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);
86 computeRegisterProperties();
88 addLegalFPImmediate(+0.0);
89 addLegalFPImmediate(+1.0);
92 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
93 static bool isFloatingPointZero(SDOperand Op) {
94 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
95 return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
96 else if (Op.getOpcode() == ISD::EXTLOAD || Op.getOpcode() == ISD::LOAD) {
97 // Maybe this has already been legalized into the constant pool?
98 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
99 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->get()))
100 return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
105 std::vector<SDOperand>
106 IA64TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
107 std::vector<SDOperand> ArgValues;
109 // add beautiful description of IA64 stack frame format
110 // here (from intel 24535803.pdf most likely)
112 MachineFunction &MF = DAG.getMachineFunction();
113 MachineFrameInfo *MFI = MF.getFrameInfo();
115 GP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
116 SP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
117 RP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
119 MachineBasicBlock& BB = MF.front();
121 unsigned args_int[] = {IA64::r32, IA64::r33, IA64::r34, IA64::r35,
122 IA64::r36, IA64::r37, IA64::r38, IA64::r39};
124 unsigned args_FP[] = {IA64::F8, IA64::F9, IA64::F10, IA64::F11,
125 IA64::F12,IA64::F13,IA64::F14, IA64::F15};
131 unsigned used_FPArgs = 0; // how many FP args have been used so far?
133 unsigned ArgOffset = 0;
136 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
138 SDOperand newroot, argt;
139 if(count < 8) { // need to fix this logic? maybe.
141 switch (getValueType(I->getType())) {
143 assert(0 && "ERROR in LowerArgs: can't lower this type of arg.\n");
145 // fixme? (well, will need to for weird FP structy stuff,
146 // see intel ABI docs)
148 //XXX BuildMI(&BB, IA64::IDEF, 0, args_FP[used_FPArgs]);
149 MF.addLiveIn(args_FP[used_FPArgs]); // mark this reg as liveIn
150 // floating point args go into f8..f15 as-needed, the increment
151 argVreg[count] = // is below..:
152 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::f64));
153 // FP args go into f8..f15 as needed: (hence the ++)
154 argPreg[count] = args_FP[used_FPArgs++];
155 argOpc[count] = IA64::FMOV;
156 argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), argVreg[count],
158 if (I->getType() == Type::FloatTy)
159 argt = DAG.getNode(ISD::FP_ROUND, MVT::f32, argt);
161 case MVT::i1: // NOTE: as far as C abi stuff goes,
162 // bools are just boring old ints
167 //XXX BuildMI(&BB, IA64::IDEF, 0, args_int[count]);
168 MF.addLiveIn(args_int[count]); // mark this register as liveIn
170 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
171 argPreg[count] = args_int[count];
172 argOpc[count] = IA64::MOV;
174 DAG.getCopyFromReg(DAG.getRoot(), argVreg[count], MVT::i64);
175 if ( getValueType(I->getType()) != MVT::i64)
176 argt = DAG.getNode(ISD::TRUNCATE, getValueType(I->getType()),
180 } else { // more than 8 args go into the frame
181 // Create the frame index object for this incoming parameter...
182 ArgOffset = 16 + 8 * (count - 8);
183 int FI = MFI->CreateFixedObject(8, ArgOffset);
185 // Create the SelectionDAG nodes corresponding to a load
186 //from this parameter
187 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i64);
188 argt = newroot = DAG.getLoad(getValueType(I->getType()),
189 DAG.getEntryNode(), FIN, DAG.getSrcValue(NULL));
192 DAG.setRoot(newroot.getValue(1));
193 ArgValues.push_back(argt);
197 // Create a vreg to hold the output of (what will become)
198 // the "alloc" instruction
199 VirtGPR = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
200 BuildMI(&BB, IA64::PSEUDO_ALLOC, 0, VirtGPR);
201 // we create a PSEUDO_ALLOC (pseudo)instruction for now
203 BuildMI(&BB, IA64::IDEF, 0, IA64::r1);
206 BuildMI(&BB, IA64::IDEF, 0, IA64::r12);
207 BuildMI(&BB, IA64::IDEF, 0, IA64::rp);
210 BuildMI(&BB, IA64::MOV, 1, GP).addReg(IA64::r1);
213 BuildMI(&BB, IA64::MOV, 1, SP).addReg(IA64::r12);
214 BuildMI(&BB, IA64::MOV, 1, RP).addReg(IA64::rp);
218 unsigned tempOffset=0;
220 // if this is a varargs function, we simply lower llvm.va_start by
221 // pointing to the first entry
224 VarArgsFrameIndex = MFI->CreateFixedObject(8, tempOffset);
227 // here we actually do the moving of args, and store them to the stack
228 // too if this is a varargs function:
229 for (int i = 0; i < count && i < 8; ++i) {
230 BuildMI(&BB, argOpc[i], 1, argVreg[i]).addReg(argPreg[i]);
232 // if this is a varargs function, we copy the input registers to the stack
233 int FI = MFI->CreateFixedObject(8, tempOffset);
234 tempOffset+=8; //XXX: is it safe to use r22 like this?
235 BuildMI(&BB, IA64::MOV, 1, IA64::r22).addFrameIndex(FI);
236 // FIXME: we should use st8.spill here, one day
237 BuildMI(&BB, IA64::ST8, 1, IA64::r22).addReg(argPreg[i]);
241 // Finally, inform the code generator which regs we return values in.
242 // (see the ISD::RET: case in the instruction selector)
243 switch (getValueType(F.getReturnType())) {
244 default: assert(0 && "i have no idea where to return this type!");
245 case MVT::isVoid: break;
251 MF.addLiveOut(IA64::r8);
255 MF.addLiveOut(IA64::F8);
262 std::pair<SDOperand, SDOperand>
263 IA64TargetLowering::LowerCallTo(SDOperand Chain,
264 const Type *RetTy, bool isVarArg,
265 unsigned CallingConv, bool isTailCall,
266 SDOperand Callee, ArgListTy &Args,
269 MachineFunction &MF = DAG.getMachineFunction();
271 unsigned NumBytes = 16;
272 unsigned outRegsUsed = 0;
274 if (Args.size() > 8) {
275 NumBytes += (Args.size() - 8) * 8;
278 outRegsUsed = Args.size();
281 // FIXME? this WILL fail if we ever try to pass around an arg that
282 // consumes more than a single output slot (a 'real' double, int128
283 // some sort of aggregate etc.), as we'll underestimate how many 'outX'
284 // registers we use. Hopefully, the assembler will notice.
285 MF.getInfo<IA64FunctionInfo>()->outRegsUsed=
286 std::max(outRegsUsed, MF.getInfo<IA64FunctionInfo>()->outRegsUsed);
288 // keep stack frame 16-byte aligned
289 //assert(NumBytes==((NumBytes+15) & ~15) && "stack frame not 16-byte aligned!");
290 NumBytes = (NumBytes+15) & ~15;
292 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
293 DAG.getConstant(NumBytes, getPointerTy()));
295 SDOperand StackPtr, NullSV;
296 std::vector<SDOperand> Stores;
297 std::vector<SDOperand> Converts;
298 std::vector<SDOperand> RegValuesToPass;
299 unsigned ArgOffset = 16;
301 for (unsigned i = 0, e = Args.size(); i != e; ++i)
303 SDOperand Val = Args[i].first;
304 MVT::ValueType ObjectVT = Val.getValueType();
305 SDOperand ValToStore(0, 0), ValToConvert(0, 0);
308 default: assert(0 && "unexpected argument type!");
313 //promote to 64-bits, sign/zero extending based on type
315 if(Args[i].second->isSigned())
316 Val = DAG.getNode(ISD::SIGN_EXTEND, MVT::i64, Val);
318 Val = DAG.getNode(ISD::ZERO_EXTEND, MVT::i64, Val);
322 if(RegValuesToPass.size() >= 8) {
325 RegValuesToPass.push_back(Val);
330 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
333 if(RegValuesToPass.size() >= 8) {
336 RegValuesToPass.push_back(Val);
337 if(1 /* TODO: if(calling external or varadic function)*/ ) {
338 ValToConvert = Val; // additionally pass this FP value as an int
346 StackPtr = DAG.getRegister(IA64::r12, MVT::i64);
347 NullSV = DAG.getSrcValue(NULL);
349 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
350 PtrOff = DAG.getNode(ISD::ADD, MVT::i64, StackPtr, PtrOff);
351 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
352 ValToStore, PtrOff, NullSV));
353 ArgOffset += ObjSize;
356 if(ValToConvert.Val) {
357 Converts.push_back(DAG.getNode(IA64ISD::GETFD, MVT::i64, ValToConvert));
361 // Emit all stores, make sure they occur before any copies into physregs.
363 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
365 static const unsigned IntArgRegs[] = {
366 IA64::out0, IA64::out1, IA64::out2, IA64::out3,
367 IA64::out4, IA64::out5, IA64::out6, IA64::out7
370 static const unsigned FPArgRegs[] = {
371 IA64::F8, IA64::F9, IA64::F10, IA64::F11,
372 IA64::F12, IA64::F13, IA64::F14, IA64::F15
377 // save the current GP, SP and RP : FIXME: do we need to do all 3 always?
378 SDOperand GPBeforeCall = DAG.getCopyFromReg(Chain, IA64::r1, MVT::i64, InFlag);
379 Chain = GPBeforeCall.getValue(1);
380 InFlag = Chain.getValue(2);
381 SDOperand SPBeforeCall = DAG.getCopyFromReg(Chain, IA64::r12, MVT::i64, InFlag);
382 Chain = SPBeforeCall.getValue(1);
383 InFlag = Chain.getValue(2);
384 SDOperand RPBeforeCall = DAG.getCopyFromReg(Chain, IA64::rp, MVT::i64, InFlag);
385 Chain = RPBeforeCall.getValue(1);
386 InFlag = Chain.getValue(2);
388 // Build a sequence of copy-to-reg nodes chained together with token chain
389 // and flag operands which copy the outgoing integer args into regs out[0-7]
390 // mapped 1:1 and the FP args into regs F8-F15 "lazily"
391 // TODO: for performance, we should only copy FP args into int regs when we
392 // know this is required (i.e. for varardic or external (unknown) functions)
394 // first to the FP->(integer representation) conversions, these are
395 // flagged for now, but shouldn't have to be (TODO)
396 unsigned seenConverts = 0;
397 for (unsigned i = 0, e = RegValuesToPass.size(); i != e; ++i) {
398 if(MVT::isFloatingPoint(RegValuesToPass[i].getValueType())) {
399 Chain = DAG.getCopyToReg(Chain, IntArgRegs[i], Converts[seenConverts++], InFlag);
400 InFlag = Chain.getValue(1);
404 // next copy args into the usual places, these are flagged
405 unsigned usedFPArgs = 0;
406 for (unsigned i = 0, e = RegValuesToPass.size(); i != e; ++i) {
407 Chain = DAG.getCopyToReg(Chain,
408 MVT::isInteger(RegValuesToPass[i].getValueType()) ?
409 IntArgRegs[i] : FPArgRegs[usedFPArgs++],
410 RegValuesToPass[i], InFlag);
411 InFlag = Chain.getValue(1);
414 // If the callee is a GlobalAddress node (quite common, every direct call is)
415 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
417 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
418 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), MVT::i64);
422 std::vector<MVT::ValueType> NodeTys;
423 std::vector<SDOperand> CallOperands;
424 NodeTys.push_back(MVT::Other); // Returns a chain
425 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
426 CallOperands.push_back(Chain);
427 CallOperands.push_back(Callee);
429 // emit the call itself
431 CallOperands.push_back(InFlag);
433 assert(0 && "this should never happen!\n");
435 /* out with the old...
436 Chain = SDOperand(DAG.getCall(NodeTys, Chain, Callee, InFlag), 0);
438 Chain = SDOperand(DAG.getCall(NodeTys, Chain, Callee), 0);
440 // to make way for a hack:
441 Chain = DAG.getNode(IA64ISD::BRCALL, NodeTys, CallOperands);
442 InFlag = Chain.getValue(1);
444 // restore the GP, SP and RP after the call
445 Chain = DAG.getCopyToReg(Chain, IA64::r1, GPBeforeCall, InFlag);
446 InFlag = Chain.getValue(1);
447 Chain = DAG.getCopyToReg(Chain, IA64::r12, SPBeforeCall, InFlag);
448 InFlag = Chain.getValue(1);
449 Chain = DAG.getCopyToReg(Chain, IA64::rp, RPBeforeCall, InFlag);
450 InFlag = Chain.getValue(1);
452 std::vector<MVT::ValueType> RetVals;
453 RetVals.push_back(MVT::Other);
454 RetVals.push_back(MVT::Flag);
456 MVT::ValueType RetTyVT = getValueType(RetTy);
458 if (RetTyVT != MVT::isVoid) {
460 default: // assert(0 && "Unknown value type to return!");
461 case MVT::i1: { // bools are just like other integers (returned in r8)
462 SDOperand boolInR8 = DAG.getCopyFromReg(Chain, IA64::r8, MVT::i64, InFlag);
463 InFlag = boolInR8.getValue(2);
464 Chain = boolInR8.getValue(1);
465 SDOperand zeroReg = DAG.getCopyFromReg(Chain, IA64::r0, MVT::i64, InFlag);
466 InFlag = zeroReg.getValue(2);
467 Chain = zeroReg.getValue(1);
469 RetVal = DAG.getSetCC(MVT::i1, boolInR8, zeroReg, ISD::SETNE);
475 RetVal = DAG.getCopyFromReg(Chain, IA64::r8, MVT::i64, InFlag);
476 Chain = RetVal.getValue(1);
478 // Add a note to keep track of whether it is sign or zero extended - TODO: bools
479 RetVal = DAG.getNode(RetTy->isSigned() ? ISD::AssertSext :ISD::AssertZext,
480 MVT::i64, RetVal, DAG.getValueType(RetTyVT));
481 RetVal = DAG.getNode(ISD::TRUNCATE, RetTyVT, RetVal);
484 RetVal = DAG.getCopyFromReg(Chain, IA64::r8, MVT::i64, InFlag);
485 Chain = RetVal.getValue(1);
486 InFlag = RetVal.getValue(2); // XXX dead
489 RetVal = DAG.getCopyFromReg(Chain, IA64::F8, MVT::f64, InFlag);
490 Chain = RetVal.getValue(1);
491 InFlag = RetVal.getValue(2); // XXX dead
496 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
497 DAG.getConstant(NumBytes, getPointerTy()));
499 return std::make_pair(RetVal, Chain);
503 IA64TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
504 Value *VAListV, SelectionDAG &DAG) {
505 // vastart just stores the address of the VarArgsFrameIndex slot.
506 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i64);
507 return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR,
508 VAListP, DAG.getSrcValue(VAListV));
511 std::pair<SDOperand,SDOperand> IA64TargetLowering::
512 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
513 const Type *ArgTy, SelectionDAG &DAG) {
515 MVT::ValueType ArgVT = getValueType(ArgTy);
516 SDOperand Val = DAG.getLoad(MVT::i64, Chain,
517 VAListP, DAG.getSrcValue(VAListV));
518 SDOperand Result = DAG.getLoad(ArgVT, DAG.getEntryNode(), Val,
519 DAG.getSrcValue(NULL));
521 if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
524 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
525 "Other types should have been promoted for varargs!");
528 Val = DAG.getNode(ISD::ADD, Val.getValueType(), Val,
529 DAG.getConstant(Amt, Val.getValueType()));
530 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
531 Val, VAListP, DAG.getSrcValue(VAListV));
532 return std::make_pair(Result, Chain);
537 std::pair<SDOperand, SDOperand> IA64TargetLowering::
538 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
540 assert(0 && "LowerFrameReturnAddress unimplemented");