)
3803 assert(0 && "Already emitted?");
3807 if (getRegPressure(TheLoad.getOperand(0)) >
3808 getRegPressure(TheLoad.getOperand(1))) {
3809 Select(TheLoad.getOperand(0));
3810 SelectAddress(TheLoad.getOperand(1), AM);
3812 SelectAddress(TheLoad.getOperand(1), AM);
3813 Select(TheLoad.getOperand(0));
3816 if (StVal.getOpcode() == ISD::ADD) {
3817 if (CN->getValue() == 1) {
3818 switch (Op0.getValueType()) {
3821 addFullAddress(BuildMI(BB, X86::INC8m, 4), AM);
3823 case MVT::i16: Opc = TabPtr[1];
3824 addFullAddress(BuildMI(BB, X86::INC16m, 4), AM);
3826 case MVT::i32: Opc = TabPtr[2];
3827 addFullAddress(BuildMI(BB, X86::INC32m, 4), AM);
3830 } else if (CN->getValue()+1 == 0) { // [X] += -1 -> DEC [X]
3831 switch (Op0.getValueType()) {
3834 addFullAddress(BuildMI(BB, X86::DEC8m, 4), AM);
3836 case MVT::i16: Opc = TabPtr[1];
3837 addFullAddress(BuildMI(BB, X86::DEC16m, 4), AM);
3839 case MVT::i32: Opc = TabPtr[2];
3840 addFullAddress(BuildMI(BB, X86::DEC32m, 4), AM);
3846 addFullAddress(BuildMI(BB, Opc, 4+1),AM).addImm(CN->getValue());
3851 // If we have [mem] = V op [mem], try to turn it into:
3852 // [mem] = [mem] op V.
3853 if (Op1 == TheLoad &&
3854 StVal.getOpcode() != ISD::SUB && StVal.getOpcode() != ISD::FSUB &&
3855 StVal.getOpcode() != ISD::SHL && StVal.getOpcode() != ISD::SRA &&
3856 StVal.getOpcode() != ISD::SRL)
3857 std::swap(Op0, Op1);
3859 if (Op0 != TheLoad) return false;
3861 switch (Op0.getValueType()) {
3862 default: return false;
3864 case MVT::i8: Opc = TabPtr[3]; break;
3865 case MVT::i16: Opc = TabPtr[4]; break;
3866 case MVT::i32: Opc = TabPtr[5]; break;
3869 // Table entry doesn't exist?
3870 if (Opc == 0) return false;
3872 if (!ExprMap.insert(std::make_pair(TheLoad.getValue(1), 1)).second)
3873 assert(0 && "Already emitted?");
3875 Select(TheLoad.getOperand(0));
3878 SelectAddress(TheLoad.getOperand(1), AM);
3879 unsigned Reg = SelectExpr(Op1);
3880 addFullAddress(BuildMI(BB, Opc, 4+1), AM).addReg(Reg);
3884 /// If node is a ret(tailcall) node, emit the specified tail call and return
3885 /// true, otherwise return false.
3887 /// FIXME: This whole thing should be a post-legalize optimization pass which
3888 /// recognizes and transforms the dag. We don't want the selection phase doing
3891 bool ISel::EmitPotentialTailCall(SDNode *RetNode) {
3892 assert(RetNode->getOpcode() == ISD::RET && "Not a return");
3894 SDOperand Chain = RetNode->getOperand(0);
3896 // If this is a token factor node where one operand is a call, dig into it.
3897 SDOperand TokFactor;
3898 unsigned TokFactorOperand = 0;
3899 if (Chain.getOpcode() == ISD::TokenFactor) {
3900 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
3901 if (Chain.getOperand(i).getOpcode() == ISD::CALLSEQ_END ||
3902 Chain.getOperand(i).getOpcode() == X86ISD::TAILCALL) {
3903 TokFactorOperand = i;
3905 Chain = Chain.getOperand(i);
3908 if (TokFactor.Val == 0) return false; // No call operand.
3911 // Skip the CALLSEQ_END node if present.
3912 if (Chain.getOpcode() == ISD::CALLSEQ_END)
3913 Chain = Chain.getOperand(0);
3915 // Is a tailcall the last control operation that occurs before the return?
3916 if (Chain.getOpcode() != X86ISD::TAILCALL)
3919 // If we return a value, is it the value produced by the call?
3920 if (RetNode->getNumOperands() > 1) {
3921 // Not returning the ret val of the call?
3922 if (Chain.Val->getNumValues() == 1 ||
3923 RetNode->getOperand(1) != Chain.getValue(1))
3926 if (RetNode->getNumOperands() > 2) {
3927 if (Chain.Val->getNumValues() == 2 ||
3928 RetNode->getOperand(2) != Chain.getValue(2))
3931 assert(RetNode->getNumOperands() <= 3);
3934 // CalleeCallArgAmt - The total number of bytes used for the callee arg area.
3935 // For FastCC, this will always be > 0.
3936 unsigned CalleeCallArgAmt =
3937 cast<ConstantSDNode>(Chain.getOperand(2))->getValue();
3939 // CalleeCallArgPopAmt - The number of bytes in the call area popped by the
3940 // callee. For FastCC this will always be > 0, for CCC this is always 0.
3941 unsigned CalleeCallArgPopAmt =
3942 cast<ConstantSDNode>(Chain.getOperand(3))->getValue();
3944 // There are several cases we can handle here. First, if the caller and
3945 // callee are both CCC functions, we can tailcall if the callee takes <= the
3946 // number of argument bytes that the caller does.
3947 if (CalleeCallArgPopAmt == 0 && // Callee is C CallingConv?
3948 X86Lowering.getBytesToPopOnReturn() == 0) { // Caller is C CallingConv?
3949 // Check to see if caller arg area size >= callee arg area size.
3950 if (X86Lowering.getBytesCallerReserves() >= CalleeCallArgAmt) {
3951 //std::cerr << "CCC TAILCALL UNIMP!\n";
3952 // If TokFactor is non-null, emit all operands.
3954 //EmitCCCToCCCTailCall(Chain.Val);
3960 // Second, if both are FastCC functions, we can always perform the tail call.
3961 if (CalleeCallArgPopAmt && X86Lowering.getBytesToPopOnReturn()) {
3962 // If TokFactor is non-null, emit all operands before the call.
3963 if (TokFactor.Val) {
3964 for (unsigned i = 0, e = TokFactor.getNumOperands(); i != e; ++i)
3965 if (i != TokFactorOperand)
3966 Select(TokFactor.getOperand(i));
3969 EmitFastCCToFastCCTailCall(Chain.Val);
3973 // We don't support mixed calls, due to issues with alignment. We could in
3974 // theory handle some mixed calls from CCC -> FastCC if the stack is properly
3975 // aligned (which depends on the number of arguments to the callee). TODO.
3979 static SDOperand GetAdjustedArgumentStores(SDOperand Chain, int Offset,
3980 SelectionDAG &DAG) {
3981 MVT::ValueType StoreVT;
3982 switch (Chain.getOpcode()) {
3983 default: assert(0 && "Unexpected node!");
3984 case ISD::CALLSEQ_START:
3985 // If we found the start of the call sequence, we're done. We actually
3986 // strip off the CALLSEQ_START node, to avoid generating the
3987 // ADJCALLSTACKDOWN marker for the tail call.
3988 return Chain.getOperand(0);
3989 case ISD::TokenFactor: {
3990 std::vector<SDOperand> Ops;
3991 Ops.reserve(Chain.getNumOperands());
3992 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
3993 Ops.push_back(GetAdjustedArgumentStores(Chain.getOperand(i), Offset,DAG));
3994 return DAG.getNode(ISD::TokenFactor, MVT::Other, Ops);
3996 case ISD::STORE: // Normal store
3997 StoreVT = Chain.getOperand(1).getValueType();
3999 case ISD::TRUNCSTORE: // FLOAT store
4000 StoreVT = cast<VTSDNode>(Chain.getOperand(4))->getVT();
4004 SDOperand OrigDest = Chain.getOperand(2);
4005 unsigned OrigOffset;
4007 if (OrigDest.getOpcode() == ISD::CopyFromReg) {
4009 assert(cast<RegisterSDNode>(OrigDest.getOperand(1))->getReg() == X86::ESP);
4011 // We expect only (ESP+C)
4012 assert(OrigDest.getOpcode() == ISD::ADD &&
4013 isa<ConstantSDNode>(OrigDest.getOperand(1)) &&
4014 OrigDest.getOperand(0).getOpcode() == ISD::CopyFromReg &&
4015 cast<RegisterSDNode>(OrigDest.getOperand(0).getOperand(1))->getReg()
4017 OrigOffset = cast<ConstantSDNode>(OrigDest.getOperand(1))->getValue();
4020 // Compute the new offset from the incoming ESP value we wish to use.
4021 unsigned NewOffset = OrigOffset + Offset;
4023 unsigned OpSize = (MVT::getSizeInBits(StoreVT)+7)/8; // Bits -> Bytes
4024 MachineFunction &MF = DAG.getMachineFunction();
4025 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, NewOffset);
4026 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
4028 SDOperand InChain = GetAdjustedArgumentStores(Chain.getOperand(0), Offset,
4030 if (Chain.getOpcode() == ISD::STORE)
4031 return DAG.getNode(ISD::STORE, MVT::Other, InChain, Chain.getOperand(1),
4033 assert(Chain.getOpcode() == ISD::TRUNCSTORE);
4034 return DAG.getNode(ISD::TRUNCSTORE, MVT::Other, InChain, Chain.getOperand(1),
4035 FIN, DAG.getSrcValue(NULL), DAG.getValueType(StoreVT));
4039 /// EmitFastCCToFastCCTailCall - Given a tailcall in the tail position to a
4040 /// fastcc function from a fastcc function, emit the code to emit a 'proper'
4042 void ISel::EmitFastCCToFastCCTailCall(SDNode *TailCallNode) {
4043 unsigned CalleeCallArgSize =
4044 cast<ConstantSDNode>(TailCallNode->getOperand(2))->getValue();
4045 unsigned CallerArgSize = X86Lowering.getBytesToPopOnReturn();
4047 //std::cerr << "****\n*** EMITTING TAIL CALL!\n****\n";
4049 // Adjust argument stores. Instead of storing to [ESP], f.e., store to frame
4050 // indexes that are relative to the incoming ESP. If the incoming and
4051 // outgoing arg sizes are the same we will store to [InESP] instead of
4052 // [CurESP] and the ESP referenced will be relative to the incoming function
4054 int ESPOffset = CallerArgSize-CalleeCallArgSize;
4055 SDOperand AdjustedArgStores =
4056 GetAdjustedArgumentStores(TailCallNode->getOperand(0), ESPOffset, *TheDAG);
4058 // Copy the return address of the caller into a virtual register so we don't
4062 SDOperand RetValAddr = X86Lowering.getReturnAddressFrameIndex(*TheDAG);
4063 RetVal = TheDAG->getLoad(MVT::i32, TheDAG->getEntryNode(),
4064 RetValAddr, TheDAG->getSrcValue(NULL));
4068 // Codegen all of the argument stores.
4069 Select(AdjustedArgStores);
4072 // Emit a store of the saved ret value to the new location.
4073 MachineFunction &MF = TheDAG->getMachineFunction();
4074 int ReturnAddrFI = MF.getFrameInfo()->CreateFixedObject(4, ESPOffset-4);
4075 SDOperand RetValAddr = TheDAG->getFrameIndex(ReturnAddrFI, MVT::i32);
4076 Select(TheDAG->getNode(ISD::STORE, MVT::Other, TheDAG->getEntryNode(),
4077 RetVal, RetValAddr));
4080 // Get the destination value.
4081 SDOperand Callee = TailCallNode->getOperand(1);
4082 bool isDirect = isa<GlobalAddressSDNode>(Callee) ||
4083 isa<ExternalSymbolSDNode>(Callee);
4084 unsigned CalleeReg = 0;
4085 if (!isDirect) CalleeReg = SelectExpr(Callee);
4087 unsigned RegOp1 = 0;
4088 unsigned RegOp2 = 0;
4090 if (TailCallNode->getNumOperands() > 4) {
4091 // The first value is passed in (a part of) EAX, the second in EDX.
4092 RegOp1 = SelectExpr(TailCallNode->getOperand(4));
4093 if (TailCallNode->getNumOperands() > 5)
4094 RegOp2 = SelectExpr(TailCallNode->getOperand(5));
4096 switch (TailCallNode->getOperand(4).getValueType()) {
4097 default: assert(0 && "Bad thing to pass in regs");
4100 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(RegOp1);
4104 BuildMI(BB, X86::MOV16rr, 1,X86::AX).addReg(RegOp1);
4108 BuildMI(BB, X86::MOV32rr, 1,X86::EAX).addReg(RegOp1);
4113 switch (TailCallNode->getOperand(5).getValueType()) {
4114 default: assert(0 && "Bad thing to pass in regs");
4117 BuildMI(BB, X86::MOV8rr, 1, X86::DL).addReg(RegOp2);
4121 BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(RegOp2);
4125 BuildMI(BB, X86::MOV32rr, 1, X86::EDX).addReg(RegOp2);
4133 BuildMI(BB, X86::ADJSTACKPTRri, 2,
4134 X86::ESP).addReg(X86::ESP).addImm(ESPOffset);
4136 // TODO: handle jmp [mem]
4138 BuildMI(BB, X86::TAILJMPr, 1).addReg(CalleeReg);
4139 } else if (GlobalAddressSDNode *GASD = dyn_cast<GlobalAddressSDNode>(Callee)){
4140 BuildMI(BB, X86::TAILJMPd, 1).addGlobalAddress(GASD->getGlobal(), true);
4142 ExternalSymbolSDNode *ESSDN = cast<ExternalSymbolSDNode>(Callee);
4143 BuildMI(BB, X86::TAILJMPd, 1).addExternalSymbol(ESSDN->getSymbol(), true);
4145 // ADD IMPLICIT USE RegOp1/RegOp2's
4149 void ISel::Select(SDOperand N) {
4150 unsigned Tmp1 = 0, Tmp2 = 0, Opc = 0;
4152 if (!ExprMap.insert(std::make_pair(N, 1)).second)
4153 return; // Already selected.
4155 SDNode *Node = N.Val;
4157 switch (Node->getOpcode()) {
4159 Node->dump(); std::cerr << "\n";
4160 assert(0 && "Node not handled yet!");
4161 case ISD::EntryToken: return; // Noop
4162 case ISD::TokenFactor:
4163 if (Node->getNumOperands() == 2) {
4165 getRegPressure(Node->getOperand(1))>getRegPressure(Node->getOperand(0));
4166 Select(Node->getOperand(OneFirst));
4167 Select(Node->getOperand(!OneFirst));
4169 std::vector<std::pair<unsigned, unsigned> > OpsP;
4170 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
4171 OpsP.push_back(std::make_pair(getRegPressure(Node->getOperand(i)), i));
4172 std::sort(OpsP.begin(), OpsP.end());
4173 std::reverse(OpsP.begin(), OpsP.end());
4174 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
4175 Select(Node->getOperand(OpsP[i].second));
4178 case ISD::CopyToReg:
4179 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(2))) {
4180 Select(N.getOperand(0));
4181 Tmp1 = SelectExpr(N.getOperand(2));
4183 Tmp1 = SelectExpr(N.getOperand(2));
4184 Select(N.getOperand(0));
4186 Tmp2 = cast<RegisterSDNode>(N.getOperand(1))->getReg();
4189 switch (N.getOperand(2).getValueType()) {
4190 default: assert(0 && "Invalid type for operation!");
4192 case MVT::i8: Opc = X86::MOV8rr; break;
4193 case MVT::i16: Opc = X86::MOV16rr; break;
4194 case MVT::i32: Opc = X86::MOV32rr; break;
4195 case MVT::f32: Opc = X86::MOVAPSrr; break;
4198 Opc = X86::MOVAPDrr;
4201 ContainsFPCode = true;
4205 BuildMI(BB, Opc, 1, Tmp2).addReg(Tmp1);
4209 if (N.getOperand(0).getOpcode() == ISD::CALLSEQ_END ||
4210 N.getOperand(0).getOpcode() == X86ISD::TAILCALL ||
4211 N.getOperand(0).getOpcode() == ISD::TokenFactor)
4212 if (EmitPotentialTailCall(Node))
4215 switch (N.getNumOperands()) {
4217 assert(0 && "Unknown return instruction!");
4219 assert(N.getOperand(1).getValueType() == MVT::i32 &&
4220 N.getOperand(2).getValueType() == MVT::i32 &&
4221 "Unknown two-register value!");
4222 if (getRegPressure(N.getOperand(1)) > getRegPressure(N.getOperand(2))) {
4223 Tmp1 = SelectExpr(N.getOperand(1));
4224 Tmp2 = SelectExpr(N.getOperand(2));
4226 Tmp2 = SelectExpr(N.getOperand(2));
4227 Tmp1 = SelectExpr(N.getOperand(1));
4229 Select(N.getOperand(0));
4231 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Tmp1);
4232 BuildMI(BB, X86::MOV32rr, 1, X86::EDX).addReg(Tmp2);
4235 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
4236 Select(N.getOperand(0));
4237 Tmp1 = SelectExpr(N.getOperand(1));
4239 Tmp1 = SelectExpr(N.getOperand(1));
4240 Select(N.getOperand(0));
4242 switch (N.getOperand(1).getValueType()) {
4243 default: assert(0 && "All other types should have been promoted!!");
4246 // Spill the value to memory and reload it into top of stack.
4247 unsigned Size = MVT::getSizeInBits(MVT::f32)/8;
4248 MachineFunction *F = BB->getParent();
4249 int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, Size);
4250 addFrameReference(BuildMI(BB, X86::MOVSSmr, 5), FrameIdx).addReg(Tmp1);
4251 addFrameReference(BuildMI(BB, X86::FLD32m, 4, X86::FP0), FrameIdx);
4252 BuildMI(BB, X86::FpSETRESULT, 1).addReg(X86::FP0);
4253 ContainsFPCode = true;
4255 assert(0 && "MVT::f32 only legal with scalar sse fp");
4261 // Spill the value to memory and reload it into top of stack.
4262 unsigned Size = MVT::getSizeInBits(MVT::f64)/8;
4263 MachineFunction *F = BB->getParent();
4264 int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, Size);
4265 addFrameReference(BuildMI(BB, X86::MOVSDmr, 5), FrameIdx).addReg(Tmp1);
4266 addFrameReference(BuildMI(BB, X86::FLD64m, 4, X86::FP0), FrameIdx);
4267 BuildMI(BB, X86::FpSETRESULT, 1).addReg(X86::FP0);
4268 ContainsFPCode = true;
4270 BuildMI(BB, X86::FpSETRESULT, 1).addReg(Tmp1);
4274 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Tmp1);
4279 Select(N.getOperand(0));
4282 if (X86Lowering.getBytesToPopOnReturn() == 0)
4283 BuildMI(BB, X86::RET, 0); // Just emit a 'ret' instruction
4285 BuildMI(BB, X86::RETI, 1).addImm(X86Lowering.getBytesToPopOnReturn());
4288 Select(N.getOperand(0));
4289 MachineBasicBlock *Dest =
4290 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
4291 BuildMI(BB, X86::JMP, 1).addMBB(Dest);
4296 MachineBasicBlock *Dest =
4297 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
4299 // Try to fold a setcc into the branch. If this fails, emit a test/jne
4301 if (EmitBranchCC(Dest, N.getOperand(0), N.getOperand(1))) {
4302 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
4303 Select(N.getOperand(0));
4304 Tmp1 = SelectExpr(N.getOperand(1));
4306 Tmp1 = SelectExpr(N.getOperand(1));
4307 Select(N.getOperand(0));
4309 BuildMI(BB, X86::TEST8rr, 2).addReg(Tmp1).addReg(Tmp1);
4310 BuildMI(BB, X86::JNE, 1).addMBB(Dest);
4317 // If this load could be folded into the only using instruction, and if it
4318 // is safe to emit the instruction here, try to do so now.
4319 if (Node->hasNUsesOfValue(1, 0)) {
4320 SDOperand TheVal = N.getValue(0);
4322 for (SDNode::use_iterator UI = Node->use_begin(); ; ++UI) {
4323 assert(UI != Node->use_end() && "Didn't find use!");
4325 for (unsigned i = 0, e = UN->getNumOperands(); i != e; ++i)
4326 if (UN->getOperand(i) == TheVal) {
4332 // Only handle unary operators right now.
4333 if (User->getNumOperands() == 1) {
4335 SelectExpr(SDOperand(User, 0));
4346 case ISD::DYNAMIC_STACKALLOC:
4347 case X86ISD::TAILCALL:
4352 case ISD::CopyFromReg:
4353 case X86ISD::FILD64m:
4355 SelectExpr(N.getValue(0));
4358 case X86ISD::FP_TO_INT16_IN_MEM:
4359 case X86ISD::FP_TO_INT32_IN_MEM:
4360 case X86ISD::FP_TO_INT64_IN_MEM: {
4361 assert(N.getOperand(1).getValueType() == MVT::f64);
4363 Select(N.getOperand(0)); // Select the token chain
4366 if (getRegPressure(N.getOperand(1)) > getRegPressure(N.getOperand(2))) {
4367 ValReg = SelectExpr(N.getOperand(1));
4368 SelectAddress(N.getOperand(2), AM);
4370 SelectAddress(N.getOperand(2), AM);
4371 ValReg = SelectExpr(N.getOperand(1));
4374 // Change the floating point control register to use "round towards zero"
4375 // mode when truncating to an integer value.
4377 MachineFunction *F = BB->getParent();
4378 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4379 addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
4381 // Load the old value of the high byte of the control word...
4382 unsigned OldCW = MakeReg(MVT::i16);
4383 addFrameReference(BuildMI(BB, X86::MOV16rm, 4, OldCW), CWFrameIdx);
4385 // Set the high part to be round to zero...
4386 addFrameReference(BuildMI(BB, X86::MOV16mi, 5), CWFrameIdx).addImm(0xC7F);
4388 // Reload the modified control word now...
4389 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
4391 // Restore the memory image of control word to original value
4392 addFrameReference(BuildMI(BB, X86::MOV16mr, 5), CWFrameIdx).addReg(OldCW);
4394 // Get the X86 opcode to use.
4395 switch (N.getOpcode()) {
4396 case X86ISD::FP_TO_INT16_IN_MEM: Tmp1 = X86::FIST16m; break;
4397 case X86ISD::FP_TO_INT32_IN_MEM: Tmp1 = X86::FIST32m; break;
4398 case X86ISD::FP_TO_INT64_IN_MEM: Tmp1 = X86::FISTP64m; break;
4401 addFullAddress(BuildMI(BB, Tmp1, 5), AM).addReg(ValReg);
4403 // Reload the original control word now.
4404 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
4408 case ISD::TRUNCSTORE: { // truncstore chain, val, ptr, SRCVALUE, storety
4410 MVT::ValueType StoredTy = cast<VTSDNode>(N.getOperand(4))->getVT();
4411 assert((StoredTy == MVT::i1 || StoredTy == MVT::f32 ||
4412 StoredTy == MVT::i16 /*FIXME: THIS IS JUST FOR TESTING!*/)
4413 && "Unsupported TRUNCSTORE for this target!");
4415 if (StoredTy == MVT::i16) {
4416 // FIXME: This is here just to allow testing. X86 doesn't really have a
4417 // TRUNCSTORE i16 operation, but this is required for targets that do not
4418 // have 16-bit integer registers. We occasionally disable 16-bit integer
4419 // registers to test the promotion code.
4420 Select(N.getOperand(0));
4421 Tmp1 = SelectExpr(N.getOperand(1));
4422 SelectAddress(N.getOperand(2), AM);
4424 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Tmp1);
4425 addFullAddress(BuildMI(BB, X86::MOV16mr, 5), AM).addReg(X86::AX);
4429 // Store of constant bool?
4430 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
4431 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(2))) {
4432 Select(N.getOperand(0));
4433 SelectAddress(N.getOperand(2), AM);
4435 SelectAddress(N.getOperand(2), AM);
4436 Select(N.getOperand(0));
4438 addFullAddress(BuildMI(BB, X86::MOV8mi, 5), AM).addImm(CN->getValue());
4443 default: assert(0 && "Cannot truncstore this type!");
4444 case MVT::i1: Opc = X86::MOV8mr; break;
4446 assert(!X86ScalarSSE && "Cannot truncstore scalar SSE regs");
4447 Opc = X86::FST32m; break;
4450 std::vector<std::pair<unsigned, unsigned> > RP;
4451 RP.push_back(std::make_pair(getRegPressure(N.getOperand(0)), 0));
4452 RP.push_back(std::make_pair(getRegPressure(N.getOperand(1)), 1));
4453 RP.push_back(std::make_pair(getRegPressure(N.getOperand(2)), 2));
4454 std::sort(RP.begin(), RP.end());
4456 Tmp1 = 0; // Silence a warning.
4457 for (unsigned i = 0; i != 3; ++i)
4458 switch (RP[2-i].second) {
4459 default: assert(0 && "Unknown operand number!");
4460 case 0: Select(N.getOperand(0)); break;
4461 case 1: Tmp1 = SelectExpr(N.getOperand(1)); break;
4462 case 2: SelectAddress(N.getOperand(2), AM); break;
4465 addFullAddress(BuildMI(BB, Opc, 4+1), AM).addReg(Tmp1);
4471 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
4473 switch (CN->getValueType(0)) {
4474 default: assert(0 && "Invalid type for operation!");
4476 case MVT::i8: Opc = X86::MOV8mi; break;
4477 case MVT::i16: Opc = X86::MOV16mi; break;
4478 case MVT::i32: Opc = X86::MOV32mi; break;
4481 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(2))) {
4482 Select(N.getOperand(0));
4483 SelectAddress(N.getOperand(2), AM);
4485 SelectAddress(N.getOperand(2), AM);
4486 Select(N.getOperand(0));
4488 addFullAddress(BuildMI(BB, Opc, 4+1), AM).addImm(CN->getValue());
4491 } else if (GlobalAddressSDNode *GA =
4492 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
4493 assert(GA->getValueType(0) == MVT::i32 && "Bad pointer operand");
4495 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(2))) {
4496 Select(N.getOperand(0));
4497 SelectAddress(N.getOperand(2), AM);
4499 SelectAddress(N.getOperand(2), AM);
4500 Select(N.getOperand(0));
4502 GlobalValue *GV = GA->getGlobal();
4503 // For Darwin, external and weak symbols are indirect, so we want to load
4504 // the value at address GV, not the value of GV itself.
4505 if (Subtarget->getIndirectExternAndWeakGlobals() &&
4506 (GV->hasWeakLinkage() || GV->isExternal())) {
4507 Tmp1 = MakeReg(MVT::i32);
4508 BuildMI(BB, X86::MOV32rm, 4, Tmp1).addReg(0).addZImm(1).addReg(0)
4509 .addGlobalAddress(GV, false, 0);
4510 addFullAddress(BuildMI(BB, X86::MOV32mr, 4+1),AM).addReg(Tmp1);
4512 addFullAddress(BuildMI(BB, X86::MOV32mi, 4+1),AM).addGlobalAddress(GV);
4517 // Check to see if this is a load/op/store combination.
4518 if (TryToFoldLoadOpStore(Node))
4521 switch (N.getOperand(1).getValueType()) {
4522 default: assert(0 && "Cannot store this type!");
4524 case MVT::i8: Opc = X86::MOV8mr; break;
4525 case MVT::i16: Opc = X86::MOV16mr; break;
4526 case MVT::i32: Opc = X86::MOV32mr; break;
4527 case MVT::f32: Opc = X86::MOVSSmr; break;
4528 case MVT::f64: Opc = X86ScalarSSE ? X86::MOVSDmr : X86::FST64m; break;
4531 std::vector<std::pair<unsigned, unsigned> > RP;
4532 RP.push_back(std::make_pair(getRegPressure(N.getOperand(0)), 0));
4533 RP.push_back(std::make_pair(getRegPressure(N.getOperand(1)), 1));
4534 RP.push_back(std::make_pair(getRegPressure(N.getOperand(2)), 2));
4535 std::sort(RP.begin(), RP.end());
4537 Tmp1 = 0; // Silence a warning.
4538 for (unsigned i = 0; i != 3; ++i)
4539 switch (RP[2-i].second) {
4540 default: assert(0 && "Unknown operand number!");
4541 case 0: Select(N.getOperand(0)); break;
4542 case 1: Tmp1 = SelectExpr(N.getOperand(1)); break;
4543 case 2: SelectAddress(N.getOperand(2), AM); break;
4546 addFullAddress(BuildMI(BB, Opc, 4+1), AM).addReg(Tmp1);
4549 case ISD::CALLSEQ_START:
4550 Select(N.getOperand(0));
4552 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
4553 BuildMI(BB, X86::ADJCALLSTACKDOWN, 1).addImm(Tmp1);
4555 case ISD::CALLSEQ_END:
4556 Select(N.getOperand(0));
4559 Select(N.getOperand(0)); // Select the chain.
4561 (unsigned)cast<ConstantSDNode>(Node->getOperand(4))->getValue();
4562 if (Align == 0) Align = 1;
4564 // Turn the byte code into # iterations
4567 if (ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Node->getOperand(2))) {
4568 unsigned Val = ValC->getValue() & 255;
4570 // If the value is a constant, then we can potentially use larger sets.
4571 switch (Align & 3) {
4572 case 2: // WORD aligned
4573 CountReg = MakeReg(MVT::i32);
4574 if (ConstantSDNode *I = dyn_cast<ConstantSDNode>(Node->getOperand(3))) {
4575 BuildMI(BB, X86::MOV32ri, 1, CountReg).addImm(I->getValue()/2);
4577 unsigned ByteReg = SelectExpr(Node->getOperand(3));
4578 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(1);
4580 BuildMI(BB, X86::MOV16ri, 1, X86::AX).addImm((Val << 8) | Val);
4581 Opcode = X86::REP_STOSW;
4583 case 0: // DWORD aligned
4584 CountReg = MakeReg(MVT::i32);
4585 if (ConstantSDNode *I = dyn_cast<ConstantSDNode>(Node->getOperand(3))) {
4586 BuildMI(BB, X86::MOV32ri, 1, CountReg).addImm(I->getValue()/4);
4588 unsigned ByteReg = SelectExpr(Node->getOperand(3));
4589 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(2);
4591 Val = (Val << 8) | Val;
4592 BuildMI(BB, X86::MOV32ri, 1, X86::EAX).addImm((Val << 16) | Val);
4593 Opcode = X86::REP_STOSD;
4595 default: // BYTE aligned
4596 CountReg = SelectExpr(Node->getOperand(3));
4597 BuildMI(BB, X86::MOV8ri, 1, X86::AL).addImm(Val);
4598 Opcode = X86::REP_STOSB;
4602 // If it's not a constant value we are storing, just fall back. We could
4603 // try to be clever to form 16 bit and 32 bit values, but we don't yet.
4604 unsigned ValReg = SelectExpr(Node->getOperand(2));
4605 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(ValReg);
4606 CountReg = SelectExpr(Node->getOperand(3));
4607 Opcode = X86::REP_STOSB;
4610 // No matter what the alignment is, we put the source in ESI, the
4611 // destination in EDI, and the count in ECX.
4612 unsigned TmpReg1 = SelectExpr(Node->getOperand(1));
4613 BuildMI(BB, X86::MOV32rr, 1, X86::ECX).addReg(CountReg);
4614 BuildMI(BB, X86::MOV32rr, 1, X86::EDI).addReg(TmpReg1);
4615 BuildMI(BB, Opcode, 0);
4619 Select(N.getOperand(0)); // Select the chain.
4621 (unsigned)cast<ConstantSDNode>(Node->getOperand(4))->getValue();
4622 if (Align == 0) Align = 1;
4624 // Turn the byte code into # iterations
4627 switch (Align & 3) {
4628 case 2: // WORD aligned
4629 CountReg = MakeReg(MVT::i32);
4630 if (ConstantSDNode *I = dyn_cast<ConstantSDNode>(Node->getOperand(3))) {
4631 BuildMI(BB, X86::MOV32ri, 1, CountReg).addImm(I->getValue()/2);
4633 unsigned ByteReg = SelectExpr(Node->getOperand(3));
4634 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(1);
4636 Opcode = X86::REP_MOVSW;
4638 case 0: // DWORD aligned
4639 CountReg = MakeReg(MVT::i32);
4640 if (ConstantSDNode *I = dyn_cast<ConstantSDNode>(Node->getOperand(3))) {
4641 BuildMI(BB, X86::MOV32ri, 1, CountReg).addImm(I->getValue()/4);
4643 unsigned ByteReg = SelectExpr(Node->getOperand(3));
4644 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(2);
4646 Opcode = X86::REP_MOVSD;
4648 default: // BYTE aligned
4649 CountReg = SelectExpr(Node->getOperand(3));
4650 Opcode = X86::REP_MOVSB;
4654 // No matter what the alignment is, we put the source in ESI, the
4655 // destination in EDI, and the count in ECX.
4656 unsigned TmpReg1 = SelectExpr(Node->getOperand(1));
4657 unsigned TmpReg2 = SelectExpr(Node->getOperand(2));
4658 BuildMI(BB, X86::MOV32rr, 1, X86::ECX).addReg(CountReg);
4659 BuildMI(BB, X86::MOV32rr, 1, X86::EDI).addReg(TmpReg1);
4660 BuildMI(BB, X86::MOV32rr, 1, X86::ESI).addReg(TmpReg2);
4661 BuildMI(BB, Opcode, 0);
4664 case ISD::WRITEPORT:
4665 if (Node->getOperand(2).getValueType() != MVT::i16) {
4666 std::cerr << "llvm.writeport: Address size is not 16 bits\n";
4669 Select(Node->getOperand(0)); // Emit the chain.
4671 Tmp1 = SelectExpr(Node->getOperand(1));
4672 switch (Node->getOperand(1).getValueType()) {
4674 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(Tmp1);
4675 Tmp2 = X86::OUT8ir; Opc = X86::OUT8rr;
4678 BuildMI(BB, X86::MOV16rr, 1, X86::AX).addReg(Tmp1);
4679 Tmp2 = X86::OUT16ir; Opc = X86::OUT16rr;
4682 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Tmp1);
4683 Tmp2 = X86::OUT32ir; Opc = X86::OUT32rr;
4686 std::cerr << "llvm.writeport: invalid data type for X86 target";
4690 // If the port is a single-byte constant, use the immediate form.
4691 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Node->getOperand(2)))
4692 if ((CN->getValue() & 255) == CN->getValue()) {
4693 BuildMI(BB, Tmp2, 1).addImm(CN->getValue());
4697 // Otherwise, move the I/O port address into the DX register.
4698 unsigned Reg = SelectExpr(Node->getOperand(2));
4699 BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(Reg);
4700 BuildMI(BB, Opc, 0);
4703 assert(0 && "Should not be reached!");
4707 /// createX86PatternInstructionSelector - This pass converts an LLVM function
4708 /// into a machine code representation using pattern matching and a machine
4709 /// description file.
4711 FunctionPass *llvm::createX86PatternInstructionSelector(TargetMachine &TM) {
4712 return new ISel(TM);