1 //===-- IA64ISelLowering.cpp - IA64 DAG Lowering Implementation -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Duraid Madina and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the IA64ISelLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "IA64ISelLowering.h"
15 #include "IA64MachineFunctionInfo.h"
16 #include "IA64TargetMachine.h"
17 #include "llvm/CodeGen/MachineFrameInfo.h"
18 #include "llvm/CodeGen/MachineFunction.h"
19 #include "llvm/CodeGen/MachineInstrBuilder.h"
20 #include "llvm/CodeGen/SelectionDAG.h"
21 #include "llvm/CodeGen/SSARegMap.h"
22 #include "llvm/Constants.h"
23 #include "llvm/Function.h"
26 IA64TargetLowering::IA64TargetLowering(TargetMachine &TM)
27 : TargetLowering(TM) {
29 // register class for general registers
30 addRegisterClass(MVT::i64, IA64::GRRegisterClass);
32 // register class for FP registers
33 addRegisterClass(MVT::f64, IA64::FPRegisterClass);
35 // register class for predicate registers
36 addRegisterClass(MVT::i1, IA64::PRRegisterClass);
38 setOperationAction(ISD::BRCONDTWOWAY , MVT::Other, Expand);
39 setOperationAction(ISD::BRTWOWAY_CC , MVT::Other, Expand);
40 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
42 setSetCCResultType(MVT::i1);
43 setShiftAmountType(MVT::i64);
45 setOperationAction(ISD::EXTLOAD , MVT::i1 , Promote);
47 setOperationAction(ISD::ZEXTLOAD , MVT::i1 , Expand);
49 setOperationAction(ISD::SEXTLOAD , MVT::i1 , Expand);
50 setOperationAction(ISD::SEXTLOAD , MVT::i8 , Expand);
51 setOperationAction(ISD::SEXTLOAD , MVT::i16 , Expand);
52 setOperationAction(ISD::SEXTLOAD , MVT::i32 , Expand);
54 setOperationAction(ISD::FREM , MVT::f32 , Expand);
55 setOperationAction(ISD::FREM , MVT::f64 , Expand);
57 setOperationAction(ISD::UREM , MVT::f32 , Expand);
58 setOperationAction(ISD::UREM , MVT::f64 , Expand);
60 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
61 setOperationAction(ISD::MEMSET , MVT::Other, Expand);
62 setOperationAction(ISD::MEMCPY , MVT::Other, Expand);
64 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
65 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
67 // We don't support sin/cos/sqrt
68 setOperationAction(ISD::FSIN , MVT::f64, Expand);
69 setOperationAction(ISD::FCOS , MVT::f64, Expand);
70 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
71 setOperationAction(ISD::FSIN , MVT::f32, Expand);
72 setOperationAction(ISD::FCOS , MVT::f32, Expand);
73 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
75 // We don't have line number support yet.
76 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
77 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
78 setOperationAction(ISD::DEBUG_LABEL, MVT::Other, Expand);
80 //IA64 has these, but they are not implemented
81 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
82 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
83 setOperationAction(ISD::ROTL , MVT::i64 , Expand);
84 setOperationAction(ISD::ROTR , MVT::i64 , Expand);
85 setOperationAction(ISD::BSWAP, MVT::i64 , Expand); // mux @rev
87 // Not implemented yet.
88 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
89 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
91 computeRegisterProperties();
93 addLegalFPImmediate(+0.0);
94 addLegalFPImmediate(+1.0);
97 const char *IA64TargetLowering::getTargetNodeName(unsigned Opcode) const {
100 case IA64ISD::GETFD: return "IA64ISD::GETFD";
101 case IA64ISD::BRCALL: return "IA64ISD::BRCALL";
106 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
107 static bool isFloatingPointZero(SDOperand Op) {
108 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
109 return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
110 else if (Op.getOpcode() == ISD::EXTLOAD || Op.getOpcode() == ISD::LOAD) {
111 // Maybe this has already been legalized into the constant pool?
112 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
113 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->get()))
114 return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
119 std::vector<SDOperand>
120 IA64TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
121 std::vector<SDOperand> ArgValues;
123 // add beautiful description of IA64 stack frame format
124 // here (from intel 24535803.pdf most likely)
126 MachineFunction &MF = DAG.getMachineFunction();
127 MachineFrameInfo *MFI = MF.getFrameInfo();
129 GP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
130 SP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
131 RP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
133 MachineBasicBlock& BB = MF.front();
135 unsigned args_int[] = {IA64::r32, IA64::r33, IA64::r34, IA64::r35,
136 IA64::r36, IA64::r37, IA64::r38, IA64::r39};
138 unsigned args_FP[] = {IA64::F8, IA64::F9, IA64::F10, IA64::F11,
139 IA64::F12,IA64::F13,IA64::F14, IA64::F15};
145 unsigned used_FPArgs = 0; // how many FP args have been used so far?
147 unsigned ArgOffset = 0;
150 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
152 SDOperand newroot, argt;
153 if(count < 8) { // need to fix this logic? maybe.
155 switch (getValueType(I->getType())) {
157 assert(0 && "ERROR in LowerArgs: can't lower this type of arg.\n");
159 // fixme? (well, will need to for weird FP structy stuff,
160 // see intel ABI docs)
162 //XXX BuildMI(&BB, IA64::IDEF, 0, args_FP[used_FPArgs]);
163 MF.addLiveIn(args_FP[used_FPArgs]); // mark this reg as liveIn
164 // floating point args go into f8..f15 as-needed, the increment
165 argVreg[count] = // is below..:
166 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::f64));
167 // FP args go into f8..f15 as needed: (hence the ++)
168 argPreg[count] = args_FP[used_FPArgs++];
169 argOpc[count] = IA64::FMOV;
170 argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), argVreg[count],
172 if (I->getType() == Type::FloatTy)
173 argt = DAG.getNode(ISD::FP_ROUND, MVT::f32, argt);
175 case MVT::i1: // NOTE: as far as C abi stuff goes,
176 // bools are just boring old ints
181 //XXX BuildMI(&BB, IA64::IDEF, 0, args_int[count]);
182 MF.addLiveIn(args_int[count]); // mark this register as liveIn
184 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
185 argPreg[count] = args_int[count];
186 argOpc[count] = IA64::MOV;
188 DAG.getCopyFromReg(DAG.getRoot(), argVreg[count], MVT::i64);
189 if ( getValueType(I->getType()) != MVT::i64)
190 argt = DAG.getNode(ISD::TRUNCATE, getValueType(I->getType()),
194 } else { // more than 8 args go into the frame
195 // Create the frame index object for this incoming parameter...
196 ArgOffset = 16 + 8 * (count - 8);
197 int FI = MFI->CreateFixedObject(8, ArgOffset);
199 // Create the SelectionDAG nodes corresponding to a load
200 //from this parameter
201 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i64);
202 argt = newroot = DAG.getLoad(getValueType(I->getType()),
203 DAG.getEntryNode(), FIN, DAG.getSrcValue(NULL));
206 DAG.setRoot(newroot.getValue(1));
207 ArgValues.push_back(argt);
211 // Create a vreg to hold the output of (what will become)
212 // the "alloc" instruction
213 VirtGPR = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
214 BuildMI(&BB, IA64::PSEUDO_ALLOC, 0, VirtGPR);
215 // we create a PSEUDO_ALLOC (pseudo)instruction for now
217 BuildMI(&BB, IA64::IDEF, 0, IA64::r1);
220 BuildMI(&BB, IA64::IDEF, 0, IA64::r12);
221 BuildMI(&BB, IA64::IDEF, 0, IA64::rp);
224 BuildMI(&BB, IA64::MOV, 1, GP).addReg(IA64::r1);
227 BuildMI(&BB, IA64::MOV, 1, SP).addReg(IA64::r12);
228 BuildMI(&BB, IA64::MOV, 1, RP).addReg(IA64::rp);
232 unsigned tempOffset=0;
234 // if this is a varargs function, we simply lower llvm.va_start by
235 // pointing to the first entry
238 VarArgsFrameIndex = MFI->CreateFixedObject(8, tempOffset);
241 // here we actually do the moving of args, and store them to the stack
242 // too if this is a varargs function:
243 for (int i = 0; i < count && i < 8; ++i) {
244 BuildMI(&BB, argOpc[i], 1, argVreg[i]).addReg(argPreg[i]);
246 // if this is a varargs function, we copy the input registers to the stack
247 int FI = MFI->CreateFixedObject(8, tempOffset);
248 tempOffset+=8; //XXX: is it safe to use r22 like this?
249 BuildMI(&BB, IA64::MOV, 1, IA64::r22).addFrameIndex(FI);
250 // FIXME: we should use st8.spill here, one day
251 BuildMI(&BB, IA64::ST8, 1, IA64::r22).addReg(argPreg[i]);
255 // Finally, inform the code generator which regs we return values in.
256 // (see the ISD::RET: case in the instruction selector)
257 switch (getValueType(F.getReturnType())) {
258 default: assert(0 && "i have no idea where to return this type!");
259 case MVT::isVoid: break;
265 MF.addLiveOut(IA64::r8);
269 MF.addLiveOut(IA64::F8);
276 std::pair<SDOperand, SDOperand>
277 IA64TargetLowering::LowerCallTo(SDOperand Chain,
278 const Type *RetTy, bool isVarArg,
279 unsigned CallingConv, bool isTailCall,
280 SDOperand Callee, ArgListTy &Args,
283 MachineFunction &MF = DAG.getMachineFunction();
285 unsigned NumBytes = 16;
286 unsigned outRegsUsed = 0;
288 if (Args.size() > 8) {
289 NumBytes += (Args.size() - 8) * 8;
292 outRegsUsed = Args.size();
295 // FIXME? this WILL fail if we ever try to pass around an arg that
296 // consumes more than a single output slot (a 'real' double, int128
297 // some sort of aggregate etc.), as we'll underestimate how many 'outX'
298 // registers we use. Hopefully, the assembler will notice.
299 MF.getInfo<IA64FunctionInfo>()->outRegsUsed=
300 std::max(outRegsUsed, MF.getInfo<IA64FunctionInfo>()->outRegsUsed);
302 // keep stack frame 16-byte aligned
303 //assert(NumBytes==((NumBytes+15) & ~15) && "stack frame not 16-byte aligned!");
304 NumBytes = (NumBytes+15) & ~15;
306 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
307 DAG.getConstant(NumBytes, getPointerTy()));
309 SDOperand StackPtr, NullSV;
310 std::vector<SDOperand> Stores;
311 std::vector<SDOperand> Converts;
312 std::vector<SDOperand> RegValuesToPass;
313 unsigned ArgOffset = 16;
315 for (unsigned i = 0, e = Args.size(); i != e; ++i)
317 SDOperand Val = Args[i].first;
318 MVT::ValueType ObjectVT = Val.getValueType();
319 SDOperand ValToStore(0, 0), ValToConvert(0, 0);
322 default: assert(0 && "unexpected argument type!");
327 //promote to 64-bits, sign/zero extending based on type
329 if(Args[i].second->isSigned())
330 Val = DAG.getNode(ISD::SIGN_EXTEND, MVT::i64, Val);
332 Val = DAG.getNode(ISD::ZERO_EXTEND, MVT::i64, Val);
336 if(RegValuesToPass.size() >= 8) {
339 RegValuesToPass.push_back(Val);
344 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
347 if(RegValuesToPass.size() >= 8) {
350 RegValuesToPass.push_back(Val);
351 if(1 /* TODO: if(calling external or varadic function)*/ ) {
352 ValToConvert = Val; // additionally pass this FP value as an int
360 StackPtr = DAG.getRegister(IA64::r12, MVT::i64);
361 NullSV = DAG.getSrcValue(NULL);
363 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
364 PtrOff = DAG.getNode(ISD::ADD, MVT::i64, StackPtr, PtrOff);
365 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
366 ValToStore, PtrOff, NullSV));
367 ArgOffset += ObjSize;
370 if(ValToConvert.Val) {
371 Converts.push_back(DAG.getNode(IA64ISD::GETFD, MVT::i64, ValToConvert));
375 // Emit all stores, make sure they occur before any copies into physregs.
377 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
379 static const unsigned IntArgRegs[] = {
380 IA64::out0, IA64::out1, IA64::out2, IA64::out3,
381 IA64::out4, IA64::out5, IA64::out6, IA64::out7
384 static const unsigned FPArgRegs[] = {
385 IA64::F8, IA64::F9, IA64::F10, IA64::F11,
386 IA64::F12, IA64::F13, IA64::F14, IA64::F15
391 // save the current GP, SP and RP : FIXME: do we need to do all 3 always?
392 SDOperand GPBeforeCall = DAG.getCopyFromReg(Chain, IA64::r1, MVT::i64, InFlag);
393 Chain = GPBeforeCall.getValue(1);
394 InFlag = Chain.getValue(2);
395 SDOperand SPBeforeCall = DAG.getCopyFromReg(Chain, IA64::r12, MVT::i64, InFlag);
396 Chain = SPBeforeCall.getValue(1);
397 InFlag = Chain.getValue(2);
398 SDOperand RPBeforeCall = DAG.getCopyFromReg(Chain, IA64::rp, MVT::i64, InFlag);
399 Chain = RPBeforeCall.getValue(1);
400 InFlag = Chain.getValue(2);
402 // Build a sequence of copy-to-reg nodes chained together with token chain
403 // and flag operands which copy the outgoing integer args into regs out[0-7]
404 // mapped 1:1 and the FP args into regs F8-F15 "lazily"
405 // TODO: for performance, we should only copy FP args into int regs when we
406 // know this is required (i.e. for varardic or external (unknown) functions)
408 // first to the FP->(integer representation) conversions, these are
409 // flagged for now, but shouldn't have to be (TODO)
410 unsigned seenConverts = 0;
411 for (unsigned i = 0, e = RegValuesToPass.size(); i != e; ++i) {
412 if(MVT::isFloatingPoint(RegValuesToPass[i].getValueType())) {
413 Chain = DAG.getCopyToReg(Chain, IntArgRegs[i], Converts[seenConverts++], InFlag);
414 InFlag = Chain.getValue(1);
418 // next copy args into the usual places, these are flagged
419 unsigned usedFPArgs = 0;
420 for (unsigned i = 0, e = RegValuesToPass.size(); i != e; ++i) {
421 Chain = DAG.getCopyToReg(Chain,
422 MVT::isInteger(RegValuesToPass[i].getValueType()) ?
423 IntArgRegs[i] : FPArgRegs[usedFPArgs++],
424 RegValuesToPass[i], InFlag);
425 InFlag = Chain.getValue(1);
428 // If the callee is a GlobalAddress node (quite common, every direct call is)
429 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
431 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
432 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), MVT::i64);
436 std::vector<MVT::ValueType> NodeTys;
437 std::vector<SDOperand> CallOperands;
438 NodeTys.push_back(MVT::Other); // Returns a chain
439 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
440 CallOperands.push_back(Chain);
441 CallOperands.push_back(Callee);
443 // emit the call itself
445 CallOperands.push_back(InFlag);
447 assert(0 && "this should never happen!\n");
449 /* out with the old...
450 Chain = SDOperand(DAG.getCall(NodeTys, Chain, Callee, InFlag), 0);
452 Chain = SDOperand(DAG.getCall(NodeTys, Chain, Callee), 0);
454 // to make way for a hack:
455 Chain = DAG.getNode(IA64ISD::BRCALL, NodeTys, CallOperands);
456 InFlag = Chain.getValue(1);
458 // restore the GP, SP and RP after the call
459 Chain = DAG.getCopyToReg(Chain, IA64::r1, GPBeforeCall, InFlag);
460 InFlag = Chain.getValue(1);
461 Chain = DAG.getCopyToReg(Chain, IA64::r12, SPBeforeCall, InFlag);
462 InFlag = Chain.getValue(1);
463 Chain = DAG.getCopyToReg(Chain, IA64::rp, RPBeforeCall, InFlag);
464 InFlag = Chain.getValue(1);
466 std::vector<MVT::ValueType> RetVals;
467 RetVals.push_back(MVT::Other);
468 RetVals.push_back(MVT::Flag);
470 MVT::ValueType RetTyVT = getValueType(RetTy);
472 if (RetTyVT != MVT::isVoid) {
474 default: // assert(0 && "Unknown value type to return!");
475 case MVT::i1: { // bools are just like other integers (returned in r8)
476 SDOperand boolInR8 = DAG.getCopyFromReg(Chain, IA64::r8, MVT::i64, InFlag);
477 InFlag = boolInR8.getValue(2);
478 Chain = boolInR8.getValue(1);
479 SDOperand zeroReg = DAG.getCopyFromReg(Chain, IA64::r0, MVT::i64, InFlag);
480 InFlag = zeroReg.getValue(2);
481 Chain = zeroReg.getValue(1);
483 RetVal = DAG.getSetCC(MVT::i1, boolInR8, zeroReg, ISD::SETNE);
489 RetVal = DAG.getCopyFromReg(Chain, IA64::r8, MVT::i64, InFlag);
490 Chain = RetVal.getValue(1);
492 // Add a note to keep track of whether it is sign or zero extended - TODO: bools
493 RetVal = DAG.getNode(RetTy->isSigned() ? ISD::AssertSext :ISD::AssertZext,
494 MVT::i64, RetVal, DAG.getValueType(RetTyVT));
495 RetVal = DAG.getNode(ISD::TRUNCATE, RetTyVT, RetVal);
498 RetVal = DAG.getCopyFromReg(Chain, IA64::r8, MVT::i64, InFlag);
499 Chain = RetVal.getValue(1);
500 InFlag = RetVal.getValue(2); // XXX dead
503 RetVal = DAG.getCopyFromReg(Chain, IA64::F8, MVT::f64, InFlag);
504 Chain = RetVal.getValue(1);
505 InFlag = RetVal.getValue(2); // XXX dead
510 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
511 DAG.getConstant(NumBytes, getPointerTy()));
513 return std::make_pair(RetVal, Chain);
517 IA64TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
518 Value *VAListV, SelectionDAG &DAG) {
519 // vastart just stores the address of the VarArgsFrameIndex slot.
520 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i64);
521 return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR,
522 VAListP, DAG.getSrcValue(VAListV));
525 std::pair<SDOperand,SDOperand> IA64TargetLowering::
526 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
527 const Type *ArgTy, SelectionDAG &DAG) {
529 MVT::ValueType ArgVT = getValueType(ArgTy);
530 SDOperand Val = DAG.getLoad(MVT::i64, Chain,
531 VAListP, DAG.getSrcValue(VAListV));
532 SDOperand Result = DAG.getLoad(ArgVT, DAG.getEntryNode(), Val,
533 DAG.getSrcValue(NULL));
535 if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
538 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
539 "Other types should have been promoted for varargs!");
542 Val = DAG.getNode(ISD::ADD, Val.getValueType(), Val,
543 DAG.getConstant(Amt, Val.getValueType()));
544 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
545 Val, VAListP, DAG.getSrcValue(VAListV));
546 return std::make_pair(Result, Chain);
551 std::pair<SDOperand, SDOperand> IA64TargetLowering::
552 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
554 assert(0 && "LowerFrameReturnAddress unimplemented");